llvm-project/llvm/lib/CodeGen/GlobalISel/LegalizerInfo.cpp

521 lines
21 KiB
C++
Raw Normal View History

//===- lib/CodeGen/GlobalISel/LegalizerInfo.cpp - Legalizer ---------------===//
//
// The LLVM Compiler Infrastructure
//
// This file is distributed under the University of Illinois Open Source
// License. See LICENSE.TXT for details.
//
//===----------------------------------------------------------------------===//
//
// Implement an interface to specify and query how an illegal operation on a
// given type should be expanded.
//
// Issues to be resolved:
// + Make it fast.
// + Support weird types like i3, <7 x i3>, ...
// + Operations with more than one type (ICMP, CMPXCHG, intrinsics, ...)
//
//===----------------------------------------------------------------------===//
#include "llvm/CodeGen/GlobalISel/LegalizerInfo.h"
#include "llvm/ADT/SmallBitVector.h"
#include "llvm/CodeGen/MachineInstr.h"
#include "llvm/CodeGen/MachineOperand.h"
#include "llvm/CodeGen/MachineRegisterInfo.h"
#include "llvm/CodeGen/TargetOpcodes.h"
#include "llvm/MC/MCInstrDesc.h"
[globalisel][legalizer] Adapt LegalizerInfo to support inter-type dependencies and other things. Summary: As discussed in D42244, we have difficulty describing the legality of some operations. We're not able to specify relationships between types. For example, declaring the following setAction({..., 0, s32}, Legal) setAction({..., 0, s64}, Legal) setAction({..., 1, s32}, Legal) setAction({..., 1, s64}, Legal) currently declares these type combinations as legal: {s32, s32} {s64, s32} {s32, s64} {s64, s64} but we currently have no means to say that, for example, {s64, s32} is not legal. Some operations such as G_INSERT/G_EXTRACT/G_MERGE_VALUES/ G_UNMERGE_VALUES have relationships between the types that are currently described incorrectly. Additionally, G_LOAD/G_STORE currently have no means to legalize non-atomics differently to atomics. The necessary information is in the MMO but we have no way to use this in the legalizer. Similarly, there is currently no way for the register type and the memory type to differ so there is no way to cleanly represent extending-load/truncating-store in a way that can't be broken by optimizers (resulting in illegal MIR). It's also difficult to control the legalization strategy. We've added support for legalizing non-power of 2 types but there's still some hardcoded assumptions about the strategy. The main one I've noticed is that type0 is always legalized before type1 which is not a good strategy for `type0 = G_EXTRACT type1, ...` if you need to widen the container. It will converge on the same result eventually but it will take a much longer route when legalizing type0 than if you legalize type1 first. Lastly, the definition of legality and the legalization strategy is kept separate which is not ideal. It's helpful to be able to look at a one piece of code and see both what is legal and the method the legalizer will use to make illegal MIR more legal. This patch adds a layer onto the LegalizerInfo (to be removed when all targets have been migrated) which resolves all these issues. Here are the rules for shift and division: for (unsigned BinOp : {G_LSHR, G_ASHR, G_SDIV, G_UDIV}) getActionDefinitions(BinOp) .legalFor({s32, s64}) // If type0 is s32/s64 then it's Legal .clampScalar(0, s32, s64) // If type0 is <s32 then WidenScalar to s32 // If type0 is >s64 then NarrowScalar to s64 .widenScalarToPow2(0) // Round type0 scalars up to powers of 2 .unsupported(); // Otherwise, it's unsupported This describes everything needed to both define legality and describe how to make illegal things legal. Here's an example of a complex rule: getActionDefinitions(G_INSERT) .unsupportedIf([=](const LegalityQuery &Query) { // If type0 is smaller than type1 then it's unsupported return Query.Types[0].getSizeInBits() <= Query.Types[1].getSizeInBits(); }) .legalIf([=](const LegalityQuery &Query) { // If type0 is s32/s64/p0 and type1 is a power of 2 other than 2 or 4 then it's legal // We don't need to worry about large type1's because unsupportedIf caught that. const LLT &Ty0 = Query.Types[0]; const LLT &Ty1 = Query.Types[1]; if (Ty0 != s32 && Ty0 != s64 && Ty0 != p0) return false; return isPowerOf2_32(Ty1.getSizeInBits()) && (Ty1.getSizeInBits() == 1 || Ty1.getSizeInBits() >= 8); }) .clampScalar(0, s32, s64) .widenScalarToPow2(0) .maxScalarIf(typeInSet(0, {s32}), 1, s16) // If type0 is s32 and type1 is bigger than s16 then NarrowScalar type1 to s16 .maxScalarIf(typeInSet(0, {s64}), 1, s32) // If type0 is s64 and type1 is bigger than s32 then NarrowScalar type1 to s32 .widenScalarToPow2(1) // Round type1 scalars up to powers of 2 .unsupported(); This uses a lambda to say that G_INSERT is unsupported when type0 is bigger than type1 (in practice, this would be a default rule for G_INSERT). It also uses one to describe the legal cases. This particular predicate is equivalent to: .legalFor({{s32, s1}, {s32, s8}, {s32, s16}, {s64, s1}, {s64, s8}, {s64, s16}, {s64, s32}}) In terms of performance, I saw a slight (~6%) performance improvement when AArch64 was around 30% ported but it's pretty much break even right now. I'm going to take a look at constexpr as a means to reduce the initialization cost. Future work: * Make it possible for opcodes to share rulesets. There's no need for G_LSHR/G_ASHR/G_SDIV/G_UDIV to have separate rule and ruleset objects. There's no technical barrier to this, it just hasn't been done yet. * Replace the type-index numbers with an enum to get .clampScalar(Type0, s32, s64) * Better names for things like .maxScalarIf() (clampMaxScalar?) and the vector rules. * Improve initialization cost using constexpr Possible future work: * It's possible to make these rulesets change the MIR directly instead of returning a description of how to change the MIR. This should remove a little overhead caused by parsing the description and routing to the right code, but the real motivation is that it removes the need for LegalizeAction::Custom. With Custom removed, there's no longer a requirement that Custom legalization change the opcode to something that's considered legal. Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar, volkan, reames, bogner Reviewed By: bogner Subscribers: hintonda, bogner, aemerson, mgorny, javed.absar, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D42251 llvm-svn: 323681
2018-01-30 03:54:49 +08:00
#include "llvm/Support/Debug.h"
#include "llvm/Support/ErrorHandling.h"
#include "llvm/Support/LowLevelTypeImpl.h"
#include "llvm/Support/MathExtras.h"
#include <algorithm>
[GlobalISel] Enable legalizing non-power-of-2 sized types. This changes the interface of how targets describe how to legalize, see the below description. 1. Interface for targets to describe how to legalize. In GlobalISel, the API in the LegalizerInfo class is the main interface for targets to specify which types are legal for which operations, and what to do to turn illegal type/operation combinations into legal ones. For each operation the type sizes that can be legalized without having to change the size of the type are specified with a call to setAction. This isn't different to how GlobalISel worked before. For example, for a target that supports 32 and 64 bit adds natively: for (auto Ty : {s32, s64}) setAction({G_ADD, 0, s32}, Legal); or for a target that needs a library call for a 32 bit division: setAction({G_SDIV, s32}, Libcall); The main conceptual change to the LegalizerInfo API, is in specifying how to legalize the type sizes for which a change of size is needed. For example, in the above example, how to specify how all types from i1 to i8388607 (apart from s32 and s64 which are legal) need to be legalized and expressed in terms of operations on the available legal sizes (again, i32 and i64 in this case). Before, the implementation only allowed specifying power-of-2-sized types (e.g. setAction({G_ADD, 0, s128}, NarrowScalar). A worse limitation was that if you'd wanted to specify how to legalize all the sized types as allowed by the LLVM-IR LangRef, i1 to i8388607, you'd have to call setAction 8388607-3 times and probably would need a lot of memory to store all of these specifications. Instead, the legalization actions that need to change the size of the type are specified now using a "SizeChangeStrategy". For example: setLegalizeScalarToDifferentSizeStrategy( G_ADD, 0, widenToLargerAndNarrowToLargest); This example indicates that for type sizes for which there is a larger size that can be legalized towards, do it by Widening the size. For example, G_ADD on s17 will be legalized by first doing WidenScalar to make it s32, after which it's legal. The "NarrowToLargest" indicates what to do if there is no larger size that can be legalized towards. E.g. G_ADD on s92 will be legalized by doing NarrowScalar to s64. Another example, taken from the ARM backend is: for (unsigned Op : {G_SDIV, G_UDIV}) { setLegalizeScalarToDifferentSizeStrategy(Op, 0, widenToLargerTypesUnsupportedOtherwise); if (ST.hasDivideInARMMode()) setAction({Op, s32}, Legal); else setAction({Op, s32}, Libcall); } For this example, G_SDIV on s8, on a target without a divide instruction, would be legalized by first doing action (WidenScalar, s32), followed by (Libcall, s32). The same principle is also followed for when the number of vector lanes on vector data types need to be changed, e.g.: setAction({G_ADD, LLT::vector(8, 8)}, LegalizerInfo::Legal); setAction({G_ADD, LLT::vector(16, 8)}, LegalizerInfo::Legal); setAction({G_ADD, LLT::vector(4, 16)}, LegalizerInfo::Legal); setAction({G_ADD, LLT::vector(8, 16)}, LegalizerInfo::Legal); setAction({G_ADD, LLT::vector(2, 32)}, LegalizerInfo::Legal); setAction({G_ADD, LLT::vector(4, 32)}, LegalizerInfo::Legal); setLegalizeVectorElementToDifferentSizeStrategy( G_ADD, 0, widenToLargerTypesUnsupportedOtherwise); As currently implemented here, vector types are legalized by first making the vector element size legal, followed by then making the number of lanes legal. The strategy to follow in the first step is set by a call to setLegalizeVectorElementToDifferentSizeStrategy, see example above. The strategy followed in the second step "moreToWiderTypesAndLessToWidest" (see code for its definition), indicating that vectors are widened to more elements so they map to natively supported vector widths, or when there isn't a legal wider vector, split the vector to map it to the widest vector supported. Therefore, for the above specification, some example legalizations are: * getAction({G_ADD, LLT::vector(3, 3)}) returns {WidenScalar, LLT::vector(3, 8)} * getAction({G_ADD, LLT::vector(3, 8)}) then returns {MoreElements, LLT::vector(8, 8)} * getAction({G_ADD, LLT::vector(20, 8)}) returns {FewerElements, LLT::vector(16, 8)} 2. Key implementation aspects. How to legalize a specific (operation, type index, size) tuple is represented by mapping intervals of integers representing a range of size types to an action to take, e.g.: setScalarAction({G_ADD, LLT:scalar(1)}, {{1, WidenScalar}, // bit sizes [ 1, 31[ {32, Legal}, // bit sizes [32, 33[ {33, WidenScalar}, // bit sizes [33, 64[ {64, Legal}, // bit sizes [64, 65[ {65, NarrowScalar} // bit sizes [65, +inf[ }); Please note that most of the code to do the actual lowering of non-power-of-2 sized types is currently missing, this is just trying to make it possible for targets to specify what is legal, and how non-legal types should be legalized. Probably quite a bit of further work is needed in the actual legalizing and the other passes in GlobalISel to support non-power-of-2 sized types. I hope the documentation in LegalizerInfo.h and the examples provided in the various {Target}LegalizerInfo.cpp and LegalizerInfoTest.cpp explains well enough how this is meant to be used. This drops the need for LLT::{half,double}...Size(). Differential Revision: https://reviews.llvm.org/D30529 llvm-svn: 317560
2017-11-07 18:34:34 +08:00
#include <map>
using namespace llvm;
using namespace LegalizeActions;
[globalisel][legalizer] Adapt LegalizerInfo to support inter-type dependencies and other things. Summary: As discussed in D42244, we have difficulty describing the legality of some operations. We're not able to specify relationships between types. For example, declaring the following setAction({..., 0, s32}, Legal) setAction({..., 0, s64}, Legal) setAction({..., 1, s32}, Legal) setAction({..., 1, s64}, Legal) currently declares these type combinations as legal: {s32, s32} {s64, s32} {s32, s64} {s64, s64} but we currently have no means to say that, for example, {s64, s32} is not legal. Some operations such as G_INSERT/G_EXTRACT/G_MERGE_VALUES/ G_UNMERGE_VALUES have relationships between the types that are currently described incorrectly. Additionally, G_LOAD/G_STORE currently have no means to legalize non-atomics differently to atomics. The necessary information is in the MMO but we have no way to use this in the legalizer. Similarly, there is currently no way for the register type and the memory type to differ so there is no way to cleanly represent extending-load/truncating-store in a way that can't be broken by optimizers (resulting in illegal MIR). It's also difficult to control the legalization strategy. We've added support for legalizing non-power of 2 types but there's still some hardcoded assumptions about the strategy. The main one I've noticed is that type0 is always legalized before type1 which is not a good strategy for `type0 = G_EXTRACT type1, ...` if you need to widen the container. It will converge on the same result eventually but it will take a much longer route when legalizing type0 than if you legalize type1 first. Lastly, the definition of legality and the legalization strategy is kept separate which is not ideal. It's helpful to be able to look at a one piece of code and see both what is legal and the method the legalizer will use to make illegal MIR more legal. This patch adds a layer onto the LegalizerInfo (to be removed when all targets have been migrated) which resolves all these issues. Here are the rules for shift and division: for (unsigned BinOp : {G_LSHR, G_ASHR, G_SDIV, G_UDIV}) getActionDefinitions(BinOp) .legalFor({s32, s64}) // If type0 is s32/s64 then it's Legal .clampScalar(0, s32, s64) // If type0 is <s32 then WidenScalar to s32 // If type0 is >s64 then NarrowScalar to s64 .widenScalarToPow2(0) // Round type0 scalars up to powers of 2 .unsupported(); // Otherwise, it's unsupported This describes everything needed to both define legality and describe how to make illegal things legal. Here's an example of a complex rule: getActionDefinitions(G_INSERT) .unsupportedIf([=](const LegalityQuery &Query) { // If type0 is smaller than type1 then it's unsupported return Query.Types[0].getSizeInBits() <= Query.Types[1].getSizeInBits(); }) .legalIf([=](const LegalityQuery &Query) { // If type0 is s32/s64/p0 and type1 is a power of 2 other than 2 or 4 then it's legal // We don't need to worry about large type1's because unsupportedIf caught that. const LLT &Ty0 = Query.Types[0]; const LLT &Ty1 = Query.Types[1]; if (Ty0 != s32 && Ty0 != s64 && Ty0 != p0) return false; return isPowerOf2_32(Ty1.getSizeInBits()) && (Ty1.getSizeInBits() == 1 || Ty1.getSizeInBits() >= 8); }) .clampScalar(0, s32, s64) .widenScalarToPow2(0) .maxScalarIf(typeInSet(0, {s32}), 1, s16) // If type0 is s32 and type1 is bigger than s16 then NarrowScalar type1 to s16 .maxScalarIf(typeInSet(0, {s64}), 1, s32) // If type0 is s64 and type1 is bigger than s32 then NarrowScalar type1 to s32 .widenScalarToPow2(1) // Round type1 scalars up to powers of 2 .unsupported(); This uses a lambda to say that G_INSERT is unsupported when type0 is bigger than type1 (in practice, this would be a default rule for G_INSERT). It also uses one to describe the legal cases. This particular predicate is equivalent to: .legalFor({{s32, s1}, {s32, s8}, {s32, s16}, {s64, s1}, {s64, s8}, {s64, s16}, {s64, s32}}) In terms of performance, I saw a slight (~6%) performance improvement when AArch64 was around 30% ported but it's pretty much break even right now. I'm going to take a look at constexpr as a means to reduce the initialization cost. Future work: * Make it possible for opcodes to share rulesets. There's no need for G_LSHR/G_ASHR/G_SDIV/G_UDIV to have separate rule and ruleset objects. There's no technical barrier to this, it just hasn't been done yet. * Replace the type-index numbers with an enum to get .clampScalar(Type0, s32, s64) * Better names for things like .maxScalarIf() (clampMaxScalar?) and the vector rules. * Improve initialization cost using constexpr Possible future work: * It's possible to make these rulesets change the MIR directly instead of returning a description of how to change the MIR. This should remove a little overhead caused by parsing the description and routing to the right code, but the real motivation is that it removes the need for LegalizeAction::Custom. With Custom removed, there's no longer a requirement that Custom legalization change the opcode to something that's considered legal. Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar, volkan, reames, bogner Reviewed By: bogner Subscribers: hintonda, bogner, aemerson, mgorny, javed.absar, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D42251 llvm-svn: 323681
2018-01-30 03:54:49 +08:00
#define DEBUG_TYPE "legalizer-info"
[GlobalISel][AArch64] Adding -disable-gisel-legality-check CL option Currently it's impossible to test InstructionSelect pass with MIR which is considered illegal by the Legalizer in Assert builds. In early stages of porting an existing backend from SelectionDAG ISel to GlobalISel, however, we would have very basic CallLowering, Legalizer, and RegBankSelect implementations, but rather functional Instruction Select with quite a few patterns selectable due to the semi-automatic porting process borrowing them from SelectionDAG ISel. As we are trying to define legality as a property of being selectable by the instruction selector, it would be nice to be able to easily check what the selector can do in its current state w/o the legality check provided by the Legalizer getting in the way. It also seems beneficial to have a regression testing set up that would not allow the selector to silently regress in its support of the MIR not supported yet by the previous passes in the GlobalISel pipeline. This commit adds -disable-gisel-legality-check command line option to llc that disables those legality checks in RegBankSelect and InstructionSelect passes. It also adds quite a few MIR test cases for AArch64's Instruction Selector. Every one of them would fail on the legality check at the moment, but will select just fine if the check is disabled. Every test MachineFunction is intended to exercise a specific selection rule and that rule only, encoded in the MachineFunction's name by the rule's number, ID, and index of its GIM_Try opcode in TableGen'erated MatchTable (-optimize-match-table=false). Reviewers: ab, dsanders, qcolombet, rovka Reviewed By: bogner Subscribers: kristof.beyls, volkan, aditya_nandakumar, aemerson, rengolin, t.p.northover, javed.absar, llvm-commits Differential Revision: https://reviews.llvm.org/D42886 llvm-svn: 326396
2018-03-01 08:27:48 +08:00
cl::opt<bool> llvm::DisableGISelLegalityCheck(
"disable-gisel-legality-check",
cl::desc("Don't verify that MIR is fully legal between GlobalISel passes"),
cl::Hidden);
[globalisel][legalizer] Adapt LegalizerInfo to support inter-type dependencies and other things. Summary: As discussed in D42244, we have difficulty describing the legality of some operations. We're not able to specify relationships between types. For example, declaring the following setAction({..., 0, s32}, Legal) setAction({..., 0, s64}, Legal) setAction({..., 1, s32}, Legal) setAction({..., 1, s64}, Legal) currently declares these type combinations as legal: {s32, s32} {s64, s32} {s32, s64} {s64, s64} but we currently have no means to say that, for example, {s64, s32} is not legal. Some operations such as G_INSERT/G_EXTRACT/G_MERGE_VALUES/ G_UNMERGE_VALUES have relationships between the types that are currently described incorrectly. Additionally, G_LOAD/G_STORE currently have no means to legalize non-atomics differently to atomics. The necessary information is in the MMO but we have no way to use this in the legalizer. Similarly, there is currently no way for the register type and the memory type to differ so there is no way to cleanly represent extending-load/truncating-store in a way that can't be broken by optimizers (resulting in illegal MIR). It's also difficult to control the legalization strategy. We've added support for legalizing non-power of 2 types but there's still some hardcoded assumptions about the strategy. The main one I've noticed is that type0 is always legalized before type1 which is not a good strategy for `type0 = G_EXTRACT type1, ...` if you need to widen the container. It will converge on the same result eventually but it will take a much longer route when legalizing type0 than if you legalize type1 first. Lastly, the definition of legality and the legalization strategy is kept separate which is not ideal. It's helpful to be able to look at a one piece of code and see both what is legal and the method the legalizer will use to make illegal MIR more legal. This patch adds a layer onto the LegalizerInfo (to be removed when all targets have been migrated) which resolves all these issues. Here are the rules for shift and division: for (unsigned BinOp : {G_LSHR, G_ASHR, G_SDIV, G_UDIV}) getActionDefinitions(BinOp) .legalFor({s32, s64}) // If type0 is s32/s64 then it's Legal .clampScalar(0, s32, s64) // If type0 is <s32 then WidenScalar to s32 // If type0 is >s64 then NarrowScalar to s64 .widenScalarToPow2(0) // Round type0 scalars up to powers of 2 .unsupported(); // Otherwise, it's unsupported This describes everything needed to both define legality and describe how to make illegal things legal. Here's an example of a complex rule: getActionDefinitions(G_INSERT) .unsupportedIf([=](const LegalityQuery &Query) { // If type0 is smaller than type1 then it's unsupported return Query.Types[0].getSizeInBits() <= Query.Types[1].getSizeInBits(); }) .legalIf([=](const LegalityQuery &Query) { // If type0 is s32/s64/p0 and type1 is a power of 2 other than 2 or 4 then it's legal // We don't need to worry about large type1's because unsupportedIf caught that. const LLT &Ty0 = Query.Types[0]; const LLT &Ty1 = Query.Types[1]; if (Ty0 != s32 && Ty0 != s64 && Ty0 != p0) return false; return isPowerOf2_32(Ty1.getSizeInBits()) && (Ty1.getSizeInBits() == 1 || Ty1.getSizeInBits() >= 8); }) .clampScalar(0, s32, s64) .widenScalarToPow2(0) .maxScalarIf(typeInSet(0, {s32}), 1, s16) // If type0 is s32 and type1 is bigger than s16 then NarrowScalar type1 to s16 .maxScalarIf(typeInSet(0, {s64}), 1, s32) // If type0 is s64 and type1 is bigger than s32 then NarrowScalar type1 to s32 .widenScalarToPow2(1) // Round type1 scalars up to powers of 2 .unsupported(); This uses a lambda to say that G_INSERT is unsupported when type0 is bigger than type1 (in practice, this would be a default rule for G_INSERT). It also uses one to describe the legal cases. This particular predicate is equivalent to: .legalFor({{s32, s1}, {s32, s8}, {s32, s16}, {s64, s1}, {s64, s8}, {s64, s16}, {s64, s32}}) In terms of performance, I saw a slight (~6%) performance improvement when AArch64 was around 30% ported but it's pretty much break even right now. I'm going to take a look at constexpr as a means to reduce the initialization cost. Future work: * Make it possible for opcodes to share rulesets. There's no need for G_LSHR/G_ASHR/G_SDIV/G_UDIV to have separate rule and ruleset objects. There's no technical barrier to this, it just hasn't been done yet. * Replace the type-index numbers with an enum to get .clampScalar(Type0, s32, s64) * Better names for things like .maxScalarIf() (clampMaxScalar?) and the vector rules. * Improve initialization cost using constexpr Possible future work: * It's possible to make these rulesets change the MIR directly instead of returning a description of how to change the MIR. This should remove a little overhead caused by parsing the description and routing to the right code, but the real motivation is that it removes the need for LegalizeAction::Custom. With Custom removed, there's no longer a requirement that Custom legalization change the opcode to something that's considered legal. Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar, volkan, reames, bogner Reviewed By: bogner Subscribers: hintonda, bogner, aemerson, mgorny, javed.absar, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D42251 llvm-svn: 323681
2018-01-30 03:54:49 +08:00
raw_ostream &LegalityQuery::print(raw_ostream &OS) const {
OS << Opcode << ", {";
for (const auto &Type : Types) {
OS << Type << ", ";
}
OS << "}";
return OS;
}
LegalizeActionStep LegalizeRuleSet::apply(const LegalityQuery &Query) const {
DEBUG(dbgs() << "Applying legalizer ruleset to: "; Query.print(dbgs());
dbgs() << "\n");
if (Rules.empty()) {
DEBUG(dbgs() << ".. fallback to legacy rules (no rules defined)\n");
return {LegalizeAction::UseLegacyRules, 0, LLT{}};
}
for (const auto &Rule : Rules) {
if (Rule.match(Query)) {
DEBUG(dbgs() << ".. match\n");
std::pair<unsigned, LLT> Mutation = Rule.determineMutation(Query);
DEBUG(dbgs() << ".. .. " << (unsigned)Rule.getAction() << ", "
<< Mutation.first << ", " << Mutation.second << "\n");
assert((Query.Types[Mutation.first] != Mutation.second ||
Rule.getAction() == MoreElements ||
Rule.getAction() == FewerElements) &&
[globalisel][legalizer] Adapt LegalizerInfo to support inter-type dependencies and other things. Summary: As discussed in D42244, we have difficulty describing the legality of some operations. We're not able to specify relationships between types. For example, declaring the following setAction({..., 0, s32}, Legal) setAction({..., 0, s64}, Legal) setAction({..., 1, s32}, Legal) setAction({..., 1, s64}, Legal) currently declares these type combinations as legal: {s32, s32} {s64, s32} {s32, s64} {s64, s64} but we currently have no means to say that, for example, {s64, s32} is not legal. Some operations such as G_INSERT/G_EXTRACT/G_MERGE_VALUES/ G_UNMERGE_VALUES have relationships between the types that are currently described incorrectly. Additionally, G_LOAD/G_STORE currently have no means to legalize non-atomics differently to atomics. The necessary information is in the MMO but we have no way to use this in the legalizer. Similarly, there is currently no way for the register type and the memory type to differ so there is no way to cleanly represent extending-load/truncating-store in a way that can't be broken by optimizers (resulting in illegal MIR). It's also difficult to control the legalization strategy. We've added support for legalizing non-power of 2 types but there's still some hardcoded assumptions about the strategy. The main one I've noticed is that type0 is always legalized before type1 which is not a good strategy for `type0 = G_EXTRACT type1, ...` if you need to widen the container. It will converge on the same result eventually but it will take a much longer route when legalizing type0 than if you legalize type1 first. Lastly, the definition of legality and the legalization strategy is kept separate which is not ideal. It's helpful to be able to look at a one piece of code and see both what is legal and the method the legalizer will use to make illegal MIR more legal. This patch adds a layer onto the LegalizerInfo (to be removed when all targets have been migrated) which resolves all these issues. Here are the rules for shift and division: for (unsigned BinOp : {G_LSHR, G_ASHR, G_SDIV, G_UDIV}) getActionDefinitions(BinOp) .legalFor({s32, s64}) // If type0 is s32/s64 then it's Legal .clampScalar(0, s32, s64) // If type0 is <s32 then WidenScalar to s32 // If type0 is >s64 then NarrowScalar to s64 .widenScalarToPow2(0) // Round type0 scalars up to powers of 2 .unsupported(); // Otherwise, it's unsupported This describes everything needed to both define legality and describe how to make illegal things legal. Here's an example of a complex rule: getActionDefinitions(G_INSERT) .unsupportedIf([=](const LegalityQuery &Query) { // If type0 is smaller than type1 then it's unsupported return Query.Types[0].getSizeInBits() <= Query.Types[1].getSizeInBits(); }) .legalIf([=](const LegalityQuery &Query) { // If type0 is s32/s64/p0 and type1 is a power of 2 other than 2 or 4 then it's legal // We don't need to worry about large type1's because unsupportedIf caught that. const LLT &Ty0 = Query.Types[0]; const LLT &Ty1 = Query.Types[1]; if (Ty0 != s32 && Ty0 != s64 && Ty0 != p0) return false; return isPowerOf2_32(Ty1.getSizeInBits()) && (Ty1.getSizeInBits() == 1 || Ty1.getSizeInBits() >= 8); }) .clampScalar(0, s32, s64) .widenScalarToPow2(0) .maxScalarIf(typeInSet(0, {s32}), 1, s16) // If type0 is s32 and type1 is bigger than s16 then NarrowScalar type1 to s16 .maxScalarIf(typeInSet(0, {s64}), 1, s32) // If type0 is s64 and type1 is bigger than s32 then NarrowScalar type1 to s32 .widenScalarToPow2(1) // Round type1 scalars up to powers of 2 .unsupported(); This uses a lambda to say that G_INSERT is unsupported when type0 is bigger than type1 (in practice, this would be a default rule for G_INSERT). It also uses one to describe the legal cases. This particular predicate is equivalent to: .legalFor({{s32, s1}, {s32, s8}, {s32, s16}, {s64, s1}, {s64, s8}, {s64, s16}, {s64, s32}}) In terms of performance, I saw a slight (~6%) performance improvement when AArch64 was around 30% ported but it's pretty much break even right now. I'm going to take a look at constexpr as a means to reduce the initialization cost. Future work: * Make it possible for opcodes to share rulesets. There's no need for G_LSHR/G_ASHR/G_SDIV/G_UDIV to have separate rule and ruleset objects. There's no technical barrier to this, it just hasn't been done yet. * Replace the type-index numbers with an enum to get .clampScalar(Type0, s32, s64) * Better names for things like .maxScalarIf() (clampMaxScalar?) and the vector rules. * Improve initialization cost using constexpr Possible future work: * It's possible to make these rulesets change the MIR directly instead of returning a description of how to change the MIR. This should remove a little overhead caused by parsing the description and routing to the right code, but the real motivation is that it removes the need for LegalizeAction::Custom. With Custom removed, there's no longer a requirement that Custom legalization change the opcode to something that's considered legal. Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar, volkan, reames, bogner Reviewed By: bogner Subscribers: hintonda, bogner, aemerson, mgorny, javed.absar, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D42251 llvm-svn: 323681
2018-01-30 03:54:49 +08:00
"Simple loop detected");
return {Rule.getAction(), Mutation.first, Mutation.second};
} else
DEBUG(dbgs() << ".. no match\n");
}
DEBUG(dbgs() << ".. unsupported\n");
return {LegalizeAction::Unsupported, 0, LLT{}};
}
[GlobalISel] Enable legalizing non-power-of-2 sized types. This changes the interface of how targets describe how to legalize, see the below description. 1. Interface for targets to describe how to legalize. In GlobalISel, the API in the LegalizerInfo class is the main interface for targets to specify which types are legal for which operations, and what to do to turn illegal type/operation combinations into legal ones. For each operation the type sizes that can be legalized without having to change the size of the type are specified with a call to setAction. This isn't different to how GlobalISel worked before. For example, for a target that supports 32 and 64 bit adds natively: for (auto Ty : {s32, s64}) setAction({G_ADD, 0, s32}, Legal); or for a target that needs a library call for a 32 bit division: setAction({G_SDIV, s32}, Libcall); The main conceptual change to the LegalizerInfo API, is in specifying how to legalize the type sizes for which a change of size is needed. For example, in the above example, how to specify how all types from i1 to i8388607 (apart from s32 and s64 which are legal) need to be legalized and expressed in terms of operations on the available legal sizes (again, i32 and i64 in this case). Before, the implementation only allowed specifying power-of-2-sized types (e.g. setAction({G_ADD, 0, s128}, NarrowScalar). A worse limitation was that if you'd wanted to specify how to legalize all the sized types as allowed by the LLVM-IR LangRef, i1 to i8388607, you'd have to call setAction 8388607-3 times and probably would need a lot of memory to store all of these specifications. Instead, the legalization actions that need to change the size of the type are specified now using a "SizeChangeStrategy". For example: setLegalizeScalarToDifferentSizeStrategy( G_ADD, 0, widenToLargerAndNarrowToLargest); This example indicates that for type sizes for which there is a larger size that can be legalized towards, do it by Widening the size. For example, G_ADD on s17 will be legalized by first doing WidenScalar to make it s32, after which it's legal. The "NarrowToLargest" indicates what to do if there is no larger size that can be legalized towards. E.g. G_ADD on s92 will be legalized by doing NarrowScalar to s64. Another example, taken from the ARM backend is: for (unsigned Op : {G_SDIV, G_UDIV}) { setLegalizeScalarToDifferentSizeStrategy(Op, 0, widenToLargerTypesUnsupportedOtherwise); if (ST.hasDivideInARMMode()) setAction({Op, s32}, Legal); else setAction({Op, s32}, Libcall); } For this example, G_SDIV on s8, on a target without a divide instruction, would be legalized by first doing action (WidenScalar, s32), followed by (Libcall, s32). The same principle is also followed for when the number of vector lanes on vector data types need to be changed, e.g.: setAction({G_ADD, LLT::vector(8, 8)}, LegalizerInfo::Legal); setAction({G_ADD, LLT::vector(16, 8)}, LegalizerInfo::Legal); setAction({G_ADD, LLT::vector(4, 16)}, LegalizerInfo::Legal); setAction({G_ADD, LLT::vector(8, 16)}, LegalizerInfo::Legal); setAction({G_ADD, LLT::vector(2, 32)}, LegalizerInfo::Legal); setAction({G_ADD, LLT::vector(4, 32)}, LegalizerInfo::Legal); setLegalizeVectorElementToDifferentSizeStrategy( G_ADD, 0, widenToLargerTypesUnsupportedOtherwise); As currently implemented here, vector types are legalized by first making the vector element size legal, followed by then making the number of lanes legal. The strategy to follow in the first step is set by a call to setLegalizeVectorElementToDifferentSizeStrategy, see example above. The strategy followed in the second step "moreToWiderTypesAndLessToWidest" (see code for its definition), indicating that vectors are widened to more elements so they map to natively supported vector widths, or when there isn't a legal wider vector, split the vector to map it to the widest vector supported. Therefore, for the above specification, some example legalizations are: * getAction({G_ADD, LLT::vector(3, 3)}) returns {WidenScalar, LLT::vector(3, 8)} * getAction({G_ADD, LLT::vector(3, 8)}) then returns {MoreElements, LLT::vector(8, 8)} * getAction({G_ADD, LLT::vector(20, 8)}) returns {FewerElements, LLT::vector(16, 8)} 2. Key implementation aspects. How to legalize a specific (operation, type index, size) tuple is represented by mapping intervals of integers representing a range of size types to an action to take, e.g.: setScalarAction({G_ADD, LLT:scalar(1)}, {{1, WidenScalar}, // bit sizes [ 1, 31[ {32, Legal}, // bit sizes [32, 33[ {33, WidenScalar}, // bit sizes [33, 64[ {64, Legal}, // bit sizes [64, 65[ {65, NarrowScalar} // bit sizes [65, +inf[ }); Please note that most of the code to do the actual lowering of non-power-of-2 sized types is currently missing, this is just trying to make it possible for targets to specify what is legal, and how non-legal types should be legalized. Probably quite a bit of further work is needed in the actual legalizing and the other passes in GlobalISel to support non-power-of-2 sized types. I hope the documentation in LegalizerInfo.h and the examples provided in the various {Target}LegalizerInfo.cpp and LegalizerInfoTest.cpp explains well enough how this is meant to be used. This drops the need for LLT::{half,double}...Size(). Differential Revision: https://reviews.llvm.org/D30529 llvm-svn: 317560
2017-11-07 18:34:34 +08:00
LegalizerInfo::LegalizerInfo() : TablesInitialized(false) {
// Set defaults.
