maxjhandsome
/
llvm-project
forked from OSchip/llvm-project
1
0
Fork 0
Code Issues Pull Requests Packages Releases Wiki Activity
llvm-project/llvm/utils/TableGen/GlobalISelEmitter.cpp

4731 lines
173 KiB
C++
Raw Normal View History

[GlobalISel] Add basic Selector-emitter tblgen backend. This adds a basic tablegen backend that analyzes the SelectionDAG patterns to find simple ones that are eligible for GlobalISel-emission. That's similar to FastISel, with one notable difference: we're not fed ISD opcodes, so we need to map the SDNode operators to generic opcodes. That's done using GINodeEquiv in TargetGlobalISel.td. Otherwise, this is mostly boilerplate, and lots of filtering of any kind of "complicated" pattern. On AArch64, this is sufficient to match G_ADD up to s64 (to ADDWrr/ADDXrr) and G_BR (to B). Differential Revision: https://reviews.llvm.org/D26878 llvm-svn: 290284
2016-12-22 07:26:20 +08:00
//===- GlobalISelEmitter.cpp - Generate an instruction selector -----------===//
//
// The LLVM Compiler Infrastructure
//
// This file is distributed under the University of Illinois Open Source
// License. See LICENSE.TXT for details.
//
//===----------------------------------------------------------------------===//
//
/// \file
/// This tablegen backend emits code for use by the GlobalISel instruction
/// selector. See include/llvm/CodeGen/TargetGlobalISel.td.
///
/// This file analyzes the patterns recognized by the SelectionDAGISel tablegen
/// backend, filters out the ones that are unsupported, maps
/// SelectionDAG-specific constructs to their GlobalISel counterpart
/// (when applicable: MVT to LLT; SDNode to generic Instruction).
///
/// Not all patterns are supported: pass the tablegen invocation
/// "-warn-on-skipped-patterns" to emit a warning when a pattern is skipped,
/// as well as why.
///
/// The generated file defines a single method:
/// bool <Target>InstructionSelector::selectImpl(MachineInstr &I) const;
/// intended to be used in InstructionSelector::select as the first-step
/// selector for the patterns that don't require complex C++.
///
/// FIXME: We'll probably want to eventually define a base
/// "TargetGenInstructionSelector" class.
///
//===----------------------------------------------------------------------===//
#include "CodeGenDAGPatterns.h"
[globalisel][tablegen] Import SelectionDAG's rule predicates and support the equivalent in GIRule. Summary: The SelectionDAG importer now imports rules with Predicate's attached via Requires, PredicateControl, etc. These predicates are implemented as bitset's to allow multiple predicates to be tested together. However, unlike the MC layer subtarget features, each target only pays for it's own predicates (e.g. AArch64 doesn't have 192 feature bits just because X86 needs a lot). Both AArch64 and X86 derive at least one predicate from the MachineFunction or Function so they must re-initialize AvailableFeatures before each function. They also declare locals in <Target>InstructionSelector so that computeAvailableFeatures() can use the code from SelectionDAG without modification. Reviewers: rovka, qcolombet, aditya_nandakumar, t.p.northover, ab Reviewed By: rovka Subscribers: aemerson, rengolin, dberris, kristof.beyls, llvm-commits, igorb Differential Revision: https://reviews.llvm.org/D31418 llvm-svn: 300993
2017-04-21 23:59:56 +08:00
#include "SubtargetFeatureInfo.h"
[GlobalISel] Add basic Selector-emitter tblgen backend. This adds a basic tablegen backend that analyzes the SelectionDAG patterns to find simple ones that are eligible for GlobalISel-emission. That's similar to FastISel, with one notable difference: we're not fed ISD opcodes, so we need to map the SDNode operators to generic opcodes. That's done using GINodeEquiv in TargetGlobalISel.td. Otherwise, this is mostly boilerplate, and lots of filtering of any kind of "complicated" pattern. On AArch64, this is sufficient to match G_ADD up to s64 (to ADDWrr/ADDXrr) and G_BR (to B). Differential Revision: https://reviews.llvm.org/D26878 llvm-svn: 290284
2016-12-22 07:26:20 +08:00
#include "llvm/ADT/Optional.h"
[globalisel][tablegen] Add experimental support for OperandWithDefaultOps, PredicateOperand, and OptionalDefOperand Summary: As far as instruction selection is concerned, all three appear to be same thing. Support for these operands is experimental since AArch64 doesn't make use of them and the in-tree targets that do use them (AMDGPU for OperandWithDefaultOps, AMDGPU/ARM/Hexagon/Lanai for PredicateOperand, and ARM for OperandWithDefaultOps) are not using tablegen-erated GlobalISel yet. Reviewers: rovka, aditya_nandakumar, t.p.northover, qcolombet, ab Reviewed By: rovka Subscribers: inglorion, aemerson, rengolin, mehdi_amini, dberris, kristof.beyls, igorb, tpr, llvm-commits Differential Revision: https://reviews.llvm.org/D31135 llvm-svn: 300037
2017-04-12 16:23:08 +08:00
#include "llvm/ADT/SmallSet.h"
[GlobalISel] Add basic Selector-emitter tblgen backend. This adds a basic tablegen backend that analyzes the SelectionDAG patterns to find simple ones that are eligible for GlobalISel-emission. That's similar to FastISel, with one notable difference: we're not fed ISD opcodes, so we need to map the SDNode operators to generic opcodes. That's done using GINodeEquiv in TargetGlobalISel.td. Otherwise, this is mostly boilerplate, and lots of filtering of any kind of "complicated" pattern. On AArch64, this is sufficient to match G_ADD up to s64 (to ADDWrr/ADDXrr) and G_BR (to B). Differential Revision: https://reviews.llvm.org/D26878 llvm-svn: 290284
2016-12-22 07:26:20 +08:00
#include "llvm/ADT/Statistic.h"
[globalisel][tablegen] Generate rule coverage and use it to identify untested rules Summary: This patch adds a LLVM_ENABLE_GISEL_COV which, like LLVM_ENABLE_DAGISEL_COV, causes TableGen to instrument the generated table to collect rule coverage information. However, LLVM_ENABLE_GISEL_COV goes a bit further than LLVM_ENABLE_DAGISEL_COV. The information is written to files (${CMAKE_BINARY_DIR}/gisel-coverage-* by default). These files can then be concatenated into ${LLVM_GISEL_COV_PREFIX}-all after which TableGen will read this information and use it to emit warnings about untested rules. This technique could also be used by SelectionDAG and can be further extended to detect hot rules and give them priority over colder rules. Usage: * Enable LLVM_ENABLE_GISEL_COV in CMake * Build the compiler and run some tests * cat gisel-coverage-[0-9]* > gisel-coverage-all * Delete lib/Target/*/*GenGlobalISel.inc* * Build the compiler Known issues: * ${LLVM_GISEL_COV_PREFIX}-all must be generated as a manual step due to a lack of a portable 'cat' command. It should be the concatenation of all ${LLVM_GISEL_COV_PREFIX}-[0-9]* files. * There's no mechanism to discard coverage information when the ruleset changes Depends on D39742 Reviewers: ab, qcolombet, t.p.northover, aditya_nandakumar, rovka Reviewed By: rovka Subscribers: vsk, arsenm, nhaehnle, mgorny, kristof.beyls, javed.absar, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D39747 llvm-svn: 318356
2017-11-16 08:46:35 +08:00
#include "llvm/Support/CodeGenCoverage.h"
[GlobalISel] Add basic Selector-emitter tblgen backend. This adds a basic tablegen backend that analyzes the SelectionDAG patterns to find simple ones that are eligible for GlobalISel-emission. That's similar to FastISel, with one notable difference: we're not fed ISD opcodes, so we need to map the SDNode operators to generic opcodes. That's done using GINodeEquiv in TargetGlobalISel.td. Otherwise, this is mostly boilerplate, and lots of filtering of any kind of "complicated" pattern. On AArch64, this is sufficient to match G_ADD up to s64 (to ADDWrr/ADDXrr) and G_BR (to B). Differential Revision: https://reviews.llvm.org/D26878 llvm-svn: 290284
2016-12-22 07:26:20 +08:00
#include "llvm/Support/CommandLine.h"
[GlobalISel] Return an Expected<RuleMatcher> for each SDAG pattern. NFC. Instead of emitting the matcher code directly, return the rule matcher and the skip reason as an Expected<RuleMatcher>. This will let us record all matchers and process them before emission. It's a somewhat unconventional use of Error, but it's nicer than, say, std::pair, because of the bool conversions. Differential Revision: https://reviews.llvm.org/D29743 llvm-svn: 294706
2017-02-10 12:00:17 +08:00
#include "llvm/Support/Error.h"
Recommit: [globalisel] Change LLT constructor string into an LLT-based object that knows how to generate it. Summary: This will allow future patches to inspect the details of the LLT. The implementation is now split between the Support and CodeGen libraries to allow TableGen to use this class without introducing layering concerns. Thanks to Ahmed Bougacha for finding a reasonable way to avoid the layering issue and providing the version of this patch without that problem. The problem with the previous commit appears to have been that TableGen was including CodeGen/LowLevelType.h instead of Support/LowLevelTypeImpl.h. Reviewers: t.p.northover, qcolombet, rovka, aditya_nandakumar, ab, javed.absar Subscribers: arsenm, nhaehnle, mgorny, dberris, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D30046 llvm-svn: 297241
2017-03-08 07:20:35 +08:00
#include "llvm/Support/LowLevelTypeImpl.h"
Fix layering of MachineValueType.h by moving it from CodeGen to Support This is used by llvm tblgen as well as by LLVM Targets, so the only common place is Support for now. (maybe we need another target for these sorts of things - but for now I'm at least making them correct & we can make them better if/when people have strong feelings) llvm-svn: 328395
2018-03-24 07:58:25 +08:00
#include "llvm/Support/MachineValueType.h"
tablegen: Fix android build use llvm::to_string instead of std:: version. llvm-svn: 295711
2017-02-21 17:19:41 +08:00
#include "llvm/Support/ScopedPrinter.h"
[GlobalISel] Add basic Selector-emitter tblgen backend. This adds a basic tablegen backend that analyzes the SelectionDAG patterns to find simple ones that are eligible for GlobalISel-emission. That's similar to FastISel, with one notable difference: we're not fed ISD opcodes, so we need to map the SDNode operators to generic opcodes. That's done using GINodeEquiv in TargetGlobalISel.td. Otherwise, this is mostly boilerplate, and lots of filtering of any kind of "complicated" pattern. On AArch64, this is sufficient to match G_ADD up to s64 (to ADDWrr/ADDXrr) and G_BR (to B). Differential Revision: https://reviews.llvm.org/D26878 llvm-svn: 290284
2016-12-22 07:26:20 +08:00
#include "llvm/TableGen/Error.h"
#include "llvm/TableGen/Record.h"
#include "llvm/TableGen/TableGenBackend.h"
[globalisel][tblgen] Add support for ComplexPatterns Summary: Adds a new kind of MachineOperand: MO_Placeholder. This operand must not appear in the MIR and only exists as a way of creating an 'uninitialized' operand until a matcher function overwrites it. Depends on D30046, D29712 Reviewers: t.p.northover, ab, rovka, aditya_nandakumar, javed.absar, qcolombet Reviewed By: qcolombet Subscribers: dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D30089 llvm-svn: 297782
2017-03-15 05:32:08 +08:00
#include <numeric>
[globalisel][tablegen] Generate rule coverage and use it to identify untested rules Summary: This patch adds a LLVM_ENABLE_GISEL_COV which, like LLVM_ENABLE_DAGISEL_COV, causes TableGen to instrument the generated table to collect rule coverage information. However, LLVM_ENABLE_GISEL_COV goes a bit further than LLVM_ENABLE_DAGISEL_COV. The information is written to files (${CMAKE_BINARY_DIR}/gisel-coverage-* by default). These files can then be concatenated into ${LLVM_GISEL_COV_PREFIX}-all after which TableGen will read this information and use it to emit warnings about untested rules. This technique could also be used by SelectionDAG and can be further extended to detect hot rules and give them priority over colder rules. Usage: * Enable LLVM_ENABLE_GISEL_COV in CMake * Build the compiler and run some tests * cat gisel-coverage-[0-9]* > gisel-coverage-all * Delete lib/Target/*/*GenGlobalISel.inc* * Build the compiler Known issues: * ${LLVM_GISEL_COV_PREFIX}-all must be generated as a manual step due to a lack of a portable 'cat' command. It should be the concatenation of all ${LLVM_GISEL_COV_PREFIX}-[0-9]* files. * There's no mechanism to discard coverage information when the ruleset changes Depends on D39742 Reviewers: ab, qcolombet, t.p.northover, aditya_nandakumar, rovka Reviewed By: rovka Subscribers: vsk, arsenm, nhaehnle, mgorny, kristof.beyls, javed.absar, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D39747 llvm-svn: 318356
2017-11-16 08:46:35 +08:00
#include <string>
[GlobalISel] Add basic Selector-emitter tblgen backend. This adds a basic tablegen backend that analyzes the SelectionDAG patterns to find simple ones that are eligible for GlobalISel-emission. That's similar to FastISel, with one notable difference: we're not fed ISD opcodes, so we need to map the SDNode operators to generic opcodes. That's done using GINodeEquiv in TargetGlobalISel.td. Otherwise, this is mostly boilerplate, and lots of filtering of any kind of "complicated" pattern. On AArch64, this is sufficient to match G_ADD up to s64 (to ADDWrr/ADDXrr) and G_BR (to B). Differential Revision: https://reviews.llvm.org/D26878 llvm-svn: 290284
2016-12-22 07:26:20 +08:00
using namespace llvm;
#define DEBUG_TYPE "gisel-emitter"
STATISTIC(NumPatternTotal, "Total number of patterns");
[globalisel] Separate the SelectionDAG importer from the emitter. NFC Summary: In the near future the rules will be sorted between these two steps to ensure that more important rules are not prevented by less important ones. Reviewers: t.p.northover, ab, rovka, qcolombet, aditya_nandakumar Reviewed By: ab Subscribers: dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D29709 llvm-svn: 295661
2017-02-20 22:31:27 +08:00
STATISTIC(NumPatternImported, "Number of patterns imported from SelectionDAG");
STATISTIC(NumPatternImportsSkipped, "Number of SelectionDAG imports skipped");
[globalisel][tablegen] Generate rule coverage and use it to identify untested rules Summary: This patch adds a LLVM_ENABLE_GISEL_COV which, like LLVM_ENABLE_DAGISEL_COV, causes TableGen to instrument the generated table to collect rule coverage information. However, LLVM_ENABLE_GISEL_COV goes a bit further than LLVM_ENABLE_DAGISEL_COV. The information is written to files (${CMAKE_BINARY_DIR}/gisel-coverage-* by default). These files can then be concatenated into ${LLVM_GISEL_COV_PREFIX}-all after which TableGen will read this information and use it to emit warnings about untested rules. This technique could also be used by SelectionDAG and can be further extended to detect hot rules and give them priority over colder rules. Usage: * Enable LLVM_ENABLE_GISEL_COV in CMake * Build the compiler and run some tests * cat gisel-coverage-[0-9]* > gisel-coverage-all * Delete lib/Target/*/*GenGlobalISel.inc* * Build the compiler Known issues: * ${LLVM_GISEL_COV_PREFIX}-all must be generated as a manual step due to a lack of a portable 'cat' command. It should be the concatenation of all ${LLVM_GISEL_COV_PREFIX}-[0-9]* files. * There's no mechanism to discard coverage information when the ruleset changes Depends on D39742 Reviewers: ab, qcolombet, t.p.northover, aditya_nandakumar, rovka Reviewed By: rovka Subscribers: vsk, arsenm, nhaehnle, mgorny, kristof.beyls, javed.absar, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D39747 llvm-svn: 318356
2017-11-16 08:46:35 +08:00
STATISTIC(NumPatternsTested, "Number of patterns executed according to coverage information");
[GlobalISel] Add basic Selector-emitter tblgen backend. This adds a basic tablegen backend that analyzes the SelectionDAG patterns to find simple ones that are eligible for GlobalISel-emission. That's similar to FastISel, with one notable difference: we're not fed ISD opcodes, so we need to map the SDNode operators to generic opcodes. That's done using GINodeEquiv in TargetGlobalISel.td. Otherwise, this is mostly boilerplate, and lots of filtering of any kind of "complicated" pattern. On AArch64, this is sufficient to match G_ADD up to s64 (to ADDWrr/ADDXrr) and G_BR (to B). Differential Revision: https://reviews.llvm.org/D26878 llvm-svn: 290284
2016-12-22 07:26:20 +08:00
STATISTIC(NumPatternEmitted, "Number of patterns emitted");
[tablegen] Use categories on options that only matter to one emitter. Summary: The categories are emitted in a strange order in this patch due to a bug in the CommandLine library. Reviewers: ab Reviewed By: ab Subscribers: ab, llvm-commits Differential Revision: https://reviews.llvm.org/D30995 llvm-svn: 298843
2017-03-27 21:15:13 +08:00
cl::OptionCategory GlobalISelEmitterCat("Options for -gen-global-isel");
[GlobalISel] Add basic Selector-emitter tblgen backend. This adds a basic tablegen backend that analyzes the SelectionDAG patterns to find simple ones that are eligible for GlobalISel-emission. That's similar to FastISel, with one notable difference: we're not fed ISD opcodes, so we need to map the SDNode operators to generic opcodes. That's done using GINodeEquiv in TargetGlobalISel.td. Otherwise, this is mostly boilerplate, and lots of filtering of any kind of "complicated" pattern. On AArch64, this is sufficient to match G_ADD up to s64 (to ADDWrr/ADDXrr) and G_BR (to B). Differential Revision: https://reviews.llvm.org/D26878 llvm-svn: 290284
2016-12-22 07:26:20 +08:00
static cl::opt<bool> WarnOnSkippedPatterns(
"warn-on-skipped-patterns",
cl::desc("Explain why a pattern was skipped for inclusion "
"in the GlobalISel selector"),
[tablegen] Use categories on options that only matter to one emitter. Summary: The categories are emitted in a strange order in this patch due to a bug in the CommandLine library. Reviewers: ab Reviewed By: ab Subscribers: ab, llvm-commits Differential Revision: https://reviews.llvm.org/D30995 llvm-svn: 298843
2017-03-27 21:15:13 +08:00
cl::init(false), cl::cat(GlobalISelEmitterCat));
[GlobalISel] Add basic Selector-emitter tblgen backend. This adds a basic tablegen backend that analyzes the SelectionDAG patterns to find simple ones that are eligible for GlobalISel-emission. That's similar to FastISel, with one notable difference: we're not fed ISD opcodes, so we need to map the SDNode operators to generic opcodes. That's done using GINodeEquiv in TargetGlobalISel.td. Otherwise, this is mostly boilerplate, and lots of filtering of any kind of "complicated" pattern. On AArch64, this is sufficient to match G_ADD up to s64 (to ADDWrr/ADDXrr) and G_BR (to B). Differential Revision: https://reviews.llvm.org/D26878 llvm-svn: 290284
2016-12-22 07:26:20 +08:00
[globalisel][tablegen] Generate rule coverage and use it to identify untested rules Summary: This patch adds a LLVM_ENABLE_GISEL_COV which, like LLVM_ENABLE_DAGISEL_COV, causes TableGen to instrument the generated table to collect rule coverage information. However, LLVM_ENABLE_GISEL_COV goes a bit further than LLVM_ENABLE_DAGISEL_COV. The information is written to files (${CMAKE_BINARY_DIR}/gisel-coverage-* by default). These files can then be concatenated into ${LLVM_GISEL_COV_PREFIX}-all after which TableGen will read this information and use it to emit warnings about untested rules. This technique could also be used by SelectionDAG and can be further extended to detect hot rules and give them priority over colder rules. Usage: * Enable LLVM_ENABLE_GISEL_COV in CMake * Build the compiler and run some tests * cat gisel-coverage-[0-9]* > gisel-coverage-all * Delete lib/Target/*/*GenGlobalISel.inc* * Build the compiler Known issues: * ${LLVM_GISEL_COV_PREFIX}-all must be generated as a manual step due to a lack of a portable 'cat' command. It should be the concatenation of all ${LLVM_GISEL_COV_PREFIX}-[0-9]* files. * There's no mechanism to discard coverage information when the ruleset changes Depends on D39742 Reviewers: ab, qcolombet, t.p.northover, aditya_nandakumar, rovka Reviewed By: rovka Subscribers: vsk, arsenm, nhaehnle, mgorny, kristof.beyls, javed.absar, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D39747 llvm-svn: 318356
2017-11-16 08:46:35 +08:00
static cl::opt<bool> GenerateCoverage(
"instrument-gisel-coverage",
cl::desc("Generate coverage instrumentation for GlobalISel"),
cl::init(false), cl::cat(GlobalISelEmitterCat));
static cl::opt<std::string> UseCoverageFile(
"gisel-coverage-file", cl::init(""),
cl::desc("Specify file to retrieve coverage information from"),
cl::cat(GlobalISelEmitterCat));
[TableGen][GlobalISel] Optimize MatchTable for faster instruction selection *** Context *** Prior to this patchw, the table generated for matching instruction was straight forward but highly inefficient. Basically, each pattern generates its own set of self contained checks and actions. E.g., TableGen generated: // First pattern CheckNumOperand 3 CheckOpcode G_ADD ... Build ADDrr // Second pattern CheckNumOperand 3 CheckOpcode G_ADD ... Build ADDri // Third pattern CheckNumOperand 3 CheckOpcode G_SUB ... Build SUBrr *** Problem *** Because of that generation, a *lot* of check were redundant between each pattern and were checked every single time until we reach the pattern that matches. E.g., Taking the previous table, let say we are matching a G_SUB, that means we were going to check all the rules for G_ADD before looking at the G_SUB rule. In particular we are going to do: check 3 operands; PASS check G_ADD; FAIL ; Next rule check 3 operands; PASS (but we already knew that!) check G_ADD; FAIL (well it is still not true) ; Next rule check 3 operands; PASS (really!!) check G_SUB; PASS (at last :P) *** Proposed Solution *** This patch introduces a concept of group of rules (GroupMatcher) that share some predicates and only get checked once for the whole group. This patch only creates groups with one nesting level. Conceptually there is nothing preventing us for having deeper nest level. However, the current implementation is not smart enough to share the recording (aka capturing) of values. That limits its ability to do more sharing. For the given example the current patch will generate: // First group CheckOpcode G_ADD // First pattern CheckNumOperand 3 ... Build ADDrr // Second pattern CheckNumOperand 3 ... Build ADDri // Second group CheckOpcode G_SUB // Third pattern CheckNumOperand 3 ... Build SUBrr But if we allowed several nesting level, it could create a sub group for the checknumoperand 3. (We would need to call optimizeRules on the rules within a group.) *** Result *** With only one level of nesting, the instruction selection pass is up to 4x faster. For instance, one instruction now takes 500 checks, instead of 24k! With more nesting we could get in the tens I believe. Differential Revision: https://reviews.llvm.org/D39034 rdar://problem/34670699 llvm-svn: 321017
2017-12-19 03:47:41 +08:00
static cl::opt<bool> OptimizeMatchTable(
"optimize-match-table",
cl::desc("Generate an optimized version of the match table"),
cl::init(true), cl::cat(GlobalISelEmitterCat));
[tablegen][globalisel] Trivial changes to reduce size of following patch. NFC. Summary: * Move namespace { * Trivial: Typo * RuleMatcher: Separate class and definition * Trivial: const findNodeEquiv Reviewers: t.p.northover, qcolombet, aditya_nandakumar, ab, rovka Reviewed By: rovka Subscribers: dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D30531 llvm-svn: 297884
2017-03-16 04:18:38 +08:00
namespace {
[GlobalISel] Add basic Selector-emitter tblgen backend. This adds a basic tablegen backend that analyzes the SelectionDAG patterns to find simple ones that are eligible for GlobalISel-emission. That's similar to FastISel, with one notable difference: we're not fed ISD opcodes, so we need to map the SDNode operators to generic opcodes. That's done using GINodeEquiv in TargetGlobalISel.td. Otherwise, this is mostly boilerplate, and lots of filtering of any kind of "complicated" pattern. On AArch64, this is sufficient to match G_ADD up to s64 (to ADDWrr/ADDXrr) and G_BR (to B). Differential Revision: https://reviews.llvm.org/D26878 llvm-svn: 290284
2016-12-22 07:26:20 +08:00
//===- Helper functions ---------------------------------------------------===//
[globalisel][tablegen] Add support for fpimm and import of APInt/APFloat based ImmLeaf. Summary: There's only a tablegen testcase for IntImmLeaf and not a CodeGen one because the relevant rules are rejected for other reasons at the moment. On AArch64, it's because there's an SDNodeXForm attached to the operand. On X86, it's because the rule either emits multiple instructions or has another predicate using PatFrag which cannot easily be supported at the same time. Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Reviewed By: qcolombet Subscribers: aemerson, javed.absar, igorb, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D36569 llvm-svn: 315761
2017-10-14 05:28:03 +08:00
/// Get the name of the enum value used to number the predicate function.
std::string getEnumNameForPredicate(const TreePredicateFn &Predicate) {
[TableGen] Avoid unnecessary std::string creations Avoid unnecessary std::string creations in the TreePredicateFn getters. llvm-svn: 315826
2017-10-15 05:27:53 +08:00
return "GIPFP_" + Predicate.getImmTypeIdentifier().str() + "_" +
[globalisel][tablegen] Add support for fpimm and import of APInt/APFloat based ImmLeaf. Summary: There's only a tablegen testcase for IntImmLeaf and not a CodeGen one because the relevant rules are rejected for other reasons at the moment. On AArch64, it's because there's an SDNodeXForm attached to the operand. On X86, it's because the rule either emits multiple instructions or has another predicate using PatFrag which cannot easily be supported at the same time. Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Reviewed By: qcolombet Subscribers: aemerson, javed.absar, igorb, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D36569 llvm-svn: 315761
2017-10-14 05:28:03 +08:00
Predicate.getFnName();
}
/// Get the opcode used to check this predicate.
std::string getMatchOpcodeForPredicate(const TreePredicateFn &Predicate) {
[TableGen] Avoid unnecessary std::string creations Avoid unnecessary std::string creations in the TreePredicateFn getters. llvm-svn: 315826
2017-10-15 05:27:53 +08:00
return "GIM_Check" + Predicate.getImmTypeIdentifier().str() + "ImmPredicate";
[globalisel][tablegen] Add support for fpimm and import of APInt/APFloat based ImmLeaf. Summary: There's only a tablegen testcase for IntImmLeaf and not a CodeGen one because the relevant rules are rejected for other reasons at the moment. On AArch64, it's because there's an SDNodeXForm attached to the operand. On X86, it's because the rule either emits multiple instructions or has another predicate using PatFrag which cannot easily be supported at the same time. Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Reviewed By: qcolombet Subscribers: aemerson, javed.absar, igorb, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D36569 llvm-svn: 315761
2017-10-14 05:28:03 +08:00
}
Recommit: [globalisel] Change LLT constructor string into an LLT-based object that knows how to generate it. Summary: This will allow future patches to inspect the details of the LLT. The implementation is now split between the Support and CodeGen libraries to allow TableGen to use this class without introducing layering concerns. Thanks to Ahmed Bougacha for finding a reasonable way to avoid the layering issue and providing the version of this patch without that problem. The problem with the previous commit appears to have been that TableGen was including CodeGen/LowLevelType.h instead of Support/LowLevelTypeImpl.h. Reviewers: t.p.northover, qcolombet, rovka, aditya_nandakumar, ab, javed.absar Subscribers: arsenm, nhaehnle, mgorny, dberris, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D30046 llvm-svn: 297241
2017-03-08 07:20:35 +08:00
/// This class stands in for LLT wherever we want to tablegen-erate an
/// equivalent at compiler run-time.
class LLTCodeGen {
private:
LLT Ty;
public:
Reapply r332907 "[GlobalISel] Improving InstructionSelect's performance by reducing MatchTable..." Apparently the compile time problem was caused by the fact that not all compilers / STL implementations can automatically convert std::unique_ptr<Derived> to std::unique_ptr<Base>. Fixed (hopefully) by making sure it's std::unique_ptr<Derived>&& (rvalue ref) to std::unique_ptr<Base> conversion instead. llvm-svn: 332917
2018-05-22 07:28:51 +08:00
LLTCodeGen() = default;
Recommit: [globalisel] Change LLT constructor string into an LLT-based object that knows how to generate it. Summary: This will allow future patches to inspect the details of the LLT. The implementation is now split between the Support and CodeGen libraries to allow TableGen to use this class without introducing layering concerns. Thanks to Ahmed Bougacha for finding a reasonable way to avoid the layering issue and providing the version of this patch without that problem. The problem with the previous commit appears to have been that TableGen was including CodeGen/LowLevelType.h instead of Support/LowLevelTypeImpl.h. Reviewers: t.p.northover, qcolombet, rovka, aditya_nandakumar, ab, javed.absar Subscribers: arsenm, nhaehnle, mgorny, dberris, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D30046 llvm-svn: 297241
2017-03-08 07:20:35 +08:00
LLTCodeGen(const LLT &Ty) : Ty(Ty) {}
[globalisel][tablegen] Add control-flow to the MatchTable. Summary: This will allow us to merge the various sub-tables into a single table. This is a compile-time saving at this point. However, this will also enable the optimization of a table so that similar instructions can be tested together, reducing the time spent on the matching the code. The bulk of this patch is a mechanical conversion to the new MatchTable object which is responsible for tracking label definitions and filling in the index of the jump targets. It is also responsible for nicely formatting the table. This was necessary to support the new GIM_Try opcode which takes the index to jump to if the match should fail. This value is unknown during table construction and is filled in during emission. To support nesting try-blocks (although we currently don't emit tables with nested try-blocks), GIM_Reject has been re-introduced to explicitly exit a try-block or fail the overall match if there are no active try-blocks. Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Reviewed By: rovka Subscribers: kristof.beyls, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D35117 llvm-svn: 308596
2017-07-20 17:25:44 +08:00
std::string getCxxEnumValue() const {
std::string Str;
raw_string_ostream OS(Str);
emitCxxEnumValue(OS);
return OS.str();
}
[globalisel][tablegen] Partially fix compile-time regressions by converting matcher to state-machine(s) Summary: Replace the matcher if-statements for each rule with a state-machine. This significantly reduces compile time, memory allocations, and cumulative memory allocation when compiling AArch64InstructionSelector.cpp.o after r303259 is recommitted. The following patches will expand on this further to fully fix the regressions. Reviewers: rovka, ab, t.p.northover, qcolombet, aditya_nandakumar Reviewed By: ab Subscribers: vitalybuka, aemerson, javed.absar, igorb, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D33758 llvm-svn: 307079
2017-07-04 22:35:06 +08:00
void emitCxxEnumValue(raw_ostream &OS) const {
if (Ty.isScalar()) {
OS << "GILLT_s" << Ty.getSizeInBits();
return;
}
if (Ty.isVector()) {
OS << "GILLT_v" << Ty.getNumElements() << "s" << Ty.getScalarSizeInBits();
return;
}
[globalisel][tablegen] Implement unindexed load, non-extending load, and MemVT checks Summary: This includes some context-sensitivity in the MVT to LLT conversion so that pointer types are tested correctly. FIXME: I'm not happy with the way this is done since everything is a special-case. I've yet to find a reasonable way to implement it. select-load.mir fails because <1 x s64> loads in tablegen get priority over s64 loads. This is fixed in the next patch and as such they should be committed together, I've posted them separately to help with the review. Depends on D37456 Reviewers: ab, qcolombet, t.p.northover, rovka, aditya_nandakumar Subscribers: kristof.beyls, javed.absar, llvm-commits, igorb Differential Revision: https://reviews.llvm.org/D37457 llvm-svn: 315884
2017-10-16 08:56:30 +08:00
if (Ty.isPointer()) {
OS << "GILLT_p" << Ty.getAddressSpace();
if (Ty.getSizeInBits() > 0)
OS << "s" << Ty.getSizeInBits();
return;
}
[globalisel][tablegen] Partially fix compile-time regressions by converting matcher to state-machine(s) Summary: Replace the matcher if-statements for each rule with a state-machine. This significantly reduces compile time, memory allocations, and cumulative memory allocation when compiling AArch64InstructionSelector.cpp.o after r303259 is recommitted. The following patches will expand on this further to fully fix the regressions. Reviewers: rovka, ab, t.p.northover, qcolombet, aditya_nandakumar Reviewed By: ab Subscribers: vitalybuka, aemerson, javed.absar, igorb, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D33758 llvm-svn: 307079
2017-07-04 22:35:06 +08:00
llvm_unreachable("Unhandled LLT");
}
Recommit: [globalisel] Change LLT constructor string into an LLT-based object that knows how to generate it. Summary: This will allow future patches to inspect the details of the LLT. The implementation is now split between the Support and CodeGen libraries to allow TableGen to use this class without introducing layering concerns. Thanks to Ahmed Bougacha for finding a reasonable way to avoid the layering issue and providing the version of this patch without that problem. The problem with the previous commit appears to have been that TableGen was including CodeGen/LowLevelType.h instead of Support/LowLevelTypeImpl.h. Reviewers: t.p.northover, qcolombet, rovka, aditya_nandakumar, ab, javed.absar Subscribers: arsenm, nhaehnle, mgorny, dberris, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D30046 llvm-svn: 297241
2017-03-08 07:20:35 +08:00
void emitCxxConstructorCall(raw_ostream &OS) const {
if (Ty.isScalar()) {
OS << "LLT::scalar(" << Ty.getSizeInBits() << ")";
return;
}
if (Ty.isVector()) {
[globalisel][tablegen] Multiple 80-col corrections. llvm-svn: 306544
2017-06-28 21:50:04 +08:00
OS << "LLT::vector(" << Ty.getNumElements() << ", "
<< Ty.getScalarSizeInBits() << ")";
Recommit: [globalisel] Change LLT constructor string into an LLT-based object that knows how to generate it. Summary: This will allow future patches to inspect the details of the LLT. The implementation is now split between the Support and CodeGen libraries to allow TableGen to use this class without introducing layering concerns. Thanks to Ahmed Bougacha for finding a reasonable way to avoid the layering issue and providing the version of this patch without that problem. The problem with the previous commit appears to have been that TableGen was including CodeGen/LowLevelType.h instead of Support/LowLevelTypeImpl.h. Reviewers: t.p.northover, qcolombet, rovka, aditya_nandakumar, ab, javed.absar Subscribers: arsenm, nhaehnle, mgorny, dberris, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D30046 llvm-svn: 297241
2017-03-08 07:20:35 +08:00
return;
}
[globalisel][tablegen] Implement unindexed load, non-extending load, and MemVT checks Summary: This includes some context-sensitivity in the MVT to LLT conversion so that pointer types are tested correctly. FIXME: I'm not happy with the way this is done since everything is a special-case. I've yet to find a reasonable way to implement it. select-load.mir fails because <1 x s64> loads in tablegen get priority over s64 loads. This is fixed in the next patch and as such they should be committed together, I've posted them separately to help with the review. Depends on D37456 Reviewers: ab, qcolombet, t.p.northover, rovka, aditya_nandakumar Subscribers: kristof.beyls, javed.absar, llvm-commits, igorb Differential Revision: https://reviews.llvm.org/D37457 llvm-svn: 315884
2017-10-16 08:56:30 +08:00
if (Ty.isPointer() && Ty.getSizeInBits() > 0) {
OS << "LLT::pointer(" << Ty.getAddressSpace() << ", "
<< Ty.getSizeInBits() << ")";
return;
}
Recommit: [globalisel] Change LLT constructor string into an LLT-based object that knows how to generate it. Summary: This will allow future patches to inspect the details of the LLT. The implementation is now split between the Support and CodeGen libraries to allow TableGen to use this class without introducing layering concerns. Thanks to Ahmed Bougacha for finding a reasonable way to avoid the layering issue and providing the version of this patch without that problem. The problem with the previous commit appears to have been that TableGen was including CodeGen/LowLevelType.h instead of Support/LowLevelTypeImpl.h. Reviewers: t.p.northover, qcolombet, rovka, aditya_nandakumar, ab, javed.absar Subscribers: arsenm, nhaehnle, mgorny, dberris, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D30046 llvm-svn: 297241
2017-03-08 07:20:35 +08:00
llvm_unreachable("Unhandled LLT");
}
[globalisel][tblgen] Add support for ComplexPatterns Summary: Adds a new kind of MachineOperand: MO_Placeholder. This operand must not appear in the MIR and only exists as a way of creating an 'uninitialized' operand until a matcher function overwrites it. Depends on D30046, D29712 Reviewers: t.p.northover, ab, rovka, aditya_nandakumar, javed.absar, qcolombet Reviewed By: qcolombet Subscribers: dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D30089 llvm-svn: 297782
2017-03-15 05:32:08 +08:00
const LLT &get() const { return Ty; }
[globalisel][tablegen] Partially fix compile-time regressions by converting matcher to state-machine(s) Summary: Replace the matcher if-statements for each rule with a state-machine. This significantly reduces compile time, memory allocations, and cumulative memory allocation when compiling AArch64InstructionSelector.cpp.o after r303259 is recommitted. The following patches will expand on this further to fully fix the regressions. Reviewers: rovka, ab, t.p.northover, qcolombet, aditya_nandakumar Reviewed By: ab Subscribers: vitalybuka, aemerson, javed.absar, igorb, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D33758 llvm-svn: 307079
2017-07-04 22:35:06 +08:00
[TableGen] Change std::sort to llvm::sort in response to r327219 Summary: r327219 added wrappers to std::sort which randomly shuffle the container before sorting. This will help in uncovering non-determinism caused due to undefined sorting order of objects having the same key. To make use of that infrastructure we need to invoke llvm::sort instead of std::sort. Note: This patch is one of a series of patches to replace *all* std::sort to llvm::sort. Refer the comments section in D44363 for a list of all the required patches. Reviewers: stoklund, kparzysz, dsanders Reviewed By: dsanders Subscribers: llvm-commits Differential Revision: https://reviews.llvm.org/D45144 llvm-svn: 329451
2018-04-07 04:18:05 +08:00
/// This ordering is used for std::unique() and llvm::sort(). There's no
[globalisel][tablegen] Generate TypeObject table. NFC Summary: Generate the type table from the types used by a target rather than hard-coding the union of types used by all targets. Depends on D36084 Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Reviewed By: rovka Subscribers: kristof.beyls, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D36085 llvm-svn: 311084
2017-08-17 21:18:35 +08:00
/// particular logic behind the order but either A < B or B < A must be
/// true if A != B.
[globalisel][tablegen] Partially fix compile-time regressions by converting matcher to state-machine(s) Summary: Replace the matcher if-statements for each rule with a state-machine. This significantly reduces compile time, memory allocations, and cumulative memory allocation when compiling AArch64InstructionSelector.cpp.o after r303259 is recommitted. The following patches will expand on this further to fully fix the regressions. Reviewers: rovka, ab, t.p.northover, qcolombet, aditya_nandakumar Reviewed By: ab Subscribers: vitalybuka, aemerson, javed.absar, igorb, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D33758 llvm-svn: 307079
2017-07-04 22:35:06 +08:00
bool operator<(const LLTCodeGen &Other) const {
[globalisel][tablegen] Generate TypeObject table. NFC Summary: Generate the type table from the types used by a target rather than hard-coding the union of types used by all targets. Depends on D36084 Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Reviewed By: rovka Subscribers: kristof.beyls, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D36085 llvm-svn: 311084
2017-08-17 21:18:35 +08:00
if (Ty.isValid() != Other.Ty.isValid())
return Ty.isValid() < Other.Ty.isValid();
[globalisel][tablegen] Partially fix compile-time regressions by converting matcher to state-machine(s) Summary: Replace the matcher if-statements for each rule with a state-machine. This significantly reduces compile time, memory allocations, and cumulative memory allocation when compiling AArch64InstructionSelector.cpp.o after r303259 is recommitted. The following patches will expand on this further to fully fix the regressions. Reviewers: rovka, ab, t.p.northover, qcolombet, aditya_nandakumar Reviewed By: ab Subscribers: vitalybuka, aemerson, javed.absar, igorb, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D33758 llvm-svn: 307079
2017-07-04 22:35:06 +08:00
if (!Ty.isValid())
Revert r310716 (and r310735): [globalisel][tablegen] Support zero-instruction emission. Two of the Windows bots are failing test\CodeGen\X86\GlobalISel\select-inc.mir which should not have been affected by the change. Reverting while I investigate. Also reverted r310735 because it builds on r310716. llvm-svn: 310745
2017-08-12 03:19:21 +08:00
return false;
[globalisel][tablegen] Generate TypeObject table. NFC Summary: Generate the type table from the types used by a target rather than hard-coding the union of types used by all targets. Depends on D36084 Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Reviewed By: rovka Subscribers: kristof.beyls, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D36085 llvm-svn: 311084
2017-08-17 21:18:35 +08:00
if (Ty.isVector() != Other.Ty.isVector())
return Ty.isVector() < Other.Ty.isVector();
if (Ty.isScalar() != Other.Ty.isScalar())
return Ty.isScalar() < Other.Ty.isScalar();
if (Ty.isPointer() != Other.Ty.isPointer())
return Ty.isPointer() < Other.Ty.isPointer();
if (Ty.isPointer() && Ty.getAddressSpace() != Other.Ty.getAddressSpace())
return Ty.getAddressSpace() < Other.Ty.getAddressSpace();
if (Ty.isVector() && Ty.getNumElements() != Other.Ty.getNumElements())
return Ty.getNumElements() < Other.Ty.getNumElements();
return Ty.getSizeInBits() < Other.Ty.getSizeInBits();
[globalisel][tablegen] Partially fix compile-time regressions by converting matcher to state-machine(s) Summary: Replace the matcher if-statements for each rule with a state-machine. This significantly reduces compile time, memory allocations, and cumulative memory allocation when compiling AArch64InstructionSelector.cpp.o after r303259 is recommitted. The following patches will expand on this further to fully fix the regressions. Reviewers: rovka, ab, t.p.northover, qcolombet, aditya_nandakumar Reviewed By: ab Subscribers: vitalybuka, aemerson, javed.absar, igorb, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D33758 llvm-svn: 307079
2017-07-04 22:35:06 +08:00
}
[TableGen][GlobalISel] Make the different Matcher comparable This opens refactoring opportunities in the match table now that we can check that two predicates are the same. NFC. llvm-svn: 320890
2017-12-16 07:24:39 +08:00
bool operator==(const LLTCodeGen &B) const { return Ty == B.Ty; }
[globalisel][tblgen] Add support for ComplexPatterns Summary: Adds a new kind of MachineOperand: MO_Placeholder. This operand must not appear in the MIR and only exists as a way of creating an 'uninitialized' operand until a matcher function overwrites it. Depends on D30046, D29712 Reviewers: t.p.northover, ab, rovka, aditya_nandakumar, javed.absar, qcolombet Reviewed By: qcolombet Subscribers: dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D30089 llvm-svn: 297782
2017-03-15 05:32:08 +08:00
};
[globalisel] Update GlobalISel emitter to match new representation of extending loads Summary: Previously, a extending load was represented at (G_*EXT (G_LOAD x)). This had a few drawbacks: * G_LOAD had to be legal for all sizes you could extend from, even if registers didn't naturally hold those sizes. * All sizes you could extend from had to be allocatable just in case the extend went missing (e.g. by optimization). * At minimum, G_*EXT and G_TRUNC had to be legal for these sizes. As we improve optimization of extends and truncates, this legality requirement would spread without considerable care w.r.t when certain combines were permitted. * The SelectionDAG importer required some ugly and fragile pattern rewriting to translate patterns into this style. This patch changes the representation to: * (G_[SZ]EXTLOAD x) * (G_LOAD x) any-extends when MMO.getSize() * 8 < ResultTy.getSizeInBits() which resolves these issues by allowing targets to work entirely in their native register sizes, and by having a more direct translation from SelectionDAG patterns. Each extending load can be lowered by the legalizer into separate extends and loads, however a target that supports s1 will need the any-extending load to extend to at least s8 since LLVM does not represent memory accesses smaller than 8 bit. The legalizer can widenScalar G_LOAD into an any-extending load but sign/zero-extending loads need help from something else like a combiner pass. A follow-up patch that adds combiner helpers for for this will follow. The new representation requires that the MMO correctly reflect the memory access so this has been corrected in a couple tests. I've also moved the extending loads to their own tests since they are (mostly) separate opcodes now. Additionally, the re-write appears to have invalidated two tests from select-with-no-legality-check.mir since the matcher table no longer contains loads that result in s1's and they aren't legal in AArch64 anymore. Depends on D45540 Reviewers: ab, aditya_nandakumar, bogner, rtereshin, volkan, rovka, javed.absar Reviewed By: rtereshin Subscribers: javed.absar, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D45541 llvm-svn: 331601
2018-05-06 04:53:24 +08:00
// Track all types that are used so we can emit the corresponding enum.
std::set<LLTCodeGen> KnownTypes;
[globalisel][tblgen] Add support for ComplexPatterns Summary: Adds a new kind of MachineOperand: MO_Placeholder. This operand must not appear in the MIR and only exists as a way of creating an 'uninitialized' operand until a matcher function overwrites it. Depends on D30046, D29712 Reviewers: t.p.northover, ab, rovka, aditya_nandakumar, javed.absar, qcolombet Reviewed By: qcolombet Subscribers: dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D30089 llvm-svn: 297782
2017-03-15 05:32:08 +08:00
class InstructionMatcher;
[GlobalISel] Add basic Selector-emitter tblgen backend. This adds a basic tablegen backend that analyzes the SelectionDAG patterns to find simple ones that are eligible for GlobalISel-emission. That's similar to FastISel, with one notable difference: we're not fed ISD opcodes, so we need to map the SDNode operators to generic opcodes. That's done using GINodeEquiv in TargetGlobalISel.td. Otherwise, this is mostly boilerplate, and lots of filtering of any kind of "complicated" pattern. On AArch64, this is sufficient to match G_ADD up to s64 (to ADDWrr/ADDXrr) and G_BR (to B). Differential Revision: https://reviews.llvm.org/D26878 llvm-svn: 290284
2016-12-22 07:26:20 +08:00
/// Convert an MVT to an equivalent LLT if possible, or the invalid LLT() for
/// MVTs that don't map cleanly to an LLT (e.g., iPTR, *any, ...).
Recommit: [globalisel] Change LLT constructor string into an LLT-based object that knows how to generate it. Summary: This will allow future patches to inspect the details of the LLT. The implementation is now split between the Support and CodeGen libraries to allow TableGen to use this class without introducing layering concerns. Thanks to Ahmed Bougacha for finding a reasonable way to avoid the layering issue and providing the version of this patch without that problem. The problem with the previous commit appears to have been that TableGen was including CodeGen/LowLevelType.h instead of Support/LowLevelTypeImpl.h. Reviewers: t.p.northover, qcolombet, rovka, aditya_nandakumar, ab, javed.absar Subscribers: arsenm, nhaehnle, mgorny, dberris, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D30046 llvm-svn: 297241
2017-03-08 07:20:35 +08:00
static Optional<LLTCodeGen> MVTToLLT(MVT::SimpleValueType SVT) {
[GlobalISel] Add basic Selector-emitter tblgen backend. This adds a basic tablegen backend that analyzes the SelectionDAG patterns to find simple ones that are eligible for GlobalISel-emission. That's similar to FastISel, with one notable difference: we're not fed ISD opcodes, so we need to map the SDNode operators to generic opcodes. That's done using GINodeEquiv in TargetGlobalISel.td. Otherwise, this is mostly boilerplate, and lots of filtering of any kind of "complicated" pattern. On AArch64, this is sufficient to match G_ADD up to s64 (to ADDWrr/ADDXrr) and G_BR (to B). Differential Revision: https://reviews.llvm.org/D26878 llvm-svn: 290284
2016-12-22 07:26:20 +08:00
MVT VT(SVT);
[globalisel][tablegen] Implement unindexed load, non-extending load, and MemVT checks Summary: This includes some context-sensitivity in the MVT to LLT conversion so that pointer types are tested correctly. FIXME: I'm not happy with the way this is done since everything is a special-case. I've yet to find a reasonable way to implement it. select-load.mir fails because <1 x s64> loads in tablegen get priority over s64 loads. This is fixed in the next patch and as such they should be committed together, I've posted them separately to help with the review. Depends on D37456 Reviewers: ab, qcolombet, t.p.northover, rovka, aditya_nandakumar Subscribers: kristof.beyls, javed.absar, llvm-commits, igorb Differential Revision: https://reviews.llvm.org/D37457 llvm-svn: 315884
2017-10-16 08:56:30 +08:00
Recommit: [globalisel] Change LLT constructor string into an LLT-based object that knows how to generate it. Summary: This will allow future patches to inspect the details of the LLT. The implementation is now split between the Support and CodeGen libraries to allow TableGen to use this class without introducing layering concerns. Thanks to Ahmed Bougacha for finding a reasonable way to avoid the layering issue and providing the version of this patch without that problem. The problem with the previous commit appears to have been that TableGen was including CodeGen/LowLevelType.h instead of Support/LowLevelTypeImpl.h. Reviewers: t.p.northover, qcolombet, rovka, aditya_nandakumar, ab, javed.absar Subscribers: arsenm, nhaehnle, mgorny, dberris, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D30046 llvm-svn: 297241
2017-03-08 07:20:35 +08:00
if (VT.isVector() && VT.getVectorNumElements() != 1)
[globalisel][tablegen] Multiple 80-col corrections. llvm-svn: 306544
2017-06-28 21:50:04 +08:00
return LLTCodeGen(
LLT::vector(VT.getVectorNumElements(), VT.getScalarSizeInBits()));
[globalisel][tablegen] Implement unindexed load, non-extending load, and MemVT checks Summary: This includes some context-sensitivity in the MVT to LLT conversion so that pointer types are tested correctly. FIXME: I'm not happy with the way this is done since everything is a special-case. I've yet to find a reasonable way to implement it. select-load.mir fails because <1 x s64> loads in tablegen get priority over s64 loads. This is fixed in the next patch and as such they should be committed together, I've posted them separately to help with the review. Depends on D37456 Reviewers: ab, qcolombet, t.p.northover, rovka, aditya_nandakumar Subscribers: kristof.beyls, javed.absar, llvm-commits, igorb Differential Revision: https://reviews.llvm.org/D37457 llvm-svn: 315884
2017-10-16 08:56:30 +08:00
Recommit: [globalisel] Change LLT constructor string into an LLT-based object that knows how to generate it. Summary: This will allow future patches to inspect the details of the LLT. The implementation is now split between the Support and CodeGen libraries to allow TableGen to use this class without introducing layering concerns. Thanks to Ahmed Bougacha for finding a reasonable way to avoid the layering issue and providing the version of this patch without that problem. The problem with the previous commit appears to have been that TableGen was including CodeGen/LowLevelType.h instead of Support/LowLevelTypeImpl.h. Reviewers: t.p.northover, qcolombet, rovka, aditya_nandakumar, ab, javed.absar Subscribers: arsenm, nhaehnle, mgorny, dberris, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D30046 llvm-svn: 297241
2017-03-08 07:20:35 +08:00
if (VT.isInteger() || VT.isFloatingPoint())
return LLTCodeGen(LLT::scalar(VT.getSizeInBits()));
return None;
[GlobalISel] Add basic Selector-emitter tblgen backend. This adds a basic tablegen backend that analyzes the SelectionDAG patterns to find simple ones that are eligible for GlobalISel-emission. That's similar to FastISel, with one notable difference: we're not fed ISD opcodes, so we need to map the SDNode operators to generic opcodes. That's done using GINodeEquiv in TargetGlobalISel.td. Otherwise, this is mostly boilerplate, and lots of filtering of any kind of "complicated" pattern. On AArch64, this is sufficient to match G_ADD up to s64 (to ADDWrr/ADDXrr) and G_BR (to B). Differential Revision: https://reviews.llvm.org/D26878 llvm-svn: 290284
2016-12-22 07:26:20 +08:00
}
[globalisel][tablegen] Report more detail in some SelectionDAG import failures. NFC Reviewers: ab, t.p.northover, qcolombet, aditya_nandakumar, rovka Reviewed By: ab Subscribers: dberris, kristof.beyls, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D31325 llvm-svn: 300186
2017-04-13 17:45:37 +08:00
static std::string explainPredicates(const TreePatternNode *N) {
std::string Explanation = "";
StringRef Separator = "";
for (const auto &P : N->getPredicateFns()) {
Explanation +=
(Separator + P.getOrigPatFragRecord()->getRecord()->getName()).str();
[globalisel][tablegen] Add support for importing G_ATOMIC_CMPXCHG, G_ATOMICRMW_* rules from SelectionDAG. GIM_CheckNonAtomic has been replaced by GIM_CheckAtomicOrdering to allow it to support a wider range of orderings. This has then been used to import patterns using nodes such as atomic_cmp_swap, atomic_swap, and atomic_load_*. llvm-svn: 319232
2017-11-29 06:07:05 +08:00
Separator = ", ";
[globalisel][tablegen] Report more detail in some SelectionDAG import failures. NFC Reviewers: ab, t.p.northover, qcolombet, aditya_nandakumar, rovka Reviewed By: ab Subscribers: dberris, kristof.beyls, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D31325 llvm-svn: 300186
2017-04-13 17:45:37 +08:00
if (P.isAlwaysTrue())
Explanation += " always-true";
if (P.isImmediatePattern())
Explanation += " immediate";
[tablegen] Handle common load/store predicates inside tablegen. NFC. Summary: GlobalISel and SelectionDAG require different code for the common load/store predicates due to differences in the representation. For example: SelectionDAG: (load<signext,i8>:i32 GPR32:$addr) // The <> denote properties of the SDNode that are not printed in the DAG GlobalISel: (G_SEXT:s32 (G_LOAD:s8 GPR32:$addr)) Even without that, differences in the IR (SDNode vs MachineInstr) require differences in the C++ predicate. This patch moves the implementation of the common load/store predicates into tablegen so that it can handle these differences. It's NFC for SelectionDAG since it emits equivalent code and it's NFC for GlobalISel since the rules involving the relevant predicates are still rejected by the importer. Depends on D36618 Reviewers: ab, qcolombet, t.p.northover, rovka, aditya_nandakumar Subscribers: llvm-commits, igorb Differential Revision: https://reviews.llvm.org/D37443 Includes a partial revert of r315826 since this patch makes it necessary for getPredCode() to return a std::string and getImmCode() should have the same interface as getPredCode(). llvm-svn: 315841
2017-10-15 10:06:44 +08:00
if (P.isUnindexed())
Explanation += " unindexed";
if (P.isNonExtLoad())
Explanation += " non-extload";
if (P.isAnyExtLoad())
Explanation += " extload";
if (P.isSignExtLoad())
Explanation += " sextload";
if (P.isZeroExtLoad())
Explanation += " zextload";
if (P.isNonTruncStore())
Explanation += " non-truncstore";
if (P.isTruncStore())
Explanation += " truncstore";
if (Record *VT = P.getMemoryVT())
Explanation += (" MemVT=" + VT->getName()).str();
if (Record *VT = P.getScalarMemoryVT())
Explanation += (" ScalarVT(MemVT)=" + VT->getName()).str();
[globalisel][tablegen] Add support for importing G_ATOMIC_CMPXCHG, G_ATOMICRMW_* rules from SelectionDAG. GIM_CheckNonAtomic has been replaced by GIM_CheckAtomicOrdering to allow it to support a wider range of orderings. This has then been used to import patterns using nodes such as atomic_cmp_swap, atomic_swap, and atomic_load_*. llvm-svn: 319232
2017-11-29 06:07:05 +08:00
if (P.isAtomicOrderingMonotonic())
Explanation += " monotonic";
if (P.isAtomicOrderingAcquire())
Explanation += " acquire";
if (P.isAtomicOrderingRelease())
Explanation += " release";
if (P.isAtomicOrderingAcquireRelease())
Explanation += " acq_rel";
if (P.isAtomicOrderingSequentiallyConsistent())
Explanation += " seq_cst";
[globalisel][tablegen] Add support for relative AtomicOrderings No test yet because the relevant rules are blocked on the atomic_load, and atomic_store nodes. llvm-svn: 319475
2017-12-01 05:05:59 +08:00
if (P.isAtomicOrderingAcquireOrStronger())
Explanation += " >=acquire";
if (P.isAtomicOrderingWeakerThanAcquire())
Explanation += " <acquire";
if (P.isAtomicOrderingReleaseOrStronger())
Explanation += " >=release";
if (P.isAtomicOrderingWeakerThanRelease())
Explanation += " <release";
[globalisel][tablegen] Report more detail in some SelectionDAG import failures. NFC Reviewers: ab, t.p.northover, qcolombet, aditya_nandakumar, rovka Reviewed By: ab Subscribers: dberris, kristof.beyls, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D31325 llvm-svn: 300186
2017-04-13 17:45:37 +08:00
}
return Explanation;
}
std::string explainOperator(Record *Operator) {
if (Operator->isSubClassOf("SDNode"))
[TableGen] Make Record::getValueAsString and getValueAsListOfStrings return StringRefs instead of std::string Internally both these methods just return the result of getValue on either a StringInit or a CodeInit object. In both cases this returns a StringRef pointing to a string allocated in the BumpPtrAllocator so its not going anywhere. So we can just pass that StringRef along. This is a fairly naive patch that targets just the build failures caused by this change. There's additional work that can be done to avoid creating std::string at call sites that still think getValueAsString returns a std::string. I'll try to clean those up in future patches. Differential Revision: https://reviews.llvm.org/D33710 llvm-svn: 304325
2017-06-01 03:01:11 +08:00
return (" (" + Operator->getValueAsString("Opcode") + ")").str();
[globalisel][tablegen] Report more detail in some SelectionDAG import failures. NFC Reviewers: ab, t.p.northover, qcolombet, aditya_nandakumar, rovka Reviewed By: ab Subscribers: dberris, kristof.beyls, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D31325 llvm-svn: 300186
2017-04-13 17:45:37 +08:00
if (Operator->isSubClassOf("Intrinsic"))
return (" (Operator is an Intrinsic, " + Operator->getName() + ")").str();
Re-commit r315863: [globalisel][tablegen] Import ComplexPattern when used as an operator Summary: It's possible for a ComplexPattern to be used as an operator in a match pattern. This is used by the load/store patterns in AArch64 to name the suboperands returned by ComplexPattern predicate so that they can be broken apart and referenced independently in the result pattern. This patch adds support for this in order to enable the import of load/store patterns. Depends on D37445 Hopefully fixed the ambiguous constructor that a large number of bots reported. Reviewers: ab, qcolombet, t.p.northover, rovka, aditya_nandakumar Reviewed By: qcolombet Subscribers: aemerson, javed.absar, igorb, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D37456 llvm-svn: 315869
2017-10-16 02:22:54 +08:00
if (Operator->isSubClassOf("ComplexPattern"))
return (" (Operator is an unmapped ComplexPattern, " + Operator->getName() +
")")
.str();
[GlobalISel][TableGen] Add support for SDNodeXForm Summary: This patch adds CustomRenderer which renders the matched operands to the specified instruction. Targets can enable the matching of SDNodeXForm by adding a definition that inherits from GICustomOperandRenderer and GISDNodeXFormEquiv as follows. def gi_imm8 : GICustomOperandRenderer<"renderImm8”>, GISDNodeXFormEquiv<imm8_xform>; Custom renderer functions should be of the form: void render(MachineInstrBuilder &MIB, const MachineInstr &I); Reviewers: dsanders, ab, rovka Reviewed By: dsanders Subscribers: kristof.beyls, javed.absar, llvm-commits, mgrang, qcolombet Differential Revision: https://reviews.llvm.org/D42012 llvm-svn: 322582
2018-01-17 02:44:05 +08:00
if (Operator->isSubClassOf("SDNodeXForm"))
return (" (Operator is an unmapped SDNodeXForm, " + Operator->getName() +
")")
.str();
Re-commit r315863: [globalisel][tablegen] Import ComplexPattern when used as an operator Summary: It's possible for a ComplexPattern to be used as an operator in a match pattern. This is used by the load/store patterns in AArch64 to name the suboperands returned by ComplexPattern predicate so that they can be broken apart and referenced independently in the result pattern. This patch adds support for this in order to enable the import of load/store patterns. Depends on D37445 Hopefully fixed the ambiguous constructor that a large number of bots reported. Reviewers: ab, qcolombet, t.p.northover, rovka, aditya_nandakumar Reviewed By: qcolombet Subscribers: aemerson, javed.absar, igorb, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D37456 llvm-svn: 315869
2017-10-16 02:22:54 +08:00
return (" (Operator " + Operator->getName() + " not understood)").str();
[globalisel][tablegen] Report more detail in some SelectionDAG import failures. NFC Reviewers: ab, t.p.northover, qcolombet, aditya_nandakumar, rovka Reviewed By: ab Subscribers: dberris, kristof.beyls, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D31325 llvm-svn: 300186
2017-04-13 17:45:37 +08:00
}
/// Helper function to let the emitter report skip reason error messages.
static Error failedImport(const Twine &Reason) {
return make_error<StringError>(Reason, inconvertibleErrorCode());
}
static Error isTrivialOperatorNode(const TreePatternNode *N) {
std::string Explanation = "";
std::string Separator = "";
Re-commit: [globalisel][tablegen] Add support for ImmLeaf without SDNodeXForm Summary: This patch adds support for predicates on imm nodes but only for ImmLeaf and not for PatLeaf or PatFrag and only where the value does not need to be transformed before being rendered into the instruction. The limitation on PatLeaf/PatFrag/SDNodeXForm is due to differences in the necessary target-supplied C++ for GlobalISel. Depends on D36085 The previous commit was reverted for breaking the build but this appears to have been the recurring problem on the Windows bots with tablegen not being re-run when llvm-tblgen is changed but the .td's aren't. If it re-occurs then forcing a build with clean=True should fix it but this string should do this in advance: Requires a clean build. Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Reviewed By: rovka Subscribers: kristof.beyls, javed.absar, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D36086 llvm-svn: 311645
2017-08-24 17:11:20 +08:00
bool HasUnsupportedPredicate = false;
for (const auto &Predicate : N->getPredicateFns()) {
if (Predicate.isAlwaysTrue())
continue;
if (Predicate.isImmediatePattern())
continue;
[globalisel] Update GlobalISel emitter to match new representation of extending loads Summary: Previously, a extending load was represented at (G_*EXT (G_LOAD x)). This had a few drawbacks: * G_LOAD had to be legal for all sizes you could extend from, even if registers didn't naturally hold those sizes. * All sizes you could extend from had to be allocatable just in case the extend went missing (e.g. by optimization). * At minimum, G_*EXT and G_TRUNC had to be legal for these sizes. As we improve optimization of extends and truncates, this legality requirement would spread without considerable care w.r.t when certain combines were permitted. * The SelectionDAG importer required some ugly and fragile pattern rewriting to translate patterns into this style. This patch changes the representation to: * (G_[SZ]EXTLOAD x) * (G_LOAD x) any-extends when MMO.getSize() * 8 < ResultTy.getSizeInBits() which resolves these issues by allowing targets to work entirely in their native register sizes, and by having a more direct translation from SelectionDAG patterns. Each extending load can be lowered by the legalizer into separate extends and loads, however a target that supports s1 will need the any-extending load to extend to at least s8 since LLVM does not represent memory accesses smaller than 8 bit. The legalizer can widenScalar G_LOAD into an any-extending load but sign/zero-extending loads need help from something else like a combiner pass. A follow-up patch that adds combiner helpers for for this will follow. The new representation requires that the MMO correctly reflect the memory access so this has been corrected in a couple tests. I've also moved the extending loads to their own tests since they are (mostly) separate opcodes now. Additionally, the re-write appears to have invalidated two tests from select-with-no-legality-check.mir since the matcher table no longer contains loads that result in s1's and they aren't legal in AArch64 anymore. Depends on D45540 Reviewers: ab, aditya_nandakumar, bogner, rtereshin, volkan, rovka, javed.absar Reviewed By: rtereshin Subscribers: javed.absar, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D45541 llvm-svn: 331601
2018-05-06 04:53:24 +08:00
if (Predicate.isNonExtLoad() || Predicate.isAnyExtLoad() ||
Predicate.isSignExtLoad() || Predicate.isZeroExtLoad())
[globalisel][tablegen] Implement unindexed load, non-extending load, and MemVT checks Summary: This includes some context-sensitivity in the MVT to LLT conversion so that pointer types are tested correctly. FIXME: I'm not happy with the way this is done since everything is a special-case. I've yet to find a reasonable way to implement it. select-load.mir fails because <1 x s64> loads in tablegen get priority over s64 loads. This is fixed in the next patch and as such they should be committed together, I've posted them separately to help with the review. Depends on D37456 Reviewers: ab, qcolombet, t.p.northover, rovka, aditya_nandakumar Subscribers: kristof.beyls, javed.absar, llvm-commits, igorb Differential Revision: https://reviews.llvm.org/D37457 llvm-svn: 315884
2017-10-16 08:56:30 +08:00
continue;
[globalisel][tablegen] Add support for importing G_ATOMIC_CMPXCHG, G_ATOMICRMW_* rules from SelectionDAG. GIM_CheckNonAtomic has been replaced by GIM_CheckAtomicOrdering to allow it to support a wider range of orderings. This has then been used to import patterns using nodes such as atomic_cmp_swap, atomic_swap, and atomic_load_*. llvm-svn: 319232
2017-11-29 06:07:05 +08:00
if (Predicate.isNonTruncStore())
[globalisel][tablegen] Implement unindexed load, non-extending load, and MemVT checks Summary: This includes some context-sensitivity in the MVT to LLT conversion so that pointer types are tested correctly. FIXME: I'm not happy with the way this is done since everything is a special-case. I've yet to find a reasonable way to implement it. select-load.mir fails because <1 x s64> loads in tablegen get priority over s64 loads. This is fixed in the next patch and as such they should be committed together, I've posted them separately to help with the review. Depends on D37456 Reviewers: ab, qcolombet, t.p.northover, rovka, aditya_nandakumar Subscribers: kristof.beyls, javed.absar, llvm-commits, igorb Differential Revision: https://reviews.llvm.org/D37457 llvm-svn: 315884
2017-10-16 08:56:30 +08:00
continue;
[globalisel][tablegen] Import stores and allow GISel to automatically substitute zero regs like WZR/XZR/$zero. This patch enables the import of stores. Unfortunately, doing so by itself, loses an optimization where storing 0 to memory makes use of WZR/XZR. To mitigate this, this patch also introduces a new feature that allows register operands to nominate a zero register. When this is done, GlobalISel will substitute (G_CONSTANT 0) with the nominated register automatically. This is currently configured to only apply to the stores. Applying it to GPR32/GPR64 register classes in general will be done after review see (https://reviews.llvm.org/D39150). llvm-svn: 316360
2017-10-24 02:19:24 +08:00
[globalisel] Update GlobalISel emitter to match new representation of extending loads Summary: Previously, a extending load was represented at (G_*EXT (G_LOAD x)). This had a few drawbacks: * G_LOAD had to be legal for all sizes you could extend from, even if registers didn't naturally hold those sizes. * All sizes you could extend from had to be allocatable just in case the extend went missing (e.g. by optimization). * At minimum, G_*EXT and G_TRUNC had to be legal for these sizes. As we improve optimization of extends and truncates, this legality requirement would spread without considerable care w.r.t when certain combines were permitted. * The SelectionDAG importer required some ugly and fragile pattern rewriting to translate patterns into this style. This patch changes the representation to: * (G_[SZ]EXTLOAD x) * (G_LOAD x) any-extends when MMO.getSize() * 8 < ResultTy.getSizeInBits() which resolves these issues by allowing targets to work entirely in their native register sizes, and by having a more direct translation from SelectionDAG patterns. Each extending load can be lowered by the legalizer into separate extends and loads, however a target that supports s1 will need the any-extending load to extend to at least s8 since LLVM does not represent memory accesses smaller than 8 bit. The legalizer can widenScalar G_LOAD into an any-extending load but sign/zero-extending loads need help from something else like a combiner pass. A follow-up patch that adds combiner helpers for for this will follow. The new representation requires that the MMO correctly reflect the memory access so this has been corrected in a couple tests. I've also moved the extending loads to their own tests since they are (mostly) separate opcodes now. Additionally, the re-write appears to have invalidated two tests from select-with-no-legality-check.mir since the matcher table no longer contains loads that result in s1's and they aren't legal in AArch64 anymore. Depends on D45540 Reviewers: ab, aditya_nandakumar, bogner, rtereshin, volkan, rovka, javed.absar Reviewed By: rtereshin Subscribers: javed.absar, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D45541 llvm-svn: 331601
2018-05-06 04:53:24 +08:00
if (Predicate.isLoad() && Predicate.getMemoryVT())
continue;
[globalisel][tablegen] Add support for importing G_ATOMIC_CMPXCHG, G_ATOMICRMW_* rules from SelectionDAG. GIM_CheckNonAtomic has been replaced by GIM_CheckAtomicOrdering to allow it to support a wider range of orderings. This has then been used to import patterns using nodes such as atomic_cmp_swap, atomic_swap, and atomic_load_*. llvm-svn: 319232
2017-11-29 06:07:05 +08:00
if (Predicate.isLoad() || Predicate.isStore()) {
if (Predicate.isUnindexed())
continue;
}
if (Predicate.isAtomic() && Predicate.getMemoryVT())
[globalisel][tablegen] Import stores and allow GISel to automatically substitute zero regs like WZR/XZR/$zero. This patch enables the import of stores. Unfortunately, doing so by itself, loses an optimization where storing 0 to memory makes use of WZR/XZR. To mitigate this, this patch also introduces a new feature that allows register operands to nominate a zero register. When this is done, GlobalISel will substitute (G_CONSTANT 0) with the nominated register automatically. This is currently configured to only apply to the stores. Applying it to GPR32/GPR64 register classes in general will be done after review see (https://reviews.llvm.org/D39150). llvm-svn: 316360
2017-10-24 02:19:24 +08:00
continue;
[globalisel][tablegen] Add support for importing G_ATOMIC_CMPXCHG, G_ATOMICRMW_* rules from SelectionDAG. GIM_CheckNonAtomic has been replaced by GIM_CheckAtomicOrdering to allow it to support a wider range of orderings. This has then been used to import patterns using nodes such as atomic_cmp_swap, atomic_swap, and atomic_load_*. llvm-svn: 319232
2017-11-29 06:07:05 +08:00
if (Predicate.isAtomic() &&
(Predicate.isAtomicOrderingMonotonic() ||
Predicate.isAtomicOrderingAcquire() ||
Predicate.isAtomicOrderingRelease() ||
Predicate.isAtomicOrderingAcquireRelease() ||
[globalisel][tablegen] Add support for relative AtomicOrderings No test yet because the relevant rules are blocked on the atomic_load, and atomic_store nodes. llvm-svn: 319475
2017-12-01 05:05:59 +08:00
Predicate.isAtomicOrderingSequentiallyConsistent() ||
Predicate.isAtomicOrderingAcquireOrStronger() ||
Predicate.isAtomicOrderingWeakerThanAcquire() ||
Predicate.isAtomicOrderingReleaseOrStronger() ||
Predicate.isAtomicOrderingWeakerThanRelease()))
[globalisel][tablegen] Import stores and allow GISel to automatically substitute zero regs like WZR/XZR/$zero. This patch enables the import of stores. Unfortunately, doing so by itself, loses an optimization where storing 0 to memory makes use of WZR/XZR. To mitigate this, this patch also introduces a new feature that allows register operands to nominate a zero register. When this is done, GlobalISel will substitute (G_CONSTANT 0) with the nominated register automatically. This is currently configured to only apply to the stores. Applying it to GPR32/GPR64 register classes in general will be done after review see (https://reviews.llvm.org/D39150). llvm-svn: 316360
2017-10-24 02:19:24 +08:00
continue;
Re-commit: [globalisel][tablegen] Add support for ImmLeaf without SDNodeXForm Summary: This patch adds support for predicates on imm nodes but only for ImmLeaf and not for PatLeaf or PatFrag and only where the value does not need to be transformed before being rendered into the instruction. The limitation on PatLeaf/PatFrag/SDNodeXForm is due to differences in the necessary target-supplied C++ for GlobalISel. Depends on D36085 The previous commit was reverted for breaking the build but this appears to have been the recurring problem on the Windows bots with tablegen not being re-run when llvm-tblgen is changed but the .td's aren't. If it re-occurs then forcing a build with clean=True should fix it but this string should do this in advance: Requires a clean build. Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Reviewed By: rovka Subscribers: kristof.beyls, javed.absar, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D36086 llvm-svn: 311645
2017-08-24 17:11:20 +08:00
HasUnsupportedPredicate = true;
[globalisel][tablegen] Report more detail in some SelectionDAG import failures. NFC Reviewers: ab, t.p.northover, qcolombet, aditya_nandakumar, rovka Reviewed By: ab Subscribers: dberris, kristof.beyls, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D31325 llvm-svn: 300186
2017-04-13 17:45:37 +08:00
Explanation = Separator + "Has a predicate (" + explainPredicates(N) + ")";
Separator = ", ";
[tablegen] Handle common load/store predicates inside tablegen. NFC. Summary: GlobalISel and SelectionDAG require different code for the common load/store predicates due to differences in the representation. For example: SelectionDAG: (load<signext,i8>:i32 GPR32:$addr) // The <> denote properties of the SDNode that are not printed in the DAG GlobalISel: (G_SEXT:s32 (G_LOAD:s8 GPR32:$addr)) Even without that, differences in the IR (SDNode vs MachineInstr) require differences in the C++ predicate. This patch moves the implementation of the common load/store predicates into tablegen so that it can handle these differences. It's NFC for SelectionDAG since it emits equivalent code and it's NFC for GlobalISel since the rules involving the relevant predicates are still rejected by the importer. Depends on D36618 Reviewers: ab, qcolombet, t.p.northover, rovka, aditya_nandakumar Subscribers: llvm-commits, igorb Differential Revision: https://reviews.llvm.org/D37443 Includes a partial revert of r315826 since this patch makes it necessary for getPredCode() to return a std::string and getImmCode() should have the same interface as getPredCode(). llvm-svn: 315841
2017-10-15 10:06:44 +08:00
Explanation += (Separator + "first-failing:" +
Predicate.getOrigPatFragRecord()->getRecord()->getName())
.str();
Re-commit: [globalisel][tablegen] Add support for ImmLeaf without SDNodeXForm Summary: This patch adds support for predicates on imm nodes but only for ImmLeaf and not for PatLeaf or PatFrag and only where the value does not need to be transformed before being rendered into the instruction. The limitation on PatLeaf/PatFrag/SDNodeXForm is due to differences in the necessary target-supplied C++ for GlobalISel. Depends on D36085 The previous commit was reverted for breaking the build but this appears to have been the recurring problem on the Windows bots with tablegen not being re-run when llvm-tblgen is changed but the .td's aren't. If it re-occurs then forcing a build with clean=True should fix it but this string should do this in advance: Requires a clean build. Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Reviewed By: rovka Subscribers: kristof.beyls, javed.absar, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D36086 llvm-svn: 311645
2017-08-24 17:11:20 +08:00
break;
[globalisel][tablegen] Report more detail in some SelectionDAG import failures. NFC Reviewers: ab, t.p.northover, qcolombet, aditya_nandakumar, rovka Reviewed By: ab Subscribers: dberris, kristof.beyls, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D31325 llvm-svn: 300186
2017-04-13 17:45:37 +08:00
}
[GlobalISel][TableGen] Add support for SDNodeXForm Summary: This patch adds CustomRenderer which renders the matched operands to the specified instruction. Targets can enable the matching of SDNodeXForm by adding a definition that inherits from GICustomOperandRenderer and GISDNodeXFormEquiv as follows. def gi_imm8 : GICustomOperandRenderer<"renderImm8”>, GISDNodeXFormEquiv<imm8_xform>; Custom renderer functions should be of the form: void render(MachineInstrBuilder &MIB, const MachineInstr &I); Reviewers: dsanders, ab, rovka Reviewed By: dsanders Subscribers: kristof.beyls, javed.absar, llvm-commits, mgrang, qcolombet Differential Revision: https://reviews.llvm.org/D42012 llvm-svn: 322582
2018-01-17 02:44:05 +08:00
if (!HasUnsupportedPredicate)
[globalisel][tablegen] Report more detail in some SelectionDAG import failures. NFC Reviewers: ab, t.p.northover, qcolombet, aditya_nandakumar, rovka Reviewed By: ab Subscribers: dberris, kristof.beyls, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D31325 llvm-svn: 300186
2017-04-13 17:45:37 +08:00
return Error::success();
return failedImport(Explanation);
[GlobalISel] Add basic Selector-emitter tblgen backend. This adds a basic tablegen backend that analyzes the SelectionDAG patterns to find simple ones that are eligible for GlobalISel-emission. That's similar to FastISel, with one notable difference: we're not fed ISD opcodes, so we need to map the SDNode operators to generic opcodes. That's done using GINodeEquiv in TargetGlobalISel.td. Otherwise, this is mostly boilerplate, and lots of filtering of any kind of "complicated" pattern. On AArch64, this is sufficient to match G_ADD up to s64 (to ADDWrr/ADDXrr) and G_BR (to B). Differential Revision: https://reviews.llvm.org/D26878 llvm-svn: 290284
2016-12-22 07:26:20 +08:00
}
[globalisel][tablegen] Add support for COPY_TO_REGCLASS. Summary: As part of this * Emitted instructions now have named MachineInstr variables associated with them. This isn't particularly important yet but it's a small step towards multiple-insn emission. * constrainSelectedInstRegOperands() is no longer hardcoded. It's now added as the ConstrainOperandsToDefinitionAction() action. COPY_TO_REGCLASS uses an alternate constraint mechanism ConstrainOperandToRegClassAction() which supports arbitrary constraints such as that defined by COPY_TO_REGCLASS. Reviewers: ab, qcolombet, t.p.northover, rovka, kristof.beyls, aditya_nandakumar Reviewed By: ab Subscribers: javed.absar, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D33590 llvm-svn: 305791
2017-06-20 20:36:34 +08:00
static Record *getInitValueAsRegClass(Init *V) {
if (DefInit *VDefInit = dyn_cast<DefInit>(V)) {
if (VDefInit->getDef()->isSubClassOf("RegisterOperand"))
return VDefInit->getDef()->getValueAsDef("RegClass");
if (VDefInit->getDef()->isSubClassOf("RegisterClass"))
return VDefInit->getDef();
}
return nullptr;
}
[globalisel][tablegen] Partially fix compile-time regressions by converting matcher to state-machine(s) Summary: Replace the matcher if-statements for each rule with a state-machine. This significantly reduces compile time, memory allocations, and cumulative memory allocation when compiling AArch64InstructionSelector.cpp.o after r303259 is recommitted. The following patches will expand on this further to fully fix the regressions. Reviewers: rovka, ab, t.p.northover, qcolombet, aditya_nandakumar Reviewed By: ab Subscribers: vitalybuka, aemerson, javed.absar, igorb, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D33758 llvm-svn: 307079
2017-07-04 22:35:06 +08:00
std::string
getNameForFeatureBitset(const std::vector<Record *> &FeatureBitset) {
std::string Name = "GIFBS";
for (const auto &Feature : FeatureBitset)
Name += ("_" + Feature->getName()).str();
return Name;
}
[globalisel][tablegen] Add control-flow to the MatchTable. Summary: This will allow us to merge the various sub-tables into a single table. This is a compile-time saving at this point. However, this will also enable the optimization of a table so that similar instructions can be tested together, reducing the time spent on the matching the code. The bulk of this patch is a mechanical conversion to the new MatchTable object which is responsible for tracking label definitions and filling in the index of the jump targets. It is also responsible for nicely formatting the table. This was necessary to support the new GIM_Try opcode which takes the index to jump to if the match should fail. This value is unknown during table construction and is filled in during emission. To support nesting try-blocks (although we currently don't emit tables with nested try-blocks), GIM_Reject has been re-introduced to explicitly exit a try-block or fail the overall match if there are no active try-blocks. Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Reviewed By: rovka Subscribers: kristof.beyls, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D35117 llvm-svn: 308596
2017-07-20 17:25:44 +08:00
//===- MatchTable Helpers -------------------------------------------------===//
class MatchTable;
/// A record to be stored in a MatchTable.
///
/// This class represents any and all output that may be required to emit the
/// MatchTable. Instances are most often configured to represent an opcode or
/// value that will be emitted to the table with some formatting but it can also
/// represent commas, comments, and other formatting instructions.
struct MatchTableRecord {
enum RecordFlagsBits {
MTRF_None = 0x0,
/// Causes EmitStr to be formatted as comment when emitted.
MTRF_Comment = 0x1,
/// Causes the record value to be followed by a comma when emitted.
MTRF_CommaFollows = 0x2,
/// Causes the record value to be followed by a line break when emitted.
MTRF_LineBreakFollows = 0x4,
/// Indicates that the record defines a label and causes an additional
/// comment to be emitted containing the index of the label.
MTRF_Label = 0x8,
/// Causes the record to be emitted as the index of the label specified by
/// LabelID along with a comment indicating where that label is.
MTRF_JumpTarget = 0x10,
/// Causes the formatter to add a level of indentation before emitting the
/// record.
MTRF_Indent = 0x20,
/// Causes the formatter to remove a level of indentation after emitting the
/// record.
MTRF_Outdent = 0x40,
};
/// When MTRF_Label or MTRF_JumpTarget is used, indicates a label id to
/// reference or define.
unsigned LabelID;
/// The string to emit. Depending on the MTRF_* flags it may be a comment, a
/// value, a label name.
std::string EmitStr;
private:
/// The number of MatchTable elements described by this record. Comments are 0
/// while values are typically 1. Values >1 may occur when we need to emit
/// values that exceed the size of a MatchTable element.
unsigned NumElements;
public:
/// A bitfield of RecordFlagsBits flags.
unsigned Flags;
Reapply r332907 "[GlobalISel] Improving InstructionSelect's performance by reducing MatchTable..." Apparently the compile time problem was caused by the fact that not all compilers / STL implementations can automatically convert std::unique_ptr<Derived> to std::unique_ptr<Base>. Fixed (hopefully) by making sure it's std::unique_ptr<Derived>&& (rvalue ref) to std::unique_ptr<Base> conversion instead. llvm-svn: 332917
2018-05-22 07:28:51 +08:00
/// The actual run-time value, if known
int64_t RawValue;
[globalisel][tablegen] Add control-flow to the MatchTable. Summary: This will allow us to merge the various sub-tables into a single table. This is a compile-time saving at this point. However, this will also enable the optimization of a table so that similar instructions can be tested together, reducing the time spent on the matching the code. The bulk of this patch is a mechanical conversion to the new MatchTable object which is responsible for tracking label definitions and filling in the index of the jump targets. It is also responsible for nicely formatting the table. This was necessary to support the new GIM_Try opcode which takes the index to jump to if the match should fail. This value is unknown during table construction and is filled in during emission. To support nesting try-blocks (although we currently don't emit tables with nested try-blocks), GIM_Reject has been re-introduced to explicitly exit a try-block or fail the overall match if there are no active try-blocks. Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Reviewed By: rovka Subscribers: kristof.beyls, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D35117 llvm-svn: 308596
2017-07-20 17:25:44 +08:00
MatchTableRecord(Optional<unsigned> LabelID_, StringRef EmitStr,
Reapply r332907 "[GlobalISel] Improving InstructionSelect's performance by reducing MatchTable..." Apparently the compile time problem was caused by the fact that not all compilers / STL implementations can automatically convert std::unique_ptr<Derived> to std::unique_ptr<Base>. Fixed (hopefully) by making sure it's std::unique_ptr<Derived>&& (rvalue ref) to std::unique_ptr<Base> conversion instead. llvm-svn: 332917
2018-05-22 07:28:51 +08:00
unsigned NumElements, unsigned Flags,
int64_t RawValue = std::numeric_limits<int64_t>::min())
[globalisel][tablegen] Add control-flow to the MatchTable. Summary: This will allow us to merge the various sub-tables into a single table. This is a compile-time saving at this point. However, this will also enable the optimization of a table so that similar instructions can be tested together, reducing the time spent on the matching the code. The bulk of this patch is a mechanical conversion to the new MatchTable object which is responsible for tracking label definitions and filling in the index of the jump targets. It is also responsible for nicely formatting the table. This was necessary to support the new GIM_Try opcode which takes the index to jump to if the match should fail. This value is unknown during table construction and is filled in during emission. To support nesting try-blocks (although we currently don't emit tables with nested try-blocks), GIM_Reject has been re-introduced to explicitly exit a try-block or fail the overall match if there are no active try-blocks. Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Reviewed By: rovka Subscribers: kristof.beyls, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D35117 llvm-svn: 308596
2017-07-20 17:25:44 +08:00
: LabelID(LabelID_.hasValue() ? LabelID_.getValue() : ~0u),
Reapply r332907 "[GlobalISel] Improving InstructionSelect's performance by reducing MatchTable..." Apparently the compile time problem was caused by the fact that not all compilers / STL implementations can automatically convert std::unique_ptr<Derived> to std::unique_ptr<Base>. Fixed (hopefully) by making sure it's std::unique_ptr<Derived>&& (rvalue ref) to std::unique_ptr<Base> conversion instead. llvm-svn: 332917
2018-05-22 07:28:51 +08:00
EmitStr(EmitStr), NumElements(NumElements), Flags(Flags),
RawValue(RawValue) {
[globalisel][tablegen] Add control-flow to the MatchTable. Summary: This will allow us to merge the various sub-tables into a single table. This is a compile-time saving at this point. However, this will also enable the optimization of a table so that similar instructions can be tested together, reducing the time spent on the matching the code. The bulk of this patch is a mechanical conversion to the new MatchTable object which is responsible for tracking label definitions and filling in the index of the jump targets. It is also responsible for nicely formatting the table. This was necessary to support the new GIM_Try opcode which takes the index to jump to if the match should fail. This value is unknown during table construction and is filled in during emission. To support nesting try-blocks (although we currently don't emit tables with nested try-blocks), GIM_Reject has been re-introduced to explicitly exit a try-block or fail the overall match if there are no active try-blocks. Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Reviewed By: rovka Subscribers: kristof.beyls, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D35117 llvm-svn: 308596
2017-07-20 17:25:44 +08:00
assert((!LabelID_.hasValue() || LabelID != ~0u) &&
"This value is reserved for non-labels");
}
Reapply r332907 "[GlobalISel] Improving InstructionSelect's performance by reducing MatchTable..." Apparently the compile time problem was caused by the fact that not all compilers / STL implementations can automatically convert std::unique_ptr<Derived> to std::unique_ptr<Base>. Fixed (hopefully) by making sure it's std::unique_ptr<Derived>&& (rvalue ref) to std::unique_ptr<Base> conversion instead. llvm-svn: 332917
2018-05-22 07:28:51 +08:00
MatchTableRecord(const MatchTableRecord &Other) = default;
MatchTableRecord(MatchTableRecord &&Other) = default;
/// Useful if a Match Table Record gets optimized out
void turnIntoComment() {
Flags |= MTRF_Comment;
Flags &= ~MTRF_CommaFollows;
NumElements = 0;
}
/// For Jump Table generation purposes
bool operator<(const MatchTableRecord &Other) const {
return RawValue < Other.RawValue;
}
int64_t getRawValue() const { return RawValue; }
[globalisel][tablegen] Add control-flow to the MatchTable. Summary: This will allow us to merge the various sub-tables into a single table. This is a compile-time saving at this point. However, this will also enable the optimization of a table so that similar instructions can be tested together, reducing the time spent on the matching the code. The bulk of this patch is a mechanical conversion to the new MatchTable object which is responsible for tracking label definitions and filling in the index of the jump targets. It is also responsible for nicely formatting the table. This was necessary to support the new GIM_Try opcode which takes the index to jump to if the match should fail. This value is unknown during table construction and is filled in during emission. To support nesting try-blocks (although we currently don't emit tables with nested try-blocks), GIM_Reject has been re-introduced to explicitly exit a try-block or fail the overall match if there are no active try-blocks. Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Reviewed By: rovka Subscribers: kristof.beyls, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D35117 llvm-svn: 308596
2017-07-20 17:25:44 +08:00
void emit(raw_ostream &OS, bool LineBreakNextAfterThis,
const MatchTable &Table) const;
unsigned size() const { return NumElements; }
};
[GlobalISel][InstructionSelect] Refactoring buildMatchTable out, NFC to share it between the Instruction Selector in optimized and non-optimized modes both and the Testgen. Reviewers: dsanders, aemerson Reviewed By: dsanders Subscribers: rovka, kristof.beyls, llvm-commits, dsanders Differential Revision: https://reviews.llvm.org/D46097 llvm-svn: 331396
2018-05-03 04:08:14 +08:00
class Matcher;
[globalisel][tablegen] Add control-flow to the MatchTable. Summary: This will allow us to merge the various sub-tables into a single table. This is a compile-time saving at this point. However, this will also enable the optimization of a table so that similar instructions can be tested together, reducing the time spent on the matching the code. The bulk of this patch is a mechanical conversion to the new MatchTable object which is responsible for tracking label definitions and filling in the index of the jump targets. It is also responsible for nicely formatting the table. This was necessary to support the new GIM_Try opcode which takes the index to jump to if the match should fail. This value is unknown during table construction and is filled in during emission. To support nesting try-blocks (although we currently don't emit tables with nested try-blocks), GIM_Reject has been re-introduced to explicitly exit a try-block or fail the overall match if there are no active try-blocks. Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Reviewed By: rovka Subscribers: kristof.beyls, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D35117 llvm-svn: 308596
2017-07-20 17:25:44 +08:00
/// Holds the contents of a generated MatchTable to enable formatting and the
/// necessary index tracking needed to support GIM_Try.
class MatchTable {
/// An unique identifier for the table. The generated table will be named
/// MatchTable${ID}.
unsigned ID;
/// The records that make up the table. Also includes comments describing the
/// values being emitted and line breaks to format it.
std::vector<MatchTableRecord> Contents;
/// The currently defined labels.
DenseMap<unsigned, unsigned> LabelMap;
/// Tracks the sum of MatchTableRecord::NumElements as the table is built.
[GlobalISel][InstructionSelect] Refactoring buildMatchTable out, NFC to share it between the Instruction Selector in optimized and non-optimized modes both and the Testgen. Reviewers: dsanders, aemerson Reviewed By: dsanders Subscribers: rovka, kristof.beyls, llvm-commits, dsanders Differential Revision: https://reviews.llvm.org/D46097 llvm-svn: 331396
2018-05-03 04:08:14 +08:00
unsigned CurrentSize = 0;
Re-commit: r309094 [globalisel][tablegen] Fuse the generated tables together. Summary: Now that we have control flow in place, fuse the per-rule tables into a single table. This is a compile-time saving at this point. However, this will also enable the optimization of a table so that similar instructions can be tested together, reducing the time spent on the matching the code. This is NFC in terms of externally visible behaviour but some internals have changed slightly. State.MIs is no longer reset between each rule that is attempted because it's not necessary to do so. As a consequence of this the restriction on the order that instructions are added to State.MIs has been relaxed to only affect recorded instructions that require new elements to be added to the vector. GIM_RecordInsn can now write to any element from 1 to State.MIs.size() instead of just State.MIs.size(). The compile-time regressions from the last commit were caused by the ARM target including a non-const variable (zero_reg) in the table and therefore generating an initializer for it. That variable is now const. Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Reviewed By: rovka Subscribers: kristof.beyls, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D35681 llvm-svn: 309264
2017-07-27 19:03:45 +08:00
/// A unique identifier for a MatchTable label.
[GlobalISel][InstructionSelect] Refactoring buildMatchTable out, NFC to share it between the Instruction Selector in optimized and non-optimized modes both and the Testgen. Reviewers: dsanders, aemerson Reviewed By: dsanders Subscribers: rovka, kristof.beyls, llvm-commits, dsanders Differential Revision: https://reviews.llvm.org/D46097 llvm-svn: 331396
2018-05-03 04:08:14 +08:00
unsigned CurrentLabelID = 0;
[GlobalISel][InstructionSelect] Making Coverage Info generation optional on per-match table basis to make sure that Testgen always has access to coverage info even if the match table used by the selector itself is stripped off that information for performance reasons. Reviewers: dsanders, aemerson Reviewed By: dsanders Subscribers: rovka, kristof.beyls, llvm-commits, dsanders Differential Revision: https://reviews.llvm.org/D46098 llvm-svn: 331398
2018-05-03 04:15:11 +08:00
/// Determines if the table should be instrumented for rule coverage tracking.
bool IsWithCoverage;
Re-commit: r309094 [globalisel][tablegen] Fuse the generated tables together. Summary: Now that we have control flow in place, fuse the per-rule tables into a single table. This is a compile-time saving at this point. However, this will also enable the optimization of a table so that similar instructions can be tested together, reducing the time spent on the matching the code. This is NFC in terms of externally visible behaviour but some internals have changed slightly. State.MIs is no longer reset between each rule that is attempted because it's not necessary to do so. As a consequence of this the restriction on the order that instructions are added to State.MIs has been relaxed to only affect recorded instructions that require new elements to be added to the vector. GIM_RecordInsn can now write to any element from 1 to State.MIs.size() instead of just State.MIs.size(). The compile-time regressions from the last commit were caused by the ARM target including a non-const variable (zero_reg) in the table and therefore generating an initializer for it. That variable is now const. Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Reviewed By: rovka Subscribers: kristof.beyls, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D35681 llvm-svn: 309264
2017-07-27 19:03:45 +08:00
[globalisel][tablegen] Add control-flow to the MatchTable. Summary: This will allow us to merge the various sub-tables into a single table. This is a compile-time saving at this point. However, this will also enable the optimization of a table so that similar instructions can be tested together, reducing the time spent on the matching the code. The bulk of this patch is a mechanical conversion to the new MatchTable object which is responsible for tracking label definitions and filling in the index of the jump targets. It is also responsible for nicely formatting the table. This was necessary to support the new GIM_Try opcode which takes the index to jump to if the match should fail. This value is unknown during table construction and is filled in during emission. To support nesting try-blocks (although we currently don't emit tables with nested try-blocks), GIM_Reject has been re-introduced to explicitly exit a try-block or fail the overall match if there are no active try-blocks. Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Reviewed By: rovka Subscribers: kristof.beyls, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D35117 llvm-svn: 308596
2017-07-20 17:25:44 +08:00
public:
static MatchTableRecord LineBreak;
static MatchTableRecord Comment(StringRef Comment) {
return MatchTableRecord(None, Comment, 0, MatchTableRecord::MTRF_Comment);
}
static MatchTableRecord Opcode(StringRef Opcode, int IndentAdjust = 0) {
unsigned ExtraFlags = 0;
if (IndentAdjust > 0)
ExtraFlags |= MatchTableRecord::MTRF_Indent;
if (IndentAdjust < 0)
ExtraFlags |= MatchTableRecord::MTRF_Outdent;
return MatchTableRecord(None, Opcode, 1,
MatchTableRecord::MTRF_CommaFollows | ExtraFlags);
}
static MatchTableRecord NamedValue(StringRef NamedValue) {
return MatchTableRecord(None, NamedValue, 1,
MatchTableRecord::MTRF_CommaFollows);
}
Reapply r332907 "[GlobalISel] Improving InstructionSelect's performance by reducing MatchTable..." Apparently the compile time problem was caused by the fact that not all compilers / STL implementations can automatically convert std::unique_ptr<Derived> to std::unique_ptr<Base>. Fixed (hopefully) by making sure it's std::unique_ptr<Derived>&& (rvalue ref) to std::unique_ptr<Base> conversion instead. llvm-svn: 332917
2018-05-22 07:28:51 +08:00
static MatchTableRecord NamedValue(StringRef NamedValue, int64_t RawValue) {
return MatchTableRecord(None, NamedValue, 1,
MatchTableRecord::MTRF_CommaFollows, RawValue);
}
[globalisel][tablegen] Add control-flow to the MatchTable. Summary: This will allow us to merge the various sub-tables into a single table. This is a compile-time saving at this point. However, this will also enable the optimization of a table so that similar instructions can be tested together, reducing the time spent on the matching the code. The bulk of this patch is a mechanical conversion to the new MatchTable object which is responsible for tracking label definitions and filling in the index of the jump targets. It is also responsible for nicely formatting the table. This was necessary to support the new GIM_Try opcode which takes the index to jump to if the match should fail. This value is unknown during table construction and is filled in during emission. To support nesting try-blocks (although we currently don't emit tables with nested try-blocks), GIM_Reject has been re-introduced to explicitly exit a try-block or fail the overall match if there are no active try-blocks. Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Reviewed By: rovka Subscribers: kristof.beyls, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D35117 llvm-svn: 308596
2017-07-20 17:25:44 +08:00
static MatchTableRecord NamedValue(StringRef Namespace,
StringRef NamedValue) {
return MatchTableRecord(None, (Namespace + "::" + NamedValue).str(), 1,
MatchTableRecord::MTRF_CommaFollows);
}
Reapply r332907 "[GlobalISel] Improving InstructionSelect's performance by reducing MatchTable..." Apparently the compile time problem was caused by the fact that not all compilers / STL implementations can automatically convert std::unique_ptr<Derived> to std::unique_ptr<Base>. Fixed (hopefully) by making sure it's std::unique_ptr<Derived>&& (rvalue ref) to std::unique_ptr<Base> conversion instead. llvm-svn: 332917
2018-05-22 07:28:51 +08:00
static MatchTableRecord NamedValue(StringRef Namespace, StringRef NamedValue,
int64_t RawValue) {
return MatchTableRecord(None, (Namespace + "::" + NamedValue).str(), 1,
MatchTableRecord::MTRF_CommaFollows, RawValue);
}
[globalisel][tablegen] Add control-flow to the MatchTable. Summary: This will allow us to merge the various sub-tables into a single table. This is a compile-time saving at this point. However, this will also enable the optimization of a table so that similar instructions can be tested together, reducing the time spent on the matching the code. The bulk of this patch is a mechanical conversion to the new MatchTable object which is responsible for tracking label definitions and filling in the index of the jump targets. It is also responsible for nicely formatting the table. This was necessary to support the new GIM_Try opcode which takes the index to jump to if the match should fail. This value is unknown during table construction and is filled in during emission. To support nesting try-blocks (although we currently don't emit tables with nested try-blocks), GIM_Reject has been re-introduced to explicitly exit a try-block or fail the overall match if there are no active try-blocks. Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Reviewed By: rovka Subscribers: kristof.beyls, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D35117 llvm-svn: 308596
2017-07-20 17:25:44 +08:00
static MatchTableRecord IntValue(int64_t IntValue) {
return MatchTableRecord(None, llvm::to_string(IntValue), 1,
MatchTableRecord::MTRF_CommaFollows);
}
static MatchTableRecord Label(unsigned LabelID) {
return MatchTableRecord(LabelID, "Label " + llvm::to_string(LabelID), 0,
MatchTableRecord::MTRF_Label |
MatchTableRecord::MTRF_Comment |
MatchTableRecord::MTRF_LineBreakFollows);
}
static MatchTableRecord JumpTarget(unsigned LabelID) {
Re-commit: r309094 [globalisel][tablegen] Fuse the generated tables together. Summary: Now that we have control flow in place, fuse the per-rule tables into a single table. This is a compile-time saving at this point. However, this will also enable the optimization of a table so that similar instructions can be tested together, reducing the time spent on the matching the code. This is NFC in terms of externally visible behaviour but some internals have changed slightly. State.MIs is no longer reset between each rule that is attempted because it's not necessary to do so. As a consequence of this the restriction on the order that instructions are added to State.MIs has been relaxed to only affect recorded instructions that require new elements to be added to the vector. GIM_RecordInsn can now write to any element from 1 to State.MIs.size() instead of just State.MIs.size(). The compile-time regressions from the last commit were caused by the ARM target including a non-const variable (zero_reg) in the table and therefore generating an initializer for it. That variable is now const. Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Reviewed By: rovka Subscribers: kristof.beyls, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D35681 llvm-svn: 309264
2017-07-27 19:03:45 +08:00
return MatchTableRecord(LabelID, "Label " + llvm::to_string(LabelID), 1,
[globalisel][tablegen] Add control-flow to the MatchTable. Summary: This will allow us to merge the various sub-tables into a single table. This is a compile-time saving at this point. However, this will also enable the optimization of a table so that similar instructions can be tested together, reducing the time spent on the matching the code. The bulk of this patch is a mechanical conversion to the new MatchTable object which is responsible for tracking label definitions and filling in the index of the jump targets. It is also responsible for nicely formatting the table. This was necessary to support the new GIM_Try opcode which takes the index to jump to if the match should fail. This value is unknown during table construction and is filled in during emission. To support nesting try-blocks (although we currently don't emit tables with nested try-blocks), GIM_Reject has been re-introduced to explicitly exit a try-block or fail the overall match if there are no active try-blocks. Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Reviewed By: rovka Subscribers: kristof.beyls, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D35117 llvm-svn: 308596
2017-07-20 17:25:44 +08:00
MatchTableRecord::MTRF_JumpTarget |
MatchTableRecord::MTRF_Comment |
MatchTableRecord::MTRF_CommaFollows);
}
[GlobalISel][InstructionSelect] Making Coverage Info generation optional on per-match table basis to make sure that Testgen always has access to coverage info even if the match table used by the selector itself is stripped off that information for performance reasons. Reviewers: dsanders, aemerson Reviewed By: dsanders Subscribers: rovka, kristof.beyls, llvm-commits, dsanders Differential Revision: https://reviews.llvm.org/D46098 llvm-svn: 331398
2018-05-03 04:15:11 +08:00
static MatchTable buildTable(ArrayRef<Matcher *> Rules, bool WithCoverage);
[GlobalISel][InstructionSelect] Refactoring buildMatchTable out, NFC to share it between the Instruction Selector in optimized and non-optimized modes both and the Testgen. Reviewers: dsanders, aemerson Reviewed By: dsanders Subscribers: rovka, kristof.beyls, llvm-commits, dsanders Differential Revision: https://reviews.llvm.org/D46097 llvm-svn: 331396
2018-05-03 04:08:14 +08:00
[GlobalISel][InstructionSelect] Making Coverage Info generation optional on per-match table basis to make sure that Testgen always has access to coverage info even if the match table used by the selector itself is stripped off that information for performance reasons. Reviewers: dsanders, aemerson Reviewed By: dsanders Subscribers: rovka, kristof.beyls, llvm-commits, dsanders Differential Revision: https://reviews.llvm.org/D46098 llvm-svn: 331398
2018-05-03 04:15:11 +08:00
MatchTable(bool WithCoverage, unsigned ID = 0)
: ID(ID), IsWithCoverage(WithCoverage) {}
bool isWithCoverage() const { return IsWithCoverage; }
[globalisel][tablegen] Add control-flow to the MatchTable. Summary: This will allow us to merge the various sub-tables into a single table. This is a compile-time saving at this point. However, this will also enable the optimization of a table so that similar instructions can be tested together, reducing the time spent on the matching the code. The bulk of this patch is a mechanical conversion to the new MatchTable object which is responsible for tracking label definitions and filling in the index of the jump targets. It is also responsible for nicely formatting the table. This was necessary to support the new GIM_Try opcode which takes the index to jump to if the match should fail. This value is unknown during table construction and is filled in during emission. To support nesting try-blocks (although we currently don't emit tables with nested try-blocks), GIM_Reject has been re-introduced to explicitly exit a try-block or fail the overall match if there are no active try-blocks. Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Reviewed By: rovka Subscribers: kristof.beyls, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D35117 llvm-svn: 308596
2017-07-20 17:25:44 +08:00
void push_back(const MatchTableRecord &Value) {
if (Value.Flags & MatchTableRecord::MTRF_Label)
defineLabel(Value.LabelID);
Contents.push_back(Value);
CurrentSize += Value.size();
}
[GlobalISel][InstructionSelect] Refactoring buildMatchTable out, NFC to share it between the Instruction Selector in optimized and non-optimized modes both and the Testgen. Reviewers: dsanders, aemerson Reviewed By: dsanders Subscribers: rovka, kristof.beyls, llvm-commits, dsanders Differential Revision: https://reviews.llvm.org/D46097 llvm-svn: 331396
2018-05-03 04:08:14 +08:00
unsigned allocateLabelID() { return CurrentLabelID++; }
Re-commit: r309094 [globalisel][tablegen] Fuse the generated tables together. Summary: Now that we have control flow in place, fuse the per-rule tables into a single table. This is a compile-time saving at this point. However, this will also enable the optimization of a table so that similar instructions can be tested together, reducing the time spent on the matching the code. This is NFC in terms of externally visible behaviour but some internals have changed slightly. State.MIs is no longer reset between each rule that is attempted because it's not necessary to do so. As a consequence of this the restriction on the order that instructions are added to State.MIs has been relaxed to only affect recorded instructions that require new elements to be added to the vector. GIM_RecordInsn can now write to any element from 1 to State.MIs.size() instead of just State.MIs.size(). The compile-time regressions from the last commit were caused by the ARM target including a non-const variable (zero_reg) in the table and therefore generating an initializer for it. That variable is now const. Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Reviewed By: rovka Subscribers: kristof.beyls, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D35681 llvm-svn: 309264
2017-07-27 19:03:45 +08:00
[globalisel][tablegen] Add control-flow to the MatchTable. Summary: This will allow us to merge the various sub-tables into a single table. This is a compile-time saving at this point. However, this will also enable the optimization of a table so that similar instructions can be tested together, reducing the time spent on the matching the code. The bulk of this patch is a mechanical conversion to the new MatchTable object which is responsible for tracking label definitions and filling in the index of the jump targets. It is also responsible for nicely formatting the table. This was necessary to support the new GIM_Try opcode which takes the index to jump to if the match should fail. This value is unknown during table construction and is filled in during emission. To support nesting try-blocks (although we currently don't emit tables with nested try-blocks), GIM_Reject has been re-introduced to explicitly exit a try-block or fail the overall match if there are no active try-blocks. Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Reviewed By: rovka Subscribers: kristof.beyls, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D35117 llvm-svn: 308596
2017-07-20 17:25:44 +08:00
void defineLabel(unsigned LabelID) {
Re-commit: r309094 [globalisel][tablegen] Fuse the generated tables together. Summary: Now that we have control flow in place, fuse the per-rule tables into a single table. This is a compile-time saving at this point. However, this will also enable the optimization of a table so that similar instructions can be tested together, reducing the time spent on the matching the code. This is NFC in terms of externally visible behaviour but some internals have changed slightly. State.MIs is no longer reset between each rule that is attempted because it's not necessary to do so. As a consequence of this the restriction on the order that instructions are added to State.MIs has been relaxed to only affect recorded instructions that require new elements to be added to the vector. GIM_RecordInsn can now write to any element from 1 to State.MIs.size() instead of just State.MIs.size(). The compile-time regressions from the last commit were caused by the ARM target including a non-const variable (zero_reg) in the table and therefore generating an initializer for it. That variable is now const. Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Reviewed By: rovka Subscribers: kristof.beyls, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D35681 llvm-svn: 309264
2017-07-27 19:03:45 +08:00
LabelMap.insert(std::make_pair(LabelID, CurrentSize));
[globalisel][tablegen] Add control-flow to the MatchTable. Summary: This will allow us to merge the various sub-tables into a single table. This is a compile-time saving at this point. However, this will also enable the optimization of a table so that similar instructions can be tested together, reducing the time spent on the matching the code. The bulk of this patch is a mechanical conversion to the new MatchTable object which is responsible for tracking label definitions and filling in the index of the jump targets. It is also responsible for nicely formatting the table. This was necessary to support the new GIM_Try opcode which takes the index to jump to if the match should fail. This value is unknown during table construction and is filled in during emission. To support nesting try-blocks (although we currently don't emit tables with nested try-blocks), GIM_Reject has been re-introduced to explicitly exit a try-block or fail the overall match if there are no active try-blocks. Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Reviewed By: rovka Subscribers: kristof.beyls, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D35117 llvm-svn: 308596
2017-07-20 17:25:44 +08:00
}
unsigned getLabelIndex(unsigned LabelID) const {
const auto I = LabelMap.find(LabelID);
assert(I != LabelMap.end() && "Use of undeclared label");
return I->second;
}
Re-commit: r309094 [globalisel][tablegen] Fuse the generated tables together. Summary: Now that we have control flow in place, fuse the per-rule tables into a single table. This is a compile-time saving at this point. However, this will also enable the optimization of a table so that similar instructions can be tested together, reducing the time spent on the matching the code. This is NFC in terms of externally visible behaviour but some internals have changed slightly. State.MIs is no longer reset between each rule that is attempted because it's not necessary to do so. As a consequence of this the restriction on the order that instructions are added to State.MIs has been relaxed to only affect recorded instructions that require new elements to be added to the vector. GIM_RecordInsn can now write to any element from 1 to State.MIs.size() instead of just State.MIs.size(). The compile-time regressions from the last commit were caused by the ARM target including a non-const variable (zero_reg) in the table and therefore generating an initializer for it. That variable is now const. Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Reviewed By: rovka Subscribers: kristof.beyls, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D35681 llvm-svn: 309264
2017-07-27 19:03:45 +08:00
void emitUse(raw_ostream &OS) const { OS << "MatchTable" << ID; }
void emitDeclaration(raw_ostream &OS) const {
[globalisel][tablegen] Add control-flow to the MatchTable. Summary: This will allow us to merge the various sub-tables into a single table. This is a compile-time saving at this point. However, this will also enable the optimization of a table so that similar instructions can be tested together, reducing the time spent on the matching the code. The bulk of this patch is a mechanical conversion to the new MatchTable object which is responsible for tracking label definitions and filling in the index of the jump targets. It is also responsible for nicely formatting the table. This was necessary to support the new GIM_Try opcode which takes the index to jump to if the match should fail. This value is unknown during table construction and is filled in during emission. To support nesting try-blocks (although we currently don't emit tables with nested try-blocks), GIM_Reject has been re-introduced to explicitly exit a try-block or fail the overall match if there are no active try-blocks. Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Reviewed By: rovka Subscribers: kristof.beyls, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D35117 llvm-svn: 308596
2017-07-20 17:25:44 +08:00
unsigned Indentation = 4;
[globalisel][tablegen] Ensure MatchTable's are compile-time constants with constexpr. NFC. This should prevent any re-occurence of the problem where the table was initialized at run-time. llvm-svn: 309267
2017-07-27 20:47:31 +08:00
OS << " constexpr static int64_t MatchTable" << ID << "[] = {";
[globalisel][tablegen] Add control-flow to the MatchTable. Summary: This will allow us to merge the various sub-tables into a single table. This is a compile-time saving at this point. However, this will also enable the optimization of a table so that similar instructions can be tested together, reducing the time spent on the matching the code. The bulk of this patch is a mechanical conversion to the new MatchTable object which is responsible for tracking label definitions and filling in the index of the jump targets. It is also responsible for nicely formatting the table. This was necessary to support the new GIM_Try opcode which takes the index to jump to if the match should fail. This value is unknown during table construction and is filled in during emission. To support nesting try-blocks (although we currently don't emit tables with nested try-blocks), GIM_Reject has been re-introduced to explicitly exit a try-block or fail the overall match if there are no active try-blocks. Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Reviewed By: rovka Subscribers: kristof.beyls, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D35117 llvm-svn: 308596
2017-07-20 17:25:44 +08:00
LineBreak.emit(OS, true, *this);
OS << std::string(Indentation, ' ');
for (auto I = Contents.begin(), E = Contents.end(); I != E;
++I) {
bool LineBreakIsNext = false;
const auto &NextI = std::next(I);
if (NextI != E) {
if (NextI->EmitStr == "" &&
NextI->Flags == MatchTableRecord::MTRF_LineBreakFollows)
LineBreakIsNext = true;
}
if (I->Flags & MatchTableRecord::MTRF_Indent)
Indentation += 2;
I->emit(OS, LineBreakIsNext, *this);
if (I->Flags & MatchTableRecord::MTRF_LineBreakFollows)
OS << std::string(Indentation, ' ');
if (I->Flags & MatchTableRecord::MTRF_Outdent)
Indentation -= 2;
}
OS << "};\n";
}
};
MatchTableRecord MatchTable::LineBreak = {
None, "" /* Emit String */, 0 /* Elements */,
MatchTableRecord::MTRF_LineBreakFollows};
void MatchTableRecord::emit(raw_ostream &OS, bool LineBreakIsNextAfterThis,
const MatchTable &Table) const {
bool UseLineComment =
LineBreakIsNextAfterThis | (Flags & MTRF_LineBreakFollows);
if (Flags & (MTRF_JumpTarget | MTRF_CommaFollows))
UseLineComment = false;
if (Flags & MTRF_Comment)
OS << (UseLineComment ? "// " : "/*");
OS << EmitStr;
if (Flags & MTRF_Label)
OS << ": @" << Table.getLabelIndex(LabelID);
if (Flags & MTRF_Comment && !UseLineComment)
OS << "*/";
if (Flags & MTRF_JumpTarget) {
if (Flags & MTRF_Comment)
OS << " ";
OS << Table.getLabelIndex(LabelID);
}
if (Flags & MTRF_CommaFollows) {
OS << ",";
if (!LineBreakIsNextAfterThis && !(Flags & MTRF_LineBreakFollows))
OS << " ";
}
if (Flags & MTRF_LineBreakFollows)
OS << "\n";
}
MatchTable &operator<<(MatchTable &Table, const MatchTableRecord &Value) {
Table.push_back(Value);
return Table;
}
[GlobalISel] Add basic Selector-emitter tblgen backend. This adds a basic tablegen backend that analyzes the SelectionDAG patterns to find simple ones that are eligible for GlobalISel-emission. That's similar to FastISel, with one notable difference: we're not fed ISD opcodes, so we need to map the SDNode operators to generic opcodes. That's done using GINodeEquiv in TargetGlobalISel.td. Otherwise, this is mostly boilerplate, and lots of filtering of any kind of "complicated" pattern. On AArch64, this is sufficient to match G_ADD up to s64 (to ADDWrr/ADDXrr) and G_BR (to B). Differential Revision: https://reviews.llvm.org/D26878 llvm-svn: 290284
2016-12-22 07:26:20 +08:00
//===- Matchers -----------------------------------------------------------===//
[tablegen][globalisel] Add support for nested instruction matching. Summary: Lift the restrictions that prevented the tree walking introduced in the previous change and add support for patterns like: (G_ADD (G_MUL (G_SEXT $src1), (G_SEXT $src2)), $src3) -> SMADDWrrr $dst, $src1, $src2, $src3 Also adds support for G_SEXT and G_ZEXT to support these cases. One particular aspect of this that I should draw attention to is that I've tried to be overly conservative in determining the safety of matches that involve non-adjacent instructions and multiple basic blocks. This is intended to be used as a cheap initial check and we may add a more expensive check in the future. The current rules are: * Reject if any instruction may load/store (we'd need to check for intervening memory operations. * Reject if any instruction has implicit operands. * Reject if any instruction has unmodelled side-effects. See isObviouslySafeToFold(). Reviewers: t.p.northover, javed.absar, qcolombet, aditya_nandakumar, ab, rovka Reviewed By: ab Subscribers: igorb, dberris, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D30539 llvm-svn: 299430
2017-04-04 21:25:23 +08:00
class OperandMatcher;
[tablegen][globalisel] Trivial changes to reduce size of following patch. NFC. Summary: * Move namespace { * Trivial: Typo * RuleMatcher: Separate class and definition * Trivial: const findNodeEquiv Reviewers: t.p.northover, qcolombet, aditya_nandakumar, ab, rovka Reviewed By: rovka Subscribers: dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D30531 llvm-svn: 297884
2017-03-16 04:18:38 +08:00
class MatchAction;
[TableGen][GlobalISel] Optimize MatchTable for faster instruction selection *** Context *** Prior to this patchw, the table generated for matching instruction was straight forward but highly inefficient. Basically, each pattern generates its own set of self contained checks and actions. E.g., TableGen generated: // First pattern CheckNumOperand 3 CheckOpcode G_ADD ... Build ADDrr // Second pattern CheckNumOperand 3 CheckOpcode G_ADD ... Build ADDri // Third pattern CheckNumOperand 3 CheckOpcode G_SUB ... Build SUBrr *** Problem *** Because of that generation, a *lot* of check were redundant between each pattern and were checked every single time until we reach the pattern that matches. E.g., Taking the previous table, let say we are matching a G_SUB, that means we were going to check all the rules for G_ADD before looking at the G_SUB rule. In particular we are going to do: check 3 operands; PASS check G_ADD; FAIL ; Next rule check 3 operands; PASS (but we already knew that!) check G_ADD; FAIL (well it is still not true) ; Next rule check 3 operands; PASS (really!!) check G_SUB; PASS (at last :P) *** Proposed Solution *** This patch introduces a concept of group of rules (GroupMatcher) that share some predicates and only get checked once for the whole group. This patch only creates groups with one nesting level. Conceptually there is nothing preventing us for having deeper nest level. However, the current implementation is not smart enough to share the recording (aka capturing) of values. That limits its ability to do more sharing. For the given example the current patch will generate: // First group CheckOpcode G_ADD // First pattern CheckNumOperand 3 ... Build ADDrr // Second pattern CheckNumOperand 3 ... Build ADDri // Second group CheckOpcode G_SUB // Third pattern CheckNumOperand 3 ... Build SUBrr But if we allowed several nesting level, it could create a sub group for the checknumoperand 3. (We would need to call optimizeRules on the rules within a group.) *** Result *** With only one level of nesting, the instruction selection pass is up to 4x faster. For instance, one instruction now takes 500 checks, instead of 24k! With more nesting we could get in the tens I believe. Differential Revision: https://reviews.llvm.org/D39034 rdar://problem/34670699 llvm-svn: 321017
2017-12-19 03:47:41 +08:00
class PredicateMatcher;
class RuleMatcher;
class Matcher {
public:
virtual ~Matcher() = default;
Reapply r332907 "[GlobalISel] Improving InstructionSelect's performance by reducing MatchTable..." Apparently the compile time problem was caused by the fact that not all compilers / STL implementations can automatically convert std::unique_ptr<Derived> to std::unique_ptr<Base>. Fixed (hopefully) by making sure it's std::unique_ptr<Derived>&& (rvalue ref) to std::unique_ptr<Base> conversion instead. llvm-svn: 332917
2018-05-22 07:28:51 +08:00
virtual void optimize() {}
[TableGen][GlobalISel] Optimize MatchTable for faster instruction selection *** Context *** Prior to this patchw, the table generated for matching instruction was straight forward but highly inefficient. Basically, each pattern generates its own set of self contained checks and actions. E.g., TableGen generated: // First pattern CheckNumOperand 3 CheckOpcode G_ADD ... Build ADDrr // Second pattern CheckNumOperand 3 CheckOpcode G_ADD ... Build ADDri // Third pattern CheckNumOperand 3 CheckOpcode G_SUB ... Build SUBrr *** Problem *** Because of that generation, a *lot* of check were redundant between each pattern and were checked every single time until we reach the pattern that matches. E.g., Taking the previous table, let say we are matching a G_SUB, that means we were going to check all the rules for G_ADD before looking at the G_SUB rule. In particular we are going to do: check 3 operands; PASS check G_ADD; FAIL ; Next rule check 3 operands; PASS (but we already knew that!) check G_ADD; FAIL (well it is still not true) ; Next rule check 3 operands; PASS (really!!) check G_SUB; PASS (at last :P) *** Proposed Solution *** This patch introduces a concept of group of rules (GroupMatcher) that share some predicates and only get checked once for the whole group. This patch only creates groups with one nesting level. Conceptually there is nothing preventing us for having deeper nest level. However, the current implementation is not smart enough to share the recording (aka capturing) of values. That limits its ability to do more sharing. For the given example the current patch will generate: // First group CheckOpcode G_ADD // First pattern CheckNumOperand 3 ... Build ADDrr // Second pattern CheckNumOperand 3 ... Build ADDri // Second group CheckOpcode G_SUB // Third pattern CheckNumOperand 3 ... Build SUBrr But if we allowed several nesting level, it could create a sub group for the checknumoperand 3. (We would need to call optimizeRules on the rules within a group.) *** Result *** With only one level of nesting, the instruction selection pass is up to 4x faster. For instance, one instruction now takes 500 checks, instead of 24k! With more nesting we could get in the tens I believe. Differential Revision: https://reviews.llvm.org/D39034 rdar://problem/34670699 llvm-svn: 321017
2017-12-19 03:47:41 +08:00
virtual void emit(MatchTable &Table) = 0;
Reapply r332907 "[GlobalISel] Improving InstructionSelect's performance by reducing MatchTable..." Apparently the compile time problem was caused by the fact that not all compilers / STL implementations can automatically convert std::unique_ptr<Derived> to std::unique_ptr<Base>. Fixed (hopefully) by making sure it's std::unique_ptr<Derived>&& (rvalue ref) to std::unique_ptr<Base> conversion instead. llvm-svn: 332917
2018-05-22 07:28:51 +08:00
virtual bool hasFirstCondition() const = 0;
virtual const PredicateMatcher &getFirstCondition() const = 0;
virtual std::unique_ptr<PredicateMatcher> popFirstCondition() = 0;
[TableGen][GlobalISel] Optimize MatchTable for faster instruction selection *** Context *** Prior to this patchw, the table generated for matching instruction was straight forward but highly inefficient. Basically, each pattern generates its own set of self contained checks and actions. E.g., TableGen generated: // First pattern CheckNumOperand 3 CheckOpcode G_ADD ... Build ADDrr // Second pattern CheckNumOperand 3 CheckOpcode G_ADD ... Build ADDri // Third pattern CheckNumOperand 3 CheckOpcode G_SUB ... Build SUBrr *** Problem *** Because of that generation, a *lot* of check were redundant between each pattern and were checked every single time until we reach the pattern that matches. E.g., Taking the previous table, let say we are matching a G_SUB, that means we were going to check all the rules for G_ADD before looking at the G_SUB rule. In particular we are going to do: check 3 operands; PASS check G_ADD; FAIL ; Next rule check 3 operands; PASS (but we already knew that!) check G_ADD; FAIL (well it is still not true) ; Next rule check 3 operands; PASS (really!!) check G_SUB; PASS (at last :P) *** Proposed Solution *** This patch introduces a concept of group of rules (GroupMatcher) that share some predicates and only get checked once for the whole group. This patch only creates groups with one nesting level. Conceptually there is nothing preventing us for having deeper nest level. However, the current implementation is not smart enough to share the recording (aka capturing) of values. That limits its ability to do more sharing. For the given example the current patch will generate: // First group CheckOpcode G_ADD // First pattern CheckNumOperand 3 ... Build ADDrr // Second pattern CheckNumOperand 3 ... Build ADDri // Second group CheckOpcode G_SUB // Third pattern CheckNumOperand 3 ... Build SUBrr But if we allowed several nesting level, it could create a sub group for the checknumoperand 3. (We would need to call optimizeRules on the rules within a group.) *** Result *** With only one level of nesting, the instruction selection pass is up to 4x faster. For instance, one instruction now takes 500 checks, instead of 24k! With more nesting we could get in the tens I believe. Differential Revision: https://reviews.llvm.org/D39034 rdar://problem/34670699 llvm-svn: 321017
2017-12-19 03:47:41 +08:00
};
[GlobalISel][InstructionSelect] Making Coverage Info generation optional on per-match table basis to make sure that Testgen always has access to coverage info even if the match table used by the selector itself is stripped off that information for performance reasons. Reviewers: dsanders, aemerson Reviewed By: dsanders Subscribers: rovka, kristof.beyls, llvm-commits, dsanders Differential Revision: https://reviews.llvm.org/D46098 llvm-svn: 331398
2018-05-03 04:15:11 +08:00
MatchTable MatchTable::buildTable(ArrayRef<Matcher *> Rules,
bool WithCoverage) {
MatchTable Table(WithCoverage);
[GlobalISel][InstructionSelect] Refactoring buildMatchTable out, NFC to share it between the Instruction Selector in optimized and non-optimized modes both and the Testgen. Reviewers: dsanders, aemerson Reviewed By: dsanders Subscribers: rovka, kristof.beyls, llvm-commits, dsanders Differential Revision: https://reviews.llvm.org/D46097 llvm-svn: 331396
2018-05-03 04:08:14 +08:00
for (Matcher *Rule : Rules)
Rule->emit(Table);
return Table << MatchTable::Opcode("GIM_Reject") << MatchTable::LineBreak;
}
Reapply r332907 "[GlobalISel] Improving InstructionSelect's performance by reducing MatchTable..." Apparently the compile time problem was caused by the fact that not all compilers / STL implementations can automatically convert std::unique_ptr<Derived> to std::unique_ptr<Base>. Fixed (hopefully) by making sure it's std::unique_ptr<Derived>&& (rvalue ref) to std::unique_ptr<Base> conversion instead. llvm-svn: 332917
2018-05-22 07:28:51 +08:00
class GroupMatcher final : public Matcher {
/// Conditions that form a common prefix of all the matchers contained.
SmallVector<std::unique_ptr<PredicateMatcher>, 1> Conditions;
/// All the nested matchers, sharing a common prefix.
std::vector<Matcher *> Matchers;
/// An owning collection for any auxiliary matchers created while optimizing
/// nested matchers contained.
std::vector<std::unique_ptr<Matcher>> MatcherStorage;
[TableGen][GlobalISel] Optimize MatchTable for faster instruction selection *** Context *** Prior to this patchw, the table generated for matching instruction was straight forward but highly inefficient. Basically, each pattern generates its own set of self contained checks and actions. E.g., TableGen generated: // First pattern CheckNumOperand 3 CheckOpcode G_ADD ... Build ADDrr // Second pattern CheckNumOperand 3 CheckOpcode G_ADD ... Build ADDri // Third pattern CheckNumOperand 3 CheckOpcode G_SUB ... Build SUBrr *** Problem *** Because of that generation, a *lot* of check were redundant between each pattern and were checked every single time until we reach the pattern that matches. E.g., Taking the previous table, let say we are matching a G_SUB, that means we were going to check all the rules for G_ADD before looking at the G_SUB rule. In particular we are going to do: check 3 operands; PASS check G_ADD; FAIL ; Next rule check 3 operands; PASS (but we already knew that!) check G_ADD; FAIL (well it is still not true) ; Next rule check 3 operands; PASS (really!!) check G_SUB; PASS (at last :P) *** Proposed Solution *** This patch introduces a concept of group of rules (GroupMatcher) that share some predicates and only get checked once for the whole group. This patch only creates groups with one nesting level. Conceptually there is nothing preventing us for having deeper nest level. However, the current implementation is not smart enough to share the recording (aka capturing) of values. That limits its ability to do more sharing. For the given example the current patch will generate: // First group CheckOpcode G_ADD // First pattern CheckNumOperand 3 ... Build ADDrr // Second pattern CheckNumOperand 3 ... Build ADDri // Second group CheckOpcode G_SUB // Third pattern CheckNumOperand 3 ... Build SUBrr But if we allowed several nesting level, it could create a sub group for the checknumoperand 3. (We would need to call optimizeRules on the rules within a group.) *** Result *** With only one level of nesting, the instruction selection pass is up to 4x faster. For instance, one instruction now takes 500 checks, instead of 24k! With more nesting we could get in the tens I believe. Differential Revision: https://reviews.llvm.org/D39034 rdar://problem/34670699 llvm-svn: 321017
2017-12-19 03:47:41 +08:00
public:
Reapply r332907 "[GlobalISel] Improving InstructionSelect's performance by reducing MatchTable..." Apparently the compile time problem was caused by the fact that not all compilers / STL implementations can automatically convert std::unique_ptr<Derived> to std::unique_ptr<Base>. Fixed (hopefully) by making sure it's std::unique_ptr<Derived>&& (rvalue ref) to std::unique_ptr<Base> conversion instead. llvm-svn: 332917
2018-05-22 07:28:51 +08:00
/// Add a matcher to the collection of nested matchers if it meets the
/// requirements, and return true. If it doesn't, do nothing and return false.
///
/// Expected to preserve its argument, so it could be moved out later on.
bool addMatcher(Matcher &Candidate);
/// Mark the matcher as fully-built and ensure any invariants expected by both
/// optimize() and emit(...) methods. Generally, both sequences of calls
/// are expected to lead to a sensible result:
///
/// addMatcher(...)*; finalize(); optimize(); emit(...); and
/// addMatcher(...)*; finalize(); emit(...);
///
/// or generally
///
/// addMatcher(...)*; finalize(); { optimize()*; emit(...); }*
///
/// Multiple calls to optimize() are expected to be handled gracefully, though
/// optimize() is not expected to be idempotent. Multiple calls to finalize()
/// aren't generally supported. emit(...) is expected to be non-mutating and
/// producing the exact same results upon repeated calls.
///
/// addMatcher() calls after the finalize() call are not supported.
///
/// finalize() and optimize() are both allowed to mutate the contained
/// matchers, so moving them out after finalize() is not supported.
void finalize();
[GlobalISel][InstructionSelect] MatchTable second level grouping, perf patch 5 This patch continues a series of patches started by r332907 (reapplied as r332917) In this commit we start grouping rules with common first condition on the second level of the table. This is expected to decrease time GlobalISel spends in its InstructionSelect pass by roughly 13% for an -O0 build as measured on sqlite3-amalgamation (http://sqlite.org/download.html) targeting AArch64. Reviewers: qcolombet, dsanders, bogner, aemerson, javed.absar Reviewed By: qcolombet Subscribers: rovka, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D44700 llvm-svn: 333053
2018-05-23 10:04:19 +08:00
void optimize() override;
Reapply r332907 "[GlobalISel] Improving InstructionSelect's performance by reducing MatchTable..." Apparently the compile time problem was caused by the fact that not all compilers / STL implementations can automatically convert std::unique_ptr<Derived> to std::unique_ptr<Base>. Fixed (hopefully) by making sure it's std::unique_ptr<Derived>&& (rvalue ref) to std::unique_ptr<Base> conversion instead. llvm-svn: 332917
2018-05-22 07:28:51 +08:00
void emit(MatchTable &Table) override;
/// Could be used to move out the matchers added previously, unless finalize()
/// has been already called. If any of the matchers are moved out, the group
/// becomes safe to destroy, but not safe to re-use for anything else.
iterator_range<std::vector<Matcher *>::iterator> matchers() {
return make_range(Matchers.begin(), Matchers.end());
[TableGen][GlobalISel] Optimize MatchTable for faster instruction selection *** Context *** Prior to this patchw, the table generated for matching instruction was straight forward but highly inefficient. Basically, each pattern generates its own set of self contained checks and actions. E.g., TableGen generated: // First pattern CheckNumOperand 3 CheckOpcode G_ADD ... Build ADDrr // Second pattern CheckNumOperand 3 CheckOpcode G_ADD ... Build ADDri // Third pattern CheckNumOperand 3 CheckOpcode G_SUB ... Build SUBrr *** Problem *** Because of that generation, a *lot* of check were redundant between each pattern and were checked every single time until we reach the pattern that matches. E.g., Taking the previous table, let say we are matching a G_SUB, that means we were going to check all the rules for G_ADD before looking at the G_SUB rule. In particular we are going to do: check 3 operands; PASS check G_ADD; FAIL ; Next rule check 3 operands; PASS (but we already knew that!) check G_ADD; FAIL (well it is still not true) ; Next rule check 3 operands; PASS (really!!) check G_SUB; PASS (at last :P) *** Proposed Solution *** This patch introduces a concept of group of rules (GroupMatcher) that share some predicates and only get checked once for the whole group. This patch only creates groups with one nesting level. Conceptually there is nothing preventing us for having deeper nest level. However, the current implementation is not smart enough to share the recording (aka capturing) of values. That limits its ability to do more sharing. For the given example the current patch will generate: // First group CheckOpcode G_ADD // First pattern CheckNumOperand 3 ... Build ADDrr // Second pattern CheckNumOperand 3 ... Build ADDri // Second group CheckOpcode G_SUB // Third pattern CheckNumOperand 3 ... Build SUBrr But if we allowed several nesting level, it could create a sub group for the checknumoperand 3. (We would need to call optimizeRules on the rules within a group.) *** Result *** With only one level of nesting, the instruction selection pass is up to 4x faster. For instance, one instruction now takes 500 checks, instead of 24k! With more nesting we could get in the tens I believe. Differential Revision: https://reviews.llvm.org/D39034 rdar://problem/34670699 llvm-svn: 321017
2017-12-19 03:47:41 +08:00
}
Reapply r332907 "[GlobalISel] Improving InstructionSelect's performance by reducing MatchTable..." Apparently the compile time problem was caused by the fact that not all compilers / STL implementations can automatically convert std::unique_ptr<Derived> to std::unique_ptr<Base>. Fixed (hopefully) by making sure it's std::unique_ptr<Derived>&& (rvalue ref) to std::unique_ptr<Base> conversion instead. llvm-svn: 332917
2018-05-22 07:28:51 +08:00
size_t size() const { return Matchers.size(); }
bool empty() const { return Matchers.empty(); }
std::unique_ptr<PredicateMatcher> popFirstCondition() override {
assert(!Conditions.empty() &&
"Trying to pop a condition from a condition-less group");
std::unique_ptr<PredicateMatcher> P = std::move(Conditions.front());
Conditions.erase(Conditions.begin());
return P;
[TableGen][GlobalISel] Optimize MatchTable for faster instruction selection *** Context *** Prior to this patchw, the table generated for matching instruction was straight forward but highly inefficient. Basically, each pattern generates its own set of self contained checks and actions. E.g., TableGen generated: // First pattern CheckNumOperand 3 CheckOpcode G_ADD ... Build ADDrr // Second pattern CheckNumOperand 3 CheckOpcode G_ADD ... Build ADDri // Third pattern CheckNumOperand 3 CheckOpcode G_SUB ... Build SUBrr *** Problem *** Because of that generation, a *lot* of check were redundant between each pattern and were checked every single time until we reach the pattern that matches. E.g., Taking the previous table, let say we are matching a G_SUB, that means we were going to check all the rules for G_ADD before looking at the G_SUB rule. In particular we are going to do: check 3 operands; PASS check G_ADD; FAIL ; Next rule check 3 operands; PASS (but we already knew that!) check G_ADD; FAIL (well it is still not true) ; Next rule check 3 operands; PASS (really!!) check G_SUB; PASS (at last :P) *** Proposed Solution *** This patch introduces a concept of group of rules (GroupMatcher) that share some predicates and only get checked once for the whole group. This patch only creates groups with one nesting level. Conceptually there is nothing preventing us for having deeper nest level. However, the current implementation is not smart enough to share the recording (aka capturing) of values. That limits its ability to do more sharing. For the given example the current patch will generate: // First group CheckOpcode G_ADD // First pattern CheckNumOperand 3 ... Build ADDrr // Second pattern CheckNumOperand 3 ... Build ADDri // Second group CheckOpcode G_SUB // Third pattern CheckNumOperand 3 ... Build SUBrr But if we allowed several nesting level, it could create a sub group for the checknumoperand 3. (We would need to call optimizeRules on the rules within a group.) *** Result *** With only one level of nesting, the instruction selection pass is up to 4x faster. For instance, one instruction now takes 500 checks, instead of 24k! With more nesting we could get in the tens I believe. Differential Revision: https://reviews.llvm.org/D39034 rdar://problem/34670699 llvm-svn: 321017
2017-12-19 03:47:41 +08:00
}
Reapply r332907 "[GlobalISel] Improving InstructionSelect's performance by reducing MatchTable..." Apparently the compile time problem was caused by the fact that not all compilers / STL implementations can automatically convert std::unique_ptr<Derived> to std::unique_ptr<Base>. Fixed (hopefully) by making sure it's std::unique_ptr<Derived>&& (rvalue ref) to std::unique_ptr<Base> conversion instead. llvm-svn: 332917
2018-05-22 07:28:51 +08:00
const PredicateMatcher &getFirstCondition() const override {
assert(!Conditions.empty() &&
"Trying to get a condition from a condition-less group");
return *Conditions.front();
[TableGen][GlobalISel] Optimize MatchTable for faster instruction selection *** Context *** Prior to this patchw, the table generated for matching instruction was straight forward but highly inefficient. Basically, each pattern generates its own set of self contained checks and actions. E.g., TableGen generated: // First pattern CheckNumOperand 3 CheckOpcode G_ADD ... Build ADDrr // Second pattern CheckNumOperand 3 CheckOpcode G_ADD ... Build ADDri // Third pattern CheckNumOperand 3 CheckOpcode G_SUB ... Build SUBrr *** Problem *** Because of that generation, a *lot* of check were redundant between each pattern and were checked every single time until we reach the pattern that matches. E.g., Taking the previous table, let say we are matching a G_SUB, that means we were going to check all the rules for G_ADD before looking at the G_SUB rule. In particular we are going to do: check 3 operands; PASS check G_ADD; FAIL ; Next rule check 3 operands; PASS (but we already knew that!) check G_ADD; FAIL (well it is still not true) ; Next rule check 3 operands; PASS (really!!) check G_SUB; PASS (at last :P) *** Proposed Solution *** This patch introduces a concept of group of rules (GroupMatcher) that share some predicates and only get checked once for the whole group. This patch only creates groups with one nesting level. Conceptually there is nothing preventing us for having deeper nest level. However, the current implementation is not smart enough to share the recording (aka capturing) of values. That limits its ability to do more sharing. For the given example the current patch will generate: // First group CheckOpcode G_ADD // First pattern CheckNumOperand 3 ... Build ADDrr // Second pattern CheckNumOperand 3 ... Build ADDri // Second group CheckOpcode G_SUB // Third pattern CheckNumOperand 3 ... Build SUBrr But if we allowed several nesting level, it could create a sub group for the checknumoperand 3. (We would need to call optimizeRules on the rules within a group.) *** Result *** With only one level of nesting, the instruction selection pass is up to 4x faster. For instance, one instruction now takes 500 checks, instead of 24k! With more nesting we could get in the tens I believe. Differential Revision: https://reviews.llvm.org/D39034 rdar://problem/34670699 llvm-svn: 321017
2017-12-19 03:47:41 +08:00
}
Reapply r332907 "[GlobalISel] Improving InstructionSelect's performance by reducing MatchTable..." Apparently the compile time problem was caused by the fact that not all compilers / STL implementations can automatically convert std::unique_ptr<Derived> to std::unique_ptr<Base>. Fixed (hopefully) by making sure it's std::unique_ptr<Derived>&& (rvalue ref) to std::unique_ptr<Base> conversion instead. llvm-svn: 332917
2018-05-22 07:28:51 +08:00
bool hasFirstCondition() const override { return !Conditions.empty(); }
[TableGen][GlobalISel] Refactor optimizeRules related bit to allow code reuse In theory, reapplying optimizeRules on each group matchers should give us a second nesting level on the matching table. In practice, we need more work to make that happen because all the predicates are actually not directly available through the predicate matchers list. NFC. llvm-svn: 321025
2017-12-19 05:25:53 +08:00
Reapply r332907 "[GlobalISel] Improving InstructionSelect's performance by reducing MatchTable..." Apparently the compile time problem was caused by the fact that not all compilers / STL implementations can automatically convert std::unique_ptr<Derived> to std::unique_ptr<Base>. Fixed (hopefully) by making sure it's std::unique_ptr<Derived>&& (rvalue ref) to std::unique_ptr<Base> conversion instead. llvm-svn: 332917
2018-05-22 07:28:51 +08:00
private:
/// See if a candidate matcher could be added to this group solely by
/// analyzing its first condition.
bool candidateConditionMatches(const PredicateMatcher &Predicate) const;
[TableGen][GlobalISel] Optimize MatchTable for faster instruction selection *** Context *** Prior to this patchw, the table generated for matching instruction was straight forward but highly inefficient. Basically, each pattern generates its own set of self contained checks and actions. E.g., TableGen generated: // First pattern CheckNumOperand 3 CheckOpcode G_ADD ... Build ADDrr // Second pattern CheckNumOperand 3 CheckOpcode G_ADD ... Build ADDri // Third pattern CheckNumOperand 3 CheckOpcode G_SUB ... Build SUBrr *** Problem *** Because of that generation, a *lot* of check were redundant between each pattern and were checked every single time until we reach the pattern that matches. E.g., Taking the previous table, let say we are matching a G_SUB, that means we were going to check all the rules for G_ADD before looking at the G_SUB rule. In particular we are going to do: check 3 operands; PASS check G_ADD; FAIL ; Next rule check 3 operands; PASS (but we already knew that!) check G_ADD; FAIL (well it is still not true) ; Next rule check 3 operands; PASS (really!!) check G_SUB; PASS (at last :P) *** Proposed Solution *** This patch introduces a concept of group of rules (GroupMatcher) that share some predicates and only get checked once for the whole group. This patch only creates groups with one nesting level. Conceptually there is nothing preventing us for having deeper nest level. However, the current implementation is not smart enough to share the recording (aka capturing) of values. That limits its ability to do more sharing. For the given example the current patch will generate: // First group CheckOpcode G_ADD // First pattern CheckNumOperand 3 ... Build ADDrr // Second pattern CheckNumOperand 3 ... Build ADDri // Second group CheckOpcode G_SUB // Third pattern CheckNumOperand 3 ... Build SUBrr But if we allowed several nesting level, it could create a sub group for the checknumoperand 3. (We would need to call optimizeRules on the rules within a group.) *** Result *** With only one level of nesting, the instruction selection pass is up to 4x faster. For instance, one instruction now takes 500 checks, instead of 24k! With more nesting we could get in the tens I believe. Differential Revision: https://reviews.llvm.org/D39034 rdar://problem/34670699 llvm-svn: 321017
2017-12-19 03:47:41 +08:00
};
[tablegen][globalisel] Trivial changes to reduce size of following patch. NFC. Summary: * Move namespace { * Trivial: Typo * RuleMatcher: Separate class and definition * Trivial: const findNodeEquiv Reviewers: t.p.northover, qcolombet, aditya_nandakumar, ab, rovka Reviewed By: rovka Subscribers: dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D30531 llvm-svn: 297884
2017-03-16 04:18:38 +08:00
[GlobalISel][InstructionSelect] Switching MatchTable over opcodes, perf patch 4 This patch continues a series of patches started by r332907 (reapplied as r332917) In this commit we introduce a new matching opcode GIM_SwitchOpcode that implements a jump table over opcodes and start emitting them for root instructions. This is expected to decrease time GlobalISel spends in its InstructionSelect pass by roughly 20% for an -O0 build as measured on sqlite3-amalgamation (http://sqlite.org/download.html) targeting AArch64. To some degree, we assume here that the opcodes form a dense set, which is true at the moment for all upstream targets given the limitations of our rule importing mechanism. It might not be true for out of tree targets, specifically due to pseudo's. If so, we might noticeably increase the size of the MatchTable with this patch due to padding zeros. This will be addressed later. Reviewers: qcolombet, dsanders, bogner, aemerson, javed.absar Reviewed By: qcolombet Subscribers: rovka, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D44700 llvm-svn: 333017
2018-05-23 03:37:59 +08:00
class SwitchMatcher : public Matcher {
/// All the nested matchers, representing distinct switch-cases. The first
/// conditions (as Matcher::getFirstCondition() reports) of all the nested
/// matchers must share the same type and path to a value they check, in other
/// words, be isIdenticalDownToValue, but have different values they check
/// against.
std::vector<Matcher *> Matchers;
/// The representative condition, with a type and a path (InsnVarID and OpIdx
/// in most cases) shared by all the matchers contained.
std::unique_ptr<PredicateMatcher> Condition = nullptr;
/// Temporary set used to check that the case values don't repeat within the
/// same switch.
std::set<MatchTableRecord> Values;
/// An owning collection for any auxiliary matchers created while optimizing
/// nested matchers contained.
std::vector<std::unique_ptr<Matcher>> MatcherStorage;
public:
bool addMatcher(Matcher &Candidate);
void finalize();
void emit(MatchTable &Table) override;
iterator_range<std::vector<Matcher *>::iterator> matchers() {
return make_range(Matchers.begin(), Matchers.end());
}
size_t size() const { return Matchers.size(); }
bool empty() const { return Matchers.empty(); }
std::unique_ptr<PredicateMatcher> popFirstCondition() override {
// SwitchMatcher doesn't have a common first condition for its cases, as all
// the cases only share a kind of a value (a type and a path to it) they
// match, but deliberately differ in the actual value they match.
llvm_unreachable("Trying to pop a condition from a condition-less group");
}
const PredicateMatcher &getFirstCondition() const override {
llvm_unreachable("Trying to pop a condition from a condition-less group");
}
bool hasFirstCondition() const override { return false; }
private:
/// See if the predicate type has a Switch-implementation for it.
static bool isSupportedPredicateType(const PredicateMatcher &Predicate);
bool candidateConditionMatches(const PredicateMatcher &Predicate) const;
/// emit()-helper
static void emitPredicateSpecificOpcodes(const PredicateMatcher &P,
MatchTable &Table);
};
[tablegen][globalisel] Trivial changes to reduce size of following patch. NFC. Summary: * Move namespace { * Trivial: Typo * RuleMatcher: Separate class and definition * Trivial: const findNodeEquiv Reviewers: t.p.northover, qcolombet, aditya_nandakumar, ab, rovka Reviewed By: rovka Subscribers: dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D30531 llvm-svn: 297884
2017-03-16 04:18:38 +08:00
/// Generates code to check that a match rule matches.
