llvm-project/llvm/lib/CodeGen/PeepholeOptimizer.cpp

1984 lines
75 KiB
C++
Raw Normal View History

//===-- PeepholeOptimizer.cpp - Peephole Optimizations --------------------===//
//
// The LLVM Compiler Infrastructure
//
// This file is distributed under the University of Illinois Open Source
// License. See LICENSE.TXT for details.
//
//===----------------------------------------------------------------------===//
//
// Perform peephole optimizations on the machine code:
//
// - Optimize Extensions
//
// Optimization of sign / zero extension instructions. It may be extended to
// handle other instructions with similar properties.
//
// On some targets, some instructions, e.g. X86 sign / zero extension, may
// leave the source value in the lower part of the result. This optimization
// will replace some uses of the pre-extension value with uses of the
// sub-register of the results.
//
// - Optimize Comparisons
//
// Optimization of comparison instructions. For instance, in this code:
//
// sub r1, 1
// cmp r1, 0
// bz L1
//
// If the "sub" instruction all ready sets (or could be modified to set) the
// same flag that the "cmp" instruction sets and that "bz" uses, then we can
// eliminate the "cmp" instruction.
//
// Another instance, in this code:
//
// sub r1, r3 | sub r1, imm
// cmp r3, r1 or cmp r1, r3 | cmp r1, imm
// bge L1
//
// If the branch instruction can use flag from "sub", then we can replace
// "sub" with "subs" and eliminate the "cmp" instruction.
//
// - Optimize Loads:
//
// Loads that can be folded into a later instruction. A load is foldable
2015-09-09 08:38:33 +08:00
// if it loads to virtual registers and the virtual register defined has
// a single use.
//
[PeepholeOptimizer] Refactor the advanced copy optimization to take advantage of the isRegSequence property. This is a follow-up of r215394 and r215404, which respectively introduces the isRegSequence property and uses it for ARM. Thanks to the property introduced by the previous commits, this patch is able to optimize the following sequence: vmov d0, r2, r3 vmov d1, r0, r1 vmov r0, s0 vmov r1, s2 udiv r0, r1, r0 vmov r1, s1 vmov r2, s3 udiv r1, r2, r1 vmov.32 d16[0], r0 vmov.32 d16[1], r1 vmov r0, r1, d16 bx lr into: udiv r0, r0, r2 udiv r1, r1, r3 vmov.32 d16[0], r0 vmov.32 d16[1], r1 vmov r0, r1, d16 bx lr This patch refactors how the copy optimizations are done in the peephole optimizer. Prior to this patch, we had one copy-related optimization that replaced a copy or bitcast by a generic, more suitable (in terms of register file), copy. With this patch, the peephole optimizer features two copy-related optimizations: 1. One for rewriting generic copies to generic copies: PeepholeOptimizer::optimizeCoalescableCopy. 2. One for replacing non-generic copies with generic copies: PeepholeOptimizer::optimizeUncoalescableCopy. The goals of these two optimizations are slightly different: one rewrite the operand of the instruction (#1), the other kills off the non-generic instruction and replace it by a (sequence of) generic instruction(s). Both optimizations rely on the ValueTracker introduced in r212100. The ValueTracker has been refactored to use the information from the TargetInstrInfo for non-generic instruction. As part of the refactoring, we switched the tracking from the index of the definition to the actual register (virtual or physical). This one change is to provide better consistency with register related APIs and to ease the use of the TargetInstrInfo. Moreover, this patch introduces a new helper class CopyRewriter used to ease the rewriting of generic copies (i.e., #1). Finally, this patch adds a dead code elimination pass right after the peephole optimizer to get rid of dead code that may appear after rewriting. This is related to <rdar://problem/12702965>. Review: http://reviews.llvm.org/D4874 llvm-svn: 216088
2014-08-21 01:41:48 +08:00
// - Optimize Copies and Bitcast (more generally, target specific copies):
//
// Rewrite copies and bitcasts to avoid cross register bank copies
// when possible.
// E.g., Consider the following example, where capital and lower
// letters denote different register file:
// b = copy A <-- cross-bank copy
// C = copy b <-- cross-bank copy
// =>
// b = copy A <-- cross-bank copy
// C = copy A <-- same-bank copy
//
// E.g., for bitcast:
// b = bitcast A <-- cross-bank copy
// C = bitcast b <-- cross-bank copy
// =>
// b = bitcast A <-- cross-bank copy
// C = copy A <-- same-bank copy
//===----------------------------------------------------------------------===//
#include "llvm/CodeGen/Passes.h"
#include "llvm/ADT/DenseMap.h"
#include "llvm/ADT/SmallPtrSet.h"
#include "llvm/ADT/SmallSet.h"
#include "llvm/ADT/SmallVector.h"
#include "llvm/ADT/Statistic.h"
#include "llvm/CodeGen/MachineBasicBlock.h"
#include "llvm/CodeGen/MachineDominators.h"
#include "llvm/CodeGen/MachineFunction.h"
#include "llvm/CodeGen/MachineInstr.h"
#include "llvm/CodeGen/MachineInstrBuilder.h"
#include "llvm/CodeGen/MachineOperand.h"
#include "llvm/CodeGen/MachineRegisterInfo.h"
#include "llvm/MC/MCInstrDesc.h"
#include "llvm/Support/CommandLine.h"
#include "llvm/Support/Debug.h"
#include "llvm/Support/ErrorHandling.h"
#include "llvm/Support/raw_ostream.h"
#include "llvm/Target/TargetInstrInfo.h"
#include "llvm/Target/TargetRegisterInfo.h"
#include "llvm/Target/TargetSubtargetInfo.h"
#include <cassert>
#include <cstdint>
#include <memory>
[PeepholeOptimizer] Refactor the advanced copy optimization to take advantage of the isRegSequence property. This is a follow-up of r215394 and r215404, which respectively introduces the isRegSequence property and uses it for ARM. Thanks to the property introduced by the previous commits, this patch is able to optimize the following sequence: vmov d0, r2, r3 vmov d1, r0, r1 vmov r0, s0 vmov r1, s2 udiv r0, r1, r0 vmov r1, s1 vmov r2, s3 udiv r1, r2, r1 vmov.32 d16[0], r0 vmov.32 d16[1], r1 vmov r0, r1, d16 bx lr into: udiv r0, r0, r2 udiv r1, r1, r3 vmov.32 d16[0], r0 vmov.32 d16[1], r1 vmov r0, r1, d16 bx lr This patch refactors how the copy optimizations are done in the peephole optimizer. Prior to this patch, we had one copy-related optimization that replaced a copy or bitcast by a generic, more suitable (in terms of register file), copy. With this patch, the peephole optimizer features two copy-related optimizations: 1. One for rewriting generic copies to generic copies: PeepholeOptimizer::optimizeCoalescableCopy. 2. One for replacing non-generic copies with generic copies: PeepholeOptimizer::optimizeUncoalescableCopy. The goals of these two optimizations are slightly different: one rewrite the operand of the instruction (#1), the other kills off the non-generic instruction and replace it by a (sequence of) generic instruction(s). Both optimizations rely on the ValueTracker introduced in r212100. The ValueTracker has been refactored to use the information from the TargetInstrInfo for non-generic instruction. As part of the refactoring, we switched the tracking from the index of the definition to the actual register (virtual or physical). This one change is to provide better consistency with register related APIs and to ease the use of the TargetInstrInfo. Moreover, this patch introduces a new helper class CopyRewriter used to ease the rewriting of generic copies (i.e., #1). Finally, this patch adds a dead code elimination pass right after the peephole optimizer to get rid of dead code that may appear after rewriting. This is related to <rdar://problem/12702965>. Review: http://reviews.llvm.org/D4874 llvm-svn: 216088
2014-08-21 01:41:48 +08:00
#include <utility>
using namespace llvm;
#define DEBUG_TYPE "peephole-opt"
// Optimize Extensions
static cl::opt<bool>
Aggressive("aggressive-ext-opt", cl::Hidden,
cl::desc("Aggressive extension optimization"));
static cl::opt<bool>
DisablePeephole("disable-peephole", cl::Hidden, cl::init(false),
cl::desc("Disable the peephole optimizer"));
static cl::opt<bool>
DisableAdvCopyOpt("disable-adv-copy-opt", cl::Hidden, cl::init(false),
cl::desc("Disable advanced copy optimization"));
CodeGen peephole: fold redundant phys reg copies Code generation often exposes redundant physical register copies through virtual registers such as: %vreg = COPY %PHYSREG ... %PHYSREG = COPY %vreg There are cases where no intervening clobber of %PHYSREG occurs, and the later copy could therefore be removed. In some cases this further allows us to remove the initial copy. This patch contains a motivating example which comes from the x86 build of Chrome, specifically cc::ResourceProvider::UnlockForRead uses libstdc++'s implementation of hash_map. That example has two tests live at the same time, and after machine sinking LLVM has confused itself enough and things spilling EFLAGS is a great idea even though it's never restored and the comparison results are both live. Before this patch we have: DEC32m %RIP, 1, %noreg, <ga:@L>, %noreg, %EFLAGS<imp-def> %vreg1<def> = COPY %EFLAGS; GR64:%vreg1 %EFLAGS<def> = COPY %vreg1; GR64:%vreg1 JNE_1 <BB#1>, %EFLAGS<imp-use> Both copies are useless. This patch tries to eliminate the later copy in a generic manner. dec is especially confusing to LLVM when compared with sub. I wrote this patch to treat all physical registers generically, but only remove redundant copies of non-allocatable physical registers because the allocatable ones caused issues (e.g. when calling conventions weren't properly modeled) and should be handled later by the register allocator anyways. The following tests used to failed when the patch also replaced allocatable registers: CodeGen/X86/StackColoring.ll CodeGen/X86/avx512-calling-conv.ll CodeGen/X86/copy-propagation.ll CodeGen/X86/inline-asm-fpstack.ll CodeGen/X86/musttail-varargs.ll CodeGen/X86/pop-stack-cleanup.ll CodeGen/X86/preserve_mostcc64.ll CodeGen/X86/tailcallstack64.ll CodeGen/X86/this-return-64.ll This happens because COPY has other special meaning for e.g. dependency breakage and x87 FP stack. Note that all other backends' tests pass. Reviewers: qcolombet Subscribers: llvm-commits Differential Revision: http://reviews.llvm.org/D15157 llvm-svn: 254665
2015-12-04 07:43:56 +08:00
static cl::opt<bool> DisableNAPhysCopyOpt(
"disable-non-allocatable-phys-copy-opt", cl::Hidden, cl::init(false),
cl::desc("Disable non-allocatable physical register copy optimization"));
// Limit the number of PHI instructions to process
// in PeepholeOptimizer::getNextSource.
static cl::opt<unsigned> RewritePHILimit(
"rewrite-phi-limit", cl::Hidden, cl::init(10),
cl::desc("Limit the length of PHI chains to lookup"));
STATISTIC(NumReuse, "Number of extension results reused");
STATISTIC(NumCmps, "Number of compares eliminated");
2012-02-25 08:46:38 +08:00
STATISTIC(NumImmFold, "Number of move immediate folded");
STATISTIC(NumLoadFold, "Number of loads folded");
STATISTIC(NumSelects, "Number of selects optimized");
[PeepholeOptimizer] Refactor the advanced copy optimization to take advantage of the isRegSequence property. This is a follow-up of r215394 and r215404, which respectively introduces the isRegSequence property and uses it for ARM. Thanks to the property introduced by the previous commits, this patch is able to optimize the following sequence: vmov d0, r2, r3 vmov d1, r0, r1 vmov r0, s0 vmov r1, s2 udiv r0, r1, r0 vmov r1, s1 vmov r2, s3 udiv r1, r2, r1 vmov.32 d16[0], r0 vmov.32 d16[1], r1 vmov r0, r1, d16 bx lr into: udiv r0, r0, r2 udiv r1, r1, r3 vmov.32 d16[0], r0 vmov.32 d16[1], r1 vmov r0, r1, d16 bx lr This patch refactors how the copy optimizations are done in the peephole optimizer. Prior to this patch, we had one copy-related optimization that replaced a copy or bitcast by a generic, more suitable (in terms of register file), copy. With this patch, the peephole optimizer features two copy-related optimizations: 1. One for rewriting generic copies to generic copies: PeepholeOptimizer::optimizeCoalescableCopy. 2. One for replacing non-generic copies with generic copies: PeepholeOptimizer::optimizeUncoalescableCopy. The goals of these two optimizations are slightly different: one rewrite the operand of the instruction (#1), the other kills off the non-generic instruction and replace it by a (sequence of) generic instruction(s). Both optimizations rely on the ValueTracker introduced in r212100. The ValueTracker has been refactored to use the information from the TargetInstrInfo for non-generic instruction. As part of the refactoring, we switched the tracking from the index of the definition to the actual register (virtual or physical). This one change is to provide better consistency with register related APIs and to ease the use of the TargetInstrInfo. Moreover, this patch introduces a new helper class CopyRewriter used to ease the rewriting of generic copies (i.e., #1). Finally, this patch adds a dead code elimination pass right after the peephole optimizer to get rid of dead code that may appear after rewriting. This is related to <rdar://problem/12702965>. Review: http://reviews.llvm.org/D4874 llvm-svn: 216088
2014-08-21 01:41:48 +08:00
STATISTIC(NumUncoalescableCopies, "Number of uncoalescable copies optimized");
STATISTIC(NumRewrittenCopies, "Number of copies rewritten");
CodeGen peephole: fold redundant phys reg copies Code generation often exposes redundant physical register copies through virtual registers such as: %vreg = COPY %PHYSREG ... %PHYSREG = COPY %vreg There are cases where no intervening clobber of %PHYSREG occurs, and the later copy could therefore be removed. In some cases this further allows us to remove the initial copy. This patch contains a motivating example which comes from the x86 build of Chrome, specifically cc::ResourceProvider::UnlockForRead uses libstdc++'s implementation of hash_map. That example has two tests live at the same time, and after machine sinking LLVM has confused itself enough and things spilling EFLAGS is a great idea even though it's never restored and the comparison results are both live. Before this patch we have: DEC32m %RIP, 1, %noreg, <ga:@L>, %noreg, %EFLAGS<imp-def> %vreg1<def> = COPY %EFLAGS; GR64:%vreg1 %EFLAGS<def> = COPY %vreg1; GR64:%vreg1 JNE_1 <BB#1>, %EFLAGS<imp-use> Both copies are useless. This patch tries to eliminate the later copy in a generic manner. dec is especially confusing to LLVM when compared with sub. I wrote this patch to treat all physical registers generically, but only remove redundant copies of non-allocatable physical registers because the allocatable ones caused issues (e.g. when calling conventions weren't properly modeled) and should be handled later by the register allocator anyways. The following tests used to failed when the patch also replaced allocatable registers: CodeGen/X86/StackColoring.ll CodeGen/X86/avx512-calling-conv.ll CodeGen/X86/copy-propagation.ll CodeGen/X86/inline-asm-fpstack.ll CodeGen/X86/musttail-varargs.ll CodeGen/X86/pop-stack-cleanup.ll CodeGen/X86/preserve_mostcc64.ll CodeGen/X86/tailcallstack64.ll CodeGen/X86/this-return-64.ll This happens because COPY has other special meaning for e.g. dependency breakage and x87 FP stack. Note that all other backends' tests pass. Reviewers: qcolombet Subscribers: llvm-commits Differential Revision: http://reviews.llvm.org/D15157 llvm-svn: 254665
2015-12-04 07:43:56 +08:00
STATISTIC(NumNAPhysCopies, "Number of non-allocatable physical copies removed");
namespace {
class ValueTrackerResult;
class PeepholeOptimizer : public MachineFunctionPass {
const TargetInstrInfo *TII;
const TargetRegisterInfo *TRI;
MachineRegisterInfo *MRI;
MachineDominatorTree *DT; // Machine dominator tree
public:
static char ID; // Pass identification
PeepholeOptimizer() : MachineFunctionPass(ID) {
initializePeepholeOptimizerPass(*PassRegistry::getPassRegistry());
}
bool runOnMachineFunction(MachineFunction &MF) override;
void getAnalysisUsage(AnalysisUsage &AU) const override {
AU.setPreservesCFG();
MachineFunctionPass::getAnalysisUsage(AU);
if (Aggressive) {
AU.addRequired<MachineDominatorTree>();
AU.addPreserved<MachineDominatorTree>();
}
}
/// \brief Track Def -> Use info used for rewriting copies.
typedef SmallDenseMap<TargetInstrInfo::RegSubRegPair, ValueTrackerResult>
RewriteMapTy;
private:
bool optimizeCmpInstr(MachineInstr *MI, MachineBasicBlock *MBB);
bool optimizeExtInstr(MachineInstr *MI, MachineBasicBlock *MBB,
SmallPtrSetImpl<MachineInstr*> &LocalMIs);
bool optimizeSelect(MachineInstr *MI,
SmallPtrSetImpl<MachineInstr *> &LocalMIs);
bool optimizeCondBranch(MachineInstr *MI);
[PeepholeOptimizer] Refactor the advanced copy optimization to take advantage of the isRegSequence property. This is a follow-up of r215394 and r215404, which respectively introduces the isRegSequence property and uses it for ARM. Thanks to the property introduced by the previous commits, this patch is able to optimize the following sequence: vmov d0, r2, r3 vmov d1, r0, r1 vmov r0, s0 vmov r1, s2 udiv r0, r1, r0 vmov r1, s1 vmov r2, s3 udiv r1, r2, r1 vmov.32 d16[0], r0 vmov.32 d16[1], r1 vmov r0, r1, d16 bx lr into: udiv r0, r0, r2 udiv r1, r1, r3 vmov.32 d16[0], r0 vmov.32 d16[1], r1 vmov r0, r1, d16 bx lr This patch refactors how the copy optimizations are done in the peephole optimizer. Prior to this patch, we had one copy-related optimization that replaced a copy or bitcast by a generic, more suitable (in terms of register file), copy. With this patch, the peephole optimizer features two copy-related optimizations: 1. One for rewriting generic copies to generic copies: PeepholeOptimizer::optimizeCoalescableCopy. 2. One for replacing non-generic copies with generic copies: PeepholeOptimizer::optimizeUncoalescableCopy. The goals of these two optimizations are slightly different: one rewrite the operand of the instruction (#1), the other kills off the non-generic instruction and replace it by a (sequence of) generic instruction(s). Both optimizations rely on the ValueTracker introduced in r212100. The ValueTracker has been refactored to use the information from the TargetInstrInfo for non-generic instruction. As part of the refactoring, we switched the tracking from the index of the definition to the actual register (virtual or physical). This one change is to provide better consistency with register related APIs and to ease the use of the TargetInstrInfo. Moreover, this patch introduces a new helper class CopyRewriter used to ease the rewriting of generic copies (i.e., #1). Finally, this patch adds a dead code elimination pass right after the peephole optimizer to get rid of dead code that may appear after rewriting. This is related to <rdar://problem/12702965>. Review: http://reviews.llvm.org/D4874 llvm-svn: 216088
2014-08-21 01:41:48 +08:00
bool optimizeCoalescableCopy(MachineInstr *MI);
bool optimizeUncoalescableCopy(MachineInstr *MI,
SmallPtrSetImpl<MachineInstr *> &LocalMIs);
bool findNextSource(unsigned Reg, unsigned SubReg,
RewriteMapTy &RewriteMap);
bool isMoveImmediate(MachineInstr *MI,
SmallSet<unsigned, 4> &ImmDefRegs,
DenseMap<unsigned, MachineInstr*> &ImmDefMIs);
bool foldImmediate(MachineInstr *MI, MachineBasicBlock *MBB,
SmallSet<unsigned, 4> &ImmDefRegs,
DenseMap<unsigned, MachineInstr*> &ImmDefMIs);
/// \brief If copy instruction \p MI is a virtual register copy, track it in
CodeGen peephole: fold redundant phys reg copies Code generation often exposes redundant physical register copies through virtual registers such as: %vreg = COPY %PHYSREG ... %PHYSREG = COPY %vreg There are cases where no intervening clobber of %PHYSREG occurs, and the later copy could therefore be removed. In some cases this further allows us to remove the initial copy. This patch contains a motivating example which comes from the x86 build of Chrome, specifically cc::ResourceProvider::UnlockForRead uses libstdc++'s implementation of hash_map. That example has two tests live at the same time, and after machine sinking LLVM has confused itself enough and things spilling EFLAGS is a great idea even though it's never restored and the comparison results are both live. Before this patch we have: DEC32m %RIP, 1, %noreg, <ga:@L>, %noreg, %EFLAGS<imp-def> %vreg1<def> = COPY %EFLAGS; GR64:%vreg1 %EFLAGS<def> = COPY %vreg1; GR64:%vreg1 JNE_1 <BB#1>, %EFLAGS<imp-use> Both copies are useless. This patch tries to eliminate the later copy in a generic manner. dec is especially confusing to LLVM when compared with sub. I wrote this patch to treat all physical registers generically, but only remove redundant copies of non-allocatable physical registers because the allocatable ones caused issues (e.g. when calling conventions weren't properly modeled) and should be handled later by the register allocator anyways. The following tests used to failed when the patch also replaced allocatable registers: CodeGen/X86/StackColoring.ll CodeGen/X86/avx512-calling-conv.ll CodeGen/X86/copy-propagation.ll CodeGen/X86/inline-asm-fpstack.ll CodeGen/X86/musttail-varargs.ll CodeGen/X86/pop-stack-cleanup.ll CodeGen/X86/preserve_mostcc64.ll CodeGen/X86/tailcallstack64.ll CodeGen/X86/this-return-64.ll This happens because COPY has other special meaning for e.g. dependency breakage and x87 FP stack. Note that all other backends' tests pass. Reviewers: qcolombet Subscribers: llvm-commits Differential Revision: http://reviews.llvm.org/D15157 llvm-svn: 254665
2015-12-04 07:43:56 +08:00
/// the set \p CopySrcRegs and \p CopyMIs. If this virtual register was
/// previously seen as a copy, replace the uses of this copy with the
/// previously seen copy's destination register.
bool foldRedundantCopy(MachineInstr *MI,
CodeGen peephole: fold redundant phys reg copies Code generation often exposes redundant physical register copies through virtual registers such as: %vreg = COPY %PHYSREG ... %PHYSREG = COPY %vreg There are cases where no intervening clobber of %PHYSREG occurs, and the later copy could therefore be removed. In some cases this further allows us to remove the initial copy. This patch contains a motivating example which comes from the x86 build of Chrome, specifically cc::ResourceProvider::UnlockForRead uses libstdc++'s implementation of hash_map. That example has two tests live at the same time, and after machine sinking LLVM has confused itself enough and things spilling EFLAGS is a great idea even though it's never restored and the comparison results are both live. Before this patch we have: DEC32m %RIP, 1, %noreg, <ga:@L>, %noreg, %EFLAGS<imp-def> %vreg1<def> = COPY %EFLAGS; GR64:%vreg1 %EFLAGS<def> = COPY %vreg1; GR64:%vreg1 JNE_1 <BB#1>, %EFLAGS<imp-use> Both copies are useless. This patch tries to eliminate the later copy in a generic manner. dec is especially confusing to LLVM when compared with sub. I wrote this patch to treat all physical registers generically, but only remove redundant copies of non-allocatable physical registers because the allocatable ones caused issues (e.g. when calling conventions weren't properly modeled) and should be handled later by the register allocator anyways. The following tests used to failed when the patch also replaced allocatable registers: CodeGen/X86/StackColoring.ll CodeGen/X86/avx512-calling-conv.ll CodeGen/X86/copy-propagation.ll CodeGen/X86/inline-asm-fpstack.ll CodeGen/X86/musttail-varargs.ll CodeGen/X86/pop-stack-cleanup.ll CodeGen/X86/preserve_mostcc64.ll CodeGen/X86/tailcallstack64.ll CodeGen/X86/this-return-64.ll This happens because COPY has other special meaning for e.g. dependency breakage and x87 FP stack. Note that all other backends' tests pass. Reviewers: qcolombet Subscribers: llvm-commits Differential Revision: http://reviews.llvm.org/D15157 llvm-svn: 254665
2015-12-04 07:43:56 +08:00
SmallSet<unsigned, 4> &CopySrcRegs,
DenseMap<unsigned, MachineInstr *> &CopyMIs);
/// \brief Is the register \p Reg a non-allocatable physical register?
bool isNAPhysCopy(unsigned Reg);
/// \brief If copy instruction \p MI is a non-allocatable virtual<->physical
/// register copy, track it in the \p NAPhysToVirtMIs map. If this
/// non-allocatable physical register was previously copied to a virtual
/// registered and hasn't been clobbered, the virt->phys copy can be
/// deleted.
bool foldRedundantNAPhysCopy(
MachineInstr *MI,
DenseMap<unsigned, MachineInstr *> &NAPhysToVirtMIs);
bool isLoadFoldable(MachineInstr *MI,
SmallSet<unsigned, 16> &FoldAsLoadDefCandidates);
[PeepholeOptimizer] Refactor the advanced copy optimization to take advantage of the isRegSequence property. This is a follow-up of r215394 and r215404, which respectively introduces the isRegSequence property and uses it for ARM. Thanks to the property introduced by the previous commits, this patch is able to optimize the following sequence: vmov d0, r2, r3 vmov d1, r0, r1 vmov r0, s0 vmov r1, s2 udiv r0, r1, r0 vmov r1, s1 vmov r2, s3 udiv r1, r2, r1 vmov.32 d16[0], r0 vmov.32 d16[1], r1 vmov r0, r1, d16 bx lr into: udiv r0, r0, r2 udiv r1, r1, r3 vmov.32 d16[0], r0 vmov.32 d16[1], r1 vmov r0, r1, d16 bx lr This patch refactors how the copy optimizations are done in the peephole optimizer. Prior to this patch, we had one copy-related optimization that replaced a copy or bitcast by a generic, more suitable (in terms of register file), copy. With this patch, the peephole optimizer features two copy-related optimizations: 1. One for rewriting generic copies to generic copies: PeepholeOptimizer::optimizeCoalescableCopy. 2. One for replacing non-generic copies with generic copies: PeepholeOptimizer::optimizeUncoalescableCopy. The goals of these two optimizations are slightly different: one rewrite the operand of the instruction (#1), the other kills off the non-generic instruction and replace it by a (sequence of) generic instruction(s). Both optimizations rely on the ValueTracker introduced in r212100. The ValueTracker has been refactored to use the information from the TargetInstrInfo for non-generic instruction. As part of the refactoring, we switched the tracking from the index of the definition to the actual register (virtual or physical). This one change is to provide better consistency with register related APIs and to ease the use of the TargetInstrInfo. Moreover, this patch introduces a new helper class CopyRewriter used to ease the rewriting of generic copies (i.e., #1). Finally, this patch adds a dead code elimination pass right after the peephole optimizer to get rid of dead code that may appear after rewriting. This is related to <rdar://problem/12702965>. Review: http://reviews.llvm.org/D4874 llvm-svn: 216088
2014-08-21 01:41:48 +08:00
/// \brief Check whether \p MI is understood by the register coalescer
/// but may require some rewriting.
bool isCoalescableCopy(const MachineInstr &MI) {
// SubregToRegs are not interesting, because they are already register
// coalescer friendly.
return MI.isCopy() || (!DisableAdvCopyOpt &&
(MI.isRegSequence() || MI.isInsertSubreg() ||
MI.isExtractSubreg()));
}
/// \brief Check whether \p MI is a copy like instruction that is
/// not recognized by the register coalescer.
bool isUncoalescableCopy(const MachineInstr &MI) {
return MI.isBitcast() ||
(!DisableAdvCopyOpt &&
(MI.isRegSequenceLike() || MI.isInsertSubregLike() ||
MI.isExtractSubregLike()));
[PeepholeOptimizer] Refactor the advanced copy optimization to take advantage of the isRegSequence property. This is a follow-up of r215394 and r215404, which respectively introduces the isRegSequence property and uses it for ARM. Thanks to the property introduced by the previous commits, this patch is able to optimize the following sequence: vmov d0, r2, r3 vmov d1, r0, r1 vmov r0, s0 vmov r1, s2 udiv r0, r1, r0 vmov r1, s1 vmov r2, s3 udiv r1, r2, r1 vmov.32 d16[0], r0 vmov.32 d16[1], r1 vmov r0, r1, d16 bx lr into: udiv r0, r0, r2 udiv r1, r1, r3 vmov.32 d16[0], r0 vmov.32 d16[1], r1 vmov r0, r1, d16 bx lr This patch refactors how the copy optimizations are done in the peephole optimizer. Prior to this patch, we had one copy-related optimization that replaced a copy or bitcast by a generic, more suitable (in terms of register file), copy. With this patch, the peephole optimizer features two copy-related optimizations: 1. One for rewriting generic copies to generic copies: PeepholeOptimizer::optimizeCoalescableCopy. 2. One for replacing non-generic copies with generic copies: PeepholeOptimizer::optimizeUncoalescableCopy. The goals of these two optimizations are slightly different: one rewrite the operand of the instruction (#1), the other kills off the non-generic instruction and replace it by a (sequence of) generic instruction(s). Both optimizations rely on the ValueTracker introduced in r212100. The ValueTracker has been refactored to use the information from the TargetInstrInfo for non-generic instruction. As part of the refactoring, we switched the tracking from the index of the definition to the actual register (virtual or physical). This one change is to provide better consistency with register related APIs and to ease the use of the TargetInstrInfo. Moreover, this patch introduces a new helper class CopyRewriter used to ease the rewriting of generic copies (i.e., #1). Finally, this patch adds a dead code elimination pass right after the peephole optimizer to get rid of dead code that may appear after rewriting. This is related to <rdar://problem/12702965>. Review: http://reviews.llvm.org/D4874 llvm-svn: 216088
2014-08-21 01:41:48 +08:00
}
};
/// \brief Helper class to hold a reply for ValueTracker queries. Contains the
/// returned sources for a given search and the instructions where the sources
/// were tracked from.
class ValueTrackerResult {
private:
/// Track all sources found by one ValueTracker query.