// FIXME: these two (G_ANYEXT and G_TRUNC?) can be legalized to the
// fundamental load/store Jakob proposed. Once loads & stores are supported.
setScalarAction(TargetOpcode::G_ANYEXT, 1, {{1, Legal}});
setScalarAction(TargetOpcode::G_ZEXT, 1, {{1, Legal}});
setScalarAction(TargetOpcode::G_SEXT, 1, {{1, Legal}});
setScalarAction(TargetOpcode::G_TRUNC, 0, {{1, Legal}});
setScalarAction(TargetOpcode::G_TRUNC, 1, {{1, Legal}});
[GlobalISel] Enable legalizing non-power-of-2 sized types. This changes the interface of how targets describe how to legalize, see the below description. 1. Interface for targets to describe how to legalize. In GlobalISel, the API in the LegalizerInfo class is the main interface for targets to specify which types are legal for which operations, and what to do to turn illegal type/operation combinations into legal ones. For each operation the type sizes that can be legalized without having to change the size of the type are specified with a call to setAction. This isn't different to how GlobalISel worked before. For example, for a target that supports 32 and 64 bit adds natively: for (auto Ty : {s32, s64}) setAction({G_ADD, 0, s32}, Legal); or for a target that needs a library call for a 32 bit division: setAction({G_SDIV, s32}, Libcall); The main conceptual change to the LegalizerInfo API, is in specifying how to legalize the type sizes for which a change of size is needed. For example, in the above example, how to specify how all types from i1 to i8388607 (apart from s32 and s64 which are legal) need to be legalized and expressed in terms of operations on the available legal sizes (again, i32 and i64 in this case). Before, the implementation only allowed specifying power-of-2-sized types (e.g. setAction({G_ADD, 0, s128}, NarrowScalar). A worse limitation was that if you'd wanted to specify how to legalize all the sized types as allowed by the LLVM-IR LangRef, i1 to i8388607, you'd have to call setAction 8388607-3 times and probably would need a lot of memory to store all of these specifications. Instead, the legalization actions that need to change the size of the type are specified now using a "SizeChangeStrategy". For example: setLegalizeScalarToDifferentSizeStrategy( G_ADD, 0, widenToLargerAndNarrowToLargest); This example indicates that for type sizes for which there is a larger size that can be legalized towards, do it by Widening the size. For example, G_ADD on s17 will be legalized by first doing WidenScalar to make it s32, after which it's legal. The "NarrowToLargest" indicates what to do if there is no larger size that can be legalized towards. E.g. G_ADD on s92 will be legalized by doing NarrowScalar to s64. Another example, taken from the ARM backend is: for (unsigned Op : {G_SDIV, G_UDIV}) { setLegalizeScalarToDifferentSizeStrategy(Op, 0, widenToLargerTypesUnsupportedOtherwise); if (ST.hasDivideInARMMode()) setAction({Op, s32}, Legal); else setAction({Op, s32}, Libcall); } For this example, G_SDIV on s8, on a target without a divide instruction, would be legalized by first doing action (WidenScalar, s32), followed by (Libcall, s32). The same principle is also followed for when the number of vector lanes on vector data types need to be changed, e.g.: setAction({G_ADD, LLT::vector(8, 8)}, LegalizerInfo::Legal); setAction({G_ADD, LLT::vector(16, 8)}, LegalizerInfo::Legal); setAction({G_ADD, LLT::vector(4, 16)}, LegalizerInfo::Legal); setAction({G_ADD, LLT::vector(8, 16)}, LegalizerInfo::Legal); setAction({G_ADD, LLT::vector(2, 32)}, LegalizerInfo::Legal); setAction({G_ADD, LLT::vector(4, 32)}, LegalizerInfo::Legal); setLegalizeVectorElementToDifferentSizeStrategy( G_ADD, 0, widenToLargerTypesUnsupportedOtherwise); As currently implemented here, vector types are legalized by first making the vector element size legal, followed by then making the number of lanes legal. The strategy to follow in the first step is set by a call to setLegalizeVectorElementToDifferentSizeStrategy, see example above. The strategy followed in the second step "moreToWiderTypesAndLessToWidest" (see code for its definition), indicating that vectors are widened to more elements so they map to natively supported vector widths, or when there isn't a legal wider vector, split the vector to map it to the widest vector supported. Therefore, for the above specification, some example legalizations are: * getAction({G_ADD, LLT::vector(3, 3)}) returns {WidenScalar, LLT::vector(3, 8)} * getAction({G_ADD, LLT::vector(3, 8)}) then returns {MoreElements, LLT::vector(8, 8)} * getAction({G_ADD, LLT::vector(20, 8)}) returns {FewerElements, LLT::vector(16, 8)} 2. Key implementation aspects. How to legalize a specific (operation, type index, size) tuple is represented by mapping intervals of integers representing a range of size types to an action to take, e.g.: setScalarAction({G_ADD, LLT:scalar(1)}, {{1, WidenScalar}, // bit sizes [ 1, 31[ {32, Legal}, // bit sizes [32, 33[ {33, WidenScalar}, // bit sizes [33, 64[ {64, Legal}, // bit sizes [64, 65[ {65, NarrowScalar} // bit sizes [65, +inf[ }); Please note that most of the code to do the actual lowering of non-power-of-2 sized types is currently missing, this is just trying to make it possible for targets to specify what is legal, and how non-legal types should be legalized. Probably quite a bit of further work is needed in the actual legalizing and the other passes in GlobalISel to support non-power-of-2 sized types. I hope the documentation in LegalizerInfo.h and the examples provided in the various {Target}LegalizerInfo.cpp and LegalizerInfoTest.cpp explains well enough how this is meant to be used. This drops the need for LLT::{half,double}...Size(). Differential Revision: https://reviews.llvm.org/D30529 llvm-svn: 317560
2017-11-07 18:34:34 +08:00
setScalarAction(TargetOpcode::G_INTRINSIC, 0, {{1, Legal}});
setScalarAction(TargetOpcode::G_INTRINSIC_W_SIDE_EFFECTS, 0, {{1, Legal}});
[GlobalISel] Enable legalizing non-power-of-2 sized types. This changes the interface of how targets describe how to legalize, see the below description. 1. Interface for targets to describe how to legalize. In GlobalISel, the API in the LegalizerInfo class is the main interface for targets to specify which types are legal for which operations, and what to do to turn illegal type/operation combinations into legal ones. For each operation the type sizes that can be legalized without having to change the size of the type are specified with a call to setAction. This isn't different to how GlobalISel worked before. For example, for a target that supports 32 and 64 bit adds natively: for (auto Ty : {s32, s64}) setAction({G_ADD, 0, s32}, Legal); or for a target that needs a library call for a 32 bit division: setAction({G_SDIV, s32}, Libcall); The main conceptual change to the LegalizerInfo API, is in specifying how to legalize the type sizes for which a change of size is needed. For example, in the above example, how to specify how all types from i1 to i8388607 (apart from s32 and s64 which are legal) need to be legalized and expressed in terms of operations on the available legal sizes (again, i32 and i64 in this case). Before, the implementation only allowed specifying power-of-2-sized types (e.g. setAction({G_ADD, 0, s128}, NarrowScalar). A worse limitation was that if you'd wanted to specify how to legalize all the sized types as allowed by the LLVM-IR LangRef, i1 to i8388607, you'd have to call setAction 8388607-3 times and probably would need a lot of memory to store all of these specifications. Instead, the legalization actions that need to change the size of the type are specified now using a "SizeChangeStrategy". For example: setLegalizeScalarToDifferentSizeStrategy( G_ADD, 0, widenToLargerAndNarrowToLargest); This example indicates that for type sizes for which there is a larger size that can be legalized towards, do it by Widening the size. For example, G_ADD on s17 will be legalized by first doing WidenScalar to make it s32, after which it's legal. The "NarrowToLargest" indicates what to do if there is no larger size that can be legalized towards. E.g. G_ADD on s92 will be legalized by doing NarrowScalar to s64. Another example, taken from the ARM backend is: for (unsigned Op : {G_SDIV, G_UDIV}) { setLegalizeScalarToDifferentSizeStrategy(Op, 0, widenToLargerTypesUnsupportedOtherwise); if (ST.hasDivideInARMMode()) setAction({Op, s32}, Legal); else setAction({Op, s32}, Libcall); } For this example, G_SDIV on s8, on a target without a divide instruction, would be legalized by first doing action (WidenScalar, s32), followed by (Libcall, s32). The same principle is also followed for when the number of vector lanes on vector data types need to be changed, e.g.: setAction({G_ADD, LLT::vector(8, 8)}, LegalizerInfo::Legal); setAction({G_ADD, LLT::vector(16, 8)}, LegalizerInfo::Legal); setAction({G_ADD, LLT::vector(4, 16)}, LegalizerInfo::Legal); setAction({G_ADD, LLT::vector(8, 16)}, LegalizerInfo::Legal); setAction({G_ADD, LLT::vector(2, 32)}, LegalizerInfo::Legal); setAction({G_ADD, LLT::vector(4, 32)}, LegalizerInfo::Legal); setLegalizeVectorElementToDifferentSizeStrategy( G_ADD, 0, widenToLargerTypesUnsupportedOtherwise); As currently implemented here, vector types are legalized by first making the vector element size legal, followed by then making the number of lanes legal. The strategy to follow in the first step is set by a call to setLegalizeVectorElementToDifferentSizeStrategy, see example above. The strategy followed in the second step "moreToWiderTypesAndLessToWidest" (see code for its definition), indicating that vectors are widened to more elements so they map to natively supported vector widths, or when there isn't a legal wider vector, split the vector to map it to the widest vector supported. Therefore, for the above specification, some example legalizations are: * getAction({G_ADD, LLT::vector(3, 3)}) returns {WidenScalar, LLT::vector(3, 8)} * getAction({G_ADD, LLT::vector(3, 8)}) then returns {MoreElements, LLT::vector(8, 8)} * getAction({G_ADD, LLT::vector(20, 8)}) returns {FewerElements, LLT::vector(16, 8)} 2. Key implementation aspects. How to legalize a specific (operation, type index, size) tuple is represented by mapping intervals of integers representing a range of size types to an action to take, e.g.: setScalarAction({G_ADD, LLT:scalar(1)}, {{1, WidenScalar}, // bit sizes [ 1, 31[ {32, Legal}, // bit sizes [32, 33[ {33, WidenScalar}, // bit sizes [33, 64[ {64, Legal}, // bit sizes [64, 65[ {65, NarrowScalar} // bit sizes [65, +inf[ }); Please note that most of the code to do the actual lowering of non-power-of-2 sized types is currently missing, this is just trying to make it possible for targets to specify what is legal, and how non-legal types should be legalized. Probably quite a bit of further work is needed in the actual legalizing and the other passes in GlobalISel to support non-power-of-2 sized types. I hope the documentation in LegalizerInfo.h and the examples provided in the various {Target}LegalizerInfo.cpp and LegalizerInfoTest.cpp explains well enough how this is meant to be used. This drops the need for LLT::{half,double}...Size(). Differential Revision: https://reviews.llvm.org/D30529 llvm-svn: 317560
2017-11-07 18:34:34 +08:00
setLegalizeScalarToDifferentSizeStrategy(
TargetOpcode::G_IMPLICIT_DEF, 0, narrowToSmallerAndUnsupportedIfTooSmall);
setLegalizeScalarToDifferentSizeStrategy(
TargetOpcode::G_ADD, 0, widenToLargerTypesAndNarrowToLargest);
setLegalizeScalarToDifferentSizeStrategy(
TargetOpcode::G_OR, 0, widenToLargerTypesAndNarrowToLargest);
setLegalizeScalarToDifferentSizeStrategy(
TargetOpcode::G_LOAD, 0, narrowToSmallerAndUnsupportedIfTooSmall);
setLegalizeScalarToDifferentSizeStrategy(
TargetOpcode::G_STORE, 0, narrowToSmallerAndUnsupportedIfTooSmall);
[GlobalISel] Enable legalizing non-power-of-2 sized types. This changes the interface of how targets describe how to legalize, see the below description. 1. Interface for targets to describe how to legalize. In GlobalISel, the API in the LegalizerInfo class is the main interface for targets to specify which types are legal for which operations, and what to do to turn illegal type/operation combinations into legal ones. For each operation the type sizes that can be legalized without having to change the size of the type are specified with a call to setAction. This isn't different to how GlobalISel worked before. For example, for a target that supports 32 and 64 bit adds natively: for (auto Ty : {s32, s64}) setAction({G_ADD, 0, s32}, Legal); or for a target that needs a library call for a 32 bit division: setAction({G_SDIV, s32}, Libcall); The main conceptual change to the LegalizerInfo API, is in specifying how to legalize the type sizes for which a change of size is needed. For example, in the above example, how to specify how all types from i1 to i8388607 (apart from s32 and s64 which are legal) need to be legalized and expressed in terms of operations on the available legal sizes (again, i32 and i64 in this case). Before, the implementation only allowed specifying power-of-2-sized types (e.g. setAction({G_ADD, 0, s128}, NarrowScalar). A worse limitation was that if you'd wanted to specify how to legalize all the sized types as allowed by the LLVM-IR LangRef, i1 to i8388607, you'd have to call setAction 8388607-3 times and probably would need a lot of memory to store all of these specifications. Instead, the legalization actions that need to change the size of the type are specified now using a "SizeChangeStrategy". For example: setLegalizeScalarToDifferentSizeStrategy( G_ADD, 0, widenToLargerAndNarrowToLargest); This example indicates that for type sizes for which there is a larger size that can be legalized towards, do it by Widening the size. For example, G_ADD on s17 will be legalized by first doing WidenScalar to make it s32, after which it's legal. The "NarrowToLargest" indicates what to do if there is no larger size that can be legalized towards. E.g. G_ADD on s92 will be legalized by doing NarrowScalar to s64. Another example, taken from the ARM backend is: for (unsigned Op : {G_SDIV, G_UDIV}) { setLegalizeScalarToDifferentSizeStrategy(Op, 0, widenToLargerTypesUnsupportedOtherwise); if (ST.hasDivideInARMMode()) setAction({Op, s32}, Legal); else setAction({Op, s32}, Libcall); } For this example, G_SDIV on s8, on a target without a divide instruction, would be legalized by first doing action (WidenScalar, s32), followed by (Libcall, s32). The same principle is also followed for when the number of vector lanes on vector data types need to be changed, e.g.: setAction({G_ADD, LLT::vector(8, 8)}, LegalizerInfo::Legal); setAction({G_ADD, LLT::vector(16, 8)}, LegalizerInfo::Legal); setAction({G_ADD, LLT::vector(4, 16)}, LegalizerInfo::Legal); setAction({G_ADD, LLT::vector(8, 16)}, LegalizerInfo::Legal); setAction({G_ADD, LLT::vector(2, 32)}, LegalizerInfo::Legal); setAction({G_ADD, LLT::vector(4, 32)}, LegalizerInfo::Legal); setLegalizeVectorElementToDifferentSizeStrategy( G_ADD, 0, widenToLargerTypesUnsupportedOtherwise); As currently implemented here, vector types are legalized by first making the vector element size legal, followed by then making the number of lanes legal. The strategy to follow in the first step is set by a call to setLegalizeVectorElementToDifferentSizeStrategy, see example above. The strategy followed in the second step "moreToWiderTypesAndLessToWidest" (see code for its definition), indicating that vectors are widened to more elements so they map to natively supported vector widths, or when there isn't a legal wider vector, split the vector to map it to the widest vector supported. Therefore, for the above specification, some example legalizations are: * getAction({G_ADD, LLT::vector(3, 3)}) returns {WidenScalar, LLT::vector(3, 8)} * getAction({G_ADD, LLT::vector(3, 8)}) then returns {MoreElements, LLT::vector(8, 8)} * getAction({G_ADD, LLT::vector(20, 8)}) returns {FewerElements, LLT::vector(16, 8)} 2. Key implementation aspects. How to legalize a specific (operation, type index, size) tuple is represented by mapping intervals of integers representing a range of size types to an action to take, e.g.: setScalarAction({G_ADD, LLT:scalar(1)}, {{1, WidenScalar}, // bit sizes [ 1, 31[ {32, Legal}, // bit sizes [32, 33[ {33, WidenScalar}, // bit sizes [33, 64[ {64, Legal}, // bit sizes [64, 65[ {65, NarrowScalar} // bit sizes [65, +inf[ }); Please note that most of the code to do the actual lowering of non-power-of-2 sized types is currently missing, this is just trying to make it possible for targets to specify what is legal, and how non-legal types should be legalized. Probably quite a bit of further work is needed in the actual legalizing and the other passes in GlobalISel to support non-power-of-2 sized types. I hope the documentation in LegalizerInfo.h and the examples provided in the various {Target}LegalizerInfo.cpp and LegalizerInfoTest.cpp explains well enough how this is meant to be used. This drops the need for LLT::{half,double}...Size(). Differential Revision: https://reviews.llvm.org/D30529 llvm-svn: 317560
2017-11-07 18:34:34 +08:00
setLegalizeScalarToDifferentSizeStrategy(
TargetOpcode::G_BRCOND, 0, widenToLargerTypesUnsupportedOtherwise);
setLegalizeScalarToDifferentSizeStrategy(
TargetOpcode::G_INSERT, 0, narrowToSmallerAndUnsupportedIfTooSmall);
setLegalizeScalarToDifferentSizeStrategy(
TargetOpcode::G_EXTRACT, 0, narrowToSmallerAndUnsupportedIfTooSmall);
setLegalizeScalarToDifferentSizeStrategy(
TargetOpcode::G_EXTRACT, 1, narrowToSmallerAndUnsupportedIfTooSmall);
setScalarAction(TargetOpcode::G_FNEG, 0, {{1, Lower}});
}
void LegalizerInfo::computeTables() {
[GlobalISel] Enable legalizing non-power-of-2 sized types. This changes the interface of how targets describe how to legalize, see the below description. 1. Interface for targets to describe how to legalize. In GlobalISel, the API in the LegalizerInfo class is the main interface for targets to specify which types are legal for which operations, and what to do to turn illegal type/operation combinations into legal ones. For each operation the type sizes that can be legalized without having to change the size of the type are specified with a call to setAction. This isn't different to how GlobalISel worked before. For example, for a target that supports 32 and 64 bit adds natively: for (auto Ty : {s32, s64}) setAction({G_ADD, 0, s32}, Legal); or for a target that needs a library call for a 32 bit division: setAction({G_SDIV, s32}, Libcall); The main conceptual change to the LegalizerInfo API, is in specifying how to legalize the type sizes for which a change of size is needed. For example, in the above example, how to specify how all types from i1 to i8388607 (apart from s32 and s64 which are legal) need to be legalized and expressed in terms of operations on the available legal sizes (again, i32 and i64 in this case). Before, the implementation only allowed specifying power-of-2-sized types (e.g. setAction({G_ADD, 0, s128}, NarrowScalar). A worse limitation was that if you'd wanted to specify how to legalize all the sized types as allowed by the LLVM-IR LangRef, i1 to i8388607, you'd have to call setAction 8388607-3 times and probably would need a lot of memory to store all of these specifications. Instead, the legalization actions that need to change the size of the type are specified now using a "SizeChangeStrategy". For example: setLegalizeScalarToDifferentSizeStrategy( G_ADD, 0, widenToLargerAndNarrowToLargest); This example indicates that for type sizes for which there is a larger size that can be legalized towards, do it by Widening the size. For example, G_ADD on s17 will be legalized by first doing WidenScalar to make it s32, after which it's legal. The "NarrowToLargest" indicates what to do if there is no larger size that can be legalized towards. E.g. G_ADD on s92 will be legalized by doing NarrowScalar to s64. Another example, taken from the ARM backend is: for (unsigned Op : {G_SDIV, G_UDIV}) { setLegalizeScalarToDifferentSizeStrategy(Op, 0, widenToLargerTypesUnsupportedOtherwise); if (ST.hasDivideInARMMode()) setAction({Op, s32}, Legal); else setAction({Op, s32}, Libcall); } For this example, G_SDIV on s8, on a target without a divide instruction, would be legalized by first doing action (WidenScalar, s32), followed by (Libcall, s32). The same principle is also followed for when the number of vector lanes on vector data types need to be changed, e.g.: setAction({G_ADD, LLT::vector(8, 8)}, LegalizerInfo::Legal); setAction({G_ADD, LLT::vector(16, 8)}, LegalizerInfo::Legal); setAction({G_ADD, LLT::vector(4, 16)}, LegalizerInfo::Legal); setAction({G_ADD, LLT::vector(8, 16)}, LegalizerInfo::Legal); setAction({G_ADD, LLT::vector(2, 32)}, LegalizerInfo::Legal); setAction({G_ADD, LLT::vector(4, 32)}, LegalizerInfo::Legal); setLegalizeVectorElementToDifferentSizeStrategy( G_ADD, 0, widenToLargerTypesUnsupportedOtherwise); As currently implemented here, vector types are legalized by first making the vector element size legal, followed by then making the number of lanes legal. The strategy to follow in the first step is set by a call to setLegalizeVectorElementToDifferentSizeStrategy, see example above. The strategy followed in the second step "moreToWiderTypesAndLessToWidest" (see code for its definition), indicating that vectors are widened to more elements so they map to natively supported vector widths, or when there isn't a legal wider vector, split the vector to map it to the widest vector supported. Therefore, for the above specification, some example legalizations are: * getAction({G_ADD, LLT::vector(3, 3)}) returns {WidenScalar, LLT::vector(3, 8)} * getAction({G_ADD, LLT::vector(3, 8)}) then returns {MoreElements, LLT::vector(8, 8)} * getAction({G_ADD, LLT::vector(20, 8)}) returns {FewerElements, LLT::vector(16, 8)} 2. Key implementation aspects. How to legalize a specific (operation, type index, size) tuple is represented by mapping intervals of integers representing a range of size types to an action to take, e.g.: setScalarAction({G_ADD, LLT:scalar(1)}, {{1, WidenScalar}, // bit sizes [ 1, 31[ {32, Legal}, // bit sizes [32, 33[ {33, WidenScalar}, // bit sizes [33, 64[ {64, Legal}, // bit sizes [64, 65[ {65, NarrowScalar} // bit sizes [65, +inf[ }); Please note that most of the code to do the actual lowering of non-power-of-2 sized types is currently missing, this is just trying to make it possible for targets to specify what is legal, and how non-legal types should be legalized. Probably quite a bit of further work is needed in the actual legalizing and the other passes in GlobalISel to support non-power-of-2 sized types. I hope the documentation in LegalizerInfo.h and the examples provided in the various {Target}LegalizerInfo.cpp and LegalizerInfoTest.cpp explains well enough how this is meant to be used. This drops the need for LLT::{half,double}...Size(). Differential Revision: https://reviews.llvm.org/D30529 llvm-svn: 317560
2017-11-07 18:34:34 +08:00
assert(TablesInitialized == false);
for (unsigned OpcodeIdx = 0; OpcodeIdx <= LastOp - FirstOp; ++OpcodeIdx) {
const unsigned Opcode = FirstOp + OpcodeIdx;
for (unsigned TypeIdx = 0; TypeIdx != SpecifiedActions[OpcodeIdx].size();
++TypeIdx) {
// 0. Collect information specified through the setAction API, i.e.