[TableGen][GlobalISel] Optimize MatchTable for faster instruction selection *** Context *** Prior to this patchw, the table generated for matching instruction was straight forward but highly inefficient. Basically, each pattern generates its own set of self contained checks and actions. E.g., TableGen generated: // First pattern CheckNumOperand 3 CheckOpcode G_ADD ... Build ADDrr // Second pattern CheckNumOperand 3 CheckOpcode G_ADD ... Build ADDri // Third pattern CheckNumOperand 3 CheckOpcode G_SUB ... Build SUBrr *** Problem *** Because of that generation, a *lot* of check were redundant between each pattern and were checked every single time until we reach the pattern that matches. E.g., Taking the previous table, let say we are matching a G_SUB, that means we were going to check all the rules for G_ADD before looking at the G_SUB rule. In particular we are going to do: check 3 operands; PASS check G_ADD; FAIL ; Next rule check 3 operands; PASS (but we already knew that!) check G_ADD; FAIL (well it is still not true) ; Next rule check 3 operands; PASS (really!!) check G_SUB; PASS (at last :P) *** Proposed Solution *** This patch introduces a concept of group of rules (GroupMatcher) that share some predicates and only get checked once for the whole group. This patch only creates groups with one nesting level. Conceptually there is nothing preventing us for having deeper nest level. However, the current implementation is not smart enough to share the recording (aka capturing) of values. That limits its ability to do more sharing. For the given example the current patch will generate: // First group CheckOpcode G_ADD // First pattern CheckNumOperand 3 ... Build ADDrr // Second pattern CheckNumOperand 3 ... Build ADDri // Second group CheckOpcode G_SUB // Third pattern CheckNumOperand 3 ... Build SUBrr But if we allowed several nesting level, it could create a sub group for the checknumoperand 3. (We would need to call optimizeRules on the rules within a group.) *** Result *** With only one level of nesting, the instruction selection pass is up to 4x faster. For instance, one instruction now takes 500 checks, instead of 24k! With more nesting we could get in the tens I believe. Differential Revision: https://reviews.llvm.org/D39034 rdar://problem/34670699 llvm-svn: 321017
2017-12-19 03:47:41 +08:00
class RuleMatcher : public Matcher {
Re-commit: [globalisel][tablegen] Keep track of the insertion point while adding BuildMIAction's. NFC Multi-instruction emission needs to ensure the the instructions are generated a depth-first fashion. For example: (ADDWrr (SUBWrr a, b), c) needs to emit the SUBWrr before the ADDWrr. However, our walk over TreePatternNode's is highly context sensitive which makes it difficult to append BuildMIActions in the order we want. To fix this, we now keep track of the insertion point as we add actions. This will allow multi-insn emission to insert BuildMI's in the correct place. The previous commit failed on the Ubuntu bots using GCC 4.8. These bots lack the const_iterator forms of insert() and emplace() that were added in C++11. As a result I've switched the const_iterators to iterators. llvm-svn: 317049
2017-11-01 07:03:18 +08:00
public:
[ARM][GISel] PR35965 Constrain RegClasses of nested instructions built from Dst Pattern Summary: Apparently, we missed on constraining register classes of VReg-operands of all the instructions built from a destination pattern but the root (top-level) one. The issue exposed itself while selecting G_FPTOSI for armv7: the corresponding pattern generates VTOSIZS wrapped into COPY_TO_REGCLASS, so top-level COPY_TO_REGCLASS gets properly constrained, while nested VTOSIZS (or rather its destination virtual register to be exact) does not. Fixing this by issuing GIR_ConstrainSelectedInstOperands for every nested GIR_BuildMI. https://bugs.llvm.org/show_bug.cgi?id=35965 rdar://problem/36886530 Patch by Roman Tereshin Reviewers: dsanders, qcolombet, rovka, bogner, aditya_nandakumar, volkan Reviewed By: dsanders, qcolombet, rovka Subscribers: aemerson, javed.absar, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D42565 llvm-svn: 323692
2018-01-30 05:09:12 +08:00
using ActionList = std::list<std::unique_ptr<MatchAction>>;
using action_iterator = ActionList::iterator;
Re-commit: [globalisel][tablegen] Keep track of the insertion point while adding BuildMIAction's. NFC Multi-instruction emission needs to ensure the the instructions are generated a depth-first fashion. For example: (ADDWrr (SUBWrr a, b), c) needs to emit the SUBWrr before the ADDWrr. However, our walk over TreePatternNode's is highly context sensitive which makes it difficult to append BuildMIActions in the order we want. To fix this, we now keep track of the insertion point as we add actions. This will allow multi-insn emission to insert BuildMI's in the correct place. The previous commit failed on the Ubuntu bots using GCC 4.8. These bots lack the const_iterator forms of insert() and emplace() that were added in C++11. As a result I've switched the const_iterators to iterators. llvm-svn: 317049
2017-11-01 07:03:18 +08:00
protected:
[tablegen][globalisel] Trivial changes to reduce size of following patch. NFC. Summary: * Move namespace { * Trivial: Typo * RuleMatcher: Separate class and definition * Trivial: const findNodeEquiv Reviewers: t.p.northover, qcolombet, aditya_nandakumar, ab, rovka Reviewed By: rovka Subscribers: dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D30531 llvm-svn: 297884
2017-03-16 04:18:38 +08:00
/// A list of matchers that all need to succeed for the current rule to match.
/// FIXME: This currently supports a single match position but could be
/// extended to support multiple positions to support div/rem fusion or
/// load-multiple instructions.
Reapply r332907 "[GlobalISel] Improving InstructionSelect's performance by reducing MatchTable..." Apparently the compile time problem was caused by the fact that not all compilers / STL implementations can automatically convert std::unique_ptr<Derived> to std::unique_ptr<Base>. Fixed (hopefully) by making sure it's std::unique_ptr<Derived>&& (rvalue ref) to std::unique_ptr<Base> conversion instead. llvm-svn: 332917
2018-05-22 07:28:51 +08:00
using MatchersTy = std::vector<std::unique_ptr<InstructionMatcher>> ;
MatchersTy Matchers;
[tablegen][globalisel] Trivial changes to reduce size of following patch. NFC. Summary: * Move namespace { * Trivial: Typo * RuleMatcher: Separate class and definition * Trivial: const findNodeEquiv Reviewers: t.p.northover, qcolombet, aditya_nandakumar, ab, rovka Reviewed By: rovka Subscribers: dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D30531 llvm-svn: 297884
2017-03-16 04:18:38 +08:00
/// A list of actions that need to be taken when all predicates in this rule
/// have succeeded.
[ARM][GISel] PR35965 Constrain RegClasses of nested instructions built from Dst Pattern Summary: Apparently, we missed on constraining register classes of VReg-operands of all the instructions built from a destination pattern but the root (top-level) one. The issue exposed itself while selecting G_FPTOSI for armv7: the corresponding pattern generates VTOSIZS wrapped into COPY_TO_REGCLASS, so top-level COPY_TO_REGCLASS gets properly constrained, while nested VTOSIZS (or rather its destination virtual register to be exact) does not. Fixing this by issuing GIR_ConstrainSelectedInstOperands for every nested GIR_BuildMI. https://bugs.llvm.org/show_bug.cgi?id=35965 rdar://problem/36886530 Patch by Roman Tereshin Reviewers: dsanders, qcolombet, rovka, bogner, aditya_nandakumar, volkan Reviewed By: dsanders, qcolombet, rovka Subscribers: aemerson, javed.absar, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D42565 llvm-svn: 323692
2018-01-30 05:09:12 +08:00
ActionList Actions;
[tablegen][globalisel] Trivial changes to reduce size of following patch. NFC. Summary: * Move namespace { * Trivial: Typo * RuleMatcher: Separate class and definition * Trivial: const findNodeEquiv Reviewers: t.p.northover, qcolombet, aditya_nandakumar, ab, rovka Reviewed By: rovka Subscribers: dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D30531 llvm-svn: 297884
2017-03-16 04:18:38 +08:00
Reapply r332907 "[GlobalISel] Improving InstructionSelect's performance by reducing MatchTable..." Apparently the compile time problem was caused by the fact that not all compilers / STL implementations can automatically convert std::unique_ptr<Derived> to std::unique_ptr<Base>. Fixed (hopefully) by making sure it's std::unique_ptr<Derived>&& (rvalue ref) to std::unique_ptr<Base> conversion instead. llvm-svn: 332917
2018-05-22 07:28:51 +08:00
using DefinedInsnVariablesMap = std::map<InstructionMatcher *, unsigned>;
[globalisel][tablegen] Add infrastructure to potentially allow BuildMIAction to choose a mutatable instruction. NFC Prepare for multiple instruction emission by allowing BuildMIAction to search for a suitable matcher that will support mutation. This patch deliberately neglects to add matchers aside from the root to preserve NFC. That said, it should be noted that until we support mutations other than just the opcode the chances of finding a non-root instruction for which canMutate() is true, is essentially zero. Furthermore in the presence of multi-instruction emission the chances of finding any instruction for which canMutate() is true is also zero. Nevertheless, we can't continue to require that all BuildMIAction's consider the root of the match to be recyclable due to the risk of recycling it twice in the same rule. llvm-svn: 317022
2017-11-01 02:50:24 +08:00
Reapply r332907 "[GlobalISel] Improving InstructionSelect's performance by reducing MatchTable..." Apparently the compile time problem was caused by the fact that not all compilers / STL implementations can automatically convert std::unique_ptr<Derived> to std::unique_ptr<Base>. Fixed (hopefully) by making sure it's std::unique_ptr<Derived>&& (rvalue ref) to std::unique_ptr<Base> conversion instead. llvm-svn: 332917
2018-05-22 07:28:51 +08:00
/// A map of instruction matchers to the local variables
[globalisel][tablegen] Do not merge memoperands from instructions that weren't in the match. Summary: Fix a bug discovered in an out-of-tree target where memoperands from pseudo-instructions that weren't part of the match were being merged into the result instructions as part of GIR_MergeMemOperands. This bug was caused by a change to the handling of State.MIs between rules when the state machine tables were fused into a single table. Previously, each rule would reset State.MIs using State.MIs.resize(1) but this is no longer done, as a result stale data is occasionally left in some elements of State.MIs. Most opcodes aren't affected by this but GIR_MergeMemOperands merges all memoperands from the intructions recorded in State.MIs into the result instruction. Suppose for example, we processed but rejected the following pattern: (signextend (load x)) at this point, State.MIs contains the signextend and the load. Now suppose we process and accept this pattern: (add x, y) at this point, State.MIs contains the add as well as the (now irrelevant) load. When GIR_MergeMemOperands is processed, the memoperands from that irrelevant load will be merged into the result instruction even though it was not part of the match. Bringing back the State.MIs.resize(1) would fix the problem but it would limit our ability to optimize the table in the future. Instead, this patch fixes the problem by explicitly stating which instructions should be merged into the result. There's no direct test case in this commit because a test case would be very brittle. However, at the time of writing this should fix the failures in http://green.lab.llvm.org/green/job/Compiler_Verifiers_GlobalISEL/ as well as a failure in test/CodeGen/ARM/GlobalISel/arm-isel.ll when expensive checks are enabled. Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Subscribers: fhahn, kristof.beyls, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D36094 llvm-svn: 309804
2017-08-02 19:03:36 +08:00
DefinedInsnVariablesMap InsnVariableIDs;
[tablegen][globalisel] Capture instructions into locals and related infrastructure for multiple instructions matches. Summary: Prepare the way for nested instruction matching support by having actions like CopyRenderer look up operands in the RuleMatcher rather than a specific InstructionMatcher. This allows actions to reference any operand from any matched instruction. It works by checking the 'shape' of the match and capturing each matched instruction to a local variable. If the shape is wrong (not enough operands, leaf nodes where non-leafs are expected, etc.), then the rule exits early without checking the predicates. Once we've captured the instructions, we then test the predicates as before (except using the local variables). If the match is successful, then we render the new instruction as before using the local variables. It's not noticable in this patch but by the time we support multiple instruction matching, this patch will also cause a significant improvement to readability of the emitted code since MRI.getVRegDef(I->getOperand(0).getReg()) will simply be MI1 after emitCxxCaptureStmts(). This isn't quite NFC because I've also fixed a bug that I'm surprised we haven't encountered yet. It now checks there are at least the expected number of operands before accessing them with getOperand(). Depends on D30531 Reviewers: t.p.northover, qcolombet, aditya_nandakumar, ab, rovka Reviewed By: rovka Subscribers: dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D30535 llvm-svn: 298257
2017-03-20 23:20:42 +08:00
Reapply r332907 "[GlobalISel] Improving InstructionSelect's performance by reducing MatchTable..." Apparently the compile time problem was caused by the fact that not all compilers / STL implementations can automatically convert std::unique_ptr<Derived> to std::unique_ptr<Base>. Fixed (hopefully) by making sure it's std::unique_ptr<Derived>&& (rvalue ref) to std::unique_ptr<Base> conversion instead. llvm-svn: 332917
2018-05-22 07:28:51 +08:00
using MutatableInsnSet = SmallPtrSet<InstructionMatcher *, 4>;
[globalisel][tablegen] Add infrastructure to potentially allow BuildMIAction to choose a mutatable instruction. NFC Prepare for multiple instruction emission by allowing BuildMIAction to search for a suitable matcher that will support mutation. This patch deliberately neglects to add matchers aside from the root to preserve NFC. That said, it should be noted that until we support mutations other than just the opcode the chances of finding a non-root instruction for which canMutate() is true, is essentially zero. Furthermore in the presence of multi-instruction emission the chances of finding any instruction for which canMutate() is true is also zero. Nevertheless, we can't continue to require that all BuildMIAction's consider the root of the match to be recyclable due to the risk of recycling it twice in the same rule. llvm-svn: 317022
2017-11-01 02:50:24 +08:00
// The set of instruction matchers that have not yet been claimed for mutation
// by a BuildMI.
MutatableInsnSet MutatableInsns;
[globalisel][tablegen] Simplify named operand/operator lookups and fix a wrong-code bug this revealed. Summary: Operand variable lookups are now performed by the RuleMatcher rather than searching the whole matcher hierarchy for a match. This revealed a wrong-code bug that currently affects ARM and X86 where patterns that use a variable more than once in the match pattern will be imported but won't check that the operands are identical. This can cause the tablegen-erated matcher to accept matches that should be rejected. Depends on D36569 Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Subscribers: aemerson, igorb, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D36618 llvm-svn: 315780
2017-10-14 08:31:58 +08:00
/// A map of named operands defined by the matchers that may be referenced by
/// the renderers.
StringMap<OperandMatcher *> DefinedOperands;
Reapply r332907 "[GlobalISel] Improving InstructionSelect's performance by reducing MatchTable..." Apparently the compile time problem was caused by the fact that not all compilers / STL implementations can automatically convert std::unique_ptr<Derived> to std::unique_ptr<Base>. Fixed (hopefully) by making sure it's std::unique_ptr<Derived>&& (rvalue ref) to std::unique_ptr<Base> conversion instead. llvm-svn: 332917
2018-05-22 07:28:51 +08:00
/// ID for the next instruction variable defined with implicitlyDefineInsnVar()
[tablegen][globalisel] Capture instructions into locals and related infrastructure for multiple instructions matches. Summary: Prepare the way for nested instruction matching support by having actions like CopyRenderer look up operands in the RuleMatcher rather than a specific InstructionMatcher. This allows actions to reference any operand from any matched instruction. It works by checking the 'shape' of the match and capturing each matched instruction to a local variable. If the shape is wrong (not enough operands, leaf nodes where non-leafs are expected, etc.), then the rule exits early without checking the predicates. Once we've captured the instructions, we then test the predicates as before (except using the local variables). If the match is successful, then we render the new instruction as before using the local variables. It's not noticable in this patch but by the time we support multiple instruction matching, this patch will also cause a significant improvement to readability of the emitted code since MRI.getVRegDef(I->getOperand(0).getReg()) will simply be MI1 after emitCxxCaptureStmts(). This isn't quite NFC because I've also fixed a bug that I'm surprised we haven't encountered yet. It now checks there are at least the expected number of operands before accessing them with getOperand(). Depends on D30531 Reviewers: t.p.northover, qcolombet, aditya_nandakumar, ab, rovka Reviewed By: rovka Subscribers: dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D30535 llvm-svn: 298257
2017-03-20 23:20:42 +08:00
unsigned NextInsnVarID;
[globalisel][tablegen] Stop hard-coding the emitted instruction ID to 0. NFC The next commit will add support for multi-instruction emission so we need to start allocating instruction ID's instead of hard-coding them to 0. llvm-svn: 317057
2017-11-01 08:29:47 +08:00
/// ID for the next output instruction allocated with allocateOutputInsnID()
unsigned NextOutputInsnID;
[globalisel][tablegen] Add support for multi-insn emission The importer will now accept nested instructions in the result pattern such as (ADDWrr $a, (SUBWrr $b, $c)). This is only valid when the nested instruction def's a single vreg and the parent instruction consumes a single vreg where a nested instruction is specified. The importer will automatically create a vreg to connect the two using the type information from the pattern. This vreg will be constrained to the register classes given in the instruction definitions*. * REG_SEQUENCE is explicitly rejected because of this. The definition doesn't constrain to a register class and it therefore needs special handling. llvm-svn: 317117
2017-11-02 03:57:57 +08:00
/// ID for the next temporary register ID allocated with allocateTempRegID()
unsigned NextTempRegID;
[globalisel][tablegen] Import SelectionDAG's rule predicates and support the equivalent in GIRule. Summary: The SelectionDAG importer now imports rules with Predicate's attached via Requires, PredicateControl, etc. These predicates are implemented as bitset's to allow multiple predicates to be tested together. However, unlike the MC layer subtarget features, each target only pays for it's own predicates (e.g. AArch64 doesn't have 192 feature bits just because X86 needs a lot). Both AArch64 and X86 derive at least one predicate from the MachineFunction or Function so they must re-initialize AvailableFeatures before each function. They also declare locals in <Target>InstructionSelector so that computeAvailableFeatures() can use the code from SelectionDAG without modification. Reviewers: rovka, qcolombet, aditya_nandakumar, t.p.northover, ab Reviewed By: rovka Subscribers: aemerson, rengolin, dberris, kristof.beyls, llvm-commits, igorb Differential Revision: https://reviews.llvm.org/D31418 llvm-svn: 300993
2017-04-21 23:59:56 +08:00
std::vector<Record *> RequiredFeatures;
Reapply r332907 "[GlobalISel] Improving InstructionSelect's performance by reducing MatchTable..." Apparently the compile time problem was caused by the fact that not all compilers / STL implementations can automatically convert std::unique_ptr<Derived> to std::unique_ptr<Base>. Fixed (hopefully) by making sure it's std::unique_ptr<Derived>&& (rvalue ref) to std::unique_ptr<Base> conversion instead. llvm-svn: 332917
2018-05-22 07:28:51 +08:00
std::vector<std::unique_ptr<PredicateMatcher>> EpilogueMatchers;
[globalisel][tablegen] Import SelectionDAG's rule predicates and support the equivalent in GIRule. Summary: The SelectionDAG importer now imports rules with Predicate's attached via Requires, PredicateControl, etc. These predicates are implemented as bitset's to allow multiple predicates to be tested together. However, unlike the MC layer subtarget features, each target only pays for it's own predicates (e.g. AArch64 doesn't have 192 feature bits just because X86 needs a lot). Both AArch64 and X86 derive at least one predicate from the MachineFunction or Function so they must re-initialize AvailableFeatures before each function. They also declare locals in <Target>InstructionSelector so that computeAvailableFeatures() can use the code from SelectionDAG without modification. Reviewers: rovka, qcolombet, aditya_nandakumar, t.p.northover, ab Reviewed By: rovka Subscribers: aemerson, rengolin, dberris, kristof.beyls, llvm-commits, igorb Differential Revision: https://reviews.llvm.org/D31418 llvm-svn: 300993
2017-04-21 23:59:56 +08:00
[globalisel][tablegen] Simplify named operand/operator lookups and fix a wrong-code bug this revealed. Summary: Operand variable lookups are now performed by the RuleMatcher rather than searching the whole matcher hierarchy for a match. This revealed a wrong-code bug that currently affects ARM and X86 where patterns that use a variable more than once in the match pattern will be imported but won't check that the operands are identical. This can cause the tablegen-erated matcher to accept matches that should be rejected. Depends on D36569 Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Subscribers: aemerson, igorb, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D36618 llvm-svn: 315780
2017-10-14 08:31:58 +08:00
ArrayRef<SMLoc> SrcLoc;
Re-commit r315863: [globalisel][tablegen] Import ComplexPattern when used as an operator Summary: It's possible for a ComplexPattern to be used as an operator in a match pattern. This is used by the load/store patterns in AArch64 to name the suboperands returned by ComplexPattern predicate so that they can be broken apart and referenced independently in the result pattern. This patch adds support for this in order to enable the import of load/store patterns. Depends on D37445 Hopefully fixed the ambiguous constructor that a large number of bots reported. Reviewers: ab, qcolombet, t.p.northover, rovka, aditya_nandakumar Reviewed By: qcolombet Subscribers: aemerson, javed.absar, igorb, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D37456 llvm-svn: 315869
2017-10-16 02:22:54 +08:00
typedef std::tuple<Record *, unsigned, unsigned>
DefinedComplexPatternSubOperand;
typedef StringMap<DefinedComplexPatternSubOperand>
DefinedComplexPatternSubOperandMap;
/// A map of Symbolic Names to ComplexPattern sub-operands.
DefinedComplexPatternSubOperandMap ComplexSubOperands;
[globalisel][tablegen] Generate rule coverage and use it to identify untested rules Summary: This patch adds a LLVM_ENABLE_GISEL_COV which, like LLVM_ENABLE_DAGISEL_COV, causes TableGen to instrument the generated table to collect rule coverage information. However, LLVM_ENABLE_GISEL_COV goes a bit further than LLVM_ENABLE_DAGISEL_COV. The information is written to files (${CMAKE_BINARY_DIR}/gisel-coverage-* by default). These files can then be concatenated into ${LLVM_GISEL_COV_PREFIX}-all after which TableGen will read this information and use it to emit warnings about untested rules. This technique could also be used by SelectionDAG and can be further extended to detect hot rules and give them priority over colder rules. Usage: * Enable LLVM_ENABLE_GISEL_COV in CMake * Build the compiler and run some tests * cat gisel-coverage-[0-9]* > gisel-coverage-all * Delete lib/Target/*/*GenGlobalISel.inc* * Build the compiler Known issues: * ${LLVM_GISEL_COV_PREFIX}-all must be generated as a manual step due to a lack of a portable 'cat' command. It should be the concatenation of all ${LLVM_GISEL_COV_PREFIX}-[0-9]* files. * There's no mechanism to discard coverage information when the ruleset changes Depends on D39742 Reviewers: ab, qcolombet, t.p.northover, aditya_nandakumar, rovka Reviewed By: rovka Subscribers: vsk, arsenm, nhaehnle, mgorny, kristof.beyls, javed.absar, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D39747 llvm-svn: 318356
2017-11-16 08:46:35 +08:00
uint64_t RuleID;
static uint64_t NextRuleID;
[tablegen][globalisel] Trivial changes to reduce size of following patch. NFC. Summary: * Move namespace { * Trivial: Typo * RuleMatcher: Separate class and definition * Trivial: const findNodeEquiv Reviewers: t.p.northover, qcolombet, aditya_nandakumar, ab, rovka Reviewed By: rovka Subscribers: dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D30531 llvm-svn: 297884
2017-03-16 04:18:38 +08:00
public:
[globalisel][tablegen] Simplify named operand/operator lookups and fix a wrong-code bug this revealed. Summary: Operand variable lookups are now performed by the RuleMatcher rather than searching the whole matcher hierarchy for a match. This revealed a wrong-code bug that currently affects ARM and X86 where patterns that use a variable more than once in the match pattern will be imported but won't check that the operands are identical. This can cause the tablegen-erated matcher to accept matches that should be rejected. Depends on D36569 Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Subscribers: aemerson, igorb, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D36618 llvm-svn: 315780
2017-10-14 08:31:58 +08:00
RuleMatcher(ArrayRef<SMLoc> SrcLoc)
[globalisel][tablegen] Add infrastructure to potentially allow BuildMIAction to choose a mutatable instruction. NFC Prepare for multiple instruction emission by allowing BuildMIAction to search for a suitable matcher that will support mutation. This patch deliberately neglects to add matchers aside from the root to preserve NFC. That said, it should be noted that until we support mutations other than just the opcode the chances of finding a non-root instruction for which canMutate() is true, is essentially zero. Furthermore in the presence of multi-instruction emission the chances of finding any instruction for which canMutate() is true is also zero. Nevertheless, we can't continue to require that all BuildMIAction's consider the root of the match to be recyclable due to the risk of recycling it twice in the same rule. llvm-svn: 317022
2017-11-01 02:50:24 +08:00
: Matchers(), Actions(), InsnVariableIDs(), MutatableInsns(),
[globalisel][tablegen] Stop hard-coding the emitted instruction ID to 0. NFC The next commit will add support for multi-instruction emission so we need to start allocating instruction ID's instead of hard-coding them to 0. llvm-svn: 317057
2017-11-01 08:29:47 +08:00
DefinedOperands(), NextInsnVarID(0), NextOutputInsnID(0),
[globalisel][tablegen] Generate rule coverage and use it to identify untested rules Summary: This patch adds a LLVM_ENABLE_GISEL_COV which, like LLVM_ENABLE_DAGISEL_COV, causes TableGen to instrument the generated table to collect rule coverage information. However, LLVM_ENABLE_GISEL_COV goes a bit further than LLVM_ENABLE_DAGISEL_COV. The information is written to files (${CMAKE_BINARY_DIR}/gisel-coverage-* by default). These files can then be concatenated into ${LLVM_GISEL_COV_PREFIX}-all after which TableGen will read this information and use it to emit warnings about untested rules. This technique could also be used by SelectionDAG and can be further extended to detect hot rules and give them priority over colder rules. Usage: * Enable LLVM_ENABLE_GISEL_COV in CMake * Build the compiler and run some tests * cat gisel-coverage-[0-9]* > gisel-coverage-all * Delete lib/Target/*/*GenGlobalISel.inc* * Build the compiler Known issues: * ${LLVM_GISEL_COV_PREFIX}-all must be generated as a manual step due to a lack of a portable 'cat' command. It should be the concatenation of all ${LLVM_GISEL_COV_PREFIX}-[0-9]* files. * There's no mechanism to discard coverage information when the ruleset changes Depends on D39742 Reviewers: ab, qcolombet, t.p.northover, aditya_nandakumar, rovka Reviewed By: rovka Subscribers: vsk, arsenm, nhaehnle, mgorny, kristof.beyls, javed.absar, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D39747 llvm-svn: 318356
2017-11-16 08:46:35 +08:00
NextTempRegID(0), SrcLoc(SrcLoc), ComplexSubOperands(),
RuleID(NextRuleID++) {}
Explicitly add move constructor/assignment operators. These are needed due to some obscure rules in the standard about how std::vector selects between copy and move constructors, which can cause a conforming implementation to attempt to select the copy constructor of RuleMatcher, which will fail since std::unique_ptr<> isn't copyable. llvm-svn: 298294
2017-03-21 03:56:52 +08:00
RuleMatcher(RuleMatcher &&Other) = default;
RuleMatcher &operator=(RuleMatcher &&Other) = default;
[tablegen][globalisel] Trivial changes to reduce size of following patch. NFC. Summary: * Move namespace { * Trivial: Typo * RuleMatcher: Separate class and definition * Trivial: const findNodeEquiv Reviewers: t.p.northover, qcolombet, aditya_nandakumar, ab, rovka Reviewed By: rovka Subscribers: dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D30531 llvm-svn: 297884
2017-03-16 04:18:38 +08:00
[globalisel][tablegen] Generate rule coverage and use it to identify untested rules Summary: This patch adds a LLVM_ENABLE_GISEL_COV which, like LLVM_ENABLE_DAGISEL_COV, causes TableGen to instrument the generated table to collect rule coverage information. However, LLVM_ENABLE_GISEL_COV goes a bit further than LLVM_ENABLE_DAGISEL_COV. The information is written to files (${CMAKE_BINARY_DIR}/gisel-coverage-* by default). These files can then be concatenated into ${LLVM_GISEL_COV_PREFIX}-all after which TableGen will read this information and use it to emit warnings about untested rules. This technique could also be used by SelectionDAG and can be further extended to detect hot rules and give them priority over colder rules. Usage: * Enable LLVM_ENABLE_GISEL_COV in CMake * Build the compiler and run some tests * cat gisel-coverage-[0-9]* > gisel-coverage-all * Delete lib/Target/*/*GenGlobalISel.inc* * Build the compiler Known issues: * ${LLVM_GISEL_COV_PREFIX}-all must be generated as a manual step due to a lack of a portable 'cat' command. It should be the concatenation of all ${LLVM_GISEL_COV_PREFIX}-[0-9]* files. * There's no mechanism to discard coverage information when the ruleset changes Depends on D39742 Reviewers: ab, qcolombet, t.p.northover, aditya_nandakumar, rovka Reviewed By: rovka Subscribers: vsk, arsenm, nhaehnle, mgorny, kristof.beyls, javed.absar, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D39747 llvm-svn: 318356
2017-11-16 08:46:35 +08:00
uint64_t getRuleID() const { return RuleID; }
[globalisel][tablegen] Add support for importing 'imm' operands. Summary: This patch enables the import of rules containing 'imm' operands that do not constrain the acceptable values using predicates. Support for ImmLeaf will arrive in a later patch. Depends on D35681 Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Reviewed By: rovka Subscribers: kristof.beyls, javed.absar, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D35833 llvm-svn: 310343
2017-08-08 18:44:31 +08:00
InstructionMatcher &addInstructionMatcher(StringRef SymbolicName);
[globalisel][tablegen] Import SelectionDAG's rule predicates and support the equivalent in GIRule. Summary: The SelectionDAG importer now imports rules with Predicate's attached via Requires, PredicateControl, etc. These predicates are implemented as bitset's to allow multiple predicates to be tested together. However, unlike the MC layer subtarget features, each target only pays for it's own predicates (e.g. AArch64 doesn't have 192 feature bits just because X86 needs a lot). Both AArch64 and X86 derive at least one predicate from the MachineFunction or Function so they must re-initialize AvailableFeatures before each function. They also declare locals in <Target>InstructionSelector so that computeAvailableFeatures() can use the code from SelectionDAG without modification. Reviewers: rovka, qcolombet, aditya_nandakumar, t.p.northover, ab Reviewed By: rovka Subscribers: aemerson, rengolin, dberris, kristof.beyls, llvm-commits, igorb Differential Revision: https://reviews.llvm.org/D31418 llvm-svn: 300993
2017-04-21 23:59:56 +08:00
void addRequiredFeature(Record *Feature);
[globalisel][tablegen] Partially fix compile-time regressions by converting matcher to state-machine(s) Summary: Replace the matcher if-statements for each rule with a state-machine. This significantly reduces compile time, memory allocations, and cumulative memory allocation when compiling AArch64InstructionSelector.cpp.o after r303259 is recommitted. The following patches will expand on this further to fully fix the regressions. Reviewers: rovka, ab, t.p.northover, qcolombet, aditya_nandakumar Reviewed By: ab Subscribers: vitalybuka, aemerson, javed.absar, igorb, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D33758 llvm-svn: 307079
2017-07-04 22:35:06 +08:00
const std::vector<Record *> &getRequiredFeatures() const;
[tablegen][globalisel] Trivial changes to reduce size of following patch. NFC. Summary: * Move namespace { * Trivial: Typo * RuleMatcher: Separate class and definition * Trivial: const findNodeEquiv Reviewers: t.p.northover, qcolombet, aditya_nandakumar, ab, rovka Reviewed By: rovka Subscribers: dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D30531 llvm-svn: 297884
2017-03-16 04:18:38 +08:00
template <class Kind, class... Args> Kind &addAction(Args &&... args);
Re-commit: [globalisel][tablegen] Keep track of the insertion point while adding BuildMIAction's. NFC Multi-instruction emission needs to ensure the the instructions are generated a depth-first fashion. For example: (ADDWrr (SUBWrr a, b), c) needs to emit the SUBWrr before the ADDWrr. However, our walk over TreePatternNode's is highly context sensitive which makes it difficult to append BuildMIActions in the order we want. To fix this, we now keep track of the insertion point as we add actions. This will allow multi-insn emission to insert BuildMI's in the correct place. The previous commit failed on the Ubuntu bots using GCC 4.8. These bots lack the const_iterator forms of insert() and emplace() that were added in C++11. As a result I've switched the const_iterators to iterators. llvm-svn: 317049
2017-11-01 07:03:18 +08:00
template <class Kind, class... Args>
action_iterator insertAction(action_iterator InsertPt, Args &&... args);
[tablegen][globalisel] Trivial changes to reduce size of following patch. NFC. Summary: * Move namespace { * Trivial: Typo * RuleMatcher: Separate class and definition * Trivial: const findNodeEquiv Reviewers: t.p.northover, qcolombet, aditya_nandakumar, ab, rovka Reviewed By: rovka Subscribers: dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D30531 llvm-svn: 297884
2017-03-16 04:18:38 +08:00
[globalisel][tablegen] Partially fix compile-time regressions by converting matcher to state-machine(s) Summary: Replace the matcher if-statements for each rule with a state-machine. This significantly reduces compile time, memory allocations, and cumulative memory allocation when compiling AArch64InstructionSelector.cpp.o after r303259 is recommitted. The following patches will expand on this further to fully fix the regressions. Reviewers: rovka, ab, t.p.northover, qcolombet, aditya_nandakumar Reviewed By: ab Subscribers: vitalybuka, aemerson, javed.absar, igorb, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D33758 llvm-svn: 307079
2017-07-04 22:35:06 +08:00
/// Define an instruction without emitting any code to do so.
Reapply r332907 "[GlobalISel] Improving InstructionSelect's performance by reducing MatchTable..." Apparently the compile time problem was caused by the fact that not all compilers / STL implementations can automatically convert std::unique_ptr<Derived> to std::unique_ptr<Base>. Fixed (hopefully) by making sure it's std::unique_ptr<Derived>&& (rvalue ref) to std::unique_ptr<Base> conversion instead. llvm-svn: 332917
2018-05-22 07:28:51 +08:00
unsigned implicitlyDefineInsnVar(InstructionMatcher &Matcher);
unsigned getInsnVarID(InstructionMatcher &InsnMatcher) const;
[globalisel][tablegen] Do not merge memoperands from instructions that weren't in the match. Summary: Fix a bug discovered in an out-of-tree target where memoperands from pseudo-instructions that weren't part of the match were being merged into the result instructions as part of GIR_MergeMemOperands. This bug was caused by a change to the handling of State.MIs between rules when the state machine tables were fused into a single table. Previously, each rule would reset State.MIs using State.MIs.resize(1) but this is no longer done, as a result stale data is occasionally left in some elements of State.MIs. Most opcodes aren't affected by this but GIR_MergeMemOperands merges all memoperands from the intructions recorded in State.MIs into the result instruction. Suppose for example, we processed but rejected the following pattern: (signextend (load x)) at this point, State.MIs contains the signextend and the load. Now suppose we process and accept this pattern: (add x, y) at this point, State.MIs contains the add as well as the (now irrelevant) load. When GIR_MergeMemOperands is processed, the memoperands from that irrelevant load will be merged into the result instruction even though it was not part of the match. Bringing back the State.MIs.resize(1) would fix the problem but it would limit our ability to optimize the table in the future. Instead, this patch fixes the problem by explicitly stating which instructions should be merged into the result. There's no direct test case in this commit because a test case would be very brittle. However, at the time of writing this should fix the failures in http://green.lab.llvm.org/green/job/Compiler_Verifiers_GlobalISEL/ as well as a failure in test/CodeGen/ARM/GlobalISel/arm-isel.ll when expensive checks are enabled. Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Subscribers: fhahn, kristof.beyls, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D36094 llvm-svn: 309804
2017-08-02 19:03:36 +08:00
DefinedInsnVariablesMap::const_iterator defined_insn_vars_begin() const {
return InsnVariableIDs.begin();
}
DefinedInsnVariablesMap::const_iterator defined_insn_vars_end() const {
return InsnVariableIDs.end();
}
iterator_range<typename DefinedInsnVariablesMap::const_iterator>
defined_insn_vars() const {
return make_range(defined_insn_vars_begin(), defined_insn_vars_end());
}
[tablegen][globalisel] Capture instructions into locals and related infrastructure for multiple instructions matches. Summary: Prepare the way for nested instruction matching support by having actions like CopyRenderer look up operands in the RuleMatcher rather than a specific InstructionMatcher. This allows actions to reference any operand from any matched instruction. It works by checking the 'shape' of the match and capturing each matched instruction to a local variable. If the shape is wrong (not enough operands, leaf nodes where non-leafs are expected, etc.), then the rule exits early without checking the predicates. Once we've captured the instructions, we then test the predicates as before (except using the local variables). If the match is successful, then we render the new instruction as before using the local variables. It's not noticable in this patch but by the time we support multiple instruction matching, this patch will also cause a significant improvement to readability of the emitted code since MRI.getVRegDef(I->getOperand(0).getReg()) will simply be MI1 after emitCxxCaptureStmts(). This isn't quite NFC because I've also fixed a bug that I'm surprised we haven't encountered yet. It now checks there are at least the expected number of operands before accessing them with getOperand(). Depends on D30531 Reviewers: t.p.northover, qcolombet, aditya_nandakumar, ab, rovka Reviewed By: rovka Subscribers: dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D30535 llvm-svn: 298257
2017-03-20 23:20:42 +08:00
[globalisel][tablegen] Add infrastructure to potentially allow BuildMIAction to choose a mutatable instruction. NFC Prepare for multiple instruction emission by allowing BuildMIAction to search for a suitable matcher that will support mutation. This patch deliberately neglects to add matchers aside from the root to preserve NFC. That said, it should be noted that until we support mutations other than just the opcode the chances of finding a non-root instruction for which canMutate() is true, is essentially zero. Furthermore in the presence of multi-instruction emission the chances of finding any instruction for which canMutate() is true is also zero. Nevertheless, we can't continue to require that all BuildMIAction's consider the root of the match to be recyclable due to the risk of recycling it twice in the same rule. llvm-svn: 317022
2017-11-01 02:50:24 +08:00
MutatableInsnSet::const_iterator mutatable_insns_begin() const {
return MutatableInsns.begin();
}
MutatableInsnSet::const_iterator mutatable_insns_end() const {
return MutatableInsns.end();
}
iterator_range<typename MutatableInsnSet::const_iterator>
mutatable_insns() const {
return make_range(mutatable_insns_begin(), mutatable_insns_end());
}
Reapply r332907 "[GlobalISel] Improving InstructionSelect's performance by reducing MatchTable..." Apparently the compile time problem was caused by the fact that not all compilers / STL implementations can automatically convert std::unique_ptr<Derived> to std::unique_ptr<Base>. Fixed (hopefully) by making sure it's std::unique_ptr<Derived>&& (rvalue ref) to std::unique_ptr<Base> conversion instead. llvm-svn: 332917
2018-05-22 07:28:51 +08:00
void reserveInsnMatcherForMutation(InstructionMatcher *InsnMatcher) {
[globalisel][tablegen] Add infrastructure to potentially allow BuildMIAction to choose a mutatable instruction. NFC Prepare for multiple instruction emission by allowing BuildMIAction to search for a suitable matcher that will support mutation. This patch deliberately neglects to add matchers aside from the root to preserve NFC. That said, it should be noted that until we support mutations other than just the opcode the chances of finding a non-root instruction for which canMutate() is true, is essentially zero. Furthermore in the presence of multi-instruction emission the chances of finding any instruction for which canMutate() is true is also zero. Nevertheless, we can't continue to require that all BuildMIAction's consider the root of the match to be recyclable due to the risk of recycling it twice in the same rule. llvm-svn: 317022
2017-11-01 02:50:24 +08:00
bool R = MutatableInsns.erase(InsnMatcher);
assert(R && "Reserving a mutatable insn that isn't available");
(void)R;
}
Re-commit: [globalisel][tablegen] Keep track of the insertion point while adding BuildMIAction's. NFC Multi-instruction emission needs to ensure the the instructions are generated a depth-first fashion. For example: (ADDWrr (SUBWrr a, b), c) needs to emit the SUBWrr before the ADDWrr. However, our walk over TreePatternNode's is highly context sensitive which makes it difficult to append BuildMIActions in the order we want. To fix this, we now keep track of the insertion point as we add actions. This will allow multi-insn emission to insert BuildMI's in the correct place. The previous commit failed on the Ubuntu bots using GCC 4.8. These bots lack the const_iterator forms of insert() and emplace() that were added in C++11. As a result I've switched the const_iterators to iterators. llvm-svn: 317049
2017-11-01 07:03:18 +08:00
action_iterator actions_begin() { return Actions.begin(); }
action_iterator actions_end() { return Actions.end(); }
iterator_range<action_iterator> actions() {
return make_range(actions_begin(), actions_end());
}
[globalisel][tablegen] Simplify named operand/operator lookups and fix a wrong-code bug this revealed. Summary: Operand variable lookups are now performed by the RuleMatcher rather than searching the whole matcher hierarchy for a match. This revealed a wrong-code bug that currently affects ARM and X86 where patterns that use a variable more than once in the match pattern will be imported but won't check that the operands are identical. This can cause the tablegen-erated matcher to accept matches that should be rejected. Depends on D36569 Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Subscribers: aemerson, igorb, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D36618 llvm-svn: 315780
2017-10-14 08:31:58 +08:00
void defineOperand(StringRef SymbolicName, OperandMatcher &OM);
Re-commit r315863: [globalisel][tablegen] Import ComplexPattern when used as an operator Summary: It's possible for a ComplexPattern to be used as an operator in a match pattern. This is used by the load/store patterns in AArch64 to name the suboperands returned by ComplexPattern predicate so that they can be broken apart and referenced independently in the result pattern. This patch adds support for this in order to enable the import of load/store patterns. Depends on D37445 Hopefully fixed the ambiguous constructor that a large number of bots reported. Reviewers: ab, qcolombet, t.p.northover, rovka, aditya_nandakumar Reviewed By: qcolombet Subscribers: aemerson, javed.absar, igorb, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D37456 llvm-svn: 315869
2017-10-16 02:22:54 +08:00
void defineComplexSubOperand(StringRef SymbolicName, Record *ComplexPattern,
unsigned RendererID, unsigned SubOperandID) {
assert(ComplexSubOperands.count(SymbolicName) == 0 && "Already defined");
ComplexSubOperands[SymbolicName] =
std::make_tuple(ComplexPattern, RendererID, SubOperandID);
}
Optional<DefinedComplexPatternSubOperand>
getComplexSubOperand(StringRef SymbolicName) const {
const auto &I = ComplexSubOperands.find(SymbolicName);
if (I == ComplexSubOperands.end())
return None;
return I->second;
}
Reapply r332907 "[GlobalISel] Improving InstructionSelect's performance by reducing MatchTable..." Apparently the compile time problem was caused by the fact that not all compilers / STL implementations can automatically convert std::unique_ptr<Derived> to std::unique_ptr<Base>. Fixed (hopefully) by making sure it's std::unique_ptr<Derived>&& (rvalue ref) to std::unique_ptr<Base> conversion instead. llvm-svn: 332917
2018-05-22 07:28:51 +08:00
InstructionMatcher &getInstructionMatcher(StringRef SymbolicName) const;
[globalisel][tablegen] Simplify named operand/operator lookups and fix a wrong-code bug this revealed. Summary: Operand variable lookups are now performed by the RuleMatcher rather than searching the whole matcher hierarchy for a match. This revealed a wrong-code bug that currently affects ARM and X86 where patterns that use a variable more than once in the match pattern will be imported but won't check that the operands are identical. This can cause the tablegen-erated matcher to accept matches that should be rejected. Depends on D36569 Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Subscribers: aemerson, igorb, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D36618 llvm-svn: 315780
2017-10-14 08:31:58 +08:00
const OperandMatcher &getOperandMatcher(StringRef Name) const;
[globalisel][tablegen] Add support for importing 'imm' operands. Summary: This patch enables the import of rules containing 'imm' operands that do not constrain the acceptable values using predicates. Support for ImmLeaf will arrive in a later patch. Depends on D35681 Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Reviewed By: rovka Subscribers: kristof.beyls, javed.absar, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D35833 llvm-svn: 310343
2017-08-08 18:44:31 +08:00
Reapply r332907 "[GlobalISel] Improving InstructionSelect's performance by reducing MatchTable..." Apparently the compile time problem was caused by the fact that not all compilers / STL implementations can automatically convert std::unique_ptr<Derived> to std::unique_ptr<Base>. Fixed (hopefully) by making sure it's std::unique_ptr<Derived>&& (rvalue ref) to std::unique_ptr<Base> conversion instead. llvm-svn: 332917
2018-05-22 07:28:51 +08:00
void optimize() override;
[TableGen][GlobalISel] Optimize MatchTable for faster instruction selection *** Context *** Prior to this patchw, the table generated for matching instruction was straight forward but highly inefficient. Basically, each pattern generates its own set of self contained checks and actions. E.g., TableGen generated: // First pattern CheckNumOperand 3 CheckOpcode G_ADD ... Build ADDrr // Second pattern CheckNumOperand 3 CheckOpcode G_ADD ... Build ADDri // Third pattern CheckNumOperand 3 CheckOpcode G_SUB ... Build SUBrr *** Problem *** Because of that generation, a *lot* of check were redundant between each pattern and were checked every single time until we reach the pattern that matches. E.g., Taking the previous table, let say we are matching a G_SUB, that means we were going to check all the rules for G_ADD before looking at the G_SUB rule. In particular we are going to do: check 3 operands; PASS check G_ADD; FAIL ; Next rule check 3 operands; PASS (but we already knew that!) check G_ADD; FAIL (well it is still not true) ; Next rule check 3 operands; PASS (really!!) check G_SUB; PASS (at last :P) *** Proposed Solution *** This patch introduces a concept of group of rules (GroupMatcher) that share some predicates and only get checked once for the whole group. This patch only creates groups with one nesting level. Conceptually there is nothing preventing us for having deeper nest level. However, the current implementation is not smart enough to share the recording (aka capturing) of values. That limits its ability to do more sharing. For the given example the current patch will generate: // First group CheckOpcode G_ADD // First pattern CheckNumOperand 3 ... Build ADDrr // Second pattern CheckNumOperand 3 ... Build ADDri // Second group CheckOpcode G_SUB // Third pattern CheckNumOperand 3 ... Build SUBrr But if we allowed several nesting level, it could create a sub group for the checknumoperand 3. (We would need to call optimizeRules on the rules within a group.) *** Result *** With only one level of nesting, the instruction selection pass is up to 4x faster. For instance, one instruction now takes 500 checks, instead of 24k! With more nesting we could get in the tens I believe. Differential Revision: https://reviews.llvm.org/D39034 rdar://problem/34670699 llvm-svn: 321017
2017-12-19 03:47:41 +08:00
void emit(MatchTable &Table) override;
[tablegen][globalisel] Trivial changes to reduce size of following patch. NFC. Summary: * Move namespace { * Trivial: Typo * RuleMatcher: Separate class and definition * Trivial: const findNodeEquiv Reviewers: t.p.northover, qcolombet, aditya_nandakumar, ab, rovka Reviewed By: rovka Subscribers: dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D30531 llvm-svn: 297884
2017-03-16 04:18:38 +08:00
[globalisel][tablegen] Partially fix compile-time regressions by converting matcher to state-machine(s) Summary: Replace the matcher if-statements for each rule with a state-machine. This significantly reduces compile time, memory allocations, and cumulative memory allocation when compiling AArch64InstructionSelector.cpp.o after r303259 is recommitted. The following patches will expand on this further to fully fix the regressions. Reviewers: rovka, ab, t.p.northover, qcolombet, aditya_nandakumar Reviewed By: ab Subscribers: vitalybuka, aemerson, javed.absar, igorb, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D33758 llvm-svn: 307079
2017-07-04 22:35:06 +08:00
/// Compare the priority of this object and B.
///
/// Returns true if this object is more important than B.
bool isHigherPriorityThan(const RuleMatcher &B) const;
[tablegen][globalisel] Trivial changes to reduce size of following patch. NFC. Summary: * Move namespace { * Trivial: Typo * RuleMatcher: Separate class and definition * Trivial: const findNodeEquiv Reviewers: t.p.northover, qcolombet, aditya_nandakumar, ab, rovka Reviewed By: rovka Subscribers: dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D30531 llvm-svn: 297884
2017-03-16 04:18:38 +08:00
[globalisel][tablegen] Partially fix compile-time regressions by converting matcher to state-machine(s) Summary: Replace the matcher if-statements for each rule with a state-machine. This significantly reduces compile time, memory allocations, and cumulative memory allocation when compiling AArch64InstructionSelector.cpp.o after r303259 is recommitted. The following patches will expand on this further to fully fix the regressions. Reviewers: rovka, ab, t.p.northover, qcolombet, aditya_nandakumar Reviewed By: ab Subscribers: vitalybuka, aemerson, javed.absar, igorb, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D33758 llvm-svn: 307079
2017-07-04 22:35:06 +08:00
/// Report the maximum number of temporary operands needed by the rule
/// matcher.
unsigned countRendererFns() const;
[globalisel][tablegen] Revise API for ComplexPattern operands to improve flexibility. Summary: Some targets need to be able to do more complex rendering than just adding an operand or two to an instruction. For example, it may need to insert an instruction to extract a subreg first, or it may need to perform an operation on the operand. In SelectionDAG, targets would create SDNode's to achieve the desired effect during the complex pattern predicate. This worked because SelectionDAG had a form of garbage collection that would take care of SDNode's that were created but not used due to a later predicate rejecting a match. This doesn't translate well to GlobalISel and the churn was wasteful. The API changes in this patch enable GlobalISel to accomplish the same thing without the waste. The API is now: InstructionSelector::OptionalComplexRendererFn selectArithImmed(MachineOperand &Root) const; where Root is the root of the match. The return value can be omitted to indicate that the predicate failed to match, or a function with the signature ComplexRendererFn can be returned. For example: return OptionalComplexRendererFn( [=](MachineInstrBuilder &MIB) { MIB.addImm(Immed).addImm(ShVal); }); adds two immediate operands to the rendered instruction. Immed and ShVal are captured from the predicate function. As an added bonus, this also reduces the amount of information we need to provide to GIComplexOperandMatcher. Depends on D31418 Reviewers: aditya_nandakumar, t.p.northover, qcolombet, rovka, ab, javed.absar Reviewed By: ab Subscribers: dberris, kristof.beyls, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D31761 llvm-svn: 301079
2017-04-22 23:11:04 +08:00
Reapply r332907 "[GlobalISel] Improving InstructionSelect's performance by reducing MatchTable..." Apparently the compile time problem was caused by the fact that not all compilers / STL implementations can automatically convert std::unique_ptr<Derived> to std::unique_ptr<Base>. Fixed (hopefully) by making sure it's std::unique_ptr<Derived>&& (rvalue ref) to std::unique_ptr<Base> conversion instead. llvm-svn: 332917
2018-05-22 07:28:51 +08:00
std::unique_ptr<PredicateMatcher> popFirstCondition() override;
const PredicateMatcher &getFirstCondition() const override;
bool hasFirstCondition() const override;
unsigned getNumOperands() const;
[GlobalISel][InstructionSelect] Removing redundant num operands and nested def operands checks, perf patch 2 This patch continues a series of patches that decrease time spent by GlobalISel in its InstructionSelect pass by roughly 60% for -O0 builds for large inputs as measured on sqlite3-amalgamation (http://sqlite.org/download.html) targeting AArch64. This commit specifically removes number of operands checks that are redundant if the instruction's opcode already guarantees that number of operands (or more), and also avoids any kind of checks on a def operand of a nested instruction as everything about it was already checked at its use. The expected performance implication is about 3% off InstructionSelect comparing to the baseline (before the series of patches) This patch also contains a bit of NFC changes required for further patches in the series. Every commit planned shares the same Phabricator Review. Reviewers: qcolombet, dsanders, bogner, aemerson, javed.absar Reviewed By: qcolombet Subscribers: rovka, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D44700 llvm-svn: 332945
2018-05-22 12:31:50 +08:00
StringRef getOpcode() const;
[TableGen][GlobalISel] Optimize MatchTable for faster instruction selection *** Context *** Prior to this patchw, the table generated for matching instruction was straight forward but highly inefficient. Basically, each pattern generates its own set of self contained checks and actions. E.g., TableGen generated: // First pattern CheckNumOperand 3 CheckOpcode G_ADD ... Build ADDrr // Second pattern CheckNumOperand 3 CheckOpcode G_ADD ... Build ADDri // Third pattern CheckNumOperand 3 CheckOpcode G_SUB ... Build SUBrr *** Problem *** Because of that generation, a *lot* of check were redundant between each pattern and were checked every single time until we reach the pattern that matches. E.g., Taking the previous table, let say we are matching a G_SUB, that means we were going to check all the rules for G_ADD before looking at the G_SUB rule. In particular we are going to do: check 3 operands; PASS check G_ADD; FAIL ; Next rule check 3 operands; PASS (but we already knew that!) check G_ADD; FAIL (well it is still not true) ; Next rule check 3 operands; PASS (really!!) check G_SUB; PASS (at last :P) *** Proposed Solution *** This patch introduces a concept of group of rules (GroupMatcher) that share some predicates and only get checked once for the whole group. This patch only creates groups with one nesting level. Conceptually there is nothing preventing us for having deeper nest level. However, the current implementation is not smart enough to share the recording (aka capturing) of values. That limits its ability to do more sharing. For the given example the current patch will generate: // First group CheckOpcode G_ADD // First pattern CheckNumOperand 3 ... Build ADDrr // Second pattern CheckNumOperand 3 ... Build ADDri // Second group CheckOpcode G_SUB // Third pattern CheckNumOperand 3 ... Build SUBrr But if we allowed several nesting level, it could create a sub group for the checknumoperand 3. (We would need to call optimizeRules on the rules within a group.) *** Result *** With only one level of nesting, the instruction selection pass is up to 4x faster. For instance, one instruction now takes 500 checks, instead of 24k! With more nesting we could get in the tens I believe. Differential Revision: https://reviews.llvm.org/D39034 rdar://problem/34670699 llvm-svn: 321017
2017-12-19 03:47:41 +08:00
[globalisel][tablegen] Partially fix compile-time regressions by converting matcher to state-machine(s) Summary: Replace the matcher if-statements for each rule with a state-machine. This significantly reduces compile time, memory allocations, and cumulative memory allocation when compiling AArch64InstructionSelector.cpp.o after r303259 is recommitted. The following patches will expand on this further to fully fix the regressions. Reviewers: rovka, ab, t.p.northover, qcolombet, aditya_nandakumar Reviewed By: ab Subscribers: vitalybuka, aemerson, javed.absar, igorb, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D33758 llvm-svn: 307079
2017-07-04 22:35:06 +08:00
// FIXME: Remove this as soon as possible
[TableGen][GlobalISel] Optimize MatchTable for faster instruction selection *** Context *** Prior to this patchw, the table generated for matching instruction was straight forward but highly inefficient. Basically, each pattern generates its own set of self contained checks and actions. E.g., TableGen generated: // First pattern CheckNumOperand 3 CheckOpcode G_ADD ... Build ADDrr // Second pattern CheckNumOperand 3 CheckOpcode G_ADD ... Build ADDri // Third pattern CheckNumOperand 3 CheckOpcode G_SUB ... Build SUBrr *** Problem *** Because of that generation, a *lot* of check were redundant between each pattern and were checked every single time until we reach the pattern that matches. E.g., Taking the previous table, let say we are matching a G_SUB, that means we were going to check all the rules for G_ADD before looking at the G_SUB rule. In particular we are going to do: check 3 operands; PASS check G_ADD; FAIL ; Next rule check 3 operands; PASS (but we already knew that!) check G_ADD; FAIL (well it is still not true) ; Next rule check 3 operands; PASS (really!!) check G_SUB; PASS (at last :P) *** Proposed Solution *** This patch introduces a concept of group of rules (GroupMatcher) that share some predicates and only get checked once for the whole group. This patch only creates groups with one nesting level. Conceptually there is nothing preventing us for having deeper nest level. However, the current implementation is not smart enough to share the recording (aka capturing) of values. That limits its ability to do more sharing. For the given example the current patch will generate: // First group CheckOpcode G_ADD // First pattern CheckNumOperand 3 ... Build ADDrr // Second pattern CheckNumOperand 3 ... Build ADDri // Second group CheckOpcode G_SUB // Third pattern CheckNumOperand 3 ... Build SUBrr But if we allowed several nesting level, it could create a sub group for the checknumoperand 3. (We would need to call optimizeRules on the rules within a group.) *** Result *** With only one level of nesting, the instruction selection pass is up to 4x faster. For instance, one instruction now takes 500 checks, instead of 24k! With more nesting we could get in the tens I believe. Differential Revision: https://reviews.llvm.org/D39034 rdar://problem/34670699 llvm-svn: 321017
2017-12-19 03:47:41 +08:00
InstructionMatcher &insnmatchers_front() const { return *Matchers.front(); }
[globalisel][tablegen] Stop hard-coding the emitted instruction ID to 0. NFC The next commit will add support for multi-instruction emission so we need to start allocating instruction ID's instead of hard-coding them to 0. llvm-svn: 317057
2017-11-01 08:29:47 +08:00
unsigned allocateOutputInsnID() { return NextOutputInsnID++; }
[globalisel][tablegen] Add support for multi-insn emission The importer will now accept nested instructions in the result pattern such as (ADDWrr $a, (SUBWrr $b, $c)). This is only valid when the nested instruction def's a single vreg and the parent instruction consumes a single vreg where a nested instruction is specified. The importer will automatically create a vreg to connect the two using the type information from the pattern. This vreg will be constrained to the register classes given in the instruction definitions*. * REG_SEQUENCE is explicitly rejected because of this. The definition doesn't constrain to a register class and it therefore needs special handling. llvm-svn: 317117
2017-11-02 03:57:57 +08:00
unsigned allocateTempRegID() { return NextTempRegID++; }
[TableGen][GlobalISel] Optimize MatchTable for faster instruction selection *** Context *** Prior to this patchw, the table generated for matching instruction was straight forward but highly inefficient. Basically, each pattern generates its own set of self contained checks and actions. E.g., TableGen generated: // First pattern CheckNumOperand 3 CheckOpcode G_ADD ... Build ADDrr // Second pattern CheckNumOperand 3 CheckOpcode G_ADD ... Build ADDri // Third pattern CheckNumOperand 3 CheckOpcode G_SUB ... Build SUBrr *** Problem *** Because of that generation, a *lot* of check were redundant between each pattern and were checked every single time until we reach the pattern that matches. E.g., Taking the previous table, let say we are matching a G_SUB, that means we were going to check all the rules for G_ADD before looking at the G_SUB rule. In particular we are going to do: check 3 operands; PASS check G_ADD; FAIL ; Next rule check 3 operands; PASS (but we already knew that!) check G_ADD; FAIL (well it is still not true) ; Next rule check 3 operands; PASS (really!!) check G_SUB; PASS (at last :P) *** Proposed Solution *** This patch introduces a concept of group of rules (GroupMatcher) that share some predicates and only get checked once for the whole group. This patch only creates groups with one nesting level. Conceptually there is nothing preventing us for having deeper nest level. However, the current implementation is not smart enough to share the recording (aka capturing) of values. That limits its ability to do more sharing. For the given example the current patch will generate: // First group CheckOpcode G_ADD // First pattern CheckNumOperand 3 ... Build ADDrr // Second pattern CheckNumOperand 3 ... Build ADDri // Second group CheckOpcode G_SUB // Third pattern CheckNumOperand 3 ... Build SUBrr But if we allowed several nesting level, it could create a sub group for the checknumoperand 3. (We would need to call optimizeRules on the rules within a group.) *** Result *** With only one level of nesting, the instruction selection pass is up to 4x faster. For instance, one instruction now takes 500 checks, instead of 24k! With more nesting we could get in the tens I believe. Differential Revision: https://reviews.llvm.org/D39034 rdar://problem/34670699 llvm-svn: 321017
2017-12-19 03:47:41 +08:00
Reapply r332907 "[GlobalISel] Improving InstructionSelect's performance by reducing MatchTable..." Apparently the compile time problem was caused by the fact that not all compilers / STL implementations can automatically convert std::unique_ptr<Derived> to std::unique_ptr<Base>. Fixed (hopefully) by making sure it's std::unique_ptr<Derived>&& (rvalue ref) to std::unique_ptr<Base> conversion instead. llvm-svn: 332917
2018-05-22 07:28:51 +08:00
iterator_range<MatchersTy::iterator> insnmatchers() {
return make_range(Matchers.begin(), Matchers.end());
}
[TableGen][GlobalISel] Optimize MatchTable for faster instruction selection *** Context *** Prior to this patchw, the table generated for matching instruction was straight forward but highly inefficient. Basically, each pattern generates its own set of self contained checks and actions. E.g., TableGen generated: // First pattern CheckNumOperand 3 CheckOpcode G_ADD ... Build ADDrr // Second pattern CheckNumOperand 3 CheckOpcode G_ADD ... Build ADDri // Third pattern CheckNumOperand 3 CheckOpcode G_SUB ... Build SUBrr *** Problem *** Because of that generation, a *lot* of check were redundant between each pattern and were checked every single time until we reach the pattern that matches. E.g., Taking the previous table, let say we are matching a G_SUB, that means we were going to check all the rules for G_ADD before looking at the G_SUB rule. In particular we are going to do: check 3 operands; PASS check G_ADD; FAIL ; Next rule check 3 operands; PASS (but we already knew that!) check G_ADD; FAIL (well it is still not true) ; Next rule check 3 operands; PASS (really!!) check G_SUB; PASS (at last :P) *** Proposed Solution *** This patch introduces a concept of group of rules (GroupMatcher) that share some predicates and only get checked once for the whole group. This patch only creates groups with one nesting level. Conceptually there is nothing preventing us for having deeper nest level. However, the current implementation is not smart enough to share the recording (aka capturing) of values. That limits its ability to do more sharing. For the given example the current patch will generate: // First group CheckOpcode G_ADD // First pattern CheckNumOperand 3 ... Build ADDrr // Second pattern CheckNumOperand 3 ... Build ADDri // Second group CheckOpcode G_SUB // Third pattern CheckNumOperand 3 ... Build SUBrr But if we allowed several nesting level, it could create a sub group for the checknumoperand 3. (We would need to call optimizeRules on the rules within a group.) *** Result *** With only one level of nesting, the instruction selection pass is up to 4x faster. For instance, one instruction now takes 500 checks, instead of 24k! With more nesting we could get in the tens I believe. Differential Revision: https://reviews.llvm.org/D39034 rdar://problem/34670699 llvm-svn: 321017
2017-12-19 03:47:41 +08:00
bool insnmatchers_empty() const { return Matchers.empty(); }
void insnmatchers_pop_front() { Matchers.erase(Matchers.begin()); }
[tablegen][globalisel] Trivial changes to reduce size of following patch. NFC. Summary: * Move namespace { * Trivial: Typo * RuleMatcher: Separate class and definition * Trivial: const findNodeEquiv Reviewers: t.p.northover, qcolombet, aditya_nandakumar, ab, rovka Reviewed By: rovka Subscribers: dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D30531 llvm-svn: 297884
2017-03-16 04:18:38 +08:00
};
[globalisel][tablegen] Generate rule coverage and use it to identify untested rules Summary: This patch adds a LLVM_ENABLE_GISEL_COV which, like LLVM_ENABLE_DAGISEL_COV, causes TableGen to instrument the generated table to collect rule coverage information. However, LLVM_ENABLE_GISEL_COV goes a bit further than LLVM_ENABLE_DAGISEL_COV. The information is written to files (${CMAKE_BINARY_DIR}/gisel-coverage-* by default). These files can then be concatenated into ${LLVM_GISEL_COV_PREFIX}-all after which TableGen will read this information and use it to emit warnings about untested rules. This technique could also be used by SelectionDAG and can be further extended to detect hot rules and give them priority over colder rules. Usage: * Enable LLVM_ENABLE_GISEL_COV in CMake * Build the compiler and run some tests * cat gisel-coverage-[0-9]* > gisel-coverage-all * Delete lib/Target/*/*GenGlobalISel.inc* * Build the compiler Known issues: * ${LLVM_GISEL_COV_PREFIX}-all must be generated as a manual step due to a lack of a portable 'cat' command. It should be the concatenation of all ${LLVM_GISEL_COV_PREFIX}-[0-9]* files. * There's no mechanism to discard coverage information when the ruleset changes Depends on D39742 Reviewers: ab, qcolombet, t.p.northover, aditya_nandakumar, rovka Reviewed By: rovka Subscribers: vsk, arsenm, nhaehnle, mgorny, kristof.beyls, javed.absar, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D39747 llvm-svn: 318356
2017-11-16 08:46:35 +08:00
uint64_t RuleMatcher::NextRuleID = 0;
Re-commit: [globalisel][tablegen] Keep track of the insertion point while adding BuildMIAction's. NFC Multi-instruction emission needs to ensure the the instructions are generated a depth-first fashion. For example: (ADDWrr (SUBWrr a, b), c) needs to emit the SUBWrr before the ADDWrr. However, our walk over TreePatternNode's is highly context sensitive which makes it difficult to append BuildMIActions in the order we want. To fix this, we now keep track of the insertion point as we add actions. This will allow multi-insn emission to insert BuildMI's in the correct place. The previous commit failed on the Ubuntu bots using GCC 4.8. These bots lack the const_iterator forms of insert() and emplace() that were added in C++11. As a result I've switched the const_iterators to iterators. llvm-svn: 317049
2017-11-01 07:03:18 +08:00
using action_iterator = RuleMatcher::action_iterator;
[globalisel] Re-factor ISel matchers into a hierarchy. NFC Summary: This should make it possible to easily add everything needed to import all the existing SelectionDAG rules. It should also serve the likely kinds of GlobalISel rules (some of which are not currently representable in SelectionDAG) once we've nailed down the tablegen definition for that. The hierarchy is as follows: MatcherRule - A matching rule. Currently used to emit C++ ISel code but will | also be used to emit test cases and tablegen definitions in the | near future. |- Instruction(s) - Represents the instruction to be matched. |- Instruction Predicate(s) - Test the opcode, arithmetic flags, etc. of an | instruction. \- Operand(s) - Represents a particular operand of the instruction. In the | future, there may be subclasses to test the same predicates | on multiple operands (including for variadic instructions). \ Operand Predicate(s) - Test the type, register bank, etc. of an operand. This is where the ComplexPattern equivalent will be represented. It's also nested-instruction matching will live as a predicate that follows the DefUse chain to the Def and tests a MatcherRule from that position. Support for multiple instruction matchers in a rule has been retained from the existing code but has been adjusted to assert when it is used. Previously it would silently drop all but the first instruction matcher. The tablegen-erated file is not functionally changed but has more parentheses and no longer attempts to format the if-statements since keeping track of the indentation is tricky in the presence of the matcher hierarchy. It would be nice to have CMakes tablegen() run the output through clang-format (when available) so we don't have to complicate TableGen with pretty-printing. It's also worth mentioning that this hierarchy will also be able to emit TableGen definitions and test cases in the near future. This is the reason for favouring explicit emit*() calls rather than the << operator. Reviewers: aditya_nandakumar, rovka, t.p.northover, qcolombet, ab Reviewed By: ab Subscribers: igorb, dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D28942 llvm-svn: 293172
2017-01-26 19:10:14 +08:00
template <class PredicateTy> class PredicateListMatcher {
private:
Re-commit: [globalisel][tablegen] Add support for ImmLeaf without SDNodeXForm Summary: This patch adds support for predicates on imm nodes but only for ImmLeaf and not for PatLeaf or PatFrag and only where the value does not need to be transformed before being rendered into the instruction. The limitation on PatLeaf/PatFrag/SDNodeXForm is due to differences in the necessary target-supplied C++ for GlobalISel. Depends on D36085 The previous commit was reverted for breaking the build but this appears to have been the recurring problem on the Windows bots with tablegen not being re-run when llvm-tblgen is changed but the .td's aren't. If it re-occurs then forcing a build with clean=True should fix it but this string should do this in advance: Requires a clean build. Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Reviewed By: rovka Subscribers: kristof.beyls, javed.absar, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D36086 llvm-svn: 311645
2017-08-24 17:11:20 +08:00
/// Template instantiations should specialize this to return a string to use
/// for the comment emitted when there are no predicates.
std::string getNoPredicateComment() const;
Reapply r332907 "[GlobalISel] Improving InstructionSelect's performance by reducing MatchTable..." Apparently the compile time problem was caused by the fact that not all compilers / STL implementations can automatically convert std::unique_ptr<Derived> to std::unique_ptr<Base>. Fixed (hopefully) by making sure it's std::unique_ptr<Derived>&& (rvalue ref) to std::unique_ptr<Base> conversion instead. llvm-svn: 332917
2018-05-22 07:28:51 +08:00
protected:
using PredicatesTy = std::deque<std::unique_ptr<PredicateTy>>;
PredicatesTy Predicates;
/// Track if the list of predicates was manipulated by one of the optimization
/// methods.
bool Optimized = false;
[globalisel] Re-factor ISel matchers into a hierarchy. NFC Summary: This should make it possible to easily add everything needed to import all the existing SelectionDAG rules. It should also serve the likely kinds of GlobalISel rules (some of which are not currently representable in SelectionDAG) once we've nailed down the tablegen definition for that. The hierarchy is as follows: MatcherRule - A matching rule. Currently used to emit C++ ISel code but will | also be used to emit test cases and tablegen definitions in the | near future. |- Instruction(s) - Represents the instruction to be matched. |- Instruction Predicate(s) - Test the opcode, arithmetic flags, etc. of an | instruction. \- Operand(s) - Represents a particular operand of the instruction. In the | future, there may be subclasses to test the same predicates | on multiple operands (including for variadic instructions). \ Operand Predicate(s) - Test the type, register bank, etc. of an operand. This is where the ComplexPattern equivalent will be represented. It's also nested-instruction matching will live as a predicate that follows the DefUse chain to the Def and tests a MatcherRule from that position. Support for multiple instruction matchers in a rule has been retained from the existing code but has been adjusted to assert when it is used. Previously it would silently drop all but the first instruction matcher. The tablegen-erated file is not functionally changed but has more parentheses and no longer attempts to format the if-statements since keeping track of the indentation is tricky in the presence of the matcher hierarchy. It would be nice to have CMakes tablegen() run the output through clang-format (when available) so we don't have to complicate TableGen with pretty-printing. It's also worth mentioning that this hierarchy will also be able to emit TableGen definitions and test cases in the near future. This is the reason for favouring explicit emit*() calls rather than the << operator. Reviewers: aditya_nandakumar, rovka, t.p.northover, qcolombet, ab Reviewed By: ab Subscribers: igorb, dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D28942 llvm-svn: 293172
2017-01-26 19:10:14 +08:00
public:
Reapply r332907 "[GlobalISel] Improving InstructionSelect's performance by reducing MatchTable..." Apparently the compile time problem was caused by the fact that not all compilers / STL implementations can automatically convert std::unique_ptr<Derived> to std::unique_ptr<Base>. Fixed (hopefully) by making sure it's std::unique_ptr<Derived>&& (rvalue ref) to std::unique_ptr<Base> conversion instead. llvm-svn: 332917
2018-05-22 07:28:51 +08:00
/// Construct a new predicate and add it to the matcher.
[globalisel] Re-factor ISel matchers into a hierarchy. NFC Summary: This should make it possible to easily add everything needed to import all the existing SelectionDAG rules. It should also serve the likely kinds of GlobalISel rules (some of which are not currently representable in SelectionDAG) once we've nailed down the tablegen definition for that. The hierarchy is as follows: MatcherRule - A matching rule. Currently used to emit C++ ISel code but will | also be used to emit test cases and tablegen definitions in the | near future. |- Instruction(s) - Represents the instruction to be matched. |- Instruction Predicate(s) - Test the opcode, arithmetic flags, etc. of an | instruction. \- Operand(s) - Represents a particular operand of the instruction. In the | future, there may be subclasses to test the same predicates | on multiple operands (including for variadic instructions). \ Operand Predicate(s) - Test the type, register bank, etc. of an operand. This is where the ComplexPattern equivalent will be represented. It's also nested-instruction matching will live as a predicate that follows the DefUse chain to the Def and tests a MatcherRule from that position. Support for multiple instruction matchers in a rule has been retained from the existing code but has been adjusted to assert when it is used. Previously it would silently drop all but the first instruction matcher. The tablegen-erated file is not functionally changed but has more parentheses and no longer attempts to format the if-statements since keeping track of the indentation is tricky in the presence of the matcher hierarchy. It would be nice to have CMakes tablegen() run the output through clang-format (when available) so we don't have to complicate TableGen with pretty-printing. It's also worth mentioning that this hierarchy will also be able to emit TableGen definitions and test cases in the near future. This is the reason for favouring explicit emit*() calls rather than the << operator. Reviewers: aditya_nandakumar, rovka, t.p.northover, qcolombet, ab Reviewed By: ab Subscribers: igorb, dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D28942 llvm-svn: 293172
2017-01-26 19:10:14 +08:00
template <class Kind, class... Args>
Reapply r332907 "[GlobalISel] Improving InstructionSelect's performance by reducing MatchTable..." Apparently the compile time problem was caused by the fact that not all compilers / STL implementations can automatically convert std::unique_ptr<Derived> to std::unique_ptr<Base>. Fixed (hopefully) by making sure it's std::unique_ptr<Derived>&& (rvalue ref) to std::unique_ptr<Base> conversion instead. llvm-svn: 332917
2018-05-22 07:28:51 +08:00
Optional<Kind *> addPredicate(Args &&... args);
[globalisel] Re-factor ISel matchers into a hierarchy. NFC Summary: This should make it possible to easily add everything needed to import all the existing SelectionDAG rules. It should also serve the likely kinds of GlobalISel rules (some of which are not currently representable in SelectionDAG) once we've nailed down the tablegen definition for that. The hierarchy is as follows: MatcherRule - A matching rule. Currently used to emit C++ ISel code but will | also be used to emit test cases and tablegen definitions in the | near future. |- Instruction(s) - Represents the instruction to be matched. |- Instruction Predicate(s) - Test the opcode, arithmetic flags, etc. of an | instruction. \- Operand(s) - Represents a particular operand of the instruction. In the | future, there may be subclasses to test the same predicates | on multiple operands (including for variadic instructions). \ Operand Predicate(s) - Test the type, register bank, etc. of an operand. This is where the ComplexPattern equivalent will be represented. It's also nested-instruction matching will live as a predicate that follows the DefUse chain to the Def and tests a MatcherRule from that position. Support for multiple instruction matchers in a rule has been retained from the existing code but has been adjusted to assert when it is used. Previously it would silently drop all but the first instruction matcher. The tablegen-erated file is not functionally changed but has more parentheses and no longer attempts to format the if-statements since keeping track of the indentation is tricky in the presence of the matcher hierarchy. It would be nice to have CMakes tablegen() run the output through clang-format (when available) so we don't have to complicate TableGen with pretty-printing. It's also worth mentioning that this hierarchy will also be able to emit TableGen definitions and test cases in the near future. This is the reason for favouring explicit emit*() calls rather than the << operator. Reviewers: aditya_nandakumar, rovka, t.p.northover, qcolombet, ab Reviewed By: ab Subscribers: igorb, dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D28942 llvm-svn: 293172
2017-01-26 19:10:14 +08:00
Reapply r332907 "[GlobalISel] Improving InstructionSelect's performance by reducing MatchTable..." Apparently the compile time problem was caused by the fact that not all compilers / STL implementations can automatically convert std::unique_ptr<Derived> to std::unique_ptr<Base>. Fixed (hopefully) by making sure it's std::unique_ptr<Derived>&& (rvalue ref) to std::unique_ptr<Base> conversion instead. llvm-svn: 332917
2018-05-22 07:28:51 +08:00
typename PredicatesTy::iterator predicates_begin() {
[globalisel][tablegen] Multiple 80-col corrections. llvm-svn: 306544
2017-06-28 21:50:04 +08:00
return Predicates.begin();
}
Reapply r332907 "[GlobalISel] Improving InstructionSelect's performance by reducing MatchTable..." Apparently the compile time problem was caused by the fact that not all compilers / STL implementations can automatically convert std::unique_ptr<Derived> to std::unique_ptr<Base>. Fixed (hopefully) by making sure it's std::unique_ptr<Derived>&& (rvalue ref) to std::unique_ptr<Base> conversion instead. llvm-svn: 332917
2018-05-22 07:28:51 +08:00
typename PredicatesTy::iterator predicates_end() {
[globalisel][tablegen] Multiple 80-col corrections. llvm-svn: 306544
2017-06-28 21:50:04 +08:00
return Predicates.end();
}
Reapply r332907 "[GlobalISel] Improving InstructionSelect's performance by reducing MatchTable..." Apparently the compile time problem was caused by the fact that not all compilers / STL implementations can automatically convert std::unique_ptr<Derived> to std::unique_ptr<Base>. Fixed (hopefully) by making sure it's std::unique_ptr<Derived>&& (rvalue ref) to std::unique_ptr<Base> conversion instead. llvm-svn: 332917
2018-05-22 07:28:51 +08:00
iterator_range<typename PredicatesTy::iterator> predicates() {
[globalisel] Re-factor ISel matchers into a hierarchy. NFC Summary: This should make it possible to easily add everything needed to import all the existing SelectionDAG rules. It should also serve the likely kinds of GlobalISel rules (some of which are not currently representable in SelectionDAG) once we've nailed down the tablegen definition for that. The hierarchy is as follows: MatcherRule - A matching rule. Currently used to emit C++ ISel code but will | also be used to emit test cases and tablegen definitions in the | near future. |- Instruction(s) - Represents the instruction to be matched. |- Instruction Predicate(s) - Test the opcode, arithmetic flags, etc. of an | instruction. \- Operand(s) - Represents a particular operand of the instruction. In the | future, there may be subclasses to test the same predicates | on multiple operands (including for variadic instructions). \ Operand Predicate(s) - Test the type, register bank, etc. of an operand. This is where the ComplexPattern equivalent will be represented. It's also nested-instruction matching will live as a predicate that follows the DefUse chain to the Def and tests a MatcherRule from that position. Support for multiple instruction matchers in a rule has been retained from the existing code but has been adjusted to assert when it is used. Previously it would silently drop all but the first instruction matcher. The tablegen-erated file is not functionally changed but has more parentheses and no longer attempts to format the if-statements since keeping track of the indentation is tricky in the presence of the matcher hierarchy. It would be nice to have CMakes tablegen() run the output through clang-format (when available) so we don't have to complicate TableGen with pretty-printing. It's also worth mentioning that this hierarchy will also be able to emit TableGen definitions and test cases in the near future. This is the reason for favouring explicit emit*() calls rather than the << operator. Reviewers: aditya_nandakumar, rovka, t.p.northover, qcolombet, ab Reviewed By: ab Subscribers: igorb, dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D28942 llvm-svn: 293172
2017-01-26 19:10:14 +08:00
return make_range(predicates_begin(), predicates_end());
[GlobalISel] Add basic Selector-emitter tblgen backend. This adds a basic tablegen backend that analyzes the SelectionDAG patterns to find simple ones that are eligible for GlobalISel-emission. That's similar to FastISel, with one notable difference: we're not fed ISD opcodes, so we need to map the SDNode operators to generic opcodes. That's done using GINodeEquiv in TargetGlobalISel.td. Otherwise, this is mostly boilerplate, and lots of filtering of any kind of "complicated" pattern. On AArch64, this is sufficient to match G_ADD up to s64 (to ADDWrr/ADDXrr) and G_BR (to B). Differential Revision: https://reviews.llvm.org/D26878 llvm-svn: 290284
2016-12-22 07:26:20 +08:00
}
Reapply r332907 "[GlobalISel] Improving InstructionSelect's performance by reducing MatchTable..." Apparently the compile time problem was caused by the fact that not all compilers / STL implementations can automatically convert std::unique_ptr<Derived> to std::unique_ptr<Base>. Fixed (hopefully) by making sure it's std::unique_ptr<Derived>&& (rvalue ref) to std::unique_ptr<Base> conversion instead. llvm-svn: 332917
2018-05-22 07:28:51 +08:00
typename PredicatesTy::size_type predicates_size() const {
[globalisel][tablegen] Multiple 80-col corrections. llvm-svn: 306544
2017-06-28 21:50:04 +08:00
return Predicates.size();
}
[TableGen][GlobalISel] Optimize MatchTable for faster instruction selection *** Context *** Prior to this patchw, the table generated for matching instruction was straight forward but highly inefficient. Basically, each pattern generates its own set of self contained checks and actions. E.g., TableGen generated: // First pattern CheckNumOperand 3 CheckOpcode G_ADD ... Build ADDrr // Second pattern CheckNumOperand 3 CheckOpcode G_ADD ... Build ADDri // Third pattern CheckNumOperand 3 CheckOpcode G_SUB ... Build SUBrr *** Problem *** Because of that generation, a *lot* of check were redundant between each pattern and were checked every single time until we reach the pattern that matches. E.g., Taking the previous table, let say we are matching a G_SUB, that means we were going to check all the rules for G_ADD before looking at the G_SUB rule. In particular we are going to do: check 3 operands; PASS check G_ADD; FAIL ; Next rule check 3 operands; PASS (but we already knew that!) check G_ADD; FAIL (well it is still not true) ; Next rule check 3 operands; PASS (really!!) check G_SUB; PASS (at last :P) *** Proposed Solution *** This patch introduces a concept of group of rules (GroupMatcher) that share some predicates and only get checked once for the whole group. This patch only creates groups with one nesting level. Conceptually there is nothing preventing us for having deeper nest level. However, the current implementation is not smart enough to share the recording (aka capturing) of values. That limits its ability to do more sharing. For the given example the current patch will generate: // First group CheckOpcode G_ADD // First pattern CheckNumOperand 3 ... Build ADDrr // Second pattern CheckNumOperand 3 ... Build ADDri // Second group CheckOpcode G_SUB // Third pattern CheckNumOperand 3 ... Build SUBrr But if we allowed several nesting level, it could create a sub group for the checknumoperand 3. (We would need to call optimizeRules on the rules within a group.) *** Result *** With only one level of nesting, the instruction selection pass is up to 4x faster. For instance, one instruction now takes 500 checks, instead of 24k! With more nesting we could get in the tens I believe. Differential Revision: https://reviews.llvm.org/D39034 rdar://problem/34670699 llvm-svn: 321017
2017-12-19 03:47:41 +08:00
bool predicates_empty() const { return Predicates.empty(); }
std::unique_ptr<PredicateTy> predicates_pop_front() {
std::unique_ptr<PredicateTy> Front = std::move(Predicates.front());
Reapply r332907 "[GlobalISel] Improving InstructionSelect's performance by reducing MatchTable..." Apparently the compile time problem was caused by the fact that not all compilers / STL implementations can automatically convert std::unique_ptr<Derived> to std::unique_ptr<Base>. Fixed (hopefully) by making sure it's std::unique_ptr<Derived>&& (rvalue ref) to std::unique_ptr<Base> conversion instead. llvm-svn: 332917
2018-05-22 07:28:51 +08:00
Predicates.pop_front();
Optimized = true;
[TableGen][GlobalISel] Optimize MatchTable for faster instruction selection *** Context *** Prior to this patchw, the table generated for matching instruction was straight forward but highly inefficient. Basically, each pattern generates its own set of self contained checks and actions. E.g., TableGen generated: // First pattern CheckNumOperand 3 CheckOpcode G_ADD ... Build ADDrr // Second pattern CheckNumOperand 3 CheckOpcode G_ADD ... Build ADDri // Third pattern CheckNumOperand 3 CheckOpcode G_SUB ... Build SUBrr *** Problem *** Because of that generation, a *lot* of check were redundant between each pattern and were checked every single time until we reach the pattern that matches. E.g., Taking the previous table, let say we are matching a G_SUB, that means we were going to check all the rules for G_ADD before looking at the G_SUB rule. In particular we are going to do: check 3 operands; PASS check G_ADD; FAIL ; Next rule check 3 operands; PASS (but we already knew that!) check G_ADD; FAIL (well it is still not true) ; Next rule check 3 operands; PASS (really!!) check G_SUB; PASS (at last :P) *** Proposed Solution *** This patch introduces a concept of group of rules (GroupMatcher) that share some predicates and only get checked once for the whole group. This patch only creates groups with one nesting level. Conceptually there is nothing preventing us for having deeper nest level. However, the current implementation is not smart enough to share the recording (aka capturing) of values. That limits its ability to do more sharing. For the given example the current patch will generate: // First group CheckOpcode G_ADD // First pattern CheckNumOperand 3 ... Build ADDrr // Second pattern CheckNumOperand 3 ... Build ADDri // Second group CheckOpcode G_SUB // Third pattern CheckNumOperand 3 ... Build SUBrr But if we allowed several nesting level, it could create a sub group for the checknumoperand 3. (We would need to call optimizeRules on the rules within a group.) *** Result *** With only one level of nesting, the instruction selection pass is up to 4x faster. For instance, one instruction now takes 500 checks, instead of 24k! With more nesting we could get in the tens I believe. Differential Revision: https://reviews.llvm.org/D39034 rdar://problem/34670699 llvm-svn: 321017
2017-12-19 03:47:41 +08:00
return Front;
}
[globalisel] Re-factor ISel matchers into a hierarchy. NFC Summary: This should make it possible to easily add everything needed to import all the existing SelectionDAG rules. It should also serve the likely kinds of GlobalISel rules (some of which are not currently representable in SelectionDAG) once we've nailed down the tablegen definition for that. The hierarchy is as follows: MatcherRule - A matching rule. Currently used to emit C++ ISel code but will | also be used to emit test cases and tablegen definitions in the | near future. |- Instruction(s) - Represents the instruction to be matched. |- Instruction Predicate(s) - Test the opcode, arithmetic flags, etc. of an | instruction. \- Operand(s) - Represents a particular operand of the instruction. In the | future, there may be subclasses to test the same predicates | on multiple operands (including for variadic instructions). \ Operand Predicate(s) - Test the type, register bank, etc. of an operand. This is where the ComplexPattern equivalent will be represented. It's also nested-instruction matching will live as a predicate that follows the DefUse chain to the Def and tests a MatcherRule from that position. Support for multiple instruction matchers in a rule has been retained from the existing code but has been adjusted to assert when it is used. Previously it would silently drop all but the first instruction matcher. The tablegen-erated file is not functionally changed but has more parentheses and no longer attempts to format the if-statements since keeping track of the indentation is tricky in the presence of the matcher hierarchy. It would be nice to have CMakes tablegen() run the output through clang-format (when available) so we don't have to complicate TableGen with pretty-printing. It's also worth mentioning that this hierarchy will also be able to emit TableGen definitions and test cases in the near future. This is the reason for favouring explicit emit*() calls rather than the << operator. Reviewers: aditya_nandakumar, rovka, t.p.northover, qcolombet, ab Reviewed By: ab Subscribers: igorb, dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D28942 llvm-svn: 293172
2017-01-26 19:10:14 +08:00
Reapply r332907 "[GlobalISel] Improving InstructionSelect's performance by reducing MatchTable..." Apparently the compile time problem was caused by the fact that not all compilers / STL implementations can automatically convert std::unique_ptr<Derived> to std::unique_ptr<Base>. Fixed (hopefully) by making sure it's std::unique_ptr<Derived>&& (rvalue ref) to std::unique_ptr<Base> conversion instead. llvm-svn: 332917
2018-05-22 07:28:51 +08:00
void prependPredicate(std::unique_ptr<PredicateTy> &&Predicate) {
Predicates.push_front(std::move(Predicate));
}
void eraseNullPredicates() {
const auto NewEnd =
std::stable_partition(Predicates.begin(), Predicates.end(),
std::logical_not<std::unique_ptr<PredicateTy>>());
if (NewEnd != Predicates.begin()) {
Predicates.erase(Predicates.begin(), NewEnd);
Optimized = true;
}
}
[globalisel][tablegen] Rename and re-comment to match the new MatchTables. NFC. The conversion to MatchTable left the function names and comments referring to C++ statements and expressions. Updated the names and comments to account for the fact that they're no longer unconstrained statements/expressions. llvm-svn: 307246
2017-07-06 18:06:12 +08:00
/// Emit MatchTable opcodes that tests whether all the predicates are met.
[GlobalISel] Remove duplicate function using variadic templates. NFC. I think the initial version of r293172 was trying: std::forward<Args...>(args)... which doesn't compile. This seems like the correct way: std::forward<Args>(args)... llvm-svn: 293214
2017-01-27 06:07:37 +08:00
template <class... Args>
Reapply r332907 "[GlobalISel] Improving InstructionSelect's performance by reducing MatchTable..." Apparently the compile time problem was caused by the fact that not all compilers / STL implementations can automatically convert std::unique_ptr<Derived> to std::unique_ptr<Base>. Fixed (hopefully) by making sure it's std::unique_ptr<Derived>&& (rvalue ref) to std::unique_ptr<Base> conversion instead. llvm-svn: 332917
2018-05-22 07:28:51 +08:00
void emitPredicateListOpcodes(MatchTable &Table, Args &&... args) {
if (Predicates.empty() && !Optimized) {
Re-commit: [globalisel][tablegen] Add support for ImmLeaf without SDNodeXForm Summary: This patch adds support for predicates on imm nodes but only for ImmLeaf and not for PatLeaf or PatFrag and only where the value does not need to be transformed before being rendered into the instruction. The limitation on PatLeaf/PatFrag/SDNodeXForm is due to differences in the necessary target-supplied C++ for GlobalISel. Depends on D36085 The previous commit was reverted for breaking the build but this appears to have been the recurring problem on the Windows bots with tablegen not being re-run when llvm-tblgen is changed but the .td's aren't. If it re-occurs then forcing a build with clean=True should fix it but this string should do this in advance: Requires a clean build. Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Reviewed By: rovka Subscribers: kristof.beyls, javed.absar, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D36086 llvm-svn: 311645
2017-08-24 17:11:20 +08:00
Table << MatchTable::Comment(getNoPredicateComment())
<< MatchTable::LineBreak;
[globalisel] Re-factor ISel matchers into a hierarchy. NFC Summary: This should make it possible to easily add everything needed to import all the existing SelectionDAG rules. It should also serve the likely kinds of GlobalISel rules (some of which are not currently representable in SelectionDAG) once we've nailed down the tablegen definition for that. The hierarchy is as follows: MatcherRule - A matching rule. Currently used to emit C++ ISel code but will | also be used to emit test cases and tablegen definitions in the | near future. |- Instruction(s) - Represents the instruction to be matched. |- Instruction Predicate(s) - Test the opcode, arithmetic flags, etc. of an | instruction. \- Operand(s) - Represents a particular operand of the instruction. In the | future, there may be subclasses to test the same predicates | on multiple operands (including for variadic instructions). \ Operand Predicate(s) - Test the type, register bank, etc. of an operand. This is where the ComplexPattern equivalent will be represented. It's also nested-instruction matching will live as a predicate that follows the DefUse chain to the Def and tests a MatcherRule from that position. Support for multiple instruction matchers in a rule has been retained from the existing code but has been adjusted to assert when it is used. Previously it would silently drop all but the first instruction matcher. The tablegen-erated file is not functionally changed but has more parentheses and no longer attempts to format the if-statements since keeping track of the indentation is tricky in the presence of the matcher hierarchy. It would be nice to have CMakes tablegen() run the output through clang-format (when available) so we don't have to complicate TableGen with pretty-printing. It's also worth mentioning that this hierarchy will also be able to emit TableGen definitions and test cases in the near future. This is the reason for favouring explicit emit*() calls rather than the << operator. Reviewers: aditya_nandakumar, rovka, t.p.northover, qcolombet, ab Reviewed By: ab Subscribers: igorb, dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D28942 llvm-svn: 293172
2017-01-26 19:10:14 +08:00
return;
}
Reapply r332907 "[GlobalISel] Improving InstructionSelect's performance by reducing MatchTable..." Apparently the compile time problem was caused by the fact that not all compilers / STL implementations can automatically convert std::unique_ptr<Derived> to std::unique_ptr<Base>. Fixed (hopefully) by making sure it's std::unique_ptr<Derived>&& (rvalue ref) to std::unique_ptr<Base> conversion instead. llvm-svn: 332917
2018-05-22 07:28:51 +08:00
for (const auto &Predicate : predicates())
[globalisel][tablegen] Add control-flow to the MatchTable. Summary: This will allow us to merge the various sub-tables into a single table. This is a compile-time saving at this point. However, this will also enable the optimization of a table so that similar instructions can be tested together, reducing the time spent on the matching the code. The bulk of this patch is a mechanical conversion to the new MatchTable object which is responsible for tracking label definitions and filling in the index of the jump targets. It is also responsible for nicely formatting the table. This was necessary to support the new GIM_Try opcode which takes the index to jump to if the match should fail. This value is unknown during table construction and is filled in during emission. To support nesting try-blocks (although we currently don't emit tables with nested try-blocks), GIM_Reject has been re-introduced to explicitly exit a try-block or fail the overall match if there are no active try-blocks. Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Reviewed By: rovka Subscribers: kristof.beyls, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D35117 llvm-svn: 308596
2017-07-20 17:25:44 +08:00
Predicate->emitPredicateOpcodes(Table, std::forward<Args>(args)...);
[globalisel] Re-factor ISel matchers into a hierarchy. NFC Summary: This should make it possible to easily add everything needed to import all the existing SelectionDAG rules. It should also serve the likely kinds of GlobalISel rules (some of which are not currently representable in SelectionDAG) once we've nailed down the tablegen definition for that. The hierarchy is as follows: MatcherRule - A matching rule. Currently used to emit C++ ISel code but will | also be used to emit test cases and tablegen definitions in the | near future. |- Instruction(s) - Represents the instruction to be matched. |- Instruction Predicate(s) - Test the opcode, arithmetic flags, etc. of an | instruction. \- Operand(s) - Represents a particular operand of the instruction. In the | future, there may be subclasses to test the same predicates | on multiple operands (including for variadic instructions). \ Operand Predicate(s) - Test the type, register bank, etc. of an operand. This is where the ComplexPattern equivalent will be represented. It's also nested-instruction matching will live as a predicate that follows the DefUse chain to the Def and tests a MatcherRule from that position. Support for multiple instruction matchers in a rule has been retained from the existing code but has been adjusted to assert when it is used. Previously it would silently drop all but the first instruction matcher. The tablegen-erated file is not functionally changed but has more parentheses and no longer attempts to format the if-statements since keeping track of the indentation is tricky in the presence of the matcher hierarchy. It would be nice to have CMakes tablegen() run the output through clang-format (when available) so we don't have to complicate TableGen with pretty-printing. It's also worth mentioning that this hierarchy will also be able to emit TableGen definitions and test cases in the near future. This is the reason for favouring explicit emit*() calls rather than the << operator. Reviewers: aditya_nandakumar, rovka, t.p.northover, qcolombet, ab Reviewed By: ab Subscribers: igorb, dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D28942 llvm-svn: 293172
2017-01-26 19:10:14 +08:00
}
[GlobalISel] Add basic Selector-emitter tblgen backend. This adds a basic tablegen backend that analyzes the SelectionDAG patterns to find simple ones that are eligible for GlobalISel-emission. That's similar to FastISel, with one notable difference: we're not fed ISD opcodes, so we need to map the SDNode operators to generic opcodes. That's done using GINodeEquiv in TargetGlobalISel.td. Otherwise, this is mostly boilerplate, and lots of filtering of any kind of "complicated" pattern. On AArch64, this is sufficient to match G_ADD up to s64 (to ADDWrr/ADDXrr) and G_BR (to B). Differential Revision: https://reviews.llvm.org/D26878 llvm-svn: 290284
2016-12-22 07:26:20 +08:00
};
[TableGen][GlobalISel] Add a common class for all PredicateMatcher NFC. llvm-svn: 320767
2017-12-15 07:44:07 +08:00
class PredicateMatcher {
[globalisel] Re-factor ISel matchers into a hierarchy. NFC Summary: This should make it possible to easily add everything needed to import all the existing SelectionDAG rules. It should also serve the likely kinds of GlobalISel rules (some of which are not currently representable in SelectionDAG) once we've nailed down the tablegen definition for that. The hierarchy is as follows: MatcherRule - A matching rule. Currently used to emit C++ ISel code but will | also be used to emit test cases and tablegen definitions in the | near future. |- Instruction(s) - Represents the instruction to be matched. |- Instruction Predicate(s) - Test the opcode, arithmetic flags, etc. of an | instruction. \- Operand(s) - Represents a particular operand of the instruction. In the | future, there may be subclasses to test the same predicates | on multiple operands (including for variadic instructions). \ Operand Predicate(s) - Test the type, register bank, etc. of an operand. This is where the ComplexPattern equivalent will be represented. It's also nested-instruction matching will live as a predicate that follows the DefUse chain to the Def and tests a MatcherRule from that position. Support for multiple instruction matchers in a rule has been retained from the existing code but has been adjusted to assert when it is used. Previously it would silently drop all but the first instruction matcher. The tablegen-erated file is not functionally changed but has more parentheses and no longer attempts to format the if-statements since keeping track of the indentation is tricky in the presence of the matcher hierarchy. It would be nice to have CMakes tablegen() run the output through clang-format (when available) so we don't have to complicate TableGen with pretty-printing. It's also worth mentioning that this hierarchy will also be able to emit TableGen definitions and test cases in the near future. This is the reason for favouring explicit emit*() calls rather than the << operator. Reviewers: aditya_nandakumar, rovka, t.p.northover, qcolombet, ab Reviewed By: ab Subscribers: igorb, dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D28942 llvm-svn: 293172
2017-01-26 19:10:14 +08:00
public:
[globalisel] Sort RuleMatchers by priority. Summary: This makes more important rules have priority over less important rules. For example, '%a = G_ADD $b:s64, $c:s64' has priority over '%a = G_ADD $b:s32, $c:s32'. Previously these rules were emitted in the correct order by chance. NFC in this patch but it is required to make the next patch work correctly. Depends on D29710 Reviewers: t.p.northover, ab, qcolombet, aditya_nandakumar, rovka Reviewed By: ab, rovka Subscribers: javed.absar, dberris, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D29711 llvm-svn: 296121
2017-02-24 21:58:11 +08:00
/// This enum is used for RTTI and also defines the priority that is given to
/// the predicate when generating the matcher code. Kinds with higher priority
/// must be tested first.
///
[globalisel] Decouple src pattern operands from dst pattern operands. Summary: This isn't testable for AArch64 by itself so this patch also adds support for constant immediates in the pattern and physical register uses in the result. The new IntOperandMatcher matches the constant in patterns such as '(set $rd:GPR32, (G_XOR $rs:GPR32, -1))'. It's always safe to fold immediates into an instruction so this is the first rule that will match across multiple BB's. The Renderer hierarchy is responsible for adding operands to the result instruction. Renderers can copy operands (CopyRenderer) or add physical registers (in particular %wzr and %xzr) to the result instruction in any order (OperandMatchers now import the operand names from SelectionDAG to allow renderers to access any operand). This allows us to emit the result instruction for: %1 = G_XOR %0, -1 --> %1 = ORNWrr %wzr, %0 %1 = G_XOR -1, %0 --> %1 = ORNWrr %wzr, %0 although the latter is untested since the matcher/importer has not been taught about commutativity yet. Added BuildMIAction which can build new instructions and mutate them where possible. W.r.t the mutation aspect, MatchActions are now told the name of an instruction they can recycle and BuildMIAction will emit mutation code when the renderers are appropriate. They are appropriate when all operands are rendered using CopyRenderer and the indices are the same as the matcher. This currently assumes that all operands have at least one matcher. Finally, this change also fixes a crash in AArch64InstructionSelector::select() caused by an immediate operand passing isImm() rather than isCImm(). This was uncovered by the other changes and was detected by existing tests. Depends on D29711 Reviewers: t.p.northover, ab, qcolombet, rovka, aditya_nandakumar, javed.absar Reviewed By: rovka Subscribers: aemerson, dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D29712 llvm-svn: 296131
2017-02-24 23:43:30 +08:00
/// The relative priority of OPM_LLT, OPM_RegBank, and OPM_MBB do not matter
/// but OPM_Int must have priority over OPM_RegBank since constant integers
/// are represented by a virtual register defined by a G_CONSTANT instruction.
[TableGen][GlobalISel] Add a common class for all PredicateMatcher NFC. llvm-svn: 320767
2017-12-15 07:44:07 +08:00
///
/// Note: The relative priority between IPM_ and OPM_ does not matter, they
/// are currently not compared between each other.
[globalisel] Sort RuleMatchers by priority. Summary: This makes more important rules have priority over less important rules. For example, '%a = G_ADD $b:s64, $c:s64' has priority over '%a = G_ADD $b:s32, $c:s32'. Previously these rules were emitted in the correct order by chance. NFC in this patch but it is required to make the next patch work correctly. Depends on D29710 Reviewers: t.p.northover, ab, qcolombet, aditya_nandakumar, rovka Reviewed By: ab, rovka Subscribers: javed.absar, dberris, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D29711 llvm-svn: 296121
2017-02-24 21:58:11 +08:00
enum PredicateKind {
[TableGen][GlobalISel] Add a common class for all PredicateMatcher NFC. llvm-svn: 320767
2017-12-15 07:44:07 +08:00
IPM_Opcode,
Reapply r332907 "[GlobalISel] Improving InstructionSelect's performance by reducing MatchTable..." Apparently the compile time problem was caused by the fact that not all compilers / STL implementations can automatically convert std::unique_ptr<Derived> to std::unique_ptr<Base>. Fixed (hopefully) by making sure it's std::unique_ptr<Derived>&& (rvalue ref) to std::unique_ptr<Base> conversion instead. llvm-svn: 332917
2018-05-22 07:28:51 +08:00
IPM_NumOperands,
[TableGen][GlobalISel] Add a common class for all PredicateMatcher NFC. llvm-svn: 320767
2017-12-15 07:44:07 +08:00
IPM_ImmPredicate,
IPM_AtomicOrderingMMO,
[globalisel] Update GlobalISel emitter to match new representation of extending loads Summary: Previously, a extending load was represented at (G_*EXT (G_LOAD x)). This had a few drawbacks: * G_LOAD had to be legal for all sizes you could extend from, even if registers didn't naturally hold those sizes. * All sizes you could extend from had to be allocatable just in case the extend went missing (e.g. by optimization). * At minimum, G_*EXT and G_TRUNC had to be legal for these sizes. As we improve optimization of extends and truncates, this legality requirement would spread without considerable care w.r.t when certain combines were permitted. * The SelectionDAG importer required some ugly and fragile pattern rewriting to translate patterns into this style. This patch changes the representation to: * (G_[SZ]EXTLOAD x) * (G_LOAD x) any-extends when MMO.getSize() * 8 < ResultTy.getSizeInBits() which resolves these issues by allowing targets to work entirely in their native register sizes, and by having a more direct translation from SelectionDAG patterns. Each extending load can be lowered by the legalizer into separate extends and loads, however a target that supports s1 will need the any-extending load to extend to at least s8 since LLVM does not represent memory accesses smaller than 8 bit. The legalizer can widenScalar G_LOAD into an any-extending load but sign/zero-extending loads need help from something else like a combiner pass. A follow-up patch that adds combiner helpers for for this will follow. The new representation requires that the MMO correctly reflect the memory access so this has been corrected in a couple tests. I've also moved the extending loads to their own tests since they are (mostly) separate opcodes now. Additionally, the re-write appears to have invalidated two tests from select-with-no-legality-check.mir since the matcher table no longer contains loads that result in s1's and they aren't legal in AArch64 anymore. Depends on D45540 Reviewers: ab, aditya_nandakumar, bogner, rtereshin, volkan, rovka, javed.absar Reviewed By: rtereshin Subscribers: javed.absar, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D45541 llvm-svn: 331601
2018-05-06 04:53:24 +08:00
IPM_MemoryLLTSize,
IPM_MemoryVsLLTSize,
[globalisel][tablegen] Fix small spelling nits. NFC ComplexRendererFn -> ComplexRendererFns Corrected a couple lingering references to tied operands that were missed. llvm-svn: 316237
2017-10-21 04:55:29 +08:00
OPM_SameOperand,
[globalisel][tblgen] Add support for ComplexPatterns Summary: Adds a new kind of MachineOperand: MO_Placeholder. This operand must not appear in the MIR and only exists as a way of creating an 'uninitialized' operand until a matcher function overwrites it. Depends on D30046, D29712 Reviewers: t.p.northover, ab, rovka, aditya_nandakumar, javed.absar, qcolombet Reviewed By: qcolombet Subscribers: dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D30089 llvm-svn: 297782
2017-03-15 05:32:08 +08:00
OPM_ComplexPattern,
[globalisel][tablegen] Correct matching of intrinsic ID's. TreePatternNode considers them to be plain integers but MachineInstr considers them to be a distinct kind of operand. The tweak to AArch64InstrInfo.td to produce a simple test case is a NFC for everything except GlobalISelEmitter (confirmed by diffing the tablegenerated files). GlobalISelEmitter is currently unable to infer the type of operands in the Dst pattern from the operands in the Src pattern. llvm-svn: 307634
2017-07-11 16:57:29 +08:00
OPM_IntrinsicID,
[globalisel][tablegen] Add support for importing 'imm' operands. Summary: This patch enables the import of rules containing 'imm' operands that do not constrain the acceptable values using predicates. Support for ImmLeaf will arrive in a later patch. Depends on D35681 Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Reviewed By: rovka Subscribers: kristof.beyls, javed.absar, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D35833 llvm-svn: 310343
2017-08-08 18:44:31 +08:00
OPM_Instruction,
[globalisel] Decouple src pattern operands from dst pattern operands. Summary: This isn't testable for AArch64 by itself so this patch also adds support for constant immediates in the pattern and physical register uses in the result. The new IntOperandMatcher matches the constant in patterns such as '(set $rd:GPR32, (G_XOR $rs:GPR32, -1))'. It's always safe to fold immediates into an instruction so this is the first rule that will match across multiple BB's. The Renderer hierarchy is responsible for adding operands to the result instruction. Renderers can copy operands (CopyRenderer) or add physical registers (in particular %wzr and %xzr) to the result instruction in any order (OperandMatchers now import the operand names from SelectionDAG to allow renderers to access any operand). This allows us to emit the result instruction for: %1 = G_XOR %0, -1 --> %1 = ORNWrr %wzr, %0 %1 = G_XOR -1, %0 --> %1 = ORNWrr %wzr, %0 although the latter is untested since the matcher/importer has not been taught about commutativity yet. Added BuildMIAction which can build new instructions and mutate them where possible. W.r.t the mutation aspect, MatchActions are now told the name of an instruction they can recycle and BuildMIAction will emit mutation code when the renderers are appropriate. They are appropriate when all operands are rendered using CopyRenderer and the indices are the same as the matcher. This currently assumes that all operands have at least one matcher. Finally, this change also fixes a crash in AArch64InstructionSelector::select() caused by an immediate operand passing isImm() rather than isCImm(). This was uncovered by the other changes and was detected by existing tests. Depends on D29711 Reviewers: t.p.northover, ab, qcolombet, rovka, aditya_nandakumar, javed.absar Reviewed By: rovka Subscribers: aemerson, dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D29712 llvm-svn: 296131
2017-02-24 23:43:30 +08:00
OPM_Int,
[globalisel][tablegen] Add support for (set $dst, 1) and test X86's OptForSize predicate. Summary: It's rare but a small number of patterns use IntInit's at the root of the match. On X86, one such rule is enabled by the OptForSize predicate and causes the compiler to use the smaller: %0 = MOV32r1 instead of the usual: %0 = MOV32ri 1 This patch adds support for matching IntInit's at the root and uses this as a test case for the optsize attribute that was implemented in r301750 Reviewers: qcolombet, ab, t.p.northover, rovka, kristof.beyls, aditya_nandakumar Reviewed By: qcolombet Subscribers: igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D32791 llvm-svn: 303678
2017-05-24 03:33:16 +08:00
OPM_LiteralInt,
[globalisel] Sort RuleMatchers by priority. Summary: This makes more important rules have priority over less important rules. For example, '%a = G_ADD $b:s64, $c:s64' has priority over '%a = G_ADD $b:s32, $c:s32'. Previously these rules were emitted in the correct order by chance. NFC in this patch but it is required to make the next patch work correctly. Depends on D29710 Reviewers: t.p.northover, ab, qcolombet, aditya_nandakumar, rovka Reviewed By: ab, rovka Subscribers: javed.absar, dberris, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D29711 llvm-svn: 296121
2017-02-24 21:58:11 +08:00
OPM_LLT,
[globalisel][tablegen] Implement unindexed load, non-extending load, and MemVT checks Summary: This includes some context-sensitivity in the MVT to LLT conversion so that pointer types are tested correctly. FIXME: I'm not happy with the way this is done since everything is a special-case. I've yet to find a reasonable way to implement it. select-load.mir fails because <1 x s64> loads in tablegen get priority over s64 loads. This is fixed in the next patch and as such they should be committed together, I've posted them separately to help with the review. Depends on D37456 Reviewers: ab, qcolombet, t.p.northover, rovka, aditya_nandakumar Subscribers: kristof.beyls, javed.absar, llvm-commits, igorb Differential Revision: https://reviews.llvm.org/D37457 llvm-svn: 315884
2017-10-16 08:56:30 +08:00
OPM_PointerToAny,
[globalisel] Sort RuleMatchers by priority. Summary: This makes more important rules have priority over less important rules. For example, '%a = G_ADD $b:s64, $c:s64' has priority over '%a = G_ADD $b:s32, $c:s32'. Previously these rules were emitted in the correct order by chance. NFC in this patch but it is required to make the next patch work correctly. Depends on D29710 Reviewers: t.p.northover, ab, qcolombet, aditya_nandakumar, rovka Reviewed By: ab, rovka Subscribers: javed.absar, dberris, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D29711 llvm-svn: 296121
2017-02-24 21:58:11 +08:00
OPM_RegBank,
OPM_MBB,
};
protected:
PredicateKind Kind;
[TableGen][GlobalISel] Have the predicate directly know which data they are dealing with Prior to this patch, a predicate wouldn't make sense outside of its rule. Indeed, it was only during emitting a rule that a predicate would be made aware of the IDs of the data it is checking. Because of that, predicates could not be moved around or compared between each other. NFC. llvm-svn: 320887
2017-12-16 07:07:42 +08:00
unsigned InsnVarID;
unsigned OpIdx;
[globalisel] Sort RuleMatchers by priority. Summary: This makes more important rules have priority over less important rules. For example, '%a = G_ADD $b:s64, $c:s64' has priority over '%a = G_ADD $b:s32, $c:s32'. Previously these rules were emitted in the correct order by chance. NFC in this patch but it is required to make the next patch work correctly. Depends on D29710 Reviewers: t.p.northover, ab, qcolombet, aditya_nandakumar, rovka Reviewed By: ab, rovka Subscribers: javed.absar, dberris, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D29711 llvm-svn: 296121
2017-02-24 21:58:11 +08:00
public:
[TableGen][GlobalISel] Have the predicate directly know which data they are dealing with Prior to this patch, a predicate wouldn't make sense outside of its rule. Indeed, it was only during emitting a rule that a predicate would be made aware of the IDs of the data it is checking. Because of that, predicates could not be moved around or compared between each other. NFC. llvm-svn: 320887
2017-12-16 07:07:42 +08:00
PredicateMatcher(PredicateKind Kind, unsigned InsnVarID, unsigned OpIdx = ~0)
: Kind(Kind), InsnVarID(InsnVarID), OpIdx(OpIdx) {}
[TableGen][GlobalISel] Add a common class for all PredicateMatcher NFC. llvm-svn: 320767
2017-12-15 07:44:07 +08:00
Reapply r332907 "[GlobalISel] Improving InstructionSelect's performance by reducing MatchTable..." Apparently the compile time problem was caused by the fact that not all compilers / STL implementations can automatically convert std::unique_ptr<Derived> to std::unique_ptr<Base>. Fixed (hopefully) by making sure it's std::unique_ptr<Derived>&& (rvalue ref) to std::unique_ptr<Base> conversion instead. llvm-svn: 332917
2018-05-22 07:28:51 +08:00
unsigned getInsnVarID() const { return InsnVarID; }
[TableGen][GlobalISel] Have the predicate directly know which data they are dealing with Prior to this patch, a predicate wouldn't make sense outside of its rule. Indeed, it was only during emitting a rule that a predicate would be made aware of the IDs of the data it is checking. Because of that, predicates could not be moved around or compared between each other. NFC. llvm-svn: 320887
2017-12-16 07:07:42 +08:00
unsigned getOpIdx() const { return OpIdx; }
Reapply r332907 "[GlobalISel] Improving InstructionSelect's performance by reducing MatchTable..." Apparently the compile time problem was caused by the fact that not all compilers / STL implementations can automatically convert std::unique_ptr<Derived> to std::unique_ptr<Base>. Fixed (hopefully) by making sure it's std::unique_ptr<Derived>&& (rvalue ref) to std::unique_ptr<Base> conversion instead. llvm-svn: 332917
2018-05-22 07:28:51 +08:00
[TableGen][GlobalISel] Add a common class for all PredicateMatcher NFC. llvm-svn: 320767
2017-12-15 07:44:07 +08:00
virtual ~PredicateMatcher() = default;
/// Emit MatchTable opcodes that check the predicate for the given operand.
[TableGen][GlobalISel] Have the predicate directly know which data they are dealing with Prior to this patch, a predicate wouldn't make sense outside of its rule. Indeed, it was only during emitting a rule that a predicate would be made aware of the IDs of the data it is checking. Because of that, predicates could not be moved around or compared between each other. NFC. llvm-svn: 320887
2017-12-16 07:07:42 +08:00
virtual void emitPredicateOpcodes(MatchTable &Table,
RuleMatcher &Rule) const = 0;
[globalisel] Re-factor ISel matchers into a hierarchy. NFC Summary: This should make it possible to easily add everything needed to import all the existing SelectionDAG rules. It should also serve the likely kinds of GlobalISel rules (some of which are not currently representable in SelectionDAG) once we've nailed down the tablegen definition for that. The hierarchy is as follows: MatcherRule - A matching rule. Currently used to emit C++ ISel code but will | also be used to emit test cases and tablegen definitions in the | near future. |- Instruction(s) - Represents the instruction to be matched. |- Instruction Predicate(s) - Test the opcode, arithmetic flags, etc. of an | instruction. \- Operand(s) - Represents a particular operand of the instruction. In the | future, there may be subclasses to test the same predicates | on multiple operands (including for variadic instructions). \ Operand Predicate(s) - Test the type, register bank, etc. of an operand. This is where the ComplexPattern equivalent will be represented. It's also nested-instruction matching will live as a predicate that follows the DefUse chain to the Def and tests a MatcherRule from that position. Support for multiple instruction matchers in a rule has been retained from the existing code but has been adjusted to assert when it is used. Previously it would silently drop all but the first instruction matcher. The tablegen-erated file is not functionally changed but has more parentheses and no longer attempts to format the if-statements since keeping track of the indentation is tricky in the presence of the matcher hierarchy. It would be nice to have CMakes tablegen() run the output through clang-format (when available) so we don't have to complicate TableGen with pretty-printing. It's also worth mentioning that this hierarchy will also be able to emit TableGen definitions and test cases in the near future. This is the reason for favouring explicit emit*() calls rather than the << operator. Reviewers: aditya_nandakumar, rovka, t.p.northover, qcolombet, ab Reviewed By: ab Subscribers: igorb, dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D28942 llvm-svn: 293172
2017-01-26 19:10:14 +08:00
[globalisel] Sort RuleMatchers by priority. Summary: This makes more important rules have priority over less important rules. For example, '%a = G_ADD $b:s64, $c:s64' has priority over '%a = G_ADD $b:s32, $c:s32'. Previously these rules were emitted in the correct order by chance. NFC in this patch but it is required to make the next patch work correctly. Depends on D29710 Reviewers: t.p.northover, ab, qcolombet, aditya_nandakumar, rovka Reviewed By: ab, rovka Subscribers: javed.absar, dberris, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D29711 llvm-svn: 296121
2017-02-24 21:58:11 +08:00
PredicateKind getKind() const { return Kind; }
[TableGen][GlobalISel] Make the different Matcher comparable This opens refactoring opportunities in the match table now that we can check that two predicates are the same. NFC. llvm-svn: 320890
2017-12-16 07:24:39 +08:00
virtual bool isIdentical(const PredicateMatcher &B) const {
return B.getKind() == getKind() && InsnVarID == B.InsnVarID &&
OpIdx == B.OpIdx;
}
Reapply r332907 "[GlobalISel] Improving InstructionSelect's performance by reducing MatchTable..." Apparently the compile time problem was caused by the fact that not all compilers / STL implementations can automatically convert std::unique_ptr<Derived> to std::unique_ptr<Base>. Fixed (hopefully) by making sure it's std::unique_ptr<Derived>&& (rvalue ref) to std::unique_ptr<Base> conversion instead. llvm-svn: 332917
2018-05-22 07:28:51 +08:00
virtual bool isIdenticalDownToValue(const PredicateMatcher &B) const {
return hasValue() && PredicateMatcher::isIdentical(B);
}
virtual MatchTableRecord getValue() const {
assert(hasValue() && "Can not get a value of a value-less predicate!");
llvm_unreachable("Not implemented yet");
}
virtual bool hasValue() const { return false; }
/// Report the maximum number of temporary operands needed by the predicate
/// matcher.
virtual unsigned countRendererFns() const { return 0; }
[TableGen][GlobalISel] Add a common class for all PredicateMatcher NFC. llvm-svn: 320767
2017-12-15 07:44:07 +08:00
};
/// Generates code to check a predicate of an operand.
///
/// Typical predicates include:
/// * Operand is a particular register.
/// * Operand is assigned a particular register bank.
/// * Operand is an MBB.
class OperandPredicateMatcher : public PredicateMatcher {
public:
[TableGen][GlobalISel] Have the predicate directly know which data they are dealing with Prior to this patch, a predicate wouldn't make sense outside of its rule. Indeed, it was only during emitting a rule that a predicate would be made aware of the IDs of the data it is checking. Because of that, predicates could not be moved around or compared between each other. NFC. llvm-svn: 320887
2017-12-16 07:07:42 +08:00
OperandPredicateMatcher(PredicateKind Kind, unsigned InsnVarID,
unsigned OpIdx)
: PredicateMatcher(Kind, InsnVarID, OpIdx) {}
[TableGen][GlobalISel] Add a common class for all PredicateMatcher NFC. llvm-svn: 320767
2017-12-15 07:44:07 +08:00
virtual ~OperandPredicateMatcher() {}
[globalisel] Sort RuleMatchers by priority. Summary: This makes more important rules have priority over less important rules. For example, '%a = G_ADD $b:s64, $c:s64' has priority over '%a = G_ADD $b:s32, $c:s32'. Previously these rules were emitted in the correct order by chance. NFC in this patch but it is required to make the next patch work correctly. Depends on D29710 Reviewers: t.p.northover, ab, qcolombet, aditya_nandakumar, rovka Reviewed By: ab, rovka Subscribers: javed.absar, dberris, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D29711 llvm-svn: 296121
2017-02-24 21:58:11 +08:00
/// Compare the priority of this object and B.
///
/// Returns true if this object is more important than B.
[globalisel][tablegen] Add support for importing 'imm' operands. Summary: This patch enables the import of rules containing 'imm' operands that do not constrain the acceptable values using predicates. Support for ImmLeaf will arrive in a later patch. Depends on D35681 Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Reviewed By: rovka Subscribers: kristof.beyls, javed.absar, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D35833 llvm-svn: 310343
2017-08-08 18:44:31 +08:00
virtual bool isHigherPriorityThan(const OperandPredicateMatcher &B) const;
[globalisel] Re-factor ISel matchers into a hierarchy. NFC Summary: This should make it possible to easily add everything needed to import all the existing SelectionDAG rules. It should also serve the likely kinds of GlobalISel rules (some of which are not currently representable in SelectionDAG) once we've nailed down the tablegen definition for that. The hierarchy is as follows: MatcherRule - A matching rule. Currently used to emit C++ ISel code but will | also be used to emit test cases and tablegen definitions in the | near future. |- Instruction(s) - Represents the instruction to be matched. |- Instruction Predicate(s) - Test the opcode, arithmetic flags, etc. of an | instruction. \- Operand(s) - Represents a particular operand of the instruction. In the | future, there may be subclasses to test the same predicates | on multiple operands (including for variadic instructions). \ Operand Predicate(s) - Test the type, register bank, etc. of an operand. This is where the ComplexPattern equivalent will be represented. It's also nested-instruction matching will live as a predicate that follows the DefUse chain to the Def and tests a MatcherRule from that position. Support for multiple instruction matchers in a rule has been retained from the existing code but has been adjusted to assert when it is used. Previously it would silently drop all but the first instruction matcher. The tablegen-erated file is not functionally changed but has more parentheses and no longer attempts to format the if-statements since keeping track of the indentation is tricky in the presence of the matcher hierarchy. It would be nice to have CMakes tablegen() run the output through clang-format (when available) so we don't have to complicate TableGen with pretty-printing. It's also worth mentioning that this hierarchy will also be able to emit TableGen definitions and test cases in the near future. This is the reason for favouring explicit emit*() calls rather than the << operator. Reviewers: aditya_nandakumar, rovka, t.p.northover, qcolombet, ab Reviewed By: ab Subscribers: igorb, dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D28942 llvm-svn: 293172
2017-01-26 19:10:14 +08:00
};
Re-commit: [globalisel][tablegen] Add support for ImmLeaf without SDNodeXForm Summary: This patch adds support for predicates on imm nodes but only for ImmLeaf and not for PatLeaf or PatFrag and only where the value does not need to be transformed before being rendered into the instruction. The limitation on PatLeaf/PatFrag/SDNodeXForm is due to differences in the necessary target-supplied C++ for GlobalISel. Depends on D36085 The previous commit was reverted for breaking the build but this appears to have been the recurring problem on the Windows bots with tablegen not being re-run when llvm-tblgen is changed but the .td's aren't. If it re-occurs then forcing a build with clean=True should fix it but this string should do this in advance: Requires a clean build. Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Reviewed By: rovka Subscribers: kristof.beyls, javed.absar, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D36086 llvm-svn: 311645
2017-08-24 17:11:20 +08:00
template <>
std::string
PredicateListMatcher<OperandPredicateMatcher>::getNoPredicateComment() const {
return "No operand predicates";
}
[globalisel][tablegen] Simplify named operand/operator lookups and fix a wrong-code bug this revealed. Summary: Operand variable lookups are now performed by the RuleMatcher rather than searching the whole matcher hierarchy for a match. This revealed a wrong-code bug that currently affects ARM and X86 where patterns that use a variable more than once in the match pattern will be imported but won't check that the operands are identical. This can cause the tablegen-erated matcher to accept matches that should be rejected. Depends on D36569 Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Subscribers: aemerson, igorb, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D36618 llvm-svn: 315780
2017-10-14 08:31:58 +08:00
/// Generates code to check that a register operand is defined by the same exact
/// one as another.
class SameOperandMatcher : public OperandPredicateMatcher {
[globalisel][tablegen] Fix small spelling nits. NFC ComplexRendererFn -> ComplexRendererFns Corrected a couple lingering references to tied operands that were missed. llvm-svn: 316237
2017-10-21 04:55:29 +08:00
std::string MatchingName;
[globalisel][tablegen] Simplify named operand/operator lookups and fix a wrong-code bug this revealed. Summary: Operand variable lookups are now performed by the RuleMatcher rather than searching the whole matcher hierarchy for a match. This revealed a wrong-code bug that currently affects ARM and X86 where patterns that use a variable more than once in the match pattern will be imported but won't check that the operands are identical. This can cause the tablegen-erated matcher to accept matches that should be rejected. Depends on D36569 Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Subscribers: aemerson, igorb, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D36618 llvm-svn: 315780
2017-10-14 08:31:58 +08:00
public:
[TableGen][GlobalISel] Make the arguments of the Instruction and Operand Matchers consistent Move InsnVarID and OpIdx at the beginning of the list of arguments for all the constructors of the OperandMatcher subclasses. This matches what we do for the InstructionMatcher. NFC. llvm-svn: 321031
2017-12-19 06:12:13 +08:00
SameOperandMatcher(unsigned InsnVarID, unsigned OpIdx, StringRef MatchingName)
[TableGen][GlobalISel] Have the predicate directly know which data they are dealing with Prior to this patch, a predicate wouldn't make sense outside of its rule. Indeed, it was only during emitting a rule that a predicate would be made aware of the IDs of the data it is checking. Because of that, predicates could not be moved around or compared between each other. NFC. llvm-svn: 320887
2017-12-16 07:07:42 +08:00
: OperandPredicateMatcher(OPM_SameOperand, InsnVarID, OpIdx),
MatchingName(MatchingName) {}
[globalisel][tablegen] Simplify named operand/operator lookups and fix a wrong-code bug this revealed. Summary: Operand variable lookups are now performed by the RuleMatcher rather than searching the whole matcher hierarchy for a match. This revealed a wrong-code bug that currently affects ARM and X86 where patterns that use a variable more than once in the match pattern will be imported but won't check that the operands are identical. This can cause the tablegen-erated matcher to accept matches that should be rejected. Depends on D36569 Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Subscribers: aemerson, igorb, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D36618 llvm-svn: 315780
2017-10-14 08:31:58 +08:00
Reapply r332907 "[GlobalISel] Improving InstructionSelect's performance by reducing MatchTable..." Apparently the compile time problem was caused by the fact that not all compilers / STL implementations can automatically convert std::unique_ptr<Derived> to std::unique_ptr<Base>. Fixed (hopefully) by making sure it's std::unique_ptr<Derived>&& (rvalue ref) to std::unique_ptr<Base> conversion instead. llvm-svn: 332917
2018-05-22 07:28:51 +08:00
static bool classof(const PredicateMatcher *P) {
[globalisel][tablegen] Fix small spelling nits. NFC ComplexRendererFn -> ComplexRendererFns Corrected a couple lingering references to tied operands that were missed. llvm-svn: 316237
2017-10-21 04:55:29 +08:00
return P->getKind() == OPM_SameOperand;
[globalisel][tablegen] Simplify named operand/operator lookups and fix a wrong-code bug this revealed. Summary: Operand variable lookups are now performed by the RuleMatcher rather than searching the whole matcher hierarchy for a match. This revealed a wrong-code bug that currently affects ARM and X86 where patterns that use a variable more than once in the match pattern will be imported but won't check that the operands are identical. This can cause the tablegen-erated matcher to accept matches that should be rejected. Depends on D36569 Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Subscribers: aemerson, igorb, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D36618 llvm-svn: 315780
2017-10-14 08:31:58 +08:00
}
[TableGen][GlobalISel] Have the predicate directly know which data they are dealing with Prior to this patch, a predicate wouldn't make sense outside of its rule. Indeed, it was only during emitting a rule that a predicate would be made aware of the IDs of the data it is checking. Because of that, predicates could not be moved around or compared between each other. NFC. llvm-svn: 320887
2017-12-16 07:07:42 +08:00
void emitPredicateOpcodes(MatchTable &Table,
RuleMatcher &Rule) const override;
Reapply r332907 "[GlobalISel] Improving InstructionSelect's performance by reducing MatchTable..." Apparently the compile time problem was caused by the fact that not all compilers / STL implementations can automatically convert std::unique_ptr<Derived> to std::unique_ptr<Base>. Fixed (hopefully) by making sure it's std::unique_ptr<Derived>&& (rvalue ref) to std::unique_ptr<Base> conversion instead. llvm-svn: 332917
2018-05-22 07:28:51 +08:00
bool isIdentical(const PredicateMatcher &B) const override {
return OperandPredicateMatcher::isIdentical(B) &&
MatchingName == cast<SameOperandMatcher>(&B)->MatchingName;
}
[globalisel][tablegen] Simplify named operand/operator lookups and fix a wrong-code bug this revealed. Summary: Operand variable lookups are now performed by the RuleMatcher rather than searching the whole matcher hierarchy for a match. This revealed a wrong-code bug that currently affects ARM and X86 where patterns that use a variable more than once in the match pattern will be imported but won't check that the operands are identical. This can cause the tablegen-erated matcher to accept matches that should be rejected. Depends on D36569 Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Subscribers: aemerson, igorb, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D36618 llvm-svn: 315780
2017-10-14 08:31:58 +08:00
};
[globalisel] Re-factor ISel matchers into a hierarchy. NFC Summary: This should make it possible to easily add everything needed to import all the existing SelectionDAG rules. It should also serve the likely kinds of GlobalISel rules (some of which are not currently representable in SelectionDAG) once we've nailed down the tablegen definition for that. The hierarchy is as follows: MatcherRule - A matching rule. Currently used to emit C++ ISel code but will | also be used to emit test cases and tablegen definitions in the | near future. |- Instruction(s) - Represents the instruction to be matched. |- Instruction Predicate(s) - Test the opcode, arithmetic flags, etc. of an | instruction. \- Operand(s) - Represents a particular operand of the instruction. In the | future, there may be subclasses to test the same predicates | on multiple operands (including for variadic instructions). \ Operand Predicate(s) - Test the type, register bank, etc. of an operand. This is where the ComplexPattern equivalent will be represented. It's also nested-instruction matching will live as a predicate that follows the DefUse chain to the Def and tests a MatcherRule from that position. Support for multiple instruction matchers in a rule has been retained from the existing code but has been adjusted to assert when it is used. Previously it would silently drop all but the first instruction matcher. The tablegen-erated file is not functionally changed but has more parentheses and no longer attempts to format the if-statements since keeping track of the indentation is tricky in the presence of the matcher hierarchy. It would be nice to have CMakes tablegen() run the output through clang-format (when available) so we don't have to complicate TableGen with pretty-printing. It's also worth mentioning that this hierarchy will also be able to emit TableGen definitions and test cases in the near future. This is the reason for favouring explicit emit*() calls rather than the << operator. Reviewers: aditya_nandakumar, rovka, t.p.northover, qcolombet, ab Reviewed By: ab Subscribers: igorb, dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D28942 llvm-svn: 293172
2017-01-26 19:10:14 +08:00
/// Generates code to check that an operand is a particular LLT.
class LLTOperandMatcher : public OperandPredicateMatcher {
protected:
Recommit: [globalisel] Change LLT constructor string into an LLT-based object that knows how to generate it. Summary: This will allow future patches to inspect the details of the LLT. The implementation is now split between the Support and CodeGen libraries to allow TableGen to use this class without introducing layering concerns. Thanks to Ahmed Bougacha for finding a reasonable way to avoid the layering issue and providing the version of this patch without that problem. The problem with the previous commit appears to have been that TableGen was including CodeGen/LowLevelType.h instead of Support/LowLevelTypeImpl.h. Reviewers: t.p.northover, qcolombet, rovka, aditya_nandakumar, ab, javed.absar Subscribers: arsenm, nhaehnle, mgorny, dberris, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D30046 llvm-svn: 297241
2017-03-08 07:20:35 +08:00
LLTCodeGen Ty;
[GlobalISel] Add basic Selector-emitter tblgen backend. This adds a basic tablegen backend that analyzes the SelectionDAG patterns to find simple ones that are eligible for GlobalISel-emission. That's similar to FastISel, with one notable difference: we're not fed ISD opcodes, so we need to map the SDNode operators to generic opcodes. That's done using GINodeEquiv in TargetGlobalISel.td. Otherwise, this is mostly boilerplate, and lots of filtering of any kind of "complicated" pattern. On AArch64, this is sufficient to match G_ADD up to s64 (to ADDWrr/ADDXrr) and G_BR (to B). Differential Revision: https://reviews.llvm.org/D26878 llvm-svn: 290284
2016-12-22 07:26:20 +08:00
[globalisel] Re-factor ISel matchers into a hierarchy. NFC Summary: This should make it possible to easily add everything needed to import all the existing SelectionDAG rules. It should also serve the likely kinds of GlobalISel rules (some of which are not currently representable in SelectionDAG) once we've nailed down the tablegen definition for that. The hierarchy is as follows: MatcherRule - A matching rule. Currently used to emit C++ ISel code but will | also be used to emit test cases and tablegen definitions in the | near future. |- Instruction(s) - Represents the instruction to be matched. |- Instruction Predicate(s) - Test the opcode, arithmetic flags, etc. of an | instruction. \- Operand(s) - Represents a particular operand of the instruction. In the | future, there may be subclasses to test the same predicates | on multiple operands (including for variadic instructions). \ Operand Predicate(s) - Test the type, register bank, etc. of an operand. This is where the ComplexPattern equivalent will be represented. It's also nested-instruction matching will live as a predicate that follows the DefUse chain to the Def and tests a MatcherRule from that position. Support for multiple instruction matchers in a rule has been retained from the existing code but has been adjusted to assert when it is used. Previously it would silently drop all but the first instruction matcher. The tablegen-erated file is not functionally changed but has more parentheses and no longer attempts to format the if-statements since keeping track of the indentation is tricky in the presence of the matcher hierarchy. It would be nice to have CMakes tablegen() run the output through clang-format (when available) so we don't have to complicate TableGen with pretty-printing. It's also worth mentioning that this hierarchy will also be able to emit TableGen definitions and test cases in the near future. This is the reason for favouring explicit emit*() calls rather than the << operator. Reviewers: aditya_nandakumar, rovka, t.p.northover, qcolombet, ab Reviewed By: ab Subscribers: igorb, dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D28942 llvm-svn: 293172
2017-01-26 19:10:14 +08:00
public:
Reapply r332907 "[GlobalISel] Improving InstructionSelect's performance by reducing MatchTable..." Apparently the compile time problem was caused by the fact that not all compilers / STL implementations can automatically convert std::unique_ptr<Derived> to std::unique_ptr<Base>. Fixed (hopefully) by making sure it's std::unique_ptr<Derived>&& (rvalue ref) to std::unique_ptr<Base> conversion instead. llvm-svn: 332917
2018-05-22 07:28:51 +08:00
static std::map<LLTCodeGen, unsigned> TypeIDValues;
static void initTypeIDValuesMap() {
TypeIDValues.clear();
unsigned ID = 0;
for (const LLTCodeGen LLTy : KnownTypes)
TypeIDValues[LLTy] = ID++;
}
[TableGen][GlobalISel] Make the arguments of the Instruction and Operand Matchers consistent Move InsnVarID and OpIdx at the beginning of the list of arguments for all the constructors of the OperandMatcher subclasses. This matches what we do for the InstructionMatcher. NFC. llvm-svn: 321031
2017-12-19 06:12:13 +08:00
LLTOperandMatcher(unsigned InsnVarID, unsigned OpIdx, const LLTCodeGen &Ty)
[TableGen][GlobalISel] Have the predicate directly know which data they are dealing with Prior to this patch, a predicate wouldn't make sense outside of its rule. Indeed, it was only during emitting a rule that a predicate would be made aware of the IDs of the data it is checking. Because of that, predicates could not be moved around or compared between each other. NFC. llvm-svn: 320887
2017-12-16 07:07:42 +08:00
: OperandPredicateMatcher(OPM_LLT, InsnVarID, OpIdx), Ty(Ty) {
[globalisel][tablegen] Generate TypeObject table. NFC Summary: Generate the type table from the types used by a target rather than hard-coding the union of types used by all targets. Depends on D36084 Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Reviewed By: rovka Subscribers: kristof.beyls, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D36085 llvm-svn: 311084
2017-08-17 21:18:35 +08:00
KnownTypes.insert(Ty);
}
[globalisel] Sort RuleMatchers by priority. Summary: This makes more important rules have priority over less important rules. For example, '%a = G_ADD $b:s64, $c:s64' has priority over '%a = G_ADD $b:s32, $c:s32'. Previously these rules were emitted in the correct order by chance. NFC in this patch but it is required to make the next patch work correctly. Depends on D29710 Reviewers: t.p.northover, ab, qcolombet, aditya_nandakumar, rovka Reviewed By: ab, rovka Subscribers: javed.absar, dberris, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D29711 llvm-svn: 296121
2017-02-24 21:58:11 +08:00
[TableGen][GlobalISel] Add a common class for all PredicateMatcher NFC. llvm-svn: 320767
2017-12-15 07:44:07 +08:00
static bool classof(const PredicateMatcher *P) {
[globalisel] Sort RuleMatchers by priority. Summary: This makes more important rules have priority over less important rules. For example, '%a = G_ADD $b:s64, $c:s64' has priority over '%a = G_ADD $b:s32, $c:s32'. Previously these rules were emitted in the correct order by chance. NFC in this patch but it is required to make the next patch work correctly. Depends on D29710 Reviewers: t.p.northover, ab, qcolombet, aditya_nandakumar, rovka Reviewed By: ab, rovka Subscribers: javed.absar, dberris, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D29711 llvm-svn: 296121
2017-02-24 21:58:11 +08:00
return P->getKind() == OPM_LLT;
}
[TableGen][GlobalISel] Make the different Matcher comparable This opens refactoring opportunities in the match table now that we can check that two predicates are the same. NFC. llvm-svn: 320890
2017-12-16 07:24:39 +08:00
bool isIdentical(const PredicateMatcher &B) const override {
return OperandPredicateMatcher::isIdentical(B) &&
Ty == cast<LLTOperandMatcher>(&B)->Ty;
}
Reapply r332907 "[GlobalISel] Improving InstructionSelect's performance by reducing MatchTable..." Apparently the compile time problem was caused by the fact that not all compilers / STL implementations can automatically convert std::unique_ptr<Derived> to std::unique_ptr<Base>. Fixed (hopefully) by making sure it's std::unique_ptr<Derived>&& (rvalue ref) to std::unique_ptr<Base> conversion instead. llvm-svn: 332917
2018-05-22 07:28:51 +08:00
MatchTableRecord getValue() const override {
const auto VI = TypeIDValues.find(Ty);
if (VI == TypeIDValues.end())
return MatchTable::NamedValue(getTy().getCxxEnumValue());
return MatchTable::NamedValue(getTy().getCxxEnumValue(), VI->second);
}
bool hasValue() const override {
if (TypeIDValues.size() != KnownTypes.size())
initTypeIDValuesMap();
return TypeIDValues.count(Ty);
}
LLTCodeGen getTy() const { return Ty; }
[globalisel] Re-factor ISel matchers into a hierarchy. NFC Summary: This should make it possible to easily add everything needed to import all the existing SelectionDAG rules. It should also serve the likely kinds of GlobalISel rules (some of which are not currently representable in SelectionDAG) once we've nailed down the tablegen definition for that. The hierarchy is as follows: MatcherRule - A matching rule. Currently used to emit C++ ISel code but will | also be used to emit test cases and tablegen definitions in the | near future. |- Instruction(s) - Represents the instruction to be matched. |- Instruction Predicate(s) - Test the opcode, arithmetic flags, etc. of an | instruction. \- Operand(s) - Represents a particular operand of the instruction. In the | future, there may be subclasses to test the same predicates | on multiple operands (including for variadic instructions). \ Operand Predicate(s) - Test the type, register bank, etc. of an operand. This is where the ComplexPattern equivalent will be represented. It's also nested-instruction matching will live as a predicate that follows the DefUse chain to the Def and tests a MatcherRule from that position. Support for multiple instruction matchers in a rule has been retained from the existing code but has been adjusted to assert when it is used. Previously it would silently drop all but the first instruction matcher. The tablegen-erated file is not functionally changed but has more parentheses and no longer attempts to format the if-statements since keeping track of the indentation is tricky in the presence of the matcher hierarchy. It would be nice to have CMakes tablegen() run the output through clang-format (when available) so we don't have to complicate TableGen with pretty-printing. It's also worth mentioning that this hierarchy will also be able to emit TableGen definitions and test cases in the near future. This is the reason for favouring explicit emit*() calls rather than the << operator. Reviewers: aditya_nandakumar, rovka, t.p.northover, qcolombet, ab Reviewed By: ab Subscribers: igorb, dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D28942 llvm-svn: 293172
2017-01-26 19:10:14 +08:00
[TableGen][GlobalISel] Have the predicate directly know which data they are dealing with Prior to this patch, a predicate wouldn't make sense outside of its rule. Indeed, it was only during emitting a rule that a predicate would be made aware of the IDs of the data it is checking. Because of that, predicates could not be moved around or compared between each other. NFC. llvm-svn: 320887
2017-12-16 07:07:42 +08:00
void emitPredicateOpcodes(MatchTable &Table,
RuleMatcher &Rule) const override {
[globalisel][tablegen] Add control-flow to the MatchTable. Summary: This will allow us to merge the various sub-tables into a single table. This is a compile-time saving at this point. However, this will also enable the optimization of a table so that similar instructions can be tested together, reducing the time spent on the matching the code. The bulk of this patch is a mechanical conversion to the new MatchTable object which is responsible for tracking label definitions and filling in the index of the jump targets. It is also responsible for nicely formatting the table. This was necessary to support the new GIM_Try opcode which takes the index to jump to if the match should fail. This value is unknown during table construction and is filled in during emission. To support nesting try-blocks (although we currently don't emit tables with nested try-blocks), GIM_Reject has been re-introduced to explicitly exit a try-block or fail the overall match if there are no active try-blocks. Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Reviewed By: rovka Subscribers: kristof.beyls, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D35117 llvm-svn: 308596
2017-07-20 17:25:44 +08:00
Table << MatchTable::Opcode("GIM_CheckType") << MatchTable::Comment("MI")
<< MatchTable::IntValue(InsnVarID) << MatchTable::Comment("Op")
<< MatchTable::IntValue(OpIdx) << MatchTable::Comment("Type")
Reapply r332907 "[GlobalISel] Improving InstructionSelect's performance by reducing MatchTable..." Apparently the compile time problem was caused by the fact that not all compilers / STL implementations can automatically convert std::unique_ptr<Derived> to std::unique_ptr<Base>. Fixed (hopefully) by making sure it's std::unique_ptr<Derived>&& (rvalue ref) to std::unique_ptr<Base> conversion instead. llvm-svn: 332917
2018-05-22 07:28:51 +08:00
<< getValue() << MatchTable::LineBreak;
[GlobalISel] Add basic Selector-emitter tblgen backend. This adds a basic tablegen backend that analyzes the SelectionDAG patterns to find simple ones that are eligible for GlobalISel-emission. That's similar to FastISel, with one notable difference: we're not fed ISD opcodes, so we need to map the SDNode operators to generic opcodes. That's done using GINodeEquiv in TargetGlobalISel.td. Otherwise, this is mostly boilerplate, and lots of filtering of any kind of "complicated" pattern. On AArch64, this is sufficient to match G_ADD up to s64 (to ADDWrr/ADDXrr) and G_BR (to B). Differential Revision: https://reviews.llvm.org/D26878 llvm-svn: 290284
2016-12-22 07:26:20 +08:00
}
};
Reapply r332907 "[GlobalISel] Improving InstructionSelect's performance by reducing MatchTable..." Apparently the compile time problem was caused by the fact that not all compilers / STL implementations can automatically convert std::unique_ptr<Derived> to std::unique_ptr<Base>. Fixed (hopefully) by making sure it's std::unique_ptr<Derived>&& (rvalue ref) to std::unique_ptr<Base> conversion instead. llvm-svn: 332917
2018-05-22 07:28:51 +08:00
std::map<LLTCodeGen, unsigned> LLTOperandMatcher::TypeIDValues;
[globalisel][tablegen] Implement unindexed load, non-extending load, and MemVT checks Summary: This includes some context-sensitivity in the MVT to LLT conversion so that pointer types are tested correctly. FIXME: I'm not happy with the way this is done since everything is a special-case. I've yet to find a reasonable way to implement it. select-load.mir fails because <1 x s64> loads in tablegen get priority over s64 loads. This is fixed in the next patch and as such they should be committed together, I've posted them separately to help with the review. Depends on D37456 Reviewers: ab, qcolombet, t.p.northover, rovka, aditya_nandakumar Subscribers: kristof.beyls, javed.absar, llvm-commits, igorb Differential Revision: https://reviews.llvm.org/D37457 llvm-svn: 315884
2017-10-16 08:56:30 +08:00
/// Generates code to check that an operand is a pointer to any address space.
