Re: [TCWG CI] 400.perlbench slowed down by 6% after llvm: [AArch64] Remove redundant ORRWrs which is generated by zero-extend

Hi David,

Thanks for looking at this!

I can’t immediately say that this is a false positive, the performance difference reproduces in several independent builds.

Looking at the save-temps -- at least 400.perlbench’es regexec.s (which hosts S_regmatch()) has 19 extra instructions, which are, if I spotted correctly, for a couple of additional stack spills. To get to the bottom of this we need to look at the runtime profiles, which are not automatically generated yet. One need to dig them up from the raw benchmarking data we have stored.

-- Maxim Kuvyrkov https://www.linaro.org

On 27 Oct 2021, at 17:14, David Spickett david.spickett@linaro.org wrote:

I think this is a false positive/one off disturbance in the benchmarking. Based on the contents of the saved temps.

FastFullPelBlockMotionSearch has not changed at all. (so unless perf is saying time spent in that function and its callees went up, it must be something other than code change)

perlbench has assorted changes that all make sense with the llvm change. Some redundant movs removed, register numbers shuffled. Potentially it hit some slow path for this core, but doesn't seem like something to be concerned about for this change.

@Maxim Kuvyrkov Please correct me if I'm wrong.

On Wed, 27 Oct 2021 at 14:33, ci_notify@linaro.org wrote:

...

After llvm commit a502436259307f95e9c95437d8a1d2d07294341c Author: Jingu Kang jingu.kang@arm.com

[AArch64] Remove redundant ORRWrs which is generated by zero-extend

the following benchmarks slowed down by more than 2%:

• 400.perlbench slowed down by 6% from 9792 to 10354 perf samples
• 464.h264ref slowed down by 4% from 11023 to 11509 perf samples
• 464.h264ref:[.] FastFullPelBlockMotionSearch slowed down by 33% from 1634 to 2180 perf samples

Below reproducer instructions can be used to re-build both "first_bad" and "last_good" cross-toolchains used in this bisection. Naturally, the scripts will fail when triggerring benchmarking jobs if you don't have access to Linaro TCWG CI.

For your convenience, we have uploaded tarballs with pre-processed source and assembly files at:

• First_bad save-temps: https://ci.linaro.org/job/tcwg_bmk_ci_llvm-bisect-tcwg_bmk_tx1-llvm-master-a...
• Last_good save-temps: https://ci.linaro.org/job/tcwg_bmk_ci_llvm-bisect-tcwg_bmk_tx1-llvm-master-a...
• Baseline save-temps: https://ci.linaro.org/job/tcwg_bmk_ci_llvm-bisect-tcwg_bmk_tx1-llvm-master-a...

Configuration:

• Benchmark: SPEC CPU2006
• Toolchain: Clang + Glibc + LLVM Linker
• Version: all components were built from their tip of trunk
• Target: aarch64-linux-gnu
• Compiler flags: -O3
• Hardware: NVidia TX1 4x Cortex-A57

This benchmarking CI is work-in-progress, and we welcome feedback and suggestions at linaro-toolchain@lists.linaro.org . In our improvement plans is to add support for SPEC CPU2017 benchmarks and provide "perf report/annotate" data behind these reports.

THIS IS THE END OF INTERESTING STUFF. BELOW ARE LINKS TO BUILDS, REPRODUCTION INSTRUCTIONS, AND THE RAW COMMIT.

This commit has regressed these CI configurations:

• tcwg_bmk_llvm_tx1/llvm-master-aarch64-spec2k6-O3

First_bad build: https://ci.linaro.org/job/tcwg_bmk_ci_llvm-bisect-tcwg_bmk_tx1-llvm-master-a... Last_good build: https://ci.linaro.org/job/tcwg_bmk_ci_llvm-bisect-tcwg_bmk_tx1-llvm-master-a... Baseline build: https://ci.linaro.org/job/tcwg_bmk_ci_llvm-bisect-tcwg_bmk_tx1-llvm-master-a... Even more details: https://ci.linaro.org/job/tcwg_bmk_ci_llvm-bisect-tcwg_bmk_tx1-llvm-master-a...