// for specific bit sizes.
// For scalar types:
SizeAndActionsVec ScalarSpecifiedActions;
// For pointer types:
std::map<uint16_t, SizeAndActionsVec> AddressSpace2SpecifiedActions;
// For vector types:
std::map<uint16_t, SizeAndActionsVec> ElemSize2SpecifiedActions;
for (auto LLT2Action : SpecifiedActions[OpcodeIdx][TypeIdx]) {
const LLT Type = LLT2Action.first;
const LegalizeAction Action = LLT2Action.second;
auto SizeAction = std::make_pair(Type.getSizeInBits(), Action);
if (Type.isPointer())
AddressSpace2SpecifiedActions[Type.getAddressSpace()].push_back(
SizeAction);
else if (Type.isVector())
ElemSize2SpecifiedActions[Type.getElementType().getSizeInBits()]
.push_back(SizeAction);
else
ScalarSpecifiedActions.push_back(SizeAction);
}
// 1. Handle scalar types
{
// Decide how to handle bit sizes for which no explicit specification
// was given.
SizeChangeStrategy S = &unsupportedForDifferentSizes;
if (TypeIdx < ScalarSizeChangeStrategies[OpcodeIdx].size() &&
ScalarSizeChangeStrategies[OpcodeIdx][TypeIdx] != nullptr)
S = ScalarSizeChangeStrategies[OpcodeIdx][TypeIdx];
std::sort(ScalarSpecifiedActions.begin(), ScalarSpecifiedActions.end());
checkPartialSizeAndActionsVector(ScalarSpecifiedActions);
setScalarAction(Opcode, TypeIdx, S(ScalarSpecifiedActions));
}
[GlobalISel] Enable legalizing non-power-of-2 sized types. This changes the interface of how targets describe how to legalize, see the below description. 1. Interface for targets to describe how to legalize. In GlobalISel, the API in the LegalizerInfo class is the main interface for targets to specify which types are legal for which operations, and what to do to turn illegal type/operation combinations into legal ones. For each operation the type sizes that can be legalized without having to change the size of the type are specified with a call to setAction. This isn't different to how GlobalISel worked before. For example, for a target that supports 32 and 64 bit adds natively: for (auto Ty : {s32, s64}) setAction({G_ADD, 0, s32}, Legal); or for a target that needs a library call for a 32 bit division: setAction({G_SDIV, s32}, Libcall); The main conceptual change to the LegalizerInfo API, is in specifying how to legalize the type sizes for which a change of size is needed. For example, in the above example, how to specify how all types from i1 to i8388607 (apart from s32 and s64 which are legal) need to be legalized and expressed in terms of operations on the available legal sizes (again, i32 and i64 in this case). Before, the implementation only allowed specifying power-of-2-sized types (e.g. setAction({G_ADD, 0, s128}, NarrowScalar). A worse limitation was that if you'd wanted to specify how to legalize all the sized types as allowed by the LLVM-IR LangRef, i1 to i8388607, you'd have to call setAction 8388607-3 times and probably would need a lot of memory to store all of these specifications. Instead, the legalization actions that need to change the size of the type are specified now using a "SizeChangeStrategy". For example: setLegalizeScalarToDifferentSizeStrategy( G_ADD, 0, widenToLargerAndNarrowToLargest); This example indicates that for type sizes for which there is a larger size that can be legalized towards, do it by Widening the size. For example, G_ADD on s17 will be legalized by first doing WidenScalar to make it s32, after which it's legal. The "NarrowToLargest" indicates what to do if there is no larger size that can be legalized towards. E.g. G_ADD on s92 will be legalized by doing NarrowScalar to s64. Another example, taken from the ARM backend is: for (unsigned Op : {G_SDIV, G_UDIV}) { setLegalizeScalarToDifferentSizeStrategy(Op, 0, widenToLargerTypesUnsupportedOtherwise); if (ST.hasDivideInARMMode()) setAction({Op, s32}, Legal); else setAction({Op, s32}, Libcall); } For this example, G_SDIV on s8, on a target without a divide instruction, would be legalized by first doing action (WidenScalar, s32), followed by (Libcall, s32). The same principle is also followed for when the number of vector lanes on vector data types need to be changed, e.g.: setAction({G_ADD, LLT::vector(8, 8)}, LegalizerInfo::Legal); setAction({G_ADD, LLT::vector(16, 8)}, LegalizerInfo::Legal); setAction({G_ADD, LLT::vector(4, 16)}, LegalizerInfo::Legal); setAction({G_ADD, LLT::vector(8, 16)}, LegalizerInfo::Legal); setAction({G_ADD, LLT::vector(2, 32)}, LegalizerInfo::Legal); setAction({G_ADD, LLT::vector(4, 32)}, LegalizerInfo::Legal); setLegalizeVectorElementToDifferentSizeStrategy( G_ADD, 0, widenToLargerTypesUnsupportedOtherwise); As currently implemented here, vector types are legalized by first making the vector element size legal, followed by then making the number of lanes legal. The strategy to follow in the first step is set by a call to setLegalizeVectorElementToDifferentSizeStrategy, see example above. The strategy followed in the second step "moreToWiderTypesAndLessToWidest" (see code for its definition), indicating that vectors are widened to more elements so they map to natively supported vector widths, or when there isn't a legal wider vector, split the vector to map it to the widest vector supported. Therefore, for the above specification, some example legalizations are: * getAction({G_ADD, LLT::vector(3, 3)}) returns {WidenScalar, LLT::vector(3, 8)} * getAction({G_ADD, LLT::vector(3, 8)}) then returns {MoreElements, LLT::vector(8, 8)} * getAction({G_ADD, LLT::vector(20, 8)}) returns {FewerElements, LLT::vector(16, 8)} 2. Key implementation aspects. How to legalize a specific (operation, type index, size) tuple is represented by mapping intervals of integers representing a range of size types to an action to take, e.g.: setScalarAction({G_ADD, LLT:scalar(1)}, {{1, WidenScalar}, // bit sizes [ 1, 31[ {32, Legal}, // bit sizes [32, 33[ {33, WidenScalar}, // bit sizes [33, 64[ {64, Legal}, // bit sizes [64, 65[ {65, NarrowScalar} // bit sizes [65, +inf[ }); Please note that most of the code to do the actual lowering of non-power-of-2 sized types is currently missing, this is just trying to make it possible for targets to specify what is legal, and how non-legal types should be legalized. Probably quite a bit of further work is needed in the actual legalizing and the other passes in GlobalISel to support non-power-of-2 sized types. I hope the documentation in LegalizerInfo.h and the examples provided in the various {Target}LegalizerInfo.cpp and LegalizerInfoTest.cpp explains well enough how this is meant to be used. This drops the need for LLT::{half,double}...Size(). Differential Revision: https://reviews.llvm.org/D30529 llvm-svn: 317560
2017-11-07 18:34:34 +08:00
// 2. Handle pointer types
for (auto PointerSpecifiedActions : AddressSpace2SpecifiedActions) {
std::sort(PointerSpecifiedActions.second.begin(),
PointerSpecifiedActions.second.end());
checkPartialSizeAndActionsVector(PointerSpecifiedActions.second);
// For pointer types, we assume that there isn't a meaningfull way
// to change the number of bits used in the pointer.
setPointerAction(
Opcode, TypeIdx, PointerSpecifiedActions.first,
unsupportedForDifferentSizes(PointerSpecifiedActions.second));
}
// 3. Handle vector types
SizeAndActionsVec ElementSizesSeen;
for (auto VectorSpecifiedActions : ElemSize2SpecifiedActions) {
std::sort(VectorSpecifiedActions.second.begin(),
VectorSpecifiedActions.second.end());
const uint16_t ElementSize = VectorSpecifiedActions.first;
ElementSizesSeen.push_back({ElementSize, Legal});
checkPartialSizeAndActionsVector(VectorSpecifiedActions.second);
// For vector types, we assume that the best way to adapt the number
// of elements is to the next larger number of elements type for which
// the vector type is legal, unless there is no such type. In that case,
// legalize towards a vector type with a smaller number of elements.
SizeAndActionsVec NumElementsActions;
for (SizeAndAction BitsizeAndAction : VectorSpecifiedActions.second) {
assert(BitsizeAndAction.first % ElementSize == 0);
const uint16_t NumElements = BitsizeAndAction.first / ElementSize;
NumElementsActions.push_back({NumElements, BitsizeAndAction.second});
}
setVectorNumElementAction(
Opcode, TypeIdx, ElementSize,
moreToWiderTypesAndLessToWidest(NumElementsActions));
}
std::sort(ElementSizesSeen.begin(), ElementSizesSeen.end());
SizeChangeStrategy VectorElementSizeChangeStrategy =
&unsupportedForDifferentSizes;
if (TypeIdx < VectorElementSizeChangeStrategies[OpcodeIdx].size() &&
VectorElementSizeChangeStrategies[OpcodeIdx][TypeIdx] != nullptr)
VectorElementSizeChangeStrategy =
VectorElementSizeChangeStrategies[OpcodeIdx][TypeIdx];
setScalarInVectorAction(
Opcode, TypeIdx, VectorElementSizeChangeStrategy(ElementSizesSeen));
}
}
TablesInitialized = true;
}
// FIXME: inefficient implementation for now. Without ComputeValueVTs we're
// probably going to need specialized lookup structures for various types before
// we have any hope of doing well with something like <13 x i3>. Even the common
// cases should do better than what we have now.
std::pair<LegalizeAction, LLT>
[globalisel] Introduce LegalityQuery to better encapsulate the legalizer decisions. NFC. Summary: `getAction(const InstrAspect &) const` breaks encapsulation by exposing the smaller components that are used to decide how to legalize an instruction. This is a problem because we need to change the implementation of LegalizerInfo so that it's able to describe particular type combinations rather than just cartesian products of types. For example, declaring the following setAction({..., 0, s32}, Legal) setAction({..., 0, s64}, Legal) setAction({..., 1, s32}, Legal) setAction({..., 1, s64}, Legal) currently declares these type combinations as legal: {s32, s32} {s64, s32} {s32, s64} {s64, s64} but we currently have no means to say that, for example, {s64, s32} is not legal. Some operations such as G_INSERT/G_EXTRACT/G_MERGE_VALUES/ G_UNMERGE_VALUES has relationships between the types that are currently described incorrectly. Additionally, G_LOAD/G_STORE currently have no means to legalize non-atomics differently to atomics. The necessary information is in the MMO but we have no way to use this in the legalizer. Similarly, there is currently no way for the register type and the memory type to differ so there is no way to cleanly represent extending-load/truncating-store in a way that can't be broken by optimizers (resulting in illegal MIR). This patch introduces LegalityQuery which provides all the information needed by the legalizer to make a decision on whether something is legal and how to legalize it. Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar, volkan, reames, bogner Reviewed By: bogner Subscribers: bogner, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D42244 llvm-svn: 323342
2018-01-25 01:17:46 +08:00
LegalizerInfo::getAspectAction(const InstrAspect &Aspect) const {
assert(TablesInitialized && "backend forgot to call computeTables");
// These *have* to be implemented for now, they're the fundamental basis of
// how everything else is transformed.
[GlobalISel] Enable legalizing non-power-of-2 sized types. This changes the interface of how targets describe how to legalize, see the below description. 1. Interface for targets to describe how to legalize. In GlobalISel, the API in the LegalizerInfo class is the main interface for targets to specify which types are legal for which operations, and what to do to turn illegal type/operation combinations into legal ones. For each operation the type sizes that can be legalized without having to change the size of the type are specified with a call to setAction. This isn't different to how GlobalISel worked before. For example, for a target that supports 32 and 64 bit adds natively: for (auto Ty : {s32, s64}) setAction({G_ADD, 0, s32}, Legal); or for a target that needs a library call for a 32 bit division: setAction({G_SDIV, s32}, Libcall); The main conceptual change to the LegalizerInfo API, is in specifying how to legalize the type sizes for which a change of size is needed. For example, in the above example, how to specify how all types from i1 to i8388607 (apart from s32 and s64 which are legal) need to be legalized and expressed in terms of operations on the available legal sizes (again, i32 and i64 in this case). Before, the implementation only allowed specifying power-of-2-sized types (e.g. setAction({G_ADD, 0, s128}, NarrowScalar). A worse limitation was that if you'd wanted to specify how to legalize all the sized types as allowed by the LLVM-IR LangRef, i1 to i8388607, you'd have to call setAction 8388607-3 times and probably would need a lot of memory to store all of these specifications. Instead, the legalization actions that need to change the size of the type are specified now using a "SizeChangeStrategy". For example: setLegalizeScalarToDifferentSizeStrategy( G_ADD, 0, widenToLargerAndNarrowToLargest); This example indicates that for type sizes for which there is a larger size that can be legalized towards, do it by Widening the size. For example, G_ADD on s17 will be legalized by first doing WidenScalar to make it s32, after which it's legal. The "NarrowToLargest" indicates what to do if there is no larger size that can be legalized towards. E.g. G_ADD on s92 will be legalized by doing NarrowScalar to s64. Another example, taken from the ARM backend is: for (unsigned Op : {G_SDIV, G_UDIV}) { setLegalizeScalarToDifferentSizeStrategy(Op, 0, widenToLargerTypesUnsupportedOtherwise); if (ST.hasDivideInARMMode()) setAction({Op, s32}, Legal); else setAction({Op, s32}, Libcall); } For this example, G_SDIV on s8, on a target without a divide instruction, would be legalized by first doing action (WidenScalar, s32), followed by (Libcall, s32). The same principle is also followed for when the number of vector lanes on vector data types need to be changed, e.g.: setAction({G_ADD, LLT::vector(8, 8)}, LegalizerInfo::Legal); setAction({G_ADD, LLT::vector(16, 8)}, LegalizerInfo::Legal); setAction({G_ADD, LLT::vector(4, 16)}, LegalizerInfo::Legal); setAction({G_ADD, LLT::vector(8, 16)}, LegalizerInfo::Legal); setAction({G_ADD, LLT::vector(2, 32)}, LegalizerInfo::Legal); setAction({G_ADD, LLT::vector(4, 32)}, LegalizerInfo::Legal); setLegalizeVectorElementToDifferentSizeStrategy( G_ADD, 0, widenToLargerTypesUnsupportedOtherwise); As currently implemented here, vector types are legalized by first making the vector element size legal, followed by then making the number of lanes legal. The strategy to follow in the first step is set by a call to setLegalizeVectorElementToDifferentSizeStrategy, see example above. The strategy followed in the second step "moreToWiderTypesAndLessToWidest" (see code for its definition), indicating that vectors are widened to more elements so they map to natively supported vector widths, or when there isn't a legal wider vector, split the vector to map it to the widest vector supported. Therefore, for the above specification, some example legalizations are: * getAction({G_ADD, LLT::vector(3, 3)}) returns {WidenScalar, LLT::vector(3, 8)} * getAction({G_ADD, LLT::vector(3, 8)}) then returns {MoreElements, LLT::vector(8, 8)} * getAction({G_ADD, LLT::vector(20, 8)}) returns {FewerElements, LLT::vector(16, 8)} 2. Key implementation aspects. How to legalize a specific (operation, type index, size) tuple is represented by mapping intervals of integers representing a range of size types to an action to take, e.g.: setScalarAction({G_ADD, LLT:scalar(1)}, {{1, WidenScalar}, // bit sizes [ 1, 31[ {32, Legal}, // bit sizes [32, 33[ {33, WidenScalar}, // bit sizes [33, 64[ {64, Legal}, // bit sizes [64, 65[ {65, NarrowScalar} // bit sizes [65, +inf[ }); Please note that most of the code to do the actual lowering of non-power-of-2 sized types is currently missing, this is just trying to make it possible for targets to specify what is legal, and how non-legal types should be legalized. Probably quite a bit of further work is needed in the actual legalizing and the other passes in GlobalISel to support non-power-of-2 sized types. I hope the documentation in LegalizerInfo.h and the examples provided in the various {Target}LegalizerInfo.cpp and LegalizerInfoTest.cpp explains well enough how this is meant to be used. This drops the need for LLT::{half,double}...Size(). Differential Revision: https://reviews.llvm.org/D30529 llvm-svn: 317560
2017-11-07 18:34:34 +08:00
if (Aspect.Type.isScalar() || Aspect.Type.isPointer())
return findScalarLegalAction(Aspect);
assert(Aspect.Type.isVector());
return findVectorLegalAction(Aspect);
}
/// Helper function to get LLT for the given type index.
static LLT getTypeFromTypeIdx(const MachineInstr &MI,
const MachineRegisterInfo &MRI, unsigned OpIdx,
unsigned TypeIdx) {
assert(TypeIdx < MI.getNumOperands() && "Unexpected TypeIdx");
// G_UNMERGE_VALUES has variable number of operands, but there is only
// one source type and one destination type as all destinations must be the
// same type. So, get the last operand if TypeIdx == 1.
