[llvm] d84ca05 - Enhance peephole optimization.
Mingming Liu via llvm-commits
llvm-commits at lists.llvm.org
Tue May 10 13:27:50 PDT 2022
Hello Philip!
A differential revision is attached in the commit message, and has more
context of the commit.
Shall I just use differential revision or still revert and re-commit?
On Tue, May 10, 2022 at 1:17 PM Philip Reames <listmail at philipreames.com>
wrote:
> Please revert and reapply with a descriptive commit message.
>
> Philip
>
> On 5/10/22 12:39, Mingming Liu via llvm-commits wrote:
> > Author: Mingming Liu
> > Date: 2022-05-10T12:35:35-07:00
> > New Revision: d84ca05ef7f897fdd51900ea07e3c5344632130a
> >
> > URL:
> https://github.com/llvm/llvm-project/commit/d84ca05ef7f897fdd51900ea07e3c5344632130a
> > DIFF:
> https://github.com/llvm/llvm-project/commit/d84ca05ef7f897fdd51900ea07e3c5344632130a.diff
> >
> > LOG: Enhance peephole optimization.
> >
> > Differential Revision: https://reviews.llvm.org/D124118
> >
> > Added:
> > llvm/test/CodeGen/X86/peephole-test-after-add.mir
> >
> > Modified:
> > llvm/lib/Target/X86/X86InstrInfo.cpp
> >
> > Removed:
> >
> >
> >
> >
> ################################################################################
> > diff --git a/llvm/lib/Target/X86/X86InstrInfo.cpp
> b/llvm/lib/Target/X86/X86InstrInfo.cpp
> > index 0c45094b6e2f6..f4ffb42d9972d 100644
> > --- a/llvm/lib/Target/X86/X86InstrInfo.cpp
> > +++ b/llvm/lib/Target/X86/X86InstrInfo.cpp
> > @@ -25,13 +25,16 @@
> > #include "llvm/CodeGen/MachineConstantPool.h"
> > #include "llvm/CodeGen/MachineDominators.h"
> > #include "llvm/CodeGen/MachineFrameInfo.h"
> > +#include "llvm/CodeGen/MachineInstr.h"
> > #include "llvm/CodeGen/MachineInstrBuilder.h"
> > #include "llvm/CodeGen/MachineModuleInfo.h"
> > +#include "llvm/CodeGen/MachineOperand.h"
> > #include "llvm/CodeGen/MachineRegisterInfo.h"
> > #include "llvm/CodeGen/StackMaps.h"
> > #include "llvm/IR/DebugInfoMetadata.h"
> > #include "llvm/IR/DerivedTypes.h"
> > #include "llvm/IR/Function.h"
> > +#include "llvm/IR/InstrTypes.h"
> > #include "llvm/MC/MCAsmInfo.h"
> > #include "llvm/MC/MCExpr.h"
> > #include "llvm/MC/MCInst.h"
> > @@ -964,6 +967,101 @@ inline static bool
> isTruncatedShiftCountForLEA(unsigned ShAmt) {
> > return ShAmt < 4 && ShAmt > 0;
> > }
> >
> > +static bool findRedundantFlagInstr(MachineInstr &CmpInstr,
> > + MachineInstr &CmpValDefInstr,
> > + const MachineRegisterInfo *MRI,
> > + MachineInstr **AndInstr,
> > + const TargetRegisterInfo *TRI,
> > + bool &NoSignFlag, bool
> &ClearsOverflowFlag) {
> > + if (CmpValDefInstr.getOpcode() != X86::SUBREG_TO_REG)
> > + return false;
> > +
> > + if (CmpInstr.getOpcode() != X86::TEST64rr)
> > + return false;
> > +
> > + // CmpInstr is a TEST64rr instruction, and
> `X86InstrInfo::analyzeCompare`
> > + // guarantees that it's analyzable only if two registers are
> identical.