Reproduce builds:

<cut> mkdir investigate-llvm-a502436259307f95e9c95437d8a1d2d07294341c cd investigate-llvm-a502436259307f95e9c95437d8a1d2d07294341c

# Fetch scripts git clone https://git.linaro.org/toolchain/jenkins-scripts

# Fetch manifests and test.sh script mkdir -p artifacts/manifests curl -o artifacts/manifests/build-baseline.sh https://ci.linaro.org/job/tcwg_bmk_ci_llvm-bisect-tcwg_bmk_tx1-llvm-master-a... --fail curl -o artifacts/manifests/build-parameters.sh https://ci.linaro.org/job/tcwg_bmk_ci_llvm-bisect-tcwg_bmk_tx1-llvm-master-a... --fail curl -o artifacts/test.sh https://ci.linaro.org/job/tcwg_bmk_ci_llvm-bisect-tcwg_bmk_tx1-llvm-master-a... --fail chmod +x artifacts/test.sh

# Reproduce the baseline build (build all pre-requisites) ./jenkins-scripts/tcwg_bmk-build.sh @@ artifacts/manifests/build-baseline.sh

# Save baseline build state (which is then restored in artifacts/test.sh) mkdir -p ./bisect rsync -a --del --delete-excluded --exclude /bisect/ --exclude /artifacts/ --exclude /llvm/ ./ ./bisect/baseline/

cd llvm

# Reproduce first_bad build git checkout --detach a502436259307f95e9c95437d8a1d2d07294341c ../artifacts/test.sh

# Reproduce last_good build git checkout --detach 6fa1b4ff4b05b9b9a432f7310802255c160c8f4f ../artifacts/test.sh

cd ..

</cut>

Full commit (up to 1000 lines):

<cut> commit a502436259307f95e9c95437d8a1d2d07294341c Author: Jingu Kang <jingu.kang@arm.com> Date: Thu Sep 30 15:39:10 2021 +0100

[AArch64] Remove redundant ORRWrs which is generated by zero-extend

%3:gpr32 = ORRWrs $wzr, %2, 0 %4:gpr64 = SUBREG_TO_REG 0, %3, %subreg.sub_32

If AArch64's 32-bit form of instruction defines the source operand of ORRWrs, we can remove the ORRWrs because the upper 32 bits of the source operand are set to zero.

Differential Revision: https://reviews.llvm.org/D110841

llvm/lib/Target/AArch64/AArch64MIPeepholeOpt.cpp | 57 ++++++++++++++++ .../test/CodeGen/AArch64/arm64-assert-zext-sext.ll | 51 +++++++++++--- .../AArch64/redundant-mov-from-zero-extend.ll | 79 ++++++++++++++++++++++ .../AArch64/redundant-orrwrs-from-zero-extend.mir | 69 +++++++++++++++++++ 4 files changed, 248 insertions(+), 8 deletions(-)

diff --git a/llvm/lib/Target/AArch64/AArch64MIPeepholeOpt.cpp b/llvm/lib/Target/AArch64/AArch64MIPeepholeOpt.cpp index 42f683613698..42db18332f1c 100644 --- a/llvm/lib/Target/AArch64/AArch64MIPeepholeOpt.cpp +++ b/llvm/lib/Target/AArch64/AArch64MIPeepholeOpt.cpp @@ -15,6 +15,16 @@ // later. In this case, we could try to split the constant operand of mov // instruction into two bitmask immediates. It makes two AND instructions // intead of multiple `mov` + `and` instructions. +// +// 2. Remove redundant ORRWrs which is generated by zero-extend. +// +// %3:gpr32 = ORRWrs $wzr, %2, 0 +// %4:gpr64 = SUBREG_TO_REG 0, %3, %subreg.sub_32 +// +// If AArch64's 32-bit form of instruction defines the source operand of +// ORRWrs, we can remove the ORRWrs because the upper 32 bits of the source +// operand are set to zero. +// //===----------------------------------------------------------------------===//

#include "AArch64ExpandImm.h" @@ -44,6 +54,8 @@ struct AArch64MIPeepholeOpt : public MachineFunctionPass { template <typename T> bool visitAND(MachineInstr &MI, SmallSetVector<MachineInstr *, 8> &ToBeRemoved);

• bool visitORR(MachineInstr &MI,

•            SmallSetVector<MachineInstr *, 8> &ToBeRemoved);
  

  bool runOnMachineFunction(MachineFunction &MF) override;