if (MI.getOpcode() == TargetOpcode::G_UNMERGE_VALUES && TypeIdx == 1)
return MRI.getType(MI.getOperand(MI.getNumOperands() - 1).getReg());
return MRI.getType(MI.getOperand(OpIdx).getReg());
}
[globalisel][legalizer] Adapt LegalizerInfo to support inter-type dependencies and other things. Summary: As discussed in D42244, we have difficulty describing the legality of some operations. We're not able to specify relationships between types. For example, declaring the following setAction({..., 0, s32}, Legal) setAction({..., 0, s64}, Legal) setAction({..., 1, s32}, Legal) setAction({..., 1, s64}, Legal) currently declares these type combinations as legal: {s32, s32} {s64, s32} {s32, s64} {s64, s64} but we currently have no means to say that, for example, {s64, s32} is not legal. Some operations such as G_INSERT/G_EXTRACT/G_MERGE_VALUES/ G_UNMERGE_VALUES have relationships between the types that are currently described incorrectly. Additionally, G_LOAD/G_STORE currently have no means to legalize non-atomics differently to atomics. The necessary information is in the MMO but we have no way to use this in the legalizer. Similarly, there is currently no way for the register type and the memory type to differ so there is no way to cleanly represent extending-load/truncating-store in a way that can't be broken by optimizers (resulting in illegal MIR). It's also difficult to control the legalization strategy. We've added support for legalizing non-power of 2 types but there's still some hardcoded assumptions about the strategy. The main one I've noticed is that type0 is always legalized before type1 which is not a good strategy for `type0 = G_EXTRACT type1, ...` if you need to widen the container. It will converge on the same result eventually but it will take a much longer route when legalizing type0 than if you legalize type1 first. Lastly, the definition of legality and the legalization strategy is kept separate which is not ideal. It's helpful to be able to look at a one piece of code and see both what is legal and the method the legalizer will use to make illegal MIR more legal. This patch adds a layer onto the LegalizerInfo (to be removed when all targets have been migrated) which resolves all these issues. Here are the rules for shift and division: for (unsigned BinOp : {G_LSHR, G_ASHR, G_SDIV, G_UDIV}) getActionDefinitions(BinOp) .legalFor({s32, s64}) // If type0 is s32/s64 then it's Legal .clampScalar(0, s32, s64) // If type0 is <s32 then WidenScalar to s32 // If type0 is >s64 then NarrowScalar to s64 .widenScalarToPow2(0) // Round type0 scalars up to powers of 2 .unsupported(); // Otherwise, it's unsupported This describes everything needed to both define legality and describe how to make illegal things legal. Here's an example of a complex rule: getActionDefinitions(G_INSERT) .unsupportedIf([=](const LegalityQuery &Query) { // If type0 is smaller than type1 then it's unsupported return Query.Types[0].getSizeInBits() <= Query.Types[1].getSizeInBits(); }) .legalIf([=](const LegalityQuery &Query) { // If type0 is s32/s64/p0 and type1 is a power of 2 other than 2 or 4 then it's legal // We don't need to worry about large type1's because unsupportedIf caught that. const LLT &Ty0 = Query.Types[0]; const LLT &Ty1 = Query.Types[1]; if (Ty0 != s32 && Ty0 != s64 && Ty0 != p0) return false; return isPowerOf2_32(Ty1.getSizeInBits()) && (Ty1.getSizeInBits() == 1 || Ty1.getSizeInBits() >= 8); }) .clampScalar(0, s32, s64) .widenScalarToPow2(0) .maxScalarIf(typeInSet(0, {s32}), 1, s16) // If type0 is s32 and type1 is bigger than s16 then NarrowScalar type1 to s16 .maxScalarIf(typeInSet(0, {s64}), 1, s32) // If type0 is s64 and type1 is bigger than s32 then NarrowScalar type1 to s32 .widenScalarToPow2(1) // Round type1 scalars up to powers of 2 .unsupported(); This uses a lambda to say that G_INSERT is unsupported when type0 is bigger than type1 (in practice, this would be a default rule for G_INSERT). It also uses one to describe the legal cases. This particular predicate is equivalent to: .legalFor({{s32, s1}, {s32, s8}, {s32, s16}, {s64, s1}, {s64, s8}, {s64, s16}, {s64, s32}}) In terms of performance, I saw a slight (~6%) performance improvement when AArch64 was around 30% ported but it's pretty much break even right now. I'm going to take a look at constexpr as a means to reduce the initialization cost. Future work: * Make it possible for opcodes to share rulesets. There's no need for G_LSHR/G_ASHR/G_SDIV/G_UDIV to have separate rule and ruleset objects. There's no technical barrier to this, it just hasn't been done yet. * Replace the type-index numbers with an enum to get .clampScalar(Type0, s32, s64) * Better names for things like .maxScalarIf() (clampMaxScalar?) and the vector rules. * Improve initialization cost using constexpr Possible future work: * It's possible to make these rulesets change the MIR directly instead of returning a description of how to change the MIR. This should remove a little overhead caused by parsing the description and routing to the right code, but the real motivation is that it removes the need for LegalizeAction::Custom. With Custom removed, there's no longer a requirement that Custom legalization change the opcode to something that's considered legal. Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar, volkan, reames, bogner Reviewed By: bogner Subscribers: hintonda, bogner, aemerson, mgorny, javed.absar, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D42251 llvm-svn: 323681
2018-01-30 03:54:49 +08:00
unsigned LegalizerInfo::getOpcodeIdxForOpcode(unsigned Opcode) const {
assert(Opcode >= FirstOp && Opcode <= LastOp && "Unsupported opcode");
return Opcode - FirstOp;
}
unsigned LegalizerInfo::getActionDefinitionsIdx(unsigned Opcode) const {
unsigned OpcodeIdx = getOpcodeIdxForOpcode(Opcode);
if (unsigned Alias = RulesForOpcode[OpcodeIdx].getAlias()) {
DEBUG(dbgs() << ".. opcode " << Opcode << " is aliased to " << Alias
<< "\n");
OpcodeIdx = getOpcodeIdxForOpcode(Alias);
DEBUG(dbgs() << ".. opcode " << Alias << " is aliased to "
<< RulesForOpcode[OpcodeIdx].getAlias() << "\n");
assert(RulesForOpcode[OpcodeIdx].getAlias() == 0 && "Cannot chain aliases");
}
return OpcodeIdx;
}
const LegalizeRuleSet &
LegalizerInfo::getActionDefinitions(unsigned Opcode) const {
unsigned OpcodeIdx = getActionDefinitionsIdx(Opcode);
return RulesForOpcode[OpcodeIdx];
}
LegalizeRuleSet &LegalizerInfo::getActionDefinitionsBuilder(unsigned Opcode) {
unsigned OpcodeIdx = getActionDefinitionsIdx(Opcode);
auto &Result = RulesForOpcode[OpcodeIdx];
assert(!Result.isAliasedByAnother() && "Modifying this opcode will modify aliases");
return Result;
}
LegalizeRuleSet &LegalizerInfo::getActionDefinitionsBuilder(
std::initializer_list<unsigned> Opcodes) {
unsigned Representative = *Opcodes.begin();
assert(Opcodes.begin() != Opcodes.end() &&
Opcodes.begin() + 1 != Opcodes.end() &&
"Initializer list must have at least two opcodes");
[globalisel][legalizer] Adapt LegalizerInfo to support inter-type dependencies and other things. Summary: As discussed in D42244, we have difficulty describing the legality of some operations. We're not able to specify relationships between types. For example, declaring the following setAction({..., 0, s32}, Legal) setAction({..., 0, s64}, Legal) setAction({..., 1, s32}, Legal) setAction({..., 1, s64}, Legal) currently declares these type combinations as legal: {s32, s32} {s64, s32} {s32, s64} {s64, s64} but we currently have no means to say that, for example, {s64, s32} is not legal. Some operations such as G_INSERT/G_EXTRACT/G_MERGE_VALUES/ G_UNMERGE_VALUES have relationships between the types that are currently described incorrectly. Additionally, G_LOAD/G_STORE currently have no means to legalize non-atomics differently to atomics. The necessary information is in the MMO but we have no way to use this in the legalizer. Similarly, there is currently no way for the register type and the memory type to differ so there is no way to cleanly represent extending-load/truncating-store in a way that can't be broken by optimizers (resulting in illegal MIR). It's also difficult to control the legalization strategy. We've added support for legalizing non-power of 2 types but there's still some hardcoded assumptions about the strategy. The main one I've noticed is that type0 is always legalized before type1 which is not a good strategy for `type0 = G_EXTRACT type1, ...` if you need to widen the container. It will converge on the same result eventually but it will take a much longer route when legalizing type0 than if you legalize type1 first. Lastly, the definition of legality and the legalization strategy is kept separate which is not ideal. It's helpful to be able to look at a one piece of code and see both what is legal and the method the legalizer will use to make illegal MIR more legal. This patch adds a layer onto the LegalizerInfo (to be removed when all targets have been migrated) which resolves all these issues. Here are the rules for shift and division: for (unsigned BinOp : {G_LSHR, G_ASHR, G_SDIV, G_UDIV}) getActionDefinitions(BinOp) .legalFor({s32, s64}) // If type0 is s32/s64 then it's Legal .clampScalar(0, s32, s64) // If type0 is <s32 then WidenScalar to s32 // If type0 is >s64 then NarrowScalar to s64 .widenScalarToPow2(0) // Round type0 scalars up to powers of 2 .unsupported(); // Otherwise, it's unsupported This describes everything needed to both define legality and describe how to make illegal things legal. Here's an example of a complex rule: getActionDefinitions(G_INSERT) .unsupportedIf([=](const LegalityQuery &Query) { // If type0 is smaller than type1 then it's unsupported return Query.Types[0].getSizeInBits() <= Query.Types[1].getSizeInBits(); }) .legalIf([=](const LegalityQuery &Query) { // If type0 is s32/s64/p0 and type1 is a power of 2 other than 2 or 4 then it's legal // We don't need to worry about large type1's because unsupportedIf caught that. const LLT &Ty0 = Query.Types[0]; const LLT &Ty1 = Query.Types[1]; if (Ty0 != s32 && Ty0 != s64 && Ty0 != p0) return false; return isPowerOf2_32(Ty1.getSizeInBits()) && (Ty1.getSizeInBits() == 1 || Ty1.getSizeInBits() >= 8); }) .clampScalar(0, s32, s64) .widenScalarToPow2(0) .maxScalarIf(typeInSet(0, {s32}), 1, s16) // If type0 is s32 and type1 is bigger than s16 then NarrowScalar type1 to s16 .maxScalarIf(typeInSet(0, {s64}), 1, s32) // If type0 is s64 and type1 is bigger than s32 then NarrowScalar type1 to s32 .widenScalarToPow2(1) // Round type1 scalars up to powers of 2 .unsupported(); This uses a lambda to say that G_INSERT is unsupported when type0 is bigger than type1 (in practice, this would be a default rule for G_INSERT). It also uses one to describe the legal cases. This particular predicate is equivalent to: .legalFor({{s32, s1}, {s32, s8}, {s32, s16}, {s64, s1}, {s64, s8}, {s64, s16}, {s64, s32}}) In terms of performance, I saw a slight (~6%) performance improvement when AArch64 was around 30% ported but it's pretty much break even right now. I'm going to take a look at constexpr as a means to reduce the initialization cost. Future work: * Make it possible for opcodes to share rulesets. There's no need for G_LSHR/G_ASHR/G_SDIV/G_UDIV to have separate rule and ruleset objects. There's no technical barrier to this, it just hasn't been done yet. * Replace the type-index numbers with an enum to get .clampScalar(Type0, s32, s64) * Better names for things like .maxScalarIf() (clampMaxScalar?) and the vector rules. * Improve initialization cost using constexpr Possible future work: * It's possible to make these rulesets change the MIR directly instead of returning a description of how to change the MIR. This should remove a little overhead caused by parsing the description and routing to the right code, but the real motivation is that it removes the need for LegalizeAction::Custom. With Custom removed, there's no longer a requirement that Custom legalization change the opcode to something that's considered legal. Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar, volkan, reames, bogner Reviewed By: bogner Subscribers: hintonda, bogner, aemerson, mgorny, javed.absar, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D42251 llvm-svn: 323681
2018-01-30 03:54:49 +08:00
for (auto I = Opcodes.begin() + 1, E = Opcodes.end(); I != E; ++I)
aliasActionDefinitions(Representative, *I);
auto &Return = getActionDefinitionsBuilder(Representative);
Return.setIsAliasedByAnother();
return Return;
}
void LegalizerInfo::aliasActionDefinitions(unsigned OpcodeTo,
unsigned OpcodeFrom) {
assert(OpcodeTo != OpcodeFrom && "Cannot alias to self");
assert(OpcodeTo >= FirstOp && OpcodeTo <= LastOp && "Unsupported opcode");
const unsigned OpcodeFromIdx = getOpcodeIdxForOpcode(OpcodeFrom);
RulesForOpcode[OpcodeFromIdx].aliasTo(OpcodeTo);
}
LegalizeActionStep
[globalisel] Introduce LegalityQuery to better encapsulate the legalizer decisions. NFC. Summary: `getAction(const InstrAspect &) const` breaks encapsulation by exposing the smaller components that are used to decide how to legalize an instruction. This is a problem because we need to change the implementation of LegalizerInfo so that it's able to describe particular type combinations rather than just cartesian products of types. For example, declaring the following setAction({..., 0, s32}, Legal) setAction({..., 0, s64}, Legal) setAction({..., 1, s32}, Legal) setAction({..., 1, s64}, Legal) currently declares these type combinations as legal: {s32, s32} {s64, s32} {s32, s64} {s64, s64} but we currently have no means to say that, for example, {s64, s32} is not legal. Some operations such as G_INSERT/G_EXTRACT/G_MERGE_VALUES/ G_UNMERGE_VALUES has relationships between the types that are currently described incorrectly. Additionally, G_LOAD/G_STORE currently have no means to legalize non-atomics differently to atomics. The necessary information is in the MMO but we have no way to use this in the legalizer. Similarly, there is currently no way for the register type and the memory type to differ so there is no way to cleanly represent extending-load/truncating-store in a way that can't be broken by optimizers (resulting in illegal MIR). This patch introduces LegalityQuery which provides all the information needed by the legalizer to make a decision on whether something is legal and how to legalize it. Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar, volkan, reames, bogner Reviewed By: bogner Subscribers: bogner, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D42244 llvm-svn: 323342
2018-01-25 01:17:46 +08:00
LegalizerInfo::getAction(const LegalityQuery &Query) const {
[globalisel][legalizer] Adapt LegalizerInfo to support inter-type dependencies and other things. Summary: As discussed in D42244, we have difficulty describing the legality of some operations. We're not able to specify relationships between types. For example, declaring the following setAction({..., 0, s32}, Legal) setAction({..., 0, s64}, Legal) setAction({..., 1, s32}, Legal) setAction({..., 1, s64}, Legal) currently declares these type combinations as legal: {s32, s32} {s64, s32} {s32, s64} {s64, s64} but we currently have no means to say that, for example, {s64, s32} is not legal. Some operations such as G_INSERT/G_EXTRACT/G_MERGE_VALUES/ G_UNMERGE_VALUES have relationships between the types that are currently described incorrectly. Additionally, G_LOAD/G_STORE currently have no means to legalize non-atomics differently to atomics. The necessary information is in the MMO but we have no way to use this in the legalizer. Similarly, there is currently no way for the register type and the memory type to differ so there is no way to cleanly represent extending-load/truncating-store in a way that can't be broken by optimizers (resulting in illegal MIR). It's also difficult to control the legalization strategy. We've added support for legalizing non-power of 2 types but there's still some hardcoded assumptions about the strategy. The main one I've noticed is that type0 is always legalized before type1 which is not a good strategy for `type0 = G_EXTRACT type1, ...` if you need to widen the container. It will converge on the same result eventually but it will take a much longer route when legalizing type0 than if you legalize type1 first. Lastly, the definition of legality and the legalization strategy is kept separate which is not ideal. It's helpful to be able to look at a one piece of code and see both what is legal and the method the legalizer will use to make illegal MIR more legal. This patch adds a layer onto the LegalizerInfo (to be removed when all targets have been migrated) which resolves all these issues. Here are the rules for shift and division: for (unsigned BinOp : {G_LSHR, G_ASHR, G_SDIV, G_UDIV}) getActionDefinitions(BinOp) .legalFor({s32, s64}) // If type0 is s32/s64 then it's Legal .clampScalar(0, s32, s64) // If type0 is <s32 then WidenScalar to s32 // If type0 is >s64 then NarrowScalar to s64 .widenScalarToPow2(0) // Round type0 scalars up to powers of 2 .unsupported(); // Otherwise, it's unsupported This describes everything needed to both define legality and describe how to make illegal things legal. Here's an example of a complex rule: getActionDefinitions(G_INSERT) .unsupportedIf([=](const LegalityQuery &Query) { // If type0 is smaller than type1 then it's unsupported return Query.Types[0].getSizeInBits() <= Query.Types[1].getSizeInBits(); }) .legalIf([=](const LegalityQuery &Query) { // If type0 is s32/s64/p0 and type1 is a power of 2 other than 2 or 4 then it's legal // We don't need to worry about large type1's because unsupportedIf caught that. const LLT &Ty0 = Query.Types[0]; const LLT &Ty1 = Query.Types[1]; if (Ty0 != s32 && Ty0 != s64 && Ty0 != p0) return false; return isPowerOf2_32(Ty1.getSizeInBits()) && (Ty1.getSizeInBits() == 1 || Ty1.getSizeInBits() >= 8); }) .clampScalar(0, s32, s64) .widenScalarToPow2(0) .maxScalarIf(typeInSet(0, {s32}), 1, s16) // If type0 is s32 and type1 is bigger than s16 then NarrowScalar type1 to s16 .maxScalarIf(typeInSet(0, {s64}), 1, s32) // If type0 is s64 and type1 is bigger than s32 then NarrowScalar type1 to s32 .widenScalarToPow2(1) // Round type1 scalars up to powers of 2 .unsupported(); This uses a lambda to say that G_INSERT is unsupported when type0 is bigger than type1 (in practice, this would be a default rule for G_INSERT). It also uses one to describe the legal cases. This particular predicate is equivalent to: .legalFor({{s32, s1}, {s32, s8}, {s32, s16}, {s64, s1}, {s64, s8}, {s64, s16}, {s64, s32}}) In terms of performance, I saw a slight (~6%) performance improvement when AArch64 was around 30% ported but it's pretty much break even right now. I'm going to take a look at constexpr as a means to reduce the initialization cost. Future work: * Make it possible for opcodes to share rulesets. There's no need for G_LSHR/G_ASHR/G_SDIV/G_UDIV to have separate rule and ruleset objects. There's no technical barrier to this, it just hasn't been done yet. * Replace the type-index numbers with an enum to get .clampScalar(Type0, s32, s64) * Better names for things like .maxScalarIf() (clampMaxScalar?) and the vector rules. * Improve initialization cost using constexpr Possible future work: * It's possible to make these rulesets change the MIR directly instead of returning a description of how to change the MIR. This should remove a little overhead caused by parsing the description and routing to the right code, but the real motivation is that it removes the need for LegalizeAction::Custom. With Custom removed, there's no longer a requirement that Custom legalization change the opcode to something that's considered legal. Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar, volkan, reames, bogner Reviewed By: bogner Subscribers: hintonda, bogner, aemerson, mgorny, javed.absar, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D42251 llvm-svn: 323681
2018-01-30 03:54:49 +08:00
LegalizeActionStep Step = getActionDefinitions(Query.Opcode).apply(Query);
if (Step.Action != LegalizeAction::UseLegacyRules) {
return Step;
}
[globalisel] Introduce LegalityQuery to better encapsulate the legalizer decisions. NFC. Summary: `getAction(const InstrAspect &) const` breaks encapsulation by exposing the smaller components that are used to decide how to legalize an instruction. This is a problem because we need to change the implementation of LegalizerInfo so that it's able to describe particular type combinations rather than just cartesian products of types. For example, declaring the following setAction({..., 0, s32}, Legal) setAction({..., 0, s64}, Legal) setAction({..., 1, s32}, Legal) setAction({..., 1, s64}, Legal) currently declares these type combinations as legal: {s32, s32} {s64, s32} {s32, s64} {s64, s64} but we currently have no means to say that, for example, {s64, s32} is not legal. Some operations such as G_INSERT/G_EXTRACT/G_MERGE_VALUES/ G_UNMERGE_VALUES has relationships between the types that are currently described incorrectly. Additionally, G_LOAD/G_STORE currently have no means to legalize non-atomics differently to atomics. The necessary information is in the MMO but we have no way to use this in the legalizer. Similarly, there is currently no way for the register type and the memory type to differ so there is no way to cleanly represent extending-load/truncating-store in a way that can't be broken by optimizers (resulting in illegal MIR). This patch introduces LegalityQuery which provides all the information needed by the legalizer to make a decision on whether something is legal and how to legalize it. Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar, volkan, reames, bogner Reviewed By: bogner Subscribers: bogner, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D42244 llvm-svn: 323342
2018-01-25 01:17:46 +08:00
for (unsigned i = 0; i < Query.Types.size(); ++i) {
auto Action = getAspectAction({Query.Opcode, i, Query.Types[i]});
[globalisel][legalizer] Adapt LegalizerInfo to support inter-type dependencies and other things. Summary: As discussed in D42244, we have difficulty describing the legality of some operations. We're not able to specify relationships between types. For example, declaring the following setAction({..., 0, s32}, Legal) setAction({..., 0, s64}, Legal) setAction({..., 1, s32}, Legal) setAction({..., 1, s64}, Legal) currently declares these type combinations as legal: {s32, s32} {s64, s32} {s32, s64} {s64, s64} but we currently have no means to say that, for example, {s64, s32} is not legal. Some operations such as G_INSERT/G_EXTRACT/G_MERGE_VALUES/ G_UNMERGE_VALUES have relationships between the types that are currently described incorrectly. Additionally, G_LOAD/G_STORE currently have no means to legalize non-atomics differently to atomics. The necessary information is in the MMO but we have no way to use this in the legalizer. Similarly, there is currently no way for the register type and the memory type to differ so there is no way to cleanly represent extending-load/truncating-store in a way that can't be broken by optimizers (resulting in illegal MIR). It's also difficult to control the legalization strategy. We've added support for legalizing non-power of 2 types but there's still some hardcoded assumptions about the strategy. The main one I've noticed is that type0 is always legalized before type1 which is not a good strategy for `type0 = G_EXTRACT type1, ...` if you need to widen the container. It will converge on the same result eventually but it will take a much longer route when legalizing type0 than if you legalize type1 first. Lastly, the definition of legality and the legalization strategy is kept separate which is not ideal. It's helpful to be able to look at a one piece of code and see both what is legal and the method the legalizer will use to make illegal MIR more legal. This patch adds a layer onto the LegalizerInfo (to be removed when all targets have been migrated) which resolves all these issues. Here are the rules for shift and division: for (unsigned BinOp : {G_LSHR, G_ASHR, G_SDIV, G_UDIV}) getActionDefinitions(BinOp) .legalFor({s32, s64}) // If type0 is s32/s64 then it's Legal .clampScalar(0, s32, s64) // If type0 is <s32 then WidenScalar to s32 // If type0 is >s64 then NarrowScalar to s64 .widenScalarToPow2(0) // Round type0 scalars up to powers of 2 .unsupported(); // Otherwise, it's unsupported This describes everything needed to both define legality and describe how to make illegal things legal. Here's an example of a complex rule: getActionDefinitions(G_INSERT) .unsupportedIf([=](const LegalityQuery &Query) { // If type0 is smaller than type1 then it's unsupported return Query.Types[0].getSizeInBits() <= Query.Types[1].getSizeInBits(); }) .legalIf([=](const LegalityQuery &Query) { // If type0 is s32/s64/p0 and type1 is a power of 2 other than 2 or 4 then it's legal // We don't need to worry about large type1's because unsupportedIf caught that. const LLT &Ty0 = Query.Types[0]; const LLT &Ty1 = Query.Types[1]; if (Ty0 != s32 && Ty0 != s64 && Ty0 != p0) return false; return isPowerOf2_32(Ty1.getSizeInBits()) && (Ty1.getSizeInBits() == 1 || Ty1.getSizeInBits() >= 8); }) .clampScalar(0, s32, s64) .widenScalarToPow2(0) .maxScalarIf(typeInSet(0, {s32}), 1, s16) // If type0 is s32 and type1 is bigger than s16 then NarrowScalar type1 to s16 .maxScalarIf(typeInSet(0, {s64}), 1, s32) // If type0 is s64 and type1 is bigger than s32 then NarrowScalar type1 to s32 .widenScalarToPow2(1) // Round type1 scalars up to powers of 2 .unsupported(); This uses a lambda to say that G_INSERT is unsupported when type0 is bigger than type1 (in practice, this would be a default rule for G_INSERT). It also uses one to describe the legal cases. This particular predicate is equivalent to: .legalFor({{s32, s1}, {s32, s8}, {s32, s16}, {s64, s1}, {s64, s8}, {s64, s16}, {s64, s32}}) In terms of performance, I saw a slight (~6%) performance improvement when AArch64 was around 30% ported but it's pretty much break even right now. I'm going to take a look at constexpr as a means to reduce the initialization cost. Future work: * Make it possible for opcodes to share rulesets. There's no need for G_LSHR/G_ASHR/G_SDIV/G_UDIV to have separate rule and ruleset objects. There's no technical barrier to this, it just hasn't been done yet. * Replace the type-index numbers with an enum to get .clampScalar(Type0, s32, s64) * Better names for things like .maxScalarIf() (clampMaxScalar?) and the vector rules. * Improve initialization cost using constexpr Possible future work: * It's possible to make these rulesets change the MIR directly instead of returning a description of how to change the MIR. This should remove a little overhead caused by parsing the description and routing to the right code, but the real motivation is that it removes the need for LegalizeAction::Custom. With Custom removed, there's no longer a requirement that Custom legalization change the opcode to something that's considered legal. Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar, volkan, reames, bogner Reviewed By: bogner Subscribers: hintonda, bogner, aemerson, mgorny, javed.absar, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D42251 llvm-svn: 323681
2018-01-30 03:54:49 +08:00
if (Action.first != Legal) {
DEBUG(dbgs() << ".. (legacy) Type " << i << " Action="
<< (unsigned)Action.first << ", " << Action.second << "\n");
[globalisel] Introduce LegalityQuery to better encapsulate the legalizer decisions. NFC. Summary: `getAction(const InstrAspect &) const` breaks encapsulation by exposing the smaller components that are used to decide how to legalize an instruction. This is a problem because we need to change the implementation of LegalizerInfo so that it's able to describe particular type combinations rather than just cartesian products of types. For example, declaring the following setAction({..., 0, s32}, Legal) setAction({..., 0, s64}, Legal) setAction({..., 1, s32}, Legal) setAction({..., 1, s64}, Legal) currently declares these type combinations as legal: {s32, s32} {s64, s32} {s32, s64} {s64, s64} but we currently have no means to say that, for example, {s64, s32} is not legal. Some operations such as G_INSERT/G_EXTRACT/G_MERGE_VALUES/ G_UNMERGE_VALUES has relationships between the types that are currently described incorrectly. Additionally, G_LOAD/G_STORE currently have no means to legalize non-atomics differently to atomics. The necessary information is in the MMO but we have no way to use this in the legalizer. Similarly, there is currently no way for the register type and the memory type to differ so there is no way to cleanly represent extending-load/truncating-store in a way that can't be broken by optimizers (resulting in illegal MIR). This patch introduces LegalityQuery which provides all the information needed by the legalizer to make a decision on whether something is legal and how to legalize it. Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar, volkan, reames, bogner Reviewed By: bogner Subscribers: bogner, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D42244 llvm-svn: 323342
2018-01-25 01:17:46 +08:00
return {Action.first, i, Action.second};