> > + assert(
> > + (CmpInstr.getOperand(0).getReg() ==
> CmpInstr.getOperand(1).getReg()) &&
> > + "CmpInstr is an analyzable TEST64rr, and
> `X86InstrInfo::analyzeCompare` "
> > + "requires two reg operands are the same.");
> > +
> > + // Caller (`X86InstrInfo::optimizeCompareInstr`) guarantees that
> > + // `CmpValDefInstr` defines the value that's used by `CmpInstr`; in
> this case
> > + // if `CmpValDefInstr` sets the EFLAGS, it is likely that `CmpInstr`
> is
> > + // redundant.
> > + assert(
> > + (MRI->getVRegDef(CmpInstr.getOperand(0).getReg()) ==
> &CmpValDefInstr) &&
> > + "Caller guarantees that TEST64rr is a user of SUBREG_TO_REG.");
> > +
> > + // As seen in X86 td files, CmpValDefInstr.getOperand(1).getImm() is
> typically
> > + // 0.
> > + if (CmpValDefInstr.getOperand(1).getImm() != 0)
> > + return false;
> > +
> > + // As seen in X86 td files, CmpValDefInstr.getOperand(3) is typically
> > + // sub_32bit or sub_xmm.
> > + if (CmpValDefInstr.getOperand(3).getImm() != X86::sub_32bit)
> > + return false;
> > +
> > + MachineInstr *VregDefInstr =
> > + MRI->getVRegDef(CmpValDefInstr.getOperand(2).getReg());
> > +
> > + assert(VregDefInstr && "Must have a definition (SSA)");
> > +
> > + // Requires `CmpValDefInstr` and `VregDefInstr` are from the same MBB
> > + // to simplify the subsequent analysis.
> > + //
> > + // FIXME: If `VregDefInstr->getParent()` is the only predecessor of
> > + // `CmpValDefInstr.getParent()`, this could be handled.
> > + if (VregDefInstr->getParent() != CmpValDefInstr.getParent())
> > + return false;
> > +
> > + if (X86::isAND(VregDefInstr->getOpcode())) {
> > + // Get a sequence of instructions like
> > + // %reg = and* ... // Set EFLAGS
> > + // ... // EFLAGS not changed
> > + // %extended_reg = subreg_to_reg 0, %reg, %subreg.sub_32bit
> > + // test64rr %extended_reg, %extended_reg, implicit-def $eflags
> > + //
> > + // If subsequent readers use a subset of bits that don't change
> > + // after `and*` instructions, it's likely that the test64rr could
> > + // be optimized away.
> > + for (const MachineInstr &Instr :
> > +
> make_range(std::next(MachineBasicBlock::iterator(VregDefInstr)),
> > + MachineBasicBlock::iterator(CmpValDefInstr))) {
> > + // There are instructions between 'VregDefInstr' and
> > + // 'CmpValDefInstr' that modifies EFLAGS.
> > + if (Instr.modifiesRegister(X86::EFLAGS, TRI))
> > + return false;
> > + }
> > +
> > + *AndInstr = VregDefInstr;
> > +
> > + // AND instruction will essentially update SF and clear OF, so
> > + // NoSignFlag should be false in the sense that SF is modified by
> `AND`.
> > + //
> > + // However, the implementation artifically sets `NoSignFlag` to true
> > + // to poison the SF bit; that is to say, if SF is looked at later,
> the
> > + // optimization (to erase TEST64rr) will be disabled.
> > + //
> > + // The reason to poison SF bit is that SF bit value could be
> > diff erent
> > + // in the `AND` and `TEST` operation; signed bit is not known for
> `AND`,
> > + // and is known to be 0 as a result of `TEST64rr`.
> > + //
> > + // FIXME: As opposed to poisoning the SF bit direclty, consider
> peeking into
> > + // the AND instruction and using the static information to guide
> peephole optimization if possible.
> > + // For example, it's possible to fold a conditional move into a copy
> > + // if the relevant EFLAG bits could be deduced from an immediate
> operand of and operation.
> > + //
> > + NoSignFlag = true;
> > + // ClearsOverflowFlag is true for AND operation (no surprise).