SmallVector<TargetInstrInfo::RegSubRegPair, 2> RegSrcs;
/// Instruction using the sources in 'RegSrcs'.
const MachineInstr *Inst;
public:
ValueTrackerResult() : Inst(nullptr) {}
ValueTrackerResult(unsigned Reg, unsigned SubReg) : Inst(nullptr) {
addSource(Reg, SubReg);
}
bool isValid() const { return getNumSources() > 0; }
void setInst(const MachineInstr *I) { Inst = I; }
const MachineInstr *getInst() const { return Inst; }
void clear() {
RegSrcs.clear();
Inst = nullptr;
}
void addSource(unsigned SrcReg, unsigned SrcSubReg) {
RegSrcs.push_back(TargetInstrInfo::RegSubRegPair(SrcReg, SrcSubReg));
}
void setSource(int Idx, unsigned SrcReg, unsigned SrcSubReg) {
assert(Idx < getNumSources() && "Reg pair source out of index");
RegSrcs[Idx] = TargetInstrInfo::RegSubRegPair(SrcReg, SrcSubReg);
}
int getNumSources() const { return RegSrcs.size(); }
unsigned getSrcReg(int Idx) const {
assert(Idx < getNumSources() && "Reg source out of index");
return RegSrcs[Idx].Reg;
}
unsigned getSrcSubReg(int Idx) const {
assert(Idx < getNumSources() && "SubReg source out of index");
return RegSrcs[Idx].SubReg;
}
bool operator==(const ValueTrackerResult &Other) {
if (Other.getInst() != getInst())
return false;
if (Other.getNumSources() != getNumSources())
return false;
for (int i = 0, e = Other.getNumSources(); i != e; ++i)
if (Other.getSrcReg(i) != getSrcReg(i) ||
Other.getSrcSubReg(i) != getSrcSubReg(i))
return false;
return true;
}
};
/// \brief Helper class to track the possible sources of a value defined by
/// a (chain of) copy related instructions.
/// Given a definition (instruction and definition index), this class
/// follows the use-def chain to find successive suitable sources.
/// The given source can be used to rewrite the definition into
/// def = COPY src.
///
/// For instance, let us consider the following snippet:
/// v0 =
/// v2 = INSERT_SUBREG v1, v0, sub0
/// def = COPY v2.sub0
///
/// Using a ValueTracker for def = COPY v2.sub0 will give the following
/// suitable sources:
/// v2.sub0 and v0.
/// Then, def can be rewritten into def = COPY v0.
class ValueTracker {
private:
/// The current point into the use-def chain.
const MachineInstr *Def;
/// The index of the definition in Def.
unsigned DefIdx;
/// The sub register index of the definition.
unsigned DefSubReg;
/// The register where the value can be found.
unsigned Reg;
/// Specifiy whether or not the value tracking looks through
/// complex instructions. When this is false, the value tracker
/// bails on everything that is not a copy or a bitcast.
///
/// Note: This could have been implemented as a specialized version of
/// the ValueTracker class but that would have complicated the code of
/// the users of this class.
bool UseAdvancedTracking;
[PeepholeOptimizer] Refactor the advanced copy optimization to take advantage of the isRegSequence property. This is a follow-up of r215394 and r215404, which respectively introduces the isRegSequence property and uses it for ARM. Thanks to the property introduced by the previous commits, this patch is able to optimize the following sequence: vmov d0, r2, r3 vmov d1, r0, r1 vmov r0, s0 vmov r1, s2 udiv r0, r1, r0 vmov r1, s1 vmov r2, s3 udiv r1, r2, r1 vmov.32 d16[0], r0 vmov.32 d16[1], r1 vmov r0, r1, d16 bx lr into: udiv r0, r0, r2 udiv r1, r1, r3 vmov.32 d16[0], r0 vmov.32 d16[1], r1 vmov r0, r1, d16 bx lr This patch refactors how the copy optimizations are done in the peephole optimizer. Prior to this patch, we had one copy-related optimization that replaced a copy or bitcast by a generic, more suitable (in terms of register file), copy. With this patch, the peephole optimizer features two copy-related optimizations: 1. One for rewriting generic copies to generic copies: PeepholeOptimizer::optimizeCoalescableCopy. 2. One for replacing non-generic copies with generic copies: PeepholeOptimizer::optimizeUncoalescableCopy. The goals of these two optimizations are slightly different: one rewrite the operand of the instruction (#1), the other kills off the non-generic instruction and replace it by a (sequence of) generic instruction(s). Both optimizations rely on the ValueTracker introduced in r212100. The ValueTracker has been refactored to use the information from the TargetInstrInfo for non-generic instruction. As part of the refactoring, we switched the tracking from the index of the definition to the actual register (virtual or physical). This one change is to provide better consistency with register related APIs and to ease the use of the TargetInstrInfo. Moreover, this patch introduces a new helper class CopyRewriter used to ease the rewriting of generic copies (i.e., #1). Finally, this patch adds a dead code elimination pass right after the peephole optimizer to get rid of dead code that may appear after rewriting. This is related to <rdar://problem/12702965>. Review: http://reviews.llvm.org/D4874 llvm-svn: 216088
2014-08-21 01:41:48 +08:00
/// MachineRegisterInfo used to perform tracking.
const MachineRegisterInfo &MRI;
/// Optional TargetInstrInfo used to perform some complex
/// tracking.
[PeepholeOptimizer] Refactor the advanced copy optimization to take advantage of the isRegSequence property. This is a follow-up of r215394 and r215404, which respectively introduces the isRegSequence property and uses it for ARM. Thanks to the property introduced by the previous commits, this patch is able to optimize the following sequence: vmov d0, r2, r3 vmov d1, r0, r1 vmov r0, s0 vmov r1, s2 udiv r0, r1, r0 vmov r1, s1 vmov r2, s3 udiv r1, r2, r1 vmov.32 d16[0], r0 vmov.32 d16[1], r1 vmov r0, r1, d16 bx lr into: udiv r0, r0, r2 udiv r1, r1, r3 vmov.32 d16[0], r0 vmov.32 d16[1], r1 vmov r0, r1, d16 bx lr This patch refactors how the copy optimizations are done in the peephole optimizer. Prior to this patch, we had one copy-related optimization that replaced a copy or bitcast by a generic, more suitable (in terms of register file), copy. With this patch, the peephole optimizer features two copy-related optimizations: 1. One for rewriting generic copies to generic copies: PeepholeOptimizer::optimizeCoalescableCopy. 2. One for replacing non-generic copies with generic copies: PeepholeOptimizer::optimizeUncoalescableCopy. The goals of these two optimizations are slightly different: one rewrite the operand of the instruction (#1), the other kills off the non-generic instruction and replace it by a (sequence of) generic instruction(s). Both optimizations rely on the ValueTracker introduced in r212100. The ValueTracker has been refactored to use the information from the TargetInstrInfo for non-generic instruction. As part of the refactoring, we switched the tracking from the index of the definition to the actual register (virtual or physical). This one change is to provide better consistency with register related APIs and to ease the use of the TargetInstrInfo. Moreover, this patch introduces a new helper class CopyRewriter used to ease the rewriting of generic copies (i.e., #1). Finally, this patch adds a dead code elimination pass right after the peephole optimizer to get rid of dead code that may appear after rewriting. This is related to <rdar://problem/12702965>. Review: http://reviews.llvm.org/D4874 llvm-svn: 216088
2014-08-21 01:41:48 +08:00
const TargetInstrInfo *TII;
/// \brief Dispatcher to the right underlying implementation of
/// getNextSource.
ValueTrackerResult getNextSourceImpl();
/// \brief Specialized version of getNextSource for Copy instructions.
ValueTrackerResult getNextSourceFromCopy();
/// \brief Specialized version of getNextSource for Bitcast instructions.
ValueTrackerResult getNextSourceFromBitcast();
/// \brief Specialized version of getNextSource for RegSequence
/// instructions.
ValueTrackerResult getNextSourceFromRegSequence();
/// \brief Specialized version of getNextSource for InsertSubreg
/// instructions.
ValueTrackerResult getNextSourceFromInsertSubreg();
/// \brief Specialized version of getNextSource for ExtractSubreg
/// instructions.
ValueTrackerResult getNextSourceFromExtractSubreg();
/// \brief Specialized version of getNextSource for SubregToReg
/// instructions.
ValueTrackerResult getNextSourceFromSubregToReg();
/// \brief Specialized version of getNextSource for PHI instructions.
ValueTrackerResult getNextSourceFromPHI();
public:
[PeepholeOptimizer] Refactor the advanced copy optimization to take advantage of the isRegSequence property. This is a follow-up of r215394 and r215404, which respectively introduces the isRegSequence property and uses it for ARM. Thanks to the property introduced by the previous commits, this patch is able to optimize the following sequence: vmov d0, r2, r3 vmov d1, r0, r1 vmov r0, s0 vmov r1, s2 udiv r0, r1, r0 vmov r1, s1 vmov r2, s3 udiv r1, r2, r1 vmov.32 d16[0], r0 vmov.32 d16[1], r1 vmov r0, r1, d16 bx lr into: udiv r0, r0, r2 udiv r1, r1, r3 vmov.32 d16[0], r0 vmov.32 d16[1], r1 vmov r0, r1, d16 bx lr This patch refactors how the copy optimizations are done in the peephole optimizer. Prior to this patch, we had one copy-related optimization that replaced a copy or bitcast by a generic, more suitable (in terms of register file), copy. With this patch, the peephole optimizer features two copy-related optimizations: 1. One for rewriting generic copies to generic copies: PeepholeOptimizer::optimizeCoalescableCopy. 2. One for replacing non-generic copies with generic copies: PeepholeOptimizer::optimizeUncoalescableCopy. The goals of these two optimizations are slightly different: one rewrite the operand of the instruction (#1), the other kills off the non-generic instruction and replace it by a (sequence of) generic instruction(s). Both optimizations rely on the ValueTracker introduced in r212100. The ValueTracker has been refactored to use the information from the TargetInstrInfo for non-generic instruction. As part of the refactoring, we switched the tracking from the index of the definition to the actual register (virtual or physical). This one change is to provide better consistency with register related APIs and to ease the use of the TargetInstrInfo. Moreover, this patch introduces a new helper class CopyRewriter used to ease the rewriting of generic copies (i.e., #1). Finally, this patch adds a dead code elimination pass right after the peephole optimizer to get rid of dead code that may appear after rewriting. This is related to <rdar://problem/12702965>. Review: http://reviews.llvm.org/D4874 llvm-svn: 216088
2014-08-21 01:41:48 +08:00
/// \brief Create a ValueTracker instance for the value defined by \p Reg.
/// \p DefSubReg represents the sub register index the value tracker will
[PeepholeOptimizer] Refactor the advanced copy optimization to take advantage of the isRegSequence property. This is a follow-up of r215394 and r215404, which respectively introduces the isRegSequence property and uses it for ARM. Thanks to the property introduced by the previous commits, this patch is able to optimize the following sequence: vmov d0, r2, r3 vmov d1, r0, r1 vmov r0, s0 vmov r1, s2 udiv r0, r1, r0 vmov r1, s1 vmov r2, s3 udiv r1, r2, r1 vmov.32 d16[0], r0 vmov.32 d16[1], r1 vmov r0, r1, d16 bx lr into: udiv r0, r0, r2 udiv r1, r1, r3 vmov.32 d16[0], r0 vmov.32 d16[1], r1 vmov r0, r1, d16 bx lr This patch refactors how the copy optimizations are done in the peephole optimizer. Prior to this patch, we had one copy-related optimization that replaced a copy or bitcast by a generic, more suitable (in terms of register file), copy. With this patch, the peephole optimizer features two copy-related optimizations: 1. One for rewriting generic copies to generic copies: PeepholeOptimizer::optimizeCoalescableCopy. 2. One for replacing non-generic copies with generic copies: PeepholeOptimizer::optimizeUncoalescableCopy. The goals of these two optimizations are slightly different: one rewrite the operand of the instruction (#1), the other kills off the non-generic instruction and replace it by a (sequence of) generic instruction(s). Both optimizations rely on the ValueTracker introduced in r212100. The ValueTracker has been refactored to use the information from the TargetInstrInfo for non-generic instruction. As part of the refactoring, we switched the tracking from the index of the definition to the actual register (virtual or physical). This one change is to provide better consistency with register related APIs and to ease the use of the TargetInstrInfo. Moreover, this patch introduces a new helper class CopyRewriter used to ease the rewriting of generic copies (i.e., #1). Finally, this patch adds a dead code elimination pass right after the peephole optimizer to get rid of dead code that may appear after rewriting. This is related to <rdar://problem/12702965>. Review: http://reviews.llvm.org/D4874 llvm-svn: 216088
2014-08-21 01:41:48 +08:00
/// track. It does not need to match the sub register index used in the
/// definition of \p Reg.
/// \p UseAdvancedTracking specifies whether or not the value tracker looks
/// through complex instructions. By default (false), it handles only copy
/// and bitcast instructions.
[PeepholeOptimizer] Refactor the advanced copy optimization to take advantage of the isRegSequence property. This is a follow-up of r215394 and r215404, which respectively introduces the isRegSequence property and uses it for ARM. Thanks to the property introduced by the previous commits, this patch is able to optimize the following sequence: vmov d0, r2, r3 vmov d1, r0, r1 vmov r0, s0 vmov r1, s2 udiv r0, r1, r0 vmov r1, s1 vmov r2, s3 udiv r1, r2, r1 vmov.32 d16[0], r0 vmov.32 d16[1], r1 vmov r0, r1, d16 bx lr into: udiv r0, r0, r2 udiv r1, r1, r3 vmov.32 d16[0], r0 vmov.32 d16[1], r1 vmov r0, r1, d16 bx lr This patch refactors how the copy optimizations are done in the peephole optimizer. Prior to this patch, we had one copy-related optimization that replaced a copy or bitcast by a generic, more suitable (in terms of register file), copy. With this patch, the peephole optimizer features two copy-related optimizations: 1. One for rewriting generic copies to generic copies: PeepholeOptimizer::optimizeCoalescableCopy. 2. One for replacing non-generic copies with generic copies: PeepholeOptimizer::optimizeUncoalescableCopy. The goals of these two optimizations are slightly different: one rewrite the operand of the instruction (#1), the other kills off the non-generic instruction and replace it by a (sequence of) generic instruction(s). Both optimizations rely on the ValueTracker introduced in r212100. The ValueTracker has been refactored to use the information from the TargetInstrInfo for non-generic instruction. As part of the refactoring, we switched the tracking from the index of the definition to the actual register (virtual or physical). This one change is to provide better consistency with register related APIs and to ease the use of the TargetInstrInfo. Moreover, this patch introduces a new helper class CopyRewriter used to ease the rewriting of generic copies (i.e., #1). Finally, this patch adds a dead code elimination pass right after the peephole optimizer to get rid of dead code that may appear after rewriting. This is related to <rdar://problem/12702965>. Review: http://reviews.llvm.org/D4874 llvm-svn: 216088
2014-08-21 01:41:48 +08:00
/// If \p Reg is a physical register, a value tracker constructed with
/// this constructor will not find any alternative source.
/// Indeed, when \p Reg is a physical register that constructor does not
/// know which definition of \p Reg it should track.
/// Use the next constructor to track a physical register.
ValueTracker(unsigned Reg, unsigned DefSubReg,
const MachineRegisterInfo &MRI,
bool UseAdvancedTracking = false,
const TargetInstrInfo *TII = nullptr)
: Def(nullptr), DefIdx(0), DefSubReg(DefSubReg), Reg(Reg),
UseAdvancedTracking(UseAdvancedTracking), MRI(MRI), TII(TII) {
if (!TargetRegisterInfo::isPhysicalRegister(Reg)) {
Def = MRI.getVRegDef(Reg);
DefIdx = MRI.def_begin(Reg).getOperandNo();
}
}
/// \brief Create a ValueTracker instance for the value defined by
/// the pair \p MI, \p DefIdx.
/// Unlike the other constructor, the value tracker produced by this one
/// may be able to find a new source when the definition is a physical
/// register.
/// This could be useful to rewrite target specific instructions into
/// generic copy instructions.
ValueTracker(const MachineInstr &MI, unsigned DefIdx, unsigned DefSubReg,
[PeepholeOptimizer] Refactor the advanced copy optimization to take advantage of the isRegSequence property. This is a follow-up of r215394 and r215404, which respectively introduces the isRegSequence property and uses it for ARM. Thanks to the property introduced by the previous commits, this patch is able to optimize the following sequence: vmov d0, r2, r3 vmov d1, r0, r1 vmov r0, s0 vmov r1, s2 udiv r0, r1, r0 vmov r1, s1 vmov r2, s3 udiv r1, r2, r1 vmov.32 d16[0], r0 vmov.32 d16[1], r1 vmov r0, r1, d16 bx lr into: udiv r0, r0, r2 udiv r1, r1, r3 vmov.32 d16[0], r0 vmov.32 d16[1], r1 vmov r0, r1, d16 bx lr This patch refactors how the copy optimizations are done in the peephole optimizer. Prior to this patch, we had one copy-related optimization that replaced a copy or bitcast by a generic, more suitable (in terms of register file), copy. With this patch, the peephole optimizer features two copy-related optimizations: 1. One for rewriting generic copies to generic copies: PeepholeOptimizer::optimizeCoalescableCopy. 2. One for replacing non-generic copies with generic copies: PeepholeOptimizer::optimizeUncoalescableCopy. The goals of these two optimizations are slightly different: one rewrite the operand of the instruction (#1), the other kills off the non-generic instruction and replace it by a (sequence of) generic instruction(s). Both optimizations rely on the ValueTracker introduced in r212100. The ValueTracker has been refactored to use the information from the TargetInstrInfo for non-generic instruction. As part of the refactoring, we switched the tracking from the index of the definition to the actual register (virtual or physical). This one change is to provide better consistency with register related APIs and to ease the use of the TargetInstrInfo. Moreover, this patch introduces a new helper class CopyRewriter used to ease the rewriting of generic copies (i.e., #1). Finally, this patch adds a dead code elimination pass right after the peephole optimizer to get rid of dead code that may appear after rewriting. This is related to <rdar://problem/12702965>. Review: http://reviews.llvm.org/D4874 llvm-svn: 216088
2014-08-21 01:41:48 +08:00
const MachineRegisterInfo &MRI,
bool UseAdvancedTracking = false,
[PeepholeOptimizer] Refactor the advanced copy optimization to take advantage of the isRegSequence property. This is a follow-up of r215394 and r215404, which respectively introduces the isRegSequence property and uses it for ARM. Thanks to the property introduced by the previous commits, this patch is able to optimize the following sequence: vmov d0, r2, r3 vmov d1, r0, r1 vmov r0, s0 vmov r1, s2 udiv r0, r1, r0 vmov r1, s1 vmov r2, s3 udiv r1, r2, r1 vmov.32 d16[0], r0 vmov.32 d16[1], r1 vmov r0, r1, d16 bx lr into: udiv r0, r0, r2 udiv r1, r1, r3 vmov.32 d16[0], r0 vmov.32 d16[1], r1 vmov r0, r1, d16 bx lr This patch refactors how the copy optimizations are done in the peephole optimizer. Prior to this patch, we had one copy-related optimization that replaced a copy or bitcast by a generic, more suitable (in terms of register file), copy. With this patch, the peephole optimizer features two copy-related optimizations: 1. One for rewriting generic copies to generic copies: PeepholeOptimizer::optimizeCoalescableCopy. 2. One for replacing non-generic copies with generic copies: PeepholeOptimizer::optimizeUncoalescableCopy. The goals of these two optimizations are slightly different: one rewrite the operand of the instruction (#1), the other kills off the non-generic instruction and replace it by a (sequence of) generic instruction(s). Both optimizations rely on the ValueTracker introduced in r212100. The ValueTracker has been refactored to use the information from the TargetInstrInfo for non-generic instruction. As part of the refactoring, we switched the tracking from the index of the definition to the actual register (virtual or physical). This one change is to provide better consistency with register related APIs and to ease the use of the TargetInstrInfo. Moreover, this patch introduces a new helper class CopyRewriter used to ease the rewriting of generic copies (i.e., #1). Finally, this patch adds a dead code elimination pass right after the peephole optimizer to get rid of dead code that may appear after rewriting. This is related to <rdar://problem/12702965>. Review: http://reviews.llvm.org/D4874 llvm-svn: 216088
2014-08-21 01:41:48 +08:00
const TargetInstrInfo *TII = nullptr)
: Def(&MI), DefIdx(DefIdx), DefSubReg(DefSubReg),
[PeepholeOptimizer] Refactor the advanced copy optimization to take advantage of the isRegSequence property. This is a follow-up of r215394 and r215404, which respectively introduces the isRegSequence property and uses it for ARM. Thanks to the property introduced by the previous commits, this patch is able to optimize the following sequence: vmov d0, r2, r3 vmov d1, r0, r1 vmov r0, s0 vmov r1, s2 udiv r0, r1, r0 vmov r1, s1 vmov r2, s3 udiv r1, r2, r1 vmov.32 d16[0], r0 vmov.32 d16[1], r1 vmov r0, r1, d16 bx lr into: udiv r0, r0, r2 udiv r1, r1, r3 vmov.32 d16[0], r0 vmov.32 d16[1], r1 vmov r0, r1, d16 bx lr This patch refactors how the copy optimizations are done in the peephole optimizer. Prior to this patch, we had one copy-related optimization that replaced a copy or bitcast by a generic, more suitable (in terms of register file), copy. With this patch, the peephole optimizer features two copy-related optimizations: 1. One for rewriting generic copies to generic copies: PeepholeOptimizer::optimizeCoalescableCopy. 2. One for replacing non-generic copies with generic copies: PeepholeOptimizer::optimizeUncoalescableCopy. The goals of these two optimizations are slightly different: one rewrite the operand of the instruction (#1), the other kills off the non-generic instruction and replace it by a (sequence of) generic instruction(s). Both optimizations rely on the ValueTracker introduced in r212100. The ValueTracker has been refactored to use the information from the TargetInstrInfo for non-generic instruction. As part of the refactoring, we switched the tracking from the index of the definition to the actual register (virtual or physical). This one change is to provide better consistency with register related APIs and to ease the use of the TargetInstrInfo. Moreover, this patch introduces a new helper class CopyRewriter used to ease the rewriting of generic copies (i.e., #1). Finally, this patch adds a dead code elimination pass right after the peephole optimizer to get rid of dead code that may appear after rewriting. This is related to <rdar://problem/12702965>. Review: http://reviews.llvm.org/D4874 llvm-svn: 216088
2014-08-21 01:41:48 +08:00
UseAdvancedTracking(UseAdvancedTracking), MRI(MRI), TII(TII) {
assert(DefIdx < Def->getDesc().getNumDefs() &&
Def->getOperand(DefIdx).isReg() && "Invalid definition");
Reg = Def->getOperand(DefIdx).getReg();
}
/// \brief Following the use-def chain, get the next available source
/// for the tracked value.
/// \return A ValueTrackerResult containing a set of registers
/// and sub registers with tracked values. A ValueTrackerResult with
/// an empty set of registers means no source was found.
ValueTrackerResult getNextSource();
/// \brief Get the last register where the initial value can be found.
/// Initially this is the register of the definition.
/// Then, after each successful call to getNextSource, this is the
/// register of the last source.
unsigned getReg() const { return Reg; }
};
} // end anonymous namespace
char PeepholeOptimizer::ID = 0;
char &llvm::PeepholeOptimizerID = PeepholeOptimizer::ID;
INITIALIZE_PASS_BEGIN(PeepholeOptimizer, DEBUG_TYPE,
"Peephole Optimizations", false, false)
INITIALIZE_PASS_DEPENDENCY(MachineDominatorTree)
INITIALIZE_PASS_END(PeepholeOptimizer, DEBUG_TYPE,
"Peephole Optimizations", false, false)
/// If instruction is a copy-like instruction, i.e. it reads a single register
/// and writes a single register and it does not modify the source, and if the
/// source value is preserved as a sub-register of the result, then replace all
/// reachable uses of the source with the subreg of the result.
2012-02-09 05:22:43 +08:00
///
/// Do not generate an EXTRACT that is used only in a debug use, as this changes
/// the code. Since this code does not currently share EXTRACTs, just ignore all
/// debug uses.
bool PeepholeOptimizer::
optimizeExtInstr(MachineInstr *MI, MachineBasicBlock *MBB,
SmallPtrSetImpl<MachineInstr*> &LocalMIs) {
unsigned SrcReg, DstReg, SubIdx;
if (!TII->isCoalescableExtInstr(*MI, SrcReg, DstReg, SubIdx))
return false;
2012-02-09 05:22:43 +08:00
if (TargetRegisterInfo::isPhysicalRegister(DstReg) ||
TargetRegisterInfo::isPhysicalRegister(SrcReg))
return false;
if (MRI->hasOneNonDBGUse(SrcReg))
// No other uses.
return false;
// Ensure DstReg can get a register class that actually supports
// sub-registers. Don't change the class until we commit.
const TargetRegisterClass *DstRC = MRI->getRegClass(DstReg);
DstRC = TRI->getSubClassWithSubReg(DstRC, SubIdx);
if (!DstRC)
return false;
// The ext instr may be operating on a sub-register of SrcReg as well.
// PPC::EXTSW is a 32 -> 64-bit sign extension, but it reads a 64-bit
// register.
// If UseSrcSubIdx is Set, SubIdx also applies to SrcReg, and only uses of
// SrcReg:SubIdx should be replaced.
bool UseSrcSubIdx =
TRI->getSubClassWithSubReg(MRI->getRegClass(SrcReg), SubIdx) != nullptr;
// The source has other uses. See if we can replace the other uses with use of
// the result of the extension.
SmallPtrSet<MachineBasicBlock*, 4> ReachedBBs;
for (MachineInstr &UI : MRI->use_nodbg_instructions(DstReg))
ReachedBBs.insert(UI.getParent());
// Uses that are in the same BB of uses of the result of the instruction.
SmallVector<MachineOperand*, 8> Uses;
// Uses that the result of the instruction can reach.
SmallVector<MachineOperand*, 8> ExtendedUses;
bool ExtendLife = true;
for (MachineOperand &UseMO : MRI->use_nodbg_operands(SrcReg)) {
MachineInstr *UseMI = UseMO.getParent();
if (UseMI == MI)
continue;
if (UseMI->isPHI()) {
ExtendLife = false;
continue;
}
// Only accept uses of SrcReg:SubIdx.
if (UseSrcSubIdx && UseMO.getSubReg() != SubIdx)
continue;
// It's an error to translate this:
//
// %reg1025 = <sext> %reg1024
// ...
// %reg1026 = SUBREG_TO_REG 0, %reg1024, 4
//
// into this:
//
// %reg1025 = <sext> %reg1024
// ...
// %reg1027 = COPY %reg1025:4
// %reg1026 = SUBREG_TO_REG 0, %reg1027, 4
//
// The problem here is that SUBREG_TO_REG is there to assert that an
// implicit zext occurs. It doesn't insert a zext instruction. If we allow
// the COPY here, it will give us the value after the <sext>, not the
// original value of %reg1024 before <sext>.
if (UseMI->getOpcode() == TargetOpcode::SUBREG_TO_REG)
continue;
MachineBasicBlock *UseMBB = UseMI->getParent();
if (UseMBB == MBB) {
// Local uses that come after the extension.
if (!LocalMIs.count(UseMI))
Uses.push_back(&UseMO);
} else if (ReachedBBs.count(UseMBB)) {
// Non-local uses where the result of the extension is used. Always
// replace these unless it's a PHI.
Uses.push_back(&UseMO);
} else if (Aggressive && DT->dominates(MBB, UseMBB)) {
// We may want to extend the live range of the extension result in order
// to replace these uses.
ExtendedUses.push_back(&UseMO);
} else {
// Both will be live out of the def MBB anyway. Don't extend live range of
// the extension result.
ExtendLife = false;
break;
}
}
if (ExtendLife && !ExtendedUses.empty())
// Extend the liveness of the extension result.
Uses.append(ExtendedUses.begin(), ExtendedUses.end());
// Now replace all uses.
bool Changed = false;
if (!Uses.empty()) {
SmallPtrSet<MachineBasicBlock*, 4> PHIBBs;
// Look for PHI uses of the extended result, we don't want to extend the
// liveness of a PHI input. It breaks all kinds of assumptions down
// stream. A PHI use is expected to be the kill of its source values.
for (MachineInstr &UI : MRI->use_nodbg_instructions(DstReg))
if (UI.isPHI())
PHIBBs.insert(UI.getParent());
const TargetRegisterClass *RC = MRI->getRegClass(SrcReg);
for (unsigned i = 0, e = Uses.size(); i != e; ++i) {
MachineOperand *UseMO = Uses[i];
MachineInstr *UseMI = UseMO->getParent();
MachineBasicBlock *UseMBB = UseMI->getParent();
if (PHIBBs.count(UseMBB))
continue;
// About to add uses of DstReg, clear DstReg's kill flags.
if (!Changed) {
MRI->clearKillFlags(DstReg);
MRI->constrainRegClass(DstReg, DstRC);
}
unsigned NewVR = MRI->createVirtualRegister(RC);
MachineInstr *Copy = BuildMI(*UseMBB, UseMI, UseMI->getDebugLoc(),
TII->get(TargetOpcode::COPY), NewVR)
.addReg(DstReg, 0, SubIdx);
// SubIdx applies to both SrcReg and DstReg when UseSrcSubIdx is set.
if (UseSrcSubIdx) {
Copy->getOperand(0).setSubReg(SubIdx);
Copy->getOperand(0).setIsUndef();
}
UseMO->setReg(NewVR);
++NumReuse;
Changed = true;
}
}
return Changed;
}
/// If the instruction is a compare and the previous instruction it's comparing
/// against already sets (or could be modified to set) the same flag as the
/// compare, then we can remove the comparison and use the flag from the
/// previous instruction.
bool PeepholeOptimizer::optimizeCmpInstr(MachineInstr *MI,
MachineBasicBlock *MBB) {
// If this instruction is a comparison against zero and isn't comparing a
// physical register, we can try to optimize it.
unsigned SrcReg, SrcReg2;
int CmpMask, CmpValue;
if (!TII->analyzeCompare(*MI, SrcReg, SrcReg2, CmpMask, CmpValue) ||
TargetRegisterInfo::isPhysicalRegister(SrcReg) ||
(SrcReg2 != 0 && TargetRegisterInfo::isPhysicalRegister(SrcReg2)))
return false;
// Attempt to optimize the comparison instruction.
if (TII->optimizeCompareInstr(*MI, SrcReg, SrcReg2, CmpMask, CmpValue, MRI)) {
++NumCmps;
return true;
}
return false;
}
/// Optimize a select instruction.
bool PeepholeOptimizer::optimizeSelect(MachineInstr *MI,
SmallPtrSetImpl<MachineInstr *> &LocalMIs) {
unsigned TrueOp = 0;
unsigned FalseOp = 0;
bool Optimizable = false;
SmallVector<MachineOperand, 4> Cond;
if (TII->analyzeSelect(*MI, Cond, TrueOp, FalseOp, Optimizable))
return false;
if (!Optimizable)
return false;
if (!TII->optimizeSelect(*MI, LocalMIs))
return false;
MI->eraseFromParent();
++NumSelects;
return true;
}
/// \brief Check if a simpler conditional branch can be
// generated
bool PeepholeOptimizer::optimizeCondBranch(MachineInstr *MI) {
return TII->optimizeCondBranch(*MI);
}
[PeepholeOptimizer] Refactor the advanced copy optimization to take advantage of the isRegSequence property. This is a follow-up of r215394 and r215404, which respectively introduces the isRegSequence property and uses it for ARM. Thanks to the property introduced by the previous commits, this patch is able to optimize the following sequence: vmov d0, r2, r3 vmov d1, r0, r1 vmov r0, s0 vmov r1, s2 udiv r0, r1, r0 vmov r1, s1 vmov r2, s3 udiv r1, r2, r1 vmov.32 d16[0], r0 vmov.32 d16[1], r1 vmov r0, r1, d16 bx lr into: udiv r0, r0, r2 udiv r1, r1, r3 vmov.32 d16[0], r0 vmov.32 d16[1], r1 vmov r0, r1, d16 bx lr This patch refactors how the copy optimizations are done in the peephole optimizer. Prior to this patch, we had one copy-related optimization that replaced a copy or bitcast by a generic, more suitable (in terms of register file), copy. With this patch, the peephole optimizer features two copy-related optimizations: 1. One for rewriting generic copies to generic copies: PeepholeOptimizer::optimizeCoalescableCopy. 2. One for replacing non-generic copies with generic copies: PeepholeOptimizer::optimizeUncoalescableCopy. The goals of these two optimizations are slightly different: one rewrite the operand of the instruction (#1), the other kills off the non-generic instruction and replace it by a (sequence of) generic instruction(s). Both optimizations rely on the ValueTracker introduced in r212100. The ValueTracker has been refactored to use the information from the TargetInstrInfo for non-generic instruction. As part of the refactoring, we switched the tracking from the index of the definition to the actual register (virtual or physical). This one change is to provide better consistency with register related APIs and to ease the use of the TargetInstrInfo. Moreover, this patch introduces a new helper class CopyRewriter used to ease the rewriting of generic copies (i.e., #1). Finally, this patch adds a dead code elimination pass right after the peephole optimizer to get rid of dead code that may appear after rewriting. This is related to <rdar://problem/12702965>. Review: http://reviews.llvm.org/D4874 llvm-svn: 216088
2014-08-21 01:41:48 +08:00
/// \brief Try to find the next source that share the same register file
/// for the value defined by \p Reg and \p SubReg.