[globalisel][legalizer] Adapt LegalizerInfo to support inter-type dependencies and other things. Summary: As discussed in D42244, we have difficulty describing the legality of some operations. We're not able to specify relationships between types. For example, declaring the following setAction({..., 0, s32}, Legal) setAction({..., 0, s64}, Legal) setAction({..., 1, s32}, Legal) setAction({..., 1, s64}, Legal) currently declares these type combinations as legal: {s32, s32} {s64, s32} {s32, s64} {s64, s64} but we currently have no means to say that, for example, {s64, s32} is not legal. Some operations such as G_INSERT/G_EXTRACT/G_MERGE_VALUES/ G_UNMERGE_VALUES have relationships between the types that are currently described incorrectly. Additionally, G_LOAD/G_STORE currently have no means to legalize non-atomics differently to atomics. The necessary information is in the MMO but we have no way to use this in the legalizer. Similarly, there is currently no way for the register type and the memory type to differ so there is no way to cleanly represent extending-load/truncating-store in a way that can't be broken by optimizers (resulting in illegal MIR). It's also difficult to control the legalization strategy. We've added support for legalizing non-power of 2 types but there's still some hardcoded assumptions about the strategy. The main one I've noticed is that type0 is always legalized before type1 which is not a good strategy for `type0 = G_EXTRACT type1, ...` if you need to widen the container. It will converge on the same result eventually but it will take a much longer route when legalizing type0 than if you legalize type1 first. Lastly, the definition of legality and the legalization strategy is kept separate which is not ideal. It's helpful to be able to look at a one piece of code and see both what is legal and the method the legalizer will use to make illegal MIR more legal. This patch adds a layer onto the LegalizerInfo (to be removed when all targets have been migrated) which resolves all these issues. Here are the rules for shift and division: for (unsigned BinOp : {G_LSHR, G_ASHR, G_SDIV, G_UDIV}) getActionDefinitions(BinOp) .legalFor({s32, s64}) // If type0 is s32/s64 then it's Legal .clampScalar(0, s32, s64) // If type0 is <s32 then WidenScalar to s32 // If type0 is >s64 then NarrowScalar to s64 .widenScalarToPow2(0) // Round type0 scalars up to powers of 2 .unsupported(); // Otherwise, it's unsupported This describes everything needed to both define legality and describe how to make illegal things legal. Here's an example of a complex rule: getActionDefinitions(G_INSERT) .unsupportedIf([=](const LegalityQuery &Query) { // If type0 is smaller than type1 then it's unsupported return Query.Types[0].getSizeInBits() <= Query.Types[1].getSizeInBits(); }) .legalIf([=](const LegalityQuery &Query) { // If type0 is s32/s64/p0 and type1 is a power of 2 other than 2 or 4 then it's legal // We don't need to worry about large type1's because unsupportedIf caught that. const LLT &Ty0 = Query.Types[0]; const LLT &Ty1 = Query.Types[1]; if (Ty0 != s32 && Ty0 != s64 && Ty0 != p0) return false; return isPowerOf2_32(Ty1.getSizeInBits()) && (Ty1.getSizeInBits() == 1 || Ty1.getSizeInBits() >= 8); }) .clampScalar(0, s32, s64) .widenScalarToPow2(0) .maxScalarIf(typeInSet(0, {s32}), 1, s16) // If type0 is s32 and type1 is bigger than s16 then NarrowScalar type1 to s16 .maxScalarIf(typeInSet(0, {s64}), 1, s32) // If type0 is s64 and type1 is bigger than s32 then NarrowScalar type1 to s32 .widenScalarToPow2(1) // Round type1 scalars up to powers of 2 .unsupported(); This uses a lambda to say that G_INSERT is unsupported when type0 is bigger than type1 (in practice, this would be a default rule for G_INSERT). It also uses one to describe the legal cases. This particular predicate is equivalent to: .legalFor({{s32, s1}, {s32, s8}, {s32, s16}, {s64, s1}, {s64, s8}, {s64, s16}, {s64, s32}}) In terms of performance, I saw a slight (~6%) performance improvement when AArch64 was around 30% ported but it's pretty much break even right now. I'm going to take a look at constexpr as a means to reduce the initialization cost. Future work: * Make it possible for opcodes to share rulesets. There's no need for G_LSHR/G_ASHR/G_SDIV/G_UDIV to have separate rule and ruleset objects. There's no technical barrier to this, it just hasn't been done yet. * Replace the type-index numbers with an enum to get .clampScalar(Type0, s32, s64) * Better names for things like .maxScalarIf() (clampMaxScalar?) and the vector rules. * Improve initialization cost using constexpr Possible future work: * It's possible to make these rulesets change the MIR directly instead of returning a description of how to change the MIR. This should remove a little overhead caused by parsing the description and routing to the right code, but the real motivation is that it removes the need for LegalizeAction::Custom. With Custom removed, there's no longer a requirement that Custom legalization change the opcode to something that's considered legal. Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar, volkan, reames, bogner Reviewed By: bogner Subscribers: hintonda, bogner, aemerson, mgorny, javed.absar, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D42251 llvm-svn: 323681
2018-01-30 03:54:49 +08:00
} else
DEBUG(dbgs() << ".. (legacy) Type " << i << " Legal\n");
[globalisel] Introduce LegalityQuery to better encapsulate the legalizer decisions. NFC. Summary: `getAction(const InstrAspect &) const` breaks encapsulation by exposing the smaller components that are used to decide how to legalize an instruction. This is a problem because we need to change the implementation of LegalizerInfo so that it's able to describe particular type combinations rather than just cartesian products of types. For example, declaring the following setAction({..., 0, s32}, Legal) setAction({..., 0, s64}, Legal) setAction({..., 1, s32}, Legal) setAction({..., 1, s64}, Legal) currently declares these type combinations as legal: {s32, s32} {s64, s32} {s32, s64} {s64, s64} but we currently have no means to say that, for example, {s64, s32} is not legal. Some operations such as G_INSERT/G_EXTRACT/G_MERGE_VALUES/ G_UNMERGE_VALUES has relationships between the types that are currently described incorrectly. Additionally, G_LOAD/G_STORE currently have no means to legalize non-atomics differently to atomics. The necessary information is in the MMO but we have no way to use this in the legalizer. Similarly, there is currently no way for the register type and the memory type to differ so there is no way to cleanly represent extending-load/truncating-store in a way that can't be broken by optimizers (resulting in illegal MIR). This patch introduces LegalityQuery which provides all the information needed by the legalizer to make a decision on whether something is legal and how to legalize it. Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar, volkan, reames, bogner Reviewed By: bogner Subscribers: bogner, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D42244 llvm-svn: 323342
2018-01-25 01:17:46 +08:00
}
[globalisel][legalizer] Adapt LegalizerInfo to support inter-type dependencies and other things. Summary: As discussed in D42244, we have difficulty describing the legality of some operations. We're not able to specify relationships between types. For example, declaring the following setAction({..., 0, s32}, Legal) setAction({..., 0, s64}, Legal) setAction({..., 1, s32}, Legal) setAction({..., 1, s64}, Legal) currently declares these type combinations as legal: {s32, s32} {s64, s32} {s32, s64} {s64, s64} but we currently have no means to say that, for example, {s64, s32} is not legal. Some operations such as G_INSERT/G_EXTRACT/G_MERGE_VALUES/ G_UNMERGE_VALUES have relationships between the types that are currently described incorrectly. Additionally, G_LOAD/G_STORE currently have no means to legalize non-atomics differently to atomics. The necessary information is in the MMO but we have no way to use this in the legalizer. Similarly, there is currently no way for the register type and the memory type to differ so there is no way to cleanly represent extending-load/truncating-store in a way that can't be broken by optimizers (resulting in illegal MIR). It's also difficult to control the legalization strategy. We've added support for legalizing non-power of 2 types but there's still some hardcoded assumptions about the strategy. The main one I've noticed is that type0 is always legalized before type1 which is not a good strategy for `type0 = G_EXTRACT type1, ...` if you need to widen the container. It will converge on the same result eventually but it will take a much longer route when legalizing type0 than if you legalize type1 first. Lastly, the definition of legality and the legalization strategy is kept separate which is not ideal. It's helpful to be able to look at a one piece of code and see both what is legal and the method the legalizer will use to make illegal MIR more legal. This patch adds a layer onto the LegalizerInfo (to be removed when all targets have been migrated) which resolves all these issues. Here are the rules for shift and division: for (unsigned BinOp : {G_LSHR, G_ASHR, G_SDIV, G_UDIV}) getActionDefinitions(BinOp) .legalFor({s32, s64}) // If type0 is s32/s64 then it's Legal .clampScalar(0, s32, s64) // If type0 is <s32 then WidenScalar to s32 // If type0 is >s64 then NarrowScalar to s64 .widenScalarToPow2(0) // Round type0 scalars up to powers of 2 .unsupported(); // Otherwise, it's unsupported This describes everything needed to both define legality and describe how to make illegal things legal. Here's an example of a complex rule: getActionDefinitions(G_INSERT) .unsupportedIf([=](const LegalityQuery &Query) { // If type0 is smaller than type1 then it's unsupported return Query.Types[0].getSizeInBits() <= Query.Types[1].getSizeInBits(); }) .legalIf([=](const LegalityQuery &Query) { // If type0 is s32/s64/p0 and type1 is a power of 2 other than 2 or 4 then it's legal // We don't need to worry about large type1's because unsupportedIf caught that. const LLT &Ty0 = Query.Types[0]; const LLT &Ty1 = Query.Types[1]; if (Ty0 != s32 && Ty0 != s64 && Ty0 != p0) return false; return isPowerOf2_32(Ty1.getSizeInBits()) && (Ty1.getSizeInBits() == 1 || Ty1.getSizeInBits() >= 8); }) .clampScalar(0, s32, s64) .widenScalarToPow2(0) .maxScalarIf(typeInSet(0, {s32}), 1, s16) // If type0 is s32 and type1 is bigger than s16 then NarrowScalar type1 to s16 .maxScalarIf(typeInSet(0, {s64}), 1, s32) // If type0 is s64 and type1 is bigger than s32 then NarrowScalar type1 to s32 .widenScalarToPow2(1) // Round type1 scalars up to powers of 2 .unsupported(); This uses a lambda to say that G_INSERT is unsupported when type0 is bigger than type1 (in practice, this would be a default rule for G_INSERT). It also uses one to describe the legal cases. This particular predicate is equivalent to: .legalFor({{s32, s1}, {s32, s8}, {s32, s16}, {s64, s1}, {s64, s8}, {s64, s16}, {s64, s32}}) In terms of performance, I saw a slight (~6%) performance improvement when AArch64 was around 30% ported but it's pretty much break even right now. I'm going to take a look at constexpr as a means to reduce the initialization cost. Future work: * Make it possible for opcodes to share rulesets. There's no need for G_LSHR/G_ASHR/G_SDIV/G_UDIV to have separate rule and ruleset objects. There's no technical barrier to this, it just hasn't been done yet. * Replace the type-index numbers with an enum to get .clampScalar(Type0, s32, s64) * Better names for things like .maxScalarIf() (clampMaxScalar?) and the vector rules. * Improve initialization cost using constexpr Possible future work: * It's possible to make these rulesets change the MIR directly instead of returning a description of how to change the MIR. This should remove a little overhead caused by parsing the description and routing to the right code, but the real motivation is that it removes the need for LegalizeAction::Custom. With Custom removed, there's no longer a requirement that Custom legalization change the opcode to something that's considered legal. Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar, volkan, reames, bogner Reviewed By: bogner Subscribers: hintonda, bogner, aemerson, mgorny, javed.absar, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D42251 llvm-svn: 323681
2018-01-30 03:54:49 +08:00
DEBUG(dbgs() << ".. (legacy) Legal\n");
[globalisel] Introduce LegalityQuery to better encapsulate the legalizer decisions. NFC. Summary: `getAction(const InstrAspect &) const` breaks encapsulation by exposing the smaller components that are used to decide how to legalize an instruction. This is a problem because we need to change the implementation of LegalizerInfo so that it's able to describe particular type combinations rather than just cartesian products of types. For example, declaring the following setAction({..., 0, s32}, Legal) setAction({..., 0, s64}, Legal) setAction({..., 1, s32}, Legal) setAction({..., 1, s64}, Legal) currently declares these type combinations as legal: {s32, s32} {s64, s32} {s32, s64} {s64, s64} but we currently have no means to say that, for example, {s64, s32} is not legal. Some operations such as G_INSERT/G_EXTRACT/G_MERGE_VALUES/ G_UNMERGE_VALUES has relationships between the types that are currently described incorrectly. Additionally, G_LOAD/G_STORE currently have no means to legalize non-atomics differently to atomics. The necessary information is in the MMO but we have no way to use this in the legalizer. Similarly, there is currently no way for the register type and the memory type to differ so there is no way to cleanly represent extending-load/truncating-store in a way that can't be broken by optimizers (resulting in illegal MIR). This patch introduces LegalityQuery which provides all the information needed by the legalizer to make a decision on whether something is legal and how to legalize it. Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar, volkan, reames, bogner Reviewed By: bogner Subscribers: bogner, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D42244 llvm-svn: 323342
2018-01-25 01:17:46 +08:00
return {Legal, 0, LLT{}};
}
[globalisel][legalizer] Adapt LegalizerInfo to support inter-type dependencies and other things. Summary: As discussed in D42244, we have difficulty describing the legality of some operations. We're not able to specify relationships between types. For example, declaring the following setAction({..., 0, s32}, Legal) setAction({..., 0, s64}, Legal) setAction({..., 1, s32}, Legal) setAction({..., 1, s64}, Legal) currently declares these type combinations as legal: {s32, s32} {s64, s32} {s32, s64} {s64, s64} but we currently have no means to say that, for example, {s64, s32} is not legal. Some operations such as G_INSERT/G_EXTRACT/G_MERGE_VALUES/ G_UNMERGE_VALUES have relationships between the types that are currently described incorrectly. Additionally, G_LOAD/G_STORE currently have no means to legalize non-atomics differently to atomics. The necessary information is in the MMO but we have no way to use this in the legalizer. Similarly, there is currently no way for the register type and the memory type to differ so there is no way to cleanly represent extending-load/truncating-store in a way that can't be broken by optimizers (resulting in illegal MIR). It's also difficult to control the legalization strategy. We've added support for legalizing non-power of 2 types but there's still some hardcoded assumptions about the strategy. The main one I've noticed is that type0 is always legalized before type1 which is not a good strategy for `type0 = G_EXTRACT type1, ...` if you need to widen the container. It will converge on the same result eventually but it will take a much longer route when legalizing type0 than if you legalize type1 first. Lastly, the definition of legality and the legalization strategy is kept separate which is not ideal. It's helpful to be able to look at a one piece of code and see both what is legal and the method the legalizer will use to make illegal MIR more legal. This patch adds a layer onto the LegalizerInfo (to be removed when all targets have been migrated) which resolves all these issues. Here are the rules for shift and division: for (unsigned BinOp : {G_LSHR, G_ASHR, G_SDIV, G_UDIV}) getActionDefinitions(BinOp) .legalFor({s32, s64}) // If type0 is s32/s64 then it's Legal .clampScalar(0, s32, s64) // If type0 is <s32 then WidenScalar to s32 // If type0 is >s64 then NarrowScalar to s64 .widenScalarToPow2(0) // Round type0 scalars up to powers of 2 .unsupported(); // Otherwise, it's unsupported This describes everything needed to both define legality and describe how to make illegal things legal. Here's an example of a complex rule: getActionDefinitions(G_INSERT) .unsupportedIf([=](const LegalityQuery &Query) { // If type0 is smaller than type1 then it's unsupported return Query.Types[0].getSizeInBits() <= Query.Types[1].getSizeInBits(); }) .legalIf([=](const LegalityQuery &Query) { // If type0 is s32/s64/p0 and type1 is a power of 2 other than 2 or 4 then it's legal // We don't need to worry about large type1's because unsupportedIf caught that. const LLT &Ty0 = Query.Types[0]; const LLT &Ty1 = Query.Types[1]; if (Ty0 != s32 && Ty0 != s64 && Ty0 != p0) return false; return isPowerOf2_32(Ty1.getSizeInBits()) && (Ty1.getSizeInBits() == 1 || Ty1.getSizeInBits() >= 8); }) .clampScalar(0, s32, s64) .widenScalarToPow2(0) .maxScalarIf(typeInSet(0, {s32}), 1, s16) // If type0 is s32 and type1 is bigger than s16 then NarrowScalar type1 to s16 .maxScalarIf(typeInSet(0, {s64}), 1, s32) // If type0 is s64 and type1 is bigger than s32 then NarrowScalar type1 to s32 .widenScalarToPow2(1) // Round type1 scalars up to powers of 2 .unsupported(); This uses a lambda to say that G_INSERT is unsupported when type0 is bigger than type1 (in practice, this would be a default rule for G_INSERT). It also uses one to describe the legal cases. This particular predicate is equivalent to: .legalFor({{s32, s1}, {s32, s8}, {s32, s16}, {s64, s1}, {s64, s8}, {s64, s16}, {s64, s32}}) In terms of performance, I saw a slight (~6%) performance improvement when AArch64 was around 30% ported but it's pretty much break even right now. I'm going to take a look at constexpr as a means to reduce the initialization cost. Future work: * Make it possible for opcodes to share rulesets. There's no need for G_LSHR/G_ASHR/G_SDIV/G_UDIV to have separate rule and ruleset objects. There's no technical barrier to this, it just hasn't been done yet. * Replace the type-index numbers with an enum to get .clampScalar(Type0, s32, s64) * Better names for things like .maxScalarIf() (clampMaxScalar?) and the vector rules. * Improve initialization cost using constexpr Possible future work: * It's possible to make these rulesets change the MIR directly instead of returning a description of how to change the MIR. This should remove a little overhead caused by parsing the description and routing to the right code, but the real motivation is that it removes the need for LegalizeAction::Custom. With Custom removed, there's no longer a requirement that Custom legalization change the opcode to something that's considered legal. Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar, volkan, reames, bogner Reviewed By: bogner Subscribers: hintonda, bogner, aemerson, mgorny, javed.absar, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D42251 llvm-svn: 323681
2018-01-30 03:54:49 +08:00
LegalizeActionStep
LegalizerInfo::getAction(const MachineInstr &MI,
const MachineRegisterInfo &MRI) const {
[globalisel] Introduce LegalityQuery to better encapsulate the legalizer decisions. NFC. Summary: `getAction(const InstrAspect &) const` breaks encapsulation by exposing the smaller components that are used to decide how to legalize an instruction. This is a problem because we need to change the implementation of LegalizerInfo so that it's able to describe particular type combinations rather than just cartesian products of types. For example, declaring the following setAction({..., 0, s32}, Legal) setAction({..., 0, s64}, Legal) setAction({..., 1, s32}, Legal) setAction({..., 1, s64}, Legal) currently declares these type combinations as legal: {s32, s32} {s64, s32} {s32, s64} {s64, s64} but we currently have no means to say that, for example, {s64, s32} is not legal. Some operations such as G_INSERT/G_EXTRACT/G_MERGE_VALUES/ G_UNMERGE_VALUES has relationships between the types that are currently described incorrectly. Additionally, G_LOAD/G_STORE currently have no means to legalize non-atomics differently to atomics. The necessary information is in the MMO but we have no way to use this in the legalizer. Similarly, there is currently no way for the register type and the memory type to differ so there is no way to cleanly represent extending-load/truncating-store in a way that can't be broken by optimizers (resulting in illegal MIR). This patch introduces LegalityQuery which provides all the information needed by the legalizer to make a decision on whether something is legal and how to legalize it. Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar, volkan, reames, bogner Reviewed By: bogner Subscribers: bogner, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D42244 llvm-svn: 323342
2018-01-25 01:17:46 +08:00
SmallVector<LLT, 2> Types;
SmallBitVector SeenTypes(8);
[GlobalISel] Enable legalizing non-power-of-2 sized types. This changes the interface of how targets describe how to legalize, see the below description. 1. Interface for targets to describe how to legalize. In GlobalISel, the API in the LegalizerInfo class is the main interface for targets to specify which types are legal for which operations, and what to do to turn illegal type/operation combinations into legal ones. For each operation the type sizes that can be legalized without having to change the size of the type are specified with a call to setAction. This isn't different to how GlobalISel worked before. For example, for a target that supports 32 and 64 bit adds natively: for (auto Ty : {s32, s64}) setAction({G_ADD, 0, s32}, Legal); or for a target that needs a library call for a 32 bit division: setAction({G_SDIV, s32}, Libcall); The main conceptual change to the LegalizerInfo API, is in specifying how to legalize the type sizes for which a change of size is needed. For example, in the above example, how to specify how all types from i1 to i8388607 (apart from s32 and s64 which are legal) need to be legalized and expressed in terms of operations on the available legal sizes (again, i32 and i64 in this case). Before, the implementation only allowed specifying power-of-2-sized types (e.g. setAction({G_ADD, 0, s128}, NarrowScalar). A worse limitation was that if you'd wanted to specify how to legalize all the sized types as allowed by the LLVM-IR LangRef, i1 to i8388607, you'd have to call setAction 8388607-3 times and probably would need a lot of memory to store all of these specifications. Instead, the legalization actions that need to change the size of the type are specified now using a "SizeChangeStrategy". For example: setLegalizeScalarToDifferentSizeStrategy( G_ADD, 0, widenToLargerAndNarrowToLargest); This example indicates that for type sizes for which there is a larger size that can be legalized towards, do it by Widening the size. For example, G_ADD on s17 will be legalized by first doing WidenScalar to make it s32, after which it's legal. The "NarrowToLargest" indicates what to do if there is no larger size that can be legalized towards. E.g. G_ADD on s92 will be legalized by doing NarrowScalar to s64. Another example, taken from the ARM backend is: for (unsigned Op : {G_SDIV, G_UDIV}) { setLegalizeScalarToDifferentSizeStrategy(Op, 0, widenToLargerTypesUnsupportedOtherwise); if (ST.hasDivideInARMMode()) setAction({Op, s32}, Legal); else setAction({Op, s32}, Libcall); } For this example, G_SDIV on s8, on a target without a divide instruction, would be legalized by first doing action (WidenScalar, s32), followed by (Libcall, s32). The same principle is also followed for when the number of vector lanes on vector data types need to be changed, e.g.: setAction({G_ADD, LLT::vector(8, 8)}, LegalizerInfo::Legal); setAction({G_ADD, LLT::vector(16, 8)}, LegalizerInfo::Legal); setAction({G_ADD, LLT::vector(4, 16)}, LegalizerInfo::Legal); setAction({G_ADD, LLT::vector(8, 16)}, LegalizerInfo::Legal); setAction({G_ADD, LLT::vector(2, 32)}, LegalizerInfo::Legal); setAction({G_ADD, LLT::vector(4, 32)}, LegalizerInfo::Legal); setLegalizeVectorElementToDifferentSizeStrategy( G_ADD, 0, widenToLargerTypesUnsupportedOtherwise); As currently implemented here, vector types are legalized by first making the vector element size legal, followed by then making the number of lanes legal. The strategy to follow in the first step is set by a call to setLegalizeVectorElementToDifferentSizeStrategy, see example above. The strategy followed in the second step "moreToWiderTypesAndLessToWidest" (see code for its definition), indicating that vectors are widened to more elements so they map to natively supported vector widths, or when there isn't a legal wider vector, split the vector to map it to the widest vector supported. Therefore, for the above specification, some example legalizations are: * getAction({G_ADD, LLT::vector(3, 3)}) returns {WidenScalar, LLT::vector(3, 8)} * getAction({G_ADD, LLT::vector(3, 8)}) then returns {MoreElements, LLT::vector(8, 8)} * getAction({G_ADD, LLT::vector(20, 8)}) returns {FewerElements, LLT::vector(16, 8)} 2. Key implementation aspects. How to legalize a specific (operation, type index, size) tuple is represented by mapping intervals of integers representing a range of size types to an action to take, e.g.: setScalarAction({G_ADD, LLT:scalar(1)}, {{1, WidenScalar}, // bit sizes [ 1, 31[ {32, Legal}, // bit sizes [32, 33[ {33, WidenScalar}, // bit sizes [33, 64[ {64, Legal}, // bit sizes [64, 65[ {65, NarrowScalar} // bit sizes [65, +inf[ }); Please note that most of the code to do the actual lowering of non-power-of-2 sized types is currently missing, this is just trying to make it possible for targets to specify what is legal, and how non-legal types should be legalized. Probably quite a bit of further work is needed in the actual legalizing and the other passes in GlobalISel to support non-power-of-2 sized types. I hope the documentation in LegalizerInfo.h and the examples provided in the various {Target}LegalizerInfo.cpp and LegalizerInfoTest.cpp explains well enough how this is meant to be used. This drops the need for LLT::{half,double}...Size(). Differential Revision: https://reviews.llvm.org/D30529 llvm-svn: 317560
2017-11-07 18:34:34 +08:00
const MCOperandInfo *OpInfo = MI.getDesc().OpInfo;
// FIXME: probably we'll need to cache the results here somehow?
for (unsigned i = 0; i < MI.getDesc().getNumOperands(); ++i) {
if (!OpInfo[i].isGenericType())
continue;
[GlobalISel] Enable legalizing non-power-of-2 sized types. This changes the interface of how targets describe how to legalize, see the below description. 1. Interface for targets to describe how to legalize. In GlobalISel, the API in the LegalizerInfo class is the main interface for targets to specify which types are legal for which operations, and what to do to turn illegal type/operation combinations into legal ones. For each operation the type sizes that can be legalized without having to change the size of the type are specified with a call to setAction. This isn't different to how GlobalISel worked before. For example, for a target that supports 32 and 64 bit adds natively: for (auto Ty : {s32, s64}) setAction({G_ADD, 0, s32}, Legal); or for a target that needs a library call for a 32 bit division: setAction({G_SDIV, s32}, Libcall); The main conceptual change to the LegalizerInfo API, is in specifying how to legalize the type sizes for which a change of size is needed. For example, in the above example, how to specify how all types from i1 to i8388607 (apart from s32 and s64 which are legal) need to be legalized and expressed in terms of operations on the available legal sizes (again, i32 and i64 in this case). Before, the implementation only allowed specifying power-of-2-sized types (e.g. setAction({G_ADD, 0, s128}, NarrowScalar). A worse limitation was that if you'd wanted to specify how to legalize all the sized types as allowed by the LLVM-IR LangRef, i1 to i8388607, you'd have to call setAction 8388607-3 times and probably would need a lot of memory to store all of these specifications. Instead, the legalization actions that need to change the size of the type are specified now using a "SizeChangeStrategy". For example: setLegalizeScalarToDifferentSizeStrategy( G_ADD, 0, widenToLargerAndNarrowToLargest); This example indicates that for type sizes for which there is a larger size that can be legalized towards, do it by Widening the size. For example, G_ADD on s17 will be legalized by first doing WidenScalar to make it s32, after which it's legal. The "NarrowToLargest" indicates what to do if there is no larger size that can be legalized towards. E.g. G_ADD on s92 will be legalized by doing NarrowScalar to s64. Another example, taken from the ARM backend is: for (unsigned Op : {G_SDIV, G_UDIV}) { setLegalizeScalarToDifferentSizeStrategy(Op, 0, widenToLargerTypesUnsupportedOtherwise); if (ST.hasDivideInARMMode()) setAction({Op, s32}, Legal); else setAction({Op, s32}, Libcall); } For this example, G_SDIV on s8, on a target without a divide instruction, would be legalized by first doing action (WidenScalar, s32), followed by (Libcall, s32). The same principle is also followed for when the number of vector lanes on vector data types need to be changed, e.g.: setAction({G_ADD, LLT::vector(8, 8)}, LegalizerInfo::Legal); setAction({G_ADD, LLT::vector(16, 8)}, LegalizerInfo::Legal); setAction({G_ADD, LLT::vector(4, 16)}, LegalizerInfo::Legal); setAction({G_ADD, LLT::vector(8, 16)}, LegalizerInfo::Legal); setAction({G_ADD, LLT::vector(2, 32)}, LegalizerInfo::Legal); setAction({G_ADD, LLT::vector(4, 32)}, LegalizerInfo::Legal); setLegalizeVectorElementToDifferentSizeStrategy( G_ADD, 0, widenToLargerTypesUnsupportedOtherwise); As currently implemented here, vector types are legalized by first making the vector element size legal, followed by then making the number of lanes legal. The strategy to follow in the first step is set by a call to setLegalizeVectorElementToDifferentSizeStrategy, see example above. The strategy followed in the second step "moreToWiderTypesAndLessToWidest" (see code for its definition), indicating that vectors are widened to more elements so they map to natively supported vector widths, or when there isn't a legal wider vector, split the vector to map it to the widest vector supported. Therefore, for the above specification, some example legalizations are: * getAction({G_ADD, LLT::vector(3, 3)}) returns {WidenScalar, LLT::vector(3, 8)} * getAction({G_ADD, LLT::vector(3, 8)}) then returns {MoreElements, LLT::vector(8, 8)} * getAction({G_ADD, LLT::vector(20, 8)}) returns {FewerElements, LLT::vector(16, 8)} 2. Key implementation aspects. How to legalize a specific (operation, type index, size) tuple is represented by mapping intervals of integers representing a range of size types to an action to take, e.g.: setScalarAction({G_ADD, LLT:scalar(1)}, {{1, WidenScalar}, // bit sizes [ 1, 31[ {32, Legal}, // bit sizes [32, 33[ {33, WidenScalar}, // bit sizes [33, 64[ {64, Legal}, // bit sizes [64, 65[ {65, NarrowScalar} // bit sizes [65, +inf[ }); Please note that most of the code to do the actual lowering of non-power-of-2 sized types is currently missing, this is just trying to make it possible for targets to specify what is legal, and how non-legal types should be legalized. Probably quite a bit of further work is needed in the actual legalizing and the other passes in GlobalISel to support non-power-of-2 sized types. I hope the documentation in LegalizerInfo.h and the examples provided in the various {Target}LegalizerInfo.cpp and LegalizerInfoTest.cpp explains well enough how this is meant to be used. This drops the need for LLT::{half,double}...Size(). Differential Revision: https://reviews.llvm.org/D30529 llvm-svn: 317560
2017-11-07 18:34:34 +08:00
// We must only record actions once for each TypeIdx; otherwise we'd
// try to legalize operands multiple times down the line.