  StringRef getPassName() const override {

@@ -196,6 +208,49 @@ bool AArch64MIPeepholeOpt::visitAND( return true; }

+bool AArch64MIPeepholeOpt::visitORR(

• MachineInstr &MI, SmallSetVector<MachineInstr *, 8> &ToBeRemoved) {
• // Check this ORR comes from below zero-extend pattern.
• //
• // def : Pat<(i64 (zext GPR32:$src)),
• // (SUBREG_TO_REG (i32 0), (ORRWrs WZR, GPR32:$src, 0), sub_32)>;
• if (MI.getOperand(3).getImm() != 0)
• return false;
• if (MI.getOperand(1).getReg() != AArch64::WZR)
• return false;
• MachineInstr *SrcMI = MRI->getUniqueVRegDef(MI.getOperand(2).getReg());
• if (!SrcMI)
• return false;
• // From https://developer.arm.com/documentation/dui0801/b/BABBGCAC
• //
• // When you use the 32-bit form of an instruction, the upper 32 bits of the
• // source registers are ignored and the upper 32 bits of the destination
• // register are set to zero.
• //
• // If AArch64's 32-bit form of instruction defines the source operand of
• // zero-extend, we do not need the zero-extend. Let's check the MI's opcode is
• // real AArch64 instruction and if it is not, do not process the opcode
• // conservatively.
• if (SrcMI->getOpcode() <= TargetOpcode::GENERIC_OP_END)
• return false;
• Register DefReg = MI.getOperand(0).getReg();
• Register SrcReg = MI.getOperand(2).getReg();
• MRI->replaceRegWith(DefReg, SrcReg);
• MRI->clearKillFlags(SrcReg);
• // replaceRegWith changes MI's definition register. Keep it for SSA form until
• // deleting MI.
• MI.getOperand(0).setReg(DefReg);
• ToBeRemoved.insert(&MI);
• LLVM_DEBUG({ dbgs() << "Removed: " << MI << "\n"; });
• return true;

+}

bool AArch64MIPeepholeOpt::runOnMachineFunction(MachineFunction &MF) { if (skipFunction(MF.getFunction())) return false; @@ -221,6 +276,8 @@ bool AArch64MIPeepholeOpt::runOnMachineFunction(MachineFunction &MF) { case AArch64::ANDXrr: Changed = visitAND<uint64_t>(MI, ToBeRemoved); break;

•  case AArch64::ORRWrs:
  

•    Changed = visitORR(MI, ToBeRemoved);
  }
  

  } }

diff --git a/llvm/test/CodeGen/AArch64/arm64-assert-zext-sext.ll b/llvm/test/CodeGen/AArch64/arm64-assert-zext-sext.ll index 07dd0b4ec56b..df4a9010dfa9 100644 --- a/llvm/test/CodeGen/AArch64/arm64-assert-zext-sext.ll +++ b/llvm/test/CodeGen/AArch64/arm64-assert-zext-sext.ll @@ -1,9 +1,32 @@ +; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py ; RUN: llc -O2 -mtriple=aarch64-linux-gnu < %s | FileCheck %s

declare i32 @test(i32) local_unnamed_addr declare i32 @test1(i64) local_unnamed_addr

define i32 @assertzext(i32 %n, i1 %a, i32* %b) local_unnamed_addr { +; CHECK-LABEL: assertzext: +; CHECK: // %bb.0: // %entry +; CHECK-NEXT: stp x30, x19, [sp, #-16]! // 16-byte Folded Spill +; CHECK-NEXT: .cfi_def_cfa_offset 16 +; CHECK-NEXT: .cfi_offset w19, -8 +; CHECK-NEXT: .cfi_offset w30, -16 +; CHECK-NEXT: mov w8, #33066 +; CHECK-NEXT: tst w1, #0x1 +; CHECK-NEXT: movk w8, #28567, lsl #16 +; CHECK-NEXT: csel w19, wzr, w8, ne +; CHECK-NEXT: cbnz w0, .LBB0_2 +; CHECK-NEXT: // %bb.1: // %if.then +; CHECK-NEXT: mov w19, wzr +; CHECK-NEXT: str wzr, [x2] +; CHECK-NEXT: .LBB0_2: // %if.end +; CHECK-NEXT: mov w0, w19 +; CHECK-NEXT: bl test +; CHECK-NEXT: mov w0, w19 +; CHECK-NEXT: bl test1 +; CHECK-NEXT: mov w0, wzr +; CHECK-NEXT: ldp x30, x19, [sp], #16 // 16-byte Folded Reload +; CHECK-NEXT: ret entry: %i = select i1 %a, i64 0, i64 66296709418 %conv.i = trunc i64 %i to i32 @@ -20,14 +43,29 @@ if.end: ; preds = %if.then, %entry %i2 = sext i32 %i1 to i64 %call1.i = tail call i32 @test1(i64 %i2) ret i32 0 -; CHECK: // %if.end -; CHECK: mov w{{[0-9]+}}, w{{[0-9]+}} -; CHECK: bl test -; CHECK: mov w{{[0-9]+}}, w{{[0-9]+}} -; CHECK: bl test1 }