/// When true is returned, the \p RewriteMap can be used by the client to
/// retrieve all Def -> Use along the way up to the next source. Any found
/// Use that is not itself a key for another entry, is the next source to
/// use. During the search for the next source, multiple sources can be found
/// given multiple incoming sources of a PHI instruction. In this case, we
/// look in each PHI source for the next source; all found next sources must
/// share the same register file as \p Reg and \p SubReg. The client should
/// then be capable to rewrite all intermediate PHIs to get the next source.
[PeepholeOptimizer] Refactor the advanced copy optimization to take advantage of the isRegSequence property. This is a follow-up of r215394 and r215404, which respectively introduces the isRegSequence property and uses it for ARM. Thanks to the property introduced by the previous commits, this patch is able to optimize the following sequence: vmov d0, r2, r3 vmov d1, r0, r1 vmov r0, s0 vmov r1, s2 udiv r0, r1, r0 vmov r1, s1 vmov r2, s3 udiv r1, r2, r1 vmov.32 d16[0], r0 vmov.32 d16[1], r1 vmov r0, r1, d16 bx lr into: udiv r0, r0, r2 udiv r1, r1, r3 vmov.32 d16[0], r0 vmov.32 d16[1], r1 vmov r0, r1, d16 bx lr This patch refactors how the copy optimizations are done in the peephole optimizer. Prior to this patch, we had one copy-related optimization that replaced a copy or bitcast by a generic, more suitable (in terms of register file), copy. With this patch, the peephole optimizer features two copy-related optimizations: 1. One for rewriting generic copies to generic copies: PeepholeOptimizer::optimizeCoalescableCopy. 2. One for replacing non-generic copies with generic copies: PeepholeOptimizer::optimizeUncoalescableCopy. The goals of these two optimizations are slightly different: one rewrite the operand of the instruction (#1), the other kills off the non-generic instruction and replace it by a (sequence of) generic instruction(s). Both optimizations rely on the ValueTracker introduced in r212100. The ValueTracker has been refactored to use the information from the TargetInstrInfo for non-generic instruction. As part of the refactoring, we switched the tracking from the index of the definition to the actual register (virtual or physical). This one change is to provide better consistency with register related APIs and to ease the use of the TargetInstrInfo. Moreover, this patch introduces a new helper class CopyRewriter used to ease the rewriting of generic copies (i.e., #1). Finally, this patch adds a dead code elimination pass right after the peephole optimizer to get rid of dead code that may appear after rewriting. This is related to <rdar://problem/12702965>. Review: http://reviews.llvm.org/D4874 llvm-svn: 216088
2014-08-21 01:41:48 +08:00
/// \return False if no alternative sources are available. True otherwise.
bool PeepholeOptimizer::findNextSource(unsigned Reg, unsigned SubReg,
RewriteMapTy &RewriteMap) {
[PeepholeOptimizer] Refactor the advanced copy optimization to take advantage of the isRegSequence property. This is a follow-up of r215394 and r215404, which respectively introduces the isRegSequence property and uses it for ARM. Thanks to the property introduced by the previous commits, this patch is able to optimize the following sequence: vmov d0, r2, r3 vmov d1, r0, r1 vmov r0, s0 vmov r1, s2 udiv r0, r1, r0 vmov r1, s1 vmov r2, s3 udiv r1, r2, r1 vmov.32 d16[0], r0 vmov.32 d16[1], r1 vmov r0, r1, d16 bx lr into: udiv r0, r0, r2 udiv r1, r1, r3 vmov.32 d16[0], r0 vmov.32 d16[1], r1 vmov r0, r1, d16 bx lr This patch refactors how the copy optimizations are done in the peephole optimizer. Prior to this patch, we had one copy-related optimization that replaced a copy or bitcast by a generic, more suitable (in terms of register file), copy. With this patch, the peephole optimizer features two copy-related optimizations: 1. One for rewriting generic copies to generic copies: PeepholeOptimizer::optimizeCoalescableCopy. 2. One for replacing non-generic copies with generic copies: PeepholeOptimizer::optimizeUncoalescableCopy. The goals of these two optimizations are slightly different: one rewrite the operand of the instruction (#1), the other kills off the non-generic instruction and replace it by a (sequence of) generic instruction(s). Both optimizations rely on the ValueTracker introduced in r212100. The ValueTracker has been refactored to use the information from the TargetInstrInfo for non-generic instruction. As part of the refactoring, we switched the tracking from the index of the definition to the actual register (virtual or physical). This one change is to provide better consistency with register related APIs and to ease the use of the TargetInstrInfo. Moreover, this patch introduces a new helper class CopyRewriter used to ease the rewriting of generic copies (i.e., #1). Finally, this patch adds a dead code elimination pass right after the peephole optimizer to get rid of dead code that may appear after rewriting. This is related to <rdar://problem/12702965>. Review: http://reviews.llvm.org/D4874 llvm-svn: 216088
2014-08-21 01:41:48 +08:00
// Do not try to find a new source for a physical register.
// So far we do not have any motivating example for doing that.
// Thus, instead of maintaining untested code, we will revisit that if
// that changes at some point.
if (TargetRegisterInfo::isPhysicalRegister(Reg))
return false;
const TargetRegisterClass *DefRC = MRI->getRegClass(Reg);
SmallVector<TargetInstrInfo::RegSubRegPair, 4> SrcToLook;
TargetInstrInfo::RegSubRegPair CurSrcPair(Reg, SubReg);
SrcToLook.push_back(CurSrcPair);
unsigned PHICount = 0;
while (!SrcToLook.empty() && PHICount < RewritePHILimit) {
TargetInstrInfo::RegSubRegPair Pair = SrcToLook.pop_back_val();
// As explained above, do not handle physical registers
if (TargetRegisterInfo::isPhysicalRegister(Pair.Reg))
return false;
CurSrcPair = Pair;
ValueTracker ValTracker(CurSrcPair.Reg, CurSrcPair.SubReg, *MRI,
!DisableAdvCopyOpt, TII);
ValueTrackerResult Res;
bool ShouldRewrite = false;
do {
// Follow the chain of copies until we reach the top of the use-def chain
// or find a more suitable source.
Res = ValTracker.getNextSource();
if (!Res.isValid())
break;
// Insert the Def -> Use entry for the recently found source.
ValueTrackerResult CurSrcRes = RewriteMap.lookup(CurSrcPair);
if (CurSrcRes.isValid()) {
assert(CurSrcRes == Res && "ValueTrackerResult found must match");
// An existent entry with multiple sources is a PHI cycle we must avoid.
// Otherwise it's an entry with a valid next source we already found.
if (CurSrcRes.getNumSources() > 1) {
DEBUG(dbgs() << "findNextSource: found PHI cycle, aborting...\n");
return false;
}
break;
}
RewriteMap.insert(std::make_pair(CurSrcPair, Res));
// ValueTrackerResult usually have one source unless it's the result from
// a PHI instruction. Add the found PHI edges to be looked up further.
unsigned NumSrcs = Res.getNumSources();
if (NumSrcs > 1) {
PHICount++;
for (unsigned i = 0; i < NumSrcs; ++i)
SrcToLook.push_back(TargetInstrInfo::RegSubRegPair(
Res.getSrcReg(i), Res.getSrcSubReg(i)));
break;
}
CurSrcPair.Reg = Res.getSrcReg(0);
CurSrcPair.SubReg = Res.getSrcSubReg(0);
// Do not extend the live-ranges of physical registers as they add
// constraints to the register allocator. Moreover, if we want to extend
// the live-range of a physical register, unlike SSA virtual register,
// we will have to check that they aren't redefine before the related use.
if (TargetRegisterInfo::isPhysicalRegister(CurSrcPair.Reg))
return false;
const TargetRegisterClass *SrcRC = MRI->getRegClass(CurSrcPair.Reg);
ShouldRewrite = TRI->shouldRewriteCopySrc(DefRC, SubReg, SrcRC,
CurSrcPair.SubReg);
} while (!ShouldRewrite);
// Continue looking for new sources...
if (Res.isValid())
continue;
// Do not continue searching for a new source if the there's at least
// one use-def which cannot be rewritten.
if (!ShouldRewrite)
return false;
}
if (PHICount >= RewritePHILimit) {
DEBUG(dbgs() << "findNextSource: PHI limit reached\n");
return false;
}
// If we did not find a more suitable source, there is nothing to optimize.
return CurSrcPair.Reg != Reg;
[PeepholeOptimizer] Refactor the advanced copy optimization to take advantage of the isRegSequence property. This is a follow-up of r215394 and r215404, which respectively introduces the isRegSequence property and uses it for ARM. Thanks to the property introduced by the previous commits, this patch is able to optimize the following sequence: vmov d0, r2, r3 vmov d1, r0, r1 vmov r0, s0 vmov r1, s2 udiv r0, r1, r0 vmov r1, s1 vmov r2, s3 udiv r1, r2, r1 vmov.32 d16[0], r0 vmov.32 d16[1], r1 vmov r0, r1, d16 bx lr into: udiv r0, r0, r2 udiv r1, r1, r3 vmov.32 d16[0], r0 vmov.32 d16[1], r1 vmov r0, r1, d16 bx lr This patch refactors how the copy optimizations are done in the peephole optimizer. Prior to this patch, we had one copy-related optimization that replaced a copy or bitcast by a generic, more suitable (in terms of register file), copy. With this patch, the peephole optimizer features two copy-related optimizations: 1. One for rewriting generic copies to generic copies: PeepholeOptimizer::optimizeCoalescableCopy. 2. One for replacing non-generic copies with generic copies: PeepholeOptimizer::optimizeUncoalescableCopy. The goals of these two optimizations are slightly different: one rewrite the operand of the instruction (#1), the other kills off the non-generic instruction and replace it by a (sequence of) generic instruction(s). Both optimizations rely on the ValueTracker introduced in r212100. The ValueTracker has been refactored to use the information from the TargetInstrInfo for non-generic instruction. As part of the refactoring, we switched the tracking from the index of the definition to the actual register (virtual or physical). This one change is to provide better consistency with register related APIs and to ease the use of the TargetInstrInfo. Moreover, this patch introduces a new helper class CopyRewriter used to ease the rewriting of generic copies (i.e., #1). Finally, this patch adds a dead code elimination pass right after the peephole optimizer to get rid of dead code that may appear after rewriting. This is related to <rdar://problem/12702965>. Review: http://reviews.llvm.org/D4874 llvm-svn: 216088
2014-08-21 01:41:48 +08:00
}
/// \brief Insert a PHI instruction with incoming edges \p SrcRegs that are
/// guaranteed to have the same register class. This is necessary whenever we
/// successfully traverse a PHI instruction and find suitable sources coming
/// from its edges. By inserting a new PHI, we provide a rewritten PHI def
/// suitable to be used in a new COPY instruction.
static MachineInstr *
insertPHI(MachineRegisterInfo *MRI, const TargetInstrInfo *TII,
const SmallVectorImpl<TargetInstrInfo::RegSubRegPair> &SrcRegs,
MachineInstr *OrigPHI) {
assert(!SrcRegs.empty() && "No sources to create a PHI instruction?");
const TargetRegisterClass *NewRC = MRI->getRegClass(SrcRegs[0].Reg);
unsigned NewVR = MRI->createVirtualRegister(NewRC);
MachineBasicBlock *MBB = OrigPHI->getParent();
MachineInstrBuilder MIB = BuildMI(*MBB, OrigPHI, OrigPHI->getDebugLoc(),
TII->get(TargetOpcode::PHI), NewVR);
unsigned MBBOpIdx = 2;
for (auto RegPair : SrcRegs) {
MIB.addReg(RegPair.Reg, 0, RegPair.SubReg);
MIB.addMBB(OrigPHI->getOperand(MBBOpIdx).getMBB());
// Since we're extended the lifetime of RegPair.Reg, clear the
// kill flags to account for that and make RegPair.Reg reaches
// the new PHI.
MRI->clearKillFlags(RegPair.Reg);
MBBOpIdx += 2;
}
return MIB;
}
[PeepholeOptimizer] Refactor the advanced copy optimization to take advantage of the isRegSequence property. This is a follow-up of r215394 and r215404, which respectively introduces the isRegSequence property and uses it for ARM. Thanks to the property introduced by the previous commits, this patch is able to optimize the following sequence: vmov d0, r2, r3 vmov d1, r0, r1 vmov r0, s0 vmov r1, s2 udiv r0, r1, r0 vmov r1, s1 vmov r2, s3 udiv r1, r2, r1 vmov.32 d16[0], r0 vmov.32 d16[1], r1 vmov r0, r1, d16 bx lr into: udiv r0, r0, r2 udiv r1, r1, r3 vmov.32 d16[0], r0 vmov.32 d16[1], r1 vmov r0, r1, d16 bx lr This patch refactors how the copy optimizations are done in the peephole optimizer. Prior to this patch, we had one copy-related optimization that replaced a copy or bitcast by a generic, more suitable (in terms of register file), copy. With this patch, the peephole optimizer features two copy-related optimizations: 1. One for rewriting generic copies to generic copies: PeepholeOptimizer::optimizeCoalescableCopy. 2. One for replacing non-generic copies with generic copies: PeepholeOptimizer::optimizeUncoalescableCopy. The goals of these two optimizations are slightly different: one rewrite the operand of the instruction (#1), the other kills off the non-generic instruction and replace it by a (sequence of) generic instruction(s). Both optimizations rely on the ValueTracker introduced in r212100. The ValueTracker has been refactored to use the information from the TargetInstrInfo for non-generic instruction. As part of the refactoring, we switched the tracking from the index of the definition to the actual register (virtual or physical). This one change is to provide better consistency with register related APIs and to ease the use of the TargetInstrInfo. Moreover, this patch introduces a new helper class CopyRewriter used to ease the rewriting of generic copies (i.e., #1). Finally, this patch adds a dead code elimination pass right after the peephole optimizer to get rid of dead code that may appear after rewriting. This is related to <rdar://problem/12702965>. Review: http://reviews.llvm.org/D4874 llvm-svn: 216088
2014-08-21 01:41:48 +08:00
namespace {
[PeepholeOptimizer] Refactor the advanced copy optimization to take advantage of the isRegSequence property. This is a follow-up of r215394 and r215404, which respectively introduces the isRegSequence property and uses it for ARM. Thanks to the property introduced by the previous commits, this patch is able to optimize the following sequence: vmov d0, r2, r3 vmov d1, r0, r1 vmov r0, s0 vmov r1, s2 udiv r0, r1, r0 vmov r1, s1 vmov r2, s3 udiv r1, r2, r1 vmov.32 d16[0], r0 vmov.32 d16[1], r1 vmov r0, r1, d16 bx lr into: udiv r0, r0, r2 udiv r1, r1, r3 vmov.32 d16[0], r0 vmov.32 d16[1], r1 vmov r0, r1, d16 bx lr This patch refactors how the copy optimizations are done in the peephole optimizer. Prior to this patch, we had one copy-related optimization that replaced a copy or bitcast by a generic, more suitable (in terms of register file), copy. With this patch, the peephole optimizer features two copy-related optimizations: 1. One for rewriting generic copies to generic copies: PeepholeOptimizer::optimizeCoalescableCopy. 2. One for replacing non-generic copies with generic copies: PeepholeOptimizer::optimizeUncoalescableCopy. The goals of these two optimizations are slightly different: one rewrite the operand of the instruction (#1), the other kills off the non-generic instruction and replace it by a (sequence of) generic instruction(s). Both optimizations rely on the ValueTracker introduced in r212100. The ValueTracker has been refactored to use the information from the TargetInstrInfo for non-generic instruction. As part of the refactoring, we switched the tracking from the index of the definition to the actual register (virtual or physical). This one change is to provide better consistency with register related APIs and to ease the use of the TargetInstrInfo. Moreover, this patch introduces a new helper class CopyRewriter used to ease the rewriting of generic copies (i.e., #1). Finally, this patch adds a dead code elimination pass right after the peephole optimizer to get rid of dead code that may appear after rewriting. This is related to <rdar://problem/12702965>. Review: http://reviews.llvm.org/D4874 llvm-svn: 216088
2014-08-21 01:41:48 +08:00
/// \brief Helper class to rewrite the arguments of a copy-like instruction.
class CopyRewriter {
protected:
/// The copy-like instruction.
MachineInstr &CopyLike;
/// The index of the source being rewritten.
unsigned CurrentSrcIdx;
public:
CopyRewriter(MachineInstr &MI) : CopyLike(MI), CurrentSrcIdx(0) {}
virtual ~CopyRewriter() {}
/// \brief Get the next rewritable source (SrcReg, SrcSubReg) and
/// the related value that it affects (TrackReg, TrackSubReg).
/// A source is considered rewritable if its register class and the
/// register class of the related TrackReg may not be register
/// coalescer friendly. In other words, given a copy-like instruction
/// not all the arguments may be returned at rewritable source, since
/// some arguments are none to be register coalescer friendly.
///
/// Each call of this method moves the current source to the next
/// rewritable source.
/// For instance, let CopyLike be the instruction to rewrite.
/// CopyLike has one definition and one source:
/// dst.dstSubIdx = CopyLike src.srcSubIdx.
///
/// The first call will give the first rewritable source, i.e.,
/// the only source this instruction has:
/// (SrcReg, SrcSubReg) = (src, srcSubIdx).
/// This source defines the whole definition, i.e.,
/// (TrackReg, TrackSubReg) = (dst, dstSubIdx).
///
2015-09-09 08:38:33 +08:00
/// The second and subsequent calls will return false, as there is only one
[PeepholeOptimizer] Refactor the advanced copy optimization to take advantage of the isRegSequence property. This is a follow-up of r215394 and r215404, which respectively introduces the isRegSequence property and uses it for ARM. Thanks to the property introduced by the previous commits, this patch is able to optimize the following sequence: vmov d0, r2, r3 vmov d1, r0, r1 vmov r0, s0 vmov r1, s2 udiv r0, r1, r0 vmov r1, s1 vmov r2, s3 udiv r1, r2, r1 vmov.32 d16[0], r0 vmov.32 d16[1], r1 vmov r0, r1, d16 bx lr into: udiv r0, r0, r2 udiv r1, r1, r3 vmov.32 d16[0], r0 vmov.32 d16[1], r1 vmov r0, r1, d16 bx lr This patch refactors how the copy optimizations are done in the peephole optimizer. Prior to this patch, we had one copy-related optimization that replaced a copy or bitcast by a generic, more suitable (in terms of register file), copy. With this patch, the peephole optimizer features two copy-related optimizations: 1. One for rewriting generic copies to generic copies: PeepholeOptimizer::optimizeCoalescableCopy. 2. One for replacing non-generic copies with generic copies: PeepholeOptimizer::optimizeUncoalescableCopy. The goals of these two optimizations are slightly different: one rewrite the operand of the instruction (#1), the other kills off the non-generic instruction and replace it by a (sequence of) generic instruction(s). Both optimizations rely on the ValueTracker introduced in r212100. The ValueTracker has been refactored to use the information from the TargetInstrInfo for non-generic instruction. As part of the refactoring, we switched the tracking from the index of the definition to the actual register (virtual or physical). This one change is to provide better consistency with register related APIs and to ease the use of the TargetInstrInfo. Moreover, this patch introduces a new helper class CopyRewriter used to ease the rewriting of generic copies (i.e., #1). Finally, this patch adds a dead code elimination pass right after the peephole optimizer to get rid of dead code that may appear after rewriting. This is related to <rdar://problem/12702965>. Review: http://reviews.llvm.org/D4874 llvm-svn: 216088
2014-08-21 01:41:48 +08:00
/// rewritable source.
///
/// \return True if a rewritable source has been found, false otherwise.
/// The output arguments are valid if and only if true is returned.
virtual bool getNextRewritableSource(unsigned &SrcReg, unsigned &SrcSubReg,
unsigned &TrackReg,
unsigned &TrackSubReg) {
2015-09-09 08:38:33 +08:00
// If CurrentSrcIdx == 1, this means this function has already been called
// once. CopyLike has one definition and one argument, thus, there is
// nothing else to rewrite.
[PeepholeOptimizer] Refactor the advanced copy optimization to take advantage of the isRegSequence property. This is a follow-up of r215394 and r215404, which respectively introduces the isRegSequence property and uses it for ARM. Thanks to the property introduced by the previous commits, this patch is able to optimize the following sequence: vmov d0, r2, r3 vmov d1, r0, r1 vmov r0, s0 vmov r1, s2 udiv r0, r1, r0 vmov r1, s1 vmov r2, s3 udiv r1, r2, r1 vmov.32 d16[0], r0 vmov.32 d16[1], r1 vmov r0, r1, d16 bx lr into: udiv r0, r0, r2 udiv r1, r1, r3 vmov.32 d16[0], r0 vmov.32 d16[1], r1 vmov r0, r1, d16 bx lr This patch refactors how the copy optimizations are done in the peephole optimizer. Prior to this patch, we had one copy-related optimization that replaced a copy or bitcast by a generic, more suitable (in terms of register file), copy. With this patch, the peephole optimizer features two copy-related optimizations: 1. One for rewriting generic copies to generic copies: PeepholeOptimizer::optimizeCoalescableCopy. 2. One for replacing non-generic copies with generic copies: PeepholeOptimizer::optimizeUncoalescableCopy. The goals of these two optimizations are slightly different: one rewrite the operand of the instruction (#1), the other kills off the non-generic instruction and replace it by a (sequence of) generic instruction(s). Both optimizations rely on the ValueTracker introduced in r212100. The ValueTracker has been refactored to use the information from the TargetInstrInfo for non-generic instruction. As part of the refactoring, we switched the tracking from the index of the definition to the actual register (virtual or physical). This one change is to provide better consistency with register related APIs and to ease the use of the TargetInstrInfo. Moreover, this patch introduces a new helper class CopyRewriter used to ease the rewriting of generic copies (i.e., #1). Finally, this patch adds a dead code elimination pass right after the peephole optimizer to get rid of dead code that may appear after rewriting. This is related to <rdar://problem/12702965>. Review: http://reviews.llvm.org/D4874 llvm-svn: 216088
2014-08-21 01:41:48 +08:00
if (!CopyLike.isCopy() || CurrentSrcIdx == 1)
return false;
// This is the first call to getNextRewritableSource.
// Move the CurrentSrcIdx to remember that we made that call.
CurrentSrcIdx = 1;
// The rewritable source is the argument.
const MachineOperand &MOSrc = CopyLike.getOperand(1);
SrcReg = MOSrc.getReg();
SrcSubReg = MOSrc.getSubReg();
// What we track are the alternative sources of the definition.
const MachineOperand &MODef = CopyLike.getOperand(0);
TrackReg = MODef.getReg();
TrackSubReg = MODef.getSubReg();
return true;
}
/// \brief Rewrite the current source with \p NewReg and \p NewSubReg
/// if possible.
/// \return True if the rewriting was possible, false otherwise.
[PeepholeOptimizer] Refactor the advanced copy optimization to take advantage of the isRegSequence property. This is a follow-up of r215394 and r215404, which respectively introduces the isRegSequence property and uses it for ARM. Thanks to the property introduced by the previous commits, this patch is able to optimize the following sequence: vmov d0, r2, r3 vmov d1, r0, r1 vmov r0, s0 vmov r1, s2 udiv r0, r1, r0 vmov r1, s1 vmov r2, s3 udiv r1, r2, r1 vmov.32 d16[0], r0 vmov.32 d16[1], r1 vmov r0, r1, d16 bx lr into: udiv r0, r0, r2 udiv r1, r1, r3 vmov.32 d16[0], r0 vmov.32 d16[1], r1 vmov r0, r1, d16 bx lr This patch refactors how the copy optimizations are done in the peephole optimizer. Prior to this patch, we had one copy-related optimization that replaced a copy or bitcast by a generic, more suitable (in terms of register file), copy. With this patch, the peephole optimizer features two copy-related optimizations: 1. One for rewriting generic copies to generic copies: PeepholeOptimizer::optimizeCoalescableCopy. 2. One for replacing non-generic copies with generic copies: PeepholeOptimizer::optimizeUncoalescableCopy. The goals of these two optimizations are slightly different: one rewrite the operand of the instruction (#1), the other kills off the non-generic instruction and replace it by a (sequence of) generic instruction(s). Both optimizations rely on the ValueTracker introduced in r212100. The ValueTracker has been refactored to use the information from the TargetInstrInfo for non-generic instruction. As part of the refactoring, we switched the tracking from the index of the definition to the actual register (virtual or physical). This one change is to provide better consistency with register related APIs and to ease the use of the TargetInstrInfo. Moreover, this patch introduces a new helper class CopyRewriter used to ease the rewriting of generic copies (i.e., #1). Finally, this patch adds a dead code elimination pass right after the peephole optimizer to get rid of dead code that may appear after rewriting. This is related to <rdar://problem/12702965>. Review: http://reviews.llvm.org/D4874 llvm-svn: 216088
2014-08-21 01:41:48 +08:00
virtual bool RewriteCurrentSource(unsigned NewReg, unsigned NewSubReg) {
if (!CopyLike.isCopy() || CurrentSrcIdx != 1)
return false;
MachineOperand &MOSrc = CopyLike.getOperand(CurrentSrcIdx);
MOSrc.setReg(NewReg);
MOSrc.setSubReg(NewSubReg);
return true;
}
/// \brief Given a \p Def.Reg and Def.SubReg pair, use \p RewriteMap to find
/// the new source to use for rewrite. If \p HandleMultipleSources is true and
/// multiple sources for a given \p Def are found along the way, we found a
/// PHI instructions that needs to be rewritten.
/// TODO: HandleMultipleSources should be removed once we test PHI handling
/// with coalescable copies.
TargetInstrInfo::RegSubRegPair
getNewSource(MachineRegisterInfo *MRI, const TargetInstrInfo *TII,
TargetInstrInfo::RegSubRegPair Def,
PeepholeOptimizer::RewriteMapTy &RewriteMap,
bool HandleMultipleSources = true) {
TargetInstrInfo::RegSubRegPair LookupSrc(Def.Reg, Def.SubReg);
do {
ValueTrackerResult Res = RewriteMap.lookup(LookupSrc);
// If there are no entries on the map, LookupSrc is the new source.
if (!Res.isValid())
return LookupSrc;
// There's only one source for this definition, keep searching...
unsigned NumSrcs = Res.getNumSources();
if (NumSrcs == 1) {
LookupSrc.Reg = Res.getSrcReg(0);
LookupSrc.SubReg = Res.getSrcSubReg(0);
continue;
}
2015-09-09 08:38:33 +08:00
// TODO: Remove once multiple srcs w/ coalescable copies are supported.
if (!HandleMultipleSources)
break;
// Multiple sources, recurse into each source to find a new source
// for it. Then, rewrite the PHI accordingly to its new edges.
SmallVector<TargetInstrInfo::RegSubRegPair, 4> NewPHISrcs;
for (unsigned i = 0; i < NumSrcs; ++i) {
TargetInstrInfo::RegSubRegPair PHISrc(Res.getSrcReg(i),
Res.getSrcSubReg(i));
NewPHISrcs.push_back(
getNewSource(MRI, TII, PHISrc, RewriteMap, HandleMultipleSources));
}
// Build the new PHI node and return its def register as the new source.
MachineInstr *OrigPHI = const_cast<MachineInstr *>(Res.getInst());
MachineInstr *NewPHI = insertPHI(MRI, TII, NewPHISrcs, OrigPHI);
DEBUG(dbgs() << "-- getNewSource\n");
DEBUG(dbgs() << " Replacing: " << *OrigPHI);
DEBUG(dbgs() << " With: " << *NewPHI);
const MachineOperand &MODef = NewPHI->getOperand(0);
return TargetInstrInfo::RegSubRegPair(MODef.getReg(), MODef.getSubReg());
} while (true);
return TargetInstrInfo::RegSubRegPair(0, 0);
}
/// \brief Rewrite the source found through \p Def, by using the \p RewriteMap
/// and create a new COPY instruction. More info about RewriteMap in
/// PeepholeOptimizer::findNextSource. Right now this is only used to handle
/// Uncoalescable copies, since they are copy like instructions that aren't
/// recognized by the register allocator.
virtual MachineInstr *
RewriteSource(TargetInstrInfo::RegSubRegPair Def,
PeepholeOptimizer::RewriteMapTy &RewriteMap) {
return nullptr;
}
};
/// \brief Helper class to rewrite uncoalescable copy like instructions
/// into new COPY (coalescable friendly) instructions.
class UncoalescableRewriter : public CopyRewriter {
protected:
const TargetInstrInfo &TII;
MachineRegisterInfo &MRI;
/// The number of defs in the bitcast
unsigned NumDefs;
public:
UncoalescableRewriter(MachineInstr &MI, const TargetInstrInfo &TII,
MachineRegisterInfo &MRI)
: CopyRewriter(MI), TII(TII), MRI(MRI) {
NumDefs = MI.getDesc().getNumDefs();
}
/// \brief Get the next rewritable def source (TrackReg, TrackSubReg)
/// All such sources need to be considered rewritable in order to
/// rewrite a uncoalescable copy-like instruction. This method return
/// each definition that must be checked if rewritable.