///
/// In SelectionDAG, the types did not describe pointers or address spaces. As a
/// result, iN is used to describe a pointer of N bits to any address space and
/// PatFrag predicates are typically used to constrain the address space. There's
/// no reliable means to derive the missing type information from the pattern so
/// imported rules must test the components of a pointer separately.
///
Re-commit r315885: [globalisel][tblgen] Add support for iPTR and implement am_unscaled* and am_indexed* Summary: iPTR is a pointer of subtarget-specific size to any address space. Therefore type checks on this size derive the SizeInBits from a subtarget hook. At this point, we can import the simplests G_LOAD rules and select load instructions using them. Further patches will support for the predicates to enable additional loads as well as the stores. The previous commit failed on MSVC due to a failure to convert an initializer_list to a std::vector. Hopefully, MSVC will accept this version. Depends on D37457 Reviewers: ab, qcolombet, t.p.northover, rovka, aditya_nandakumar Reviewed By: qcolombet Subscribers: kristof.beyls, javed.absar, llvm-commits, igorb Differential Revision: https://reviews.llvm.org/D37458 llvm-svn: 315887
2017-10-16 11:36:29 +08:00
/// If SizeInBits is zero, then the pointer size will be obtained from the
/// subtarget.
[globalisel][tablegen] Implement unindexed load, non-extending load, and MemVT checks Summary: This includes some context-sensitivity in the MVT to LLT conversion so that pointer types are tested correctly. FIXME: I'm not happy with the way this is done since everything is a special-case. I've yet to find a reasonable way to implement it. select-load.mir fails because <1 x s64> loads in tablegen get priority over s64 loads. This is fixed in the next patch and as such they should be committed together, I've posted them separately to help with the review. Depends on D37456 Reviewers: ab, qcolombet, t.p.northover, rovka, aditya_nandakumar Subscribers: kristof.beyls, javed.absar, llvm-commits, igorb Differential Revision: https://reviews.llvm.org/D37457 llvm-svn: 315884
2017-10-16 08:56:30 +08:00
class PointerToAnyOperandMatcher : public OperandPredicateMatcher {
protected:
unsigned SizeInBits;
public:
[TableGen][GlobalISel] Make the arguments of the Instruction and Operand Matchers consistent Move InsnVarID and OpIdx at the beginning of the list of arguments for all the constructors of the OperandMatcher subclasses. This matches what we do for the InstructionMatcher. NFC. llvm-svn: 321031
2017-12-19 06:12:13 +08:00
PointerToAnyOperandMatcher(unsigned InsnVarID, unsigned OpIdx,
unsigned SizeInBits)
[TableGen][GlobalISel] Have the predicate directly know which data they are dealing with Prior to this patch, a predicate wouldn't make sense outside of its rule. Indeed, it was only during emitting a rule that a predicate would be made aware of the IDs of the data it is checking. Because of that, predicates could not be moved around or compared between each other. NFC. llvm-svn: 320887
2017-12-16 07:07:42 +08:00
: OperandPredicateMatcher(OPM_PointerToAny, InsnVarID, OpIdx),
SizeInBits(SizeInBits) {}
[globalisel][tablegen] Implement unindexed load, non-extending load, and MemVT checks Summary: This includes some context-sensitivity in the MVT to LLT conversion so that pointer types are tested correctly. FIXME: I'm not happy with the way this is done since everything is a special-case. I've yet to find a reasonable way to implement it. select-load.mir fails because <1 x s64> loads in tablegen get priority over s64 loads. This is fixed in the next patch and as such they should be committed together, I've posted them separately to help with the review. Depends on D37456 Reviewers: ab, qcolombet, t.p.northover, rovka, aditya_nandakumar Subscribers: kristof.beyls, javed.absar, llvm-commits, igorb Differential Revision: https://reviews.llvm.org/D37457 llvm-svn: 315884
2017-10-16 08:56:30 +08:00
static bool classof(const OperandPredicateMatcher *P) {
return P->getKind() == OPM_PointerToAny;
}
[TableGen][GlobalISel] Have the predicate directly know which data they are dealing with Prior to this patch, a predicate wouldn't make sense outside of its rule. Indeed, it was only during emitting a rule that a predicate would be made aware of the IDs of the data it is checking. Because of that, predicates could not be moved around or compared between each other. NFC. llvm-svn: 320887
2017-12-16 07:07:42 +08:00
void emitPredicateOpcodes(MatchTable &Table,
RuleMatcher &Rule) const override {
Table << MatchTable::Opcode("GIM_CheckPointerToAny")
<< MatchTable::Comment("MI") << MatchTable::IntValue(InsnVarID)
<< MatchTable::Comment("Op") << MatchTable::IntValue(OpIdx)
<< MatchTable::Comment("SizeInBits")
[globalisel][tablegen] Implement unindexed load, non-extending load, and MemVT checks Summary: This includes some context-sensitivity in the MVT to LLT conversion so that pointer types are tested correctly. FIXME: I'm not happy with the way this is done since everything is a special-case. I've yet to find a reasonable way to implement it. select-load.mir fails because <1 x s64> loads in tablegen get priority over s64 loads. This is fixed in the next patch and as such they should be committed together, I've posted them separately to help with the review. Depends on D37456 Reviewers: ab, qcolombet, t.p.northover, rovka, aditya_nandakumar Subscribers: kristof.beyls, javed.absar, llvm-commits, igorb Differential Revision: https://reviews.llvm.org/D37457 llvm-svn: 315884
2017-10-16 08:56:30 +08:00
<< MatchTable::IntValue(SizeInBits) << MatchTable::LineBreak;
}
};
[globalisel][tblgen] Add support for ComplexPatterns Summary: Adds a new kind of MachineOperand: MO_Placeholder. This operand must not appear in the MIR and only exists as a way of creating an 'uninitialized' operand until a matcher function overwrites it. Depends on D30046, D29712 Reviewers: t.p.northover, ab, rovka, aditya_nandakumar, javed.absar, qcolombet Reviewed By: qcolombet Subscribers: dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D30089 llvm-svn: 297782
2017-03-15 05:32:08 +08:00
/// Generates code to check that an operand is a particular target constant.
class ComplexPatternOperandMatcher : public OperandPredicateMatcher {
protected:
[globalisel][tablegen] Fix patterns involving multiple ComplexPatterns. Summary: Temporaries are now allocated to operands instead of predicates and this allocation is used to correctly pair up the rendered operands with the matched operands. Previously, ComplexPatterns were allocated temporaries independently in the Src Pattern and Dst Pattern, leading to mismatches. Additionally, the Dst Pattern failed to account for the allocated index and therefore always used temporary 0, 1, ... when it should have used base+0, base+1, ... Thanks to Aditya Nandakumar for noticing the bug. Depends on D30539 Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Reviewed By: rovka Subscribers: igorb, dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D31054 llvm-svn: 299538
2017-04-05 21:14:03 +08:00
const OperandMatcher &Operand;
[globalisel][tblgen] Add support for ComplexPatterns Summary: Adds a new kind of MachineOperand: MO_Placeholder. This operand must not appear in the MIR and only exists as a way of creating an 'uninitialized' operand until a matcher function overwrites it. Depends on D30046, D29712 Reviewers: t.p.northover, ab, rovka, aditya_nandakumar, javed.absar, qcolombet Reviewed By: qcolombet Subscribers: dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D30089 llvm-svn: 297782
2017-03-15 05:32:08 +08:00
const Record &TheDef;
[globalisel][tablegen] Fix patterns involving multiple ComplexPatterns. Summary: Temporaries are now allocated to operands instead of predicates and this allocation is used to correctly pair up the rendered operands with the matched operands. Previously, ComplexPatterns were allocated temporaries independently in the Src Pattern and Dst Pattern, leading to mismatches. Additionally, the Dst Pattern failed to account for the allocated index and therefore always used temporary 0, 1, ... when it should have used base+0, base+1, ... Thanks to Aditya Nandakumar for noticing the bug. Depends on D30539 Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Reviewed By: rovka Subscribers: igorb, dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D31054 llvm-svn: 299538
2017-04-05 21:14:03 +08:00
unsigned getAllocatedTemporariesBaseID() const;
[globalisel][tblgen] Add support for ComplexPatterns Summary: Adds a new kind of MachineOperand: MO_Placeholder. This operand must not appear in the MIR and only exists as a way of creating an 'uninitialized' operand until a matcher function overwrites it. Depends on D30046, D29712 Reviewers: t.p.northover, ab, rovka, aditya_nandakumar, javed.absar, qcolombet Reviewed By: qcolombet Subscribers: dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D30089 llvm-svn: 297782
2017-03-15 05:32:08 +08:00
public:
[TableGen][GlobalISel] Make the different Matcher comparable This opens refactoring opportunities in the match table now that we can check that two predicates are the same. NFC. llvm-svn: 320890
2017-12-16 07:24:39 +08:00
bool isIdentical(const PredicateMatcher &B) const override { return false; }
[TableGen][GlobalISel] Make the arguments of the Instruction and Operand Matchers consistent Move InsnVarID and OpIdx at the beginning of the list of arguments for all the constructors of the OperandMatcher subclasses. This matches what we do for the InstructionMatcher. NFC. llvm-svn: 321031
2017-12-19 06:12:13 +08:00
ComplexPatternOperandMatcher(unsigned InsnVarID, unsigned OpIdx,
const OperandMatcher &Operand,
const Record &TheDef)
[TableGen][GlobalISel] Have the predicate directly know which data they are dealing with Prior to this patch, a predicate wouldn't make sense outside of its rule. Indeed, it was only during emitting a rule that a predicate would be made aware of the IDs of the data it is checking. Because of that, predicates could not be moved around or compared between each other. NFC. llvm-svn: 320887
2017-12-16 07:07:42 +08:00
: OperandPredicateMatcher(OPM_ComplexPattern, InsnVarID, OpIdx),
Operand(Operand), TheDef(TheDef) {}
[globalisel][tablegen] Fix patterns involving multiple ComplexPatterns. Summary: Temporaries are now allocated to operands instead of predicates and this allocation is used to correctly pair up the rendered operands with the matched operands. Previously, ComplexPatterns were allocated temporaries independently in the Src Pattern and Dst Pattern, leading to mismatches. Additionally, the Dst Pattern failed to account for the allocated index and therefore always used temporary 0, 1, ... when it should have used base+0, base+1, ... Thanks to Aditya Nandakumar for noticing the bug. Depends on D30539 Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Reviewed By: rovka Subscribers: igorb, dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D31054 llvm-svn: 299538
2017-04-05 21:14:03 +08:00
[TableGen][GlobalISel] Add a common class for all PredicateMatcher NFC. llvm-svn: 320767
2017-12-15 07:44:07 +08:00
static bool classof(const PredicateMatcher *P) {
[globalisel][tablegen] Fix patterns involving multiple ComplexPatterns. Summary: Temporaries are now allocated to operands instead of predicates and this allocation is used to correctly pair up the rendered operands with the matched operands. Previously, ComplexPatterns were allocated temporaries independently in the Src Pattern and Dst Pattern, leading to mismatches. Additionally, the Dst Pattern failed to account for the allocated index and therefore always used temporary 0, 1, ... when it should have used base+0, base+1, ... Thanks to Aditya Nandakumar for noticing the bug. Depends on D30539 Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Reviewed By: rovka Subscribers: igorb, dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D31054 llvm-svn: 299538
2017-04-05 21:14:03 +08:00
return P->getKind() == OPM_ComplexPattern;
}
[globalisel][tblgen] Add support for ComplexPatterns Summary: Adds a new kind of MachineOperand: MO_Placeholder. This operand must not appear in the MIR and only exists as a way of creating an 'uninitialized' operand until a matcher function overwrites it. Depends on D30046, D29712 Reviewers: t.p.northover, ab, rovka, aditya_nandakumar, javed.absar, qcolombet Reviewed By: qcolombet Subscribers: dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D30089 llvm-svn: 297782
2017-03-15 05:32:08 +08:00
[TableGen][GlobalISel] Have the predicate directly know which data they are dealing with Prior to this patch, a predicate wouldn't make sense outside of its rule. Indeed, it was only during emitting a rule that a predicate would be made aware of the IDs of the data it is checking. Because of that, predicates could not be moved around or compared between each other. NFC. llvm-svn: 320887
2017-12-16 07:07:42 +08:00
void emitPredicateOpcodes(MatchTable &Table,
RuleMatcher &Rule) const override {
[globalisel][tablegen] Revise API for ComplexPattern operands to improve flexibility. Summary: Some targets need to be able to do more complex rendering than just adding an operand or two to an instruction. For example, it may need to insert an instruction to extract a subreg first, or it may need to perform an operation on the operand. In SelectionDAG, targets would create SDNode's to achieve the desired effect during the complex pattern predicate. This worked because SelectionDAG had a form of garbage collection that would take care of SDNode's that were created but not used due to a later predicate rejecting a match. This doesn't translate well to GlobalISel and the churn was wasteful. The API changes in this patch enable GlobalISel to accomplish the same thing without the waste. The API is now: InstructionSelector::OptionalComplexRendererFn selectArithImmed(MachineOperand &Root) const; where Root is the root of the match. The return value can be omitted to indicate that the predicate failed to match, or a function with the signature ComplexRendererFn can be returned. For example: return OptionalComplexRendererFn( [=](MachineInstrBuilder &MIB) { MIB.addImm(Immed).addImm(ShVal); }); adds two immediate operands to the rendered instruction. Immed and ShVal are captured from the predicate function. As an added bonus, this also reduces the amount of information we need to provide to GIComplexOperandMatcher. Depends on D31418 Reviewers: aditya_nandakumar, t.p.northover, qcolombet, rovka, ab, javed.absar Reviewed By: ab Subscribers: dberris, kristof.beyls, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D31761 llvm-svn: 301079
2017-04-22 23:11:04 +08:00
unsigned ID = getAllocatedTemporariesBaseID();
[globalisel][tablegen] Add control-flow to the MatchTable. Summary: This will allow us to merge the various sub-tables into a single table. This is a compile-time saving at this point. However, this will also enable the optimization of a table so that similar instructions can be tested together, reducing the time spent on the matching the code. The bulk of this patch is a mechanical conversion to the new MatchTable object which is responsible for tracking label definitions and filling in the index of the jump targets. It is also responsible for nicely formatting the table. This was necessary to support the new GIM_Try opcode which takes the index to jump to if the match should fail. This value is unknown during table construction and is filled in during emission. To support nesting try-blocks (although we currently don't emit tables with nested try-blocks), GIM_Reject has been re-introduced to explicitly exit a try-block or fail the overall match if there are no active try-blocks. Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Reviewed By: rovka Subscribers: kristof.beyls, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D35117 llvm-svn: 308596
2017-07-20 17:25:44 +08:00
Table << MatchTable::Opcode("GIM_CheckComplexPattern")
<< MatchTable::Comment("MI") << MatchTable::IntValue(InsnVarID)
<< MatchTable::Comment("Op") << MatchTable::IntValue(OpIdx)
<< MatchTable::Comment("Renderer") << MatchTable::IntValue(ID)
<< MatchTable::NamedValue(("GICP_" + TheDef.getName()).str())
<< MatchTable::LineBreak;
[globalisel][tblgen] Add support for ComplexPatterns Summary: Adds a new kind of MachineOperand: MO_Placeholder. This operand must not appear in the MIR and only exists as a way of creating an 'uninitialized' operand until a matcher function overwrites it. Depends on D30046, D29712 Reviewers: t.p.northover, ab, rovka, aditya_nandakumar, javed.absar, qcolombet Reviewed By: qcolombet Subscribers: dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D30089 llvm-svn: 297782
2017-03-15 05:32:08 +08:00
}
[globalisel][tablegen] Revise API for ComplexPattern operands to improve flexibility. Summary: Some targets need to be able to do more complex rendering than just adding an operand or two to an instruction. For example, it may need to insert an instruction to extract a subreg first, or it may need to perform an operation on the operand. In SelectionDAG, targets would create SDNode's to achieve the desired effect during the complex pattern predicate. This worked because SelectionDAG had a form of garbage collection that would take care of SDNode's that were created but not used due to a later predicate rejecting a match. This doesn't translate well to GlobalISel and the churn was wasteful. The API changes in this patch enable GlobalISel to accomplish the same thing without the waste. The API is now: InstructionSelector::OptionalComplexRendererFn selectArithImmed(MachineOperand &Root) const; where Root is the root of the match. The return value can be omitted to indicate that the predicate failed to match, or a function with the signature ComplexRendererFn can be returned. For example: return OptionalComplexRendererFn( [=](MachineInstrBuilder &MIB) { MIB.addImm(Immed).addImm(ShVal); }); adds two immediate operands to the rendered instruction. Immed and ShVal are captured from the predicate function. As an added bonus, this also reduces the amount of information we need to provide to GIComplexOperandMatcher. Depends on D31418 Reviewers: aditya_nandakumar, t.p.northover, qcolombet, rovka, ab, javed.absar Reviewed By: ab Subscribers: dberris, kristof.beyls, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D31761 llvm-svn: 301079
2017-04-22 23:11:04 +08:00
unsigned countRendererFns() const override {
return 1;
[globalisel][tblgen] Add support for ComplexPatterns Summary: Adds a new kind of MachineOperand: MO_Placeholder. This operand must not appear in the MIR and only exists as a way of creating an 'uninitialized' operand until a matcher function overwrites it. Depends on D30046, D29712 Reviewers: t.p.northover, ab, rovka, aditya_nandakumar, javed.absar, qcolombet Reviewed By: qcolombet Subscribers: dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D30089 llvm-svn: 297782
2017-03-15 05:32:08 +08:00
}
};
[globalisel] Re-factor ISel matchers into a hierarchy. NFC Summary: This should make it possible to easily add everything needed to import all the existing SelectionDAG rules. It should also serve the likely kinds of GlobalISel rules (some of which are not currently representable in SelectionDAG) once we've nailed down the tablegen definition for that. The hierarchy is as follows: MatcherRule - A matching rule. Currently used to emit C++ ISel code but will | also be used to emit test cases and tablegen definitions in the | near future. |- Instruction(s) - Represents the instruction to be matched. |- Instruction Predicate(s) - Test the opcode, arithmetic flags, etc. of an | instruction. \- Operand(s) - Represents a particular operand of the instruction. In the | future, there may be subclasses to test the same predicates | on multiple operands (including for variadic instructions). \ Operand Predicate(s) - Test the type, register bank, etc. of an operand. This is where the ComplexPattern equivalent will be represented. It's also nested-instruction matching will live as a predicate that follows the DefUse chain to the Def and tests a MatcherRule from that position. Support for multiple instruction matchers in a rule has been retained from the existing code but has been adjusted to assert when it is used. Previously it would silently drop all but the first instruction matcher. The tablegen-erated file is not functionally changed but has more parentheses and no longer attempts to format the if-statements since keeping track of the indentation is tricky in the presence of the matcher hierarchy. It would be nice to have CMakes tablegen() run the output through clang-format (when available) so we don't have to complicate TableGen with pretty-printing. It's also worth mentioning that this hierarchy will also be able to emit TableGen definitions and test cases in the near future. This is the reason for favouring explicit emit*() calls rather than the << operator. Reviewers: aditya_nandakumar, rovka, t.p.northover, qcolombet, ab Reviewed By: ab Subscribers: igorb, dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D28942 llvm-svn: 293172
2017-01-26 19:10:14 +08:00
/// Generates code to check that an operand is in a particular register bank.
class RegisterBankOperandMatcher : public OperandPredicateMatcher {
protected:
[GlobalISel] Add basic Selector-emitter tblgen backend. This adds a basic tablegen backend that analyzes the SelectionDAG patterns to find simple ones that are eligible for GlobalISel-emission. That's similar to FastISel, with one notable difference: we're not fed ISD opcodes, so we need to map the SDNode operators to generic opcodes. That's done using GINodeEquiv in TargetGlobalISel.td. Otherwise, this is mostly boilerplate, and lots of filtering of any kind of "complicated" pattern. On AArch64, this is sufficient to match G_ADD up to s64 (to ADDWrr/ADDXrr) and G_BR (to B). Differential Revision: https://reviews.llvm.org/D26878 llvm-svn: 290284
2016-12-22 07:26:20 +08:00
const CodeGenRegisterClass &RC;
[globalisel] Re-factor ISel matchers into a hierarchy. NFC Summary: This should make it possible to easily add everything needed to import all the existing SelectionDAG rules. It should also serve the likely kinds of GlobalISel rules (some of which are not currently representable in SelectionDAG) once we've nailed down the tablegen definition for that. The hierarchy is as follows: MatcherRule - A matching rule. Currently used to emit C++ ISel code but will | also be used to emit test cases and tablegen definitions in the | near future. |- Instruction(s) - Represents the instruction to be matched. |- Instruction Predicate(s) - Test the opcode, arithmetic flags, etc. of an | instruction. \- Operand(s) - Represents a particular operand of the instruction. In the | future, there may be subclasses to test the same predicates | on multiple operands (including for variadic instructions). \ Operand Predicate(s) - Test the type, register bank, etc. of an operand. This is where the ComplexPattern equivalent will be represented. It's also nested-instruction matching will live as a predicate that follows the DefUse chain to the Def and tests a MatcherRule from that position. Support for multiple instruction matchers in a rule has been retained from the existing code but has been adjusted to assert when it is used. Previously it would silently drop all but the first instruction matcher. The tablegen-erated file is not functionally changed but has more parentheses and no longer attempts to format the if-statements since keeping track of the indentation is tricky in the presence of the matcher hierarchy. It would be nice to have CMakes tablegen() run the output through clang-format (when available) so we don't have to complicate TableGen with pretty-printing. It's also worth mentioning that this hierarchy will also be able to emit TableGen definitions and test cases in the near future. This is the reason for favouring explicit emit*() calls rather than the << operator. Reviewers: aditya_nandakumar, rovka, t.p.northover, qcolombet, ab Reviewed By: ab Subscribers: igorb, dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D28942 llvm-svn: 293172
2017-01-26 19:10:14 +08:00
public:
[TableGen][GlobalISel] Make the arguments of the Instruction and Operand Matchers consistent Move InsnVarID and OpIdx at the beginning of the list of arguments for all the constructors of the OperandMatcher subclasses. This matches what we do for the InstructionMatcher. NFC. llvm-svn: 321031
2017-12-19 06:12:13 +08:00
RegisterBankOperandMatcher(unsigned InsnVarID, unsigned OpIdx,
const CodeGenRegisterClass &RC)
[TableGen][GlobalISel] Have the predicate directly know which data they are dealing with Prior to this patch, a predicate wouldn't make sense outside of its rule. Indeed, it was only during emitting a rule that a predicate would be made aware of the IDs of the data it is checking. Because of that, predicates could not be moved around or compared between each other. NFC. llvm-svn: 320887
2017-12-16 07:07:42 +08:00
: OperandPredicateMatcher(OPM_RegBank, InsnVarID, OpIdx), RC(RC) {}
[globalisel] Sort RuleMatchers by priority. Summary: This makes more important rules have priority over less important rules. For example, '%a = G_ADD $b:s64, $c:s64' has priority over '%a = G_ADD $b:s32, $c:s32'. Previously these rules were emitted in the correct order by chance. NFC in this patch but it is required to make the next patch work correctly. Depends on D29710 Reviewers: t.p.northover, ab, qcolombet, aditya_nandakumar, rovka Reviewed By: ab, rovka Subscribers: javed.absar, dberris, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D29711 llvm-svn: 296121
2017-02-24 21:58:11 +08:00
[TableGen][GlobalISel] Make the different Matcher comparable This opens refactoring opportunities in the match table now that we can check that two predicates are the same. NFC. llvm-svn: 320890
2017-12-16 07:24:39 +08:00
bool isIdentical(const PredicateMatcher &B) const override {
return OperandPredicateMatcher::isIdentical(B) &&
RC.getDef() == cast<RegisterBankOperandMatcher>(&B)->RC.getDef();
}
[TableGen][GlobalISel] Add a common class for all PredicateMatcher NFC. llvm-svn: 320767
2017-12-15 07:44:07 +08:00
static bool classof(const PredicateMatcher *P) {
[globalisel] Sort RuleMatchers by priority. Summary: This makes more important rules have priority over less important rules. For example, '%a = G_ADD $b:s64, $c:s64' has priority over '%a = G_ADD $b:s32, $c:s32'. Previously these rules were emitted in the correct order by chance. NFC in this patch but it is required to make the next patch work correctly. Depends on D29710 Reviewers: t.p.northover, ab, qcolombet, aditya_nandakumar, rovka Reviewed By: ab, rovka Subscribers: javed.absar, dberris, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D29711 llvm-svn: 296121
2017-02-24 21:58:11 +08:00
return P->getKind() == OPM_RegBank;
}
[globalisel] Re-factor ISel matchers into a hierarchy. NFC Summary: This should make it possible to easily add everything needed to import all the existing SelectionDAG rules. It should also serve the likely kinds of GlobalISel rules (some of which are not currently representable in SelectionDAG) once we've nailed down the tablegen definition for that. The hierarchy is as follows: MatcherRule - A matching rule. Currently used to emit C++ ISel code but will | also be used to emit test cases and tablegen definitions in the | near future. |- Instruction(s) - Represents the instruction to be matched. |- Instruction Predicate(s) - Test the opcode, arithmetic flags, etc. of an | instruction. \- Operand(s) - Represents a particular operand of the instruction. In the | future, there may be subclasses to test the same predicates | on multiple operands (including for variadic instructions). \ Operand Predicate(s) - Test the type, register bank, etc. of an operand. This is where the ComplexPattern equivalent will be represented. It's also nested-instruction matching will live as a predicate that follows the DefUse chain to the Def and tests a MatcherRule from that position. Support for multiple instruction matchers in a rule has been retained from the existing code but has been adjusted to assert when it is used. Previously it would silently drop all but the first instruction matcher. The tablegen-erated file is not functionally changed but has more parentheses and no longer attempts to format the if-statements since keeping track of the indentation is tricky in the presence of the matcher hierarchy. It would be nice to have CMakes tablegen() run the output through clang-format (when available) so we don't have to complicate TableGen with pretty-printing. It's also worth mentioning that this hierarchy will also be able to emit TableGen definitions and test cases in the near future. This is the reason for favouring explicit emit*() calls rather than the << operator. Reviewers: aditya_nandakumar, rovka, t.p.northover, qcolombet, ab Reviewed By: ab Subscribers: igorb, dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D28942 llvm-svn: 293172
2017-01-26 19:10:14 +08:00
[TableGen][GlobalISel] Have the predicate directly know which data they are dealing with Prior to this patch, a predicate wouldn't make sense outside of its rule. Indeed, it was only during emitting a rule that a predicate would be made aware of the IDs of the data it is checking. Because of that, predicates could not be moved around or compared between each other. NFC. llvm-svn: 320887
2017-12-16 07:07:42 +08:00
void emitPredicateOpcodes(MatchTable &Table,
RuleMatcher &Rule) const override {
[globalisel][tablegen] Add control-flow to the MatchTable. Summary: This will allow us to merge the various sub-tables into a single table. This is a compile-time saving at this point. However, this will also enable the optimization of a table so that similar instructions can be tested together, reducing the time spent on the matching the code. The bulk of this patch is a mechanical conversion to the new MatchTable object which is responsible for tracking label definitions and filling in the index of the jump targets. It is also responsible for nicely formatting the table. This was necessary to support the new GIM_Try opcode which takes the index to jump to if the match should fail. This value is unknown during table construction and is filled in during emission. To support nesting try-blocks (although we currently don't emit tables with nested try-blocks), GIM_Reject has been re-introduced to explicitly exit a try-block or fail the overall match if there are no active try-blocks. Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Reviewed By: rovka Subscribers: kristof.beyls, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D35117 llvm-svn: 308596
2017-07-20 17:25:44 +08:00
Table << MatchTable::Opcode("GIM_CheckRegBankForClass")
<< MatchTable::Comment("MI") << MatchTable::IntValue(InsnVarID)
<< MatchTable::Comment("Op") << MatchTable::IntValue(OpIdx)
<< MatchTable::Comment("RC")
<< MatchTable::NamedValue(RC.getQualifiedName() + "RegClassID")
<< MatchTable::LineBreak;
[GlobalISel] Add basic Selector-emitter tblgen backend. This adds a basic tablegen backend that analyzes the SelectionDAG patterns to find simple ones that are eligible for GlobalISel-emission. That's similar to FastISel, with one notable difference: we're not fed ISD opcodes, so we need to map the SDNode operators to generic opcodes. That's done using GINodeEquiv in TargetGlobalISel.td. Otherwise, this is mostly boilerplate, and lots of filtering of any kind of "complicated" pattern. On AArch64, this is sufficient to match G_ADD up to s64 (to ADDWrr/ADDXrr) and G_BR (to B). Differential Revision: https://reviews.llvm.org/D26878 llvm-svn: 290284
2016-12-22 07:26:20 +08:00
}
};
[globalisel] Re-factor ISel matchers into a hierarchy. NFC Summary: This should make it possible to easily add everything needed to import all the existing SelectionDAG rules. It should also serve the likely kinds of GlobalISel rules (some of which are not currently representable in SelectionDAG) once we've nailed down the tablegen definition for that. The hierarchy is as follows: MatcherRule - A matching rule. Currently used to emit C++ ISel code but will | also be used to emit test cases and tablegen definitions in the | near future. |- Instruction(s) - Represents the instruction to be matched. |- Instruction Predicate(s) - Test the opcode, arithmetic flags, etc. of an | instruction. \- Operand(s) - Represents a particular operand of the instruction. In the | future, there may be subclasses to test the same predicates | on multiple operands (including for variadic instructions). \ Operand Predicate(s) - Test the type, register bank, etc. of an operand. This is where the ComplexPattern equivalent will be represented. It's also nested-instruction matching will live as a predicate that follows the DefUse chain to the Def and tests a MatcherRule from that position. Support for multiple instruction matchers in a rule has been retained from the existing code but has been adjusted to assert when it is used. Previously it would silently drop all but the first instruction matcher. The tablegen-erated file is not functionally changed but has more parentheses and no longer attempts to format the if-statements since keeping track of the indentation is tricky in the presence of the matcher hierarchy. It would be nice to have CMakes tablegen() run the output through clang-format (when available) so we don't have to complicate TableGen with pretty-printing. It's also worth mentioning that this hierarchy will also be able to emit TableGen definitions and test cases in the near future. This is the reason for favouring explicit emit*() calls rather than the << operator. Reviewers: aditya_nandakumar, rovka, t.p.northover, qcolombet, ab Reviewed By: ab Subscribers: igorb, dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D28942 llvm-svn: 293172
2017-01-26 19:10:14 +08:00
/// Generates code to check that an operand is a basic block.
class MBBOperandMatcher : public OperandPredicateMatcher {
public:
[TableGen][GlobalISel] Have the predicate directly know which data they are dealing with Prior to this patch, a predicate wouldn't make sense outside of its rule. Indeed, it was only during emitting a rule that a predicate would be made aware of the IDs of the data it is checking. Because of that, predicates could not be moved around or compared between each other. NFC. llvm-svn: 320887
2017-12-16 07:07:42 +08:00
MBBOperandMatcher(unsigned InsnVarID, unsigned OpIdx)
: OperandPredicateMatcher(OPM_MBB, InsnVarID, OpIdx) {}
[globalisel] Sort RuleMatchers by priority. Summary: This makes more important rules have priority over less important rules. For example, '%a = G_ADD $b:s64, $c:s64' has priority over '%a = G_ADD $b:s32, $c:s32'. Previously these rules were emitted in the correct order by chance. NFC in this patch but it is required to make the next patch work correctly. Depends on D29710 Reviewers: t.p.northover, ab, qcolombet, aditya_nandakumar, rovka Reviewed By: ab, rovka Subscribers: javed.absar, dberris, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D29711 llvm-svn: 296121
2017-02-24 21:58:11 +08:00
[TableGen][GlobalISel] Add a common class for all PredicateMatcher NFC. llvm-svn: 320767
2017-12-15 07:44:07 +08:00
static bool classof(const PredicateMatcher *P) {
[globalisel] Sort RuleMatchers by priority. Summary: This makes more important rules have priority over less important rules. For example, '%a = G_ADD $b:s64, $c:s64' has priority over '%a = G_ADD $b:s32, $c:s32'. Previously these rules were emitted in the correct order by chance. NFC in this patch but it is required to make the next patch work correctly. Depends on D29710 Reviewers: t.p.northover, ab, qcolombet, aditya_nandakumar, rovka Reviewed By: ab, rovka Subscribers: javed.absar, dberris, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D29711 llvm-svn: 296121
2017-02-24 21:58:11 +08:00
return P->getKind() == OPM_MBB;
}
[TableGen][GlobalISel] Have the predicate directly know which data they are dealing with Prior to this patch, a predicate wouldn't make sense outside of its rule. Indeed, it was only during emitting a rule that a predicate would be made aware of the IDs of the data it is checking. Because of that, predicates could not be moved around or compared between each other. NFC. llvm-svn: 320887
2017-12-16 07:07:42 +08:00
void emitPredicateOpcodes(MatchTable &Table,
RuleMatcher &Rule) const override {
[globalisel][tablegen] Add control-flow to the MatchTable. Summary: This will allow us to merge the various sub-tables into a single table. This is a compile-time saving at this point. However, this will also enable the optimization of a table so that similar instructions can be tested together, reducing the time spent on the matching the code. The bulk of this patch is a mechanical conversion to the new MatchTable object which is responsible for tracking label definitions and filling in the index of the jump targets. It is also responsible for nicely formatting the table. This was necessary to support the new GIM_Try opcode which takes the index to jump to if the match should fail. This value is unknown during table construction and is filled in during emission. To support nesting try-blocks (although we currently don't emit tables with nested try-blocks), GIM_Reject has been re-introduced to explicitly exit a try-block or fail the overall match if there are no active try-blocks. Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Reviewed By: rovka Subscribers: kristof.beyls, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D35117 llvm-svn: 308596
2017-07-20 17:25:44 +08:00
Table << MatchTable::Opcode("GIM_CheckIsMBB") << MatchTable::Comment("MI")
<< MatchTable::IntValue(InsnVarID) << MatchTable::Comment("Op")
<< MatchTable::IntValue(OpIdx) << MatchTable::LineBreak;
[globalisel] Re-factor ISel matchers into a hierarchy. NFC Summary: This should make it possible to easily add everything needed to import all the existing SelectionDAG rules. It should also serve the likely kinds of GlobalISel rules (some of which are not currently representable in SelectionDAG) once we've nailed down the tablegen definition for that. The hierarchy is as follows: MatcherRule - A matching rule. Currently used to emit C++ ISel code but will | also be used to emit test cases and tablegen definitions in the | near future. |- Instruction(s) - Represents the instruction to be matched. |- Instruction Predicate(s) - Test the opcode, arithmetic flags, etc. of an | instruction. \- Operand(s) - Represents a particular operand of the instruction. In the | future, there may be subclasses to test the same predicates | on multiple operands (including for variadic instructions). \ Operand Predicate(s) - Test the type, register bank, etc. of an operand. This is where the ComplexPattern equivalent will be represented. It's also nested-instruction matching will live as a predicate that follows the DefUse chain to the Def and tests a MatcherRule from that position. Support for multiple instruction matchers in a rule has been retained from the existing code but has been adjusted to assert when it is used. Previously it would silently drop all but the first instruction matcher. The tablegen-erated file is not functionally changed but has more parentheses and no longer attempts to format the if-statements since keeping track of the indentation is tricky in the presence of the matcher hierarchy. It would be nice to have CMakes tablegen() run the output through clang-format (when available) so we don't have to complicate TableGen with pretty-printing. It's also worth mentioning that this hierarchy will also be able to emit TableGen definitions and test cases in the near future. This is the reason for favouring explicit emit*() calls rather than the << operator. Reviewers: aditya_nandakumar, rovka, t.p.northover, qcolombet, ab Reviewed By: ab Subscribers: igorb, dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D28942 llvm-svn: 293172
2017-01-26 19:10:14 +08:00
}
};
[globalisel][tablegen] Add support for (set $dst, 1) and test X86's OptForSize predicate. Summary: It's rare but a small number of patterns use IntInit's at the root of the match. On X86, one such rule is enabled by the OptForSize predicate and causes the compiler to use the smaller: %0 = MOV32r1 instead of the usual: %0 = MOV32ri 1 This patch adds support for matching IntInit's at the root and uses this as a test case for the optsize attribute that was implemented in r301750 Reviewers: qcolombet, ab, t.p.northover, rovka, kristof.beyls, aditya_nandakumar Reviewed By: qcolombet Subscribers: igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D32791 llvm-svn: 303678
2017-05-24 03:33:16 +08:00
/// Generates code to check that an operand is a G_CONSTANT with a particular
/// int.
class ConstantIntOperandMatcher : public OperandPredicateMatcher {
[globalisel] Decouple src pattern operands from dst pattern operands. Summary: This isn't testable for AArch64 by itself so this patch also adds support for constant immediates in the pattern and physical register uses in the result. The new IntOperandMatcher matches the constant in patterns such as '(set $rd:GPR32, (G_XOR $rs:GPR32, -1))'. It's always safe to fold immediates into an instruction so this is the first rule that will match across multiple BB's. The Renderer hierarchy is responsible for adding operands to the result instruction. Renderers can copy operands (CopyRenderer) or add physical registers (in particular %wzr and %xzr) to the result instruction in any order (OperandMatchers now import the operand names from SelectionDAG to allow renderers to access any operand). This allows us to emit the result instruction for: %1 = G_XOR %0, -1 --> %1 = ORNWrr %wzr, %0 %1 = G_XOR -1, %0 --> %1 = ORNWrr %wzr, %0 although the latter is untested since the matcher/importer has not been taught about commutativity yet. Added BuildMIAction which can build new instructions and mutate them where possible. W.r.t the mutation aspect, MatchActions are now told the name of an instruction they can recycle and BuildMIAction will emit mutation code when the renderers are appropriate. They are appropriate when all operands are rendered using CopyRenderer and the indices are the same as the matcher. This currently assumes that all operands have at least one matcher. Finally, this change also fixes a crash in AArch64InstructionSelector::select() caused by an immediate operand passing isImm() rather than isCImm(). This was uncovered by the other changes and was detected by existing tests. Depends on D29711 Reviewers: t.p.northover, ab, qcolombet, rovka, aditya_nandakumar, javed.absar Reviewed By: rovka Subscribers: aemerson, dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D29712 llvm-svn: 296131
2017-02-24 23:43:30 +08:00
protected:
int64_t Value;
public:
[TableGen][GlobalISel] Make the arguments of the Instruction and Operand Matchers consistent Move InsnVarID and OpIdx at the beginning of the list of arguments for all the constructors of the OperandMatcher subclasses. This matches what we do for the InstructionMatcher. NFC. llvm-svn: 321031
2017-12-19 06:12:13 +08:00
ConstantIntOperandMatcher(unsigned InsnVarID, unsigned OpIdx, int64_t Value)
[TableGen][GlobalISel] Have the predicate directly know which data they are dealing with Prior to this patch, a predicate wouldn't make sense outside of its rule. Indeed, it was only during emitting a rule that a predicate would be made aware of the IDs of the data it is checking. Because of that, predicates could not be moved around or compared between each other. NFC. llvm-svn: 320887
2017-12-16 07:07:42 +08:00
: OperandPredicateMatcher(OPM_Int, InsnVarID, OpIdx), Value(Value) {}
[globalisel] Decouple src pattern operands from dst pattern operands. Summary: This isn't testable for AArch64 by itself so this patch also adds support for constant immediates in the pattern and physical register uses in the result. The new IntOperandMatcher matches the constant in patterns such as '(set $rd:GPR32, (G_XOR $rs:GPR32, -1))'. It's always safe to fold immediates into an instruction so this is the first rule that will match across multiple BB's. The Renderer hierarchy is responsible for adding operands to the result instruction. Renderers can copy operands (CopyRenderer) or add physical registers (in particular %wzr and %xzr) to the result instruction in any order (OperandMatchers now import the operand names from SelectionDAG to allow renderers to access any operand). This allows us to emit the result instruction for: %1 = G_XOR %0, -1 --> %1 = ORNWrr %wzr, %0 %1 = G_XOR -1, %0 --> %1 = ORNWrr %wzr, %0 although the latter is untested since the matcher/importer has not been taught about commutativity yet. Added BuildMIAction which can build new instructions and mutate them where possible. W.r.t the mutation aspect, MatchActions are now told the name of an instruction they can recycle and BuildMIAction will emit mutation code when the renderers are appropriate. They are appropriate when all operands are rendered using CopyRenderer and the indices are the same as the matcher. This currently assumes that all operands have at least one matcher. Finally, this change also fixes a crash in AArch64InstructionSelector::select() caused by an immediate operand passing isImm() rather than isCImm(). This was uncovered by the other changes and was detected by existing tests. Depends on D29711 Reviewers: t.p.northover, ab, qcolombet, rovka, aditya_nandakumar, javed.absar Reviewed By: rovka Subscribers: aemerson, dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D29712 llvm-svn: 296131
2017-02-24 23:43:30 +08:00
[TableGen][GlobalISel] Make the different Matcher comparable This opens refactoring opportunities in the match table now that we can check that two predicates are the same. NFC. llvm-svn: 320890
2017-12-16 07:24:39 +08:00
bool isIdentical(const PredicateMatcher &B) const override {
return OperandPredicateMatcher::isIdentical(B) &&
Value == cast<ConstantIntOperandMatcher>(&B)->Value;
}
[TableGen][GlobalISel] Add a common class for all PredicateMatcher NFC. llvm-svn: 320767
2017-12-15 07:44:07 +08:00
static bool classof(const PredicateMatcher *P) {
[globalisel] Decouple src pattern operands from dst pattern operands. Summary: This isn't testable for AArch64 by itself so this patch also adds support for constant immediates in the pattern and physical register uses in the result. The new IntOperandMatcher matches the constant in patterns such as '(set $rd:GPR32, (G_XOR $rs:GPR32, -1))'. It's always safe to fold immediates into an instruction so this is the first rule that will match across multiple BB's. The Renderer hierarchy is responsible for adding operands to the result instruction. Renderers can copy operands (CopyRenderer) or add physical registers (in particular %wzr and %xzr) to the result instruction in any order (OperandMatchers now import the operand names from SelectionDAG to allow renderers to access any operand). This allows us to emit the result instruction for: %1 = G_XOR %0, -1 --> %1 = ORNWrr %wzr, %0 %1 = G_XOR -1, %0 --> %1 = ORNWrr %wzr, %0 although the latter is untested since the matcher/importer has not been taught about commutativity yet. Added BuildMIAction which can build new instructions and mutate them where possible. W.r.t the mutation aspect, MatchActions are now told the name of an instruction they can recycle and BuildMIAction will emit mutation code when the renderers are appropriate. They are appropriate when all operands are rendered using CopyRenderer and the indices are the same as the matcher. This currently assumes that all operands have at least one matcher. Finally, this change also fixes a crash in AArch64InstructionSelector::select() caused by an immediate operand passing isImm() rather than isCImm(). This was uncovered by the other changes and was detected by existing tests. Depends on D29711 Reviewers: t.p.northover, ab, qcolombet, rovka, aditya_nandakumar, javed.absar Reviewed By: rovka Subscribers: aemerson, dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D29712 llvm-svn: 296131
2017-02-24 23:43:30 +08:00
return P->getKind() == OPM_Int;
}
[TableGen][GlobalISel] Have the predicate directly know which data they are dealing with Prior to this patch, a predicate wouldn't make sense outside of its rule. Indeed, it was only during emitting a rule that a predicate would be made aware of the IDs of the data it is checking. Because of that, predicates could not be moved around or compared between each other. NFC. llvm-svn: 320887
2017-12-16 07:07:42 +08:00
void emitPredicateOpcodes(MatchTable &Table,
RuleMatcher &Rule) const override {
[globalisel][tablegen] Add control-flow to the MatchTable. Summary: This will allow us to merge the various sub-tables into a single table. This is a compile-time saving at this point. However, this will also enable the optimization of a table so that similar instructions can be tested together, reducing the time spent on the matching the code. The bulk of this patch is a mechanical conversion to the new MatchTable object which is responsible for tracking label definitions and filling in the index of the jump targets. It is also responsible for nicely formatting the table. This was necessary to support the new GIM_Try opcode which takes the index to jump to if the match should fail. This value is unknown during table construction and is filled in during emission. To support nesting try-blocks (although we currently don't emit tables with nested try-blocks), GIM_Reject has been re-introduced to explicitly exit a try-block or fail the overall match if there are no active try-blocks. Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Reviewed By: rovka Subscribers: kristof.beyls, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D35117 llvm-svn: 308596
2017-07-20 17:25:44 +08:00
Table << MatchTable::Opcode("GIM_CheckConstantInt")
<< MatchTable::Comment("MI") << MatchTable::IntValue(InsnVarID)
<< MatchTable::Comment("Op") << MatchTable::IntValue(OpIdx)
<< MatchTable::IntValue(Value) << MatchTable::LineBreak;
[globalisel] Decouple src pattern operands from dst pattern operands. Summary: This isn't testable for AArch64 by itself so this patch also adds support for constant immediates in the pattern and physical register uses in the result. The new IntOperandMatcher matches the constant in patterns such as '(set $rd:GPR32, (G_XOR $rs:GPR32, -1))'. It's always safe to fold immediates into an instruction so this is the first rule that will match across multiple BB's. The Renderer hierarchy is responsible for adding operands to the result instruction. Renderers can copy operands (CopyRenderer) or add physical registers (in particular %wzr and %xzr) to the result instruction in any order (OperandMatchers now import the operand names from SelectionDAG to allow renderers to access any operand). This allows us to emit the result instruction for: %1 = G_XOR %0, -1 --> %1 = ORNWrr %wzr, %0 %1 = G_XOR -1, %0 --> %1 = ORNWrr %wzr, %0 although the latter is untested since the matcher/importer has not been taught about commutativity yet. Added BuildMIAction which can build new instructions and mutate them where possible. W.r.t the mutation aspect, MatchActions are now told the name of an instruction they can recycle and BuildMIAction will emit mutation code when the renderers are appropriate. They are appropriate when all operands are rendered using CopyRenderer and the indices are the same as the matcher. This currently assumes that all operands have at least one matcher. Finally, this change also fixes a crash in AArch64InstructionSelector::select() caused by an immediate operand passing isImm() rather than isCImm(). This was uncovered by the other changes and was detected by existing tests. Depends on D29711 Reviewers: t.p.northover, ab, qcolombet, rovka, aditya_nandakumar, javed.absar Reviewed By: rovka Subscribers: aemerson, dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D29712 llvm-svn: 296131
2017-02-24 23:43:30 +08:00
}
};
[globalisel][tablegen] Add support for (set $dst, 1) and test X86's OptForSize predicate. Summary: It's rare but a small number of patterns use IntInit's at the root of the match. On X86, one such rule is enabled by the OptForSize predicate and causes the compiler to use the smaller: %0 = MOV32r1 instead of the usual: %0 = MOV32ri 1 This patch adds support for matching IntInit's at the root and uses this as a test case for the optsize attribute that was implemented in r301750 Reviewers: qcolombet, ab, t.p.northover, rovka, kristof.beyls, aditya_nandakumar Reviewed By: qcolombet Subscribers: igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D32791 llvm-svn: 303678
2017-05-24 03:33:16 +08:00
/// Generates code to check that an operand is a raw int (where MO.isImm() or
/// MO.isCImm() is true).
class LiteralIntOperandMatcher : public OperandPredicateMatcher {
protected:
int64_t Value;
public:
[TableGen][GlobalISel] Make the arguments of the Instruction and Operand Matchers consistent Move InsnVarID and OpIdx at the beginning of the list of arguments for all the constructors of the OperandMatcher subclasses. This matches what we do for the InstructionMatcher. NFC. llvm-svn: 321031
2017-12-19 06:12:13 +08:00
LiteralIntOperandMatcher(unsigned InsnVarID, unsigned OpIdx, int64_t Value)
[TableGen][GlobalISel] Have the predicate directly know which data they are dealing with Prior to this patch, a predicate wouldn't make sense outside of its rule. Indeed, it was only during emitting a rule that a predicate would be made aware of the IDs of the data it is checking. Because of that, predicates could not be moved around or compared between each other. NFC. llvm-svn: 320887
2017-12-16 07:07:42 +08:00
: OperandPredicateMatcher(OPM_LiteralInt, InsnVarID, OpIdx),
Value(Value) {}
[globalisel][tablegen] Add support for (set $dst, 1) and test X86's OptForSize predicate. Summary: It's rare but a small number of patterns use IntInit's at the root of the match. On X86, one such rule is enabled by the OptForSize predicate and causes the compiler to use the smaller: %0 = MOV32r1 instead of the usual: %0 = MOV32ri 1 This patch adds support for matching IntInit's at the root and uses this as a test case for the optsize attribute that was implemented in r301750 Reviewers: qcolombet, ab, t.p.northover, rovka, kristof.beyls, aditya_nandakumar Reviewed By: qcolombet Subscribers: igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D32791 llvm-svn: 303678
2017-05-24 03:33:16 +08:00
[TableGen][GlobalISel] Make the different Matcher comparable This opens refactoring opportunities in the match table now that we can check that two predicates are the same. NFC. llvm-svn: 320890
2017-12-16 07:24:39 +08:00
bool isIdentical(const PredicateMatcher &B) const override {
return OperandPredicateMatcher::isIdentical(B) &&
Value == cast<LiteralIntOperandMatcher>(&B)->Value;
}
[TableGen][GlobalISel] Add a common class for all PredicateMatcher NFC. llvm-svn: 320767
2017-12-15 07:44:07 +08:00
static bool classof(const PredicateMatcher *P) {
[globalisel][tablegen] Add support for (set $dst, 1) and test X86's OptForSize predicate. Summary: It's rare but a small number of patterns use IntInit's at the root of the match. On X86, one such rule is enabled by the OptForSize predicate and causes the compiler to use the smaller: %0 = MOV32r1 instead of the usual: %0 = MOV32ri 1 This patch adds support for matching IntInit's at the root and uses this as a test case for the optsize attribute that was implemented in r301750 Reviewers: qcolombet, ab, t.p.northover, rovka, kristof.beyls, aditya_nandakumar Reviewed By: qcolombet Subscribers: igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D32791 llvm-svn: 303678
2017-05-24 03:33:16 +08:00
return P->getKind() == OPM_LiteralInt;
}
[TableGen][GlobalISel] Have the predicate directly know which data they are dealing with Prior to this patch, a predicate wouldn't make sense outside of its rule. Indeed, it was only during emitting a rule that a predicate would be made aware of the IDs of the data it is checking. Because of that, predicates could not be moved around or compared between each other. NFC. llvm-svn: 320887
2017-12-16 07:07:42 +08:00
void emitPredicateOpcodes(MatchTable &Table,
RuleMatcher &Rule) const override {
[globalisel][tablegen] Add control-flow to the MatchTable. Summary: This will allow us to merge the various sub-tables into a single table. This is a compile-time saving at this point. However, this will also enable the optimization of a table so that similar instructions can be tested together, reducing the time spent on the matching the code. The bulk of this patch is a mechanical conversion to the new MatchTable object which is responsible for tracking label definitions and filling in the index of the jump targets. It is also responsible for nicely formatting the table. This was necessary to support the new GIM_Try opcode which takes the index to jump to if the match should fail. This value is unknown during table construction and is filled in during emission. To support nesting try-blocks (although we currently don't emit tables with nested try-blocks), GIM_Reject has been re-introduced to explicitly exit a try-block or fail the overall match if there are no active try-blocks. Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Reviewed By: rovka Subscribers: kristof.beyls, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D35117 llvm-svn: 308596
2017-07-20 17:25:44 +08:00
Table << MatchTable::Opcode("GIM_CheckLiteralInt")
<< MatchTable::Comment("MI") << MatchTable::IntValue(InsnVarID)
<< MatchTable::Comment("Op") << MatchTable::IntValue(OpIdx)
<< MatchTable::IntValue(Value) << MatchTable::LineBreak;
[globalisel][tablegen] Add support for (set $dst, 1) and test X86's OptForSize predicate. Summary: It's rare but a small number of patterns use IntInit's at the root of the match. On X86, one such rule is enabled by the OptForSize predicate and causes the compiler to use the smaller: %0 = MOV32r1 instead of the usual: %0 = MOV32ri 1 This patch adds support for matching IntInit's at the root and uses this as a test case for the optsize attribute that was implemented in r301750 Reviewers: qcolombet, ab, t.p.northover, rovka, kristof.beyls, aditya_nandakumar Reviewed By: qcolombet Subscribers: igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D32791 llvm-svn: 303678
2017-05-24 03:33:16 +08:00
}
};
[globalisel][tablegen] Correct matching of intrinsic ID's. TreePatternNode considers them to be plain integers but MachineInstr considers them to be a distinct kind of operand. The tweak to AArch64InstrInfo.td to produce a simple test case is a NFC for everything except GlobalISelEmitter (confirmed by diffing the tablegenerated files). GlobalISelEmitter is currently unable to infer the type of operands in the Dst pattern from the operands in the Src pattern. llvm-svn: 307634
2017-07-11 16:57:29 +08:00
/// Generates code to check that an operand is an intrinsic ID.
class IntrinsicIDOperandMatcher : public OperandPredicateMatcher {
protected:
const CodeGenIntrinsic *II;
public:
[TableGen][GlobalISel] Make the arguments of the Instruction and Operand Matchers consistent Move InsnVarID and OpIdx at the beginning of the list of arguments for all the constructors of the OperandMatcher subclasses. This matches what we do for the InstructionMatcher. NFC. llvm-svn: 321031
2017-12-19 06:12:13 +08:00
IntrinsicIDOperandMatcher(unsigned InsnVarID, unsigned OpIdx,
const CodeGenIntrinsic *II)
[TableGen][GlobalISel] Have the predicate directly know which data they are dealing with Prior to this patch, a predicate wouldn't make sense outside of its rule. Indeed, it was only during emitting a rule that a predicate would be made aware of the IDs of the data it is checking. Because of that, predicates could not be moved around or compared between each other. NFC. llvm-svn: 320887
2017-12-16 07:07:42 +08:00
: OperandPredicateMatcher(OPM_IntrinsicID, InsnVarID, OpIdx), II(II) {}
[globalisel][tablegen] Correct matching of intrinsic ID's. TreePatternNode considers them to be plain integers but MachineInstr considers them to be a distinct kind of operand. The tweak to AArch64InstrInfo.td to produce a simple test case is a NFC for everything except GlobalISelEmitter (confirmed by diffing the tablegenerated files). GlobalISelEmitter is currently unable to infer the type of operands in the Dst pattern from the operands in the Src pattern. llvm-svn: 307634
2017-07-11 16:57:29 +08:00
[TableGen][GlobalISel] Make the different Matcher comparable This opens refactoring opportunities in the match table now that we can check that two predicates are the same. NFC. llvm-svn: 320890
2017-12-16 07:24:39 +08:00
bool isIdentical(const PredicateMatcher &B) const override {
return OperandPredicateMatcher::isIdentical(B) &&
II == cast<IntrinsicIDOperandMatcher>(&B)->II;
}
[TableGen][GlobalISel] Add a common class for all PredicateMatcher NFC. llvm-svn: 320767
2017-12-15 07:44:07 +08:00
static bool classof(const PredicateMatcher *P) {
[globalisel][tablegen] Correct matching of intrinsic ID's. TreePatternNode considers them to be plain integers but MachineInstr considers them to be a distinct kind of operand. The tweak to AArch64InstrInfo.td to produce a simple test case is a NFC for everything except GlobalISelEmitter (confirmed by diffing the tablegenerated files). GlobalISelEmitter is currently unable to infer the type of operands in the Dst pattern from the operands in the Src pattern. llvm-svn: 307634
2017-07-11 16:57:29 +08:00
return P->getKind() == OPM_IntrinsicID;
}
[TableGen][GlobalISel] Have the predicate directly know which data they are dealing with Prior to this patch, a predicate wouldn't make sense outside of its rule. Indeed, it was only during emitting a rule that a predicate would be made aware of the IDs of the data it is checking. Because of that, predicates could not be moved around or compared between each other. NFC. llvm-svn: 320887
2017-12-16 07:07:42 +08:00
void emitPredicateOpcodes(MatchTable &Table,
RuleMatcher &Rule) const override {
[globalisel][tablegen] Add control-flow to the MatchTable. Summary: This will allow us to merge the various sub-tables into a single table. This is a compile-time saving at this point. However, this will also enable the optimization of a table so that similar instructions can be tested together, reducing the time spent on the matching the code. The bulk of this patch is a mechanical conversion to the new MatchTable object which is responsible for tracking label definitions and filling in the index of the jump targets. It is also responsible for nicely formatting the table. This was necessary to support the new GIM_Try opcode which takes the index to jump to if the match should fail. This value is unknown during table construction and is filled in during emission. To support nesting try-blocks (although we currently don't emit tables with nested try-blocks), GIM_Reject has been re-introduced to explicitly exit a try-block or fail the overall match if there are no active try-blocks. Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Reviewed By: rovka Subscribers: kristof.beyls, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D35117 llvm-svn: 308596
2017-07-20 17:25:44 +08:00
Table << MatchTable::Opcode("GIM_CheckIntrinsicID")
<< MatchTable::Comment("MI") << MatchTable::IntValue(InsnVarID)
<< MatchTable::Comment("Op") << MatchTable::IntValue(OpIdx)
<< MatchTable::NamedValue("Intrinsic::" + II->EnumName)
<< MatchTable::LineBreak;
[globalisel][tablegen] Correct matching of intrinsic ID's. TreePatternNode considers them to be plain integers but MachineInstr considers them to be a distinct kind of operand. The tweak to AArch64InstrInfo.td to produce a simple test case is a NFC for everything except GlobalISelEmitter (confirmed by diffing the tablegenerated files). GlobalISelEmitter is currently unable to infer the type of operands in the Dst pattern from the operands in the Src pattern. llvm-svn: 307634
2017-07-11 16:57:29 +08:00
}
};
[globalisel] Re-factor ISel matchers into a hierarchy. NFC Summary: This should make it possible to easily add everything needed to import all the existing SelectionDAG rules. It should also serve the likely kinds of GlobalISel rules (some of which are not currently representable in SelectionDAG) once we've nailed down the tablegen definition for that. The hierarchy is as follows: MatcherRule - A matching rule. Currently used to emit C++ ISel code but will | also be used to emit test cases and tablegen definitions in the | near future. |- Instruction(s) - Represents the instruction to be matched. |- Instruction Predicate(s) - Test the opcode, arithmetic flags, etc. of an | instruction. \- Operand(s) - Represents a particular operand of the instruction. In the | future, there may be subclasses to test the same predicates | on multiple operands (including for variadic instructions). \ Operand Predicate(s) - Test the type, register bank, etc. of an operand. This is where the ComplexPattern equivalent will be represented. It's also nested-instruction matching will live as a predicate that follows the DefUse chain to the Def and tests a MatcherRule from that position. Support for multiple instruction matchers in a rule has been retained from the existing code but has been adjusted to assert when it is used. Previously it would silently drop all but the first instruction matcher. The tablegen-erated file is not functionally changed but has more parentheses and no longer attempts to format the if-statements since keeping track of the indentation is tricky in the presence of the matcher hierarchy. It would be nice to have CMakes tablegen() run the output through clang-format (when available) so we don't have to complicate TableGen with pretty-printing. It's also worth mentioning that this hierarchy will also be able to emit TableGen definitions and test cases in the near future. This is the reason for favouring explicit emit*() calls rather than the << operator. Reviewers: aditya_nandakumar, rovka, t.p.northover, qcolombet, ab Reviewed By: ab Subscribers: igorb, dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D28942 llvm-svn: 293172
2017-01-26 19:10:14 +08:00
/// Generates code to check that a set of predicates match for a particular
/// operand.
class OperandMatcher : public PredicateListMatcher<OperandPredicateMatcher> {
protected:
[tablegen][globalisel] Capture instructions into locals and related infrastructure for multiple instructions matches. Summary: Prepare the way for nested instruction matching support by having actions like CopyRenderer look up operands in the RuleMatcher rather than a specific InstructionMatcher. This allows actions to reference any operand from any matched instruction. It works by checking the 'shape' of the match and capturing each matched instruction to a local variable. If the shape is wrong (not enough operands, leaf nodes where non-leafs are expected, etc.), then the rule exits early without checking the predicates. Once we've captured the instructions, we then test the predicates as before (except using the local variables). If the match is successful, then we render the new instruction as before using the local variables. It's not noticable in this patch but by the time we support multiple instruction matching, this patch will also cause a significant improvement to readability of the emitted code since MRI.getVRegDef(I->getOperand(0).getReg()) will simply be MI1 after emitCxxCaptureStmts(). This isn't quite NFC because I've also fixed a bug that I'm surprised we haven't encountered yet. It now checks there are at least the expected number of operands before accessing them with getOperand(). Depends on D30531 Reviewers: t.p.northover, qcolombet, aditya_nandakumar, ab, rovka Reviewed By: rovka Subscribers: dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D30535 llvm-svn: 298257
2017-03-20 23:20:42 +08:00
InstructionMatcher &Insn;
[GlobalISel] Add basic Selector-emitter tblgen backend. This adds a basic tablegen backend that analyzes the SelectionDAG patterns to find simple ones that are eligible for GlobalISel-emission. That's similar to FastISel, with one notable difference: we're not fed ISD opcodes, so we need to map the SDNode operators to generic opcodes. That's done using GINodeEquiv in TargetGlobalISel.td. Otherwise, this is mostly boilerplate, and lots of filtering of any kind of "complicated" pattern. On AArch64, this is sufficient to match G_ADD up to s64 (to ADDWrr/ADDXrr) and G_BR (to B). Differential Revision: https://reviews.llvm.org/D26878 llvm-svn: 290284
2016-12-22 07:26:20 +08:00
unsigned OpIdx;
[globalisel] Decouple src pattern operands from dst pattern operands. Summary: This isn't testable for AArch64 by itself so this patch also adds support for constant immediates in the pattern and physical register uses in the result. The new IntOperandMatcher matches the constant in patterns such as '(set $rd:GPR32, (G_XOR $rs:GPR32, -1))'. It's always safe to fold immediates into an instruction so this is the first rule that will match across multiple BB's. The Renderer hierarchy is responsible for adding operands to the result instruction. Renderers can copy operands (CopyRenderer) or add physical registers (in particular %wzr and %xzr) to the result instruction in any order (OperandMatchers now import the operand names from SelectionDAG to allow renderers to access any operand). This allows us to emit the result instruction for: %1 = G_XOR %0, -1 --> %1 = ORNWrr %wzr, %0 %1 = G_XOR -1, %0 --> %1 = ORNWrr %wzr, %0 although the latter is untested since the matcher/importer has not been taught about commutativity yet. Added BuildMIAction which can build new instructions and mutate them where possible. W.r.t the mutation aspect, MatchActions are now told the name of an instruction they can recycle and BuildMIAction will emit mutation code when the renderers are appropriate. They are appropriate when all operands are rendered using CopyRenderer and the indices are the same as the matcher. This currently assumes that all operands have at least one matcher. Finally, this change also fixes a crash in AArch64InstructionSelector::select() caused by an immediate operand passing isImm() rather than isCImm(). This was uncovered by the other changes and was detected by existing tests. Depends on D29711 Reviewers: t.p.northover, ab, qcolombet, rovka, aditya_nandakumar, javed.absar Reviewed By: rovka Subscribers: aemerson, dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D29712 llvm-svn: 296131
2017-02-24 23:43:30 +08:00
std::string SymbolicName;
[GlobalISel] Add basic Selector-emitter tblgen backend. This adds a basic tablegen backend that analyzes the SelectionDAG patterns to find simple ones that are eligible for GlobalISel-emission. That's similar to FastISel, with one notable difference: we're not fed ISD opcodes, so we need to map the SDNode operators to generic opcodes. That's done using GINodeEquiv in TargetGlobalISel.td. Otherwise, this is mostly boilerplate, and lots of filtering of any kind of "complicated" pattern. On AArch64, this is sufficient to match G_ADD up to s64 (to ADDWrr/ADDXrr) and G_BR (to B). Differential Revision: https://reviews.llvm.org/D26878 llvm-svn: 290284
2016-12-22 07:26:20 +08:00
[globalisel][tablegen] Fix patterns involving multiple ComplexPatterns. Summary: Temporaries are now allocated to operands instead of predicates and this allocation is used to correctly pair up the rendered operands with the matched operands. Previously, ComplexPatterns were allocated temporaries independently in the Src Pattern and Dst Pattern, leading to mismatches. Additionally, the Dst Pattern failed to account for the allocated index and therefore always used temporary 0, 1, ... when it should have used base+0, base+1, ... Thanks to Aditya Nandakumar for noticing the bug. Depends on D30539 Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Reviewed By: rovka Subscribers: igorb, dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D31054 llvm-svn: 299538
2017-04-05 21:14:03 +08:00
/// The index of the first temporary variable allocated to this operand. The
/// number of allocated temporaries can be found with
[globalisel][tablegen] Revise API for ComplexPattern operands to improve flexibility. Summary: Some targets need to be able to do more complex rendering than just adding an operand or two to an instruction. For example, it may need to insert an instruction to extract a subreg first, or it may need to perform an operation on the operand. In SelectionDAG, targets would create SDNode's to achieve the desired effect during the complex pattern predicate. This worked because SelectionDAG had a form of garbage collection that would take care of SDNode's that were created but not used due to a later predicate rejecting a match. This doesn't translate well to GlobalISel and the churn was wasteful. The API changes in this patch enable GlobalISel to accomplish the same thing without the waste. The API is now: InstructionSelector::OptionalComplexRendererFn selectArithImmed(MachineOperand &Root) const; where Root is the root of the match. The return value can be omitted to indicate that the predicate failed to match, or a function with the signature ComplexRendererFn can be returned. For example: return OptionalComplexRendererFn( [=](MachineInstrBuilder &MIB) { MIB.addImm(Immed).addImm(ShVal); }); adds two immediate operands to the rendered instruction. Immed and ShVal are captured from the predicate function. As an added bonus, this also reduces the amount of information we need to provide to GIComplexOperandMatcher. Depends on D31418 Reviewers: aditya_nandakumar, t.p.northover, qcolombet, rovka, ab, javed.absar Reviewed By: ab Subscribers: dberris, kristof.beyls, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D31761 llvm-svn: 301079
2017-04-22 23:11:04 +08:00
/// countRendererFns().
[globalisel][tablegen] Fix patterns involving multiple ComplexPatterns. Summary: Temporaries are now allocated to operands instead of predicates and this allocation is used to correctly pair up the rendered operands with the matched operands. Previously, ComplexPatterns were allocated temporaries independently in the Src Pattern and Dst Pattern, leading to mismatches. Additionally, the Dst Pattern failed to account for the allocated index and therefore always used temporary 0, 1, ... when it should have used base+0, base+1, ... Thanks to Aditya Nandakumar for noticing the bug. Depends on D30539 Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Reviewed By: rovka Subscribers: igorb, dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D31054 llvm-svn: 299538
2017-04-05 21:14:03 +08:00
unsigned AllocatedTemporariesBaseID;
[globalisel] Re-factor ISel matchers into a hierarchy. NFC Summary: This should make it possible to easily add everything needed to import all the existing SelectionDAG rules. It should also serve the likely kinds of GlobalISel rules (some of which are not currently representable in SelectionDAG) once we've nailed down the tablegen definition for that. The hierarchy is as follows: MatcherRule - A matching rule. Currently used to emit C++ ISel code but will | also be used to emit test cases and tablegen definitions in the | near future. |- Instruction(s) - Represents the instruction to be matched. |- Instruction Predicate(s) - Test the opcode, arithmetic flags, etc. of an | instruction. \- Operand(s) - Represents a particular operand of the instruction. In the | future, there may be subclasses to test the same predicates | on multiple operands (including for variadic instructions). \ Operand Predicate(s) - Test the type, register bank, etc. of an operand. This is where the ComplexPattern equivalent will be represented. It's also nested-instruction matching will live as a predicate that follows the DefUse chain to the Def and tests a MatcherRule from that position. Support for multiple instruction matchers in a rule has been retained from the existing code but has been adjusted to assert when it is used. Previously it would silently drop all but the first instruction matcher. The tablegen-erated file is not functionally changed but has more parentheses and no longer attempts to format the if-statements since keeping track of the indentation is tricky in the presence of the matcher hierarchy. It would be nice to have CMakes tablegen() run the output through clang-format (when available) so we don't have to complicate TableGen with pretty-printing. It's also worth mentioning that this hierarchy will also be able to emit TableGen definitions and test cases in the near future. This is the reason for favouring explicit emit*() calls rather than the << operator. Reviewers: aditya_nandakumar, rovka, t.p.northover, qcolombet, ab Reviewed By: ab Subscribers: igorb, dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D28942 llvm-svn: 293172
2017-01-26 19:10:14 +08:00
public:
[tablegen][globalisel] Capture instructions into locals and related infrastructure for multiple instructions matches. Summary: Prepare the way for nested instruction matching support by having actions like CopyRenderer look up operands in the RuleMatcher rather than a specific InstructionMatcher. This allows actions to reference any operand from any matched instruction. It works by checking the 'shape' of the match and capturing each matched instruction to a local variable. If the shape is wrong (not enough operands, leaf nodes where non-leafs are expected, etc.), then the rule exits early without checking the predicates. Once we've captured the instructions, we then test the predicates as before (except using the local variables). If the match is successful, then we render the new instruction as before using the local variables. It's not noticable in this patch but by the time we support multiple instruction matching, this patch will also cause a significant improvement to readability of the emitted code since MRI.getVRegDef(I->getOperand(0).getReg()) will simply be MI1 after emitCxxCaptureStmts(). This isn't quite NFC because I've also fixed a bug that I'm surprised we haven't encountered yet. It now checks there are at least the expected number of operands before accessing them with getOperand(). Depends on D30531 Reviewers: t.p.northover, qcolombet, aditya_nandakumar, ab, rovka Reviewed By: rovka Subscribers: dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D30535 llvm-svn: 298257
2017-03-20 23:20:42 +08:00
OperandMatcher(InstructionMatcher &Insn, unsigned OpIdx,
[globalisel][tablegen] Fix patterns involving multiple ComplexPatterns. Summary: Temporaries are now allocated to operands instead of predicates and this allocation is used to correctly pair up the rendered operands with the matched operands. Previously, ComplexPatterns were allocated temporaries independently in the Src Pattern and Dst Pattern, leading to mismatches. Additionally, the Dst Pattern failed to account for the allocated index and therefore always used temporary 0, 1, ... when it should have used base+0, base+1, ... Thanks to Aditya Nandakumar for noticing the bug. Depends on D30539 Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Reviewed By: rovka Subscribers: igorb, dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D31054 llvm-svn: 299538
2017-04-05 21:14:03 +08:00
const std::string &SymbolicName,
unsigned AllocatedTemporariesBaseID)
: Insn(Insn), OpIdx(OpIdx), SymbolicName(SymbolicName),
AllocatedTemporariesBaseID(AllocatedTemporariesBaseID) {}
[globalisel] Decouple src pattern operands from dst pattern operands. Summary: This isn't testable for AArch64 by itself so this patch also adds support for constant immediates in the pattern and physical register uses in the result. The new IntOperandMatcher matches the constant in patterns such as '(set $rd:GPR32, (G_XOR $rs:GPR32, -1))'. It's always safe to fold immediates into an instruction so this is the first rule that will match across multiple BB's. The Renderer hierarchy is responsible for adding operands to the result instruction. Renderers can copy operands (CopyRenderer) or add physical registers (in particular %wzr and %xzr) to the result instruction in any order (OperandMatchers now import the operand names from SelectionDAG to allow renderers to access any operand). This allows us to emit the result instruction for: %1 = G_XOR %0, -1 --> %1 = ORNWrr %wzr, %0 %1 = G_XOR -1, %0 --> %1 = ORNWrr %wzr, %0 although the latter is untested since the matcher/importer has not been taught about commutativity yet. Added BuildMIAction which can build new instructions and mutate them where possible. W.r.t the mutation aspect, MatchActions are now told the name of an instruction they can recycle and BuildMIAction will emit mutation code when the renderers are appropriate. They are appropriate when all operands are rendered using CopyRenderer and the indices are the same as the matcher. This currently assumes that all operands have at least one matcher. Finally, this change also fixes a crash in AArch64InstructionSelector::select() caused by an immediate operand passing isImm() rather than isCImm(). This was uncovered by the other changes and was detected by existing tests. Depends on D29711 Reviewers: t.p.northover, ab, qcolombet, rovka, aditya_nandakumar, javed.absar Reviewed By: rovka Subscribers: aemerson, dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D29712 llvm-svn: 296131
2017-02-24 23:43:30 +08:00
bool hasSymbolicName() const { return !SymbolicName.empty(); }
const StringRef getSymbolicName() const { return SymbolicName; }
[tablegen][globalisel] Convert the SelectionDAG importer to a tree walking approach. NFC Summary: But don't actually inspect the tree any deeper than we already do. This change is NFC but the next one will enable full traversal of the source/destination patterns. Depends on D30535 Reviewers: t.p.northover, qcolombet, aditya_nandakumar, rovka, ab Subscribers: igorb, dberris, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D30536 llvm-svn: 299001
2017-03-29 23:37:18 +08:00
void setSymbolicName(StringRef Name) {
assert(SymbolicName.empty() && "Operand already has a symbolic name");
SymbolicName = Name;
}
Reapply r332907 "[GlobalISel] Improving InstructionSelect's performance by reducing MatchTable..." Apparently the compile time problem was caused by the fact that not all compilers / STL implementations can automatically convert std::unique_ptr<Derived> to std::unique_ptr<Base>. Fixed (hopefully) by making sure it's std::unique_ptr<Derived>&& (rvalue ref) to std::unique_ptr<Base> conversion instead. llvm-svn: 332917
2018-05-22 07:28:51 +08:00
/// Construct a new operand predicate and add it to the matcher.
template <class Kind, class... Args>
Optional<Kind *> addPredicate(Args &&... args) {
if (isSameAsAnotherOperand())
return None;
Predicates.emplace_back(llvm::make_unique<Kind>(
getInsnVarID(), getOpIdx(), std::forward<Args>(args)...));
return static_cast<Kind *>(Predicates.back().get());
}
unsigned getOpIdx() const { return OpIdx; }
[TableGen][GlobalISel] Have the predicate directly know which data they are dealing with Prior to this patch, a predicate wouldn't make sense outside of its rule. Indeed, it was only during emitting a rule that a predicate would be made aware of the IDs of the data it is checking. Because of that, predicates could not be moved around or compared between each other. NFC. llvm-svn: 320887
2017-12-16 07:07:42 +08:00
unsigned getInsnVarID() const;
[globalisel] Decouple src pattern operands from dst pattern operands. Summary: This isn't testable for AArch64 by itself so this patch also adds support for constant immediates in the pattern and physical register uses in the result. The new IntOperandMatcher matches the constant in patterns such as '(set $rd:GPR32, (G_XOR $rs:GPR32, -1))'. It's always safe to fold immediates into an instruction so this is the first rule that will match across multiple BB's. The Renderer hierarchy is responsible for adding operands to the result instruction. Renderers can copy operands (CopyRenderer) or add physical registers (in particular %wzr and %xzr) to the result instruction in any order (OperandMatchers now import the operand names from SelectionDAG to allow renderers to access any operand). This allows us to emit the result instruction for: %1 = G_XOR %0, -1 --> %1 = ORNWrr %wzr, %0 %1 = G_XOR -1, %0 --> %1 = ORNWrr %wzr, %0 although the latter is untested since the matcher/importer has not been taught about commutativity yet. Added BuildMIAction which can build new instructions and mutate them where possible. W.r.t the mutation aspect, MatchActions are now told the name of an instruction they can recycle and BuildMIAction will emit mutation code when the renderers are appropriate. They are appropriate when all operands are rendered using CopyRenderer and the indices are the same as the matcher. This currently assumes that all operands have at least one matcher. Finally, this change also fixes a crash in AArch64InstructionSelector::select() caused by an immediate operand passing isImm() rather than isCImm(). This was uncovered by the other changes and was detected by existing tests. Depends on D29711 Reviewers: t.p.northover, ab, qcolombet, rovka, aditya_nandakumar, javed.absar Reviewed By: rovka Subscribers: aemerson, dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D29712 llvm-svn: 296131
2017-02-24 23:43:30 +08:00
[globalisel][tablegen] Partially fix compile-time regressions by converting matcher to state-machine(s) Summary: Replace the matcher if-statements for each rule with a state-machine. This significantly reduces compile time, memory allocations, and cumulative memory allocation when compiling AArch64InstructionSelector.cpp.o after r303259 is recommitted. The following patches will expand on this further to fully fix the regressions. Reviewers: rovka, ab, t.p.northover, qcolombet, aditya_nandakumar Reviewed By: ab Subscribers: vitalybuka, aemerson, javed.absar, igorb, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D33758 llvm-svn: 307079
2017-07-04 22:35:06 +08:00
std::string getOperandExpr(unsigned InsnVarID) const {
return "State.MIs[" + llvm::to_string(InsnVarID) + "]->getOperand(" +
llvm::to_string(OpIdx) + ")";
[globalisel] OperandPredicateMatcher's shouldn't need to generate the MachineOperand expr. NFC Summary: Each OperandPredicateMatcher shouldn't need to know how to generate the expression to reference a MachineOperand. The OperandMatcher should provide it. In addition to separating responsibilities, this also lays some groundwork for decoupling source patterns from destination patterns to allow invented operands or operands provided by GlobalISel's equivalent to the ComplexPattern<> class. Depends on D29709 Reviewers: t.p.northover, ab, rovka, qcolombet, aditya_nandakumar Reviewed By: ab Subscribers: dberris, kristof.beyls, llvm-commits, igorb Differential Revision: https://reviews.llvm.org/D29710 llvm-svn: 295668
2017-02-20 23:30:43 +08:00
}
[globalisel] Re-factor ISel matchers into a hierarchy. NFC Summary: This should make it possible to easily add everything needed to import all the existing SelectionDAG rules. It should also serve the likely kinds of GlobalISel rules (some of which are not currently representable in SelectionDAG) once we've nailed down the tablegen definition for that. The hierarchy is as follows: MatcherRule - A matching rule. Currently used to emit C++ ISel code but will | also be used to emit test cases and tablegen definitions in the | near future. |- Instruction(s) - Represents the instruction to be matched. |- Instruction Predicate(s) - Test the opcode, arithmetic flags, etc. of an | instruction. \- Operand(s) - Represents a particular operand of the instruction. In the | future, there may be subclasses to test the same predicates | on multiple operands (including for variadic instructions). \ Operand Predicate(s) - Test the type, register bank, etc. of an operand. This is where the ComplexPattern equivalent will be represented. It's also nested-instruction matching will live as a predicate that follows the DefUse chain to the Def and tests a MatcherRule from that position. Support for multiple instruction matchers in a rule has been retained from the existing code but has been adjusted to assert when it is used. Previously it would silently drop all but the first instruction matcher. The tablegen-erated file is not functionally changed but has more parentheses and no longer attempts to format the if-statements since keeping track of the indentation is tricky in the presence of the matcher hierarchy. It would be nice to have CMakes tablegen() run the output through clang-format (when available) so we don't have to complicate TableGen with pretty-printing. It's also worth mentioning that this hierarchy will also be able to emit TableGen definitions and test cases in the near future. This is the reason for favouring explicit emit*() calls rather than the << operator. Reviewers: aditya_nandakumar, rovka, t.p.northover, qcolombet, ab Reviewed By: ab Subscribers: igorb, dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D28942 llvm-svn: 293172
2017-01-26 19:10:14 +08:00
[tablegen][globalisel] Capture instructions into locals and related infrastructure for multiple instructions matches. Summary: Prepare the way for nested instruction matching support by having actions like CopyRenderer look up operands in the RuleMatcher rather than a specific InstructionMatcher. This allows actions to reference any operand from any matched instruction. It works by checking the 'shape' of the match and capturing each matched instruction to a local variable. If the shape is wrong (not enough operands, leaf nodes where non-leafs are expected, etc.), then the rule exits early without checking the predicates. Once we've captured the instructions, we then test the predicates as before (except using the local variables). If the match is successful, then we render the new instruction as before using the local variables. It's not noticable in this patch but by the time we support multiple instruction matching, this patch will also cause a significant improvement to readability of the emitted code since MRI.getVRegDef(I->getOperand(0).getReg()) will simply be MI1 after emitCxxCaptureStmts(). This isn't quite NFC because I've also fixed a bug that I'm surprised we haven't encountered yet. It now checks there are at least the expected number of operands before accessing them with getOperand(). Depends on D30531 Reviewers: t.p.northover, qcolombet, aditya_nandakumar, ab, rovka Reviewed By: rovka Subscribers: dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D30535 llvm-svn: 298257
2017-03-20 23:20:42 +08:00
InstructionMatcher &getInstructionMatcher() const { return Insn; }
[globalisel][tablegen] Implement unindexed load, non-extending load, and MemVT checks Summary: This includes some context-sensitivity in the MVT to LLT conversion so that pointer types are tested correctly. FIXME: I'm not happy with the way this is done since everything is a special-case. I've yet to find a reasonable way to implement it. select-load.mir fails because <1 x s64> loads in tablegen get priority over s64 loads. This is fixed in the next patch and as such they should be committed together, I've posted them separately to help with the review. Depends on D37456 Reviewers: ab, qcolombet, t.p.northover, rovka, aditya_nandakumar Subscribers: kristof.beyls, javed.absar, llvm-commits, igorb Differential Revision: https://reviews.llvm.org/D37457 llvm-svn: 315884
2017-10-16 08:56:30 +08:00
Error addTypeCheckPredicate(const TypeSetByHwMode &VTy,
Revert MSVC 2017 build fix and fix it by moving the method that implicitly instantiates addPredicate out of line llvm-svn: 315932
2017-10-17 04:31:16 +08:00
bool OperandIsAPointer);
[globalisel][tablegen] Implement unindexed load, non-extending load, and MemVT checks Summary: This includes some context-sensitivity in the MVT to LLT conversion so that pointer types are tested correctly. FIXME: I'm not happy with the way this is done since everything is a special-case. I've yet to find a reasonable way to implement it. select-load.mir fails because <1 x s64> loads in tablegen get priority over s64 loads. This is fixed in the next patch and as such they should be committed together, I've posted them separately to help with the review. Depends on D37456 Reviewers: ab, qcolombet, t.p.northover, rovka, aditya_nandakumar Subscribers: kristof.beyls, javed.absar, llvm-commits, igorb Differential Revision: https://reviews.llvm.org/D37457 llvm-svn: 315884
2017-10-16 08:56:30 +08:00
[globalisel][tablegen] Rename and re-comment to match the new MatchTables. NFC. The conversion to MatchTable left the function names and comments referring to C++ statements and expressions. Updated the names and comments to account for the fact that they're no longer unconstrained statements/expressions. llvm-svn: 307246
2017-07-06 18:06:12 +08:00
/// Emit MatchTable opcodes that test whether the instruction named in
[globalisel][tablegen] Partially fix compile-time regressions by converting matcher to state-machine(s) Summary: Replace the matcher if-statements for each rule with a state-machine. This significantly reduces compile time, memory allocations, and cumulative memory allocation when compiling AArch64InstructionSelector.cpp.o after r303259 is recommitted. The following patches will expand on this further to fully fix the regressions. Reviewers: rovka, ab, t.p.northover, qcolombet, aditya_nandakumar Reviewed By: ab Subscribers: vitalybuka, aemerson, javed.absar, igorb, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D33758 llvm-svn: 307079
2017-07-04 22:35:06 +08:00
/// InsnVarID matches all the predicates and all the operands.
Reapply r332907 "[GlobalISel] Improving InstructionSelect's performance by reducing MatchTable..." Apparently the compile time problem was caused by the fact that not all compilers / STL implementations can automatically convert std::unique_ptr<Derived> to std::unique_ptr<Base>. Fixed (hopefully) by making sure it's std::unique_ptr<Derived>&& (rvalue ref) to std::unique_ptr<Base> conversion instead. llvm-svn: 332917
2018-05-22 07:28:51 +08:00
void emitPredicateOpcodes(MatchTable &Table, RuleMatcher &Rule) {
if (!Optimized) {
std::string Comment;
raw_string_ostream CommentOS(Comment);
CommentOS << "MIs[" << getInsnVarID() << "] ";
if (SymbolicName.empty())
CommentOS << "Operand " << OpIdx;
else
CommentOS << SymbolicName;
Table << MatchTable::Comment(CommentOS.str()) << MatchTable::LineBreak;
}
[globalisel][tablegen] Add control-flow to the MatchTable. Summary: This will allow us to merge the various sub-tables into a single table. This is a compile-time saving at this point. However, this will also enable the optimization of a table so that similar instructions can be tested together, reducing the time spent on the matching the code. The bulk of this patch is a mechanical conversion to the new MatchTable object which is responsible for tracking label definitions and filling in the index of the jump targets. It is also responsible for nicely formatting the table. This was necessary to support the new GIM_Try opcode which takes the index to jump to if the match should fail. This value is unknown during table construction and is filled in during emission. To support nesting try-blocks (although we currently don't emit tables with nested try-blocks), GIM_Reject has been re-introduced to explicitly exit a try-block or fail the overall match if there are no active try-blocks. Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Reviewed By: rovka Subscribers: kristof.beyls, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D35117 llvm-svn: 308596
2017-07-20 17:25:44 +08:00
[TableGen][GlobalISel] Have the predicate directly know which data they are dealing with Prior to this patch, a predicate wouldn't make sense outside of its rule. Indeed, it was only during emitting a rule that a predicate would be made aware of the IDs of the data it is checking. Because of that, predicates could not be moved around or compared between each other. NFC. llvm-svn: 320887
2017-12-16 07:07:42 +08:00
emitPredicateListOpcodes(Table, Rule);
[globalisel] Re-factor ISel matchers into a hierarchy. NFC Summary: This should make it possible to easily add everything needed to import all the existing SelectionDAG rules. It should also serve the likely kinds of GlobalISel rules (some of which are not currently representable in SelectionDAG) once we've nailed down the tablegen definition for that. The hierarchy is as follows: MatcherRule - A matching rule. Currently used to emit C++ ISel code but will | also be used to emit test cases and tablegen definitions in the | near future. |- Instruction(s) - Represents the instruction to be matched. |- Instruction Predicate(s) - Test the opcode, arithmetic flags, etc. of an | instruction. \- Operand(s) - Represents a particular operand of the instruction. In the | future, there may be subclasses to test the same predicates | on multiple operands (including for variadic instructions). \ Operand Predicate(s) - Test the type, register bank, etc. of an operand. This is where the ComplexPattern equivalent will be represented. It's also nested-instruction matching will live as a predicate that follows the DefUse chain to the Def and tests a MatcherRule from that position. Support for multiple instruction matchers in a rule has been retained from the existing code but has been adjusted to assert when it is used. Previously it would silently drop all but the first instruction matcher. The tablegen-erated file is not functionally changed but has more parentheses and no longer attempts to format the if-statements since keeping track of the indentation is tricky in the presence of the matcher hierarchy. It would be nice to have CMakes tablegen() run the output through clang-format (when available) so we don't have to complicate TableGen with pretty-printing. It's also worth mentioning that this hierarchy will also be able to emit TableGen definitions and test cases in the near future. This is the reason for favouring explicit emit*() calls rather than the << operator. Reviewers: aditya_nandakumar, rovka, t.p.northover, qcolombet, ab Reviewed By: ab Subscribers: igorb, dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D28942 llvm-svn: 293172
2017-01-26 19:10:14 +08:00
}
[globalisel] Sort RuleMatchers by priority. Summary: This makes more important rules have priority over less important rules. For example, '%a = G_ADD $b:s64, $c:s64' has priority over '%a = G_ADD $b:s32, $c:s32'. Previously these rules were emitted in the correct order by chance. NFC in this patch but it is required to make the next patch work correctly. Depends on D29710 Reviewers: t.p.northover, ab, qcolombet, aditya_nandakumar, rovka Reviewed By: ab, rovka Subscribers: javed.absar, dberris, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D29711 llvm-svn: 296121
2017-02-24 21:58:11 +08:00
/// Compare the priority of this object and B.
///
/// Returns true if this object is more important than B.
Reapply r332907 "[GlobalISel] Improving InstructionSelect's performance by reducing MatchTable..." Apparently the compile time problem was caused by the fact that not all compilers / STL implementations can automatically convert std::unique_ptr<Derived> to std::unique_ptr<Base>. Fixed (hopefully) by making sure it's std::unique_ptr<Derived>&& (rvalue ref) to std::unique_ptr<Base> conversion instead. llvm-svn: 332917
2018-05-22 07:28:51 +08:00
bool isHigherPriorityThan(OperandMatcher &B) {
[globalisel] Sort RuleMatchers by priority. Summary: This makes more important rules have priority over less important rules. For example, '%a = G_ADD $b:s64, $c:s64' has priority over '%a = G_ADD $b:s32, $c:s32'. Previously these rules were emitted in the correct order by chance. NFC in this patch but it is required to make the next patch work correctly. Depends on D29710 Reviewers: t.p.northover, ab, qcolombet, aditya_nandakumar, rovka Reviewed By: ab, rovka Subscribers: javed.absar, dberris, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D29711 llvm-svn: 296121
2017-02-24 21:58:11 +08:00
// Operand matchers involving more predicates have higher priority.
if (predicates_size() > B.predicates_size())
return true;
if (predicates_size() < B.predicates_size())
return false;
// This assumes that predicates are added in a consistent order.
Reapply r332907 "[GlobalISel] Improving InstructionSelect's performance by reducing MatchTable..." Apparently the compile time problem was caused by the fact that not all compilers / STL implementations can automatically convert std::unique_ptr<Derived> to std::unique_ptr<Base>. Fixed (hopefully) by making sure it's std::unique_ptr<Derived>&& (rvalue ref) to std::unique_ptr<Base> conversion instead. llvm-svn: 332917
2018-05-22 07:28:51 +08:00
for (auto &&Predicate : zip(predicates(), B.predicates())) {
[globalisel] Sort RuleMatchers by priority. Summary: This makes more important rules have priority over less important rules. For example, '%a = G_ADD $b:s64, $c:s64' has priority over '%a = G_ADD $b:s32, $c:s32'. Previously these rules were emitted in the correct order by chance. NFC in this patch but it is required to make the next patch work correctly. Depends on D29710 Reviewers: t.p.northover, ab, qcolombet, aditya_nandakumar, rovka Reviewed By: ab, rovka Subscribers: javed.absar, dberris, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D29711 llvm-svn: 296121
2017-02-24 21:58:11 +08:00
if (std::get<0>(Predicate)->isHigherPriorityThan(*std::get<1>(Predicate)))
return true;
if (std::get<1>(Predicate)->isHigherPriorityThan(*std::get<0>(Predicate)))
return false;
}
return false;
};
[globalisel][tblgen] Add support for ComplexPatterns Summary: Adds a new kind of MachineOperand: MO_Placeholder. This operand must not appear in the MIR and only exists as a way of creating an 'uninitialized' operand until a matcher function overwrites it. Depends on D30046, D29712 Reviewers: t.p.northover, ab, rovka, aditya_nandakumar, javed.absar, qcolombet Reviewed By: qcolombet Subscribers: dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D30089 llvm-svn: 297782
2017-03-15 05:32:08 +08:00
/// Report the maximum number of temporary operands needed by the operand
/// matcher.
Reapply r332907 "[GlobalISel] Improving InstructionSelect's performance by reducing MatchTable..." Apparently the compile time problem was caused by the fact that not all compilers / STL implementations can automatically convert std::unique_ptr<Derived> to std::unique_ptr<Base>. Fixed (hopefully) by making sure it's std::unique_ptr<Derived>&& (rvalue ref) to std::unique_ptr<Base> conversion instead. llvm-svn: 332917
2018-05-22 07:28:51 +08:00
unsigned countRendererFns() {
[globalisel][tblgen] Add support for ComplexPatterns Summary: Adds a new kind of MachineOperand: MO_Placeholder. This operand must not appear in the MIR and only exists as a way of creating an 'uninitialized' operand until a matcher function overwrites it. Depends on D30046, D29712 Reviewers: t.p.northover, ab, rovka, aditya_nandakumar, javed.absar, qcolombet Reviewed By: qcolombet Subscribers: dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D30089 llvm-svn: 297782
2017-03-15 05:32:08 +08:00
return std::accumulate(
predicates().begin(), predicates().end(), 0,
[](unsigned A,
const std::unique_ptr<OperandPredicateMatcher> &Predicate) {
[globalisel][tablegen] Revise API for ComplexPattern operands to improve flexibility. Summary: Some targets need to be able to do more complex rendering than just adding an operand or two to an instruction. For example, it may need to insert an instruction to extract a subreg first, or it may need to perform an operation on the operand. In SelectionDAG, targets would create SDNode's to achieve the desired effect during the complex pattern predicate. This worked because SelectionDAG had a form of garbage collection that would take care of SDNode's that were created but not used due to a later predicate rejecting a match. This doesn't translate well to GlobalISel and the churn was wasteful. The API changes in this patch enable GlobalISel to accomplish the same thing without the waste. The API is now: InstructionSelector::OptionalComplexRendererFn selectArithImmed(MachineOperand &Root) const; where Root is the root of the match. The return value can be omitted to indicate that the predicate failed to match, or a function with the signature ComplexRendererFn can be returned. For example: return OptionalComplexRendererFn( [=](MachineInstrBuilder &MIB) { MIB.addImm(Immed).addImm(ShVal); }); adds two immediate operands to the rendered instruction. Immed and ShVal are captured from the predicate function. As an added bonus, this also reduces the amount of information we need to provide to GIComplexOperandMatcher. Depends on D31418 Reviewers: aditya_nandakumar, t.p.northover, qcolombet, rovka, ab, javed.absar Reviewed By: ab Subscribers: dberris, kristof.beyls, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D31761 llvm-svn: 301079
2017-04-22 23:11:04 +08:00
return A + Predicate->countRendererFns();
[globalisel][tblgen] Add support for ComplexPatterns Summary: Adds a new kind of MachineOperand: MO_Placeholder. This operand must not appear in the MIR and only exists as a way of creating an 'uninitialized' operand until a matcher function overwrites it. Depends on D30046, D29712 Reviewers: t.p.northover, ab, rovka, aditya_nandakumar, javed.absar, qcolombet Reviewed By: qcolombet Subscribers: dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D30089 llvm-svn: 297782
2017-03-15 05:32:08 +08:00
});
}
[globalisel][tablegen] Fix patterns involving multiple ComplexPatterns. Summary: Temporaries are now allocated to operands instead of predicates and this allocation is used to correctly pair up the rendered operands with the matched operands. Previously, ComplexPatterns were allocated temporaries independently in the Src Pattern and Dst Pattern, leading to mismatches. Additionally, the Dst Pattern failed to account for the allocated index and therefore always used temporary 0, 1, ... when it should have used base+0, base+1, ... Thanks to Aditya Nandakumar for noticing the bug. Depends on D30539 Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Reviewed By: rovka Subscribers: igorb, dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D31054 llvm-svn: 299538
2017-04-05 21:14:03 +08:00
unsigned getAllocatedTemporariesBaseID() const {
return AllocatedTemporariesBaseID;
}
[globalisel][tablegen] Simplify named operand/operator lookups and fix a wrong-code bug this revealed. Summary: Operand variable lookups are now performed by the RuleMatcher rather than searching the whole matcher hierarchy for a match. This revealed a wrong-code bug that currently affects ARM and X86 where patterns that use a variable more than once in the match pattern will be imported but won't check that the operands are identical. This can cause the tablegen-erated matcher to accept matches that should be rejected. Depends on D36569 Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Subscribers: aemerson, igorb, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D36618 llvm-svn: 315780
2017-10-14 08:31:58 +08:00
Reapply r332907 "[GlobalISel] Improving InstructionSelect's performance by reducing MatchTable..." Apparently the compile time problem was caused by the fact that not all compilers / STL implementations can automatically convert std::unique_ptr<Derived> to std::unique_ptr<Base>. Fixed (hopefully) by making sure it's std::unique_ptr<Derived>&& (rvalue ref) to std::unique_ptr<Base> conversion instead. llvm-svn: 332917
2018-05-22 07:28:51 +08:00
bool isSameAsAnotherOperand() {
[globalisel][tablegen] Simplify named operand/operator lookups and fix a wrong-code bug this revealed. Summary: Operand variable lookups are now performed by the RuleMatcher rather than searching the whole matcher hierarchy for a match. This revealed a wrong-code bug that currently affects ARM and X86 where patterns that use a variable more than once in the match pattern will be imported but won't check that the operands are identical. This can cause the tablegen-erated matcher to accept matches that should be rejected. Depends on D36569 Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Subscribers: aemerson, igorb, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D36618 llvm-svn: 315780
2017-10-14 08:31:58 +08:00
for (const auto &Predicate : predicates())
if (isa<SameOperandMatcher>(Predicate))
return true;
return false;
}
[globalisel] Re-factor ISel matchers into a hierarchy. NFC Summary: This should make it possible to easily add everything needed to import all the existing SelectionDAG rules. It should also serve the likely kinds of GlobalISel rules (some of which are not currently representable in SelectionDAG) once we've nailed down the tablegen definition for that. The hierarchy is as follows: MatcherRule - A matching rule. Currently used to emit C++ ISel code but will | also be used to emit test cases and tablegen definitions in the | near future. |- Instruction(s) - Represents the instruction to be matched. |- Instruction Predicate(s) - Test the opcode, arithmetic flags, etc. of an | instruction. \- Operand(s) - Represents a particular operand of the instruction. In the | future, there may be subclasses to test the same predicates | on multiple operands (including for variadic instructions). \ Operand Predicate(s) - Test the type, register bank, etc. of an operand. This is where the ComplexPattern equivalent will be represented. It's also nested-instruction matching will live as a predicate that follows the DefUse chain to the Def and tests a MatcherRule from that position. Support for multiple instruction matchers in a rule has been retained from the existing code but has been adjusted to assert when it is used. Previously it would silently drop all but the first instruction matcher. The tablegen-erated file is not functionally changed but has more parentheses and no longer attempts to format the if-statements since keeping track of the indentation is tricky in the presence of the matcher hierarchy. It would be nice to have CMakes tablegen() run the output through clang-format (when available) so we don't have to complicate TableGen with pretty-printing. It's also worth mentioning that this hierarchy will also be able to emit TableGen definitions and test cases in the near future. This is the reason for favouring explicit emit*() calls rather than the << operator. Reviewers: aditya_nandakumar, rovka, t.p.northover, qcolombet, ab Reviewed By: ab Subscribers: igorb, dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D28942 llvm-svn: 293172
2017-01-26 19:10:14 +08:00
};
Revert MSVC 2017 build fix and fix it by moving the method that implicitly instantiates addPredicate out of line llvm-svn: 315932
2017-10-17 04:31:16 +08:00
Error OperandMatcher::addTypeCheckPredicate(const TypeSetByHwMode &VTy,
[TableGen][GlobalISel] Have the predicate directly know which data they are dealing with Prior to this patch, a predicate wouldn't make sense outside of its rule. Indeed, it was only during emitting a rule that a predicate would be made aware of the IDs of the data it is checking. Because of that, predicates could not be moved around or compared between each other. NFC. llvm-svn: 320887
2017-12-16 07:07:42 +08:00
bool OperandIsAPointer) {
Revert MSVC 2017 build fix and fix it by moving the method that implicitly instantiates addPredicate out of line llvm-svn: 315932
2017-10-17 04:31:16 +08:00
if (!VTy.isMachineValueType())
return failedImport("unsupported typeset");
if (VTy.getMachineValueType() == MVT::iPTR && OperandIsAPointer) {
addPredicate<PointerToAnyOperandMatcher>(0);
return Error::success();
}
auto OpTyOrNone = MVTToLLT(VTy.getMachineValueType().SimpleTy);
if (!OpTyOrNone)
return failedImport("unsupported type");
if (OperandIsAPointer)
addPredicate<PointerToAnyOperandMatcher>(OpTyOrNone->get().getSizeInBits());
else
addPredicate<LLTOperandMatcher>(*OpTyOrNone);
return Error::success();
}
[globalisel][tablegen] Fix patterns involving multiple ComplexPatterns. Summary: Temporaries are now allocated to operands instead of predicates and this allocation is used to correctly pair up the rendered operands with the matched operands. Previously, ComplexPatterns were allocated temporaries independently in the Src Pattern and Dst Pattern, leading to mismatches. Additionally, the Dst Pattern failed to account for the allocated index and therefore always used temporary 0, 1, ... when it should have used base+0, base+1, ... Thanks to Aditya Nandakumar for noticing the bug. Depends on D30539 Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Reviewed By: rovka Subscribers: igorb, dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D31054 llvm-svn: 299538
2017-04-05 21:14:03 +08:00
unsigned ComplexPatternOperandMatcher::getAllocatedTemporariesBaseID() const {
return Operand.getAllocatedTemporariesBaseID();
}
[globalisel] Re-factor ISel matchers into a hierarchy. NFC Summary: This should make it possible to easily add everything needed to import all the existing SelectionDAG rules. It should also serve the likely kinds of GlobalISel rules (some of which are not currently representable in SelectionDAG) once we've nailed down the tablegen definition for that. The hierarchy is as follows: MatcherRule - A matching rule. Currently used to emit C++ ISel code but will | also be used to emit test cases and tablegen definitions in the | near future. |- Instruction(s) - Represents the instruction to be matched. |- Instruction Predicate(s) - Test the opcode, arithmetic flags, etc. of an | instruction. \- Operand(s) - Represents a particular operand of the instruction. In the | future, there may be subclasses to test the same predicates | on multiple operands (including for variadic instructions). \ Operand Predicate(s) - Test the type, register bank, etc. of an operand. This is where the ComplexPattern equivalent will be represented. It's also nested-instruction matching will live as a predicate that follows the DefUse chain to the Def and tests a MatcherRule from that position. Support for multiple instruction matchers in a rule has been retained from the existing code but has been adjusted to assert when it is used. Previously it would silently drop all but the first instruction matcher. The tablegen-erated file is not functionally changed but has more parentheses and no longer attempts to format the if-statements since keeping track of the indentation is tricky in the presence of the matcher hierarchy. It would be nice to have CMakes tablegen() run the output through clang-format (when available) so we don't have to complicate TableGen with pretty-printing. It's also worth mentioning that this hierarchy will also be able to emit TableGen definitions and test cases in the near future. This is the reason for favouring explicit emit*() calls rather than the << operator. Reviewers: aditya_nandakumar, rovka, t.p.northover, qcolombet, ab Reviewed By: ab Subscribers: igorb, dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D28942 llvm-svn: 293172
2017-01-26 19:10:14 +08:00
/// Generates code to check a predicate on an instruction.
///
/// Typical predicates include:
/// * The opcode of the instruction is a particular value.
/// * The nsw/nuw flag is/isn't set.
[TableGen][GlobalISel] Add a common class for all PredicateMatcher NFC. llvm-svn: 320767
2017-12-15 07:44:07 +08:00
class InstructionPredicateMatcher : public PredicateMatcher {
[globalisel] Re-factor ISel matchers into a hierarchy. NFC Summary: This should make it possible to easily add everything needed to import all the existing SelectionDAG rules. It should also serve the likely kinds of GlobalISel rules (some of which are not currently representable in SelectionDAG) once we've nailed down the tablegen definition for that. The hierarchy is as follows: MatcherRule - A matching rule. Currently used to emit C++ ISel code but will | also be used to emit test cases and tablegen definitions in the | near future. |- Instruction(s) - Represents the instruction to be matched. |- Instruction Predicate(s) - Test the opcode, arithmetic flags, etc. of an | instruction. \- Operand(s) - Represents a particular operand of the instruction. In the | future, there may be subclasses to test the same predicates | on multiple operands (including for variadic instructions). \ Operand Predicate(s) - Test the type, register bank, etc. of an operand. This is where the ComplexPattern equivalent will be represented. It's also nested-instruction matching will live as a predicate that follows the DefUse chain to the Def and tests a MatcherRule from that position. Support for multiple instruction matchers in a rule has been retained from the existing code but has been adjusted to assert when it is used. Previously it would silently drop all but the first instruction matcher. The tablegen-erated file is not functionally changed but has more parentheses and no longer attempts to format the if-statements since keeping track of the indentation is tricky in the presence of the matcher hierarchy. It would be nice to have CMakes tablegen() run the output through clang-format (when available) so we don't have to complicate TableGen with pretty-printing. It's also worth mentioning that this hierarchy will also be able to emit TableGen definitions and test cases in the near future. This is the reason for favouring explicit emit*() calls rather than the << operator. Reviewers: aditya_nandakumar, rovka, t.p.northover, qcolombet, ab Reviewed By: ab Subscribers: igorb, dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D28942 llvm-svn: 293172
2017-01-26 19:10:14 +08:00
public:
[TableGen][GlobalISel] Have the predicate directly know which data they are dealing with Prior to this patch, a predicate wouldn't make sense outside of its rule. Indeed, it was only during emitting a rule that a predicate would be made aware of the IDs of the data it is checking. Because of that, predicates could not be moved around or compared between each other. NFC. llvm-svn: 320887
2017-12-16 07:07:42 +08:00
InstructionPredicateMatcher(PredicateKind Kind, unsigned InsnVarID)
: PredicateMatcher(Kind, InsnVarID) {}
[globalisel] Re-factor ISel matchers into a hierarchy. NFC Summary: This should make it possible to easily add everything needed to import all the existing SelectionDAG rules. It should also serve the likely kinds of GlobalISel rules (some of which are not currently representable in SelectionDAG) once we've nailed down the tablegen definition for that. The hierarchy is as follows: MatcherRule - A matching rule. Currently used to emit C++ ISel code but will | also be used to emit test cases and tablegen definitions in the | near future. |- Instruction(s) - Represents the instruction to be matched. |- Instruction Predicate(s) - Test the opcode, arithmetic flags, etc. of an | instruction. \- Operand(s) - Represents a particular operand of the instruction. In the | future, there may be subclasses to test the same predicates | on multiple operands (including for variadic instructions). \ Operand Predicate(s) - Test the type, register bank, etc. of an operand. This is where the ComplexPattern equivalent will be represented. It's also nested-instruction matching will live as a predicate that follows the DefUse chain to the Def and tests a MatcherRule from that position. Support for multiple instruction matchers in a rule has been retained from the existing code but has been adjusted to assert when it is used. Previously it would silently drop all but the first instruction matcher. The tablegen-erated file is not functionally changed but has more parentheses and no longer attempts to format the if-statements since keeping track of the indentation is tricky in the presence of the matcher hierarchy. It would be nice to have CMakes tablegen() run the output through clang-format (when available) so we don't have to complicate TableGen with pretty-printing. It's also worth mentioning that this hierarchy will also be able to emit TableGen definitions and test cases in the near future. This is the reason for favouring explicit emit*() calls rather than the << operator. Reviewers: aditya_nandakumar, rovka, t.p.northover, qcolombet, ab Reviewed By: ab Subscribers: igorb, dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D28942 llvm-svn: 293172
2017-01-26 19:10:14 +08:00
virtual ~InstructionPredicateMatcher() {}
[globalisel] Sort RuleMatchers by priority. Summary: This makes more important rules have priority over less important rules. For example, '%a = G_ADD $b:s64, $c:s64' has priority over '%a = G_ADD $b:s32, $c:s32'. Previously these rules were emitted in the correct order by chance. NFC in this patch but it is required to make the next patch work correctly. Depends on D29710 Reviewers: t.p.northover, ab, qcolombet, aditya_nandakumar, rovka Reviewed By: ab, rovka Subscribers: javed.absar, dberris, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D29711 llvm-svn: 296121
2017-02-24 21:58:11 +08:00
/// Compare the priority of this object and B.
///
/// Returns true if this object is more important than B.
[globalisel][tablegen] Multiple 80-col corrections. llvm-svn: 306544
2017-06-28 21:50:04 +08:00
virtual bool
isHigherPriorityThan(const InstructionPredicateMatcher &B) const {
[globalisel] Sort RuleMatchers by priority. Summary: This makes more important rules have priority over less important rules. For example, '%a = G_ADD $b:s64, $c:s64' has priority over '%a = G_ADD $b:s32, $c:s32'. Previously these rules were emitted in the correct order by chance. NFC in this patch but it is required to make the next patch work correctly. Depends on D29710 Reviewers: t.p.northover, ab, qcolombet, aditya_nandakumar, rovka Reviewed By: ab, rovka Subscribers: javed.absar, dberris, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D29711 llvm-svn: 296121
2017-02-24 21:58:11 +08:00
return Kind < B.Kind;
};
[globalisel] Re-factor ISel matchers into a hierarchy. NFC Summary: This should make it possible to easily add everything needed to import all the existing SelectionDAG rules. It should also serve the likely kinds of GlobalISel rules (some of which are not currently representable in SelectionDAG) once we've nailed down the tablegen definition for that. The hierarchy is as follows: MatcherRule - A matching rule. Currently used to emit C++ ISel code but will | also be used to emit test cases and tablegen definitions in the | near future. |- Instruction(s) - Represents the instruction to be matched. |- Instruction Predicate(s) - Test the opcode, arithmetic flags, etc. of an | instruction. \- Operand(s) - Represents a particular operand of the instruction. In the | future, there may be subclasses to test the same predicates | on multiple operands (including for variadic instructions). \ Operand Predicate(s) - Test the type, register bank, etc. of an operand. This is where the ComplexPattern equivalent will be represented. It's also nested-instruction matching will live as a predicate that follows the DefUse chain to the Def and tests a MatcherRule from that position. Support for multiple instruction matchers in a rule has been retained from the existing code but has been adjusted to assert when it is used. Previously it would silently drop all but the first instruction matcher. The tablegen-erated file is not functionally changed but has more parentheses and no longer attempts to format the if-statements since keeping track of the indentation is tricky in the presence of the matcher hierarchy. It would be nice to have CMakes tablegen() run the output through clang-format (when available) so we don't have to complicate TableGen with pretty-printing. It's also worth mentioning that this hierarchy will also be able to emit TableGen definitions and test cases in the near future. This is the reason for favouring explicit emit*() calls rather than the << operator. Reviewers: aditya_nandakumar, rovka, t.p.northover, qcolombet, ab Reviewed By: ab Subscribers: igorb, dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D28942 llvm-svn: 293172
2017-01-26 19:10:14 +08:00
};
Re-commit: [globalisel][tablegen] Add support for ImmLeaf without SDNodeXForm Summary: This patch adds support for predicates on imm nodes but only for ImmLeaf and not for PatLeaf or PatFrag and only where the value does not need to be transformed before being rendered into the instruction. The limitation on PatLeaf/PatFrag/SDNodeXForm is due to differences in the necessary target-supplied C++ for GlobalISel. Depends on D36085 The previous commit was reverted for breaking the build but this appears to have been the recurring problem on the Windows bots with tablegen not being re-run when llvm-tblgen is changed but the .td's aren't. If it re-occurs then forcing a build with clean=True should fix it but this string should do this in advance: Requires a clean build. Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Reviewed By: rovka Subscribers: kristof.beyls, javed.absar, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D36086 llvm-svn: 311645
2017-08-24 17:11:20 +08:00
template <>
std::string
Reapply r332907 "[GlobalISel] Improving InstructionSelect's performance by reducing MatchTable..." Apparently the compile time problem was caused by the fact that not all compilers / STL implementations can automatically convert std::unique_ptr<Derived> to std::unique_ptr<Base>. Fixed (hopefully) by making sure it's std::unique_ptr<Derived>&& (rvalue ref) to std::unique_ptr<Base> conversion instead. llvm-svn: 332917
2018-05-22 07:28:51 +08:00
PredicateListMatcher<PredicateMatcher>::getNoPredicateComment() const {
Re-commit: [globalisel][tablegen] Add support for ImmLeaf without SDNodeXForm Summary: This patch adds support for predicates on imm nodes but only for ImmLeaf and not for PatLeaf or PatFrag and only where the value does not need to be transformed before being rendered into the instruction. The limitation on PatLeaf/PatFrag/SDNodeXForm is due to differences in the necessary target-supplied C++ for GlobalISel. Depends on D36085 The previous commit was reverted for breaking the build but this appears to have been the recurring problem on the Windows bots with tablegen not being re-run when llvm-tblgen is changed but the .td's aren't. If it re-occurs then forcing a build with clean=True should fix it but this string should do this in advance: Requires a clean build. Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Reviewed By: rovka Subscribers: kristof.beyls, javed.absar, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D36086 llvm-svn: 311645
2017-08-24 17:11:20 +08:00
return "No instruction predicates";
}
[globalisel] Re-factor ISel matchers into a hierarchy. NFC Summary: This should make it possible to easily add everything needed to import all the existing SelectionDAG rules. It should also serve the likely kinds of GlobalISel rules (some of which are not currently representable in SelectionDAG) once we've nailed down the tablegen definition for that. The hierarchy is as follows: MatcherRule - A matching rule. Currently used to emit C++ ISel code but will | also be used to emit test cases and tablegen definitions in the | near future. |- Instruction(s) - Represents the instruction to be matched. |- Instruction Predicate(s) - Test the opcode, arithmetic flags, etc. of an | instruction. \- Operand(s) - Represents a particular operand of the instruction. In the | future, there may be subclasses to test the same predicates | on multiple operands (including for variadic instructions). \ Operand Predicate(s) - Test the type, register bank, etc. of an operand. This is where the ComplexPattern equivalent will be represented. It's also nested-instruction matching will live as a predicate that follows the DefUse chain to the Def and tests a MatcherRule from that position. Support for multiple instruction matchers in a rule has been retained from the existing code but has been adjusted to assert when it is used. Previously it would silently drop all but the first instruction matcher. The tablegen-erated file is not functionally changed but has more parentheses and no longer attempts to format the if-statements since keeping track of the indentation is tricky in the presence of the matcher hierarchy. It would be nice to have CMakes tablegen() run the output through clang-format (when available) so we don't have to complicate TableGen with pretty-printing. It's also worth mentioning that this hierarchy will also be able to emit TableGen definitions and test cases in the near future. This is the reason for favouring explicit emit*() calls rather than the << operator. Reviewers: aditya_nandakumar, rovka, t.p.northover, qcolombet, ab Reviewed By: ab Subscribers: igorb, dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D28942 llvm-svn: 293172
2017-01-26 19:10:14 +08:00
/// Generates code to check the opcode of an instruction.
class InstructionOpcodeMatcher : public InstructionPredicateMatcher {
protected:
const CodeGenInstruction *I;
Reapply r332907 "[GlobalISel] Improving InstructionSelect's performance by reducing MatchTable..." Apparently the compile time problem was caused by the fact that not all compilers / STL implementations can automatically convert std::unique_ptr<Derived> to std::unique_ptr<Base>. Fixed (hopefully) by making sure it's std::unique_ptr<Derived>&& (rvalue ref) to std::unique_ptr<Base> conversion instead. llvm-svn: 332917
2018-05-22 07:28:51 +08:00
static DenseMap<const CodeGenInstruction *, unsigned> OpcodeValues;
[globalisel] Re-factor ISel matchers into a hierarchy. NFC Summary: This should make it possible to easily add everything needed to import all the existing SelectionDAG rules. It should also serve the likely kinds of GlobalISel rules (some of which are not currently representable in SelectionDAG) once we've nailed down the tablegen definition for that. The hierarchy is as follows: MatcherRule - A matching rule. Currently used to emit C++ ISel code but will | also be used to emit test cases and tablegen definitions in the | near future. |- Instruction(s) - Represents the instruction to be matched. |- Instruction Predicate(s) - Test the opcode, arithmetic flags, etc. of an | instruction. \- Operand(s) - Represents a particular operand of the instruction. In the | future, there may be subclasses to test the same predicates | on multiple operands (including for variadic instructions). \ Operand Predicate(s) - Test the type, register bank, etc. of an operand. This is where the ComplexPattern equivalent will be represented. It's also nested-instruction matching will live as a predicate that follows the DefUse chain to the Def and tests a MatcherRule from that position. Support for multiple instruction matchers in a rule has been retained from the existing code but has been adjusted to assert when it is used. Previously it would silently drop all but the first instruction matcher. The tablegen-erated file is not functionally changed but has more parentheses and no longer attempts to format the if-statements since keeping track of the indentation is tricky in the presence of the matcher hierarchy. It would be nice to have CMakes tablegen() run the output through clang-format (when available) so we don't have to complicate TableGen with pretty-printing. It's also worth mentioning that this hierarchy will also be able to emit TableGen definitions and test cases in the near future. This is the reason for favouring explicit emit*() calls rather than the << operator. Reviewers: aditya_nandakumar, rovka, t.p.northover, qcolombet, ab Reviewed By: ab Subscribers: igorb, dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D28942 llvm-svn: 293172
2017-01-26 19:10:14 +08:00
public:
Reapply r332907 "[GlobalISel] Improving InstructionSelect's performance by reducing MatchTable..." Apparently the compile time problem was caused by the fact that not all compilers / STL implementations can automatically convert std::unique_ptr<Derived> to std::unique_ptr<Base>. Fixed (hopefully) by making sure it's std::unique_ptr<Derived>&& (rvalue ref) to std::unique_ptr<Base> conversion instead. llvm-svn: 332917
2018-05-22 07:28:51 +08:00
static void initOpcodeValuesMap(const CodeGenTarget &Target) {
OpcodeValues.clear();
unsigned OpcodeValue = 0;
for (const CodeGenInstruction *I : Target.getInstructionsByEnumValue())
OpcodeValues[I] = OpcodeValue++;
}
[TableGen][GlobalISel] Have the predicate directly know which data they are dealing with Prior to this patch, a predicate wouldn't make sense outside of its rule. Indeed, it was only during emitting a rule that a predicate would be made aware of the IDs of the data it is checking. Because of that, predicates could not be moved around or compared between each other. NFC. llvm-svn: 320887
2017-12-16 07:07:42 +08:00
InstructionOpcodeMatcher(unsigned InsnVarID, const CodeGenInstruction *I)
: InstructionPredicateMatcher(IPM_Opcode, InsnVarID), I(I) {}
[globalisel] Re-factor ISel matchers into a hierarchy. NFC Summary: This should make it possible to easily add everything needed to import all the existing SelectionDAG rules. It should also serve the likely kinds of GlobalISel rules (some of which are not currently representable in SelectionDAG) once we've nailed down the tablegen definition for that. The hierarchy is as follows: MatcherRule - A matching rule. Currently used to emit C++ ISel code but will | also be used to emit test cases and tablegen definitions in the | near future. |- Instruction(s) - Represents the instruction to be matched. |- Instruction Predicate(s) - Test the opcode, arithmetic flags, etc. of an | instruction. \- Operand(s) - Represents a particular operand of the instruction. In the | future, there may be subclasses to test the same predicates | on multiple operands (including for variadic instructions). \ Operand Predicate(s) - Test the type, register bank, etc. of an operand. This is where the ComplexPattern equivalent will be represented. It's also nested-instruction matching will live as a predicate that follows the DefUse chain to the Def and tests a MatcherRule from that position. Support for multiple instruction matchers in a rule has been retained from the existing code but has been adjusted to assert when it is used. Previously it would silently drop all but the first instruction matcher. The tablegen-erated file is not functionally changed but has more parentheses and no longer attempts to format the if-statements since keeping track of the indentation is tricky in the presence of the matcher hierarchy. It would be nice to have CMakes tablegen() run the output through clang-format (when available) so we don't have to complicate TableGen with pretty-printing. It's also worth mentioning that this hierarchy will also be able to emit TableGen definitions and test cases in the near future. This is the reason for favouring explicit emit*() calls rather than the << operator. Reviewers: aditya_nandakumar, rovka, t.p.northover, qcolombet, ab Reviewed By: ab Subscribers: igorb, dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D28942 llvm-svn: 293172
2017-01-26 19:10:14 +08:00
[TableGen][GlobalISel] Add a common class for all PredicateMatcher NFC. llvm-svn: 320767
2017-12-15 07:44:07 +08:00
static bool classof(const PredicateMatcher *P) {
[globalisel] Sort RuleMatchers by priority. Summary: This makes more important rules have priority over less important rules. For example, '%a = G_ADD $b:s64, $c:s64' has priority over '%a = G_ADD $b:s32, $c:s32'. Previously these rules were emitted in the correct order by chance. NFC in this patch but it is required to make the next patch work correctly. Depends on D29710 Reviewers: t.p.northover, ab, qcolombet, aditya_nandakumar, rovka Reviewed By: ab, rovka Subscribers: javed.absar, dberris, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D29711 llvm-svn: 296121
2017-02-24 21:58:11 +08:00
return P->getKind() == IPM_Opcode;
}
[TableGen][GlobalISel] Make the different Matcher comparable This opens refactoring opportunities in the match table now that we can check that two predicates are the same. NFC. llvm-svn: 320890
2017-12-16 07:24:39 +08:00
bool isIdentical(const PredicateMatcher &B) const override {
return InstructionPredicateMatcher::isIdentical(B) &&
I == cast<InstructionOpcodeMatcher>(&B)->I;
}
Reapply r332907 "[GlobalISel] Improving InstructionSelect's performance by reducing MatchTable..." Apparently the compile time problem was caused by the fact that not all compilers / STL implementations can automatically convert std::unique_ptr<Derived> to std::unique_ptr<Base>. Fixed (hopefully) by making sure it's std::unique_ptr<Derived>&& (rvalue ref) to std::unique_ptr<Base> conversion instead. llvm-svn: 332917
2018-05-22 07:28:51 +08:00
MatchTableRecord getValue() const override {
const auto VI = OpcodeValues.find(I);
if (VI != OpcodeValues.end())
return MatchTable::NamedValue(I->Namespace, I->TheDef->getName(),
VI->second);
return MatchTable::NamedValue(I->Namespace, I->TheDef->getName());
}
bool hasValue() const override { return OpcodeValues.count(I); }
[TableGen][GlobalISel] Make the different Matcher comparable This opens refactoring opportunities in the match table now that we can check that two predicates are the same. NFC. llvm-svn: 320890
2017-12-16 07:24:39 +08:00
[TableGen][GlobalISel] Have the predicate directly know which data they are dealing with Prior to this patch, a predicate wouldn't make sense outside of its rule. Indeed, it was only during emitting a rule that a predicate would be made aware of the IDs of the data it is checking. Because of that, predicates could not be moved around or compared between each other. NFC. llvm-svn: 320887
2017-12-16 07:07:42 +08:00
void emitPredicateOpcodes(MatchTable &Table,
RuleMatcher &Rule) const override {
[globalisel][tablegen] Add control-flow to the MatchTable. Summary: This will allow us to merge the various sub-tables into a single table. This is a compile-time saving at this point. However, this will also enable the optimization of a table so that similar instructions can be tested together, reducing the time spent on the matching the code. The bulk of this patch is a mechanical conversion to the new MatchTable object which is responsible for tracking label definitions and filling in the index of the jump targets. It is also responsible for nicely formatting the table. This was necessary to support the new GIM_Try opcode which takes the index to jump to if the match should fail. This value is unknown during table construction and is filled in during emission. To support nesting try-blocks (although we currently don't emit tables with nested try-blocks), GIM_Reject has been re-introduced to explicitly exit a try-block or fail the overall match if there are no active try-blocks. Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Reviewed By: rovka Subscribers: kristof.beyls, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D35117 llvm-svn: 308596
2017-07-20 17:25:44 +08:00
Table << MatchTable::Opcode("GIM_CheckOpcode") << MatchTable::Comment("MI")
Reapply r332907 "[GlobalISel] Improving InstructionSelect's performance by reducing MatchTable..." Apparently the compile time problem was caused by the fact that not all compilers / STL implementations can automatically convert std::unique_ptr<Derived> to std::unique_ptr<Base>. Fixed (hopefully) by making sure it's std::unique_ptr<Derived>&& (rvalue ref) to std::unique_ptr<Base> conversion instead. llvm-svn: 332917
2018-05-22 07:28:51 +08:00
<< MatchTable::IntValue(InsnVarID) << getValue()
[globalisel][tablegen] Add control-flow to the MatchTable. Summary: This will allow us to merge the various sub-tables into a single table. This is a compile-time saving at this point. However, this will also enable the optimization of a table so that similar instructions can be tested together, reducing the time spent on the matching the code. The bulk of this patch is a mechanical conversion to the new MatchTable object which is responsible for tracking label definitions and filling in the index of the jump targets. It is also responsible for nicely formatting the table. This was necessary to support the new GIM_Try opcode which takes the index to jump to if the match should fail. This value is unknown during table construction and is filled in during emission. To support nesting try-blocks (although we currently don't emit tables with nested try-blocks), GIM_Reject has been re-introduced to explicitly exit a try-block or fail the overall match if there are no active try-blocks. Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Reviewed By: rovka Subscribers: kristof.beyls, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D35117 llvm-svn: 308596
2017-07-20 17:25:44 +08:00
<< MatchTable::LineBreak;
[globalisel] Re-factor ISel matchers into a hierarchy. NFC Summary: This should make it possible to easily add everything needed to import all the existing SelectionDAG rules. It should also serve the likely kinds of GlobalISel rules (some of which are not currently representable in SelectionDAG) once we've nailed down the tablegen definition for that. The hierarchy is as follows: MatcherRule - A matching rule. Currently used to emit C++ ISel code but will | also be used to emit test cases and tablegen definitions in the | near future. |- Instruction(s) - Represents the instruction to be matched. |- Instruction Predicate(s) - Test the opcode, arithmetic flags, etc. of an | instruction. \- Operand(s) - Represents a particular operand of the instruction. In the | future, there may be subclasses to test the same predicates | on multiple operands (including for variadic instructions). \ Operand Predicate(s) - Test the type, register bank, etc. of an operand. This is where the ComplexPattern equivalent will be represented. It's also nested-instruction matching will live as a predicate that follows the DefUse chain to the Def and tests a MatcherRule from that position. Support for multiple instruction matchers in a rule has been retained from the existing code but has been adjusted to assert when it is used. Previously it would silently drop all but the first instruction matcher. The tablegen-erated file is not functionally changed but has more parentheses and no longer attempts to format the if-statements since keeping track of the indentation is tricky in the presence of the matcher hierarchy. It would be nice to have CMakes tablegen() run the output through clang-format (when available) so we don't have to complicate TableGen with pretty-printing. It's also worth mentioning that this hierarchy will also be able to emit TableGen definitions and test cases in the near future. This is the reason for favouring explicit emit*() calls rather than the << operator. Reviewers: aditya_nandakumar, rovka, t.p.northover, qcolombet, ab Reviewed By: ab Subscribers: igorb, dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D28942 llvm-svn: 293172
2017-01-26 19:10:14 +08:00
}
[globalisel] Sort RuleMatchers by priority. Summary: This makes more important rules have priority over less important rules. For example, '%a = G_ADD $b:s64, $c:s64' has priority over '%a = G_ADD $b:s32, $c:s32'. Previously these rules were emitted in the correct order by chance. NFC in this patch but it is required to make the next patch work correctly. Depends on D29710 Reviewers: t.p.northover, ab, qcolombet, aditya_nandakumar, rovka Reviewed By: ab, rovka Subscribers: javed.absar, dberris, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D29711 llvm-svn: 296121
2017-02-24 21:58:11 +08:00
/// Compare the priority of this object and B.
///
[globalisel] Decouple src pattern operands from dst pattern operands. Summary: This isn't testable for AArch64 by itself so this patch also adds support for constant immediates in the pattern and physical register uses in the result. The new IntOperandMatcher matches the constant in patterns such as '(set $rd:GPR32, (G_XOR $rs:GPR32, -1))'. It's always safe to fold immediates into an instruction so this is the first rule that will match across multiple BB's. The Renderer hierarchy is responsible for adding operands to the result instruction. Renderers can copy operands (CopyRenderer) or add physical registers (in particular %wzr and %xzr) to the result instruction in any order (OperandMatchers now import the operand names from SelectionDAG to allow renderers to access any operand). This allows us to emit the result instruction for: %1 = G_XOR %0, -1 --> %1 = ORNWrr %wzr, %0 %1 = G_XOR -1, %0 --> %1 = ORNWrr %wzr, %0 although the latter is untested since the matcher/importer has not been taught about commutativity yet. Added BuildMIAction which can build new instructions and mutate them where possible. W.r.t the mutation aspect, MatchActions are now told the name of an instruction they can recycle and BuildMIAction will emit mutation code when the renderers are appropriate. They are appropriate when all operands are rendered using CopyRenderer and the indices are the same as the matcher. This currently assumes that all operands have at least one matcher. Finally, this change also fixes a crash in AArch64InstructionSelector::select() caused by an immediate operand passing isImm() rather than isCImm(). This was uncovered by the other changes and was detected by existing tests. Depends on D29711 Reviewers: t.p.northover, ab, qcolombet, rovka, aditya_nandakumar, javed.absar Reviewed By: rovka Subscribers: aemerson, dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D29712 llvm-svn: 296131
2017-02-24 23:43:30 +08:00
/// Returns true if this object is more important than B.
[globalisel][tablegen] Multiple 80-col corrections. llvm-svn: 306544
2017-06-28 21:50:04 +08:00
bool
isHigherPriorityThan(const InstructionPredicateMatcher &B) const override {
[globalisel] Sort RuleMatchers by priority. Summary: This makes more important rules have priority over less important rules. For example, '%a = G_ADD $b:s64, $c:s64' has priority over '%a = G_ADD $b:s32, $c:s32'. Previously these rules were emitted in the correct order by chance. NFC in this patch but it is required to make the next patch work correctly. Depends on D29710 Reviewers: t.p.northover, ab, qcolombet, aditya_nandakumar, rovka Reviewed By: ab, rovka Subscribers: javed.absar, dberris, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D29711 llvm-svn: 296121
2017-02-24 21:58:11 +08:00
if (InstructionPredicateMatcher::isHigherPriorityThan(B))
return true;
if (B.InstructionPredicateMatcher::isHigherPriorityThan(*this))
return false;
// Prioritize opcodes for cosmetic reasons in the generated source. Although
// this is cosmetic at the moment, we may want to drive a similar ordering
// using instruction frequency information to improve compile time.
if (const InstructionOpcodeMatcher *BO =
dyn_cast<InstructionOpcodeMatcher>(&B))
return I->TheDef->getName() < BO->I->TheDef->getName();
return false;
};
[globalisel][tablegen] Add support for importing 'imm' operands. Summary: This patch enables the import of rules containing 'imm' operands that do not constrain the acceptable values using predicates. Support for ImmLeaf will arrive in a later patch. Depends on D35681 Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Reviewed By: rovka Subscribers: kristof.beyls, javed.absar, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D35833 llvm-svn: 310343
2017-08-08 18:44:31 +08:00
bool isConstantInstruction() const {
return I->TheDef->getName() == "G_CONSTANT";
}
Reapply r332907 "[GlobalISel] Improving InstructionSelect's performance by reducing MatchTable..." Apparently the compile time problem was caused by the fact that not all compilers / STL implementations can automatically convert std::unique_ptr<Derived> to std::unique_ptr<Base>. Fixed (hopefully) by making sure it's std::unique_ptr<Derived>&& (rvalue ref) to std::unique_ptr<Base> conversion instead. llvm-svn: 332917
2018-05-22 07:28:51 +08:00
[GlobalISel][InstructionSelect] Removing redundant num operands and nested def operands checks, perf patch 2 This patch continues a series of patches that decrease time spent by GlobalISel in its InstructionSelect pass by roughly 60% for -O0 builds for large inputs as measured on sqlite3-amalgamation (http://sqlite.org/download.html) targeting AArch64. This commit specifically removes number of operands checks that are redundant if the instruction's opcode already guarantees that number of operands (or more), and also avoids any kind of checks on a def operand of a nested instruction as everything about it was already checked at its use. The expected performance implication is about 3% off InstructionSelect comparing to the baseline (before the series of patches) This patch also contains a bit of NFC changes required for further patches in the series. Every commit planned shares the same Phabricator Review. Reviewers: qcolombet, dsanders, bogner, aemerson, javed.absar Reviewed By: qcolombet Subscribers: rovka, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D44700 llvm-svn: 332945
2018-05-22 12:31:50 +08:00
StringRef getOpcode() const { return I->TheDef->getName(); }
Reapply r332907 "[GlobalISel] Improving InstructionSelect's performance by reducing MatchTable..." Apparently the compile time problem was caused by the fact that not all compilers / STL implementations can automatically convert std::unique_ptr<Derived> to std::unique_ptr<Base>. Fixed (hopefully) by making sure it's std::unique_ptr<Derived>&& (rvalue ref) to std::unique_ptr<Base> conversion instead. llvm-svn: 332917
2018-05-22 07:28:51 +08:00
unsigned getNumOperands() const { return I->Operands.size(); }
StringRef getOperandType(unsigned OpIdx) const {
return I->Operands[OpIdx].OperandType;
}
[globalisel] Re-factor ISel matchers into a hierarchy. NFC Summary: This should make it possible to easily add everything needed to import all the existing SelectionDAG rules. It should also serve the likely kinds of GlobalISel rules (some of which are not currently representable in SelectionDAG) once we've nailed down the tablegen definition for that. The hierarchy is as follows: MatcherRule - A matching rule. Currently used to emit C++ ISel code but will | also be used to emit test cases and tablegen definitions in the | near future. |- Instruction(s) - Represents the instruction to be matched. |- Instruction Predicate(s) - Test the opcode, arithmetic flags, etc. of an | instruction. \- Operand(s) - Represents a particular operand of the instruction. In the | future, there may be subclasses to test the same predicates | on multiple operands (including for variadic instructions). \ Operand Predicate(s) - Test the type, register bank, etc. of an operand. This is where the ComplexPattern equivalent will be represented. It's also nested-instruction matching will live as a predicate that follows the DefUse chain to the Def and tests a MatcherRule from that position. Support for multiple instruction matchers in a rule has been retained from the existing code but has been adjusted to assert when it is used. Previously it would silently drop all but the first instruction matcher. The tablegen-erated file is not functionally changed but has more parentheses and no longer attempts to format the if-statements since keeping track of the indentation is tricky in the presence of the matcher hierarchy. It would be nice to have CMakes tablegen() run the output through clang-format (when available) so we don't have to complicate TableGen with pretty-printing. It's also worth mentioning that this hierarchy will also be able to emit TableGen definitions and test cases in the near future. This is the reason for favouring explicit emit*() calls rather than the << operator. Reviewers: aditya_nandakumar, rovka, t.p.northover, qcolombet, ab Reviewed By: ab Subscribers: igorb, dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D28942 llvm-svn: 293172
2017-01-26 19:10:14 +08:00
};
Reapply r332907 "[GlobalISel] Improving InstructionSelect's performance by reducing MatchTable..." Apparently the compile time problem was caused by the fact that not all compilers / STL implementations can automatically convert std::unique_ptr<Derived> to std::unique_ptr<Base>. Fixed (hopefully) by making sure it's std::unique_ptr<Derived>&& (rvalue ref) to std::unique_ptr<Base> conversion instead. llvm-svn: 332917
2018-05-22 07:28:51 +08:00
DenseMap<const CodeGenInstruction *, unsigned>
InstructionOpcodeMatcher::OpcodeValues;
[GlobalISel][InstructionSelect] Removing redundant num operands and nested def operands checks, perf patch 2 This patch continues a series of patches that decrease time spent by GlobalISel in its InstructionSelect pass by roughly 60% for -O0 builds for large inputs as measured on sqlite3-amalgamation (http://sqlite.org/download.html) targeting AArch64. This commit specifically removes number of operands checks that are redundant if the instruction's opcode already guarantees that number of operands (or more), and also avoids any kind of checks on a def operand of a nested instruction as everything about it was already checked at its use. The expected performance implication is about 3% off InstructionSelect comparing to the baseline (before the series of patches) This patch also contains a bit of NFC changes required for further patches in the series. Every commit planned shares the same Phabricator Review. Reviewers: qcolombet, dsanders, bogner, aemerson, javed.absar Reviewed By: qcolombet Subscribers: rovka, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D44700 llvm-svn: 332945
2018-05-22 12:31:50 +08:00
class InstructionNumOperandsMatcher final : public InstructionPredicateMatcher {
unsigned NumOperands = 0;
public:
InstructionNumOperandsMatcher(unsigned InsnVarID, unsigned NumOperands)
: InstructionPredicateMatcher(IPM_NumOperands, InsnVarID),
NumOperands(NumOperands) {}
static bool classof(const PredicateMatcher *P) {
return P->getKind() == IPM_NumOperands;
}
bool isIdentical(const PredicateMatcher &B) const override {
return InstructionPredicateMatcher::isIdentical(B) &&
NumOperands == cast<InstructionNumOperandsMatcher>(&B)->NumOperands;
}
void emitPredicateOpcodes(MatchTable &Table,
RuleMatcher &Rule) const override {
Table << MatchTable::Opcode("GIM_CheckNumOperands")
<< MatchTable::Comment("MI") << MatchTable::IntValue(InsnVarID)
<< MatchTable::Comment("Expected")
<< MatchTable::IntValue(NumOperands) << MatchTable::LineBreak;
}
};
Re-commit: [globalisel][tablegen] Add support for ImmLeaf without SDNodeXForm Summary: This patch adds support for predicates on imm nodes but only for ImmLeaf and not for PatLeaf or PatFrag and only where the value does not need to be transformed before being rendered into the instruction. The limitation on PatLeaf/PatFrag/SDNodeXForm is due to differences in the necessary target-supplied C++ for GlobalISel. Depends on D36085 The previous commit was reverted for breaking the build but this appears to have been the recurring problem on the Windows bots with tablegen not being re-run when llvm-tblgen is changed but the .td's aren't. If it re-occurs then forcing a build with clean=True should fix it but this string should do this in advance: Requires a clean build. Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Reviewed By: rovka Subscribers: kristof.beyls, javed.absar, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D36086 llvm-svn: 311645
2017-08-24 17:11:20 +08:00
/// Generates code to check that this instruction is a constant whose value
/// meets an immediate predicate.
///
/// Immediates are slightly odd since they are typically used like an operand
/// but are represented as an operator internally. We typically write simm8:$src
/// in a tablegen pattern, but this is just syntactic sugar for
/// (imm:i32)<<P:Predicate_simm8>>:$imm which more directly describes the nodes
/// that will be matched and the predicate (which is attached to the imm
/// operator) that will be tested. In SelectionDAG this describes a
/// ConstantSDNode whose internal value will be tested using the simm8 predicate.
///
/// The corresponding GlobalISel representation is %1 = G_CONSTANT iN Value. In
/// this representation, the immediate could be tested with an
/// InstructionMatcher, InstructionOpcodeMatcher, OperandMatcher, and a
/// OperandPredicateMatcher-subclass to check the Value meets the predicate but
/// there are two implementation issues with producing that matcher
/// configuration from the SelectionDAG pattern:
/// * ImmLeaf is a PatFrag whose root is an InstructionMatcher. This means that
/// were we to sink the immediate predicate to the operand we would have to
/// have two partial implementations of PatFrag support, one for immediates
/// and one for non-immediates.
/// * At the point we handle the predicate, the OperandMatcher hasn't been
/// created yet. If we were to sink the predicate to the OperandMatcher we
/// would also have to complicate (or duplicate) the code that descends and
/// creates matchers for the subtree.