unsigned TypeIdx = OpInfo[i].getGenericTypeIndex();
if (SeenTypes[TypeIdx])
continue;
SeenTypes.set(TypeIdx);
LLT Ty = getTypeFromTypeIdx(MI, MRI, i, TypeIdx);
[globalisel] Introduce LegalityQuery to better encapsulate the legalizer decisions. NFC. Summary: `getAction(const InstrAspect &) const` breaks encapsulation by exposing the smaller components that are used to decide how to legalize an instruction. This is a problem because we need to change the implementation of LegalizerInfo so that it's able to describe particular type combinations rather than just cartesian products of types. For example, declaring the following setAction({..., 0, s32}, Legal) setAction({..., 0, s64}, Legal) setAction({..., 1, s32}, Legal) setAction({..., 1, s64}, Legal) currently declares these type combinations as legal: {s32, s32} {s64, s32} {s32, s64} {s64, s64} but we currently have no means to say that, for example, {s64, s32} is not legal. Some operations such as G_INSERT/G_EXTRACT/G_MERGE_VALUES/ G_UNMERGE_VALUES has relationships between the types that are currently described incorrectly. Additionally, G_LOAD/G_STORE currently have no means to legalize non-atomics differently to atomics. The necessary information is in the MMO but we have no way to use this in the legalizer. Similarly, there is currently no way for the register type and the memory type to differ so there is no way to cleanly represent extending-load/truncating-store in a way that can't be broken by optimizers (resulting in illegal MIR). This patch introduces LegalityQuery which provides all the information needed by the legalizer to make a decision on whether something is legal and how to legalize it. Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar, volkan, reames, bogner Reviewed By: bogner Subscribers: bogner, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D42244 llvm-svn: 323342
2018-01-25 01:17:46 +08:00
Types.push_back(Ty);
}
[globalisel] Introduce LegalityQuery to better encapsulate the legalizer decisions. NFC. Summary: `getAction(const InstrAspect &) const` breaks encapsulation by exposing the smaller components that are used to decide how to legalize an instruction. This is a problem because we need to change the implementation of LegalizerInfo so that it's able to describe particular type combinations rather than just cartesian products of types. For example, declaring the following setAction({..., 0, s32}, Legal) setAction({..., 0, s64}, Legal) setAction({..., 1, s32}, Legal) setAction({..., 1, s64}, Legal) currently declares these type combinations as legal: {s32, s32} {s64, s32} {s32, s64} {s64, s64} but we currently have no means to say that, for example, {s64, s32} is not legal. Some operations such as G_INSERT/G_EXTRACT/G_MERGE_VALUES/ G_UNMERGE_VALUES has relationships between the types that are currently described incorrectly. Additionally, G_LOAD/G_STORE currently have no means to legalize non-atomics differently to atomics. The necessary information is in the MMO but we have no way to use this in the legalizer. Similarly, there is currently no way for the register type and the memory type to differ so there is no way to cleanly represent extending-load/truncating-store in a way that can't be broken by optimizers (resulting in illegal MIR). This patch introduces LegalityQuery which provides all the information needed by the legalizer to make a decision on whether something is legal and how to legalize it. Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar, volkan, reames, bogner Reviewed By: bogner Subscribers: bogner, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D42244 llvm-svn: 323342
2018-01-25 01:17:46 +08:00
return getAction({MI.getOpcode(), Types});
}
bool LegalizerInfo::isLegal(const MachineInstr &MI,
const MachineRegisterInfo &MRI) const {
[globalisel] Introduce LegalityQuery to better encapsulate the legalizer decisions. NFC. Summary: `getAction(const InstrAspect &) const` breaks encapsulation by exposing the smaller components that are used to decide how to legalize an instruction. This is a problem because we need to change the implementation of LegalizerInfo so that it's able to describe particular type combinations rather than just cartesian products of types. For example, declaring the following setAction({..., 0, s32}, Legal) setAction({..., 0, s64}, Legal) setAction({..., 1, s32}, Legal) setAction({..., 1, s64}, Legal) currently declares these type combinations as legal: {s32, s32} {s64, s32} {s32, s64} {s64, s64} but we currently have no means to say that, for example, {s64, s32} is not legal. Some operations such as G_INSERT/G_EXTRACT/G_MERGE_VALUES/ G_UNMERGE_VALUES has relationships between the types that are currently described incorrectly. Additionally, G_LOAD/G_STORE currently have no means to legalize non-atomics differently to atomics. The necessary information is in the MMO but we have no way to use this in the legalizer. Similarly, there is currently no way for the register type and the memory type to differ so there is no way to cleanly represent extending-load/truncating-store in a way that can't be broken by optimizers (resulting in illegal MIR). This patch introduces LegalityQuery which provides all the information needed by the legalizer to make a decision on whether something is legal and how to legalize it. Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar, volkan, reames, bogner Reviewed By: bogner Subscribers: bogner, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D42244 llvm-svn: 323342
2018-01-25 01:17:46 +08:00
return getAction(MI, MRI).Action == Legal;
}
[GlobalISel] Enable legalizing non-power-of-2 sized types. This changes the interface of how targets describe how to legalize, see the below description. 1. Interface for targets to describe how to legalize. In GlobalISel, the API in the LegalizerInfo class is the main interface for targets to specify which types are legal for which operations, and what to do to turn illegal type/operation combinations into legal ones. For each operation the type sizes that can be legalized without having to change the size of the type are specified with a call to setAction. This isn't different to how GlobalISel worked before. For example, for a target that supports 32 and 64 bit adds natively: for (auto Ty : {s32, s64}) setAction({G_ADD, 0, s32}, Legal); or for a target that needs a library call for a 32 bit division: setAction({G_SDIV, s32}, Libcall); The main conceptual change to the LegalizerInfo API, is in specifying how to legalize the type sizes for which a change of size is needed. For example, in the above example, how to specify how all types from i1 to i8388607 (apart from s32 and s64 which are legal) need to be legalized and expressed in terms of operations on the available legal sizes (again, i32 and i64 in this case). Before, the implementation only allowed specifying power-of-2-sized types (e.g. setAction({G_ADD, 0, s128}, NarrowScalar). A worse limitation was that if you'd wanted to specify how to legalize all the sized types as allowed by the LLVM-IR LangRef, i1 to i8388607, you'd have to call setAction 8388607-3 times and probably would need a lot of memory to store all of these specifications. Instead, the legalization actions that need to change the size of the type are specified now using a "SizeChangeStrategy". For example: setLegalizeScalarToDifferentSizeStrategy( G_ADD, 0, widenToLargerAndNarrowToLargest); This example indicates that for type sizes for which there is a larger size that can be legalized towards, do it by Widening the size. For example, G_ADD on s17 will be legalized by first doing WidenScalar to make it s32, after which it's legal. The "NarrowToLargest" indicates what to do if there is no larger size that can be legalized towards. E.g. G_ADD on s92 will be legalized by doing NarrowScalar to s64. Another example, taken from the ARM backend is: for (unsigned Op : {G_SDIV, G_UDIV}) { setLegalizeScalarToDifferentSizeStrategy(Op, 0, widenToLargerTypesUnsupportedOtherwise); if (ST.hasDivideInARMMode()) setAction({Op, s32}, Legal); else setAction({Op, s32}, Libcall); } For this example, G_SDIV on s8, on a target without a divide instruction, would be legalized by first doing action (WidenScalar, s32), followed by (Libcall, s32). The same principle is also followed for when the number of vector lanes on vector data types need to be changed, e.g.: setAction({G_ADD, LLT::vector(8, 8)}, LegalizerInfo::Legal); setAction({G_ADD, LLT::vector(16, 8)}, LegalizerInfo::Legal); setAction({G_ADD, LLT::vector(4, 16)}, LegalizerInfo::Legal); setAction({G_ADD, LLT::vector(8, 16)}, LegalizerInfo::Legal); setAction({G_ADD, LLT::vector(2, 32)}, LegalizerInfo::Legal); setAction({G_ADD, LLT::vector(4, 32)}, LegalizerInfo::Legal); setLegalizeVectorElementToDifferentSizeStrategy( G_ADD, 0, widenToLargerTypesUnsupportedOtherwise); As currently implemented here, vector types are legalized by first making the vector element size legal, followed by then making the number of lanes legal. The strategy to follow in the first step is set by a call to setLegalizeVectorElementToDifferentSizeStrategy, see example above. The strategy followed in the second step "moreToWiderTypesAndLessToWidest" (see code for its definition), indicating that vectors are widened to more elements so they map to natively supported vector widths, or when there isn't a legal wider vector, split the vector to map it to the widest vector supported. Therefore, for the above specification, some example legalizations are: * getAction({G_ADD, LLT::vector(3, 3)}) returns {WidenScalar, LLT::vector(3, 8)} * getAction({G_ADD, LLT::vector(3, 8)}) then returns {MoreElements, LLT::vector(8, 8)} * getAction({G_ADD, LLT::vector(20, 8)}) returns {FewerElements, LLT::vector(16, 8)} 2. Key implementation aspects. How to legalize a specific (operation, type index, size) tuple is represented by mapping intervals of integers representing a range of size types to an action to take, e.g.: setScalarAction({G_ADD, LLT:scalar(1)}, {{1, WidenScalar}, // bit sizes [ 1, 31[ {32, Legal}, // bit sizes [32, 33[ {33, WidenScalar}, // bit sizes [33, 64[ {64, Legal}, // bit sizes [64, 65[ {65, NarrowScalar} // bit sizes [65, +inf[ }); Please note that most of the code to do the actual lowering of non-power-of-2 sized types is currently missing, this is just trying to make it possible for targets to specify what is legal, and how non-legal types should be legalized. Probably quite a bit of further work is needed in the actual legalizing and the other passes in GlobalISel to support non-power-of-2 sized types. I hope the documentation in LegalizerInfo.h and the examples provided in the various {Target}LegalizerInfo.cpp and LegalizerInfoTest.cpp explains well enough how this is meant to be used. This drops the need for LLT::{half,double}...Size(). Differential Revision: https://reviews.llvm.org/D30529 llvm-svn: 317560
2017-11-07 18:34:34 +08:00
bool LegalizerInfo::legalizeCustom(MachineInstr &MI, MachineRegisterInfo &MRI,
MachineIRBuilder &MIRBuilder) const {
return false;
}
LegalizerInfo::SizeAndActionsVec
LegalizerInfo::increaseToLargerTypesAndDecreaseToLargest(
const SizeAndActionsVec &v, LegalizeAction IncreaseAction,
LegalizeAction DecreaseAction) {
SizeAndActionsVec result;
unsigned LargestSizeSoFar = 0;
if (v.size() >= 1 && v[0].first != 1)
result.push_back({1, IncreaseAction});
for (size_t i = 0; i < v.size(); ++i) {
result.push_back(v[i]);
LargestSizeSoFar = v[i].first;
if (i + 1 < v.size() && v[i + 1].first != v[i].first + 1) {
result.push_back({LargestSizeSoFar + 1, IncreaseAction});
LargestSizeSoFar = v[i].first + 1;
}
}
result.push_back({LargestSizeSoFar + 1, DecreaseAction});
return result;
}
LegalizerInfo::SizeAndActionsVec
LegalizerInfo::decreaseToSmallerTypesAndIncreaseToSmallest(
const SizeAndActionsVec &v, LegalizeAction DecreaseAction,
LegalizeAction IncreaseAction) {
SizeAndActionsVec result;
if (v.size() == 0 || v[0].first != 1)
result.push_back({1, IncreaseAction});
for (size_t i = 0; i < v.size(); ++i) {
result.push_back(v[i]);
if (i + 1 == v.size() || v[i + 1].first != v[i].first + 1) {
result.push_back({v[i].first + 1, DecreaseAction});
}
}
return result;
}
LegalizerInfo::SizeAndAction
LegalizerInfo::findAction(const SizeAndActionsVec &Vec, const uint32_t Size) {
assert(Size >= 1);
// Find the last element in Vec that has a bitsize equal to or smaller than
// the requested bit size.
// That is the element just before the first element that is bigger than Size.
auto VecIt = std::upper_bound(
Vec.begin(), Vec.end(), Size,
[](const uint32_t Size, const SizeAndAction lhs) -> bool {
return Size < lhs.first;
});
assert(VecIt != Vec.begin() && "Does Vec not start with size 1?");
--VecIt;
int VecIdx = VecIt - Vec.begin();
LegalizeAction Action = Vec[VecIdx].second;
switch (Action) {
case Legal:
case Lower:
case Libcall:
case Custom:
[GlobalISel] Enable legalizing non-power-of-2 sized types. This changes the interface of how targets describe how to legalize, see the below description. 1. Interface for targets to describe how to legalize. In GlobalISel, the API in the LegalizerInfo class is the main interface for targets to specify which types are legal for which operations, and what to do to turn illegal type/operation combinations into legal ones. For each operation the type sizes that can be legalized without having to change the size of the type are specified with a call to setAction. This isn't different to how GlobalISel worked before. For example, for a target that supports 32 and 64 bit adds natively: for (auto Ty : {s32, s64}) setAction({G_ADD, 0, s32}, Legal); or for a target that needs a library call for a 32 bit division: setAction({G_SDIV, s32}, Libcall); The main conceptual change to the LegalizerInfo API, is in specifying how to legalize the type sizes for which a change of size is needed. For example, in the above example, how to specify how all types from i1 to i8388607 (apart from s32 and s64 which are legal) need to be legalized and expressed in terms of operations on the available legal sizes (again, i32 and i64 in this case). Before, the implementation only allowed specifying power-of-2-sized types (e.g. setAction({G_ADD, 0, s128}, NarrowScalar). A worse limitation was that if you'd wanted to specify how to legalize all the sized types as allowed by the LLVM-IR LangRef, i1 to i8388607, you'd have to call setAction 8388607-3 times and probably would need a lot of memory to store all of these specifications. Instead, the legalization actions that need to change the size of the type are specified now using a "SizeChangeStrategy". For example: setLegalizeScalarToDifferentSizeStrategy( G_ADD, 0, widenToLargerAndNarrowToLargest); This example indicates that for type sizes for which there is a larger size that can be legalized towards, do it by Widening the size. For example, G_ADD on s17 will be legalized by first doing WidenScalar to make it s32, after which it's legal. The "NarrowToLargest" indicates what to do if there is no larger size that can be legalized towards. E.g. G_ADD on s92 will be legalized by doing NarrowScalar to s64. Another example, taken from the ARM backend is: for (unsigned Op : {G_SDIV, G_UDIV}) { setLegalizeScalarToDifferentSizeStrategy(Op, 0, widenToLargerTypesUnsupportedOtherwise); if (ST.hasDivideInARMMode()) setAction({Op, s32}, Legal); else setAction({Op, s32}, Libcall); } For this example, G_SDIV on s8, on a target without a divide instruction, would be legalized by first doing action (WidenScalar, s32), followed by (Libcall, s32). The same principle is also followed for when the number of vector lanes on vector data types need to be changed, e.g.: setAction({G_ADD, LLT::vector(8, 8)}, LegalizerInfo::Legal); setAction({G_ADD, LLT::vector(16, 8)}, LegalizerInfo::Legal); setAction({G_ADD, LLT::vector(4, 16)}, LegalizerInfo::Legal); setAction({G_ADD, LLT::vector(8, 16)}, LegalizerInfo::Legal); setAction({G_ADD, LLT::vector(2, 32)}, LegalizerInfo::Legal); setAction({G_ADD, LLT::vector(4, 32)}, LegalizerInfo::Legal); setLegalizeVectorElementToDifferentSizeStrategy( G_ADD, 0, widenToLargerTypesUnsupportedOtherwise); As currently implemented here, vector types are legalized by first making the vector element size legal, followed by then making the number of lanes legal. The strategy to follow in the first step is set by a call to setLegalizeVectorElementToDifferentSizeStrategy, see example above. The strategy followed in the second step "moreToWiderTypesAndLessToWidest" (see code for its definition), indicating that vectors are widened to more elements so they map to natively supported vector widths, or when there isn't a legal wider vector, split the vector to map it to the widest vector supported. Therefore, for the above specification, some example legalizations are: * getAction({G_ADD, LLT::vector(3, 3)}) returns {WidenScalar, LLT::vector(3, 8)} * getAction({G_ADD, LLT::vector(3, 8)}) then returns {MoreElements, LLT::vector(8, 8)} * getAction({G_ADD, LLT::vector(20, 8)}) returns {FewerElements, LLT::vector(16, 8)} 2. Key implementation aspects. How to legalize a specific (operation, type index, size) tuple is represented by mapping intervals of integers representing a range of size types to an action to take, e.g.: setScalarAction({G_ADD, LLT:scalar(1)}, {{1, WidenScalar}, // bit sizes [ 1, 31[ {32, Legal}, // bit sizes [32, 33[ {33, WidenScalar}, // bit sizes [33, 64[ {64, Legal}, // bit sizes [64, 65[ {65, NarrowScalar} // bit sizes [65, +inf[ }); Please note that most of the code to do the actual lowering of non-power-of-2 sized types is currently missing, this is just trying to make it possible for targets to specify what is legal, and how non-legal types should be legalized. Probably quite a bit of further work is needed in the actual legalizing and the other passes in GlobalISel to support non-power-of-2 sized types. I hope the documentation in LegalizerInfo.h and the examples provided in the various {Target}LegalizerInfo.cpp and LegalizerInfoTest.cpp explains well enough how this is meant to be used. This drops the need for LLT::{half,double}...Size(). Differential Revision: https://reviews.llvm.org/D30529 llvm-svn: 317560
2017-11-07 18:34:34 +08:00
return {Size, Action};
case FewerElements:
// FIXME: is this special case still needed and correct?
// Special case for scalarization:
if (Vec == SizeAndActionsVec({{1, FewerElements}}))
return {1, FewerElements};
LLVM_FALLTHROUGH;
case NarrowScalar: {
[GlobalISel] Enable legalizing non-power-of-2 sized types. This changes the interface of how targets describe how to legalize, see the below description. 1. Interface for targets to describe how to legalize. In GlobalISel, the API in the LegalizerInfo class is the main interface for targets to specify which types are legal for which operations, and what to do to turn illegal type/operation combinations into legal ones. For each operation the type sizes that can be legalized without having to change the size of the type are specified with a call to setAction. This isn't different to how GlobalISel worked before. For example, for a target that supports 32 and 64 bit adds natively: for (auto Ty : {s32, s64}) setAction({G_ADD, 0, s32}, Legal); or for a target that needs a library call for a 32 bit division: setAction({G_SDIV, s32}, Libcall); The main conceptual change to the LegalizerInfo API, is in specifying how to legalize the type sizes for which a change of size is needed. For example, in the above example, how to specify how all types from i1 to i8388607 (apart from s32 and s64 which are legal) need to be legalized and expressed in terms of operations on the available legal sizes (again, i32 and i64 in this case). Before, the implementation only allowed specifying power-of-2-sized types (e.g. setAction({G_ADD, 0, s128}, NarrowScalar). A worse limitation was that if you'd wanted to specify how to legalize all the sized types as allowed by the LLVM-IR LangRef, i1 to i8388607, you'd have to call setAction 8388607-3 times and probably would need a lot of memory to store all of these specifications. Instead, the legalization actions that need to change the size of the type are specified now using a "SizeChangeStrategy". For example: setLegalizeScalarToDifferentSizeStrategy( G_ADD, 0, widenToLargerAndNarrowToLargest); This example indicates that for type sizes for which there is a larger size that can be legalized towards, do it by Widening the size. For example, G_ADD on s17 will be legalized by first doing WidenScalar to make it s32, after which it's legal. The "NarrowToLargest" indicates what to do if there is no larger size that can be legalized towards. E.g. G_ADD on s92 will be legalized by doing NarrowScalar to s64. Another example, taken from the ARM backend is: for (unsigned Op : {G_SDIV, G_UDIV}) { setLegalizeScalarToDifferentSizeStrategy(Op, 0, widenToLargerTypesUnsupportedOtherwise); if (ST.hasDivideInARMMode()) setAction({Op, s32}, Legal); else setAction({Op, s32}, Libcall); } For this example, G_SDIV on s8, on a target without a divide instruction, would be legalized by first doing action (WidenScalar, s32), followed by (Libcall, s32). The same principle is also followed for when the number of vector lanes on vector data types need to be changed, e.g.: setAction({G_ADD, LLT::vector(8, 8)}, LegalizerInfo::Legal); setAction({G_ADD, LLT::vector(16, 8)}, LegalizerInfo::Legal); setAction({G_ADD, LLT::vector(4, 16)}, LegalizerInfo::Legal); setAction({G_ADD, LLT::vector(8, 16)}, LegalizerInfo::Legal); setAction({G_ADD, LLT::vector(2, 32)}, LegalizerInfo::Legal); setAction({G_ADD, LLT::vector(4, 32)}, LegalizerInfo::Legal); setLegalizeVectorElementToDifferentSizeStrategy( G_ADD, 0, widenToLargerTypesUnsupportedOtherwise); As currently implemented here, vector types are legalized by first making the vector element size legal, followed by then making the number of lanes legal. The strategy to follow in the first step is set by a call to setLegalizeVectorElementToDifferentSizeStrategy, see example above. The strategy followed in the second step "moreToWiderTypesAndLessToWidest" (see code for its definition), indicating that vectors are widened to more elements so they map to natively supported vector widths, or when there isn't a legal wider vector, split the vector to map it to the widest vector supported. Therefore, for the above specification, some example legalizations are: * getAction({G_ADD, LLT::vector(3, 3)}) returns {WidenScalar, LLT::vector(3, 8)} * getAction({G_ADD, LLT::vector(3, 8)}) then returns {MoreElements, LLT::vector(8, 8)} * getAction({G_ADD, LLT::vector(20, 8)}) returns {FewerElements, LLT::vector(16, 8)} 2. Key implementation aspects. How to legalize a specific (operation, type index, size) tuple is represented by mapping intervals of integers representing a range of size types to an action to take, e.g.: setScalarAction({G_ADD, LLT:scalar(1)}, {{1, WidenScalar}, // bit sizes [ 1, 31[ {32, Legal}, // bit sizes [32, 33[ {33, WidenScalar}, // bit sizes [33, 64[ {64, Legal}, // bit sizes [64, 65[ {65, NarrowScalar} // bit sizes [65, +inf[ }); Please note that most of the code to do the actual lowering of non-power-of-2 sized types is currently missing, this is just trying to make it possible for targets to specify what is legal, and how non-legal types should be legalized. Probably quite a bit of further work is needed in the actual legalizing and the other passes in GlobalISel to support non-power-of-2 sized types. I hope the documentation in LegalizerInfo.h and the examples provided in the various {Target}LegalizerInfo.cpp and LegalizerInfoTest.cpp explains well enough how this is meant to be used. This drops the need for LLT::{half,double}...Size(). Differential Revision: https://reviews.llvm.org/D30529 llvm-svn: 317560
2017-11-07 18:34:34 +08:00
// The following needs to be a loop, as for now, we do allow needing to
// go over "Unsupported" bit sizes before finding a legalizable bit size.
// e.g. (s8, WidenScalar), (s9, Unsupported), (s32, Legal). if Size==8,
// we need to iterate over s9, and then to s32 to return (s32, Legal).
// If we want to get rid of the below loop, we should have stronger asserts
// when building the SizeAndActionsVecs, probably not allowing
// "Unsupported" unless at the ends of the vector.
for (int i = VecIdx - 1; i >= 0; --i)
if (!needsLegalizingToDifferentSize(Vec[i].second) &&
Vec[i].second != Unsupported)
return {Vec[i].first, Action};
llvm_unreachable("");
}
[GlobalISel] Enable legalizing non-power-of-2 sized types. This changes the interface of how targets describe how to legalize, see the below description. 1. Interface for targets to describe how to legalize. In GlobalISel, the API in the LegalizerInfo class is the main interface for targets to specify which types are legal for which operations, and what to do to turn illegal type/operation combinations into legal ones. For each operation the type sizes that can be legalized without having to change the size of the type are specified with a call to setAction. This isn't different to how GlobalISel worked before. For example, for a target that supports 32 and 64 bit adds natively: for (auto Ty : {s32, s64}) setAction({G_ADD, 0, s32}, Legal); or for a target that needs a library call for a 32 bit division: setAction({G_SDIV, s32}, Libcall); The main conceptual change to the LegalizerInfo API, is in specifying how to legalize the type sizes for which a change of size is needed. For example, in the above example, how to specify how all types from i1 to i8388607 (apart from s32 and s64 which are legal) need to be legalized and expressed in terms of operations on the available legal sizes (again, i32 and i64 in this case). Before, the implementation only allowed specifying power-of-2-sized types (e.g. setAction({G_ADD, 0, s128}, NarrowScalar). A worse limitation was that if you'd wanted to specify how to legalize all the sized types as allowed by the LLVM-IR LangRef, i1 to i8388607, you'd have to call setAction 8388607-3 times and probably would need a lot of memory to store all of these specifications. Instead, the legalization actions that need to change the size of the type are specified now using a "SizeChangeStrategy". For example: setLegalizeScalarToDifferentSizeStrategy( G_ADD, 0, widenToLargerAndNarrowToLargest); This example indicates that for type sizes for which there is a larger size that can be legalized towards, do it by Widening the size. For example, G_ADD on s17 will be legalized by first doing WidenScalar to make it s32, after which it's legal. The "NarrowToLargest" indicates what to do if there is no larger size that can be legalized towards. E.g. G_ADD on s92 will be legalized by doing NarrowScalar to s64. Another example, taken from the ARM backend is: for (unsigned Op : {G_SDIV, G_UDIV}) { setLegalizeScalarToDifferentSizeStrategy(Op, 0, widenToLargerTypesUnsupportedOtherwise); if (ST.hasDivideInARMMode()) setAction({Op, s32}, Legal); else setAction({Op, s32}, Libcall); } For this example, G_SDIV on s8, on a target without a divide instruction, would be legalized by first doing action (WidenScalar, s32), followed by (Libcall, s32). The same principle is also followed for when the number of vector lanes on vector data types need to be changed, e.g.: setAction({G_ADD, LLT::vector(8, 8)}, LegalizerInfo::Legal); setAction({G_ADD, LLT::vector(16, 8)}, LegalizerInfo::Legal); setAction({G_ADD, LLT::vector(4, 16)}, LegalizerInfo::Legal); setAction({G_ADD, LLT::vector(8, 16)}, LegalizerInfo::Legal); setAction({G_ADD, LLT::vector(2, 32)}, LegalizerInfo::Legal); setAction({G_ADD, LLT::vector(4, 32)}, LegalizerInfo::Legal); setLegalizeVectorElementToDifferentSizeStrategy( G_ADD, 0, widenToLargerTypesUnsupportedOtherwise); As currently implemented here, vector types are legalized by first making the vector element size legal, followed by then making the number of lanes legal. The strategy to follow in the first step is set by a call to setLegalizeVectorElementToDifferentSizeStrategy, see example above. The strategy followed in the second step "moreToWiderTypesAndLessToWidest" (see code for its definition), indicating that vectors are widened to more elements so they map to natively supported vector widths, or when there isn't a legal wider vector, split the vector to map it to the widest vector supported. Therefore, for the above specification, some example legalizations are: * getAction({G_ADD, LLT::vector(3, 3)}) returns {WidenScalar, LLT::vector(3, 8)} * getAction({G_ADD, LLT::vector(3, 8)}) then returns {MoreElements, LLT::vector(8, 8)} * getAction({G_ADD, LLT::vector(20, 8)}) returns {FewerElements, LLT::vector(16, 8)} 2. Key implementation aspects. How to legalize a specific (operation, type index, size) tuple is represented by mapping intervals of integers representing a range of size types to an action to take, e.g.: setScalarAction({G_ADD, LLT:scalar(1)}, {{1, WidenScalar}, // bit sizes [ 1, 31[ {32, Legal}, // bit sizes [32, 33[ {33, WidenScalar}, // bit sizes [33, 64[ {64, Legal}, // bit sizes [64, 65[ {65, NarrowScalar} // bit sizes [65, +inf[ }); Please note that most of the code to do the actual lowering of non-power-of-2 sized types is currently missing, this is just trying to make it possible for targets to specify what is legal, and how non-legal types should be legalized. Probably quite a bit of further work is needed in the actual legalizing and the other passes in GlobalISel to support non-power-of-2 sized types. I hope the documentation in LegalizerInfo.h and the examples provided in the various {Target}LegalizerInfo.cpp and LegalizerInfoTest.cpp explains well enough how this is meant to be used. This drops the need for LLT::{half,double}...Size(). Differential Revision: https://reviews.llvm.org/D30529 llvm-svn: 317560
2017-11-07 18:34:34 +08:00
case WidenScalar:
case MoreElements: {
[GlobalISel] Enable legalizing non-power-of-2 sized types. This changes the interface of how targets describe how to legalize, see the below description. 1. Interface for targets to describe how to legalize. In GlobalISel, the API in the LegalizerInfo class is the main interface for targets to specify which types are legal for which operations, and what to do to turn illegal type/operation combinations into legal ones. For each operation the type sizes that can be legalized without having to change the size of the type are specified with a call to setAction. This isn't different to how GlobalISel worked before. For example, for a target that supports 32 and 64 bit adds natively: for (auto Ty : {s32, s64}) setAction({G_ADD, 0, s32}, Legal); or for a target that needs a library call for a 32 bit division: setAction({G_SDIV, s32}, Libcall); The main conceptual change to the LegalizerInfo API, is in specifying how to legalize the type sizes for which a change of size is needed. For example, in the above example, how to specify how all types from i1 to i8388607 (apart from s32 and s64 which are legal) need to be legalized and expressed in terms of operations on the available legal sizes (again, i32 and i64 in this case). Before, the implementation only allowed specifying power-of-2-sized types (e.g. setAction({G_ADD, 0, s128}, NarrowScalar). A worse limitation was that if you'd wanted to specify how to legalize all the sized types as allowed by the LLVM-IR LangRef, i1 to i8388607, you'd have to call setAction 8388607-3 times and probably would need a lot of memory to store all of these specifications. Instead, the legalization actions that need to change the size of the type are specified now using a "SizeChangeStrategy". For example: setLegalizeScalarToDifferentSizeStrategy( G_ADD, 0, widenToLargerAndNarrowToLargest); This example indicates that for type sizes for which there is a larger size that can be legalized towards, do it by Widening the size. For example, G_ADD on s17 will be legalized by first doing WidenScalar to make it s32, after which it's legal. The "NarrowToLargest" indicates what to do if there is no larger size that can be legalized towards. E.g. G_ADD on s92 will be legalized by doing NarrowScalar to s64. Another example, taken from the ARM backend is: for (unsigned Op : {G_SDIV, G_UDIV}) { setLegalizeScalarToDifferentSizeStrategy(Op, 0, widenToLargerTypesUnsupportedOtherwise); if (ST.hasDivideInARMMode()) setAction({Op, s32}, Legal); else setAction({Op, s32}, Libcall); } For this example, G_SDIV on s8, on a target without a divide instruction, would be legalized by first doing action (WidenScalar, s32), followed by (Libcall, s32). The same principle is also followed for when the number of vector lanes on vector data types need to be changed, e.g.: setAction({G_ADD, LLT::vector(8, 8)}, LegalizerInfo::Legal); setAction({G_ADD, LLT::vector(16, 8)}, LegalizerInfo::Legal); setAction({G_ADD, LLT::vector(4, 16)}, LegalizerInfo::Legal); setAction({G_ADD, LLT::vector(8, 16)}, LegalizerInfo::Legal); setAction({G_ADD, LLT::vector(2, 32)}, LegalizerInfo::Legal); setAction({G_ADD, LLT::vector(4, 32)}, LegalizerInfo::Legal); setLegalizeVectorElementToDifferentSizeStrategy( G_ADD, 0, widenToLargerTypesUnsupportedOtherwise); As currently implemented here, vector types are legalized by first making the vector element size legal, followed by then making the number of lanes legal. The strategy to follow in the first step is set by a call to setLegalizeVectorElementToDifferentSizeStrategy, see example above. The strategy followed in the second step "moreToWiderTypesAndLessToWidest" (see code for its definition), indicating that vectors are widened to more elements so they map to natively supported vector widths, or when there isn't a legal wider vector, split the vector to map it to the widest vector supported. Therefore, for the above specification, some example legalizations are: * getAction({G_ADD, LLT::vector(3, 3)}) returns {WidenScalar, LLT::vector(3, 8)} * getAction({G_ADD, LLT::vector(3, 8)}) then returns {MoreElements, LLT::vector(8, 8)} * getAction({G_ADD, LLT::vector(20, 8)}) returns {FewerElements, LLT::vector(16, 8)} 2. Key implementation aspects. How to legalize a specific (operation, type index, size) tuple is represented by mapping intervals of integers representing a range of size types to an action to take, e.g.: setScalarAction({G_ADD, LLT:scalar(1)}, {{1, WidenScalar}, // bit sizes [ 1, 31[ {32, Legal}, // bit sizes [32, 33[ {33, WidenScalar}, // bit sizes [33, 64[ {64, Legal}, // bit sizes [64, 65[ {65, NarrowScalar} // bit sizes [65, +inf[ }); Please note that most of the code to do the actual lowering of non-power-of-2 sized types is currently missing, this is just trying to make it possible for targets to specify what is legal, and how non-legal types should be legalized. Probably quite a bit of further work is needed in the actual legalizing and the other passes in GlobalISel to support non-power-of-2 sized types. I hope the documentation in LegalizerInfo.h and the examples provided in the various {Target}LegalizerInfo.cpp and LegalizerInfoTest.cpp explains well enough how this is meant to be used. This drops the need for LLT::{half,double}...Size(). Differential Revision: https://reviews.llvm.org/D30529 llvm-svn: 317560
2017-11-07 18:34:34 +08:00
// See above, the following needs to be a loop, at least for now.