///
bool getNextRewritableSource(unsigned &SrcReg, unsigned &SrcSubReg,
unsigned &TrackReg,
unsigned &TrackSubReg) override {
// Find the next non-dead definition and continue from there.
if (CurrentSrcIdx == NumDefs)
return false;
while (CopyLike.getOperand(CurrentSrcIdx).isDead()) {
++CurrentSrcIdx;
if (CurrentSrcIdx == NumDefs)
return false;
}
// What we track are the alternative sources of the definition.
const MachineOperand &MODef = CopyLike.getOperand(CurrentSrcIdx);
TrackReg = MODef.getReg();
TrackSubReg = MODef.getSubReg();
CurrentSrcIdx++;
return true;
}
/// \brief Rewrite the source found through \p Def, by using the \p RewriteMap
/// and create a new COPY instruction. More info about RewriteMap in
/// PeepholeOptimizer::findNextSource. Right now this is only used to handle
/// Uncoalescable copies, since they are copy like instructions that aren't
/// recognized by the register allocator.
MachineInstr *
RewriteSource(TargetInstrInfo::RegSubRegPair Def,
PeepholeOptimizer::RewriteMapTy &RewriteMap) override {
assert(!TargetRegisterInfo::isPhysicalRegister(Def.Reg) &&
"We do not rewrite physical registers");
// Find the new source to use in the COPY rewrite.
TargetInstrInfo::RegSubRegPair NewSrc =
getNewSource(&MRI, &TII, Def, RewriteMap);
// Insert the COPY.
const TargetRegisterClass *DefRC = MRI.getRegClass(Def.Reg);
unsigned NewVR = MRI.createVirtualRegister(DefRC);
MachineInstr *NewCopy =
BuildMI(*CopyLike.getParent(), &CopyLike, CopyLike.getDebugLoc(),
TII.get(TargetOpcode::COPY), NewVR)
.addReg(NewSrc.Reg, 0, NewSrc.SubReg);
NewCopy->getOperand(0).setSubReg(Def.SubReg);
if (Def.SubReg)
NewCopy->getOperand(0).setIsUndef();
DEBUG(dbgs() << "-- RewriteSource\n");
DEBUG(dbgs() << " Replacing: " << CopyLike);
DEBUG(dbgs() << " With: " << *NewCopy);
MRI.replaceRegWith(Def.Reg, NewVR);
MRI.clearKillFlags(NewVR);
// We extended the lifetime of NewSrc.Reg, clear the kill flags to
// account for that.
MRI.clearKillFlags(NewSrc.Reg);
return NewCopy;
}
[PeepholeOptimizer] Refactor the advanced copy optimization to take advantage of the isRegSequence property. This is a follow-up of r215394 and r215404, which respectively introduces the isRegSequence property and uses it for ARM. Thanks to the property introduced by the previous commits, this patch is able to optimize the following sequence: vmov d0, r2, r3 vmov d1, r0, r1 vmov r0, s0 vmov r1, s2 udiv r0, r1, r0 vmov r1, s1 vmov r2, s3 udiv r1, r2, r1 vmov.32 d16[0], r0 vmov.32 d16[1], r1 vmov r0, r1, d16 bx lr into: udiv r0, r0, r2 udiv r1, r1, r3 vmov.32 d16[0], r0 vmov.32 d16[1], r1 vmov r0, r1, d16 bx lr This patch refactors how the copy optimizations are done in the peephole optimizer. Prior to this patch, we had one copy-related optimization that replaced a copy or bitcast by a generic, more suitable (in terms of register file), copy. With this patch, the peephole optimizer features two copy-related optimizations: 1. One for rewriting generic copies to generic copies: PeepholeOptimizer::optimizeCoalescableCopy. 2. One for replacing non-generic copies with generic copies: PeepholeOptimizer::optimizeUncoalescableCopy. The goals of these two optimizations are slightly different: one rewrite the operand of the instruction (#1), the other kills off the non-generic instruction and replace it by a (sequence of) generic instruction(s). Both optimizations rely on the ValueTracker introduced in r212100. The ValueTracker has been refactored to use the information from the TargetInstrInfo for non-generic instruction. As part of the refactoring, we switched the tracking from the index of the definition to the actual register (virtual or physical). This one change is to provide better consistency with register related APIs and to ease the use of the TargetInstrInfo. Moreover, this patch introduces a new helper class CopyRewriter used to ease the rewriting of generic copies (i.e., #1). Finally, this patch adds a dead code elimination pass right after the peephole optimizer to get rid of dead code that may appear after rewriting. This is related to <rdar://problem/12702965>. Review: http://reviews.llvm.org/D4874 llvm-svn: 216088
2014-08-21 01:41:48 +08:00
};
/// \brief Specialized rewriter for INSERT_SUBREG instruction.
class InsertSubregRewriter : public CopyRewriter {
public:
InsertSubregRewriter(MachineInstr &MI) : CopyRewriter(MI) {
assert(MI.isInsertSubreg() && "Invalid instruction");
}
/// \brief See CopyRewriter::getNextRewritableSource.
/// Here CopyLike has the following form:
/// dst = INSERT_SUBREG Src1, Src2.src2SubIdx, subIdx.
/// Src1 has the same register class has dst, hence, there is
/// nothing to rewrite.
/// Src2.src2SubIdx, may not be register coalescer friendly.
/// Therefore, the first call to this method returns:
/// (SrcReg, SrcSubReg) = (Src2, src2SubIdx).
/// (TrackReg, TrackSubReg) = (dst, subIdx).
///
/// Subsequence calls will return false.
bool getNextRewritableSource(unsigned &SrcReg, unsigned &SrcSubReg,
unsigned &TrackReg,
unsigned &TrackSubReg) override {
// If we already get the only source we can rewrite, return false.
if (CurrentSrcIdx == 2)
return false;
// We are looking at v2 = INSERT_SUBREG v0, v1, sub0.
CurrentSrcIdx = 2;
const MachineOperand &MOInsertedReg = CopyLike.getOperand(2);
SrcReg = MOInsertedReg.getReg();
SrcSubReg = MOInsertedReg.getSubReg();
const MachineOperand &MODef = CopyLike.getOperand(0);
// We want to track something that is compatible with the
// partial definition.
TrackReg = MODef.getReg();
if (MODef.getSubReg())
2015-09-09 08:38:33 +08:00
// Bail if we have to compose sub-register indices.
[PeepholeOptimizer] Refactor the advanced copy optimization to take advantage of the isRegSequence property. This is a follow-up of r215394 and r215404, which respectively introduces the isRegSequence property and uses it for ARM. Thanks to the property introduced by the previous commits, this patch is able to optimize the following sequence: vmov d0, r2, r3 vmov d1, r0, r1 vmov r0, s0 vmov r1, s2 udiv r0, r1, r0 vmov r1, s1 vmov r2, s3 udiv r1, r2, r1 vmov.32 d16[0], r0 vmov.32 d16[1], r1 vmov r0, r1, d16 bx lr into: udiv r0, r0, r2 udiv r1, r1, r3 vmov.32 d16[0], r0 vmov.32 d16[1], r1 vmov r0, r1, d16 bx lr This patch refactors how the copy optimizations are done in the peephole optimizer. Prior to this patch, we had one copy-related optimization that replaced a copy or bitcast by a generic, more suitable (in terms of register file), copy. With this patch, the peephole optimizer features two copy-related optimizations: 1. One for rewriting generic copies to generic copies: PeepholeOptimizer::optimizeCoalescableCopy. 2. One for replacing non-generic copies with generic copies: PeepholeOptimizer::optimizeUncoalescableCopy. The goals of these two optimizations are slightly different: one rewrite the operand of the instruction (#1), the other kills off the non-generic instruction and replace it by a (sequence of) generic instruction(s). Both optimizations rely on the ValueTracker introduced in r212100. The ValueTracker has been refactored to use the information from the TargetInstrInfo for non-generic instruction. As part of the refactoring, we switched the tracking from the index of the definition to the actual register (virtual or physical). This one change is to provide better consistency with register related APIs and to ease the use of the TargetInstrInfo. Moreover, this patch introduces a new helper class CopyRewriter used to ease the rewriting of generic copies (i.e., #1). Finally, this patch adds a dead code elimination pass right after the peephole optimizer to get rid of dead code that may appear after rewriting. This is related to <rdar://problem/12702965>. Review: http://reviews.llvm.org/D4874 llvm-svn: 216088
2014-08-21 01:41:48 +08:00
return false;
TrackSubReg = (unsigned)CopyLike.getOperand(3).getImm();
return true;
}
[PeepholeOptimizer] Refactor the advanced copy optimization to take advantage of the isRegSequence property. This is a follow-up of r215394 and r215404, which respectively introduces the isRegSequence property and uses it for ARM. Thanks to the property introduced by the previous commits, this patch is able to optimize the following sequence: vmov d0, r2, r3 vmov d1, r0, r1 vmov r0, s0 vmov r1, s2 udiv r0, r1, r0 vmov r1, s1 vmov r2, s3 udiv r1, r2, r1 vmov.32 d16[0], r0 vmov.32 d16[1], r1 vmov r0, r1, d16 bx lr into: udiv r0, r0, r2 udiv r1, r1, r3 vmov.32 d16[0], r0 vmov.32 d16[1], r1 vmov r0, r1, d16 bx lr This patch refactors how the copy optimizations are done in the peephole optimizer. Prior to this patch, we had one copy-related optimization that replaced a copy or bitcast by a generic, more suitable (in terms of register file), copy. With this patch, the peephole optimizer features two copy-related optimizations: 1. One for rewriting generic copies to generic copies: PeepholeOptimizer::optimizeCoalescableCopy. 2. One for replacing non-generic copies with generic copies: PeepholeOptimizer::optimizeUncoalescableCopy. The goals of these two optimizations are slightly different: one rewrite the operand of the instruction (#1), the other kills off the non-generic instruction and replace it by a (sequence of) generic instruction(s). Both optimizations rely on the ValueTracker introduced in r212100. The ValueTracker has been refactored to use the information from the TargetInstrInfo for non-generic instruction. As part of the refactoring, we switched the tracking from the index of the definition to the actual register (virtual or physical). This one change is to provide better consistency with register related APIs and to ease the use of the TargetInstrInfo. Moreover, this patch introduces a new helper class CopyRewriter used to ease the rewriting of generic copies (i.e., #1). Finally, this patch adds a dead code elimination pass right after the peephole optimizer to get rid of dead code that may appear after rewriting. This is related to <rdar://problem/12702965>. Review: http://reviews.llvm.org/D4874 llvm-svn: 216088
2014-08-21 01:41:48 +08:00
bool RewriteCurrentSource(unsigned NewReg, unsigned NewSubReg) override {
if (CurrentSrcIdx != 2)
return false;
// We are rewriting the inserted reg.
MachineOperand &MO = CopyLike.getOperand(CurrentSrcIdx);
MO.setReg(NewReg);
MO.setSubReg(NewSubReg);
return true;
}
};
/// \brief Specialized rewriter for EXTRACT_SUBREG instruction.
class ExtractSubregRewriter : public CopyRewriter {
const TargetInstrInfo &TII;
public:
ExtractSubregRewriter(MachineInstr &MI, const TargetInstrInfo &TII)
: CopyRewriter(MI), TII(TII) {
assert(MI.isExtractSubreg() && "Invalid instruction");
}
/// \brief See CopyRewriter::getNextRewritableSource.
/// Here CopyLike has the following form:
/// dst.dstSubIdx = EXTRACT_SUBREG Src, subIdx.
/// There is only one rewritable source: Src.subIdx,
/// which defines dst.dstSubIdx.
bool getNextRewritableSource(unsigned &SrcReg, unsigned &SrcSubReg,
unsigned &TrackReg,
unsigned &TrackSubReg) override {
// If we already get the only source we can rewrite, return false.
if (CurrentSrcIdx == 1)
return false;
// We are looking at v1 = EXTRACT_SUBREG v0, sub0.
CurrentSrcIdx = 1;
const MachineOperand &MOExtractedReg = CopyLike.getOperand(1);
SrcReg = MOExtractedReg.getReg();
2015-09-09 08:38:33 +08:00
// If we have to compose sub-register indices, bail out.
[PeepholeOptimizer] Refactor the advanced copy optimization to take advantage of the isRegSequence property. This is a follow-up of r215394 and r215404, which respectively introduces the isRegSequence property and uses it for ARM. Thanks to the property introduced by the previous commits, this patch is able to optimize the following sequence: vmov d0, r2, r3 vmov d1, r0, r1 vmov r0, s0 vmov r1, s2 udiv r0, r1, r0 vmov r1, s1 vmov r2, s3 udiv r1, r2, r1 vmov.32 d16[0], r0 vmov.32 d16[1], r1 vmov r0, r1, d16 bx lr into: udiv r0, r0, r2 udiv r1, r1, r3 vmov.32 d16[0], r0 vmov.32 d16[1], r1 vmov r0, r1, d16 bx lr This patch refactors how the copy optimizations are done in the peephole optimizer. Prior to this patch, we had one copy-related optimization that replaced a copy or bitcast by a generic, more suitable (in terms of register file), copy. With this patch, the peephole optimizer features two copy-related optimizations: 1. One for rewriting generic copies to generic copies: PeepholeOptimizer::optimizeCoalescableCopy. 2. One for replacing non-generic copies with generic copies: PeepholeOptimizer::optimizeUncoalescableCopy. The goals of these two optimizations are slightly different: one rewrite the operand of the instruction (#1), the other kills off the non-generic instruction and replace it by a (sequence of) generic instruction(s). Both optimizations rely on the ValueTracker introduced in r212100. The ValueTracker has been refactored to use the information from the TargetInstrInfo for non-generic instruction. As part of the refactoring, we switched the tracking from the index of the definition to the actual register (virtual or physical). This one change is to provide better consistency with register related APIs and to ease the use of the TargetInstrInfo. Moreover, this patch introduces a new helper class CopyRewriter used to ease the rewriting of generic copies (i.e., #1). Finally, this patch adds a dead code elimination pass right after the peephole optimizer to get rid of dead code that may appear after rewriting. This is related to <rdar://problem/12702965>. Review: http://reviews.llvm.org/D4874 llvm-svn: 216088
2014-08-21 01:41:48 +08:00
if (MOExtractedReg.getSubReg())
return false;
SrcSubReg = CopyLike.getOperand(2).getImm();
// We want to track something that is compatible with the definition.
const MachineOperand &MODef = CopyLike.getOperand(0);
TrackReg = MODef.getReg();
TrackSubReg = MODef.getSubReg();
return true;
}
bool RewriteCurrentSource(unsigned NewReg, unsigned NewSubReg) override {
// The only source we can rewrite is the input register.
if (CurrentSrcIdx != 1)
return false;
CopyLike.getOperand(CurrentSrcIdx).setReg(NewReg);
// If we find a source that does not require to extract something,
// rewrite the operation with a copy.
if (!NewSubReg) {
// Move the current index to an invalid position.
// We do not want another call to this method to be able
// to do any change.
CurrentSrcIdx = -1;
// Rewrite the operation as a COPY.
// Get rid of the sub-register index.
CopyLike.RemoveOperand(2);
// Morph the operation into a COPY.
CopyLike.setDesc(TII.get(TargetOpcode::COPY));
return true;
}
CopyLike.getOperand(CurrentSrcIdx + 1).setImm(NewSubReg);
return true;
}
};
/// \brief Specialized rewriter for REG_SEQUENCE instruction.
class RegSequenceRewriter : public CopyRewriter {
public:
RegSequenceRewriter(MachineInstr &MI) : CopyRewriter(MI) {
assert(MI.isRegSequence() && "Invalid instruction");
}
/// \brief See CopyRewriter::getNextRewritableSource.
/// Here CopyLike has the following form:
/// dst = REG_SEQUENCE Src1.src1SubIdx, subIdx1, Src2.src2SubIdx, subIdx2.
/// Each call will return a different source, walking all the available
/// source.
///
/// The first call returns:
/// (SrcReg, SrcSubReg) = (Src1, src1SubIdx).
/// (TrackReg, TrackSubReg) = (dst, subIdx1).
///
/// The second call returns:
/// (SrcReg, SrcSubReg) = (Src2, src2SubIdx).
/// (TrackReg, TrackSubReg) = (dst, subIdx2).
///
/// And so on, until all the sources have been traversed, then
/// it returns false.
bool getNextRewritableSource(unsigned &SrcReg, unsigned &SrcSubReg,
unsigned &TrackReg,
unsigned &TrackSubReg) override {
// We are looking at v0 = REG_SEQUENCE v1, sub1, v2, sub2, etc.
// If this is the first call, move to the first argument.
if (CurrentSrcIdx == 0) {
CurrentSrcIdx = 1;
} else {
// Otherwise, move to the next argument and check that it is valid.
CurrentSrcIdx += 2;
if (CurrentSrcIdx >= CopyLike.getNumOperands())
return false;
}
const MachineOperand &MOInsertedReg = CopyLike.getOperand(CurrentSrcIdx);
SrcReg = MOInsertedReg.getReg();
2015-09-09 08:38:33 +08:00
// If we have to compose sub-register indices, bail out.
[PeepholeOptimizer] Refactor the advanced copy optimization to take advantage of the isRegSequence property. This is a follow-up of r215394 and r215404, which respectively introduces the isRegSequence property and uses it for ARM. Thanks to the property introduced by the previous commits, this patch is able to optimize the following sequence: vmov d0, r2, r3 vmov d1, r0, r1 vmov r0, s0 vmov r1, s2 udiv r0, r1, r0 vmov r1, s1 vmov r2, s3 udiv r1, r2, r1 vmov.32 d16[0], r0 vmov.32 d16[1], r1 vmov r0, r1, d16 bx lr into: udiv r0, r0, r2 udiv r1, r1, r3 vmov.32 d16[0], r0 vmov.32 d16[1], r1 vmov r0, r1, d16 bx lr This patch refactors how the copy optimizations are done in the peephole optimizer. Prior to this patch, we had one copy-related optimization that replaced a copy or bitcast by a generic, more suitable (in terms of register file), copy. With this patch, the peephole optimizer features two copy-related optimizations: 1. One for rewriting generic copies to generic copies: PeepholeOptimizer::optimizeCoalescableCopy. 2. One for replacing non-generic copies with generic copies: PeepholeOptimizer::optimizeUncoalescableCopy. The goals of these two optimizations are slightly different: one rewrite the operand of the instruction (#1), the other kills off the non-generic instruction and replace it by a (sequence of) generic instruction(s). Both optimizations rely on the ValueTracker introduced in r212100. The ValueTracker has been refactored to use the information from the TargetInstrInfo for non-generic instruction. As part of the refactoring, we switched the tracking from the index of the definition to the actual register (virtual or physical). This one change is to provide better consistency with register related APIs and to ease the use of the TargetInstrInfo. Moreover, this patch introduces a new helper class CopyRewriter used to ease the rewriting of generic copies (i.e., #1). Finally, this patch adds a dead code elimination pass right after the peephole optimizer to get rid of dead code that may appear after rewriting. This is related to <rdar://problem/12702965>. Review: http://reviews.llvm.org/D4874 llvm-svn: 216088
2014-08-21 01:41:48 +08:00
if ((SrcSubReg = MOInsertedReg.getSubReg()))
return false;
// We want to track something that is compatible with the related
// partial definition.
TrackSubReg = CopyLike.getOperand(CurrentSrcIdx + 1).getImm();
const MachineOperand &MODef = CopyLike.getOperand(0);
TrackReg = MODef.getReg();
2015-09-09 08:38:33 +08:00
// If we have to compose sub-registers, bail.
[PeepholeOptimizer] Refactor the advanced copy optimization to take advantage of the isRegSequence property. This is a follow-up of r215394 and r215404, which respectively introduces the isRegSequence property and uses it for ARM. Thanks to the property introduced by the previous commits, this patch is able to optimize the following sequence: vmov d0, r2, r3 vmov d1, r0, r1 vmov r0, s0 vmov r1, s2 udiv r0, r1, r0 vmov r1, s1 vmov r2, s3 udiv r1, r2, r1 vmov.32 d16[0], r0 vmov.32 d16[1], r1 vmov r0, r1, d16 bx lr into: udiv r0, r0, r2 udiv r1, r1, r3 vmov.32 d16[0], r0 vmov.32 d16[1], r1 vmov r0, r1, d16 bx lr This patch refactors how the copy optimizations are done in the peephole optimizer. Prior to this patch, we had one copy-related optimization that replaced a copy or bitcast by a generic, more suitable (in terms of register file), copy. With this patch, the peephole optimizer features two copy-related optimizations: 1. One for rewriting generic copies to generic copies: PeepholeOptimizer::optimizeCoalescableCopy. 2. One for replacing non-generic copies with generic copies: PeepholeOptimizer::optimizeUncoalescableCopy. The goals of these two optimizations are slightly different: one rewrite the operand of the instruction (#1), the other kills off the non-generic instruction and replace it by a (sequence of) generic instruction(s). Both optimizations rely on the ValueTracker introduced in r212100. The ValueTracker has been refactored to use the information from the TargetInstrInfo for non-generic instruction. As part of the refactoring, we switched the tracking from the index of the definition to the actual register (virtual or physical). This one change is to provide better consistency with register related APIs and to ease the use of the TargetInstrInfo. Moreover, this patch introduces a new helper class CopyRewriter used to ease the rewriting of generic copies (i.e., #1). Finally, this patch adds a dead code elimination pass right after the peephole optimizer to get rid of dead code that may appear after rewriting. This is related to <rdar://problem/12702965>. Review: http://reviews.llvm.org/D4874 llvm-svn: 216088
2014-08-21 01:41:48 +08:00
return MODef.getSubReg() == 0;
}
bool RewriteCurrentSource(unsigned NewReg, unsigned NewSubReg) override {
// We cannot rewrite out of bound operands.
// Moreover, rewritable sources are at odd positions.
if ((CurrentSrcIdx & 1) != 1 || CurrentSrcIdx > CopyLike.getNumOperands())
return false;
MachineOperand &MO = CopyLike.getOperand(CurrentSrcIdx);
MO.setReg(NewReg);
MO.setSubReg(NewSubReg);
return true;
}
};
} // end anonymous namespace
[PeepholeOptimizer] Refactor the advanced copy optimization to take advantage of the isRegSequence property. This is a follow-up of r215394 and r215404, which respectively introduces the isRegSequence property and uses it for ARM. Thanks to the property introduced by the previous commits, this patch is able to optimize the following sequence: vmov d0, r2, r3 vmov d1, r0, r1 vmov r0, s0 vmov r1, s2 udiv r0, r1, r0 vmov r1, s1 vmov r2, s3 udiv r1, r2, r1 vmov.32 d16[0], r0 vmov.32 d16[1], r1 vmov r0, r1, d16 bx lr into: udiv r0, r0, r2 udiv r1, r1, r3 vmov.32 d16[0], r0 vmov.32 d16[1], r1 vmov r0, r1, d16 bx lr This patch refactors how the copy optimizations are done in the peephole optimizer. Prior to this patch, we had one copy-related optimization that replaced a copy or bitcast by a generic, more suitable (in terms of register file), copy. With this patch, the peephole optimizer features two copy-related optimizations: 1. One for rewriting generic copies to generic copies: PeepholeOptimizer::optimizeCoalescableCopy. 2. One for replacing non-generic copies with generic copies: PeepholeOptimizer::optimizeUncoalescableCopy. The goals of these two optimizations are slightly different: one rewrite the operand of the instruction (#1), the other kills off the non-generic instruction and replace it by a (sequence of) generic instruction(s). Both optimizations rely on the ValueTracker introduced in r212100. The ValueTracker has been refactored to use the information from the TargetInstrInfo for non-generic instruction. As part of the refactoring, we switched the tracking from the index of the definition to the actual register (virtual or physical). This one change is to provide better consistency with register related APIs and to ease the use of the TargetInstrInfo. Moreover, this patch introduces a new helper class CopyRewriter used to ease the rewriting of generic copies (i.e., #1). Finally, this patch adds a dead code elimination pass right after the peephole optimizer to get rid of dead code that may appear after rewriting. This is related to <rdar://problem/12702965>. Review: http://reviews.llvm.org/D4874 llvm-svn: 216088
2014-08-21 01:41:48 +08:00
/// \brief Get the appropriated CopyRewriter for \p MI.
/// \return A pointer to a dynamically allocated CopyRewriter or nullptr
/// if no rewriter works for \p MI.
static CopyRewriter *getCopyRewriter(MachineInstr &MI,
const TargetInstrInfo &TII,
MachineRegisterInfo &MRI) {
// Handle uncoalescable copy-like instructions.
if (MI.isBitcast() || (MI.isRegSequenceLike() || MI.isInsertSubregLike() ||
MI.isExtractSubregLike()))
return new UncoalescableRewriter(MI, TII, MRI);
[PeepholeOptimizer] Refactor the advanced copy optimization to take advantage of the isRegSequence property. This is a follow-up of r215394 and r215404, which respectively introduces the isRegSequence property and uses it for ARM. Thanks to the property introduced by the previous commits, this patch is able to optimize the following sequence: vmov d0, r2, r3 vmov d1, r0, r1 vmov r0, s0 vmov r1, s2 udiv r0, r1, r0 vmov r1, s1 vmov r2, s3 udiv r1, r2, r1 vmov.32 d16[0], r0 vmov.32 d16[1], r1 vmov r0, r1, d16 bx lr into: udiv r0, r0, r2 udiv r1, r1, r3 vmov.32 d16[0], r0 vmov.32 d16[1], r1 vmov r0, r1, d16 bx lr This patch refactors how the copy optimizations are done in the peephole optimizer. Prior to this patch, we had one copy-related optimization that replaced a copy or bitcast by a generic, more suitable (in terms of register file), copy. With this patch, the peephole optimizer features two copy-related optimizations: 1. One for rewriting generic copies to generic copies: PeepholeOptimizer::optimizeCoalescableCopy. 2. One for replacing non-generic copies with generic copies: PeepholeOptimizer::optimizeUncoalescableCopy. The goals of these two optimizations are slightly different: one rewrite the operand of the instruction (#1), the other kills off the non-generic instruction and replace it by a (sequence of) generic instruction(s). Both optimizations rely on the ValueTracker introduced in r212100. The ValueTracker has been refactored to use the information from the TargetInstrInfo for non-generic instruction. As part of the refactoring, we switched the tracking from the index of the definition to the actual register (virtual or physical). This one change is to provide better consistency with register related APIs and to ease the use of the TargetInstrInfo. Moreover, this patch introduces a new helper class CopyRewriter used to ease the rewriting of generic copies (i.e., #1). Finally, this patch adds a dead code elimination pass right after the peephole optimizer to get rid of dead code that may appear after rewriting. This is related to <rdar://problem/12702965>. Review: http://reviews.llvm.org/D4874 llvm-svn: 216088
2014-08-21 01:41:48 +08:00
switch (MI.getOpcode()) {
default:
return nullptr;
case TargetOpcode::COPY:
return new CopyRewriter(MI);
case TargetOpcode::INSERT_SUBREG:
return new InsertSubregRewriter(MI);
case TargetOpcode::EXTRACT_SUBREG:
return new ExtractSubregRewriter(MI, TII);
case TargetOpcode::REG_SEQUENCE:
return new RegSequenceRewriter(MI);
}
llvm_unreachable(nullptr);
}
/// \brief Optimize generic copy instructions to avoid cross
/// register bank copy. The optimization looks through a chain of
/// copies and tries to find a source that has a compatible register
/// class.
/// Two register classes are considered to be compatible if they share
/// the same register bank.
/// New copies issued by this optimization are register allocator
/// friendly. This optimization does not remove any copy as it may
2015-09-09 08:38:33 +08:00
/// overconstrain the register allocator, but replaces some operands
[PeepholeOptimizer] Refactor the advanced copy optimization to take advantage of the isRegSequence property. This is a follow-up of r215394 and r215404, which respectively introduces the isRegSequence property and uses it for ARM. Thanks to the property introduced by the previous commits, this patch is able to optimize the following sequence: vmov d0, r2, r3 vmov d1, r0, r1 vmov r0, s0 vmov r1, s2 udiv r0, r1, r0 vmov r1, s1 vmov r2, s3 udiv r1, r2, r1 vmov.32 d16[0], r0 vmov.32 d16[1], r1 vmov r0, r1, d16 bx lr into: udiv r0, r0, r2 udiv r1, r1, r3 vmov.32 d16[0], r0 vmov.32 d16[1], r1 vmov r0, r1, d16 bx lr This patch refactors how the copy optimizations are done in the peephole optimizer. Prior to this patch, we had one copy-related optimization that replaced a copy or bitcast by a generic, more suitable (in terms of register file), copy. With this patch, the peephole optimizer features two copy-related optimizations: 1. One for rewriting generic copies to generic copies: PeepholeOptimizer::optimizeCoalescableCopy. 2. One for replacing non-generic copies with generic copies: PeepholeOptimizer::optimizeUncoalescableCopy. The goals of these two optimizations are slightly different: one rewrite the operand of the instruction (#1), the other kills off the non-generic instruction and replace it by a (sequence of) generic instruction(s). Both optimizations rely on the ValueTracker introduced in r212100. The ValueTracker has been refactored to use the information from the TargetInstrInfo for non-generic instruction. As part of the refactoring, we switched the tracking from the index of the definition to the actual register (virtual or physical). This one change is to provide better consistency with register related APIs and to ease the use of the TargetInstrInfo. Moreover, this patch introduces a new helper class CopyRewriter used to ease the rewriting of generic copies (i.e., #1). Finally, this patch adds a dead code elimination pass right after the peephole optimizer to get rid of dead code that may appear after rewriting. This is related to <rdar://problem/12702965>. Review: http://reviews.llvm.org/D4874 llvm-svn: 216088
2014-08-21 01:41:48 +08:00
/// when possible.