/// Overall, it's simpler to handle it in the place it was found.
class InstructionImmPredicateMatcher : public InstructionPredicateMatcher {
protected:
TreePredicateFn Predicate;
public:
[TableGen][GlobalISel] Have the predicate directly know which data they are dealing with Prior to this patch, a predicate wouldn't make sense outside of its rule. Indeed, it was only during emitting a rule that a predicate would be made aware of the IDs of the data it is checking. Because of that, predicates could not be moved around or compared between each other. NFC. llvm-svn: 320887
2017-12-16 07:07:42 +08:00
InstructionImmPredicateMatcher(unsigned InsnVarID,
const TreePredicateFn &Predicate)
: InstructionPredicateMatcher(IPM_ImmPredicate, InsnVarID),
Predicate(Predicate) {}
Re-commit: [globalisel][tablegen] Add support for ImmLeaf without SDNodeXForm Summary: This patch adds support for predicates on imm nodes but only for ImmLeaf and not for PatLeaf or PatFrag and only where the value does not need to be transformed before being rendered into the instruction. The limitation on PatLeaf/PatFrag/SDNodeXForm is due to differences in the necessary target-supplied C++ for GlobalISel. Depends on D36085 The previous commit was reverted for breaking the build but this appears to have been the recurring problem on the Windows bots with tablegen not being re-run when llvm-tblgen is changed but the .td's aren't. If it re-occurs then forcing a build with clean=True should fix it but this string should do this in advance: Requires a clean build. Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Reviewed By: rovka Subscribers: kristof.beyls, javed.absar, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D36086 llvm-svn: 311645
2017-08-24 17:11:20 +08:00
[TableGen][GlobalISel] Make the different Matcher comparable This opens refactoring opportunities in the match table now that we can check that two predicates are the same. NFC. llvm-svn: 320890
2017-12-16 07:24:39 +08:00
bool isIdentical(const PredicateMatcher &B) const override {
return InstructionPredicateMatcher::isIdentical(B) &&
Predicate.getOrigPatFragRecord() ==
cast<InstructionImmPredicateMatcher>(&B)
->Predicate.getOrigPatFragRecord();
}
[TableGen][GlobalISel] Add a common class for all PredicateMatcher NFC. llvm-svn: 320767
2017-12-15 07:44:07 +08:00
static bool classof(const PredicateMatcher *P) {
Re-commit: [globalisel][tablegen] Add support for ImmLeaf without SDNodeXForm Summary: This patch adds support for predicates on imm nodes but only for ImmLeaf and not for PatLeaf or PatFrag and only where the value does not need to be transformed before being rendered into the instruction. The limitation on PatLeaf/PatFrag/SDNodeXForm is due to differences in the necessary target-supplied C++ for GlobalISel. Depends on D36085 The previous commit was reverted for breaking the build but this appears to have been the recurring problem on the Windows bots with tablegen not being re-run when llvm-tblgen is changed but the .td's aren't. If it re-occurs then forcing a build with clean=True should fix it but this string should do this in advance: Requires a clean build. Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Reviewed By: rovka Subscribers: kristof.beyls, javed.absar, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D36086 llvm-svn: 311645
2017-08-24 17:11:20 +08:00
return P->getKind() == IPM_ImmPredicate;
}
[TableGen][GlobalISel] Have the predicate directly know which data they are dealing with Prior to this patch, a predicate wouldn't make sense outside of its rule. Indeed, it was only during emitting a rule that a predicate would be made aware of the IDs of the data it is checking. Because of that, predicates could not be moved around or compared between each other. NFC. llvm-svn: 320887
2017-12-16 07:07:42 +08:00
void emitPredicateOpcodes(MatchTable &Table,
RuleMatcher &Rule) const override {
[globalisel][tablegen] Add support for fpimm and import of APInt/APFloat based ImmLeaf. Summary: There's only a tablegen testcase for IntImmLeaf and not a CodeGen one because the relevant rules are rejected for other reasons at the moment. On AArch64, it's because there's an SDNodeXForm attached to the operand. On X86, it's because the rule either emits multiple instructions or has another predicate using PatFrag which cannot easily be supported at the same time. Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Reviewed By: qcolombet Subscribers: aemerson, javed.absar, igorb, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D36569 llvm-svn: 315761
2017-10-14 05:28:03 +08:00
Table << MatchTable::Opcode(getMatchOpcodeForPredicate(Predicate))
Re-commit: [globalisel][tablegen] Add support for ImmLeaf without SDNodeXForm Summary: This patch adds support for predicates on imm nodes but only for ImmLeaf and not for PatLeaf or PatFrag and only where the value does not need to be transformed before being rendered into the instruction. The limitation on PatLeaf/PatFrag/SDNodeXForm is due to differences in the necessary target-supplied C++ for GlobalISel. Depends on D36085 The previous commit was reverted for breaking the build but this appears to have been the recurring problem on the Windows bots with tablegen not being re-run when llvm-tblgen is changed but the .td's aren't. If it re-occurs then forcing a build with clean=True should fix it but this string should do this in advance: Requires a clean build. Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Reviewed By: rovka Subscribers: kristof.beyls, javed.absar, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D36086 llvm-svn: 311645
2017-08-24 17:11:20 +08:00
<< MatchTable::Comment("MI") << MatchTable::IntValue(InsnVarID)
<< MatchTable::Comment("Predicate")
[globalisel][tablegen] Add support for fpimm and import of APInt/APFloat based ImmLeaf. Summary: There's only a tablegen testcase for IntImmLeaf and not a CodeGen one because the relevant rules are rejected for other reasons at the moment. On AArch64, it's because there's an SDNodeXForm attached to the operand. On X86, it's because the rule either emits multiple instructions or has another predicate using PatFrag which cannot easily be supported at the same time. Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Reviewed By: qcolombet Subscribers: aemerson, javed.absar, igorb, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D36569 llvm-svn: 315761
2017-10-14 05:28:03 +08:00
<< MatchTable::NamedValue(getEnumNameForPredicate(Predicate))
Re-commit: [globalisel][tablegen] Add support for ImmLeaf without SDNodeXForm Summary: This patch adds support for predicates on imm nodes but only for ImmLeaf and not for PatLeaf or PatFrag and only where the value does not need to be transformed before being rendered into the instruction. The limitation on PatLeaf/PatFrag/SDNodeXForm is due to differences in the necessary target-supplied C++ for GlobalISel. Depends on D36085 The previous commit was reverted for breaking the build but this appears to have been the recurring problem on the Windows bots with tablegen not being re-run when llvm-tblgen is changed but the .td's aren't. If it re-occurs then forcing a build with clean=True should fix it but this string should do this in advance: Requires a clean build. Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Reviewed By: rovka Subscribers: kristof.beyls, javed.absar, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D36086 llvm-svn: 311645
2017-08-24 17:11:20 +08:00
<< MatchTable::LineBreak;
}
};
[globalisel][tablegen] Add support for importing G_ATOMIC_CMPXCHG, G_ATOMICRMW_* rules from SelectionDAG. GIM_CheckNonAtomic has been replaced by GIM_CheckAtomicOrdering to allow it to support a wider range of orderings. This has then been used to import patterns using nodes such as atomic_cmp_swap, atomic_swap, and atomic_load_*. llvm-svn: 319232
2017-11-29 06:07:05 +08:00
/// Generates code to check that a memory instruction has a atomic ordering
/// MachineMemoryOperand.
class AtomicOrderingMMOPredicateMatcher : public InstructionPredicateMatcher {
[globalisel][tablegen] Add support for relative AtomicOrderings No test yet because the relevant rules are blocked on the atomic_load, and atomic_store nodes. llvm-svn: 319475
2017-12-01 05:05:59 +08:00
public:
enum AOComparator {
AO_Exactly,
AO_OrStronger,
AO_WeakerThan,
};
protected:
[globalisel][tablegen] Add support for importing G_ATOMIC_CMPXCHG, G_ATOMICRMW_* rules from SelectionDAG. GIM_CheckNonAtomic has been replaced by GIM_CheckAtomicOrdering to allow it to support a wider range of orderings. This has then been used to import patterns using nodes such as atomic_cmp_swap, atomic_swap, and atomic_load_*. llvm-svn: 319232
2017-11-29 06:07:05 +08:00
StringRef Order;
[globalisel][tablegen] Add support for relative AtomicOrderings No test yet because the relevant rules are blocked on the atomic_load, and atomic_store nodes. llvm-svn: 319475
2017-12-01 05:05:59 +08:00
AOComparator Comparator;
[globalisel][tablegen] Add support for importing G_ATOMIC_CMPXCHG, G_ATOMICRMW_* rules from SelectionDAG. GIM_CheckNonAtomic has been replaced by GIM_CheckAtomicOrdering to allow it to support a wider range of orderings. This has then been used to import patterns using nodes such as atomic_cmp_swap, atomic_swap, and atomic_load_*. llvm-svn: 319232
2017-11-29 06:07:05 +08:00
[globalisel][tablegen] Map ld and st to G_LOAD and G_STORE. NFC Summary: There is an important mismatch between ISD::LOAD and G_LOAD (and likewise for ISD::STORE and G_STORE). In SelectionDAG, ISD::LOAD is a non-atomic load and atomic loads are handled by a separate node. However, this is not true of GlobalISel's G_LOAD. For G_LOAD, the MachineMemOperand indicates the atomicity of the operation. As a result, this mapping must also add a predicate that checks for non-atomic MachineMemOperands. This is NFC since these nodes always have predicates in practice and are therefore always rejected at the moment. Depends on D37443 Reviewers: ab, qcolombet, t.p.northover, rovka, aditya_nandakumar Reviewed By: qcolombet Subscribers: kristof.beyls, llvm-commits, igorb Differential Revision: https://reviews.llvm.org/D37445 llvm-svn: 315843
2017-10-15 10:41:12 +08:00
public:
[TableGen][GlobalISel] Have the predicate directly know which data they are dealing with Prior to this patch, a predicate wouldn't make sense outside of its rule. Indeed, it was only during emitting a rule that a predicate would be made aware of the IDs of the data it is checking. Because of that, predicates could not be moved around or compared between each other. NFC. llvm-svn: 320887
2017-12-16 07:07:42 +08:00
AtomicOrderingMMOPredicateMatcher(unsigned InsnVarID, StringRef Order,
[globalisel][tablegen] Add support for relative AtomicOrderings No test yet because the relevant rules are blocked on the atomic_load, and atomic_store nodes. llvm-svn: 319475
2017-12-01 05:05:59 +08:00
AOComparator Comparator = AO_Exactly)
[TableGen][GlobalISel] Have the predicate directly know which data they are dealing with Prior to this patch, a predicate wouldn't make sense outside of its rule. Indeed, it was only during emitting a rule that a predicate would be made aware of the IDs of the data it is checking. Because of that, predicates could not be moved around or compared between each other. NFC. llvm-svn: 320887
2017-12-16 07:07:42 +08:00
: InstructionPredicateMatcher(IPM_AtomicOrderingMMO, InsnVarID),
Order(Order), Comparator(Comparator) {}
[globalisel][tablegen] Map ld and st to G_LOAD and G_STORE. NFC Summary: There is an important mismatch between ISD::LOAD and G_LOAD (and likewise for ISD::STORE and G_STORE). In SelectionDAG, ISD::LOAD is a non-atomic load and atomic loads are handled by a separate node. However, this is not true of GlobalISel's G_LOAD. For G_LOAD, the MachineMemOperand indicates the atomicity of the operation. As a result, this mapping must also add a predicate that checks for non-atomic MachineMemOperands. This is NFC since these nodes always have predicates in practice and are therefore always rejected at the moment. Depends on D37443 Reviewers: ab, qcolombet, t.p.northover, rovka, aditya_nandakumar Reviewed By: qcolombet Subscribers: kristof.beyls, llvm-commits, igorb Differential Revision: https://reviews.llvm.org/D37445 llvm-svn: 315843
2017-10-15 10:41:12 +08:00
Reapply r332907 "[GlobalISel] Improving InstructionSelect's performance by reducing MatchTable..." Apparently the compile time problem was caused by the fact that not all compilers / STL implementations can automatically convert std::unique_ptr<Derived> to std::unique_ptr<Base>. Fixed (hopefully) by making sure it's std::unique_ptr<Derived>&& (rvalue ref) to std::unique_ptr<Base> conversion instead. llvm-svn: 332917
2018-05-22 07:28:51 +08:00
static bool classof(const PredicateMatcher *P) {
[globalisel][tablegen] Add support for importing G_ATOMIC_CMPXCHG, G_ATOMICRMW_* rules from SelectionDAG. GIM_CheckNonAtomic has been replaced by GIM_CheckAtomicOrdering to allow it to support a wider range of orderings. This has then been used to import patterns using nodes such as atomic_cmp_swap, atomic_swap, and atomic_load_*. llvm-svn: 319232
2017-11-29 06:07:05 +08:00
return P->getKind() == IPM_AtomicOrderingMMO;
[globalisel][tablegen] Map ld and st to G_LOAD and G_STORE. NFC Summary: There is an important mismatch between ISD::LOAD and G_LOAD (and likewise for ISD::STORE and G_STORE). In SelectionDAG, ISD::LOAD is a non-atomic load and atomic loads are handled by a separate node. However, this is not true of GlobalISel's G_LOAD. For G_LOAD, the MachineMemOperand indicates the atomicity of the operation. As a result, this mapping must also add a predicate that checks for non-atomic MachineMemOperands. This is NFC since these nodes always have predicates in practice and are therefore always rejected at the moment. Depends on D37443 Reviewers: ab, qcolombet, t.p.northover, rovka, aditya_nandakumar Reviewed By: qcolombet Subscribers: kristof.beyls, llvm-commits, igorb Differential Revision: https://reviews.llvm.org/D37445 llvm-svn: 315843
2017-10-15 10:41:12 +08:00
}
Reapply r332907 "[GlobalISel] Improving InstructionSelect's performance by reducing MatchTable..." Apparently the compile time problem was caused by the fact that not all compilers / STL implementations can automatically convert std::unique_ptr<Derived> to std::unique_ptr<Base>. Fixed (hopefully) by making sure it's std::unique_ptr<Derived>&& (rvalue ref) to std::unique_ptr<Base> conversion instead. llvm-svn: 332917
2018-05-22 07:28:51 +08:00
bool isIdentical(const PredicateMatcher &B) const override {
if (!InstructionPredicateMatcher::isIdentical(B))
return false;
const auto &R = *cast<AtomicOrderingMMOPredicateMatcher>(&B);
return Order == R.Order && Comparator == R.Comparator;
}
[TableGen][GlobalISel] Have the predicate directly know which data they are dealing with Prior to this patch, a predicate wouldn't make sense outside of its rule. Indeed, it was only during emitting a rule that a predicate would be made aware of the IDs of the data it is checking. Because of that, predicates could not be moved around or compared between each other. NFC. llvm-svn: 320887
2017-12-16 07:07:42 +08:00
void emitPredicateOpcodes(MatchTable &Table,
RuleMatcher &Rule) const override {
[globalisel][tablegen] Add support for relative AtomicOrderings No test yet because the relevant rules are blocked on the atomic_load, and atomic_store nodes. llvm-svn: 319475
2017-12-01 05:05:59 +08:00
StringRef Opcode = "GIM_CheckAtomicOrdering";
if (Comparator == AO_OrStronger)
Opcode = "GIM_CheckAtomicOrderingOrStrongerThan";
if (Comparator == AO_WeakerThan)
Opcode = "GIM_CheckAtomicOrderingWeakerThan";
Table << MatchTable::Opcode(Opcode) << MatchTable::Comment("MI")
<< MatchTable::IntValue(InsnVarID) << MatchTable::Comment("Order")
[globalisel][tablegen] Add support for importing G_ATOMIC_CMPXCHG, G_ATOMICRMW_* rules from SelectionDAG. GIM_CheckNonAtomic has been replaced by GIM_CheckAtomicOrdering to allow it to support a wider range of orderings. This has then been used to import patterns using nodes such as atomic_cmp_swap, atomic_swap, and atomic_load_*. llvm-svn: 319232
2017-11-29 06:07:05 +08:00
<< MatchTable::NamedValue(("(int64_t)AtomicOrdering::" + Order).str())
[globalisel][tablegen] Map ld and st to G_LOAD and G_STORE. NFC Summary: There is an important mismatch between ISD::LOAD and G_LOAD (and likewise for ISD::STORE and G_STORE). In SelectionDAG, ISD::LOAD is a non-atomic load and atomic loads are handled by a separate node. However, this is not true of GlobalISel's G_LOAD. For G_LOAD, the MachineMemOperand indicates the atomicity of the operation. As a result, this mapping must also add a predicate that checks for non-atomic MachineMemOperands. This is NFC since these nodes always have predicates in practice and are therefore always rejected at the moment. Depends on D37443 Reviewers: ab, qcolombet, t.p.northover, rovka, aditya_nandakumar Reviewed By: qcolombet Subscribers: kristof.beyls, llvm-commits, igorb Differential Revision: https://reviews.llvm.org/D37445 llvm-svn: 315843
2017-10-15 10:41:12 +08:00
<< MatchTable::LineBreak;
}
};
[globalisel] Update GlobalISel emitter to match new representation of extending loads Summary: Previously, a extending load was represented at (G_*EXT (G_LOAD x)). This had a few drawbacks: * G_LOAD had to be legal for all sizes you could extend from, even if registers didn't naturally hold those sizes. * All sizes you could extend from had to be allocatable just in case the extend went missing (e.g. by optimization). * At minimum, G_*EXT and G_TRUNC had to be legal for these sizes. As we improve optimization of extends and truncates, this legality requirement would spread without considerable care w.r.t when certain combines were permitted. * The SelectionDAG importer required some ugly and fragile pattern rewriting to translate patterns into this style. This patch changes the representation to: * (G_[SZ]EXTLOAD x) * (G_LOAD x) any-extends when MMO.getSize() * 8 < ResultTy.getSizeInBits() which resolves these issues by allowing targets to work entirely in their native register sizes, and by having a more direct translation from SelectionDAG patterns. Each extending load can be lowered by the legalizer into separate extends and loads, however a target that supports s1 will need the any-extending load to extend to at least s8 since LLVM does not represent memory accesses smaller than 8 bit. The legalizer can widenScalar G_LOAD into an any-extending load but sign/zero-extending loads need help from something else like a combiner pass. A follow-up patch that adds combiner helpers for for this will follow. The new representation requires that the MMO correctly reflect the memory access so this has been corrected in a couple tests. I've also moved the extending loads to their own tests since they are (mostly) separate opcodes now. Additionally, the re-write appears to have invalidated two tests from select-with-no-legality-check.mir since the matcher table no longer contains loads that result in s1's and they aren't legal in AArch64 anymore. Depends on D45540 Reviewers: ab, aditya_nandakumar, bogner, rtereshin, volkan, rovka, javed.absar Reviewed By: rtereshin Subscribers: javed.absar, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D45541 llvm-svn: 331601
2018-05-06 04:53:24 +08:00
/// Generates code to check that the size of an MMO is exactly N bytes.
class MemorySizePredicateMatcher : public InstructionPredicateMatcher {
protected:
unsigned MMOIdx;
uint64_t Size;
public:
MemorySizePredicateMatcher(unsigned InsnVarID, unsigned MMOIdx, unsigned Size)
: InstructionPredicateMatcher(IPM_MemoryLLTSize, InsnVarID),
MMOIdx(MMOIdx), Size(Size) {}
static bool classof(const PredicateMatcher *P) {
return P->getKind() == IPM_MemoryLLTSize;
}
bool isIdentical(const PredicateMatcher &B) const override {
return InstructionPredicateMatcher::isIdentical(B) &&
MMOIdx == cast<MemorySizePredicateMatcher>(&B)->MMOIdx &&
Size == cast<MemorySizePredicateMatcher>(&B)->Size;
}
void emitPredicateOpcodes(MatchTable &Table,
RuleMatcher &Rule) const override {
Table << MatchTable::Opcode("GIM_CheckMemorySizeEqualTo")
<< MatchTable::Comment("MI") << MatchTable::IntValue(InsnVarID)
<< MatchTable::Comment("MMO") << MatchTable::IntValue(MMOIdx)
<< MatchTable::Comment("Size") << MatchTable::IntValue(Size)
<< MatchTable::LineBreak;
}
};
/// Generates code to check that the size of an MMO is less-than, equal-to, or
/// greater than a given LLT.
class MemoryVsLLTSizePredicateMatcher : public InstructionPredicateMatcher {
public:
enum RelationKind {
GreaterThan,
EqualTo,
LessThan,
};
protected:
unsigned MMOIdx;
RelationKind Relation;
unsigned OpIdx;
public:
MemoryVsLLTSizePredicateMatcher(unsigned InsnVarID, unsigned MMOIdx,
enum RelationKind Relation,
unsigned OpIdx)
: InstructionPredicateMatcher(IPM_MemoryVsLLTSize, InsnVarID),
MMOIdx(MMOIdx), Relation(Relation), OpIdx(OpIdx) {}
static bool classof(const PredicateMatcher *P) {
return P->getKind() == IPM_MemoryVsLLTSize;
}
bool isIdentical(const PredicateMatcher &B) const override {
return InstructionPredicateMatcher::isIdentical(B) &&
MMOIdx == cast<MemoryVsLLTSizePredicateMatcher>(&B)->MMOIdx &&
Relation == cast<MemoryVsLLTSizePredicateMatcher>(&B)->Relation &&
OpIdx == cast<MemoryVsLLTSizePredicateMatcher>(&B)->OpIdx;
}
void emitPredicateOpcodes(MatchTable &Table,
RuleMatcher &Rule) const override {
Table << MatchTable::Opcode(Relation == EqualTo
? "GIM_CheckMemorySizeEqualToLLT"
: Relation == GreaterThan
? "GIM_CheckMemorySizeGreaterThanLLT"
: "GIM_CheckMemorySizeLessThanLLT")
<< MatchTable::Comment("MI") << MatchTable::IntValue(InsnVarID)
<< MatchTable::Comment("MMO") << MatchTable::IntValue(MMOIdx)
<< MatchTable::Comment("OpIdx") << MatchTable::IntValue(OpIdx)
<< MatchTable::LineBreak;
}
};
[globalisel] Re-factor ISel matchers into a hierarchy. NFC Summary: This should make it possible to easily add everything needed to import all the existing SelectionDAG rules. It should also serve the likely kinds of GlobalISel rules (some of which are not currently representable in SelectionDAG) once we've nailed down the tablegen definition for that. The hierarchy is as follows: MatcherRule - A matching rule. Currently used to emit C++ ISel code but will | also be used to emit test cases and tablegen definitions in the | near future. |- Instruction(s) - Represents the instruction to be matched. |- Instruction Predicate(s) - Test the opcode, arithmetic flags, etc. of an | instruction. \- Operand(s) - Represents a particular operand of the instruction. In the | future, there may be subclasses to test the same predicates | on multiple operands (including for variadic instructions). \ Operand Predicate(s) - Test the type, register bank, etc. of an operand. This is where the ComplexPattern equivalent will be represented. It's also nested-instruction matching will live as a predicate that follows the DefUse chain to the Def and tests a MatcherRule from that position. Support for multiple instruction matchers in a rule has been retained from the existing code but has been adjusted to assert when it is used. Previously it would silently drop all but the first instruction matcher. The tablegen-erated file is not functionally changed but has more parentheses and no longer attempts to format the if-statements since keeping track of the indentation is tricky in the presence of the matcher hierarchy. It would be nice to have CMakes tablegen() run the output through clang-format (when available) so we don't have to complicate TableGen with pretty-printing. It's also worth mentioning that this hierarchy will also be able to emit TableGen definitions and test cases in the near future. This is the reason for favouring explicit emit*() calls rather than the << operator. Reviewers: aditya_nandakumar, rovka, t.p.northover, qcolombet, ab Reviewed By: ab Subscribers: igorb, dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D28942 llvm-svn: 293172
2017-01-26 19:10:14 +08:00
/// Generates code to check that a set of predicates and operands match for a
/// particular instruction.
///
/// Typical predicates include:
/// * Has a specific opcode.
/// * Has an nsw/nuw flag or doesn't.
Reapply r332907 "[GlobalISel] Improving InstructionSelect's performance by reducing MatchTable..." Apparently the compile time problem was caused by the fact that not all compilers / STL implementations can automatically convert std::unique_ptr<Derived> to std::unique_ptr<Base>. Fixed (hopefully) by making sure it's std::unique_ptr<Derived>&& (rvalue ref) to std::unique_ptr<Base> conversion instead. llvm-svn: 332917
2018-05-22 07:28:51 +08:00
class InstructionMatcher final : public PredicateListMatcher<PredicateMatcher> {
[globalisel] Re-factor ISel matchers into a hierarchy. NFC Summary: This should make it possible to easily add everything needed to import all the existing SelectionDAG rules. It should also serve the likely kinds of GlobalISel rules (some of which are not currently representable in SelectionDAG) once we've nailed down the tablegen definition for that. The hierarchy is as follows: MatcherRule - A matching rule. Currently used to emit C++ ISel code but will | also be used to emit test cases and tablegen definitions in the | near future. |- Instruction(s) - Represents the instruction to be matched. |- Instruction Predicate(s) - Test the opcode, arithmetic flags, etc. of an | instruction. \- Operand(s) - Represents a particular operand of the instruction. In the | future, there may be subclasses to test the same predicates | on multiple operands (including for variadic instructions). \ Operand Predicate(s) - Test the type, register bank, etc. of an operand. This is where the ComplexPattern equivalent will be represented. It's also nested-instruction matching will live as a predicate that follows the DefUse chain to the Def and tests a MatcherRule from that position. Support for multiple instruction matchers in a rule has been retained from the existing code but has been adjusted to assert when it is used. Previously it would silently drop all but the first instruction matcher. The tablegen-erated file is not functionally changed but has more parentheses and no longer attempts to format the if-statements since keeping track of the indentation is tricky in the presence of the matcher hierarchy. It would be nice to have CMakes tablegen() run the output through clang-format (when available) so we don't have to complicate TableGen with pretty-printing. It's also worth mentioning that this hierarchy will also be able to emit TableGen definitions and test cases in the near future. This is the reason for favouring explicit emit*() calls rather than the << operator. Reviewers: aditya_nandakumar, rovka, t.p.northover, qcolombet, ab Reviewed By: ab Subscribers: igorb, dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D28942 llvm-svn: 293172
2017-01-26 19:10:14 +08:00
protected:
[globalisel][tablegen] Fix patterns involving multiple ComplexPatterns. Summary: Temporaries are now allocated to operands instead of predicates and this allocation is used to correctly pair up the rendered operands with the matched operands. Previously, ComplexPatterns were allocated temporaries independently in the Src Pattern and Dst Pattern, leading to mismatches. Additionally, the Dst Pattern failed to account for the allocated index and therefore always used temporary 0, 1, ... when it should have used base+0, base+1, ... Thanks to Aditya Nandakumar for noticing the bug. Depends on D30539 Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Reviewed By: rovka Subscribers: igorb, dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D31054 llvm-svn: 299538
2017-04-05 21:14:03 +08:00
typedef std::vector<std::unique_ptr<OperandMatcher>> OperandVec;
[globalisel] Decouple src pattern operands from dst pattern operands. Summary: This isn't testable for AArch64 by itself so this patch also adds support for constant immediates in the pattern and physical register uses in the result. The new IntOperandMatcher matches the constant in patterns such as '(set $rd:GPR32, (G_XOR $rs:GPR32, -1))'. It's always safe to fold immediates into an instruction so this is the first rule that will match across multiple BB's. The Renderer hierarchy is responsible for adding operands to the result instruction. Renderers can copy operands (CopyRenderer) or add physical registers (in particular %wzr and %xzr) to the result instruction in any order (OperandMatchers now import the operand names from SelectionDAG to allow renderers to access any operand). This allows us to emit the result instruction for: %1 = G_XOR %0, -1 --> %1 = ORNWrr %wzr, %0 %1 = G_XOR -1, %0 --> %1 = ORNWrr %wzr, %0 although the latter is untested since the matcher/importer has not been taught about commutativity yet. Added BuildMIAction which can build new instructions and mutate them where possible. W.r.t the mutation aspect, MatchActions are now told the name of an instruction they can recycle and BuildMIAction will emit mutation code when the renderers are appropriate. They are appropriate when all operands are rendered using CopyRenderer and the indices are the same as the matcher. This currently assumes that all operands have at least one matcher. Finally, this change also fixes a crash in AArch64InstructionSelector::select() caused by an immediate operand passing isImm() rather than isCImm(). This was uncovered by the other changes and was detected by existing tests. Depends on D29711 Reviewers: t.p.northover, ab, qcolombet, rovka, aditya_nandakumar, javed.absar Reviewed By: rovka Subscribers: aemerson, dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D29712 llvm-svn: 296131
2017-02-24 23:43:30 +08:00
[globalisel][tablegen] Simplify named operand/operator lookups and fix a wrong-code bug this revealed. Summary: Operand variable lookups are now performed by the RuleMatcher rather than searching the whole matcher hierarchy for a match. This revealed a wrong-code bug that currently affects ARM and X86 where patterns that use a variable more than once in the match pattern will be imported but won't check that the operands are identical. This can cause the tablegen-erated matcher to accept matches that should be rejected. Depends on D36569 Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Subscribers: aemerson, igorb, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D36618 llvm-svn: 315780
2017-10-14 08:31:58 +08:00
RuleMatcher &Rule;
[globalisel] Decouple src pattern operands from dst pattern operands. Summary: This isn't testable for AArch64 by itself so this patch also adds support for constant immediates in the pattern and physical register uses in the result. The new IntOperandMatcher matches the constant in patterns such as '(set $rd:GPR32, (G_XOR $rs:GPR32, -1))'. It's always safe to fold immediates into an instruction so this is the first rule that will match across multiple BB's. The Renderer hierarchy is responsible for adding operands to the result instruction. Renderers can copy operands (CopyRenderer) or add physical registers (in particular %wzr and %xzr) to the result instruction in any order (OperandMatchers now import the operand names from SelectionDAG to allow renderers to access any operand). This allows us to emit the result instruction for: %1 = G_XOR %0, -1 --> %1 = ORNWrr %wzr, %0 %1 = G_XOR -1, %0 --> %1 = ORNWrr %wzr, %0 although the latter is untested since the matcher/importer has not been taught about commutativity yet. Added BuildMIAction which can build new instructions and mutate them where possible. W.r.t the mutation aspect, MatchActions are now told the name of an instruction they can recycle and BuildMIAction will emit mutation code when the renderers are appropriate. They are appropriate when all operands are rendered using CopyRenderer and the indices are the same as the matcher. This currently assumes that all operands have at least one matcher. Finally, this change also fixes a crash in AArch64InstructionSelector::select() caused by an immediate operand passing isImm() rather than isCImm(). This was uncovered by the other changes and was detected by existing tests. Depends on D29711 Reviewers: t.p.northover, ab, qcolombet, rovka, aditya_nandakumar, javed.absar Reviewed By: rovka Subscribers: aemerson, dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D29712 llvm-svn: 296131
2017-02-24 23:43:30 +08:00
/// The operands to match. All rendered operands must be present even if the
/// condition is always true.
OperandVec Operands;
[GlobalISel][InstructionSelect] Removing redundant num operands and nested def operands checks, perf patch 2 This patch continues a series of patches that decrease time spent by GlobalISel in its InstructionSelect pass by roughly 60% for -O0 builds for large inputs as measured on sqlite3-amalgamation (http://sqlite.org/download.html) targeting AArch64. This commit specifically removes number of operands checks that are redundant if the instruction's opcode already guarantees that number of operands (or more), and also avoids any kind of checks on a def operand of a nested instruction as everything about it was already checked at its use. The expected performance implication is about 3% off InstructionSelect comparing to the baseline (before the series of patches) This patch also contains a bit of NFC changes required for further patches in the series. Every commit planned shares the same Phabricator Review. Reviewers: qcolombet, dsanders, bogner, aemerson, javed.absar Reviewed By: qcolombet Subscribers: rovka, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D44700 llvm-svn: 332945
2018-05-22 12:31:50 +08:00
bool NumOperandsCheck = true;
[globalisel] Re-factor ISel matchers into a hierarchy. NFC Summary: This should make it possible to easily add everything needed to import all the existing SelectionDAG rules. It should also serve the likely kinds of GlobalISel rules (some of which are not currently representable in SelectionDAG) once we've nailed down the tablegen definition for that. The hierarchy is as follows: MatcherRule - A matching rule. Currently used to emit C++ ISel code but will | also be used to emit test cases and tablegen definitions in the | near future. |- Instruction(s) - Represents the instruction to be matched. |- Instruction Predicate(s) - Test the opcode, arithmetic flags, etc. of an | instruction. \- Operand(s) - Represents a particular operand of the instruction. In the | future, there may be subclasses to test the same predicates | on multiple operands (including for variadic instructions). \ Operand Predicate(s) - Test the type, register bank, etc. of an operand. This is where the ComplexPattern equivalent will be represented. It's also nested-instruction matching will live as a predicate that follows the DefUse chain to the Def and tests a MatcherRule from that position. Support for multiple instruction matchers in a rule has been retained from the existing code but has been adjusted to assert when it is used. Previously it would silently drop all but the first instruction matcher. The tablegen-erated file is not functionally changed but has more parentheses and no longer attempts to format the if-statements since keeping track of the indentation is tricky in the presence of the matcher hierarchy. It would be nice to have CMakes tablegen() run the output through clang-format (when available) so we don't have to complicate TableGen with pretty-printing. It's also worth mentioning that this hierarchy will also be able to emit TableGen definitions and test cases in the near future. This is the reason for favouring explicit emit*() calls rather than the << operator. Reviewers: aditya_nandakumar, rovka, t.p.northover, qcolombet, ab Reviewed By: ab Subscribers: igorb, dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D28942 llvm-svn: 293172
2017-01-26 19:10:14 +08:00
[globalisel][tablegen] Add support for importing 'imm' operands. Summary: This patch enables the import of rules containing 'imm' operands that do not constrain the acceptable values using predicates. Support for ImmLeaf will arrive in a later patch. Depends on D35681 Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Reviewed By: rovka Subscribers: kristof.beyls, javed.absar, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D35833 llvm-svn: 310343
2017-08-08 18:44:31 +08:00
std::string SymbolicName;
[TableGen][GlobalISel] Have the predicate directly know which data they are dealing with Prior to this patch, a predicate wouldn't make sense outside of its rule. Indeed, it was only during emitting a rule that a predicate would be made aware of the IDs of the data it is checking. Because of that, predicates could not be moved around or compared between each other. NFC. llvm-svn: 320887
2017-12-16 07:07:42 +08:00
unsigned InsnVarID;
[globalisel][tablegen] Add support for importing 'imm' operands. Summary: This patch enables the import of rules containing 'imm' operands that do not constrain the acceptable values using predicates. Support for ImmLeaf will arrive in a later patch. Depends on D35681 Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Reviewed By: rovka Subscribers: kristof.beyls, javed.absar, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D35833 llvm-svn: 310343
2017-08-08 18:44:31 +08:00
[globalisel] Re-factor ISel matchers into a hierarchy. NFC Summary: This should make it possible to easily add everything needed to import all the existing SelectionDAG rules. It should also serve the likely kinds of GlobalISel rules (some of which are not currently representable in SelectionDAG) once we've nailed down the tablegen definition for that. The hierarchy is as follows: MatcherRule - A matching rule. Currently used to emit C++ ISel code but will | also be used to emit test cases and tablegen definitions in the | near future. |- Instruction(s) - Represents the instruction to be matched. |- Instruction Predicate(s) - Test the opcode, arithmetic flags, etc. of an | instruction. \- Operand(s) - Represents a particular operand of the instruction. In the | future, there may be subclasses to test the same predicates | on multiple operands (including for variadic instructions). \ Operand Predicate(s) - Test the type, register bank, etc. of an operand. This is where the ComplexPattern equivalent will be represented. It's also nested-instruction matching will live as a predicate that follows the DefUse chain to the Def and tests a MatcherRule from that position. Support for multiple instruction matchers in a rule has been retained from the existing code but has been adjusted to assert when it is used. Previously it would silently drop all but the first instruction matcher. The tablegen-erated file is not functionally changed but has more parentheses and no longer attempts to format the if-statements since keeping track of the indentation is tricky in the presence of the matcher hierarchy. It would be nice to have CMakes tablegen() run the output through clang-format (when available) so we don't have to complicate TableGen with pretty-printing. It's also worth mentioning that this hierarchy will also be able to emit TableGen definitions and test cases in the near future. This is the reason for favouring explicit emit*() calls rather than the << operator. Reviewers: aditya_nandakumar, rovka, t.p.northover, qcolombet, ab Reviewed By: ab Subscribers: igorb, dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D28942 llvm-svn: 293172
2017-01-26 19:10:14 +08:00
public:
[globalisel][tablegen] Simplify named operand/operator lookups and fix a wrong-code bug this revealed. Summary: Operand variable lookups are now performed by the RuleMatcher rather than searching the whole matcher hierarchy for a match. This revealed a wrong-code bug that currently affects ARM and X86 where patterns that use a variable more than once in the match pattern will be imported but won't check that the operands are identical. This can cause the tablegen-erated matcher to accept matches that should be rejected. Depends on D36569 Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Subscribers: aemerson, igorb, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D36618 llvm-svn: 315780
2017-10-14 08:31:58 +08:00
InstructionMatcher(RuleMatcher &Rule, StringRef SymbolicName)
[TableGen][GlobalISel] Have the predicate directly know which data they are dealing with Prior to this patch, a predicate wouldn't make sense outside of its rule. Indeed, it was only during emitting a rule that a predicate would be made aware of the IDs of the data it is checking. Because of that, predicates could not be moved around or compared between each other. NFC. llvm-svn: 320887
2017-12-16 07:07:42 +08:00
: Rule(Rule), SymbolicName(SymbolicName) {
// We create a new instruction matcher.
// Get a new ID for that instruction.
InsnVarID = Rule.implicitlyDefineInsnVar(*this);
}
[globalisel][tablegen] Simplify named operand/operator lookups and fix a wrong-code bug this revealed. Summary: Operand variable lookups are now performed by the RuleMatcher rather than searching the whole matcher hierarchy for a match. This revealed a wrong-code bug that currently affects ARM and X86 where patterns that use a variable more than once in the match pattern will be imported but won't check that the operands are identical. This can cause the tablegen-erated matcher to accept matches that should be rejected. Depends on D36569 Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Subscribers: aemerson, igorb, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D36618 llvm-svn: 315780
2017-10-14 08:31:58 +08:00
Reapply r332907 "[GlobalISel] Improving InstructionSelect's performance by reducing MatchTable..." Apparently the compile time problem was caused by the fact that not all compilers / STL implementations can automatically convert std::unique_ptr<Derived> to std::unique_ptr<Base>. Fixed (hopefully) by making sure it's std::unique_ptr<Derived>&& (rvalue ref) to std::unique_ptr<Base> conversion instead. llvm-svn: 332917
2018-05-22 07:28:51 +08:00
/// Construct a new instruction predicate and add it to the matcher.
template <class Kind, class... Args>
Optional<Kind *> addPredicate(Args &&... args) {
Predicates.emplace_back(
llvm::make_unique<Kind>(getInsnVarID(), std::forward<Args>(args)...));
return static_cast<Kind *>(Predicates.back().get());
}
[globalisel][tablegen] Simplify named operand/operator lookups and fix a wrong-code bug this revealed. Summary: Operand variable lookups are now performed by the RuleMatcher rather than searching the whole matcher hierarchy for a match. This revealed a wrong-code bug that currently affects ARM and X86 where patterns that use a variable more than once in the match pattern will be imported but won't check that the operands are identical. This can cause the tablegen-erated matcher to accept matches that should be rejected. Depends on D36569 Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Subscribers: aemerson, igorb, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D36618 llvm-svn: 315780
2017-10-14 08:31:58 +08:00
RuleMatcher &getRuleMatcher() const { return Rule; }
[globalisel][tablegen] Add support for importing 'imm' operands. Summary: This patch enables the import of rules containing 'imm' operands that do not constrain the acceptable values using predicates. Support for ImmLeaf will arrive in a later patch. Depends on D35681 Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Reviewed By: rovka Subscribers: kristof.beyls, javed.absar, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D35833 llvm-svn: 310343
2017-08-08 18:44:31 +08:00
Reapply r332907 "[GlobalISel] Improving InstructionSelect's performance by reducing MatchTable..." Apparently the compile time problem was caused by the fact that not all compilers / STL implementations can automatically convert std::unique_ptr<Derived> to std::unique_ptr<Base>. Fixed (hopefully) by making sure it's std::unique_ptr<Derived>&& (rvalue ref) to std::unique_ptr<Base> conversion instead. llvm-svn: 332917
2018-05-22 07:28:51 +08:00
unsigned getInsnVarID() const { return InsnVarID; }
[TableGen][GlobalISel] Have the predicate directly know which data they are dealing with Prior to this patch, a predicate wouldn't make sense outside of its rule. Indeed, it was only during emitting a rule that a predicate would be made aware of the IDs of the data it is checking. Because of that, predicates could not be moved around or compared between each other. NFC. llvm-svn: 320887
2017-12-16 07:07:42 +08:00
[globalisel] Re-factor ISel matchers into a hierarchy. NFC Summary: This should make it possible to easily add everything needed to import all the existing SelectionDAG rules. It should also serve the likely kinds of GlobalISel rules (some of which are not currently representable in SelectionDAG) once we've nailed down the tablegen definition for that. The hierarchy is as follows: MatcherRule - A matching rule. Currently used to emit C++ ISel code but will | also be used to emit test cases and tablegen definitions in the | near future. |- Instruction(s) - Represents the instruction to be matched. |- Instruction Predicate(s) - Test the opcode, arithmetic flags, etc. of an | instruction. \- Operand(s) - Represents a particular operand of the instruction. In the | future, there may be subclasses to test the same predicates | on multiple operands (including for variadic instructions). \ Operand Predicate(s) - Test the type, register bank, etc. of an operand. This is where the ComplexPattern equivalent will be represented. It's also nested-instruction matching will live as a predicate that follows the DefUse chain to the Def and tests a MatcherRule from that position. Support for multiple instruction matchers in a rule has been retained from the existing code but has been adjusted to assert when it is used. Previously it would silently drop all but the first instruction matcher. The tablegen-erated file is not functionally changed but has more parentheses and no longer attempts to format the if-statements since keeping track of the indentation is tricky in the presence of the matcher hierarchy. It would be nice to have CMakes tablegen() run the output through clang-format (when available) so we don't have to complicate TableGen with pretty-printing. It's also worth mentioning that this hierarchy will also be able to emit TableGen definitions and test cases in the near future. This is the reason for favouring explicit emit*() calls rather than the << operator. Reviewers: aditya_nandakumar, rovka, t.p.northover, qcolombet, ab Reviewed By: ab Subscribers: igorb, dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D28942 llvm-svn: 293172
2017-01-26 19:10:14 +08:00
/// Add an operand to the matcher.
[globalisel][tablegen] Fix patterns involving multiple ComplexPatterns. Summary: Temporaries are now allocated to operands instead of predicates and this allocation is used to correctly pair up the rendered operands with the matched operands. Previously, ComplexPatterns were allocated temporaries independently in the Src Pattern and Dst Pattern, leading to mismatches. Additionally, the Dst Pattern failed to account for the allocated index and therefore always used temporary 0, 1, ... when it should have used base+0, base+1, ... Thanks to Aditya Nandakumar for noticing the bug. Depends on D30539 Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Reviewed By: rovka Subscribers: igorb, dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D31054 llvm-svn: 299538
2017-04-05 21:14:03 +08:00
OperandMatcher &addOperand(unsigned OpIdx, const std::string &SymbolicName,
unsigned AllocatedTemporariesBaseID) {
Operands.emplace_back(new OperandMatcher(*this, OpIdx, SymbolicName,
AllocatedTemporariesBaseID));
[globalisel][tablegen] Simplify named operand/operator lookups and fix a wrong-code bug this revealed. Summary: Operand variable lookups are now performed by the RuleMatcher rather than searching the whole matcher hierarchy for a match. This revealed a wrong-code bug that currently affects ARM and X86 where patterns that use a variable more than once in the match pattern will be imported but won't check that the operands are identical. This can cause the tablegen-erated matcher to accept matches that should be rejected. Depends on D36569 Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Subscribers: aemerson, igorb, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D36618 llvm-svn: 315780
2017-10-14 08:31:58 +08:00
if (!SymbolicName.empty())
Rule.defineOperand(SymbolicName, *Operands.back());
[globalisel][tablegen] Fix patterns involving multiple ComplexPatterns. Summary: Temporaries are now allocated to operands instead of predicates and this allocation is used to correctly pair up the rendered operands with the matched operands. Previously, ComplexPatterns were allocated temporaries independently in the Src Pattern and Dst Pattern, leading to mismatches. Additionally, the Dst Pattern failed to account for the allocated index and therefore always used temporary 0, 1, ... when it should have used base+0, base+1, ... Thanks to Aditya Nandakumar for noticing the bug. Depends on D30539 Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Reviewed By: rovka Subscribers: igorb, dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D31054 llvm-svn: 299538
2017-04-05 21:14:03 +08:00
return *Operands.back();
[globalisel] Re-factor ISel matchers into a hierarchy. NFC Summary: This should make it possible to easily add everything needed to import all the existing SelectionDAG rules. It should also serve the likely kinds of GlobalISel rules (some of which are not currently representable in SelectionDAG) once we've nailed down the tablegen definition for that. The hierarchy is as follows: MatcherRule - A matching rule. Currently used to emit C++ ISel code but will | also be used to emit test cases and tablegen definitions in the | near future. |- Instruction(s) - Represents the instruction to be matched. |- Instruction Predicate(s) - Test the opcode, arithmetic flags, etc. of an | instruction. \- Operand(s) - Represents a particular operand of the instruction. In the | future, there may be subclasses to test the same predicates | on multiple operands (including for variadic instructions). \ Operand Predicate(s) - Test the type, register bank, etc. of an operand. This is where the ComplexPattern equivalent will be represented. It's also nested-instruction matching will live as a predicate that follows the DefUse chain to the Def and tests a MatcherRule from that position. Support for multiple instruction matchers in a rule has been retained from the existing code but has been adjusted to assert when it is used. Previously it would silently drop all but the first instruction matcher. The tablegen-erated file is not functionally changed but has more parentheses and no longer attempts to format the if-statements since keeping track of the indentation is tricky in the presence of the matcher hierarchy. It would be nice to have CMakes tablegen() run the output through clang-format (when available) so we don't have to complicate TableGen with pretty-printing. It's also worth mentioning that this hierarchy will also be able to emit TableGen definitions and test cases in the near future. This is the reason for favouring explicit emit*() calls rather than the << operator. Reviewers: aditya_nandakumar, rovka, t.p.northover, qcolombet, ab Reviewed By: ab Subscribers: igorb, dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D28942 llvm-svn: 293172
2017-01-26 19:10:14 +08:00
}
[tablegen][globalisel] Convert the SelectionDAG importer to a tree walking approach. NFC Summary: But don't actually inspect the tree any deeper than we already do. This change is NFC but the next one will enable full traversal of the source/destination patterns. Depends on D30535 Reviewers: t.p.northover, qcolombet, aditya_nandakumar, rovka, ab Subscribers: igorb, dberris, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D30536 llvm-svn: 299001
2017-03-29 23:37:18 +08:00
OperandMatcher &getOperand(unsigned OpIdx) {
auto I = std::find_if(Operands.begin(), Operands.end(),
[globalisel][tablegen] Fix patterns involving multiple ComplexPatterns. Summary: Temporaries are now allocated to operands instead of predicates and this allocation is used to correctly pair up the rendered operands with the matched operands. Previously, ComplexPatterns were allocated temporaries independently in the Src Pattern and Dst Pattern, leading to mismatches. Additionally, the Dst Pattern failed to account for the allocated index and therefore always used temporary 0, 1, ... when it should have used base+0, base+1, ... Thanks to Aditya Nandakumar for noticing the bug. Depends on D30539 Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Reviewed By: rovka Subscribers: igorb, dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D31054 llvm-svn: 299538
2017-04-05 21:14:03 +08:00
[&OpIdx](const std::unique_ptr<OperandMatcher> &X) {
Reapply r332907 "[GlobalISel] Improving InstructionSelect's performance by reducing MatchTable..." Apparently the compile time problem was caused by the fact that not all compilers / STL implementations can automatically convert std::unique_ptr<Derived> to std::unique_ptr<Base>. Fixed (hopefully) by making sure it's std::unique_ptr<Derived>&& (rvalue ref) to std::unique_ptr<Base> conversion instead. llvm-svn: 332917
2018-05-22 07:28:51 +08:00
return X->getOpIdx() == OpIdx;
[tablegen][globalisel] Convert the SelectionDAG importer to a tree walking approach. NFC Summary: But don't actually inspect the tree any deeper than we already do. This change is NFC but the next one will enable full traversal of the source/destination patterns. Depends on D30535 Reviewers: t.p.northover, qcolombet, aditya_nandakumar, rovka, ab Subscribers: igorb, dberris, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D30536 llvm-svn: 299001
2017-03-29 23:37:18 +08:00
});
if (I != Operands.end())
[globalisel][tablegen] Fix patterns involving multiple ComplexPatterns. Summary: Temporaries are now allocated to operands instead of predicates and this allocation is used to correctly pair up the rendered operands with the matched operands. Previously, ComplexPatterns were allocated temporaries independently in the Src Pattern and Dst Pattern, leading to mismatches. Additionally, the Dst Pattern failed to account for the allocated index and therefore always used temporary 0, 1, ... when it should have used base+0, base+1, ... Thanks to Aditya Nandakumar for noticing the bug. Depends on D30539 Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Reviewed By: rovka Subscribers: igorb, dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D31054 llvm-svn: 299538
2017-04-05 21:14:03 +08:00
return **I;
[tablegen][globalisel] Convert the SelectionDAG importer to a tree walking approach. NFC Summary: But don't actually inspect the tree any deeper than we already do. This change is NFC but the next one will enable full traversal of the source/destination patterns. Depends on D30535 Reviewers: t.p.northover, qcolombet, aditya_nandakumar, rovka, ab Subscribers: igorb, dberris, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D30536 llvm-svn: 299001
2017-03-29 23:37:18 +08:00
llvm_unreachable("Failed to lookup operand");
}
[globalisel][tablegen] Add support for importing 'imm' operands. Summary: This patch enables the import of rules containing 'imm' operands that do not constrain the acceptable values using predicates. Support for ImmLeaf will arrive in a later patch. Depends on D35681 Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Reviewed By: rovka Subscribers: kristof.beyls, javed.absar, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D35833 llvm-svn: 310343
2017-08-08 18:44:31 +08:00
StringRef getSymbolicName() const { return SymbolicName; }
[globalisel] Decouple src pattern operands from dst pattern operands. Summary: This isn't testable for AArch64 by itself so this patch also adds support for constant immediates in the pattern and physical register uses in the result. The new IntOperandMatcher matches the constant in patterns such as '(set $rd:GPR32, (G_XOR $rs:GPR32, -1))'. It's always safe to fold immediates into an instruction so this is the first rule that will match across multiple BB's. The Renderer hierarchy is responsible for adding operands to the result instruction. Renderers can copy operands (CopyRenderer) or add physical registers (in particular %wzr and %xzr) to the result instruction in any order (OperandMatchers now import the operand names from SelectionDAG to allow renderers to access any operand). This allows us to emit the result instruction for: %1 = G_XOR %0, -1 --> %1 = ORNWrr %wzr, %0 %1 = G_XOR -1, %0 --> %1 = ORNWrr %wzr, %0 although the latter is untested since the matcher/importer has not been taught about commutativity yet. Added BuildMIAction which can build new instructions and mutate them where possible. W.r.t the mutation aspect, MatchActions are now told the name of an instruction they can recycle and BuildMIAction will emit mutation code when the renderers are appropriate. They are appropriate when all operands are rendered using CopyRenderer and the indices are the same as the matcher. This currently assumes that all operands have at least one matcher. Finally, this change also fixes a crash in AArch64InstructionSelector::select() caused by an immediate operand passing isImm() rather than isCImm(). This was uncovered by the other changes and was detected by existing tests. Depends on D29711 Reviewers: t.p.northover, ab, qcolombet, rovka, aditya_nandakumar, javed.absar Reviewed By: rovka Subscribers: aemerson, dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D29712 llvm-svn: 296131
2017-02-24 23:43:30 +08:00
unsigned getNumOperands() const { return Operands.size(); }
[tablegen][globalisel] Add support for nested instruction matching. Summary: Lift the restrictions that prevented the tree walking introduced in the previous change and add support for patterns like: (G_ADD (G_MUL (G_SEXT $src1), (G_SEXT $src2)), $src3) -> SMADDWrrr $dst, $src1, $src2, $src3 Also adds support for G_SEXT and G_ZEXT to support these cases. One particular aspect of this that I should draw attention to is that I've tried to be overly conservative in determining the safety of matches that involve non-adjacent instructions and multiple basic blocks. This is intended to be used as a cheap initial check and we may add a more expensive check in the future. The current rules are: * Reject if any instruction may load/store (we'd need to check for intervening memory operations. * Reject if any instruction has implicit operands. * Reject if any instruction has unmodelled side-effects. See isObviouslySafeToFold(). Reviewers: t.p.northover, javed.absar, qcolombet, aditya_nandakumar, ab, rovka Reviewed By: ab Subscribers: igorb, dberris, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D30539 llvm-svn: 299430
2017-04-04 21:25:23 +08:00
OperandVec::iterator operands_begin() { return Operands.begin(); }
OperandVec::iterator operands_end() { return Operands.end(); }
iterator_range<OperandVec::iterator> operands() {
return make_range(operands_begin(), operands_end());
}
[tablegen][globalisel] Convert the SelectionDAG importer to a tree walking approach. NFC Summary: But don't actually inspect the tree any deeper than we already do. This change is NFC but the next one will enable full traversal of the source/destination patterns. Depends on D30535 Reviewers: t.p.northover, qcolombet, aditya_nandakumar, rovka, ab Subscribers: igorb, dberris, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D30536 llvm-svn: 299001
2017-03-29 23:37:18 +08:00
OperandVec::const_iterator operands_begin() const { return Operands.begin(); }
OperandVec::const_iterator operands_end() const { return Operands.end(); }
[globalisel] Decouple src pattern operands from dst pattern operands. Summary: This isn't testable for AArch64 by itself so this patch also adds support for constant immediates in the pattern and physical register uses in the result. The new IntOperandMatcher matches the constant in patterns such as '(set $rd:GPR32, (G_XOR $rs:GPR32, -1))'. It's always safe to fold immediates into an instruction so this is the first rule that will match across multiple BB's. The Renderer hierarchy is responsible for adding operands to the result instruction. Renderers can copy operands (CopyRenderer) or add physical registers (in particular %wzr and %xzr) to the result instruction in any order (OperandMatchers now import the operand names from SelectionDAG to allow renderers to access any operand). This allows us to emit the result instruction for: %1 = G_XOR %0, -1 --> %1 = ORNWrr %wzr, %0 %1 = G_XOR -1, %0 --> %1 = ORNWrr %wzr, %0 although the latter is untested since the matcher/importer has not been taught about commutativity yet. Added BuildMIAction which can build new instructions and mutate them where possible. W.r.t the mutation aspect, MatchActions are now told the name of an instruction they can recycle and BuildMIAction will emit mutation code when the renderers are appropriate. They are appropriate when all operands are rendered using CopyRenderer and the indices are the same as the matcher. This currently assumes that all operands have at least one matcher. Finally, this change also fixes a crash in AArch64InstructionSelector::select() caused by an immediate operand passing isImm() rather than isCImm(). This was uncovered by the other changes and was detected by existing tests. Depends on D29711 Reviewers: t.p.northover, ab, qcolombet, rovka, aditya_nandakumar, javed.absar Reviewed By: rovka Subscribers: aemerson, dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D29712 llvm-svn: 296131
2017-02-24 23:43:30 +08:00
iterator_range<OperandVec::const_iterator> operands() const {
return make_range(operands_begin(), operands_end());
}
[TableGen][GlobalISel] Optimize MatchTable for faster instruction selection *** Context *** Prior to this patchw, the table generated for matching instruction was straight forward but highly inefficient. Basically, each pattern generates its own set of self contained checks and actions. E.g., TableGen generated: // First pattern CheckNumOperand 3 CheckOpcode G_ADD ... Build ADDrr // Second pattern CheckNumOperand 3 CheckOpcode G_ADD ... Build ADDri // Third pattern CheckNumOperand 3 CheckOpcode G_SUB ... Build SUBrr *** Problem *** Because of that generation, a *lot* of check were redundant between each pattern and were checked every single time until we reach the pattern that matches. E.g., Taking the previous table, let say we are matching a G_SUB, that means we were going to check all the rules for G_ADD before looking at the G_SUB rule. In particular we are going to do: check 3 operands; PASS check G_ADD; FAIL ; Next rule check 3 operands; PASS (but we already knew that!) check G_ADD; FAIL (well it is still not true) ; Next rule check 3 operands; PASS (really!!) check G_SUB; PASS (at last :P) *** Proposed Solution *** This patch introduces a concept of group of rules (GroupMatcher) that share some predicates and only get checked once for the whole group. This patch only creates groups with one nesting level. Conceptually there is nothing preventing us for having deeper nest level. However, the current implementation is not smart enough to share the recording (aka capturing) of values. That limits its ability to do more sharing. For the given example the current patch will generate: // First group CheckOpcode G_ADD // First pattern CheckNumOperand 3 ... Build ADDrr // Second pattern CheckNumOperand 3 ... Build ADDri // Second group CheckOpcode G_SUB // Third pattern CheckNumOperand 3 ... Build SUBrr But if we allowed several nesting level, it could create a sub group for the checknumoperand 3. (We would need to call optimizeRules on the rules within a group.) *** Result *** With only one level of nesting, the instruction selection pass is up to 4x faster. For instance, one instruction now takes 500 checks, instead of 24k! With more nesting we could get in the tens I believe. Differential Revision: https://reviews.llvm.org/D39034 rdar://problem/34670699 llvm-svn: 321017
2017-12-19 03:47:41 +08:00
bool operands_empty() const { return Operands.empty(); }
void pop_front() { Operands.erase(Operands.begin()); }
[globalisel] Decouple src pattern operands from dst pattern operands. Summary: This isn't testable for AArch64 by itself so this patch also adds support for constant immediates in the pattern and physical register uses in the result. The new IntOperandMatcher matches the constant in patterns such as '(set $rd:GPR32, (G_XOR $rs:GPR32, -1))'. It's always safe to fold immediates into an instruction so this is the first rule that will match across multiple BB's. The Renderer hierarchy is responsible for adding operands to the result instruction. Renderers can copy operands (CopyRenderer) or add physical registers (in particular %wzr and %xzr) to the result instruction in any order (OperandMatchers now import the operand names from SelectionDAG to allow renderers to access any operand). This allows us to emit the result instruction for: %1 = G_XOR %0, -1 --> %1 = ORNWrr %wzr, %0 %1 = G_XOR -1, %0 --> %1 = ORNWrr %wzr, %0 although the latter is untested since the matcher/importer has not been taught about commutativity yet. Added BuildMIAction which can build new instructions and mutate them where possible. W.r.t the mutation aspect, MatchActions are now told the name of an instruction they can recycle and BuildMIAction will emit mutation code when the renderers are appropriate. They are appropriate when all operands are rendered using CopyRenderer and the indices are the same as the matcher. This currently assumes that all operands have at least one matcher. Finally, this change also fixes a crash in AArch64InstructionSelector::select() caused by an immediate operand passing isImm() rather than isCImm(). This was uncovered by the other changes and was detected by existing tests. Depends on D29711 Reviewers: t.p.northover, ab, qcolombet, rovka, aditya_nandakumar, javed.absar Reviewed By: rovka Subscribers: aemerson, dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D29712 llvm-svn: 296131
2017-02-24 23:43:30 +08:00
[GlobalISel][InstructionSelect] Removing redundant num operands and nested def operands checks, perf patch 2 This patch continues a series of patches that decrease time spent by GlobalISel in its InstructionSelect pass by roughly 60% for -O0 builds for large inputs as measured on sqlite3-amalgamation (http://sqlite.org/download.html) targeting AArch64. This commit specifically removes number of operands checks that are redundant if the instruction's opcode already guarantees that number of operands (or more), and also avoids any kind of checks on a def operand of a nested instruction as everything about it was already checked at its use. The expected performance implication is about 3% off InstructionSelect comparing to the baseline (before the series of patches) This patch also contains a bit of NFC changes required for further patches in the series. Every commit planned shares the same Phabricator Review. Reviewers: qcolombet, dsanders, bogner, aemerson, javed.absar Reviewed By: qcolombet Subscribers: rovka, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D44700 llvm-svn: 332945
2018-05-22 12:31:50 +08:00
void optimize();
Reapply r332907 "[GlobalISel] Improving InstructionSelect's performance by reducing MatchTable..." Apparently the compile time problem was caused by the fact that not all compilers / STL implementations can automatically convert std::unique_ptr<Derived> to std::unique_ptr<Base>. Fixed (hopefully) by making sure it's std::unique_ptr<Derived>&& (rvalue ref) to std::unique_ptr<Base> conversion instead. llvm-svn: 332917
2018-05-22 07:28:51 +08:00
/// Emit MatchTable opcodes that test whether the instruction named in
/// InsnVarName matches all the predicates and all the operands.
void emitPredicateOpcodes(MatchTable &Table, RuleMatcher &Rule) {
[GlobalISel][InstructionSelect] Removing redundant num operands and nested def operands checks, perf patch 2 This patch continues a series of patches that decrease time spent by GlobalISel in its InstructionSelect pass by roughly 60% for -O0 builds for large inputs as measured on sqlite3-amalgamation (http://sqlite.org/download.html) targeting AArch64. This commit specifically removes number of operands checks that are redundant if the instruction's opcode already guarantees that number of operands (or more), and also avoids any kind of checks on a def operand of a nested instruction as everything about it was already checked at its use. The expected performance implication is about 3% off InstructionSelect comparing to the baseline (before the series of patches) This patch also contains a bit of NFC changes required for further patches in the series. Every commit planned shares the same Phabricator Review. Reviewers: qcolombet, dsanders, bogner, aemerson, javed.absar Reviewed By: qcolombet Subscribers: rovka, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D44700 llvm-svn: 332945
2018-05-22 12:31:50 +08:00
if (NumOperandsCheck)
InstructionNumOperandsMatcher(InsnVarID, getNumOperands())
.emitPredicateOpcodes(Table, Rule);
[tablegen][globalisel] Capture instructions into locals and related infrastructure for multiple instructions matches. Summary: Prepare the way for nested instruction matching support by having actions like CopyRenderer look up operands in the RuleMatcher rather than a specific InstructionMatcher. This allows actions to reference any operand from any matched instruction. It works by checking the 'shape' of the match and capturing each matched instruction to a local variable. If the shape is wrong (not enough operands, leaf nodes where non-leafs are expected, etc.), then the rule exits early without checking the predicates. Once we've captured the instructions, we then test the predicates as before (except using the local variables). If the match is successful, then we render the new instruction as before using the local variables. It's not noticable in this patch but by the time we support multiple instruction matching, this patch will also cause a significant improvement to readability of the emitted code since MRI.getVRegDef(I->getOperand(0).getReg()) will simply be MI1 after emitCxxCaptureStmts(). This isn't quite NFC because I've also fixed a bug that I'm surprised we haven't encountered yet. It now checks there are at least the expected number of operands before accessing them with getOperand(). Depends on D30531 Reviewers: t.p.northover, qcolombet, aditya_nandakumar, ab, rovka Reviewed By: rovka Subscribers: dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D30535 llvm-svn: 298257
2017-03-20 23:20:42 +08:00
[TableGen][GlobalISel] Have the predicate directly know which data they are dealing with Prior to this patch, a predicate wouldn't make sense outside of its rule. Indeed, it was only during emitting a rule that a predicate would be made aware of the IDs of the data it is checking. Because of that, predicates could not be moved around or compared between each other. NFC. llvm-svn: 320887
2017-12-16 07:07:42 +08:00
emitPredicateListOpcodes(Table, Rule);
Reapply r332907 "[GlobalISel] Improving InstructionSelect's performance by reducing MatchTable..." Apparently the compile time problem was caused by the fact that not all compilers / STL implementations can automatically convert std::unique_ptr<Derived> to std::unique_ptr<Base>. Fixed (hopefully) by making sure it's std::unique_ptr<Derived>&& (rvalue ref) to std::unique_ptr<Base> conversion instead. llvm-svn: 332917
2018-05-22 07:28:51 +08:00
[globalisel][tablegen] Partially fix compile-time regressions by converting matcher to state-machine(s) Summary: Replace the matcher if-statements for each rule with a state-machine. This significantly reduces compile time, memory allocations, and cumulative memory allocation when compiling AArch64InstructionSelector.cpp.o after r303259 is recommitted. The following patches will expand on this further to fully fix the regressions. Reviewers: rovka, ab, t.p.northover, qcolombet, aditya_nandakumar Reviewed By: ab Subscribers: vitalybuka, aemerson, javed.absar, igorb, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D33758 llvm-svn: 307079
2017-07-04 22:35:06 +08:00
for (const auto &Operand : Operands)
[TableGen][GlobalISel] Have the predicate directly know which data they are dealing with Prior to this patch, a predicate wouldn't make sense outside of its rule. Indeed, it was only during emitting a rule that a predicate would be made aware of the IDs of the data it is checking. Because of that, predicates could not be moved around or compared between each other. NFC. llvm-svn: 320887
2017-12-16 07:07:42 +08:00
Operand->emitPredicateOpcodes(Table, Rule);
[GlobalISel] Add basic Selector-emitter tblgen backend. This adds a basic tablegen backend that analyzes the SelectionDAG patterns to find simple ones that are eligible for GlobalISel-emission. That's similar to FastISel, with one notable difference: we're not fed ISD opcodes, so we need to map the SDNode operators to generic opcodes. That's done using GINodeEquiv in TargetGlobalISel.td. Otherwise, this is mostly boilerplate, and lots of filtering of any kind of "complicated" pattern. On AArch64, this is sufficient to match G_ADD up to s64 (to ADDWrr/ADDXrr) and G_BR (to B). Differential Revision: https://reviews.llvm.org/D26878 llvm-svn: 290284
2016-12-22 07:26:20 +08:00
}
[globalisel] Sort RuleMatchers by priority. Summary: This makes more important rules have priority over less important rules. For example, '%a = G_ADD $b:s64, $c:s64' has priority over '%a = G_ADD $b:s32, $c:s32'. Previously these rules were emitted in the correct order by chance. NFC in this patch but it is required to make the next patch work correctly. Depends on D29710 Reviewers: t.p.northover, ab, qcolombet, aditya_nandakumar, rovka Reviewed By: ab, rovka Subscribers: javed.absar, dberris, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D29711 llvm-svn: 296121
2017-02-24 21:58:11 +08:00
/// Compare the priority of this object and B.
///
/// Returns true if this object is more important than B.
Reapply r332907 "[GlobalISel] Improving InstructionSelect's performance by reducing MatchTable..." Apparently the compile time problem was caused by the fact that not all compilers / STL implementations can automatically convert std::unique_ptr<Derived> to std::unique_ptr<Base>. Fixed (hopefully) by making sure it's std::unique_ptr<Derived>&& (rvalue ref) to std::unique_ptr<Base> conversion instead. llvm-svn: 332917
2018-05-22 07:28:51 +08:00
bool isHigherPriorityThan(InstructionMatcher &B) {
[globalisel] Sort RuleMatchers by priority. Summary: This makes more important rules have priority over less important rules. For example, '%a = G_ADD $b:s64, $c:s64' has priority over '%a = G_ADD $b:s32, $c:s32'. Previously these rules were emitted in the correct order by chance. NFC in this patch but it is required to make the next patch work correctly. Depends on D29710 Reviewers: t.p.northover, ab, qcolombet, aditya_nandakumar, rovka Reviewed By: ab, rovka Subscribers: javed.absar, dberris, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D29711 llvm-svn: 296121
2017-02-24 21:58:11 +08:00
// Instruction matchers involving more operands have higher priority.
if (Operands.size() > B.Operands.size())
return true;
if (Operands.size() < B.Operands.size())
return false;
Reapply r332907 "[GlobalISel] Improving InstructionSelect's performance by reducing MatchTable..." Apparently the compile time problem was caused by the fact that not all compilers / STL implementations can automatically convert std::unique_ptr<Derived> to std::unique_ptr<Base>. Fixed (hopefully) by making sure it's std::unique_ptr<Derived>&& (rvalue ref) to std::unique_ptr<Base> conversion instead. llvm-svn: 332917
2018-05-22 07:28:51 +08:00
for (auto &&P : zip(predicates(), B.predicates())) {
auto L = static_cast<InstructionPredicateMatcher *>(std::get<0>(P).get());
auto R = static_cast<InstructionPredicateMatcher *>(std::get<1>(P).get());
if (L->isHigherPriorityThan(*R))
[globalisel] Sort RuleMatchers by priority. Summary: This makes more important rules have priority over less important rules. For example, '%a = G_ADD $b:s64, $c:s64' has priority over '%a = G_ADD $b:s32, $c:s32'. Previously these rules were emitted in the correct order by chance. NFC in this patch but it is required to make the next patch work correctly. Depends on D29710 Reviewers: t.p.northover, ab, qcolombet, aditya_nandakumar, rovka Reviewed By: ab, rovka Subscribers: javed.absar, dberris, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D29711 llvm-svn: 296121
2017-02-24 21:58:11 +08:00
return true;
Reapply r332907 "[GlobalISel] Improving InstructionSelect's performance by reducing MatchTable..." Apparently the compile time problem was caused by the fact that not all compilers / STL implementations can automatically convert std::unique_ptr<Derived> to std::unique_ptr<Base>. Fixed (hopefully) by making sure it's std::unique_ptr<Derived>&& (rvalue ref) to std::unique_ptr<Base> conversion instead. llvm-svn: 332917
2018-05-22 07:28:51 +08:00
if (R->isHigherPriorityThan(*L))
[globalisel] Sort RuleMatchers by priority. Summary: This makes more important rules have priority over less important rules. For example, '%a = G_ADD $b:s64, $c:s64' has priority over '%a = G_ADD $b:s32, $c:s32'. Previously these rules were emitted in the correct order by chance. NFC in this patch but it is required to make the next patch work correctly. Depends on D29710 Reviewers: t.p.northover, ab, qcolombet, aditya_nandakumar, rovka Reviewed By: ab, rovka Subscribers: javed.absar, dberris, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D29711 llvm-svn: 296121
2017-02-24 21:58:11 +08:00
return false;
}
for (const auto &Operand : zip(Operands, B.Operands)) {
[globalisel][tablegen] Fix patterns involving multiple ComplexPatterns. Summary: Temporaries are now allocated to operands instead of predicates and this allocation is used to correctly pair up the rendered operands with the matched operands. Previously, ComplexPatterns were allocated temporaries independently in the Src Pattern and Dst Pattern, leading to mismatches. Additionally, the Dst Pattern failed to account for the allocated index and therefore always used temporary 0, 1, ... when it should have used base+0, base+1, ... Thanks to Aditya Nandakumar for noticing the bug. Depends on D30539 Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Reviewed By: rovka Subscribers: igorb, dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D31054 llvm-svn: 299538
2017-04-05 21:14:03 +08:00
if (std::get<0>(Operand)->isHigherPriorityThan(*std::get<1>(Operand)))
[globalisel] Sort RuleMatchers by priority. Summary: This makes more important rules have priority over less important rules. For example, '%a = G_ADD $b:s64, $c:s64' has priority over '%a = G_ADD $b:s32, $c:s32'. Previously these rules were emitted in the correct order by chance. NFC in this patch but it is required to make the next patch work correctly. Depends on D29710 Reviewers: t.p.northover, ab, qcolombet, aditya_nandakumar, rovka Reviewed By: ab, rovka Subscribers: javed.absar, dberris, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D29711 llvm-svn: 296121
2017-02-24 21:58:11 +08:00
return true;
[globalisel][tablegen] Fix patterns involving multiple ComplexPatterns. Summary: Temporaries are now allocated to operands instead of predicates and this allocation is used to correctly pair up the rendered operands with the matched operands. Previously, ComplexPatterns were allocated temporaries independently in the Src Pattern and Dst Pattern, leading to mismatches. Additionally, the Dst Pattern failed to account for the allocated index and therefore always used temporary 0, 1, ... when it should have used base+0, base+1, ... Thanks to Aditya Nandakumar for noticing the bug. Depends on D30539 Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Reviewed By: rovka Subscribers: igorb, dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D31054 llvm-svn: 299538
2017-04-05 21:14:03 +08:00
if (std::get<1>(Operand)->isHigherPriorityThan(*std::get<0>(Operand)))
[globalisel] Sort RuleMatchers by priority. Summary: This makes more important rules have priority over less important rules. For example, '%a = G_ADD $b:s64, $c:s64' has priority over '%a = G_ADD $b:s32, $c:s32'. Previously these rules were emitted in the correct order by chance. NFC in this patch but it is required to make the next patch work correctly. Depends on D29710 Reviewers: t.p.northover, ab, qcolombet, aditya_nandakumar, rovka Reviewed By: ab, rovka Subscribers: javed.absar, dberris, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D29711 llvm-svn: 296121
2017-02-24 21:58:11 +08:00
return false;
}
return false;
};
[globalisel][tblgen] Add support for ComplexPatterns Summary: Adds a new kind of MachineOperand: MO_Placeholder. This operand must not appear in the MIR and only exists as a way of creating an 'uninitialized' operand until a matcher function overwrites it. Depends on D30046, D29712 Reviewers: t.p.northover, ab, rovka, aditya_nandakumar, javed.absar, qcolombet Reviewed By: qcolombet Subscribers: dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D30089 llvm-svn: 297782
2017-03-15 05:32:08 +08:00
/// Report the maximum number of temporary operands needed by the instruction
/// matcher.
Reapply r332907 "[GlobalISel] Improving InstructionSelect's performance by reducing MatchTable..." Apparently the compile time problem was caused by the fact that not all compilers / STL implementations can automatically convert std::unique_ptr<Derived> to std::unique_ptr<Base>. Fixed (hopefully) by making sure it's std::unique_ptr<Derived>&& (rvalue ref) to std::unique_ptr<Base> conversion instead. llvm-svn: 332917
2018-05-22 07:28:51 +08:00
unsigned countRendererFns() {
return std::accumulate(
predicates().begin(), predicates().end(), 0,
[](unsigned A,
const std::unique_ptr<PredicateMatcher> &Predicate) {
return A + Predicate->countRendererFns();
}) +
[globalisel][tablegen] Fix patterns involving multiple ComplexPatterns. Summary: Temporaries are now allocated to operands instead of predicates and this allocation is used to correctly pair up the rendered operands with the matched operands. Previously, ComplexPatterns were allocated temporaries independently in the Src Pattern and Dst Pattern, leading to mismatches. Additionally, the Dst Pattern failed to account for the allocated index and therefore always used temporary 0, 1, ... when it should have used base+0, base+1, ... Thanks to Aditya Nandakumar for noticing the bug. Depends on D30539 Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Reviewed By: rovka Subscribers: igorb, dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D31054 llvm-svn: 299538
2017-04-05 21:14:03 +08:00
std::accumulate(
Operands.begin(), Operands.end(), 0,
[](unsigned A, const std::unique_ptr<OperandMatcher> &Operand) {
[globalisel][tablegen] Revise API for ComplexPattern operands to improve flexibility. Summary: Some targets need to be able to do more complex rendering than just adding an operand or two to an instruction. For example, it may need to insert an instruction to extract a subreg first, or it may need to perform an operation on the operand. In SelectionDAG, targets would create SDNode's to achieve the desired effect during the complex pattern predicate. This worked because SelectionDAG had a form of garbage collection that would take care of SDNode's that were created but not used due to a later predicate rejecting a match. This doesn't translate well to GlobalISel and the churn was wasteful. The API changes in this patch enable GlobalISel to accomplish the same thing without the waste. The API is now: InstructionSelector::OptionalComplexRendererFn selectArithImmed(MachineOperand &Root) const; where Root is the root of the match. The return value can be omitted to indicate that the predicate failed to match, or a function with the signature ComplexRendererFn can be returned. For example: return OptionalComplexRendererFn( [=](MachineInstrBuilder &MIB) { MIB.addImm(Immed).addImm(ShVal); }); adds two immediate operands to the rendered instruction. Immed and ShVal are captured from the predicate function. As an added bonus, this also reduces the amount of information we need to provide to GIComplexOperandMatcher. Depends on D31418 Reviewers: aditya_nandakumar, t.p.northover, qcolombet, rovka, ab, javed.absar Reviewed By: ab Subscribers: dberris, kristof.beyls, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D31761 llvm-svn: 301079
2017-04-22 23:11:04 +08:00
return A + Operand->countRendererFns();
[globalisel][tablegen] Fix patterns involving multiple ComplexPatterns. Summary: Temporaries are now allocated to operands instead of predicates and this allocation is used to correctly pair up the rendered operands with the matched operands. Previously, ComplexPatterns were allocated temporaries independently in the Src Pattern and Dst Pattern, leading to mismatches. Additionally, the Dst Pattern failed to account for the allocated index and therefore always used temporary 0, 1, ... when it should have used base+0, base+1, ... Thanks to Aditya Nandakumar for noticing the bug. Depends on D30539 Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Reviewed By: rovka Subscribers: igorb, dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D31054 llvm-svn: 299538
2017-04-05 21:14:03 +08:00
});
[globalisel][tblgen] Add support for ComplexPatterns Summary: Adds a new kind of MachineOperand: MO_Placeholder. This operand must not appear in the MIR and only exists as a way of creating an 'uninitialized' operand until a matcher function overwrites it. Depends on D30046, D29712 Reviewers: t.p.northover, ab, rovka, aditya_nandakumar, javed.absar, qcolombet Reviewed By: qcolombet Subscribers: dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D30089 llvm-svn: 297782
2017-03-15 05:32:08 +08:00
}
[globalisel][tablegen] Add support for importing 'imm' operands. Summary: This patch enables the import of rules containing 'imm' operands that do not constrain the acceptable values using predicates. Support for ImmLeaf will arrive in a later patch. Depends on D35681 Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Reviewed By: rovka Subscribers: kristof.beyls, javed.absar, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D35833 llvm-svn: 310343
2017-08-08 18:44:31 +08:00
Reapply r332907 "[GlobalISel] Improving InstructionSelect's performance by reducing MatchTable..." Apparently the compile time problem was caused by the fact that not all compilers / STL implementations can automatically convert std::unique_ptr<Derived> to std::unique_ptr<Base>. Fixed (hopefully) by making sure it's std::unique_ptr<Derived>&& (rvalue ref) to std::unique_ptr<Base> conversion instead. llvm-svn: 332917
2018-05-22 07:28:51 +08:00
InstructionOpcodeMatcher &getOpcodeMatcher() {
for (auto &P : predicates())
if (auto *OpMatcher = dyn_cast<InstructionOpcodeMatcher>(P.get()))
return *OpMatcher;
llvm_unreachable("Didn't find an opcode matcher");
}
bool isConstantInstruction() {
return getOpcodeMatcher().isConstantInstruction();
[globalisel][tablegen] Add support for importing 'imm' operands. Summary: This patch enables the import of rules containing 'imm' operands that do not constrain the acceptable values using predicates. Support for ImmLeaf will arrive in a later patch. Depends on D35681 Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Reviewed By: rovka Subscribers: kristof.beyls, javed.absar, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D35833 llvm-svn: 310343
2017-08-08 18:44:31 +08:00
}
[GlobalISel][InstructionSelect] Removing redundant num operands and nested def operands checks, perf patch 2 This patch continues a series of patches that decrease time spent by GlobalISel in its InstructionSelect pass by roughly 60% for -O0 builds for large inputs as measured on sqlite3-amalgamation (http://sqlite.org/download.html) targeting AArch64. This commit specifically removes number of operands checks that are redundant if the instruction's opcode already guarantees that number of operands (or more), and also avoids any kind of checks on a def operand of a nested instruction as everything about it was already checked at its use. The expected performance implication is about 3% off InstructionSelect comparing to the baseline (before the series of patches) This patch also contains a bit of NFC changes required for further patches in the series. Every commit planned shares the same Phabricator Review. Reviewers: qcolombet, dsanders, bogner, aemerson, javed.absar Reviewed By: qcolombet Subscribers: rovka, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D44700 llvm-svn: 332945
2018-05-22 12:31:50 +08:00
StringRef getOpcode() { return getOpcodeMatcher().getOpcode(); }
[GlobalISel] Add basic Selector-emitter tblgen backend. This adds a basic tablegen backend that analyzes the SelectionDAG patterns to find simple ones that are eligible for GlobalISel-emission. That's similar to FastISel, with one notable difference: we're not fed ISD opcodes, so we need to map the SDNode operators to generic opcodes. That's done using GINodeEquiv in TargetGlobalISel.td. Otherwise, this is mostly boilerplate, and lots of filtering of any kind of "complicated" pattern. On AArch64, this is sufficient to match G_ADD up to s64 (to ADDWrr/ADDXrr) and G_BR (to B). Differential Revision: https://reviews.llvm.org/D26878 llvm-svn: 290284
2016-12-22 07:26:20 +08:00
};
[GlobalISel][InstructionSelect] Removing redundant num operands and nested def operands checks, perf patch 2 This patch continues a series of patches that decrease time spent by GlobalISel in its InstructionSelect pass by roughly 60% for -O0 builds for large inputs as measured on sqlite3-amalgamation (http://sqlite.org/download.html) targeting AArch64. This commit specifically removes number of operands checks that are redundant if the instruction's opcode already guarantees that number of operands (or more), and also avoids any kind of checks on a def operand of a nested instruction as everything about it was already checked at its use. The expected performance implication is about 3% off InstructionSelect comparing to the baseline (before the series of patches) This patch also contains a bit of NFC changes required for further patches in the series. Every commit planned shares the same Phabricator Review. Reviewers: qcolombet, dsanders, bogner, aemerson, javed.absar Reviewed By: qcolombet Subscribers: rovka, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D44700 llvm-svn: 332945
2018-05-22 12:31:50 +08:00
StringRef RuleMatcher::getOpcode() const {
return Matchers.front()->getOpcode();
}
Reapply r332907 "[GlobalISel] Improving InstructionSelect's performance by reducing MatchTable..." Apparently the compile time problem was caused by the fact that not all compilers / STL implementations can automatically convert std::unique_ptr<Derived> to std::unique_ptr<Base>. Fixed (hopefully) by making sure it's std::unique_ptr<Derived>&& (rvalue ref) to std::unique_ptr<Base> conversion instead. llvm-svn: 332917
2018-05-22 07:28:51 +08:00
unsigned RuleMatcher::getNumOperands() const {
return Matchers.front()->getNumOperands();
}
[tablegen][globalisel] Add support for nested instruction matching. Summary: Lift the restrictions that prevented the tree walking introduced in the previous change and add support for patterns like: (G_ADD (G_MUL (G_SEXT $src1), (G_SEXT $src2)), $src3) -> SMADDWrrr $dst, $src1, $src2, $src3 Also adds support for G_SEXT and G_ZEXT to support these cases. One particular aspect of this that I should draw attention to is that I've tried to be overly conservative in determining the safety of matches that involve non-adjacent instructions and multiple basic blocks. This is intended to be used as a cheap initial check and we may add a more expensive check in the future. The current rules are: * Reject if any instruction may load/store (we'd need to check for intervening memory operations. * Reject if any instruction has implicit operands. * Reject if any instruction has unmodelled side-effects. See isObviouslySafeToFold(). Reviewers: t.p.northover, javed.absar, qcolombet, aditya_nandakumar, ab, rovka Reviewed By: ab Subscribers: igorb, dberris, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D30539 llvm-svn: 299430
2017-04-04 21:25:23 +08:00
/// Generates code to check that the operand is a register defined by an
/// instruction that matches the given instruction matcher.
///
/// For example, the pattern:
/// (set $dst, (G_MUL (G_ADD $src1, $src2), $src3))
/// would use an InstructionOperandMatcher for operand 1 of the G_MUL to match
/// the:
/// (G_ADD $src1, $src2)
/// subpattern.
class InstructionOperandMatcher : public OperandPredicateMatcher {
protected:
std::unique_ptr<InstructionMatcher> InsnMatcher;
public:
[TableGen][GlobalISel] Make the arguments of the Instruction and Operand Matchers consistent Move InsnVarID and OpIdx at the beginning of the list of arguments for all the constructors of the OperandMatcher subclasses. This matches what we do for the InstructionMatcher. NFC. llvm-svn: 321031
2017-12-19 06:12:13 +08:00
InstructionOperandMatcher(unsigned InsnVarID, unsigned OpIdx,
RuleMatcher &Rule, StringRef SymbolicName)
[TableGen][GlobalISel] Have the predicate directly know which data they are dealing with Prior to this patch, a predicate wouldn't make sense outside of its rule. Indeed, it was only during emitting a rule that a predicate would be made aware of the IDs of the data it is checking. Because of that, predicates could not be moved around or compared between each other. NFC. llvm-svn: 320887
2017-12-16 07:07:42 +08:00
: OperandPredicateMatcher(OPM_Instruction, InsnVarID, OpIdx),
[globalisel][tablegen] Simplify named operand/operator lookups and fix a wrong-code bug this revealed. Summary: Operand variable lookups are now performed by the RuleMatcher rather than searching the whole matcher hierarchy for a match. This revealed a wrong-code bug that currently affects ARM and X86 where patterns that use a variable more than once in the match pattern will be imported but won't check that the operands are identical. This can cause the tablegen-erated matcher to accept matches that should be rejected. Depends on D36569 Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Subscribers: aemerson, igorb, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D36618 llvm-svn: 315780
2017-10-14 08:31:58 +08:00
InsnMatcher(new InstructionMatcher(Rule, SymbolicName)) {}
[tablegen][globalisel] Add support for nested instruction matching. Summary: Lift the restrictions that prevented the tree walking introduced in the previous change and add support for patterns like: (G_ADD (G_MUL (G_SEXT $src1), (G_SEXT $src2)), $src3) -> SMADDWrrr $dst, $src1, $src2, $src3 Also adds support for G_SEXT and G_ZEXT to support these cases. One particular aspect of this that I should draw attention to is that I've tried to be overly conservative in determining the safety of matches that involve non-adjacent instructions and multiple basic blocks. This is intended to be used as a cheap initial check and we may add a more expensive check in the future. The current rules are: * Reject if any instruction may load/store (we'd need to check for intervening memory operations. * Reject if any instruction has implicit operands. * Reject if any instruction has unmodelled side-effects. See isObviouslySafeToFold(). Reviewers: t.p.northover, javed.absar, qcolombet, aditya_nandakumar, ab, rovka Reviewed By: ab Subscribers: igorb, dberris, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D30539 llvm-svn: 299430
2017-04-04 21:25:23 +08:00
[TableGen][GlobalISel] Add a common class for all PredicateMatcher NFC. llvm-svn: 320767
2017-12-15 07:44:07 +08:00
static bool classof(const PredicateMatcher *P) {
[tablegen][globalisel] Add support for nested instruction matching. Summary: Lift the restrictions that prevented the tree walking introduced in the previous change and add support for patterns like: (G_ADD (G_MUL (G_SEXT $src1), (G_SEXT $src2)), $src3) -> SMADDWrrr $dst, $src1, $src2, $src3 Also adds support for G_SEXT and G_ZEXT to support these cases. One particular aspect of this that I should draw attention to is that I've tried to be overly conservative in determining the safety of matches that involve non-adjacent instructions and multiple basic blocks. This is intended to be used as a cheap initial check and we may add a more expensive check in the future. The current rules are: * Reject if any instruction may load/store (we'd need to check for intervening memory operations. * Reject if any instruction has implicit operands. * Reject if any instruction has unmodelled side-effects. See isObviouslySafeToFold(). Reviewers: t.p.northover, javed.absar, qcolombet, aditya_nandakumar, ab, rovka Reviewed By: ab Subscribers: igorb, dberris, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D30539 llvm-svn: 299430
2017-04-04 21:25:23 +08:00
return P->getKind() == OPM_Instruction;
}
InstructionMatcher &getInsnMatcher() const { return *InsnMatcher; }
Reapply r332907 "[GlobalISel] Improving InstructionSelect's performance by reducing MatchTable..." Apparently the compile time problem was caused by the fact that not all compilers / STL implementations can automatically convert std::unique_ptr<Derived> to std::unique_ptr<Base>. Fixed (hopefully) by making sure it's std::unique_ptr<Derived>&& (rvalue ref) to std::unique_ptr<Base> conversion instead. llvm-svn: 332917
2018-05-22 07:28:51 +08:00
void emitCaptureOpcodes(MatchTable &Table, RuleMatcher &Rule) const {
const unsigned NewInsnVarID = InsnMatcher->getInsnVarID();
Table << MatchTable::Opcode("GIM_RecordInsn")
<< MatchTable::Comment("DefineMI")
<< MatchTable::IntValue(NewInsnVarID) << MatchTable::Comment("MI")
<< MatchTable::IntValue(getInsnVarID())
<< MatchTable::Comment("OpIdx") << MatchTable::IntValue(getOpIdx())
<< MatchTable::Comment("MIs[" + llvm::to_string(NewInsnVarID) + "]")
<< MatchTable::LineBreak;
[tablegen][globalisel] Add support for nested instruction matching. Summary: Lift the restrictions that prevented the tree walking introduced in the previous change and add support for patterns like: (G_ADD (G_MUL (G_SEXT $src1), (G_SEXT $src2)), $src3) -> SMADDWrrr $dst, $src1, $src2, $src3 Also adds support for G_SEXT and G_ZEXT to support these cases. One particular aspect of this that I should draw attention to is that I've tried to be overly conservative in determining the safety of matches that involve non-adjacent instructions and multiple basic blocks. This is intended to be used as a cheap initial check and we may add a more expensive check in the future. The current rules are: * Reject if any instruction may load/store (we'd need to check for intervening memory operations. * Reject if any instruction has implicit operands. * Reject if any instruction has unmodelled side-effects. See isObviouslySafeToFold(). Reviewers: t.p.northover, javed.absar, qcolombet, aditya_nandakumar, ab, rovka Reviewed By: ab Subscribers: igorb, dberris, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D30539 llvm-svn: 299430
2017-04-04 21:25:23 +08:00
}
[TableGen][GlobalISel] Have the predicate directly know which data they are dealing with Prior to this patch, a predicate wouldn't make sense outside of its rule. Indeed, it was only during emitting a rule that a predicate would be made aware of the IDs of the data it is checking. Because of that, predicates could not be moved around or compared between each other. NFC. llvm-svn: 320887
2017-12-16 07:07:42 +08:00
void emitPredicateOpcodes(MatchTable &Table,
RuleMatcher &Rule) const override {
Reapply r332907 "[GlobalISel] Improving InstructionSelect's performance by reducing MatchTable..." Apparently the compile time problem was caused by the fact that not all compilers / STL implementations can automatically convert std::unique_ptr<Derived> to std::unique_ptr<Base>. Fixed (hopefully) by making sure it's std::unique_ptr<Derived>&& (rvalue ref) to std::unique_ptr<Base> conversion instead. llvm-svn: 332917
2018-05-22 07:28:51 +08:00
emitCaptureOpcodes(Table, Rule);
[TableGen][GlobalISel] Have the predicate directly know which data they are dealing with Prior to this patch, a predicate wouldn't make sense outside of its rule. Indeed, it was only during emitting a rule that a predicate would be made aware of the IDs of the data it is checking. Because of that, predicates could not be moved around or compared between each other. NFC. llvm-svn: 320887
2017-12-16 07:07:42 +08:00
InsnMatcher->emitPredicateOpcodes(Table, Rule);
[tablegen][globalisel] Add support for nested instruction matching. Summary: Lift the restrictions that prevented the tree walking introduced in the previous change and add support for patterns like: (G_ADD (G_MUL (G_SEXT $src1), (G_SEXT $src2)), $src3) -> SMADDWrrr $dst, $src1, $src2, $src3 Also adds support for G_SEXT and G_ZEXT to support these cases. One particular aspect of this that I should draw attention to is that I've tried to be overly conservative in determining the safety of matches that involve non-adjacent instructions and multiple basic blocks. This is intended to be used as a cheap initial check and we may add a more expensive check in the future. The current rules are: * Reject if any instruction may load/store (we'd need to check for intervening memory operations. * Reject if any instruction has implicit operands. * Reject if any instruction has unmodelled side-effects. See isObviouslySafeToFold(). Reviewers: t.p.northover, javed.absar, qcolombet, aditya_nandakumar, ab, rovka Reviewed By: ab Subscribers: igorb, dberris, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D30539 llvm-svn: 299430
2017-04-04 21:25:23 +08:00
}
[globalisel][tablegen] Honour priority order within nested instructions. It appears that we haven't been prioritizing rules that contain nested instructions properly. InstructionOperandMatcher didn't override isHigherPriorityThan so it never compared the instructions/operands/predicates inside nested instructions. Fixes PR35926. Thanks to Diana Picus for the bug report. llvm-svn: 322754
2018-01-18 04:34:29 +08:00
bool isHigherPriorityThan(const OperandPredicateMatcher &B) const override {
if (OperandPredicateMatcher::isHigherPriorityThan(B))
return true;
if (B.OperandPredicateMatcher::isHigherPriorityThan(*this))
return false;
if (const InstructionOperandMatcher *BP =
dyn_cast<InstructionOperandMatcher>(&B))
if (InsnMatcher->isHigherPriorityThan(*BP->InsnMatcher))
return true;
return false;
}
[tablegen][globalisel] Add support for nested instruction matching. Summary: Lift the restrictions that prevented the tree walking introduced in the previous change and add support for patterns like: (G_ADD (G_MUL (G_SEXT $src1), (G_SEXT $src2)), $src3) -> SMADDWrrr $dst, $src1, $src2, $src3 Also adds support for G_SEXT and G_ZEXT to support these cases. One particular aspect of this that I should draw attention to is that I've tried to be overly conservative in determining the safety of matches that involve non-adjacent instructions and multiple basic blocks. This is intended to be used as a cheap initial check and we may add a more expensive check in the future. The current rules are: * Reject if any instruction may load/store (we'd need to check for intervening memory operations. * Reject if any instruction has implicit operands. * Reject if any instruction has unmodelled side-effects. See isObviouslySafeToFold(). Reviewers: t.p.northover, javed.absar, qcolombet, aditya_nandakumar, ab, rovka Reviewed By: ab Subscribers: igorb, dberris, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D30539 llvm-svn: 299430
2017-04-04 21:25:23 +08:00
};
[GlobalISel][InstructionSelect] Removing redundant num operands and nested def operands checks, perf patch 2 This patch continues a series of patches that decrease time spent by GlobalISel in its InstructionSelect pass by roughly 60% for -O0 builds for large inputs as measured on sqlite3-amalgamation (http://sqlite.org/download.html) targeting AArch64. This commit specifically removes number of operands checks that are redundant if the instruction's opcode already guarantees that number of operands (or more), and also avoids any kind of checks on a def operand of a nested instruction as everything about it was already checked at its use. The expected performance implication is about 3% off InstructionSelect comparing to the baseline (before the series of patches) This patch also contains a bit of NFC changes required for further patches in the series. Every commit planned shares the same Phabricator Review. Reviewers: qcolombet, dsanders, bogner, aemerson, javed.absar Reviewed By: qcolombet Subscribers: rovka, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D44700 llvm-svn: 332945
2018-05-22 12:31:50 +08:00
void InstructionMatcher::optimize() {
SmallVector<std::unique_ptr<PredicateMatcher>, 8> Stash;
const auto &OpcMatcher = getOpcodeMatcher();
Stash.push_back(predicates_pop_front());
if (Stash.back().get() == &OpcMatcher) {
if (NumOperandsCheck && OpcMatcher.getNumOperands() < getNumOperands())
Stash.emplace_back(
new InstructionNumOperandsMatcher(InsnVarID, getNumOperands()));
NumOperandsCheck = false;
[GlobalISel][InstructionSelect] MatchTable second level grouping, perf patch 5 This patch continues a series of patches started by r332907 (reapplied as r332917) In this commit we start grouping rules with common first condition on the second level of the table. This is expected to decrease time GlobalISel spends in its InstructionSelect pass by roughly 13% for an -O0 build as measured on sqlite3-amalgamation (http://sqlite.org/download.html) targeting AArch64. Reviewers: qcolombet, dsanders, bogner, aemerson, javed.absar Reviewed By: qcolombet Subscribers: rovka, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D44700 llvm-svn: 333053
2018-05-23 10:04:19 +08:00
for (auto &OM : Operands)
for (auto &OP : OM->predicates())
if (isa<IntrinsicIDOperandMatcher>(OP)) {
Stash.push_back(std::move(OP));
OM->eraseNullPredicates();
break;
}
[GlobalISel][InstructionSelect] Removing redundant num operands and nested def operands checks, perf patch 2 This patch continues a series of patches that decrease time spent by GlobalISel in its InstructionSelect pass by roughly 60% for -O0 builds for large inputs as measured on sqlite3-amalgamation (http://sqlite.org/download.html) targeting AArch64. This commit specifically removes number of operands checks that are redundant if the instruction's opcode already guarantees that number of operands (or more), and also avoids any kind of checks on a def operand of a nested instruction as everything about it was already checked at its use. The expected performance implication is about 3% off InstructionSelect comparing to the baseline (before the series of patches) This patch also contains a bit of NFC changes required for further patches in the series. Every commit planned shares the same Phabricator Review. Reviewers: qcolombet, dsanders, bogner, aemerson, javed.absar Reviewed By: qcolombet Subscribers: rovka, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D44700 llvm-svn: 332945
2018-05-22 12:31:50 +08:00
}
if (InsnVarID > 0) {
assert(!Operands.empty() && "Nested instruction is expected to def a vreg");
for (auto &OP : Operands[0]->predicates())
OP.reset();
Operands[0]->eraseNullPredicates();
}
while (!Stash.empty())
prependPredicate(Stash.pop_back_val());
}
[globalisel] Make the MatchAction hierarchy consistent with the matchers. NFC. Reviewers: aditya_nandakumar, rovka, t.p.northover, qcolombet, ab Subscribers: dberris, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D29321 llvm-svn: 293760
2017-02-01 18:53:10 +08:00
//===- Actions ------------------------------------------------------------===//
[globalisel] Decouple src pattern operands from dst pattern operands. Summary: This isn't testable for AArch64 by itself so this patch also adds support for constant immediates in the pattern and physical register uses in the result. The new IntOperandMatcher matches the constant in patterns such as '(set $rd:GPR32, (G_XOR $rs:GPR32, -1))'. It's always safe to fold immediates into an instruction so this is the first rule that will match across multiple BB's. The Renderer hierarchy is responsible for adding operands to the result instruction. Renderers can copy operands (CopyRenderer) or add physical registers (in particular %wzr and %xzr) to the result instruction in any order (OperandMatchers now import the operand names from SelectionDAG to allow renderers to access any operand). This allows us to emit the result instruction for: %1 = G_XOR %0, -1 --> %1 = ORNWrr %wzr, %0 %1 = G_XOR -1, %0 --> %1 = ORNWrr %wzr, %0 although the latter is untested since the matcher/importer has not been taught about commutativity yet. Added BuildMIAction which can build new instructions and mutate them where possible. W.r.t the mutation aspect, MatchActions are now told the name of an instruction they can recycle and BuildMIAction will emit mutation code when the renderers are appropriate. They are appropriate when all operands are rendered using CopyRenderer and the indices are the same as the matcher. This currently assumes that all operands have at least one matcher. Finally, this change also fixes a crash in AArch64InstructionSelector::select() caused by an immediate operand passing isImm() rather than isCImm(). This was uncovered by the other changes and was detected by existing tests. Depends on D29711 Reviewers: t.p.northover, ab, qcolombet, rovka, aditya_nandakumar, javed.absar Reviewed By: rovka Subscribers: aemerson, dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D29712 llvm-svn: 296131
2017-02-24 23:43:30 +08:00
class OperandRenderer {
public:
[globalisel][tablegen] Add support for EXTRACT_SUBREG. Summary: After this patch, we finally have test cases that require multiple instruction emission. Depends on D33590 Reviewers: ab, qcolombet, t.p.northover, rovka, kristof.beyls Subscribers: javed.absar, llvm-commits, igorb Differential Revision: https://reviews.llvm.org/D33596 llvm-svn: 306388
2017-06-27 18:11:39 +08:00
enum RendererKind {
OR_Copy,
[globalisel][tablegen] Import stores and allow GISel to automatically substitute zero regs like WZR/XZR/$zero. This patch enables the import of stores. Unfortunately, doing so by itself, loses an optimization where storing 0 to memory makes use of WZR/XZR. To mitigate this, this patch also introduces a new feature that allows register operands to nominate a zero register. When this is done, GlobalISel will substitute (G_CONSTANT 0) with the nominated register automatically. This is currently configured to only apply to the stores. Applying it to GPR32/GPR64 register classes in general will be done after review see (https://reviews.llvm.org/D39150). llvm-svn: 316360
2017-10-24 02:19:24 +08:00
OR_CopyOrAddZeroReg,
[globalisel][tablegen] Add support for EXTRACT_SUBREG. Summary: After this patch, we finally have test cases that require multiple instruction emission. Depends on D33590 Reviewers: ab, qcolombet, t.p.northover, rovka, kristof.beyls Subscribers: javed.absar, llvm-commits, igorb Differential Revision: https://reviews.llvm.org/D33596 llvm-svn: 306388
2017-06-27 18:11:39 +08:00
OR_CopySubReg,
[globalisel][tablegen] Add support for importing 'imm' operands. Summary: This patch enables the import of rules containing 'imm' operands that do not constrain the acceptable values using predicates. Support for ImmLeaf will arrive in a later patch. Depends on D35681 Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Reviewed By: rovka Subscribers: kristof.beyls, javed.absar, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D35833 llvm-svn: 310343
2017-08-08 18:44:31 +08:00
OR_CopyConstantAsImm,
[globalisel][tablegen] Add support for fpimm and import of APInt/APFloat based ImmLeaf. Summary: There's only a tablegen testcase for IntImmLeaf and not a CodeGen one because the relevant rules are rejected for other reasons at the moment. On AArch64, it's because there's an SDNodeXForm attached to the operand. On X86, it's because the rule either emits multiple instructions or has another predicate using PatFrag which cannot easily be supported at the same time. Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Reviewed By: qcolombet Subscribers: aemerson, javed.absar, igorb, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D36569 llvm-svn: 315761
2017-10-14 05:28:03 +08:00
OR_CopyFConstantAsFPImm,
[globalisel][tablegen] Add support for EXTRACT_SUBREG. Summary: After this patch, we finally have test cases that require multiple instruction emission. Depends on D33590 Reviewers: ab, qcolombet, t.p.northover, rovka, kristof.beyls Subscribers: javed.absar, llvm-commits, igorb Differential Revision: https://reviews.llvm.org/D33596 llvm-svn: 306388
2017-06-27 18:11:39 +08:00
OR_Imm,
OR_Register,
[globalisel][tablegen] Add support for multi-insn emission The importer will now accept nested instructions in the result pattern such as (ADDWrr $a, (SUBWrr $b, $c)). This is only valid when the nested instruction def's a single vreg and the parent instruction consumes a single vreg where a nested instruction is specified. The importer will automatically create a vreg to connect the two using the type information from the pattern. This vreg will be constrained to the register classes given in the instruction definitions*. * REG_SEQUENCE is explicitly rejected because of this. The definition doesn't constrain to a register class and it therefore needs special handling. llvm-svn: 317117
2017-11-02 03:57:57 +08:00
OR_TempRegister,
[GlobalISel][TableGen] Add support for SDNodeXForm Summary: This patch adds CustomRenderer which renders the matched operands to the specified instruction. Targets can enable the matching of SDNodeXForm by adding a definition that inherits from GICustomOperandRenderer and GISDNodeXFormEquiv as follows. def gi_imm8 : GICustomOperandRenderer<"renderImm8”>, GISDNodeXFormEquiv<imm8_xform>; Custom renderer functions should be of the form: void render(MachineInstrBuilder &MIB, const MachineInstr &I); Reviewers: dsanders, ab, rovka Reviewed By: dsanders Subscribers: kristof.beyls, javed.absar, llvm-commits, mgrang, qcolombet Differential Revision: https://reviews.llvm.org/D42012 llvm-svn: 322582
2018-01-17 02:44:05 +08:00
OR_ComplexPattern,
OR_Custom
[globalisel][tablegen] Add support for EXTRACT_SUBREG. Summary: After this patch, we finally have test cases that require multiple instruction emission. Depends on D33590 Reviewers: ab, qcolombet, t.p.northover, rovka, kristof.beyls Subscribers: javed.absar, llvm-commits, igorb Differential Revision: https://reviews.llvm.org/D33596 llvm-svn: 306388
2017-06-27 18:11:39 +08:00
};
[globalisel] Decouple src pattern operands from dst pattern operands. Summary: This isn't testable for AArch64 by itself so this patch also adds support for constant immediates in the pattern and physical register uses in the result. The new IntOperandMatcher matches the constant in patterns such as '(set $rd:GPR32, (G_XOR $rs:GPR32, -1))'. It's always safe to fold immediates into an instruction so this is the first rule that will match across multiple BB's. The Renderer hierarchy is responsible for adding operands to the result instruction. Renderers can copy operands (CopyRenderer) or add physical registers (in particular %wzr and %xzr) to the result instruction in any order (OperandMatchers now import the operand names from SelectionDAG to allow renderers to access any operand). This allows us to emit the result instruction for: %1 = G_XOR %0, -1 --> %1 = ORNWrr %wzr, %0 %1 = G_XOR -1, %0 --> %1 = ORNWrr %wzr, %0 although the latter is untested since the matcher/importer has not been taught about commutativity yet. Added BuildMIAction which can build new instructions and mutate them where possible. W.r.t the mutation aspect, MatchActions are now told the name of an instruction they can recycle and BuildMIAction will emit mutation code when the renderers are appropriate. They are appropriate when all operands are rendered using CopyRenderer and the indices are the same as the matcher. This currently assumes that all operands have at least one matcher. Finally, this change also fixes a crash in AArch64InstructionSelector::select() caused by an immediate operand passing isImm() rather than isCImm(). This was uncovered by the other changes and was detected by existing tests. Depends on D29711 Reviewers: t.p.northover, ab, qcolombet, rovka, aditya_nandakumar, javed.absar Reviewed By: rovka Subscribers: aemerson, dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D29712 llvm-svn: 296131
2017-02-24 23:43:30 +08:00
protected:
RendererKind Kind;
public:
OperandRenderer(RendererKind Kind) : Kind(Kind) {}
virtual ~OperandRenderer() {}
RendererKind getKind() const { return Kind; }
[globalisel][tablegen] Add control-flow to the MatchTable. Summary: This will allow us to merge the various sub-tables into a single table. This is a compile-time saving at this point. However, this will also enable the optimization of a table so that similar instructions can be tested together, reducing the time spent on the matching the code. The bulk of this patch is a mechanical conversion to the new MatchTable object which is responsible for tracking label definitions and filling in the index of the jump targets. It is also responsible for nicely formatting the table. This was necessary to support the new GIM_Try opcode which takes the index to jump to if the match should fail. This value is unknown during table construction and is filled in during emission. To support nesting try-blocks (although we currently don't emit tables with nested try-blocks), GIM_Reject has been re-introduced to explicitly exit a try-block or fail the overall match if there are no active try-blocks. Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Reviewed By: rovka Subscribers: kristof.beyls, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D35117 llvm-svn: 308596
2017-07-20 17:25:44 +08:00
virtual void emitRenderOpcodes(MatchTable &Table,
RuleMatcher &Rule) const = 0;
[globalisel] Decouple src pattern operands from dst pattern operands. Summary: This isn't testable for AArch64 by itself so this patch also adds support for constant immediates in the pattern and physical register uses in the result. The new IntOperandMatcher matches the constant in patterns such as '(set $rd:GPR32, (G_XOR $rs:GPR32, -1))'. It's always safe to fold immediates into an instruction so this is the first rule that will match across multiple BB's. The Renderer hierarchy is responsible for adding operands to the result instruction. Renderers can copy operands (CopyRenderer) or add physical registers (in particular %wzr and %xzr) to the result instruction in any order (OperandMatchers now import the operand names from SelectionDAG to allow renderers to access any operand). This allows us to emit the result instruction for: %1 = G_XOR %0, -1 --> %1 = ORNWrr %wzr, %0 %1 = G_XOR -1, %0 --> %1 = ORNWrr %wzr, %0 although the latter is untested since the matcher/importer has not been taught about commutativity yet. Added BuildMIAction which can build new instructions and mutate them where possible. W.r.t the mutation aspect, MatchActions are now told the name of an instruction they can recycle and BuildMIAction will emit mutation code when the renderers are appropriate. They are appropriate when all operands are rendered using CopyRenderer and the indices are the same as the matcher. This currently assumes that all operands have at least one matcher. Finally, this change also fixes a crash in AArch64InstructionSelector::select() caused by an immediate operand passing isImm() rather than isCImm(). This was uncovered by the other changes and was detected by existing tests. Depends on D29711 Reviewers: t.p.northover, ab, qcolombet, rovka, aditya_nandakumar, javed.absar Reviewed By: rovka Subscribers: aemerson, dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D29712 llvm-svn: 296131
2017-02-24 23:43:30 +08:00
};
/// A CopyRenderer emits code to copy a single operand from an existing
/// instruction to the one being built.
class CopyRenderer : public OperandRenderer {
protected:
[globalisel][tablegen] Added instruction emission to the state-machine-based matcher. Summary: This further improves the compile-time regressions that will be caused by a re-commit of r303259. Also added included preliminary work in preparation for the multi-insn emitter since I needed to change the relevant part of the API for this patch anyway. Depends on D33758 Reviewers: rovka, vitalybuka, ab, t.p.northover, qcolombet, aditya_nandakumar Reviewed By: ab Subscribers: kristof.beyls, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D33764 llvm-svn: 307133
2017-07-05 17:39:33 +08:00
unsigned NewInsnID;
[globalisel] Decouple src pattern operands from dst pattern operands. Summary: This isn't testable for AArch64 by itself so this patch also adds support for constant immediates in the pattern and physical register uses in the result. The new IntOperandMatcher matches the constant in patterns such as '(set $rd:GPR32, (G_XOR $rs:GPR32, -1))'. It's always safe to fold immediates into an instruction so this is the first rule that will match across multiple BB's. The Renderer hierarchy is responsible for adding operands to the result instruction. Renderers can copy operands (CopyRenderer) or add physical registers (in particular %wzr and %xzr) to the result instruction in any order (OperandMatchers now import the operand names from SelectionDAG to allow renderers to access any operand). This allows us to emit the result instruction for: %1 = G_XOR %0, -1 --> %1 = ORNWrr %wzr, %0 %1 = G_XOR -1, %0 --> %1 = ORNWrr %wzr, %0 although the latter is untested since the matcher/importer has not been taught about commutativity yet. Added BuildMIAction which can build new instructions and mutate them where possible. W.r.t the mutation aspect, MatchActions are now told the name of an instruction they can recycle and BuildMIAction will emit mutation code when the renderers are appropriate. They are appropriate when all operands are rendered using CopyRenderer and the indices are the same as the matcher. This currently assumes that all operands have at least one matcher. Finally, this change also fixes a crash in AArch64InstructionSelector::select() caused by an immediate operand passing isImm() rather than isCImm(). This was uncovered by the other changes and was detected by existing tests. Depends on D29711 Reviewers: t.p.northover, ab, qcolombet, rovka, aditya_nandakumar, javed.absar Reviewed By: rovka Subscribers: aemerson, dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D29712 llvm-svn: 296131
2017-02-24 23:43:30 +08:00
/// The name of the operand.
const StringRef SymbolicName;
public:
[globalisel][tablegen] Remove unused InstructionMatcher's. NFC llvm-svn: 316407
2017-10-24 09:48:34 +08:00
CopyRenderer(unsigned NewInsnID, StringRef SymbolicName)
: OperandRenderer(OR_Copy), NewInsnID(NewInsnID),
[globalisel][tablegen] Add support for importing 'imm' operands. Summary: This patch enables the import of rules containing 'imm' operands that do not constrain the acceptable values using predicates. Support for ImmLeaf will arrive in a later patch. Depends on D35681 Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Reviewed By: rovka Subscribers: kristof.beyls, javed.absar, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D35833 llvm-svn: 310343
2017-08-08 18:44:31 +08:00
SymbolicName(SymbolicName) {
assert(!SymbolicName.empty() && "Cannot copy from an unspecified source");
}
[globalisel] Decouple src pattern operands from dst pattern operands. Summary: This isn't testable for AArch64 by itself so this patch also adds support for constant immediates in the pattern and physical register uses in the result. The new IntOperandMatcher matches the constant in patterns such as '(set $rd:GPR32, (G_XOR $rs:GPR32, -1))'. It's always safe to fold immediates into an instruction so this is the first rule that will match across multiple BB's. The Renderer hierarchy is responsible for adding operands to the result instruction. Renderers can copy operands (CopyRenderer) or add physical registers (in particular %wzr and %xzr) to the result instruction in any order (OperandMatchers now import the operand names from SelectionDAG to allow renderers to access any operand). This allows us to emit the result instruction for: %1 = G_XOR %0, -1 --> %1 = ORNWrr %wzr, %0 %1 = G_XOR -1, %0 --> %1 = ORNWrr %wzr, %0 although the latter is untested since the matcher/importer has not been taught about commutativity yet. Added BuildMIAction which can build new instructions and mutate them where possible. W.r.t the mutation aspect, MatchActions are now told the name of an instruction they can recycle and BuildMIAction will emit mutation code when the renderers are appropriate. They are appropriate when all operands are rendered using CopyRenderer and the indices are the same as the matcher. This currently assumes that all operands have at least one matcher. Finally, this change also fixes a crash in AArch64InstructionSelector::select() caused by an immediate operand passing isImm() rather than isCImm(). This was uncovered by the other changes and was detected by existing tests. Depends on D29711 Reviewers: t.p.northover, ab, qcolombet, rovka, aditya_nandakumar, javed.absar Reviewed By: rovka Subscribers: aemerson, dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D29712 llvm-svn: 296131
2017-02-24 23:43:30 +08:00
static bool classof(const OperandRenderer *R) {
return R->getKind() == OR_Copy;
}
const StringRef getSymbolicName() const { return SymbolicName; }
[globalisel][tablegen] Add control-flow to the MatchTable. Summary: This will allow us to merge the various sub-tables into a single table. This is a compile-time saving at this point. However, this will also enable the optimization of a table so that similar instructions can be tested together, reducing the time spent on the matching the code. The bulk of this patch is a mechanical conversion to the new MatchTable object which is responsible for tracking label definitions and filling in the index of the jump targets. It is also responsible for nicely formatting the table. This was necessary to support the new GIM_Try opcode which takes the index to jump to if the match should fail. This value is unknown during table construction and is filled in during emission. To support nesting try-blocks (although we currently don't emit tables with nested try-blocks), GIM_Reject has been re-introduced to explicitly exit a try-block or fail the overall match if there are no active try-blocks. Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Reviewed By: rovka Subscribers: kristof.beyls, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D35117 llvm-svn: 308596
2017-07-20 17:25:44 +08:00
void emitRenderOpcodes(MatchTable &Table, RuleMatcher &Rule) const override {
[globalisel][tablegen] Simplify named operand/operator lookups and fix a wrong-code bug this revealed. Summary: Operand variable lookups are now performed by the RuleMatcher rather than searching the whole matcher hierarchy for a match. This revealed a wrong-code bug that currently affects ARM and X86 where patterns that use a variable more than once in the match pattern will be imported but won't check that the operands are identical. This can cause the tablegen-erated matcher to accept matches that should be rejected. Depends on D36569 Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Subscribers: aemerson, igorb, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D36618 llvm-svn: 315780
2017-10-14 08:31:58 +08:00
const OperandMatcher &Operand = Rule.getOperandMatcher(SymbolicName);
[globalisel][tablegen] Added instruction emission to the state-machine-based matcher. Summary: This further improves the compile-time regressions that will be caused by a re-commit of r303259. Also added included preliminary work in preparation for the multi-insn emitter since I needed to change the relevant part of the API for this patch anyway. Depends on D33758 Reviewers: rovka, vitalybuka, ab, t.p.northover, qcolombet, aditya_nandakumar Reviewed By: ab Subscribers: kristof.beyls, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D33764 llvm-svn: 307133
2017-07-05 17:39:33 +08:00
unsigned OldInsnVarID = Rule.getInsnVarID(Operand.getInstructionMatcher());
[globalisel][tablegen] Add control-flow to the MatchTable. Summary: This will allow us to merge the various sub-tables into a single table. This is a compile-time saving at this point. However, this will also enable the optimization of a table so that similar instructions can be tested together, reducing the time spent on the matching the code. The bulk of this patch is a mechanical conversion to the new MatchTable object which is responsible for tracking label definitions and filling in the index of the jump targets. It is also responsible for nicely formatting the table. This was necessary to support the new GIM_Try opcode which takes the index to jump to if the match should fail. This value is unknown during table construction and is filled in during emission. To support nesting try-blocks (although we currently don't emit tables with nested try-blocks), GIM_Reject has been re-introduced to explicitly exit a try-block or fail the overall match if there are no active try-blocks. Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Reviewed By: rovka Subscribers: kristof.beyls, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D35117 llvm-svn: 308596
2017-07-20 17:25:44 +08:00
Table << MatchTable::Opcode("GIR_Copy") << MatchTable::Comment("NewInsnID")
<< MatchTable::IntValue(NewInsnID) << MatchTable::Comment("OldInsnID")
<< MatchTable::IntValue(OldInsnVarID) << MatchTable::Comment("OpIdx")
Reapply r332907 "[GlobalISel] Improving InstructionSelect's performance by reducing MatchTable..." Apparently the compile time problem was caused by the fact that not all compilers / STL implementations can automatically convert std::unique_ptr<Derived> to std::unique_ptr<Base>. Fixed (hopefully) by making sure it's std::unique_ptr<Derived>&& (rvalue ref) to std::unique_ptr<Base> conversion instead. llvm-svn: 332917
2018-05-22 07:28:51 +08:00
<< MatchTable::IntValue(Operand.getOpIdx())
[globalisel][tablegen] Add control-flow to the MatchTable. Summary: This will allow us to merge the various sub-tables into a single table. This is a compile-time saving at this point. However, this will also enable the optimization of a table so that similar instructions can be tested together, reducing the time spent on the matching the code. The bulk of this patch is a mechanical conversion to the new MatchTable object which is responsible for tracking label definitions and filling in the index of the jump targets. It is also responsible for nicely formatting the table. This was necessary to support the new GIM_Try opcode which takes the index to jump to if the match should fail. This value is unknown during table construction and is filled in during emission. To support nesting try-blocks (although we currently don't emit tables with nested try-blocks), GIM_Reject has been re-introduced to explicitly exit a try-block or fail the overall match if there are no active try-blocks. Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Reviewed By: rovka Subscribers: kristof.beyls, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D35117 llvm-svn: 308596
2017-07-20 17:25:44 +08:00
<< MatchTable::Comment(SymbolicName) << MatchTable::LineBreak;
[globalisel] Decouple src pattern operands from dst pattern operands. Summary: This isn't testable for AArch64 by itself so this patch also adds support for constant immediates in the pattern and physical register uses in the result. The new IntOperandMatcher matches the constant in patterns such as '(set $rd:GPR32, (G_XOR $rs:GPR32, -1))'. It's always safe to fold immediates into an instruction so this is the first rule that will match across multiple BB's. The Renderer hierarchy is responsible for adding operands to the result instruction. Renderers can copy operands (CopyRenderer) or add physical registers (in particular %wzr and %xzr) to the result instruction in any order (OperandMatchers now import the operand names from SelectionDAG to allow renderers to access any operand). This allows us to emit the result instruction for: %1 = G_XOR %0, -1 --> %1 = ORNWrr %wzr, %0 %1 = G_XOR -1, %0 --> %1 = ORNWrr %wzr, %0 although the latter is untested since the matcher/importer has not been taught about commutativity yet. Added BuildMIAction which can build new instructions and mutate them where possible. W.r.t the mutation aspect, MatchActions are now told the name of an instruction they can recycle and BuildMIAction will emit mutation code when the renderers are appropriate. They are appropriate when all operands are rendered using CopyRenderer and the indices are the same as the matcher. This currently assumes that all operands have at least one matcher. Finally, this change also fixes a crash in AArch64InstructionSelector::select() caused by an immediate operand passing isImm() rather than isCImm(). This was uncovered by the other changes and was detected by existing tests. Depends on D29711 Reviewers: t.p.northover, ab, qcolombet, rovka, aditya_nandakumar, javed.absar Reviewed By: rovka Subscribers: aemerson, dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D29712 llvm-svn: 296131
2017-02-24 23:43:30 +08:00
}
};
[globalisel][tablegen] Import stores and allow GISel to automatically substitute zero regs like WZR/XZR/$zero. This patch enables the import of stores. Unfortunately, doing so by itself, loses an optimization where storing 0 to memory makes use of WZR/XZR. To mitigate this, this patch also introduces a new feature that allows register operands to nominate a zero register. When this is done, GlobalISel will substitute (G_CONSTANT 0) with the nominated register automatically. This is currently configured to only apply to the stores. Applying it to GPR32/GPR64 register classes in general will be done after review see (https://reviews.llvm.org/D39150). llvm-svn: 316360
2017-10-24 02:19:24 +08:00
/// A CopyOrAddZeroRegRenderer emits code to copy a single operand from an
/// existing instruction to the one being built. If the operand turns out to be
/// a 'G_CONSTANT 0' then it replaces the operand with a zero register.
class CopyOrAddZeroRegRenderer : public OperandRenderer {
protected:
unsigned NewInsnID;
/// The name of the operand.
const StringRef SymbolicName;
const Record *ZeroRegisterDef;
public:
CopyOrAddZeroRegRenderer(unsigned NewInsnID,
StringRef SymbolicName, Record *ZeroRegisterDef)
: OperandRenderer(OR_CopyOrAddZeroReg), NewInsnID(NewInsnID),
SymbolicName(SymbolicName), ZeroRegisterDef(ZeroRegisterDef) {
assert(!SymbolicName.empty() && "Cannot copy from an unspecified source");
}
static bool classof(const OperandRenderer *R) {
return R->getKind() == OR_CopyOrAddZeroReg;
}
const StringRef getSymbolicName() const { return SymbolicName; }
void emitRenderOpcodes(MatchTable &Table, RuleMatcher &Rule) const override {
const OperandMatcher &Operand = Rule.getOperandMatcher(SymbolicName);
unsigned OldInsnVarID = Rule.getInsnVarID(Operand.getInstructionMatcher());
Table << MatchTable::Opcode("GIR_CopyOrAddZeroReg")
<< MatchTable::Comment("NewInsnID") << MatchTable::IntValue(NewInsnID)
<< MatchTable::Comment("OldInsnID")
<< MatchTable::IntValue(OldInsnVarID) << MatchTable::Comment("OpIdx")
Reapply r332907 "[GlobalISel] Improving InstructionSelect's performance by reducing MatchTable..." Apparently the compile time problem was caused by the fact that not all compilers / STL implementations can automatically convert std::unique_ptr<Derived> to std::unique_ptr<Base>. Fixed (hopefully) by making sure it's std::unique_ptr<Derived>&& (rvalue ref) to std::unique_ptr<Base> conversion instead. llvm-svn: 332917
2018-05-22 07:28:51 +08:00
<< MatchTable::IntValue(Operand.getOpIdx())
[globalisel][tablegen] Import stores and allow GISel to automatically substitute zero regs like WZR/XZR/$zero. This patch enables the import of stores. Unfortunately, doing so by itself, loses an optimization where storing 0 to memory makes use of WZR/XZR. To mitigate this, this patch also introduces a new feature that allows register operands to nominate a zero register. When this is done, GlobalISel will substitute (G_CONSTANT 0) with the nominated register automatically. This is currently configured to only apply to the stores. Applying it to GPR32/GPR64 register classes in general will be done after review see (https://reviews.llvm.org/D39150). llvm-svn: 316360
2017-10-24 02:19:24 +08:00
<< MatchTable::NamedValue(
(ZeroRegisterDef->getValue("Namespace")
? ZeroRegisterDef->getValueAsString("Namespace")
: ""),
ZeroRegisterDef->getName())
<< MatchTable::Comment(SymbolicName) << MatchTable::LineBreak;
}
};
[globalisel][tablegen] Add support for importing 'imm' operands. Summary: This patch enables the import of rules containing 'imm' operands that do not constrain the acceptable values using predicates. Support for ImmLeaf will arrive in a later patch. Depends on D35681 Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Reviewed By: rovka Subscribers: kristof.beyls, javed.absar, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D35833 llvm-svn: 310343
2017-08-08 18:44:31 +08:00
/// A CopyConstantAsImmRenderer emits code to render a G_CONSTANT instruction to
/// an extended immediate operand.
class CopyConstantAsImmRenderer : public OperandRenderer {
protected:
unsigned NewInsnID;
/// The name of the operand.
const std::string SymbolicName;
bool Signed;
public:
CopyConstantAsImmRenderer(unsigned NewInsnID, StringRef SymbolicName)
: OperandRenderer(OR_CopyConstantAsImm), NewInsnID(NewInsnID),
SymbolicName(SymbolicName), Signed(true) {}
static bool classof(const OperandRenderer *R) {
return R->getKind() == OR_CopyConstantAsImm;
}
const StringRef getSymbolicName() const { return SymbolicName; }
void emitRenderOpcodes(MatchTable &Table, RuleMatcher &Rule) const override {
Reapply r332907 "[GlobalISel] Improving InstructionSelect's performance by reducing MatchTable..." Apparently the compile time problem was caused by the fact that not all compilers / STL implementations can automatically convert std::unique_ptr<Derived> to std::unique_ptr<Base>. Fixed (hopefully) by making sure it's std::unique_ptr<Derived>&& (rvalue ref) to std::unique_ptr<Base> conversion instead. llvm-svn: 332917
2018-05-22 07:28:51 +08:00
InstructionMatcher &InsnMatcher = Rule.getInstructionMatcher(SymbolicName);
[globalisel][tablegen] Add support for importing 'imm' operands. Summary: This patch enables the import of rules containing 'imm' operands that do not constrain the acceptable values using predicates. Support for ImmLeaf will arrive in a later patch. Depends on D35681 Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Reviewed By: rovka Subscribers: kristof.beyls, javed.absar, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D35833 llvm-svn: 310343
2017-08-08 18:44:31 +08:00
unsigned OldInsnVarID = Rule.getInsnVarID(InsnMatcher);
Table << MatchTable::Opcode(Signed ? "GIR_CopyConstantAsSImm"
: "GIR_CopyConstantAsUImm")
<< MatchTable::Comment("NewInsnID") << MatchTable::IntValue(NewInsnID)
<< MatchTable::Comment("OldInsnID")
<< MatchTable::IntValue(OldInsnVarID)
<< MatchTable::Comment(SymbolicName) << MatchTable::LineBreak;
}
};
[globalisel][tablegen] Add support for fpimm and import of APInt/APFloat based ImmLeaf. Summary: There's only a tablegen testcase for IntImmLeaf and not a CodeGen one because the relevant rules are rejected for other reasons at the moment. On AArch64, it's because there's an SDNodeXForm attached to the operand. On X86, it's because the rule either emits multiple instructions or has another predicate using PatFrag which cannot easily be supported at the same time. Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Reviewed By: qcolombet Subscribers: aemerson, javed.absar, igorb, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D36569 llvm-svn: 315761
2017-10-14 05:28:03 +08:00
/// A CopyFConstantAsFPImmRenderer emits code to render a G_FCONSTANT
/// instruction to an extended immediate operand.
class CopyFConstantAsFPImmRenderer : public OperandRenderer {
protected:
unsigned NewInsnID;
/// The name of the operand.
const std::string SymbolicName;
public:
CopyFConstantAsFPImmRenderer(unsigned NewInsnID, StringRef SymbolicName)
: OperandRenderer(OR_CopyFConstantAsFPImm), NewInsnID(NewInsnID),
SymbolicName(SymbolicName) {}
static bool classof(const OperandRenderer *R) {
return R->getKind() == OR_CopyFConstantAsFPImm;
}
const StringRef getSymbolicName() const { return SymbolicName; }
void emitRenderOpcodes(MatchTable &Table, RuleMatcher &Rule) const override {
Reapply r332907 "[GlobalISel] Improving InstructionSelect's performance by reducing MatchTable..." Apparently the compile time problem was caused by the fact that not all compilers / STL implementations can automatically convert std::unique_ptr<Derived> to std::unique_ptr<Base>. Fixed (hopefully) by making sure it's std::unique_ptr<Derived>&& (rvalue ref) to std::unique_ptr<Base> conversion instead. llvm-svn: 332917
2018-05-22 07:28:51 +08:00
InstructionMatcher &InsnMatcher = Rule.getInstructionMatcher(SymbolicName);
[globalisel][tablegen] Add support for fpimm and import of APInt/APFloat based ImmLeaf. Summary: There's only a tablegen testcase for IntImmLeaf and not a CodeGen one because the relevant rules are rejected for other reasons at the moment. On AArch64, it's because there's an SDNodeXForm attached to the operand. On X86, it's because the rule either emits multiple instructions or has another predicate using PatFrag which cannot easily be supported at the same time. Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Reviewed By: qcolombet Subscribers: aemerson, javed.absar, igorb, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D36569 llvm-svn: 315761
2017-10-14 05:28:03 +08:00
unsigned OldInsnVarID = Rule.getInsnVarID(InsnMatcher);
Table << MatchTable::Opcode("GIR_CopyFConstantAsFPImm")
<< MatchTable::Comment("NewInsnID") << MatchTable::IntValue(NewInsnID)
<< MatchTable::Comment("OldInsnID")
<< MatchTable::IntValue(OldInsnVarID)
<< MatchTable::Comment(SymbolicName) << MatchTable::LineBreak;
}
};
[globalisel][tablegen] Add support for EXTRACT_SUBREG. Summary: After this patch, we finally have test cases that require multiple instruction emission. Depends on D33590 Reviewers: ab, qcolombet, t.p.northover, rovka, kristof.beyls Subscribers: javed.absar, llvm-commits, igorb Differential Revision: https://reviews.llvm.org/D33596 llvm-svn: 306388
2017-06-27 18:11:39 +08:00
/// A CopySubRegRenderer emits code to copy a single register operand from an
/// existing instruction to the one being built and indicate that only a
/// subregister should be copied.
class CopySubRegRenderer : public OperandRenderer {
protected:
[globalisel][tablegen] Added instruction emission to the state-machine-based matcher. Summary: This further improves the compile-time regressions that will be caused by a re-commit of r303259. Also added included preliminary work in preparation for the multi-insn emitter since I needed to change the relevant part of the API for this patch anyway. Depends on D33758 Reviewers: rovka, vitalybuka, ab, t.p.northover, qcolombet, aditya_nandakumar Reviewed By: ab Subscribers: kristof.beyls, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D33764 llvm-svn: 307133
2017-07-05 17:39:33 +08:00
unsigned NewInsnID;
[globalisel][tablegen] Add support for EXTRACT_SUBREG. Summary: After this patch, we finally have test cases that require multiple instruction emission. Depends on D33590 Reviewers: ab, qcolombet, t.p.northover, rovka, kristof.beyls Subscribers: javed.absar, llvm-commits, igorb Differential Revision: https://reviews.llvm.org/D33596 llvm-svn: 306388
2017-06-27 18:11:39 +08:00
/// The name of the operand.
const StringRef SymbolicName;
/// The subregister to extract.
const CodeGenSubRegIndex *SubReg;
public:
[globalisel][tablegen] Remove unused InstructionMatcher's. NFC llvm-svn: 316407
2017-10-24 09:48:34 +08:00
CopySubRegRenderer(unsigned NewInsnID, StringRef SymbolicName,
const CodeGenSubRegIndex *SubReg)
: OperandRenderer(OR_CopySubReg), NewInsnID(NewInsnID),
[globalisel][tablegen] Add support for EXTRACT_SUBREG. Summary: After this patch, we finally have test cases that require multiple instruction emission. Depends on D33590 Reviewers: ab, qcolombet, t.p.northover, rovka, kristof.beyls Subscribers: javed.absar, llvm-commits, igorb Differential Revision: https://reviews.llvm.org/D33596 llvm-svn: 306388
2017-06-27 18:11:39 +08:00
SymbolicName(SymbolicName), SubReg(SubReg) {}
static bool classof(const OperandRenderer *R) {
return R->getKind() == OR_CopySubReg;
}
const StringRef getSymbolicName() const { return SymbolicName; }
[globalisel][tablegen] Add control-flow to the MatchTable. Summary: This will allow us to merge the various sub-tables into a single table. This is a compile-time saving at this point. However, this will also enable the optimization of a table so that similar instructions can be tested together, reducing the time spent on the matching the code. The bulk of this patch is a mechanical conversion to the new MatchTable object which is responsible for tracking label definitions and filling in the index of the jump targets. It is also responsible for nicely formatting the table. This was necessary to support the new GIM_Try opcode which takes the index to jump to if the match should fail. This value is unknown during table construction and is filled in during emission. To support nesting try-blocks (although we currently don't emit tables with nested try-blocks), GIM_Reject has been re-introduced to explicitly exit a try-block or fail the overall match if there are no active try-blocks. Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Reviewed By: rovka Subscribers: kristof.beyls, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D35117 llvm-svn: 308596
2017-07-20 17:25:44 +08:00
void emitRenderOpcodes(MatchTable &Table, RuleMatcher &Rule) const override {
[globalisel][tablegen] Simplify named operand/operator lookups and fix a wrong-code bug this revealed. Summary: Operand variable lookups are now performed by the RuleMatcher rather than searching the whole matcher hierarchy for a match. This revealed a wrong-code bug that currently affects ARM and X86 where patterns that use a variable more than once in the match pattern will be imported but won't check that the operands are identical. This can cause the tablegen-erated matcher to accept matches that should be rejected. Depends on D36569 Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Subscribers: aemerson, igorb, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D36618 llvm-svn: 315780
2017-10-14 08:31:58 +08:00
const OperandMatcher &Operand = Rule.getOperandMatcher(SymbolicName);
[globalisel][tablegen] Added instruction emission to the state-machine-based matcher. Summary: This further improves the compile-time regressions that will be caused by a re-commit of r303259. Also added included preliminary work in preparation for the multi-insn emitter since I needed to change the relevant part of the API for this patch anyway. Depends on D33758 Reviewers: rovka, vitalybuka, ab, t.p.northover, qcolombet, aditya_nandakumar Reviewed By: ab Subscribers: kristof.beyls, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D33764 llvm-svn: 307133
2017-07-05 17:39:33 +08:00
unsigned OldInsnVarID = Rule.getInsnVarID(Operand.getInstructionMatcher());
[globalisel][tablegen] Add control-flow to the MatchTable. Summary: This will allow us to merge the various sub-tables into a single table. This is a compile-time saving at this point. However, this will also enable the optimization of a table so that similar instructions can be tested together, reducing the time spent on the matching the code. The bulk of this patch is a mechanical conversion to the new MatchTable object which is responsible for tracking label definitions and filling in the index of the jump targets. It is also responsible for nicely formatting the table. This was necessary to support the new GIM_Try opcode which takes the index to jump to if the match should fail. This value is unknown during table construction and is filled in during emission. To support nesting try-blocks (although we currently don't emit tables with nested try-blocks), GIM_Reject has been re-introduced to explicitly exit a try-block or fail the overall match if there are no active try-blocks. Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Reviewed By: rovka Subscribers: kristof.beyls, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D35117 llvm-svn: 308596
2017-07-20 17:25:44 +08:00
Table << MatchTable::Opcode("GIR_CopySubReg")
<< MatchTable::Comment("NewInsnID") << MatchTable::IntValue(NewInsnID)
<< MatchTable::Comment("OldInsnID")
<< MatchTable::IntValue(OldInsnVarID) << MatchTable::Comment("OpIdx")
Reapply r332907 "[GlobalISel] Improving InstructionSelect's performance by reducing MatchTable..." Apparently the compile time problem was caused by the fact that not all compilers / STL implementations can automatically convert std::unique_ptr<Derived> to std::unique_ptr<Base>. Fixed (hopefully) by making sure it's std::unique_ptr<Derived>&& (rvalue ref) to std::unique_ptr<Base> conversion instead. llvm-svn: 332917
2018-05-22 07:28:51 +08:00
<< MatchTable::IntValue(Operand.getOpIdx())
[globalisel][tablegen] Add control-flow to the MatchTable. Summary: This will allow us to merge the various sub-tables into a single table. This is a compile-time saving at this point. However, this will also enable the optimization of a table so that similar instructions can be tested together, reducing the time spent on the matching the code. The bulk of this patch is a mechanical conversion to the new MatchTable object which is responsible for tracking label definitions and filling in the index of the jump targets. It is also responsible for nicely formatting the table. This was necessary to support the new GIM_Try opcode which takes the index to jump to if the match should fail. This value is unknown during table construction and is filled in during emission. To support nesting try-blocks (although we currently don't emit tables with nested try-blocks), GIM_Reject has been re-introduced to explicitly exit a try-block or fail the overall match if there are no active try-blocks. Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Reviewed By: rovka Subscribers: kristof.beyls, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D35117 llvm-svn: 308596
2017-07-20 17:25:44 +08:00
<< MatchTable::Comment("SubRegIdx")
<< MatchTable::IntValue(SubReg->EnumValue)
<< MatchTable::Comment(SymbolicName) << MatchTable::LineBreak;
[globalisel][tablegen] Add support for EXTRACT_SUBREG. Summary: After this patch, we finally have test cases that require multiple instruction emission. Depends on D33590 Reviewers: ab, qcolombet, t.p.northover, rovka, kristof.beyls Subscribers: javed.absar, llvm-commits, igorb Differential Revision: https://reviews.llvm.org/D33596 llvm-svn: 306388
2017-06-27 18:11:39 +08:00
}
};
[globalisel] Decouple src pattern operands from dst pattern operands. Summary: This isn't testable for AArch64 by itself so this patch also adds support for constant immediates in the pattern and physical register uses in the result. The new IntOperandMatcher matches the constant in patterns such as '(set $rd:GPR32, (G_XOR $rs:GPR32, -1))'. It's always safe to fold immediates into an instruction so this is the first rule that will match across multiple BB's. The Renderer hierarchy is responsible for adding operands to the result instruction. Renderers can copy operands (CopyRenderer) or add physical registers (in particular %wzr and %xzr) to the result instruction in any order (OperandMatchers now import the operand names from SelectionDAG to allow renderers to access any operand). This allows us to emit the result instruction for: %1 = G_XOR %0, -1 --> %1 = ORNWrr %wzr, %0 %1 = G_XOR -1, %0 --> %1 = ORNWrr %wzr, %0 although the latter is untested since the matcher/importer has not been taught about commutativity yet. Added BuildMIAction which can build new instructions and mutate them where possible. W.r.t the mutation aspect, MatchActions are now told the name of an instruction they can recycle and BuildMIAction will emit mutation code when the renderers are appropriate. They are appropriate when all operands are rendered using CopyRenderer and the indices are the same as the matcher. This currently assumes that all operands have at least one matcher. Finally, this change also fixes a crash in AArch64InstructionSelector::select() caused by an immediate operand passing isImm() rather than isCImm(). This was uncovered by the other changes and was detected by existing tests. Depends on D29711 Reviewers: t.p.northover, ab, qcolombet, rovka, aditya_nandakumar, javed.absar Reviewed By: rovka Subscribers: aemerson, dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D29712 llvm-svn: 296131
2017-02-24 23:43:30 +08:00
/// Adds a specific physical register to the instruction being built.
/// This is typically useful for WZR/XZR on AArch64.
class AddRegisterRenderer : public OperandRenderer {
protected:
[globalisel][tablegen] Added instruction emission to the state-machine-based matcher. Summary: This further improves the compile-time regressions that will be caused by a re-commit of r303259. Also added included preliminary work in preparation for the multi-insn emitter since I needed to change the relevant part of the API for this patch anyway. Depends on D33758 Reviewers: rovka, vitalybuka, ab, t.p.northover, qcolombet, aditya_nandakumar Reviewed By: ab Subscribers: kristof.beyls, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D33764 llvm-svn: 307133
2017-07-05 17:39:33 +08:00
unsigned InsnID;
[globalisel] Decouple src pattern operands from dst pattern operands. Summary: This isn't testable for AArch64 by itself so this patch also adds support for constant immediates in the pattern and physical register uses in the result. The new IntOperandMatcher matches the constant in patterns such as '(set $rd:GPR32, (G_XOR $rs:GPR32, -1))'. It's always safe to fold immediates into an instruction so this is the first rule that will match across multiple BB's. The Renderer hierarchy is responsible for adding operands to the result instruction. Renderers can copy operands (CopyRenderer) or add physical registers (in particular %wzr and %xzr) to the result instruction in any order (OperandMatchers now import the operand names from SelectionDAG to allow renderers to access any operand). This allows us to emit the result instruction for: %1 = G_XOR %0, -1 --> %1 = ORNWrr %wzr, %0 %1 = G_XOR -1, %0 --> %1 = ORNWrr %wzr, %0 although the latter is untested since the matcher/importer has not been taught about commutativity yet. Added BuildMIAction which can build new instructions and mutate them where possible. W.r.t the mutation aspect, MatchActions are now told the name of an instruction they can recycle and BuildMIAction will emit mutation code when the renderers are appropriate. They are appropriate when all operands are rendered using CopyRenderer and the indices are the same as the matcher. This currently assumes that all operands have at least one matcher. Finally, this change also fixes a crash in AArch64InstructionSelector::select() caused by an immediate operand passing isImm() rather than isCImm(). This was uncovered by the other changes and was detected by existing tests. Depends on D29711 Reviewers: t.p.northover, ab, qcolombet, rovka, aditya_nandakumar, javed.absar Reviewed By: rovka Subscribers: aemerson, dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D29712 llvm-svn: 296131
2017-02-24 23:43:30 +08:00
const Record *RegisterDef;
public:
[globalisel][tablegen] Added instruction emission to the state-machine-based matcher. Summary: This further improves the compile-time regressions that will be caused by a re-commit of r303259. Also added included preliminary work in preparation for the multi-insn emitter since I needed to change the relevant part of the API for this patch anyway. Depends on D33758 Reviewers: rovka, vitalybuka, ab, t.p.northover, qcolombet, aditya_nandakumar Reviewed By: ab Subscribers: kristof.beyls, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D33764 llvm-svn: 307133
2017-07-05 17:39:33 +08:00
AddRegisterRenderer(unsigned InsnID, const Record *RegisterDef)
: OperandRenderer(OR_Register), InsnID(InsnID), RegisterDef(RegisterDef) {
}
[globalisel] Decouple src pattern operands from dst pattern operands. Summary: This isn't testable for AArch64 by itself so this patch also adds support for constant immediates in the pattern and physical register uses in the result. The new IntOperandMatcher matches the constant in patterns such as '(set $rd:GPR32, (G_XOR $rs:GPR32, -1))'. It's always safe to fold immediates into an instruction so this is the first rule that will match across multiple BB's. The Renderer hierarchy is responsible for adding operands to the result instruction. Renderers can copy operands (CopyRenderer) or add physical registers (in particular %wzr and %xzr) to the result instruction in any order (OperandMatchers now import the operand names from SelectionDAG to allow renderers to access any operand). This allows us to emit the result instruction for: %1 = G_XOR %0, -1 --> %1 = ORNWrr %wzr, %0 %1 = G_XOR -1, %0 --> %1 = ORNWrr %wzr, %0 although the latter is untested since the matcher/importer has not been taught about commutativity yet. Added BuildMIAction which can build new instructions and mutate them where possible. W.r.t the mutation aspect, MatchActions are now told the name of an instruction they can recycle and BuildMIAction will emit mutation code when the renderers are appropriate. They are appropriate when all operands are rendered using CopyRenderer and the indices are the same as the matcher. This currently assumes that all operands have at least one matcher. Finally, this change also fixes a crash in AArch64InstructionSelector::select() caused by an immediate operand passing isImm() rather than isCImm(). This was uncovered by the other changes and was detected by existing tests. Depends on D29711 Reviewers: t.p.northover, ab, qcolombet, rovka, aditya_nandakumar, javed.absar Reviewed By: rovka Subscribers: aemerson, dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D29712 llvm-svn: 296131
2017-02-24 23:43:30 +08:00
static bool classof(const OperandRenderer *R) {
return R->getKind() == OR_Register;
}
[globalisel][tablegen] Add control-flow to the MatchTable. Summary: This will allow us to merge the various sub-tables into a single table. This is a compile-time saving at this point. However, this will also enable the optimization of a table so that similar instructions can be tested together, reducing the time spent on the matching the code. The bulk of this patch is a mechanical conversion to the new MatchTable object which is responsible for tracking label definitions and filling in the index of the jump targets. It is also responsible for nicely formatting the table. This was necessary to support the new GIM_Try opcode which takes the index to jump to if the match should fail. This value is unknown during table construction and is filled in during emission. To support nesting try-blocks (although we currently don't emit tables with nested try-blocks), GIM_Reject has been re-introduced to explicitly exit a try-block or fail the overall match if there are no active try-blocks. Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Reviewed By: rovka Subscribers: kristof.beyls, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D35117 llvm-svn: 308596
2017-07-20 17:25:44 +08:00
void emitRenderOpcodes(MatchTable &Table, RuleMatcher &Rule) const override {
Table << MatchTable::Opcode("GIR_AddRegister")
<< MatchTable::Comment("InsnID") << MatchTable::IntValue(InsnID)
<< MatchTable::NamedValue(
(RegisterDef->getValue("Namespace")
? RegisterDef->getValueAsString("Namespace")
: ""),
RegisterDef->getName())
<< MatchTable::LineBreak;
[globalisel] Decouple src pattern operands from dst pattern operands. Summary: This isn't testable for AArch64 by itself so this patch also adds support for constant immediates in the pattern and physical register uses in the result. The new IntOperandMatcher matches the constant in patterns such as '(set $rd:GPR32, (G_XOR $rs:GPR32, -1))'. It's always safe to fold immediates into an instruction so this is the first rule that will match across multiple BB's. The Renderer hierarchy is responsible for adding operands to the result instruction. Renderers can copy operands (CopyRenderer) or add physical registers (in particular %wzr and %xzr) to the result instruction in any order (OperandMatchers now import the operand names from SelectionDAG to allow renderers to access any operand). This allows us to emit the result instruction for: %1 = G_XOR %0, -1 --> %1 = ORNWrr %wzr, %0 %1 = G_XOR -1, %0 --> %1 = ORNWrr %wzr, %0 although the latter is untested since the matcher/importer has not been taught about commutativity yet. Added BuildMIAction which can build new instructions and mutate them where possible. W.r.t the mutation aspect, MatchActions are now told the name of an instruction they can recycle and BuildMIAction will emit mutation code when the renderers are appropriate. They are appropriate when all operands are rendered using CopyRenderer and the indices are the same as the matcher. This currently assumes that all operands have at least one matcher. Finally, this change also fixes a crash in AArch64InstructionSelector::select() caused by an immediate operand passing isImm() rather than isCImm(). This was uncovered by the other changes and was detected by existing tests. Depends on D29711 Reviewers: t.p.northover, ab, qcolombet, rovka, aditya_nandakumar, javed.absar Reviewed By: rovka Subscribers: aemerson, dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D29712 llvm-svn: 296131
2017-02-24 23:43:30 +08:00
}
};
[globalisel][tablegen] Add support for multi-insn emission The importer will now accept nested instructions in the result pattern such as (ADDWrr $a, (SUBWrr $b, $c)). This is only valid when the nested instruction def's a single vreg and the parent instruction consumes a single vreg where a nested instruction is specified. The importer will automatically create a vreg to connect the two using the type information from the pattern. This vreg will be constrained to the register classes given in the instruction definitions*. * REG_SEQUENCE is explicitly rejected because of this. The definition doesn't constrain to a register class and it therefore needs special handling. llvm-svn: 317117
2017-11-02 03:57:57 +08:00
/// Adds a specific temporary virtual register to the instruction being built.