> > + ClearsOverflowFlag = true;
> > + return true;
> > + }
> > + return false;
> > +}
> > +
> > bool X86InstrInfo::classifyLEAReg(MachineInstr &MI, const
> MachineOperand &Src,
> > unsigned Opc, bool AllowSP, Register
> &NewSrc,
> > bool &isKill, MachineOperand
> &ImplicitOp,
> > @@ -4226,6 +4324,23 @@ bool
> X86InstrInfo::optimizeCompareInstr(MachineInstr &CmpInstr, Register SrcReg,
> > MI = &Inst;
> > break;
> > }
> > +
> > + // Look back for the following pattern, in which case the
> test64rr
> > + // instruction could be erased.
> > + //
> > + // Example:
> > + // %reg = and32ri %in_reg, 5
> > + // ... // EFLAGS not changed.
> > + // %src_reg = subreg_to_reg 0, %reg, %subreg.sub_index
> > + // test64rr %src_reg, %src_reg, implicit-def $eflags
> > + MachineInstr *AndInstr = nullptr;
> > + if (IsCmpZero &&
> > + findRedundantFlagInstr(CmpInstr, Inst, MRI, &AndInstr, TRI,
> > + NoSignFlag, ClearsOverflowFlag)) {
> > + assert(AndInstr != nullptr &&
> X86::isAND(AndInstr->getOpcode()));
> > + MI = AndInstr;
> > + break;
> > + }
> > // Cannot find other candidates before definition of SrcReg.
> > return false;
> > }
> >
> > diff --git a/llvm/test/CodeGen/X86/peephole-test-after-add.mir
> b/llvm/test/CodeGen/X86/peephole-test-after-add.mir
> > new file mode 100644
> > index 0000000000000..65a90fafa655c
> > --- /dev/null
> > +++ b/llvm/test/CodeGen/X86/peephole-test-after-add.mir
> > @@ -0,0 +1,196 @@
> > +# NOTE: Assertions have been autogenerated by
> utils/update_mir_test_checks.py
> > +# RUN: llc -o - %s -mtriple=x86_64-unknown-linux-gnu
> --run-pass=peephole-opt | FileCheck %s
> > +
> > +# Test that TEST64rr is erased in `test_erased`, and kept in
> `test_not_erased_when_sf_used`
> > +# and `test_not_erased_when_eflags_change`.
> > +
> > +--- |
> > + ; ModuleID = 'tmp.ll'
> > + source_filename = "tmp.ll"
> > + target datalayout =
> "e-m:e-p270:32:32-p271:32:32-p272:64:64-i64:64-f80:128-n8:16:32:64-S128"
> > +
> > + define i64 @test_erased(ptr %0, i64 %1, i64 %2) {
> > + %4 = load i64, ptr %0, align 8
> > + %5 = and i64 %4, 3
> > + %6 = icmp eq i64 %5, 0
> > + %7 = select i1 %6, i64 %1, i64 %5
> > + store i64 %7, ptr %0, align 8
> > + ret i64 %5
> > + }
> > +
> > + define i64 @test_not_erased_when_sf_used(ptr %0, i64 %1, i64 %2, i64
> %3) {
> > + %5 = load i64, ptr %0, align 8
> > + %6 = and i64 %5, 3
> > + %7 = icmp slt i64 %6, 0
> > + %8 = select i1 %7, i64 %1, i64 %6
> > + store i64 %8, ptr %0, align 8
> > + ret i64 %5
> > + }
> > +
> > + define void @test_not_erased_when_eflags_change(ptr %0, i64 %1, i64
> %2, i64 %3, ptr %4) {
> > + %6 = load i64, ptr %0, align 8
> > + %7 = and i64 %6, 3
> > + %8 = xor i64 %3, 5
> > + %9 = icmp eq i64 %7, 0
> > + %10 = select i1 %9, i64 %1, i64 %7
> > + store i64 %10, ptr %0, align 8
> > + store i64 %8, ptr %4, align 8
> > + ret void
> > + }
> > +
> > +...