/// \pre isCoalescableCopy(*MI) is true.
/// \return True, when \p MI has been rewritten. False otherwise.
bool PeepholeOptimizer::optimizeCoalescableCopy(MachineInstr *MI) {
assert(MI && isCoalescableCopy(*MI) && "Invalid argument");
assert(MI->getDesc().getNumDefs() == 1 &&
"Coalescer can understand multiple defs?!");
const MachineOperand &MODef = MI->getOperand(0);
// Do not rewrite physical definitions.
if (TargetRegisterInfo::isPhysicalRegister(MODef.getReg()))
return false;
bool Changed = false;
// Get the right rewriter for the current copy.
std::unique_ptr<CopyRewriter> CpyRewriter(getCopyRewriter(*MI, *TII, *MRI));
2015-09-09 08:38:33 +08:00
// If none exists, bail out.
[PeepholeOptimizer] Refactor the advanced copy optimization to take advantage of the isRegSequence property. This is a follow-up of r215394 and r215404, which respectively introduces the isRegSequence property and uses it for ARM. Thanks to the property introduced by the previous commits, this patch is able to optimize the following sequence: vmov d0, r2, r3 vmov d1, r0, r1 vmov r0, s0 vmov r1, s2 udiv r0, r1, r0 vmov r1, s1 vmov r2, s3 udiv r1, r2, r1 vmov.32 d16[0], r0 vmov.32 d16[1], r1 vmov r0, r1, d16 bx lr into: udiv r0, r0, r2 udiv r1, r1, r3 vmov.32 d16[0], r0 vmov.32 d16[1], r1 vmov r0, r1, d16 bx lr This patch refactors how the copy optimizations are done in the peephole optimizer. Prior to this patch, we had one copy-related optimization that replaced a copy or bitcast by a generic, more suitable (in terms of register file), copy. With this patch, the peephole optimizer features two copy-related optimizations: 1. One for rewriting generic copies to generic copies: PeepholeOptimizer::optimizeCoalescableCopy. 2. One for replacing non-generic copies with generic copies: PeepholeOptimizer::optimizeUncoalescableCopy. The goals of these two optimizations are slightly different: one rewrite the operand of the instruction (#1), the other kills off the non-generic instruction and replace it by a (sequence of) generic instruction(s). Both optimizations rely on the ValueTracker introduced in r212100. The ValueTracker has been refactored to use the information from the TargetInstrInfo for non-generic instruction. As part of the refactoring, we switched the tracking from the index of the definition to the actual register (virtual or physical). This one change is to provide better consistency with register related APIs and to ease the use of the TargetInstrInfo. Moreover, this patch introduces a new helper class CopyRewriter used to ease the rewriting of generic copies (i.e., #1). Finally, this patch adds a dead code elimination pass right after the peephole optimizer to get rid of dead code that may appear after rewriting. This is related to <rdar://problem/12702965>. Review: http://reviews.llvm.org/D4874 llvm-svn: 216088
2014-08-21 01:41:48 +08:00
if (!CpyRewriter)
return false;
// Rewrite each rewritable source.
unsigned SrcReg, SrcSubReg, TrackReg, TrackSubReg;
while (CpyRewriter->getNextRewritableSource(SrcReg, SrcSubReg, TrackReg,
TrackSubReg)) {
// Keep track of PHI nodes and its incoming edges when looking for sources.
RewriteMapTy RewriteMap;
// Try to find a more suitable source. If we failed to do so, or get the
// actual source, move to the next source.
if (!findNextSource(TrackReg, TrackSubReg, RewriteMap))
continue;
// Get the new source to rewrite. TODO: Only enable handling of multiple
// sources (PHIs) once we have a motivating example and testcases for it.
TargetInstrInfo::RegSubRegPair TrackPair(TrackReg, TrackSubReg);
TargetInstrInfo::RegSubRegPair NewSrc = CpyRewriter->getNewSource(
MRI, TII, TrackPair, RewriteMap, false /* multiple sources */);
if (SrcReg == NewSrc.Reg || NewSrc.Reg == 0)
[PeepholeOptimizer] Refactor the advanced copy optimization to take advantage of the isRegSequence property. This is a follow-up of r215394 and r215404, which respectively introduces the isRegSequence property and uses it for ARM. Thanks to the property introduced by the previous commits, this patch is able to optimize the following sequence: vmov d0, r2, r3 vmov d1, r0, r1 vmov r0, s0 vmov r1, s2 udiv r0, r1, r0 vmov r1, s1 vmov r2, s3 udiv r1, r2, r1 vmov.32 d16[0], r0 vmov.32 d16[1], r1 vmov r0, r1, d16 bx lr into: udiv r0, r0, r2 udiv r1, r1, r3 vmov.32 d16[0], r0 vmov.32 d16[1], r1 vmov r0, r1, d16 bx lr This patch refactors how the copy optimizations are done in the peephole optimizer. Prior to this patch, we had one copy-related optimization that replaced a copy or bitcast by a generic, more suitable (in terms of register file), copy. With this patch, the peephole optimizer features two copy-related optimizations: 1. One for rewriting generic copies to generic copies: PeepholeOptimizer::optimizeCoalescableCopy. 2. One for replacing non-generic copies with generic copies: PeepholeOptimizer::optimizeUncoalescableCopy. The goals of these two optimizations are slightly different: one rewrite the operand of the instruction (#1), the other kills off the non-generic instruction and replace it by a (sequence of) generic instruction(s). Both optimizations rely on the ValueTracker introduced in r212100. The ValueTracker has been refactored to use the information from the TargetInstrInfo for non-generic instruction. As part of the refactoring, we switched the tracking from the index of the definition to the actual register (virtual or physical). This one change is to provide better consistency with register related APIs and to ease the use of the TargetInstrInfo. Moreover, this patch introduces a new helper class CopyRewriter used to ease the rewriting of generic copies (i.e., #1). Finally, this patch adds a dead code elimination pass right after the peephole optimizer to get rid of dead code that may appear after rewriting. This is related to <rdar://problem/12702965>. Review: http://reviews.llvm.org/D4874 llvm-svn: 216088
2014-08-21 01:41:48 +08:00
continue;
[PeepholeOptimizer] Refactor the advanced copy optimization to take advantage of the isRegSequence property. This is a follow-up of r215394 and r215404, which respectively introduces the isRegSequence property and uses it for ARM. Thanks to the property introduced by the previous commits, this patch is able to optimize the following sequence: vmov d0, r2, r3 vmov d1, r0, r1 vmov r0, s0 vmov r1, s2 udiv r0, r1, r0 vmov r1, s1 vmov r2, s3 udiv r1, r2, r1 vmov.32 d16[0], r0 vmov.32 d16[1], r1 vmov r0, r1, d16 bx lr into: udiv r0, r0, r2 udiv r1, r1, r3 vmov.32 d16[0], r0 vmov.32 d16[1], r1 vmov r0, r1, d16 bx lr This patch refactors how the copy optimizations are done in the peephole optimizer. Prior to this patch, we had one copy-related optimization that replaced a copy or bitcast by a generic, more suitable (in terms of register file), copy. With this patch, the peephole optimizer features two copy-related optimizations: 1. One for rewriting generic copies to generic copies: PeepholeOptimizer::optimizeCoalescableCopy. 2. One for replacing non-generic copies with generic copies: PeepholeOptimizer::optimizeUncoalescableCopy. The goals of these two optimizations are slightly different: one rewrite the operand of the instruction (#1), the other kills off the non-generic instruction and replace it by a (sequence of) generic instruction(s). Both optimizations rely on the ValueTracker introduced in r212100. The ValueTracker has been refactored to use the information from the TargetInstrInfo for non-generic instruction. As part of the refactoring, we switched the tracking from the index of the definition to the actual register (virtual or physical). This one change is to provide better consistency with register related APIs and to ease the use of the TargetInstrInfo. Moreover, this patch introduces a new helper class CopyRewriter used to ease the rewriting of generic copies (i.e., #1). Finally, this patch adds a dead code elimination pass right after the peephole optimizer to get rid of dead code that may appear after rewriting. This is related to <rdar://problem/12702965>. Review: http://reviews.llvm.org/D4874 llvm-svn: 216088
2014-08-21 01:41:48 +08:00
// Rewrite source.
if (CpyRewriter->RewriteCurrentSource(NewSrc.Reg, NewSrc.SubReg)) {
// We may have extended the live-range of NewSrc, account for that.
MRI->clearKillFlags(NewSrc.Reg);
Changed = true;
}
[PeepholeOptimizer] Refactor the advanced copy optimization to take advantage of the isRegSequence property. This is a follow-up of r215394 and r215404, which respectively introduces the isRegSequence property and uses it for ARM. Thanks to the property introduced by the previous commits, this patch is able to optimize the following sequence: vmov d0, r2, r3 vmov d1, r0, r1 vmov r0, s0 vmov r1, s2 udiv r0, r1, r0 vmov r1, s1 vmov r2, s3 udiv r1, r2, r1 vmov.32 d16[0], r0 vmov.32 d16[1], r1 vmov r0, r1, d16 bx lr into: udiv r0, r0, r2 udiv r1, r1, r3 vmov.32 d16[0], r0 vmov.32 d16[1], r1 vmov r0, r1, d16 bx lr This patch refactors how the copy optimizations are done in the peephole optimizer. Prior to this patch, we had one copy-related optimization that replaced a copy or bitcast by a generic, more suitable (in terms of register file), copy. With this patch, the peephole optimizer features two copy-related optimizations: 1. One for rewriting generic copies to generic copies: PeepholeOptimizer::optimizeCoalescableCopy. 2. One for replacing non-generic copies with generic copies: PeepholeOptimizer::optimizeUncoalescableCopy. The goals of these two optimizations are slightly different: one rewrite the operand of the instruction (#1), the other kills off the non-generic instruction and replace it by a (sequence of) generic instruction(s). Both optimizations rely on the ValueTracker introduced in r212100. The ValueTracker has been refactored to use the information from the TargetInstrInfo for non-generic instruction. As part of the refactoring, we switched the tracking from the index of the definition to the actual register (virtual or physical). This one change is to provide better consistency with register related APIs and to ease the use of the TargetInstrInfo. Moreover, this patch introduces a new helper class CopyRewriter used to ease the rewriting of generic copies (i.e., #1). Finally, this patch adds a dead code elimination pass right after the peephole optimizer to get rid of dead code that may appear after rewriting. This is related to <rdar://problem/12702965>. Review: http://reviews.llvm.org/D4874 llvm-svn: 216088
2014-08-21 01:41:48 +08:00
}
// TODO: We could have a clean-up method to tidy the instruction.
// E.g., v0 = INSERT_SUBREG v1, v1.sub0, sub0
// => v0 = COPY v1
// Currently we haven't seen motivating example for that and we
// want to avoid untested code.
NumRewrittenCopies += Changed;
[PeepholeOptimizer] Refactor the advanced copy optimization to take advantage of the isRegSequence property. This is a follow-up of r215394 and r215404, which respectively introduces the isRegSequence property and uses it for ARM. Thanks to the property introduced by the previous commits, this patch is able to optimize the following sequence: vmov d0, r2, r3 vmov d1, r0, r1 vmov r0, s0 vmov r1, s2 udiv r0, r1, r0 vmov r1, s1 vmov r2, s3 udiv r1, r2, r1 vmov.32 d16[0], r0 vmov.32 d16[1], r1 vmov r0, r1, d16 bx lr into: udiv r0, r0, r2 udiv r1, r1, r3 vmov.32 d16[0], r0 vmov.32 d16[1], r1 vmov r0, r1, d16 bx lr This patch refactors how the copy optimizations are done in the peephole optimizer. Prior to this patch, we had one copy-related optimization that replaced a copy or bitcast by a generic, more suitable (in terms of register file), copy. With this patch, the peephole optimizer features two copy-related optimizations: 1. One for rewriting generic copies to generic copies: PeepholeOptimizer::optimizeCoalescableCopy. 2. One for replacing non-generic copies with generic copies: PeepholeOptimizer::optimizeUncoalescableCopy. The goals of these two optimizations are slightly different: one rewrite the operand of the instruction (#1), the other kills off the non-generic instruction and replace it by a (sequence of) generic instruction(s). Both optimizations rely on the ValueTracker introduced in r212100. The ValueTracker has been refactored to use the information from the TargetInstrInfo for non-generic instruction. As part of the refactoring, we switched the tracking from the index of the definition to the actual register (virtual or physical). This one change is to provide better consistency with register related APIs and to ease the use of the TargetInstrInfo. Moreover, this patch introduces a new helper class CopyRewriter used to ease the rewriting of generic copies (i.e., #1). Finally, this patch adds a dead code elimination pass right after the peephole optimizer to get rid of dead code that may appear after rewriting. This is related to <rdar://problem/12702965>. Review: http://reviews.llvm.org/D4874 llvm-svn: 216088
2014-08-21 01:41:48 +08:00
return Changed;
}
/// \brief Optimize copy-like instructions to create
/// register coalescer friendly instruction.
/// The optimization tries to kill-off the \p MI by looking
/// through a chain of copies to find a source that has a compatible
/// register class.
/// If such a source is found, it replace \p MI by a generic COPY
/// operation.
/// \pre isUncoalescableCopy(*MI) is true.
/// \return True, when \p MI has been optimized. In that case, \p MI has
/// been removed from its parent.
/// All COPY instructions created, are inserted in \p LocalMIs.
bool PeepholeOptimizer::optimizeUncoalescableCopy(
MachineInstr *MI, SmallPtrSetImpl<MachineInstr *> &LocalMIs) {
assert(MI && isUncoalescableCopy(*MI) && "Invalid argument");
// Check if we can rewrite all the values defined by this instruction.
SmallVector<TargetInstrInfo::RegSubRegPair, 4> RewritePairs;
// Get the right rewriter for the current copy.
std::unique_ptr<CopyRewriter> CpyRewriter(getCopyRewriter(*MI, *TII, *MRI));
2015-09-09 08:38:33 +08:00
// If none exists, bail out.
if (!CpyRewriter)
return false;
[PeepholeOptimizer] Refactor the advanced copy optimization to take advantage of the isRegSequence property. This is a follow-up of r215394 and r215404, which respectively introduces the isRegSequence property and uses it for ARM. Thanks to the property introduced by the previous commits, this patch is able to optimize the following sequence: vmov d0, r2, r3 vmov d1, r0, r1 vmov r0, s0 vmov r1, s2 udiv r0, r1, r0 vmov r1, s1 vmov r2, s3 udiv r1, r2, r1 vmov.32 d16[0], r0 vmov.32 d16[1], r1 vmov r0, r1, d16 bx lr into: udiv r0, r0, r2 udiv r1, r1, r3 vmov.32 d16[0], r0 vmov.32 d16[1], r1 vmov r0, r1, d16 bx lr This patch refactors how the copy optimizations are done in the peephole optimizer. Prior to this patch, we had one copy-related optimization that replaced a copy or bitcast by a generic, more suitable (in terms of register file), copy. With this patch, the peephole optimizer features two copy-related optimizations: 1. One for rewriting generic copies to generic copies: PeepholeOptimizer::optimizeCoalescableCopy. 2. One for replacing non-generic copies with generic copies: PeepholeOptimizer::optimizeUncoalescableCopy. The goals of these two optimizations are slightly different: one rewrite the operand of the instruction (#1), the other kills off the non-generic instruction and replace it by a (sequence of) generic instruction(s). Both optimizations rely on the ValueTracker introduced in r212100. The ValueTracker has been refactored to use the information from the TargetInstrInfo for non-generic instruction. As part of the refactoring, we switched the tracking from the index of the definition to the actual register (virtual or physical). This one change is to provide better consistency with register related APIs and to ease the use of the TargetInstrInfo. Moreover, this patch introduces a new helper class CopyRewriter used to ease the rewriting of generic copies (i.e., #1). Finally, this patch adds a dead code elimination pass right after the peephole optimizer to get rid of dead code that may appear after rewriting. This is related to <rdar://problem/12702965>. Review: http://reviews.llvm.org/D4874 llvm-svn: 216088
2014-08-21 01:41:48 +08:00
// Rewrite each rewritable source by generating new COPYs. This works
// differently from optimizeCoalescableCopy since it first makes sure that all
// definitions can be rewritten.
RewriteMapTy RewriteMap;
unsigned Reg, SubReg, CopyDefReg, CopyDefSubReg;
while (CpyRewriter->getNextRewritableSource(Reg, SubReg, CopyDefReg,
CopyDefSubReg)) {
[PeepholeOptimizer] Refactor the advanced copy optimization to take advantage of the isRegSequence property. This is a follow-up of r215394 and r215404, which respectively introduces the isRegSequence property and uses it for ARM. Thanks to the property introduced by the previous commits, this patch is able to optimize the following sequence: vmov d0, r2, r3 vmov d1, r0, r1 vmov r0, s0 vmov r1, s2 udiv r0, r1, r0 vmov r1, s1 vmov r2, s3 udiv r1, r2, r1 vmov.32 d16[0], r0 vmov.32 d16[1], r1 vmov r0, r1, d16 bx lr into: udiv r0, r0, r2 udiv r1, r1, r3 vmov.32 d16[0], r0 vmov.32 d16[1], r1 vmov r0, r1, d16 bx lr This patch refactors how the copy optimizations are done in the peephole optimizer. Prior to this patch, we had one copy-related optimization that replaced a copy or bitcast by a generic, more suitable (in terms of register file), copy. With this patch, the peephole optimizer features two copy-related optimizations: 1. One for rewriting generic copies to generic copies: PeepholeOptimizer::optimizeCoalescableCopy. 2. One for replacing non-generic copies with generic copies: PeepholeOptimizer::optimizeUncoalescableCopy. The goals of these two optimizations are slightly different: one rewrite the operand of the instruction (#1), the other kills off the non-generic instruction and replace it by a (sequence of) generic instruction(s). Both optimizations rely on the ValueTracker introduced in r212100. The ValueTracker has been refactored to use the information from the TargetInstrInfo for non-generic instruction. As part of the refactoring, we switched the tracking from the index of the definition to the actual register (virtual or physical). This one change is to provide better consistency with register related APIs and to ease the use of the TargetInstrInfo. Moreover, this patch introduces a new helper class CopyRewriter used to ease the rewriting of generic copies (i.e., #1). Finally, this patch adds a dead code elimination pass right after the peephole optimizer to get rid of dead code that may appear after rewriting. This is related to <rdar://problem/12702965>. Review: http://reviews.llvm.org/D4874 llvm-svn: 216088
2014-08-21 01:41:48 +08:00
// If a physical register is here, this is probably for a good reason.
// Do not rewrite that.
if (TargetRegisterInfo::isPhysicalRegister(CopyDefReg))
[PeepholeOptimizer] Refactor the advanced copy optimization to take advantage of the isRegSequence property. This is a follow-up of r215394 and r215404, which respectively introduces the isRegSequence property and uses it for ARM. Thanks to the property introduced by the previous commits, this patch is able to optimize the following sequence: vmov d0, r2, r3 vmov d1, r0, r1 vmov r0, s0 vmov r1, s2 udiv r0, r1, r0 vmov r1, s1 vmov r2, s3 udiv r1, r2, r1 vmov.32 d16[0], r0 vmov.32 d16[1], r1 vmov r0, r1, d16 bx lr into: udiv r0, r0, r2 udiv r1, r1, r3 vmov.32 d16[0], r0 vmov.32 d16[1], r1 vmov r0, r1, d16 bx lr This patch refactors how the copy optimizations are done in the peephole optimizer. Prior to this patch, we had one copy-related optimization that replaced a copy or bitcast by a generic, more suitable (in terms of register file), copy. With this patch, the peephole optimizer features two copy-related optimizations: 1. One for rewriting generic copies to generic copies: PeepholeOptimizer::optimizeCoalescableCopy. 2. One for replacing non-generic copies with generic copies: PeepholeOptimizer::optimizeUncoalescableCopy. The goals of these two optimizations are slightly different: one rewrite the operand of the instruction (#1), the other kills off the non-generic instruction and replace it by a (sequence of) generic instruction(s). Both optimizations rely on the ValueTracker introduced in r212100. The ValueTracker has been refactored to use the information from the TargetInstrInfo for non-generic instruction. As part of the refactoring, we switched the tracking from the index of the definition to the actual register (virtual or physical). This one change is to provide better consistency with register related APIs and to ease the use of the TargetInstrInfo. Moreover, this patch introduces a new helper class CopyRewriter used to ease the rewriting of generic copies (i.e., #1). Finally, this patch adds a dead code elimination pass right after the peephole optimizer to get rid of dead code that may appear after rewriting. This is related to <rdar://problem/12702965>. Review: http://reviews.llvm.org/D4874 llvm-svn: 216088
2014-08-21 01:41:48 +08:00
return false;
// If we do not know how to rewrite this definition, there is no point
// in trying to kill this instruction.
TargetInstrInfo::RegSubRegPair Def(CopyDefReg, CopyDefSubReg);
if (!findNextSource(Def.Reg, Def.SubReg, RewriteMap))
[PeepholeOptimizer] Refactor the advanced copy optimization to take advantage of the isRegSequence property. This is a follow-up of r215394 and r215404, which respectively introduces the isRegSequence property and uses it for ARM. Thanks to the property introduced by the previous commits, this patch is able to optimize the following sequence: vmov d0, r2, r3 vmov d1, r0, r1 vmov r0, s0 vmov r1, s2 udiv r0, r1, r0 vmov r1, s1 vmov r2, s3 udiv r1, r2, r1 vmov.32 d16[0], r0 vmov.32 d16[1], r1 vmov r0, r1, d16 bx lr into: udiv r0, r0, r2 udiv r1, r1, r3 vmov.32 d16[0], r0 vmov.32 d16[1], r1 vmov r0, r1, d16 bx lr This patch refactors how the copy optimizations are done in the peephole optimizer. Prior to this patch, we had one copy-related optimization that replaced a copy or bitcast by a generic, more suitable (in terms of register file), copy. With this patch, the peephole optimizer features two copy-related optimizations: 1. One for rewriting generic copies to generic copies: PeepholeOptimizer::optimizeCoalescableCopy. 2. One for replacing non-generic copies with generic copies: PeepholeOptimizer::optimizeUncoalescableCopy. The goals of these two optimizations are slightly different: one rewrite the operand of the instruction (#1), the other kills off the non-generic instruction and replace it by a (sequence of) generic instruction(s). Both optimizations rely on the ValueTracker introduced in r212100. The ValueTracker has been refactored to use the information from the TargetInstrInfo for non-generic instruction. As part of the refactoring, we switched the tracking from the index of the definition to the actual register (virtual or physical). This one change is to provide better consistency with register related APIs and to ease the use of the TargetInstrInfo. Moreover, this patch introduces a new helper class CopyRewriter used to ease the rewriting of generic copies (i.e., #1). Finally, this patch adds a dead code elimination pass right after the peephole optimizer to get rid of dead code that may appear after rewriting. This is related to <rdar://problem/12702965>. Review: http://reviews.llvm.org/D4874 llvm-svn: 216088
2014-08-21 01:41:48 +08:00
return false;
RewritePairs.push_back(Def);
[PeepholeOptimizer] Refactor the advanced copy optimization to take advantage of the isRegSequence property. This is a follow-up of r215394 and r215404, which respectively introduces the isRegSequence property and uses it for ARM. Thanks to the property introduced by the previous commits, this patch is able to optimize the following sequence: vmov d0, r2, r3 vmov d1, r0, r1 vmov r0, s0 vmov r1, s2 udiv r0, r1, r0 vmov r1, s1 vmov r2, s3 udiv r1, r2, r1 vmov.32 d16[0], r0 vmov.32 d16[1], r1 vmov r0, r1, d16 bx lr into: udiv r0, r0, r2 udiv r1, r1, r3 vmov.32 d16[0], r0 vmov.32 d16[1], r1 vmov r0, r1, d16 bx lr This patch refactors how the copy optimizations are done in the peephole optimizer. Prior to this patch, we had one copy-related optimization that replaced a copy or bitcast by a generic, more suitable (in terms of register file), copy. With this patch, the peephole optimizer features two copy-related optimizations: 1. One for rewriting generic copies to generic copies: PeepholeOptimizer::optimizeCoalescableCopy. 2. One for replacing non-generic copies with generic copies: PeepholeOptimizer::optimizeUncoalescableCopy. The goals of these two optimizations are slightly different: one rewrite the operand of the instruction (#1), the other kills off the non-generic instruction and replace it by a (sequence of) generic instruction(s). Both optimizations rely on the ValueTracker introduced in r212100. The ValueTracker has been refactored to use the information from the TargetInstrInfo for non-generic instruction. As part of the refactoring, we switched the tracking from the index of the definition to the actual register (virtual or physical). This one change is to provide better consistency with register related APIs and to ease the use of the TargetInstrInfo. Moreover, this patch introduces a new helper class CopyRewriter used to ease the rewriting of generic copies (i.e., #1). Finally, this patch adds a dead code elimination pass right after the peephole optimizer to get rid of dead code that may appear after rewriting. This is related to <rdar://problem/12702965>. Review: http://reviews.llvm.org/D4874 llvm-svn: 216088
2014-08-21 01:41:48 +08:00
}
[PeepholeOptimizer] Refactor the advanced copy optimization to take advantage of the isRegSequence property. This is a follow-up of r215394 and r215404, which respectively introduces the isRegSequence property and uses it for ARM. Thanks to the property introduced by the previous commits, this patch is able to optimize the following sequence: vmov d0, r2, r3 vmov d1, r0, r1 vmov r0, s0 vmov r1, s2 udiv r0, r1, r0 vmov r1, s1 vmov r2, s3 udiv r1, r2, r1 vmov.32 d16[0], r0 vmov.32 d16[1], r1 vmov r0, r1, d16 bx lr into: udiv r0, r0, r2 udiv r1, r1, r3 vmov.32 d16[0], r0 vmov.32 d16[1], r1 vmov r0, r1, d16 bx lr This patch refactors how the copy optimizations are done in the peephole optimizer. Prior to this patch, we had one copy-related optimization that replaced a copy or bitcast by a generic, more suitable (in terms of register file), copy. With this patch, the peephole optimizer features two copy-related optimizations: 1. One for rewriting generic copies to generic copies: PeepholeOptimizer::optimizeCoalescableCopy. 2. One for replacing non-generic copies with generic copies: PeepholeOptimizer::optimizeUncoalescableCopy. The goals of these two optimizations are slightly different: one rewrite the operand of the instruction (#1), the other kills off the non-generic instruction and replace it by a (sequence of) generic instruction(s). Both optimizations rely on the ValueTracker introduced in r212100. The ValueTracker has been refactored to use the information from the TargetInstrInfo for non-generic instruction. As part of the refactoring, we switched the tracking from the index of the definition to the actual register (virtual or physical). This one change is to provide better consistency with register related APIs and to ease the use of the TargetInstrInfo. Moreover, this patch introduces a new helper class CopyRewriter used to ease the rewriting of generic copies (i.e., #1). Finally, this patch adds a dead code elimination pass right after the peephole optimizer to get rid of dead code that may appear after rewriting. This is related to <rdar://problem/12702965>. Review: http://reviews.llvm.org/D4874 llvm-svn: 216088
2014-08-21 01:41:48 +08:00
// The change is possible for all defs, do it.
for (const auto &Def : RewritePairs) {
[PeepholeOptimizer] Refactor the advanced copy optimization to take advantage of the isRegSequence property. This is a follow-up of r215394 and r215404, which respectively introduces the isRegSequence property and uses it for ARM. Thanks to the property introduced by the previous commits, this patch is able to optimize the following sequence: vmov d0, r2, r3 vmov d1, r0, r1 vmov r0, s0 vmov r1, s2 udiv r0, r1, r0 vmov r1, s1 vmov r2, s3 udiv r1, r2, r1 vmov.32 d16[0], r0 vmov.32 d16[1], r1 vmov r0, r1, d16 bx lr into: udiv r0, r0, r2 udiv r1, r1, r3 vmov.32 d16[0], r0 vmov.32 d16[1], r1 vmov r0, r1, d16 bx lr This patch refactors how the copy optimizations are done in the peephole optimizer. Prior to this patch, we had one copy-related optimization that replaced a copy or bitcast by a generic, more suitable (in terms of register file), copy. With this patch, the peephole optimizer features two copy-related optimizations: 1. One for rewriting generic copies to generic copies: PeepholeOptimizer::optimizeCoalescableCopy. 2. One for replacing non-generic copies with generic copies: PeepholeOptimizer::optimizeUncoalescableCopy. The goals of these two optimizations are slightly different: one rewrite the operand of the instruction (#1), the other kills off the non-generic instruction and replace it by a (sequence of) generic instruction(s). Both optimizations rely on the ValueTracker introduced in r212100. The ValueTracker has been refactored to use the information from the TargetInstrInfo for non-generic instruction. As part of the refactoring, we switched the tracking from the index of the definition to the actual register (virtual or physical). This one change is to provide better consistency with register related APIs and to ease the use of the TargetInstrInfo. Moreover, this patch introduces a new helper class CopyRewriter used to ease the rewriting of generic copies (i.e., #1). Finally, this patch adds a dead code elimination pass right after the peephole optimizer to get rid of dead code that may appear after rewriting. This is related to <rdar://problem/12702965>. Review: http://reviews.llvm.org/D4874 llvm-svn: 216088
2014-08-21 01:41:48 +08:00
// Rewrite the "copy" in a way the register coalescer understands.