/// This is used to chain instructions together when emitting multiple
/// instructions.
class TempRegRenderer : public OperandRenderer {
protected:
unsigned InsnID;
unsigned TempRegID;
bool IsDef;
public:
TempRegRenderer(unsigned InsnID, unsigned TempRegID, bool IsDef = false)
: OperandRenderer(OR_Register), InsnID(InsnID), TempRegID(TempRegID),
IsDef(IsDef) {}
static bool classof(const OperandRenderer *R) {
return R->getKind() == OR_TempRegister;
}
void emitRenderOpcodes(MatchTable &Table, RuleMatcher &Rule) const override {
Table << MatchTable::Opcode("GIR_AddTempRegister")
<< MatchTable::Comment("InsnID") << MatchTable::IntValue(InsnID)
<< MatchTable::Comment("TempRegID") << MatchTable::IntValue(TempRegID)
<< MatchTable::Comment("TempRegFlags");
if (IsDef)
Table << MatchTable::NamedValue("RegState::Define");
else
Table << MatchTable::IntValue(0);
Table << MatchTable::LineBreak;
}
};
[globalisel][tablegen] Add experimental support for OperandWithDefaultOps, PredicateOperand, and OptionalDefOperand Summary: As far as instruction selection is concerned, all three appear to be same thing. Support for these operands is experimental since AArch64 doesn't make use of them and the in-tree targets that do use them (AMDGPU for OperandWithDefaultOps, AMDGPU/ARM/Hexagon/Lanai for PredicateOperand, and ARM for OperandWithDefaultOps) are not using tablegen-erated GlobalISel yet. Reviewers: rovka, aditya_nandakumar, t.p.northover, qcolombet, ab Reviewed By: rovka Subscribers: inglorion, aemerson, rengolin, mehdi_amini, dberris, kristof.beyls, igorb, tpr, llvm-commits Differential Revision: https://reviews.llvm.org/D31135 llvm-svn: 300037
2017-04-12 16:23:08 +08:00
/// Adds a specific immediate to the instruction being built.
class ImmRenderer : public OperandRenderer {
protected:
[globalisel][tablegen] Added instruction emission to the state-machine-based matcher. Summary: This further improves the compile-time regressions that will be caused by a re-commit of r303259. Also added included preliminary work in preparation for the multi-insn emitter since I needed to change the relevant part of the API for this patch anyway. Depends on D33758 Reviewers: rovka, vitalybuka, ab, t.p.northover, qcolombet, aditya_nandakumar Reviewed By: ab Subscribers: kristof.beyls, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D33764 llvm-svn: 307133
2017-07-05 17:39:33 +08:00
unsigned InsnID;
[globalisel][tablegen] Add experimental support for OperandWithDefaultOps, PredicateOperand, and OptionalDefOperand Summary: As far as instruction selection is concerned, all three appear to be same thing. Support for these operands is experimental since AArch64 doesn't make use of them and the in-tree targets that do use them (AMDGPU for OperandWithDefaultOps, AMDGPU/ARM/Hexagon/Lanai for PredicateOperand, and ARM for OperandWithDefaultOps) are not using tablegen-erated GlobalISel yet. Reviewers: rovka, aditya_nandakumar, t.p.northover, qcolombet, ab Reviewed By: rovka Subscribers: inglorion, aemerson, rengolin, mehdi_amini, dberris, kristof.beyls, igorb, tpr, llvm-commits Differential Revision: https://reviews.llvm.org/D31135 llvm-svn: 300037
2017-04-12 16:23:08 +08:00
int64_t Imm;
public:
[globalisel][tablegen] Added instruction emission to the state-machine-based matcher. Summary: This further improves the compile-time regressions that will be caused by a re-commit of r303259. Also added included preliminary work in preparation for the multi-insn emitter since I needed to change the relevant part of the API for this patch anyway. Depends on D33758 Reviewers: rovka, vitalybuka, ab, t.p.northover, qcolombet, aditya_nandakumar Reviewed By: ab Subscribers: kristof.beyls, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D33764 llvm-svn: 307133
2017-07-05 17:39:33 +08:00
ImmRenderer(unsigned InsnID, int64_t Imm)
: OperandRenderer(OR_Imm), InsnID(InsnID), Imm(Imm) {}
[globalisel][tablegen] Add experimental support for OperandWithDefaultOps, PredicateOperand, and OptionalDefOperand Summary: As far as instruction selection is concerned, all three appear to be same thing. Support for these operands is experimental since AArch64 doesn't make use of them and the in-tree targets that do use them (AMDGPU for OperandWithDefaultOps, AMDGPU/ARM/Hexagon/Lanai for PredicateOperand, and ARM for OperandWithDefaultOps) are not using tablegen-erated GlobalISel yet. Reviewers: rovka, aditya_nandakumar, t.p.northover, qcolombet, ab Reviewed By: rovka Subscribers: inglorion, aemerson, rengolin, mehdi_amini, dberris, kristof.beyls, igorb, tpr, llvm-commits Differential Revision: https://reviews.llvm.org/D31135 llvm-svn: 300037
2017-04-12 16:23:08 +08:00
static bool classof(const OperandRenderer *R) {
return R->getKind() == OR_Imm;
}
[globalisel][tablegen] Add control-flow to the MatchTable. Summary: This will allow us to merge the various sub-tables into a single table. This is a compile-time saving at this point. However, this will also enable the optimization of a table so that similar instructions can be tested together, reducing the time spent on the matching the code. The bulk of this patch is a mechanical conversion to the new MatchTable object which is responsible for tracking label definitions and filling in the index of the jump targets. It is also responsible for nicely formatting the table. This was necessary to support the new GIM_Try opcode which takes the index to jump to if the match should fail. This value is unknown during table construction and is filled in during emission. To support nesting try-blocks (although we currently don't emit tables with nested try-blocks), GIM_Reject has been re-introduced to explicitly exit a try-block or fail the overall match if there are no active try-blocks. Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Reviewed By: rovka Subscribers: kristof.beyls, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D35117 llvm-svn: 308596
2017-07-20 17:25:44 +08:00
void emitRenderOpcodes(MatchTable &Table, RuleMatcher &Rule) const override {
Table << MatchTable::Opcode("GIR_AddImm") << MatchTable::Comment("InsnID")
<< MatchTable::IntValue(InsnID) << MatchTable::Comment("Imm")
<< MatchTable::IntValue(Imm) << MatchTable::LineBreak;
[globalisel][tablegen] Add experimental support for OperandWithDefaultOps, PredicateOperand, and OptionalDefOperand Summary: As far as instruction selection is concerned, all three appear to be same thing. Support for these operands is experimental since AArch64 doesn't make use of them and the in-tree targets that do use them (AMDGPU for OperandWithDefaultOps, AMDGPU/ARM/Hexagon/Lanai for PredicateOperand, and ARM for OperandWithDefaultOps) are not using tablegen-erated GlobalISel yet. Reviewers: rovka, aditya_nandakumar, t.p.northover, qcolombet, ab Reviewed By: rovka Subscribers: inglorion, aemerson, rengolin, mehdi_amini, dberris, kristof.beyls, igorb, tpr, llvm-commits Differential Revision: https://reviews.llvm.org/D31135 llvm-svn: 300037
2017-04-12 16:23:08 +08:00
}
};
[globalisel][tablegen] Revise API for ComplexPattern operands to improve flexibility. Summary: Some targets need to be able to do more complex rendering than just adding an operand or two to an instruction. For example, it may need to insert an instruction to extract a subreg first, or it may need to perform an operation on the operand. In SelectionDAG, targets would create SDNode's to achieve the desired effect during the complex pattern predicate. This worked because SelectionDAG had a form of garbage collection that would take care of SDNode's that were created but not used due to a later predicate rejecting a match. This doesn't translate well to GlobalISel and the churn was wasteful. The API changes in this patch enable GlobalISel to accomplish the same thing without the waste. The API is now: InstructionSelector::OptionalComplexRendererFn selectArithImmed(MachineOperand &Root) const; where Root is the root of the match. The return value can be omitted to indicate that the predicate failed to match, or a function with the signature ComplexRendererFn can be returned. For example: return OptionalComplexRendererFn( [=](MachineInstrBuilder &MIB) { MIB.addImm(Immed).addImm(ShVal); }); adds two immediate operands to the rendered instruction. Immed and ShVal are captured from the predicate function. As an added bonus, this also reduces the amount of information we need to provide to GIComplexOperandMatcher. Depends on D31418 Reviewers: aditya_nandakumar, t.p.northover, qcolombet, rovka, ab, javed.absar Reviewed By: ab Subscribers: dberris, kristof.beyls, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D31761 llvm-svn: 301079
2017-04-22 23:11:04 +08:00
/// Adds operands by calling a renderer function supplied by the ComplexPattern
/// matcher function.
[globalisel][tblgen] Add support for ComplexPatterns Summary: Adds a new kind of MachineOperand: MO_Placeholder. This operand must not appear in the MIR and only exists as a way of creating an 'uninitialized' operand until a matcher function overwrites it. Depends on D30046, D29712 Reviewers: t.p.northover, ab, rovka, aditya_nandakumar, javed.absar, qcolombet Reviewed By: qcolombet Subscribers: dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D30089 llvm-svn: 297782
2017-03-15 05:32:08 +08:00
class RenderComplexPatternOperand : public OperandRenderer {
private:
[globalisel][tablegen] Added instruction emission to the state-machine-based matcher. Summary: This further improves the compile-time regressions that will be caused by a re-commit of r303259. Also added included preliminary work in preparation for the multi-insn emitter since I needed to change the relevant part of the API for this patch anyway. Depends on D33758 Reviewers: rovka, vitalybuka, ab, t.p.northover, qcolombet, aditya_nandakumar Reviewed By: ab Subscribers: kristof.beyls, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D33764 llvm-svn: 307133
2017-07-05 17:39:33 +08:00
unsigned InsnID;
[globalisel][tblgen] Add support for ComplexPatterns Summary: Adds a new kind of MachineOperand: MO_Placeholder. This operand must not appear in the MIR and only exists as a way of creating an 'uninitialized' operand until a matcher function overwrites it. Depends on D30046, D29712 Reviewers: t.p.northover, ab, rovka, aditya_nandakumar, javed.absar, qcolombet Reviewed By: qcolombet Subscribers: dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D30089 llvm-svn: 297782
2017-03-15 05:32:08 +08:00
const Record &TheDef;
[globalisel][tablegen] Revise API for ComplexPattern operands to improve flexibility. Summary: Some targets need to be able to do more complex rendering than just adding an operand or two to an instruction. For example, it may need to insert an instruction to extract a subreg first, or it may need to perform an operation on the operand. In SelectionDAG, targets would create SDNode's to achieve the desired effect during the complex pattern predicate. This worked because SelectionDAG had a form of garbage collection that would take care of SDNode's that were created but not used due to a later predicate rejecting a match. This doesn't translate well to GlobalISel and the churn was wasteful. The API changes in this patch enable GlobalISel to accomplish the same thing without the waste. The API is now: InstructionSelector::OptionalComplexRendererFn selectArithImmed(MachineOperand &Root) const; where Root is the root of the match. The return value can be omitted to indicate that the predicate failed to match, or a function with the signature ComplexRendererFn can be returned. For example: return OptionalComplexRendererFn( [=](MachineInstrBuilder &MIB) { MIB.addImm(Immed).addImm(ShVal); }); adds two immediate operands to the rendered instruction. Immed and ShVal are captured from the predicate function. As an added bonus, this also reduces the amount of information we need to provide to GIComplexOperandMatcher. Depends on D31418 Reviewers: aditya_nandakumar, t.p.northover, qcolombet, rovka, ab, javed.absar Reviewed By: ab Subscribers: dberris, kristof.beyls, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D31761 llvm-svn: 301079
2017-04-22 23:11:04 +08:00
/// The name of the operand.
const StringRef SymbolicName;
/// The renderer number. This must be unique within a rule since it's used to
/// identify a temporary variable to hold the renderer function.
unsigned RendererID;
Re-commit r315863: [globalisel][tablegen] Import ComplexPattern when used as an operator Summary: It's possible for a ComplexPattern to be used as an operator in a match pattern. This is used by the load/store patterns in AArch64 to name the suboperands returned by ComplexPattern predicate so that they can be broken apart and referenced independently in the result pattern. This patch adds support for this in order to enable the import of load/store patterns. Depends on D37445 Hopefully fixed the ambiguous constructor that a large number of bots reported. Reviewers: ab, qcolombet, t.p.northover, rovka, aditya_nandakumar Reviewed By: qcolombet Subscribers: aemerson, javed.absar, igorb, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D37456 llvm-svn: 315869
2017-10-16 02:22:54 +08:00
/// When provided, this is the suboperand of the ComplexPattern operand to
/// render. Otherwise all the suboperands will be rendered.
Optional<unsigned> SubOperand;
[globalisel][tblgen] Add support for ComplexPatterns Summary: Adds a new kind of MachineOperand: MO_Placeholder. This operand must not appear in the MIR and only exists as a way of creating an 'uninitialized' operand until a matcher function overwrites it. Depends on D30046, D29712 Reviewers: t.p.northover, ab, rovka, aditya_nandakumar, javed.absar, qcolombet Reviewed By: qcolombet Subscribers: dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D30089 llvm-svn: 297782
2017-03-15 05:32:08 +08:00
unsigned getNumOperands() const {
return TheDef.getValueAsDag("Operands")->getNumArgs();
}
public:
[globalisel][tablegen] Added instruction emission to the state-machine-based matcher. Summary: This further improves the compile-time regressions that will be caused by a re-commit of r303259. Also added included preliminary work in preparation for the multi-insn emitter since I needed to change the relevant part of the API for this patch anyway. Depends on D33758 Reviewers: rovka, vitalybuka, ab, t.p.northover, qcolombet, aditya_nandakumar Reviewed By: ab Subscribers: kristof.beyls, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D33764 llvm-svn: 307133
2017-07-05 17:39:33 +08:00
RenderComplexPatternOperand(unsigned InsnID, const Record &TheDef,
Re-commit r315863: [globalisel][tablegen] Import ComplexPattern when used as an operator Summary: It's possible for a ComplexPattern to be used as an operator in a match pattern. This is used by the load/store patterns in AArch64 to name the suboperands returned by ComplexPattern predicate so that they can be broken apart and referenced independently in the result pattern. This patch adds support for this in order to enable the import of load/store patterns. Depends on D37445 Hopefully fixed the ambiguous constructor that a large number of bots reported. Reviewers: ab, qcolombet, t.p.northover, rovka, aditya_nandakumar Reviewed By: qcolombet Subscribers: aemerson, javed.absar, igorb, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D37456 llvm-svn: 315869
2017-10-16 02:22:54 +08:00
StringRef SymbolicName, unsigned RendererID,
Optional<unsigned> SubOperand = None)
[globalisel][tablegen] Added instruction emission to the state-machine-based matcher. Summary: This further improves the compile-time regressions that will be caused by a re-commit of r303259. Also added included preliminary work in preparation for the multi-insn emitter since I needed to change the relevant part of the API for this patch anyway. Depends on D33758 Reviewers: rovka, vitalybuka, ab, t.p.northover, qcolombet, aditya_nandakumar Reviewed By: ab Subscribers: kristof.beyls, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D33764 llvm-svn: 307133
2017-07-05 17:39:33 +08:00
: OperandRenderer(OR_ComplexPattern), InsnID(InsnID), TheDef(TheDef),
Re-commit r315863: [globalisel][tablegen] Import ComplexPattern when used as an operator Summary: It's possible for a ComplexPattern to be used as an operator in a match pattern. This is used by the load/store patterns in AArch64 to name the suboperands returned by ComplexPattern predicate so that they can be broken apart and referenced independently in the result pattern. This patch adds support for this in order to enable the import of load/store patterns. Depends on D37445 Hopefully fixed the ambiguous constructor that a large number of bots reported. Reviewers: ab, qcolombet, t.p.northover, rovka, aditya_nandakumar Reviewed By: qcolombet Subscribers: aemerson, javed.absar, igorb, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D37456 llvm-svn: 315869
2017-10-16 02:22:54 +08:00
SymbolicName(SymbolicName), RendererID(RendererID),
SubOperand(SubOperand) {}
[globalisel][tblgen] Add support for ComplexPatterns Summary: Adds a new kind of MachineOperand: MO_Placeholder. This operand must not appear in the MIR and only exists as a way of creating an 'uninitialized' operand until a matcher function overwrites it. Depends on D30046, D29712 Reviewers: t.p.northover, ab, rovka, aditya_nandakumar, javed.absar, qcolombet Reviewed By: qcolombet Subscribers: dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D30089 llvm-svn: 297782
2017-03-15 05:32:08 +08:00
static bool classof(const OperandRenderer *R) {
return R->getKind() == OR_ComplexPattern;
}
[globalisel][tablegen] Add control-flow to the MatchTable. Summary: This will allow us to merge the various sub-tables into a single table. This is a compile-time saving at this point. However, this will also enable the optimization of a table so that similar instructions can be tested together, reducing the time spent on the matching the code. The bulk of this patch is a mechanical conversion to the new MatchTable object which is responsible for tracking label definitions and filling in the index of the jump targets. It is also responsible for nicely formatting the table. This was necessary to support the new GIM_Try opcode which takes the index to jump to if the match should fail. This value is unknown during table construction and is filled in during emission. To support nesting try-blocks (although we currently don't emit tables with nested try-blocks), GIM_Reject has been re-introduced to explicitly exit a try-block or fail the overall match if there are no active try-blocks. Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Reviewed By: rovka Subscribers: kristof.beyls, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D35117 llvm-svn: 308596
2017-07-20 17:25:44 +08:00
void emitRenderOpcodes(MatchTable &Table, RuleMatcher &Rule) const override {
Re-commit r315863: [globalisel][tablegen] Import ComplexPattern when used as an operator Summary: It's possible for a ComplexPattern to be used as an operator in a match pattern. This is used by the load/store patterns in AArch64 to name the suboperands returned by ComplexPattern predicate so that they can be broken apart and referenced independently in the result pattern. This patch adds support for this in order to enable the import of load/store patterns. Depends on D37445 Hopefully fixed the ambiguous constructor that a large number of bots reported. Reviewers: ab, qcolombet, t.p.northover, rovka, aditya_nandakumar Reviewed By: qcolombet Subscribers: aemerson, javed.absar, igorb, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D37456 llvm-svn: 315869
2017-10-16 02:22:54 +08:00
Table << MatchTable::Opcode(SubOperand.hasValue() ? "GIR_ComplexSubOperandRenderer"
: "GIR_ComplexRenderer")
[globalisel][tablegen] Add control-flow to the MatchTable. Summary: This will allow us to merge the various sub-tables into a single table. This is a compile-time saving at this point. However, this will also enable the optimization of a table so that similar instructions can be tested together, reducing the time spent on the matching the code. The bulk of this patch is a mechanical conversion to the new MatchTable object which is responsible for tracking label definitions and filling in the index of the jump targets. It is also responsible for nicely formatting the table. This was necessary to support the new GIM_Try opcode which takes the index to jump to if the match should fail. This value is unknown during table construction and is filled in during emission. To support nesting try-blocks (although we currently don't emit tables with nested try-blocks), GIM_Reject has been re-introduced to explicitly exit a try-block or fail the overall match if there are no active try-blocks. Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Reviewed By: rovka Subscribers: kristof.beyls, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D35117 llvm-svn: 308596
2017-07-20 17:25:44 +08:00
<< MatchTable::Comment("InsnID") << MatchTable::IntValue(InsnID)
<< MatchTable::Comment("RendererID")
Re-commit r315863: [globalisel][tablegen] Import ComplexPattern when used as an operator Summary: It's possible for a ComplexPattern to be used as an operator in a match pattern. This is used by the load/store patterns in AArch64 to name the suboperands returned by ComplexPattern predicate so that they can be broken apart and referenced independently in the result pattern. This patch adds support for this in order to enable the import of load/store patterns. Depends on D37445 Hopefully fixed the ambiguous constructor that a large number of bots reported. Reviewers: ab, qcolombet, t.p.northover, rovka, aditya_nandakumar Reviewed By: qcolombet Subscribers: aemerson, javed.absar, igorb, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D37456 llvm-svn: 315869
2017-10-16 02:22:54 +08:00
<< MatchTable::IntValue(RendererID);
if (SubOperand.hasValue())
Table << MatchTable::Comment("SubOperand")
<< MatchTable::IntValue(SubOperand.getValue());
Table << MatchTable::Comment(SymbolicName) << MatchTable::LineBreak;
[globalisel][tblgen] Add support for ComplexPatterns Summary: Adds a new kind of MachineOperand: MO_Placeholder. This operand must not appear in the MIR and only exists as a way of creating an 'uninitialized' operand until a matcher function overwrites it. Depends on D30046, D29712 Reviewers: t.p.northover, ab, rovka, aditya_nandakumar, javed.absar, qcolombet Reviewed By: qcolombet Subscribers: dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D30089 llvm-svn: 297782
2017-03-15 05:32:08 +08:00
}
};
[GlobalISel][TableGen] Add support for SDNodeXForm Summary: This patch adds CustomRenderer which renders the matched operands to the specified instruction. Targets can enable the matching of SDNodeXForm by adding a definition that inherits from GICustomOperandRenderer and GISDNodeXFormEquiv as follows. def gi_imm8 : GICustomOperandRenderer<"renderImm8”>, GISDNodeXFormEquiv<imm8_xform>; Custom renderer functions should be of the form: void render(MachineInstrBuilder &MIB, const MachineInstr &I); Reviewers: dsanders, ab, rovka Reviewed By: dsanders Subscribers: kristof.beyls, javed.absar, llvm-commits, mgrang, qcolombet Differential Revision: https://reviews.llvm.org/D42012 llvm-svn: 322582
2018-01-17 02:44:05 +08:00
class CustomRenderer : public OperandRenderer {
protected:
unsigned InsnID;
const Record &Renderer;
/// The name of the operand.
const std::string SymbolicName;
public:
CustomRenderer(unsigned InsnID, const Record &Renderer,
StringRef SymbolicName)
: OperandRenderer(OR_Custom), InsnID(InsnID), Renderer(Renderer),
SymbolicName(SymbolicName) {}
static bool classof(const OperandRenderer *R) {
return R->getKind() == OR_Custom;
}
void emitRenderOpcodes(MatchTable &Table, RuleMatcher &Rule) const override {
Reapply r332907 "[GlobalISel] Improving InstructionSelect's performance by reducing MatchTable..." Apparently the compile time problem was caused by the fact that not all compilers / STL implementations can automatically convert std::unique_ptr<Derived> to std::unique_ptr<Base>. Fixed (hopefully) by making sure it's std::unique_ptr<Derived>&& (rvalue ref) to std::unique_ptr<Base> conversion instead. llvm-svn: 332917
2018-05-22 07:28:51 +08:00
InstructionMatcher &InsnMatcher = Rule.getInstructionMatcher(SymbolicName);
[GlobalISel][TableGen] Add support for SDNodeXForm Summary: This patch adds CustomRenderer which renders the matched operands to the specified instruction. Targets can enable the matching of SDNodeXForm by adding a definition that inherits from GICustomOperandRenderer and GISDNodeXFormEquiv as follows. def gi_imm8 : GICustomOperandRenderer<"renderImm8”>, GISDNodeXFormEquiv<imm8_xform>; Custom renderer functions should be of the form: void render(MachineInstrBuilder &MIB, const MachineInstr &I); Reviewers: dsanders, ab, rovka Reviewed By: dsanders Subscribers: kristof.beyls, javed.absar, llvm-commits, mgrang, qcolombet Differential Revision: https://reviews.llvm.org/D42012 llvm-svn: 322582
2018-01-17 02:44:05 +08:00
unsigned OldInsnVarID = Rule.getInsnVarID(InsnMatcher);
Table << MatchTable::Opcode("GIR_CustomRenderer")
<< MatchTable::Comment("InsnID") << MatchTable::IntValue(InsnID)
<< MatchTable::Comment("OldInsnID")
<< MatchTable::IntValue(OldInsnVarID)
<< MatchTable::Comment("Renderer")
<< MatchTable::NamedValue(
"GICR_" + Renderer.getValueAsString("RendererFn").str())
<< MatchTable::Comment(SymbolicName) << MatchTable::LineBreak;
}
};
[GlobalISel] Add a few comments to the tablegen backend. NFC. Based on a review of D29478 by Kristof Beyls. llvm-svn: 294077
2017-02-04 08:47:10 +08:00
/// An action taken when all Matcher predicates succeeded for a parent rule.
///
/// Typical actions include:
/// * Changing the opcode of an instruction.
/// * Adding an operand to an instruction.
[globalisel] Make the MatchAction hierarchy consistent with the matchers. NFC. Reviewers: aditya_nandakumar, rovka, t.p.northover, qcolombet, ab Subscribers: dberris, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D29321 llvm-svn: 293760
2017-02-01 18:53:10 +08:00
class MatchAction {
public:
virtual ~MatchAction() {}
[globalisel] Decouple src pattern operands from dst pattern operands. Summary: This isn't testable for AArch64 by itself so this patch also adds support for constant immediates in the pattern and physical register uses in the result. The new IntOperandMatcher matches the constant in patterns such as '(set $rd:GPR32, (G_XOR $rs:GPR32, -1))'. It's always safe to fold immediates into an instruction so this is the first rule that will match across multiple BB's. The Renderer hierarchy is responsible for adding operands to the result instruction. Renderers can copy operands (CopyRenderer) or add physical registers (in particular %wzr and %xzr) to the result instruction in any order (OperandMatchers now import the operand names from SelectionDAG to allow renderers to access any operand). This allows us to emit the result instruction for: %1 = G_XOR %0, -1 --> %1 = ORNWrr %wzr, %0 %1 = G_XOR -1, %0 --> %1 = ORNWrr %wzr, %0 although the latter is untested since the matcher/importer has not been taught about commutativity yet. Added BuildMIAction which can build new instructions and mutate them where possible. W.r.t the mutation aspect, MatchActions are now told the name of an instruction they can recycle and BuildMIAction will emit mutation code when the renderers are appropriate. They are appropriate when all operands are rendered using CopyRenderer and the indices are the same as the matcher. This currently assumes that all operands have at least one matcher. Finally, this change also fixes a crash in AArch64InstructionSelector::select() caused by an immediate operand passing isImm() rather than isCImm(). This was uncovered by the other changes and was detected by existing tests. Depends on D29711 Reviewers: t.p.northover, ab, qcolombet, rovka, aditya_nandakumar, javed.absar Reviewed By: rovka Subscribers: aemerson, dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D29712 llvm-svn: 296131
2017-02-24 23:43:30 +08:00
[globalisel][tablegen] Add control-flow to the MatchTable. Summary: This will allow us to merge the various sub-tables into a single table. This is a compile-time saving at this point. However, this will also enable the optimization of a table so that similar instructions can be tested together, reducing the time spent on the matching the code. The bulk of this patch is a mechanical conversion to the new MatchTable object which is responsible for tracking label definitions and filling in the index of the jump targets. It is also responsible for nicely formatting the table. This was necessary to support the new GIM_Try opcode which takes the index to jump to if the match should fail. This value is unknown during table construction and is filled in during emission. To support nesting try-blocks (although we currently don't emit tables with nested try-blocks), GIM_Reject has been re-introduced to explicitly exit a try-block or fail the overall match if there are no active try-blocks. Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Reviewed By: rovka Subscribers: kristof.beyls, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D35117 llvm-svn: 308596
2017-07-20 17:25:44 +08:00
/// Emit the MatchTable opcodes to implement the action.
[globalisel][tablegen] Add infrastructure to potentially allow BuildMIAction to choose a mutatable instruction. NFC Prepare for multiple instruction emission by allowing BuildMIAction to search for a suitable matcher that will support mutation. This patch deliberately neglects to add matchers aside from the root to preserve NFC. That said, it should be noted that until we support mutations other than just the opcode the chances of finding a non-root instruction for which canMutate() is true, is essentially zero. Furthermore in the presence of multi-instruction emission the chances of finding any instruction for which canMutate() is true is also zero. Nevertheless, we can't continue to require that all BuildMIAction's consider the root of the match to be recyclable due to the risk of recycling it twice in the same rule. llvm-svn: 317022
2017-11-01 02:50:24 +08:00
virtual void emitActionOpcodes(MatchTable &Table,
RuleMatcher &Rule) const = 0;
[globalisel] Make the MatchAction hierarchy consistent with the matchers. NFC. Reviewers: aditya_nandakumar, rovka, t.p.northover, qcolombet, ab Subscribers: dberris, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D29321 llvm-svn: 293760
2017-02-01 18:53:10 +08:00
};
[GlobalISel] Print the matched patterns using an action. This lets us split out PatternToMatch from the top-level RuleMatcher, where it doesn't really belong. That, in turn, lets us eventually generate RuleMatchers from non-SelectionDAG sources. llvm-svn: 294076
2017-02-04 08:47:08 +08:00
/// Generates a comment describing the matched rule being acted upon.
class DebugCommentAction : public MatchAction {
private:
[globalisel][tablegen] Allow any comment in DebugCommentAction. NFC llvm-svn: 317017
2017-11-01 02:07:03 +08:00
std::string S;
[GlobalISel] Print the matched patterns using an action. This lets us split out PatternToMatch from the top-level RuleMatcher, where it doesn't really belong. That, in turn, lets us eventually generate RuleMatchers from non-SelectionDAG sources. llvm-svn: 294076
2017-02-04 08:47:08 +08:00
public:
[globalisel][tablegen] Allow any comment in DebugCommentAction. NFC llvm-svn: 317017
2017-11-01 02:07:03 +08:00
DebugCommentAction(StringRef S) : S(S) {}
[GlobalISel] Print the matched patterns using an action. This lets us split out PatternToMatch from the top-level RuleMatcher, where it doesn't really belong. That, in turn, lets us eventually generate RuleMatchers from non-SelectionDAG sources. llvm-svn: 294076
2017-02-04 08:47:08 +08:00
[globalisel][tablegen] Add infrastructure to potentially allow BuildMIAction to choose a mutatable instruction. NFC Prepare for multiple instruction emission by allowing BuildMIAction to search for a suitable matcher that will support mutation. This patch deliberately neglects to add matchers aside from the root to preserve NFC. That said, it should be noted that until we support mutations other than just the opcode the chances of finding a non-root instruction for which canMutate() is true, is essentially zero. Furthermore in the presence of multi-instruction emission the chances of finding any instruction for which canMutate() is true is also zero. Nevertheless, we can't continue to require that all BuildMIAction's consider the root of the match to be recyclable due to the risk of recycling it twice in the same rule. llvm-svn: 317022
2017-11-01 02:50:24 +08:00
void emitActionOpcodes(MatchTable &Table, RuleMatcher &Rule) const override {
[globalisel][tablegen] Allow any comment in DebugCommentAction. NFC llvm-svn: 317017
2017-11-01 02:07:03 +08:00
Table << MatchTable::Comment(S) << MatchTable::LineBreak;
[GlobalISel] Print the matched patterns using an action. This lets us split out PatternToMatch from the top-level RuleMatcher, where it doesn't really belong. That, in turn, lets us eventually generate RuleMatchers from non-SelectionDAG sources. llvm-svn: 294076
2017-02-04 08:47:08 +08:00
}
};
[globalisel] Decouple src pattern operands from dst pattern operands. Summary: This isn't testable for AArch64 by itself so this patch also adds support for constant immediates in the pattern and physical register uses in the result. The new IntOperandMatcher matches the constant in patterns such as '(set $rd:GPR32, (G_XOR $rs:GPR32, -1))'. It's always safe to fold immediates into an instruction so this is the first rule that will match across multiple BB's. The Renderer hierarchy is responsible for adding operands to the result instruction. Renderers can copy operands (CopyRenderer) or add physical registers (in particular %wzr and %xzr) to the result instruction in any order (OperandMatchers now import the operand names from SelectionDAG to allow renderers to access any operand). This allows us to emit the result instruction for: %1 = G_XOR %0, -1 --> %1 = ORNWrr %wzr, %0 %1 = G_XOR -1, %0 --> %1 = ORNWrr %wzr, %0 although the latter is untested since the matcher/importer has not been taught about commutativity yet. Added BuildMIAction which can build new instructions and mutate them where possible. W.r.t the mutation aspect, MatchActions are now told the name of an instruction they can recycle and BuildMIAction will emit mutation code when the renderers are appropriate. They are appropriate when all operands are rendered using CopyRenderer and the indices are the same as the matcher. This currently assumes that all operands have at least one matcher. Finally, this change also fixes a crash in AArch64InstructionSelector::select() caused by an immediate operand passing isImm() rather than isCImm(). This was uncovered by the other changes and was detected by existing tests. Depends on D29711 Reviewers: t.p.northover, ab, qcolombet, rovka, aditya_nandakumar, javed.absar Reviewed By: rovka Subscribers: aemerson, dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D29712 llvm-svn: 296131
2017-02-24 23:43:30 +08:00
/// Generates code to build an instruction or mutate an existing instruction
/// into the desired instruction when this is possible.
class BuildMIAction : public MatchAction {
[globalisel] Make the MatchAction hierarchy consistent with the matchers. NFC. Reviewers: aditya_nandakumar, rovka, t.p.northover, qcolombet, ab Subscribers: dberris, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D29321 llvm-svn: 293760
2017-02-01 18:53:10 +08:00
private:
[globalisel][tablegen] Added instruction emission to the state-machine-based matcher. Summary: This further improves the compile-time regressions that will be caused by a re-commit of r303259. Also added included preliminary work in preparation for the multi-insn emitter since I needed to change the relevant part of the API for this patch anyway. Depends on D33758 Reviewers: rovka, vitalybuka, ab, t.p.northover, qcolombet, aditya_nandakumar Reviewed By: ab Subscribers: kristof.beyls, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D33764 llvm-svn: 307133
2017-07-05 17:39:33 +08:00
unsigned InsnID;
[GlobalISel] Add basic Selector-emitter tblgen backend. This adds a basic tablegen backend that analyzes the SelectionDAG patterns to find simple ones that are eligible for GlobalISel-emission. That's similar to FastISel, with one notable difference: we're not fed ISD opcodes, so we need to map the SDNode operators to generic opcodes. That's done using GINodeEquiv in TargetGlobalISel.td. Otherwise, this is mostly boilerplate, and lots of filtering of any kind of "complicated" pattern. On AArch64, this is sufficient to match G_ADD up to s64 (to ADDWrr/ADDXrr) and G_BR (to B). Differential Revision: https://reviews.llvm.org/D26878 llvm-svn: 290284
2016-12-22 07:26:20 +08:00
const CodeGenInstruction *I;
Reapply r332907 "[GlobalISel] Improving InstructionSelect's performance by reducing MatchTable..." Apparently the compile time problem was caused by the fact that not all compilers / STL implementations can automatically convert std::unique_ptr<Derived> to std::unique_ptr<Base>. Fixed (hopefully) by making sure it's std::unique_ptr<Derived>&& (rvalue ref) to std::unique_ptr<Base> conversion instead. llvm-svn: 332917
2018-05-22 07:28:51 +08:00
InstructionMatcher *Matched;
[globalisel] Decouple src pattern operands from dst pattern operands. Summary: This isn't testable for AArch64 by itself so this patch also adds support for constant immediates in the pattern and physical register uses in the result. The new IntOperandMatcher matches the constant in patterns such as '(set $rd:GPR32, (G_XOR $rs:GPR32, -1))'. It's always safe to fold immediates into an instruction so this is the first rule that will match across multiple BB's. The Renderer hierarchy is responsible for adding operands to the result instruction. Renderers can copy operands (CopyRenderer) or add physical registers (in particular %wzr and %xzr) to the result instruction in any order (OperandMatchers now import the operand names from SelectionDAG to allow renderers to access any operand). This allows us to emit the result instruction for: %1 = G_XOR %0, -1 --> %1 = ORNWrr %wzr, %0 %1 = G_XOR -1, %0 --> %1 = ORNWrr %wzr, %0 although the latter is untested since the matcher/importer has not been taught about commutativity yet. Added BuildMIAction which can build new instructions and mutate them where possible. W.r.t the mutation aspect, MatchActions are now told the name of an instruction they can recycle and BuildMIAction will emit mutation code when the renderers are appropriate. They are appropriate when all operands are rendered using CopyRenderer and the indices are the same as the matcher. This currently assumes that all operands have at least one matcher. Finally, this change also fixes a crash in AArch64InstructionSelector::select() caused by an immediate operand passing isImm() rather than isCImm(). This was uncovered by the other changes and was detected by existing tests. Depends on D29711 Reviewers: t.p.northover, ab, qcolombet, rovka, aditya_nandakumar, javed.absar Reviewed By: rovka Subscribers: aemerson, dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D29712 llvm-svn: 296131
2017-02-24 23:43:30 +08:00
std::vector<std::unique_ptr<OperandRenderer>> OperandRenderers;
/// True if the instruction can be built solely by mutating the opcode.
[globalisel][tablegen] Add infrastructure to potentially allow BuildMIAction to choose a mutatable instruction. NFC Prepare for multiple instruction emission by allowing BuildMIAction to search for a suitable matcher that will support mutation. This patch deliberately neglects to add matchers aside from the root to preserve NFC. That said, it should be noted that until we support mutations other than just the opcode the chances of finding a non-root instruction for which canMutate() is true, is essentially zero. Furthermore in the presence of multi-instruction emission the chances of finding any instruction for which canMutate() is true is also zero. Nevertheless, we can't continue to require that all BuildMIAction's consider the root of the match to be recyclable due to the risk of recycling it twice in the same rule. llvm-svn: 317022
2017-11-01 02:50:24 +08:00
bool canMutate(RuleMatcher &Rule, const InstructionMatcher *Insn) const {
if (!Insn)
[globalisel][tablegen] Fix future undefined behaviour in r316463. I missed a dereference of `Matched` that preceeded the new check. Thanks to Justin Bogner for spotting it. llvm-svn: 316480
2017-10-25 02:11:54 +08:00
return false;
[globalisel][tablegen] Add infrastructure to potentially allow BuildMIAction to choose a mutatable instruction. NFC Prepare for multiple instruction emission by allowing BuildMIAction to search for a suitable matcher that will support mutation. This patch deliberately neglects to add matchers aside from the root to preserve NFC. That said, it should be noted that until we support mutations other than just the opcode the chances of finding a non-root instruction for which canMutate() is true, is essentially zero. Furthermore in the presence of multi-instruction emission the chances of finding any instruction for which canMutate() is true is also zero. Nevertheless, we can't continue to require that all BuildMIAction's consider the root of the match to be recyclable due to the risk of recycling it twice in the same rule. llvm-svn: 317022
2017-11-01 02:50:24 +08:00
if (OperandRenderers.size() != Insn->getNumOperands())
[globalisel][tablegen] Compute available feature bits correctly. Summary: Predicate<> now has a field to indicate how often it must be recomputed. Currently, there are two frequencies, per-module (RecomputePerFunction==0) and per-function (RecomputePerFunction==1). Per-function predicates are currently recomputed more frequently than necessary since the only predicate in this category is cheap to test. Per-module predicates are now computed in getSubtargetImpl() while per-function predicates are computed in selectImpl(). Tablegen now manages the PredicateBitset internally. It should only be necessary to add the required includes. Also fixed a problem revealed by the test case where constrainSelectedInstRegOperands() would attempt to tie operands that BuildMI had already tied. Reviewers: ab, qcolombet, t.p.northover, rovka, aditya_nandakumar Reviewed By: rovka Subscribers: kristof.beyls, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D32491 llvm-svn: 301750
2017-04-30 01:30:09 +08:00
return false;
[globalisel] Decouple src pattern operands from dst pattern operands. Summary: This isn't testable for AArch64 by itself so this patch also adds support for constant immediates in the pattern and physical register uses in the result. The new IntOperandMatcher matches the constant in patterns such as '(set $rd:GPR32, (G_XOR $rs:GPR32, -1))'. It's always safe to fold immediates into an instruction so this is the first rule that will match across multiple BB's. The Renderer hierarchy is responsible for adding operands to the result instruction. Renderers can copy operands (CopyRenderer) or add physical registers (in particular %wzr and %xzr) to the result instruction in any order (OperandMatchers now import the operand names from SelectionDAG to allow renderers to access any operand). This allows us to emit the result instruction for: %1 = G_XOR %0, -1 --> %1 = ORNWrr %wzr, %0 %1 = G_XOR -1, %0 --> %1 = ORNWrr %wzr, %0 although the latter is untested since the matcher/importer has not been taught about commutativity yet. Added BuildMIAction which can build new instructions and mutate them where possible. W.r.t the mutation aspect, MatchActions are now told the name of an instruction they can recycle and BuildMIAction will emit mutation code when the renderers are appropriate. They are appropriate when all operands are rendered using CopyRenderer and the indices are the same as the matcher. This currently assumes that all operands have at least one matcher. Finally, this change also fixes a crash in AArch64InstructionSelector::select() caused by an immediate operand passing isImm() rather than isCImm(). This was uncovered by the other changes and was detected by existing tests. Depends on D29711 Reviewers: t.p.northover, ab, qcolombet, rovka, aditya_nandakumar, javed.absar Reviewed By: rovka Subscribers: aemerson, dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D29712 llvm-svn: 296131
2017-02-24 23:43:30 +08:00
for (const auto &Renderer : enumerate(OperandRenderers)) {
[ADT] Improve the genericity of llvm::enumerate(). There were some issues in the implementation of enumerate() preventing it from being used in various contexts. These were all related to the fact that it did not supporter llvm's iterator_facade_base class. So this patch adds support for that and additionally exposes a new helper method to_vector() that will evaluate an entire range and store the results in a vector. Differential Revision: https://reviews.llvm.org/D30853 llvm-svn: 297633
2017-03-14 00:24:10 +08:00
if (const auto *Copy = dyn_cast<CopyRenderer>(&*Renderer.value())) {
[globalisel][tablegen] Simplify named operand/operator lookups and fix a wrong-code bug this revealed. Summary: Operand variable lookups are now performed by the RuleMatcher rather than searching the whole matcher hierarchy for a match. This revealed a wrong-code bug that currently affects ARM and X86 where patterns that use a variable more than once in the match pattern will be imported but won't check that the operands are identical. This can cause the tablegen-erated matcher to accept matches that should be rejected. Depends on D36569 Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Subscribers: aemerson, igorb, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D36618 llvm-svn: 315780
2017-10-14 08:31:58 +08:00
const OperandMatcher &OM = Rule.getOperandMatcher(Copy->getSymbolicName());
[globalisel][tablegen] Add infrastructure to potentially allow BuildMIAction to choose a mutatable instruction. NFC Prepare for multiple instruction emission by allowing BuildMIAction to search for a suitable matcher that will support mutation. This patch deliberately neglects to add matchers aside from the root to preserve NFC. That said, it should be noted that until we support mutations other than just the opcode the chances of finding a non-root instruction for which canMutate() is true, is essentially zero. Furthermore in the presence of multi-instruction emission the chances of finding any instruction for which canMutate() is true is also zero. Nevertheless, we can't continue to require that all BuildMIAction's consider the root of the match to be recyclable due to the risk of recycling it twice in the same rule. llvm-svn: 317022
2017-11-01 02:50:24 +08:00
if (Insn != &OM.getInstructionMatcher() ||
Reapply r332907 "[GlobalISel] Improving InstructionSelect's performance by reducing MatchTable..." Apparently the compile time problem was caused by the fact that not all compilers / STL implementations can automatically convert std::unique_ptr<Derived> to std::unique_ptr<Base>. Fixed (hopefully) by making sure it's std::unique_ptr<Derived>&& (rvalue ref) to std::unique_ptr<Base> conversion instead. llvm-svn: 332917
2018-05-22 07:28:51 +08:00
OM.getOpIdx() != Renderer.index())
[globalisel] Decouple src pattern operands from dst pattern operands. Summary: This isn't testable for AArch64 by itself so this patch also adds support for constant immediates in the pattern and physical register uses in the result. The new IntOperandMatcher matches the constant in patterns such as '(set $rd:GPR32, (G_XOR $rs:GPR32, -1))'. It's always safe to fold immediates into an instruction so this is the first rule that will match across multiple BB's. The Renderer hierarchy is responsible for adding operands to the result instruction. Renderers can copy operands (CopyRenderer) or add physical registers (in particular %wzr and %xzr) to the result instruction in any order (OperandMatchers now import the operand names from SelectionDAG to allow renderers to access any operand). This allows us to emit the result instruction for: %1 = G_XOR %0, -1 --> %1 = ORNWrr %wzr, %0 %1 = G_XOR -1, %0 --> %1 = ORNWrr %wzr, %0 although the latter is untested since the matcher/importer has not been taught about commutativity yet. Added BuildMIAction which can build new instructions and mutate them where possible. W.r.t the mutation aspect, MatchActions are now told the name of an instruction they can recycle and BuildMIAction will emit mutation code when the renderers are appropriate. They are appropriate when all operands are rendered using CopyRenderer and the indices are the same as the matcher. This currently assumes that all operands have at least one matcher. Finally, this change also fixes a crash in AArch64InstructionSelector::select() caused by an immediate operand passing isImm() rather than isCImm(). This was uncovered by the other changes and was detected by existing tests. Depends on D29711 Reviewers: t.p.northover, ab, qcolombet, rovka, aditya_nandakumar, javed.absar Reviewed By: rovka Subscribers: aemerson, dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D29712 llvm-svn: 296131
2017-02-24 23:43:30 +08:00
return false;
} else
return false;
}
return true;
}
[GlobalISel] Add basic Selector-emitter tblgen backend. This adds a basic tablegen backend that analyzes the SelectionDAG patterns to find simple ones that are eligible for GlobalISel-emission. That's similar to FastISel, with one notable difference: we're not fed ISD opcodes, so we need to map the SDNode operators to generic opcodes. That's done using GINodeEquiv in TargetGlobalISel.td. Otherwise, this is mostly boilerplate, and lots of filtering of any kind of "complicated" pattern. On AArch64, this is sufficient to match G_ADD up to s64 (to ADDWrr/ADDXrr) and G_BR (to B). Differential Revision: https://reviews.llvm.org/D26878 llvm-svn: 290284
2016-12-22 07:26:20 +08:00
[globalisel] Make the MatchAction hierarchy consistent with the matchers. NFC. Reviewers: aditya_nandakumar, rovka, t.p.northover, qcolombet, ab Subscribers: dberris, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D29321 llvm-svn: 293760
2017-02-01 18:53:10 +08:00
public:
[globalisel][tablegen] Add infrastructure to potentially allow BuildMIAction to choose a mutatable instruction. NFC Prepare for multiple instruction emission by allowing BuildMIAction to search for a suitable matcher that will support mutation. This patch deliberately neglects to add matchers aside from the root to preserve NFC. That said, it should be noted that until we support mutations other than just the opcode the chances of finding a non-root instruction for which canMutate() is true, is essentially zero. Furthermore in the presence of multi-instruction emission the chances of finding any instruction for which canMutate() is true is also zero. Nevertheless, we can't continue to require that all BuildMIAction's consider the root of the match to be recyclable due to the risk of recycling it twice in the same rule. llvm-svn: 317022
2017-11-01 02:50:24 +08:00
BuildMIAction(unsigned InsnID, const CodeGenInstruction *I)
: InsnID(InsnID), I(I), Matched(nullptr) {}
[ARM][GISel] PR35965 Constrain RegClasses of nested instructions built from Dst Pattern Summary: Apparently, we missed on constraining register classes of VReg-operands of all the instructions built from a destination pattern but the root (top-level) one. The issue exposed itself while selecting G_FPTOSI for armv7: the corresponding pattern generates VTOSIZS wrapped into COPY_TO_REGCLASS, so top-level COPY_TO_REGCLASS gets properly constrained, while nested VTOSIZS (or rather its destination virtual register to be exact) does not. Fixing this by issuing GIR_ConstrainSelectedInstOperands for every nested GIR_BuildMI. https://bugs.llvm.org/show_bug.cgi?id=35965 rdar://problem/36886530 Patch by Roman Tereshin Reviewers: dsanders, qcolombet, rovka, bogner, aditya_nandakumar, volkan Reviewed By: dsanders, qcolombet, rovka Subscribers: aemerson, javed.absar, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D42565 llvm-svn: 323692
2018-01-30 05:09:12 +08:00
unsigned getInsnID() const { return InsnID; }
[globalisel][tablegen] Factor out implicit def/use renderers from createAndImportInstructionRenderer(). NFC Multi-instruction emission will require that we have separate handling for the defs between the implicitly created temporaries and the rule outputs. The former require new temporary vregs while the latter should copy existing operands. Factor out the implicit def/use renderers to minimize the code duplication when we implement that. llvm-svn: 317025
2017-11-01 03:09:29 +08:00
const CodeGenInstruction *getCGI() const { return I; }
[globalisel][tablegen] Add infrastructure to potentially allow BuildMIAction to choose a mutatable instruction. NFC Prepare for multiple instruction emission by allowing BuildMIAction to search for a suitable matcher that will support mutation. This patch deliberately neglects to add matchers aside from the root to preserve NFC. That said, it should be noted that until we support mutations other than just the opcode the chances of finding a non-root instruction for which canMutate() is true, is essentially zero. Furthermore in the presence of multi-instruction emission the chances of finding any instruction for which canMutate() is true is also zero. Nevertheless, we can't continue to require that all BuildMIAction's consider the root of the match to be recyclable due to the risk of recycling it twice in the same rule. llvm-svn: 317022
2017-11-01 02:50:24 +08:00
void chooseInsnToMutate(RuleMatcher &Rule) {
Reapply r332907 "[GlobalISel] Improving InstructionSelect's performance by reducing MatchTable..." Apparently the compile time problem was caused by the fact that not all compilers / STL implementations can automatically convert std::unique_ptr<Derived> to std::unique_ptr<Base>. Fixed (hopefully) by making sure it's std::unique_ptr<Derived>&& (rvalue ref) to std::unique_ptr<Base> conversion instead. llvm-svn: 332917
2018-05-22 07:28:51 +08:00
for (auto *MutateCandidate : Rule.mutatable_insns()) {
[globalisel][tablegen] Add infrastructure to potentially allow BuildMIAction to choose a mutatable instruction. NFC Prepare for multiple instruction emission by allowing BuildMIAction to search for a suitable matcher that will support mutation. This patch deliberately neglects to add matchers aside from the root to preserve NFC. That said, it should be noted that until we support mutations other than just the opcode the chances of finding a non-root instruction for which canMutate() is true, is essentially zero. Furthermore in the presence of multi-instruction emission the chances of finding any instruction for which canMutate() is true is also zero. Nevertheless, we can't continue to require that all BuildMIAction's consider the root of the match to be recyclable due to the risk of recycling it twice in the same rule. llvm-svn: 317022
2017-11-01 02:50:24 +08:00
if (canMutate(Rule, MutateCandidate)) {
// Take the first one we're offered that we're able to mutate.
Rule.reserveInsnMatcherForMutation(MutateCandidate);
Matched = MutateCandidate;
return;
}
}
}
[globalisel] Decouple src pattern operands from dst pattern operands. Summary: This isn't testable for AArch64 by itself so this patch also adds support for constant immediates in the pattern and physical register uses in the result. The new IntOperandMatcher matches the constant in patterns such as '(set $rd:GPR32, (G_XOR $rs:GPR32, -1))'. It's always safe to fold immediates into an instruction so this is the first rule that will match across multiple BB's. The Renderer hierarchy is responsible for adding operands to the result instruction. Renderers can copy operands (CopyRenderer) or add physical registers (in particular %wzr and %xzr) to the result instruction in any order (OperandMatchers now import the operand names from SelectionDAG to allow renderers to access any operand). This allows us to emit the result instruction for: %1 = G_XOR %0, -1 --> %1 = ORNWrr %wzr, %0 %1 = G_XOR -1, %0 --> %1 = ORNWrr %wzr, %0 although the latter is untested since the matcher/importer has not been taught about commutativity yet. Added BuildMIAction which can build new instructions and mutate them where possible. W.r.t the mutation aspect, MatchActions are now told the name of an instruction they can recycle and BuildMIAction will emit mutation code when the renderers are appropriate. They are appropriate when all operands are rendered using CopyRenderer and the indices are the same as the matcher. This currently assumes that all operands have at least one matcher. Finally, this change also fixes a crash in AArch64InstructionSelector::select() caused by an immediate operand passing isImm() rather than isCImm(). This was uncovered by the other changes and was detected by existing tests. Depends on D29711 Reviewers: t.p.northover, ab, qcolombet, rovka, aditya_nandakumar, javed.absar Reviewed By: rovka Subscribers: aemerson, dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D29712 llvm-svn: 296131
2017-02-24 23:43:30 +08:00
template <class Kind, class... Args>
Kind &addRenderer(Args&&... args) {
OperandRenderers.emplace_back(
[globalisel][tablegen] Stop hard-coding the emitted instruction ID to 0. NFC The next commit will add support for multi-instruction emission so we need to start allocating instruction ID's instead of hard-coding them to 0. llvm-svn: 317057
2017-11-01 08:29:47 +08:00
llvm::make_unique<Kind>(InsnID, std::forward<Args>(args)...));
[globalisel] Decouple src pattern operands from dst pattern operands. Summary: This isn't testable for AArch64 by itself so this patch also adds support for constant immediates in the pattern and physical register uses in the result. The new IntOperandMatcher matches the constant in patterns such as '(set $rd:GPR32, (G_XOR $rs:GPR32, -1))'. It's always safe to fold immediates into an instruction so this is the first rule that will match across multiple BB's. The Renderer hierarchy is responsible for adding operands to the result instruction. Renderers can copy operands (CopyRenderer) or add physical registers (in particular %wzr and %xzr) to the result instruction in any order (OperandMatchers now import the operand names from SelectionDAG to allow renderers to access any operand). This allows us to emit the result instruction for: %1 = G_XOR %0, -1 --> %1 = ORNWrr %wzr, %0 %1 = G_XOR -1, %0 --> %1 = ORNWrr %wzr, %0 although the latter is untested since the matcher/importer has not been taught about commutativity yet. Added BuildMIAction which can build new instructions and mutate them where possible. W.r.t the mutation aspect, MatchActions are now told the name of an instruction they can recycle and BuildMIAction will emit mutation code when the renderers are appropriate. They are appropriate when all operands are rendered using CopyRenderer and the indices are the same as the matcher. This currently assumes that all operands have at least one matcher. Finally, this change also fixes a crash in AArch64InstructionSelector::select() caused by an immediate operand passing isImm() rather than isCImm(). This was uncovered by the other changes and was detected by existing tests. Depends on D29711 Reviewers: t.p.northover, ab, qcolombet, rovka, aditya_nandakumar, javed.absar Reviewed By: rovka Subscribers: aemerson, dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D29712 llvm-svn: 296131
2017-02-24 23:43:30 +08:00
return *static_cast<Kind *>(OperandRenderers.back().get());
}
[globalisel][tablegen] Add infrastructure to potentially allow BuildMIAction to choose a mutatable instruction. NFC Prepare for multiple instruction emission by allowing BuildMIAction to search for a suitable matcher that will support mutation. This patch deliberately neglects to add matchers aside from the root to preserve NFC. That said, it should be noted that until we support mutations other than just the opcode the chances of finding a non-root instruction for which canMutate() is true, is essentially zero. Furthermore in the presence of multi-instruction emission the chances of finding any instruction for which canMutate() is true is also zero. Nevertheless, we can't continue to require that all BuildMIAction's consider the root of the match to be recyclable due to the risk of recycling it twice in the same rule. llvm-svn: 317022
2017-11-01 02:50:24 +08:00
void emitActionOpcodes(MatchTable &Table, RuleMatcher &Rule) const override {
if (Matched) {
assert(canMutate(Rule, Matched) &&
"Arranged to mutate an insn that isn't mutatable");
unsigned RecycleInsnID = Rule.getInsnVarID(*Matched);
[globalisel][tablegen] Add control-flow to the MatchTable. Summary: This will allow us to merge the various sub-tables into a single table. This is a compile-time saving at this point. However, this will also enable the optimization of a table so that similar instructions can be tested together, reducing the time spent on the matching the code. The bulk of this patch is a mechanical conversion to the new MatchTable object which is responsible for tracking label definitions and filling in the index of the jump targets. It is also responsible for nicely formatting the table. This was necessary to support the new GIM_Try opcode which takes the index to jump to if the match should fail. This value is unknown during table construction and is filled in during emission. To support nesting try-blocks (although we currently don't emit tables with nested try-blocks), GIM_Reject has been re-introduced to explicitly exit a try-block or fail the overall match if there are no active try-blocks. Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Reviewed By: rovka Subscribers: kristof.beyls, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D35117 llvm-svn: 308596
2017-07-20 17:25:44 +08:00
Table << MatchTable::Opcode("GIR_MutateOpcode")
<< MatchTable::Comment("InsnID") << MatchTable::IntValue(InsnID)
<< MatchTable::Comment("RecycleInsnID")
<< MatchTable::IntValue(RecycleInsnID)
<< MatchTable::Comment("Opcode")
<< MatchTable::NamedValue(I->Namespace, I->TheDef->getName())
<< MatchTable::LineBreak;
GlobalISel: add implicit defs & uses when mutating an instruction. Otherwise a scheduler might do bad things to the code we produce. llvm-svn: 298311
2017-03-21 05:58:23 +08:00
if (!I->ImplicitDefs.empty() || !I->ImplicitUses.empty()) {
for (auto Def : I->ImplicitDefs) {
[ARM] GlobalISel: Use TableGen instruction selector Emit and use the TableGen instruction selector for ARM. At the moment, this allows us to remove the hand-written code for selecting G_SDIV and G_UDIV. Future commits will focus on increasing the code coverage for it and removing more dead code from the current instruction selector. llvm-svn: 301905
2017-05-02 17:40:49 +08:00
auto Namespace = Def->getValue("Namespace")
? Def->getValueAsString("Namespace")
: "";
[globalisel][tablegen] Add control-flow to the MatchTable. Summary: This will allow us to merge the various sub-tables into a single table. This is a compile-time saving at this point. However, this will also enable the optimization of a table so that similar instructions can be tested together, reducing the time spent on the matching the code. The bulk of this patch is a mechanical conversion to the new MatchTable object which is responsible for tracking label definitions and filling in the index of the jump targets. It is also responsible for nicely formatting the table. This was necessary to support the new GIM_Try opcode which takes the index to jump to if the match should fail. This value is unknown during table construction and is filled in during emission. To support nesting try-blocks (although we currently don't emit tables with nested try-blocks), GIM_Reject has been re-introduced to explicitly exit a try-block or fail the overall match if there are no active try-blocks. Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Reviewed By: rovka Subscribers: kristof.beyls, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D35117 llvm-svn: 308596
2017-07-20 17:25:44 +08:00
Table << MatchTable::Opcode("GIR_AddImplicitDef")
<< MatchTable::Comment("InsnID") << MatchTable::IntValue(InsnID)
<< MatchTable::NamedValue(Namespace, Def->getName())
<< MatchTable::LineBreak;
GlobalISel: add implicit defs & uses when mutating an instruction. Otherwise a scheduler might do bad things to the code we produce. llvm-svn: 298311
2017-03-21 05:58:23 +08:00
}
for (auto Use : I->ImplicitUses) {
[ARM] GlobalISel: Use TableGen instruction selector Emit and use the TableGen instruction selector for ARM. At the moment, this allows us to remove the hand-written code for selecting G_SDIV and G_UDIV. Future commits will focus on increasing the code coverage for it and removing more dead code from the current instruction selector. llvm-svn: 301905
2017-05-02 17:40:49 +08:00
auto Namespace = Use->getValue("Namespace")
? Use->getValueAsString("Namespace")
: "";
[globalisel][tablegen] Add control-flow to the MatchTable. Summary: This will allow us to merge the various sub-tables into a single table. This is a compile-time saving at this point. However, this will also enable the optimization of a table so that similar instructions can be tested together, reducing the time spent on the matching the code. The bulk of this patch is a mechanical conversion to the new MatchTable object which is responsible for tracking label definitions and filling in the index of the jump targets. It is also responsible for nicely formatting the table. This was necessary to support the new GIM_Try opcode which takes the index to jump to if the match should fail. This value is unknown during table construction and is filled in during emission. To support nesting try-blocks (although we currently don't emit tables with nested try-blocks), GIM_Reject has been re-introduced to explicitly exit a try-block or fail the overall match if there are no active try-blocks. Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Reviewed By: rovka Subscribers: kristof.beyls, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D35117 llvm-svn: 308596
2017-07-20 17:25:44 +08:00
Table << MatchTable::Opcode("GIR_AddImplicitUse")
<< MatchTable::Comment("InsnID") << MatchTable::IntValue(InsnID)
<< MatchTable::NamedValue(Namespace, Use->getName())
<< MatchTable::LineBreak;
GlobalISel: add implicit defs & uses when mutating an instruction. Otherwise a scheduler might do bad things to the code we produce. llvm-svn: 298311
2017-03-21 05:58:23 +08:00
}
}
[globalisel] Decouple src pattern operands from dst pattern operands. Summary: This isn't testable for AArch64 by itself so this patch also adds support for constant immediates in the pattern and physical register uses in the result. The new IntOperandMatcher matches the constant in patterns such as '(set $rd:GPR32, (G_XOR $rs:GPR32, -1))'. It's always safe to fold immediates into an instruction so this is the first rule that will match across multiple BB's. The Renderer hierarchy is responsible for adding operands to the result instruction. Renderers can copy operands (CopyRenderer) or add physical registers (in particular %wzr and %xzr) to the result instruction in any order (OperandMatchers now import the operand names from SelectionDAG to allow renderers to access any operand). This allows us to emit the result instruction for: %1 = G_XOR %0, -1 --> %1 = ORNWrr %wzr, %0 %1 = G_XOR -1, %0 --> %1 = ORNWrr %wzr, %0 although the latter is untested since the matcher/importer has not been taught about commutativity yet. Added BuildMIAction which can build new instructions and mutate them where possible. W.r.t the mutation aspect, MatchActions are now told the name of an instruction they can recycle and BuildMIAction will emit mutation code when the renderers are appropriate. They are appropriate when all operands are rendered using CopyRenderer and the indices are the same as the matcher. This currently assumes that all operands have at least one matcher. Finally, this change also fixes a crash in AArch64InstructionSelector::select() caused by an immediate operand passing isImm() rather than isCImm(). This was uncovered by the other changes and was detected by existing tests. Depends on D29711 Reviewers: t.p.northover, ab, qcolombet, rovka, aditya_nandakumar, javed.absar Reviewed By: rovka Subscribers: aemerson, dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D29712 llvm-svn: 296131
2017-02-24 23:43:30 +08:00
return;
}
[globalisel] Make the MatchAction hierarchy consistent with the matchers. NFC. Reviewers: aditya_nandakumar, rovka, t.p.northover, qcolombet, ab Subscribers: dberris, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D29321 llvm-svn: 293760
2017-02-01 18:53:10 +08:00
[globalisel] Decouple src pattern operands from dst pattern operands. Summary: This isn't testable for AArch64 by itself so this patch also adds support for constant immediates in the pattern and physical register uses in the result. The new IntOperandMatcher matches the constant in patterns such as '(set $rd:GPR32, (G_XOR $rs:GPR32, -1))'. It's always safe to fold immediates into an instruction so this is the first rule that will match across multiple BB's. The Renderer hierarchy is responsible for adding operands to the result instruction. Renderers can copy operands (CopyRenderer) or add physical registers (in particular %wzr and %xzr) to the result instruction in any order (OperandMatchers now import the operand names from SelectionDAG to allow renderers to access any operand). This allows us to emit the result instruction for: %1 = G_XOR %0, -1 --> %1 = ORNWrr %wzr, %0 %1 = G_XOR -1, %0 --> %1 = ORNWrr %wzr, %0 although the latter is untested since the matcher/importer has not been taught about commutativity yet. Added BuildMIAction which can build new instructions and mutate them where possible. W.r.t the mutation aspect, MatchActions are now told the name of an instruction they can recycle and BuildMIAction will emit mutation code when the renderers are appropriate. They are appropriate when all operands are rendered using CopyRenderer and the indices are the same as the matcher. This currently assumes that all operands have at least one matcher. Finally, this change also fixes a crash in AArch64InstructionSelector::select() caused by an immediate operand passing isImm() rather than isCImm(). This was uncovered by the other changes and was detected by existing tests. Depends on D29711 Reviewers: t.p.northover, ab, qcolombet, rovka, aditya_nandakumar, javed.absar Reviewed By: rovka Subscribers: aemerson, dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D29712 llvm-svn: 296131
2017-02-24 23:43:30 +08:00
// TODO: Simple permutation looks like it could be almost as common as
// mutation due to commutative operations.
[globalisel][tablegen] Add control-flow to the MatchTable. Summary: This will allow us to merge the various sub-tables into a single table. This is a compile-time saving at this point. However, this will also enable the optimization of a table so that similar instructions can be tested together, reducing the time spent on the matching the code. The bulk of this patch is a mechanical conversion to the new MatchTable object which is responsible for tracking label definitions and filling in the index of the jump targets. It is also responsible for nicely formatting the table. This was necessary to support the new GIM_Try opcode which takes the index to jump to if the match should fail. This value is unknown during table construction and is filled in during emission. To support nesting try-blocks (although we currently don't emit tables with nested try-blocks), GIM_Reject has been re-introduced to explicitly exit a try-block or fail the overall match if there are no active try-blocks. Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Reviewed By: rovka Subscribers: kristof.beyls, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D35117 llvm-svn: 308596
2017-07-20 17:25:44 +08:00
Table << MatchTable::Opcode("GIR_BuildMI") << MatchTable::Comment("InsnID")
<< MatchTable::IntValue(InsnID) << MatchTable::Comment("Opcode")
<< MatchTable::NamedValue(I->Namespace, I->TheDef->getName())
<< MatchTable::LineBreak;
[globalisel] Decouple src pattern operands from dst pattern operands. Summary: This isn't testable for AArch64 by itself so this patch also adds support for constant immediates in the pattern and physical register uses in the result. The new IntOperandMatcher matches the constant in patterns such as '(set $rd:GPR32, (G_XOR $rs:GPR32, -1))'. It's always safe to fold immediates into an instruction so this is the first rule that will match across multiple BB's. The Renderer hierarchy is responsible for adding operands to the result instruction. Renderers can copy operands (CopyRenderer) or add physical registers (in particular %wzr and %xzr) to the result instruction in any order (OperandMatchers now import the operand names from SelectionDAG to allow renderers to access any operand). This allows us to emit the result instruction for: %1 = G_XOR %0, -1 --> %1 = ORNWrr %wzr, %0 %1 = G_XOR -1, %0 --> %1 = ORNWrr %wzr, %0 although the latter is untested since the matcher/importer has not been taught about commutativity yet. Added BuildMIAction which can build new instructions and mutate them where possible. W.r.t the mutation aspect, MatchActions are now told the name of an instruction they can recycle and BuildMIAction will emit mutation code when the renderers are appropriate. They are appropriate when all operands are rendered using CopyRenderer and the indices are the same as the matcher. This currently assumes that all operands have at least one matcher. Finally, this change also fixes a crash in AArch64InstructionSelector::select() caused by an immediate operand passing isImm() rather than isCImm(). This was uncovered by the other changes and was detected by existing tests. Depends on D29711 Reviewers: t.p.northover, ab, qcolombet, rovka, aditya_nandakumar, javed.absar Reviewed By: rovka Subscribers: aemerson, dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D29712 llvm-svn: 296131
2017-02-24 23:43:30 +08:00
for (const auto &Renderer : OperandRenderers)
[globalisel][tablegen] Add control-flow to the MatchTable. Summary: This will allow us to merge the various sub-tables into a single table. This is a compile-time saving at this point. However, this will also enable the optimization of a table so that similar instructions can be tested together, reducing the time spent on the matching the code. The bulk of this patch is a mechanical conversion to the new MatchTable object which is responsible for tracking label definitions and filling in the index of the jump targets. It is also responsible for nicely formatting the table. This was necessary to support the new GIM_Try opcode which takes the index to jump to if the match should fail. This value is unknown during table construction and is filled in during emission. To support nesting try-blocks (although we currently don't emit tables with nested try-blocks), GIM_Reject has been re-introduced to explicitly exit a try-block or fail the overall match if there are no active try-blocks. Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Reviewed By: rovka Subscribers: kristof.beyls, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D35117 llvm-svn: 308596
2017-07-20 17:25:44 +08:00
Renderer->emitRenderOpcodes(Table, Rule);
[globalisel][tablegen] Added instruction emission to the state-machine-based matcher. Summary: This further improves the compile-time regressions that will be caused by a re-commit of r303259. Also added included preliminary work in preparation for the multi-insn emitter since I needed to change the relevant part of the API for this patch anyway. Depends on D33758 Reviewers: rovka, vitalybuka, ab, t.p.northover, qcolombet, aditya_nandakumar Reviewed By: ab Subscribers: kristof.beyls, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D33764 llvm-svn: 307133
2017-07-05 17:39:33 +08:00
[GlobalISel] Only merge memory ops for mayLoad or mayStore instrs. Summary: We only need to merge memory operands for instructions that access memory. This slightly reduces the number of actions executed. Reviewers: MatzeB, rovka, dsanders Reviewed By: dsanders Subscribers: aemerson, igorb, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D36151 llvm-svn: 309944
2017-08-03 22:48:22 +08:00
if (I->mayLoad || I->mayStore) {
Table << MatchTable::Opcode("GIR_MergeMemOperands")
<< MatchTable::Comment("InsnID") << MatchTable::IntValue(InsnID)
<< MatchTable::Comment("MergeInsnID's");
// Emit the ID's for all the instructions that are matched by this rule.
// TODO: Limit this to matched instructions that mayLoad/mayStore or have
// some other means of having a memoperand. Also limit this to
// emitted instructions that expect to have a memoperand too. For
// example, (G_SEXT (G_LOAD x)) that results in separate load and
// sign-extend instructions shouldn't put the memoperand on the
// sign-extend since it has no effect there.
std::vector<unsigned> MergeInsnIDs;
for (const auto &IDMatcherPair : Rule.defined_insn_vars())
MergeInsnIDs.push_back(IDMatcherPair.second);
[TableGen] Change std::sort to llvm::sort in response to r327219 Summary: r327219 added wrappers to std::sort which randomly shuffle the container before sorting. This will help in uncovering non-determinism caused due to undefined sorting order of objects having the same key. To make use of that infrastructure we need to invoke llvm::sort instead of std::sort. Note: This patch is one of a series of patches to replace *all* std::sort to llvm::sort. Refer the comments section in D44363 for a list of all the required patches. Reviewers: stoklund, kparzysz, dsanders Reviewed By: dsanders Subscribers: llvm-commits Differential Revision: https://reviews.llvm.org/D45144 llvm-svn: 329451
2018-04-07 04:18:05 +08:00
llvm::sort(MergeInsnIDs.begin(), MergeInsnIDs.end());
[GlobalISel] Only merge memory ops for mayLoad or mayStore instrs. Summary: We only need to merge memory operands for instructions that access memory. This slightly reduces the number of actions executed. Reviewers: MatzeB, rovka, dsanders Reviewed By: dsanders Subscribers: aemerson, igorb, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D36151 llvm-svn: 309944
2017-08-03 22:48:22 +08:00
for (const auto &MergeInsnID : MergeInsnIDs)
Table << MatchTable::IntValue(MergeInsnID);
[globalisel][tablegen] Add support for importing 'imm' operands. Summary: This patch enables the import of rules containing 'imm' operands that do not constrain the acceptable values using predicates. Support for ImmLeaf will arrive in a later patch. Depends on D35681 Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Reviewed By: rovka Subscribers: kristof.beyls, javed.absar, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D35833 llvm-svn: 310343
2017-08-08 18:44:31 +08:00
Table << MatchTable::NamedValue("GIU_MergeMemOperands_EndOfList")
<< MatchTable::LineBreak;
[GlobalISel] Only merge memory ops for mayLoad or mayStore instrs. Summary: We only need to merge memory operands for instructions that access memory. This slightly reduces the number of actions executed. Reviewers: MatzeB, rovka, dsanders Reviewed By: dsanders Subscribers: aemerson, igorb, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D36151 llvm-svn: 309944
2017-08-03 22:48:22 +08:00
}
[globalisel][tablegen] Add support for multi-insn emission The importer will now accept nested instructions in the result pattern such as (ADDWrr $a, (SUBWrr $b, $c)). This is only valid when the nested instruction def's a single vreg and the parent instruction consumes a single vreg where a nested instruction is specified. The importer will automatically create a vreg to connect the two using the type information from the pattern. This vreg will be constrained to the register classes given in the instruction definitions*. * REG_SEQUENCE is explicitly rejected because of this. The definition doesn't constrain to a register class and it therefore needs special handling. llvm-svn: 317117
2017-11-02 03:57:57 +08:00
// FIXME: This is a hack but it's sufficient for ISel. We'll need to do
// better for combines. Particularly when there are multiple match
// roots.
if (InsnID == 0)
Table << MatchTable::Opcode("GIR_EraseFromParent")
<< MatchTable::Comment("InsnID") << MatchTable::IntValue(InsnID)
<< MatchTable::LineBreak;
[globalisel][tablegen] Add support for COPY_TO_REGCLASS. Summary: As part of this * Emitted instructions now have named MachineInstr variables associated with them. This isn't particularly important yet but it's a small step towards multiple-insn emission. * constrainSelectedInstRegOperands() is no longer hardcoded. It's now added as the ConstrainOperandsToDefinitionAction() action. COPY_TO_REGCLASS uses an alternate constraint mechanism ConstrainOperandToRegClassAction() which supports arbitrary constraints such as that defined by COPY_TO_REGCLASS. Reviewers: ab, qcolombet, t.p.northover, rovka, kristof.beyls, aditya_nandakumar Reviewed By: ab Subscribers: javed.absar, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D33590 llvm-svn: 305791
2017-06-20 20:36:34 +08:00
}
};
/// Generates code to constrain the operands of an output instruction to the
/// register classes specified by the definition of that instruction.
class ConstrainOperandsToDefinitionAction : public MatchAction {
[globalisel][tablegen] Added instruction emission to the state-machine-based matcher. Summary: This further improves the compile-time regressions that will be caused by a re-commit of r303259. Also added included preliminary work in preparation for the multi-insn emitter since I needed to change the relevant part of the API for this patch anyway. Depends on D33758 Reviewers: rovka, vitalybuka, ab, t.p.northover, qcolombet, aditya_nandakumar Reviewed By: ab Subscribers: kristof.beyls, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D33764 llvm-svn: 307133
2017-07-05 17:39:33 +08:00
unsigned InsnID;
[globalisel][tablegen] Add support for COPY_TO_REGCLASS. Summary: As part of this * Emitted instructions now have named MachineInstr variables associated with them. This isn't particularly important yet but it's a small step towards multiple-insn emission. * constrainSelectedInstRegOperands() is no longer hardcoded. It's now added as the ConstrainOperandsToDefinitionAction() action. COPY_TO_REGCLASS uses an alternate constraint mechanism ConstrainOperandToRegClassAction() which supports arbitrary constraints such as that defined by COPY_TO_REGCLASS. Reviewers: ab, qcolombet, t.p.northover, rovka, kristof.beyls, aditya_nandakumar Reviewed By: ab Subscribers: javed.absar, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D33590 llvm-svn: 305791
2017-06-20 20:36:34 +08:00
public:
[globalisel][tablegen] Added instruction emission to the state-machine-based matcher. Summary: This further improves the compile-time regressions that will be caused by a re-commit of r303259. Also added included preliminary work in preparation for the multi-insn emitter since I needed to change the relevant part of the API for this patch anyway. Depends on D33758 Reviewers: rovka, vitalybuka, ab, t.p.northover, qcolombet, aditya_nandakumar Reviewed By: ab Subscribers: kristof.beyls, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D33764 llvm-svn: 307133
2017-07-05 17:39:33 +08:00
ConstrainOperandsToDefinitionAction(unsigned InsnID) : InsnID(InsnID) {}
[globalisel][tablegen] Add support for COPY_TO_REGCLASS. Summary: As part of this * Emitted instructions now have named MachineInstr variables associated with them. This isn't particularly important yet but it's a small step towards multiple-insn emission. * constrainSelectedInstRegOperands() is no longer hardcoded. It's now added as the ConstrainOperandsToDefinitionAction() action. COPY_TO_REGCLASS uses an alternate constraint mechanism ConstrainOperandToRegClassAction() which supports arbitrary constraints such as that defined by COPY_TO_REGCLASS. Reviewers: ab, qcolombet, t.p.northover, rovka, kristof.beyls, aditya_nandakumar Reviewed By: ab Subscribers: javed.absar, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D33590 llvm-svn: 305791
2017-06-20 20:36:34 +08:00
[globalisel][tablegen] Add infrastructure to potentially allow BuildMIAction to choose a mutatable instruction. NFC Prepare for multiple instruction emission by allowing BuildMIAction to search for a suitable matcher that will support mutation. This patch deliberately neglects to add matchers aside from the root to preserve NFC. That said, it should be noted that until we support mutations other than just the opcode the chances of finding a non-root instruction for which canMutate() is true, is essentially zero. Furthermore in the presence of multi-instruction emission the chances of finding any instruction for which canMutate() is true is also zero. Nevertheless, we can't continue to require that all BuildMIAction's consider the root of the match to be recyclable due to the risk of recycling it twice in the same rule. llvm-svn: 317022
2017-11-01 02:50:24 +08:00
void emitActionOpcodes(MatchTable &Table, RuleMatcher &Rule) const override {
[globalisel][tablegen] Add control-flow to the MatchTable. Summary: This will allow us to merge the various sub-tables into a single table. This is a compile-time saving at this point. However, this will also enable the optimization of a table so that similar instructions can be tested together, reducing the time spent on the matching the code. The bulk of this patch is a mechanical conversion to the new MatchTable object which is responsible for tracking label definitions and filling in the index of the jump targets. It is also responsible for nicely formatting the table. This was necessary to support the new GIM_Try opcode which takes the index to jump to if the match should fail. This value is unknown during table construction and is filled in during emission. To support nesting try-blocks (although we currently don't emit tables with nested try-blocks), GIM_Reject has been re-introduced to explicitly exit a try-block or fail the overall match if there are no active try-blocks. Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Reviewed By: rovka Subscribers: kristof.beyls, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D35117 llvm-svn: 308596
2017-07-20 17:25:44 +08:00
Table << MatchTable::Opcode("GIR_ConstrainSelectedInstOperands")
<< MatchTable::Comment("InsnID") << MatchTable::IntValue(InsnID)
<< MatchTable::LineBreak;
[globalisel][tablegen] Add support for COPY_TO_REGCLASS. Summary: As part of this * Emitted instructions now have named MachineInstr variables associated with them. This isn't particularly important yet but it's a small step towards multiple-insn emission. * constrainSelectedInstRegOperands() is no longer hardcoded. It's now added as the ConstrainOperandsToDefinitionAction() action. COPY_TO_REGCLASS uses an alternate constraint mechanism ConstrainOperandToRegClassAction() which supports arbitrary constraints such as that defined by COPY_TO_REGCLASS. Reviewers: ab, qcolombet, t.p.northover, rovka, kristof.beyls, aditya_nandakumar Reviewed By: ab Subscribers: javed.absar, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D33590 llvm-svn: 305791
2017-06-20 20:36:34 +08:00
}
};
/// Generates code to constrain the specified operand of an output instruction
/// to the specified register class.
class ConstrainOperandToRegClassAction : public MatchAction {
[globalisel][tablegen] Added instruction emission to the state-machine-based matcher. Summary: This further improves the compile-time regressions that will be caused by a re-commit of r303259. Also added included preliminary work in preparation for the multi-insn emitter since I needed to change the relevant part of the API for this patch anyway. Depends on D33758 Reviewers: rovka, vitalybuka, ab, t.p.northover, qcolombet, aditya_nandakumar Reviewed By: ab Subscribers: kristof.beyls, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D33764 llvm-svn: 307133
2017-07-05 17:39:33 +08:00
unsigned InsnID;
[globalisel][tablegen] Add support for COPY_TO_REGCLASS. Summary: As part of this * Emitted instructions now have named MachineInstr variables associated with them. This isn't particularly important yet but it's a small step towards multiple-insn emission. * constrainSelectedInstRegOperands() is no longer hardcoded. It's now added as the ConstrainOperandsToDefinitionAction() action. COPY_TO_REGCLASS uses an alternate constraint mechanism ConstrainOperandToRegClassAction() which supports arbitrary constraints such as that defined by COPY_TO_REGCLASS. Reviewers: ab, qcolombet, t.p.northover, rovka, kristof.beyls, aditya_nandakumar Reviewed By: ab Subscribers: javed.absar, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D33590 llvm-svn: 305791
2017-06-20 20:36:34 +08:00
unsigned OpIdx;
const CodeGenRegisterClass &RC;
public:
[globalisel][tablegen] Added instruction emission to the state-machine-based matcher. Summary: This further improves the compile-time regressions that will be caused by a re-commit of r303259. Also added included preliminary work in preparation for the multi-insn emitter since I needed to change the relevant part of the API for this patch anyway. Depends on D33758 Reviewers: rovka, vitalybuka, ab, t.p.northover, qcolombet, aditya_nandakumar Reviewed By: ab Subscribers: kristof.beyls, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D33764 llvm-svn: 307133
2017-07-05 17:39:33 +08:00
ConstrainOperandToRegClassAction(unsigned InsnID, unsigned OpIdx,
[globalisel][tablegen] Add support for COPY_TO_REGCLASS. Summary: As part of this * Emitted instructions now have named MachineInstr variables associated with them. This isn't particularly important yet but it's a small step towards multiple-insn emission. * constrainSelectedInstRegOperands() is no longer hardcoded. It's now added as the ConstrainOperandsToDefinitionAction() action. COPY_TO_REGCLASS uses an alternate constraint mechanism ConstrainOperandToRegClassAction() which supports arbitrary constraints such as that defined by COPY_TO_REGCLASS. Reviewers: ab, qcolombet, t.p.northover, rovka, kristof.beyls, aditya_nandakumar Reviewed By: ab Subscribers: javed.absar, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D33590 llvm-svn: 305791
2017-06-20 20:36:34 +08:00
const CodeGenRegisterClass &RC)
[globalisel][tablegen] Added instruction emission to the state-machine-based matcher. Summary: This further improves the compile-time regressions that will be caused by a re-commit of r303259. Also added included preliminary work in preparation for the multi-insn emitter since I needed to change the relevant part of the API for this patch anyway. Depends on D33758 Reviewers: rovka, vitalybuka, ab, t.p.northover, qcolombet, aditya_nandakumar Reviewed By: ab Subscribers: kristof.beyls, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D33764 llvm-svn: 307133
2017-07-05 17:39:33 +08:00
: InsnID(InsnID), OpIdx(OpIdx), RC(RC) {}
[globalisel][tablegen] Add support for COPY_TO_REGCLASS. Summary: As part of this * Emitted instructions now have named MachineInstr variables associated with them. This isn't particularly important yet but it's a small step towards multiple-insn emission. * constrainSelectedInstRegOperands() is no longer hardcoded. It's now added as the ConstrainOperandsToDefinitionAction() action. COPY_TO_REGCLASS uses an alternate constraint mechanism ConstrainOperandToRegClassAction() which supports arbitrary constraints such as that defined by COPY_TO_REGCLASS. Reviewers: ab, qcolombet, t.p.northover, rovka, kristof.beyls, aditya_nandakumar Reviewed By: ab Subscribers: javed.absar, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D33590 llvm-svn: 305791
2017-06-20 20:36:34 +08:00
[globalisel][tablegen] Add infrastructure to potentially allow BuildMIAction to choose a mutatable instruction. NFC Prepare for multiple instruction emission by allowing BuildMIAction to search for a suitable matcher that will support mutation. This patch deliberately neglects to add matchers aside from the root to preserve NFC. That said, it should be noted that until we support mutations other than just the opcode the chances of finding a non-root instruction for which canMutate() is true, is essentially zero. Furthermore in the presence of multi-instruction emission the chances of finding any instruction for which canMutate() is true is also zero. Nevertheless, we can't continue to require that all BuildMIAction's consider the root of the match to be recyclable due to the risk of recycling it twice in the same rule. llvm-svn: 317022
2017-11-01 02:50:24 +08:00
void emitActionOpcodes(MatchTable &Table, RuleMatcher &Rule) const override {
[globalisel][tablegen] Add control-flow to the MatchTable. Summary: This will allow us to merge the various sub-tables into a single table. This is a compile-time saving at this point. However, this will also enable the optimization of a table so that similar instructions can be tested together, reducing the time spent on the matching the code. The bulk of this patch is a mechanical conversion to the new MatchTable object which is responsible for tracking label definitions and filling in the index of the jump targets. It is also responsible for nicely formatting the table. This was necessary to support the new GIM_Try opcode which takes the index to jump to if the match should fail. This value is unknown during table construction and is filled in during emission. To support nesting try-blocks (although we currently don't emit tables with nested try-blocks), GIM_Reject has been re-introduced to explicitly exit a try-block or fail the overall match if there are no active try-blocks. Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Reviewed By: rovka Subscribers: kristof.beyls, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D35117 llvm-svn: 308596
2017-07-20 17:25:44 +08:00
Table << MatchTable::Opcode("GIR_ConstrainOperandRC")
<< MatchTable::Comment("InsnID") << MatchTable::IntValue(InsnID)
<< MatchTable::Comment("Op") << MatchTable::IntValue(OpIdx)
<< MatchTable::Comment("RC " + RC.getName())
<< MatchTable::IntValue(RC.EnumValue) << MatchTable::LineBreak;
[GlobalISel] Add basic Selector-emitter tblgen backend. This adds a basic tablegen backend that analyzes the SelectionDAG patterns to find simple ones that are eligible for GlobalISel-emission. That's similar to FastISel, with one notable difference: we're not fed ISD opcodes, so we need to map the SDNode operators to generic opcodes. That's done using GINodeEquiv in TargetGlobalISel.td. Otherwise, this is mostly boilerplate, and lots of filtering of any kind of "complicated" pattern. On AArch64, this is sufficient to match G_ADD up to s64 (to ADDWrr/ADDXrr) and G_BR (to B). Differential Revision: https://reviews.llvm.org/D26878 llvm-svn: 290284
2016-12-22 07:26:20 +08:00
}
};
[globalisel][tablegen] Add support for multi-insn emission The importer will now accept nested instructions in the result pattern such as (ADDWrr $a, (SUBWrr $b, $c)). This is only valid when the nested instruction def's a single vreg and the parent instruction consumes a single vreg where a nested instruction is specified. The importer will automatically create a vreg to connect the two using the type information from the pattern. This vreg will be constrained to the register classes given in the instruction definitions*. * REG_SEQUENCE is explicitly rejected because of this. The definition doesn't constrain to a register class and it therefore needs special handling. llvm-svn: 317117
2017-11-02 03:57:57 +08:00
/// Generates code to create a temporary register which can be used to chain
/// instructions together.
class MakeTempRegisterAction : public MatchAction {
private:
LLTCodeGen Ty;
unsigned TempRegID;
public:
MakeTempRegisterAction(const LLTCodeGen &Ty, unsigned TempRegID)
: Ty(Ty), TempRegID(TempRegID) {}
void emitActionOpcodes(MatchTable &Table, RuleMatcher &Rule) const override {
Table << MatchTable::Opcode("GIR_MakeTempReg")
<< MatchTable::Comment("TempRegID") << MatchTable::IntValue(TempRegID)
<< MatchTable::Comment("TypeID")
<< MatchTable::NamedValue(Ty.getCxxEnumValue())
<< MatchTable::LineBreak;
}
};
[globalisel][tablegen] Add support for importing 'imm' operands. Summary: This patch enables the import of rules containing 'imm' operands that do not constrain the acceptable values using predicates. Support for ImmLeaf will arrive in a later patch. Depends on D35681 Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Reviewed By: rovka Subscribers: kristof.beyls, javed.absar, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D35833 llvm-svn: 310343
2017-08-08 18:44:31 +08:00
InstructionMatcher &RuleMatcher::addInstructionMatcher(StringRef SymbolicName) {
[globalisel][tablegen] Simplify named operand/operator lookups and fix a wrong-code bug this revealed. Summary: Operand variable lookups are now performed by the RuleMatcher rather than searching the whole matcher hierarchy for a match. This revealed a wrong-code bug that currently affects ARM and X86 where patterns that use a variable more than once in the match pattern will be imported but won't check that the operands are identical. This can cause the tablegen-erated matcher to accept matches that should be rejected. Depends on D36569 Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Subscribers: aemerson, igorb, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D36618 llvm-svn: 315780
2017-10-14 08:31:58 +08:00
Matchers.emplace_back(new InstructionMatcher(*this, SymbolicName));
[globalisel][tablegen] Add infrastructure to potentially allow BuildMIAction to choose a mutatable instruction. NFC Prepare for multiple instruction emission by allowing BuildMIAction to search for a suitable matcher that will support mutation. This patch deliberately neglects to add matchers aside from the root to preserve NFC. That said, it should be noted that until we support mutations other than just the opcode the chances of finding a non-root instruction for which canMutate() is true, is essentially zero. Furthermore in the presence of multi-instruction emission the chances of finding any instruction for which canMutate() is true is also zero. Nevertheless, we can't continue to require that all BuildMIAction's consider the root of the match to be recyclable due to the risk of recycling it twice in the same rule. llvm-svn: 317022
2017-11-01 02:50:24 +08:00
MutatableInsns.insert(Matchers.back().get());
[tablegen][globalisel] Trivial changes to reduce size of following patch. NFC. Summary: * Move namespace { * Trivial: Typo * RuleMatcher: Separate class and definition * Trivial: const findNodeEquiv Reviewers: t.p.northover, qcolombet, aditya_nandakumar, ab, rovka Reviewed By: rovka Subscribers: dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D30531 llvm-svn: 297884
2017-03-16 04:18:38 +08:00
return *Matchers.back();
}
[globalisel] Re-factor ISel matchers into a hierarchy. NFC Summary: This should make it possible to easily add everything needed to import all the existing SelectionDAG rules. It should also serve the likely kinds of GlobalISel rules (some of which are not currently representable in SelectionDAG) once we've nailed down the tablegen definition for that. The hierarchy is as follows: MatcherRule - A matching rule. Currently used to emit C++ ISel code but will | also be used to emit test cases and tablegen definitions in the | near future. |- Instruction(s) - Represents the instruction to be matched. |- Instruction Predicate(s) - Test the opcode, arithmetic flags, etc. of an | instruction. \- Operand(s) - Represents a particular operand of the instruction. In the | future, there may be subclasses to test the same predicates | on multiple operands (including for variadic instructions). \ Operand Predicate(s) - Test the type, register bank, etc. of an operand. This is where the ComplexPattern equivalent will be represented. It's also nested-instruction matching will live as a predicate that follows the DefUse chain to the Def and tests a MatcherRule from that position. Support for multiple instruction matchers in a rule has been retained from the existing code but has been adjusted to assert when it is used. Previously it would silently drop all but the first instruction matcher. The tablegen-erated file is not functionally changed but has more parentheses and no longer attempts to format the if-statements since keeping track of the indentation is tricky in the presence of the matcher hierarchy. It would be nice to have CMakes tablegen() run the output through clang-format (when available) so we don't have to complicate TableGen with pretty-printing. It's also worth mentioning that this hierarchy will also be able to emit TableGen definitions and test cases in the near future. This is the reason for favouring explicit emit*() calls rather than the << operator. Reviewers: aditya_nandakumar, rovka, t.p.northover, qcolombet, ab Reviewed By: ab Subscribers: igorb, dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D28942 llvm-svn: 293172
2017-01-26 19:10:14 +08:00
[globalisel][tablegen] Import SelectionDAG's rule predicates and support the equivalent in GIRule. Summary: The SelectionDAG importer now imports rules with Predicate's attached via Requires, PredicateControl, etc. These predicates are implemented as bitset's to allow multiple predicates to be tested together. However, unlike the MC layer subtarget features, each target only pays for it's own predicates (e.g. AArch64 doesn't have 192 feature bits just because X86 needs a lot). Both AArch64 and X86 derive at least one predicate from the MachineFunction or Function so they must re-initialize AvailableFeatures before each function. They also declare locals in <Target>InstructionSelector so that computeAvailableFeatures() can use the code from SelectionDAG without modification. Reviewers: rovka, qcolombet, aditya_nandakumar, t.p.northover, ab Reviewed By: rovka Subscribers: aemerson, rengolin, dberris, kristof.beyls, llvm-commits, igorb Differential Revision: https://reviews.llvm.org/D31418 llvm-svn: 300993
2017-04-21 23:59:56 +08:00
void RuleMatcher::addRequiredFeature(Record *Feature) {
RequiredFeatures.push_back(Feature);
}
[globalisel][tablegen] Partially fix compile-time regressions by converting matcher to state-machine(s) Summary: Replace the matcher if-statements for each rule with a state-machine. This significantly reduces compile time, memory allocations, and cumulative memory allocation when compiling AArch64InstructionSelector.cpp.o after r303259 is recommitted. The following patches will expand on this further to fully fix the regressions. Reviewers: rovka, ab, t.p.northover, qcolombet, aditya_nandakumar Reviewed By: ab Subscribers: vitalybuka, aemerson, javed.absar, igorb, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D33758 llvm-svn: 307079
2017-07-04 22:35:06 +08:00
const std::vector<Record *> &RuleMatcher::getRequiredFeatures() const {
return RequiredFeatures;
}
Re-commit: [globalisel][tablegen] Keep track of the insertion point while adding BuildMIAction's. NFC Multi-instruction emission needs to ensure the the instructions are generated a depth-first fashion. For example: (ADDWrr (SUBWrr a, b), c) needs to emit the SUBWrr before the ADDWrr. However, our walk over TreePatternNode's is highly context sensitive which makes it difficult to append BuildMIActions in the order we want. To fix this, we now keep track of the insertion point as we add actions. This will allow multi-insn emission to insert BuildMI's in the correct place. The previous commit failed on the Ubuntu bots using GCC 4.8. These bots lack the const_iterator forms of insert() and emplace() that were added in C++11. As a result I've switched the const_iterators to iterators. llvm-svn: 317049
2017-11-01 07:03:18 +08:00
// Emplaces an action of the specified Kind at the end of the action list.
//
// Returns a reference to the newly created action.
//
// Like std::vector::emplace_back(), may invalidate all iterators if the new
// size exceeds the capacity. Otherwise, only invalidates the past-the-end
// iterator.
[tablegen][globalisel] Trivial changes to reduce size of following patch. NFC. Summary: * Move namespace { * Trivial: Typo * RuleMatcher: Separate class and definition * Trivial: const findNodeEquiv Reviewers: t.p.northover, qcolombet, aditya_nandakumar, ab, rovka Reviewed By: rovka Subscribers: dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D30531 llvm-svn: 297884
2017-03-16 04:18:38 +08:00
template <class Kind, class... Args>
Kind &RuleMatcher::addAction(Args &&... args) {
Actions.emplace_back(llvm::make_unique<Kind>(std::forward<Args>(args)...));
return *static_cast<Kind *>(Actions.back().get());
}
[globalisel][tablegen] Add support for multi-insn emission The importer will now accept nested instructions in the result pattern such as (ADDWrr $a, (SUBWrr $b, $c)). This is only valid when the nested instruction def's a single vreg and the parent instruction consumes a single vreg where a nested instruction is specified. The importer will automatically create a vreg to connect the two using the type information from the pattern. This vreg will be constrained to the register classes given in the instruction definitions*. * REG_SEQUENCE is explicitly rejected because of this. The definition doesn't constrain to a register class and it therefore needs special handling. llvm-svn: 317117
2017-11-02 03:57:57 +08:00
Re-commit: [globalisel][tablegen] Keep track of the insertion point while adding BuildMIAction's. NFC Multi-instruction emission needs to ensure the the instructions are generated a depth-first fashion. For example: (ADDWrr (SUBWrr a, b), c) needs to emit the SUBWrr before the ADDWrr. However, our walk over TreePatternNode's is highly context sensitive which makes it difficult to append BuildMIActions in the order we want. To fix this, we now keep track of the insertion point as we add actions. This will allow multi-insn emission to insert BuildMI's in the correct place. The previous commit failed on the Ubuntu bots using GCC 4.8. These bots lack the const_iterator forms of insert() and emplace() that were added in C++11. As a result I've switched the const_iterators to iterators. llvm-svn: 317049
2017-11-01 07:03:18 +08:00
// Emplaces an action of the specified Kind before the given insertion point.
//
// Returns an iterator pointing at the newly created instruction.
//
// Like std::vector::insert(), may invalidate all iterators if the new size
// exceeds the capacity. Otherwise, only invalidates the iterators from the
// insertion point onwards.
template <class Kind, class... Args>
action_iterator RuleMatcher::insertAction(action_iterator InsertPt,
Args &&... args) {
[globalisel][tablegen] Add support for multi-insn emission The importer will now accept nested instructions in the result pattern such as (ADDWrr $a, (SUBWrr $b, $c)). This is only valid when the nested instruction def's a single vreg and the parent instruction consumes a single vreg where a nested instruction is specified. The importer will automatically create a vreg to connect the two using the type information from the pattern. This vreg will be constrained to the register classes given in the instruction definitions*. * REG_SEQUENCE is explicitly rejected because of this. The definition doesn't constrain to a register class and it therefore needs special handling. llvm-svn: 317117
2017-11-02 03:57:57 +08:00
return Actions.emplace(InsertPt,
llvm::make_unique<Kind>(std::forward<Args>(args)...));
Re-commit: [globalisel][tablegen] Keep track of the insertion point while adding BuildMIAction's. NFC Multi-instruction emission needs to ensure the the instructions are generated a depth-first fashion. For example: (ADDWrr (SUBWrr a, b), c) needs to emit the SUBWrr before the ADDWrr. However, our walk over TreePatternNode's is highly context sensitive which makes it difficult to append BuildMIActions in the order we want. To fix this, we now keep track of the insertion point as we add actions. This will allow multi-insn emission to insert BuildMI's in the correct place. The previous commit failed on the Ubuntu bots using GCC 4.8. These bots lack the const_iterator forms of insert() and emplace() that were added in C++11. As a result I've switched the const_iterators to iterators. llvm-svn: 317049
2017-11-01 07:03:18 +08:00
}
[GlobalISel] Add basic Selector-emitter tblgen backend. This adds a basic tablegen backend that analyzes the SelectionDAG patterns to find simple ones that are eligible for GlobalISel-emission. That's similar to FastISel, with one notable difference: we're not fed ISD opcodes, so we need to map the SDNode operators to generic opcodes. That's done using GINodeEquiv in TargetGlobalISel.td. Otherwise, this is mostly boilerplate, and lots of filtering of any kind of "complicated" pattern. On AArch64, this is sufficient to match G_ADD up to s64 (to ADDWrr/ADDXrr) and G_BR (to B). Differential Revision: https://reviews.llvm.org/D26878 llvm-svn: 290284
2016-12-22 07:26:20 +08:00
Reapply r332907 "[GlobalISel] Improving InstructionSelect's performance by reducing MatchTable..." Apparently the compile time problem was caused by the fact that not all compilers / STL implementations can automatically convert std::unique_ptr<Derived> to std::unique_ptr<Base>. Fixed (hopefully) by making sure it's std::unique_ptr<Derived>&& (rvalue ref) to std::unique_ptr<Base> conversion instead. llvm-svn: 332917
2018-05-22 07:28:51 +08:00
unsigned RuleMatcher::implicitlyDefineInsnVar(InstructionMatcher &Matcher) {
[globalisel][tablegen] Partially fix compile-time regressions by converting matcher to state-machine(s) Summary: Replace the matcher if-statements for each rule with a state-machine. This significantly reduces compile time, memory allocations, and cumulative memory allocation when compiling AArch64InstructionSelector.cpp.o after r303259 is recommitted. The following patches will expand on this further to fully fix the regressions. Reviewers: rovka, ab, t.p.northover, qcolombet, aditya_nandakumar Reviewed By: ab Subscribers: vitalybuka, aemerson, javed.absar, igorb, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D33758 llvm-svn: 307079
2017-07-04 22:35:06 +08:00
unsigned NewInsnVarID = NextInsnVarID++;
InsnVariableIDs[&Matcher] = NewInsnVarID;
return NewInsnVarID;
[tablegen][globalisel] Capture instructions into locals and related infrastructure for multiple instructions matches. Summary: Prepare the way for nested instruction matching support by having actions like CopyRenderer look up operands in the RuleMatcher rather than a specific InstructionMatcher. This allows actions to reference any operand from any matched instruction. It works by checking the 'shape' of the match and capturing each matched instruction to a local variable. If the shape is wrong (not enough operands, leaf nodes where non-leafs are expected, etc.), then the rule exits early without checking the predicates. Once we've captured the instructions, we then test the predicates as before (except using the local variables). If the match is successful, then we render the new instruction as before using the local variables. It's not noticable in this patch but by the time we support multiple instruction matching, this patch will also cause a significant improvement to readability of the emitted code since MRI.getVRegDef(I->getOperand(0).getReg()) will simply be MI1 after emitCxxCaptureStmts(). This isn't quite NFC because I've also fixed a bug that I'm surprised we haven't encountered yet. It now checks there are at least the expected number of operands before accessing them with getOperand(). Depends on D30531 Reviewers: t.p.northover, qcolombet, aditya_nandakumar, ab, rovka Reviewed By: rovka Subscribers: dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D30535 llvm-svn: 298257
2017-03-20 23:20:42 +08:00
}
Reapply r332907 "[GlobalISel] Improving InstructionSelect's performance by reducing MatchTable..." Apparently the compile time problem was caused by the fact that not all compilers / STL implementations can automatically convert std::unique_ptr<Derived> to std::unique_ptr<Base>. Fixed (hopefully) by making sure it's std::unique_ptr<Derived>&& (rvalue ref) to std::unique_ptr<Base> conversion instead. llvm-svn: 332917
2018-05-22 07:28:51 +08:00
unsigned RuleMatcher::getInsnVarID(InstructionMatcher &InsnMatcher) const {
[globalisel][tablegen] Partially fix compile-time regressions by converting matcher to state-machine(s) Summary: Replace the matcher if-statements for each rule with a state-machine. This significantly reduces compile time, memory allocations, and cumulative memory allocation when compiling AArch64InstructionSelector.cpp.o after r303259 is recommitted. The following patches will expand on this further to fully fix the regressions. Reviewers: rovka, ab, t.p.northover, qcolombet, aditya_nandakumar Reviewed By: ab Subscribers: vitalybuka, aemerson, javed.absar, igorb, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D33758 llvm-svn: 307079
2017-07-04 22:35:06 +08:00
const auto &I = InsnVariableIDs.find(&InsnMatcher);
if (I != InsnVariableIDs.end())
[tablegen][globalisel] Capture instructions into locals and related infrastructure for multiple instructions matches. Summary: Prepare the way for nested instruction matching support by having actions like CopyRenderer look up operands in the RuleMatcher rather than a specific InstructionMatcher. This allows actions to reference any operand from any matched instruction. It works by checking the 'shape' of the match and capturing each matched instruction to a local variable. If the shape is wrong (not enough operands, leaf nodes where non-leafs are expected, etc.), then the rule exits early without checking the predicates. Once we've captured the instructions, we then test the predicates as before (except using the local variables). If the match is successful, then we render the new instruction as before using the local variables. It's not noticable in this patch but by the time we support multiple instruction matching, this patch will also cause a significant improvement to readability of the emitted code since MRI.getVRegDef(I->getOperand(0).getReg()) will simply be MI1 after emitCxxCaptureStmts(). This isn't quite NFC because I've also fixed a bug that I'm surprised we haven't encountered yet. It now checks there are at least the expected number of operands before accessing them with getOperand(). Depends on D30531 Reviewers: t.p.northover, qcolombet, aditya_nandakumar, ab, rovka Reviewed By: rovka Subscribers: dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D30535 llvm-svn: 298257
2017-03-20 23:20:42 +08:00
return I->second;
llvm_unreachable("Matched Insn was not captured in a local variable");
}
[globalisel][tablegen] Simplify named operand/operator lookups and fix a wrong-code bug this revealed. Summary: Operand variable lookups are now performed by the RuleMatcher rather than searching the whole matcher hierarchy for a match. This revealed a wrong-code bug that currently affects ARM and X86 where patterns that use a variable more than once in the match pattern will be imported but won't check that the operands are identical. This can cause the tablegen-erated matcher to accept matches that should be rejected. Depends on D36569 Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Subscribers: aemerson, igorb, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D36618 llvm-svn: 315780
2017-10-14 08:31:58 +08:00
void RuleMatcher::defineOperand(StringRef SymbolicName, OperandMatcher &OM) {
if (DefinedOperands.find(SymbolicName) == DefinedOperands.end()) {
DefinedOperands[SymbolicName] = &OM;
return;
}
// If the operand is already defined, then we must ensure both references in
// the matcher have the exact same node.
OM.addPredicate<SameOperandMatcher>(OM.getSymbolicName());
}
Reapply r332907 "[GlobalISel] Improving InstructionSelect's performance by reducing MatchTable..." Apparently the compile time problem was caused by the fact that not all compilers / STL implementations can automatically convert std::unique_ptr<Derived> to std::unique_ptr<Base>. Fixed (hopefully) by making sure it's std::unique_ptr<Derived>&& (rvalue ref) to std::unique_ptr<Base> conversion instead. llvm-svn: 332917
2018-05-22 07:28:51 +08:00
InstructionMatcher &
[globalisel][tablegen] Add support for importing 'imm' operands. Summary: This patch enables the import of rules containing 'imm' operands that do not constrain the acceptable values using predicates. Support for ImmLeaf will arrive in a later patch. Depends on D35681 Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Reviewed By: rovka Subscribers: kristof.beyls, javed.absar, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D35833 llvm-svn: 310343
2017-08-08 18:44:31 +08:00
RuleMatcher::getInstructionMatcher(StringRef SymbolicName) const {
for (const auto &I : InsnVariableIDs)
if (I.first->getSymbolicName() == SymbolicName)
return *I.first;
llvm_unreachable(
("Failed to lookup instruction " + SymbolicName).str().c_str());
}
[globalisel][tablegen] Simplify named operand/operator lookups and fix a wrong-code bug this revealed. Summary: Operand variable lookups are now performed by the RuleMatcher rather than searching the whole matcher hierarchy for a match. This revealed a wrong-code bug that currently affects ARM and X86 where patterns that use a variable more than once in the match pattern will be imported but won't check that the operands are identical. This can cause the tablegen-erated matcher to accept matches that should be rejected. Depends on D36569 Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Subscribers: aemerson, igorb, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D36618 llvm-svn: 315780
2017-10-14 08:31:58 +08:00
const OperandMatcher &
RuleMatcher::getOperandMatcher(StringRef Name) const {
const auto &I = DefinedOperands.find(Name);
if (I == DefinedOperands.end())
PrintFatalError(SrcLoc, "Operand " + Name + " was not declared in matcher");
return *I->second;
}
Re-commit: r309094 [globalisel][tablegen] Fuse the generated tables together. Summary: Now that we have control flow in place, fuse the per-rule tables into a single table. This is a compile-time saving at this point. However, this will also enable the optimization of a table so that similar instructions can be tested together, reducing the time spent on the matching the code. This is NFC in terms of externally visible behaviour but some internals have changed slightly. State.MIs is no longer reset between each rule that is attempted because it's not necessary to do so. As a consequence of this the restriction on the order that instructions are added to State.MIs has been relaxed to only affect recorded instructions that require new elements to be added to the vector. GIM_RecordInsn can now write to any element from 1 to State.MIs.size() instead of just State.MIs.size(). The compile-time regressions from the last commit were caused by the ARM target including a non-const variable (zero_reg) in the table and therefore generating an initializer for it. That variable is now const. Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Reviewed By: rovka Subscribers: kristof.beyls, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D35681 llvm-svn: 309264
2017-07-27 19:03:45 +08:00
void RuleMatcher::emit(MatchTable &Table) {
[tablegen][globalisel] Trivial changes to reduce size of following patch. NFC. Summary: * Move namespace { * Trivial: Typo * RuleMatcher: Separate class and definition * Trivial: const findNodeEquiv Reviewers: t.p.northover, qcolombet, aditya_nandakumar, ab, rovka Reviewed By: rovka Subscribers: dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D30531 llvm-svn: 297884
2017-03-16 04:18:38 +08:00
if (Matchers.empty())
llvm_unreachable("Unexpected empty matcher!");
// The representation supports rules that require multiple roots such as:
// %ptr(p0) = ...
// %elt0(s32) = G_LOAD %ptr
// %1(p0) = G_ADD %ptr, 4
// %elt1(s32) = G_LOAD p0 %1
// which could be usefully folded into:
// %ptr(p0) = ...
// %elt0(s32), %elt1(s32) = TGT_LOAD_PAIR %ptr
// on some targets but we don't need to make use of that yet.
assert(Matchers.size() == 1 && "Cannot handle multi-root matchers yet");
[globalisel][tablegen] Import SelectionDAG's rule predicates and support the equivalent in GIRule. Summary: The SelectionDAG importer now imports rules with Predicate's attached via Requires, PredicateControl, etc. These predicates are implemented as bitset's to allow multiple predicates to be tested together. However, unlike the MC layer subtarget features, each target only pays for it's own predicates (e.g. AArch64 doesn't have 192 feature bits just because X86 needs a lot). Both AArch64 and X86 derive at least one predicate from the MachineFunction or Function so they must re-initialize AvailableFeatures before each function. They also declare locals in <Target>InstructionSelector so that computeAvailableFeatures() can use the code from SelectionDAG without modification. Reviewers: rovka, qcolombet, aditya_nandakumar, t.p.northover, ab Reviewed By: rovka Subscribers: aemerson, rengolin, dberris, kristof.beyls, llvm-commits, igorb Differential Revision: https://reviews.llvm.org/D31418 llvm-svn: 300993
2017-04-21 23:59:56 +08:00
Re-commit: r309094 [globalisel][tablegen] Fuse the generated tables together. Summary: Now that we have control flow in place, fuse the per-rule tables into a single table. This is a compile-time saving at this point. However, this will also enable the optimization of a table so that similar instructions can be tested together, reducing the time spent on the matching the code. This is NFC in terms of externally visible behaviour but some internals have changed slightly. State.MIs is no longer reset between each rule that is attempted because it's not necessary to do so. As a consequence of this the restriction on the order that instructions are added to State.MIs has been relaxed to only affect recorded instructions that require new elements to be added to the vector. GIM_RecordInsn can now write to any element from 1 to State.MIs.size() instead of just State.MIs.size(). The compile-time regressions from the last commit were caused by the ARM target including a non-const variable (zero_reg) in the table and therefore generating an initializer for it. That variable is now const. Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Reviewed By: rovka Subscribers: kristof.beyls, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D35681 llvm-svn: 309264
2017-07-27 19:03:45 +08:00
unsigned LabelID = Table.allocateLabelID();
[globalisel][tablegen] Add control-flow to the MatchTable. Summary: This will allow us to merge the various sub-tables into a single table. This is a compile-time saving at this point. However, this will also enable the optimization of a table so that similar instructions can be tested together, reducing the time spent on the matching the code. The bulk of this patch is a mechanical conversion to the new MatchTable object which is responsible for tracking label definitions and filling in the index of the jump targets. It is also responsible for nicely formatting the table. This was necessary to support the new GIM_Try opcode which takes the index to jump to if the match should fail. This value is unknown during table construction and is filled in during emission. To support nesting try-blocks (although we currently don't emit tables with nested try-blocks), GIM_Reject has been re-introduced to explicitly exit a try-block or fail the overall match if there are no active try-blocks. Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Reviewed By: rovka Subscribers: kristof.beyls, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D35117 llvm-svn: 308596
2017-07-20 17:25:44 +08:00
Table << MatchTable::Opcode("GIM_Try", +1)
Reapply r332907 "[GlobalISel] Improving InstructionSelect's performance by reducing MatchTable..." Apparently the compile time problem was caused by the fact that not all compilers / STL implementations can automatically convert std::unique_ptr<Derived> to std::unique_ptr<Base>. Fixed (hopefully) by making sure it's std::unique_ptr<Derived>&& (rvalue ref) to std::unique_ptr<Base> conversion instead. llvm-svn: 332917
2018-05-22 07:28:51 +08:00
<< MatchTable::Comment("On fail goto")
<< MatchTable::JumpTarget(LabelID)
<< MatchTable::Comment(("Rule ID " + Twine(RuleID) + " //").str())
[globalisel][tablegen] Add control-flow to the MatchTable. Summary: This will allow us to merge the various sub-tables into a single table. This is a compile-time saving at this point. However, this will also enable the optimization of a table so that similar instructions can be tested together, reducing the time spent on the matching the code. The bulk of this patch is a mechanical conversion to the new MatchTable object which is responsible for tracking label definitions and filling in the index of the jump targets. It is also responsible for nicely formatting the table. This was necessary to support the new GIM_Try opcode which takes the index to jump to if the match should fail. This value is unknown during table construction and is filled in during emission. To support nesting try-blocks (although we currently don't emit tables with nested try-blocks), GIM_Reject has been re-introduced to explicitly exit a try-block or fail the overall match if there are no active try-blocks. Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Reviewed By: rovka Subscribers: kristof.beyls, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D35117 llvm-svn: 308596
2017-07-20 17:25:44 +08:00
<< MatchTable::LineBreak;
[globalisel][tablegen] Import SelectionDAG's rule predicates and support the equivalent in GIRule. Summary: The SelectionDAG importer now imports rules with Predicate's attached via Requires, PredicateControl, etc. These predicates are implemented as bitset's to allow multiple predicates to be tested together. However, unlike the MC layer subtarget features, each target only pays for it's own predicates (e.g. AArch64 doesn't have 192 feature bits just because X86 needs a lot). Both AArch64 and X86 derive at least one predicate from the MachineFunction or Function so they must re-initialize AvailableFeatures before each function. They also declare locals in <Target>InstructionSelector so that computeAvailableFeatures() can use the code from SelectionDAG without modification. Reviewers: rovka, qcolombet, aditya_nandakumar, t.p.northover, ab Reviewed By: rovka Subscribers: aemerson, rengolin, dberris, kristof.beyls, llvm-commits, igorb Differential Revision: https://reviews.llvm.org/D31418 llvm-svn: 300993
2017-04-21 23:59:56 +08:00
if (!RequiredFeatures.empty()) {
[globalisel][tablegen] Add control-flow to the MatchTable. Summary: This will allow us to merge the various sub-tables into a single table. This is a compile-time saving at this point. However, this will also enable the optimization of a table so that similar instructions can be tested together, reducing the time spent on the matching the code. The bulk of this patch is a mechanical conversion to the new MatchTable object which is responsible for tracking label definitions and filling in the index of the jump targets. It is also responsible for nicely formatting the table. This was necessary to support the new GIM_Try opcode which takes the index to jump to if the match should fail. This value is unknown during table construction and is filled in during emission. To support nesting try-blocks (although we currently don't emit tables with nested try-blocks), GIM_Reject has been re-introduced to explicitly exit a try-block or fail the overall match if there are no active try-blocks. Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Reviewed By: rovka Subscribers: kristof.beyls, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D35117 llvm-svn: 308596
2017-07-20 17:25:44 +08:00
Table << MatchTable::Opcode("GIM_CheckFeatures")
<< MatchTable::NamedValue(getNameForFeatureBitset(RequiredFeatures))
<< MatchTable::LineBreak;
[globalisel][tablegen] Import SelectionDAG's rule predicates and support the equivalent in GIRule. Summary: The SelectionDAG importer now imports rules with Predicate's attached via Requires, PredicateControl, etc. These predicates are implemented as bitset's to allow multiple predicates to be tested together. However, unlike the MC layer subtarget features, each target only pays for it's own predicates (e.g. AArch64 doesn't have 192 feature bits just because X86 needs a lot). Both AArch64 and X86 derive at least one predicate from the MachineFunction or Function so they must re-initialize AvailableFeatures before each function. They also declare locals in <Target>InstructionSelector so that computeAvailableFeatures() can use the code from SelectionDAG without modification. Reviewers: rovka, qcolombet, aditya_nandakumar, t.p.northover, ab Reviewed By: rovka Subscribers: aemerson, rengolin, dberris, kristof.beyls, llvm-commits, igorb Differential Revision: https://reviews.llvm.org/D31418 llvm-svn: 300993
2017-04-21 23:59:56 +08:00
}
[tablegen][globalisel] Capture instructions into locals and related infrastructure for multiple instructions matches. Summary: Prepare the way for nested instruction matching support by having actions like CopyRenderer look up operands in the RuleMatcher rather than a specific InstructionMatcher. This allows actions to reference any operand from any matched instruction. It works by checking the 'shape' of the match and capturing each matched instruction to a local variable. If the shape is wrong (not enough operands, leaf nodes where non-leafs are expected, etc.), then the rule exits early without checking the predicates. Once we've captured the instructions, we then test the predicates as before (except using the local variables). If the match is successful, then we render the new instruction as before using the local variables. It's not noticable in this patch but by the time we support multiple instruction matching, this patch will also cause a significant improvement to readability of the emitted code since MRI.getVRegDef(I->getOperand(0).getReg()) will simply be MI1 after emitCxxCaptureStmts(). This isn't quite NFC because I've also fixed a bug that I'm surprised we haven't encountered yet. It now checks there are at least the expected number of operands before accessing them with getOperand(). Depends on D30531 Reviewers: t.p.northover, qcolombet, aditya_nandakumar, ab, rovka Reviewed By: rovka Subscribers: dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D30535 llvm-svn: 298257
2017-03-20 23:20:42 +08:00
[TableGen][GlobalISel] Have the predicate directly know which data they are dealing with Prior to this patch, a predicate wouldn't make sense outside of its rule. Indeed, it was only during emitting a rule that a predicate would be made aware of the IDs of the data it is checking. Because of that, predicates could not be moved around or compared between each other. NFC. llvm-svn: 320887
2017-12-16 07:07:42 +08:00
Matchers.front()->emitPredicateOpcodes(Table, *this);
[globalisel][tablegen] Partially fix compile-time regressions by converting matcher to state-machine(s) Summary: Replace the matcher if-statements for each rule with a state-machine. This significantly reduces compile time, memory allocations, and cumulative memory allocation when compiling AArch64InstructionSelector.cpp.o after r303259 is recommitted. The following patches will expand on this further to fully fix the regressions. Reviewers: rovka, ab, t.p.northover, qcolombet, aditya_nandakumar Reviewed By: ab Subscribers: vitalybuka, aemerson, javed.absar, igorb, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D33758 llvm-svn: 307079
2017-07-04 22:35:06 +08:00
[tablegen][globalisel] Add support for nested instruction matching. Summary: Lift the restrictions that prevented the tree walking introduced in the previous change and add support for patterns like: (G_ADD (G_MUL (G_SEXT $src1), (G_SEXT $src2)), $src3) -> SMADDWrrr $dst, $src1, $src2, $src3 Also adds support for G_SEXT and G_ZEXT to support these cases. One particular aspect of this that I should draw attention to is that I've tried to be overly conservative in determining the safety of matches that involve non-adjacent instructions and multiple basic blocks. This is intended to be used as a cheap initial check and we may add a more expensive check in the future. The current rules are: * Reject if any instruction may load/store (we'd need to check for intervening memory operations. * Reject if any instruction has implicit operands. * Reject if any instruction has unmodelled side-effects. See isObviouslySafeToFold(). Reviewers: t.p.northover, javed.absar, qcolombet, aditya_nandakumar, ab, rovka Reviewed By: ab Subscribers: igorb, dberris, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D30539 llvm-svn: 299430
2017-04-04 21:25:23 +08:00
// We must also check if it's safe to fold the matched instructions.
[globalisel][tablegen] Partially fix compile-time regressions by converting matcher to state-machine(s) Summary: Replace the matcher if-statements for each rule with a state-machine. This significantly reduces compile time, memory allocations, and cumulative memory allocation when compiling AArch64InstructionSelector.cpp.o after r303259 is recommitted. The following patches will expand on this further to fully fix the regressions. Reviewers: rovka, ab, t.p.northover, qcolombet, aditya_nandakumar Reviewed By: ab Subscribers: vitalybuka, aemerson, javed.absar, igorb, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D33758 llvm-svn: 307079
2017-07-04 22:35:06 +08:00
if (InsnVariableIDs.size() >= 2) {
Fixed nondeterminism in RuleMatcher::emit. llvm-svn: 303829
2017-05-25 09:51:53 +08:00
// Invert the map to create stable ordering (by var names)
[globalisel][tablegen] Partially fix compile-time regressions by converting matcher to state-machine(s) Summary: Replace the matcher if-statements for each rule with a state-machine. This significantly reduces compile time, memory allocations, and cumulative memory allocation when compiling AArch64InstructionSelector.cpp.o after r303259 is recommitted. The following patches will expand on this further to fully fix the regressions. Reviewers: rovka, ab, t.p.northover, qcolombet, aditya_nandakumar Reviewed By: ab Subscribers: vitalybuka, aemerson, javed.absar, igorb, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D33758 llvm-svn: 307079
2017-07-04 22:35:06 +08:00
SmallVector<unsigned, 2> InsnIDs;
for (const auto &Pair : InsnVariableIDs) {
[tablegen][globalisel] Add support for nested instruction matching. Summary: Lift the restrictions that prevented the tree walking introduced in the previous change and add support for patterns like: (G_ADD (G_MUL (G_SEXT $src1), (G_SEXT $src2)), $src3) -> SMADDWrrr $dst, $src1, $src2, $src3 Also adds support for G_SEXT and G_ZEXT to support these cases. One particular aspect of this that I should draw attention to is that I've tried to be overly conservative in determining the safety of matches that involve non-adjacent instructions and multiple basic blocks. This is intended to be used as a cheap initial check and we may add a more expensive check in the future. The current rules are: * Reject if any instruction may load/store (we'd need to check for intervening memory operations. * Reject if any instruction has implicit operands. * Reject if any instruction has unmodelled side-effects. See isObviouslySafeToFold(). Reviewers: t.p.northover, javed.absar, qcolombet, aditya_nandakumar, ab, rovka Reviewed By: ab Subscribers: igorb, dberris, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D30539 llvm-svn: 299430
2017-04-04 21:25:23 +08:00
// Skip the root node since it isn't moving anywhere. Everything else is
// sinking to meet it.
if (Pair.first == Matchers.front().get())
continue;
[globalisel][tablegen] Partially fix compile-time regressions by converting matcher to state-machine(s) Summary: Replace the matcher if-statements for each rule with a state-machine. This significantly reduces compile time, memory allocations, and cumulative memory allocation when compiling AArch64InstructionSelector.cpp.o after r303259 is recommitted. The following patches will expand on this further to fully fix the regressions. Reviewers: rovka, ab, t.p.northover, qcolombet, aditya_nandakumar Reviewed By: ab Subscribers: vitalybuka, aemerson, javed.absar, igorb, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D33758 llvm-svn: 307079
2017-07-04 22:35:06 +08:00
InsnIDs.push_back(Pair.second);
Fixed nondeterminism in RuleMatcher::emit. llvm-svn: 303829
2017-05-25 09:51:53 +08:00
}
[TableGen] Change std::sort to llvm::sort in response to r327219 Summary: r327219 added wrappers to std::sort which randomly shuffle the container before sorting. This will help in uncovering non-determinism caused due to undefined sorting order of objects having the same key. To make use of that infrastructure we need to invoke llvm::sort instead of std::sort. Note: This patch is one of a series of patches to replace *all* std::sort to llvm::sort. Refer the comments section in D44363 for a list of all the required patches. Reviewers: stoklund, kparzysz, dsanders Reviewed By: dsanders Subscribers: llvm-commits Differential Revision: https://reviews.llvm.org/D45144 llvm-svn: 329451
2018-04-07 04:18:05 +08:00
llvm::sort(InsnIDs.begin(), InsnIDs.end());
Fixed nondeterminism in RuleMatcher::emit. llvm-svn: 303829
2017-05-25 09:51:53 +08:00
[globalisel][tablegen] Partially fix compile-time regressions by converting matcher to state-machine(s) Summary: Replace the matcher if-statements for each rule with a state-machine. This significantly reduces compile time, memory allocations, and cumulative memory allocation when compiling AArch64InstructionSelector.cpp.o after r303259 is recommitted. The following patches will expand on this further to fully fix the regressions. Reviewers: rovka, ab, t.p.northover, qcolombet, aditya_nandakumar Reviewed By: ab Subscribers: vitalybuka, aemerson, javed.absar, igorb, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D33758 llvm-svn: 307079
2017-07-04 22:35:06 +08:00
for (const auto &InsnID : InsnIDs) {
[tablegen][globalisel] Add support for nested instruction matching. Summary: Lift the restrictions that prevented the tree walking introduced in the previous change and add support for patterns like: (G_ADD (G_MUL (G_SEXT $src1), (G_SEXT $src2)), $src3) -> SMADDWrrr $dst, $src1, $src2, $src3 Also adds support for G_SEXT and G_ZEXT to support these cases. One particular aspect of this that I should draw attention to is that I've tried to be overly conservative in determining the safety of matches that involve non-adjacent instructions and multiple basic blocks. This is intended to be used as a cheap initial check and we may add a more expensive check in the future. The current rules are: * Reject if any instruction may load/store (we'd need to check for intervening memory operations. * Reject if any instruction has implicit operands. * Reject if any instruction has unmodelled side-effects. See isObviouslySafeToFold(). Reviewers: t.p.northover, javed.absar, qcolombet, aditya_nandakumar, ab, rovka Reviewed By: ab Subscribers: igorb, dberris, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D30539 llvm-svn: 299430
2017-04-04 21:25:23 +08:00
// Reject the difficult cases until we have a more accurate check.
[globalisel][tablegen] Add control-flow to the MatchTable. Summary: This will allow us to merge the various sub-tables into a single table. This is a compile-time saving at this point. However, this will also enable the optimization of a table so that similar instructions can be tested together, reducing the time spent on the matching the code. The bulk of this patch is a mechanical conversion to the new MatchTable object which is responsible for tracking label definitions and filling in the index of the jump targets. It is also responsible for nicely formatting the table. This was necessary to support the new GIM_Try opcode which takes the index to jump to if the match should fail. This value is unknown during table construction and is filled in during emission. To support nesting try-blocks (although we currently don't emit tables with nested try-blocks), GIM_Reject has been re-introduced to explicitly exit a try-block or fail the overall match if there are no active try-blocks. Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Reviewed By: rovka Subscribers: kristof.beyls, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D35117 llvm-svn: 308596
2017-07-20 17:25:44 +08:00
Table << MatchTable::Opcode("GIM_CheckIsSafeToFold")
<< MatchTable::Comment("InsnID") << MatchTable::IntValue(InsnID)
<< MatchTable::LineBreak;
[tablegen][globalisel] Add support for nested instruction matching. Summary: Lift the restrictions that prevented the tree walking introduced in the previous change and add support for patterns like: (G_ADD (G_MUL (G_SEXT $src1), (G_SEXT $src2)), $src3) -> SMADDWrrr $dst, $src1, $src2, $src3 Also adds support for G_SEXT and G_ZEXT to support these cases. One particular aspect of this that I should draw attention to is that I've tried to be overly conservative in determining the safety of matches that involve non-adjacent instructions and multiple basic blocks. This is intended to be used as a cheap initial check and we may add a more expensive check in the future. The current rules are: * Reject if any instruction may load/store (we'd need to check for intervening memory operations. * Reject if any instruction has implicit operands. * Reject if any instruction has unmodelled side-effects. See isObviouslySafeToFold(). Reviewers: t.p.northover, javed.absar, qcolombet, aditya_nandakumar, ab, rovka Reviewed By: ab Subscribers: igorb, dberris, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D30539 llvm-svn: 299430
2017-04-04 21:25:23 +08:00
// FIXME: Emit checks to determine it's _actually_ safe to fold and/or
// account for unsafe cases.
//
// Example:
// MI1--> %0 = ...
// %1 = ... %0
// MI0--> %2 = ... %0
// It's not safe to erase MI1. We currently handle this by not
// erasing %0 (even when it's dead).
//
// Example:
// MI1--> %0 = load volatile @a
// %1 = load volatile @a
// MI0--> %2 = ... %0
// It's not safe to sink %0's def past %1. We currently handle
// this by rejecting all loads.
//
// Example:
// MI1--> %0 = load @a
// %1 = store @a
// MI0--> %2 = ... %0
// It's not safe to sink %0's def past %1. We currently handle
// this by rejecting all loads.
//
// Example:
// G_CONDBR %cond, @BB1
// BB0:
// MI1--> %0 = load @a
// G_BR @BB1
// BB1:
// MI0--> %2 = ... %0
// It's not always safe to sink %0 across control flow. In this
// case it may introduce a memory fault. We currentl handle this
// by rejecting all loads.
}
}
Reapply r332907 "[GlobalISel] Improving InstructionSelect's performance by reducing MatchTable..." Apparently the compile time problem was caused by the fact that not all compilers / STL implementations can automatically convert std::unique_ptr<Derived> to std::unique_ptr<Base>. Fixed (hopefully) by making sure it's std::unique_ptr<Derived>&& (rvalue ref) to std::unique_ptr<Base> conversion instead. llvm-svn: 332917
2018-05-22 07:28:51 +08:00
for (const auto &PM : EpilogueMatchers)
PM->emitPredicateOpcodes(Table, *this);
[globalisel][tablegen] Added instruction emission to the state-machine-based matcher. Summary: This further improves the compile-time regressions that will be caused by a re-commit of r303259. Also added included preliminary work in preparation for the multi-insn emitter since I needed to change the relevant part of the API for this patch anyway. Depends on D33758 Reviewers: rovka, vitalybuka, ab, t.p.northover, qcolombet, aditya_nandakumar Reviewed By: ab Subscribers: kristof.beyls, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D33764 llvm-svn: 307133
2017-07-05 17:39:33 +08:00
for (const auto &MA : Actions)
[globalisel][tablegen] Add infrastructure to potentially allow BuildMIAction to choose a mutatable instruction. NFC Prepare for multiple instruction emission by allowing BuildMIAction to search for a suitable matcher that will support mutation. This patch deliberately neglects to add matchers aside from the root to preserve NFC. That said, it should be noted that until we support mutations other than just the opcode the chances of finding a non-root instruction for which canMutate() is true, is essentially zero. Furthermore in the presence of multi-instruction emission the chances of finding any instruction for which canMutate() is true is also zero. Nevertheless, we can't continue to require that all BuildMIAction's consider the root of the match to be recyclable due to the risk of recycling it twice in the same rule. llvm-svn: 317022
2017-11-01 02:50:24 +08:00
MA->emitActionOpcodes(Table, *this);
[globalisel][tablegen] Generate rule coverage and use it to identify untested rules Summary: This patch adds a LLVM_ENABLE_GISEL_COV which, like LLVM_ENABLE_DAGISEL_COV, causes TableGen to instrument the generated table to collect rule coverage information. However, LLVM_ENABLE_GISEL_COV goes a bit further than LLVM_ENABLE_DAGISEL_COV. The information is written to files (${CMAKE_BINARY_DIR}/gisel-coverage-* by default). These files can then be concatenated into ${LLVM_GISEL_COV_PREFIX}-all after which TableGen will read this information and use it to emit warnings about untested rules. This technique could also be used by SelectionDAG and can be further extended to detect hot rules and give them priority over colder rules. Usage: * Enable LLVM_ENABLE_GISEL_COV in CMake * Build the compiler and run some tests * cat gisel-coverage-[0-9]* > gisel-coverage-all * Delete lib/Target/*/*GenGlobalISel.inc* * Build the compiler Known issues: * ${LLVM_GISEL_COV_PREFIX}-all must be generated as a manual step due to a lack of a portable 'cat' command. It should be the concatenation of all ${LLVM_GISEL_COV_PREFIX}-[0-9]* files. * There's no mechanism to discard coverage information when the ruleset changes Depends on D39742 Reviewers: ab, qcolombet, t.p.northover, aditya_nandakumar, rovka Reviewed By: rovka Subscribers: vsk, arsenm, nhaehnle, mgorny, kristof.beyls, javed.absar, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D39747 llvm-svn: 318356
2017-11-16 08:46:35 +08:00
[GlobalISel][InstructionSelect] Making Coverage Info generation optional on per-match table basis to make sure that Testgen always has access to coverage info even if the match table used by the selector itself is stripped off that information for performance reasons. Reviewers: dsanders, aemerson Reviewed By: dsanders Subscribers: rovka, kristof.beyls, llvm-commits, dsanders Differential Revision: https://reviews.llvm.org/D46098 llvm-svn: 331398
2018-05-03 04:15:11 +08:00
if (Table.isWithCoverage())
[globalisel][tablegen] Generate rule coverage and use it to identify untested rules Summary: This patch adds a LLVM_ENABLE_GISEL_COV which, like LLVM_ENABLE_DAGISEL_COV, causes TableGen to instrument the generated table to collect rule coverage information. However, LLVM_ENABLE_GISEL_COV goes a bit further than LLVM_ENABLE_DAGISEL_COV. The information is written to files (${CMAKE_BINARY_DIR}/gisel-coverage-* by default). These files can then be concatenated into ${LLVM_GISEL_COV_PREFIX}-all after which TableGen will read this information and use it to emit warnings about untested rules. This technique could also be used by SelectionDAG and can be further extended to detect hot rules and give them priority over colder rules. Usage: * Enable LLVM_ENABLE_GISEL_COV in CMake * Build the compiler and run some tests * cat gisel-coverage-[0-9]* > gisel-coverage-all * Delete lib/Target/*/*GenGlobalISel.inc* * Build the compiler Known issues: * ${LLVM_GISEL_COV_PREFIX}-all must be generated as a manual step due to a lack of a portable 'cat' command. It should be the concatenation of all ${LLVM_GISEL_COV_PREFIX}-[0-9]* files. * There's no mechanism to discard coverage information when the ruleset changes Depends on D39742 Reviewers: ab, qcolombet, t.p.northover, aditya_nandakumar, rovka Reviewed By: rovka Subscribers: vsk, arsenm, nhaehnle, mgorny, kristof.beyls, javed.absar, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D39747 llvm-svn: 318356
2017-11-16 08:46:35 +08:00
Table << MatchTable::Opcode("GIR_Coverage") << MatchTable::IntValue(RuleID)
<< MatchTable::LineBreak;
Reapply r332907 "[GlobalISel] Improving InstructionSelect's performance by reducing MatchTable..." Apparently the compile time problem was caused by the fact that not all compilers / STL implementations can automatically convert std::unique_ptr<Derived> to std::unique_ptr<Base>. Fixed (hopefully) by making sure it's std::unique_ptr<Derived>&& (rvalue ref) to std::unique_ptr<Base> conversion instead. llvm-svn: 332917
2018-05-22 07:28:51 +08:00
else
Table << MatchTable::Comment(("GIR_Coverage, " + Twine(RuleID) + ",").str())
<< MatchTable::LineBreak;
[globalisel][tablegen] Generate rule coverage and use it to identify untested rules Summary: This patch adds a LLVM_ENABLE_GISEL_COV which, like LLVM_ENABLE_DAGISEL_COV, causes TableGen to instrument the generated table to collect rule coverage information. However, LLVM_ENABLE_GISEL_COV goes a bit further than LLVM_ENABLE_DAGISEL_COV. The information is written to files (${CMAKE_BINARY_DIR}/gisel-coverage-* by default). These files can then be concatenated into ${LLVM_GISEL_COV_PREFIX}-all after which TableGen will read this information and use it to emit warnings about untested rules. This technique could also be used by SelectionDAG and can be further extended to detect hot rules and give them priority over colder rules. Usage: * Enable LLVM_ENABLE_GISEL_COV in CMake * Build the compiler and run some tests * cat gisel-coverage-[0-9]* > gisel-coverage-all * Delete lib/Target/*/*GenGlobalISel.inc* * Build the compiler Known issues: * ${LLVM_GISEL_COV_PREFIX}-all must be generated as a manual step due to a lack of a portable 'cat' command. It should be the concatenation of all ${LLVM_GISEL_COV_PREFIX}-[0-9]* files. * There's no mechanism to discard coverage information when the ruleset changes Depends on D39742 Reviewers: ab, qcolombet, t.p.northover, aditya_nandakumar, rovka Reviewed By: rovka Subscribers: vsk, arsenm, nhaehnle, mgorny, kristof.beyls, javed.absar, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D39747 llvm-svn: 318356
2017-11-16 08:46:35 +08:00
[globalisel][tablegen] Add control-flow to the MatchTable. Summary: This will allow us to merge the various sub-tables into a single table. This is a compile-time saving at this point. However, this will also enable the optimization of a table so that similar instructions can be tested together, reducing the time spent on the matching the code. The bulk of this patch is a mechanical conversion to the new MatchTable object which is responsible for tracking label definitions and filling in the index of the jump targets. It is also responsible for nicely formatting the table. This was necessary to support the new GIM_Try opcode which takes the index to jump to if the match should fail. This value is unknown during table construction and is filled in during emission. To support nesting try-blocks (although we currently don't emit tables with nested try-blocks), GIM_Reject has been re-introduced to explicitly exit a try-block or fail the overall match if there are no active try-blocks. Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Reviewed By: rovka Subscribers: kristof.beyls, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D35117 llvm-svn: 308596
2017-07-20 17:25:44 +08:00
Table << MatchTable::Opcode("GIR_Done", -1) << MatchTable::LineBreak
Re-commit: r309094 [globalisel][tablegen] Fuse the generated tables together. Summary: Now that we have control flow in place, fuse the per-rule tables into a single table. This is a compile-time saving at this point. However, this will also enable the optimization of a table so that similar instructions can be tested together, reducing the time spent on the matching the code. This is NFC in terms of externally visible behaviour but some internals have changed slightly. State.MIs is no longer reset between each rule that is attempted because it's not necessary to do so. As a consequence of this the restriction on the order that instructions are added to State.MIs has been relaxed to only affect recorded instructions that require new elements to be added to the vector. GIM_RecordInsn can now write to any element from 1 to State.MIs.size() instead of just State.MIs.size(). The compile-time regressions from the last commit were caused by the ARM target including a non-const variable (zero_reg) in the table and therefore generating an initializer for it. That variable is now const. Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Reviewed By: rovka Subscribers: kristof.beyls, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D35681 llvm-svn: 309264
2017-07-27 19:03:45 +08:00
<< MatchTable::Label(LabelID);
[GlobalISel][TableGen] Fix the statistics for emitted patters Collected statistics for the number of patterns emitted can be incorrect because rules can be grouped if OptimizeMatchTable is enabled. Increase the counter in RuleMatcher::emit(...) to avoid that. llvm-svn: 323391
2018-01-25 08:18:52 +08:00
++NumPatternEmitted;
[tablegen][globalisel] Trivial changes to reduce size of following patch. NFC. Summary: * Move namespace { * Trivial: Typo * RuleMatcher: Separate class and definition * Trivial: const findNodeEquiv Reviewers: t.p.northover, qcolombet, aditya_nandakumar, ab, rovka Reviewed By: rovka Subscribers: dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D30531 llvm-svn: 297884
2017-03-16 04:18:38 +08:00
}
[GlobalISel] Add basic Selector-emitter tblgen backend. This adds a basic tablegen backend that analyzes the SelectionDAG patterns to find simple ones that are eligible for GlobalISel-emission. That's similar to FastISel, with one notable difference: we're not fed ISD opcodes, so we need to map the SDNode operators to generic opcodes. That's done using GINodeEquiv in TargetGlobalISel.td. Otherwise, this is mostly boilerplate, and lots of filtering of any kind of "complicated" pattern. On AArch64, this is sufficient to match G_ADD up to s64 (to ADDWrr/ADDXrr) and G_BR (to B). Differential Revision: https://reviews.llvm.org/D26878 llvm-svn: 290284
2016-12-22 07:26:20 +08:00
[tablegen][globalisel] Trivial changes to reduce size of following patch. NFC. Summary: * Move namespace { * Trivial: Typo * RuleMatcher: Separate class and definition * Trivial: const findNodeEquiv Reviewers: t.p.northover, qcolombet, aditya_nandakumar, ab, rovka Reviewed By: rovka Subscribers: dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D30531 llvm-svn: 297884
2017-03-16 04:18:38 +08:00
bool RuleMatcher::isHigherPriorityThan(const RuleMatcher &B) const {
// Rules involving more match roots have higher priority.
if (Matchers.size() > B.Matchers.size())
return true;
if (Matchers.size() < B.Matchers.size())
return false;
[globalisel] Sort RuleMatchers by priority. Summary: This makes more important rules have priority over less important rules. For example, '%a = G_ADD $b:s64, $c:s64' has priority over '%a = G_ADD $b:s32, $c:s32'. Previously these rules were emitted in the correct order by chance. NFC in this patch but it is required to make the next patch work correctly. Depends on D29710 Reviewers: t.p.northover, ab, qcolombet, aditya_nandakumar, rovka Reviewed By: ab, rovka Subscribers: javed.absar, dberris, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D29711 llvm-svn: 296121
2017-02-24 21:58:11 +08:00
[tablegen][globalisel] Trivial changes to reduce size of following patch. NFC. Summary: * Move namespace { * Trivial: Typo * RuleMatcher: Separate class and definition * Trivial: const findNodeEquiv Reviewers: t.p.northover, qcolombet, aditya_nandakumar, ab, rovka Reviewed By: rovka Subscribers: dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D30531 llvm-svn: 297884
2017-03-16 04:18:38 +08:00
for (const auto &Matcher : zip(Matchers, B.Matchers)) {
if (std::get<0>(Matcher)->isHigherPriorityThan(*std::get<1>(Matcher)))
[globalisel] Sort RuleMatchers by priority. Summary: This makes more important rules have priority over less important rules. For example, '%a = G_ADD $b:s64, $c:s64' has priority over '%a = G_ADD $b:s32, $c:s32'. Previously these rules were emitted in the correct order by chance. NFC in this patch but it is required to make the next patch work correctly. Depends on D29710 Reviewers: t.p.northover, ab, qcolombet, aditya_nandakumar, rovka Reviewed By: ab, rovka Subscribers: javed.absar, dberris, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D29711 llvm-svn: 296121
2017-02-24 21:58:11 +08:00
return true;
[tablegen][globalisel] Trivial changes to reduce size of following patch. NFC. Summary: * Move namespace { * Trivial: Typo * RuleMatcher: Separate class and definition * Trivial: const findNodeEquiv Reviewers: t.p.northover, qcolombet, aditya_nandakumar, ab, rovka Reviewed By: rovka Subscribers: dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D30531 llvm-svn: 297884
2017-03-16 04:18:38 +08:00
if (std::get<1>(Matcher)->isHigherPriorityThan(*std::get<0>(Matcher)))
[globalisel] Sort RuleMatchers by priority. Summary: This makes more important rules have priority over less important rules. For example, '%a = G_ADD $b:s64, $c:s64' has priority over '%a = G_ADD $b:s32, $c:s32'. Previously these rules were emitted in the correct order by chance. NFC in this patch but it is required to make the next patch work correctly. Depends on D29710 Reviewers: t.p.northover, ab, qcolombet, aditya_nandakumar, rovka Reviewed By: ab, rovka Subscribers: javed.absar, dberris, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D29711 llvm-svn: 296121
2017-02-24 21:58:11 +08:00
return false;
[globalisel][tblgen] Add support for ComplexPatterns Summary: Adds a new kind of MachineOperand: MO_Placeholder. This operand must not appear in the MIR and only exists as a way of creating an 'uninitialized' operand until a matcher function overwrites it. Depends on D30046, D29712 Reviewers: t.p.northover, ab, rovka, aditya_nandakumar, javed.absar, qcolombet Reviewed By: qcolombet Subscribers: dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D30089 llvm-svn: 297782
2017-03-15 05:32:08 +08:00
}
[tablegen][globalisel] Trivial changes to reduce size of following patch. NFC. Summary: * Move namespace { * Trivial: Typo * RuleMatcher: Separate class and definition * Trivial: const findNodeEquiv Reviewers: t.p.northover, qcolombet, aditya_nandakumar, ab, rovka Reviewed By: rovka Subscribers: dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D30531 llvm-svn: 297884
2017-03-16 04:18:38 +08:00
return false;
Removed unnecessary semicolon. NFCI. llvm-svn: 297906
2017-03-16 06:50:47 +08:00
}
[GlobalISel] Add basic Selector-emitter tblgen backend. This adds a basic tablegen backend that analyzes the SelectionDAG patterns to find simple ones that are eligible for GlobalISel-emission. That's similar to FastISel, with one notable difference: we're not fed ISD opcodes, so we need to map the SDNode operators to generic opcodes. That's done using GINodeEquiv in TargetGlobalISel.td. Otherwise, this is mostly boilerplate, and lots of filtering of any kind of "complicated" pattern. On AArch64, this is sufficient to match G_ADD up to s64 (to ADDWrr/ADDXrr) and G_BR (to B). Differential Revision: https://reviews.llvm.org/D26878 llvm-svn: 290284
2016-12-22 07:26:20 +08:00
[globalisel][tablegen] Revise API for ComplexPattern operands to improve flexibility. Summary: Some targets need to be able to do more complex rendering than just adding an operand or two to an instruction. For example, it may need to insert an instruction to extract a subreg first, or it may need to perform an operation on the operand. In SelectionDAG, targets would create SDNode's to achieve the desired effect during the complex pattern predicate. This worked because SelectionDAG had a form of garbage collection that would take care of SDNode's that were created but not used due to a later predicate rejecting a match. This doesn't translate well to GlobalISel and the churn was wasteful. The API changes in this patch enable GlobalISel to accomplish the same thing without the waste. The API is now: InstructionSelector::OptionalComplexRendererFn selectArithImmed(MachineOperand &Root) const; where Root is the root of the match. The return value can be omitted to indicate that the predicate failed to match, or a function with the signature ComplexRendererFn can be returned. For example: return OptionalComplexRendererFn( [=](MachineInstrBuilder &MIB) { MIB.addImm(Immed).addImm(ShVal); }); adds two immediate operands to the rendered instruction. Immed and ShVal are captured from the predicate function. As an added bonus, this also reduces the amount of information we need to provide to GIComplexOperandMatcher. Depends on D31418 Reviewers: aditya_nandakumar, t.p.northover, qcolombet, rovka, ab, javed.absar Reviewed By: ab Subscribers: dberris, kristof.beyls, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D31761 llvm-svn: 301079
2017-04-22 23:11:04 +08:00
unsigned RuleMatcher::countRendererFns() const {
[tablegen][globalisel] Trivial changes to reduce size of following patch. NFC. Summary: * Move namespace { * Trivial: Typo * RuleMatcher: Separate class and definition * Trivial: const findNodeEquiv Reviewers: t.p.northover, qcolombet, aditya_nandakumar, ab, rovka Reviewed By: rovka Subscribers: dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D30531 llvm-svn: 297884
2017-03-16 04:18:38 +08:00
return std::accumulate(
Matchers.begin(), Matchers.end(), 0,
[](unsigned A, const std::unique_ptr<InstructionMatcher> &Matcher) {
[globalisel][tablegen] Revise API for ComplexPattern operands to improve flexibility. Summary: Some targets need to be able to do more complex rendering than just adding an operand or two to an instruction. For example, it may need to insert an instruction to extract a subreg first, or it may need to perform an operation on the operand. In SelectionDAG, targets would create SDNode's to achieve the desired effect during the complex pattern predicate. This worked because SelectionDAG had a form of garbage collection that would take care of SDNode's that were created but not used due to a later predicate rejecting a match. This doesn't translate well to GlobalISel and the churn was wasteful. The API changes in this patch enable GlobalISel to accomplish the same thing without the waste. The API is now: InstructionSelector::OptionalComplexRendererFn selectArithImmed(MachineOperand &Root) const; where Root is the root of the match. The return value can be omitted to indicate that the predicate failed to match, or a function with the signature ComplexRendererFn can be returned. For example: return OptionalComplexRendererFn( [=](MachineInstrBuilder &MIB) { MIB.addImm(Immed).addImm(ShVal); }); adds two immediate operands to the rendered instruction. Immed and ShVal are captured from the predicate function. As an added bonus, this also reduces the amount of information we need to provide to GIComplexOperandMatcher. Depends on D31418 Reviewers: aditya_nandakumar, t.p.northover, qcolombet, rovka, ab, javed.absar Reviewed By: ab Subscribers: dberris, kristof.beyls, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D31761 llvm-svn: 301079
2017-04-22 23:11:04 +08:00
return A + Matcher->countRendererFns();
[tablegen][globalisel] Trivial changes to reduce size of following patch. NFC. Summary: * Move namespace { * Trivial: Typo * RuleMatcher: Separate class and definition * Trivial: const findNodeEquiv Reviewers: t.p.northover, qcolombet, aditya_nandakumar, ab, rovka Reviewed By: rovka Subscribers: dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D30531 llvm-svn: 297884
2017-03-16 04:18:38 +08:00
});
}
[globalisel][tablegen] Add support for importing 'imm' operands. Summary: This patch enables the import of rules containing 'imm' operands that do not constrain the acceptable values using predicates. Support for ImmLeaf will arrive in a later patch. Depends on D35681 Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Reviewed By: rovka Subscribers: kristof.beyls, javed.absar, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D35833 llvm-svn: 310343
2017-08-08 18:44:31 +08:00
bool OperandPredicateMatcher::isHigherPriorityThan(
const OperandPredicateMatcher &B) const {
// Generally speaking, an instruction is more important than an Int or a
// LiteralInt because it can cover more nodes but theres an exception to
// this. G_CONSTANT's are less important than either of those two because they
// are more permissive.