for (std::size_t i = VecIdx + 1; i < Vec.size(); ++i)
if (!needsLegalizingToDifferentSize(Vec[i].second) &&
Vec[i].second != Unsupported)
return {Vec[i].first, Action};
llvm_unreachable("");
}
[GlobalISel] Enable legalizing non-power-of-2 sized types. This changes the interface of how targets describe how to legalize, see the below description. 1. Interface for targets to describe how to legalize. In GlobalISel, the API in the LegalizerInfo class is the main interface for targets to specify which types are legal for which operations, and what to do to turn illegal type/operation combinations into legal ones. For each operation the type sizes that can be legalized without having to change the size of the type are specified with a call to setAction. This isn't different to how GlobalISel worked before. For example, for a target that supports 32 and 64 bit adds natively: for (auto Ty : {s32, s64}) setAction({G_ADD, 0, s32}, Legal); or for a target that needs a library call for a 32 bit division: setAction({G_SDIV, s32}, Libcall); The main conceptual change to the LegalizerInfo API, is in specifying how to legalize the type sizes for which a change of size is needed. For example, in the above example, how to specify how all types from i1 to i8388607 (apart from s32 and s64 which are legal) need to be legalized and expressed in terms of operations on the available legal sizes (again, i32 and i64 in this case). Before, the implementation only allowed specifying power-of-2-sized types (e.g. setAction({G_ADD, 0, s128}, NarrowScalar). A worse limitation was that if you'd wanted to specify how to legalize all the sized types as allowed by the LLVM-IR LangRef, i1 to i8388607, you'd have to call setAction 8388607-3 times and probably would need a lot of memory to store all of these specifications. Instead, the legalization actions that need to change the size of the type are specified now using a "SizeChangeStrategy". For example: setLegalizeScalarToDifferentSizeStrategy( G_ADD, 0, widenToLargerAndNarrowToLargest); This example indicates that for type sizes for which there is a larger size that can be legalized towards, do it by Widening the size. For example, G_ADD on s17 will be legalized by first doing WidenScalar to make it s32, after which it's legal. The "NarrowToLargest" indicates what to do if there is no larger size that can be legalized towards. E.g. G_ADD on s92 will be legalized by doing NarrowScalar to s64. Another example, taken from the ARM backend is: for (unsigned Op : {G_SDIV, G_UDIV}) { setLegalizeScalarToDifferentSizeStrategy(Op, 0, widenToLargerTypesUnsupportedOtherwise); if (ST.hasDivideInARMMode()) setAction({Op, s32}, Legal); else setAction({Op, s32}, Libcall); } For this example, G_SDIV on s8, on a target without a divide instruction, would be legalized by first doing action (WidenScalar, s32), followed by (Libcall, s32). The same principle is also followed for when the number of vector lanes on vector data types need to be changed, e.g.: setAction({G_ADD, LLT::vector(8, 8)}, LegalizerInfo::Legal); setAction({G_ADD, LLT::vector(16, 8)}, LegalizerInfo::Legal); setAction({G_ADD, LLT::vector(4, 16)}, LegalizerInfo::Legal); setAction({G_ADD, LLT::vector(8, 16)}, LegalizerInfo::Legal); setAction({G_ADD, LLT::vector(2, 32)}, LegalizerInfo::Legal); setAction({G_ADD, LLT::vector(4, 32)}, LegalizerInfo::Legal); setLegalizeVectorElementToDifferentSizeStrategy( G_ADD, 0, widenToLargerTypesUnsupportedOtherwise); As currently implemented here, vector types are legalized by first making the vector element size legal, followed by then making the number of lanes legal. The strategy to follow in the first step is set by a call to setLegalizeVectorElementToDifferentSizeStrategy, see example above. The strategy followed in the second step "moreToWiderTypesAndLessToWidest" (see code for its definition), indicating that vectors are widened to more elements so they map to natively supported vector widths, or when there isn't a legal wider vector, split the vector to map it to the widest vector supported. Therefore, for the above specification, some example legalizations are: * getAction({G_ADD, LLT::vector(3, 3)}) returns {WidenScalar, LLT::vector(3, 8)} * getAction({G_ADD, LLT::vector(3, 8)}) then returns {MoreElements, LLT::vector(8, 8)} * getAction({G_ADD, LLT::vector(20, 8)}) returns {FewerElements, LLT::vector(16, 8)} 2. Key implementation aspects. How to legalize a specific (operation, type index, size) tuple is represented by mapping intervals of integers representing a range of size types to an action to take, e.g.: setScalarAction({G_ADD, LLT:scalar(1)}, {{1, WidenScalar}, // bit sizes [ 1, 31[ {32, Legal}, // bit sizes [32, 33[ {33, WidenScalar}, // bit sizes [33, 64[ {64, Legal}, // bit sizes [64, 65[ {65, NarrowScalar} // bit sizes [65, +inf[ }); Please note that most of the code to do the actual lowering of non-power-of-2 sized types is currently missing, this is just trying to make it possible for targets to specify what is legal, and how non-legal types should be legalized. Probably quite a bit of further work is needed in the actual legalizing and the other passes in GlobalISel to support non-power-of-2 sized types. I hope the documentation in LegalizerInfo.h and the examples provided in the various {Target}LegalizerInfo.cpp and LegalizerInfoTest.cpp explains well enough how this is meant to be used. This drops the need for LLT::{half,double}...Size(). Differential Revision: https://reviews.llvm.org/D30529 llvm-svn: 317560
2017-11-07 18:34:34 +08:00
case Unsupported:
return {Size, Unsupported};
case NotFound:
[globalisel][legalizer] Adapt LegalizerInfo to support inter-type dependencies and other things. Summary: As discussed in D42244, we have difficulty describing the legality of some operations. We're not able to specify relationships between types. For example, declaring the following setAction({..., 0, s32}, Legal) setAction({..., 0, s64}, Legal) setAction({..., 1, s32}, Legal) setAction({..., 1, s64}, Legal) currently declares these type combinations as legal: {s32, s32} {s64, s32} {s32, s64} {s64, s64} but we currently have no means to say that, for example, {s64, s32} is not legal. Some operations such as G_INSERT/G_EXTRACT/G_MERGE_VALUES/ G_UNMERGE_VALUES have relationships between the types that are currently described incorrectly. Additionally, G_LOAD/G_STORE currently have no means to legalize non-atomics differently to atomics. The necessary information is in the MMO but we have no way to use this in the legalizer. Similarly, there is currently no way for the register type and the memory type to differ so there is no way to cleanly represent extending-load/truncating-store in a way that can't be broken by optimizers (resulting in illegal MIR). It's also difficult to control the legalization strategy. We've added support for legalizing non-power of 2 types but there's still some hardcoded assumptions about the strategy. The main one I've noticed is that type0 is always legalized before type1 which is not a good strategy for `type0 = G_EXTRACT type1, ...` if you need to widen the container. It will converge on the same result eventually but it will take a much longer route when legalizing type0 than if you legalize type1 first. Lastly, the definition of legality and the legalization strategy is kept separate which is not ideal. It's helpful to be able to look at a one piece of code and see both what is legal and the method the legalizer will use to make illegal MIR more legal. This patch adds a layer onto the LegalizerInfo (to be removed when all targets have been migrated) which resolves all these issues. Here are the rules for shift and division: for (unsigned BinOp : {G_LSHR, G_ASHR, G_SDIV, G_UDIV}) getActionDefinitions(BinOp) .legalFor({s32, s64}) // If type0 is s32/s64 then it's Legal .clampScalar(0, s32, s64) // If type0 is <s32 then WidenScalar to s32 // If type0 is >s64 then NarrowScalar to s64 .widenScalarToPow2(0) // Round type0 scalars up to powers of 2 .unsupported(); // Otherwise, it's unsupported This describes everything needed to both define legality and describe how to make illegal things legal. Here's an example of a complex rule: getActionDefinitions(G_INSERT) .unsupportedIf([=](const LegalityQuery &Query) { // If type0 is smaller than type1 then it's unsupported return Query.Types[0].getSizeInBits() <= Query.Types[1].getSizeInBits(); }) .legalIf([=](const LegalityQuery &Query) { // If type0 is s32/s64/p0 and type1 is a power of 2 other than 2 or 4 then it's legal // We don't need to worry about large type1's because unsupportedIf caught that. const LLT &Ty0 = Query.Types[0]; const LLT &Ty1 = Query.Types[1]; if (Ty0 != s32 && Ty0 != s64 && Ty0 != p0) return false; return isPowerOf2_32(Ty1.getSizeInBits()) && (Ty1.getSizeInBits() == 1 || Ty1.getSizeInBits() >= 8); }) .clampScalar(0, s32, s64) .widenScalarToPow2(0) .maxScalarIf(typeInSet(0, {s32}), 1, s16) // If type0 is s32 and type1 is bigger than s16 then NarrowScalar type1 to s16 .maxScalarIf(typeInSet(0, {s64}), 1, s32) // If type0 is s64 and type1 is bigger than s32 then NarrowScalar type1 to s32 .widenScalarToPow2(1) // Round type1 scalars up to powers of 2 .unsupported(); This uses a lambda to say that G_INSERT is unsupported when type0 is bigger than type1 (in practice, this would be a default rule for G_INSERT). It also uses one to describe the legal cases. This particular predicate is equivalent to: .legalFor({{s32, s1}, {s32, s8}, {s32, s16}, {s64, s1}, {s64, s8}, {s64, s16}, {s64, s32}}) In terms of performance, I saw a slight (~6%) performance improvement when AArch64 was around 30% ported but it's pretty much break even right now. I'm going to take a look at constexpr as a means to reduce the initialization cost. Future work: * Make it possible for opcodes to share rulesets. There's no need for G_LSHR/G_ASHR/G_SDIV/G_UDIV to have separate rule and ruleset objects. There's no technical barrier to this, it just hasn't been done yet. * Replace the type-index numbers with an enum to get .clampScalar(Type0, s32, s64) * Better names for things like .maxScalarIf() (clampMaxScalar?) and the vector rules. * Improve initialization cost using constexpr Possible future work: * It's possible to make these rulesets change the MIR directly instead of returning a description of how to change the MIR. This should remove a little overhead caused by parsing the description and routing to the right code, but the real motivation is that it removes the need for LegalizeAction::Custom. With Custom removed, there's no longer a requirement that Custom legalization change the opcode to something that's considered legal. Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar, volkan, reames, bogner Reviewed By: bogner Subscribers: hintonda, bogner, aemerson, mgorny, javed.absar, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D42251 llvm-svn: 323681
2018-01-30 03:54:49 +08:00
case UseLegacyRules:
[GlobalISel] Enable legalizing non-power-of-2 sized types. This changes the interface of how targets describe how to legalize, see the below description. 1. Interface for targets to describe how to legalize. In GlobalISel, the API in the LegalizerInfo class is the main interface for targets to specify which types are legal for which operations, and what to do to turn illegal type/operation combinations into legal ones. For each operation the type sizes that can be legalized without having to change the size of the type are specified with a call to setAction. This isn't different to how GlobalISel worked before. For example, for a target that supports 32 and 64 bit adds natively: for (auto Ty : {s32, s64}) setAction({G_ADD, 0, s32}, Legal); or for a target that needs a library call for a 32 bit division: setAction({G_SDIV, s32}, Libcall); The main conceptual change to the LegalizerInfo API, is in specifying how to legalize the type sizes for which a change of size is needed. For example, in the above example, how to specify how all types from i1 to i8388607 (apart from s32 and s64 which are legal) need to be legalized and expressed in terms of operations on the available legal sizes (again, i32 and i64 in this case). Before, the implementation only allowed specifying power-of-2-sized types (e.g. setAction({G_ADD, 0, s128}, NarrowScalar). A worse limitation was that if you'd wanted to specify how to legalize all the sized types as allowed by the LLVM-IR LangRef, i1 to i8388607, you'd have to call setAction 8388607-3 times and probably would need a lot of memory to store all of these specifications. Instead, the legalization actions that need to change the size of the type are specified now using a "SizeChangeStrategy". For example: setLegalizeScalarToDifferentSizeStrategy( G_ADD, 0, widenToLargerAndNarrowToLargest); This example indicates that for type sizes for which there is a larger size that can be legalized towards, do it by Widening the size. For example, G_ADD on s17 will be legalized by first doing WidenScalar to make it s32, after which it's legal. The "NarrowToLargest" indicates what to do if there is no larger size that can be legalized towards. E.g. G_ADD on s92 will be legalized by doing NarrowScalar to s64. Another example, taken from the ARM backend is: for (unsigned Op : {G_SDIV, G_UDIV}) { setLegalizeScalarToDifferentSizeStrategy(Op, 0, widenToLargerTypesUnsupportedOtherwise); if (ST.hasDivideInARMMode()) setAction({Op, s32}, Legal); else setAction({Op, s32}, Libcall); } For this example, G_SDIV on s8, on a target without a divide instruction, would be legalized by first doing action (WidenScalar, s32), followed by (Libcall, s32). The same principle is also followed for when the number of vector lanes on vector data types need to be changed, e.g.: setAction({G_ADD, LLT::vector(8, 8)}, LegalizerInfo::Legal); setAction({G_ADD, LLT::vector(16, 8)}, LegalizerInfo::Legal); setAction({G_ADD, LLT::vector(4, 16)}, LegalizerInfo::Legal); setAction({G_ADD, LLT::vector(8, 16)}, LegalizerInfo::Legal); setAction({G_ADD, LLT::vector(2, 32)}, LegalizerInfo::Legal); setAction({G_ADD, LLT::vector(4, 32)}, LegalizerInfo::Legal); setLegalizeVectorElementToDifferentSizeStrategy( G_ADD, 0, widenToLargerTypesUnsupportedOtherwise); As currently implemented here, vector types are legalized by first making the vector element size legal, followed by then making the number of lanes legal. The strategy to follow in the first step is set by a call to setLegalizeVectorElementToDifferentSizeStrategy, see example above. The strategy followed in the second step "moreToWiderTypesAndLessToWidest" (see code for its definition), indicating that vectors are widened to more elements so they map to natively supported vector widths, or when there isn't a legal wider vector, split the vector to map it to the widest vector supported. Therefore, for the above specification, some example legalizations are: * getAction({G_ADD, LLT::vector(3, 3)}) returns {WidenScalar, LLT::vector(3, 8)} * getAction({G_ADD, LLT::vector(3, 8)}) then returns {MoreElements, LLT::vector(8, 8)} * getAction({G_ADD, LLT::vector(20, 8)}) returns {FewerElements, LLT::vector(16, 8)} 2. Key implementation aspects. How to legalize a specific (operation, type index, size) tuple is represented by mapping intervals of integers representing a range of size types to an action to take, e.g.: setScalarAction({G_ADD, LLT:scalar(1)}, {{1, WidenScalar}, // bit sizes [ 1, 31[ {32, Legal}, // bit sizes [32, 33[ {33, WidenScalar}, // bit sizes [33, 64[ {64, Legal}, // bit sizes [64, 65[ {65, NarrowScalar} // bit sizes [65, +inf[ }); Please note that most of the code to do the actual lowering of non-power-of-2 sized types is currently missing, this is just trying to make it possible for targets to specify what is legal, and how non-legal types should be legalized. Probably quite a bit of further work is needed in the actual legalizing and the other passes in GlobalISel to support non-power-of-2 sized types. I hope the documentation in LegalizerInfo.h and the examples provided in the various {Target}LegalizerInfo.cpp and LegalizerInfoTest.cpp explains well enough how this is meant to be used. This drops the need for LLT::{half,double}...Size(). Differential Revision: https://reviews.llvm.org/D30529 llvm-svn: 317560
2017-11-07 18:34:34 +08:00
llvm_unreachable("NotFound");
}
llvm_unreachable("Action has an unknown enum value");
}
std::pair<LegalizeAction, LLT>
[GlobalISel] Enable legalizing non-power-of-2 sized types. This changes the interface of how targets describe how to legalize, see the below description. 1. Interface for targets to describe how to legalize. In GlobalISel, the API in the LegalizerInfo class is the main interface for targets to specify which types are legal for which operations, and what to do to turn illegal type/operation combinations into legal ones. For each operation the type sizes that can be legalized without having to change the size of the type are specified with a call to setAction. This isn't different to how GlobalISel worked before. For example, for a target that supports 32 and 64 bit adds natively: for (auto Ty : {s32, s64}) setAction({G_ADD, 0, s32}, Legal); or for a target that needs a library call for a 32 bit division: setAction({G_SDIV, s32}, Libcall); The main conceptual change to the LegalizerInfo API, is in specifying how to legalize the type sizes for which a change of size is needed. For example, in the above example, how to specify how all types from i1 to i8388607 (apart from s32 and s64 which are legal) need to be legalized and expressed in terms of operations on the available legal sizes (again, i32 and i64 in this case). Before, the implementation only allowed specifying power-of-2-sized types (e.g. setAction({G_ADD, 0, s128}, NarrowScalar). A worse limitation was that if you'd wanted to specify how to legalize all the sized types as allowed by the LLVM-IR LangRef, i1 to i8388607, you'd have to call setAction 8388607-3 times and probably would need a lot of memory to store all of these specifications. Instead, the legalization actions that need to change the size of the type are specified now using a "SizeChangeStrategy". For example: setLegalizeScalarToDifferentSizeStrategy( G_ADD, 0, widenToLargerAndNarrowToLargest); This example indicates that for type sizes for which there is a larger size that can be legalized towards, do it by Widening the size. For example, G_ADD on s17 will be legalized by first doing WidenScalar to make it s32, after which it's legal. The "NarrowToLargest" indicates what to do if there is no larger size that can be legalized towards. E.g. G_ADD on s92 will be legalized by doing NarrowScalar to s64. Another example, taken from the ARM backend is: for (unsigned Op : {G_SDIV, G_UDIV}) { setLegalizeScalarToDifferentSizeStrategy(Op, 0, widenToLargerTypesUnsupportedOtherwise); if (ST.hasDivideInARMMode()) setAction({Op, s32}, Legal); else setAction({Op, s32}, Libcall); } For this example, G_SDIV on s8, on a target without a divide instruction, would be legalized by first doing action (WidenScalar, s32), followed by (Libcall, s32). The same principle is also followed for when the number of vector lanes on vector data types need to be changed, e.g.: setAction({G_ADD, LLT::vector(8, 8)}, LegalizerInfo::Legal); setAction({G_ADD, LLT::vector(16, 8)}, LegalizerInfo::Legal); setAction({G_ADD, LLT::vector(4, 16)}, LegalizerInfo::Legal); setAction({G_ADD, LLT::vector(8, 16)}, LegalizerInfo::Legal); setAction({G_ADD, LLT::vector(2, 32)}, LegalizerInfo::Legal); setAction({G_ADD, LLT::vector(4, 32)}, LegalizerInfo::Legal); setLegalizeVectorElementToDifferentSizeStrategy( G_ADD, 0, widenToLargerTypesUnsupportedOtherwise); As currently implemented here, vector types are legalized by first making the vector element size legal, followed by then making the number of lanes legal. The strategy to follow in the first step is set by a call to setLegalizeVectorElementToDifferentSizeStrategy, see example above. The strategy followed in the second step "moreToWiderTypesAndLessToWidest" (see code for its definition), indicating that vectors are widened to more elements so they map to natively supported vector widths, or when there isn't a legal wider vector, split the vector to map it to the widest vector supported. Therefore, for the above specification, some example legalizations are: * getAction({G_ADD, LLT::vector(3, 3)}) returns {WidenScalar, LLT::vector(3, 8)} * getAction({G_ADD, LLT::vector(3, 8)}) then returns {MoreElements, LLT::vector(8, 8)} * getAction({G_ADD, LLT::vector(20, 8)}) returns {FewerElements, LLT::vector(16, 8)} 2. Key implementation aspects. How to legalize a specific (operation, type index, size) tuple is represented by mapping intervals of integers representing a range of size types to an action to take, e.g.: setScalarAction({G_ADD, LLT:scalar(1)}, {{1, WidenScalar}, // bit sizes [ 1, 31[ {32, Legal}, // bit sizes [32, 33[ {33, WidenScalar}, // bit sizes [33, 64[ {64, Legal}, // bit sizes [64, 65[ {65, NarrowScalar} // bit sizes [65, +inf[ }); Please note that most of the code to do the actual lowering of non-power-of-2 sized types is currently missing, this is just trying to make it possible for targets to specify what is legal, and how non-legal types should be legalized. Probably quite a bit of further work is needed in the actual legalizing and the other passes in GlobalISel to support non-power-of-2 sized types. I hope the documentation in LegalizerInfo.h and the examples provided in the various {Target}LegalizerInfo.cpp and LegalizerInfoTest.cpp explains well enough how this is meant to be used. This drops the need for LLT::{half,double}...Size(). Differential Revision: https://reviews.llvm.org/D30529 llvm-svn: 317560
2017-11-07 18:34:34 +08:00
LegalizerInfo::findScalarLegalAction(const InstrAspect &Aspect) const {
assert(Aspect.Type.isScalar() || Aspect.Type.isPointer());
if (Aspect.Opcode < FirstOp || Aspect.Opcode > LastOp)
return {NotFound, LLT()};
[globalisel][legalizer] Adapt LegalizerInfo to support inter-type dependencies and other things. Summary: As discussed in D42244, we have difficulty describing the legality of some operations. We're not able to specify relationships between types. For example, declaring the following setAction({..., 0, s32}, Legal) setAction({..., 0, s64}, Legal) setAction({..., 1, s32}, Legal) setAction({..., 1, s64}, Legal) currently declares these type combinations as legal: {s32, s32} {s64, s32} {s32, s64} {s64, s64} but we currently have no means to say that, for example, {s64, s32} is not legal. Some operations such as G_INSERT/G_EXTRACT/G_MERGE_VALUES/ G_UNMERGE_VALUES have relationships between the types that are currently described incorrectly. Additionally, G_LOAD/G_STORE currently have no means to legalize non-atomics differently to atomics. The necessary information is in the MMO but we have no way to use this in the legalizer. Similarly, there is currently no way for the register type and the memory type to differ so there is no way to cleanly represent extending-load/truncating-store in a way that can't be broken by optimizers (resulting in illegal MIR). It's also difficult to control the legalization strategy. We've added support for legalizing non-power of 2 types but there's still some hardcoded assumptions about the strategy. The main one I've noticed is that type0 is always legalized before type1 which is not a good strategy for `type0 = G_EXTRACT type1, ...` if you need to widen the container. It will converge on the same result eventually but it will take a much longer route when legalizing type0 than if you legalize type1 first. Lastly, the definition of legality and the legalization strategy is kept separate which is not ideal. It's helpful to be able to look at a one piece of code and see both what is legal and the method the legalizer will use to make illegal MIR more legal. This patch adds a layer onto the LegalizerInfo (to be removed when all targets have been migrated) which resolves all these issues. Here are the rules for shift and division: for (unsigned BinOp : {G_LSHR, G_ASHR, G_SDIV, G_UDIV}) getActionDefinitions(BinOp) .legalFor({s32, s64}) // If type0 is s32/s64 then it's Legal .clampScalar(0, s32, s64) // If type0 is <s32 then WidenScalar to s32 // If type0 is >s64 then NarrowScalar to s64 .widenScalarToPow2(0) // Round type0 scalars up to powers of 2 .unsupported(); // Otherwise, it's unsupported This describes everything needed to both define legality and describe how to make illegal things legal. Here's an example of a complex rule: getActionDefinitions(G_INSERT) .unsupportedIf([=](const LegalityQuery &Query) { // If type0 is smaller than type1 then it's unsupported return Query.Types[0].getSizeInBits() <= Query.Types[1].getSizeInBits(); }) .legalIf([=](const LegalityQuery &Query) { // If type0 is s32/s64/p0 and type1 is a power of 2 other than 2 or 4 then it's legal // We don't need to worry about large type1's because unsupportedIf caught that. const LLT &Ty0 = Query.Types[0]; const LLT &Ty1 = Query.Types[1]; if (Ty0 != s32 && Ty0 != s64 && Ty0 != p0) return false; return isPowerOf2_32(Ty1.getSizeInBits()) && (Ty1.getSizeInBits() == 1 || Ty1.getSizeInBits() >= 8); }) .clampScalar(0, s32, s64) .widenScalarToPow2(0) .maxScalarIf(typeInSet(0, {s32}), 1, s16) // If type0 is s32 and type1 is bigger than s16 then NarrowScalar type1 to s16 .maxScalarIf(typeInSet(0, {s64}), 1, s32) // If type0 is s64 and type1 is bigger than s32 then NarrowScalar type1 to s32 .widenScalarToPow2(1) // Round type1 scalars up to powers of 2 .unsupported(); This uses a lambda to say that G_INSERT is unsupported when type0 is bigger than type1 (in practice, this would be a default rule for G_INSERT). It also uses one to describe the legal cases. This particular predicate is equivalent to: .legalFor({{s32, s1}, {s32, s8}, {s32, s16}, {s64, s1}, {s64, s8}, {s64, s16}, {s64, s32}}) In terms of performance, I saw a slight (~6%) performance improvement when AArch64 was around 30% ported but it's pretty much break even right now. I'm going to take a look at constexpr as a means to reduce the initialization cost. Future work: * Make it possible for opcodes to share rulesets. There's no need for G_LSHR/G_ASHR/G_SDIV/G_UDIV to have separate rule and ruleset objects. There's no technical barrier to this, it just hasn't been done yet. * Replace the type-index numbers with an enum to get .clampScalar(Type0, s32, s64) * Better names for things like .maxScalarIf() (clampMaxScalar?) and the vector rules. * Improve initialization cost using constexpr Possible future work: * It's possible to make these rulesets change the MIR directly instead of returning a description of how to change the MIR. This should remove a little overhead caused by parsing the description and routing to the right code, but the real motivation is that it removes the need for LegalizeAction::Custom. With Custom removed, there's no longer a requirement that Custom legalization change the opcode to something that's considered legal. Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar, volkan, reames, bogner Reviewed By: bogner Subscribers: hintonda, bogner, aemerson, mgorny, javed.absar, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D42251 llvm-svn: 323681
2018-01-30 03:54:49 +08:00
const unsigned OpcodeIdx = getOpcodeIdxForOpcode(Aspect.Opcode);
[GlobalISel] Enable legalizing non-power-of-2 sized types. This changes the interface of how targets describe how to legalize, see the below description. 1. Interface for targets to describe how to legalize. In GlobalISel, the API in the LegalizerInfo class is the main interface for targets to specify which types are legal for which operations, and what to do to turn illegal type/operation combinations into legal ones. For each operation the type sizes that can be legalized without having to change the size of the type are specified with a call to setAction. This isn't different to how GlobalISel worked before. For example, for a target that supports 32 and 64 bit adds natively: for (auto Ty : {s32, s64}) setAction({G_ADD, 0, s32}, Legal); or for a target that needs a library call for a 32 bit division: setAction({G_SDIV, s32}, Libcall); The main conceptual change to the LegalizerInfo API, is in specifying how to legalize the type sizes for which a change of size is needed. For example, in the above example, how to specify how all types from i1 to i8388607 (apart from s32 and s64 which are legal) need to be legalized and expressed in terms of operations on the available legal sizes (again, i32 and i64 in this case). Before, the implementation only allowed specifying power-of-2-sized types (e.g. setAction({G_ADD, 0, s128}, NarrowScalar). A worse limitation was that if you'd wanted to specify how to legalize all the sized types as allowed by the LLVM-IR LangRef, i1 to i8388607, you'd have to call setAction 8388607-3 times and probably would need a lot of memory to store all of these specifications. Instead, the legalization actions that need to change the size of the type are specified now using a "SizeChangeStrategy". For example: setLegalizeScalarToDifferentSizeStrategy( G_ADD, 0, widenToLargerAndNarrowToLargest); This example indicates that for type sizes for which there is a larger size that can be legalized towards, do it by Widening the size. For example, G_ADD on s17 will be legalized by first doing WidenScalar to make it s32, after which it's legal. The "NarrowToLargest" indicates what to do if there is no larger size that can be legalized towards. E.g. G_ADD on s92 will be legalized by doing NarrowScalar to s64. Another example, taken from the ARM backend is: for (unsigned Op : {G_SDIV, G_UDIV}) { setLegalizeScalarToDifferentSizeStrategy(Op, 0, widenToLargerTypesUnsupportedOtherwise); if (ST.hasDivideInARMMode()) setAction({Op, s32}, Legal); else setAction({Op, s32}, Libcall); } For this example, G_SDIV on s8, on a target without a divide instruction, would be legalized by first doing action (WidenScalar, s32), followed by (Libcall, s32). The same principle is also followed for when the number of vector lanes on vector data types need to be changed, e.g.: setAction({G_ADD, LLT::vector(8, 8)}, LegalizerInfo::Legal); setAction({G_ADD, LLT::vector(16, 8)}, LegalizerInfo::Legal); setAction({G_ADD, LLT::vector(4, 16)}, LegalizerInfo::Legal); setAction({G_ADD, LLT::vector(8, 16)}, LegalizerInfo::Legal); setAction({G_ADD, LLT::vector(2, 32)}, LegalizerInfo::Legal); setAction({G_ADD, LLT::vector(4, 32)}, LegalizerInfo::Legal); setLegalizeVectorElementToDifferentSizeStrategy( G_ADD, 0, widenToLargerTypesUnsupportedOtherwise); As currently implemented here, vector types are legalized by first making the vector element size legal, followed by then making the number of lanes legal. The strategy to follow in the first step is set by a call to setLegalizeVectorElementToDifferentSizeStrategy, see example above. The strategy followed in the second step "moreToWiderTypesAndLessToWidest" (see code for its definition), indicating that vectors are widened to more elements so they map to natively supported vector widths, or when there isn't a legal wider vector, split the vector to map it to the widest vector supported. Therefore, for the above specification, some example legalizations are: * getAction({G_ADD, LLT::vector(3, 3)}) returns {WidenScalar, LLT::vector(3, 8)} * getAction({G_ADD, LLT::vector(3, 8)}) then returns {MoreElements, LLT::vector(8, 8)} * getAction({G_ADD, LLT::vector(20, 8)}) returns {FewerElements, LLT::vector(16, 8)} 2. Key implementation aspects. How to legalize a specific (operation, type index, size) tuple is represented by mapping intervals of integers representing a range of size types to an action to take, e.g.: setScalarAction({G_ADD, LLT:scalar(1)}, {{1, WidenScalar}, // bit sizes [ 1, 31[ {32, Legal}, // bit sizes [32, 33[ {33, WidenScalar}, // bit sizes [33, 64[ {64, Legal}, // bit sizes [64, 65[ {65, NarrowScalar} // bit sizes [65, +inf[ }); Please note that most of the code to do the actual lowering of non-power-of-2 sized types is currently missing, this is just trying to make it possible for targets to specify what is legal, and how non-legal types should be legalized. Probably quite a bit of further work is needed in the actual legalizing and the other passes in GlobalISel to support non-power-of-2 sized types. I hope the documentation in LegalizerInfo.h and the examples provided in the various {Target}LegalizerInfo.cpp and LegalizerInfoTest.cpp explains well enough how this is meant to be used. This drops the need for LLT::{half,double}...Size(). Differential Revision: https://reviews.llvm.org/D30529 llvm-svn: 317560
2017-11-07 18:34:34 +08:00
if (Aspect.Type.isPointer() &&
AddrSpace2PointerActions[OpcodeIdx].find(Aspect.Type.getAddressSpace()) ==
AddrSpace2PointerActions[OpcodeIdx].end()) {
return {NotFound, LLT()};
}
const SmallVector<SizeAndActionsVec, 1> &Actions =
Aspect.Type.isPointer()
? AddrSpace2PointerActions[OpcodeIdx]
.find(Aspect.Type.getAddressSpace())
->second
: ScalarActions[OpcodeIdx];
if (Aspect.Idx >= Actions.size())
return {NotFound, LLT()};
const SizeAndActionsVec &Vec = Actions[Aspect.Idx];
// FIXME: speed up this search, e.g. by using a results cache for repeated
// queries?