MachineInstr *NewCopy = CpyRewriter->RewriteSource(Def, RewriteMap);
assert(NewCopy && "Should be able to always generate a new copy");
LocalMIs.insert(NewCopy);
[PeepholeOptimizer] Refactor the advanced copy optimization to take advantage of the isRegSequence property. This is a follow-up of r215394 and r215404, which respectively introduces the isRegSequence property and uses it for ARM. Thanks to the property introduced by the previous commits, this patch is able to optimize the following sequence: vmov d0, r2, r3 vmov d1, r0, r1 vmov r0, s0 vmov r1, s2 udiv r0, r1, r0 vmov r1, s1 vmov r2, s3 udiv r1, r2, r1 vmov.32 d16[0], r0 vmov.32 d16[1], r1 vmov r0, r1, d16 bx lr into: udiv r0, r0, r2 udiv r1, r1, r3 vmov.32 d16[0], r0 vmov.32 d16[1], r1 vmov r0, r1, d16 bx lr This patch refactors how the copy optimizations are done in the peephole optimizer. Prior to this patch, we had one copy-related optimization that replaced a copy or bitcast by a generic, more suitable (in terms of register file), copy. With this patch, the peephole optimizer features two copy-related optimizations: 1. One for rewriting generic copies to generic copies: PeepholeOptimizer::optimizeCoalescableCopy. 2. One for replacing non-generic copies with generic copies: PeepholeOptimizer::optimizeUncoalescableCopy. The goals of these two optimizations are slightly different: one rewrite the operand of the instruction (#1), the other kills off the non-generic instruction and replace it by a (sequence of) generic instruction(s). Both optimizations rely on the ValueTracker introduced in r212100. The ValueTracker has been refactored to use the information from the TargetInstrInfo for non-generic instruction. As part of the refactoring, we switched the tracking from the index of the definition to the actual register (virtual or physical). This one change is to provide better consistency with register related APIs and to ease the use of the TargetInstrInfo. Moreover, this patch introduces a new helper class CopyRewriter used to ease the rewriting of generic copies (i.e., #1). Finally, this patch adds a dead code elimination pass right after the peephole optimizer to get rid of dead code that may appear after rewriting. This is related to <rdar://problem/12702965>. Review: http://reviews.llvm.org/D4874 llvm-svn: 216088
2014-08-21 01:41:48 +08:00
}
[PeepholeOptimizer] Refactor the advanced copy optimization to take advantage of the isRegSequence property. This is a follow-up of r215394 and r215404, which respectively introduces the isRegSequence property and uses it for ARM. Thanks to the property introduced by the previous commits, this patch is able to optimize the following sequence: vmov d0, r2, r3 vmov d1, r0, r1 vmov r0, s0 vmov r1, s2 udiv r0, r1, r0 vmov r1, s1 vmov r2, s3 udiv r1, r2, r1 vmov.32 d16[0], r0 vmov.32 d16[1], r1 vmov r0, r1, d16 bx lr into: udiv r0, r0, r2 udiv r1, r1, r3 vmov.32 d16[0], r0 vmov.32 d16[1], r1 vmov r0, r1, d16 bx lr This patch refactors how the copy optimizations are done in the peephole optimizer. Prior to this patch, we had one copy-related optimization that replaced a copy or bitcast by a generic, more suitable (in terms of register file), copy. With this patch, the peephole optimizer features two copy-related optimizations: 1. One for rewriting generic copies to generic copies: PeepholeOptimizer::optimizeCoalescableCopy. 2. One for replacing non-generic copies with generic copies: PeepholeOptimizer::optimizeUncoalescableCopy. The goals of these two optimizations are slightly different: one rewrite the operand of the instruction (#1), the other kills off the non-generic instruction and replace it by a (sequence of) generic instruction(s). Both optimizations rely on the ValueTracker introduced in r212100. The ValueTracker has been refactored to use the information from the TargetInstrInfo for non-generic instruction. As part of the refactoring, we switched the tracking from the index of the definition to the actual register (virtual or physical). This one change is to provide better consistency with register related APIs and to ease the use of the TargetInstrInfo. Moreover, this patch introduces a new helper class CopyRewriter used to ease the rewriting of generic copies (i.e., #1). Finally, this patch adds a dead code elimination pass right after the peephole optimizer to get rid of dead code that may appear after rewriting. This is related to <rdar://problem/12702965>. Review: http://reviews.llvm.org/D4874 llvm-svn: 216088
2014-08-21 01:41:48 +08:00
// MI is now dead.
MI->eraseFromParent();
[PeepholeOptimizer] Refactor the advanced copy optimization to take advantage of the isRegSequence property. This is a follow-up of r215394 and r215404, which respectively introduces the isRegSequence property and uses it for ARM. Thanks to the property introduced by the previous commits, this patch is able to optimize the following sequence: vmov d0, r2, r3 vmov d1, r0, r1 vmov r0, s0 vmov r1, s2 udiv r0, r1, r0 vmov r1, s1 vmov r2, s3 udiv r1, r2, r1 vmov.32 d16[0], r0 vmov.32 d16[1], r1 vmov r0, r1, d16 bx lr into: udiv r0, r0, r2 udiv r1, r1, r3 vmov.32 d16[0], r0 vmov.32 d16[1], r1 vmov r0, r1, d16 bx lr This patch refactors how the copy optimizations are done in the peephole optimizer. Prior to this patch, we had one copy-related optimization that replaced a copy or bitcast by a generic, more suitable (in terms of register file), copy. With this patch, the peephole optimizer features two copy-related optimizations: 1. One for rewriting generic copies to generic copies: PeepholeOptimizer::optimizeCoalescableCopy. 2. One for replacing non-generic copies with generic copies: PeepholeOptimizer::optimizeUncoalescableCopy. The goals of these two optimizations are slightly different: one rewrite the operand of the instruction (#1), the other kills off the non-generic instruction and replace it by a (sequence of) generic instruction(s). Both optimizations rely on the ValueTracker introduced in r212100. The ValueTracker has been refactored to use the information from the TargetInstrInfo for non-generic instruction. As part of the refactoring, we switched the tracking from the index of the definition to the actual register (virtual or physical). This one change is to provide better consistency with register related APIs and to ease the use of the TargetInstrInfo. Moreover, this patch introduces a new helper class CopyRewriter used to ease the rewriting of generic copies (i.e., #1). Finally, this patch adds a dead code elimination pass right after the peephole optimizer to get rid of dead code that may appear after rewriting. This is related to <rdar://problem/12702965>. Review: http://reviews.llvm.org/D4874 llvm-svn: 216088
2014-08-21 01:41:48 +08:00
++NumUncoalescableCopies;
return true;
}
/// Check whether MI is a candidate for folding into a later instruction.
/// We only fold loads to virtual registers and the virtual register defined
/// has a single use.
bool PeepholeOptimizer::isLoadFoldable(
2015-12-30 03:34:53 +08:00
MachineInstr *MI, SmallSet<unsigned, 16> &FoldAsLoadDefCandidates) {
if (!MI->canFoldAsLoad() || !MI->mayLoad())
return false;
const MCInstrDesc &MCID = MI->getDesc();
if (MCID.getNumDefs() != 1)
return false;
unsigned Reg = MI->getOperand(0).getReg();
// To reduce compilation time, we check MRI->hasOneNonDBGUse when inserting
// loads. It should be checked when processing uses of the load, since
// uses can be removed during peephole.
if (!MI->getOperand(0).getSubReg() &&
TargetRegisterInfo::isVirtualRegister(Reg) &&
MRI->hasOneNonDBGUse(Reg)) {
FoldAsLoadDefCandidates.insert(Reg);
return true;
}
return false;
}
2015-12-30 03:34:53 +08:00
bool PeepholeOptimizer::isMoveImmediate(
MachineInstr *MI, SmallSet<unsigned, 4> &ImmDefRegs,
DenseMap<unsigned, MachineInstr *> &ImmDefMIs) {
const MCInstrDesc &MCID = MI->getDesc();
if (!MI->isMoveImmediate())
return false;
if (MCID.getNumDefs() != 1)
return false;
unsigned Reg = MI->getOperand(0).getReg();
if (TargetRegisterInfo::isVirtualRegister(Reg)) {
ImmDefMIs.insert(std::make_pair(Reg, MI));
ImmDefRegs.insert(Reg);
return true;
}
2012-02-09 05:22:43 +08:00
return false;
}
/// Try folding register operands that are defined by move immediate
/// instructions, i.e. a trivial constant folding optimization, if
/// and only if the def and use are in the same BB.
2015-12-30 03:34:53 +08:00
bool PeepholeOptimizer::foldImmediate(
MachineInstr *MI, MachineBasicBlock *MBB, SmallSet<unsigned, 4> &ImmDefRegs,
DenseMap<unsigned, MachineInstr *> &ImmDefMIs) {
for (unsigned i = 0, e = MI->getDesc().getNumOperands(); i != e; ++i) {
MachineOperand &MO = MI->getOperand(i);
if (!MO.isReg() || MO.isDef())
continue;
// Ignore dead implicit defs.
if (MO.isImplicit() && MO.isDead())
continue;
unsigned Reg = MO.getReg();
if (!TargetRegisterInfo::isVirtualRegister(Reg))
continue;
if (ImmDefRegs.count(Reg) == 0)
continue;
DenseMap<unsigned, MachineInstr*>::iterator II = ImmDefMIs.find(Reg);
CodeGen peephole: fold redundant phys reg copies Code generation often exposes redundant physical register copies through virtual registers such as: %vreg = COPY %PHYSREG ... %PHYSREG = COPY %vreg There are cases where no intervening clobber of %PHYSREG occurs, and the later copy could therefore be removed. In some cases this further allows us to remove the initial copy. This patch contains a motivating example which comes from the x86 build of Chrome, specifically cc::ResourceProvider::UnlockForRead uses libstdc++'s implementation of hash_map. That example has two tests live at the same time, and after machine sinking LLVM has confused itself enough and things spilling EFLAGS is a great idea even though it's never restored and the comparison results are both live. Before this patch we have: DEC32m %RIP, 1, %noreg, <ga:@L>, %noreg, %EFLAGS<imp-def> %vreg1<def> = COPY %EFLAGS; GR64:%vreg1 %EFLAGS<def> = COPY %vreg1; GR64:%vreg1 JNE_1 <BB#1>, %EFLAGS<imp-use> Both copies are useless. This patch tries to eliminate the later copy in a generic manner. dec is especially confusing to LLVM when compared with sub. I wrote this patch to treat all physical registers generically, but only remove redundant copies of non-allocatable physical registers because the allocatable ones caused issues (e.g. when calling conventions weren't properly modeled) and should be handled later by the register allocator anyways. The following tests used to failed when the patch also replaced allocatable registers: CodeGen/X86/StackColoring.ll CodeGen/X86/avx512-calling-conv.ll CodeGen/X86/copy-propagation.ll CodeGen/X86/inline-asm-fpstack.ll CodeGen/X86/musttail-varargs.ll CodeGen/X86/pop-stack-cleanup.ll CodeGen/X86/preserve_mostcc64.ll CodeGen/X86/tailcallstack64.ll CodeGen/X86/this-return-64.ll This happens because COPY has other special meaning for e.g. dependency breakage and x87 FP stack. Note that all other backends' tests pass. Reviewers: qcolombet Subscribers: llvm-commits Differential Revision: http://reviews.llvm.org/D15157 llvm-svn: 254665
2015-12-04 07:43:56 +08:00
assert(II != ImmDefMIs.end() && "couldn't find immediate definition");
if (TII->FoldImmediate(*MI, *II->second, Reg, MRI)) {
++NumImmFold;
return true;
}
}
return false;
}
// FIXME: This is very simple and misses some cases which should be handled when
// motivating examples are found.
//
// The copy rewriting logic should look at uses as well as defs and be able to
// eliminate copies across blocks.
//
// Later copies that are subregister extracts will also not be eliminated since
// only the first copy is considered.
//
// e.g.
// %vreg1 = COPY %vreg0
// %vreg2 = COPY %vreg0:sub1
//
// Should replace %vreg2 uses with %vreg1:sub1
bool PeepholeOptimizer::foldRedundantCopy(
2015-12-30 03:34:53 +08:00
MachineInstr *MI, SmallSet<unsigned, 4> &CopySrcRegs,
CodeGen peephole: fold redundant phys reg copies Code generation often exposes redundant physical register copies through virtual registers such as: %vreg = COPY %PHYSREG ... %PHYSREG = COPY %vreg There are cases where no intervening clobber of %PHYSREG occurs, and the later copy could therefore be removed. In some cases this further allows us to remove the initial copy. This patch contains a motivating example which comes from the x86 build of Chrome, specifically cc::ResourceProvider::UnlockForRead uses libstdc++'s implementation of hash_map. That example has two tests live at the same time, and after machine sinking LLVM has confused itself enough and things spilling EFLAGS is a great idea even though it's never restored and the comparison results are both live. Before this patch we have: DEC32m %RIP, 1, %noreg, <ga:@L>, %noreg, %EFLAGS<imp-def> %vreg1<def> = COPY %EFLAGS; GR64:%vreg1 %EFLAGS<def> = COPY %vreg1; GR64:%vreg1 JNE_1 <BB#1>, %EFLAGS<imp-use> Both copies are useless. This patch tries to eliminate the later copy in a generic manner. dec is especially confusing to LLVM when compared with sub. I wrote this patch to treat all physical registers generically, but only remove redundant copies of non-allocatable physical registers because the allocatable ones caused issues (e.g. when calling conventions weren't properly modeled) and should be handled later by the register allocator anyways. The following tests used to failed when the patch also replaced allocatable registers: CodeGen/X86/StackColoring.ll CodeGen/X86/avx512-calling-conv.ll CodeGen/X86/copy-propagation.ll CodeGen/X86/inline-asm-fpstack.ll CodeGen/X86/musttail-varargs.ll CodeGen/X86/pop-stack-cleanup.ll CodeGen/X86/preserve_mostcc64.ll CodeGen/X86/tailcallstack64.ll CodeGen/X86/this-return-64.ll This happens because COPY has other special meaning for e.g. dependency breakage and x87 FP stack. Note that all other backends' tests pass. Reviewers: qcolombet Subscribers: llvm-commits Differential Revision: http://reviews.llvm.org/D15157 llvm-svn: 254665
2015-12-04 07:43:56 +08:00
DenseMap<unsigned, MachineInstr *> &CopyMIs) {
assert(MI->isCopy() && "expected a COPY machine instruction");
unsigned SrcReg = MI->getOperand(1).getReg();
if (!TargetRegisterInfo::isVirtualRegister(SrcReg))
return false;
unsigned DstReg = MI->getOperand(0).getReg();
if (!TargetRegisterInfo::isVirtualRegister(DstReg))
return false;
if (CopySrcRegs.insert(SrcReg).second) {
// First copy of this reg seen.
CopyMIs.insert(std::make_pair(SrcReg, MI));
return false;
}
MachineInstr *PrevCopy = CopyMIs.find(SrcReg)->second;
unsigned SrcSubReg = MI->getOperand(1).getSubReg();
unsigned PrevSrcSubReg = PrevCopy->getOperand(1).getSubReg();
// Can't replace different subregister extracts.
if (SrcSubReg != PrevSrcSubReg)
return false;
unsigned PrevDstReg = PrevCopy->getOperand(0).getReg();
// Only replace if the copy register class is the same.
//
// TODO: If we have multiple copies to different register classes, we may want
// to track multiple copies of the same source register.
if (MRI->getRegClass(DstReg) != MRI->getRegClass(PrevDstReg))
return false;
MRI->replaceRegWith(DstReg, PrevDstReg);
// Lifetime of the previous copy has been extended.
MRI->clearKillFlags(PrevDstReg);
return true;
}
CodeGen peephole: fold redundant phys reg copies Code generation often exposes redundant physical register copies through virtual registers such as: %vreg = COPY %PHYSREG ... %PHYSREG = COPY %vreg There are cases where no intervening clobber of %PHYSREG occurs, and the later copy could therefore be removed. In some cases this further allows us to remove the initial copy. This patch contains a motivating example which comes from the x86 build of Chrome, specifically cc::ResourceProvider::UnlockForRead uses libstdc++'s implementation of hash_map. That example has two tests live at the same time, and after machine sinking LLVM has confused itself enough and things spilling EFLAGS is a great idea even though it's never restored and the comparison results are both live. Before this patch we have: DEC32m %RIP, 1, %noreg, <ga:@L>, %noreg, %EFLAGS<imp-def> %vreg1<def> = COPY %EFLAGS; GR64:%vreg1 %EFLAGS<def> = COPY %vreg1; GR64:%vreg1 JNE_1 <BB#1>, %EFLAGS<imp-use> Both copies are useless. This patch tries to eliminate the later copy in a generic manner. dec is especially confusing to LLVM when compared with sub. I wrote this patch to treat all physical registers generically, but only remove redundant copies of non-allocatable physical registers because the allocatable ones caused issues (e.g. when calling conventions weren't properly modeled) and should be handled later by the register allocator anyways. The following tests used to failed when the patch also replaced allocatable registers: CodeGen/X86/StackColoring.ll CodeGen/X86/avx512-calling-conv.ll CodeGen/X86/copy-propagation.ll CodeGen/X86/inline-asm-fpstack.ll CodeGen/X86/musttail-varargs.ll CodeGen/X86/pop-stack-cleanup.ll CodeGen/X86/preserve_mostcc64.ll CodeGen/X86/tailcallstack64.ll CodeGen/X86/this-return-64.ll This happens because COPY has other special meaning for e.g. dependency breakage and x87 FP stack. Note that all other backends' tests pass. Reviewers: qcolombet Subscribers: llvm-commits Differential Revision: http://reviews.llvm.org/D15157 llvm-svn: 254665
2015-12-04 07:43:56 +08:00
bool PeepholeOptimizer::isNAPhysCopy(unsigned Reg) {
return TargetRegisterInfo::isPhysicalRegister(Reg) &&
!MRI->isAllocatable(Reg);
}
bool PeepholeOptimizer::foldRedundantNAPhysCopy(
MachineInstr *MI, DenseMap<unsigned, MachineInstr *> &NAPhysToVirtMIs) {
assert(MI->isCopy() && "expected a COPY machine instruction");
if (DisableNAPhysCopyOpt)
return false;
unsigned DstReg = MI->getOperand(0).getReg();
unsigned SrcReg = MI->getOperand(1).getReg();
if (isNAPhysCopy(SrcReg) && TargetRegisterInfo::isVirtualRegister(DstReg)) {
// %vreg = COPY %PHYSREG
// Avoid using a datastructure which can track multiple live non-allocatable
// phys->virt copies since LLVM doesn't seem to do this.
NAPhysToVirtMIs.insert({SrcReg, MI});
return false;
}
if (!(TargetRegisterInfo::isVirtualRegister(SrcReg) && isNAPhysCopy(DstReg)))
return false;
// %PHYSREG = COPY %vreg
auto PrevCopy = NAPhysToVirtMIs.find(DstReg);
if (PrevCopy == NAPhysToVirtMIs.end()) {
// We can't remove the copy: there was an intervening clobber of the
// non-allocatable physical register after the copy to virtual.
DEBUG(dbgs() << "NAPhysCopy: intervening clobber forbids erasing " << *MI
<< '\n');
return false;
}
unsigned PrevDstReg = PrevCopy->second->getOperand(0).getReg();
if (PrevDstReg == SrcReg) {
// Remove the virt->phys copy: we saw the virtual register definition, and
// the non-allocatable physical register's state hasn't changed since then.
DEBUG(dbgs() << "NAPhysCopy: erasing " << *MI << '\n');
++NumNAPhysCopies;
return true;
}
// Potential missed optimization opportunity: we saw a different virtual
// register get a copy of the non-allocatable physical register, and we only
// track one such copy. Avoid getting confused by this new non-allocatable
// physical register definition, and remove it from the tracked copies.
DEBUG(dbgs() << "NAPhysCopy: missed opportunity " << *MI << '\n');
NAPhysToVirtMIs.erase(PrevCopy);
return false;
}
bool PeepholeOptimizer::runOnMachineFunction(MachineFunction &MF) {
if (skipFunction(*MF.getFunction()))
return false;
DEBUG(dbgs() << "********** PEEPHOLE OPTIMIZER **********\n");
DEBUG(dbgs() << "********** Function: " << MF.getName() << '\n');
if (DisablePeephole)
return false;
2012-02-09 05:22:43 +08:00
TII = MF.getSubtarget().getInstrInfo();
TRI = MF.getSubtarget().getRegisterInfo();
MRI = &MF.getRegInfo();
DT = Aggressive ? &getAnalysis<MachineDominatorTree>() : nullptr;
bool Changed = false;
for (MachineBasicBlock &MBB : MF) {
bool SeenMoveImm = false;
// During this forward scan, at some point it needs to answer the question
// "given a pointer to an MI in the current BB, is it located before or
// after the current instruction".
// To perform this, the following set keeps track of the MIs already seen
// during the scan, if a MI is not in the set, it is assumed to be located
// after. Newly created MIs have to be inserted in the set as well.
SmallPtrSet<MachineInstr*, 16> LocalMIs;
SmallSet<unsigned, 4> ImmDefRegs;
DenseMap<unsigned, MachineInstr*> ImmDefMIs;
SmallSet<unsigned, 16> FoldAsLoadDefCandidates;
CodeGen peephole: fold redundant phys reg copies Code generation often exposes redundant physical register copies through virtual registers such as: %vreg = COPY %PHYSREG ... %PHYSREG = COPY %vreg There are cases where no intervening clobber of %PHYSREG occurs, and the later copy could therefore be removed. In some cases this further allows us to remove the initial copy. This patch contains a motivating example which comes from the x86 build of Chrome, specifically cc::ResourceProvider::UnlockForRead uses libstdc++'s implementation of hash_map. That example has two tests live at the same time, and after machine sinking LLVM has confused itself enough and things spilling EFLAGS is a great idea even though it's never restored and the comparison results are both live. Before this patch we have: DEC32m %RIP, 1, %noreg, <ga:@L>, %noreg, %EFLAGS<imp-def> %vreg1<def> = COPY %EFLAGS; GR64:%vreg1 %EFLAGS<def> = COPY %vreg1; GR64:%vreg1 JNE_1 <BB#1>, %EFLAGS<imp-use> Both copies are useless. This patch tries to eliminate the later copy in a generic manner. dec is especially confusing to LLVM when compared with sub. I wrote this patch to treat all physical registers generically, but only remove redundant copies of non-allocatable physical registers because the allocatable ones caused issues (e.g. when calling conventions weren't properly modeled) and should be handled later by the register allocator anyways. The following tests used to failed when the patch also replaced allocatable registers: CodeGen/X86/StackColoring.ll CodeGen/X86/avx512-calling-conv.ll CodeGen/X86/copy-propagation.ll CodeGen/X86/inline-asm-fpstack.ll CodeGen/X86/musttail-varargs.ll CodeGen/X86/pop-stack-cleanup.ll CodeGen/X86/preserve_mostcc64.ll CodeGen/X86/tailcallstack64.ll CodeGen/X86/this-return-64.ll This happens because COPY has other special meaning for e.g. dependency breakage and x87 FP stack. Note that all other backends' tests pass. Reviewers: qcolombet Subscribers: llvm-commits Differential Revision: http://reviews.llvm.org/D15157 llvm-svn: 254665
2015-12-04 07:43:56 +08:00
// Track when a non-allocatable physical register is copied to a virtual
// register so that useless moves can be removed.
//
// %PHYSREG is the map index; MI is the last valid `%vreg = COPY %PHYSREG`
// without any intervening re-definition of %PHYSREG.
DenseMap<unsigned, MachineInstr *> NAPhysToVirtMIs;
// Set of virtual registers that are copied from.
SmallSet<unsigned, 4> CopySrcRegs;
DenseMap<unsigned, MachineInstr *> CopySrcMIs;
for (MachineBasicBlock::iterator MII = MBB.begin(), MIE = MBB.end();
MII != MIE; ) {
MachineInstr *MI = &*MII;
// We may be erasing MI below, increment MII now.
++MII;
LocalMIs.insert(MI);
// Skip debug values. They should not affect this peephole optimization.
if (MI->isDebugValue())
continue;
CodeGen peephole: fold redundant phys reg copies Code generation often exposes redundant physical register copies through virtual registers such as: %vreg = COPY %PHYSREG ... %PHYSREG = COPY %vreg There are cases where no intervening clobber of %PHYSREG occurs, and the later copy could therefore be removed. In some cases this further allows us to remove the initial copy. This patch contains a motivating example which comes from the x86 build of Chrome, specifically cc::ResourceProvider::UnlockForRead uses libstdc++'s implementation of hash_map. That example has two tests live at the same time, and after machine sinking LLVM has confused itself enough and things spilling EFLAGS is a great idea even though it's never restored and the comparison results are both live. Before this patch we have: DEC32m %RIP, 1, %noreg, <ga:@L>, %noreg, %EFLAGS<imp-def> %vreg1<def> = COPY %EFLAGS; GR64:%vreg1 %EFLAGS<def> = COPY %vreg1; GR64:%vreg1 JNE_1 <BB#1>, %EFLAGS<imp-use> Both copies are useless. This patch tries to eliminate the later copy in a generic manner. dec is especially confusing to LLVM when compared with sub. I wrote this patch to treat all physical registers generically, but only remove redundant copies of non-allocatable physical registers because the allocatable ones caused issues (e.g. when calling conventions weren't properly modeled) and should be handled later by the register allocator anyways. The following tests used to failed when the patch also replaced allocatable registers: CodeGen/X86/StackColoring.ll CodeGen/X86/avx512-calling-conv.ll CodeGen/X86/copy-propagation.ll CodeGen/X86/inline-asm-fpstack.ll CodeGen/X86/musttail-varargs.ll CodeGen/X86/pop-stack-cleanup.ll CodeGen/X86/preserve_mostcc64.ll CodeGen/X86/tailcallstack64.ll CodeGen/X86/this-return-64.ll This happens because COPY has other special meaning for e.g. dependency breakage and x87 FP stack. Note that all other backends' tests pass. Reviewers: qcolombet Subscribers: llvm-commits Differential Revision: http://reviews.llvm.org/D15157 llvm-svn: 254665
2015-12-04 07:43:56 +08:00
if (MI->isPosition() || MI->isPHI())
continue;
CodeGen peephole: fold redundant phys reg copies Code generation often exposes redundant physical register copies through virtual registers such as: %vreg = COPY %PHYSREG ... %PHYSREG = COPY %vreg There are cases where no intervening clobber of %PHYSREG occurs, and the later copy could therefore be removed. In some cases this further allows us to remove the initial copy. This patch contains a motivating example which comes from the x86 build of Chrome, specifically cc::ResourceProvider::UnlockForRead uses libstdc++'s implementation of hash_map. That example has two tests live at the same time, and after machine sinking LLVM has confused itself enough and things spilling EFLAGS is a great idea even though it's never restored and the comparison results are both live. Before this patch we have: DEC32m %RIP, 1, %noreg, <ga:@L>, %noreg, %EFLAGS<imp-def> %vreg1<def> = COPY %EFLAGS; GR64:%vreg1 %EFLAGS<def> = COPY %vreg1; GR64:%vreg1 JNE_1 <BB#1>, %EFLAGS<imp-use> Both copies are useless. This patch tries to eliminate the later copy in a generic manner. dec is especially confusing to LLVM when compared with sub. I wrote this patch to treat all physical registers generically, but only remove redundant copies of non-allocatable physical registers because the allocatable ones caused issues (e.g. when calling conventions weren't properly modeled) and should be handled later by the register allocator anyways. The following tests used to failed when the patch also replaced allocatable registers: CodeGen/X86/StackColoring.ll CodeGen/X86/avx512-calling-conv.ll CodeGen/X86/copy-propagation.ll CodeGen/X86/inline-asm-fpstack.ll CodeGen/X86/musttail-varargs.ll CodeGen/X86/pop-stack-cleanup.ll CodeGen/X86/preserve_mostcc64.ll CodeGen/X86/tailcallstack64.ll CodeGen/X86/this-return-64.ll This happens because COPY has other special meaning for e.g. dependency breakage and x87 FP stack. Note that all other backends' tests pass. Reviewers: qcolombet Subscribers: llvm-commits Differential Revision: http://reviews.llvm.org/D15157 llvm-svn: 254665
2015-12-04 07:43:56 +08:00
if (!MI->isCopy()) {
for (const auto &Op : MI->operands()) {
// Visit all operands: definitions can be implicit or explicit.
if (Op.isReg()) {
unsigned Reg = Op.getReg();
if (Op.isDef() && isNAPhysCopy(Reg)) {
const auto &Def = NAPhysToVirtMIs.find(Reg);
if (Def != NAPhysToVirtMIs.end()) {
// A new definition of the non-allocatable physical register
// invalidates previous copies.
DEBUG(dbgs() << "NAPhysCopy: invalidating because of " << *MI
<< '\n');
NAPhysToVirtMIs.erase(Def);
}
}
} else if (Op.isRegMask()) {
const uint32_t *RegMask = Op.getRegMask();
for (auto &RegMI : NAPhysToVirtMIs) {
unsigned Def = RegMI.first;
if (MachineOperand::clobbersPhysReg(RegMask, Def)) {
DEBUG(dbgs() << "NAPhysCopy: invalidating because of " << *MI
<< '\n');
NAPhysToVirtMIs.erase(Def);
}
}
}
}
}
if (MI->isImplicitDef() || MI->isKill())
continue;
if (MI->isInlineAsm() || MI->hasUnmodeledSideEffects()) {
// Blow away all non-allocatable physical registers knowledge since we
// don't know what's correct anymore.
//
// FIXME: handle explicit asm clobbers.
DEBUG(dbgs() << "NAPhysCopy: blowing away all info due to " << *MI
<< '\n');
NAPhysToVirtMIs.clear();
}
[PeepholeOptimizer] Refactor the advanced copy optimization to take advantage of the isRegSequence property. This is a follow-up of r215394 and r215404, which respectively introduces the isRegSequence property and uses it for ARM. Thanks to the property introduced by the previous commits, this patch is able to optimize the following sequence: vmov d0, r2, r3 vmov d1, r0, r1 vmov r0, s0 vmov r1, s2 udiv r0, r1, r0 vmov r1, s1 vmov r2, s3 udiv r1, r2, r1 vmov.32 d16[0], r0 vmov.32 d16[1], r1 vmov r0, r1, d16 bx lr into: udiv r0, r0, r2 udiv r1, r1, r3 vmov.32 d16[0], r0 vmov.32 d16[1], r1 vmov r0, r1, d16 bx lr This patch refactors how the copy optimizations are done in the peephole optimizer. Prior to this patch, we had one copy-related optimization that replaced a copy or bitcast by a generic, more suitable (in terms of register file), copy. With this patch, the peephole optimizer features two copy-related optimizations: 1. One for rewriting generic copies to generic copies: PeepholeOptimizer::optimizeCoalescableCopy. 2. One for replacing non-generic copies with generic copies: PeepholeOptimizer::optimizeUncoalescableCopy. The goals of these two optimizations are slightly different: one rewrite the operand of the instruction (#1), the other kills off the non-generic instruction and replace it by a (sequence of) generic instruction(s). Both optimizations rely on the ValueTracker introduced in r212100. The ValueTracker has been refactored to use the information from the TargetInstrInfo for non-generic instruction. As part of the refactoring, we switched the tracking from the index of the definition to the actual register (virtual or physical). This one change is to provide better consistency with register related APIs and to ease the use of the TargetInstrInfo. Moreover, this patch introduces a new helper class CopyRewriter used to ease the rewriting of generic copies (i.e., #1). Finally, this patch adds a dead code elimination pass right after the peephole optimizer to get rid of dead code that may appear after rewriting. This is related to <rdar://problem/12702965>. Review: http://reviews.llvm.org/D4874 llvm-svn: 216088
2014-08-21 01:41:48 +08:00
if ((isUncoalescableCopy(*MI) &&
optimizeUncoalescableCopy(MI, LocalMIs)) ||
(MI->isCompare() && optimizeCmpInstr(MI, &MBB)) ||
(MI->isSelect() && optimizeSelect(MI, LocalMIs))) {
// MI is deleted.