Re-commit: [globalisel][tablegen] Support zero-instruction emission. Summary: Support the case where an operand of a pattern is also the whole of the result pattern. In this case the original result and all its uses must be replaced by the operand. However, register class restrictions can require a COPY. This patch handles both cases by always emitting the copy and leaving it for the register allocator to optimize. The previous commit failed on Windows machines due to a flaw in the sort predicate which allowed both A < B < C and B == C to be satisfied simultaneously. The cause of this was some sloppiness in the priority order of G_CONSTANT instructions compared to other instructions. These had equal priority because it makes no difference, however there were operands had higher priority than G_CONSTANT but lower priority than any other instruction. As a result, a priority order between G_CONSTANT and other instructions must be enforced to ensure the predicate defines a strict weak order. Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Subscribers: javed.absar, kristof.beyls, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D36084 llvm-svn: 311076
2017-08-17 17:26:14 +08:00
const InstructionOperandMatcher *AOM =
dyn_cast<InstructionOperandMatcher>(this);
const InstructionOperandMatcher *BOM =
dyn_cast<InstructionOperandMatcher>(&B);
bool AIsConstantInsn = AOM && AOM->getInsnMatcher().isConstantInstruction();
bool BIsConstantInsn = BOM && BOM->getInsnMatcher().isConstantInstruction();
if (AOM && BOM) {
// The relative priorities between a G_CONSTANT and any other instruction
// don't actually matter but this code is needed to ensure a strict weak
// ordering. This is particularly important on Windows where the rules will
// be incorrectly sorted without it.
if (AIsConstantInsn != BIsConstantInsn)
return AIsConstantInsn < BIsConstantInsn;
return false;
[globalisel][tablegen] Add support for importing 'imm' operands. Summary: This patch enables the import of rules containing 'imm' operands that do not constrain the acceptable values using predicates. Support for ImmLeaf will arrive in a later patch. Depends on D35681 Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Reviewed By: rovka Subscribers: kristof.beyls, javed.absar, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D35833 llvm-svn: 310343
2017-08-08 18:44:31 +08:00
}
Re-commit: [globalisel][tablegen] Support zero-instruction emission. Summary: Support the case where an operand of a pattern is also the whole of the result pattern. In this case the original result and all its uses must be replaced by the operand. However, register class restrictions can require a COPY. This patch handles both cases by always emitting the copy and leaving it for the register allocator to optimize. The previous commit failed on Windows machines due to a flaw in the sort predicate which allowed both A < B < C and B == C to be satisfied simultaneously. The cause of this was some sloppiness in the priority order of G_CONSTANT instructions compared to other instructions. These had equal priority because it makes no difference, however there were operands had higher priority than G_CONSTANT but lower priority than any other instruction. As a result, a priority order between G_CONSTANT and other instructions must be enforced to ensure the predicate defines a strict weak order. Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Subscribers: javed.absar, kristof.beyls, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D36084 llvm-svn: 311076
2017-08-17 17:26:14 +08:00
if (AOM && AIsConstantInsn && (B.Kind == OPM_Int || B.Kind == OPM_LiteralInt))
return false;
if (BOM && BIsConstantInsn && (Kind == OPM_Int || Kind == OPM_LiteralInt))
return true;
[globalisel][tablegen] Add support for importing 'imm' operands. Summary: This patch enables the import of rules containing 'imm' operands that do not constrain the acceptable values using predicates. Support for ImmLeaf will arrive in a later patch. Depends on D35681 Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Reviewed By: rovka Subscribers: kristof.beyls, javed.absar, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D35833 llvm-svn: 310343
2017-08-08 18:44:31 +08:00
return Kind < B.Kind;
[globalisel][tablegen] Remove unnecessary ; to satisfy ubuntu-gcc7.1-werror. llvm-svn: 310357
2017-08-08 21:21:26 +08:00
}
[globalisel][tablegen] Add support for importing 'imm' operands. Summary: This patch enables the import of rules containing 'imm' operands that do not constrain the acceptable values using predicates. Support for ImmLeaf will arrive in a later patch. Depends on D35681 Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Reviewed By: rovka Subscribers: kristof.beyls, javed.absar, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D35833 llvm-svn: 310343
2017-08-08 18:44:31 +08:00
[globalisel][tablegen] Simplify named operand/operator lookups and fix a wrong-code bug this revealed. Summary: Operand variable lookups are now performed by the RuleMatcher rather than searching the whole matcher hierarchy for a match. This revealed a wrong-code bug that currently affects ARM and X86 where patterns that use a variable more than once in the match pattern will be imported but won't check that the operands are identical. This can cause the tablegen-erated matcher to accept matches that should be rejected. Depends on D36569 Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Subscribers: aemerson, igorb, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D36618 llvm-svn: 315780
2017-10-14 08:31:58 +08:00
void SameOperandMatcher::emitPredicateOpcodes(MatchTable &Table,
[TableGen][GlobalISel] Have the predicate directly know which data they are dealing with Prior to this patch, a predicate wouldn't make sense outside of its rule. Indeed, it was only during emitting a rule that a predicate would be made aware of the IDs of the data it is checking. Because of that, predicates could not be moved around or compared between each other. NFC. llvm-svn: 320887
2017-12-16 07:07:42 +08:00
RuleMatcher &Rule) const {
[globalisel][tablegen] Fix small spelling nits. NFC ComplexRendererFn -> ComplexRendererFns Corrected a couple lingering references to tied operands that were missed. llvm-svn: 316237
2017-10-21 04:55:29 +08:00
const OperandMatcher &OtherOM = Rule.getOperandMatcher(MatchingName);
[globalisel][tablegen] Simplify named operand/operator lookups and fix a wrong-code bug this revealed. Summary: Operand variable lookups are now performed by the RuleMatcher rather than searching the whole matcher hierarchy for a match. This revealed a wrong-code bug that currently affects ARM and X86 where patterns that use a variable more than once in the match pattern will be imported but won't check that the operands are identical. This can cause the tablegen-erated matcher to accept matches that should be rejected. Depends on D36569 Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Subscribers: aemerson, igorb, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D36618 llvm-svn: 315780
2017-10-14 08:31:58 +08:00
unsigned OtherInsnVarID = Rule.getInsnVarID(OtherOM.getInstructionMatcher());
Reapply r332907 "[GlobalISel] Improving InstructionSelect's performance by reducing MatchTable..." Apparently the compile time problem was caused by the fact that not all compilers / STL implementations can automatically convert std::unique_ptr<Derived> to std::unique_ptr<Base>. Fixed (hopefully) by making sure it's std::unique_ptr<Derived>&& (rvalue ref) to std::unique_ptr<Base> conversion instead. llvm-svn: 332917
2018-05-22 07:28:51 +08:00
assert(OtherInsnVarID == OtherOM.getInstructionMatcher().getInsnVarID());
[globalisel][tablegen] Simplify named operand/operator lookups and fix a wrong-code bug this revealed. Summary: Operand variable lookups are now performed by the RuleMatcher rather than searching the whole matcher hierarchy for a match. This revealed a wrong-code bug that currently affects ARM and X86 where patterns that use a variable more than once in the match pattern will be imported but won't check that the operands are identical. This can cause the tablegen-erated matcher to accept matches that should be rejected. Depends on D36569 Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Subscribers: aemerson, igorb, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D36618 llvm-svn: 315780
2017-10-14 08:31:58 +08:00
Table << MatchTable::Opcode("GIM_CheckIsSameOperand")
<< MatchTable::Comment("MI") << MatchTable::IntValue(InsnVarID)
<< MatchTable::Comment("OpIdx") << MatchTable::IntValue(OpIdx)
<< MatchTable::Comment("OtherMI")
<< MatchTable::IntValue(OtherInsnVarID)
<< MatchTable::Comment("OtherOpIdx")
Reapply r332907 "[GlobalISel] Improving InstructionSelect's performance by reducing MatchTable..." Apparently the compile time problem was caused by the fact that not all compilers / STL implementations can automatically convert std::unique_ptr<Derived> to std::unique_ptr<Base>. Fixed (hopefully) by making sure it's std::unique_ptr<Derived>&& (rvalue ref) to std::unique_ptr<Base> conversion instead. llvm-svn: 332917
2018-05-22 07:28:51 +08:00
<< MatchTable::IntValue(OtherOM.getOpIdx())
[globalisel][tablegen] Simplify named operand/operator lookups and fix a wrong-code bug this revealed. Summary: Operand variable lookups are now performed by the RuleMatcher rather than searching the whole matcher hierarchy for a match. This revealed a wrong-code bug that currently affects ARM and X86 where patterns that use a variable more than once in the match pattern will be imported but won't check that the operands are identical. This can cause the tablegen-erated matcher to accept matches that should be rejected. Depends on D36569 Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Subscribers: aemerson, igorb, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D36618 llvm-svn: 315780
2017-10-14 08:31:58 +08:00
<< MatchTable::LineBreak;
}
[GlobalISel] Add basic Selector-emitter tblgen backend. This adds a basic tablegen backend that analyzes the SelectionDAG patterns to find simple ones that are eligible for GlobalISel-emission. That's similar to FastISel, with one notable difference: we're not fed ISD opcodes, so we need to map the SDNode operators to generic opcodes. That's done using GINodeEquiv in TargetGlobalISel.td. Otherwise, this is mostly boilerplate, and lots of filtering of any kind of "complicated" pattern. On AArch64, this is sufficient to match G_ADD up to s64 (to ADDWrr/ADDXrr) and G_BR (to B). Differential Revision: https://reviews.llvm.org/D26878 llvm-svn: 290284
2016-12-22 07:26:20 +08:00
//===- GlobalISelEmitter class --------------------------------------------===//
[GlobalISel] Return an Expected<RuleMatcher> for each SDAG pattern. NFC. Instead of emitting the matcher code directly, return the rule matcher and the skip reason as an Expected<RuleMatcher>. This will let us record all matchers and process them before emission. It's a somewhat unconventional use of Error, but it's nicer than, say, std::pair, because of the bool conversions. Differential Revision: https://reviews.llvm.org/D29743 llvm-svn: 294706
2017-02-10 12:00:17 +08:00
class GlobalISelEmitter {
public:
explicit GlobalISelEmitter(RecordKeeper &RK);
void run(raw_ostream &OS);
private:
const RecordKeeper &RK;
const CodeGenDAGPatterns CGP;
const CodeGenTarget &Target;
[globalisel][tablegen] Add support for EXTRACT_SUBREG. Summary: After this patch, we finally have test cases that require multiple instruction emission. Depends on D33590 Reviewers: ab, qcolombet, t.p.northover, rovka, kristof.beyls Subscribers: javed.absar, llvm-commits, igorb Differential Revision: https://reviews.llvm.org/D33596 llvm-svn: 306388
2017-06-27 18:11:39 +08:00
CodeGenRegBank CGRegs;
[GlobalISel] Return an Expected<RuleMatcher> for each SDAG pattern. NFC. Instead of emitting the matcher code directly, return the rule matcher and the skip reason as an Expected<RuleMatcher>. This will let us record all matchers and process them before emission. It's a somewhat unconventional use of Error, but it's nicer than, say, std::pair, because of the bool conversions. Differential Revision: https://reviews.llvm.org/D29743 llvm-svn: 294706
2017-02-10 12:00:17 +08:00
[globalisel][tablegen] Map ld and st to G_LOAD and G_STORE. NFC Summary: There is an important mismatch between ISD::LOAD and G_LOAD (and likewise for ISD::STORE and G_STORE). In SelectionDAG, ISD::LOAD is a non-atomic load and atomic loads are handled by a separate node. However, this is not true of GlobalISel's G_LOAD. For G_LOAD, the MachineMemOperand indicates the atomicity of the operation. As a result, this mapping must also add a predicate that checks for non-atomic MachineMemOperands. This is NFC since these nodes always have predicates in practice and are therefore always rejected at the moment. Depends on D37443 Reviewers: ab, qcolombet, t.p.northover, rovka, aditya_nandakumar Reviewed By: qcolombet Subscribers: kristof.beyls, llvm-commits, igorb Differential Revision: https://reviews.llvm.org/D37445 llvm-svn: 315843
2017-10-15 10:41:12 +08:00
/// Keep track of the equivalence between SDNodes and Instruction by mapping
Revert r319691: [globalisel][tablegen] Split atomic load/store into separate opcode and enable for AArch64. Some concerns were raised with the direction. Revert while we discuss it and look into an alternative llvm-svn: 319739
2017-12-05 13:52:07 +08:00
/// SDNodes to the GINodeEquiv mapping. We need to map to the GINodeEquiv to
/// check for attributes on the relation such as CheckMMOIsNonAtomic.
[GlobalISel] Return an Expected<RuleMatcher> for each SDAG pattern. NFC. Instead of emitting the matcher code directly, return the rule matcher and the skip reason as an Expected<RuleMatcher>. This will let us record all matchers and process them before emission. It's a somewhat unconventional use of Error, but it's nicer than, say, std::pair, because of the bool conversions. Differential Revision: https://reviews.llvm.org/D29743 llvm-svn: 294706
2017-02-10 12:00:17 +08:00
/// This is defined using 'GINodeEquiv' in the target description.
[globalisel][tablegen] Map ld and st to G_LOAD and G_STORE. NFC Summary: There is an important mismatch between ISD::LOAD and G_LOAD (and likewise for ISD::STORE and G_STORE). In SelectionDAG, ISD::LOAD is a non-atomic load and atomic loads are handled by a separate node. However, this is not true of GlobalISel's G_LOAD. For G_LOAD, the MachineMemOperand indicates the atomicity of the operation. As a result, this mapping must also add a predicate that checks for non-atomic MachineMemOperands. This is NFC since these nodes always have predicates in practice and are therefore always rejected at the moment. Depends on D37443 Reviewers: ab, qcolombet, t.p.northover, rovka, aditya_nandakumar Reviewed By: qcolombet Subscribers: kristof.beyls, llvm-commits, igorb Differential Revision: https://reviews.llvm.org/D37445 llvm-svn: 315843
2017-10-15 10:41:12 +08:00
DenseMap<Record *, Record *> NodeEquivs;
[GlobalISel] Return an Expected<RuleMatcher> for each SDAG pattern. NFC. Instead of emitting the matcher code directly, return the rule matcher and the skip reason as an Expected<RuleMatcher>. This will let us record all matchers and process them before emission. It's a somewhat unconventional use of Error, but it's nicer than, say, std::pair, because of the bool conversions. Differential Revision: https://reviews.llvm.org/D29743 llvm-svn: 294706
2017-02-10 12:00:17 +08:00
[globalisel][tblgen] Add support for ComplexPatterns Summary: Adds a new kind of MachineOperand: MO_Placeholder. This operand must not appear in the MIR and only exists as a way of creating an 'uninitialized' operand until a matcher function overwrites it. Depends on D30046, D29712 Reviewers: t.p.northover, ab, rovka, aditya_nandakumar, javed.absar, qcolombet Reviewed By: qcolombet Subscribers: dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D30089 llvm-svn: 297782
2017-03-15 05:32:08 +08:00
/// Keep track of the equivalence between ComplexPattern's and
/// GIComplexOperandMatcher. Map entries are specified by subclassing
/// GIComplexPatternEquiv.
DenseMap<const Record *, const Record *> ComplexPatternEquivs;
[GlobalISel][TableGen] Add support for SDNodeXForm Summary: This patch adds CustomRenderer which renders the matched operands to the specified instruction. Targets can enable the matching of SDNodeXForm by adding a definition that inherits from GICustomOperandRenderer and GISDNodeXFormEquiv as follows. def gi_imm8 : GICustomOperandRenderer<"renderImm8”>, GISDNodeXFormEquiv<imm8_xform>; Custom renderer functions should be of the form: void render(MachineInstrBuilder &MIB, const MachineInstr &I); Reviewers: dsanders, ab, rovka Reviewed By: dsanders Subscribers: kristof.beyls, javed.absar, llvm-commits, mgrang, qcolombet Differential Revision: https://reviews.llvm.org/D42012 llvm-svn: 322582
2018-01-17 02:44:05 +08:00
/// Keep track of the equivalence between SDNodeXForm's and
/// GICustomOperandRenderer. Map entries are specified by subclassing
/// GISDNodeXFormEquiv.
DenseMap<const Record *, const Record *> SDNodeXFormEquivs;
[GISel]: Make GlobalISelEmitter rule prioritization compatible with selectionDAG This patch changes GlobalISelEmitter to rank patterns similar to how the DAG does it (ie it computes a score for a pattern and adds the added complexity to it). This is so that the decision tree for GISelSelector remains compatible with that of SelectionDAG. https://reviews.llvm.org/D43270 llvm-svn: 325401
2018-02-17 06:37:15 +08:00
/// Keep track of Scores of PatternsToMatch similar to how the DAG does.
/// This adds compatibility for RuleMatchers to use this for ordering rules.
DenseMap<uint64_t, int> RuleMatcherScores;
[globalisel][tablegen] Import SelectionDAG's rule predicates and support the equivalent in GIRule. Summary: The SelectionDAG importer now imports rules with Predicate's attached via Requires, PredicateControl, etc. These predicates are implemented as bitset's to allow multiple predicates to be tested together. However, unlike the MC layer subtarget features, each target only pays for it's own predicates (e.g. AArch64 doesn't have 192 feature bits just because X86 needs a lot). Both AArch64 and X86 derive at least one predicate from the MachineFunction or Function so they must re-initialize AvailableFeatures before each function. They also declare locals in <Target>InstructionSelector so that computeAvailableFeatures() can use the code from SelectionDAG without modification. Reviewers: rovka, qcolombet, aditya_nandakumar, t.p.northover, ab Reviewed By: rovka Subscribers: aemerson, rengolin, dberris, kristof.beyls, llvm-commits, igorb Differential Revision: https://reviews.llvm.org/D31418 llvm-svn: 300993
2017-04-21 23:59:56 +08:00
// Map of predicates to their subtarget features.
[globalisel][tablegen] Compute available feature bits correctly. Summary: Predicate<> now has a field to indicate how often it must be recomputed. Currently, there are two frequencies, per-module (RecomputePerFunction==0) and per-function (RecomputePerFunction==1). Per-function predicates are currently recomputed more frequently than necessary since the only predicate in this category is cheap to test. Per-module predicates are now computed in getSubtargetImpl() while per-function predicates are computed in selectImpl(). Tablegen now manages the PredicateBitset internally. It should only be necessary to add the required includes. Also fixed a problem revealed by the test case where constrainSelectedInstRegOperands() would attempt to tie operands that BuildMI had already tied. Reviewers: ab, qcolombet, t.p.northover, rovka, aditya_nandakumar Reviewed By: rovka Subscribers: kristof.beyls, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D32491 llvm-svn: 301750
2017-04-30 01:30:09 +08:00
SubtargetFeatureInfoMap SubtargetFeatures;
[globalisel][tablegen] Import SelectionDAG's rule predicates and support the equivalent in GIRule. Summary: The SelectionDAG importer now imports rules with Predicate's attached via Requires, PredicateControl, etc. These predicates are implemented as bitset's to allow multiple predicates to be tested together. However, unlike the MC layer subtarget features, each target only pays for it's own predicates (e.g. AArch64 doesn't have 192 feature bits just because X86 needs a lot). Both AArch64 and X86 derive at least one predicate from the MachineFunction or Function so they must re-initialize AvailableFeatures before each function. They also declare locals in <Target>InstructionSelector so that computeAvailableFeatures() can use the code from SelectionDAG without modification. Reviewers: rovka, qcolombet, aditya_nandakumar, t.p.northover, ab Reviewed By: rovka Subscribers: aemerson, rengolin, dberris, kristof.beyls, llvm-commits, igorb Differential Revision: https://reviews.llvm.org/D31418 llvm-svn: 300993
2017-04-21 23:59:56 +08:00
[globalisel][tablegen] Generate rule coverage and use it to identify untested rules Summary: This patch adds a LLVM_ENABLE_GISEL_COV which, like LLVM_ENABLE_DAGISEL_COV, causes TableGen to instrument the generated table to collect rule coverage information. However, LLVM_ENABLE_GISEL_COV goes a bit further than LLVM_ENABLE_DAGISEL_COV. The information is written to files (${CMAKE_BINARY_DIR}/gisel-coverage-* by default). These files can then be concatenated into ${LLVM_GISEL_COV_PREFIX}-all after which TableGen will read this information and use it to emit warnings about untested rules. This technique could also be used by SelectionDAG and can be further extended to detect hot rules and give them priority over colder rules. Usage: * Enable LLVM_ENABLE_GISEL_COV in CMake * Build the compiler and run some tests * cat gisel-coverage-[0-9]* > gisel-coverage-all * Delete lib/Target/*/*GenGlobalISel.inc* * Build the compiler Known issues: * ${LLVM_GISEL_COV_PREFIX}-all must be generated as a manual step due to a lack of a portable 'cat' command. It should be the concatenation of all ${LLVM_GISEL_COV_PREFIX}-[0-9]* files. * There's no mechanism to discard coverage information when the ruleset changes Depends on D39742 Reviewers: ab, qcolombet, t.p.northover, aditya_nandakumar, rovka Reviewed By: rovka Subscribers: vsk, arsenm, nhaehnle, mgorny, kristof.beyls, javed.absar, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D39747 llvm-svn: 318356
2017-11-16 08:46:35 +08:00
// Rule coverage information.
Optional<CodeGenCoverage> RuleCoverage;
Reapply r332907 "[GlobalISel] Improving InstructionSelect's performance by reducing MatchTable..." Apparently the compile time problem was caused by the fact that not all compilers / STL implementations can automatically convert std::unique_ptr<Derived> to std::unique_ptr<Base>. Fixed (hopefully) by making sure it's std::unique_ptr<Derived>&& (rvalue ref) to std::unique_ptr<Base> conversion instead. llvm-svn: 332917
2018-05-22 07:28:51 +08:00
void gatherOpcodeValues();
void gatherTypeIDValues();
[GlobalISel] Return an Expected<RuleMatcher> for each SDAG pattern. NFC. Instead of emitting the matcher code directly, return the rule matcher and the skip reason as an Expected<RuleMatcher>. This will let us record all matchers and process them before emission. It's a somewhat unconventional use of Error, but it's nicer than, say, std::pair, because of the bool conversions. Differential Revision: https://reviews.llvm.org/D29743 llvm-svn: 294706
2017-02-10 12:00:17 +08:00
void gatherNodeEquivs();
[globalisel][tablegen] Map ld and st to G_LOAD and G_STORE. NFC Summary: There is an important mismatch between ISD::LOAD and G_LOAD (and likewise for ISD::STORE and G_STORE). In SelectionDAG, ISD::LOAD is a non-atomic load and atomic loads are handled by a separate node. However, this is not true of GlobalISel's G_LOAD. For G_LOAD, the MachineMemOperand indicates the atomicity of the operation. As a result, this mapping must also add a predicate that checks for non-atomic MachineMemOperands. This is NFC since these nodes always have predicates in practice and are therefore always rejected at the moment. Depends on D37443 Reviewers: ab, qcolombet, t.p.northover, rovka, aditya_nandakumar Reviewed By: qcolombet Subscribers: kristof.beyls, llvm-commits, igorb Differential Revision: https://reviews.llvm.org/D37445 llvm-svn: 315843
2017-10-15 10:41:12 +08:00
Record *findNodeEquiv(Record *N) const;
[globalisel] Update GlobalISel emitter to match new representation of extending loads Summary: Previously, a extending load was represented at (G_*EXT (G_LOAD x)). This had a few drawbacks: * G_LOAD had to be legal for all sizes you could extend from, even if registers didn't naturally hold those sizes. * All sizes you could extend from had to be allocatable just in case the extend went missing (e.g. by optimization). * At minimum, G_*EXT and G_TRUNC had to be legal for these sizes. As we improve optimization of extends and truncates, this legality requirement would spread without considerable care w.r.t when certain combines were permitted. * The SelectionDAG importer required some ugly and fragile pattern rewriting to translate patterns into this style. This patch changes the representation to: * (G_[SZ]EXTLOAD x) * (G_LOAD x) any-extends when MMO.getSize() * 8 < ResultTy.getSizeInBits() which resolves these issues by allowing targets to work entirely in their native register sizes, and by having a more direct translation from SelectionDAG patterns. Each extending load can be lowered by the legalizer into separate extends and loads, however a target that supports s1 will need the any-extending load to extend to at least s8 since LLVM does not represent memory accesses smaller than 8 bit. The legalizer can widenScalar G_LOAD into an any-extending load but sign/zero-extending loads need help from something else like a combiner pass. A follow-up patch that adds combiner helpers for for this will follow. The new representation requires that the MMO correctly reflect the memory access so this has been corrected in a couple tests. I've also moved the extending loads to their own tests since they are (mostly) separate opcodes now. Additionally, the re-write appears to have invalidated two tests from select-with-no-legality-check.mir since the matcher table no longer contains loads that result in s1's and they aren't legal in AArch64 anymore. Depends on D45540 Reviewers: ab, aditya_nandakumar, bogner, rtereshin, volkan, rovka, javed.absar Reviewed By: rtereshin Subscribers: javed.absar, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D45541 llvm-svn: 331601
2018-05-06 04:53:24 +08:00
const CodeGenInstruction *getEquivNode(Record &Equiv,
const TreePatternNode *N) const;
[GlobalISel] Return an Expected<RuleMatcher> for each SDAG pattern. NFC. Instead of emitting the matcher code directly, return the rule matcher and the skip reason as an Expected<RuleMatcher>. This will let us record all matchers and process them before emission. It's a somewhat unconventional use of Error, but it's nicer than, say, std::pair, because of the bool conversions. Differential Revision: https://reviews.llvm.org/D29743 llvm-svn: 294706
2017-02-10 12:00:17 +08:00
TableGen support for parameterized register class information This replaces TableGen's type inference to operate on parameterized types instead of MVTs, and as a consequence, some interfaces have changed: - Uses of MVTs are replaced by ValueTypeByHwMode. - EEVT::TypeSet is replaced by TypeSetByHwMode. This affects the way that types and type sets are printed, and the tests relying on that have been updated. There are certain users of the inferred types outside of TableGen itself, namely FastISel and GlobalISel. For those users, the way that the types are accessed have changed. For typical scenarios, these replacements can be used: - TreePatternNode::getType(ResNo) -> getSimpleType(ResNo) - TreePatternNode::hasTypeSet(ResNo) -> hasConcreteType(ResNo) - TypeSet::isConcrete -> TypeSetByHwMode::isValueTypeByHwMode(false) For more information, please refer to the review page. Differential Revision: https://reviews.llvm.org/D31951 llvm-svn: 313271
2017-09-15 00:56:21 +08:00
Error importRulePredicates(RuleMatcher &M, ArrayRef<Predicate> Predicates);
Re-commit r315863: [globalisel][tablegen] Import ComplexPattern when used as an operator Summary: It's possible for a ComplexPattern to be used as an operator in a match pattern. This is used by the load/store patterns in AArch64 to name the suboperands returned by ComplexPattern predicate so that they can be broken apart and referenced independently in the result pattern. This patch adds support for this in order to enable the import of load/store patterns. Depends on D37445 Hopefully fixed the ambiguous constructor that a large number of bots reported. Reviewers: ab, qcolombet, t.p.northover, rovka, aditya_nandakumar Reviewed By: qcolombet Subscribers: aemerson, javed.absar, igorb, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D37456 llvm-svn: 315869
2017-10-16 02:22:54 +08:00
Expected<InstructionMatcher &> createAndImportSelDAGMatcher(
RuleMatcher &Rule, InstructionMatcher &InsnMatcher,
const TreePatternNode *Src, unsigned &TempOpIdx) const;
Error importComplexPatternOperandMatcher(OperandMatcher &OM, Record *R,
unsigned &TempOpIdx) const;
Error importChildMatcher(RuleMatcher &Rule, InstructionMatcher &InsnMatcher,
[globalisel][tablegen] Implement unindexed load, non-extending load, and MemVT checks Summary: This includes some context-sensitivity in the MVT to LLT conversion so that pointer types are tested correctly. FIXME: I'm not happy with the way this is done since everything is a special-case. I've yet to find a reasonable way to implement it. select-load.mir fails because <1 x s64> loads in tablegen get priority over s64 loads. This is fixed in the next patch and as such they should be committed together, I've posted them separately to help with the review. Depends on D37456 Reviewers: ab, qcolombet, t.p.northover, rovka, aditya_nandakumar Subscribers: kristof.beyls, javed.absar, llvm-commits, igorb Differential Revision: https://reviews.llvm.org/D37457 llvm-svn: 315884
2017-10-16 08:56:30 +08:00
const TreePatternNode *SrcChild,
bool OperandIsAPointer, unsigned OpIdx,
[globalisel][tablegen] Change Expected<bool> to Error and rename functions. Functions that still return Expected<X> are now called createAndImport*() Changing the return type was requested in the review comments for r299001 llvm-svn: 299063
2017-03-30 17:36:33 +08:00
unsigned &TempOpIdx) const;
[globalisel][tablegen] Factor out implicit def/use renderers from createAndImportInstructionRenderer(). NFC Multi-instruction emission will require that we have separate handling for the defs between the implicitly created temporaries and the rule outputs. The former require new temporary vregs while the latter should copy existing operands. Factor out the implicit def/use renderers to minimize the code duplication when we implement that. llvm-svn: 317025
2017-11-01 03:09:29 +08:00
[globalisel][tablegen] Add support for EXTRACT_SUBREG. Summary: After this patch, we finally have test cases that require multiple instruction emission. Depends on D33590 Reviewers: ab, qcolombet, t.p.northover, rovka, kristof.beyls Subscribers: javed.absar, llvm-commits, igorb Differential Revision: https://reviews.llvm.org/D33596 llvm-svn: 306388
2017-06-27 18:11:39 +08:00
Expected<BuildMIAction &>
[globalisel][tablegen] Add infrastructure to potentially allow BuildMIAction to choose a mutatable instruction. NFC Prepare for multiple instruction emission by allowing BuildMIAction to search for a suitable matcher that will support mutation. This patch deliberately neglects to add matchers aside from the root to preserve NFC. That said, it should be noted that until we support mutations other than just the opcode the chances of finding a non-root instruction for which canMutate() is true, is essentially zero. Furthermore in the presence of multi-instruction emission the chances of finding any instruction for which canMutate() is true is also zero. Nevertheless, we can't continue to require that all BuildMIAction's consider the root of the match to be recyclable due to the risk of recycling it twice in the same rule. llvm-svn: 317022
2017-11-01 02:50:24 +08:00
createAndImportInstructionRenderer(RuleMatcher &M,
const TreePatternNode *Dst);
[globalisel][tablegen] Add support for multi-insn emission The importer will now accept nested instructions in the result pattern such as (ADDWrr $a, (SUBWrr $b, $c)). This is only valid when the nested instruction def's a single vreg and the parent instruction consumes a single vreg where a nested instruction is specified. The importer will automatically create a vreg to connect the two using the type information from the pattern. This vreg will be constrained to the register classes given in the instruction definitions*. * REG_SEQUENCE is explicitly rejected because of this. The definition doesn't constrain to a register class and it therefore needs special handling. llvm-svn: 317117
2017-11-02 03:57:57 +08:00
Expected<action_iterator> createAndImportSubInstructionRenderer(
action_iterator InsertPt, RuleMatcher &M, const TreePatternNode *Dst,
unsigned TempReg);
Re-commit: [globalisel][tablegen] Keep track of the insertion point while adding BuildMIAction's. NFC Multi-instruction emission needs to ensure the the instructions are generated a depth-first fashion. For example: (ADDWrr (SUBWrr a, b), c) needs to emit the SUBWrr before the ADDWrr. However, our walk over TreePatternNode's is highly context sensitive which makes it difficult to append BuildMIActions in the order we want. To fix this, we now keep track of the insertion point as we add actions. This will allow multi-insn emission to insert BuildMI's in the correct place. The previous commit failed on the Ubuntu bots using GCC 4.8. These bots lack the const_iterator forms of insert() and emplace() that were added in C++11. As a result I've switched the const_iterators to iterators. llvm-svn: 317049
2017-11-01 07:03:18 +08:00
Expected<action_iterator>
createInstructionRenderer(action_iterator InsertPt, RuleMatcher &M,
const TreePatternNode *Dst);
[globalisel][tablegen] Factor out implicit def/use renderers from createAndImportInstructionRenderer(). NFC Multi-instruction emission will require that we have separate handling for the defs between the implicitly created temporaries and the rule outputs. The former require new temporary vregs while the latter should copy existing operands. Factor out the implicit def/use renderers to minimize the code duplication when we implement that. llvm-svn: 317025
2017-11-01 03:09:29 +08:00
void importExplicitDefRenderers(BuildMIAction &DstMIBuilder);
Re-commit: [globalisel][tablegen] Keep track of the insertion point while adding BuildMIAction's. NFC Multi-instruction emission needs to ensure the the instructions are generated a depth-first fashion. For example: (ADDWrr (SUBWrr a, b), c) needs to emit the SUBWrr before the ADDWrr. However, our walk over TreePatternNode's is highly context sensitive which makes it difficult to append BuildMIActions in the order we want. To fix this, we now keep track of the insertion point as we add actions. This will allow multi-insn emission to insert BuildMI's in the correct place. The previous commit failed on the Ubuntu bots using GCC 4.8. These bots lack the const_iterator forms of insert() and emplace() that were added in C++11. As a result I've switched the const_iterators to iterators. llvm-svn: 317049
2017-11-01 07:03:18 +08:00
Expected<action_iterator>
importExplicitUseRenderers(action_iterator InsertPt, RuleMatcher &M,
BuildMIAction &DstMIBuilder,
[globalisel][tablegen] Factor out implicit def/use renderers from createAndImportInstructionRenderer(). NFC Multi-instruction emission will require that we have separate handling for the defs between the implicitly created temporaries and the rule outputs. The former require new temporary vregs while the latter should copy existing operands. Factor out the implicit def/use renderers to minimize the code duplication when we implement that. llvm-svn: 317025
2017-11-01 03:09:29 +08:00
const llvm::TreePatternNode *Dst);
Re-commit: [globalisel][tablegen] Keep track of the insertion point while adding BuildMIAction's. NFC Multi-instruction emission needs to ensure the the instructions are generated a depth-first fashion. For example: (ADDWrr (SUBWrr a, b), c) needs to emit the SUBWrr before the ADDWrr. However, our walk over TreePatternNode's is highly context sensitive which makes it difficult to append BuildMIActions in the order we want. To fix this, we now keep track of the insertion point as we add actions. This will allow multi-insn emission to insert BuildMI's in the correct place. The previous commit failed on the Ubuntu bots using GCC 4.8. These bots lack the const_iterator forms of insert() and emplace() that were added in C++11. As a result I've switched the const_iterators to iterators. llvm-svn: 317049
2017-11-01 07:03:18 +08:00
Expected<action_iterator>
importExplicitUseRenderer(action_iterator InsertPt, RuleMatcher &Rule,
BuildMIAction &DstMIBuilder,
[globalisel][tablegen] Add support for multi-insn emission The importer will now accept nested instructions in the result pattern such as (ADDWrr $a, (SUBWrr $b, $c)). This is only valid when the nested instruction def's a single vreg and the parent instruction consumes a single vreg where a nested instruction is specified. The importer will automatically create a vreg to connect the two using the type information from the pattern. This vreg will be constrained to the register classes given in the instruction definitions*. * REG_SEQUENCE is explicitly rejected because of this. The definition doesn't constrain to a register class and it therefore needs special handling. llvm-svn: 317117
2017-11-02 03:57:57 +08:00
TreePatternNode *DstChild);
[GlobalISel][TableGen] Fix handling of default operands When looping through a destination pattern's operands to decide how many default operands we need to introduce, we used to count the "expanded" number of operands. So if one default operand would be rendered as 2 values, we'd count it as 2 operands, when in fact it needs to count as only 1 operand regardless of how many values it expands to. This turns out to be a problem only in some very specific cases, e.g. when we have one operand with multiple default values followed by more operands with default values (see the new test). In such a situation we'd stop looping before looking at all the operands, and then error out assuming that we don't have enough default operands to make up the shortfall. At the moment this only affects ARM. The patch removes the loop counting default operands entirely and assumes that we'll have to introduce values for any default operand that we find (i.e. we're assuming it cannot be given as a child at all). It also extracts the code for adding renderers for default operands into a helper method. Differential Revision: https://reviews.llvm.org/D33031 llvm-svn: 303240
2017-05-17 16:57:28 +08:00
Error importDefaultOperandRenderers(BuildMIAction &DstMIBuilder,
DagInit *DefaultOps) const;
[globalisel][tablegen] Change Expected<bool> to Error and rename functions. Functions that still return Expected<X> are now called createAndImport*() Changing the return type was requested in the review comments for r299001 llvm-svn: 299063
2017-03-30 17:36:33 +08:00
Error
[tablegen][globalisel] Convert the SelectionDAG importer to a tree walking approach. NFC Summary: But don't actually inspect the tree any deeper than we already do. This change is NFC but the next one will enable full traversal of the source/destination patterns. Depends on D30535 Reviewers: t.p.northover, qcolombet, aditya_nandakumar, rovka, ab Subscribers: igorb, dberris, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D30536 llvm-svn: 299001
2017-03-29 23:37:18 +08:00
importImplicitDefRenderers(BuildMIAction &DstMIBuilder,
const std::vector<Record *> &ImplicitDefs) const;
[globalisel][tablegen] Add support for fpimm and import of APInt/APFloat based ImmLeaf. Summary: There's only a tablegen testcase for IntImmLeaf and not a CodeGen one because the relevant rules are rejected for other reasons at the moment. On AArch64, it's because there's an SDNodeXForm attached to the operand. On X86, it's because the rule either emits multiple instructions or has another predicate using PatFrag which cannot easily be supported at the same time. Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Reviewed By: qcolombet Subscribers: aemerson, javed.absar, igorb, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D36569 llvm-svn: 315761
2017-10-14 05:28:03 +08:00
void emitImmPredicates(raw_ostream &OS, StringRef TypeIdentifier,
StringRef Type,
[aarch64] Support APInt and APFloat in ImmLeaf subclasses and make AArch64 use them. Summary: The purpose of this patch is to expose more information about ImmLeaf-like PatLeaf's so that GlobalISel can learn to import them. Previously, ImmLeaf could only be used to test int64_t's produced by sign-extending an APInt. Other tests on immediates had to use the generic PatLeaf and extract the constant using C++. With this patch, tablegen will know how to generate predicates for APInt, and APFloat. This will allow it to 'do the right thing' for both SelectionDAG and GlobalISel which require different methods of extracting the immediate from the IR. This is NFC for SelectionDAG since the new code is equivalent to the previous code. It's also NFC for FastISel because FastIselShouldIgnore is 1 for the ImmLeaf subclasses. Enabling FastIselShouldIgnore == 0 for these new subclasses will require a significant re-factor of FastISel. For GlobalISel, it's currently NFC because the relevant code to import the affected rules is not yet present. This will be added in a later patch. Depends on D36086 Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Reviewed By: qcolombet Subscribers: bjope, aemerson, rengolin, javed.absar, igorb, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D36534 llvm-svn: 315747
2017-10-14 04:42:18 +08:00
std::function<bool(const Record *R)> Filter);
[GlobalISel] Return an Expected<RuleMatcher> for each SDAG pattern. NFC. Instead of emitting the matcher code directly, return the rule matcher and the skip reason as an Expected<RuleMatcher>. This will let us record all matchers and process them before emission. It's a somewhat unconventional use of Error, but it's nicer than, say, std::pair, because of the bool conversions. Differential Revision: https://reviews.llvm.org/D29743 llvm-svn: 294706
2017-02-10 12:00:17 +08:00
/// Analyze pattern \p P, returning a matcher for it if possible.
/// Otherwise, return an Error explaining why we don't support it.
Expected<RuleMatcher> runOnPattern(const PatternToMatch &P);
[globalisel][tablegen] Import SelectionDAG's rule predicates and support the equivalent in GIRule. Summary: The SelectionDAG importer now imports rules with Predicate's attached via Requires, PredicateControl, etc. These predicates are implemented as bitset's to allow multiple predicates to be tested together. However, unlike the MC layer subtarget features, each target only pays for it's own predicates (e.g. AArch64 doesn't have 192 feature bits just because X86 needs a lot). Both AArch64 and X86 derive at least one predicate from the MachineFunction or Function so they must re-initialize AvailableFeatures before each function. They also declare locals in <Target>InstructionSelector so that computeAvailableFeatures() can use the code from SelectionDAG without modification. Reviewers: rovka, qcolombet, aditya_nandakumar, t.p.northover, ab Reviewed By: rovka Subscribers: aemerson, rengolin, dberris, kristof.beyls, llvm-commits, igorb Differential Revision: https://reviews.llvm.org/D31418 llvm-svn: 300993
2017-04-21 23:59:56 +08:00
void declareSubtargetFeature(Record *Predicate);
[globalisel][tablegen] Import signextload and zeroextload. Allow a pattern rewriter to be installed in CodeGenDAGPatterns and use it to correct situations where SelectionDAG and GlobalISel disagree on representation. For example, it would rewrite: (sextload:i32 $ptr)<<unindexedload>><<sextload>><<sextloadi16> to: (sext:i32 (load:i16 $ptr)<<unindexedload>>) I'd have preferred to replace the fragments and have the expansion happen naturally as part of PatFrag expansion but the type inferencing system can't cope with loads of types narrower than those mentioned in register classes. This is because the SDTCisInt's on the sext constrain both the result and operand to the 'legal' integer types (where legal is defined as 'a register class can contain the type') which immediately rules the narrower types out. Several targets (those with only one legal integer type) would then go on to crash on the SDTCisOpSmallerThanOp<> when it removes all the possible types for the result of the extend. Also, improve isObviouslySafeToFold() slightly to automatically return true for neighbouring instructions. There can't be any re-ordering problems if re-ordering isn't happenning. We'll need to improve it further to handle sign/zero-extending loads when the extend and load aren't immediate neighbours though. llvm-svn: 317971
2017-11-11 11:23:44 +08:00
Reapply r332907 "[GlobalISel] Improving InstructionSelect's performance by reducing MatchTable..." Apparently the compile time problem was caused by the fact that not all compilers / STL implementations can automatically convert std::unique_ptr<Derived> to std::unique_ptr<Base>. Fixed (hopefully) by making sure it's std::unique_ptr<Derived>&& (rvalue ref) to std::unique_ptr<Base> conversion instead. llvm-svn: 332917
2018-05-22 07:28:51 +08:00
MatchTable buildMatchTable(MutableArrayRef<RuleMatcher> Rules, bool Optimize,
bool WithCoverage);
public:
[TableGen][GlobalISel] Optimize MatchTable for faster instruction selection *** Context *** Prior to this patchw, the table generated for matching instruction was straight forward but highly inefficient. Basically, each pattern generates its own set of self contained checks and actions. E.g., TableGen generated: // First pattern CheckNumOperand 3 CheckOpcode G_ADD ... Build ADDrr // Second pattern CheckNumOperand 3 CheckOpcode G_ADD ... Build ADDri // Third pattern CheckNumOperand 3 CheckOpcode G_SUB ... Build SUBrr *** Problem *** Because of that generation, a *lot* of check were redundant between each pattern and were checked every single time until we reach the pattern that matches. E.g., Taking the previous table, let say we are matching a G_SUB, that means we were going to check all the rules for G_ADD before looking at the G_SUB rule. In particular we are going to do: check 3 operands; PASS check G_ADD; FAIL ; Next rule check 3 operands; PASS (but we already knew that!) check G_ADD; FAIL (well it is still not true) ; Next rule check 3 operands; PASS (really!!) check G_SUB; PASS (at last :P) *** Proposed Solution *** This patch introduces a concept of group of rules (GroupMatcher) that share some predicates and only get checked once for the whole group. This patch only creates groups with one nesting level. Conceptually there is nothing preventing us for having deeper nest level. However, the current implementation is not smart enough to share the recording (aka capturing) of values. That limits its ability to do more sharing. For the given example the current patch will generate: // First group CheckOpcode G_ADD // First pattern CheckNumOperand 3 ... Build ADDrr // Second pattern CheckNumOperand 3 ... Build ADDri // Second group CheckOpcode G_SUB // Third pattern CheckNumOperand 3 ... Build SUBrr But if we allowed several nesting level, it could create a sub group for the checknumoperand 3. (We would need to call optimizeRules on the rules within a group.) *** Result *** With only one level of nesting, the instruction selection pass is up to 4x faster. For instance, one instruction now takes 500 checks, instead of 24k! With more nesting we could get in the tens I believe. Differential Revision: https://reviews.llvm.org/D39034 rdar://problem/34670699 llvm-svn: 321017
2017-12-19 03:47:41 +08:00
/// Takes a sequence of \p Rules and group them based on the predicates
Reapply r332907 "[GlobalISel] Improving InstructionSelect's performance by reducing MatchTable..." Apparently the compile time problem was caused by the fact that not all compilers / STL implementations can automatically convert std::unique_ptr<Derived> to std::unique_ptr<Base>. Fixed (hopefully) by making sure it's std::unique_ptr<Derived>&& (rvalue ref) to std::unique_ptr<Base> conversion instead. llvm-svn: 332917
2018-05-22 07:28:51 +08:00
/// they share. \p MatcherStorage is used as a memory container
[NFC] fix trivial typos in comments "the the" -> "the" llvm-svn: 323302
2018-01-24 13:04:35 +08:00
/// for the group that are created as part of this process.
[TableGen][GlobalISel] Optimize MatchTable for faster instruction selection *** Context *** Prior to this patchw, the table generated for matching instruction was straight forward but highly inefficient. Basically, each pattern generates its own set of self contained checks and actions. E.g., TableGen generated: // First pattern CheckNumOperand 3 CheckOpcode G_ADD ... Build ADDrr // Second pattern CheckNumOperand 3 CheckOpcode G_ADD ... Build ADDri // Third pattern CheckNumOperand 3 CheckOpcode G_SUB ... Build SUBrr *** Problem *** Because of that generation, a *lot* of check were redundant between each pattern and were checked every single time until we reach the pattern that matches. E.g., Taking the previous table, let say we are matching a G_SUB, that means we were going to check all the rules for G_ADD before looking at the G_SUB rule. In particular we are going to do: check 3 operands; PASS check G_ADD; FAIL ; Next rule check 3 operands; PASS (but we already knew that!) check G_ADD; FAIL (well it is still not true) ; Next rule check 3 operands; PASS (really!!) check G_SUB; PASS (at last :P) *** Proposed Solution *** This patch introduces a concept of group of rules (GroupMatcher) that share some predicates and only get checked once for the whole group. This patch only creates groups with one nesting level. Conceptually there is nothing preventing us for having deeper nest level. However, the current implementation is not smart enough to share the recording (aka capturing) of values. That limits its ability to do more sharing. For the given example the current patch will generate: // First group CheckOpcode G_ADD // First pattern CheckNumOperand 3 ... Build ADDrr // Second pattern CheckNumOperand 3 ... Build ADDri // Second group CheckOpcode G_SUB // Third pattern CheckNumOperand 3 ... Build SUBrr But if we allowed several nesting level, it could create a sub group for the checknumoperand 3. (We would need to call optimizeRules on the rules within a group.) *** Result *** With only one level of nesting, the instruction selection pass is up to 4x faster. For instance, one instruction now takes 500 checks, instead of 24k! With more nesting we could get in the tens I believe. Differential Revision: https://reviews.llvm.org/D39034 rdar://problem/34670699 llvm-svn: 321017
2017-12-19 03:47:41 +08:00
///
Reapply r332907 "[GlobalISel] Improving InstructionSelect's performance by reducing MatchTable..." Apparently the compile time problem was caused by the fact that not all compilers / STL implementations can automatically convert std::unique_ptr<Derived> to std::unique_ptr<Base>. Fixed (hopefully) by making sure it's std::unique_ptr<Derived>&& (rvalue ref) to std::unique_ptr<Base> conversion instead. llvm-svn: 332917
2018-05-22 07:28:51 +08:00
/// What this optimization does looks like if GroupT = GroupMatcher:
[TableGen][GlobalISel] Optimize MatchTable for faster instruction selection *** Context *** Prior to this patchw, the table generated for matching instruction was straight forward but highly inefficient. Basically, each pattern generates its own set of self contained checks and actions. E.g., TableGen generated: // First pattern CheckNumOperand 3 CheckOpcode G_ADD ... Build ADDrr // Second pattern CheckNumOperand 3 CheckOpcode G_ADD ... Build ADDri // Third pattern CheckNumOperand 3 CheckOpcode G_SUB ... Build SUBrr *** Problem *** Because of that generation, a *lot* of check were redundant between each pattern and were checked every single time until we reach the pattern that matches. E.g., Taking the previous table, let say we are matching a G_SUB, that means we were going to check all the rules for G_ADD before looking at the G_SUB rule. In particular we are going to do: check 3 operands; PASS check G_ADD; FAIL ; Next rule check 3 operands; PASS (but we already knew that!) check G_ADD; FAIL (well it is still not true) ; Next rule check 3 operands; PASS (really!!) check G_SUB; PASS (at last :P) *** Proposed Solution *** This patch introduces a concept of group of rules (GroupMatcher) that share some predicates and only get checked once for the whole group. This patch only creates groups with one nesting level. Conceptually there is nothing preventing us for having deeper nest level. However, the current implementation is not smart enough to share the recording (aka capturing) of values. That limits its ability to do more sharing. For the given example the current patch will generate: // First group CheckOpcode G_ADD // First pattern CheckNumOperand 3 ... Build ADDrr // Second pattern CheckNumOperand 3 ... Build ADDri // Second group CheckOpcode G_SUB // Third pattern CheckNumOperand 3 ... Build SUBrr But if we allowed several nesting level, it could create a sub group for the checknumoperand 3. (We would need to call optimizeRules on the rules within a group.) *** Result *** With only one level of nesting, the instruction selection pass is up to 4x faster. For instance, one instruction now takes 500 checks, instead of 24k! With more nesting we could get in the tens I believe. Differential Revision: https://reviews.llvm.org/D39034 rdar://problem/34670699 llvm-svn: 321017
2017-12-19 03:47:41 +08:00
/// Output without optimization:
/// \verbatim
/// # R1
/// # predicate A
/// # predicate B
/// ...
/// # R2
/// # predicate A // <-- effectively this is going to be checked twice.
/// // Once in R1 and once in R2.
/// # predicate C
/// \endverbatim
/// Output with optimization:
/// \verbatim
/// # Group1_2
/// # predicate A // <-- Check is now shared.
/// # R1
/// # predicate B
/// # R2
/// # predicate C
/// \endverbatim
Reapply r332907 "[GlobalISel] Improving InstructionSelect's performance by reducing MatchTable..." Apparently the compile time problem was caused by the fact that not all compilers / STL implementations can automatically convert std::unique_ptr<Derived> to std::unique_ptr<Base>. Fixed (hopefully) by making sure it's std::unique_ptr<Derived>&& (rvalue ref) to std::unique_ptr<Base> conversion instead. llvm-svn: 332917
2018-05-22 07:28:51 +08:00
template <class GroupT>
static std::vector<Matcher *> optimizeRules(
[GlobalISel][InstructionSelect] Refactoring buildMatchTable out, NFC to share it between the Instruction Selector in optimized and non-optimized modes both and the Testgen. Reviewers: dsanders, aemerson Reviewed By: dsanders Subscribers: rovka, kristof.beyls, llvm-commits, dsanders Differential Revision: https://reviews.llvm.org/D46097 llvm-svn: 331396
2018-05-03 04:08:14 +08:00
ArrayRef<Matcher *> Rules,
Reapply r332907 "[GlobalISel] Improving InstructionSelect's performance by reducing MatchTable..." Apparently the compile time problem was caused by the fact that not all compilers / STL implementations can automatically convert std::unique_ptr<Derived> to std::unique_ptr<Base>. Fixed (hopefully) by making sure it's std::unique_ptr<Derived>&& (rvalue ref) to std::unique_ptr<Base> conversion instead. llvm-svn: 332917
2018-05-22 07:28:51 +08:00
std::vector<std::unique_ptr<Matcher>> &MatcherStorage);
Revert r332907 "[GlobalISel] Improving InstructionSelect's performance by reducing MatchTable..." There is a compile time error I didn't see locally, investigating now. llvm-svn: 332912
2018-05-22 06:21:24 +08:00
};
[GlobalISel] Improving InstructionSelect's performance by reducing MatchTable, mostly NFC, perf patch 1 This patch starts a series of patches that decrease time spent by GlobalISel in its InstructionSelect pass by roughly 60% for -O0 builds for large inputs as measured on sqlite3-amalgamation (http://sqlite.org/download.html) targeting AArch64. The performance improvements are achieved solely by reducing the number of matching GIM_* opcodes executed by the MatchTable's interpreter during the selection by approx. a factor of 30, which also brings contribution of this particular part of the selection process to the overall runtime of InstructionSelect pass down from approx. 60-70% to 5-7%, thus making further improvements in this particular direction not very profitable. The improvements described above are expected for any target that doesn't have many complex patterns. The targets that do should strictly benefit from the changes, but by how much exactly is hard to estimate beforehand. It's also likely that such target WILL benefit from further improvements to MatchTable, most likely the ones that bring it closer to a perfect decision tree. This commit specifically is rather large mostly NFC commit that does necessary preparation work and refactoring, there will be a following series of small patches introducing a specific optimization each shortly after. This commit specifically is expected to cause a small compile time regression (around 2.5% of InstructionSelect pass time), which should be fixed by the next commit of the series. Every commit planned shares the same Phabricator Review. Reviewers: qcolombet, dsanders, bogner, aemerson, javed.absar Reviewed By: qcolombet Subscribers: rovka, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D44700 llvm-svn: 332907
2018-05-22 06:04:39 +08:00
Reapply r332907 "[GlobalISel] Improving InstructionSelect's performance by reducing MatchTable..." Apparently the compile time problem was caused by the fact that not all compilers / STL implementations can automatically convert std::unique_ptr<Derived> to std::unique_ptr<Base>. Fixed (hopefully) by making sure it's std::unique_ptr<Derived>&& (rvalue ref) to std::unique_ptr<Base> conversion instead. llvm-svn: 332917
2018-05-22 07:28:51 +08:00
void GlobalISelEmitter::gatherOpcodeValues() {
InstructionOpcodeMatcher::initOpcodeValuesMap(Target);
}
void GlobalISelEmitter::gatherTypeIDValues() {
LLTOperandMatcher::initTypeIDValuesMap();
}
[GlobalISel] Add basic Selector-emitter tblgen backend. This adds a basic tablegen backend that analyzes the SelectionDAG patterns to find simple ones that are eligible for GlobalISel-emission. That's similar to FastISel, with one notable difference: we're not fed ISD opcodes, so we need to map the SDNode operators to generic opcodes. That's done using GINodeEquiv in TargetGlobalISel.td. Otherwise, this is mostly boilerplate, and lots of filtering of any kind of "complicated" pattern. On AArch64, this is sufficient to match G_ADD up to s64 (to ADDWrr/ADDXrr) and G_BR (to B). Differential Revision: https://reviews.llvm.org/D26878 llvm-svn: 290284
2016-12-22 07:26:20 +08:00
void GlobalISelEmitter::gatherNodeEquivs() {
assert(NodeEquivs.empty());
for (Record *Equiv : RK.getAllDerivedDefinitions("GINodeEquiv"))
[globalisel][tablegen] Map ld and st to G_LOAD and G_STORE. NFC Summary: There is an important mismatch between ISD::LOAD and G_LOAD (and likewise for ISD::STORE and G_STORE). In SelectionDAG, ISD::LOAD is a non-atomic load and atomic loads are handled by a separate node. However, this is not true of GlobalISel's G_LOAD. For G_LOAD, the MachineMemOperand indicates the atomicity of the operation. As a result, this mapping must also add a predicate that checks for non-atomic MachineMemOperands. This is NFC since these nodes always have predicates in practice and are therefore always rejected at the moment. Depends on D37443 Reviewers: ab, qcolombet, t.p.northover, rovka, aditya_nandakumar Reviewed By: qcolombet Subscribers: kristof.beyls, llvm-commits, igorb Differential Revision: https://reviews.llvm.org/D37445 llvm-svn: 315843
2017-10-15 10:41:12 +08:00
NodeEquivs[Equiv->getValueAsDef("Node")] = Equiv;
[globalisel][tblgen] Add support for ComplexPatterns Summary: Adds a new kind of MachineOperand: MO_Placeholder. This operand must not appear in the MIR and only exists as a way of creating an 'uninitialized' operand until a matcher function overwrites it. Depends on D30046, D29712 Reviewers: t.p.northover, ab, rovka, aditya_nandakumar, javed.absar, qcolombet Reviewed By: qcolombet Subscribers: dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D30089 llvm-svn: 297782
2017-03-15 05:32:08 +08:00
assert(ComplexPatternEquivs.empty());
for (Record *Equiv : RK.getAllDerivedDefinitions("GIComplexPatternEquiv")) {
Record *SelDAGEquiv = Equiv->getValueAsDef("SelDAGEquivalent");
if (!SelDAGEquiv)
continue;
ComplexPatternEquivs[SelDAGEquiv] = Equiv;
}
[GlobalISel][TableGen] Add support for SDNodeXForm Summary: This patch adds CustomRenderer which renders the matched operands to the specified instruction. Targets can enable the matching of SDNodeXForm by adding a definition that inherits from GICustomOperandRenderer and GISDNodeXFormEquiv as follows. def gi_imm8 : GICustomOperandRenderer<"renderImm8”>, GISDNodeXFormEquiv<imm8_xform>; Custom renderer functions should be of the form: void render(MachineInstrBuilder &MIB, const MachineInstr &I); Reviewers: dsanders, ab, rovka Reviewed By: dsanders Subscribers: kristof.beyls, javed.absar, llvm-commits, mgrang, qcolombet Differential Revision: https://reviews.llvm.org/D42012 llvm-svn: 322582
2018-01-17 02:44:05 +08:00
assert(SDNodeXFormEquivs.empty());
for (Record *Equiv : RK.getAllDerivedDefinitions("GISDNodeXFormEquiv")) {
Record *SelDAGEquiv = Equiv->getValueAsDef("SelDAGEquivalent");
if (!SelDAGEquiv)
continue;
SDNodeXFormEquivs[SelDAGEquiv] = Equiv;
}
[GlobalISel] Add basic Selector-emitter tblgen backend. This adds a basic tablegen backend that analyzes the SelectionDAG patterns to find simple ones that are eligible for GlobalISel-emission. That's similar to FastISel, with one notable difference: we're not fed ISD opcodes, so we need to map the SDNode operators to generic opcodes. That's done using GINodeEquiv in TargetGlobalISel.td. Otherwise, this is mostly boilerplate, and lots of filtering of any kind of "complicated" pattern. On AArch64, this is sufficient to match G_ADD up to s64 (to ADDWrr/ADDXrr) and G_BR (to B). Differential Revision: https://reviews.llvm.org/D26878 llvm-svn: 290284
2016-12-22 07:26:20 +08:00
}
[globalisel][tablegen] Map ld and st to G_LOAD and G_STORE. NFC Summary: There is an important mismatch between ISD::LOAD and G_LOAD (and likewise for ISD::STORE and G_STORE). In SelectionDAG, ISD::LOAD is a non-atomic load and atomic loads are handled by a separate node. However, this is not true of GlobalISel's G_LOAD. For G_LOAD, the MachineMemOperand indicates the atomicity of the operation. As a result, this mapping must also add a predicate that checks for non-atomic MachineMemOperands. This is NFC since these nodes always have predicates in practice and are therefore always rejected at the moment. Depends on D37443 Reviewers: ab, qcolombet, t.p.northover, rovka, aditya_nandakumar Reviewed By: qcolombet Subscribers: kristof.beyls, llvm-commits, igorb Differential Revision: https://reviews.llvm.org/D37445 llvm-svn: 315843
2017-10-15 10:41:12 +08:00
Record *GlobalISelEmitter::findNodeEquiv(Record *N) const {
[GlobalISel] Add basic Selector-emitter tblgen backend. This adds a basic tablegen backend that analyzes the SelectionDAG patterns to find simple ones that are eligible for GlobalISel-emission. That's similar to FastISel, with one notable difference: we're not fed ISD opcodes, so we need to map the SDNode operators to generic opcodes. That's done using GINodeEquiv in TargetGlobalISel.td. Otherwise, this is mostly boilerplate, and lots of filtering of any kind of "complicated" pattern. On AArch64, this is sufficient to match G_ADD up to s64 (to ADDWrr/ADDXrr) and G_BR (to B). Differential Revision: https://reviews.llvm.org/D26878 llvm-svn: 290284
2016-12-22 07:26:20 +08:00
return NodeEquivs.lookup(N);
}
[globalisel] Update GlobalISel emitter to match new representation of extending loads Summary: Previously, a extending load was represented at (G_*EXT (G_LOAD x)). This had a few drawbacks: * G_LOAD had to be legal for all sizes you could extend from, even if registers didn't naturally hold those sizes. * All sizes you could extend from had to be allocatable just in case the extend went missing (e.g. by optimization). * At minimum, G_*EXT and G_TRUNC had to be legal for these sizes. As we improve optimization of extends and truncates, this legality requirement would spread without considerable care w.r.t when certain combines were permitted. * The SelectionDAG importer required some ugly and fragile pattern rewriting to translate patterns into this style. This patch changes the representation to: * (G_[SZ]EXTLOAD x) * (G_LOAD x) any-extends when MMO.getSize() * 8 < ResultTy.getSizeInBits() which resolves these issues by allowing targets to work entirely in their native register sizes, and by having a more direct translation from SelectionDAG patterns. Each extending load can be lowered by the legalizer into separate extends and loads, however a target that supports s1 will need the any-extending load to extend to at least s8 since LLVM does not represent memory accesses smaller than 8 bit. The legalizer can widenScalar G_LOAD into an any-extending load but sign/zero-extending loads need help from something else like a combiner pass. A follow-up patch that adds combiner helpers for for this will follow. The new representation requires that the MMO correctly reflect the memory access so this has been corrected in a couple tests. I've also moved the extending loads to their own tests since they are (mostly) separate opcodes now. Additionally, the re-write appears to have invalidated two tests from select-with-no-legality-check.mir since the matcher table no longer contains loads that result in s1's and they aren't legal in AArch64 anymore. Depends on D45540 Reviewers: ab, aditya_nandakumar, bogner, rtereshin, volkan, rovka, javed.absar Reviewed By: rtereshin Subscribers: javed.absar, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D45541 llvm-svn: 331601
2018-05-06 04:53:24 +08:00
const CodeGenInstruction *
GlobalISelEmitter::getEquivNode(Record &Equiv, const TreePatternNode *N) const {
for (const auto &Predicate : N->getPredicateFns()) {
if (!Equiv.isValueUnset("IfSignExtend") && Predicate.isLoad() &&
Predicate.isSignExtLoad())
return &Target.getInstruction(Equiv.getValueAsDef("IfSignExtend"));
if (!Equiv.isValueUnset("IfZeroExtend") && Predicate.isLoad() &&
Predicate.isZeroExtLoad())
return &Target.getInstruction(Equiv.getValueAsDef("IfZeroExtend"));
}
return &Target.getInstruction(Equiv.getValueAsDef("I"));
}
[GlobalISel] Add basic Selector-emitter tblgen backend. This adds a basic tablegen backend that analyzes the SelectionDAG patterns to find simple ones that are eligible for GlobalISel-emission. That's similar to FastISel, with one notable difference: we're not fed ISD opcodes, so we need to map the SDNode operators to generic opcodes. That's done using GINodeEquiv in TargetGlobalISel.td. Otherwise, this is mostly boilerplate, and lots of filtering of any kind of "complicated" pattern. On AArch64, this is sufficient to match G_ADD up to s64 (to ADDWrr/ADDXrr) and G_BR (to B). Differential Revision: https://reviews.llvm.org/D26878 llvm-svn: 290284
2016-12-22 07:26:20 +08:00
GlobalISelEmitter::GlobalISelEmitter(RecordKeeper &RK)
[globalisel] Update GlobalISel emitter to match new representation of extending loads Summary: Previously, a extending load was represented at (G_*EXT (G_LOAD x)). This had a few drawbacks: * G_LOAD had to be legal for all sizes you could extend from, even if registers didn't naturally hold those sizes. * All sizes you could extend from had to be allocatable just in case the extend went missing (e.g. by optimization). * At minimum, G_*EXT and G_TRUNC had to be legal for these sizes. As we improve optimization of extends and truncates, this legality requirement would spread without considerable care w.r.t when certain combines were permitted. * The SelectionDAG importer required some ugly and fragile pattern rewriting to translate patterns into this style. This patch changes the representation to: * (G_[SZ]EXTLOAD x) * (G_LOAD x) any-extends when MMO.getSize() * 8 < ResultTy.getSizeInBits() which resolves these issues by allowing targets to work entirely in their native register sizes, and by having a more direct translation from SelectionDAG patterns. Each extending load can be lowered by the legalizer into separate extends and loads, however a target that supports s1 will need the any-extending load to extend to at least s8 since LLVM does not represent memory accesses smaller than 8 bit. The legalizer can widenScalar G_LOAD into an any-extending load but sign/zero-extending loads need help from something else like a combiner pass. A follow-up patch that adds combiner helpers for for this will follow. The new representation requires that the MMO correctly reflect the memory access so this has been corrected in a couple tests. I've also moved the extending loads to their own tests since they are (mostly) separate opcodes now. Additionally, the re-write appears to have invalidated two tests from select-with-no-legality-check.mir since the matcher table no longer contains loads that result in s1's and they aren't legal in AArch64 anymore. Depends on D45540 Reviewers: ab, aditya_nandakumar, bogner, rtereshin, volkan, rovka, javed.absar Reviewed By: rtereshin Subscribers: javed.absar, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D45541 llvm-svn: 331601
2018-05-06 04:53:24 +08:00
: RK(RK), CGP(RK), Target(CGP.getTargetInfo()),
CGRegs(RK, Target.getHwModes()) {}
[GlobalISel] Add basic Selector-emitter tblgen backend. This adds a basic tablegen backend that analyzes the SelectionDAG patterns to find simple ones that are eligible for GlobalISel-emission. That's similar to FastISel, with one notable difference: we're not fed ISD opcodes, so we need to map the SDNode operators to generic opcodes. That's done using GINodeEquiv in TargetGlobalISel.td. Otherwise, this is mostly boilerplate, and lots of filtering of any kind of "complicated" pattern. On AArch64, this is sufficient to match G_ADD up to s64 (to ADDWrr/ADDXrr) and G_BR (to B). Differential Revision: https://reviews.llvm.org/D26878 llvm-svn: 290284
2016-12-22 07:26:20 +08:00
//===- Emitter ------------------------------------------------------------===//
[globalisel][tablegen] Change Expected<bool> to Error and rename functions. Functions that still return Expected<X> are now called createAndImport*() Changing the return type was requested in the review comments for r299001 llvm-svn: 299063
2017-03-30 17:36:33 +08:00
Error
[tablegen][globalisel] Convert the SelectionDAG importer to a tree walking approach. NFC Summary: But don't actually inspect the tree any deeper than we already do. This change is NFC but the next one will enable full traversal of the source/destination patterns. Depends on D30535 Reviewers: t.p.northover, qcolombet, aditya_nandakumar, rovka, ab Subscribers: igorb, dberris, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D30536 llvm-svn: 299001
2017-03-29 23:37:18 +08:00
GlobalISelEmitter::importRulePredicates(RuleMatcher &M,
TableGen support for parameterized register class information This replaces TableGen's type inference to operate on parameterized types instead of MVTs, and as a consequence, some interfaces have changed: - Uses of MVTs are replaced by ValueTypeByHwMode. - EEVT::TypeSet is replaced by TypeSetByHwMode. This affects the way that types and type sets are printed, and the tests relying on that have been updated. There are certain users of the inferred types outside of TableGen itself, namely FastISel and GlobalISel. For those users, the way that the types are accessed have changed. For typical scenarios, these replacements can be used: - TreePatternNode::getType(ResNo) -> getSimpleType(ResNo) - TreePatternNode::hasTypeSet(ResNo) -> hasConcreteType(ResNo) - TypeSet::isConcrete -> TypeSetByHwMode::isValueTypeByHwMode(false) For more information, please refer to the review page. Differential Revision: https://reviews.llvm.org/D31951 llvm-svn: 313271
2017-09-15 00:56:21 +08:00
ArrayRef<Predicate> Predicates) {
for (const Predicate &P : Predicates) {
if (!P.Def)
continue;
declareSubtargetFeature(P.Def);
M.addRequiredFeature(P.Def);
[globalisel][tablegen] Import SelectionDAG's rule predicates and support the equivalent in GIRule. Summary: The SelectionDAG importer now imports rules with Predicate's attached via Requires, PredicateControl, etc. These predicates are implemented as bitset's to allow multiple predicates to be tested together. However, unlike the MC layer subtarget features, each target only pays for it's own predicates (e.g. AArch64 doesn't have 192 feature bits just because X86 needs a lot). Both AArch64 and X86 derive at least one predicate from the MachineFunction or Function so they must re-initialize AvailableFeatures before each function. They also declare locals in <Target>InstructionSelector so that computeAvailableFeatures() can use the code from SelectionDAG without modification. Reviewers: rovka, qcolombet, aditya_nandakumar, t.p.northover, ab Reviewed By: rovka Subscribers: aemerson, rengolin, dberris, kristof.beyls, llvm-commits, igorb Differential Revision: https://reviews.llvm.org/D31418 llvm-svn: 300993
2017-04-21 23:59:56 +08:00
}
[globalisel][tablegen] Change Expected<bool> to Error and rename functions. Functions that still return Expected<X> are now called createAndImport*() Changing the return type was requested in the review comments for r299001 llvm-svn: 299063
2017-03-30 17:36:33 +08:00
return Error::success();
[tablegen][globalisel] Convert the SelectionDAG importer to a tree walking approach. NFC Summary: But don't actually inspect the tree any deeper than we already do. This change is NFC but the next one will enable full traversal of the source/destination patterns. Depends on D30535 Reviewers: t.p.northover, qcolombet, aditya_nandakumar, rovka, ab Subscribers: igorb, dberris, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D30536 llvm-svn: 299001
2017-03-29 23:37:18 +08:00
}
[GlobalISel] Add basic Selector-emitter tblgen backend. This adds a basic tablegen backend that analyzes the SelectionDAG patterns to find simple ones that are eligible for GlobalISel-emission. That's similar to FastISel, with one notable difference: we're not fed ISD opcodes, so we need to map the SDNode operators to generic opcodes. That's done using GINodeEquiv in TargetGlobalISel.td. Otherwise, this is mostly boilerplate, and lots of filtering of any kind of "complicated" pattern. On AArch64, this is sufficient to match G_ADD up to s64 (to ADDWrr/ADDXrr) and G_BR (to B). Differential Revision: https://reviews.llvm.org/D26878 llvm-svn: 290284
2016-12-22 07:26:20 +08:00
Revert r319691: [globalisel][tablegen] Split atomic load/store into separate opcode and enable for AArch64. Some concerns were raised with the direction. Revert while we discuss it and look into an alternative llvm-svn: 319739
2017-12-05 13:52:07 +08:00
Expected<InstructionMatcher &> GlobalISelEmitter::createAndImportSelDAGMatcher(
RuleMatcher &Rule, InstructionMatcher &InsnMatcher,
const TreePatternNode *Src, unsigned &TempOpIdx) const {
Record *SrcGIEquivOrNull = nullptr;
const CodeGenInstruction *SrcGIOrNull = nullptr;
// Start with the defined operands (i.e., the results of the root operator).
if (Src->getExtTypes().size() > 1)
return failedImport("Src pattern has multiple results");
if (Src->isLeaf()) {
Init *SrcInit = Src->getLeafValue();
if (isa<IntInit>(SrcInit)) {
InsnMatcher.addPredicate<InstructionOpcodeMatcher>(
&Target.getInstruction(RK.getDef("G_CONSTANT")));
} else
return failedImport(
"Unable to deduce gMIR opcode to handle Src (which is a leaf)");
} else {
SrcGIEquivOrNull = findNodeEquiv(Src->getOperator());
if (!SrcGIEquivOrNull)
return failedImport("Pattern operator lacks an equivalent Instruction" +
explainOperator(Src->getOperator()));
[globalisel] Update GlobalISel emitter to match new representation of extending loads Summary: Previously, a extending load was represented at (G_*EXT (G_LOAD x)). This had a few drawbacks: * G_LOAD had to be legal for all sizes you could extend from, even if registers didn't naturally hold those sizes. * All sizes you could extend from had to be allocatable just in case the extend went missing (e.g. by optimization). * At minimum, G_*EXT and G_TRUNC had to be legal for these sizes. As we improve optimization of extends and truncates, this legality requirement would spread without considerable care w.r.t when certain combines were permitted. * The SelectionDAG importer required some ugly and fragile pattern rewriting to translate patterns into this style. This patch changes the representation to: * (G_[SZ]EXTLOAD x) * (G_LOAD x) any-extends when MMO.getSize() * 8 < ResultTy.getSizeInBits() which resolves these issues by allowing targets to work entirely in their native register sizes, and by having a more direct translation from SelectionDAG patterns. Each extending load can be lowered by the legalizer into separate extends and loads, however a target that supports s1 will need the any-extending load to extend to at least s8 since LLVM does not represent memory accesses smaller than 8 bit. The legalizer can widenScalar G_LOAD into an any-extending load but sign/zero-extending loads need help from something else like a combiner pass. A follow-up patch that adds combiner helpers for for this will follow. The new representation requires that the MMO correctly reflect the memory access so this has been corrected in a couple tests. I've also moved the extending loads to their own tests since they are (mostly) separate opcodes now. Additionally, the re-write appears to have invalidated two tests from select-with-no-legality-check.mir since the matcher table no longer contains loads that result in s1's and they aren't legal in AArch64 anymore. Depends on D45540 Reviewers: ab, aditya_nandakumar, bogner, rtereshin, volkan, rovka, javed.absar Reviewed By: rtereshin Subscribers: javed.absar, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D45541 llvm-svn: 331601
2018-05-06 04:53:24 +08:00
SrcGIOrNull = getEquivNode(*SrcGIEquivOrNull, Src);
Revert r319691: [globalisel][tablegen] Split atomic load/store into separate opcode and enable for AArch64. Some concerns were raised with the direction. Revert while we discuss it and look into an alternative llvm-svn: 319739
2017-12-05 13:52:07 +08:00
// The operators look good: match the opcode
InsnMatcher.addPredicate<InstructionOpcodeMatcher>(SrcGIOrNull);
}
unsigned OpIdx = 0;
for (const TypeSetByHwMode &VTy : Src->getExtTypes()) {
// Results don't have a name unless they are the root node. The caller will
// set the name if appropriate.
OperandMatcher &OM = InsnMatcher.addOperand(OpIdx++, "", TempOpIdx);
if (auto Error = OM.addTypeCheckPredicate(VTy, false /* OperandIsAPointer */))
return failedImport(toString(std::move(Error)) +
" for result of Src pattern operator");
}
Re-commit: [globalisel][tablegen] Add support for ImmLeaf without SDNodeXForm Summary: This patch adds support for predicates on imm nodes but only for ImmLeaf and not for PatLeaf or PatFrag and only where the value does not need to be transformed before being rendered into the instruction. The limitation on PatLeaf/PatFrag/SDNodeXForm is due to differences in the necessary target-supplied C++ for GlobalISel. Depends on D36085 The previous commit was reverted for breaking the build but this appears to have been the recurring problem on the Windows bots with tablegen not being re-run when llvm-tblgen is changed but the .td's aren't. If it re-occurs then forcing a build with clean=True should fix it but this string should do this in advance: Requires a clean build. Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Reviewed By: rovka Subscribers: kristof.beyls, javed.absar, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D36086 llvm-svn: 311645
2017-08-24 17:11:20 +08:00
for (const auto &Predicate : Src->getPredicateFns()) {
if (Predicate.isAlwaysTrue())
continue;
if (Predicate.isImmediatePattern()) {
InsnMatcher.addPredicate<InstructionImmPredicateMatcher>(Predicate);
continue;
}
[globalisel] Update GlobalISel emitter to match new representation of extending loads Summary: Previously, a extending load was represented at (G_*EXT (G_LOAD x)). This had a few drawbacks: * G_LOAD had to be legal for all sizes you could extend from, even if registers didn't naturally hold those sizes. * All sizes you could extend from had to be allocatable just in case the extend went missing (e.g. by optimization). * At minimum, G_*EXT and G_TRUNC had to be legal for these sizes. As we improve optimization of extends and truncates, this legality requirement would spread without considerable care w.r.t when certain combines were permitted. * The SelectionDAG importer required some ugly and fragile pattern rewriting to translate patterns into this style. This patch changes the representation to: * (G_[SZ]EXTLOAD x) * (G_LOAD x) any-extends when MMO.getSize() * 8 < ResultTy.getSizeInBits() which resolves these issues by allowing targets to work entirely in their native register sizes, and by having a more direct translation from SelectionDAG patterns. Each extending load can be lowered by the legalizer into separate extends and loads, however a target that supports s1 will need the any-extending load to extend to at least s8 since LLVM does not represent memory accesses smaller than 8 bit. The legalizer can widenScalar G_LOAD into an any-extending load but sign/zero-extending loads need help from something else like a combiner pass. A follow-up patch that adds combiner helpers for for this will follow. The new representation requires that the MMO correctly reflect the memory access so this has been corrected in a couple tests. I've also moved the extending loads to their own tests since they are (mostly) separate opcodes now. Additionally, the re-write appears to have invalidated two tests from select-with-no-legality-check.mir since the matcher table no longer contains loads that result in s1's and they aren't legal in AArch64 anymore. Depends on D45540 Reviewers: ab, aditya_nandakumar, bogner, rtereshin, volkan, rovka, javed.absar Reviewed By: rtereshin Subscribers: javed.absar, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D45541 llvm-svn: 331601
2018-05-06 04:53:24 +08:00
// G_LOAD is used for both non-extending and any-extending loads.
if (Predicate.isLoad() && Predicate.isNonExtLoad()) {
InsnMatcher.addPredicate<MemoryVsLLTSizePredicateMatcher>(
0, MemoryVsLLTSizePredicateMatcher::EqualTo, 0);
continue;
}
if (Predicate.isLoad() && Predicate.isAnyExtLoad()) {
InsnMatcher.addPredicate<MemoryVsLLTSizePredicateMatcher>(
0, MemoryVsLLTSizePredicateMatcher::LessThan, 0);
continue;
}
// No check required. We already did it by swapping the opcode.
if (!SrcGIEquivOrNull->isValueUnset("IfSignExtend") &&
Predicate.isSignExtLoad())
continue;
// No check required. We already did it by swapping the opcode.
if (!SrcGIEquivOrNull->isValueUnset("IfZeroExtend") &&
Predicate.isZeroExtLoad())
[globalisel][tablegen] Implement unindexed load, non-extending load, and MemVT checks Summary: This includes some context-sensitivity in the MVT to LLT conversion so that pointer types are tested correctly. FIXME: I'm not happy with the way this is done since everything is a special-case. I've yet to find a reasonable way to implement it. select-load.mir fails because <1 x s64> loads in tablegen get priority over s64 loads. This is fixed in the next patch and as such they should be committed together, I've posted them separately to help with the review. Depends on D37456 Reviewers: ab, qcolombet, t.p.northover, rovka, aditya_nandakumar Subscribers: kristof.beyls, javed.absar, llvm-commits, igorb Differential Revision: https://reviews.llvm.org/D37457 llvm-svn: 315884
2017-10-16 08:56:30 +08:00
continue;
[globalisel][tablegen] Import stores and allow GISel to automatically substitute zero regs like WZR/XZR/$zero. This patch enables the import of stores. Unfortunately, doing so by itself, loses an optimization where storing 0 to memory makes use of WZR/XZR. To mitigate this, this patch also introduces a new feature that allows register operands to nominate a zero register. When this is done, GlobalISel will substitute (G_CONSTANT 0) with the nominated register automatically. This is currently configured to only apply to the stores. Applying it to GPR32/GPR64 register classes in general will be done after review see (https://reviews.llvm.org/D39150). llvm-svn: 316360
2017-10-24 02:19:24 +08:00
// No check required. G_STORE by itself is a non-extending store.
if (Predicate.isNonTruncStore())
continue;
[globalisel][tablegen] Add support for importing G_ATOMIC_CMPXCHG, G_ATOMICRMW_* rules from SelectionDAG. GIM_CheckNonAtomic has been replaced by GIM_CheckAtomicOrdering to allow it to support a wider range of orderings. This has then been used to import patterns using nodes such as atomic_cmp_swap, atomic_swap, and atomic_load_*. llvm-svn: 319232
2017-11-29 06:07:05 +08:00
if (Predicate.isLoad() || Predicate.isStore() || Predicate.isAtomic()) {
if (Predicate.getMemoryVT() != nullptr) {
Optional<LLTCodeGen> MemTyOrNone =
MVTToLLT(getValueType(Predicate.getMemoryVT()));
[globalisel][tablegen] Import stores and allow GISel to automatically substitute zero regs like WZR/XZR/$zero. This patch enables the import of stores. Unfortunately, doing so by itself, loses an optimization where storing 0 to memory makes use of WZR/XZR. To mitigate this, this patch also introduces a new feature that allows register operands to nominate a zero register. When this is done, GlobalISel will substitute (G_CONSTANT 0) with the nominated register automatically. This is currently configured to only apply to the stores. Applying it to GPR32/GPR64 register classes in general will be done after review see (https://reviews.llvm.org/D39150). llvm-svn: 316360
2017-10-24 02:19:24 +08:00
[globalisel][tablegen] Add support for importing G_ATOMIC_CMPXCHG, G_ATOMICRMW_* rules from SelectionDAG. GIM_CheckNonAtomic has been replaced by GIM_CheckAtomicOrdering to allow it to support a wider range of orderings. This has then been used to import patterns using nodes such as atomic_cmp_swap, atomic_swap, and atomic_load_*. llvm-svn: 319232
2017-11-29 06:07:05 +08:00
if (!MemTyOrNone)
return failedImport("MemVT could not be converted to LLT");
[globalisel][tablegen] Import stores and allow GISel to automatically substitute zero regs like WZR/XZR/$zero. This patch enables the import of stores. Unfortunately, doing so by itself, loses an optimization where storing 0 to memory makes use of WZR/XZR. To mitigate this, this patch also introduces a new feature that allows register operands to nominate a zero register. When this is done, GlobalISel will substitute (G_CONSTANT 0) with the nominated register automatically. This is currently configured to only apply to the stores. Applying it to GPR32/GPR64 register classes in general will be done after review see (https://reviews.llvm.org/D39150). llvm-svn: 316360
2017-10-24 02:19:24 +08:00
[globalisel] Update GlobalISel emitter to match new representation of extending loads Summary: Previously, a extending load was represented at (G_*EXT (G_LOAD x)). This had a few drawbacks: * G_LOAD had to be legal for all sizes you could extend from, even if registers didn't naturally hold those sizes. * All sizes you could extend from had to be allocatable just in case the extend went missing (e.g. by optimization). * At minimum, G_*EXT and G_TRUNC had to be legal for these sizes. As we improve optimization of extends and truncates, this legality requirement would spread without considerable care w.r.t when certain combines were permitted. * The SelectionDAG importer required some ugly and fragile pattern rewriting to translate patterns into this style. This patch changes the representation to: * (G_[SZ]EXTLOAD x) * (G_LOAD x) any-extends when MMO.getSize() * 8 < ResultTy.getSizeInBits() which resolves these issues by allowing targets to work entirely in their native register sizes, and by having a more direct translation from SelectionDAG patterns. Each extending load can be lowered by the legalizer into separate extends and loads, however a target that supports s1 will need the any-extending load to extend to at least s8 since LLVM does not represent memory accesses smaller than 8 bit. The legalizer can widenScalar G_LOAD into an any-extending load but sign/zero-extending loads need help from something else like a combiner pass. A follow-up patch that adds combiner helpers for for this will follow. The new representation requires that the MMO correctly reflect the memory access so this has been corrected in a couple tests. I've also moved the extending loads to their own tests since they are (mostly) separate opcodes now. Additionally, the re-write appears to have invalidated two tests from select-with-no-legality-check.mir since the matcher table no longer contains loads that result in s1's and they aren't legal in AArch64 anymore. Depends on D45540 Reviewers: ab, aditya_nandakumar, bogner, rtereshin, volkan, rovka, javed.absar Reviewed By: rtereshin Subscribers: javed.absar, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D45541 llvm-svn: 331601
2018-05-06 04:53:24 +08:00
// MMO's work in bytes so we must take care of unusual types like i1
// don't round down.
unsigned MemSizeInBits =
llvm::alignTo(MemTyOrNone->get().getSizeInBits(), 8);
InsnMatcher.addPredicate<MemorySizePredicateMatcher>(
0, MemSizeInBits / 8);
[globalisel][tablegen] Add support for importing G_ATOMIC_CMPXCHG, G_ATOMICRMW_* rules from SelectionDAG. GIM_CheckNonAtomic has been replaced by GIM_CheckAtomicOrdering to allow it to support a wider range of orderings. This has then been used to import patterns using nodes such as atomic_cmp_swap, atomic_swap, and atomic_load_*. llvm-svn: 319232
2017-11-29 06:07:05 +08:00
continue;
}
}
if (Predicate.isLoad() || Predicate.isStore()) {
// No check required. A G_LOAD/G_STORE is an unindexed load.
if (Predicate.isUnindexed())
continue;
}
if (Predicate.isAtomic()) {
if (Predicate.isAtomicOrderingMonotonic()) {
InsnMatcher.addPredicate<AtomicOrderingMMOPredicateMatcher>(
"Monotonic");
continue;
}
if (Predicate.isAtomicOrderingAcquire()) {
InsnMatcher.addPredicate<AtomicOrderingMMOPredicateMatcher>("Acquire");
continue;
}
if (Predicate.isAtomicOrderingRelease()) {
InsnMatcher.addPredicate<AtomicOrderingMMOPredicateMatcher>("Release");
continue;
}
if (Predicate.isAtomicOrderingAcquireRelease()) {
InsnMatcher.addPredicate<AtomicOrderingMMOPredicateMatcher>(
"AcquireRelease");
continue;
}
if (Predicate.isAtomicOrderingSequentiallyConsistent()) {
InsnMatcher.addPredicate<AtomicOrderingMMOPredicateMatcher>(
"SequentiallyConsistent");
continue;
}
[globalisel][tablegen] Add support for relative AtomicOrderings No test yet because the relevant rules are blocked on the atomic_load, and atomic_store nodes. llvm-svn: 319475
2017-12-01 05:05:59 +08:00
if (Predicate.isAtomicOrderingAcquireOrStronger()) {
InsnMatcher.addPredicate<AtomicOrderingMMOPredicateMatcher>(
"Acquire", AtomicOrderingMMOPredicateMatcher::AO_OrStronger);
continue;
}
if (Predicate.isAtomicOrderingWeakerThanAcquire()) {
InsnMatcher.addPredicate<AtomicOrderingMMOPredicateMatcher>(
"Acquire", AtomicOrderingMMOPredicateMatcher::AO_WeakerThan);
continue;
}
if (Predicate.isAtomicOrderingReleaseOrStronger()) {
InsnMatcher.addPredicate<AtomicOrderingMMOPredicateMatcher>(
"Release", AtomicOrderingMMOPredicateMatcher::AO_OrStronger);
continue;
}
if (Predicate.isAtomicOrderingWeakerThanRelease()) {
InsnMatcher.addPredicate<AtomicOrderingMMOPredicateMatcher>(
"Release", AtomicOrderingMMOPredicateMatcher::AO_WeakerThan);
continue;
}
[globalisel][tablegen] Import stores and allow GISel to automatically substitute zero regs like WZR/XZR/$zero. This patch enables the import of stores. Unfortunately, doing so by itself, loses an optimization where storing 0 to memory makes use of WZR/XZR. To mitigate this, this patch also introduces a new feature that allows register operands to nominate a zero register. When this is done, GlobalISel will substitute (G_CONSTANT 0) with the nominated register automatically. This is currently configured to only apply to the stores. Applying it to GPR32/GPR64 register classes in general will be done after review see (https://reviews.llvm.org/D39150). llvm-svn: 316360
2017-10-24 02:19:24 +08:00
}
Re-commit: [globalisel][tablegen] Add support for ImmLeaf without SDNodeXForm Summary: This patch adds support for predicates on imm nodes but only for ImmLeaf and not for PatLeaf or PatFrag and only where the value does not need to be transformed before being rendered into the instruction. The limitation on PatLeaf/PatFrag/SDNodeXForm is due to differences in the necessary target-supplied C++ for GlobalISel. Depends on D36085 The previous commit was reverted for breaking the build but this appears to have been the recurring problem on the Windows bots with tablegen not being re-run when llvm-tblgen is changed but the .td's aren't. If it re-occurs then forcing a build with clean=True should fix it but this string should do this in advance: Requires a clean build. Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Reviewed By: rovka Subscribers: kristof.beyls, javed.absar, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D36086 llvm-svn: 311645
2017-08-24 17:11:20 +08:00
return failedImport("Src pattern child has predicate (" +
explainPredicates(Src) + ")");
}
Revert r319691: [globalisel][tablegen] Split atomic load/store into separate opcode and enable for AArch64. Some concerns were raised with the direction. Revert while we discuss it and look into an alternative llvm-svn: 319739
2017-12-05 13:52:07 +08:00
if (SrcGIEquivOrNull && SrcGIEquivOrNull->getValueAsBit("CheckMMOIsNonAtomic"))
InsnMatcher.addPredicate<AtomicOrderingMMOPredicateMatcher>("NotAtomic");
Re-commit: [globalisel][tablegen] Add support for ImmLeaf without SDNodeXForm Summary: This patch adds support for predicates on imm nodes but only for ImmLeaf and not for PatLeaf or PatFrag and only where the value does not need to be transformed before being rendered into the instruction. The limitation on PatLeaf/PatFrag/SDNodeXForm is due to differences in the necessary target-supplied C++ for GlobalISel. Depends on D36085 The previous commit was reverted for breaking the build but this appears to have been the recurring problem on the Windows bots with tablegen not being re-run when llvm-tblgen is changed but the .td's aren't. If it re-occurs then forcing a build with clean=True should fix it but this string should do this in advance: Requires a clean build. Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Reviewed By: rovka Subscribers: kristof.beyls, javed.absar, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D36086 llvm-svn: 311645
2017-08-24 17:11:20 +08:00
[globalisel][tablegen] Add support for (set $dst, 1) and test X86's OptForSize predicate. Summary: It's rare but a small number of patterns use IntInit's at the root of the match. On X86, one such rule is enabled by the OptForSize predicate and causes the compiler to use the smaller: %0 = MOV32r1 instead of the usual: %0 = MOV32ri 1 This patch adds support for matching IntInit's at the root and uses this as a test case for the optsize attribute that was implemented in r301750 Reviewers: qcolombet, ab, t.p.northover, rovka, kristof.beyls, aditya_nandakumar Reviewed By: qcolombet Subscribers: igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D32791 llvm-svn: 303678
2017-05-24 03:33:16 +08:00
if (Src->isLeaf()) {
Init *SrcInit = Src->getLeafValue();
if (IntInit *SrcIntInit = dyn_cast<IntInit>(SrcInit)) {
[globalisel][tablegen] Simplify named operand/operator lookups and fix a wrong-code bug this revealed. Summary: Operand variable lookups are now performed by the RuleMatcher rather than searching the whole matcher hierarchy for a match. This revealed a wrong-code bug that currently affects ARM and X86 where patterns that use a variable more than once in the match pattern will be imported but won't check that the operands are identical. This can cause the tablegen-erated matcher to accept matches that should be rejected. Depends on D36569 Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Subscribers: aemerson, igorb, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D36618 llvm-svn: 315780
2017-10-14 08:31:58 +08:00
OperandMatcher &OM =
InsnMatcher.addOperand(OpIdx++, Src->getName(), TempOpIdx);
[globalisel][tablegen] Add support for (set $dst, 1) and test X86's OptForSize predicate. Summary: It's rare but a small number of patterns use IntInit's at the root of the match. On X86, one such rule is enabled by the OptForSize predicate and causes the compiler to use the smaller: %0 = MOV32r1 instead of the usual: %0 = MOV32ri 1 This patch adds support for matching IntInit's at the root and uses this as a test case for the optsize attribute that was implemented in r301750 Reviewers: qcolombet, ab, t.p.northover, rovka, kristof.beyls, aditya_nandakumar Reviewed By: qcolombet Subscribers: igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D32791 llvm-svn: 303678
2017-05-24 03:33:16 +08:00
OM.addPredicate<LiteralIntOperandMatcher>(SrcIntInit->getValue());
} else
[globalisel][tablegen] Multiple 80-col corrections. llvm-svn: 306544
2017-06-28 21:50:04 +08:00
return failedImport(
"Unable to deduce gMIR opcode to handle Src (which is a leaf)");
[globalisel][tablegen] Add support for (set $dst, 1) and test X86's OptForSize predicate. Summary: It's rare but a small number of patterns use IntInit's at the root of the match. On X86, one such rule is enabled by the OptForSize predicate and causes the compiler to use the smaller: %0 = MOV32r1 instead of the usual: %0 = MOV32ri 1 This patch adds support for matching IntInit's at the root and uses this as a test case for the optsize attribute that was implemented in r301750 Reviewers: qcolombet, ab, t.p.northover, rovka, kristof.beyls, aditya_nandakumar Reviewed By: qcolombet Subscribers: igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D32791 llvm-svn: 303678
2017-05-24 03:33:16 +08:00
} else {
[globalisel][tablegen] Import rules containing intrinsic_wo_chain. Summary: As of this patch, 1018 out of 3938 rules are currently imported. Depends on D32275 Reviewers: qcolombet, kristof.beyls, rovka, t.p.northover, ab, aditya_nandakumar Reviewed By: qcolombet Subscribers: dberris, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D32278 llvm-svn: 307240
2017-07-06 16:12:20 +08:00
assert(SrcGIOrNull &&
"Expected to have already found an equivalent Instruction");
[globalisel][tablegen] Add support for fpimm and import of APInt/APFloat based ImmLeaf. Summary: There's only a tablegen testcase for IntImmLeaf and not a CodeGen one because the relevant rules are rejected for other reasons at the moment. On AArch64, it's because there's an SDNodeXForm attached to the operand. On X86, it's because the rule either emits multiple instructions or has another predicate using PatFrag which cannot easily be supported at the same time. Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Reviewed By: qcolombet Subscribers: aemerson, javed.absar, igorb, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D36569 llvm-svn: 315761
2017-10-14 05:28:03 +08:00
if (SrcGIOrNull->TheDef->getName() == "G_CONSTANT" ||
SrcGIOrNull->TheDef->getName() == "G_FCONSTANT") {
// imm/fpimm still have operands but we don't need to do anything with it
[globalisel][tablegen] Add support for importing 'imm' operands. Summary: This patch enables the import of rules containing 'imm' operands that do not constrain the acceptable values using predicates. Support for ImmLeaf will arrive in a later patch. Depends on D35681 Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Reviewed By: rovka Subscribers: kristof.beyls, javed.absar, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D35833 llvm-svn: 310343
2017-08-08 18:44:31 +08:00
// here since we don't support ImmLeaf predicates yet. However, we still
// need to note the hidden operand to get GIM_CheckNumOperands correct.
InsnMatcher.addOperand(OpIdx++, "", TempOpIdx);
return InsnMatcher;
}
[globalisel][tablegen] Add support for (set $dst, 1) and test X86's OptForSize predicate. Summary: It's rare but a small number of patterns use IntInit's at the root of the match. On X86, one such rule is enabled by the OptForSize predicate and causes the compiler to use the smaller: %0 = MOV32r1 instead of the usual: %0 = MOV32ri 1 This patch adds support for matching IntInit's at the root and uses this as a test case for the optsize attribute that was implemented in r301750 Reviewers: qcolombet, ab, t.p.northover, rovka, kristof.beyls, aditya_nandakumar Reviewed By: qcolombet Subscribers: igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D32791 llvm-svn: 303678
2017-05-24 03:33:16 +08:00
// Match the used operands (i.e. the children of the operator).
for (unsigned i = 0, e = Src->getNumChildren(); i != e; ++i) {
[globalisel][tablegen] Import rules containing intrinsic_wo_chain. Summary: As of this patch, 1018 out of 3938 rules are currently imported. Depends on D32275 Reviewers: qcolombet, kristof.beyls, rovka, t.p.northover, ab, aditya_nandakumar Reviewed By: qcolombet Subscribers: dberris, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D32278 llvm-svn: 307240
2017-07-06 16:12:20 +08:00
TreePatternNode *SrcChild = Src->getChild(i);
[globalisel][tablegen] Implement unindexed load, non-extending load, and MemVT checks Summary: This includes some context-sensitivity in the MVT to LLT conversion so that pointer types are tested correctly. FIXME: I'm not happy with the way this is done since everything is a special-case. I've yet to find a reasonable way to implement it. select-load.mir fails because <1 x s64> loads in tablegen get priority over s64 loads. This is fixed in the next patch and as such they should be committed together, I've posted them separately to help with the review. Depends on D37456 Reviewers: ab, qcolombet, t.p.northover, rovka, aditya_nandakumar Subscribers: kristof.beyls, javed.absar, llvm-commits, igorb Differential Revision: https://reviews.llvm.org/D37457 llvm-svn: 315884
2017-10-16 08:56:30 +08:00
// SelectionDAG allows pointers to be represented with iN since it doesn't
// distinguish between pointers and integers but they are different types in GlobalISel.
// Coerce integers to pointers to address space 0 if the context indicates a pointer.
[globalisel][tablegen] Generalize pointer-type inference by introducing ptypeN. NFC ptypeN is functionally the same as typeN except that it informs the SelectionDAG importer that an operand should be treated as a pointer even if it was written as iN. This is important for patterns that use iN instead of iPTR to represent pointers. E.g.: (set GPR64:$dst, (load GPR64:$addr)) Previously, this was handled as a hardcoded special case for the appropriate operands to G_LOAD and G_STORE. llvm-svn: 318574
2017-11-18 08:16:44 +08:00
bool OperandIsAPointer = SrcGIOrNull->isOperandAPointer(i);
[globalisel][tablegen] Implement unindexed load, non-extending load, and MemVT checks Summary: This includes some context-sensitivity in the MVT to LLT conversion so that pointer types are tested correctly. FIXME: I'm not happy with the way this is done since everything is a special-case. I've yet to find a reasonable way to implement it. select-load.mir fails because <1 x s64> loads in tablegen get priority over s64 loads. This is fixed in the next patch and as such they should be committed together, I've posted them separately to help with the review. Depends on D37456 Reviewers: ab, qcolombet, t.p.northover, rovka, aditya_nandakumar Subscribers: kristof.beyls, javed.absar, llvm-commits, igorb Differential Revision: https://reviews.llvm.org/D37457 llvm-svn: 315884
2017-10-16 08:56:30 +08:00
[globalisel] Add support for intrinsic_w_chain. This maps directly to G_INTRINSIC_W_SIDE_EFFECTS. llvm-svn: 313627
2017-09-19 20:56:36 +08:00
// For G_INTRINSIC/G_INTRINSIC_W_SIDE_EFFECTS, the operand immediately
// following the defs is an intrinsic ID.
if ((SrcGIOrNull->TheDef->getName() == "G_INTRINSIC" ||
SrcGIOrNull->TheDef->getName() == "G_INTRINSIC_W_SIDE_EFFECTS") &&
i == 0) {
[globalisel][tablegen] Correct matching of intrinsic ID's. TreePatternNode considers them to be plain integers but MachineInstr considers them to be a distinct kind of operand. The tweak to AArch64InstrInfo.td to produce a simple test case is a NFC for everything except GlobalISelEmitter (confirmed by diffing the tablegenerated files). GlobalISelEmitter is currently unable to infer the type of operands in the Dst pattern from the operands in the Src pattern. llvm-svn: 307634
2017-07-11 16:57:29 +08:00
if (const CodeGenIntrinsic *II = Src->getIntrinsicInfo(CGP)) {
[globalisel][tablegen] Import rules containing intrinsic_wo_chain. Summary: As of this patch, 1018 out of 3938 rules are currently imported. Depends on D32275 Reviewers: qcolombet, kristof.beyls, rovka, t.p.northover, ab, aditya_nandakumar Reviewed By: qcolombet Subscribers: dberris, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D32278 llvm-svn: 307240
2017-07-06 16:12:20 +08:00
OperandMatcher &OM =
InsnMatcher.addOperand(OpIdx++, SrcChild->getName(), TempOpIdx);
[globalisel][tablegen] Correct matching of intrinsic ID's. TreePatternNode considers them to be plain integers but MachineInstr considers them to be a distinct kind of operand. The tweak to AArch64InstrInfo.td to produce a simple test case is a NFC for everything except GlobalISelEmitter (confirmed by diffing the tablegenerated files). GlobalISelEmitter is currently unable to infer the type of operands in the Dst pattern from the operands in the Src pattern. llvm-svn: 307634
2017-07-11 16:57:29 +08:00
OM.addPredicate<IntrinsicIDOperandMatcher>(II);
[globalisel][tablegen] Import rules containing intrinsic_wo_chain. Summary: As of this patch, 1018 out of 3938 rules are currently imported. Depends on D32275 Reviewers: qcolombet, kristof.beyls, rovka, t.p.northover, ab, aditya_nandakumar Reviewed By: qcolombet Subscribers: dberris, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D32278 llvm-svn: 307240
2017-07-06 16:12:20 +08:00
continue;
}
return failedImport("Expected IntInit containing instrinsic ID)");
}
[globalisel][tablegen] Implement unindexed load, non-extending load, and MemVT checks Summary: This includes some context-sensitivity in the MVT to LLT conversion so that pointer types are tested correctly. FIXME: I'm not happy with the way this is done since everything is a special-case. I've yet to find a reasonable way to implement it. select-load.mir fails because <1 x s64> loads in tablegen get priority over s64 loads. This is fixed in the next patch and as such they should be committed together, I've posted them separately to help with the review. Depends on D37456 Reviewers: ab, qcolombet, t.p.northover, rovka, aditya_nandakumar Subscribers: kristof.beyls, javed.absar, llvm-commits, igorb Differential Revision: https://reviews.llvm.org/D37457 llvm-svn: 315884
2017-10-16 08:56:30 +08:00
if (auto Error =
importChildMatcher(Rule, InsnMatcher, SrcChild, OperandIsAPointer,
OpIdx++, TempOpIdx))
[globalisel][tablegen] Add support for (set $dst, 1) and test X86's OptForSize predicate. Summary: It's rare but a small number of patterns use IntInit's at the root of the match. On X86, one such rule is enabled by the OptForSize predicate and causes the compiler to use the smaller: %0 = MOV32r1 instead of the usual: %0 = MOV32ri 1 This patch adds support for matching IntInit's at the root and uses this as a test case for the optsize attribute that was implemented in r301750 Reviewers: qcolombet, ab, t.p.northover, rovka, kristof.beyls, aditya_nandakumar Reviewed By: qcolombet Subscribers: igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D32791 llvm-svn: 303678
2017-05-24 03:33:16 +08:00
return std::move(Error);
}
[tablegen][globalisel] Convert the SelectionDAG importer to a tree walking approach. NFC Summary: But don't actually inspect the tree any deeper than we already do. This change is NFC but the next one will enable full traversal of the source/destination patterns. Depends on D30535 Reviewers: t.p.northover, qcolombet, aditya_nandakumar, rovka, ab Subscribers: igorb, dberris, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D30536 llvm-svn: 299001
2017-03-29 23:37:18 +08:00
}
return InsnMatcher;
}
Re-commit r315863: [globalisel][tablegen] Import ComplexPattern when used as an operator Summary: It's possible for a ComplexPattern to be used as an operator in a match pattern. This is used by the load/store patterns in AArch64 to name the suboperands returned by ComplexPattern predicate so that they can be broken apart and referenced independently in the result pattern. This patch adds support for this in order to enable the import of load/store patterns. Depends on D37445 Hopefully fixed the ambiguous constructor that a large number of bots reported. Reviewers: ab, qcolombet, t.p.northover, rovka, aditya_nandakumar Reviewed By: qcolombet Subscribers: aemerson, javed.absar, igorb, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D37456 llvm-svn: 315869
2017-10-16 02:22:54 +08:00
Error GlobalISelEmitter::importComplexPatternOperandMatcher(
OperandMatcher &OM, Record *R, unsigned &TempOpIdx) const {
const auto &ComplexPattern = ComplexPatternEquivs.find(R);
if (ComplexPattern == ComplexPatternEquivs.end())
return failedImport("SelectionDAG ComplexPattern (" + R->getName() +
") not mapped to GlobalISel");
OM.addPredicate<ComplexPatternOperandMatcher>(OM, *ComplexPattern->second);
TempOpIdx++;
return Error::success();
}
Error GlobalISelEmitter::importChildMatcher(RuleMatcher &Rule,
InstructionMatcher &InsnMatcher,
[globalisel][tablegen] Add support for (set $dst, 1) and test X86's OptForSize predicate. Summary: It's rare but a small number of patterns use IntInit's at the root of the match. On X86, one such rule is enabled by the OptForSize predicate and causes the compiler to use the smaller: %0 = MOV32r1 instead of the usual: %0 = MOV32ri 1 This patch adds support for matching IntInit's at the root and uses this as a test case for the optsize attribute that was implemented in r301750 Reviewers: qcolombet, ab, t.p.northover, rovka, kristof.beyls, aditya_nandakumar Reviewed By: qcolombet Subscribers: igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D32791 llvm-svn: 303678
2017-05-24 03:33:16 +08:00
const TreePatternNode *SrcChild,
[globalisel][tablegen] Implement unindexed load, non-extending load, and MemVT checks Summary: This includes some context-sensitivity in the MVT to LLT conversion so that pointer types are tested correctly. FIXME: I'm not happy with the way this is done since everything is a special-case. I've yet to find a reasonable way to implement it. select-load.mir fails because <1 x s64> loads in tablegen get priority over s64 loads. This is fixed in the next patch and as such they should be committed together, I've posted them separately to help with the review. Depends on D37456 Reviewers: ab, qcolombet, t.p.northover, rovka, aditya_nandakumar Subscribers: kristof.beyls, javed.absar, llvm-commits, igorb Differential Revision: https://reviews.llvm.org/D37457 llvm-svn: 315884
2017-10-16 08:56:30 +08:00
bool OperandIsAPointer,
[globalisel][tablegen] Change Expected<bool> to Error and rename functions. Functions that still return Expected<X> are now called createAndImport*() Changing the return type was requested in the review comments for r299001 llvm-svn: 299063
2017-03-30 17:36:33 +08:00
unsigned OpIdx,
unsigned &TempOpIdx) const {
[globalisel][tablegen] Fix patterns involving multiple ComplexPatterns. Summary: Temporaries are now allocated to operands instead of predicates and this allocation is used to correctly pair up the rendered operands with the matched operands. Previously, ComplexPatterns were allocated temporaries independently in the Src Pattern and Dst Pattern, leading to mismatches. Additionally, the Dst Pattern failed to account for the allocated index and therefore always used temporary 0, 1, ... when it should have used base+0, base+1, ... Thanks to Aditya Nandakumar for noticing the bug. Depends on D30539 Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Reviewed By: rovka Subscribers: igorb, dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D31054 llvm-svn: 299538
2017-04-05 21:14:03 +08:00
OperandMatcher &OM =
InsnMatcher.addOperand(OpIdx, SrcChild->getName(), TempOpIdx);
[globalisel][tablegen] Simplify named operand/operator lookups and fix a wrong-code bug this revealed. Summary: Operand variable lookups are now performed by the RuleMatcher rather than searching the whole matcher hierarchy for a match. This revealed a wrong-code bug that currently affects ARM and X86 where patterns that use a variable more than once in the match pattern will be imported but won't check that the operands are identical. This can cause the tablegen-erated matcher to accept matches that should be rejected. Depends on D36569 Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Subscribers: aemerson, igorb, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D36618 llvm-svn: 315780
2017-10-14 08:31:58 +08:00
if (OM.isSameAsAnotherOperand())
return Error::success();
[tablegen][globalisel] Convert the SelectionDAG importer to a tree walking approach. NFC Summary: But don't actually inspect the tree any deeper than we already do. This change is NFC but the next one will enable full traversal of the source/destination patterns. Depends on D30535 Reviewers: t.p.northover, qcolombet, aditya_nandakumar, rovka, ab Subscribers: igorb, dberris, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D30536 llvm-svn: 299001
2017-03-29 23:37:18 +08:00
TableGen support for parameterized register class information This replaces TableGen's type inference to operate on parameterized types instead of MVTs, and as a consequence, some interfaces have changed: - Uses of MVTs are replaced by ValueTypeByHwMode. - EEVT::TypeSet is replaced by TypeSetByHwMode. This affects the way that types and type sets are printed, and the tests relying on that have been updated. There are certain users of the inferred types outside of TableGen itself, namely FastISel and GlobalISel. For those users, the way that the types are accessed have changed. For typical scenarios, these replacements can be used: - TreePatternNode::getType(ResNo) -> getSimpleType(ResNo) - TreePatternNode::hasTypeSet(ResNo) -> hasConcreteType(ResNo) - TypeSet::isConcrete -> TypeSetByHwMode::isValueTypeByHwMode(false) For more information, please refer to the review page. Differential Revision: https://reviews.llvm.org/D31951 llvm-svn: 313271
2017-09-15 00:56:21 +08:00
ArrayRef<TypeSetByHwMode> ChildTypes = SrcChild->getExtTypes();
[tablegen][globalisel] Convert the SelectionDAG importer to a tree walking approach. NFC Summary: But don't actually inspect the tree any deeper than we already do. This change is NFC but the next one will enable full traversal of the source/destination patterns. Depends on D30535 Reviewers: t.p.northover, qcolombet, aditya_nandakumar, rovka, ab Subscribers: igorb, dberris, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D30536 llvm-svn: 299001
2017-03-29 23:37:18 +08:00
if (ChildTypes.size() != 1)
return failedImport("Src pattern child has multiple results");
// Check MBB's before the type check since they are not a known type.
if (!SrcChild->isLeaf()) {
if (SrcChild->getOperator()->isSubClassOf("SDNode")) {
auto &ChildSDNI = CGP.getSDNodeInfo(SrcChild->getOperator());
if (ChildSDNI.getSDClassName() == "BasicBlockSDNode") {
OM.addPredicate<MBBOperandMatcher>();
[globalisel][tablegen] Change Expected<bool> to Error and rename functions. Functions that still return Expected<X> are now called createAndImport*() Changing the return type was requested in the review comments for r299001 llvm-svn: 299063
2017-03-30 17:36:33 +08:00
return Error::success();
[tablegen][globalisel] Convert the SelectionDAG importer to a tree walking approach. NFC Summary: But don't actually inspect the tree any deeper than we already do. This change is NFC but the next one will enable full traversal of the source/destination patterns. Depends on D30535 Reviewers: t.p.northover, qcolombet, aditya_nandakumar, rovka, ab Subscribers: igorb, dberris, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D30536 llvm-svn: 299001
2017-03-29 23:37:18 +08:00
}
}
}
[globalisel][tablegen] Implement unindexed load, non-extending load, and MemVT checks Summary: This includes some context-sensitivity in the MVT to LLT conversion so that pointer types are tested correctly. FIXME: I'm not happy with the way this is done since everything is a special-case. I've yet to find a reasonable way to implement it. select-load.mir fails because <1 x s64> loads in tablegen get priority over s64 loads. This is fixed in the next patch and as such they should be committed together, I've posted them separately to help with the review. Depends on D37456 Reviewers: ab, qcolombet, t.p.northover, rovka, aditya_nandakumar Subscribers: kristof.beyls, javed.absar, llvm-commits, igorb Differential Revision: https://reviews.llvm.org/D37457 llvm-svn: 315884
2017-10-16 08:56:30 +08:00
if (auto Error =
OM.addTypeCheckPredicate(ChildTypes.front(), OperandIsAPointer))
return failedImport(toString(std::move(Error)) + " for Src operand (" +
to_string(*SrcChild) + ")");
[tablegen][globalisel] Convert the SelectionDAG importer to a tree walking approach. NFC Summary: But don't actually inspect the tree any deeper than we already do. This change is NFC but the next one will enable full traversal of the source/destination patterns. Depends on D30535 Reviewers: t.p.northover, qcolombet, aditya_nandakumar, rovka, ab Subscribers: igorb, dberris, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D30536 llvm-svn: 299001
2017-03-29 23:37:18 +08:00
[tablegen][globalisel] Add support for nested instruction matching. Summary: Lift the restrictions that prevented the tree walking introduced in the previous change and add support for patterns like: (G_ADD (G_MUL (G_SEXT $src1), (G_SEXT $src2)), $src3) -> SMADDWrrr $dst, $src1, $src2, $src3 Also adds support for G_SEXT and G_ZEXT to support these cases. One particular aspect of this that I should draw attention to is that I've tried to be overly conservative in determining the safety of matches that involve non-adjacent instructions and multiple basic blocks. This is intended to be used as a cheap initial check and we may add a more expensive check in the future. The current rules are: * Reject if any instruction may load/store (we'd need to check for intervening memory operations. * Reject if any instruction has implicit operands. * Reject if any instruction has unmodelled side-effects. See isObviouslySafeToFold(). Reviewers: t.p.northover, javed.absar, qcolombet, aditya_nandakumar, ab, rovka Reviewed By: ab Subscribers: igorb, dberris, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D30539 llvm-svn: 299430
2017-04-04 21:25:23 +08:00
// Check for nested instructions.
if (!SrcChild->isLeaf()) {
Re-commit r315863: [globalisel][tablegen] Import ComplexPattern when used as an operator Summary: It's possible for a ComplexPattern to be used as an operator in a match pattern. This is used by the load/store patterns in AArch64 to name the suboperands returned by ComplexPattern predicate so that they can be broken apart and referenced independently in the result pattern. This patch adds support for this in order to enable the import of load/store patterns. Depends on D37445 Hopefully fixed the ambiguous constructor that a large number of bots reported. Reviewers: ab, qcolombet, t.p.northover, rovka, aditya_nandakumar Reviewed By: qcolombet Subscribers: aemerson, javed.absar, igorb, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D37456 llvm-svn: 315869
2017-10-16 02:22:54 +08:00
if (SrcChild->getOperator()->isSubClassOf("ComplexPattern")) {
// When a ComplexPattern is used as an operator, it should do the same
// thing as when used as a leaf. However, the children of the operator
// name the sub-operands that make up the complex operand and we must
// prepare to reference them in the renderer too.
unsigned RendererID = TempOpIdx;
if (auto Error = importComplexPatternOperandMatcher(
OM, SrcChild->getOperator(), TempOpIdx))
return Error;
for (unsigned i = 0, e = SrcChild->getNumChildren(); i != e; ++i) {
auto *SubOperand = SrcChild->getChild(i);
if (!SubOperand->getName().empty())
Rule.defineComplexSubOperand(SubOperand->getName(),
SrcChild->getOperator(), RendererID, i);
}
return Error::success();
}
[globalisel][tablegen] Simplify named operand/operator lookups and fix a wrong-code bug this revealed. Summary: Operand variable lookups are now performed by the RuleMatcher rather than searching the whole matcher hierarchy for a match. This revealed a wrong-code bug that currently affects ARM and X86 where patterns that use a variable more than once in the match pattern will be imported but won't check that the operands are identical. This can cause the tablegen-erated matcher to accept matches that should be rejected. Depends on D36569 Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Subscribers: aemerson, igorb, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D36618 llvm-svn: 315780
2017-10-14 08:31:58 +08:00
auto MaybeInsnOperand = OM.addPredicate<InstructionOperandMatcher>(
InsnMatcher.getRuleMatcher(), SrcChild->getName());
if (!MaybeInsnOperand.hasValue()) {
// This isn't strictly true. If the user were to provide exactly the same
// matchers as the original operand then we could allow it. However, it's
// simpler to not permit the redundant specification.
return failedImport("Nested instruction cannot be the same as another operand");
}
[tablegen][globalisel] Add support for nested instruction matching. Summary: Lift the restrictions that prevented the tree walking introduced in the previous change and add support for patterns like: (G_ADD (G_MUL (G_SEXT $src1), (G_SEXT $src2)), $src3) -> SMADDWrrr $dst, $src1, $src2, $src3 Also adds support for G_SEXT and G_ZEXT to support these cases. One particular aspect of this that I should draw attention to is that I've tried to be overly conservative in determining the safety of matches that involve non-adjacent instructions and multiple basic blocks. This is intended to be used as a cheap initial check and we may add a more expensive check in the future. The current rules are: * Reject if any instruction may load/store (we'd need to check for intervening memory operations. * Reject if any instruction has implicit operands. * Reject if any instruction has unmodelled side-effects. See isObviouslySafeToFold(). Reviewers: t.p.northover, javed.absar, qcolombet, aditya_nandakumar, ab, rovka Reviewed By: ab Subscribers: igorb, dberris, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D30539 llvm-svn: 299430
2017-04-04 21:25:23 +08:00
// Map the node to a gMIR instruction.
[globalisel][tablegen] Simplify named operand/operator lookups and fix a wrong-code bug this revealed. Summary: Operand variable lookups are now performed by the RuleMatcher rather than searching the whole matcher hierarchy for a match. This revealed a wrong-code bug that currently affects ARM and X86 where patterns that use a variable more than once in the match pattern will be imported but won't check that the operands are identical. This can cause the tablegen-erated matcher to accept matches that should be rejected. Depends on D36569 Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Subscribers: aemerson, igorb, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D36618 llvm-svn: 315780
2017-10-14 08:31:58 +08:00
InstructionOperandMatcher &InsnOperand = **MaybeInsnOperand;
[globalisel][tablegen] Fix an multi-insn match bug where ComplexPattern is used on multiple insns. In each rule, each use of ComplexPattern is assigned an element in the Renderers array. The matcher then collects renderer functions in this array and they are used to render instructions. This works well for a single instruction but a bug in the allocation mechanism causes the elements to be assigned on a per-instruction basis rather than a per-rule basis. So in the case of: (set GPR32:$dst, (Op complex:$src1, complex:$src2)) tablegen currently assigns elements 0 and 1 to $src1 and $src2 respectively, but for: (set GPR32:$dst, (Op complex:$src1, (Op complex:$src2))) it currently assigned both $src1 and $src2 the same element (0). This results in one complex operand being rendered twice and the other being forgotten. This patch corrects the allocation such that $src1 and $src2 are still allocated different elements in this case. llvm-svn: 307646
2017-07-11 18:40:18 +08:00
auto InsnMatcherOrError = createAndImportSelDAGMatcher(
Re-commit r315863: [globalisel][tablegen] Import ComplexPattern when used as an operator Summary: It's possible for a ComplexPattern to be used as an operator in a match pattern. This is used by the load/store patterns in AArch64 to name the suboperands returned by ComplexPattern predicate so that they can be broken apart and referenced independently in the result pattern. This patch adds support for this in order to enable the import of load/store patterns. Depends on D37445 Hopefully fixed the ambiguous constructor that a large number of bots reported. Reviewers: ab, qcolombet, t.p.northover, rovka, aditya_nandakumar Reviewed By: qcolombet Subscribers: aemerson, javed.absar, igorb, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D37456 llvm-svn: 315869
2017-10-16 02:22:54 +08:00
Rule, InsnOperand.getInsnMatcher(), SrcChild, TempOpIdx);
[tablegen][globalisel] Add support for nested instruction matching. Summary: Lift the restrictions that prevented the tree walking introduced in the previous change and add support for patterns like: (G_ADD (G_MUL (G_SEXT $src1), (G_SEXT $src2)), $src3) -> SMADDWrrr $dst, $src1, $src2, $src3 Also adds support for G_SEXT and G_ZEXT to support these cases. One particular aspect of this that I should draw attention to is that I've tried to be overly conservative in determining the safety of matches that involve non-adjacent instructions and multiple basic blocks. This is intended to be used as a cheap initial check and we may add a more expensive check in the future. The current rules are: * Reject if any instruction may load/store (we'd need to check for intervening memory operations. * Reject if any instruction has implicit operands. * Reject if any instruction has unmodelled side-effects. See isObviouslySafeToFold(). Reviewers: t.p.northover, javed.absar, qcolombet, aditya_nandakumar, ab, rovka Reviewed By: ab Subscribers: igorb, dberris, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D30539 llvm-svn: 299430
2017-04-04 21:25:23 +08:00
if (auto Error = InsnMatcherOrError.takeError())
return Error;
return Error::success();
}
[globalisel][tablegen] Skip src child predicates The GlobalISel TableGen backend didn't check for predicates on the source children. This caused it to generate code for ARM patterns such as SMLABB or similar, but without properly checking for the sext_16_node part of the operands. This in turn meant that we would select SMLABB instead of MLA for simple sequences such as s32 + s32 * s32, which is wrong (we want a MLA on the full operands, not just their bottom 16 bits). This patch forces TableGen to skip patterns with predicates on the src children, so it doesn't generate code for SMLABB and other similar ARM instructions at all anymore. AArch64 and X86 are not affected. Differential Revision: https://reviews.llvm.org/D39554 llvm-svn: 317313
2017-11-03 18:30:19 +08:00
if (SrcChild->hasAnyPredicate())
return failedImport("Src pattern child has unsupported predicate");
[tablegen][globalisel] Convert the SelectionDAG importer to a tree walking approach. NFC Summary: But don't actually inspect the tree any deeper than we already do. This change is NFC but the next one will enable full traversal of the source/destination patterns. Depends on D30535 Reviewers: t.p.northover, qcolombet, aditya_nandakumar, rovka, ab Subscribers: igorb, dberris, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D30536 llvm-svn: 299001
2017-03-29 23:37:18 +08:00
// Check for constant immediates.
if (auto *ChildInt = dyn_cast<IntInit>(SrcChild->getLeafValue())) {
[globalisel][tablegen] Add support for (set $dst, 1) and test X86's OptForSize predicate. Summary: It's rare but a small number of patterns use IntInit's at the root of the match. On X86, one such rule is enabled by the OptForSize predicate and causes the compiler to use the smaller: %0 = MOV32r1 instead of the usual: %0 = MOV32ri 1 This patch adds support for matching IntInit's at the root and uses this as a test case for the optsize attribute that was implemented in r301750 Reviewers: qcolombet, ab, t.p.northover, rovka, kristof.beyls, aditya_nandakumar Reviewed By: qcolombet Subscribers: igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D32791 llvm-svn: 303678
2017-05-24 03:33:16 +08:00
OM.addPredicate<ConstantIntOperandMatcher>(ChildInt->getValue());
[globalisel][tablegen] Change Expected<bool> to Error and rename functions. Functions that still return Expected<X> are now called createAndImport*() Changing the return type was requested in the review comments for r299001 llvm-svn: 299063
2017-03-30 17:36:33 +08:00
return Error::success();
[tablegen][globalisel] Convert the SelectionDAG importer to a tree walking approach. NFC Summary: But don't actually inspect the tree any deeper than we already do. This change is NFC but the next one will enable full traversal of the source/destination patterns. Depends on D30535 Reviewers: t.p.northover, qcolombet, aditya_nandakumar, rovka, ab Subscribers: igorb, dberris, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D30536 llvm-svn: 299001
2017-03-29 23:37:18 +08:00
}
// Check for def's like register classes or ComplexPattern's.
if (auto *ChildDefInit = dyn_cast<DefInit>(SrcChild->getLeafValue())) {
auto *ChildRec = ChildDefInit->getDef();
// Check for register classes.
[globalisel][tablegen] Add support for COPY_TO_REGCLASS. Summary: As part of this * Emitted instructions now have named MachineInstr variables associated with them. This isn't particularly important yet but it's a small step towards multiple-insn emission. * constrainSelectedInstRegOperands() is no longer hardcoded. It's now added as the ConstrainOperandsToDefinitionAction() action. COPY_TO_REGCLASS uses an alternate constraint mechanism ConstrainOperandToRegClassAction() which supports arbitrary constraints such as that defined by COPY_TO_REGCLASS. Reviewers: ab, qcolombet, t.p.northover, rovka, kristof.beyls, aditya_nandakumar Reviewed By: ab Subscribers: javed.absar, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D33590 llvm-svn: 305791
2017-06-20 20:36:34 +08:00
if (ChildRec->isSubClassOf("RegisterClass") ||
ChildRec->isSubClassOf("RegisterOperand")) {
[globalisel][tablegen] Add support for RegisterOperand. Summary: It functions just like RegisterClass except that the class is obtained from a field. Depends on D31761. Reviewers: ab, qcolombet, t.p.northover, rovka, kristof.beyls, aditya_nandakumar Reviewed By: ab Subscribers: dberris, llvm-commits, igorb Differential Revision: https://reviews.llvm.org/D32229 llvm-svn: 301080
2017-04-22 23:53:21 +08:00
OM.addPredicate<RegisterBankOperandMatcher>(
[globalisel][tablegen] Add support for COPY_TO_REGCLASS. Summary: As part of this * Emitted instructions now have named MachineInstr variables associated with them. This isn't particularly important yet but it's a small step towards multiple-insn emission. * constrainSelectedInstRegOperands() is no longer hardcoded. It's now added as the ConstrainOperandsToDefinitionAction() action. COPY_TO_REGCLASS uses an alternate constraint mechanism ConstrainOperandToRegClassAction() which supports arbitrary constraints such as that defined by COPY_TO_REGCLASS. Reviewers: ab, qcolombet, t.p.northover, rovka, kristof.beyls, aditya_nandakumar Reviewed By: ab Subscribers: javed.absar, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D33590 llvm-svn: 305791
2017-06-20 20:36:34 +08:00
Target.getRegisterClass(getInitValueAsRegClass(ChildDefInit)));
[globalisel][tablegen] Add support for RegisterOperand. Summary: It functions just like RegisterClass except that the class is obtained from a field. Depends on D31761. Reviewers: ab, qcolombet, t.p.northover, rovka, kristof.beyls, aditya_nandakumar Reviewed By: ab Subscribers: dberris, llvm-commits, igorb Differential Revision: https://reviews.llvm.org/D32229 llvm-svn: 301080
2017-04-22 23:53:21 +08:00
return Error::success();
}
[globalisel] Add support for ValueType operands in patterns. It's rare but there are a small number of patterns like this: (set i64:$dst, (add i64:$src1, i64:$src2)) These should be equivalent to register classes except they shouldn't check for a specific register bank. This doesn't occur in AArch64/ARM/X86 but does occasionally come up in other in-tree targets such as BPF. llvm-svn: 315226
2017-10-10 02:14:53 +08:00
// Check for ValueType.
if (ChildRec->isSubClassOf("ValueType")) {
// We already added a type check as standard practice so this doesn't need
// to do anything.
return Error::success();
}
[tablegen][globalisel] Convert the SelectionDAG importer to a tree walking approach. NFC Summary: But don't actually inspect the tree any deeper than we already do. This change is NFC but the next one will enable full traversal of the source/destination patterns. Depends on D30535 Reviewers: t.p.northover, qcolombet, aditya_nandakumar, rovka, ab Subscribers: igorb, dberris, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D30536 llvm-svn: 299001
2017-03-29 23:37:18 +08:00
// Check for ComplexPattern's.
Re-commit r315863: [globalisel][tablegen] Import ComplexPattern when used as an operator Summary: It's possible for a ComplexPattern to be used as an operator in a match pattern. This is used by the load/store patterns in AArch64 to name the suboperands returned by ComplexPattern predicate so that they can be broken apart and referenced independently in the result pattern. This patch adds support for this in order to enable the import of load/store patterns. Depends on D37445 Hopefully fixed the ambiguous constructor that a large number of bots reported. Reviewers: ab, qcolombet, t.p.northover, rovka, aditya_nandakumar Reviewed By: qcolombet Subscribers: aemerson, javed.absar, igorb, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D37456 llvm-svn: 315869
2017-10-16 02:22:54 +08:00
if (ChildRec->isSubClassOf("ComplexPattern"))
return importComplexPatternOperandMatcher(OM, ChildRec, TempOpIdx);
[tablegen][globalisel] Convert the SelectionDAG importer to a tree walking approach. NFC Summary: But don't actually inspect the tree any deeper than we already do. This change is NFC but the next one will enable full traversal of the source/destination patterns. Depends on D30535 Reviewers: t.p.northover, qcolombet, aditya_nandakumar, rovka, ab Subscribers: igorb, dberris, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D30536 llvm-svn: 299001
2017-03-29 23:37:18 +08:00
[globalisel][tablegen] Report more detail in some SelectionDAG import failures. NFC Reviewers: ab, t.p.northover, qcolombet, aditya_nandakumar, rovka Reviewed By: ab Subscribers: dberris, kristof.beyls, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D31325 llvm-svn: 300186
2017-04-13 17:45:37 +08:00
if (ChildRec->isSubClassOf("ImmLeaf")) {
return failedImport(
"Src pattern child def is an unsupported tablegen class (ImmLeaf)");
}
[tablegen][globalisel] Convert the SelectionDAG importer to a tree walking approach. NFC Summary: But don't actually inspect the tree any deeper than we already do. This change is NFC but the next one will enable full traversal of the source/destination patterns. Depends on D30535 Reviewers: t.p.northover, qcolombet, aditya_nandakumar, rovka, ab Subscribers: igorb, dberris, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D30536 llvm-svn: 299001
2017-03-29 23:37:18 +08:00
return failedImport(
"Src pattern child def is an unsupported tablegen class");
}
return failedImport("Src pattern child is an unsupported kind");
}
Re-commit: [globalisel][tablegen] Keep track of the insertion point while adding BuildMIAction's. NFC Multi-instruction emission needs to ensure the the instructions are generated a depth-first fashion. For example: (ADDWrr (SUBWrr a, b), c) needs to emit the SUBWrr before the ADDWrr. However, our walk over TreePatternNode's is highly context sensitive which makes it difficult to append BuildMIActions in the order we want. To fix this, we now keep track of the insertion point as we add actions. This will allow multi-insn emission to insert BuildMI's in the correct place. The previous commit failed on the Ubuntu bots using GCC 4.8. These bots lack the const_iterator forms of insert() and emplace() that were added in C++11. As a result I've switched the const_iterators to iterators. llvm-svn: 317049
2017-11-01 07:03:18 +08:00
Expected<action_iterator> GlobalISelEmitter::importExplicitUseRenderer(
action_iterator InsertPt, RuleMatcher &Rule, BuildMIAction &DstMIBuilder,
[globalisel][tablegen] Add support for multi-insn emission The importer will now accept nested instructions in the result pattern such as (ADDWrr $a, (SUBWrr $b, $c)). This is only valid when the nested instruction def's a single vreg and the parent instruction consumes a single vreg where a nested instruction is specified. The importer will automatically create a vreg to connect the two using the type information from the pattern. This vreg will be constrained to the register classes given in the instruction definitions*. * REG_SEQUENCE is explicitly rejected because of this. The definition doesn't constrain to a register class and it therefore needs special handling. llvm-svn: 317117
2017-11-02 03:57:57 +08:00
TreePatternNode *DstChild) {
Re-commit: [globalisel][tablegen] Add support for ImmLeaf without SDNodeXForm Summary: This patch adds support for predicates on imm nodes but only for ImmLeaf and not for PatLeaf or PatFrag and only where the value does not need to be transformed before being rendered into the instruction. The limitation on PatLeaf/PatFrag/SDNodeXForm is due to differences in the necessary target-supplied C++ for GlobalISel. Depends on D36085 The previous commit was reverted for breaking the build but this appears to have been the recurring problem on the Windows bots with tablegen not being re-run when llvm-tblgen is changed but the .td's aren't. If it re-occurs then forcing a build with clean=True should fix it but this string should do this in advance: Requires a clean build. Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Reviewed By: rovka Subscribers: kristof.beyls, javed.absar, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D36086 llvm-svn: 311645
2017-08-24 17:11:20 +08:00
Re-commit r315863: [globalisel][tablegen] Import ComplexPattern when used as an operator Summary: It's possible for a ComplexPattern to be used as an operator in a match pattern. This is used by the load/store patterns in AArch64 to name the suboperands returned by ComplexPattern predicate so that they can be broken apart and referenced independently in the result pattern. This patch adds support for this in order to enable the import of load/store patterns. Depends on D37445 Hopefully fixed the ambiguous constructor that a large number of bots reported. Reviewers: ab, qcolombet, t.p.northover, rovka, aditya_nandakumar Reviewed By: qcolombet Subscribers: aemerson, javed.absar, igorb, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D37456 llvm-svn: 315869
2017-10-16 02:22:54 +08:00
const auto &SubOperand = Rule.getComplexSubOperand(DstChild->getName());
if (SubOperand.hasValue()) {
DstMIBuilder.addRenderer<RenderComplexPatternOperand>(
[globalisel][tablegen] Stop hard-coding the emitted instruction ID to 0. NFC The next commit will add support for multi-instruction emission so we need to start allocating instruction ID's instead of hard-coding them to 0. llvm-svn: 317057
2017-11-01 08:29:47 +08:00
*std::get<0>(*SubOperand), DstChild->getName(),
Re-commit r315863: [globalisel][tablegen] Import ComplexPattern when used as an operator Summary: It's possible for a ComplexPattern to be used as an operator in a match pattern. This is used by the load/store patterns in AArch64 to name the suboperands returned by ComplexPattern predicate so that they can be broken apart and referenced independently in the result pattern. This patch adds support for this in order to enable the import of load/store patterns. Depends on D37445 Hopefully fixed the ambiguous constructor that a large number of bots reported. Reviewers: ab, qcolombet, t.p.northover, rovka, aditya_nandakumar Reviewed By: qcolombet Subscribers: aemerson, javed.absar, igorb, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D37456 llvm-svn: 315869
2017-10-16 02:22:54 +08:00
std::get<1>(*SubOperand), std::get<2>(*SubOperand));
Re-commit: [globalisel][tablegen] Keep track of the insertion point while adding BuildMIAction's. NFC Multi-instruction emission needs to ensure the the instructions are generated a depth-first fashion. For example: (ADDWrr (SUBWrr a, b), c) needs to emit the SUBWrr before the ADDWrr. However, our walk over TreePatternNode's is highly context sensitive which makes it difficult to append BuildMIActions in the order we want. To fix this, we now keep track of the insertion point as we add actions. This will allow multi-insn emission to insert BuildMI's in the correct place. The previous commit failed on the Ubuntu bots using GCC 4.8. These bots lack the const_iterator forms of insert() and emplace() that were added in C++11. As a result I've switched the const_iterators to iterators. llvm-svn: 317049
2017-11-01 07:03:18 +08:00
return InsertPt;
Re-commit r315863: [globalisel][tablegen] Import ComplexPattern when used as an operator Summary: It's possible for a ComplexPattern to be used as an operator in a match pattern. This is used by the load/store patterns in AArch64 to name the suboperands returned by ComplexPattern predicate so that they can be broken apart and referenced independently in the result pattern. This patch adds support for this in order to enable the import of load/store patterns. Depends on D37445 Hopefully fixed the ambiguous constructor that a large number of bots reported. Reviewers: ab, qcolombet, t.p.northover, rovka, aditya_nandakumar Reviewed By: qcolombet Subscribers: aemerson, javed.absar, igorb, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D37456 llvm-svn: 315869
2017-10-16 02:22:54 +08:00
}
[tablegen][globalisel] Convert the SelectionDAG importer to a tree walking approach. NFC Summary: But don't actually inspect the tree any deeper than we already do. This change is NFC but the next one will enable full traversal of the source/destination patterns. Depends on D30535 Reviewers: t.p.northover, qcolombet, aditya_nandakumar, rovka, ab Subscribers: igorb, dberris, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D30536 llvm-svn: 299001
2017-03-29 23:37:18 +08:00
if (!DstChild->isLeaf()) {
[GlobalISel][TableGen] Add support for SDNodeXForm Summary: This patch adds CustomRenderer which renders the matched operands to the specified instruction. Targets can enable the matching of SDNodeXForm by adding a definition that inherits from GICustomOperandRenderer and GISDNodeXFormEquiv as follows. def gi_imm8 : GICustomOperandRenderer<"renderImm8”>, GISDNodeXFormEquiv<imm8_xform>; Custom renderer functions should be of the form: void render(MachineInstrBuilder &MIB, const MachineInstr &I); Reviewers: dsanders, ab, rovka Reviewed By: dsanders Subscribers: kristof.beyls, javed.absar, llvm-commits, mgrang, qcolombet Differential Revision: https://reviews.llvm.org/D42012 llvm-svn: 322582
2018-01-17 02:44:05 +08:00
if (DstChild->getOperator()->isSubClassOf("SDNodeXForm")) {
auto Child = DstChild->getChild(0);
auto I = SDNodeXFormEquivs.find(DstChild->getOperator());
if (I != SDNodeXFormEquivs.end()) {
DstMIBuilder.addRenderer<CustomRenderer>(*I->second, Child->getName());
return InsertPt;
}
return failedImport("SDNodeXForm " + Child->getName() +
" has no custom renderer");
}
[globalisel][tablegen] Add support for importing 'imm' operands. Summary: This patch enables the import of rules containing 'imm' operands that do not constrain the acceptable values using predicates. Support for ImmLeaf will arrive in a later patch. Depends on D35681 Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Reviewed By: rovka Subscribers: kristof.beyls, javed.absar, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D35833 llvm-svn: 310343
2017-08-08 18:44:31 +08:00
// We accept 'bb' here. It's an operator because BasicBlockSDNode isn't
// inline, but in MI it's just another operand.