auto SizeAndAction = findAction(Vec, Aspect.Type.getSizeInBits());
return {SizeAndAction.second,
Aspect.Type.isScalar() ? LLT::scalar(SizeAndAction.first)
: LLT::pointer(Aspect.Type.getAddressSpace(),
SizeAndAction.first)};
}
std::pair<LegalizeAction, LLT>
[GlobalISel] Enable legalizing non-power-of-2 sized types. This changes the interface of how targets describe how to legalize, see the below description. 1. Interface for targets to describe how to legalize. In GlobalISel, the API in the LegalizerInfo class is the main interface for targets to specify which types are legal for which operations, and what to do to turn illegal type/operation combinations into legal ones. For each operation the type sizes that can be legalized without having to change the size of the type are specified with a call to setAction. This isn't different to how GlobalISel worked before. For example, for a target that supports 32 and 64 bit adds natively: for (auto Ty : {s32, s64}) setAction({G_ADD, 0, s32}, Legal); or for a target that needs a library call for a 32 bit division: setAction({G_SDIV, s32}, Libcall); The main conceptual change to the LegalizerInfo API, is in specifying how to legalize the type sizes for which a change of size is needed. For example, in the above example, how to specify how all types from i1 to i8388607 (apart from s32 and s64 which are legal) need to be legalized and expressed in terms of operations on the available legal sizes (again, i32 and i64 in this case). Before, the implementation only allowed specifying power-of-2-sized types (e.g. setAction({G_ADD, 0, s128}, NarrowScalar). A worse limitation was that if you'd wanted to specify how to legalize all the sized types as allowed by the LLVM-IR LangRef, i1 to i8388607, you'd have to call setAction 8388607-3 times and probably would need a lot of memory to store all of these specifications. Instead, the legalization actions that need to change the size of the type are specified now using a "SizeChangeStrategy". For example: setLegalizeScalarToDifferentSizeStrategy( G_ADD, 0, widenToLargerAndNarrowToLargest); This example indicates that for type sizes for which there is a larger size that can be legalized towards, do it by Widening the size. For example, G_ADD on s17 will be legalized by first doing WidenScalar to make it s32, after which it's legal. The "NarrowToLargest" indicates what to do if there is no larger size that can be legalized towards. E.g. G_ADD on s92 will be legalized by doing NarrowScalar to s64. Another example, taken from the ARM backend is: for (unsigned Op : {G_SDIV, G_UDIV}) { setLegalizeScalarToDifferentSizeStrategy(Op, 0, widenToLargerTypesUnsupportedOtherwise); if (ST.hasDivideInARMMode()) setAction({Op, s32}, Legal); else setAction({Op, s32}, Libcall); } For this example, G_SDIV on s8, on a target without a divide instruction, would be legalized by first doing action (WidenScalar, s32), followed by (Libcall, s32). The same principle is also followed for when the number of vector lanes on vector data types need to be changed, e.g.: setAction({G_ADD, LLT::vector(8, 8)}, LegalizerInfo::Legal); setAction({G_ADD, LLT::vector(16, 8)}, LegalizerInfo::Legal); setAction({G_ADD, LLT::vector(4, 16)}, LegalizerInfo::Legal); setAction({G_ADD, LLT::vector(8, 16)}, LegalizerInfo::Legal); setAction({G_ADD, LLT::vector(2, 32)}, LegalizerInfo::Legal); setAction({G_ADD, LLT::vector(4, 32)}, LegalizerInfo::Legal); setLegalizeVectorElementToDifferentSizeStrategy( G_ADD, 0, widenToLargerTypesUnsupportedOtherwise); As currently implemented here, vector types are legalized by first making the vector element size legal, followed by then making the number of lanes legal. The strategy to follow in the first step is set by a call to setLegalizeVectorElementToDifferentSizeStrategy, see example above. The strategy followed in the second step "moreToWiderTypesAndLessToWidest" (see code for its definition), indicating that vectors are widened to more elements so they map to natively supported vector widths, or when there isn't a legal wider vector, split the vector to map it to the widest vector supported. Therefore, for the above specification, some example legalizations are: * getAction({G_ADD, LLT::vector(3, 3)}) returns {WidenScalar, LLT::vector(3, 8)} * getAction({G_ADD, LLT::vector(3, 8)}) then returns {MoreElements, LLT::vector(8, 8)} * getAction({G_ADD, LLT::vector(20, 8)}) returns {FewerElements, LLT::vector(16, 8)} 2. Key implementation aspects. How to legalize a specific (operation, type index, size) tuple is represented by mapping intervals of integers representing a range of size types to an action to take, e.g.: setScalarAction({G_ADD, LLT:scalar(1)}, {{1, WidenScalar}, // bit sizes [ 1, 31[ {32, Legal}, // bit sizes [32, 33[ {33, WidenScalar}, // bit sizes [33, 64[ {64, Legal}, // bit sizes [64, 65[ {65, NarrowScalar} // bit sizes [65, +inf[ }); Please note that most of the code to do the actual lowering of non-power-of-2 sized types is currently missing, this is just trying to make it possible for targets to specify what is legal, and how non-legal types should be legalized. Probably quite a bit of further work is needed in the actual legalizing and the other passes in GlobalISel to support non-power-of-2 sized types. I hope the documentation in LegalizerInfo.h and the examples provided in the various {Target}LegalizerInfo.cpp and LegalizerInfoTest.cpp explains well enough how this is meant to be used. This drops the need for LLT::{half,double}...Size(). Differential Revision: https://reviews.llvm.org/D30529 llvm-svn: 317560
2017-11-07 18:34:34 +08:00
LegalizerInfo::findVectorLegalAction(const InstrAspect &Aspect) const {
assert(Aspect.Type.isVector());
// First legalize the vector element size, then legalize the number of
// lanes in the vector.
if (Aspect.Opcode < FirstOp || Aspect.Opcode > LastOp)
return {NotFound, Aspect.Type};
[globalisel][legalizer] Adapt LegalizerInfo to support inter-type dependencies and other things. Summary: As discussed in D42244, we have difficulty describing the legality of some operations. We're not able to specify relationships between types. For example, declaring the following setAction({..., 0, s32}, Legal) setAction({..., 0, s64}, Legal) setAction({..., 1, s32}, Legal) setAction({..., 1, s64}, Legal) currently declares these type combinations as legal: {s32, s32} {s64, s32} {s32, s64} {s64, s64} but we currently have no means to say that, for example, {s64, s32} is not legal. Some operations such as G_INSERT/G_EXTRACT/G_MERGE_VALUES/ G_UNMERGE_VALUES have relationships between the types that are currently described incorrectly. Additionally, G_LOAD/G_STORE currently have no means to legalize non-atomics differently to atomics. The necessary information is in the MMO but we have no way to use this in the legalizer. Similarly, there is currently no way for the register type and the memory type to differ so there is no way to cleanly represent extending-load/truncating-store in a way that can't be broken by optimizers (resulting in illegal MIR). It's also difficult to control the legalization strategy. We've added support for legalizing non-power of 2 types but there's still some hardcoded assumptions about the strategy. The main one I've noticed is that type0 is always legalized before type1 which is not a good strategy for `type0 = G_EXTRACT type1, ...` if you need to widen the container. It will converge on the same result eventually but it will take a much longer route when legalizing type0 than if you legalize type1 first. Lastly, the definition of legality and the legalization strategy is kept separate which is not ideal. It's helpful to be able to look at a one piece of code and see both what is legal and the method the legalizer will use to make illegal MIR more legal. This patch adds a layer onto the LegalizerInfo (to be removed when all targets have been migrated) which resolves all these issues. Here are the rules for shift and division: for (unsigned BinOp : {G_LSHR, G_ASHR, G_SDIV, G_UDIV}) getActionDefinitions(BinOp) .legalFor({s32, s64}) // If type0 is s32/s64 then it's Legal .clampScalar(0, s32, s64) // If type0 is <s32 then WidenScalar to s32 // If type0 is >s64 then NarrowScalar to s64 .widenScalarToPow2(0) // Round type0 scalars up to powers of 2 .unsupported(); // Otherwise, it's unsupported This describes everything needed to both define legality and describe how to make illegal things legal. Here's an example of a complex rule: getActionDefinitions(G_INSERT) .unsupportedIf([=](const LegalityQuery &Query) { // If type0 is smaller than type1 then it's unsupported return Query.Types[0].getSizeInBits() <= Query.Types[1].getSizeInBits(); }) .legalIf([=](const LegalityQuery &Query) { // If type0 is s32/s64/p0 and type1 is a power of 2 other than 2 or 4 then it's legal // We don't need to worry about large type1's because unsupportedIf caught that. const LLT &Ty0 = Query.Types[0]; const LLT &Ty1 = Query.Types[1]; if (Ty0 != s32 && Ty0 != s64 && Ty0 != p0) return false; return isPowerOf2_32(Ty1.getSizeInBits()) && (Ty1.getSizeInBits() == 1 || Ty1.getSizeInBits() >= 8); }) .clampScalar(0, s32, s64) .widenScalarToPow2(0) .maxScalarIf(typeInSet(0, {s32}), 1, s16) // If type0 is s32 and type1 is bigger than s16 then NarrowScalar type1 to s16 .maxScalarIf(typeInSet(0, {s64}), 1, s32) // If type0 is s64 and type1 is bigger than s32 then NarrowScalar type1 to s32 .widenScalarToPow2(1) // Round type1 scalars up to powers of 2 .unsupported(); This uses a lambda to say that G_INSERT is unsupported when type0 is bigger than type1 (in practice, this would be a default rule for G_INSERT). It also uses one to describe the legal cases. This particular predicate is equivalent to: .legalFor({{s32, s1}, {s32, s8}, {s32, s16}, {s64, s1}, {s64, s8}, {s64, s16}, {s64, s32}}) In terms of performance, I saw a slight (~6%) performance improvement when AArch64 was around 30% ported but it's pretty much break even right now. I'm going to take a look at constexpr as a means to reduce the initialization cost. Future work: * Make it possible for opcodes to share rulesets. There's no need for G_LSHR/G_ASHR/G_SDIV/G_UDIV to have separate rule and ruleset objects. There's no technical barrier to this, it just hasn't been done yet. * Replace the type-index numbers with an enum to get .clampScalar(Type0, s32, s64) * Better names for things like .maxScalarIf() (clampMaxScalar?) and the vector rules. * Improve initialization cost using constexpr Possible future work: * It's possible to make these rulesets change the MIR directly instead of returning a description of how to change the MIR. This should remove a little overhead caused by parsing the description and routing to the right code, but the real motivation is that it removes the need for LegalizeAction::Custom. With Custom removed, there's no longer a requirement that Custom legalization change the opcode to something that's considered legal. Reviewers: ab, t.p.northover, qcolombet, rovka, aditya_nandakumar, volkan, reames, bogner Reviewed By: bogner Subscribers: hintonda, bogner, aemerson, mgorny, javed.absar, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D42251 llvm-svn: 323681
2018-01-30 03:54:49 +08:00
const unsigned OpcodeIdx = getOpcodeIdxForOpcode(Aspect.Opcode);
[GlobalISel] Enable legalizing non-power-of-2 sized types. This changes the interface of how targets describe how to legalize, see the below description. 1. Interface for targets to describe how to legalize. In GlobalISel, the API in the LegalizerInfo class is the main interface for targets to specify which types are legal for which operations, and what to do to turn illegal type/operation combinations into legal ones. For each operation the type sizes that can be legalized without having to change the size of the type are specified with a call to setAction. This isn't different to how GlobalISel worked before. For example, for a target that supports 32 and 64 bit adds natively: for (auto Ty : {s32, s64}) setAction({G_ADD, 0, s32}, Legal); or for a target that needs a library call for a 32 bit division: setAction({G_SDIV, s32}, Libcall); The main conceptual change to the LegalizerInfo API, is in specifying how to legalize the type sizes for which a change of size is needed. For example, in the above example, how to specify how all types from i1 to i8388607 (apart from s32 and s64 which are legal) need to be legalized and expressed in terms of operations on the available legal sizes (again, i32 and i64 in this case). Before, the implementation only allowed specifying power-of-2-sized types (e.g. setAction({G_ADD, 0, s128}, NarrowScalar). A worse limitation was that if you'd wanted to specify how to legalize all the sized types as allowed by the LLVM-IR LangRef, i1 to i8388607, you'd have to call setAction 8388607-3 times and probably would need a lot of memory to store all of these specifications. Instead, the legalization actions that need to change the size of the type are specified now using a "SizeChangeStrategy". For example: setLegalizeScalarToDifferentSizeStrategy( G_ADD, 0, widenToLargerAndNarrowToLargest); This example indicates that for type sizes for which there is a larger size that can be legalized towards, do it by Widening the size. For example, G_ADD on s17 will be legalized by first doing WidenScalar to make it s32, after which it's legal. The "NarrowToLargest" indicates what to do if there is no larger size that can be legalized towards. E.g. G_ADD on s92 will be legalized by doing NarrowScalar to s64. Another example, taken from the ARM backend is: for (unsigned Op : {G_SDIV, G_UDIV}) { setLegalizeScalarToDifferentSizeStrategy(Op, 0, widenToLargerTypesUnsupportedOtherwise); if (ST.hasDivideInARMMode()) setAction({Op, s32}, Legal); else setAction({Op, s32}, Libcall); } For this example, G_SDIV on s8, on a target without a divide instruction, would be legalized by first doing action (WidenScalar, s32), followed by (Libcall, s32). The same principle is also followed for when the number of vector lanes on vector data types need to be changed, e.g.: setAction({G_ADD, LLT::vector(8, 8)}, LegalizerInfo::Legal); setAction({G_ADD, LLT::vector(16, 8)}, LegalizerInfo::Legal); setAction({G_ADD, LLT::vector(4, 16)}, LegalizerInfo::Legal); setAction({G_ADD, LLT::vector(8, 16)}, LegalizerInfo::Legal); setAction({G_ADD, LLT::vector(2, 32)}, LegalizerInfo::Legal); setAction({G_ADD, LLT::vector(4, 32)}, LegalizerInfo::Legal); setLegalizeVectorElementToDifferentSizeStrategy( G_ADD, 0, widenToLargerTypesUnsupportedOtherwise); As currently implemented here, vector types are legalized by first making the vector element size legal, followed by then making the number of lanes legal. The strategy to follow in the first step is set by a call to setLegalizeVectorElementToDifferentSizeStrategy, see example above. The strategy followed in the second step "moreToWiderTypesAndLessToWidest" (see code for its definition), indicating that vectors are widened to more elements so they map to natively supported vector widths, or when there isn't a legal wider vector, split the vector to map it to the widest vector supported. Therefore, for the above specification, some example legalizations are: * getAction({G_ADD, LLT::vector(3, 3)}) returns {WidenScalar, LLT::vector(3, 8)} * getAction({G_ADD, LLT::vector(3, 8)}) then returns {MoreElements, LLT::vector(8, 8)} * getAction({G_ADD, LLT::vector(20, 8)}) returns {FewerElements, LLT::vector(16, 8)} 2. Key implementation aspects. How to legalize a specific (operation, type index, size) tuple is represented by mapping intervals of integers representing a range of size types to an action to take, e.g.: setScalarAction({G_ADD, LLT:scalar(1)}, {{1, WidenScalar}, // bit sizes [ 1, 31[ {32, Legal}, // bit sizes [32, 33[ {33, WidenScalar}, // bit sizes [33, 64[ {64, Legal}, // bit sizes [64, 65[ {65, NarrowScalar} // bit sizes [65, +inf[ }); Please note that most of the code to do the actual lowering of non-power-of-2 sized types is currently missing, this is just trying to make it possible for targets to specify what is legal, and how non-legal types should be legalized. Probably quite a bit of further work is needed in the actual legalizing and the other passes in GlobalISel to support non-power-of-2 sized types. I hope the documentation in LegalizerInfo.h and the examples provided in the various {Target}LegalizerInfo.cpp and LegalizerInfoTest.cpp explains well enough how this is meant to be used. This drops the need for LLT::{half,double}...Size(). Differential Revision: https://reviews.llvm.org/D30529 llvm-svn: 317560
2017-11-07 18:34:34 +08:00
const unsigned TypeIdx = Aspect.Idx;
if (TypeIdx >= ScalarInVectorActions[OpcodeIdx].size())
return {NotFound, Aspect.Type};
const SizeAndActionsVec &ElemSizeVec =
ScalarInVectorActions[OpcodeIdx][TypeIdx];
LLT IntermediateType;
auto ElementSizeAndAction =
findAction(ElemSizeVec, Aspect.Type.getScalarSizeInBits());
IntermediateType =
LLT::vector(Aspect.Type.getNumElements(), ElementSizeAndAction.first);
if (ElementSizeAndAction.second != Legal)
return {ElementSizeAndAction.second, IntermediateType};
auto i = NumElements2Actions[OpcodeIdx].find(
IntermediateType.getScalarSizeInBits());
if (i == NumElements2Actions[OpcodeIdx].end()) {
return {NotFound, IntermediateType};
}
const SizeAndActionsVec &NumElementsVec = (*i).second[TypeIdx];
auto NumElementsAndAction =
findAction(NumElementsVec, IntermediateType.getNumElements());
return {NumElementsAndAction.second,
LLT::vector(NumElementsAndAction.first,
IntermediateType.getScalarSizeInBits())};
}
[GlobalISel][AArch64] Adding -disable-gisel-legality-check CL option Currently it's impossible to test InstructionSelect pass with MIR which is considered illegal by the Legalizer in Assert builds. In early stages of porting an existing backend from SelectionDAG ISel to GlobalISel, however, we would have very basic CallLowering, Legalizer, and RegBankSelect implementations, but rather functional Instruction Select with quite a few patterns selectable due to the semi-automatic porting process borrowing them from SelectionDAG ISel. As we are trying to define legality as a property of being selectable by the instruction selector, it would be nice to be able to easily check what the selector can do in its current state w/o the legality check provided by the Legalizer getting in the way. It also seems beneficial to have a regression testing set up that would not allow the selector to silently regress in its support of the MIR not supported yet by the previous passes in the GlobalISel pipeline. This commit adds -disable-gisel-legality-check command line option to llc that disables those legality checks in RegBankSelect and InstructionSelect passes. It also adds quite a few MIR test cases for AArch64's Instruction Selector. Every one of them would fail on the legality check at the moment, but will select just fine if the check is disabled. Every test MachineFunction is intended to exercise a specific selection rule and that rule only, encoded in the MachineFunction's name by the rule's number, ID, and index of its GIM_Try opcode in TableGen'erated MatchTable (-optimize-match-table=false). Reviewers: ab, dsanders, qcolombet, rovka Reviewed By: bogner Subscribers: kristof.beyls, volkan, aditya_nandakumar, aemerson, rengolin, t.p.northover, javed.absar, llvm-commits Differential Revision: https://reviews.llvm.org/D42886 llvm-svn: 326396
2018-03-01 08:27:48 +08:00
#ifndef NDEBUG
// FIXME: This should be in the MachineVerifier, but it can't use the
// LegalizerInfo as it's currently in the separate GlobalISel library.
// Note that RegBankSelected property already checked in the verifier
// has the same layering problem, but we only use inline methods so
// end up not needing to link against the GlobalISel library.
const MachineInstr *llvm::machineFunctionIsIllegal(const MachineFunction &MF) {
if (const LegalizerInfo *MLI = MF.getSubtarget().getLegalizerInfo()) {
const MachineRegisterInfo &MRI = MF.getRegInfo();
for (const MachineBasicBlock &MBB : MF)
for (const MachineInstr &MI : MBB)
if (isPreISelGenericOpcode(MI.getOpcode()) && !MLI->isLegal(MI, MRI))
return &MI;
}
return nullptr;
}
#endif