LocalMIs.erase(MI);
Changed = true;
continue;
}
if (MI->isConditionalBranch() && optimizeCondBranch(MI)) {
Changed = true;
continue;
}
[PeepholeOptimizer] Refactor the advanced copy optimization to take advantage of the isRegSequence property. This is a follow-up of r215394 and r215404, which respectively introduces the isRegSequence property and uses it for ARM. Thanks to the property introduced by the previous commits, this patch is able to optimize the following sequence: vmov d0, r2, r3 vmov d1, r0, r1 vmov r0, s0 vmov r1, s2 udiv r0, r1, r0 vmov r1, s1 vmov r2, s3 udiv r1, r2, r1 vmov.32 d16[0], r0 vmov.32 d16[1], r1 vmov r0, r1, d16 bx lr into: udiv r0, r0, r2 udiv r1, r1, r3 vmov.32 d16[0], r0 vmov.32 d16[1], r1 vmov r0, r1, d16 bx lr This patch refactors how the copy optimizations are done in the peephole optimizer. Prior to this patch, we had one copy-related optimization that replaced a copy or bitcast by a generic, more suitable (in terms of register file), copy. With this patch, the peephole optimizer features two copy-related optimizations: 1. One for rewriting generic copies to generic copies: PeepholeOptimizer::optimizeCoalescableCopy. 2. One for replacing non-generic copies with generic copies: PeepholeOptimizer::optimizeUncoalescableCopy. The goals of these two optimizations are slightly different: one rewrite the operand of the instruction (#1), the other kills off the non-generic instruction and replace it by a (sequence of) generic instruction(s). Both optimizations rely on the ValueTracker introduced in r212100. The ValueTracker has been refactored to use the information from the TargetInstrInfo for non-generic instruction. As part of the refactoring, we switched the tracking from the index of the definition to the actual register (virtual or physical). This one change is to provide better consistency with register related APIs and to ease the use of the TargetInstrInfo. Moreover, this patch introduces a new helper class CopyRewriter used to ease the rewriting of generic copies (i.e., #1). Finally, this patch adds a dead code elimination pass right after the peephole optimizer to get rid of dead code that may appear after rewriting. This is related to <rdar://problem/12702965>. Review: http://reviews.llvm.org/D4874 llvm-svn: 216088
2014-08-21 01:41:48 +08:00
if (isCoalescableCopy(*MI) && optimizeCoalescableCopy(MI)) {
// MI is just rewritten.
Changed = true;
continue;
}
CodeGen peephole: fold redundant phys reg copies Code generation often exposes redundant physical register copies through virtual registers such as: %vreg = COPY %PHYSREG ... %PHYSREG = COPY %vreg There are cases where no intervening clobber of %PHYSREG occurs, and the later copy could therefore be removed. In some cases this further allows us to remove the initial copy. This patch contains a motivating example which comes from the x86 build of Chrome, specifically cc::ResourceProvider::UnlockForRead uses libstdc++'s implementation of hash_map. That example has two tests live at the same time, and after machine sinking LLVM has confused itself enough and things spilling EFLAGS is a great idea even though it's never restored and the comparison results are both live. Before this patch we have: DEC32m %RIP, 1, %noreg, <ga:@L>, %noreg, %EFLAGS<imp-def> %vreg1<def> = COPY %EFLAGS; GR64:%vreg1 %EFLAGS<def> = COPY %vreg1; GR64:%vreg1 JNE_1 <BB#1>, %EFLAGS<imp-use> Both copies are useless. This patch tries to eliminate the later copy in a generic manner. dec is especially confusing to LLVM when compared with sub. I wrote this patch to treat all physical registers generically, but only remove redundant copies of non-allocatable physical registers because the allocatable ones caused issues (e.g. when calling conventions weren't properly modeled) and should be handled later by the register allocator anyways. The following tests used to failed when the patch also replaced allocatable registers: CodeGen/X86/StackColoring.ll CodeGen/X86/avx512-calling-conv.ll CodeGen/X86/copy-propagation.ll CodeGen/X86/inline-asm-fpstack.ll CodeGen/X86/musttail-varargs.ll CodeGen/X86/pop-stack-cleanup.ll CodeGen/X86/preserve_mostcc64.ll CodeGen/X86/tailcallstack64.ll CodeGen/X86/this-return-64.ll This happens because COPY has other special meaning for e.g. dependency breakage and x87 FP stack. Note that all other backends' tests pass. Reviewers: qcolombet Subscribers: llvm-commits Differential Revision: http://reviews.llvm.org/D15157 llvm-svn: 254665
2015-12-04 07:43:56 +08:00
if (MI->isCopy() &&
(foldRedundantCopy(MI, CopySrcRegs, CopySrcMIs) ||
foldRedundantNAPhysCopy(MI, NAPhysToVirtMIs))) {
LocalMIs.erase(MI);
MI->eraseFromParent();
Changed = true;
continue;
}
if (isMoveImmediate(MI, ImmDefRegs, ImmDefMIs)) {
SeenMoveImm = true;
} else {
Changed |= optimizeExtInstr(MI, &MBB, LocalMIs);
// optimizeExtInstr might have created new instructions after MI
// and before the already incremented MII. Adjust MII so that the
// next iteration sees the new instructions.
MII = MI;
++MII;
if (SeenMoveImm)
Changed |= foldImmediate(MI, &MBB, ImmDefRegs, ImmDefMIs);
}
// Check whether MI is a load candidate for folding into a later
// instruction. If MI is not a candidate, check whether we can fold an
// earlier load into MI.
if (!isLoadFoldable(MI, FoldAsLoadDefCandidates) &&
!FoldAsLoadDefCandidates.empty()) {
// We visit each operand even after successfully folding a previous
// one. This allows us to fold multiple loads into a single
// instruction. We do assume that optimizeLoadInstr doesn't insert
// foldable uses earlier in the argument list. Since we don't restart
// iteration, we'd miss such cases.
const MCInstrDesc &MIDesc = MI->getDesc();
for (unsigned i = MIDesc.getNumDefs(); i != MI->getNumOperands();
++i) {
const MachineOperand &MOp = MI->getOperand(i);
if (!MOp.isReg())
continue;
unsigned FoldAsLoadDefReg = MOp.getReg();
if (FoldAsLoadDefCandidates.count(FoldAsLoadDefReg)) {
// We need to fold load after optimizeCmpInstr, since
// optimizeCmpInstr can enable folding by converting SUB to CMP.
// Save FoldAsLoadDefReg because optimizeLoadInstr() resets it and
// we need it for markUsesInDebugValueAsUndef().
unsigned FoldedReg = FoldAsLoadDefReg;
MachineInstr *DefMI = nullptr;
if (MachineInstr *FoldMI =
TII->optimizeLoadInstr(*MI, MRI, FoldAsLoadDefReg, DefMI)) {
// Update LocalMIs since we replaced MI with FoldMI and deleted
// DefMI.
DEBUG(dbgs() << "Replacing: " << *MI);
DEBUG(dbgs() << " With: " << *FoldMI);
LocalMIs.erase(MI);
LocalMIs.erase(DefMI);
LocalMIs.insert(FoldMI);
MI->eraseFromParent();
DefMI->eraseFromParent();
MRI->markUsesInDebugValueAsUndef(FoldedReg);
FoldAsLoadDefCandidates.erase(FoldedReg);
++NumLoadFold;
// MI is replaced with FoldMI so we can continue trying to fold
Changed = true;
MI = FoldMI;
}
}
}
}
// If we run into an instruction we can't fold across, discard
// the load candidates. Note: We might be able to fold *into* this
// instruction, so this needs to be after the folding logic.
if (MI->isLoadFoldBarrier()) {
DEBUG(dbgs() << "Encountered load fold barrier on " << *MI << "\n");
FoldAsLoadDefCandidates.clear();
}
}
}
return Changed;
}
ValueTrackerResult ValueTracker::getNextSourceFromCopy() {
assert(Def->isCopy() && "Invalid definition");
// Copy instruction are supposed to be: Def = Src.
// If someone breaks this assumption, bad things will happen everywhere.
[PeepholeOptimizer] Refactor the advanced copy optimization to take advantage of the isRegSequence property. This is a follow-up of r215394 and r215404, which respectively introduces the isRegSequence property and uses it for ARM. Thanks to the property introduced by the previous commits, this patch is able to optimize the following sequence: vmov d0, r2, r3 vmov d1, r0, r1 vmov r0, s0 vmov r1, s2 udiv r0, r1, r0 vmov r1, s1 vmov r2, s3 udiv r1, r2, r1 vmov.32 d16[0], r0 vmov.32 d16[1], r1 vmov r0, r1, d16 bx lr into: udiv r0, r0, r2 udiv r1, r1, r3 vmov.32 d16[0], r0 vmov.32 d16[1], r1 vmov r0, r1, d16 bx lr This patch refactors how the copy optimizations are done in the peephole optimizer. Prior to this patch, we had one copy-related optimization that replaced a copy or bitcast by a generic, more suitable (in terms of register file), copy. With this patch, the peephole optimizer features two copy-related optimizations: 1. One for rewriting generic copies to generic copies: PeepholeOptimizer::optimizeCoalescableCopy. 2. One for replacing non-generic copies with generic copies: PeepholeOptimizer::optimizeUncoalescableCopy. The goals of these two optimizations are slightly different: one rewrite the operand of the instruction (#1), the other kills off the non-generic instruction and replace it by a (sequence of) generic instruction(s). Both optimizations rely on the ValueTracker introduced in r212100. The ValueTracker has been refactored to use the information from the TargetInstrInfo for non-generic instruction. As part of the refactoring, we switched the tracking from the index of the definition to the actual register (virtual or physical). This one change is to provide better consistency with register related APIs and to ease the use of the TargetInstrInfo. Moreover, this patch introduces a new helper class CopyRewriter used to ease the rewriting of generic copies (i.e., #1). Finally, this patch adds a dead code elimination pass right after the peephole optimizer to get rid of dead code that may appear after rewriting. This is related to <rdar://problem/12702965>. Review: http://reviews.llvm.org/D4874 llvm-svn: 216088
2014-08-21 01:41:48 +08:00
assert(Def->getNumOperands() == 2 && "Invalid number of operands");
if (Def->getOperand(DefIdx).getSubReg() != DefSubReg)
// If we look for a different subreg, it means we want a subreg of src.
2015-09-09 08:38:33 +08:00
// Bails as we do not support composing subregs yet.
return ValueTrackerResult();
// Otherwise, we want the whole source.
[PeepholeOptimizer] Refactor the advanced copy optimization to take advantage of the isRegSequence property. This is a follow-up of r215394 and r215404, which respectively introduces the isRegSequence property and uses it for ARM. Thanks to the property introduced by the previous commits, this patch is able to optimize the following sequence: vmov d0, r2, r3 vmov d1, r0, r1 vmov r0, s0 vmov r1, s2 udiv r0, r1, r0 vmov r1, s1 vmov r2, s3 udiv r1, r2, r1 vmov.32 d16[0], r0 vmov.32 d16[1], r1 vmov r0, r1, d16 bx lr into: udiv r0, r0, r2 udiv r1, r1, r3 vmov.32 d16[0], r0 vmov.32 d16[1], r1 vmov r0, r1, d16 bx lr This patch refactors how the copy optimizations are done in the peephole optimizer. Prior to this patch, we had one copy-related optimization that replaced a copy or bitcast by a generic, more suitable (in terms of register file), copy. With this patch, the peephole optimizer features two copy-related optimizations: 1. One for rewriting generic copies to generic copies: PeepholeOptimizer::optimizeCoalescableCopy. 2. One for replacing non-generic copies with generic copies: PeepholeOptimizer::optimizeUncoalescableCopy. The goals of these two optimizations are slightly different: one rewrite the operand of the instruction (#1), the other kills off the non-generic instruction and replace it by a (sequence of) generic instruction(s). Both optimizations rely on the ValueTracker introduced in r212100. The ValueTracker has been refactored to use the information from the TargetInstrInfo for non-generic instruction. As part of the refactoring, we switched the tracking from the index of the definition to the actual register (virtual or physical). This one change is to provide better consistency with register related APIs and to ease the use of the TargetInstrInfo. Moreover, this patch introduces a new helper class CopyRewriter used to ease the rewriting of generic copies (i.e., #1). Finally, this patch adds a dead code elimination pass right after the peephole optimizer to get rid of dead code that may appear after rewriting. This is related to <rdar://problem/12702965>. Review: http://reviews.llvm.org/D4874 llvm-svn: 216088
2014-08-21 01:41:48 +08:00
const MachineOperand &Src = Def->getOperand(1);
return ValueTrackerResult(Src.getReg(), Src.getSubReg());
}
ValueTrackerResult ValueTracker::getNextSourceFromBitcast() {
assert(Def->isBitcast() && "Invalid definition");
// Bail if there are effects that a plain copy will not expose.
if (Def->hasUnmodeledSideEffects())
return ValueTrackerResult();
// Bitcasts with more than one def are not supported.
if (Def->getDesc().getNumDefs() != 1)
return ValueTrackerResult();
const MachineOperand DefOp = Def->getOperand(DefIdx);
if (DefOp.getSubReg() != DefSubReg)
// If we look for a different subreg, it means we want a subreg of the src.
2015-09-09 08:38:33 +08:00
// Bails as we do not support composing subregs yet.
return ValueTrackerResult();
[PeepholeOptimizer] Refactor the advanced copy optimization to take advantage of the isRegSequence property. This is a follow-up of r215394 and r215404, which respectively introduces the isRegSequence property and uses it for ARM. Thanks to the property introduced by the previous commits, this patch is able to optimize the following sequence: vmov d0, r2, r3 vmov d1, r0, r1 vmov r0, s0 vmov r1, s2 udiv r0, r1, r0 vmov r1, s1 vmov r2, s3 udiv r1, r2, r1 vmov.32 d16[0], r0 vmov.32 d16[1], r1 vmov r0, r1, d16 bx lr into: udiv r0, r0, r2 udiv r1, r1, r3 vmov.32 d16[0], r0 vmov.32 d16[1], r1 vmov r0, r1, d16 bx lr This patch refactors how the copy optimizations are done in the peephole optimizer. Prior to this patch, we had one copy-related optimization that replaced a copy or bitcast by a generic, more suitable (in terms of register file), copy. With this patch, the peephole optimizer features two copy-related optimizations: 1. One for rewriting generic copies to generic copies: PeepholeOptimizer::optimizeCoalescableCopy. 2. One for replacing non-generic copies with generic copies: PeepholeOptimizer::optimizeUncoalescableCopy. The goals of these two optimizations are slightly different: one rewrite the operand of the instruction (#1), the other kills off the non-generic instruction and replace it by a (sequence of) generic instruction(s). Both optimizations rely on the ValueTracker introduced in r212100. The ValueTracker has been refactored to use the information from the TargetInstrInfo for non-generic instruction. As part of the refactoring, we switched the tracking from the index of the definition to the actual register (virtual or physical). This one change is to provide better consistency with register related APIs and to ease the use of the TargetInstrInfo. Moreover, this patch introduces a new helper class CopyRewriter used to ease the rewriting of generic copies (i.e., #1). Finally, this patch adds a dead code elimination pass right after the peephole optimizer to get rid of dead code that may appear after rewriting. This is related to <rdar://problem/12702965>. Review: http://reviews.llvm.org/D4874 llvm-svn: 216088
2014-08-21 01:41:48 +08:00
unsigned SrcIdx = Def->getNumOperands();
for (unsigned OpIdx = DefIdx + 1, EndOpIdx = SrcIdx; OpIdx != EndOpIdx;
++OpIdx) {
const MachineOperand &MO = Def->getOperand(OpIdx);
if (!MO.isReg() || !MO.getReg())
continue;
// Ignore dead implicit defs.
if (MO.isImplicit() && MO.isDead())
continue;
assert(!MO.isDef() && "We should have skipped all the definitions by now");
if (SrcIdx != EndOpIdx)
// Multiple sources?
return ValueTrackerResult();
SrcIdx = OpIdx;
}
// Stop when any user of the bitcast is a SUBREG_TO_REG, replacing with a COPY
// will break the assumed guarantees for the upper bits.
for (const MachineInstr &UseMI : MRI.use_nodbg_instructions(DefOp.getReg())) {
if (UseMI.isSubregToReg())
return ValueTrackerResult();
}
[PeepholeOptimizer] Refactor the advanced copy optimization to take advantage of the isRegSequence property. This is a follow-up of r215394 and r215404, which respectively introduces the isRegSequence property and uses it for ARM. Thanks to the property introduced by the previous commits, this patch is able to optimize the following sequence: vmov d0, r2, r3 vmov d1, r0, r1 vmov r0, s0 vmov r1, s2 udiv r0, r1, r0 vmov r1, s1 vmov r2, s3 udiv r1, r2, r1 vmov.32 d16[0], r0 vmov.32 d16[1], r1 vmov r0, r1, d16 bx lr into: udiv r0, r0, r2 udiv r1, r1, r3 vmov.32 d16[0], r0 vmov.32 d16[1], r1 vmov r0, r1, d16 bx lr This patch refactors how the copy optimizations are done in the peephole optimizer. Prior to this patch, we had one copy-related optimization that replaced a copy or bitcast by a generic, more suitable (in terms of register file), copy. With this patch, the peephole optimizer features two copy-related optimizations: 1. One for rewriting generic copies to generic copies: PeepholeOptimizer::optimizeCoalescableCopy. 2. One for replacing non-generic copies with generic copies: PeepholeOptimizer::optimizeUncoalescableCopy. The goals of these two optimizations are slightly different: one rewrite the operand of the instruction (#1), the other kills off the non-generic instruction and replace it by a (sequence of) generic instruction(s). Both optimizations rely on the ValueTracker introduced in r212100. The ValueTracker has been refactored to use the information from the TargetInstrInfo for non-generic instruction. As part of the refactoring, we switched the tracking from the index of the definition to the actual register (virtual or physical). This one change is to provide better consistency with register related APIs and to ease the use of the TargetInstrInfo. Moreover, this patch introduces a new helper class CopyRewriter used to ease the rewriting of generic copies (i.e., #1). Finally, this patch adds a dead code elimination pass right after the peephole optimizer to get rid of dead code that may appear after rewriting. This is related to <rdar://problem/12702965>. Review: http://reviews.llvm.org/D4874 llvm-svn: 216088
2014-08-21 01:41:48 +08:00
const MachineOperand &Src = Def->getOperand(SrcIdx);
return ValueTrackerResult(Src.getReg(), Src.getSubReg());
}
ValueTrackerResult ValueTracker::getNextSourceFromRegSequence() {
[PeepholeOptimizer] Refactor the advanced copy optimization to take advantage of the isRegSequence property. This is a follow-up of r215394 and r215404, which respectively introduces the isRegSequence property and uses it for ARM. Thanks to the property introduced by the previous commits, this patch is able to optimize the following sequence: vmov d0, r2, r3 vmov d1, r0, r1 vmov r0, s0 vmov r1, s2 udiv r0, r1, r0 vmov r1, s1 vmov r2, s3 udiv r1, r2, r1 vmov.32 d16[0], r0 vmov.32 d16[1], r1 vmov r0, r1, d16 bx lr into: udiv r0, r0, r2 udiv r1, r1, r3 vmov.32 d16[0], r0 vmov.32 d16[1], r1 vmov r0, r1, d16 bx lr This patch refactors how the copy optimizations are done in the peephole optimizer. Prior to this patch, we had one copy-related optimization that replaced a copy or bitcast by a generic, more suitable (in terms of register file), copy. With this patch, the peephole optimizer features two copy-related optimizations: 1. One for rewriting generic copies to generic copies: PeepholeOptimizer::optimizeCoalescableCopy. 2. One for replacing non-generic copies with generic copies: PeepholeOptimizer::optimizeUncoalescableCopy. The goals of these two optimizations are slightly different: one rewrite the operand of the instruction (#1), the other kills off the non-generic instruction and replace it by a (sequence of) generic instruction(s). Both optimizations rely on the ValueTracker introduced in r212100. The ValueTracker has been refactored to use the information from the TargetInstrInfo for non-generic instruction. As part of the refactoring, we switched the tracking from the index of the definition to the actual register (virtual or physical). This one change is to provide better consistency with register related APIs and to ease the use of the TargetInstrInfo. Moreover, this patch introduces a new helper class CopyRewriter used to ease the rewriting of generic copies (i.e., #1). Finally, this patch adds a dead code elimination pass right after the peephole optimizer to get rid of dead code that may appear after rewriting. This is related to <rdar://problem/12702965>. Review: http://reviews.llvm.org/D4874 llvm-svn: 216088
2014-08-21 01:41:48 +08:00
assert((Def->isRegSequence() || Def->isRegSequenceLike()) &&
"Invalid definition");
if (Def->getOperand(DefIdx).getSubReg())
2015-09-09 08:38:33 +08:00
// If we are composing subregs, bail out.
// The case we are checking is Def.<subreg> = REG_SEQUENCE.
// This should almost never happen as the SSA property is tracked at
// the register level (as opposed to the subreg level).
// I.e.,
// Def.sub0 =
// Def.sub1 =
// is a valid SSA representation for Def.sub0 and Def.sub1, but not for
// Def. Thus, it must not be generated.
// However, some code could theoretically generates a single
// Def.sub0 (i.e, not defining the other subregs) and we would
// have this case.
// If we can ascertain (or force) that this never happens, we could
// turn that into an assertion.
return ValueTrackerResult();
[PeepholeOptimizer] Refactor the advanced copy optimization to take advantage of the isRegSequence property. This is a follow-up of r215394 and r215404, which respectively introduces the isRegSequence property and uses it for ARM. Thanks to the property introduced by the previous commits, this patch is able to optimize the following sequence: vmov d0, r2, r3 vmov d1, r0, r1 vmov r0, s0 vmov r1, s2 udiv r0, r1, r0 vmov r1, s1 vmov r2, s3 udiv r1, r2, r1 vmov.32 d16[0], r0 vmov.32 d16[1], r1 vmov r0, r1, d16 bx lr into: udiv r0, r0, r2 udiv r1, r1, r3 vmov.32 d16[0], r0 vmov.32 d16[1], r1 vmov r0, r1, d16 bx lr This patch refactors how the copy optimizations are done in the peephole optimizer. Prior to this patch, we had one copy-related optimization that replaced a copy or bitcast by a generic, more suitable (in terms of register file), copy. With this patch, the peephole optimizer features two copy-related optimizations: 1. One for rewriting generic copies to generic copies: PeepholeOptimizer::optimizeCoalescableCopy. 2. One for replacing non-generic copies with generic copies: PeepholeOptimizer::optimizeUncoalescableCopy. The goals of these two optimizations are slightly different: one rewrite the operand of the instruction (#1), the other kills off the non-generic instruction and replace it by a (sequence of) generic instruction(s). Both optimizations rely on the ValueTracker introduced in r212100. The ValueTracker has been refactored to use the information from the TargetInstrInfo for non-generic instruction. As part of the refactoring, we switched the tracking from the index of the definition to the actual register (virtual or physical). This one change is to provide better consistency with register related APIs and to ease the use of the TargetInstrInfo. Moreover, this patch introduces a new helper class CopyRewriter used to ease the rewriting of generic copies (i.e., #1). Finally, this patch adds a dead code elimination pass right after the peephole optimizer to get rid of dead code that may appear after rewriting. This is related to <rdar://problem/12702965>. Review: http://reviews.llvm.org/D4874 llvm-svn: 216088
2014-08-21 01:41:48 +08:00
if (!TII)
// We could handle the REG_SEQUENCE here, but we do not want to
// duplicate the code from the generic TII.
return ValueTrackerResult();
[PeepholeOptimizer] Refactor the advanced copy optimization to take advantage of the isRegSequence property. This is a follow-up of r215394 and r215404, which respectively introduces the isRegSequence property and uses it for ARM. Thanks to the property introduced by the previous commits, this patch is able to optimize the following sequence: vmov d0, r2, r3 vmov d1, r0, r1 vmov r0, s0 vmov r1, s2 udiv r0, r1, r0 vmov r1, s1 vmov r2, s3 udiv r1, r2, r1 vmov.32 d16[0], r0 vmov.32 d16[1], r1 vmov r0, r1, d16 bx lr into: udiv r0, r0, r2 udiv r1, r1, r3 vmov.32 d16[0], r0 vmov.32 d16[1], r1 vmov r0, r1, d16 bx lr This patch refactors how the copy optimizations are done in the peephole optimizer. Prior to this patch, we had one copy-related optimization that replaced a copy or bitcast by a generic, more suitable (in terms of register file), copy. With this patch, the peephole optimizer features two copy-related optimizations: 1. One for rewriting generic copies to generic copies: PeepholeOptimizer::optimizeCoalescableCopy. 2. One for replacing non-generic copies with generic copies: PeepholeOptimizer::optimizeUncoalescableCopy. The goals of these two optimizations are slightly different: one rewrite the operand of the instruction (#1), the other kills off the non-generic instruction and replace it by a (sequence of) generic instruction(s). Both optimizations rely on the ValueTracker introduced in r212100. The ValueTracker has been refactored to use the information from the TargetInstrInfo for non-generic instruction. As part of the refactoring, we switched the tracking from the index of the definition to the actual register (virtual or physical). This one change is to provide better consistency with register related APIs and to ease the use of the TargetInstrInfo. Moreover, this patch introduces a new helper class CopyRewriter used to ease the rewriting of generic copies (i.e., #1). Finally, this patch adds a dead code elimination pass right after the peephole optimizer to get rid of dead code that may appear after rewriting. This is related to <rdar://problem/12702965>. Review: http://reviews.llvm.org/D4874 llvm-svn: 216088
2014-08-21 01:41:48 +08:00
SmallVector<TargetInstrInfo::RegSubRegPairAndIdx, 8> RegSeqInputRegs;
if (!TII->getRegSequenceInputs(*Def, DefIdx, RegSeqInputRegs))
return ValueTrackerResult();
[PeepholeOptimizer] Refactor the advanced copy optimization to take advantage of the isRegSequence property. This is a follow-up of r215394 and r215404, which respectively introduces the isRegSequence property and uses it for ARM. Thanks to the property introduced by the previous commits, this patch is able to optimize the following sequence: vmov d0, r2, r3 vmov d1, r0, r1 vmov r0, s0 vmov r1, s2 udiv r0, r1, r0 vmov r1, s1 vmov r2, s3 udiv r1, r2, r1 vmov.32 d16[0], r0 vmov.32 d16[1], r1 vmov r0, r1, d16 bx lr into: udiv r0, r0, r2 udiv r1, r1, r3 vmov.32 d16[0], r0 vmov.32 d16[1], r1 vmov r0, r1, d16 bx lr This patch refactors how the copy optimizations are done in the peephole optimizer. Prior to this patch, we had one copy-related optimization that replaced a copy or bitcast by a generic, more suitable (in terms of register file), copy. With this patch, the peephole optimizer features two copy-related optimizations: 1. One for rewriting generic copies to generic copies: PeepholeOptimizer::optimizeCoalescableCopy. 2. One for replacing non-generic copies with generic copies: PeepholeOptimizer::optimizeUncoalescableCopy. The goals of these two optimizations are slightly different: one rewrite the operand of the instruction (#1), the other kills off the non-generic instruction and replace it by a (sequence of) generic instruction(s). Both optimizations rely on the ValueTracker introduced in r212100. The ValueTracker has been refactored to use the information from the TargetInstrInfo for non-generic instruction. As part of the refactoring, we switched the tracking from the index of the definition to the actual register (virtual or physical). This one change is to provide better consistency with register related APIs and to ease the use of the TargetInstrInfo. Moreover, this patch introduces a new helper class CopyRewriter used to ease the rewriting of generic copies (i.e., #1). Finally, this patch adds a dead code elimination pass right after the peephole optimizer to get rid of dead code that may appear after rewriting. This is related to <rdar://problem/12702965>. Review: http://reviews.llvm.org/D4874 llvm-svn: 216088
2014-08-21 01:41:48 +08:00
// We are looking at:
// Def = REG_SEQUENCE v0, sub0, v1, sub1, ...
// Check if one of the operand defines the subreg we are interested in.