[tablegen][globalisel] Convert the SelectionDAG importer to a tree walking approach. NFC Summary: But don't actually inspect the tree any deeper than we already do. This change is NFC but the next one will enable full traversal of the source/destination patterns. Depends on D30535 Reviewers: t.p.northover, qcolombet, aditya_nandakumar, rovka, ab Subscribers: igorb, dberris, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D30536 llvm-svn: 299001
2017-03-29 23:37:18 +08:00
if (DstChild->getOperator()->isSubClassOf("SDNode")) {
auto &ChildSDNI = CGP.getSDNodeInfo(DstChild->getOperator());
if (ChildSDNI.getSDClassName() == "BasicBlockSDNode") {
[globalisel][tablegen] Stop hard-coding the emitted instruction ID to 0. NFC The next commit will add support for multi-instruction emission so we need to start allocating instruction ID's instead of hard-coding them to 0. llvm-svn: 317057
2017-11-01 08:29:47 +08:00
DstMIBuilder.addRenderer<CopyRenderer>(DstChild->getName());
Re-commit: [globalisel][tablegen] Keep track of the insertion point while adding BuildMIAction's. NFC Multi-instruction emission needs to ensure the the instructions are generated a depth-first fashion. For example: (ADDWrr (SUBWrr a, b), c) needs to emit the SUBWrr before the ADDWrr. However, our walk over TreePatternNode's is highly context sensitive which makes it difficult to append BuildMIActions in the order we want. To fix this, we now keep track of the insertion point as we add actions. This will allow multi-insn emission to insert BuildMI's in the correct place. The previous commit failed on the Ubuntu bots using GCC 4.8. These bots lack the const_iterator forms of insert() and emplace() that were added in C++11. As a result I've switched the const_iterators to iterators. llvm-svn: 317049
2017-11-01 07:03:18 +08:00
return InsertPt;
[GlobalISel] Add basic Selector-emitter tblgen backend. This adds a basic tablegen backend that analyzes the SelectionDAG patterns to find simple ones that are eligible for GlobalISel-emission. That's similar to FastISel, with one notable difference: we're not fed ISD opcodes, so we need to map the SDNode operators to generic opcodes. That's done using GINodeEquiv in TargetGlobalISel.td. Otherwise, this is mostly boilerplate, and lots of filtering of any kind of "complicated" pattern. On AArch64, this is sufficient to match G_ADD up to s64 (to ADDWrr/ADDXrr) and G_BR (to B). Differential Revision: https://reviews.llvm.org/D26878 llvm-svn: 290284
2016-12-22 07:26:20 +08:00
}
}
[globalisel][tablegen] Add support for importing 'imm' operands. Summary: This patch enables the import of rules containing 'imm' operands that do not constrain the acceptable values using predicates. Support for ImmLeaf will arrive in a later patch. Depends on D35681 Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Reviewed By: rovka Subscribers: kristof.beyls, javed.absar, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D35833 llvm-svn: 310343
2017-08-08 18:44:31 +08:00
// Similarly, imm is an operator in TreePatternNode's view but must be
// rendered as operands.
// FIXME: The target should be able to choose sign-extended when appropriate
// (e.g. on Mips).
if (DstChild->getOperator()->getName() == "imm") {
[globalisel][tablegen] Stop hard-coding the emitted instruction ID to 0. NFC The next commit will add support for multi-instruction emission so we need to start allocating instruction ID's instead of hard-coding them to 0. llvm-svn: 317057
2017-11-01 08:29:47 +08:00
DstMIBuilder.addRenderer<CopyConstantAsImmRenderer>(DstChild->getName());
Re-commit: [globalisel][tablegen] Keep track of the insertion point while adding BuildMIAction's. NFC Multi-instruction emission needs to ensure the the instructions are generated a depth-first fashion. For example: (ADDWrr (SUBWrr a, b), c) needs to emit the SUBWrr before the ADDWrr. However, our walk over TreePatternNode's is highly context sensitive which makes it difficult to append BuildMIActions in the order we want. To fix this, we now keep track of the insertion point as we add actions. This will allow multi-insn emission to insert BuildMI's in the correct place. The previous commit failed on the Ubuntu bots using GCC 4.8. These bots lack the const_iterator forms of insert() and emplace() that were added in C++11. As a result I've switched the const_iterators to iterators. llvm-svn: 317049
2017-11-01 07:03:18 +08:00
return InsertPt;
[globalisel][tablegen] Add support for fpimm and import of APInt/APFloat based ImmLeaf. Summary: There's only a tablegen testcase for IntImmLeaf and not a CodeGen one because the relevant rules are rejected for other reasons at the moment. On AArch64, it's because there's an SDNodeXForm attached to the operand. On X86, it's because the rule either emits multiple instructions or has another predicate using PatFrag which cannot easily be supported at the same time. Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Reviewed By: qcolombet Subscribers: aemerson, javed.absar, igorb, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D36569 llvm-svn: 315761
2017-10-14 05:28:03 +08:00
} else if (DstChild->getOperator()->getName() == "fpimm") {
DstMIBuilder.addRenderer<CopyFConstantAsFPImmRenderer>(
[globalisel][tablegen] Stop hard-coding the emitted instruction ID to 0. NFC The next commit will add support for multi-instruction emission so we need to start allocating instruction ID's instead of hard-coding them to 0. llvm-svn: 317057
2017-11-01 08:29:47 +08:00
DstChild->getName());
Re-commit: [globalisel][tablegen] Keep track of the insertion point while adding BuildMIAction's. NFC Multi-instruction emission needs to ensure the the instructions are generated a depth-first fashion. For example: (ADDWrr (SUBWrr a, b), c) needs to emit the SUBWrr before the ADDWrr. However, our walk over TreePatternNode's is highly context sensitive which makes it difficult to append BuildMIActions in the order we want. To fix this, we now keep track of the insertion point as we add actions. This will allow multi-insn emission to insert BuildMI's in the correct place. The previous commit failed on the Ubuntu bots using GCC 4.8. These bots lack the const_iterator forms of insert() and emplace() that were added in C++11. As a result I've switched the const_iterators to iterators. llvm-svn: 317049
2017-11-01 07:03:18 +08:00
return InsertPt;
[globalisel][tablegen] Add support for importing 'imm' operands. Summary: This patch enables the import of rules containing 'imm' operands that do not constrain the acceptable values using predicates. Support for ImmLeaf will arrive in a later patch. Depends on D35681 Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Reviewed By: rovka Subscribers: kristof.beyls, javed.absar, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D35833 llvm-svn: 310343
2017-08-08 18:44:31 +08:00
}
[globalisel][tablegen] Add support for multi-insn emission The importer will now accept nested instructions in the result pattern such as (ADDWrr $a, (SUBWrr $b, $c)). This is only valid when the nested instruction def's a single vreg and the parent instruction consumes a single vreg where a nested instruction is specified. The importer will automatically create a vreg to connect the two using the type information from the pattern. This vreg will be constrained to the register classes given in the instruction definitions*. * REG_SEQUENCE is explicitly rejected because of this. The definition doesn't constrain to a register class and it therefore needs special handling. llvm-svn: 317117
2017-11-02 03:57:57 +08:00
if (DstChild->getOperator()->isSubClassOf("Instruction")) {
ArrayRef<TypeSetByHwMode> ChildTypes = DstChild->getExtTypes();
if (ChildTypes.size() != 1)
return failedImport("Dst pattern child has multiple results");
Optional<LLTCodeGen> OpTyOrNone = None;
if (ChildTypes.front().isMachineValueType())
OpTyOrNone =
MVTToLLT(ChildTypes.front().getMachineValueType().SimpleTy);
if (!OpTyOrNone)
return failedImport("Dst operand has an unsupported type");
unsigned TempRegID = Rule.allocateTempRegID();
InsertPt = Rule.insertAction<MakeTempRegisterAction>(
InsertPt, OpTyOrNone.getValue(), TempRegID);
DstMIBuilder.addRenderer<TempRegRenderer>(TempRegID);
auto InsertPtOrError = createAndImportSubInstructionRenderer(
++InsertPt, Rule, DstChild, TempRegID);
if (auto Error = InsertPtOrError.takeError())
return std::move(Error);
return InsertPtOrError.get();
}
Re-commit: [globalisel][tablegen] Add support for ImmLeaf without SDNodeXForm Summary: This patch adds support for predicates on imm nodes but only for ImmLeaf and not for PatLeaf or PatFrag and only where the value does not need to be transformed before being rendered into the instruction. The limitation on PatLeaf/PatFrag/SDNodeXForm is due to differences in the necessary target-supplied C++ for GlobalISel. Depends on D36085 The previous commit was reverted for breaking the build but this appears to have been the recurring problem on the Windows bots with tablegen not being re-run when llvm-tblgen is changed but the .td's aren't. If it re-occurs then forcing a build with clean=True should fix it but this string should do this in advance: Requires a clean build. Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Reviewed By: rovka Subscribers: kristof.beyls, javed.absar, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D36086 llvm-svn: 311645
2017-08-24 17:11:20 +08:00
return failedImport("Dst pattern child isn't a leaf node or an MBB" + llvm::to_string(*DstChild));
[tablegen][globalisel] Convert the SelectionDAG importer to a tree walking approach. NFC Summary: But don't actually inspect the tree any deeper than we already do. This change is NFC but the next one will enable full traversal of the source/destination patterns. Depends on D30535 Reviewers: t.p.northover, qcolombet, aditya_nandakumar, rovka, ab Subscribers: igorb, dberris, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D30536 llvm-svn: 299001
2017-03-29 23:37:18 +08:00
}
[GlobalISel] Add basic Selector-emitter tblgen backend. This adds a basic tablegen backend that analyzes the SelectionDAG patterns to find simple ones that are eligible for GlobalISel-emission. That's similar to FastISel, with one notable difference: we're not fed ISD opcodes, so we need to map the SDNode operators to generic opcodes. That's done using GINodeEquiv in TargetGlobalISel.td. Otherwise, this is mostly boilerplate, and lots of filtering of any kind of "complicated" pattern. On AArch64, this is sufficient to match G_ADD up to s64 (to ADDWrr/ADDXrr) and G_BR (to B). Differential Revision: https://reviews.llvm.org/D26878 llvm-svn: 290284
2016-12-22 07:26:20 +08:00
[globalisel][tablegen] Add support for specific immediates in the match pattern This enables a few rules such as ARM's uxtb instruction. llvm-svn: 319457
2017-12-01 02:48:35 +08:00
// It could be a specific immediate in which case we should just check for
// that immediate.
if (const IntInit *ChildIntInit =
dyn_cast<IntInit>(DstChild->getLeafValue())) {
DstMIBuilder.addRenderer<ImmRenderer>(ChildIntInit->getValue());
return InsertPt;
}
[tablegen][globalisel] Convert the SelectionDAG importer to a tree walking approach. NFC Summary: But don't actually inspect the tree any deeper than we already do. This change is NFC but the next one will enable full traversal of the source/destination patterns. Depends on D30535 Reviewers: t.p.northover, qcolombet, aditya_nandakumar, rovka, ab Subscribers: igorb, dberris, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D30536 llvm-svn: 299001
2017-03-29 23:37:18 +08:00
// Otherwise, we're looking for a bog-standard RegisterClass operand.
if (auto *ChildDefInit = dyn_cast<DefInit>(DstChild->getLeafValue())) {
auto *ChildRec = ChildDefInit->getDef();
TableGen support for parameterized register class information This replaces TableGen's type inference to operate on parameterized types instead of MVTs, and as a consequence, some interfaces have changed: - Uses of MVTs are replaced by ValueTypeByHwMode. - EEVT::TypeSet is replaced by TypeSetByHwMode. This affects the way that types and type sets are printed, and the tests relying on that have been updated. There are certain users of the inferred types outside of TableGen itself, namely FastISel and GlobalISel. For those users, the way that the types are accessed have changed. For typical scenarios, these replacements can be used: - TreePatternNode::getType(ResNo) -> getSimpleType(ResNo) - TreePatternNode::hasTypeSet(ResNo) -> hasConcreteType(ResNo) - TypeSet::isConcrete -> TypeSetByHwMode::isValueTypeByHwMode(false) For more information, please refer to the review page. Differential Revision: https://reviews.llvm.org/D31951 llvm-svn: 313271
2017-09-15 00:56:21 +08:00
ArrayRef<TypeSetByHwMode> ChildTypes = DstChild->getExtTypes();
[GlobalISel] Add basic Selector-emitter tblgen backend. This adds a basic tablegen backend that analyzes the SelectionDAG patterns to find simple ones that are eligible for GlobalISel-emission. That's similar to FastISel, with one notable difference: we're not fed ISD opcodes, so we need to map the SDNode operators to generic opcodes. That's done using GINodeEquiv in TargetGlobalISel.td. Otherwise, this is mostly boilerplate, and lots of filtering of any kind of "complicated" pattern. On AArch64, this is sufficient to match G_ADD up to s64 (to ADDWrr/ADDXrr) and G_BR (to B). Differential Revision: https://reviews.llvm.org/D26878 llvm-svn: 290284
2016-12-22 07:26:20 +08:00
if (ChildTypes.size() != 1)
[tablegen][globalisel] Convert the SelectionDAG importer to a tree walking approach. NFC Summary: But don't actually inspect the tree any deeper than we already do. This change is NFC but the next one will enable full traversal of the source/destination patterns. Depends on D30535 Reviewers: t.p.northover, qcolombet, aditya_nandakumar, rovka, ab Subscribers: igorb, dberris, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D30536 llvm-svn: 299001
2017-03-29 23:37:18 +08:00
return failedImport("Dst pattern child has multiple results");
[GlobalISel] Add basic Selector-emitter tblgen backend. This adds a basic tablegen backend that analyzes the SelectionDAG patterns to find simple ones that are eligible for GlobalISel-emission. That's similar to FastISel, with one notable difference: we're not fed ISD opcodes, so we need to map the SDNode operators to generic opcodes. That's done using GINodeEquiv in TargetGlobalISel.td. Otherwise, this is mostly boilerplate, and lots of filtering of any kind of "complicated" pattern. On AArch64, this is sufficient to match G_ADD up to s64 (to ADDWrr/ADDXrr) and G_BR (to B). Differential Revision: https://reviews.llvm.org/D26878 llvm-svn: 290284
2016-12-22 07:26:20 +08:00
TableGen support for parameterized register class information This replaces TableGen's type inference to operate on parameterized types instead of MVTs, and as a consequence, some interfaces have changed: - Uses of MVTs are replaced by ValueTypeByHwMode. - EEVT::TypeSet is replaced by TypeSetByHwMode. This affects the way that types and type sets are printed, and the tests relying on that have been updated. There are certain users of the inferred types outside of TableGen itself, namely FastISel and GlobalISel. For those users, the way that the types are accessed have changed. For typical scenarios, these replacements can be used: - TreePatternNode::getType(ResNo) -> getSimpleType(ResNo) - TreePatternNode::hasTypeSet(ResNo) -> hasConcreteType(ResNo) - TypeSet::isConcrete -> TypeSetByHwMode::isValueTypeByHwMode(false) For more information, please refer to the review page. Differential Revision: https://reviews.llvm.org/D31951 llvm-svn: 313271
2017-09-15 00:56:21 +08:00
Optional<LLTCodeGen> OpTyOrNone = None;
if (ChildTypes.front().isMachineValueType())
OpTyOrNone = MVTToLLT(ChildTypes.front().getMachineValueType().SimpleTy);
[GlobalISel] Add basic Selector-emitter tblgen backend. This adds a basic tablegen backend that analyzes the SelectionDAG patterns to find simple ones that are eligible for GlobalISel-emission. That's similar to FastISel, with one notable difference: we're not fed ISD opcodes, so we need to map the SDNode operators to generic opcodes. That's done using GINodeEquiv in TargetGlobalISel.td. Otherwise, this is mostly boilerplate, and lots of filtering of any kind of "complicated" pattern. On AArch64, this is sufficient to match G_ADD up to s64 (to ADDWrr/ADDXrr) and G_BR (to B). Differential Revision: https://reviews.llvm.org/D26878 llvm-svn: 290284
2016-12-22 07:26:20 +08:00
if (!OpTyOrNone)
[tablegen][globalisel] Convert the SelectionDAG importer to a tree walking approach. NFC Summary: But don't actually inspect the tree any deeper than we already do. This change is NFC but the next one will enable full traversal of the source/destination patterns. Depends on D30535 Reviewers: t.p.northover, qcolombet, aditya_nandakumar, rovka, ab Subscribers: igorb, dberris, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D30536 llvm-svn: 299001
2017-03-29 23:37:18 +08:00
return failedImport("Dst operand has an unsupported type");
[globalisel] Decouple src pattern operands from dst pattern operands. Summary: This isn't testable for AArch64 by itself so this patch also adds support for constant immediates in the pattern and physical register uses in the result. The new IntOperandMatcher matches the constant in patterns such as '(set $rd:GPR32, (G_XOR $rs:GPR32, -1))'. It's always safe to fold immediates into an instruction so this is the first rule that will match across multiple BB's. The Renderer hierarchy is responsible for adding operands to the result instruction. Renderers can copy operands (CopyRenderer) or add physical registers (in particular %wzr and %xzr) to the result instruction in any order (OperandMatchers now import the operand names from SelectionDAG to allow renderers to access any operand). This allows us to emit the result instruction for: %1 = G_XOR %0, -1 --> %1 = ORNWrr %wzr, %0 %1 = G_XOR -1, %0 --> %1 = ORNWrr %wzr, %0 although the latter is untested since the matcher/importer has not been taught about commutativity yet. Added BuildMIAction which can build new instructions and mutate them where possible. W.r.t the mutation aspect, MatchActions are now told the name of an instruction they can recycle and BuildMIAction will emit mutation code when the renderers are appropriate. They are appropriate when all operands are rendered using CopyRenderer and the indices are the same as the matcher. This currently assumes that all operands have at least one matcher. Finally, this change also fixes a crash in AArch64InstructionSelector::select() caused by an immediate operand passing isImm() rather than isCImm(). This was uncovered by the other changes and was detected by existing tests. Depends on D29711 Reviewers: t.p.northover, ab, qcolombet, rovka, aditya_nandakumar, javed.absar Reviewed By: rovka Subscribers: aemerson, dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D29712 llvm-svn: 296131
2017-02-24 23:43:30 +08:00
[tablegen][globalisel] Convert the SelectionDAG importer to a tree walking approach. NFC Summary: But don't actually inspect the tree any deeper than we already do. This change is NFC but the next one will enable full traversal of the source/destination patterns. Depends on D30535 Reviewers: t.p.northover, qcolombet, aditya_nandakumar, rovka, ab Subscribers: igorb, dberris, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D30536 llvm-svn: 299001
2017-03-29 23:37:18 +08:00
if (ChildRec->isSubClassOf("Register")) {
[globalisel][tablegen] Stop hard-coding the emitted instruction ID to 0. NFC The next commit will add support for multi-instruction emission so we need to start allocating instruction ID's instead of hard-coding them to 0. llvm-svn: 317057
2017-11-01 08:29:47 +08:00
DstMIBuilder.addRenderer<AddRegisterRenderer>(ChildRec);
Re-commit: [globalisel][tablegen] Keep track of the insertion point while adding BuildMIAction's. NFC Multi-instruction emission needs to ensure the the instructions are generated a depth-first fashion. For example: (ADDWrr (SUBWrr a, b), c) needs to emit the SUBWrr before the ADDWrr. However, our walk over TreePatternNode's is highly context sensitive which makes it difficult to append BuildMIActions in the order we want. To fix this, we now keep track of the insertion point as we add actions. This will allow multi-insn emission to insert BuildMI's in the correct place. The previous commit failed on the Ubuntu bots using GCC 4.8. These bots lack the const_iterator forms of insert() and emplace() that were added in C++11. As a result I've switched the const_iterators to iterators. llvm-svn: 317049
2017-11-01 07:03:18 +08:00
return InsertPt;
[globalisel] Decouple src pattern operands from dst pattern operands. Summary: This isn't testable for AArch64 by itself so this patch also adds support for constant immediates in the pattern and physical register uses in the result. The new IntOperandMatcher matches the constant in patterns such as '(set $rd:GPR32, (G_XOR $rs:GPR32, -1))'. It's always safe to fold immediates into an instruction so this is the first rule that will match across multiple BB's. The Renderer hierarchy is responsible for adding operands to the result instruction. Renderers can copy operands (CopyRenderer) or add physical registers (in particular %wzr and %xzr) to the result instruction in any order (OperandMatchers now import the operand names from SelectionDAG to allow renderers to access any operand). This allows us to emit the result instruction for: %1 = G_XOR %0, -1 --> %1 = ORNWrr %wzr, %0 %1 = G_XOR -1, %0 --> %1 = ORNWrr %wzr, %0 although the latter is untested since the matcher/importer has not been taught about commutativity yet. Added BuildMIAction which can build new instructions and mutate them where possible. W.r.t the mutation aspect, MatchActions are now told the name of an instruction they can recycle and BuildMIAction will emit mutation code when the renderers are appropriate. They are appropriate when all operands are rendered using CopyRenderer and the indices are the same as the matcher. This currently assumes that all operands have at least one matcher. Finally, this change also fixes a crash in AArch64InstructionSelector::select() caused by an immediate operand passing isImm() rather than isCImm(). This was uncovered by the other changes and was detected by existing tests. Depends on D29711 Reviewers: t.p.northover, ab, qcolombet, rovka, aditya_nandakumar, javed.absar Reviewed By: rovka Subscribers: aemerson, dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D29712 llvm-svn: 296131
2017-02-24 23:43:30 +08:00
}
[globalisel][tablegen] Add support for RegisterOperand. Summary: It functions just like RegisterClass except that the class is obtained from a field. Depends on D31761. Reviewers: ab, qcolombet, t.p.northover, rovka, kristof.beyls, aditya_nandakumar Reviewed By: ab Subscribers: dberris, llvm-commits, igorb Differential Revision: https://reviews.llvm.org/D32229 llvm-svn: 301080
2017-04-22 23:53:21 +08:00
if (ChildRec->isSubClassOf("RegisterClass") ||
[globalisel] Add support for ValueType operands in patterns. It's rare but there are a small number of patterns like this: (set i64:$dst, (add i64:$src1, i64:$src2)) These should be equivalent to register classes except they shouldn't check for a specific register bank. This doesn't occur in AArch64/ARM/X86 but does occasionally come up in other in-tree targets such as BPF. llvm-svn: 315226
2017-10-10 02:14:53 +08:00
ChildRec->isSubClassOf("RegisterOperand") ||
ChildRec->isSubClassOf("ValueType")) {
[globalisel][tablegen] Import stores and allow GISel to automatically substitute zero regs like WZR/XZR/$zero. This patch enables the import of stores. Unfortunately, doing so by itself, loses an optimization where storing 0 to memory makes use of WZR/XZR. To mitigate this, this patch also introduces a new feature that allows register operands to nominate a zero register. When this is done, GlobalISel will substitute (G_CONSTANT 0) with the nominated register automatically. This is currently configured to only apply to the stores. Applying it to GPR32/GPR64 register classes in general will be done after review see (https://reviews.llvm.org/D39150). llvm-svn: 316360
2017-10-24 02:19:24 +08:00
if (ChildRec->isSubClassOf("RegisterOperand") &&
!ChildRec->isValueUnset("GIZeroRegister")) {
DstMIBuilder.addRenderer<CopyOrAddZeroRegRenderer>(
[globalisel][tablegen] Stop hard-coding the emitted instruction ID to 0. NFC The next commit will add support for multi-instruction emission so we need to start allocating instruction ID's instead of hard-coding them to 0. llvm-svn: 317057
2017-11-01 08:29:47 +08:00
DstChild->getName(), ChildRec->getValueAsDef("GIZeroRegister"));
Re-commit: [globalisel][tablegen] Keep track of the insertion point while adding BuildMIAction's. NFC Multi-instruction emission needs to ensure the the instructions are generated a depth-first fashion. For example: (ADDWrr (SUBWrr a, b), c) needs to emit the SUBWrr before the ADDWrr. However, our walk over TreePatternNode's is highly context sensitive which makes it difficult to append BuildMIActions in the order we want. To fix this, we now keep track of the insertion point as we add actions. This will allow multi-insn emission to insert BuildMI's in the correct place. The previous commit failed on the Ubuntu bots using GCC 4.8. These bots lack the const_iterator forms of insert() and emplace() that were added in C++11. As a result I've switched the const_iterators to iterators. llvm-svn: 317049
2017-11-01 07:03:18 +08:00
return InsertPt;
[globalisel][tablegen] Import stores and allow GISel to automatically substitute zero regs like WZR/XZR/$zero. This patch enables the import of stores. Unfortunately, doing so by itself, loses an optimization where storing 0 to memory makes use of WZR/XZR. To mitigate this, this patch also introduces a new feature that allows register operands to nominate a zero register. When this is done, GlobalISel will substitute (G_CONSTANT 0) with the nominated register automatically. This is currently configured to only apply to the stores. Applying it to GPR32/GPR64 register classes in general will be done after review see (https://reviews.llvm.org/D39150). llvm-svn: 316360
2017-10-24 02:19:24 +08:00
}
[globalisel][tablegen] Stop hard-coding the emitted instruction ID to 0. NFC The next commit will add support for multi-instruction emission so we need to start allocating instruction ID's instead of hard-coding them to 0. llvm-svn: 317057
2017-11-01 08:29:47 +08:00
DstMIBuilder.addRenderer<CopyRenderer>(DstChild->getName());
Re-commit: [globalisel][tablegen] Keep track of the insertion point while adding BuildMIAction's. NFC Multi-instruction emission needs to ensure the the instructions are generated a depth-first fashion. For example: (ADDWrr (SUBWrr a, b), c) needs to emit the SUBWrr before the ADDWrr. However, our walk over TreePatternNode's is highly context sensitive which makes it difficult to append BuildMIActions in the order we want. To fix this, we now keep track of the insertion point as we add actions. This will allow multi-insn emission to insert BuildMI's in the correct place. The previous commit failed on the Ubuntu bots using GCC 4.8. These bots lack the const_iterator forms of insert() and emplace() that were added in C++11. As a result I've switched the const_iterators to iterators. llvm-svn: 317049
2017-11-01 07:03:18 +08:00
return InsertPt;
[tablegen][globalisel] Convert the SelectionDAG importer to a tree walking approach. NFC Summary: But don't actually inspect the tree any deeper than we already do. This change is NFC but the next one will enable full traversal of the source/destination patterns. Depends on D30535 Reviewers: t.p.northover, qcolombet, aditya_nandakumar, rovka, ab Subscribers: igorb, dberris, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D30536 llvm-svn: 299001
2017-03-29 23:37:18 +08:00
}
[globalisel] Decouple src pattern operands from dst pattern operands. Summary: This isn't testable for AArch64 by itself so this patch also adds support for constant immediates in the pattern and physical register uses in the result. The new IntOperandMatcher matches the constant in patterns such as '(set $rd:GPR32, (G_XOR $rs:GPR32, -1))'. It's always safe to fold immediates into an instruction so this is the first rule that will match across multiple BB's. The Renderer hierarchy is responsible for adding operands to the result instruction. Renderers can copy operands (CopyRenderer) or add physical registers (in particular %wzr and %xzr) to the result instruction in any order (OperandMatchers now import the operand names from SelectionDAG to allow renderers to access any operand). This allows us to emit the result instruction for: %1 = G_XOR %0, -1 --> %1 = ORNWrr %wzr, %0 %1 = G_XOR -1, %0 --> %1 = ORNWrr %wzr, %0 although the latter is untested since the matcher/importer has not been taught about commutativity yet. Added BuildMIAction which can build new instructions and mutate them where possible. W.r.t the mutation aspect, MatchActions are now told the name of an instruction they can recycle and BuildMIAction will emit mutation code when the renderers are appropriate. They are appropriate when all operands are rendered using CopyRenderer and the indices are the same as the matcher. This currently assumes that all operands have at least one matcher. Finally, this change also fixes a crash in AArch64InstructionSelector::select() caused by an immediate operand passing isImm() rather than isCImm(). This was uncovered by the other changes and was detected by existing tests. Depends on D29711 Reviewers: t.p.northover, ab, qcolombet, rovka, aditya_nandakumar, javed.absar Reviewed By: rovka Subscribers: aemerson, dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D29712 llvm-svn: 296131
2017-02-24 23:43:30 +08:00
[tablegen][globalisel] Convert the SelectionDAG importer to a tree walking approach. NFC Summary: But don't actually inspect the tree any deeper than we already do. This change is NFC but the next one will enable full traversal of the source/destination patterns. Depends on D30535 Reviewers: t.p.northover, qcolombet, aditya_nandakumar, rovka, ab Subscribers: igorb, dberris, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D30536 llvm-svn: 299001
2017-03-29 23:37:18 +08:00
if (ChildRec->isSubClassOf("ComplexPattern")) {
const auto &ComplexPattern = ComplexPatternEquivs.find(ChildRec);
if (ComplexPattern == ComplexPatternEquivs.end())
return failedImport(
"SelectionDAG ComplexPattern not mapped to GlobalISel");
[globalisel][tablegen] Simplify named operand/operator lookups and fix a wrong-code bug this revealed. Summary: Operand variable lookups are now performed by the RuleMatcher rather than searching the whole matcher hierarchy for a match. This revealed a wrong-code bug that currently affects ARM and X86 where patterns that use a variable more than once in the match pattern will be imported but won't check that the operands are identical. This can cause the tablegen-erated matcher to accept matches that should be rejected. Depends on D36569 Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Subscribers: aemerson, igorb, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D36618 llvm-svn: 315780
2017-10-14 08:31:58 +08:00
const OperandMatcher &OM = Rule.getOperandMatcher(DstChild->getName());
[tablegen][globalisel] Convert the SelectionDAG importer to a tree walking approach. NFC Summary: But don't actually inspect the tree any deeper than we already do. This change is NFC but the next one will enable full traversal of the source/destination patterns. Depends on D30535 Reviewers: t.p.northover, qcolombet, aditya_nandakumar, rovka, ab Subscribers: igorb, dberris, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D30536 llvm-svn: 299001
2017-03-29 23:37:18 +08:00
DstMIBuilder.addRenderer<RenderComplexPatternOperand>(
[globalisel][tablegen] Stop hard-coding the emitted instruction ID to 0. NFC The next commit will add support for multi-instruction emission so we need to start allocating instruction ID's instead of hard-coding them to 0. llvm-svn: 317057
2017-11-01 08:29:47 +08:00
*ComplexPattern->second, DstChild->getName(),
[globalisel][tablegen] Revise API for ComplexPattern operands to improve flexibility. Summary: Some targets need to be able to do more complex rendering than just adding an operand or two to an instruction. For example, it may need to insert an instruction to extract a subreg first, or it may need to perform an operation on the operand. In SelectionDAG, targets would create SDNode's to achieve the desired effect during the complex pattern predicate. This worked because SelectionDAG had a form of garbage collection that would take care of SDNode's that were created but not used due to a later predicate rejecting a match. This doesn't translate well to GlobalISel and the churn was wasteful. The API changes in this patch enable GlobalISel to accomplish the same thing without the waste. The API is now: InstructionSelector::OptionalComplexRendererFn selectArithImmed(MachineOperand &Root) const; where Root is the root of the match. The return value can be omitted to indicate that the predicate failed to match, or a function with the signature ComplexRendererFn can be returned. For example: return OptionalComplexRendererFn( [=](MachineInstrBuilder &MIB) { MIB.addImm(Immed).addImm(ShVal); }); adds two immediate operands to the rendered instruction. Immed and ShVal are captured from the predicate function. As an added bonus, this also reduces the amount of information we need to provide to GIComplexOperandMatcher. Depends on D31418 Reviewers: aditya_nandakumar, t.p.northover, qcolombet, rovka, ab, javed.absar Reviewed By: ab Subscribers: dberris, kristof.beyls, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D31761 llvm-svn: 301079
2017-04-22 23:11:04 +08:00
OM.getAllocatedTemporariesBaseID());
Re-commit: [globalisel][tablegen] Keep track of the insertion point while adding BuildMIAction's. NFC Multi-instruction emission needs to ensure the the instructions are generated a depth-first fashion. For example: (ADDWrr (SUBWrr a, b), c) needs to emit the SUBWrr before the ADDWrr. However, our walk over TreePatternNode's is highly context sensitive which makes it difficult to append BuildMIActions in the order we want. To fix this, we now keep track of the insertion point as we add actions. This will allow multi-insn emission to insert BuildMI's in the correct place. The previous commit failed on the Ubuntu bots using GCC 4.8. These bots lack the const_iterator forms of insert() and emplace() that were added in C++11. As a result I've switched the const_iterators to iterators. llvm-svn: 317049
2017-11-01 07:03:18 +08:00
return InsertPt;
[tablegen][globalisel] Convert the SelectionDAG importer to a tree walking approach. NFC Summary: But don't actually inspect the tree any deeper than we already do. This change is NFC but the next one will enable full traversal of the source/destination patterns. Depends on D30535 Reviewers: t.p.northover, qcolombet, aditya_nandakumar, rovka, ab Subscribers: igorb, dberris, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D30536 llvm-svn: 299001
2017-03-29 23:37:18 +08:00
}
[globalisel] Decouple src pattern operands from dst pattern operands. Summary: This isn't testable for AArch64 by itself so this patch also adds support for constant immediates in the pattern and physical register uses in the result. The new IntOperandMatcher matches the constant in patterns such as '(set $rd:GPR32, (G_XOR $rs:GPR32, -1))'. It's always safe to fold immediates into an instruction so this is the first rule that will match across multiple BB's. The Renderer hierarchy is responsible for adding operands to the result instruction. Renderers can copy operands (CopyRenderer) or add physical registers (in particular %wzr and %xzr) to the result instruction in any order (OperandMatchers now import the operand names from SelectionDAG to allow renderers to access any operand). This allows us to emit the result instruction for: %1 = G_XOR %0, -1 --> %1 = ORNWrr %wzr, %0 %1 = G_XOR -1, %0 --> %1 = ORNWrr %wzr, %0 although the latter is untested since the matcher/importer has not been taught about commutativity yet. Added BuildMIAction which can build new instructions and mutate them where possible. W.r.t the mutation aspect, MatchActions are now told the name of an instruction they can recycle and BuildMIAction will emit mutation code when the renderers are appropriate. They are appropriate when all operands are rendered using CopyRenderer and the indices are the same as the matcher. This currently assumes that all operands have at least one matcher. Finally, this change also fixes a crash in AArch64InstructionSelector::select() caused by an immediate operand passing isImm() rather than isCImm(). This was uncovered by the other changes and was detected by existing tests. Depends on D29711 Reviewers: t.p.northover, ab, qcolombet, rovka, aditya_nandakumar, javed.absar Reviewed By: rovka Subscribers: aemerson, dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D29712 llvm-svn: 296131
2017-02-24 23:43:30 +08:00
[tablegen][globalisel] Convert the SelectionDAG importer to a tree walking approach. NFC Summary: But don't actually inspect the tree any deeper than we already do. This change is NFC but the next one will enable full traversal of the source/destination patterns. Depends on D30535 Reviewers: t.p.northover, qcolombet, aditya_nandakumar, rovka, ab Subscribers: igorb, dberris, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D30536 llvm-svn: 299001
2017-03-29 23:37:18 +08:00
return failedImport(
"Dst pattern child def is an unsupported tablegen class");
}
[globalisel][tblgen] Add support for ComplexPatterns Summary: Adds a new kind of MachineOperand: MO_Placeholder. This operand must not appear in the MIR and only exists as a way of creating an 'uninitialized' operand until a matcher function overwrites it. Depends on D30046, D29712 Reviewers: t.p.northover, ab, rovka, aditya_nandakumar, javed.absar, qcolombet Reviewed By: qcolombet Subscribers: dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D30089 llvm-svn: 297782
2017-03-15 05:32:08 +08:00
[tablegen][globalisel] Convert the SelectionDAG importer to a tree walking approach. NFC Summary: But don't actually inspect the tree any deeper than we already do. This change is NFC but the next one will enable full traversal of the source/destination patterns. Depends on D30535 Reviewers: t.p.northover, qcolombet, aditya_nandakumar, rovka, ab Subscribers: igorb, dberris, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D30536 llvm-svn: 299001
2017-03-29 23:37:18 +08:00
return failedImport("Dst pattern child is an unsupported kind");
}
[globalisel][tblgen] Add support for ComplexPatterns Summary: Adds a new kind of MachineOperand: MO_Placeholder. This operand must not appear in the MIR and only exists as a way of creating an 'uninitialized' operand until a matcher function overwrites it. Depends on D30046, D29712 Reviewers: t.p.northover, ab, rovka, aditya_nandakumar, javed.absar, qcolombet Reviewed By: qcolombet Subscribers: dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D30089 llvm-svn: 297782
2017-03-15 05:32:08 +08:00
[globalisel][tablegen] Change Expected<bool> to Error and rename functions. Functions that still return Expected<X> are now called createAndImport*() Changing the return type was requested in the review comments for r299001 llvm-svn: 299063
2017-03-30 17:36:33 +08:00
Expected<BuildMIAction &> GlobalISelEmitter::createAndImportInstructionRenderer(
[globalisel][tablegen] Add infrastructure to potentially allow BuildMIAction to choose a mutatable instruction. NFC Prepare for multiple instruction emission by allowing BuildMIAction to search for a suitable matcher that will support mutation. This patch deliberately neglects to add matchers aside from the root to preserve NFC. That said, it should be noted that until we support mutations other than just the opcode the chances of finding a non-root instruction for which canMutate() is true, is essentially zero. Furthermore in the presence of multi-instruction emission the chances of finding any instruction for which canMutate() is true is also zero. Nevertheless, we can't continue to require that all BuildMIAction's consider the root of the match to be recyclable due to the risk of recycling it twice in the same rule. llvm-svn: 317022
2017-11-01 02:50:24 +08:00
RuleMatcher &M, const TreePatternNode *Dst) {
Re-commit: [globalisel][tablegen] Keep track of the insertion point while adding BuildMIAction's. NFC Multi-instruction emission needs to ensure the the instructions are generated a depth-first fashion. For example: (ADDWrr (SUBWrr a, b), c) needs to emit the SUBWrr before the ADDWrr. However, our walk over TreePatternNode's is highly context sensitive which makes it difficult to append BuildMIActions in the order we want. To fix this, we now keep track of the insertion point as we add actions. This will allow multi-insn emission to insert BuildMI's in the correct place. The previous commit failed on the Ubuntu bots using GCC 4.8. These bots lack the const_iterator forms of insert() and emplace() that were added in C++11. As a result I've switched the const_iterators to iterators. llvm-svn: 317049
2017-11-01 07:03:18 +08:00
auto InsertPtOrError = createInstructionRenderer(M.actions_end(), M, Dst);
if (auto Error = InsertPtOrError.takeError())
[globalisel][tablegen] Factor out implicit def/use renderers from createAndImportInstructionRenderer(). NFC Multi-instruction emission will require that we have separate handling for the defs between the implicitly created temporaries and the rule outputs. The former require new temporary vregs while the latter should copy existing operands. Factor out the implicit def/use renderers to minimize the code duplication when we implement that. llvm-svn: 317025
2017-11-01 03:09:29 +08:00
return std::move(Error);
Re-commit: [globalisel][tablegen] Keep track of the insertion point while adding BuildMIAction's. NFC Multi-instruction emission needs to ensure the the instructions are generated a depth-first fashion. For example: (ADDWrr (SUBWrr a, b), c) needs to emit the SUBWrr before the ADDWrr. However, our walk over TreePatternNode's is highly context sensitive which makes it difficult to append BuildMIActions in the order we want. To fix this, we now keep track of the insertion point as we add actions. This will allow multi-insn emission to insert BuildMI's in the correct place. The previous commit failed on the Ubuntu bots using GCC 4.8. These bots lack the const_iterator forms of insert() and emplace() that were added in C++11. As a result I've switched the const_iterators to iterators. llvm-svn: 317049
2017-11-01 07:03:18 +08:00
action_iterator InsertPt = InsertPtOrError.get();
BuildMIAction &DstMIBuilder = *static_cast<BuildMIAction *>(InsertPt->get());
[globalisel][tablegen] Factor out implicit def/use renderers from createAndImportInstructionRenderer(). NFC Multi-instruction emission will require that we have separate handling for the defs between the implicitly created temporaries and the rule outputs. The former require new temporary vregs while the latter should copy existing operands. Factor out the implicit def/use renderers to minimize the code duplication when we implement that. llvm-svn: 317025
2017-11-01 03:09:29 +08:00
importExplicitDefRenderers(DstMIBuilder);
Re-commit: [globalisel][tablegen] Keep track of the insertion point while adding BuildMIAction's. NFC Multi-instruction emission needs to ensure the the instructions are generated a depth-first fashion. For example: (ADDWrr (SUBWrr a, b), c) needs to emit the SUBWrr before the ADDWrr. However, our walk over TreePatternNode's is highly context sensitive which makes it difficult to append BuildMIActions in the order we want. To fix this, we now keep track of the insertion point as we add actions. This will allow multi-insn emission to insert BuildMI's in the correct place. The previous commit failed on the Ubuntu bots using GCC 4.8. These bots lack the const_iterator forms of insert() and emplace() that were added in C++11. As a result I've switched the const_iterators to iterators. llvm-svn: 317049
2017-11-01 07:03:18 +08:00
if (auto Error = importExplicitUseRenderers(InsertPt, M, DstMIBuilder, Dst)
.takeError())
[globalisel][tablegen] Factor out implicit def/use renderers from createAndImportInstructionRenderer(). NFC Multi-instruction emission will require that we have separate handling for the defs between the implicitly created temporaries and the rule outputs. The former require new temporary vregs while the latter should copy existing operands. Factor out the implicit def/use renderers to minimize the code duplication when we implement that. llvm-svn: 317025
2017-11-01 03:09:29 +08:00
return std::move(Error);
return DstMIBuilder;
}
[globalisel][tablegen] Add support for multi-insn emission The importer will now accept nested instructions in the result pattern such as (ADDWrr $a, (SUBWrr $b, $c)). This is only valid when the nested instruction def's a single vreg and the parent instruction consumes a single vreg where a nested instruction is specified. The importer will automatically create a vreg to connect the two using the type information from the pattern. This vreg will be constrained to the register classes given in the instruction definitions*. * REG_SEQUENCE is explicitly rejected because of this. The definition doesn't constrain to a register class and it therefore needs special handling. llvm-svn: 317117
2017-11-02 03:57:57 +08:00
Expected<action_iterator>
GlobalISelEmitter::createAndImportSubInstructionRenderer(
[ARM][GISel] PR35965 Constrain RegClasses of nested instructions built from Dst Pattern Summary: Apparently, we missed on constraining register classes of VReg-operands of all the instructions built from a destination pattern but the root (top-level) one. The issue exposed itself while selecting G_FPTOSI for armv7: the corresponding pattern generates VTOSIZS wrapped into COPY_TO_REGCLASS, so top-level COPY_TO_REGCLASS gets properly constrained, while nested VTOSIZS (or rather its destination virtual register to be exact) does not. Fixing this by issuing GIR_ConstrainSelectedInstOperands for every nested GIR_BuildMI. https://bugs.llvm.org/show_bug.cgi?id=35965 rdar://problem/36886530 Patch by Roman Tereshin Reviewers: dsanders, qcolombet, rovka, bogner, aditya_nandakumar, volkan Reviewed By: dsanders, qcolombet, rovka Subscribers: aemerson, javed.absar, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D42565 llvm-svn: 323692
2018-01-30 05:09:12 +08:00
const action_iterator InsertPt, RuleMatcher &M, const TreePatternNode *Dst,
[globalisel][tablegen] Add support for multi-insn emission The importer will now accept nested instructions in the result pattern such as (ADDWrr $a, (SUBWrr $b, $c)). This is only valid when the nested instruction def's a single vreg and the parent instruction consumes a single vreg where a nested instruction is specified. The importer will automatically create a vreg to connect the two using the type information from the pattern. This vreg will be constrained to the register classes given in the instruction definitions*. * REG_SEQUENCE is explicitly rejected because of this. The definition doesn't constrain to a register class and it therefore needs special handling. llvm-svn: 317117
2017-11-02 03:57:57 +08:00
unsigned TempRegID) {
auto InsertPtOrError = createInstructionRenderer(InsertPt, M, Dst);
// TODO: Assert there's exactly one result.
if (auto Error = InsertPtOrError.takeError())
return std::move(Error);
BuildMIAction &DstMIBuilder =
*static_cast<BuildMIAction *>(InsertPtOrError.get()->get());
// Assign the result to TempReg.
DstMIBuilder.addRenderer<TempRegRenderer>(TempRegID, true);
[ARM][GISel] PR35965 Constrain RegClasses of nested instructions built from Dst Pattern Summary: Apparently, we missed on constraining register classes of VReg-operands of all the instructions built from a destination pattern but the root (top-level) one. The issue exposed itself while selecting G_FPTOSI for armv7: the corresponding pattern generates VTOSIZS wrapped into COPY_TO_REGCLASS, so top-level COPY_TO_REGCLASS gets properly constrained, while nested VTOSIZS (or rather its destination virtual register to be exact) does not. Fixing this by issuing GIR_ConstrainSelectedInstOperands for every nested GIR_BuildMI. https://bugs.llvm.org/show_bug.cgi?id=35965 rdar://problem/36886530 Patch by Roman Tereshin Reviewers: dsanders, qcolombet, rovka, bogner, aditya_nandakumar, volkan Reviewed By: dsanders, qcolombet, rovka Subscribers: aemerson, javed.absar, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D42565 llvm-svn: 323692
2018-01-30 05:09:12 +08:00
InsertPtOrError =
importExplicitUseRenderers(InsertPtOrError.get(), M, DstMIBuilder, Dst);
[globalisel][tablegen] Add support for multi-insn emission The importer will now accept nested instructions in the result pattern such as (ADDWrr $a, (SUBWrr $b, $c)). This is only valid when the nested instruction def's a single vreg and the parent instruction consumes a single vreg where a nested instruction is specified. The importer will automatically create a vreg to connect the two using the type information from the pattern. This vreg will be constrained to the register classes given in the instruction definitions*. * REG_SEQUENCE is explicitly rejected because of this. The definition doesn't constrain to a register class and it therefore needs special handling. llvm-svn: 317117
2017-11-02 03:57:57 +08:00
if (auto Error = InsertPtOrError.takeError())
return std::move(Error);
[ARM][GISel] PR35965 Constrain RegClasses of nested instructions built from Dst Pattern Summary: Apparently, we missed on constraining register classes of VReg-operands of all the instructions built from a destination pattern but the root (top-level) one. The issue exposed itself while selecting G_FPTOSI for armv7: the corresponding pattern generates VTOSIZS wrapped into COPY_TO_REGCLASS, so top-level COPY_TO_REGCLASS gets properly constrained, while nested VTOSIZS (or rather its destination virtual register to be exact) does not. Fixing this by issuing GIR_ConstrainSelectedInstOperands for every nested GIR_BuildMI. https://bugs.llvm.org/show_bug.cgi?id=35965 rdar://problem/36886530 Patch by Roman Tereshin Reviewers: dsanders, qcolombet, rovka, bogner, aditya_nandakumar, volkan Reviewed By: dsanders, qcolombet, rovka Subscribers: aemerson, javed.absar, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D42565 llvm-svn: 323692
2018-01-30 05:09:12 +08:00
M.insertAction<ConstrainOperandsToDefinitionAction>(InsertPt,
DstMIBuilder.getInsnID());
[globalisel][tablegen] Add support for multi-insn emission The importer will now accept nested instructions in the result pattern such as (ADDWrr $a, (SUBWrr $b, $c)). This is only valid when the nested instruction def's a single vreg and the parent instruction consumes a single vreg where a nested instruction is specified. The importer will automatically create a vreg to connect the two using the type information from the pattern. This vreg will be constrained to the register classes given in the instruction definitions*. * REG_SEQUENCE is explicitly rejected because of this. The definition doesn't constrain to a register class and it therefore needs special handling. llvm-svn: 317117
2017-11-02 03:57:57 +08:00
return InsertPtOrError.get();
}
Re-commit: [globalisel][tablegen] Keep track of the insertion point while adding BuildMIAction's. NFC Multi-instruction emission needs to ensure the the instructions are generated a depth-first fashion. For example: (ADDWrr (SUBWrr a, b), c) needs to emit the SUBWrr before the ADDWrr. However, our walk over TreePatternNode's is highly context sensitive which makes it difficult to append BuildMIActions in the order we want. To fix this, we now keep track of the insertion point as we add actions. This will allow multi-insn emission to insert BuildMI's in the correct place. The previous commit failed on the Ubuntu bots using GCC 4.8. These bots lack the const_iterator forms of insert() and emplace() that were added in C++11. As a result I've switched the const_iterators to iterators. llvm-svn: 317049
2017-11-01 07:03:18 +08:00
Expected<action_iterator> GlobalISelEmitter::createInstructionRenderer(
action_iterator InsertPt, RuleMatcher &M, const TreePatternNode *Dst) {
[tablegen][globalisel] Convert the SelectionDAG importer to a tree walking approach. NFC Summary: But don't actually inspect the tree any deeper than we already do. This change is NFC but the next one will enable full traversal of the source/destination patterns. Depends on D30535 Reviewers: t.p.northover, qcolombet, aditya_nandakumar, rovka, ab Subscribers: igorb, dberris, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D30536 llvm-svn: 299001
2017-03-29 23:37:18 +08:00
Record *DstOp = Dst->getOperator();
[globalisel][tablegen] Report more detail in some SelectionDAG import failures. NFC Reviewers: ab, t.p.northover, qcolombet, aditya_nandakumar, rovka Reviewed By: ab Subscribers: dberris, kristof.beyls, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D31325 llvm-svn: 300186
2017-04-13 17:45:37 +08:00
if (!DstOp->isSubClassOf("Instruction")) {
if (DstOp->isSubClassOf("ValueType"))
return failedImport(
"Pattern operator isn't an instruction (it's a ValueType)");
[tablegen][globalisel] Convert the SelectionDAG importer to a tree walking approach. NFC Summary: But don't actually inspect the tree any deeper than we already do. This change is NFC but the next one will enable full traversal of the source/destination patterns. Depends on D30535 Reviewers: t.p.northover, qcolombet, aditya_nandakumar, rovka, ab Subscribers: igorb, dberris, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D30536 llvm-svn: 299001
2017-03-29 23:37:18 +08:00
return failedImport("Pattern operator isn't an instruction");
[globalisel][tablegen] Report more detail in some SelectionDAG import failures. NFC Reviewers: ab, t.p.northover, qcolombet, aditya_nandakumar, rovka Reviewed By: ab Subscribers: dberris, kristof.beyls, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D31325 llvm-svn: 300186
2017-04-13 17:45:37 +08:00
}
[globalisel][tablegen] Add support for COPY_TO_REGCLASS. Summary: As part of this * Emitted instructions now have named MachineInstr variables associated with them. This isn't particularly important yet but it's a small step towards multiple-insn emission. * constrainSelectedInstRegOperands() is no longer hardcoded. It's now added as the ConstrainOperandsToDefinitionAction() action. COPY_TO_REGCLASS uses an alternate constraint mechanism ConstrainOperandToRegClassAction() which supports arbitrary constraints such as that defined by COPY_TO_REGCLASS. Reviewers: ab, qcolombet, t.p.northover, rovka, kristof.beyls, aditya_nandakumar Reviewed By: ab Subscribers: javed.absar, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D33590 llvm-svn: 305791
2017-06-20 20:36:34 +08:00
CodeGenInstruction *DstI = &Target.getInstruction(DstOp);
// COPY_TO_REGCLASS is just a copy with a ConstrainOperandToRegClassAction
[globalisel][tablegen] Add support for EXTRACT_SUBREG. Summary: After this patch, we finally have test cases that require multiple instruction emission. Depends on D33590 Reviewers: ab, qcolombet, t.p.northover, rovka, kristof.beyls Subscribers: javed.absar, llvm-commits, igorb Differential Revision: https://reviews.llvm.org/D33596 llvm-svn: 306388
2017-06-27 18:11:39 +08:00
// attached. Similarly for EXTRACT_SUBREG except that's a subregister copy.
[globalisel][tablegen] Factor out implicit def/use renderers from createAndImportInstructionRenderer(). NFC Multi-instruction emission will require that we have separate handling for the defs between the implicitly created temporaries and the rule outputs. The former require new temporary vregs while the latter should copy existing operands. Factor out the implicit def/use renderers to minimize the code duplication when we implement that. llvm-svn: 317025
2017-11-01 03:09:29 +08:00
if (DstI->TheDef->getName() == "COPY_TO_REGCLASS")
[globalisel][tablegen] Add support for COPY_TO_REGCLASS. Summary: As part of this * Emitted instructions now have named MachineInstr variables associated with them. This isn't particularly important yet but it's a small step towards multiple-insn emission. * constrainSelectedInstRegOperands() is no longer hardcoded. It's now added as the ConstrainOperandsToDefinitionAction() action. COPY_TO_REGCLASS uses an alternate constraint mechanism ConstrainOperandToRegClassAction() which supports arbitrary constraints such as that defined by COPY_TO_REGCLASS. Reviewers: ab, qcolombet, t.p.northover, rovka, kristof.beyls, aditya_nandakumar Reviewed By: ab Subscribers: javed.absar, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D33590 llvm-svn: 305791
2017-06-20 20:36:34 +08:00
DstI = &Target.getInstruction(RK.getDef("COPY"));
[globalisel][tablegen] Factor out implicit def/use renderers from createAndImportInstructionRenderer(). NFC Multi-instruction emission will require that we have separate handling for the defs between the implicitly created temporaries and the rule outputs. The former require new temporary vregs while the latter should copy existing operands. Factor out the implicit def/use renderers to minimize the code duplication when we implement that. llvm-svn: 317025
2017-11-01 03:09:29 +08:00
else if (DstI->TheDef->getName() == "EXTRACT_SUBREG")
[globalisel][tablegen] Add support for EXTRACT_SUBREG. Summary: After this patch, we finally have test cases that require multiple instruction emission. Depends on D33590 Reviewers: ab, qcolombet, t.p.northover, rovka, kristof.beyls Subscribers: javed.absar, llvm-commits, igorb Differential Revision: https://reviews.llvm.org/D33596 llvm-svn: 306388
2017-06-27 18:11:39 +08:00
DstI = &Target.getInstruction(RK.getDef("COPY"));
[globalisel][tablegen] Add support for multi-insn emission The importer will now accept nested instructions in the result pattern such as (ADDWrr $a, (SUBWrr $b, $c)). This is only valid when the nested instruction def's a single vreg and the parent instruction consumes a single vreg where a nested instruction is specified. The importer will automatically create a vreg to connect the two using the type information from the pattern. This vreg will be constrained to the register classes given in the instruction definitions*. * REG_SEQUENCE is explicitly rejected because of this. The definition doesn't constrain to a register class and it therefore needs special handling. llvm-svn: 317117
2017-11-02 03:57:57 +08:00
else if (DstI->TheDef->getName() == "REG_SEQUENCE")
return failedImport("Unable to emit REG_SEQUENCE");
[tablegen][globalisel] Convert the SelectionDAG importer to a tree walking approach. NFC Summary: But don't actually inspect the tree any deeper than we already do. This change is NFC but the next one will enable full traversal of the source/destination patterns. Depends on D30535 Reviewers: t.p.northover, qcolombet, aditya_nandakumar, rovka, ab Subscribers: igorb, dberris, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D30536 llvm-svn: 299001
2017-03-29 23:37:18 +08:00
[globalisel][tablegen] Stop hard-coding the emitted instruction ID to 0. NFC The next commit will add support for multi-instruction emission so we need to start allocating instruction ID's instead of hard-coding them to 0. llvm-svn: 317057
2017-11-01 08:29:47 +08:00
return M.insertAction<BuildMIAction>(InsertPt, M.allocateOutputInsnID(),
DstI);
[globalisel][tablegen] Factor out implicit def/use renderers from createAndImportInstructionRenderer(). NFC Multi-instruction emission will require that we have separate handling for the defs between the implicitly created temporaries and the rule outputs. The former require new temporary vregs while the latter should copy existing operands. Factor out the implicit def/use renderers to minimize the code duplication when we implement that. llvm-svn: 317025
2017-11-01 03:09:29 +08:00
}
void GlobalISelEmitter::importExplicitDefRenderers(
BuildMIAction &DstMIBuilder) {
const CodeGenInstruction *DstI = DstMIBuilder.getCGI();
[globalisel][tablegen] Add support for COPY_TO_REGCLASS. Summary: As part of this * Emitted instructions now have named MachineInstr variables associated with them. This isn't particularly important yet but it's a small step towards multiple-insn emission. * constrainSelectedInstRegOperands() is no longer hardcoded. It's now added as the ConstrainOperandsToDefinitionAction() action. COPY_TO_REGCLASS uses an alternate constraint mechanism ConstrainOperandToRegClassAction() which supports arbitrary constraints such as that defined by COPY_TO_REGCLASS. Reviewers: ab, qcolombet, t.p.northover, rovka, kristof.beyls, aditya_nandakumar Reviewed By: ab Subscribers: javed.absar, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D33590 llvm-svn: 305791
2017-06-20 20:36:34 +08:00
for (unsigned I = 0; I < DstI->Operands.NumDefs; ++I) {
const CGIOperandList::OperandInfo &DstIOperand = DstI->Operands[I];
[globalisel][tablegen] Stop hard-coding the emitted instruction ID to 0. NFC The next commit will add support for multi-instruction emission so we need to start allocating instruction ID's instead of hard-coding them to 0. llvm-svn: 317057
2017-11-01 08:29:47 +08:00
DstMIBuilder.addRenderer<CopyRenderer>(DstIOperand.Name);
[globalisel] Decouple src pattern operands from dst pattern operands. Summary: This isn't testable for AArch64 by itself so this patch also adds support for constant immediates in the pattern and physical register uses in the result. The new IntOperandMatcher matches the constant in patterns such as '(set $rd:GPR32, (G_XOR $rs:GPR32, -1))'. It's always safe to fold immediates into an instruction so this is the first rule that will match across multiple BB's. The Renderer hierarchy is responsible for adding operands to the result instruction. Renderers can copy operands (CopyRenderer) or add physical registers (in particular %wzr and %xzr) to the result instruction in any order (OperandMatchers now import the operand names from SelectionDAG to allow renderers to access any operand). This allows us to emit the result instruction for: %1 = G_XOR %0, -1 --> %1 = ORNWrr %wzr, %0 %1 = G_XOR -1, %0 --> %1 = ORNWrr %wzr, %0 although the latter is untested since the matcher/importer has not been taught about commutativity yet. Added BuildMIAction which can build new instructions and mutate them where possible. W.r.t the mutation aspect, MatchActions are now told the name of an instruction they can recycle and BuildMIAction will emit mutation code when the renderers are appropriate. They are appropriate when all operands are rendered using CopyRenderer and the indices are the same as the matcher. This currently assumes that all operands have at least one matcher. Finally, this change also fixes a crash in AArch64InstructionSelector::select() caused by an immediate operand passing isImm() rather than isCImm(). This was uncovered by the other changes and was detected by existing tests. Depends on D29711 Reviewers: t.p.northover, ab, qcolombet, rovka, aditya_nandakumar, javed.absar Reviewed By: rovka Subscribers: aemerson, dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D29712 llvm-svn: 296131
2017-02-24 23:43:30 +08:00
}
[globalisel][tablegen] Factor out implicit def/use renderers from createAndImportInstructionRenderer(). NFC Multi-instruction emission will require that we have separate handling for the defs between the implicitly created temporaries and the rule outputs. The former require new temporary vregs while the latter should copy existing operands. Factor out the implicit def/use renderers to minimize the code duplication when we implement that. llvm-svn: 317025
2017-11-01 03:09:29 +08:00
}
Re-commit: [globalisel][tablegen] Keep track of the insertion point while adding BuildMIAction's. NFC Multi-instruction emission needs to ensure the the instructions are generated a depth-first fashion. For example: (ADDWrr (SUBWrr a, b), c) needs to emit the SUBWrr before the ADDWrr. However, our walk over TreePatternNode's is highly context sensitive which makes it difficult to append BuildMIActions in the order we want. To fix this, we now keep track of the insertion point as we add actions. This will allow multi-insn emission to insert BuildMI's in the correct place. The previous commit failed on the Ubuntu bots using GCC 4.8. These bots lack the const_iterator forms of insert() and emplace() that were added in C++11. As a result I've switched the const_iterators to iterators. llvm-svn: 317049
2017-11-01 07:03:18 +08:00
Expected<action_iterator> GlobalISelEmitter::importExplicitUseRenderers(
action_iterator InsertPt, RuleMatcher &M, BuildMIAction &DstMIBuilder,
[globalisel][tablegen] Factor out implicit def/use renderers from createAndImportInstructionRenderer(). NFC Multi-instruction emission will require that we have separate handling for the defs between the implicitly created temporaries and the rule outputs. The former require new temporary vregs while the latter should copy existing operands. Factor out the implicit def/use renderers to minimize the code duplication when we implement that. llvm-svn: 317025
2017-11-01 03:09:29 +08:00
const llvm::TreePatternNode *Dst) {
const CodeGenInstruction *DstI = DstMIBuilder.getCGI();
CodeGenInstruction *OrigDstI = &Target.getInstruction(Dst->getOperator());
[globalisel] Decouple src pattern operands from dst pattern operands. Summary: This isn't testable for AArch64 by itself so this patch also adds support for constant immediates in the pattern and physical register uses in the result. The new IntOperandMatcher matches the constant in patterns such as '(set $rd:GPR32, (G_XOR $rs:GPR32, -1))'. It's always safe to fold immediates into an instruction so this is the first rule that will match across multiple BB's. The Renderer hierarchy is responsible for adding operands to the result instruction. Renderers can copy operands (CopyRenderer) or add physical registers (in particular %wzr and %xzr) to the result instruction in any order (OperandMatchers now import the operand names from SelectionDAG to allow renderers to access any operand). This allows us to emit the result instruction for: %1 = G_XOR %0, -1 --> %1 = ORNWrr %wzr, %0 %1 = G_XOR -1, %0 --> %1 = ORNWrr %wzr, %0 although the latter is untested since the matcher/importer has not been taught about commutativity yet. Added BuildMIAction which can build new instructions and mutate them where possible. W.r.t the mutation aspect, MatchActions are now told the name of an instruction they can recycle and BuildMIAction will emit mutation code when the renderers are appropriate. They are appropriate when all operands are rendered using CopyRenderer and the indices are the same as the matcher. This currently assumes that all operands have at least one matcher. Finally, this change also fixes a crash in AArch64InstructionSelector::select() caused by an immediate operand passing isImm() rather than isCImm(). This was uncovered by the other changes and was detected by existing tests. Depends on D29711 Reviewers: t.p.northover, ab, qcolombet, rovka, aditya_nandakumar, javed.absar Reviewed By: rovka Subscribers: aemerson, dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D29712 llvm-svn: 296131
2017-02-24 23:43:30 +08:00
[globalisel][tablegen] Add support for EXTRACT_SUBREG. Summary: After this patch, we finally have test cases that require multiple instruction emission. Depends on D33590 Reviewers: ab, qcolombet, t.p.northover, rovka, kristof.beyls Subscribers: javed.absar, llvm-commits, igorb Differential Revision: https://reviews.llvm.org/D33596 llvm-svn: 306388
2017-06-27 18:11:39 +08:00
// EXTRACT_SUBREG needs to use a subregister COPY.
[globalisel][tablegen] Factor out implicit def/use renderers from createAndImportInstructionRenderer(). NFC Multi-instruction emission will require that we have separate handling for the defs between the implicitly created temporaries and the rule outputs. The former require new temporary vregs while the latter should copy existing operands. Factor out the implicit def/use renderers to minimize the code duplication when we implement that. llvm-svn: 317025
2017-11-01 03:09:29 +08:00
if (OrigDstI->TheDef->getName() == "EXTRACT_SUBREG") {
[globalisel][tablegen] Add support for EXTRACT_SUBREG. Summary: After this patch, we finally have test cases that require multiple instruction emission. Depends on D33590 Reviewers: ab, qcolombet, t.p.northover, rovka, kristof.beyls Subscribers: javed.absar, llvm-commits, igorb Differential Revision: https://reviews.llvm.org/D33596 llvm-svn: 306388
2017-06-27 18:11:39 +08:00
if (!Dst->getChild(0)->isLeaf())
return failedImport("EXTRACT_SUBREG child #1 is not a leaf");
[globalisel][tablegen] Multiple 80-col corrections. llvm-svn: 306544
2017-06-28 21:50:04 +08:00
if (DefInit *SubRegInit =
dyn_cast<DefInit>(Dst->getChild(1)->getLeafValue())) {
[globalisel][tablegen] Add support for multi-insn emission The importer will now accept nested instructions in the result pattern such as (ADDWrr $a, (SUBWrr $b, $c)). This is only valid when the nested instruction def's a single vreg and the parent instruction consumes a single vreg where a nested instruction is specified. The importer will automatically create a vreg to connect the two using the type information from the pattern. This vreg will be constrained to the register classes given in the instruction definitions*. * REG_SEQUENCE is explicitly rejected because of this. The definition doesn't constrain to a register class and it therefore needs special handling. llvm-svn: 317117
2017-11-02 03:57:57 +08:00
Record *RCDef = getInitValueAsRegClass(Dst->getChild(0)->getLeafValue());
if (!RCDef)
return failedImport("EXTRACT_SUBREG child #0 could not "
"be coerced to a register class");
CodeGenRegisterClass *RC = CGRegs.getRegClass(RCDef);
[globalisel][tablegen] Add support for EXTRACT_SUBREG. Summary: After this patch, we finally have test cases that require multiple instruction emission. Depends on D33590 Reviewers: ab, qcolombet, t.p.northover, rovka, kristof.beyls Subscribers: javed.absar, llvm-commits, igorb Differential Revision: https://reviews.llvm.org/D33596 llvm-svn: 306388
2017-06-27 18:11:39 +08:00
CodeGenSubRegIndex *SubIdx = CGRegs.getSubRegIdx(SubRegInit->getDef());
const auto &SrcRCDstRCPair =
RC->getMatchingSubClassWithSubRegs(CGRegs, SubIdx);
if (SrcRCDstRCPair.hasValue()) {
assert(SrcRCDstRCPair->second && "Couldn't find a matching subclass");
if (SrcRCDstRCPair->first != RC)
return failedImport("EXTRACT_SUBREG requires an additional COPY");
}
[globalisel][tablegen] Stop hard-coding the emitted instruction ID to 0. NFC The next commit will add support for multi-instruction emission so we need to start allocating instruction ID's instead of hard-coding them to 0. llvm-svn: 317057
2017-11-01 08:29:47 +08:00
DstMIBuilder.addRenderer<CopySubRegRenderer>(Dst->getChild(0)->getName(),
SubIdx);
Re-commit: [globalisel][tablegen] Keep track of the insertion point while adding BuildMIAction's. NFC Multi-instruction emission needs to ensure the the instructions are generated a depth-first fashion. For example: (ADDWrr (SUBWrr a, b), c) needs to emit the SUBWrr before the ADDWrr. However, our walk over TreePatternNode's is highly context sensitive which makes it difficult to append BuildMIActions in the order we want. To fix this, we now keep track of the insertion point as we add actions. This will allow multi-insn emission to insert BuildMI's in the correct place. The previous commit failed on the Ubuntu bots using GCC 4.8. These bots lack the const_iterator forms of insert() and emplace() that were added in C++11. As a result I've switched the const_iterators to iterators. llvm-svn: 317049
2017-11-01 07:03:18 +08:00
return InsertPt;
[globalisel][tablegen] Add support for EXTRACT_SUBREG. Summary: After this patch, we finally have test cases that require multiple instruction emission. Depends on D33590 Reviewers: ab, qcolombet, t.p.northover, rovka, kristof.beyls Subscribers: javed.absar, llvm-commits, igorb Differential Revision: https://reviews.llvm.org/D33596 llvm-svn: 306388
2017-06-27 18:11:39 +08:00
}
return failedImport("EXTRACT_SUBREG child #1 is not a subreg index");
}
[tablegen][globalisel] Convert the SelectionDAG importer to a tree walking approach. NFC Summary: But don't actually inspect the tree any deeper than we already do. This change is NFC but the next one will enable full traversal of the source/destination patterns. Depends on D30535 Reviewers: t.p.northover, qcolombet, aditya_nandakumar, rovka, ab Subscribers: igorb, dberris, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D30536 llvm-svn: 299001
2017-03-29 23:37:18 +08:00
// Render the explicit uses.
[globalisel][tablegen] Factor out implicit def/use renderers from createAndImportInstructionRenderer(). NFC Multi-instruction emission will require that we have separate handling for the defs between the implicitly created temporaries and the rule outputs. The former require new temporary vregs while the latter should copy existing operands. Factor out the implicit def/use renderers to minimize the code duplication when we implement that. llvm-svn: 317025
2017-11-01 03:09:29 +08:00
unsigned DstINumUses = OrigDstI->Operands.size() - OrigDstI->Operands.NumDefs;
unsigned ExpectedDstINumUses = Dst->getNumChildren();
if (OrigDstI->TheDef->getName() == "COPY_TO_REGCLASS") {
DstINumUses--; // Ignore the class constraint.
ExpectedDstINumUses--;
}
[globalisel][tablegen] Add experimental support for OperandWithDefaultOps, PredicateOperand, and OptionalDefOperand Summary: As far as instruction selection is concerned, all three appear to be same thing. Support for these operands is experimental since AArch64 doesn't make use of them and the in-tree targets that do use them (AMDGPU for OperandWithDefaultOps, AMDGPU/ARM/Hexagon/Lanai for PredicateOperand, and ARM for OperandWithDefaultOps) are not using tablegen-erated GlobalISel yet. Reviewers: rovka, aditya_nandakumar, t.p.northover, qcolombet, ab Reviewed By: rovka Subscribers: inglorion, aemerson, rengolin, mehdi_amini, dberris, kristof.beyls, igorb, tpr, llvm-commits Differential Revision: https://reviews.llvm.org/D31135 llvm-svn: 300037
2017-04-12 16:23:08 +08:00
unsigned Child = 0;
[GlobalISel][TableGen] Fix handling of default operands When looping through a destination pattern's operands to decide how many default operands we need to introduce, we used to count the "expanded" number of operands. So if one default operand would be rendered as 2 values, we'd count it as 2 operands, when in fact it needs to count as only 1 operand regardless of how many values it expands to. This turns out to be a problem only in some very specific cases, e.g. when we have one operand with multiple default values followed by more operands with default values (see the new test). In such a situation we'd stop looping before looking at all the operands, and then error out assuming that we don't have enough default operands to make up the shortfall. At the moment this only affects ARM. The patch removes the loop counting default operands entirely and assumes that we'll have to introduce values for any default operand that we find (i.e. we're assuming it cannot be given as a child at all). It also extracts the code for adding renderers for default operands into a helper method. Differential Revision: https://reviews.llvm.org/D33031 llvm-svn: 303240
2017-05-17 16:57:28 +08:00
unsigned NumDefaultOps = 0;
[globalisel][tablegen] Add experimental support for OperandWithDefaultOps, PredicateOperand, and OptionalDefOperand Summary: As far as instruction selection is concerned, all three appear to be same thing. Support for these operands is experimental since AArch64 doesn't make use of them and the in-tree targets that do use them (AMDGPU for OperandWithDefaultOps, AMDGPU/ARM/Hexagon/Lanai for PredicateOperand, and ARM for OperandWithDefaultOps) are not using tablegen-erated GlobalISel yet. Reviewers: rovka, aditya_nandakumar, t.p.northover, qcolombet, ab Reviewed By: rovka Subscribers: inglorion, aemerson, rengolin, mehdi_amini, dberris, kristof.beyls, igorb, tpr, llvm-commits Differential Revision: https://reviews.llvm.org/D31135 llvm-svn: 300037
2017-04-12 16:23:08 +08:00
for (unsigned I = 0; I != DstINumUses; ++I) {
[globalisel][tablegen] Add support for COPY_TO_REGCLASS. Summary: As part of this * Emitted instructions now have named MachineInstr variables associated with them. This isn't particularly important yet but it's a small step towards multiple-insn emission. * constrainSelectedInstRegOperands() is no longer hardcoded. It's now added as the ConstrainOperandsToDefinitionAction() action. COPY_TO_REGCLASS uses an alternate constraint mechanism ConstrainOperandToRegClassAction() which supports arbitrary constraints such as that defined by COPY_TO_REGCLASS. Reviewers: ab, qcolombet, t.p.northover, rovka, kristof.beyls, aditya_nandakumar Reviewed By: ab Subscribers: javed.absar, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D33590 llvm-svn: 305791
2017-06-20 20:36:34 +08:00
const CGIOperandList::OperandInfo &DstIOperand =
DstI->Operands[DstI->Operands.NumDefs + I];
[globalisel][tablegen] Add experimental support for OperandWithDefaultOps, PredicateOperand, and OptionalDefOperand Summary: As far as instruction selection is concerned, all three appear to be same thing. Support for these operands is experimental since AArch64 doesn't make use of them and the in-tree targets that do use them (AMDGPU for OperandWithDefaultOps, AMDGPU/ARM/Hexagon/Lanai for PredicateOperand, and ARM for OperandWithDefaultOps) are not using tablegen-erated GlobalISel yet. Reviewers: rovka, aditya_nandakumar, t.p.northover, qcolombet, ab Reviewed By: rovka Subscribers: inglorion, aemerson, rengolin, mehdi_amini, dberris, kristof.beyls, igorb, tpr, llvm-commits Differential Revision: https://reviews.llvm.org/D31135 llvm-svn: 300037
2017-04-12 16:23:08 +08:00
[GlobalISel][TableGen] Fix handling of default operands When looping through a destination pattern's operands to decide how many default operands we need to introduce, we used to count the "expanded" number of operands. So if one default operand would be rendered as 2 values, we'd count it as 2 operands, when in fact it needs to count as only 1 operand regardless of how many values it expands to. This turns out to be a problem only in some very specific cases, e.g. when we have one operand with multiple default values followed by more operands with default values (see the new test). In such a situation we'd stop looping before looking at all the operands, and then error out assuming that we don't have enough default operands to make up the shortfall. At the moment this only affects ARM. The patch removes the loop counting default operands entirely and assumes that we'll have to introduce values for any default operand that we find (i.e. we're assuming it cannot be given as a child at all). It also extracts the code for adding renderers for default operands into a helper method. Differential Revision: https://reviews.llvm.org/D33031 llvm-svn: 303240
2017-05-17 16:57:28 +08:00
// If the operand has default values, introduce them now.
// FIXME: Until we have a decent test case that dictates we should do
// otherwise, we're going to assume that operands with default values cannot
// be specified in the patterns. Therefore, adding them will not cause us to
// end up with too many rendered operands.
if (DstIOperand.Rec->isSubClassOf("OperandWithDefaultOps")) {
[globalisel][tablegen] Add experimental support for OperandWithDefaultOps, PredicateOperand, and OptionalDefOperand Summary: As far as instruction selection is concerned, all three appear to be same thing. Support for these operands is experimental since AArch64 doesn't make use of them and the in-tree targets that do use them (AMDGPU for OperandWithDefaultOps, AMDGPU/ARM/Hexagon/Lanai for PredicateOperand, and ARM for OperandWithDefaultOps) are not using tablegen-erated GlobalISel yet. Reviewers: rovka, aditya_nandakumar, t.p.northover, qcolombet, ab Reviewed By: rovka Subscribers: inglorion, aemerson, rengolin, mehdi_amini, dberris, kristof.beyls, igorb, tpr, llvm-commits Differential Revision: https://reviews.llvm.org/D31135 llvm-svn: 300037
2017-04-12 16:23:08 +08:00
DagInit *DefaultOps = DstIOperand.Rec->getValueAsDag("DefaultOps");
[GlobalISel][TableGen] Fix handling of default operands When looping through a destination pattern's operands to decide how many default operands we need to introduce, we used to count the "expanded" number of operands. So if one default operand would be rendered as 2 values, we'd count it as 2 operands, when in fact it needs to count as only 1 operand regardless of how many values it expands to. This turns out to be a problem only in some very specific cases, e.g. when we have one operand with multiple default values followed by more operands with default values (see the new test). In such a situation we'd stop looping before looking at all the operands, and then error out assuming that we don't have enough default operands to make up the shortfall. At the moment this only affects ARM. The patch removes the loop counting default operands entirely and assumes that we'll have to introduce values for any default operand that we find (i.e. we're assuming it cannot be given as a child at all). It also extracts the code for adding renderers for default operands into a helper method. Differential Revision: https://reviews.llvm.org/D33031 llvm-svn: 303240
2017-05-17 16:57:28 +08:00
if (auto Error = importDefaultOperandRenderers(DstMIBuilder, DefaultOps))
return std::move(Error);
++NumDefaultOps;
[globalisel][tablegen] Add experimental support for OperandWithDefaultOps, PredicateOperand, and OptionalDefOperand Summary: As far as instruction selection is concerned, all three appear to be same thing. Support for these operands is experimental since AArch64 doesn't make use of them and the in-tree targets that do use them (AMDGPU for OperandWithDefaultOps, AMDGPU/ARM/Hexagon/Lanai for PredicateOperand, and ARM for OperandWithDefaultOps) are not using tablegen-erated GlobalISel yet. Reviewers: rovka, aditya_nandakumar, t.p.northover, qcolombet, ab Reviewed By: rovka Subscribers: inglorion, aemerson, rengolin, mehdi_amini, dberris, kristof.beyls, igorb, tpr, llvm-commits Differential Revision: https://reviews.llvm.org/D31135 llvm-svn: 300037
2017-04-12 16:23:08 +08:00
continue;
}
Re-commit: [globalisel][tablegen] Keep track of the insertion point while adding BuildMIAction's. NFC Multi-instruction emission needs to ensure the the instructions are generated a depth-first fashion. For example: (ADDWrr (SUBWrr a, b), c) needs to emit the SUBWrr before the ADDWrr. However, our walk over TreePatternNode's is highly context sensitive which makes it difficult to append BuildMIActions in the order we want. To fix this, we now keep track of the insertion point as we add actions. This will allow multi-insn emission to insert BuildMI's in the correct place. The previous commit failed on the Ubuntu bots using GCC 4.8. These bots lack the const_iterator forms of insert() and emplace() that were added in C++11. As a result I've switched the const_iterators to iterators. llvm-svn: 317049
2017-11-01 07:03:18 +08:00
auto InsertPtOrError = importExplicitUseRenderer(InsertPt, M, DstMIBuilder,
Dst->getChild(Child));
if (auto Error = InsertPtOrError.takeError())
[tablegen][globalisel] Convert the SelectionDAG importer to a tree walking approach. NFC Summary: But don't actually inspect the tree any deeper than we already do. This change is NFC but the next one will enable full traversal of the source/destination patterns. Depends on D30535 Reviewers: t.p.northover, qcolombet, aditya_nandakumar, rovka, ab Subscribers: igorb, dberris, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D30536 llvm-svn: 299001
2017-03-29 23:37:18 +08:00
return std::move(Error);
Re-commit: [globalisel][tablegen] Keep track of the insertion point while adding BuildMIAction's. NFC Multi-instruction emission needs to ensure the the instructions are generated a depth-first fashion. For example: (ADDWrr (SUBWrr a, b), c) needs to emit the SUBWrr before the ADDWrr. However, our walk over TreePatternNode's is highly context sensitive which makes it difficult to append BuildMIActions in the order we want. To fix this, we now keep track of the insertion point as we add actions. This will allow multi-insn emission to insert BuildMI's in the correct place. The previous commit failed on the Ubuntu bots using GCC 4.8. These bots lack the const_iterator forms of insert() and emplace() that were added in C++11. As a result I've switched the const_iterators to iterators. llvm-svn: 317049
2017-11-01 07:03:18 +08:00
InsertPt = InsertPtOrError.get();
[globalisel][tablegen] Add experimental support for OperandWithDefaultOps, PredicateOperand, and OptionalDefOperand Summary: As far as instruction selection is concerned, all three appear to be same thing. Support for these operands is experimental since AArch64 doesn't make use of them and the in-tree targets that do use them (AMDGPU for OperandWithDefaultOps, AMDGPU/ARM/Hexagon/Lanai for PredicateOperand, and ARM for OperandWithDefaultOps) are not using tablegen-erated GlobalISel yet. Reviewers: rovka, aditya_nandakumar, t.p.northover, qcolombet, ab Reviewed By: rovka Subscribers: inglorion, aemerson, rengolin, mehdi_amini, dberris, kristof.beyls, igorb, tpr, llvm-commits Differential Revision: https://reviews.llvm.org/D31135 llvm-svn: 300037
2017-04-12 16:23:08 +08:00
++Child;
[tablegen][globalisel] Convert the SelectionDAG importer to a tree walking approach. NFC Summary: But don't actually inspect the tree any deeper than we already do. This change is NFC but the next one will enable full traversal of the source/destination patterns. Depends on D30535 Reviewers: t.p.northover, qcolombet, aditya_nandakumar, rovka, ab Subscribers: igorb, dberris, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D30536 llvm-svn: 299001
2017-03-29 23:37:18 +08:00
}
[globalisel] Decouple src pattern operands from dst pattern operands. Summary: This isn't testable for AArch64 by itself so this patch also adds support for constant immediates in the pattern and physical register uses in the result. The new IntOperandMatcher matches the constant in patterns such as '(set $rd:GPR32, (G_XOR $rs:GPR32, -1))'. It's always safe to fold immediates into an instruction so this is the first rule that will match across multiple BB's. The Renderer hierarchy is responsible for adding operands to the result instruction. Renderers can copy operands (CopyRenderer) or add physical registers (in particular %wzr and %xzr) to the result instruction in any order (OperandMatchers now import the operand names from SelectionDAG to allow renderers to access any operand). This allows us to emit the result instruction for: %1 = G_XOR %0, -1 --> %1 = ORNWrr %wzr, %0 %1 = G_XOR -1, %0 --> %1 = ORNWrr %wzr, %0 although the latter is untested since the matcher/importer has not been taught about commutativity yet. Added BuildMIAction which can build new instructions and mutate them where possible. W.r.t the mutation aspect, MatchActions are now told the name of an instruction they can recycle and BuildMIAction will emit mutation code when the renderers are appropriate. They are appropriate when all operands are rendered using CopyRenderer and the indices are the same as the matcher. This currently assumes that all operands have at least one matcher. Finally, this change also fixes a crash in AArch64InstructionSelector::select() caused by an immediate operand passing isImm() rather than isCImm(). This was uncovered by the other changes and was detected by existing tests. Depends on D29711 Reviewers: t.p.northover, ab, qcolombet, rovka, aditya_nandakumar, javed.absar Reviewed By: rovka Subscribers: aemerson, dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D29712 llvm-svn: 296131
2017-02-24 23:43:30 +08:00
[globalisel][tablegen] Add support for COPY_TO_REGCLASS. Summary: As part of this * Emitted instructions now have named MachineInstr variables associated with them. This isn't particularly important yet but it's a small step towards multiple-insn emission. * constrainSelectedInstRegOperands() is no longer hardcoded. It's now added as the ConstrainOperandsToDefinitionAction() action. COPY_TO_REGCLASS uses an alternate constraint mechanism ConstrainOperandToRegClassAction() which supports arbitrary constraints such as that defined by COPY_TO_REGCLASS. Reviewers: ab, qcolombet, t.p.northover, rovka, kristof.beyls, aditya_nandakumar Reviewed By: ab Subscribers: javed.absar, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D33590 llvm-svn: 305791
2017-06-20 20:36:34 +08:00
if (NumDefaultOps + ExpectedDstINumUses != DstINumUses)
Fixup r303240: Use llvm::to_string instead of std::to_string It turns out some of the buildbots don't have std::to_string around, even in this day and age... llvm-svn: 303243
2017-05-17 17:25:08 +08:00
return failedImport("Expected " + llvm::to_string(DstINumUses) +
[GlobalISel][TableGen] Fix handling of default operands When looping through a destination pattern's operands to decide how many default operands we need to introduce, we used to count the "expanded" number of operands. So if one default operand would be rendered as 2 values, we'd count it as 2 operands, when in fact it needs to count as only 1 operand regardless of how many values it expands to. This turns out to be a problem only in some very specific cases, e.g. when we have one operand with multiple default values followed by more operands with default values (see the new test). In such a situation we'd stop looping before looking at all the operands, and then error out assuming that we don't have enough default operands to make up the shortfall. At the moment this only affects ARM. The patch removes the loop counting default operands entirely and assumes that we'll have to introduce values for any default operand that we find (i.e. we're assuming it cannot be given as a child at all). It also extracts the code for adding renderers for default operands into a helper method. Differential Revision: https://reviews.llvm.org/D33031 llvm-svn: 303240
2017-05-17 16:57:28 +08:00
" used operands but found " +
[globalisel][tablegen] Add support for COPY_TO_REGCLASS. Summary: As part of this * Emitted instructions now have named MachineInstr variables associated with them. This isn't particularly important yet but it's a small step towards multiple-insn emission. * constrainSelectedInstRegOperands() is no longer hardcoded. It's now added as the ConstrainOperandsToDefinitionAction() action. COPY_TO_REGCLASS uses an alternate constraint mechanism ConstrainOperandToRegClassAction() which supports arbitrary constraints such as that defined by COPY_TO_REGCLASS. Reviewers: ab, qcolombet, t.p.northover, rovka, kristof.beyls, aditya_nandakumar Reviewed By: ab Subscribers: javed.absar, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D33590 llvm-svn: 305791
2017-06-20 20:36:34 +08:00
llvm::to_string(ExpectedDstINumUses) +
Fixup r303240: Use llvm::to_string instead of std::to_string It turns out some of the buildbots don't have std::to_string around, even in this day and age... llvm-svn: 303243
2017-05-17 17:25:08 +08:00
" explicit ones and " + llvm::to_string(NumDefaultOps) +
[GlobalISel][TableGen] Fix handling of default operands When looping through a destination pattern's operands to decide how many default operands we need to introduce, we used to count the "expanded" number of operands. So if one default operand would be rendered as 2 values, we'd count it as 2 operands, when in fact it needs to count as only 1 operand regardless of how many values it expands to. This turns out to be a problem only in some very specific cases, e.g. when we have one operand with multiple default values followed by more operands with default values (see the new test). In such a situation we'd stop looping before looking at all the operands, and then error out assuming that we don't have enough default operands to make up the shortfall. At the moment this only affects ARM. The patch removes the loop counting default operands entirely and assumes that we'll have to introduce values for any default operand that we find (i.e. we're assuming it cannot be given as a child at all). It also extracts the code for adding renderers for default operands into a helper method. Differential Revision: https://reviews.llvm.org/D33031 llvm-svn: 303240
2017-05-17 16:57:28 +08:00
" default ones");
Re-commit: [globalisel][tablegen] Keep track of the insertion point while adding BuildMIAction's. NFC Multi-instruction emission needs to ensure the the instructions are generated a depth-first fashion. For example: (ADDWrr (SUBWrr a, b), c) needs to emit the SUBWrr before the ADDWrr. However, our walk over TreePatternNode's is highly context sensitive which makes it difficult to append BuildMIActions in the order we want. To fix this, we now keep track of the insertion point as we add actions. This will allow multi-insn emission to insert BuildMI's in the correct place. The previous commit failed on the Ubuntu bots using GCC 4.8. These bots lack the const_iterator forms of insert() and emplace() that were added in C++11. As a result I've switched the const_iterators to iterators. llvm-svn: 317049
2017-11-01 07:03:18 +08:00
return InsertPt;
[tablegen][globalisel] Convert the SelectionDAG importer to a tree walking approach. NFC Summary: But don't actually inspect the tree any deeper than we already do. This change is NFC but the next one will enable full traversal of the source/destination patterns. Depends on D30535 Reviewers: t.p.northover, qcolombet, aditya_nandakumar, rovka, ab Subscribers: igorb, dberris, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D30536 llvm-svn: 299001
2017-03-29 23:37:18 +08:00
}
[globalisel] Decouple src pattern operands from dst pattern operands. Summary: This isn't testable for AArch64 by itself so this patch also adds support for constant immediates in the pattern and physical register uses in the result. The new IntOperandMatcher matches the constant in patterns such as '(set $rd:GPR32, (G_XOR $rs:GPR32, -1))'. It's always safe to fold immediates into an instruction so this is the first rule that will match across multiple BB's. The Renderer hierarchy is responsible for adding operands to the result instruction. Renderers can copy operands (CopyRenderer) or add physical registers (in particular %wzr and %xzr) to the result instruction in any order (OperandMatchers now import the operand names from SelectionDAG to allow renderers to access any operand). This allows us to emit the result instruction for: %1 = G_XOR %0, -1 --> %1 = ORNWrr %wzr, %0 %1 = G_XOR -1, %0 --> %1 = ORNWrr %wzr, %0 although the latter is untested since the matcher/importer has not been taught about commutativity yet. Added BuildMIAction which can build new instructions and mutate them where possible. W.r.t the mutation aspect, MatchActions are now told the name of an instruction they can recycle and BuildMIAction will emit mutation code when the renderers are appropriate. They are appropriate when all operands are rendered using CopyRenderer and the indices are the same as the matcher. This currently assumes that all operands have at least one matcher. Finally, this change also fixes a crash in AArch64InstructionSelector::select() caused by an immediate operand passing isImm() rather than isCImm(). This was uncovered by the other changes and was detected by existing tests. Depends on D29711 Reviewers: t.p.northover, ab, qcolombet, rovka, aditya_nandakumar, javed.absar Reviewed By: rovka Subscribers: aemerson, dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D29712 llvm-svn: 296131
2017-02-24 23:43:30 +08:00
[GlobalISel][TableGen] Fix handling of default operands When looping through a destination pattern's operands to decide how many default operands we need to introduce, we used to count the "expanded" number of operands. So if one default operand would be rendered as 2 values, we'd count it as 2 operands, when in fact it needs to count as only 1 operand regardless of how many values it expands to. This turns out to be a problem only in some very specific cases, e.g. when we have one operand with multiple default values followed by more operands with default values (see the new test). In such a situation we'd stop looping before looking at all the operands, and then error out assuming that we don't have enough default operands to make up the shortfall. At the moment this only affects ARM. The patch removes the loop counting default operands entirely and assumes that we'll have to introduce values for any default operand that we find (i.e. we're assuming it cannot be given as a child at all). It also extracts the code for adding renderers for default operands into a helper method. Differential Revision: https://reviews.llvm.org/D33031 llvm-svn: 303240
2017-05-17 16:57:28 +08:00
Error GlobalISelEmitter::importDefaultOperandRenderers(
BuildMIAction &DstMIBuilder, DagInit *DefaultOps) const {
[TableGen] Introduce DagInit::getArgs that returns an ArrayRef. Use it to fix 80 column violations in arg_begin/arg_end. Remove DagInit::args and use getArgs instead. NFC llvm-svn: 304177
2017-05-30 05:49:34 +08:00
for (const auto *DefaultOp : DefaultOps->getArgs()) {
[GlobalISel][TableGen] Fix handling of default operands When looping through a destination pattern's operands to decide how many default operands we need to introduce, we used to count the "expanded" number of operands. So if one default operand would be rendered as 2 values, we'd count it as 2 operands, when in fact it needs to count as only 1 operand regardless of how many values it expands to. This turns out to be a problem only in some very specific cases, e.g. when we have one operand with multiple default values followed by more operands with default values (see the new test). In such a situation we'd stop looping before looking at all the operands, and then error out assuming that we don't have enough default operands to make up the shortfall. At the moment this only affects ARM. The patch removes the loop counting default operands entirely and assumes that we'll have to introduce values for any default operand that we find (i.e. we're assuming it cannot be given as a child at all). It also extracts the code for adding renderers for default operands into a helper method. Differential Revision: https://reviews.llvm.org/D33031 llvm-svn: 303240
2017-05-17 16:57:28 +08:00
// Look through ValueType operators.
if (const DagInit *DefaultDagOp = dyn_cast<DagInit>(DefaultOp)) {
if (const DefInit *DefaultDagOperator =
dyn_cast<DefInit>(DefaultDagOp->getOperator())) {
if (DefaultDagOperator->getDef()->isSubClassOf("ValueType"))
DefaultOp = DefaultDagOp->getArg(0);
}
}
if (const DefInit *DefaultDefOp = dyn_cast<DefInit>(DefaultOp)) {
[globalisel][tablegen] Stop hard-coding the emitted instruction ID to 0. NFC The next commit will add support for multi-instruction emission so we need to start allocating instruction ID's instead of hard-coding them to 0. llvm-svn: 317057
2017-11-01 08:29:47 +08:00
DstMIBuilder.addRenderer<AddRegisterRenderer>(DefaultDefOp->getDef());
[GlobalISel][TableGen] Fix handling of default operands When looping through a destination pattern's operands to decide how many default operands we need to introduce, we used to count the "expanded" number of operands. So if one default operand would be rendered as 2 values, we'd count it as 2 operands, when in fact it needs to count as only 1 operand regardless of how many values it expands to. This turns out to be a problem only in some very specific cases, e.g. when we have one operand with multiple default values followed by more operands with default values (see the new test). In such a situation we'd stop looping before looking at all the operands, and then error out assuming that we don't have enough default operands to make up the shortfall. At the moment this only affects ARM. The patch removes the loop counting default operands entirely and assumes that we'll have to introduce values for any default operand that we find (i.e. we're assuming it cannot be given as a child at all). It also extracts the code for adding renderers for default operands into a helper method. Differential Revision: https://reviews.llvm.org/D33031 llvm-svn: 303240
2017-05-17 16:57:28 +08:00
continue;
}
if (const IntInit *DefaultIntOp = dyn_cast<IntInit>(DefaultOp)) {
[globalisel][tablegen] Stop hard-coding the emitted instruction ID to 0. NFC The next commit will add support for multi-instruction emission so we need to start allocating instruction ID's instead of hard-coding them to 0. llvm-svn: 317057
2017-11-01 08:29:47 +08:00
DstMIBuilder.addRenderer<ImmRenderer>(DefaultIntOp->getValue());
[GlobalISel][TableGen] Fix handling of default operands When looping through a destination pattern's operands to decide how many default operands we need to introduce, we used to count the "expanded" number of operands. So if one default operand would be rendered as 2 values, we'd count it as 2 operands, when in fact it needs to count as only 1 operand regardless of how many values it expands to. This turns out to be a problem only in some very specific cases, e.g. when we have one operand with multiple default values followed by more operands with default values (see the new test). In such a situation we'd stop looping before looking at all the operands, and then error out assuming that we don't have enough default operands to make up the shortfall. At the moment this only affects ARM. The patch removes the loop counting default operands entirely and assumes that we'll have to introduce values for any default operand that we find (i.e. we're assuming it cannot be given as a child at all). It also extracts the code for adding renderers for default operands into a helper method. Differential Revision: https://reviews.llvm.org/D33031 llvm-svn: 303240
2017-05-17 16:57:28 +08:00
continue;
}
return failedImport("Could not add default op");
}
return Error::success();
}
[globalisel][tablegen] Change Expected<bool> to Error and rename functions. Functions that still return Expected<X> are now called createAndImport*() Changing the return type was requested in the review comments for r299001 llvm-svn: 299063
2017-03-30 17:36:33 +08:00
Error GlobalISelEmitter::importImplicitDefRenderers(
[tablegen][globalisel] Convert the SelectionDAG importer to a tree walking approach. NFC Summary: But don't actually inspect the tree any deeper than we already do. This change is NFC but the next one will enable full traversal of the source/destination patterns. Depends on D30535 Reviewers: t.p.northover, qcolombet, aditya_nandakumar, rovka, ab Subscribers: igorb, dberris, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D30536 llvm-svn: 299001
2017-03-29 23:37:18 +08:00
BuildMIAction &DstMIBuilder,
const std::vector<Record *> &ImplicitDefs) const {
if (!ImplicitDefs.empty())
return failedImport("Pattern defines a physical register");
[globalisel][tablegen] Change Expected<bool> to Error and rename functions. Functions that still return Expected<X> are now called createAndImport*() Changing the return type was requested in the review comments for r299001 llvm-svn: 299063
2017-03-30 17:36:33 +08:00
return Error::success();
[tablegen][globalisel] Convert the SelectionDAG importer to a tree walking approach. NFC Summary: But don't actually inspect the tree any deeper than we already do. This change is NFC but the next one will enable full traversal of the source/destination patterns. Depends on D30535 Reviewers: t.p.northover, qcolombet, aditya_nandakumar, rovka, ab Subscribers: igorb, dberris, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D30536 llvm-svn: 299001
2017-03-29 23:37:18 +08:00
}
[globalisel] Decouple src pattern operands from dst pattern operands. Summary: This isn't testable for AArch64 by itself so this patch also adds support for constant immediates in the pattern and physical register uses in the result. The new IntOperandMatcher matches the constant in patterns such as '(set $rd:GPR32, (G_XOR $rs:GPR32, -1))'. It's always safe to fold immediates into an instruction so this is the first rule that will match across multiple BB's. The Renderer hierarchy is responsible for adding operands to the result instruction. Renderers can copy operands (CopyRenderer) or add physical registers (in particular %wzr and %xzr) to the result instruction in any order (OperandMatchers now import the operand names from SelectionDAG to allow renderers to access any operand). This allows us to emit the result instruction for: %1 = G_XOR %0, -1 --> %1 = ORNWrr %wzr, %0 %1 = G_XOR -1, %0 --> %1 = ORNWrr %wzr, %0 although the latter is untested since the matcher/importer has not been taught about commutativity yet. Added BuildMIAction which can build new instructions and mutate them where possible. W.r.t the mutation aspect, MatchActions are now told the name of an instruction they can recycle and BuildMIAction will emit mutation code when the renderers are appropriate. They are appropriate when all operands are rendered using CopyRenderer and the indices are the same as the matcher. This currently assumes that all operands have at least one matcher. Finally, this change also fixes a crash in AArch64InstructionSelector::select() caused by an immediate operand passing isImm() rather than isCImm(). This was uncovered by the other changes and was detected by existing tests. Depends on D29711 Reviewers: t.p.northover, ab, qcolombet, rovka, aditya_nandakumar, javed.absar Reviewed By: rovka Subscribers: aemerson, dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D29712 llvm-svn: 296131
2017-02-24 23:43:30 +08:00
[tablegen][globalisel] Convert the SelectionDAG importer to a tree walking approach. NFC Summary: But don't actually inspect the tree any deeper than we already do. This change is NFC but the next one will enable full traversal of the source/destination patterns. Depends on D30535 Reviewers: t.p.northover, qcolombet, aditya_nandakumar, rovka, ab Subscribers: igorb, dberris, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D30536 llvm-svn: 299001
2017-03-29 23:37:18 +08:00
Expected<RuleMatcher> GlobalISelEmitter::runOnPattern(const PatternToMatch &P) {
// Keep track of the matchers and actions to emit.
[GISel]: Make GlobalISelEmitter rule prioritization compatible with selectionDAG This patch changes GlobalISelEmitter to rank patterns similar to how the DAG does it (ie it computes a score for a pattern and adds the added complexity to it). This is so that the decision tree for GISelSelector remains compatible with that of SelectionDAG. https://reviews.llvm.org/D43270 llvm-svn: 325401
2018-02-17 06:37:15 +08:00
int Score = P.getPatternComplexity(CGP);
[globalisel][tablegen] Simplify named operand/operator lookups and fix a wrong-code bug this revealed. Summary: Operand variable lookups are now performed by the RuleMatcher rather than searching the whole matcher hierarchy for a match. This revealed a wrong-code bug that currently affects ARM and X86 where patterns that use a variable more than once in the match pattern will be imported but won't check that the operands are identical. This can cause the tablegen-erated matcher to accept matches that should be rejected. Depends on D36569 Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Subscribers: aemerson, igorb, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D36618 llvm-svn: 315780
2017-10-14 08:31:58 +08:00
RuleMatcher M(P.getSrcRecord()->getLoc());
[GISel]: Make GlobalISelEmitter rule prioritization compatible with selectionDAG This patch changes GlobalISelEmitter to rank patterns similar to how the DAG does it (ie it computes a score for a pattern and adds the added complexity to it). This is so that the decision tree for GISelSelector remains compatible with that of SelectionDAG. https://reviews.llvm.org/D43270 llvm-svn: 325401
2018-02-17 06:37:15 +08:00
RuleMatcherScores[M.getRuleID()] = Score;
[globalisel][tablegen] Allow any comment in DebugCommentAction. NFC llvm-svn: 317017
2017-11-01 02:07:03 +08:00
M.addAction<DebugCommentAction>(llvm::to_string(*P.getSrcPattern()) +
" => " +
llvm::to_string(*P.getDstPattern()));
[globalisel] Decouple src pattern operands from dst pattern operands. Summary: This isn't testable for AArch64 by itself so this patch also adds support for constant immediates in the pattern and physical register uses in the result. The new IntOperandMatcher matches the constant in patterns such as '(set $rd:GPR32, (G_XOR $rs:GPR32, -1))'. It's always safe to fold immediates into an instruction so this is the first rule that will match across multiple BB's. The Renderer hierarchy is responsible for adding operands to the result instruction. Renderers can copy operands (CopyRenderer) or add physical registers (in particular %wzr and %xzr) to the result instruction in any order (OperandMatchers now import the operand names from SelectionDAG to allow renderers to access any operand). This allows us to emit the result instruction for: %1 = G_XOR %0, -1 --> %1 = ORNWrr %wzr, %0 %1 = G_XOR -1, %0 --> %1 = ORNWrr %wzr, %0 although the latter is untested since the matcher/importer has not been taught about commutativity yet. Added BuildMIAction which can build new instructions and mutate them where possible. W.r.t the mutation aspect, MatchActions are now told the name of an instruction they can recycle and BuildMIAction will emit mutation code when the renderers are appropriate. They are appropriate when all operands are rendered using CopyRenderer and the indices are the same as the matcher. This currently assumes that all operands have at least one matcher. Finally, this change also fixes a crash in AArch64InstructionSelector::select() caused by an immediate operand passing isImm() rather than isCImm(). This was uncovered by the other changes and was detected by existing tests. Depends on D29711 Reviewers: t.p.northover, ab, qcolombet, rovka, aditya_nandakumar, javed.absar Reviewed By: rovka Subscribers: aemerson, dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D29712 llvm-svn: 296131
2017-02-24 23:43:30 +08:00
TableGen support for parameterized register class information This replaces TableGen's type inference to operate on parameterized types instead of MVTs, and as a consequence, some interfaces have changed: - Uses of MVTs are replaced by ValueTypeByHwMode. - EEVT::TypeSet is replaced by TypeSetByHwMode. This affects the way that types and type sets are printed, and the tests relying on that have been updated. There are certain users of the inferred types outside of TableGen itself, namely FastISel and GlobalISel. For those users, the way that the types are accessed have changed. For typical scenarios, these replacements can be used: - TreePatternNode::getType(ResNo) -> getSimpleType(ResNo) - TreePatternNode::hasTypeSet(ResNo) -> hasConcreteType(ResNo) - TypeSet::isConcrete -> TypeSetByHwMode::isValueTypeByHwMode(false) For more information, please refer to the review page. Differential Revision: https://reviews.llvm.org/D31951 llvm-svn: 313271
2017-09-15 00:56:21 +08:00
if (auto Error = importRulePredicates(M, P.getPredicates()))
[tablegen][globalisel] Convert the SelectionDAG importer to a tree walking approach. NFC Summary: But don't actually inspect the tree any deeper than we already do. This change is NFC but the next one will enable full traversal of the source/destination patterns. Depends on D30535 Reviewers: t.p.northover, qcolombet, aditya_nandakumar, rovka, ab Subscribers: igorb, dberris, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D30536 llvm-svn: 299001
2017-03-29 23:37:18 +08:00
return std::move(Error);
[globalisel] Decouple src pattern operands from dst pattern operands. Summary: This isn't testable for AArch64 by itself so this patch also adds support for constant immediates in the pattern and physical register uses in the result. The new IntOperandMatcher matches the constant in patterns such as '(set $rd:GPR32, (G_XOR $rs:GPR32, -1))'. It's always safe to fold immediates into an instruction so this is the first rule that will match across multiple BB's. The Renderer hierarchy is responsible for adding operands to the result instruction. Renderers can copy operands (CopyRenderer) or add physical registers (in particular %wzr and %xzr) to the result instruction in any order (OperandMatchers now import the operand names from SelectionDAG to allow renderers to access any operand). This allows us to emit the result instruction for: %1 = G_XOR %0, -1 --> %1 = ORNWrr %wzr, %0 %1 = G_XOR -1, %0 --> %1 = ORNWrr %wzr, %0 although the latter is untested since the matcher/importer has not been taught about commutativity yet. Added BuildMIAction which can build new instructions and mutate them where possible. W.r.t the mutation aspect, MatchActions are now told the name of an instruction they can recycle and BuildMIAction will emit mutation code when the renderers are appropriate. They are appropriate when all operands are rendered using CopyRenderer and the indices are the same as the matcher. This currently assumes that all operands have at least one matcher. Finally, this change also fixes a crash in AArch64InstructionSelector::select() caused by an immediate operand passing isImm() rather than isCImm(). This was uncovered by the other changes and was detected by existing tests. Depends on D29711 Reviewers: t.p.northover, ab, qcolombet, rovka, aditya_nandakumar, javed.absar Reviewed By: rovka Subscribers: aemerson, dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D29712 llvm-svn: 296131
2017-02-24 23:43:30 +08:00
[tablegen][globalisel] Convert the SelectionDAG importer to a tree walking approach. NFC Summary: But don't actually inspect the tree any deeper than we already do. This change is NFC but the next one will enable full traversal of the source/destination patterns. Depends on D30535 Reviewers: t.p.northover, qcolombet, aditya_nandakumar, rovka, ab Subscribers: igorb, dberris, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D30536 llvm-svn: 299001
2017-03-29 23:37:18 +08:00
// Next, analyze the pattern operators.
TreePatternNode *Src = P.getSrcPattern();
TreePatternNode *Dst = P.getDstPattern();
[globalisel] Decouple src pattern operands from dst pattern operands. Summary: This isn't testable for AArch64 by itself so this patch also adds support for constant immediates in the pattern and physical register uses in the result. The new IntOperandMatcher matches the constant in patterns such as '(set $rd:GPR32, (G_XOR $rs:GPR32, -1))'. It's always safe to fold immediates into an instruction so this is the first rule that will match across multiple BB's. The Renderer hierarchy is responsible for adding operands to the result instruction. Renderers can copy operands (CopyRenderer) or add physical registers (in particular %wzr and %xzr) to the result instruction in any order (OperandMatchers now import the operand names from SelectionDAG to allow renderers to access any operand). This allows us to emit the result instruction for: %1 = G_XOR %0, -1 --> %1 = ORNWrr %wzr, %0 %1 = G_XOR -1, %0 --> %1 = ORNWrr %wzr, %0 although the latter is untested since the matcher/importer has not been taught about commutativity yet. Added BuildMIAction which can build new instructions and mutate them where possible. W.r.t the mutation aspect, MatchActions are now told the name of an instruction they can recycle and BuildMIAction will emit mutation code when the renderers are appropriate. They are appropriate when all operands are rendered using CopyRenderer and the indices are the same as the matcher. This currently assumes that all operands have at least one matcher. Finally, this change also fixes a crash in AArch64InstructionSelector::select() caused by an immediate operand passing isImm() rather than isCImm(). This was uncovered by the other changes and was detected by existing tests. Depends on D29711 Reviewers: t.p.northover, ab, qcolombet, rovka, aditya_nandakumar, javed.absar Reviewed By: rovka Subscribers: aemerson, dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D29712 llvm-svn: 296131
2017-02-24 23:43:30 +08:00
[tablegen][globalisel] Convert the SelectionDAG importer to a tree walking approach. NFC Summary: But don't actually inspect the tree any deeper than we already do. This change is NFC but the next one will enable full traversal of the source/destination patterns. Depends on D30535 Reviewers: t.p.northover, qcolombet, aditya_nandakumar, rovka, ab Subscribers: igorb, dberris, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D30536 llvm-svn: 299001
2017-03-29 23:37:18 +08:00
// If the root of either pattern isn't a simple operator, ignore it.
[globalisel][tablegen] Report more detail in some SelectionDAG import failures. NFC Reviewers: ab, t.p.northover, qcolombet, aditya_nandakumar, rovka Reviewed By: ab Subscribers: dberris, kristof.beyls, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D31325 llvm-svn: 300186
2017-04-13 17:45:37 +08:00
if (auto Err = isTrivialOperatorNode(Dst))
return failedImport("Dst pattern root isn't a trivial operator (" +
toString(std::move(Err)) + ")");
if (auto Err = isTrivialOperatorNode(Src))
return failedImport("Src pattern root isn't a trivial operator (" +
toString(std::move(Err)) + ")");
[globalisel] Decouple src pattern operands from dst pattern operands. Summary: This isn't testable for AArch64 by itself so this patch also adds support for constant immediates in the pattern and physical register uses in the result. The new IntOperandMatcher matches the constant in patterns such as '(set $rd:GPR32, (G_XOR $rs:GPR32, -1))'. It's always safe to fold immediates into an instruction so this is the first rule that will match across multiple BB's. The Renderer hierarchy is responsible for adding operands to the result instruction. Renderers can copy operands (CopyRenderer) or add physical registers (in particular %wzr and %xzr) to the result instruction in any order (OperandMatchers now import the operand names from SelectionDAG to allow renderers to access any operand). This allows us to emit the result instruction for: %1 = G_XOR %0, -1 --> %1 = ORNWrr %wzr, %0 %1 = G_XOR -1, %0 --> %1 = ORNWrr %wzr, %0 although the latter is untested since the matcher/importer has not been taught about commutativity yet. Added BuildMIAction which can build new instructions and mutate them where possible. W.r.t the mutation aspect, MatchActions are now told the name of an instruction they can recycle and BuildMIAction will emit mutation code when the renderers are appropriate. They are appropriate when all operands are rendered using CopyRenderer and the indices are the same as the matcher. This currently assumes that all operands have at least one matcher. Finally, this change also fixes a crash in AArch64InstructionSelector::select() caused by an immediate operand passing isImm() rather than isCImm(). This was uncovered by the other changes and was detected by existing tests. Depends on D29711 Reviewers: t.p.northover, ab, qcolombet, rovka, aditya_nandakumar, javed.absar Reviewed By: rovka Subscribers: aemerson, dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D29712 llvm-svn: 296131
2017-02-24 23:43:30 +08:00
[TableGen][GlobalISel] Have the predicate directly know which data they are dealing with Prior to this patch, a predicate wouldn't make sense outside of its rule. Indeed, it was only during emitting a rule that a predicate would be made aware of the IDs of the data it is checking. Because of that, predicates could not be moved around or compared between each other. NFC. llvm-svn: 320887
2017-12-16 07:07:42 +08:00
// The different predicates and matchers created during
// addInstructionMatcher use the RuleMatcher M to set up their
// instruction ID (InsnVarID) that are going to be used when
// M is going to be emitted.
// However, the code doing the emission still relies on the IDs
// returned during that process by the RuleMatcher when issuing
// the recordInsn opcodes.
// Because of that:
// 1. The order in which we created the predicates
// and such must be the same as the order in which we emit them,
// and
// 2. We need to reset the generation of the IDs in M somewhere between
// addInstructionMatcher and emit
//
// FIXME: Long term, we don't want to have to rely on this implicit
// naming being the same. One possible solution would be to have
// explicit operator for operation capture and reference those.
// The plus side is that it would expose opportunities to share
// the capture accross rules. The downside is that it would
// introduce a dependency between predicates (captures must happen
// before their first use.)
Re-commit: [globalisel][tablegen] Support zero-instruction emission. Summary: Support the case where an operand of a pattern is also the whole of the result pattern. In this case the original result and all its uses must be replaced by the operand. However, register class restrictions can require a COPY. This patch handles both cases by always emitting the copy and leaving it for the register allocator to optimize. The previous commit failed on Windows machines due to a flaw in the sort predicate which allowed both A < B < C and B == C to be satisfied simultaneously. The cause of this was some sloppiness in the priority order of G_CONSTANT instructions compared to other instructions. These had equal priority because it makes no difference, however there were operands had higher priority than G_CONSTANT but lower priority than any other instruction. As a result, a priority order between G_CONSTANT and other instructions must be enforced to ensure the predicate defines a strict weak order. Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Subscribers: javed.absar, kristof.beyls, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D36084 llvm-svn: 311076
2017-08-17 17:26:14 +08:00
InstructionMatcher &InsnMatcherTemp = M.addInstructionMatcher(Src->getName());
unsigned TempOpIdx = 0;
auto InsnMatcherOrError =
Re-commit r315863: [globalisel][tablegen] Import ComplexPattern when used as an operator Summary: It's possible for a ComplexPattern to be used as an operator in a match pattern. This is used by the load/store patterns in AArch64 to name the suboperands returned by ComplexPattern predicate so that they can be broken apart and referenced independently in the result pattern. This patch adds support for this in order to enable the import of load/store patterns. Depends on D37445 Hopefully fixed the ambiguous constructor that a large number of bots reported. Reviewers: ab, qcolombet, t.p.northover, rovka, aditya_nandakumar Reviewed By: qcolombet Subscribers: aemerson, javed.absar, igorb, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D37456 llvm-svn: 315869
2017-10-16 02:22:54 +08:00
createAndImportSelDAGMatcher(M, InsnMatcherTemp, Src, TempOpIdx);
Re-commit: [globalisel][tablegen] Support zero-instruction emission. Summary: Support the case where an operand of a pattern is also the whole of the result pattern. In this case the original result and all its uses must be replaced by the operand. However, register class restrictions can require a COPY. This patch handles both cases by always emitting the copy and leaving it for the register allocator to optimize. The previous commit failed on Windows machines due to a flaw in the sort predicate which allowed both A < B < C and B == C to be satisfied simultaneously. The cause of this was some sloppiness in the priority order of G_CONSTANT instructions compared to other instructions. These had equal priority because it makes no difference, however there were operands had higher priority than G_CONSTANT but lower priority than any other instruction. As a result, a priority order between G_CONSTANT and other instructions must be enforced to ensure the predicate defines a strict weak order. Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Subscribers: javed.absar, kristof.beyls, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D36084 llvm-svn: 311076
2017-08-17 17:26:14 +08:00
if (auto Error = InsnMatcherOrError.takeError())
return std::move(Error);
InstructionMatcher &InsnMatcher = InsnMatcherOrError.get();
if (Dst->isLeaf()) {
Record *RCDef = getInitValueAsRegClass(Dst->getLeafValue());
const CodeGenRegisterClass &RC = Target.getRegisterClass(RCDef);
if (RCDef) {
// We need to replace the def and all its uses with the specified
// operand. However, we must also insert COPY's wherever needed.
// For now, emit a copy and let the register allocator clean up.
auto &DstI = Target.getInstruction(RK.getDef("COPY"));
const auto &DstIOperand = DstI.Operands[0];
OperandMatcher &OM0 = InsnMatcher.getOperand(0);
OM0.setSymbolicName(DstIOperand.Name);
[globalisel][tablegen] Simplify named operand/operator lookups and fix a wrong-code bug this revealed. Summary: Operand variable lookups are now performed by the RuleMatcher rather than searching the whole matcher hierarchy for a match. This revealed a wrong-code bug that currently affects ARM and X86 where patterns that use a variable more than once in the match pattern will be imported but won't check that the operands are identical. This can cause the tablegen-erated matcher to accept matches that should be rejected. Depends on D36569 Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Subscribers: aemerson, igorb, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D36618 llvm-svn: 315780
2017-10-14 08:31:58 +08:00
M.defineOperand(OM0.getSymbolicName(), OM0);
Re-commit: [globalisel][tablegen] Support zero-instruction emission. Summary: Support the case where an operand of a pattern is also the whole of the result pattern. In this case the original result and all its uses must be replaced by the operand. However, register class restrictions can require a COPY. This patch handles both cases by always emitting the copy and leaving it for the register allocator to optimize. The previous commit failed on Windows machines due to a flaw in the sort predicate which allowed both A < B < C and B == C to be satisfied simultaneously. The cause of this was some sloppiness in the priority order of G_CONSTANT instructions compared to other instructions. These had equal priority because it makes no difference, however there were operands had higher priority than G_CONSTANT but lower priority than any other instruction. As a result, a priority order between G_CONSTANT and other instructions must be enforced to ensure the predicate defines a strict weak order. Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Subscribers: javed.absar, kristof.beyls, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D36084 llvm-svn: 311076
2017-08-17 17:26:14 +08:00
OM0.addPredicate<RegisterBankOperandMatcher>(RC);
[globalisel][tablegen] Stop hard-coding the emitted instruction ID to 0. NFC The next commit will add support for multi-instruction emission so we need to start allocating instruction ID's instead of hard-coding them to 0. llvm-svn: 317057
2017-11-01 08:29:47 +08:00
auto &DstMIBuilder =
M.addAction<BuildMIAction>(M.allocateOutputInsnID(), &DstI);
DstMIBuilder.addRenderer<CopyRenderer>(DstIOperand.Name);
DstMIBuilder.addRenderer<CopyRenderer>(Dst->getName());
Re-commit: [globalisel][tablegen] Support zero-instruction emission. Summary: Support the case where an operand of a pattern is also the whole of the result pattern. In this case the original result and all its uses must be replaced by the operand. However, register class restrictions can require a COPY. This patch handles both cases by always emitting the copy and leaving it for the register allocator to optimize. The previous commit failed on Windows machines due to a flaw in the sort predicate which allowed both A < B < C and B == C to be satisfied simultaneously. The cause of this was some sloppiness in the priority order of G_CONSTANT instructions compared to other instructions. These had equal priority because it makes no difference, however there were operands had higher priority than G_CONSTANT but lower priority than any other instruction. As a result, a priority order between G_CONSTANT and other instructions must be enforced to ensure the predicate defines a strict weak order. Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Subscribers: javed.absar, kristof.beyls, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D36084 llvm-svn: 311076
2017-08-17 17:26:14 +08:00
M.addAction<ConstrainOperandToRegClassAction>(0, 0, RC);
// We're done with this pattern! It's eligible for GISel emission; return
// it.
++NumPatternImported;
return std::move(M);
}
[globalisel][tablegen] Add support for (set $dst, 1) and test X86's OptForSize predicate. Summary: It's rare but a small number of patterns use IntInit's at the root of the match. On X86, one such rule is enabled by the OptForSize predicate and causes the compiler to use the smaller: %0 = MOV32r1 instead of the usual: %0 = MOV32ri 1 This patch adds support for matching IntInit's at the root and uses this as a test case for the optsize attribute that was implemented in r301750 Reviewers: qcolombet, ab, t.p.northover, rovka, kristof.beyls, aditya_nandakumar Reviewed By: qcolombet Subscribers: igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D32791 llvm-svn: 303678
2017-05-24 03:33:16 +08:00
return failedImport("Dst pattern root isn't a known leaf");
Re-commit: [globalisel][tablegen] Support zero-instruction emission. Summary: Support the case where an operand of a pattern is also the whole of the result pattern. In this case the original result and all its uses must be replaced by the operand. However, register class restrictions can require a COPY. This patch handles both cases by always emitting the copy and leaving it for the register allocator to optimize. The previous commit failed on Windows machines due to a flaw in the sort predicate which allowed both A < B < C and B == C to be satisfied simultaneously. The cause of this was some sloppiness in the priority order of G_CONSTANT instructions compared to other instructions. These had equal priority because it makes no difference, however there were operands had higher priority than G_CONSTANT but lower priority than any other instruction. As a result, a priority order between G_CONSTANT and other instructions must be enforced to ensure the predicate defines a strict weak order. Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Subscribers: javed.absar, kristof.beyls, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D36084 llvm-svn: 311076
2017-08-17 17:26:14 +08:00
}
[globalisel][tablegen] Add support for (set $dst, 1) and test X86's OptForSize predicate. Summary: It's rare but a small number of patterns use IntInit's at the root of the match. On X86, one such rule is enabled by the OptForSize predicate and causes the compiler to use the smaller: %0 = MOV32r1 instead of the usual: %0 = MOV32ri 1 This patch adds support for matching IntInit's at the root and uses this as a test case for the optsize attribute that was implemented in r301750 Reviewers: qcolombet, ab, t.p.northover, rovka, kristof.beyls, aditya_nandakumar Reviewed By: qcolombet Subscribers: igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D32791 llvm-svn: 303678
2017-05-24 03:33:16 +08:00
[tablegen][globalisel] Add support for nested instruction matching. Summary: Lift the restrictions that prevented the tree walking introduced in the previous change and add support for patterns like: (G_ADD (G_MUL (G_SEXT $src1), (G_SEXT $src2)), $src3) -> SMADDWrrr $dst, $src1, $src2, $src3 Also adds support for G_SEXT and G_ZEXT to support these cases. One particular aspect of this that I should draw attention to is that I've tried to be overly conservative in determining the safety of matches that involve non-adjacent instructions and multiple basic blocks. This is intended to be used as a cheap initial check and we may add a more expensive check in the future. The current rules are: * Reject if any instruction may load/store (we'd need to check for intervening memory operations. * Reject if any instruction has implicit operands. * Reject if any instruction has unmodelled side-effects. See isObviouslySafeToFold(). Reviewers: t.p.northover, javed.absar, qcolombet, aditya_nandakumar, ab, rovka Reviewed By: ab Subscribers: igorb, dberris, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D30539 llvm-svn: 299430
2017-04-04 21:25:23 +08:00
// Start with the defined operands (i.e., the results of the root operator).
[tablegen][globalisel] Convert the SelectionDAG importer to a tree walking approach. NFC Summary: But don't actually inspect the tree any deeper than we already do. This change is NFC but the next one will enable full traversal of the source/destination patterns. Depends on D30535 Reviewers: t.p.northover, qcolombet, aditya_nandakumar, rovka, ab Subscribers: igorb, dberris, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D30536 llvm-svn: 299001
2017-03-29 23:37:18 +08:00
Record *DstOp = Dst->getOperator();
if (!DstOp->isSubClassOf("Instruction"))
return failedImport("Pattern operator isn't an instruction");
[globalisel][tblgen] Add support for ComplexPatterns Summary: Adds a new kind of MachineOperand: MO_Placeholder. This operand must not appear in the MIR and only exists as a way of creating an 'uninitialized' operand until a matcher function overwrites it. Depends on D30046, D29712 Reviewers: t.p.northover, ab, rovka, aditya_nandakumar, javed.absar, qcolombet Reviewed By: qcolombet Subscribers: dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D30089 llvm-svn: 297782
2017-03-15 05:32:08 +08:00
[tablegen][globalisel] Convert the SelectionDAG importer to a tree walking approach. NFC Summary: But don't actually inspect the tree any deeper than we already do. This change is NFC but the next one will enable full traversal of the source/destination patterns. Depends on D30535 Reviewers: t.p.northover, qcolombet, aditya_nandakumar, rovka, ab Subscribers: igorb, dberris, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D30536 llvm-svn: 299001
2017-03-29 23:37:18 +08:00
auto &DstI = Target.getInstruction(DstOp);
if (DstI.Operands.NumDefs != Src->getExtTypes().size())
[globalisel][tablegen] Report more detail in some SelectionDAG import failures. NFC Reviewers: ab, t.p.northover, qcolombet, aditya_nandakumar, rovka Reviewed By: ab Subscribers: dberris, kristof.beyls, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D31325 llvm-svn: 300186
2017-04-13 17:45:37 +08:00
return failedImport("Src pattern results and dst MI defs are different (" +
to_string(Src->getExtTypes().size()) + " def(s) vs " +
to_string(DstI.Operands.NumDefs) + " def(s))");
[globalisel] Decouple src pattern operands from dst pattern operands. Summary: This isn't testable for AArch64 by itself so this patch also adds support for constant immediates in the pattern and physical register uses in the result. The new IntOperandMatcher matches the constant in patterns such as '(set $rd:GPR32, (G_XOR $rs:GPR32, -1))'. It's always safe to fold immediates into an instruction so this is the first rule that will match across multiple BB's. The Renderer hierarchy is responsible for adding operands to the result instruction. Renderers can copy operands (CopyRenderer) or add physical registers (in particular %wzr and %xzr) to the result instruction in any order (OperandMatchers now import the operand names from SelectionDAG to allow renderers to access any operand). This allows us to emit the result instruction for: %1 = G_XOR %0, -1 --> %1 = ORNWrr %wzr, %0 %1 = G_XOR -1, %0 --> %1 = ORNWrr %wzr, %0 although the latter is untested since the matcher/importer has not been taught about commutativity yet. Added BuildMIAction which can build new instructions and mutate them where possible. W.r.t the mutation aspect, MatchActions are now told the name of an instruction they can recycle and BuildMIAction will emit mutation code when the renderers are appropriate. They are appropriate when all operands are rendered using CopyRenderer and the indices are the same as the matcher. This currently assumes that all operands have at least one matcher. Finally, this change also fixes a crash in AArch64InstructionSelector::select() caused by an immediate operand passing isImm() rather than isCImm(). This was uncovered by the other changes and was detected by existing tests. Depends on D29711 Reviewers: t.p.northover, ab, qcolombet, rovka, aditya_nandakumar, javed.absar Reviewed By: rovka Subscribers: aemerson, dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D29712 llvm-svn: 296131
2017-02-24 23:43:30 +08:00
[tablegen][globalisel] Convert the SelectionDAG importer to a tree walking approach. NFC Summary: But don't actually inspect the tree any deeper than we already do. This change is NFC but the next one will enable full traversal of the source/destination patterns. Depends on D30535 Reviewers: t.p.northover, qcolombet, aditya_nandakumar, rovka, ab Subscribers: igorb, dberris, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D30536 llvm-svn: 299001
2017-03-29 23:37:18 +08:00
// The root of the match also has constraints on the register bank so that it
// matches the result instruction.
unsigned OpIdx = 0;
TableGen support for parameterized register class information This replaces TableGen's type inference to operate on parameterized types instead of MVTs, and as a consequence, some interfaces have changed: - Uses of MVTs are replaced by ValueTypeByHwMode. - EEVT::TypeSet is replaced by TypeSetByHwMode. This affects the way that types and type sets are printed, and the tests relying on that have been updated. There are certain users of the inferred types outside of TableGen itself, namely FastISel and GlobalISel. For those users, the way that the types are accessed have changed. For typical scenarios, these replacements can be used: - TreePatternNode::getType(ResNo) -> getSimpleType(ResNo) - TreePatternNode::hasTypeSet(ResNo) -> hasConcreteType(ResNo) - TypeSet::isConcrete -> TypeSetByHwMode::isValueTypeByHwMode(false) For more information, please refer to the review page. Differential Revision: https://reviews.llvm.org/D31951 llvm-svn: 313271
2017-09-15 00:56:21 +08:00
for (const TypeSetByHwMode &VTy : Src->getExtTypes()) {
(void)VTy;
[tablegen][globalisel] Convert the SelectionDAG importer to a tree walking approach. NFC Summary: But don't actually inspect the tree any deeper than we already do. This change is NFC but the next one will enable full traversal of the source/destination patterns. Depends on D30535 Reviewers: t.p.northover, qcolombet, aditya_nandakumar, rovka, ab Subscribers: igorb, dberris, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D30536 llvm-svn: 299001
2017-03-29 23:37:18 +08:00
const auto &DstIOperand = DstI.Operands[OpIdx];
Record *DstIOpRec = DstIOperand.Rec;
[globalisel][tablegen] Add support for COPY_TO_REGCLASS. Summary: As part of this * Emitted instructions now have named MachineInstr variables associated with them. This isn't particularly important yet but it's a small step towards multiple-insn emission. * constrainSelectedInstRegOperands() is no longer hardcoded. It's now added as the ConstrainOperandsToDefinitionAction() action. COPY_TO_REGCLASS uses an alternate constraint mechanism ConstrainOperandToRegClassAction() which supports arbitrary constraints such as that defined by COPY_TO_REGCLASS. Reviewers: ab, qcolombet, t.p.northover, rovka, kristof.beyls, aditya_nandakumar Reviewed By: ab Subscribers: javed.absar, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D33590 llvm-svn: 305791
2017-06-20 20:36:34 +08:00
if (DstI.TheDef->getName() == "COPY_TO_REGCLASS") {
DstIOpRec = getInitValueAsRegClass(Dst->getChild(1)->getLeafValue());
if (DstIOpRec == nullptr)
return failedImport(
"COPY_TO_REGCLASS operand #1 isn't a register class");
[globalisel][tablegen] Add support for EXTRACT_SUBREG. Summary: After this patch, we finally have test cases that require multiple instruction emission. Depends on D33590 Reviewers: ab, qcolombet, t.p.northover, rovka, kristof.beyls Subscribers: javed.absar, llvm-commits, igorb Differential Revision: https://reviews.llvm.org/D33596 llvm-svn: 306388
2017-06-27 18:11:39 +08:00
} else if (DstI.TheDef->getName() == "EXTRACT_SUBREG") {
if (!Dst->getChild(0)->isLeaf())
return failedImport("EXTRACT_SUBREG operand #0 isn't a leaf");
[globalisel][tablegen] Multiple 80-col corrections. llvm-svn: 306544
2017-06-28 21:50:04 +08:00
// We can assume that a subregister is in the same bank as it's super
// register.
[globalisel][tablegen] Add support for EXTRACT_SUBREG. Summary: After this patch, we finally have test cases that require multiple instruction emission. Depends on D33590 Reviewers: ab, qcolombet, t.p.northover, rovka, kristof.beyls Subscribers: javed.absar, llvm-commits, igorb Differential Revision: https://reviews.llvm.org/D33596 llvm-svn: 306388
2017-06-27 18:11:39 +08:00
DstIOpRec = getInitValueAsRegClass(Dst->getChild(0)->getLeafValue());
if (DstIOpRec == nullptr)
return failedImport(
"EXTRACT_SUBREG operand #0 isn't a register class");
[globalisel][tablegen] Add support for COPY_TO_REGCLASS. Summary: As part of this * Emitted instructions now have named MachineInstr variables associated with them. This isn't particularly important yet but it's a small step towards multiple-insn emission. * constrainSelectedInstRegOperands() is no longer hardcoded. It's now added as the ConstrainOperandsToDefinitionAction() action. COPY_TO_REGCLASS uses an alternate constraint mechanism ConstrainOperandToRegClassAction() which supports arbitrary constraints such as that defined by COPY_TO_REGCLASS. Reviewers: ab, qcolombet, t.p.northover, rovka, kristof.beyls, aditya_nandakumar Reviewed By: ab Subscribers: javed.absar, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D33590 llvm-svn: 305791
2017-06-20 20:36:34 +08:00
} else if (DstIOpRec->isSubClassOf("RegisterOperand"))
[globalisel][tablegen] Add support for RegisterOperand. Summary: It functions just like RegisterClass except that the class is obtained from a field. Depends on D31761. Reviewers: ab, qcolombet, t.p.northover, rovka, kristof.beyls, aditya_nandakumar Reviewed By: ab Subscribers: dberris, llvm-commits, igorb Differential Revision: https://reviews.llvm.org/D32229 llvm-svn: 301080
2017-04-22 23:53:21 +08:00
DstIOpRec = DstIOpRec->getValueAsDef("RegClass");
[globalisel][tablegen] Add support for COPY_TO_REGCLASS. Summary: As part of this * Emitted instructions now have named MachineInstr variables associated with them. This isn't particularly important yet but it's a small step towards multiple-insn emission. * constrainSelectedInstRegOperands() is no longer hardcoded. It's now added as the ConstrainOperandsToDefinitionAction() action. COPY_TO_REGCLASS uses an alternate constraint mechanism ConstrainOperandToRegClassAction() which supports arbitrary constraints such as that defined by COPY_TO_REGCLASS. Reviewers: ab, qcolombet, t.p.northover, rovka, kristof.beyls, aditya_nandakumar Reviewed By: ab Subscribers: javed.absar, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D33590 llvm-svn: 305791
2017-06-20 20:36:34 +08:00
else if (!DstIOpRec->isSubClassOf("RegisterClass"))
[globalisel][tablegen] Multiple 80-col corrections. llvm-svn: 306544
2017-06-28 21:50:04 +08:00
return failedImport("Dst MI def isn't a register class" +
to_string(*Dst));
[tablegen][globalisel] Convert the SelectionDAG importer to a tree walking approach. NFC Summary: But don't actually inspect the tree any deeper than we already do. This change is NFC but the next one will enable full traversal of the source/destination patterns. Depends on D30535 Reviewers: t.p.northover, qcolombet, aditya_nandakumar, rovka, ab Subscribers: igorb, dberris, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D30536 llvm-svn: 299001
2017-03-29 23:37:18 +08:00
OperandMatcher &OM = InsnMatcher.getOperand(OpIdx);
OM.setSymbolicName(DstIOperand.Name);
[globalisel][tablegen] Simplify named operand/operator lookups and fix a wrong-code bug this revealed. Summary: Operand variable lookups are now performed by the RuleMatcher rather than searching the whole matcher hierarchy for a match. This revealed a wrong-code bug that currently affects ARM and X86 where patterns that use a variable more than once in the match pattern will be imported but won't check that the operands are identical. This can cause the tablegen-erated matcher to accept matches that should be rejected. Depends on D36569 Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Subscribers: aemerson, igorb, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D36618 llvm-svn: 315780
2017-10-14 08:31:58 +08:00
M.defineOperand(OM.getSymbolicName(), OM);
[tablegen][globalisel] Convert the SelectionDAG importer to a tree walking approach. NFC Summary: But don't actually inspect the tree any deeper than we already do. This change is NFC but the next one will enable full traversal of the source/destination patterns. Depends on D30535 Reviewers: t.p.northover, qcolombet, aditya_nandakumar, rovka, ab Subscribers: igorb, dberris, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D30536 llvm-svn: 299001
2017-03-29 23:37:18 +08:00
OM.addPredicate<RegisterBankOperandMatcher>(
Target.getRegisterClass(DstIOpRec));
++OpIdx;
[GlobalISel] Add basic Selector-emitter tblgen backend. This adds a basic tablegen backend that analyzes the SelectionDAG patterns to find simple ones that are eligible for GlobalISel-emission. That's similar to FastISel, with one notable difference: we're not fed ISD opcodes, so we need to map the SDNode operators to generic opcodes. That's done using GINodeEquiv in TargetGlobalISel.td. Otherwise, this is mostly boilerplate, and lots of filtering of any kind of "complicated" pattern. On AArch64, this is sufficient to match G_ADD up to s64 (to ADDWrr/ADDXrr) and G_BR (to B). Differential Revision: https://reviews.llvm.org/D26878 llvm-svn: 290284
2016-12-22 07:26:20 +08:00
}
[globalisel][tablegen] Add infrastructure to potentially allow BuildMIAction to choose a mutatable instruction. NFC Prepare for multiple instruction emission by allowing BuildMIAction to search for a suitable matcher that will support mutation. This patch deliberately neglects to add matchers aside from the root to preserve NFC. That said, it should be noted that until we support mutations other than just the opcode the chances of finding a non-root instruction for which canMutate() is true, is essentially zero. Furthermore in the presence of multi-instruction emission the chances of finding any instruction for which canMutate() is true is also zero. Nevertheless, we can't continue to require that all BuildMIAction's consider the root of the match to be recyclable due to the risk of recycling it twice in the same rule. llvm-svn: 317022
2017-11-01 02:50:24 +08:00
auto DstMIBuilderOrError = createAndImportInstructionRenderer(M, Dst);
[tablegen][globalisel] Convert the SelectionDAG importer to a tree walking approach. NFC Summary: But don't actually inspect the tree any deeper than we already do. This change is NFC but the next one will enable full traversal of the source/destination patterns. Depends on D30535 Reviewers: t.p.northover, qcolombet, aditya_nandakumar, rovka, ab Subscribers: igorb, dberris, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D30536 llvm-svn: 299001
2017-03-29 23:37:18 +08:00
if (auto Error = DstMIBuilderOrError.takeError())
return std::move(Error);
BuildMIAction &DstMIBuilder = DstMIBuilderOrError.get();
// Render the implicit defs.
// These are only added to the root of the result.
[globalisel][tablegen] Change Expected<bool> to Error and rename functions. Functions that still return Expected<X> are now called createAndImport*() Changing the return type was requested in the review comments for r299001 llvm-svn: 299063
2017-03-30 17:36:33 +08:00
if (auto Error = importImplicitDefRenderers(DstMIBuilder, P.getDstRegs()))
[tablegen][globalisel] Convert the SelectionDAG importer to a tree walking approach. NFC Summary: But don't actually inspect the tree any deeper than we already do. This change is NFC but the next one will enable full traversal of the source/destination patterns. Depends on D30535 Reviewers: t.p.northover, qcolombet, aditya_nandakumar, rovka, ab Subscribers: igorb, dberris, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D30536 llvm-svn: 299001
2017-03-29 23:37:18 +08:00
return std::move(Error);
[globalisel][tablegen] Add infrastructure to potentially allow BuildMIAction to choose a mutatable instruction. NFC Prepare for multiple instruction emission by allowing BuildMIAction to search for a suitable matcher that will support mutation. This patch deliberately neglects to add matchers aside from the root to preserve NFC. That said, it should be noted that until we support mutations other than just the opcode the chances of finding a non-root instruction for which canMutate() is true, is essentially zero. Furthermore in the presence of multi-instruction emission the chances of finding any instruction for which canMutate() is true is also zero. Nevertheless, we can't continue to require that all BuildMIAction's consider the root of the match to be recyclable due to the risk of recycling it twice in the same rule. llvm-svn: 317022
2017-11-01 02:50:24 +08:00
DstMIBuilder.chooseInsnToMutate(M);
[globalisel][tablegen] Add support for COPY_TO_REGCLASS. Summary: As part of this * Emitted instructions now have named MachineInstr variables associated with them. This isn't particularly important yet but it's a small step towards multiple-insn emission. * constrainSelectedInstRegOperands() is no longer hardcoded. It's now added as the ConstrainOperandsToDefinitionAction() action. COPY_TO_REGCLASS uses an alternate constraint mechanism ConstrainOperandToRegClassAction() which supports arbitrary constraints such as that defined by COPY_TO_REGCLASS. Reviewers: ab, qcolombet, t.p.northover, rovka, kristof.beyls, aditya_nandakumar Reviewed By: ab Subscribers: javed.absar, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D33590 llvm-svn: 305791
2017-06-20 20:36:34 +08:00
// Constrain the registers to classes. This is normally derived from the
// emitted instruction but a few instructions require special handling.
if (DstI.TheDef->getName() == "COPY_TO_REGCLASS") {
// COPY_TO_REGCLASS does not provide operand constraints itself but the
// result is constrained to the class given by the second child.
Record *DstIOpRec =
getInitValueAsRegClass(Dst->getChild(1)->getLeafValue());
if (DstIOpRec == nullptr)
return failedImport("COPY_TO_REGCLASS operand #1 isn't a register class");
M.addAction<ConstrainOperandToRegClassAction>(
[globalisel][tablegen] Added instruction emission to the state-machine-based matcher. Summary: This further improves the compile-time regressions that will be caused by a re-commit of r303259. Also added included preliminary work in preparation for the multi-insn emitter since I needed to change the relevant part of the API for this patch anyway. Depends on D33758 Reviewers: rovka, vitalybuka, ab, t.p.northover, qcolombet, aditya_nandakumar Reviewed By: ab Subscribers: kristof.beyls, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D33764 llvm-svn: 307133
2017-07-05 17:39:33 +08:00
0, 0, Target.getRegisterClass(DstIOpRec));
[globalisel][tablegen] Add support for EXTRACT_SUBREG. Summary: After this patch, we finally have test cases that require multiple instruction emission. Depends on D33590 Reviewers: ab, qcolombet, t.p.northover, rovka, kristof.beyls Subscribers: javed.absar, llvm-commits, igorb Differential Revision: https://reviews.llvm.org/D33596 llvm-svn: 306388
2017-06-27 18:11:39 +08:00
// We're done with this pattern! It's eligible for GISel emission; return
// it.