> > +---
> > +name: test_erased
> > +alignment: 16
> > +tracksDebugUserValues: false
> > +registers:
> > + - { id: 0, class: gr64, preferred-register: '' }
> > + - { id: 1, class: gr64, preferred-register: '' }
> > + - { id: 2, class: gr64, preferred-register: '' }
> > + - { id: 3, class: gr64, preferred-register: '' }
> > + - { id: 4, class: gr32, preferred-register: '' }
> > + - { id: 5, class: gr32, preferred-register: '' }
> > + - { id: 6, class: gr64, preferred-register: '' }
> > + - { id: 7, class: gr64, preferred-register: '' }
> > +liveins:
> > + - { reg: '$rdi', virtual-reg: '%0' }
> > + - { reg: '$rsi', virtual-reg: '%1' }
> > +frameInfo:
> > + maxAlignment: 1
> > +machineFunctionInfo: {}
> > +body: |
> > + bb.0 (%ir-block.3):
> > + liveins: $rdi, $rsi
> > +
> > + ; CHECK-LABEL: name: test_erased
> > + ; CHECK: [[COPY:%[0-9]+]]:gr64 = COPY $rsi
> > + ; CHECK-NEXT: [[COPY1:%[0-9]+]]:gr64 = COPY $rdi
> > + ; CHECK-NEXT: [[MOV64rm:%[0-9]+]]:gr64 = MOV64rm [[COPY1]], 1,
> $noreg, 0, $noreg :: (load (s64) from %ir.0)
> > + ; CHECK-NEXT: [[COPY2:%[0-9]+]]:gr32 = COPY [[MOV64rm]].sub_32bit
> > + ; CHECK-NEXT: [[AND32ri8_:%[0-9]+]]:gr32 = AND32ri8 [[COPY2]], 3,
> implicit-def $eflags
> > + ; CHECK-NEXT: [[SUBREG_TO_REG:%[0-9]+]]:gr64 = SUBREG_TO_REG 0,
> killed [[AND32ri8_]], %subreg.sub_32bit
> > + ; CHECK-NEXT: [[CMOV64rr:%[0-9]+]]:gr64 = CMOV64rr
> [[SUBREG_TO_REG]], [[COPY]], 4, implicit $eflags
> > + ; CHECK-NEXT: MOV64mr [[COPY1]], 1, $noreg, 0, $noreg, killed
> [[CMOV64rr]] :: (store (s64) into %ir.0)
> > + ; CHECK-NEXT: $rax = COPY [[SUBREG_TO_REG]]
> > + ; CHECK-NEXT: RET 0, $rax
> > + %1:gr64 = COPY $rsi
> > + %0:gr64 = COPY $rdi
> > + %3:gr64 = MOV64rm %0, 1, $noreg, 0, $noreg :: (load (s64) from
> %ir.0)
> > + %4:gr32 = COPY %3.sub_32bit
> > + %5:gr32 = AND32ri8 %4, 3, implicit-def dead $eflags
> > + %6:gr64 = SUBREG_TO_REG 0, killed %5, %subreg.sub_32bit
> > + TEST64rr %6, %6, implicit-def $eflags
> > + %7:gr64 = CMOV64rr %6, %1, 4, implicit $eflags
> > + MOV64mr %0, 1, $noreg, 0, $noreg, killed %7 :: (store (s64) into
> %ir.0)
> > + $rax = COPY %6
> > + RET 0, $rax
> > +
> > +...