[PeepholeOptimizer] Refactor the advanced copy optimization to take advantage of the isRegSequence property. This is a follow-up of r215394 and r215404, which respectively introduces the isRegSequence property and uses it for ARM. Thanks to the property introduced by the previous commits, this patch is able to optimize the following sequence: vmov d0, r2, r3 vmov d1, r0, r1 vmov r0, s0 vmov r1, s2 udiv r0, r1, r0 vmov r1, s1 vmov r2, s3 udiv r1, r2, r1 vmov.32 d16[0], r0 vmov.32 d16[1], r1 vmov r0, r1, d16 bx lr into: udiv r0, r0, r2 udiv r1, r1, r3 vmov.32 d16[0], r0 vmov.32 d16[1], r1 vmov r0, r1, d16 bx lr This patch refactors how the copy optimizations are done in the peephole optimizer. Prior to this patch, we had one copy-related optimization that replaced a copy or bitcast by a generic, more suitable (in terms of register file), copy. With this patch, the peephole optimizer features two copy-related optimizations: 1. One for rewriting generic copies to generic copies: PeepholeOptimizer::optimizeCoalescableCopy. 2. One for replacing non-generic copies with generic copies: PeepholeOptimizer::optimizeUncoalescableCopy. The goals of these two optimizations are slightly different: one rewrite the operand of the instruction (#1), the other kills off the non-generic instruction and replace it by a (sequence of) generic instruction(s). Both optimizations rely on the ValueTracker introduced in r212100. The ValueTracker has been refactored to use the information from the TargetInstrInfo for non-generic instruction. As part of the refactoring, we switched the tracking from the index of the definition to the actual register (virtual or physical). This one change is to provide better consistency with register related APIs and to ease the use of the TargetInstrInfo. Moreover, this patch introduces a new helper class CopyRewriter used to ease the rewriting of generic copies (i.e., #1). Finally, this patch adds a dead code elimination pass right after the peephole optimizer to get rid of dead code that may appear after rewriting. This is related to <rdar://problem/12702965>. Review: http://reviews.llvm.org/D4874 llvm-svn: 216088
2014-08-21 01:41:48 +08:00
for (auto &RegSeqInput : RegSeqInputRegs) {
if (RegSeqInput.SubIdx == DefSubReg) {
if (RegSeqInput.SubReg)
2015-09-09 08:38:33 +08:00
// Bail if we have to compose sub registers.
return ValueTrackerResult();
[PeepholeOptimizer] Refactor the advanced copy optimization to take advantage of the isRegSequence property. This is a follow-up of r215394 and r215404, which respectively introduces the isRegSequence property and uses it for ARM. Thanks to the property introduced by the previous commits, this patch is able to optimize the following sequence: vmov d0, r2, r3 vmov d1, r0, r1 vmov r0, s0 vmov r1, s2 udiv r0, r1, r0 vmov r1, s1 vmov r2, s3 udiv r1, r2, r1 vmov.32 d16[0], r0 vmov.32 d16[1], r1 vmov r0, r1, d16 bx lr into: udiv r0, r0, r2 udiv r1, r1, r3 vmov.32 d16[0], r0 vmov.32 d16[1], r1 vmov r0, r1, d16 bx lr This patch refactors how the copy optimizations are done in the peephole optimizer. Prior to this patch, we had one copy-related optimization that replaced a copy or bitcast by a generic, more suitable (in terms of register file), copy. With this patch, the peephole optimizer features two copy-related optimizations: 1. One for rewriting generic copies to generic copies: PeepholeOptimizer::optimizeCoalescableCopy. 2. One for replacing non-generic copies with generic copies: PeepholeOptimizer::optimizeUncoalescableCopy. The goals of these two optimizations are slightly different: one rewrite the operand of the instruction (#1), the other kills off the non-generic instruction and replace it by a (sequence of) generic instruction(s). Both optimizations rely on the ValueTracker introduced in r212100. The ValueTracker has been refactored to use the information from the TargetInstrInfo for non-generic instruction. As part of the refactoring, we switched the tracking from the index of the definition to the actual register (virtual or physical). This one change is to provide better consistency with register related APIs and to ease the use of the TargetInstrInfo. Moreover, this patch introduces a new helper class CopyRewriter used to ease the rewriting of generic copies (i.e., #1). Finally, this patch adds a dead code elimination pass right after the peephole optimizer to get rid of dead code that may appear after rewriting. This is related to <rdar://problem/12702965>. Review: http://reviews.llvm.org/D4874 llvm-svn: 216088
2014-08-21 01:41:48 +08:00
return ValueTrackerResult(RegSeqInput.Reg, RegSeqInput.SubReg);
}
}
// If the subreg we are tracking is super-defined by another subreg,
// we could follow this value. However, this would require to compose
// the subreg and we do not do that for now.
return ValueTrackerResult();
}
ValueTrackerResult ValueTracker::getNextSourceFromInsertSubreg() {
assert((Def->isInsertSubreg() || Def->isInsertSubregLike()) &&
"Invalid definition");
if (Def->getOperand(DefIdx).getSubReg())
2015-09-09 08:38:33 +08:00
// If we are composing subreg, bail out.
// Same remark as getNextSourceFromRegSequence.
// I.e., this may be turned into an assert.
return ValueTrackerResult();
if (!TII)
// We could handle the REG_SEQUENCE here, but we do not want to
// duplicate the code from the generic TII.
return ValueTrackerResult();
[PeepholeOptimizer] Refactor the advanced copy optimization to take advantage of the isRegSequence property. This is a follow-up of r215394 and r215404, which respectively introduces the isRegSequence property and uses it for ARM. Thanks to the property introduced by the previous commits, this patch is able to optimize the following sequence: vmov d0, r2, r3 vmov d1, r0, r1 vmov r0, s0 vmov r1, s2 udiv r0, r1, r0 vmov r1, s1 vmov r2, s3 udiv r1, r2, r1 vmov.32 d16[0], r0 vmov.32 d16[1], r1 vmov r0, r1, d16 bx lr into: udiv r0, r0, r2 udiv r1, r1, r3 vmov.32 d16[0], r0 vmov.32 d16[1], r1 vmov r0, r1, d16 bx lr This patch refactors how the copy optimizations are done in the peephole optimizer. Prior to this patch, we had one copy-related optimization that replaced a copy or bitcast by a generic, more suitable (in terms of register file), copy. With this patch, the peephole optimizer features two copy-related optimizations: 1. One for rewriting generic copies to generic copies: PeepholeOptimizer::optimizeCoalescableCopy. 2. One for replacing non-generic copies with generic copies: PeepholeOptimizer::optimizeUncoalescableCopy. The goals of these two optimizations are slightly different: one rewrite the operand of the instruction (#1), the other kills off the non-generic instruction and replace it by a (sequence of) generic instruction(s). Both optimizations rely on the ValueTracker introduced in r212100. The ValueTracker has been refactored to use the information from the TargetInstrInfo for non-generic instruction. As part of the refactoring, we switched the tracking from the index of the definition to the actual register (virtual or physical). This one change is to provide better consistency with register related APIs and to ease the use of the TargetInstrInfo. Moreover, this patch introduces a new helper class CopyRewriter used to ease the rewriting of generic copies (i.e., #1). Finally, this patch adds a dead code elimination pass right after the peephole optimizer to get rid of dead code that may appear after rewriting. This is related to <rdar://problem/12702965>. Review: http://reviews.llvm.org/D4874 llvm-svn: 216088
2014-08-21 01:41:48 +08:00
TargetInstrInfo::RegSubRegPair BaseReg;
TargetInstrInfo::RegSubRegPairAndIdx InsertedReg;
if (!TII->getInsertSubregInputs(*Def, DefIdx, BaseReg, InsertedReg))
return ValueTrackerResult();
[PeepholeOptimizer] Refactor the advanced copy optimization to take advantage of the isRegSequence property. This is a follow-up of r215394 and r215404, which respectively introduces the isRegSequence property and uses it for ARM. Thanks to the property introduced by the previous commits, this patch is able to optimize the following sequence: vmov d0, r2, r3 vmov d1, r0, r1 vmov r0, s0 vmov r1, s2 udiv r0, r1, r0 vmov r1, s1 vmov r2, s3 udiv r1, r2, r1 vmov.32 d16[0], r0 vmov.32 d16[1], r1 vmov r0, r1, d16 bx lr into: udiv r0, r0, r2 udiv r1, r1, r3 vmov.32 d16[0], r0 vmov.32 d16[1], r1 vmov r0, r1, d16 bx lr This patch refactors how the copy optimizations are done in the peephole optimizer. Prior to this patch, we had one copy-related optimization that replaced a copy or bitcast by a generic, more suitable (in terms of register file), copy. With this patch, the peephole optimizer features two copy-related optimizations: 1. One for rewriting generic copies to generic copies: PeepholeOptimizer::optimizeCoalescableCopy. 2. One for replacing non-generic copies with generic copies: PeepholeOptimizer::optimizeUncoalescableCopy. The goals of these two optimizations are slightly different: one rewrite the operand of the instruction (#1), the other kills off the non-generic instruction and replace it by a (sequence of) generic instruction(s). Both optimizations rely on the ValueTracker introduced in r212100. The ValueTracker has been refactored to use the information from the TargetInstrInfo for non-generic instruction. As part of the refactoring, we switched the tracking from the index of the definition to the actual register (virtual or physical). This one change is to provide better consistency with register related APIs and to ease the use of the TargetInstrInfo. Moreover, this patch introduces a new helper class CopyRewriter used to ease the rewriting of generic copies (i.e., #1). Finally, this patch adds a dead code elimination pass right after the peephole optimizer to get rid of dead code that may appear after rewriting. This is related to <rdar://problem/12702965>. Review: http://reviews.llvm.org/D4874 llvm-svn: 216088
2014-08-21 01:41:48 +08:00
// We are looking at:
// Def = INSERT_SUBREG v0, v1, sub1
// There are two cases:
// 1. DefSubReg == sub1, get v1.
// 2. DefSubReg != sub1, the value may be available through v0.
[PeepholeOptimizer] Refactor the advanced copy optimization to take advantage of the isRegSequence property. This is a follow-up of r215394 and r215404, which respectively introduces the isRegSequence property and uses it for ARM. Thanks to the property introduced by the previous commits, this patch is able to optimize the following sequence: vmov d0, r2, r3 vmov d1, r0, r1 vmov r0, s0 vmov r1, s2 udiv r0, r1, r0 vmov r1, s1 vmov r2, s3 udiv r1, r2, r1 vmov.32 d16[0], r0 vmov.32 d16[1], r1 vmov r0, r1, d16 bx lr into: udiv r0, r0, r2 udiv r1, r1, r3 vmov.32 d16[0], r0 vmov.32 d16[1], r1 vmov r0, r1, d16 bx lr This patch refactors how the copy optimizations are done in the peephole optimizer. Prior to this patch, we had one copy-related optimization that replaced a copy or bitcast by a generic, more suitable (in terms of register file), copy. With this patch, the peephole optimizer features two copy-related optimizations: 1. One for rewriting generic copies to generic copies: PeepholeOptimizer::optimizeCoalescableCopy. 2. One for replacing non-generic copies with generic copies: PeepholeOptimizer::optimizeUncoalescableCopy. The goals of these two optimizations are slightly different: one rewrite the operand of the instruction (#1), the other kills off the non-generic instruction and replace it by a (sequence of) generic instruction(s). Both optimizations rely on the ValueTracker introduced in r212100. The ValueTracker has been refactored to use the information from the TargetInstrInfo for non-generic instruction. As part of the refactoring, we switched the tracking from the index of the definition to the actual register (virtual or physical). This one change is to provide better consistency with register related APIs and to ease the use of the TargetInstrInfo. Moreover, this patch introduces a new helper class CopyRewriter used to ease the rewriting of generic copies (i.e., #1). Finally, this patch adds a dead code elimination pass right after the peephole optimizer to get rid of dead code that may appear after rewriting. This is related to <rdar://problem/12702965>. Review: http://reviews.llvm.org/D4874 llvm-svn: 216088
2014-08-21 01:41:48 +08:00
// #1 Check if the inserted register matches the required sub index.
if (InsertedReg.SubIdx == DefSubReg) {
return ValueTrackerResult(InsertedReg.Reg, InsertedReg.SubReg);
}
// #2 Otherwise, if the sub register we are looking for is not partial
// defined by the inserted element, we can look through the main
// register (v0).
const MachineOperand &MODef = Def->getOperand(DefIdx);
// If the result register (Def) and the base register (v0) do not
// have the same register class or if we have to compose
2015-09-09 08:38:33 +08:00
// subregisters, bail out.
[PeepholeOptimizer] Refactor the advanced copy optimization to take advantage of the isRegSequence property. This is a follow-up of r215394 and r215404, which respectively introduces the isRegSequence property and uses it for ARM. Thanks to the property introduced by the previous commits, this patch is able to optimize the following sequence: vmov d0, r2, r3 vmov d1, r0, r1 vmov r0, s0 vmov r1, s2 udiv r0, r1, r0 vmov r1, s1 vmov r2, s3 udiv r1, r2, r1 vmov.32 d16[0], r0 vmov.32 d16[1], r1 vmov r0, r1, d16 bx lr into: udiv r0, r0, r2 udiv r1, r1, r3 vmov.32 d16[0], r0 vmov.32 d16[1], r1 vmov r0, r1, d16 bx lr This patch refactors how the copy optimizations are done in the peephole optimizer. Prior to this patch, we had one copy-related optimization that replaced a copy or bitcast by a generic, more suitable (in terms of register file), copy. With this patch, the peephole optimizer features two copy-related optimizations: 1. One for rewriting generic copies to generic copies: PeepholeOptimizer::optimizeCoalescableCopy. 2. One for replacing non-generic copies with generic copies: PeepholeOptimizer::optimizeUncoalescableCopy. The goals of these two optimizations are slightly different: one rewrite the operand of the instruction (#1), the other kills off the non-generic instruction and replace it by a (sequence of) generic instruction(s). Both optimizations rely on the ValueTracker introduced in r212100. The ValueTracker has been refactored to use the information from the TargetInstrInfo for non-generic instruction. As part of the refactoring, we switched the tracking from the index of the definition to the actual register (virtual or physical). This one change is to provide better consistency with register related APIs and to ease the use of the TargetInstrInfo. Moreover, this patch introduces a new helper class CopyRewriter used to ease the rewriting of generic copies (i.e., #1). Finally, this patch adds a dead code elimination pass right after the peephole optimizer to get rid of dead code that may appear after rewriting. This is related to <rdar://problem/12702965>. Review: http://reviews.llvm.org/D4874 llvm-svn: 216088
2014-08-21 01:41:48 +08:00
if (MRI.getRegClass(MODef.getReg()) != MRI.getRegClass(BaseReg.Reg) ||
BaseReg.SubReg)
return ValueTrackerResult();
[PeepholeOptimizer] Refactor the advanced copy optimization to take advantage of the isRegSequence property. This is a follow-up of r215394 and r215404, which respectively introduces the isRegSequence property and uses it for ARM. Thanks to the property introduced by the previous commits, this patch is able to optimize the following sequence: vmov d0, r2, r3 vmov d1, r0, r1 vmov r0, s0 vmov r1, s2 udiv r0, r1, r0 vmov r1, s1 vmov r2, s3 udiv r1, r2, r1 vmov.32 d16[0], r0 vmov.32 d16[1], r1 vmov r0, r1, d16 bx lr into: udiv r0, r0, r2 udiv r1, r1, r3 vmov.32 d16[0], r0 vmov.32 d16[1], r1 vmov r0, r1, d16 bx lr This patch refactors how the copy optimizations are done in the peephole optimizer. Prior to this patch, we had one copy-related optimization that replaced a copy or bitcast by a generic, more suitable (in terms of register file), copy. With this patch, the peephole optimizer features two copy-related optimizations: 1. One for rewriting generic copies to generic copies: PeepholeOptimizer::optimizeCoalescableCopy. 2. One for replacing non-generic copies with generic copies: PeepholeOptimizer::optimizeUncoalescableCopy. The goals of these two optimizations are slightly different: one rewrite the operand of the instruction (#1), the other kills off the non-generic instruction and replace it by a (sequence of) generic instruction(s). Both optimizations rely on the ValueTracker introduced in r212100. The ValueTracker has been refactored to use the information from the TargetInstrInfo for non-generic instruction. As part of the refactoring, we switched the tracking from the index of the definition to the actual register (virtual or physical). This one change is to provide better consistency with register related APIs and to ease the use of the TargetInstrInfo. Moreover, this patch introduces a new helper class CopyRewriter used to ease the rewriting of generic copies (i.e., #1). Finally, this patch adds a dead code elimination pass right after the peephole optimizer to get rid of dead code that may appear after rewriting. This is related to <rdar://problem/12702965>. Review: http://reviews.llvm.org/D4874 llvm-svn: 216088
2014-08-21 01:41:48 +08:00
// Get the TRI and check if the inserted sub-register overlaps with the
// sub-register we are tracking.
const TargetRegisterInfo *TRI = MRI.getTargetRegisterInfo();
if (!TRI ||
!(TRI->getSubRegIndexLaneMask(DefSubReg) &
TRI->getSubRegIndexLaneMask(InsertedReg.SubIdx)).none())
return ValueTrackerResult();
// At this point, the value is available in v0 via the same subreg
// we used for Def.
return ValueTrackerResult(BaseReg.Reg, DefSubReg);
}
ValueTrackerResult ValueTracker::getNextSourceFromExtractSubreg() {
assert((Def->isExtractSubreg() ||
Def->isExtractSubregLike()) && "Invalid definition");
// We are looking at:
// Def = EXTRACT_SUBREG v0, sub0
2015-09-09 08:38:33 +08:00
// Bail if we have to compose sub registers.
// Indeed, if DefSubReg != 0, we would have to compose it with sub0.
if (DefSubReg)
return ValueTrackerResult();
if (!TII)
// We could handle the EXTRACT_SUBREG here, but we do not want to
// duplicate the code from the generic TII.
return ValueTrackerResult();
[PeepholeOptimizer] Refactor the advanced copy optimization to take advantage of the isRegSequence property. This is a follow-up of r215394 and r215404, which respectively introduces the isRegSequence property and uses it for ARM. Thanks to the property introduced by the previous commits, this patch is able to optimize the following sequence: vmov d0, r2, r3 vmov d1, r0, r1 vmov r0, s0 vmov r1, s2 udiv r0, r1, r0 vmov r1, s1 vmov r2, s3 udiv r1, r2, r1 vmov.32 d16[0], r0 vmov.32 d16[1], r1 vmov r0, r1, d16 bx lr into: udiv r0, r0, r2 udiv r1, r1, r3 vmov.32 d16[0], r0 vmov.32 d16[1], r1 vmov r0, r1, d16 bx lr This patch refactors how the copy optimizations are done in the peephole optimizer. Prior to this patch, we had one copy-related optimization that replaced a copy or bitcast by a generic, more suitable (in terms of register file), copy. With this patch, the peephole optimizer features two copy-related optimizations: 1. One for rewriting generic copies to generic copies: PeepholeOptimizer::optimizeCoalescableCopy. 2. One for replacing non-generic copies with generic copies: PeepholeOptimizer::optimizeUncoalescableCopy. The goals of these two optimizations are slightly different: one rewrite the operand of the instruction (#1), the other kills off the non-generic instruction and replace it by a (sequence of) generic instruction(s). Both optimizations rely on the ValueTracker introduced in r212100. The ValueTracker has been refactored to use the information from the TargetInstrInfo for non-generic instruction. As part of the refactoring, we switched the tracking from the index of the definition to the actual register (virtual or physical). This one change is to provide better consistency with register related APIs and to ease the use of the TargetInstrInfo. Moreover, this patch introduces a new helper class CopyRewriter used to ease the rewriting of generic copies (i.e., #1). Finally, this patch adds a dead code elimination pass right after the peephole optimizer to get rid of dead code that may appear after rewriting. This is related to <rdar://problem/12702965>. Review: http://reviews.llvm.org/D4874 llvm-svn: 216088
2014-08-21 01:41:48 +08:00
TargetInstrInfo::RegSubRegPairAndIdx ExtractSubregInputReg;
if (!TII->getExtractSubregInputs(*Def, DefIdx, ExtractSubregInputReg))
return ValueTrackerResult();
[PeepholeOptimizer] Refactor the advanced copy optimization to take advantage of the isRegSequence property. This is a follow-up of r215394 and r215404, which respectively introduces the isRegSequence property and uses it for ARM. Thanks to the property introduced by the previous commits, this patch is able to optimize the following sequence: vmov d0, r2, r3 vmov d1, r0, r1 vmov r0, s0 vmov r1, s2 udiv r0, r1, r0 vmov r1, s1 vmov r2, s3 udiv r1, r2, r1 vmov.32 d16[0], r0 vmov.32 d16[1], r1 vmov r0, r1, d16 bx lr into: udiv r0, r0, r2 udiv r1, r1, r3 vmov.32 d16[0], r0 vmov.32 d16[1], r1 vmov r0, r1, d16 bx lr This patch refactors how the copy optimizations are done in the peephole optimizer. Prior to this patch, we had one copy-related optimization that replaced a copy or bitcast by a generic, more suitable (in terms of register file), copy. With this patch, the peephole optimizer features two copy-related optimizations: 1. One for rewriting generic copies to generic copies: PeepholeOptimizer::optimizeCoalescableCopy. 2. One for replacing non-generic copies with generic copies: PeepholeOptimizer::optimizeUncoalescableCopy. The goals of these two optimizations are slightly different: one rewrite the operand of the instruction (#1), the other kills off the non-generic instruction and replace it by a (sequence of) generic instruction(s). Both optimizations rely on the ValueTracker introduced in r212100. The ValueTracker has been refactored to use the information from the TargetInstrInfo for non-generic instruction. As part of the refactoring, we switched the tracking from the index of the definition to the actual register (virtual or physical). This one change is to provide better consistency with register related APIs and to ease the use of the TargetInstrInfo. Moreover, this patch introduces a new helper class CopyRewriter used to ease the rewriting of generic copies (i.e., #1). Finally, this patch adds a dead code elimination pass right after the peephole optimizer to get rid of dead code that may appear after rewriting. This is related to <rdar://problem/12702965>. Review: http://reviews.llvm.org/D4874 llvm-svn: 216088
2014-08-21 01:41:48 +08:00
2015-09-09 08:38:33 +08:00
// Bail if we have to compose sub registers.
// Likewise, if v0.subreg != 0, we would have to compose v0.subreg with sub0.
[PeepholeOptimizer] Refactor the advanced copy optimization to take advantage of the isRegSequence property. This is a follow-up of r215394 and r215404, which respectively introduces the isRegSequence property and uses it for ARM. Thanks to the property introduced by the previous commits, this patch is able to optimize the following sequence: vmov d0, r2, r3 vmov d1, r0, r1 vmov r0, s0 vmov r1, s2 udiv r0, r1, r0 vmov r1, s1 vmov r2, s3 udiv r1, r2, r1 vmov.32 d16[0], r0 vmov.32 d16[1], r1 vmov r0, r1, d16 bx lr into: udiv r0, r0, r2 udiv r1, r1, r3 vmov.32 d16[0], r0 vmov.32 d16[1], r1 vmov r0, r1, d16 bx lr This patch refactors how the copy optimizations are done in the peephole optimizer. Prior to this patch, we had one copy-related optimization that replaced a copy or bitcast by a generic, more suitable (in terms of register file), copy. With this patch, the peephole optimizer features two copy-related optimizations: 1. One for rewriting generic copies to generic copies: PeepholeOptimizer::optimizeCoalescableCopy. 2. One for replacing non-generic copies with generic copies: PeepholeOptimizer::optimizeUncoalescableCopy. The goals of these two optimizations are slightly different: one rewrite the operand of the instruction (#1), the other kills off the non-generic instruction and replace it by a (sequence of) generic instruction(s). Both optimizations rely on the ValueTracker introduced in r212100. The ValueTracker has been refactored to use the information from the TargetInstrInfo for non-generic instruction. As part of the refactoring, we switched the tracking from the index of the definition to the actual register (virtual or physical). This one change is to provide better consistency with register related APIs and to ease the use of the TargetInstrInfo. Moreover, this patch introduces a new helper class CopyRewriter used to ease the rewriting of generic copies (i.e., #1). Finally, this patch adds a dead code elimination pass right after the peephole optimizer to get rid of dead code that may appear after rewriting. This is related to <rdar://problem/12702965>. Review: http://reviews.llvm.org/D4874 llvm-svn: 216088
2014-08-21 01:41:48 +08:00
if (ExtractSubregInputReg.SubReg)
return ValueTrackerResult();
// Otherwise, the value is available in the v0.sub0.
2015-12-30 03:34:53 +08:00
return ValueTrackerResult(ExtractSubregInputReg.Reg,
ExtractSubregInputReg.SubIdx);
}
ValueTrackerResult ValueTracker::getNextSourceFromSubregToReg() {
assert(Def->isSubregToReg() && "Invalid definition");
// We are looking at:
// Def = SUBREG_TO_REG Imm, v0, sub0
2015-09-09 08:38:33 +08:00
// Bail if we have to compose sub registers.
// If DefSubReg != sub0, we would have to check that all the bits
// we track are included in sub0 and if yes, we would have to
// determine the right subreg in v0.
if (DefSubReg != Def->getOperand(3).getImm())
return ValueTrackerResult();
2015-09-09 08:38:33 +08:00
// Bail if we have to compose sub registers.
// Likewise, if v0.subreg != 0, we would have to compose it with sub0.
if (Def->getOperand(2).getSubReg())
return ValueTrackerResult();
return ValueTrackerResult(Def->getOperand(2).getReg(),
Def->getOperand(3).getImm());
}
/// \brief Explore each PHI incoming operand and return its sources
ValueTrackerResult ValueTracker::getNextSourceFromPHI() {
assert(Def->isPHI() && "Invalid definition");
ValueTrackerResult Res;
2015-09-09 08:38:33 +08:00
// If we look for a different subreg, bail as we do not support composing
// subregs yet.
if (Def->getOperand(0).getSubReg() != DefSubReg)
return ValueTrackerResult();
// Return all register sources for PHI instructions.
for (unsigned i = 1, e = Def->getNumOperands(); i < e; i += 2) {
auto &MO = Def->getOperand(i);
assert(MO.isReg() && "Invalid PHI instruction");
Res.addSource(MO.getReg(), MO.getSubReg());
}
return Res;
}
ValueTrackerResult ValueTracker::getNextSourceImpl() {
assert(Def && "This method needs a valid definition");
assert(((Def->getOperand(DefIdx).isDef() &&
(DefIdx < Def->getDesc().getNumDefs() ||
Def->getDesc().isVariadic())) ||
Def->getOperand(DefIdx).isImplicit()) &&
"Invalid DefIdx");
if (Def->isCopy())
return getNextSourceFromCopy();
if (Def->isBitcast())
return getNextSourceFromBitcast();
// All the remaining cases involve "complex" instructions.
2015-09-09 08:38:33 +08:00
// Bail if we did not ask for the advanced tracking.
if (!UseAdvancedTracking)
return ValueTrackerResult();
[PeepholeOptimizer] Refactor the advanced copy optimization to take advantage of the isRegSequence property. This is a follow-up of r215394 and r215404, which respectively introduces the isRegSequence property and uses it for ARM. Thanks to the property introduced by the previous commits, this patch is able to optimize the following sequence: vmov d0, r2, r3 vmov d1, r0, r1 vmov r0, s0 vmov r1, s2 udiv r0, r1, r0 vmov r1, s1 vmov r2, s3 udiv r1, r2, r1 vmov.32 d16[0], r0 vmov.32 d16[1], r1 vmov r0, r1, d16 bx lr into: udiv r0, r0, r2 udiv r1, r1, r3 vmov.32 d16[0], r0 vmov.32 d16[1], r1 vmov r0, r1, d16 bx lr This patch refactors how the copy optimizations are done in the peephole optimizer. Prior to this patch, we had one copy-related optimization that replaced a copy or bitcast by a generic, more suitable (in terms of register file), copy. With this patch, the peephole optimizer features two copy-related optimizations: 1. One for rewriting generic copies to generic copies: PeepholeOptimizer::optimizeCoalescableCopy. 2. One for replacing non-generic copies with generic copies: PeepholeOptimizer::optimizeUncoalescableCopy. The goals of these two optimizations are slightly different: one rewrite the operand of the instruction (#1), the other kills off the non-generic instruction and replace it by a (sequence of) generic instruction(s). Both optimizations rely on the ValueTracker introduced in r212100. The ValueTracker has been refactored to use the information from the TargetInstrInfo for non-generic instruction. As part of the refactoring, we switched the tracking from the index of the definition to the actual register (virtual or physical). This one change is to provide better consistency with register related APIs and to ease the use of the TargetInstrInfo. Moreover, this patch introduces a new helper class CopyRewriter used to ease the rewriting of generic copies (i.e., #1). Finally, this patch adds a dead code elimination pass right after the peephole optimizer to get rid of dead code that may appear after rewriting. This is related to <rdar://problem/12702965>. Review: http://reviews.llvm.org/D4874 llvm-svn: 216088
2014-08-21 01:41:48 +08:00
if (Def->isRegSequence() || Def->isRegSequenceLike())
return getNextSourceFromRegSequence();
if (Def->isInsertSubreg() || Def->isInsertSubregLike())
return getNextSourceFromInsertSubreg();
if (Def->isExtractSubreg() || Def->isExtractSubregLike())
return getNextSourceFromExtractSubreg();
if (Def->isSubregToReg())
return getNextSourceFromSubregToReg();
if (Def->isPHI())
return getNextSourceFromPHI();
return ValueTrackerResult();
}
ValueTrackerResult ValueTracker::getNextSource() {
// If we reach a point where we cannot move up in the use-def chain,
// there is nothing we can get.
if (!Def)
return ValueTrackerResult();
ValueTrackerResult Res = getNextSourceImpl();
if (Res.isValid()) {
// Update definition, definition index, and subregister for the
// next call of getNextSource.
// Update the current register.
bool OneRegSrc = Res.getNumSources() == 1;
if (OneRegSrc)
Reg = Res.getSrcReg(0);
// Update the result before moving up in the use-def chain
// with the instruction containing the last found sources.
Res.setInst(Def);
// If we can still move up in the use-def chain, move to the next
// definition.
if (!TargetRegisterInfo::isPhysicalRegister(Reg) && OneRegSrc) {
[PeepholeOptimizer] Refactor the advanced copy optimization to take advantage of the isRegSequence property. This is a follow-up of r215394 and r215404, which respectively introduces the isRegSequence property and uses it for ARM. Thanks to the property introduced by the previous commits, this patch is able to optimize the following sequence: vmov d0, r2, r3 vmov d1, r0, r1 vmov r0, s0 vmov r1, s2 udiv r0, r1, r0 vmov r1, s1 vmov r2, s3 udiv r1, r2, r1 vmov.32 d16[0], r0 vmov.32 d16[1], r1 vmov r0, r1, d16 bx lr into: udiv r0, r0, r2 udiv r1, r1, r3 vmov.32 d16[0], r0 vmov.32 d16[1], r1 vmov r0, r1, d16 bx lr This patch refactors how the copy optimizations are done in the peephole optimizer. Prior to this patch, we had one copy-related optimization that replaced a copy or bitcast by a generic, more suitable (in terms of register file), copy. With this patch, the peephole optimizer features two copy-related optimizations: 1. One for rewriting generic copies to generic copies: PeepholeOptimizer::optimizeCoalescableCopy. 2. One for replacing non-generic copies with generic copies: PeepholeOptimizer::optimizeUncoalescableCopy. The goals of these two optimizations are slightly different: one rewrite the operand of the instruction (#1), the other kills off the non-generic instruction and replace it by a (sequence of) generic instruction(s). Both optimizations rely on the ValueTracker introduced in r212100. The ValueTracker has been refactored to use the information from the TargetInstrInfo for non-generic instruction. As part of the refactoring, we switched the tracking from the index of the definition to the actual register (virtual or physical). This one change is to provide better consistency with register related APIs and to ease the use of the TargetInstrInfo. Moreover, this patch introduces a new helper class CopyRewriter used to ease the rewriting of generic copies (i.e., #1). Finally, this patch adds a dead code elimination pass right after the peephole optimizer to get rid of dead code that may appear after rewriting. This is related to <rdar://problem/12702965>. Review: http://reviews.llvm.org/D4874 llvm-svn: 216088
2014-08-21 01:41:48 +08:00
Def = MRI.getVRegDef(Reg);
DefIdx = MRI.def_begin(Reg).getOperandNo();
DefSubReg = Res.getSrcSubReg(0);
return Res;
}
}
// If we end up here, this means we will not be able to find another source
// for the next iteration. Make sure any new call to getNextSource bails out
// early by cutting the use-def chain.
Def = nullptr;
return Res;
}