++NumPatternImported;
return std::move(M);
}
if (DstI.TheDef->getName() == "EXTRACT_SUBREG") {
// EXTRACT_SUBREG selects into a subregister COPY but unlike most
// instructions, the result register class is controlled by the
// subregisters of the operand. As a result, we must constrain the result
// class rather than check that it's already the right one.
if (!Dst->getChild(0)->isLeaf())
return failedImport("EXTRACT_SUBREG child #1 is not a leaf");
[globalisel][tablegen] Post-commit review nits for r306388. NFC One early exit and a missing assert string. llvm-svn: 306552
2017-06-28 23:16:03 +08:00
DefInit *SubRegInit = dyn_cast<DefInit>(Dst->getChild(1)->getLeafValue());
if (!SubRegInit)
return failedImport("EXTRACT_SUBREG child #1 is not a subreg index");
[globalisel][tablegen] Add support for EXTRACT_SUBREG. Summary: After this patch, we finally have test cases that require multiple instruction emission. Depends on D33590 Reviewers: ab, qcolombet, t.p.northover, rovka, kristof.beyls Subscribers: javed.absar, llvm-commits, igorb Differential Revision: https://reviews.llvm.org/D33596 llvm-svn: 306388
2017-06-27 18:11:39 +08:00
[globalisel][tablegen] Post-commit review nits for r306388. NFC One early exit and a missing assert string. llvm-svn: 306552
2017-06-28 23:16:03 +08:00
// Constrain the result to the same register bank as the operand.
Record *DstIOpRec =
getInitValueAsRegClass(Dst->getChild(0)->getLeafValue());
[globalisel][tablegen] Add support for EXTRACT_SUBREG. Summary: After this patch, we finally have test cases that require multiple instruction emission. Depends on D33590 Reviewers: ab, qcolombet, t.p.northover, rovka, kristof.beyls Subscribers: javed.absar, llvm-commits, igorb Differential Revision: https://reviews.llvm.org/D33596 llvm-svn: 306388
2017-06-27 18:11:39 +08:00
[globalisel][tablegen] Post-commit review nits for r306388. NFC One early exit and a missing assert string. llvm-svn: 306552
2017-06-28 23:16:03 +08:00
if (DstIOpRec == nullptr)
return failedImport("EXTRACT_SUBREG operand #1 isn't a register class");
CodeGenSubRegIndex *SubIdx = CGRegs.getSubRegIdx(SubRegInit->getDef());
[globalisel][tablegen] Added instruction emission to the state-machine-based matcher. Summary: This further improves the compile-time regressions that will be caused by a re-commit of r303259. Also added included preliminary work in preparation for the multi-insn emitter since I needed to change the relevant part of the API for this patch anyway. Depends on D33758 Reviewers: rovka, vitalybuka, ab, t.p.northover, qcolombet, aditya_nandakumar Reviewed By: ab Subscribers: kristof.beyls, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D33764 llvm-svn: 307133
2017-07-05 17:39:33 +08:00
CodeGenRegisterClass *SrcRC = CGRegs.getRegClass(DstIOpRec);
[globalisel][tablegen] Post-commit review nits for r306388. NFC One early exit and a missing assert string. llvm-svn: 306552
2017-06-28 23:16:03 +08:00
// It would be nice to leave this constraint implicit but we're required
// to pick a register class so constrain the result to a register class
// that can hold the correct MVT.
//
// FIXME: This may introduce an extra copy if the chosen class doesn't
// actually contain the subregisters.
assert(Src->getExtTypes().size() == 1 &&
"Expected Src of EXTRACT_SUBREG to have one result type");
const auto &SrcRCDstRCPair =
SrcRC->getMatchingSubClassWithSubRegs(CGRegs, SubIdx);
assert(SrcRCDstRCPair->second && "Couldn't find a matching subclass");
[globalisel][tablegen] Added instruction emission to the state-machine-based matcher. Summary: This further improves the compile-time regressions that will be caused by a re-commit of r303259. Also added included preliminary work in preparation for the multi-insn emitter since I needed to change the relevant part of the API for this patch anyway. Depends on D33758 Reviewers: rovka, vitalybuka, ab, t.p.northover, qcolombet, aditya_nandakumar Reviewed By: ab Subscribers: kristof.beyls, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D33764 llvm-svn: 307133
2017-07-05 17:39:33 +08:00
M.addAction<ConstrainOperandToRegClassAction>(0, 0, *SrcRCDstRCPair->second);
M.addAction<ConstrainOperandToRegClassAction>(0, 1, *SrcRCDstRCPair->first);
// We're done with this pattern! It's eligible for GISel emission; return
// it.
++NumPatternImported;
return std::move(M);
}
M.addAction<ConstrainOperandsToDefinitionAction>(0);
[globalisel][tablegen] Add support for COPY_TO_REGCLASS. Summary: As part of this * Emitted instructions now have named MachineInstr variables associated with them. This isn't particularly important yet but it's a small step towards multiple-insn emission. * constrainSelectedInstRegOperands() is no longer hardcoded. It's now added as the ConstrainOperandsToDefinitionAction() action. COPY_TO_REGCLASS uses an alternate constraint mechanism ConstrainOperandToRegClassAction() which supports arbitrary constraints such as that defined by COPY_TO_REGCLASS. Reviewers: ab, qcolombet, t.p.northover, rovka, kristof.beyls, aditya_nandakumar Reviewed By: ab Subscribers: javed.absar, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D33590 llvm-svn: 305791
2017-06-20 20:36:34 +08:00
[GlobalISel] Return an Expected<RuleMatcher> for each SDAG pattern. NFC. Instead of emitting the matcher code directly, return the rule matcher and the skip reason as an Expected<RuleMatcher>. This will let us record all matchers and process them before emission. It's a somewhat unconventional use of Error, but it's nicer than, say, std::pair, because of the bool conversions. Differential Revision: https://reviews.llvm.org/D29743 llvm-svn: 294706
2017-02-10 12:00:17 +08:00
// We're done with this pattern! It's eligible for GISel emission; return it.
[globalisel] Separate the SelectionDAG importer from the emitter. NFC Summary: In the near future the rules will be sorted between these two steps to ensure that more important rules are not prevented by less important ones. Reviewers: t.p.northover, ab, rovka, qcolombet, aditya_nandakumar Reviewed By: ab Subscribers: dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D29709 llvm-svn: 295661
2017-02-20 22:31:27 +08:00
++NumPatternImported;
[GlobalISel] Return an Expected<RuleMatcher> for each SDAG pattern. NFC. Instead of emitting the matcher code directly, return the rule matcher and the skip reason as an Expected<RuleMatcher>. This will let us record all matchers and process them before emission. It's a somewhat unconventional use of Error, but it's nicer than, say, std::pair, because of the bool conversions. Differential Revision: https://reviews.llvm.org/D29743 llvm-svn: 294706
2017-02-10 12:00:17 +08:00
return std::move(M);
[GlobalISel] Add basic Selector-emitter tblgen backend. This adds a basic tablegen backend that analyzes the SelectionDAG patterns to find simple ones that are eligible for GlobalISel-emission. That's similar to FastISel, with one notable difference: we're not fed ISD opcodes, so we need to map the SDNode operators to generic opcodes. That's done using GINodeEquiv in TargetGlobalISel.td. Otherwise, this is mostly boilerplate, and lots of filtering of any kind of "complicated" pattern. On AArch64, this is sufficient to match G_ADD up to s64 (to ADDWrr/ADDXrr) and G_BR (to B). Differential Revision: https://reviews.llvm.org/D26878 llvm-svn: 290284
2016-12-22 07:26:20 +08:00
}
[aarch64] Support APInt and APFloat in ImmLeaf subclasses and make AArch64 use them. Summary: The purpose of this patch is to expose more information about ImmLeaf-like PatLeaf's so that GlobalISel can learn to import them. Previously, ImmLeaf could only be used to test int64_t's produced by sign-extending an APInt. Other tests on immediates had to use the generic PatLeaf and extract the constant using C++. With this patch, tablegen will know how to generate predicates for APInt, and APFloat. This will allow it to 'do the right thing' for both SelectionDAG and GlobalISel which require different methods of extracting the immediate from the IR. This is NFC for SelectionDAG since the new code is equivalent to the previous code. It's also NFC for FastISel because FastIselShouldIgnore is 1 for the ImmLeaf subclasses. Enabling FastIselShouldIgnore == 0 for these new subclasses will require a significant re-factor of FastISel. For GlobalISel, it's currently NFC because the relevant code to import the affected rules is not yet present. This will be added in a later patch. Depends on D36086 Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Reviewed By: qcolombet Subscribers: bjope, aemerson, rengolin, javed.absar, igorb, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D36534 llvm-svn: 315747
2017-10-14 04:42:18 +08:00
// Emit imm predicate table and an enum to reference them with.
// The 'Predicate_' part of the name is redundant but eliminating it is more
// trouble than it's worth.
void GlobalISelEmitter::emitImmPredicates(
[globalisel][tablegen] Add support for fpimm and import of APInt/APFloat based ImmLeaf. Summary: There's only a tablegen testcase for IntImmLeaf and not a CodeGen one because the relevant rules are rejected for other reasons at the moment. On AArch64, it's because there's an SDNodeXForm attached to the operand. On X86, it's because the rule either emits multiple instructions or has another predicate using PatFrag which cannot easily be supported at the same time. Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Reviewed By: qcolombet Subscribers: aemerson, javed.absar, igorb, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D36569 llvm-svn: 315761
2017-10-14 05:28:03 +08:00
raw_ostream &OS, StringRef TypeIdentifier, StringRef Type,
std::function<bool(const Record *R)> Filter) {
[aarch64] Support APInt and APFloat in ImmLeaf subclasses and make AArch64 use them. Summary: The purpose of this patch is to expose more information about ImmLeaf-like PatLeaf's so that GlobalISel can learn to import them. Previously, ImmLeaf could only be used to test int64_t's produced by sign-extending an APInt. Other tests on immediates had to use the generic PatLeaf and extract the constant using C++. With this patch, tablegen will know how to generate predicates for APInt, and APFloat. This will allow it to 'do the right thing' for both SelectionDAG and GlobalISel which require different methods of extracting the immediate from the IR. This is NFC for SelectionDAG since the new code is equivalent to the previous code. It's also NFC for FastISel because FastIselShouldIgnore is 1 for the ImmLeaf subclasses. Enabling FastIselShouldIgnore == 0 for these new subclasses will require a significant re-factor of FastISel. For GlobalISel, it's currently NFC because the relevant code to import the affected rules is not yet present. This will be added in a later patch. Depends on D36086 Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Reviewed By: qcolombet Subscribers: bjope, aemerson, rengolin, javed.absar, igorb, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D36534 llvm-svn: 315747
2017-10-14 04:42:18 +08:00
std::vector<const Record *> MatchedRecords;
const auto &Defs = RK.getAllDerivedDefinitions("PatFrag");
std::copy_if(Defs.begin(), Defs.end(), std::back_inserter(MatchedRecords),
[&](Record *Record) {
return !Record->getValueAsString("ImmediateCode").empty() &&
Filter(Record);
});
[globalisel][tablegen] Add support for fpimm and import of APInt/APFloat based ImmLeaf. Summary: There's only a tablegen testcase for IntImmLeaf and not a CodeGen one because the relevant rules are rejected for other reasons at the moment. On AArch64, it's because there's an SDNodeXForm attached to the operand. On X86, it's because the rule either emits multiple instructions or has another predicate using PatFrag which cannot easily be supported at the same time. Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Reviewed By: qcolombet Subscribers: aemerson, javed.absar, igorb, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D36569 llvm-svn: 315761
2017-10-14 05:28:03 +08:00
if (!MatchedRecords.empty()) {
OS << "// PatFrag predicates.\n"
<< "enum {\n";
[globalisel][tablegen] Fix a use-after free bug that manifests on non-macOS machines. llvm-svn: 315765
2017-10-14 05:51:20 +08:00
std::string EnumeratorSeparator =
[globalisel][tablegen] Add support for fpimm and import of APInt/APFloat based ImmLeaf. Summary: There's only a tablegen testcase for IntImmLeaf and not a CodeGen one because the relevant rules are rejected for other reasons at the moment. On AArch64, it's because there's an SDNodeXForm attached to the operand. On X86, it's because the rule either emits multiple instructions or has another predicate using PatFrag which cannot easily be supported at the same time. Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Reviewed By: qcolombet Subscribers: aemerson, javed.absar, igorb, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D36569 llvm-svn: 315761
2017-10-14 05:28:03 +08:00
(" = GIPFP_" + TypeIdentifier + "_Invalid + 1,\n").str();
for (const auto *Record : MatchedRecords) {
OS << " GIPFP_" << TypeIdentifier << "_Predicate_" << Record->getName()
<< EnumeratorSeparator;
EnumeratorSeparator = ",\n";
}
OS << "};\n";
[aarch64] Support APInt and APFloat in ImmLeaf subclasses and make AArch64 use them. Summary: The purpose of this patch is to expose more information about ImmLeaf-like PatLeaf's so that GlobalISel can learn to import them. Previously, ImmLeaf could only be used to test int64_t's produced by sign-extending an APInt. Other tests on immediates had to use the generic PatLeaf and extract the constant using C++. With this patch, tablegen will know how to generate predicates for APInt, and APFloat. This will allow it to 'do the right thing' for both SelectionDAG and GlobalISel which require different methods of extracting the immediate from the IR. This is NFC for SelectionDAG since the new code is equivalent to the previous code. It's also NFC for FastISel because FastIselShouldIgnore is 1 for the ImmLeaf subclasses. Enabling FastIselShouldIgnore == 0 for these new subclasses will require a significant re-factor of FastISel. For GlobalISel, it's currently NFC because the relevant code to import the affected rules is not yet present. This will be added in a later patch. Depends on D36086 Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Reviewed By: qcolombet Subscribers: bjope, aemerson, rengolin, javed.absar, igorb, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D36534 llvm-svn: 315747
2017-10-14 04:42:18 +08:00
}
[globalisel][tablegen] Add support for fpimm and import of APInt/APFloat based ImmLeaf. Summary: There's only a tablegen testcase for IntImmLeaf and not a CodeGen one because the relevant rules are rejected for other reasons at the moment. On AArch64, it's because there's an SDNodeXForm attached to the operand. On X86, it's because the rule either emits multiple instructions or has another predicate using PatFrag which cannot easily be supported at the same time. Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Reviewed By: qcolombet Subscribers: aemerson, javed.absar, igorb, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D36569 llvm-svn: 315761
2017-10-14 05:28:03 +08:00
[globalisel][tablegen] Allow ImmLeaf predicates to use InstructionSelector members NFC for currently supported targets. This resolves a problem encountered by targets such as RISCV that reference `Subtarget` in ImmLeaf predicates. llvm-svn: 321176
2017-12-20 22:41:51 +08:00
OS << "bool " << Target.getName() << "InstructionSelector::testImmPredicate_"
Do not generate an empty switch statement as it causes MSVC to issue diagnostics about switch statements without case or default labels. llvm-svn: 321217
2017-12-21 04:09:30 +08:00
<< TypeIdentifier << "(unsigned PredicateID, " << Type
<< " Imm) const {\n";
if (!MatchedRecords.empty())
OS << " switch (PredicateID) {\n";
[globalisel][tablegen] Allow ImmLeaf predicates to use InstructionSelector members NFC for currently supported targets. This resolves a problem encountered by targets such as RISCV that reference `Subtarget` in ImmLeaf predicates. llvm-svn: 321176
2017-12-20 22:41:51 +08:00
for (const auto *Record : MatchedRecords) {
OS << " case GIPFP_" << TypeIdentifier << "_Predicate_"
<< Record->getName() << ": {\n"
<< " " << Record->getValueAsString("ImmediateCode") << "\n"
<< " llvm_unreachable(\"ImmediateCode should have returned\");\n"
<< " return false;\n"
<< " }\n";
}
Do not generate an empty switch statement as it causes MSVC to issue diagnostics about switch statements without case or default labels. llvm-svn: 321217
2017-12-21 04:09:30 +08:00
if (!MatchedRecords.empty())
OS << " }\n";
OS << " llvm_unreachable(\"Unknown predicate\");\n"
[globalisel][tablegen] Allow ImmLeaf predicates to use InstructionSelector members NFC for currently supported targets. This resolves a problem encountered by targets such as RISCV that reference `Subtarget` in ImmLeaf predicates. llvm-svn: 321176
2017-12-20 22:41:51 +08:00
<< " return false;\n"
<< "}\n";
[aarch64] Support APInt and APFloat in ImmLeaf subclasses and make AArch64 use them. Summary: The purpose of this patch is to expose more information about ImmLeaf-like PatLeaf's so that GlobalISel can learn to import them. Previously, ImmLeaf could only be used to test int64_t's produced by sign-extending an APInt. Other tests on immediates had to use the generic PatLeaf and extract the constant using C++. With this patch, tablegen will know how to generate predicates for APInt, and APFloat. This will allow it to 'do the right thing' for both SelectionDAG and GlobalISel which require different methods of extracting the immediate from the IR. This is NFC for SelectionDAG since the new code is equivalent to the previous code. It's also NFC for FastISel because FastIselShouldIgnore is 1 for the ImmLeaf subclasses. Enabling FastIselShouldIgnore == 0 for these new subclasses will require a significant re-factor of FastISel. For GlobalISel, it's currently NFC because the relevant code to import the affected rules is not yet present. This will be added in a later patch. Depends on D36086 Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Reviewed By: qcolombet Subscribers: bjope, aemerson, rengolin, javed.absar, igorb, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D36534 llvm-svn: 315747
2017-10-14 04:42:18 +08:00
}
Reapply r332907 "[GlobalISel] Improving InstructionSelect's performance by reducing MatchTable..." Apparently the compile time problem was caused by the fact that not all compilers / STL implementations can automatically convert std::unique_ptr<Derived> to std::unique_ptr<Base>. Fixed (hopefully) by making sure it's std::unique_ptr<Derived>&& (rvalue ref) to std::unique_ptr<Base> conversion instead. llvm-svn: 332917
2018-05-22 07:28:51 +08:00
template <class GroupT>
[TableGen][GlobalISel] Optimize MatchTable for faster instruction selection *** Context *** Prior to this patchw, the table generated for matching instruction was straight forward but highly inefficient. Basically, each pattern generates its own set of self contained checks and actions. E.g., TableGen generated: // First pattern CheckNumOperand 3 CheckOpcode G_ADD ... Build ADDrr // Second pattern CheckNumOperand 3 CheckOpcode G_ADD ... Build ADDri // Third pattern CheckNumOperand 3 CheckOpcode G_SUB ... Build SUBrr *** Problem *** Because of that generation, a *lot* of check were redundant between each pattern and were checked every single time until we reach the pattern that matches. E.g., Taking the previous table, let say we are matching a G_SUB, that means we were going to check all the rules for G_ADD before looking at the G_SUB rule. In particular we are going to do: check 3 operands; PASS check G_ADD; FAIL ; Next rule check 3 operands; PASS (but we already knew that!) check G_ADD; FAIL (well it is still not true) ; Next rule check 3 operands; PASS (really!!) check G_SUB; PASS (at last :P) *** Proposed Solution *** This patch introduces a concept of group of rules (GroupMatcher) that share some predicates and only get checked once for the whole group. This patch only creates groups with one nesting level. Conceptually there is nothing preventing us for having deeper nest level. However, the current implementation is not smart enough to share the recording (aka capturing) of values. That limits its ability to do more sharing. For the given example the current patch will generate: // First group CheckOpcode G_ADD // First pattern CheckNumOperand 3 ... Build ADDrr // Second pattern CheckNumOperand 3 ... Build ADDri // Second group CheckOpcode G_SUB // Third pattern CheckNumOperand 3 ... Build SUBrr But if we allowed several nesting level, it could create a sub group for the checknumoperand 3. (We would need to call optimizeRules on the rules within a group.) *** Result *** With only one level of nesting, the instruction selection pass is up to 4x faster. For instance, one instruction now takes 500 checks, instead of 24k! With more nesting we could get in the tens I believe. Differential Revision: https://reviews.llvm.org/D39034 rdar://problem/34670699 llvm-svn: 321017
2017-12-19 03:47:41 +08:00
std::vector<Matcher *> GlobalISelEmitter::optimizeRules(
[GlobalISel][InstructionSelect] Refactoring buildMatchTable out, NFC to share it between the Instruction Selector in optimized and non-optimized modes both and the Testgen. Reviewers: dsanders, aemerson Reviewed By: dsanders Subscribers: rovka, kristof.beyls, llvm-commits, dsanders Differential Revision: https://reviews.llvm.org/D46097 llvm-svn: 331396
2018-05-03 04:08:14 +08:00
ArrayRef<Matcher *> Rules,
Reapply r332907 "[GlobalISel] Improving InstructionSelect's performance by reducing MatchTable..." Apparently the compile time problem was caused by the fact that not all compilers / STL implementations can automatically convert std::unique_ptr<Derived> to std::unique_ptr<Base>. Fixed (hopefully) by making sure it's std::unique_ptr<Derived>&& (rvalue ref) to std::unique_ptr<Base> conversion instead. llvm-svn: 332917
2018-05-22 07:28:51 +08:00
std::vector<std::unique_ptr<Matcher>> &MatcherStorage) {
[TableGen][GlobalISel] Optimize MatchTable for faster instruction selection *** Context *** Prior to this patchw, the table generated for matching instruction was straight forward but highly inefficient. Basically, each pattern generates its own set of self contained checks and actions. E.g., TableGen generated: // First pattern CheckNumOperand 3 CheckOpcode G_ADD ... Build ADDrr // Second pattern CheckNumOperand 3 CheckOpcode G_ADD ... Build ADDri // Third pattern CheckNumOperand 3 CheckOpcode G_SUB ... Build SUBrr *** Problem *** Because of that generation, a *lot* of check were redundant between each pattern and were checked every single time until we reach the pattern that matches. E.g., Taking the previous table, let say we are matching a G_SUB, that means we were going to check all the rules for G_ADD before looking at the G_SUB rule. In particular we are going to do: check 3 operands; PASS check G_ADD; FAIL ; Next rule check 3 operands; PASS (but we already knew that!) check G_ADD; FAIL (well it is still not true) ; Next rule check 3 operands; PASS (really!!) check G_SUB; PASS (at last :P) *** Proposed Solution *** This patch introduces a concept of group of rules (GroupMatcher) that share some predicates and only get checked once for the whole group. This patch only creates groups with one nesting level. Conceptually there is nothing preventing us for having deeper nest level. However, the current implementation is not smart enough to share the recording (aka capturing) of values. That limits its ability to do more sharing. For the given example the current patch will generate: // First group CheckOpcode G_ADD // First pattern CheckNumOperand 3 ... Build ADDrr // Second pattern CheckNumOperand 3 ... Build ADDri // Second group CheckOpcode G_SUB // Third pattern CheckNumOperand 3 ... Build SUBrr But if we allowed several nesting level, it could create a sub group for the checknumoperand 3. (We would need to call optimizeRules on the rules within a group.) *** Result *** With only one level of nesting, the instruction selection pass is up to 4x faster. For instance, one instruction now takes 500 checks, instead of 24k! With more nesting we could get in the tens I believe. Differential Revision: https://reviews.llvm.org/D39034 rdar://problem/34670699 llvm-svn: 321017
2017-12-19 03:47:41 +08:00
std::vector<Matcher *> OptRules;
Reapply r332907 "[GlobalISel] Improving InstructionSelect's performance by reducing MatchTable..." Apparently the compile time problem was caused by the fact that not all compilers / STL implementations can automatically convert std::unique_ptr<Derived> to std::unique_ptr<Base>. Fixed (hopefully) by making sure it's std::unique_ptr<Derived>&& (rvalue ref) to std::unique_ptr<Base> conversion instead. llvm-svn: 332917
2018-05-22 07:28:51 +08:00
std::unique_ptr<GroupT> CurrentGroup = make_unique<GroupT>();
assert(CurrentGroup->empty() && "Newly created group isn't empty!");
unsigned NumGroups = 0;
auto ProcessCurrentGroup = [&]() {
if (CurrentGroup->empty())
// An empty group is good to be reused:
return;
// If the group isn't large enough to provide any benefit, move all the
// added rules out of it and make sure to re-create the group to properly
// re-initialize it:
if (CurrentGroup->size() < 2)
for (Matcher *M : CurrentGroup->matchers())
OptRules.push_back(M);
else {
CurrentGroup->finalize();
[TableGen][GlobalISel] Optimize MatchTable for faster instruction selection *** Context *** Prior to this patchw, the table generated for matching instruction was straight forward but highly inefficient. Basically, each pattern generates its own set of self contained checks and actions. E.g., TableGen generated: // First pattern CheckNumOperand 3 CheckOpcode G_ADD ... Build ADDrr // Second pattern CheckNumOperand 3 CheckOpcode G_ADD ... Build ADDri // Third pattern CheckNumOperand 3 CheckOpcode G_SUB ... Build SUBrr *** Problem *** Because of that generation, a *lot* of check were redundant between each pattern and were checked every single time until we reach the pattern that matches. E.g., Taking the previous table, let say we are matching a G_SUB, that means we were going to check all the rules for G_ADD before looking at the G_SUB rule. In particular we are going to do: check 3 operands; PASS check G_ADD; FAIL ; Next rule check 3 operands; PASS (but we already knew that!) check G_ADD; FAIL (well it is still not true) ; Next rule check 3 operands; PASS (really!!) check G_SUB; PASS (at last :P) *** Proposed Solution *** This patch introduces a concept of group of rules (GroupMatcher) that share some predicates and only get checked once for the whole group. This patch only creates groups with one nesting level. Conceptually there is nothing preventing us for having deeper nest level. However, the current implementation is not smart enough to share the recording (aka capturing) of values. That limits its ability to do more sharing. For the given example the current patch will generate: // First group CheckOpcode G_ADD // First pattern CheckNumOperand 3 ... Build ADDrr // Second pattern CheckNumOperand 3 ... Build ADDri // Second group CheckOpcode G_SUB // Third pattern CheckNumOperand 3 ... Build SUBrr But if we allowed several nesting level, it could create a sub group for the checknumoperand 3. (We would need to call optimizeRules on the rules within a group.) *** Result *** With only one level of nesting, the instruction selection pass is up to 4x faster. For instance, one instruction now takes 500 checks, instead of 24k! With more nesting we could get in the tens I believe. Differential Revision: https://reviews.llvm.org/D39034 rdar://problem/34670699 llvm-svn: 321017
2017-12-19 03:47:41 +08:00
OptRules.push_back(CurrentGroup.get());
Reapply r332907 "[GlobalISel] Improving InstructionSelect's performance by reducing MatchTable..." Apparently the compile time problem was caused by the fact that not all compilers / STL implementations can automatically convert std::unique_ptr<Derived> to std::unique_ptr<Base>. Fixed (hopefully) by making sure it's std::unique_ptr<Derived>&& (rvalue ref) to std::unique_ptr<Base> conversion instead. llvm-svn: 332917
2018-05-22 07:28:51 +08:00
MatcherStorage.emplace_back(std::move(CurrentGroup));
++NumGroups;
[TableGen][GlobalISel] Optimize MatchTable for faster instruction selection *** Context *** Prior to this patchw, the table generated for matching instruction was straight forward but highly inefficient. Basically, each pattern generates its own set of self contained checks and actions. E.g., TableGen generated: // First pattern CheckNumOperand 3 CheckOpcode G_ADD ... Build ADDrr // Second pattern CheckNumOperand 3 CheckOpcode G_ADD ... Build ADDri // Third pattern CheckNumOperand 3 CheckOpcode G_SUB ... Build SUBrr *** Problem *** Because of that generation, a *lot* of check were redundant between each pattern and were checked every single time until we reach the pattern that matches. E.g., Taking the previous table, let say we are matching a G_SUB, that means we were going to check all the rules for G_ADD before looking at the G_SUB rule. In particular we are going to do: check 3 operands; PASS check G_ADD; FAIL ; Next rule check 3 operands; PASS (but we already knew that!) check G_ADD; FAIL (well it is still not true) ; Next rule check 3 operands; PASS (really!!) check G_SUB; PASS (at last :P) *** Proposed Solution *** This patch introduces a concept of group of rules (GroupMatcher) that share some predicates and only get checked once for the whole group. This patch only creates groups with one nesting level. Conceptually there is nothing preventing us for having deeper nest level. However, the current implementation is not smart enough to share the recording (aka capturing) of values. That limits its ability to do more sharing. For the given example the current patch will generate: // First group CheckOpcode G_ADD // First pattern CheckNumOperand 3 ... Build ADDrr // Second pattern CheckNumOperand 3 ... Build ADDri // Second group CheckOpcode G_SUB // Third pattern CheckNumOperand 3 ... Build SUBrr But if we allowed several nesting level, it could create a sub group for the checknumoperand 3. (We would need to call optimizeRules on the rules within a group.) *** Result *** With only one level of nesting, the instruction selection pass is up to 4x faster. For instance, one instruction now takes 500 checks, instead of 24k! With more nesting we could get in the tens I believe. Differential Revision: https://reviews.llvm.org/D39034 rdar://problem/34670699 llvm-svn: 321017
2017-12-19 03:47:41 +08:00
}
Reapply r332907 "[GlobalISel] Improving InstructionSelect's performance by reducing MatchTable..." Apparently the compile time problem was caused by the fact that not all compilers / STL implementations can automatically convert std::unique_ptr<Derived> to std::unique_ptr<Base>. Fixed (hopefully) by making sure it's std::unique_ptr<Derived>&& (rvalue ref) to std::unique_ptr<Base> conversion instead. llvm-svn: 332917
2018-05-22 07:28:51 +08:00
CurrentGroup = make_unique<GroupT>();
};
for (Matcher *Rule : Rules) {
// Greedily add as many matchers as possible to the current group:
if (CurrentGroup->addMatcher(*Rule))
continue;
ProcessCurrentGroup();
assert(CurrentGroup->empty() && "A group wasn't properly re-initialized");
// Try to add the pending matcher to a newly created empty group:
if (!CurrentGroup->addMatcher(*Rule))
// If we couldn't add the matcher to an empty group, that group type
// doesn't support that kind of matchers at all, so just skip it:
OptRules.push_back(Rule);
Revert r332907 "[GlobalISel] Improving InstructionSelect's performance by reducing MatchTable..." There is a compile time error I didn't see locally, investigating now. llvm-svn: 332912
2018-05-22 06:21:24 +08:00
}
Reapply r332907 "[GlobalISel] Improving InstructionSelect's performance by reducing MatchTable..." Apparently the compile time problem was caused by the fact that not all compilers / STL implementations can automatically convert std::unique_ptr<Derived> to std::unique_ptr<Base>. Fixed (hopefully) by making sure it's std::unique_ptr<Derived>&& (rvalue ref) to std::unique_ptr<Base> conversion instead. llvm-svn: 332917
2018-05-22 07:28:51 +08:00
ProcessCurrentGroup();
Remove DEBUG macro. Now that the LLVM_DEBUG() macro landed on the various sub-projects the DEBUG macro can be removed. Also change the new uses of DEBUG to LLVM_DEBUG. Differential Revision: https://reviews.llvm.org/D46952 llvm-svn: 333091
2018-05-23 23:09:29 +08:00
LLVM_DEBUG(dbgs() << "NumGroups: " << NumGroups << "\n");
Reapply r332907 "[GlobalISel] Improving InstructionSelect's performance by reducing MatchTable..." Apparently the compile time problem was caused by the fact that not all compilers / STL implementations can automatically convert std::unique_ptr<Derived> to std::unique_ptr<Base>. Fixed (hopefully) by making sure it's std::unique_ptr<Derived>&& (rvalue ref) to std::unique_ptr<Base> conversion instead. llvm-svn: 332917
2018-05-22 07:28:51 +08:00
assert(CurrentGroup->empty() && "The last group wasn't properly processed");
[TableGen][GlobalISel] Optimize MatchTable for faster instruction selection *** Context *** Prior to this patchw, the table generated for matching instruction was straight forward but highly inefficient. Basically, each pattern generates its own set of self contained checks and actions. E.g., TableGen generated: // First pattern CheckNumOperand 3 CheckOpcode G_ADD ... Build ADDrr // Second pattern CheckNumOperand 3 CheckOpcode G_ADD ... Build ADDri // Third pattern CheckNumOperand 3 CheckOpcode G_SUB ... Build SUBrr *** Problem *** Because of that generation, a *lot* of check were redundant between each pattern and were checked every single time until we reach the pattern that matches. E.g., Taking the previous table, let say we are matching a G_SUB, that means we were going to check all the rules for G_ADD before looking at the G_SUB rule. In particular we are going to do: check 3 operands; PASS check G_ADD; FAIL ; Next rule check 3 operands; PASS (but we already knew that!) check G_ADD; FAIL (well it is still not true) ; Next rule check 3 operands; PASS (really!!) check G_SUB; PASS (at last :P) *** Proposed Solution *** This patch introduces a concept of group of rules (GroupMatcher) that share some predicates and only get checked once for the whole group. This patch only creates groups with one nesting level. Conceptually there is nothing preventing us for having deeper nest level. However, the current implementation is not smart enough to share the recording (aka capturing) of values. That limits its ability to do more sharing. For the given example the current patch will generate: // First group CheckOpcode G_ADD // First pattern CheckNumOperand 3 ... Build ADDrr // Second pattern CheckNumOperand 3 ... Build ADDri // Second group CheckOpcode G_SUB // Third pattern CheckNumOperand 3 ... Build SUBrr But if we allowed several nesting level, it could create a sub group for the checknumoperand 3. (We would need to call optimizeRules on the rules within a group.) *** Result *** With only one level of nesting, the instruction selection pass is up to 4x faster. For instance, one instruction now takes 500 checks, instead of 24k! With more nesting we could get in the tens I believe. Differential Revision: https://reviews.llvm.org/D39034 rdar://problem/34670699 llvm-svn: 321017
2017-12-19 03:47:41 +08:00
return OptRules;
}
[GlobalISel][InstructionSelect] Refactoring buildMatchTable out, NFC to share it between the Instruction Selector in optimized and non-optimized modes both and the Testgen. Reviewers: dsanders, aemerson Reviewed By: dsanders Subscribers: rovka, kristof.beyls, llvm-commits, dsanders Differential Revision: https://reviews.llvm.org/D46097 llvm-svn: 331396
2018-05-03 04:08:14 +08:00
MatchTable
GlobalISelEmitter::buildMatchTable(MutableArrayRef<RuleMatcher> Rules,
[GlobalISel][InstructionSelect] Making Coverage Info generation optional on per-match table basis to make sure that Testgen always has access to coverage info even if the match table used by the selector itself is stripped off that information for performance reasons. Reviewers: dsanders, aemerson Reviewed By: dsanders Subscribers: rovka, kristof.beyls, llvm-commits, dsanders Differential Revision: https://reviews.llvm.org/D46098 llvm-svn: 331398
2018-05-03 04:15:11 +08:00
bool Optimize, bool WithCoverage) {
[GlobalISel][InstructionSelect] Refactoring buildMatchTable out, NFC to share it between the Instruction Selector in optimized and non-optimized modes both and the Testgen. Reviewers: dsanders, aemerson Reviewed By: dsanders Subscribers: rovka, kristof.beyls, llvm-commits, dsanders Differential Revision: https://reviews.llvm.org/D46097 llvm-svn: 331396
2018-05-03 04:08:14 +08:00
std::vector<Matcher *> InputRules;
for (Matcher &Rule : Rules)
InputRules.push_back(&Rule);
if (!Optimize)
[GlobalISel][InstructionSelect] Making Coverage Info generation optional on per-match table basis to make sure that Testgen always has access to coverage info even if the match table used by the selector itself is stripped off that information for performance reasons. Reviewers: dsanders, aemerson Reviewed By: dsanders Subscribers: rovka, kristof.beyls, llvm-commits, dsanders Differential Revision: https://reviews.llvm.org/D46098 llvm-svn: 331398
2018-05-03 04:15:11 +08:00
return MatchTable::buildTable(InputRules, WithCoverage);
[GlobalISel][InstructionSelect] Refactoring buildMatchTable out, NFC to share it between the Instruction Selector in optimized and non-optimized modes both and the Testgen. Reviewers: dsanders, aemerson Reviewed By: dsanders Subscribers: rovka, kristof.beyls, llvm-commits, dsanders Differential Revision: https://reviews.llvm.org/D46097 llvm-svn: 331396
2018-05-03 04:08:14 +08:00
[GlobalISel][InstructionSelect] Sorting MatchTable's first level by opcodes and num operands, perf patch 3 This patch continues a series of patches started by r332907 (reapplied as r332917) In this commit we start sorting the rules by the opcode first, and if the same, by the number of operands of the root instructions. This allows better grouping and safe as patterns with different opcodes are mutually exclusive. This is expected to decrease time GlobalISel spends in its InstructionSelect pass by roughly 18% for an -O0 build as measured on sqlite3-amalgamation (http://sqlite.org/download.html) targeting AArch64. I'm also removing RuleMatcher::getFirstConditionAsRootType() function here and moving it to a later patch within the series as it's not used yet and was causing a warning on sanitizer-ppc64le-linux bot. Reviewers: qcolombet, dsanders, bogner, aemerson, javed.absar Reviewed By: qcolombet Subscribers: rovka, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D44700 llvm-svn: 333001
2018-05-23 00:54:27 +08:00
unsigned CurrentOrdering = 0;
StringMap<unsigned> OpcodeOrder;
for (RuleMatcher &Rule : Rules) {
const StringRef Opcode = Rule.getOpcode();
assert(!Opcode.empty() && "Didn't expect an undefined opcode");
if (OpcodeOrder.count(Opcode) == 0)
OpcodeOrder[Opcode] = CurrentOrdering++;
}
std::stable_sort(InputRules.begin(), InputRules.end(),
[&OpcodeOrder](const Matcher *A, const Matcher *B) {
auto *L = static_cast<const RuleMatcher *>(A);
auto *R = static_cast<const RuleMatcher *>(B);
return std::make_tuple(OpcodeOrder[L->getOpcode()],
L->getNumOperands()) <
std::make_tuple(OpcodeOrder[R->getOpcode()],
R->getNumOperands());
});
Reapply r332907 "[GlobalISel] Improving InstructionSelect's performance by reducing MatchTable..." Apparently the compile time problem was caused by the fact that not all compilers / STL implementations can automatically convert std::unique_ptr<Derived> to std::unique_ptr<Base>. Fixed (hopefully) by making sure it's std::unique_ptr<Derived>&& (rvalue ref) to std::unique_ptr<Base> conversion instead. llvm-svn: 332917
2018-05-22 07:28:51 +08:00
for (Matcher *Rule : InputRules)
Rule->optimize();
std::vector<std::unique_ptr<Matcher>> MatcherStorage;
[GlobalISel][InstructionSelect] Refactoring buildMatchTable out, NFC to share it between the Instruction Selector in optimized and non-optimized modes both and the Testgen. Reviewers: dsanders, aemerson Reviewed By: dsanders Subscribers: rovka, kristof.beyls, llvm-commits, dsanders Differential Revision: https://reviews.llvm.org/D46097 llvm-svn: 331396
2018-05-03 04:08:14 +08:00
std::vector<Matcher *> OptRules =
Reapply r332907 "[GlobalISel] Improving InstructionSelect's performance by reducing MatchTable..." Apparently the compile time problem was caused by the fact that not all compilers / STL implementations can automatically convert std::unique_ptr<Derived> to std::unique_ptr<Base>. Fixed (hopefully) by making sure it's std::unique_ptr<Derived>&& (rvalue ref) to std::unique_ptr<Base> conversion instead. llvm-svn: 332917
2018-05-22 07:28:51 +08:00
optimizeRules<GroupMatcher>(InputRules, MatcherStorage);
for (Matcher *Rule : OptRules)
Rule->optimize();
[GlobalISel][InstructionSelect] Refactoring buildMatchTable out, NFC to share it between the Instruction Selector in optimized and non-optimized modes both and the Testgen. Reviewers: dsanders, aemerson Reviewed By: dsanders Subscribers: rovka, kristof.beyls, llvm-commits, dsanders Differential Revision: https://reviews.llvm.org/D46097 llvm-svn: 331396
2018-05-03 04:08:14 +08:00
[GlobalISel][InstructionSelect] Switching MatchTable over opcodes, perf patch 4 This patch continues a series of patches started by r332907 (reapplied as r332917) In this commit we introduce a new matching opcode GIM_SwitchOpcode that implements a jump table over opcodes and start emitting them for root instructions. This is expected to decrease time GlobalISel spends in its InstructionSelect pass by roughly 20% for an -O0 build as measured on sqlite3-amalgamation (http://sqlite.org/download.html) targeting AArch64. To some degree, we assume here that the opcodes form a dense set, which is true at the moment for all upstream targets given the limitations of our rule importing mechanism. It might not be true for out of tree targets, specifically due to pseudo's. If so, we might noticeably increase the size of the MatchTable with this patch due to padding zeros. This will be addressed later. Reviewers: qcolombet, dsanders, bogner, aemerson, javed.absar Reviewed By: qcolombet Subscribers: rovka, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D44700 llvm-svn: 333017
2018-05-23 03:37:59 +08:00
OptRules = optimizeRules<SwitchMatcher>(OptRules, MatcherStorage);
[GlobalISel][InstructionSelect] Making Coverage Info generation optional on per-match table basis to make sure that Testgen always has access to coverage info even if the match table used by the selector itself is stripped off that information for performance reasons. Reviewers: dsanders, aemerson Reviewed By: dsanders Subscribers: rovka, kristof.beyls, llvm-commits, dsanders Differential Revision: https://reviews.llvm.org/D46098 llvm-svn: 331398
2018-05-03 04:15:11 +08:00
return MatchTable::buildTable(OptRules, WithCoverage);
[GlobalISel][InstructionSelect] Refactoring buildMatchTable out, NFC to share it between the Instruction Selector in optimized and non-optimized modes both and the Testgen. Reviewers: dsanders, aemerson Reviewed By: dsanders Subscribers: rovka, kristof.beyls, llvm-commits, dsanders Differential Revision: https://reviews.llvm.org/D46097 llvm-svn: 331396
2018-05-03 04:08:14 +08:00
}
[GlobalISel][InstructionSelect] MatchTable second level grouping, perf patch 5 This patch continues a series of patches started by r332907 (reapplied as r332917) In this commit we start grouping rules with common first condition on the second level of the table. This is expected to decrease time GlobalISel spends in its InstructionSelect pass by roughly 13% for an -O0 build as measured on sqlite3-amalgamation (http://sqlite.org/download.html) targeting AArch64. Reviewers: qcolombet, dsanders, bogner, aemerson, javed.absar Reviewed By: qcolombet Subscribers: rovka, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D44700 llvm-svn: 333053
2018-05-23 10:04:19 +08:00
void GroupMatcher::optimize() {
GlobalISelEmitter::optimizeRules<GroupMatcher>(Matchers, MatcherStorage)
.swap(Matchers);
}
[GlobalISel] Add basic Selector-emitter tblgen backend. This adds a basic tablegen backend that analyzes the SelectionDAG patterns to find simple ones that are eligible for GlobalISel-emission. That's similar to FastISel, with one notable difference: we're not fed ISD opcodes, so we need to map the SDNode operators to generic opcodes. That's done using GINodeEquiv in TargetGlobalISel.td. Otherwise, this is mostly boilerplate, and lots of filtering of any kind of "complicated" pattern. On AArch64, this is sufficient to match G_ADD up to s64 (to ADDWrr/ADDXrr) and G_BR (to B). Differential Revision: https://reviews.llvm.org/D26878 llvm-svn: 290284
2016-12-22 07:26:20 +08:00
void GlobalISelEmitter::run(raw_ostream &OS) {
[globalisel][tablegen] Generate rule coverage and use it to identify untested rules Summary: This patch adds a LLVM_ENABLE_GISEL_COV which, like LLVM_ENABLE_DAGISEL_COV, causes TableGen to instrument the generated table to collect rule coverage information. However, LLVM_ENABLE_GISEL_COV goes a bit further than LLVM_ENABLE_DAGISEL_COV. The information is written to files (${CMAKE_BINARY_DIR}/gisel-coverage-* by default). These files can then be concatenated into ${LLVM_GISEL_COV_PREFIX}-all after which TableGen will read this information and use it to emit warnings about untested rules. This technique could also be used by SelectionDAG and can be further extended to detect hot rules and give them priority over colder rules. Usage: * Enable LLVM_ENABLE_GISEL_COV in CMake * Build the compiler and run some tests * cat gisel-coverage-[0-9]* > gisel-coverage-all * Delete lib/Target/*/*GenGlobalISel.inc* * Build the compiler Known issues: * ${LLVM_GISEL_COV_PREFIX}-all must be generated as a manual step due to a lack of a portable 'cat' command. It should be the concatenation of all ${LLVM_GISEL_COV_PREFIX}-[0-9]* files. * There's no mechanism to discard coverage information when the ruleset changes Depends on D39742 Reviewers: ab, qcolombet, t.p.northover, aditya_nandakumar, rovka Reviewed By: rovka Subscribers: vsk, arsenm, nhaehnle, mgorny, kristof.beyls, javed.absar, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D39747 llvm-svn: 318356
2017-11-16 08:46:35 +08:00
if (!UseCoverageFile.empty()) {
RuleCoverage = CodeGenCoverage();
auto RuleCoverageBufOrErr = MemoryBuffer::getFile(UseCoverageFile);
if (!RuleCoverageBufOrErr) {
PrintWarning(SMLoc(), "Missing rule coverage data");
RuleCoverage = None;
} else {
if (!RuleCoverage->parse(*RuleCoverageBufOrErr.get(), Target.getName())) {
PrintWarning(SMLoc(), "Ignoring invalid or missing rule coverage data");
RuleCoverage = None;
}
}
}
Reapply r332907 "[GlobalISel] Improving InstructionSelect's performance by reducing MatchTable..." Apparently the compile time problem was caused by the fact that not all compilers / STL implementations can automatically convert std::unique_ptr<Derived> to std::unique_ptr<Base>. Fixed (hopefully) by making sure it's std::unique_ptr<Derived>&& (rvalue ref) to std::unique_ptr<Base> conversion instead. llvm-svn: 332917
2018-05-22 07:28:51 +08:00
// Track the run-time opcode values
gatherOpcodeValues();
// Track the run-time LLT ID values
gatherTypeIDValues();
[GlobalISel] Add basic Selector-emitter tblgen backend. This adds a basic tablegen backend that analyzes the SelectionDAG patterns to find simple ones that are eligible for GlobalISel-emission. That's similar to FastISel, with one notable difference: we're not fed ISD opcodes, so we need to map the SDNode operators to generic opcodes. That's done using GINodeEquiv in TargetGlobalISel.td. Otherwise, this is mostly boilerplate, and lots of filtering of any kind of "complicated" pattern. On AArch64, this is sufficient to match G_ADD up to s64 (to ADDWrr/ADDXrr) and G_BR (to B). Differential Revision: https://reviews.llvm.org/D26878 llvm-svn: 290284
2016-12-22 07:26:20 +08:00
// Track the GINodeEquiv definitions.
gatherNodeEquivs();
emitSourceFileHeader(("Global Instruction Selector for the " +
Target.getName() + " target").str(), OS);
[globalisel] Separate the SelectionDAG importer from the emitter. NFC Summary: In the near future the rules will be sorted between these two steps to ensure that more important rules are not prevented by less important ones. Reviewers: t.p.northover, ab, rovka, qcolombet, aditya_nandakumar Reviewed By: ab Subscribers: dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D29709 llvm-svn: 295661
2017-02-20 22:31:27 +08:00
std::vector<RuleMatcher> Rules;
[GlobalISel] Add basic Selector-emitter tblgen backend. This adds a basic tablegen backend that analyzes the SelectionDAG patterns to find simple ones that are eligible for GlobalISel-emission. That's similar to FastISel, with one notable difference: we're not fed ISD opcodes, so we need to map the SDNode operators to generic opcodes. That's done using GINodeEquiv in TargetGlobalISel.td. Otherwise, this is mostly boilerplate, and lots of filtering of any kind of "complicated" pattern. On AArch64, this is sufficient to match G_ADD up to s64 (to ADDWrr/ADDXrr) and G_BR (to B). Differential Revision: https://reviews.llvm.org/D26878 llvm-svn: 290284
2016-12-22 07:26:20 +08:00
// Look through the SelectionDAG patterns we found, possibly emitting some.
for (const PatternToMatch &Pat : CGP.ptms()) {
++NumPatternTotal;
[globalisel][tablegen] Import signextload and zeroextload. Allow a pattern rewriter to be installed in CodeGenDAGPatterns and use it to correct situations where SelectionDAG and GlobalISel disagree on representation. For example, it would rewrite: (sextload:i32 $ptr)<<unindexedload>><<sextload>><<sextloadi16> to: (sext:i32 (load:i16 $ptr)<<unindexedload>>) I'd have preferred to replace the fragments and have the expansion happen naturally as part of PatFrag expansion but the type inferencing system can't cope with loads of types narrower than those mentioned in register classes. This is because the SDTCisInt's on the sext constrain both the result and operand to the 'legal' integer types (where legal is defined as 'a register class can contain the type') which immediately rules the narrower types out. Several targets (those with only one legal integer type) would then go on to crash on the SDTCisOpSmallerThanOp<> when it removes all the possible types for the result of the extend. Also, improve isObviouslySafeToFold() slightly to automatically return true for neighbouring instructions. There can't be any re-ordering problems if re-ordering isn't happenning. We'll need to improve it further to handle sign/zero-extending loads when the extend and load aren't immediate neighbours though. llvm-svn: 317971
2017-11-11 11:23:44 +08:00
[GlobalISel] Return an Expected<RuleMatcher> for each SDAG pattern. NFC. Instead of emitting the matcher code directly, return the rule matcher and the skip reason as an Expected<RuleMatcher>. This will let us record all matchers and process them before emission. It's a somewhat unconventional use of Error, but it's nicer than, say, std::pair, because of the bool conversions. Differential Revision: https://reviews.llvm.org/D29743 llvm-svn: 294706
2017-02-10 12:00:17 +08:00
auto MatcherOrErr = runOnPattern(Pat);
// The pattern analysis can fail, indicating an unsupported pattern.
// Report that if we've been asked to do so.
if (auto Err = MatcherOrErr.takeError()) {
[GlobalISel] Add basic Selector-emitter tblgen backend. This adds a basic tablegen backend that analyzes the SelectionDAG patterns to find simple ones that are eligible for GlobalISel-emission. That's similar to FastISel, with one notable difference: we're not fed ISD opcodes, so we need to map the SDNode operators to generic opcodes. That's done using GINodeEquiv in TargetGlobalISel.td. Otherwise, this is mostly boilerplate, and lots of filtering of any kind of "complicated" pattern. On AArch64, this is sufficient to match G_ADD up to s64 (to ADDWrr/ADDXrr) and G_BR (to B). Differential Revision: https://reviews.llvm.org/D26878 llvm-svn: 290284
2016-12-22 07:26:20 +08:00
if (WarnOnSkippedPatterns) {
PrintWarning(Pat.getSrcRecord()->getLoc(),
[GlobalISel] Return an Expected<RuleMatcher> for each SDAG pattern. NFC. Instead of emitting the matcher code directly, return the rule matcher and the skip reason as an Expected<RuleMatcher>. This will let us record all matchers and process them before emission. It's a somewhat unconventional use of Error, but it's nicer than, say, std::pair, because of the bool conversions. Differential Revision: https://reviews.llvm.org/D29743 llvm-svn: 294706
2017-02-10 12:00:17 +08:00
"Skipped pattern: " + toString(std::move(Err)));
} else {
consumeError(std::move(Err));
[GlobalISel] Add basic Selector-emitter tblgen backend. This adds a basic tablegen backend that analyzes the SelectionDAG patterns to find simple ones that are eligible for GlobalISel-emission. That's similar to FastISel, with one notable difference: we're not fed ISD opcodes, so we need to map the SDNode operators to generic opcodes. That's done using GINodeEquiv in TargetGlobalISel.td. Otherwise, this is mostly boilerplate, and lots of filtering of any kind of "complicated" pattern. On AArch64, this is sufficient to match G_ADD up to s64 (to ADDWrr/ADDXrr) and G_BR (to B). Differential Revision: https://reviews.llvm.org/D26878 llvm-svn: 290284
2016-12-22 07:26:20 +08:00
}
[globalisel] Separate the SelectionDAG importer from the emitter. NFC Summary: In the near future the rules will be sorted between these two steps to ensure that more important rules are not prevented by less important ones. Reviewers: t.p.northover, ab, rovka, qcolombet, aditya_nandakumar Reviewed By: ab Subscribers: dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D29709 llvm-svn: 295661
2017-02-20 22:31:27 +08:00
++NumPatternImportsSkipped;
[GlobalISel] Return an Expected<RuleMatcher> for each SDAG pattern. NFC. Instead of emitting the matcher code directly, return the rule matcher and the skip reason as an Expected<RuleMatcher>. This will let us record all matchers and process them before emission. It's a somewhat unconventional use of Error, but it's nicer than, say, std::pair, because of the bool conversions. Differential Revision: https://reviews.llvm.org/D29743 llvm-svn: 294706
2017-02-10 12:00:17 +08:00
continue;
[GlobalISel] Add basic Selector-emitter tblgen backend. This adds a basic tablegen backend that analyzes the SelectionDAG patterns to find simple ones that are eligible for GlobalISel-emission. That's similar to FastISel, with one notable difference: we're not fed ISD opcodes, so we need to map the SDNode operators to generic opcodes. That's done using GINodeEquiv in TargetGlobalISel.td. Otherwise, this is mostly boilerplate, and lots of filtering of any kind of "complicated" pattern. On AArch64, this is sufficient to match G_ADD up to s64 (to ADDWrr/ADDXrr) and G_BR (to B). Differential Revision: https://reviews.llvm.org/D26878 llvm-svn: 290284
2016-12-22 07:26:20 +08:00
}
[GlobalISel] Return an Expected<RuleMatcher> for each SDAG pattern. NFC. Instead of emitting the matcher code directly, return the rule matcher and the skip reason as an Expected<RuleMatcher>. This will let us record all matchers and process them before emission. It's a somewhat unconventional use of Error, but it's nicer than, say, std::pair, because of the bool conversions. Differential Revision: https://reviews.llvm.org/D29743 llvm-svn: 294706
2017-02-10 12:00:17 +08:00
[globalisel][tablegen] Generate rule coverage and use it to identify untested rules Summary: This patch adds a LLVM_ENABLE_GISEL_COV which, like LLVM_ENABLE_DAGISEL_COV, causes TableGen to instrument the generated table to collect rule coverage information. However, LLVM_ENABLE_GISEL_COV goes a bit further than LLVM_ENABLE_DAGISEL_COV. The information is written to files (${CMAKE_BINARY_DIR}/gisel-coverage-* by default). These files can then be concatenated into ${LLVM_GISEL_COV_PREFIX}-all after which TableGen will read this information and use it to emit warnings about untested rules. This technique could also be used by SelectionDAG and can be further extended to detect hot rules and give them priority over colder rules. Usage: * Enable LLVM_ENABLE_GISEL_COV in CMake * Build the compiler and run some tests * cat gisel-coverage-[0-9]* > gisel-coverage-all * Delete lib/Target/*/*GenGlobalISel.inc* * Build the compiler Known issues: * ${LLVM_GISEL_COV_PREFIX}-all must be generated as a manual step due to a lack of a portable 'cat' command. It should be the concatenation of all ${LLVM_GISEL_COV_PREFIX}-[0-9]* files. * There's no mechanism to discard coverage information when the ruleset changes Depends on D39742 Reviewers: ab, qcolombet, t.p.northover, aditya_nandakumar, rovka Reviewed By: rovka Subscribers: vsk, arsenm, nhaehnle, mgorny, kristof.beyls, javed.absar, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D39747 llvm-svn: 318356
2017-11-16 08:46:35 +08:00
if (RuleCoverage) {
if (RuleCoverage->isCovered(MatcherOrErr->getRuleID()))
++NumPatternsTested;
else
PrintWarning(Pat.getSrcRecord()->getLoc(),
"Pattern is not covered by a test");
}
[globalisel] Separate the SelectionDAG importer from the emitter. NFC Summary: In the near future the rules will be sorted between these two steps to ensure that more important rules are not prevented by less important ones. Reviewers: t.p.northover, ab, rovka, qcolombet, aditya_nandakumar Reviewed By: ab Subscribers: dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D29709 llvm-svn: 295661
2017-02-20 22:31:27 +08:00
Rules.push_back(std::move(MatcherOrErr.get()));
}
[GlobalISel][TableGen] Add support for SDNodeXForm Summary: This patch adds CustomRenderer which renders the matched operands to the specified instruction. Targets can enable the matching of SDNodeXForm by adding a definition that inherits from GICustomOperandRenderer and GISDNodeXFormEquiv as follows. def gi_imm8 : GICustomOperandRenderer<"renderImm8”>, GISDNodeXFormEquiv<imm8_xform>; Custom renderer functions should be of the form: void render(MachineInstrBuilder &MIB, const MachineInstr &I); Reviewers: dsanders, ab, rovka Reviewed By: dsanders Subscribers: kristof.beyls, javed.absar, llvm-commits, mgrang, qcolombet Differential Revision: https://reviews.llvm.org/D42012 llvm-svn: 322582
2018-01-17 02:44:05 +08:00
// Comparison function to order records by name.
auto orderByName = [](const Record *A, const Record *B) {
return A->getName() < B->getName();
};
[globalisel][tablegen] Partially fix compile-time regressions by converting matcher to state-machine(s) Summary: Replace the matcher if-statements for each rule with a state-machine. This significantly reduces compile time, memory allocations, and cumulative memory allocation when compiling AArch64InstructionSelector.cpp.o after r303259 is recommitted. The following patches will expand on this further to fully fix the regressions. Reviewers: rovka, ab, t.p.northover, qcolombet, aditya_nandakumar Reviewed By: ab Subscribers: vitalybuka, aemerson, javed.absar, igorb, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D33758 llvm-svn: 307079
2017-07-04 22:35:06 +08:00
std::vector<Record *> ComplexPredicates =
RK.getAllDerivedDefinitions("GIComplexOperandMatcher");
[TableGen] Change std::sort to llvm::sort in response to r327219 Summary: r327219 added wrappers to std::sort which randomly shuffle the container before sorting. This will help in uncovering non-determinism caused due to undefined sorting order of objects having the same key. To make use of that infrastructure we need to invoke llvm::sort instead of std::sort. Note: This patch is one of a series of patches to replace *all* std::sort to llvm::sort. Refer the comments section in D44363 for a list of all the required patches. Reviewers: stoklund, kparzysz, dsanders Reviewed By: dsanders Subscribers: llvm-commits Differential Revision: https://reviews.llvm.org/D45144 llvm-svn: 329451
2018-04-07 04:18:05 +08:00
llvm::sort(ComplexPredicates.begin(), ComplexPredicates.end(), orderByName);
[GlobalISel][TableGen] Add support for SDNodeXForm Summary: This patch adds CustomRenderer which renders the matched operands to the specified instruction. Targets can enable the matching of SDNodeXForm by adding a definition that inherits from GICustomOperandRenderer and GISDNodeXFormEquiv as follows. def gi_imm8 : GICustomOperandRenderer<"renderImm8”>, GISDNodeXFormEquiv<imm8_xform>; Custom renderer functions should be of the form: void render(MachineInstrBuilder &MIB, const MachineInstr &I); Reviewers: dsanders, ab, rovka Reviewed By: dsanders Subscribers: kristof.beyls, javed.absar, llvm-commits, mgrang, qcolombet Differential Revision: https://reviews.llvm.org/D42012 llvm-svn: 322582
2018-01-17 02:44:05 +08:00
std::vector<Record *> CustomRendererFns =
RK.getAllDerivedDefinitions("GICustomOperandRenderer");
[TableGen] Change std::sort to llvm::sort in response to r327219 Summary: r327219 added wrappers to std::sort which randomly shuffle the container before sorting. This will help in uncovering non-determinism caused due to undefined sorting order of objects having the same key. To make use of that infrastructure we need to invoke llvm::sort instead of std::sort. Note: This patch is one of a series of patches to replace *all* std::sort to llvm::sort. Refer the comments section in D44363 for a list of all the required patches. Reviewers: stoklund, kparzysz, dsanders Reviewed By: dsanders Subscribers: llvm-commits Differential Revision: https://reviews.llvm.org/D45144 llvm-svn: 329451
2018-04-07 04:18:05 +08:00
llvm::sort(CustomRendererFns.begin(), CustomRendererFns.end(), orderByName);
[GlobalISel][TableGen] Add support for SDNodeXForm Summary: This patch adds CustomRenderer which renders the matched operands to the specified instruction. Targets can enable the matching of SDNodeXForm by adding a definition that inherits from GICustomOperandRenderer and GISDNodeXFormEquiv as follows. def gi_imm8 : GICustomOperandRenderer<"renderImm8”>, GISDNodeXFormEquiv<imm8_xform>; Custom renderer functions should be of the form: void render(MachineInstrBuilder &MIB, const MachineInstr &I); Reviewers: dsanders, ab, rovka Reviewed By: dsanders Subscribers: kristof.beyls, javed.absar, llvm-commits, mgrang, qcolombet Differential Revision: https://reviews.llvm.org/D42012 llvm-svn: 322582
2018-01-17 02:44:05 +08:00
[globalisel][tblgen] Add support for ComplexPatterns Summary: Adds a new kind of MachineOperand: MO_Placeholder. This operand must not appear in the MIR and only exists as a way of creating an 'uninitialized' operand until a matcher function overwrites it. Depends on D30046, D29712 Reviewers: t.p.northover, ab, rovka, aditya_nandakumar, javed.absar, qcolombet Reviewed By: qcolombet Subscribers: dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D30089 llvm-svn: 297782
2017-03-15 05:32:08 +08:00
unsigned MaxTemporaries = 0;
for (const auto &Rule : Rules)
[globalisel][tablegen] Revise API for ComplexPattern operands to improve flexibility. Summary: Some targets need to be able to do more complex rendering than just adding an operand or two to an instruction. For example, it may need to insert an instruction to extract a subreg first, or it may need to perform an operation on the operand. In SelectionDAG, targets would create SDNode's to achieve the desired effect during the complex pattern predicate. This worked because SelectionDAG had a form of garbage collection that would take care of SDNode's that were created but not used due to a later predicate rejecting a match. This doesn't translate well to GlobalISel and the churn was wasteful. The API changes in this patch enable GlobalISel to accomplish the same thing without the waste. The API is now: InstructionSelector::OptionalComplexRendererFn selectArithImmed(MachineOperand &Root) const; where Root is the root of the match. The return value can be omitted to indicate that the predicate failed to match, or a function with the signature ComplexRendererFn can be returned. For example: return OptionalComplexRendererFn( [=](MachineInstrBuilder &MIB) { MIB.addImm(Immed).addImm(ShVal); }); adds two immediate operands to the rendered instruction. Immed and ShVal are captured from the predicate function. As an added bonus, this also reduces the amount of information we need to provide to GIComplexOperandMatcher. Depends on D31418 Reviewers: aditya_nandakumar, t.p.northover, qcolombet, rovka, ab, javed.absar Reviewed By: ab Subscribers: dberris, kristof.beyls, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D31761 llvm-svn: 301079
2017-04-22 23:11:04 +08:00
MaxTemporaries = std::max(MaxTemporaries, Rule.countRendererFns());
[globalisel][tblgen] Add support for ComplexPatterns Summary: Adds a new kind of MachineOperand: MO_Placeholder. This operand must not appear in the MIR and only exists as a way of creating an 'uninitialized' operand until a matcher function overwrites it. Depends on D30046, D29712 Reviewers: t.p.northover, ab, rovka, aditya_nandakumar, javed.absar, qcolombet Reviewed By: qcolombet Subscribers: dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D30089 llvm-svn: 297782
2017-03-15 05:32:08 +08:00
[globalisel][tablegen] Import SelectionDAG's rule predicates and support the equivalent in GIRule. Summary: The SelectionDAG importer now imports rules with Predicate's attached via Requires, PredicateControl, etc. These predicates are implemented as bitset's to allow multiple predicates to be tested together. However, unlike the MC layer subtarget features, each target only pays for it's own predicates (e.g. AArch64 doesn't have 192 feature bits just because X86 needs a lot). Both AArch64 and X86 derive at least one predicate from the MachineFunction or Function so they must re-initialize AvailableFeatures before each function. They also declare locals in <Target>InstructionSelector so that computeAvailableFeatures() can use the code from SelectionDAG without modification. Reviewers: rovka, qcolombet, aditya_nandakumar, t.p.northover, ab Reviewed By: rovka Subscribers: aemerson, rengolin, dberris, kristof.beyls, llvm-commits, igorb Differential Revision: https://reviews.llvm.org/D31418 llvm-svn: 300993
2017-04-21 23:59:56 +08:00
OS << "#ifdef GET_GLOBALISEL_PREDICATE_BITSET\n"
<< "const unsigned MAX_SUBTARGET_PREDICATES = " << SubtargetFeatures.size()
<< ";\n"
<< "using PredicateBitset = "
"llvm::PredicateBitsetImpl<MAX_SUBTARGET_PREDICATES>;\n"
<< "#endif // ifdef GET_GLOBALISEL_PREDICATE_BITSET\n\n";
[globalisel][tablegen] Partially fix compile-time regressions by converting matcher to state-machine(s) Summary: Replace the matcher if-statements for each rule with a state-machine. This significantly reduces compile time, memory allocations, and cumulative memory allocation when compiling AArch64InstructionSelector.cpp.o after r303259 is recommitted. The following patches will expand on this further to fully fix the regressions. Reviewers: rovka, ab, t.p.northover, qcolombet, aditya_nandakumar Reviewed By: ab Subscribers: vitalybuka, aemerson, javed.absar, igorb, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D33758 llvm-svn: 307079
2017-07-04 22:35:06 +08:00
OS << "#ifdef GET_GLOBALISEL_TEMPORARIES_DECL\n"
<< " mutable MatcherState State;\n"
<< " typedef "
[globalisel][tablegen] Fix small spelling nits. NFC ComplexRendererFn -> ComplexRendererFns Corrected a couple lingering references to tied operands that were missed. llvm-svn: 316237
2017-10-21 04:55:29 +08:00
"ComplexRendererFns("
[globalisel][tablegen] Partially fix compile-time regressions by converting matcher to state-machine(s) Summary: Replace the matcher if-statements for each rule with a state-machine. This significantly reduces compile time, memory allocations, and cumulative memory allocation when compiling AArch64InstructionSelector.cpp.o after r303259 is recommitted. The following patches will expand on this further to fully fix the regressions. Reviewers: rovka, ab, t.p.northover, qcolombet, aditya_nandakumar Reviewed By: ab Subscribers: vitalybuka, aemerson, javed.absar, igorb, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D33758 llvm-svn: 307079
2017-07-04 22:35:06 +08:00
<< Target.getName()
<< "InstructionSelector::*ComplexMatcherMemFn)(MachineOperand &) const;\n"
[GlobalISel][TableGen] Add support for SDNodeXForm Summary: This patch adds CustomRenderer which renders the matched operands to the specified instruction. Targets can enable the matching of SDNodeXForm by adding a definition that inherits from GICustomOperandRenderer and GISDNodeXFormEquiv as follows. def gi_imm8 : GICustomOperandRenderer<"renderImm8”>, GISDNodeXFormEquiv<imm8_xform>; Custom renderer functions should be of the form: void render(MachineInstrBuilder &MIB, const MachineInstr &I); Reviewers: dsanders, ab, rovka Reviewed By: dsanders Subscribers: kristof.beyls, javed.absar, llvm-commits, mgrang, qcolombet Differential Revision: https://reviews.llvm.org/D42012 llvm-svn: 322582
2018-01-17 02:44:05 +08:00
<< " typedef void(" << Target.getName()
<< "InstructionSelector::*CustomRendererFn)(MachineInstrBuilder &, const "
"MachineInstr&) "
"const;\n"
<< " const ISelInfoTy<PredicateBitset, ComplexMatcherMemFn, "
"CustomRendererFn> "
"ISelInfo;\n";
OS << " static " << Target.getName()
Re-commit r315885: [globalisel][tblgen] Add support for iPTR and implement am_unscaled* and am_indexed* Summary: iPTR is a pointer of subtarget-specific size to any address space. Therefore type checks on this size derive the SizeInBits from a subtarget hook. At this point, we can import the simplests G_LOAD rules and select load instructions using them. Further patches will support for the predicates to enable additional loads as well as the stores. The previous commit failed on MSVC due to a failure to convert an initializer_list to a std::vector. Hopefully, MSVC will accept this version. Depends on D37457 Reviewers: ab, qcolombet, t.p.northover, rovka, aditya_nandakumar Reviewed By: qcolombet Subscribers: kristof.beyls, javed.absar, llvm-commits, igorb Differential Revision: https://reviews.llvm.org/D37458 llvm-svn: 315887
2017-10-16 11:36:29 +08:00
<< "InstructionSelector::ComplexMatcherMemFn ComplexPredicateFns[];\n"
[GlobalISel][TableGen] Add support for SDNodeXForm Summary: This patch adds CustomRenderer which renders the matched operands to the specified instruction. Targets can enable the matching of SDNodeXForm by adding a definition that inherits from GICustomOperandRenderer and GISDNodeXFormEquiv as follows. def gi_imm8 : GICustomOperandRenderer<"renderImm8”>, GISDNodeXFormEquiv<imm8_xform>; Custom renderer functions should be of the form: void render(MachineInstrBuilder &MIB, const MachineInstr &I); Reviewers: dsanders, ab, rovka Reviewed By: dsanders Subscribers: kristof.beyls, javed.absar, llvm-commits, mgrang, qcolombet Differential Revision: https://reviews.llvm.org/D42012 llvm-svn: 322582
2018-01-17 02:44:05 +08:00
<< " static " << Target.getName()
<< "InstructionSelector::CustomRendererFn CustomRenderers[];\n"
[GlobalISel][InstructionSelect] Refactoring out a getMatchTable virtual method + other small NFC's The main goal is to share getMatchTable between the Instruction Selector and the Testgen. The commit also contains some NFC only loosely related to refactoring out the getMatchTable, but strongly related to the initial Testgen patch (see https://reviews.llvm.org/D43962) Reviewers: dsanders, aemerson Reviewed By: dsanders Subscribers: rovka, kristof.beyls, llvm-commits, dsanders Differential Revision: https://reviews.llvm.org/D46096 llvm-svn: 331395
2018-05-03 04:07:15 +08:00
<< " bool testImmPredicate_I64(unsigned PredicateID, int64_t Imm) const "
[globalisel][tablegen] Allow ImmLeaf predicates to use InstructionSelector members NFC for currently supported targets. This resolves a problem encountered by targets such as RISCV that reference `Subtarget` in ImmLeaf predicates. llvm-svn: 321176
2017-12-20 22:41:51 +08:00
"override;\n"
[GlobalISel][InstructionSelect] Refactoring out a getMatchTable virtual method + other small NFC's The main goal is to share getMatchTable between the Instruction Selector and the Testgen. The commit also contains some NFC only loosely related to refactoring out the getMatchTable, but strongly related to the initial Testgen patch (see https://reviews.llvm.org/D43962) Reviewers: dsanders, aemerson Reviewed By: dsanders Subscribers: rovka, kristof.beyls, llvm-commits, dsanders Differential Revision: https://reviews.llvm.org/D46096 llvm-svn: 331395
2018-05-03 04:07:15 +08:00
<< " bool testImmPredicate_APInt(unsigned PredicateID, const APInt &Imm) "
[globalisel][tablegen] Allow ImmLeaf predicates to use InstructionSelector members NFC for currently supported targets. This resolves a problem encountered by targets such as RISCV that reference `Subtarget` in ImmLeaf predicates. llvm-svn: 321176
2017-12-20 22:41:51 +08:00
"const override;\n"
[GlobalISel][InstructionSelect] Refactoring out a getMatchTable virtual method + other small NFC's The main goal is to share getMatchTable between the Instruction Selector and the Testgen. The commit also contains some NFC only loosely related to refactoring out the getMatchTable, but strongly related to the initial Testgen patch (see https://reviews.llvm.org/D43962) Reviewers: dsanders, aemerson Reviewed By: dsanders Subscribers: rovka, kristof.beyls, llvm-commits, dsanders Differential Revision: https://reviews.llvm.org/D46096 llvm-svn: 331395
2018-05-03 04:07:15 +08:00
<< " bool testImmPredicate_APFloat(unsigned PredicateID, const APFloat "
[globalisel][tablegen] Allow ImmLeaf predicates to use InstructionSelector members NFC for currently supported targets. This resolves a problem encountered by targets such as RISCV that reference `Subtarget` in ImmLeaf predicates. llvm-svn: 321176
2017-12-20 22:41:51 +08:00
"&Imm) const override;\n"
[GlobalISel][InstructionSelect] Refactoring out a getMatchTable virtual method + other small NFC's The main goal is to share getMatchTable between the Instruction Selector and the Testgen. The commit also contains some NFC only loosely related to refactoring out the getMatchTable, but strongly related to the initial Testgen patch (see https://reviews.llvm.org/D43962) Reviewers: dsanders, aemerson Reviewed By: dsanders Subscribers: rovka, kristof.beyls, llvm-commits, dsanders Differential Revision: https://reviews.llvm.org/D46096 llvm-svn: 331395
2018-05-03 04:07:15 +08:00
<< " const int64_t *getMatchTable() const override;\n"
[globalisel][tablegen] Partially fix compile-time regressions by converting matcher to state-machine(s) Summary: Replace the matcher if-statements for each rule with a state-machine. This significantly reduces compile time, memory allocations, and cumulative memory allocation when compiling AArch64InstructionSelector.cpp.o after r303259 is recommitted. The following patches will expand on this further to fully fix the regressions. Reviewers: rovka, ab, t.p.northover, qcolombet, aditya_nandakumar Reviewed By: ab Subscribers: vitalybuka, aemerson, javed.absar, igorb, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D33758 llvm-svn: 307079
2017-07-04 22:35:06 +08:00
<< "#endif // ifdef GET_GLOBALISEL_TEMPORARIES_DECL\n\n";
OS << "#ifdef GET_GLOBALISEL_TEMPORARIES_INIT\n"
<< ", State(" << MaxTemporaries << "),\n"
Reapply r332907 "[GlobalISel] Improving InstructionSelect's performance by reducing MatchTable..." Apparently the compile time problem was caused by the fact that not all compilers / STL implementations can automatically convert std::unique_ptr<Derived> to std::unique_ptr<Base>. Fixed (hopefully) by making sure it's std::unique_ptr<Derived>&& (rvalue ref) to std::unique_ptr<Base> conversion instead. llvm-svn: 332917
2018-05-22 07:28:51 +08:00
<< "ISelInfo(TypeObjects, NumTypeObjects, FeatureBitsets"
<< ", ComplexPredicateFns, CustomRenderers)\n"
[globalisel][tablegen] Partially fix compile-time regressions by converting matcher to state-machine(s) Summary: Replace the matcher if-statements for each rule with a state-machine. This significantly reduces compile time, memory allocations, and cumulative memory allocation when compiling AArch64InstructionSelector.cpp.o after r303259 is recommitted. The following patches will expand on this further to fully fix the regressions. Reviewers: rovka, ab, t.p.northover, qcolombet, aditya_nandakumar Reviewed By: ab Subscribers: vitalybuka, aemerson, javed.absar, igorb, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D33758 llvm-svn: 307079
2017-07-04 22:35:06 +08:00
<< "#endif // ifdef GET_GLOBALISEL_TEMPORARIES_INIT\n\n";
[globalisel][tblgen] Add support for ComplexPatterns Summary: Adds a new kind of MachineOperand: MO_Placeholder. This operand must not appear in the MIR and only exists as a way of creating an 'uninitialized' operand until a matcher function overwrites it. Depends on D30046, D29712 Reviewers: t.p.northover, ab, rovka, aditya_nandakumar, javed.absar, qcolombet Reviewed By: qcolombet Subscribers: dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D30089 llvm-svn: 297782
2017-03-15 05:32:08 +08:00
[globalisel][tablegen] Import SelectionDAG's rule predicates and support the equivalent in GIRule. Summary: The SelectionDAG importer now imports rules with Predicate's attached via Requires, PredicateControl, etc. These predicates are implemented as bitset's to allow multiple predicates to be tested together. However, unlike the MC layer subtarget features, each target only pays for it's own predicates (e.g. AArch64 doesn't have 192 feature bits just because X86 needs a lot). Both AArch64 and X86 derive at least one predicate from the MachineFunction or Function so they must re-initialize AvailableFeatures before each function. They also declare locals in <Target>InstructionSelector so that computeAvailableFeatures() can use the code from SelectionDAG without modification. Reviewers: rovka, qcolombet, aditya_nandakumar, t.p.northover, ab Reviewed By: rovka Subscribers: aemerson, rengolin, dberris, kristof.beyls, llvm-commits, igorb Differential Revision: https://reviews.llvm.org/D31418 llvm-svn: 300993
2017-04-21 23:59:56 +08:00
OS << "#ifdef GET_GLOBALISEL_IMPL\n";
SubtargetFeatureInfo::emitSubtargetFeatureBitEnumeration(SubtargetFeatures,
OS);
[globalisel][tablegen] Compute available feature bits correctly. Summary: Predicate<> now has a field to indicate how often it must be recomputed. Currently, there are two frequencies, per-module (RecomputePerFunction==0) and per-function (RecomputePerFunction==1). Per-function predicates are currently recomputed more frequently than necessary since the only predicate in this category is cheap to test. Per-module predicates are now computed in getSubtargetImpl() while per-function predicates are computed in selectImpl(). Tablegen now manages the PredicateBitset internally. It should only be necessary to add the required includes. Also fixed a problem revealed by the test case where constrainSelectedInstRegOperands() would attempt to tie operands that BuildMI had already tied. Reviewers: ab, qcolombet, t.p.northover, rovka, aditya_nandakumar Reviewed By: rovka Subscribers: kristof.beyls, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D32491 llvm-svn: 301750
2017-04-30 01:30:09 +08:00
// Separate subtarget features by how often they must be recomputed.
SubtargetFeatureInfoMap ModuleFeatures;
std::copy_if(SubtargetFeatures.begin(), SubtargetFeatures.end(),
std::inserter(ModuleFeatures, ModuleFeatures.end()),
[](const SubtargetFeatureInfoMap::value_type &X) {
return !X.second.mustRecomputePerFunction();
});
SubtargetFeatureInfoMap FunctionFeatures;
std::copy_if(SubtargetFeatures.begin(), SubtargetFeatures.end(),
std::inserter(FunctionFeatures, FunctionFeatures.end()),
[](const SubtargetFeatureInfoMap::value_type &X) {
return X.second.mustRecomputePerFunction();
});
[globalisel][tablegen] Import SelectionDAG's rule predicates and support the equivalent in GIRule. Summary: The SelectionDAG importer now imports rules with Predicate's attached via Requires, PredicateControl, etc. These predicates are implemented as bitset's to allow multiple predicates to be tested together. However, unlike the MC layer subtarget features, each target only pays for it's own predicates (e.g. AArch64 doesn't have 192 feature bits just because X86 needs a lot). Both AArch64 and X86 derive at least one predicate from the MachineFunction or Function so they must re-initialize AvailableFeatures before each function. They also declare locals in <Target>InstructionSelector so that computeAvailableFeatures() can use the code from SelectionDAG without modification. Reviewers: rovka, qcolombet, aditya_nandakumar, t.p.northover, ab Reviewed By: rovka Subscribers: aemerson, rengolin, dberris, kristof.beyls, llvm-commits, igorb Differential Revision: https://reviews.llvm.org/D31418 llvm-svn: 300993
2017-04-21 23:59:56 +08:00
SubtargetFeatureInfo::emitComputeAvailableFeatures(
[globalisel][tablegen] Compute available feature bits correctly. Summary: Predicate<> now has a field to indicate how often it must be recomputed. Currently, there are two frequencies, per-module (RecomputePerFunction==0) and per-function (RecomputePerFunction==1). Per-function predicates are currently recomputed more frequently than necessary since the only predicate in this category is cheap to test. Per-module predicates are now computed in getSubtargetImpl() while per-function predicates are computed in selectImpl(). Tablegen now manages the PredicateBitset internally. It should only be necessary to add the required includes. Also fixed a problem revealed by the test case where constrainSelectedInstRegOperands() would attempt to tie operands that BuildMI had already tied. Reviewers: ab, qcolombet, t.p.northover, rovka, aditya_nandakumar Reviewed By: rovka Subscribers: kristof.beyls, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D32491 llvm-svn: 301750
2017-04-30 01:30:09 +08:00
Target.getName(), "InstructionSelector", "computeAvailableModuleFeatures",
ModuleFeatures, OS);
SubtargetFeatureInfo::emitComputeAvailableFeatures(
Target.getName(), "InstructionSelector",
"computeAvailableFunctionFeatures", FunctionFeatures, OS,
"const MachineFunction *MF");
[globalisel][tablegen] Import SelectionDAG's rule predicates and support the equivalent in GIRule. Summary: The SelectionDAG importer now imports rules with Predicate's attached via Requires, PredicateControl, etc. These predicates are implemented as bitset's to allow multiple predicates to be tested together. However, unlike the MC layer subtarget features, each target only pays for it's own predicates (e.g. AArch64 doesn't have 192 feature bits just because X86 needs a lot). Both AArch64 and X86 derive at least one predicate from the MachineFunction or Function so they must re-initialize AvailableFeatures before each function. They also declare locals in <Target>InstructionSelector so that computeAvailableFeatures() can use the code from SelectionDAG without modification. Reviewers: rovka, qcolombet, aditya_nandakumar, t.p.northover, ab Reviewed By: rovka Subscribers: aemerson, rengolin, dberris, kristof.beyls, llvm-commits, igorb Differential Revision: https://reviews.llvm.org/D31418 llvm-svn: 300993
2017-04-21 23:59:56 +08:00
[globalisel][tablegen] Partially fix compile-time regressions by converting matcher to state-machine(s) Summary: Replace the matcher if-statements for each rule with a state-machine. This significantly reduces compile time, memory allocations, and cumulative memory allocation when compiling AArch64InstructionSelector.cpp.o after r303259 is recommitted. The following patches will expand on this further to fully fix the regressions. Reviewers: rovka, ab, t.p.northover, qcolombet, aditya_nandakumar Reviewed By: ab Subscribers: vitalybuka, aemerson, javed.absar, igorb, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D33758 llvm-svn: 307079
2017-07-04 22:35:06 +08:00
// Emit a table containing the LLT objects needed by the matcher and an enum
// for the matcher to reference them with.
[globalisel][tablegen] Generate TypeObject table. NFC Summary: Generate the type table from the types used by a target rather than hard-coding the union of types used by all targets. Depends on D36084 Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Reviewed By: rovka Subscribers: kristof.beyls, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D36085 llvm-svn: 311084
2017-08-17 21:18:35 +08:00
std::vector<LLTCodeGen> TypeObjects;
[globalisel] Update GlobalISel emitter to match new representation of extending loads Summary: Previously, a extending load was represented at (G_*EXT (G_LOAD x)). This had a few drawbacks: * G_LOAD had to be legal for all sizes you could extend from, even if registers didn't naturally hold those sizes. * All sizes you could extend from had to be allocatable just in case the extend went missing (e.g. by optimization). * At minimum, G_*EXT and G_TRUNC had to be legal for these sizes. As we improve optimization of extends and truncates, this legality requirement would spread without considerable care w.r.t when certain combines were permitted. * The SelectionDAG importer required some ugly and fragile pattern rewriting to translate patterns into this style. This patch changes the representation to: * (G_[SZ]EXTLOAD x) * (G_LOAD x) any-extends when MMO.getSize() * 8 < ResultTy.getSizeInBits() which resolves these issues by allowing targets to work entirely in their native register sizes, and by having a more direct translation from SelectionDAG patterns. Each extending load can be lowered by the legalizer into separate extends and loads, however a target that supports s1 will need the any-extending load to extend to at least s8 since LLVM does not represent memory accesses smaller than 8 bit. The legalizer can widenScalar G_LOAD into an any-extending load but sign/zero-extending loads need help from something else like a combiner pass. A follow-up patch that adds combiner helpers for for this will follow. The new representation requires that the MMO correctly reflect the memory access so this has been corrected in a couple tests. I've also moved the extending loads to their own tests since they are (mostly) separate opcodes now. Additionally, the re-write appears to have invalidated two tests from select-with-no-legality-check.mir since the matcher table no longer contains loads that result in s1's and they aren't legal in AArch64 anymore. Depends on D45540 Reviewers: ab, aditya_nandakumar, bogner, rtereshin, volkan, rovka, javed.absar Reviewed By: rtereshin Subscribers: javed.absar, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D45541 llvm-svn: 331601
2018-05-06 04:53:24 +08:00
for (const auto &Ty : KnownTypes)
[globalisel][tablegen] Generate TypeObject table. NFC Summary: Generate the type table from the types used by a target rather than hard-coding the union of types used by all targets. Depends on D36084 Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Reviewed By: rovka Subscribers: kristof.beyls, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D36085 llvm-svn: 311084
2017-08-17 21:18:35 +08:00
TypeObjects.push_back(Ty);
[TableGen] Change std::sort to llvm::sort in response to r327219 Summary: r327219 added wrappers to std::sort which randomly shuffle the container before sorting. This will help in uncovering non-determinism caused due to undefined sorting order of objects having the same key. To make use of that infrastructure we need to invoke llvm::sort instead of std::sort. Note: This patch is one of a series of patches to replace *all* std::sort to llvm::sort. Refer the comments section in D44363 for a list of all the required patches. Reviewers: stoklund, kparzysz, dsanders Reviewed By: dsanders Subscribers: llvm-commits Differential Revision: https://reviews.llvm.org/D45144 llvm-svn: 329451
2018-04-07 04:18:05 +08:00
llvm::sort(TypeObjects.begin(), TypeObjects.end());
[globalisel][tablegen] Add tests for FeatureBitsets and ComplexPattern predicates. llvm-svn: 311542
2017-08-23 18:09:25 +08:00
OS << "// LLT Objects.\n"
<< "enum {\n";
[globalisel][tablegen] Partially fix compile-time regressions by converting matcher to state-machine(s) Summary: Replace the matcher if-statements for each rule with a state-machine. This significantly reduces compile time, memory allocations, and cumulative memory allocation when compiling AArch64InstructionSelector.cpp.o after r303259 is recommitted. The following patches will expand on this further to fully fix the regressions. Reviewers: rovka, ab, t.p.northover, qcolombet, aditya_nandakumar Reviewed By: ab Subscribers: vitalybuka, aemerson, javed.absar, igorb, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D33758 llvm-svn: 307079
2017-07-04 22:35:06 +08:00
for (const auto &TypeObject : TypeObjects) {
OS << " ";
TypeObject.emitCxxEnumValue(OS);
OS << ",\n";
}
Reapply r332907 "[GlobalISel] Improving InstructionSelect's performance by reducing MatchTable..." Apparently the compile time problem was caused by the fact that not all compilers / STL implementations can automatically convert std::unique_ptr<Derived> to std::unique_ptr<Base>. Fixed (hopefully) by making sure it's std::unique_ptr<Derived>&& (rvalue ref) to std::unique_ptr<Base> conversion instead. llvm-svn: 332917
2018-05-22 07:28:51 +08:00
OS << "};\n";
OS << "const static size_t NumTypeObjects = " << TypeObjects.size() << ";\n"
[globalisel][tablegen] Partially fix compile-time regressions by converting matcher to state-machine(s) Summary: Replace the matcher if-statements for each rule with a state-machine. This significantly reduces compile time, memory allocations, and cumulative memory allocation when compiling AArch64InstructionSelector.cpp.o after r303259 is recommitted. The following patches will expand on this further to fully fix the regressions. Reviewers: rovka, ab, t.p.northover, qcolombet, aditya_nandakumar Reviewed By: ab Subscribers: vitalybuka, aemerson, javed.absar, igorb, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D33758 llvm-svn: 307079
2017-07-04 22:35:06 +08:00
<< "const static LLT TypeObjects[] = {\n";
for (const auto &TypeObject : TypeObjects) {
OS << " ";
TypeObject.emitCxxConstructorCall(OS);
OS << ",\n";
}
OS << "};\n\n";
// Emit a table containing the PredicateBitsets objects needed by the matcher
// and an enum for the matcher to reference them with.
std::vector<std::vector<Record *>> FeatureBitsets;
for (auto &Rule : Rules)
FeatureBitsets.push_back(Rule.getRequiredFeatures());
[TableGen] Change std::sort to llvm::sort in response to r327219 Summary: r327219 added wrappers to std::sort which randomly shuffle the container before sorting. This will help in uncovering non-determinism caused due to undefined sorting order of objects having the same key. To make use of that infrastructure we need to invoke llvm::sort instead of std::sort. Note: This patch is one of a series of patches to replace *all* std::sort to llvm::sort. Refer the comments section in D44363 for a list of all the required patches. Reviewers: stoklund, kparzysz, dsanders Reviewed By: dsanders Subscribers: llvm-commits Differential Revision: https://reviews.llvm.org/D45144 llvm-svn: 329451
2018-04-07 04:18:05 +08:00
llvm::sort(
[globalisel][tablegen] Partially fix compile-time regressions by converting matcher to state-machine(s) Summary: Replace the matcher if-statements for each rule with a state-machine. This significantly reduces compile time, memory allocations, and cumulative memory allocation when compiling AArch64InstructionSelector.cpp.o after r303259 is recommitted. The following patches will expand on this further to fully fix the regressions. Reviewers: rovka, ab, t.p.northover, qcolombet, aditya_nandakumar Reviewed By: ab Subscribers: vitalybuka, aemerson, javed.absar, igorb, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D33758 llvm-svn: 307079
2017-07-04 22:35:06 +08:00
FeatureBitsets.begin(), FeatureBitsets.end(),
[&](const std::vector<Record *> &A, const std::vector<Record *> &B) {
if (A.size() < B.size())
return true;
if (A.size() > B.size())
return false;
for (const auto &Pair : zip(A, B)) {
if (std::get<0>(Pair)->getName() < std::get<1>(Pair)->getName())
return true;
if (std::get<0>(Pair)->getName() > std::get<1>(Pair)->getName())
return false;
}
return false;
});
FeatureBitsets.erase(
std::unique(FeatureBitsets.begin(), FeatureBitsets.end()),
FeatureBitsets.end());
[globalisel][tablegen] Add tests for FeatureBitsets and ComplexPattern predicates. llvm-svn: 311542
2017-08-23 18:09:25 +08:00
OS << "// Feature bitsets.\n"
<< "enum {\n"
[globalisel][tablegen] Partially fix compile-time regressions by converting matcher to state-machine(s) Summary: Replace the matcher if-statements for each rule with a state-machine. This significantly reduces compile time, memory allocations, and cumulative memory allocation when compiling AArch64InstructionSelector.cpp.o after r303259 is recommitted. The following patches will expand on this further to fully fix the regressions. Reviewers: rovka, ab, t.p.northover, qcolombet, aditya_nandakumar Reviewed By: ab Subscribers: vitalybuka, aemerson, javed.absar, igorb, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D33758 llvm-svn: 307079
2017-07-04 22:35:06 +08:00
<< " GIFBS_Invalid,\n";
for (const auto &FeatureBitset : FeatureBitsets) {
if (FeatureBitset.empty())
continue;
OS << " " << getNameForFeatureBitset(FeatureBitset) << ",\n";
}
OS << "};\n"
<< "const static PredicateBitset FeatureBitsets[] {\n"
<< " {}, // GIFBS_Invalid\n";
for (const auto &FeatureBitset : FeatureBitsets) {
if (FeatureBitset.empty())
continue;
OS << " {";
for (const auto &Feature : FeatureBitset) {
const auto &I = SubtargetFeatures.find(Feature);
assert(I != SubtargetFeatures.end() && "Didn't import predicate?");
OS << I->second.getEnumBitName() << ", ";
}
OS << "},\n";
}
OS << "};\n\n";
// Emit complex predicate table and an enum to reference them with.
[globalisel][tablegen] Add tests for FeatureBitsets and ComplexPattern predicates. llvm-svn: 311542
2017-08-23 18:09:25 +08:00
OS << "// ComplexPattern predicates.\n"
<< "enum {\n"
[globalisel][tablegen] Partially fix compile-time regressions by converting matcher to state-machine(s) Summary: Replace the matcher if-statements for each rule with a state-machine. This significantly reduces compile time, memory allocations, and cumulative memory allocation when compiling AArch64InstructionSelector.cpp.o after r303259 is recommitted. The following patches will expand on this further to fully fix the regressions. Reviewers: rovka, ab, t.p.northover, qcolombet, aditya_nandakumar Reviewed By: ab Subscribers: vitalybuka, aemerson, javed.absar, igorb, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D33758 llvm-svn: 307079
2017-07-04 22:35:06 +08:00
<< " GICP_Invalid,\n";
for (const auto &Record : ComplexPredicates)
OS << " GICP_" << Record->getName() << ",\n";
OS << "};\n"
<< "// See constructor for table contents\n\n";
[globalisel][tablegen] Add support for fpimm and import of APInt/APFloat based ImmLeaf. Summary: There's only a tablegen testcase for IntImmLeaf and not a CodeGen one because the relevant rules are rejected for other reasons at the moment. On AArch64, it's because there's an SDNodeXForm attached to the operand. On X86, it's because the rule either emits multiple instructions or has another predicate using PatFrag which cannot easily be supported at the same time. Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Reviewed By: qcolombet Subscribers: aemerson, javed.absar, igorb, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D36569 llvm-svn: 315761
2017-10-14 05:28:03 +08:00
emitImmPredicates(OS, "I64", "int64_t", [](const Record *R) {
[aarch64] Support APInt and APFloat in ImmLeaf subclasses and make AArch64 use them. Summary: The purpose of this patch is to expose more information about ImmLeaf-like PatLeaf's so that GlobalISel can learn to import them. Previously, ImmLeaf could only be used to test int64_t's produced by sign-extending an APInt. Other tests on immediates had to use the generic PatLeaf and extract the constant using C++. With this patch, tablegen will know how to generate predicates for APInt, and APFloat. This will allow it to 'do the right thing' for both SelectionDAG and GlobalISel which require different methods of extracting the immediate from the IR. This is NFC for SelectionDAG since the new code is equivalent to the previous code. It's also NFC for FastISel because FastIselShouldIgnore is 1 for the ImmLeaf subclasses. Enabling FastIselShouldIgnore == 0 for these new subclasses will require a significant re-factor of FastISel. For GlobalISel, it's currently NFC because the relevant code to import the affected rules is not yet present. This will be added in a later patch. Depends on D36086 Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Reviewed By: qcolombet Subscribers: bjope, aemerson, rengolin, javed.absar, igorb, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D36534 llvm-svn: 315747
2017-10-14 04:42:18 +08:00
bool Unset;
return !R->getValueAsBitOrUnset("IsAPFloat", Unset) &&
!R->getValueAsBit("IsAPInt");
});
[globalisel][tablegen] Add support for fpimm and import of APInt/APFloat based ImmLeaf. Summary: There's only a tablegen testcase for IntImmLeaf and not a CodeGen one because the relevant rules are rejected for other reasons at the moment. On AArch64, it's because there's an SDNodeXForm attached to the operand. On X86, it's because the rule either emits multiple instructions or has another predicate using PatFrag which cannot easily be supported at the same time. Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Reviewed By: qcolombet Subscribers: aemerson, javed.absar, igorb, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D36569 llvm-svn: 315761
2017-10-14 05:28:03 +08:00
emitImmPredicates(OS, "APFloat", "const APFloat &", [](const Record *R) {
bool Unset;
return R->getValueAsBitOrUnset("IsAPFloat", Unset);
});
emitImmPredicates(OS, "APInt", "const APInt &", [](const Record *R) {
return R->getValueAsBit("IsAPInt");
});
Re-commit r315885: [globalisel][tblgen] Add support for iPTR and implement am_unscaled* and am_indexed* Summary: iPTR is a pointer of subtarget-specific size to any address space. Therefore type checks on this size derive the SizeInBits from a subtarget hook. At this point, we can import the simplests G_LOAD rules and select load instructions using them. Further patches will support for the predicates to enable additional loads as well as the stores. The previous commit failed on MSVC due to a failure to convert an initializer_list to a std::vector. Hopefully, MSVC will accept this version. Depends on D37457 Reviewers: ab, qcolombet, t.p.northover, rovka, aditya_nandakumar Reviewed By: qcolombet Subscribers: kristof.beyls, javed.absar, llvm-commits, igorb Differential Revision: https://reviews.llvm.org/D37458 llvm-svn: 315887
2017-10-16 11:36:29 +08:00
OS << "\n";
OS << Target.getName() << "InstructionSelector::ComplexMatcherMemFn\n"
<< Target.getName() << "InstructionSelector::ComplexPredicateFns[] = {\n"
<< " nullptr, // GICP_Invalid\n";
for (const auto &Record : ComplexPredicates)
OS << " &" << Target.getName()
<< "InstructionSelector::" << Record->getValueAsString("MatcherFn")
<< ", // " << Record->getName() << "\n";
OS << "};\n\n";
Re-commit: [globalisel][tablegen] Add support for ImmLeaf without SDNodeXForm Summary: This patch adds support for predicates on imm nodes but only for ImmLeaf and not for PatLeaf or PatFrag and only where the value does not need to be transformed before being rendered into the instruction. The limitation on PatLeaf/PatFrag/SDNodeXForm is due to differences in the necessary target-supplied C++ for GlobalISel. Depends on D36085 The previous commit was reverted for breaking the build but this appears to have been the recurring problem on the Windows bots with tablegen not being re-run when llvm-tblgen is changed but the .td's aren't. If it re-occurs then forcing a build with clean=True should fix it but this string should do this in advance: Requires a clean build. Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Reviewed By: rovka Subscribers: kristof.beyls, javed.absar, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D36086 llvm-svn: 311645
2017-08-24 17:11:20 +08:00
[GlobalISel][TableGen] Add support for SDNodeXForm Summary: This patch adds CustomRenderer which renders the matched operands to the specified instruction. Targets can enable the matching of SDNodeXForm by adding a definition that inherits from GICustomOperandRenderer and GISDNodeXFormEquiv as follows. def gi_imm8 : GICustomOperandRenderer<"renderImm8”>, GISDNodeXFormEquiv<imm8_xform>; Custom renderer functions should be of the form: void render(MachineInstrBuilder &MIB, const MachineInstr &I); Reviewers: dsanders, ab, rovka Reviewed By: dsanders Subscribers: kristof.beyls, javed.absar, llvm-commits, mgrang, qcolombet Differential Revision: https://reviews.llvm.org/D42012 llvm-svn: 322582
2018-01-17 02:44:05 +08:00
OS << "// Custom renderers.\n"
<< "enum {\n"
<< " GICR_Invalid,\n";
for (const auto &Record : CustomRendererFns)
OS << " GICR_" << Record->getValueAsString("RendererFn") << ", \n";
OS << "};\n";
OS << Target.getName() << "InstructionSelector::CustomRendererFn\n"
<< Target.getName() << "InstructionSelector::CustomRenderers[] = {\n"
<< " nullptr, // GICP_Invalid\n";
for (const auto &Record : CustomRendererFns)
OS << " &" << Target.getName()
<< "InstructionSelector::" << Record->getValueAsString("RendererFn")
<< ", // " << Record->getName() << "\n";
OS << "};\n\n";
[TableGen][GlobalISel] Optimize MatchTable for faster instruction selection *** Context *** Prior to this patchw, the table generated for matching instruction was straight forward but highly inefficient. Basically, each pattern generates its own set of self contained checks and actions. E.g., TableGen generated: // First pattern CheckNumOperand 3 CheckOpcode G_ADD ... Build ADDrr // Second pattern CheckNumOperand 3 CheckOpcode G_ADD ... Build ADDri // Third pattern CheckNumOperand 3 CheckOpcode G_SUB ... Build SUBrr *** Problem *** Because of that generation, a *lot* of check were redundant between each pattern and were checked every single time until we reach the pattern that matches. E.g., Taking the previous table, let say we are matching a G_SUB, that means we were going to check all the rules for G_ADD before looking at the G_SUB rule. In particular we are going to do: check 3 operands; PASS check G_ADD; FAIL ; Next rule check 3 operands; PASS (but we already knew that!) check G_ADD; FAIL (well it is still not true) ; Next rule check 3 operands; PASS (really!!) check G_SUB; PASS (at last :P) *** Proposed Solution *** This patch introduces a concept of group of rules (GroupMatcher) that share some predicates and only get checked once for the whole group. This patch only creates groups with one nesting level. Conceptually there is nothing preventing us for having deeper nest level. However, the current implementation is not smart enough to share the recording (aka capturing) of values. That limits its ability to do more sharing. For the given example the current patch will generate: // First group CheckOpcode G_ADD // First pattern CheckNumOperand 3 ... Build ADDrr // Second pattern CheckNumOperand 3 ... Build ADDri // Second group CheckOpcode G_SUB // Third pattern CheckNumOperand 3 ... Build SUBrr But if we allowed several nesting level, it could create a sub group for the checknumoperand 3. (We would need to call optimizeRules on the rules within a group.) *** Result *** With only one level of nesting, the instruction selection pass is up to 4x faster. For instance, one instruction now takes 500 checks, instead of 24k! With more nesting we could get in the tens I believe. Differential Revision: https://reviews.llvm.org/D39034 rdar://problem/34670699 llvm-svn: 321017
2017-12-19 03:47:41 +08:00
std::stable_sort(Rules.begin(), Rules.end(), [&](const RuleMatcher &A,
const RuleMatcher &B) {
[GISel]: Make GlobalISelEmitter rule prioritization compatible with selectionDAG This patch changes GlobalISelEmitter to rank patterns similar to how the DAG does it (ie it computes a score for a pattern and adds the added complexity to it). This is so that the decision tree for GISelSelector remains compatible with that of SelectionDAG. https://reviews.llvm.org/D43270 llvm-svn: 325401
2018-02-17 06:37:15 +08:00
int ScoreA = RuleMatcherScores[A.getRuleID()];
int ScoreB = RuleMatcherScores[B.getRuleID()];
if (ScoreA > ScoreB)
return true;
if (ScoreB > ScoreA)
return false;
[TableGen][GlobalISel] Optimize MatchTable for faster instruction selection *** Context *** Prior to this patchw, the table generated for matching instruction was straight forward but highly inefficient. Basically, each pattern generates its own set of self contained checks and actions. E.g., TableGen generated: // First pattern CheckNumOperand 3 CheckOpcode G_ADD ... Build ADDrr // Second pattern CheckNumOperand 3 CheckOpcode G_ADD ... Build ADDri // Third pattern CheckNumOperand 3 CheckOpcode G_SUB ... Build SUBrr *** Problem *** Because of that generation, a *lot* of check were redundant between each pattern and were checked every single time until we reach the pattern that matches. E.g., Taking the previous table, let say we are matching a G_SUB, that means we were going to check all the rules for G_ADD before looking at the G_SUB rule. In particular we are going to do: check 3 operands; PASS check G_ADD; FAIL ; Next rule check 3 operands; PASS (but we already knew that!) check G_ADD; FAIL (well it is still not true) ; Next rule check 3 operands; PASS (really!!) check G_SUB; PASS (at last :P) *** Proposed Solution *** This patch introduces a concept of group of rules (GroupMatcher) that share some predicates and only get checked once for the whole group. This patch only creates groups with one nesting level. Conceptually there is nothing preventing us for having deeper nest level. However, the current implementation is not smart enough to share the recording (aka capturing) of values. That limits its ability to do more sharing. For the given example the current patch will generate: // First group CheckOpcode G_ADD // First pattern CheckNumOperand 3 ... Build ADDrr // Second pattern CheckNumOperand 3 ... Build ADDri // Second group CheckOpcode G_SUB // Third pattern CheckNumOperand 3 ... Build SUBrr But if we allowed several nesting level, it could create a sub group for the checknumoperand 3. (We would need to call optimizeRules on the rules within a group.) *** Result *** With only one level of nesting, the instruction selection pass is up to 4x faster. For instance, one instruction now takes 500 checks, instead of 24k! With more nesting we could get in the tens I believe. Differential Revision: https://reviews.llvm.org/D39034 rdar://problem/34670699 llvm-svn: 321017
2017-12-19 03:47:41 +08:00
if (A.isHigherPriorityThan(B)) {
assert(!B.isHigherPriorityThan(A) && "Cannot be more important "
"and less important at "
"the same time");
return true;
}
return false;
});
[GlobalISel][InstructionSelect] Refactoring out a getMatchTable virtual method + other small NFC's The main goal is to share getMatchTable between the Instruction Selector and the Testgen. The commit also contains some NFC only loosely related to refactoring out the getMatchTable, but strongly related to the initial Testgen patch (see https://reviews.llvm.org/D43962) Reviewers: dsanders, aemerson Reviewed By: dsanders Subscribers: rovka, kristof.beyls, llvm-commits, dsanders Differential Revision: https://reviews.llvm.org/D46096 llvm-svn: 331395
2018-05-03 04:07:15 +08:00
OS << "bool " << Target.getName()
<< "InstructionSelector::selectImpl(MachineInstr &I, CodeGenCoverage "
"&CoverageInfo) const {\n"
<< " MachineFunction &MF = *I.getParent()->getParent();\n"
<< " MachineRegisterInfo &MRI = MF.getRegInfo();\n"
<< " // FIXME: This should be computed on a per-function basis rather "
"than per-insn.\n"
<< " AvailableFunctionFeatures = computeAvailableFunctionFeatures(&STI, "
"&MF);\n"
<< " const PredicateBitset AvailableFeatures = getAvailableFeatures();\n"
<< " NewMIVector OutMIs;\n"
<< " State.MIs.clear();\n"
<< " State.MIs.push_back(&I);\n\n"
<< " if (executeMatchTable(*this, OutMIs, State, ISelInfo"
<< ", getMatchTable(), TII, MRI, TRI, RBI, AvailableFeatures"
<< ", CoverageInfo)) {\n"
Re-commit: r309094 [globalisel][tablegen] Fuse the generated tables together. Summary: Now that we have control flow in place, fuse the per-rule tables into a single table. This is a compile-time saving at this point. However, this will also enable the optimization of a table so that similar instructions can be tested together, reducing the time spent on the matching the code. This is NFC in terms of externally visible behaviour but some internals have changed slightly. State.MIs is no longer reset between each rule that is attempted because it's not necessary to do so. As a consequence of this the restriction on the order that instructions are added to State.MIs has been relaxed to only affect recorded instructions that require new elements to be added to the vector. GIM_RecordInsn can now write to any element from 1 to State.MIs.size() instead of just State.MIs.size(). The compile-time regressions from the last commit were caused by the ARM target including a non-const variable (zero_reg) in the table and therefore generating an initializer for it. That variable is now const. Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar Reviewed By: rovka Subscribers: kristof.beyls, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D35681 llvm-svn: 309264
2017-07-27 19:03:45 +08:00
<< " return true;\n"
[GlobalISel][InstructionSelect] Refactoring out a getMatchTable virtual method + other small NFC's The main goal is to share getMatchTable between the Instruction Selector and the Testgen. The commit also contains some NFC only loosely related to refactoring out the getMatchTable, but strongly related to the initial Testgen patch (see https://reviews.llvm.org/D43962) Reviewers: dsanders, aemerson Reviewed By: dsanders Subscribers: rovka, kristof.beyls, llvm-commits, dsanders Differential Revision: https://reviews.llvm.org/D46096 llvm-svn: 331395
2018-05-03 04:07:15 +08:00
<< " }\n\n"
<< " return false;\n"
<< "}\n\n";
[GlobalISel] Add basic Selector-emitter tblgen backend. This adds a basic tablegen backend that analyzes the SelectionDAG patterns to find simple ones that are eligible for GlobalISel-emission. That's similar to FastISel, with one notable difference: we're not fed ISD opcodes, so we need to map the SDNode operators to generic opcodes. That's done using GINodeEquiv in TargetGlobalISel.td. Otherwise, this is mostly boilerplate, and lots of filtering of any kind of "complicated" pattern. On AArch64, this is sufficient to match G_ADD up to s64 (to ADDWrr/ADDXrr) and G_BR (to B). Differential Revision: https://reviews.llvm.org/D26878 llvm-svn: 290284
2016-12-22 07:26:20 +08:00
[GlobalISel][InstructionSelect] Making Coverage Info generation optional on per-match table basis to make sure that Testgen always has access to coverage info even if the match table used by the selector itself is stripped off that information for performance reasons. Reviewers: dsanders, aemerson Reviewed By: dsanders Subscribers: rovka, kristof.beyls, llvm-commits, dsanders Differential Revision: https://reviews.llvm.org/D46098 llvm-svn: 331398
2018-05-03 04:15:11 +08:00
const MatchTable Table =
buildMatchTable(Rules, OptimizeMatchTable, GenerateCoverage);
[GlobalISel][InstructionSelect] Refactoring out a getMatchTable virtual method + other small NFC's The main goal is to share getMatchTable between the Instruction Selector and the Testgen. The commit also contains some NFC only loosely related to refactoring out the getMatchTable, but strongly related to the initial Testgen patch (see https://reviews.llvm.org/D43962) Reviewers: dsanders, aemerson Reviewed By: dsanders Subscribers: rovka, kristof.beyls, llvm-commits, dsanders Differential Revision: https://reviews.llvm.org/D46096 llvm-svn: 331395
2018-05-03 04:07:15 +08:00
OS << "const int64_t *" << Target.getName()
<< "InstructionSelector::getMatchTable() const {\n";
Table.emitDeclaration(OS);
OS << " return ";
Table.emitUse(OS);
OS << ";\n}\n";
OS << "#endif // ifdef GET_GLOBALISEL_IMPL\n";
[globalisel][tablegen] Compute available feature bits correctly. Summary: Predicate<> now has a field to indicate how often it must be recomputed. Currently, there are two frequencies, per-module (RecomputePerFunction==0) and per-function (RecomputePerFunction==1). Per-function predicates are currently recomputed more frequently than necessary since the only predicate in this category is cheap to test. Per-module predicates are now computed in getSubtargetImpl() while per-function predicates are computed in selectImpl(). Tablegen now manages the PredicateBitset internally. It should only be necessary to add the required includes. Also fixed a problem revealed by the test case where constrainSelectedInstRegOperands() would attempt to tie operands that BuildMI had already tied. Reviewers: ab, qcolombet, t.p.northover, rovka, aditya_nandakumar Reviewed By: rovka Subscribers: kristof.beyls, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D32491 llvm-svn: 301750
2017-04-30 01:30:09 +08:00
OS << "#ifdef GET_GLOBALISEL_PREDICATES_DECL\n"
<< "PredicateBitset AvailableModuleFeatures;\n"
<< "mutable PredicateBitset AvailableFunctionFeatures;\n"
<< "PredicateBitset getAvailableFeatures() const {\n"
<< " return AvailableModuleFeatures | AvailableFunctionFeatures;\n"
<< "}\n"
<< "PredicateBitset\n"
<< "computeAvailableModuleFeatures(const " << Target.getName()
<< "Subtarget *Subtarget) const;\n"
<< "PredicateBitset\n"
<< "computeAvailableFunctionFeatures(const " << Target.getName()
<< "Subtarget *Subtarget,\n"
<< " const MachineFunction *MF) const;\n"
<< "#endif // ifdef GET_GLOBALISEL_PREDICATES_DECL\n";
OS << "#ifdef GET_GLOBALISEL_PREDICATES_INIT\n"
<< "AvailableModuleFeatures(computeAvailableModuleFeatures(&STI)),\n"
<< "AvailableFunctionFeatures()\n"
<< "#endif // ifdef GET_GLOBALISEL_PREDICATES_INIT\n";
[GlobalISel] Add basic Selector-emitter tblgen backend. This adds a basic tablegen backend that analyzes the SelectionDAG patterns to find simple ones that are eligible for GlobalISel-emission. That's similar to FastISel, with one notable difference: we're not fed ISD opcodes, so we need to map the SDNode operators to generic opcodes. That's done using GINodeEquiv in TargetGlobalISel.td. Otherwise, this is mostly boilerplate, and lots of filtering of any kind of "complicated" pattern. On AArch64, this is sufficient to match G_ADD up to s64 (to ADDWrr/ADDXrr) and G_BR (to B). Differential Revision: https://reviews.llvm.org/D26878 llvm-svn: 290284
2016-12-22 07:26:20 +08:00
}
[globalisel][tablegen] Import SelectionDAG's rule predicates and support the equivalent in GIRule. Summary: The SelectionDAG importer now imports rules with Predicate's attached via Requires, PredicateControl, etc. These predicates are implemented as bitset's to allow multiple predicates to be tested together. However, unlike the MC layer subtarget features, each target only pays for it's own predicates (e.g. AArch64 doesn't have 192 feature bits just because X86 needs a lot). Both AArch64 and X86 derive at least one predicate from the MachineFunction or Function so they must re-initialize AvailableFeatures before each function. They also declare locals in <Target>InstructionSelector so that computeAvailableFeatures() can use the code from SelectionDAG without modification. Reviewers: rovka, qcolombet, aditya_nandakumar, t.p.northover, ab Reviewed By: rovka Subscribers: aemerson, rengolin, dberris, kristof.beyls, llvm-commits, igorb Differential Revision: https://reviews.llvm.org/D31418 llvm-svn: 300993
2017-04-21 23:59:56 +08:00
void GlobalISelEmitter::declareSubtargetFeature(Record *Predicate) {
if (SubtargetFeatures.count(Predicate) == 0)
SubtargetFeatures.emplace(
Predicate, SubtargetFeatureInfo(Predicate, SubtargetFeatures.size()));
}
Reapply r332907 "[GlobalISel] Improving InstructionSelect's performance by reducing MatchTable..." Apparently the compile time problem was caused by the fact that not all compilers / STL implementations can automatically convert std::unique_ptr<Derived> to std::unique_ptr<Base>. Fixed (hopefully) by making sure it's std::unique_ptr<Derived>&& (rvalue ref) to std::unique_ptr<Base> conversion instead. llvm-svn: 332917
2018-05-22 07:28:51 +08:00
void RuleMatcher::optimize() {
for (auto &Item : InsnVariableIDs) {
InstructionMatcher &InsnMatcher = *Item.first;
for (auto &OM : InsnMatcher.operands()) {
// Register Banks checks rarely fail, but often crash as targets usually
// provide only partially defined RegisterBankInfo::getRegBankFromRegClass
// method. Often the problem is hidden as non-optimized MatchTable checks
// banks rather late, most notably after checking target / function /
// module features and a few opcodes. That makes these checks a)
// beneficial to delay until the very end (we don't want to perform a lot
// of checks that all pass and then fail at the very end) b) not safe to
// have as early checks.
for (auto &OP : OM->predicates())
if (isa<RegisterBankOperandMatcher>(OP) ||
isa<ComplexPatternOperandMatcher>(OP))
EpilogueMatchers.emplace_back(std::move(OP));
OM->eraseNullPredicates();
}
InsnMatcher.optimize();
}
llvm::sort(
EpilogueMatchers.begin(), EpilogueMatchers.end(),
[](const std::unique_ptr<PredicateMatcher> &L,
const std::unique_ptr<PredicateMatcher> &R) {
return std::make_tuple(L->getKind(), L->getInsnVarID(), L->getOpIdx()) <
std::make_tuple(R->getKind(), R->getInsnVarID(), R->getOpIdx());
});
}
bool RuleMatcher::hasFirstCondition() const {
if (insnmatchers_empty())
return false;
InstructionMatcher &Matcher = insnmatchers_front();
if (!Matcher.predicates_empty())
return true;
for (auto &OM : Matcher.operands())
for (auto &OP : OM->predicates())
if (!isa<InstructionOperandMatcher>(OP))
return true;
return false;
}
const PredicateMatcher &RuleMatcher::getFirstCondition() const {
[TableGen][GlobalISel] Optimize MatchTable for faster instruction selection *** Context *** Prior to this patchw, the table generated for matching instruction was straight forward but highly inefficient. Basically, each pattern generates its own set of self contained checks and actions. E.g., TableGen generated: // First pattern CheckNumOperand 3 CheckOpcode G_ADD ... Build ADDrr // Second pattern CheckNumOperand 3 CheckOpcode G_ADD ... Build ADDri // Third pattern CheckNumOperand 3 CheckOpcode G_SUB ... Build SUBrr *** Problem *** Because of that generation, a *lot* of check were redundant between each pattern and were checked every single time until we reach the pattern that matches. E.g., Taking the previous table, let say we are matching a G_SUB, that means we were going to check all the rules for G_ADD before looking at the G_SUB rule. In particular we are going to do: check 3 operands; PASS check G_ADD; FAIL ; Next rule check 3 operands; PASS (but we already knew that!) check G_ADD; FAIL (well it is still not true) ; Next rule check 3 operands; PASS (really!!) check G_SUB; PASS (at last :P) *** Proposed Solution *** This patch introduces a concept of group of rules (GroupMatcher) that share some predicates and only get checked once for the whole group. This patch only creates groups with one nesting level. Conceptually there is nothing preventing us for having deeper nest level. However, the current implementation is not smart enough to share the recording (aka capturing) of values. That limits its ability to do more sharing. For the given example the current patch will generate: // First group CheckOpcode G_ADD // First pattern CheckNumOperand 3 ... Build ADDrr // Second pattern CheckNumOperand 3 ... Build ADDri // Second group CheckOpcode G_SUB // Third pattern CheckNumOperand 3 ... Build SUBrr But if we allowed several nesting level, it could create a sub group for the checknumoperand 3. (We would need to call optimizeRules on the rules within a group.) *** Result *** With only one level of nesting, the instruction selection pass is up to 4x faster. For instance, one instruction now takes 500 checks, instead of 24k! With more nesting we could get in the tens I believe. Differential Revision: https://reviews.llvm.org/D39034 rdar://problem/34670699 llvm-svn: 321017
2017-12-19 03:47:41 +08:00
assert(!insnmatchers_empty() &&
Reapply r332907 "[GlobalISel] Improving InstructionSelect's performance by reducing MatchTable..." Apparently the compile time problem was caused by the fact that not all compilers / STL implementations can automatically convert std::unique_ptr<Derived> to std::unique_ptr<Base>. Fixed (hopefully) by making sure it's std::unique_ptr<Derived>&& (rvalue ref) to std::unique_ptr<Base> conversion instead. llvm-svn: 332917
2018-05-22 07:28:51 +08:00
"Trying to get a condition from an empty RuleMatcher");
[TableGen][GlobalISel] Optimize MatchTable for faster instruction selection *** Context *** Prior to this patchw, the table generated for matching instruction was straight forward but highly inefficient. Basically, each pattern generates its own set of self contained checks and actions. E.g., TableGen generated: // First pattern CheckNumOperand 3 CheckOpcode G_ADD ... Build ADDrr // Second pattern CheckNumOperand 3 CheckOpcode G_ADD ... Build ADDri // Third pattern CheckNumOperand 3 CheckOpcode G_SUB ... Build SUBrr *** Problem *** Because of that generation, a *lot* of check were redundant between each pattern and were checked every single time until we reach the pattern that matches. E.g., Taking the previous table, let say we are matching a G_SUB, that means we were going to check all the rules for G_ADD before looking at the G_SUB rule. In particular we are going to do: check 3 operands; PASS check G_ADD; FAIL ; Next rule check 3 operands; PASS (but we already knew that!) check G_ADD; FAIL (well it is still not true) ; Next rule check 3 operands; PASS (really!!) check G_SUB; PASS (at last :P) *** Proposed Solution *** This patch introduces a concept of group of rules (GroupMatcher) that share some predicates and only get checked once for the whole group. This patch only creates groups with one nesting level. Conceptually there is nothing preventing us for having deeper nest level. However, the current implementation is not smart enough to share the recording (aka capturing) of values. That limits its ability to do more sharing. For the given example the current patch will generate: // First group CheckOpcode G_ADD // First pattern CheckNumOperand 3 ... Build ADDrr // Second pattern CheckNumOperand 3 ... Build ADDri // Second group CheckOpcode G_SUB // Third pattern CheckNumOperand 3 ... Build SUBrr But if we allowed several nesting level, it could create a sub group for the checknumoperand 3. (We would need to call optimizeRules on the rules within a group.) *** Result *** With only one level of nesting, the instruction selection pass is up to 4x faster. For instance, one instruction now takes 500 checks, instead of 24k! With more nesting we could get in the tens I believe. Differential Revision: https://reviews.llvm.org/D39034 rdar://problem/34670699 llvm-svn: 321017
2017-12-19 03:47:41 +08:00
InstructionMatcher &Matcher = insnmatchers_front();
if (!Matcher.predicates_empty())
Reapply r332907 "[GlobalISel] Improving InstructionSelect's performance by reducing MatchTable..." Apparently the compile time problem was caused by the fact that not all compilers / STL implementations can automatically convert std::unique_ptr<Derived> to std::unique_ptr<Base>. Fixed (hopefully) by making sure it's std::unique_ptr<Derived>&& (rvalue ref) to std::unique_ptr<Base> conversion instead. llvm-svn: 332917
2018-05-22 07:28:51 +08:00
return **Matcher.predicates_begin();
// If there is no more predicate on the instruction itself, look at its
// operands.
for (auto &OM : Matcher.operands())
for (auto &OP : OM->predicates())
if (!isa<InstructionOperandMatcher>(OP))
return *OP;
llvm_unreachable("Trying to get a condition from an InstructionMatcher with "
"no conditions");
[TableGen][GlobalISel] Optimize MatchTable for faster instruction selection *** Context *** Prior to this patchw, the table generated for matching instruction was straight forward but highly inefficient. Basically, each pattern generates its own set of self contained checks and actions. E.g., TableGen generated: // First pattern CheckNumOperand 3 CheckOpcode G_ADD ... Build ADDrr // Second pattern CheckNumOperand 3 CheckOpcode G_ADD ... Build ADDri // Third pattern CheckNumOperand 3 CheckOpcode G_SUB ... Build SUBrr *** Problem *** Because of that generation, a *lot* of check were redundant between each pattern and were checked every single time until we reach the pattern that matches. E.g., Taking the previous table, let say we are matching a G_SUB, that means we were going to check all the rules for G_ADD before looking at the G_SUB rule. In particular we are going to do: check 3 operands; PASS check G_ADD; FAIL ; Next rule check 3 operands; PASS (but we already knew that!) check G_ADD; FAIL (well it is still not true) ; Next rule check 3 operands; PASS (really!!) check G_SUB; PASS (at last :P) *** Proposed Solution *** This patch introduces a concept of group of rules (GroupMatcher) that share some predicates and only get checked once for the whole group. This patch only creates groups with one nesting level. Conceptually there is nothing preventing us for having deeper nest level. However, the current implementation is not smart enough to share the recording (aka capturing) of values. That limits its ability to do more sharing. For the given example the current patch will generate: // First group CheckOpcode G_ADD // First pattern CheckNumOperand 3 ... Build ADDrr // Second pattern CheckNumOperand 3 ... Build ADDri // Second group CheckOpcode G_SUB // Third pattern CheckNumOperand 3 ... Build SUBrr But if we allowed several nesting level, it could create a sub group for the checknumoperand 3. (We would need to call optimizeRules on the rules within a group.) *** Result *** With only one level of nesting, the instruction selection pass is up to 4x faster. For instance, one instruction now takes 500 checks, instead of 24k! With more nesting we could get in the tens I believe. Differential Revision: https://reviews.llvm.org/D39034 rdar://problem/34670699 llvm-svn: 321017
2017-12-19 03:47:41 +08:00
}
Reapply r332907 "[GlobalISel] Improving InstructionSelect's performance by reducing MatchTable..." Apparently the compile time problem was caused by the fact that not all compilers / STL implementations can automatically convert std::unique_ptr<Derived> to std::unique_ptr<Base>. Fixed (hopefully) by making sure it's std::unique_ptr<Derived>&& (rvalue ref) to std::unique_ptr<Base> conversion instead. llvm-svn: 332917
2018-05-22 07:28:51 +08:00
std::unique_ptr<PredicateMatcher> RuleMatcher::popFirstCondition() {
assert(!insnmatchers_empty() &&
"Trying to pop a condition from an empty RuleMatcher");
InstructionMatcher &Matcher = insnmatchers_front();
if (!Matcher.predicates_empty())
return Matcher.predicates_pop_front();
// If there is no more predicate on the instruction itself, look at its
// operands.
for (auto &OM : Matcher.operands())
for (auto &OP : OM->predicates())
if (!isa<InstructionOperandMatcher>(OP)) {
std::unique_ptr<PredicateMatcher> Result = std::move(OP);
OM->eraseNullPredicates();
return Result;
}
llvm_unreachable("Trying to pop a condition from an InstructionMatcher with "
"no conditions");
}
bool GroupMatcher::candidateConditionMatches(
[TableGen][GlobalISel] Optimize MatchTable for faster instruction selection *** Context *** Prior to this patchw, the table generated for matching instruction was straight forward but highly inefficient. Basically, each pattern generates its own set of self contained checks and actions. E.g., TableGen generated: // First pattern CheckNumOperand 3 CheckOpcode G_ADD ... Build ADDrr // Second pattern CheckNumOperand 3 CheckOpcode G_ADD ... Build ADDri // Third pattern CheckNumOperand 3 CheckOpcode G_SUB ... Build SUBrr *** Problem *** Because of that generation, a *lot* of check were redundant between each pattern and were checked every single time until we reach the pattern that matches. E.g., Taking the previous table, let say we are matching a G_SUB, that means we were going to check all the rules for G_ADD before looking at the G_SUB rule. In particular we are going to do: check 3 operands; PASS check G_ADD; FAIL ; Next rule check 3 operands; PASS (but we already knew that!) check G_ADD; FAIL (well it is still not true) ; Next rule check 3 operands; PASS (really!!) check G_SUB; PASS (at last :P) *** Proposed Solution *** This patch introduces a concept of group of rules (GroupMatcher) that share some predicates and only get checked once for the whole group. This patch only creates groups with one nesting level. Conceptually there is nothing preventing us for having deeper nest level. However, the current implementation is not smart enough to share the recording (aka capturing) of values. That limits its ability to do more sharing. For the given example the current patch will generate: // First group CheckOpcode G_ADD // First pattern CheckNumOperand 3 ... Build ADDrr // Second pattern CheckNumOperand 3 ... Build ADDri // Second group CheckOpcode G_SUB // Third pattern CheckNumOperand 3 ... Build SUBrr But if we allowed several nesting level, it could create a sub group for the checknumoperand 3. (We would need to call optimizeRules on the rules within a group.) *** Result *** With only one level of nesting, the instruction selection pass is up to 4x faster. For instance, one instruction now takes 500 checks, instead of 24k! With more nesting we could get in the tens I believe. Differential Revision: https://reviews.llvm.org/D39034 rdar://problem/34670699 llvm-svn: 321017
2017-12-19 03:47:41 +08:00
const PredicateMatcher &Predicate) const {
Reapply r332907 "[GlobalISel] Improving InstructionSelect's performance by reducing MatchTable..." Apparently the compile time problem was caused by the fact that not all compilers / STL implementations can automatically convert std::unique_ptr<Derived> to std::unique_ptr<Base>. Fixed (hopefully) by making sure it's std::unique_ptr<Derived>&& (rvalue ref) to std::unique_ptr<Base> conversion instead. llvm-svn: 332917
2018-05-22 07:28:51 +08:00
if (empty()) {
// Sharing predicates for nested instructions is not supported yet as we
// currently don't hoist the GIM_RecordInsn's properly, therefore we can
// only work on the original root instruction (InsnVarID == 0):
if (Predicate.getInsnVarID() != 0)
return false;
// ... otherwise an empty group can handle any predicate with no specific
// requirements:
return true;
}
const Matcher &Representative = **Matchers.begin();
const auto &RepresentativeCondition = Representative.getFirstCondition();
// ... if not empty, the group can only accomodate matchers with the exact
// same first condition:
return Predicate.isIdentical(RepresentativeCondition);
}
bool GroupMatcher::addMatcher(Matcher &Candidate) {
if (!Candidate.hasFirstCondition())
return false;
const PredicateMatcher &Predicate = Candidate.getFirstCondition();
if (!candidateConditionMatches(Predicate))
return false;
Matchers.push_back(&Candidate);
return true;
}
void GroupMatcher::finalize() {
assert(Conditions.empty() && "Already finalized?");
if (empty())
return;
Matcher &FirstRule = **Matchers.begin();
Conditions.push_back(FirstRule.popFirstCondition());
for (unsigned I = 1, E = Matchers.size(); I < E; ++I)
Matchers[I]->popFirstCondition();
[TableGen][GlobalISel] Optimize MatchTable for faster instruction selection *** Context *** Prior to this patchw, the table generated for matching instruction was straight forward but highly inefficient. Basically, each pattern generates its own set of self contained checks and actions. E.g., TableGen generated: // First pattern CheckNumOperand 3 CheckOpcode G_ADD ... Build ADDrr // Second pattern CheckNumOperand 3 CheckOpcode G_ADD ... Build ADDri // Third pattern CheckNumOperand 3 CheckOpcode G_SUB ... Build SUBrr *** Problem *** Because of that generation, a *lot* of check were redundant between each pattern and were checked every single time until we reach the pattern that matches. E.g., Taking the previous table, let say we are matching a G_SUB, that means we were going to check all the rules for G_ADD before looking at the G_SUB rule. In particular we are going to do: check 3 operands; PASS check G_ADD; FAIL ; Next rule check 3 operands; PASS (but we already knew that!) check G_ADD; FAIL (well it is still not true) ; Next rule check 3 operands; PASS (really!!) check G_SUB; PASS (at last :P) *** Proposed Solution *** This patch introduces a concept of group of rules (GroupMatcher) that share some predicates and only get checked once for the whole group. This patch only creates groups with one nesting level. Conceptually there is nothing preventing us for having deeper nest level. However, the current implementation is not smart enough to share the recording (aka capturing) of values. That limits its ability to do more sharing. For the given example the current patch will generate: // First group CheckOpcode G_ADD // First pattern CheckNumOperand 3 ... Build ADDrr // Second pattern CheckNumOperand 3 ... Build ADDri // Second group CheckOpcode G_SUB // Third pattern CheckNumOperand 3 ... Build SUBrr But if we allowed several nesting level, it could create a sub group for the checknumoperand 3. (We would need to call optimizeRules on the rules within a group.) *** Result *** With only one level of nesting, the instruction selection pass is up to 4x faster. For instance, one instruction now takes 500 checks, instead of 24k! With more nesting we could get in the tens I believe. Differential Revision: https://reviews.llvm.org/D39034 rdar://problem/34670699 llvm-svn: 321017
2017-12-19 03:47:41 +08:00
}
void GroupMatcher::emit(MatchTable &Table) {
Reapply r332907 "[GlobalISel] Improving InstructionSelect's performance by reducing MatchTable..." Apparently the compile time problem was caused by the fact that not all compilers / STL implementations can automatically convert std::unique_ptr<Derived> to std::unique_ptr<Base>. Fixed (hopefully) by making sure it's std::unique_ptr<Derived>&& (rvalue ref) to std::unique_ptr<Base> conversion instead. llvm-svn: 332917
2018-05-22 07:28:51 +08:00
unsigned LabelID = ~0U;
if (!Conditions.empty()) {
LabelID = Table.allocateLabelID();
[TableGen][GlobalISel] Optimize MatchTable for faster instruction selection *** Context *** Prior to this patchw, the table generated for matching instruction was straight forward but highly inefficient. Basically, each pattern generates its own set of self contained checks and actions. E.g., TableGen generated: // First pattern CheckNumOperand 3 CheckOpcode G_ADD ... Build ADDrr // Second pattern CheckNumOperand 3 CheckOpcode G_ADD ... Build ADDri // Third pattern CheckNumOperand 3 CheckOpcode G_SUB ... Build SUBrr *** Problem *** Because of that generation, a *lot* of check were redundant between each pattern and were checked every single time until we reach the pattern that matches. E.g., Taking the previous table, let say we are matching a G_SUB, that means we were going to check all the rules for G_ADD before looking at the G_SUB rule. In particular we are going to do: check 3 operands; PASS check G_ADD; FAIL ; Next rule check 3 operands; PASS (but we already knew that!) check G_ADD; FAIL (well it is still not true) ; Next rule check 3 operands; PASS (really!!) check G_SUB; PASS (at last :P) *** Proposed Solution *** This patch introduces a concept of group of rules (GroupMatcher) that share some predicates and only get checked once for the whole group. This patch only creates groups with one nesting level. Conceptually there is nothing preventing us for having deeper nest level. However, the current implementation is not smart enough to share the recording (aka capturing) of values. That limits its ability to do more sharing. For the given example the current patch will generate: // First group CheckOpcode G_ADD // First pattern CheckNumOperand 3 ... Build ADDrr // Second pattern CheckNumOperand 3 ... Build ADDri // Second group CheckOpcode G_SUB // Third pattern CheckNumOperand 3 ... Build SUBrr But if we allowed several nesting level, it could create a sub group for the checknumoperand 3. (We would need to call optimizeRules on the rules within a group.) *** Result *** With only one level of nesting, the instruction selection pass is up to 4x faster. For instance, one instruction now takes 500 checks, instead of 24k! With more nesting we could get in the tens I believe. Differential Revision: https://reviews.llvm.org/D39034 rdar://problem/34670699 llvm-svn: 321017
2017-12-19 03:47:41 +08:00
Table << MatchTable::Opcode("GIM_Try", +1)
<< MatchTable::Comment("On fail goto")
<< MatchTable::JumpTarget(LabelID) << MatchTable::LineBreak;
}
Reapply r332907 "[GlobalISel] Improving InstructionSelect's performance by reducing MatchTable..." Apparently the compile time problem was caused by the fact that not all compilers / STL implementations can automatically convert std::unique_ptr<Derived> to std::unique_ptr<Base>. Fixed (hopefully) by making sure it's std::unique_ptr<Derived>&& (rvalue ref) to std::unique_ptr<Base> conversion instead. llvm-svn: 332917
2018-05-22 07:28:51 +08:00
for (auto &Condition : Conditions)
Condition->emitPredicateOpcodes(
Table, *static_cast<RuleMatcher *>(*Matchers.begin()));
for (const auto &M : Matchers)
M->emit(Table);
// Exit the group
if (!Conditions.empty())
Table << MatchTable::Opcode("GIM_Reject", -1) << MatchTable::LineBreak
[TableGen][GlobalISel] Optimize MatchTable for faster instruction selection *** Context *** Prior to this patchw, the table generated for matching instruction was straight forward but highly inefficient. Basically, each pattern generates its own set of self contained checks and actions. E.g., TableGen generated: // First pattern CheckNumOperand 3 CheckOpcode G_ADD ... Build ADDrr // Second pattern CheckNumOperand 3 CheckOpcode G_ADD ... Build ADDri // Third pattern CheckNumOperand 3 CheckOpcode G_SUB ... Build SUBrr *** Problem *** Because of that generation, a *lot* of check were redundant between each pattern and were checked every single time until we reach the pattern that matches. E.g., Taking the previous table, let say we are matching a G_SUB, that means we were going to check all the rules for G_ADD before looking at the G_SUB rule. In particular we are going to do: check 3 operands; PASS check G_ADD; FAIL ; Next rule check 3 operands; PASS (but we already knew that!) check G_ADD; FAIL (well it is still not true) ; Next rule check 3 operands; PASS (really!!) check G_SUB; PASS (at last :P) *** Proposed Solution *** This patch introduces a concept of group of rules (GroupMatcher) that share some predicates and only get checked once for the whole group. This patch only creates groups with one nesting level. Conceptually there is nothing preventing us for having deeper nest level. However, the current implementation is not smart enough to share the recording (aka capturing) of values. That limits its ability to do more sharing. For the given example the current patch will generate: // First group CheckOpcode G_ADD // First pattern CheckNumOperand 3 ... Build ADDrr // Second pattern CheckNumOperand 3 ... Build ADDri // Second group CheckOpcode G_SUB // Third pattern CheckNumOperand 3 ... Build SUBrr But if we allowed several nesting level, it could create a sub group for the checknumoperand 3. (We would need to call optimizeRules on the rules within a group.) *** Result *** With only one level of nesting, the instruction selection pass is up to 4x faster. For instance, one instruction now takes 500 checks, instead of 24k! With more nesting we could get in the tens I believe. Differential Revision: https://reviews.llvm.org/D39034 rdar://problem/34670699 llvm-svn: 321017
2017-12-19 03:47:41 +08:00
<< MatchTable::Label(LabelID);
}
[GlobalISel][InstructionSelect] Switching MatchTable over opcodes, perf patch 4 This patch continues a series of patches started by r332907 (reapplied as r332917) In this commit we introduce a new matching opcode GIM_SwitchOpcode that implements a jump table over opcodes and start emitting them for root instructions. This is expected to decrease time GlobalISel spends in its InstructionSelect pass by roughly 20% for an -O0 build as measured on sqlite3-amalgamation (http://sqlite.org/download.html) targeting AArch64. To some degree, we assume here that the opcodes form a dense set, which is true at the moment for all upstream targets given the limitations of our rule importing mechanism. It might not be true for out of tree targets, specifically due to pseudo's. If so, we might noticeably increase the size of the MatchTable with this patch due to padding zeros. This will be addressed later. Reviewers: qcolombet, dsanders, bogner, aemerson, javed.absar Reviewed By: qcolombet Subscribers: rovka, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D44700 llvm-svn: 333017
2018-05-23 03:37:59 +08:00
bool SwitchMatcher::isSupportedPredicateType(const PredicateMatcher &P) {
return isa<InstructionOpcodeMatcher>(P);
}
bool SwitchMatcher::candidateConditionMatches(
const PredicateMatcher &Predicate) const {
if (empty()) {
// Sharing predicates for nested instructions is not supported yet as we
// currently don't hoist the GIM_RecordInsn's properly, therefore we can
// only work on the original root instruction (InsnVarID == 0):
if (Predicate.getInsnVarID() != 0)
return false;
// ... while an attempt to add even a root matcher to an empty SwitchMatcher
// could fail as not all the types of conditions are supported:
if (!isSupportedPredicateType(Predicate))
return false;
// ... or the condition might not have a proper implementation of
// getValue() / isIdenticalDownToValue() yet:
if (!Predicate.hasValue())
return false;
// ... otherwise an empty Switch can accomodate the condition with no
// further requirements:
return true;
}
const Matcher &CaseRepresentative = **Matchers.begin();
const auto &RepresentativeCondition = CaseRepresentative.getFirstCondition();
// Switch-cases must share the same kind of condition and path to the value it
// checks:
if (!Predicate.isIdenticalDownToValue(RepresentativeCondition))
return false;
const auto Value = Predicate.getValue();
// ... but be unique with respect to the actual value they check:
return Values.count(Value) == 0;
}
bool SwitchMatcher::addMatcher(Matcher &Candidate) {
if (!Candidate.hasFirstCondition())
return false;
const PredicateMatcher &Predicate = Candidate.getFirstCondition();
if (!candidateConditionMatches(Predicate))
return false;
const auto Value = Predicate.getValue();
Values.insert(Value);
Matchers.push_back(&Candidate);
return true;
}
void SwitchMatcher::finalize() {
assert(Condition == nullptr && "Already finalized");
assert(Values.size() == Matchers.size() && "Broken SwitchMatcher");
if (empty())
return;
std::stable_sort(Matchers.begin(), Matchers.end(),
[](const Matcher *L, const Matcher *R) {
return L->getFirstCondition().getValue() <
R->getFirstCondition().getValue();
});
Condition = Matchers[0]->popFirstCondition();
for (unsigned I = 1, E = Values.size(); I < E; ++I)
Matchers[I]->popFirstCondition();
}
void SwitchMatcher::emitPredicateSpecificOpcodes(const PredicateMatcher &P,
MatchTable &Table) {
assert(isSupportedPredicateType(P) && "Predicate type is not supported");
if (const auto *Condition = dyn_cast<InstructionOpcodeMatcher>(&P)) {
Table << MatchTable::Opcode("GIM_SwitchOpcode") << MatchTable::Comment("MI")
<< MatchTable::IntValue(Condition->getInsnVarID());
return;
}
llvm_unreachable("emitPredicateSpecificOpcodes is broken: can not handle a "
"predicate type that is claimed to be supported");
}
void SwitchMatcher::emit(MatchTable &Table) {
assert(Values.size() == Matchers.size() && "Broken SwitchMatcher");
if (empty())
return;
assert(Condition != nullptr &&
"Broken SwitchMatcher, hasn't been finalized?");
std::vector<unsigned> LabelIDs(Values.size());
std::generate(LabelIDs.begin(), LabelIDs.end(),
[&Table]() { return Table.allocateLabelID(); });
const unsigned Default = Table.allocateLabelID();
const int64_t LowerBound = Values.begin()->getRawValue();
const int64_t UpperBound = Values.rbegin()->getRawValue() + 1;
emitPredicateSpecificOpcodes(*Condition, Table);
Table << MatchTable::Comment("[") << MatchTable::IntValue(LowerBound)
<< MatchTable::IntValue(UpperBound) << MatchTable::Comment(")")
<< MatchTable::Comment("default:") << MatchTable::JumpTarget(Default);
int64_t J = LowerBound;
auto VI = Values.begin();
for (unsigned I = 0, E = Values.size(); I < E; ++I) {
auto V = *VI++;
while (J++ < V.getRawValue())
Table << MatchTable::IntValue(0);
V.turnIntoComment();
Table << MatchTable::LineBreak << V << MatchTable::JumpTarget(LabelIDs[I]);
}
Table << MatchTable::LineBreak;
for (unsigned I = 0, E = Values.size(); I < E; ++I) {
Table << MatchTable::Label(LabelIDs[I]);
Matchers[I]->emit(Table);
Table << MatchTable::Opcode("GIM_Reject") << MatchTable::LineBreak;
}
Table << MatchTable::Label(Default);
}
Reapply r332907 "[GlobalISel] Improving InstructionSelect's performance by reducing MatchTable..." Apparently the compile time problem was caused by the fact that not all compilers / STL implementations can automatically convert std::unique_ptr<Derived> to std::unique_ptr<Base>. Fixed (hopefully) by making sure it's std::unique_ptr<Derived>&& (rvalue ref) to std::unique_ptr<Base> conversion instead. llvm-svn: 332917
2018-05-22 07:28:51 +08:00
unsigned OperandMatcher::getInsnVarID() const { return Insn.getInsnVarID(); }
[TableGen][GlobalISel] Have the predicate directly know which data they are dealing with Prior to this patch, a predicate wouldn't make sense outside of its rule. Indeed, it was only during emitting a rule that a predicate would be made aware of the IDs of the data it is checking. Because of that, predicates could not be moved around or compared between each other. NFC. llvm-svn: 320887
2017-12-16 07:07:42 +08:00
[GlobalISel] Return an Expected<RuleMatcher> for each SDAG pattern. NFC. Instead of emitting the matcher code directly, return the rule matcher and the skip reason as an Expected<RuleMatcher>. This will let us record all matchers and process them before emission. It's a somewhat unconventional use of Error, but it's nicer than, say, std::pair, because of the bool conversions. Differential Revision: https://reviews.llvm.org/D29743 llvm-svn: 294706
2017-02-10 12:00:17 +08:00
} // end anonymous namespace
[GlobalISel] Add basic Selector-emitter tblgen backend. This adds a basic tablegen backend that analyzes the SelectionDAG patterns to find simple ones that are eligible for GlobalISel-emission. That's similar to FastISel, with one notable difference: we're not fed ISD opcodes, so we need to map the SDNode operators to generic opcodes. That's done using GINodeEquiv in TargetGlobalISel.td. Otherwise, this is mostly boilerplate, and lots of filtering of any kind of "complicated" pattern. On AArch64, this is sufficient to match G_ADD up to s64 (to ADDWrr/ADDXrr) and G_BR (to B). Differential Revision: https://reviews.llvm.org/D26878 llvm-svn: 290284
2016-12-22 07:26:20 +08:00
//===----------------------------------------------------------------------===//
namespace llvm {
void EmitGlobalISel(RecordKeeper &RK, raw_ostream &OS) {
GlobalISelEmitter(RK).run(OS);
}
} // End llvm namespace