> > +---
> > +name: test_not_erased_when_sf_used
> > +alignment: 16
> > +tracksDebugUserValues: false
> > +registers:
> > + - { id: 0, class: gr64, preferred-register: '' }
> > + - { id: 1, class: gr64, preferred-register: '' }
> > + - { id: 2, class: gr64, preferred-register: '' }
> > + - { id: 3, class: gr64, preferred-register: '' }
> > + - { id: 4, class: gr64, preferred-register: '' }
> > + - { id: 5, class: gr32, preferred-register: '' }
> > + - { id: 6, class: gr32, preferred-register: '' }
> > + - { id: 7, class: gr64, preferred-register: '' }
> > + - { id: 8, class: gr64, preferred-register: '' }
> > +liveins:
> > + - { reg: '$rdi', virtual-reg: '%0' }
> > + - { reg: '$rsi', virtual-reg: '%1' }
> > +frameInfo:
> > + maxAlignment: 1
> > +machineFunctionInfo: {}
> > +body: |
> > + bb.0 (%ir-block.4):
> > + liveins: $rdi, $rsi
> > +
> > + ; CHECK-LABEL: name: test_not_erased_when_sf_used
> > + ; CHECK: [[COPY:%[0-9]+]]:gr64 = COPY $rsi
> > + ; CHECK-NEXT: [[COPY1:%[0-9]+]]:gr64 = COPY $rdi
> > + ; CHECK-NEXT: [[MOV64rm:%[0-9]+]]:gr64 = MOV64rm [[COPY1]], 1,
> $noreg, 0, $noreg :: (load (s64) from %ir.0)
> > + ; CHECK-NEXT: [[COPY2:%[0-9]+]]:gr32 = COPY [[MOV64rm]].sub_32bit
> > + ; CHECK-NEXT: [[AND32ri8_:%[0-9]+]]:gr32 = AND32ri8 [[COPY2]], 3,
> implicit-def dead $eflags
> > + ; CHECK-NEXT: [[SUBREG_TO_REG:%[0-9]+]]:gr64 = SUBREG_TO_REG 0,
> killed [[AND32ri8_]], %subreg.sub_32bit
> > + ; CHECK-NEXT: TEST64rr [[SUBREG_TO_REG]], [[SUBREG_TO_REG]],
> implicit-def $eflags
> > + ; CHECK-NEXT: [[CMOV64rr:%[0-9]+]]:gr64 = CMOV64rr
> [[SUBREG_TO_REG]], [[COPY]], 8, implicit $eflags
> > + ; CHECK-NEXT: MOV64mr [[COPY1]], 1, $noreg, 0, $noreg, killed
> [[CMOV64rr]] :: (store (s64) into %ir.0)
> > + ; CHECK-NEXT: $rax = COPY [[MOV64rm]]
> > + ; CHECK-NEXT: RET 0, $rax
> > + %1:gr64 = COPY $rsi
> > + %0:gr64 = COPY $rdi
> > + %4:gr64 = MOV64rm %0, 1, $noreg, 0, $noreg :: (load (s64) from
> %ir.0)
> > + %5:gr32 = COPY %4.sub_32bit
> > + %6:gr32 = AND32ri8 %5, 3, implicit-def dead $eflags
> > + %7:gr64 = SUBREG_TO_REG 0, killed %6, %subreg.sub_32bit
> > + TEST64rr %7, %7, implicit-def $eflags
> > + %8:gr64 = CMOV64rr %7, %1, 8, implicit $eflags
> > + MOV64mr %0, 1, $noreg, 0, $noreg, killed %8 :: (store (s64) into
> %ir.0)
> > + $rax = COPY %4
> > + RET 0, $rax
> > +
> > +...
> > +---
> > +name: test_not_erased_when_eflags_change
> > +alignment: 16
> > +tracksDebugUserValues: false
> > +registers:
> > + - { id: 0, class: gr64, preferred-register: '' }
> > + - { id: 1, class: gr64, preferred-register: '' }
> > + - { id: 2, class: gr64, preferred-register: '' }
> > + - { id: 3, class: gr64, preferred-register: '' }
> > + - { id: 4, class: gr64, preferred-register: '' }
> > + - { id: 5, class: gr64, preferred-register: '' }
> > + - { id: 6, class: gr32, preferred-register: '' }
> > + - { id: 7, class: gr32, preferred-register: '' }
> > + - { id: 8, class: gr64, preferred-register: '' }
> > + - { id: 9, class: gr64, preferred-register: '' }
> > + - { id: 10, class: gr64, preferred-register: '' }
> > +liveins:
> > + - { reg: '$rdi', virtual-reg: '%0' }
> > + - { reg: '$rsi', virtual-reg: '%1' }
> > + - { reg: '$rcx', virtual-reg: '%3' }
> > + - { reg: '$r8', virtual-reg: '%4' }
> > +frameInfo:
> > + maxAlignment: 1
> > +machineFunctionInfo: {}
> > +body: |
> > + bb.0 (%ir-block.5):
> > + liveins: $rdi, $rsi, $rcx, $r8
> > +
> > + ; CHECK-LABEL: name: test_not_erased_when_eflags_change
> > + ; CHECK: [[COPY:%[0-9]+]]:gr64 = COPY $r8
> > + ; CHECK-NEXT: [[COPY1:%[0-9]+]]:gr64 = COPY $rcx
> > + ; CHECK-NEXT: [[COPY2:%[0-9]+]]:gr64 = COPY $rsi
> > + ; CHECK-NEXT: [[COPY3:%[0-9]+]]:gr64 = COPY $rdi
> > + ; CHECK-NEXT: [[MOV64rm:%[0-9]+]]:gr64 = MOV64rm [[COPY3]], 1,
> $noreg, 0, $noreg :: (load (s64) from %ir.0)
> > + ; CHECK-NEXT: [[COPY4:%[0-9]+]]:gr32 = COPY [[MOV64rm]].sub_32bit
> > + ; CHECK-NEXT: [[AND32ri8_:%[0-9]+]]:gr32 = AND32ri8 [[COPY4]], 3,
> implicit-def dead $eflags
> > + ; CHECK-NEXT: [[SUBREG_TO_REG:%[0-9]+]]:gr64 = SUBREG_TO_REG 0,
> killed [[AND32ri8_]], %subreg.sub_32bit
> > + ; CHECK-NEXT: [[XOR64ri8_:%[0-9]+]]:gr64 = XOR64ri8 [[COPY1]], 5,
> implicit-def dead $eflags
> > + ; CHECK-NEXT: TEST64rr [[SUBREG_TO_REG]], [[SUBREG_TO_REG]],
> implicit-def $eflags
> > + ; CHECK-NEXT: [[CMOV64rr:%[0-9]+]]:gr64 = CMOV64rr
> [[SUBREG_TO_REG]], [[COPY2]], 4, implicit $eflags
> > + ; CHECK-NEXT: MOV64mr [[COPY3]], 1, $noreg, 0, $noreg, killed
> [[CMOV64rr]] :: (store (s64) into %ir.0)
> > + ; CHECK-NEXT: MOV64mr [[COPY]], 1, $noreg, 0, $noreg, killed
> [[XOR64ri8_]] :: (store (s64) into %ir.4)
> > + ; CHECK-NEXT: RET 0
> > + %4:gr64 = COPY $r8
> > + %3:gr64 = COPY $rcx
> > + %1:gr64 = COPY $rsi
> > + %0:gr64 = COPY $rdi
> > + %5:gr64 = MOV64rm %0, 1, $noreg, 0, $noreg :: (load (s64) from
> %ir.0)
> > + %6:gr32 = COPY %5.sub_32bit
> > + %7:gr32 = AND32ri8 %6, 3, implicit-def dead $eflags
> > + %8:gr64 = SUBREG_TO_REG 0, killed %7, %subreg.sub_32bit
> > + %9:gr64 = XOR64ri8 %3, 5, implicit-def dead $eflags
> > + TEST64rr %8, %8, implicit-def $eflags
> > + %10:gr64 = CMOV64rr %8, %1, 4, implicit $eflags
> > + MOV64mr %0, 1, $noreg, 0, $noreg, killed %10 :: (store (s64) into
> %ir.0)
> > + MOV64mr %4, 1, $noreg, 0, $noreg, killed %9 :: (store (s64) into
> %ir.4)
> > + RET 0
> > +
> > +...
> >
> >
> >
> > _______________________________________________
> > llvm-commits mailing list
> > llvm-commits at lists.llvm.org
> > https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-commits
>
--
Thanks,
Mingming
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20220510/5e8c1526/attachment-0001.html>
More information about the llvm-commits
mailing list