define i32 @assertsext(i32 %n, i8 %a) local_unnamed_addr { +; CHECK-LABEL: assertsext: +; CHECK: // %bb.0: // %entry +; CHECK-NEXT: cbz w0, .LBB1_2 +; CHECK-NEXT: // %bb.1: +; CHECK-NEXT: mov x0, xzr +; CHECK-NEXT: b .LBB1_3 +; CHECK-NEXT: .LBB1_2: // %if.then +; CHECK-NEXT: mov x9, #24575 +; CHECK-NEXT: sxtb w8, w1 +; CHECK-NEXT: movk x9, #15873, lsl #16 +; CHECK-NEXT: movk x9, #474, lsl #32 +; CHECK-NEXT: udiv x0, x9, x8 +; CHECK-NEXT: .LBB1_3: // %if.end +; CHECK-NEXT: str x30, [sp, #-16]! // 8-byte Folded Spill +; CHECK-NEXT: .cfi_def_cfa_offset 16 +; CHECK-NEXT: .cfi_offset w30, -16 +; CHECK-NEXT: bl test1 +; CHECK-NEXT: mov w0, wzr +; CHECK-NEXT: ldr x30, [sp], #16 // 8-byte Folded Reload +; CHECK-NEXT: ret entry: %conv.i = sext i8 %a to i32 %cmp = icmp eq i32 %n, 0 @@ -37,9 +75,6 @@ if.then: ; preds = %entry %conv1 = zext i32 %conv.i to i64 %div = udiv i64 2036854775807, %conv1 br label %if.end -; CHECK: // %if.then -; CHECK: mov w{{[0-9]+}}, w{{[0-9]+}} -; CHECK: udiv x{{[0-9]+}}, x{{[0-9]+}}, x{{[0-9]+}}

if.end: ; preds = %if.then, %entry %i1 = phi i64 [ %div, %if.then ], [ 0, %entry ] diff --git a/llvm/test/CodeGen/AArch64/redundant-mov-from-zero-extend.ll b/llvm/test/CodeGen/AArch64/redundant-mov-from-zero-extend.ll new file mode 100644 index 000000000000..42b9838acef2 --- /dev/null +++ b/llvm/test/CodeGen/AArch64/redundant-mov-from-zero-extend.ll @@ -0,0 +1,79 @@ +; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py +; RUN: llc -O3 -mtriple=aarch64-linux-gnu < %s | FileCheck %s

+define i32 @test(i32 %input, i32 %n, i32 %a) { +; CHECK-LABEL: test: +; CHECK: // %bb.0: // %entry +; CHECK-NEXT: cbz w1, .LBB0_2 +; CHECK-NEXT: // %bb.1: +; CHECK-NEXT: mov w0, wzr +; CHECK-NEXT: ret +; CHECK-NEXT: .LBB0_2: // %bb.0 +; CHECK-NEXT: add w8, w0, w1 +; CHECK-NEXT: mov w0, #100 +; CHECK-NEXT: cmp w8, #4 +; CHECK-NEXT: b.hi .LBB0_5 +; CHECK-NEXT: // %bb.3: // %bb.0 +; CHECK-NEXT: adrp x9, .LJTI0_0 +; CHECK-NEXT: add x9, x9, :lo12:.LJTI0_0 +; CHECK-NEXT: adr x10, .LBB0_4 +; CHECK-NEXT: ldrb w11, [x9, x8] +; CHECK-NEXT: add x10, x10, x11, lsl #2 +; CHECK-NEXT: br x10 +; CHECK-NEXT: .LBB0_4: // %sw.bb +; CHECK-NEXT: add w0, w2, #1 +; CHECK-NEXT: ret +; CHECK-NEXT: .LBB0_5: // %bb.0 +; CHECK-NEXT: cmp w8, #200 +; CHECK-NEXT: b.ne .LBB0_10 +; CHECK-NEXT: // %bb.6: // %sw.bb7 +; CHECK-NEXT: add w0, w2, #7 +; CHECK-NEXT: ret +; CHECK-NEXT: .LBB0_7: // %sw.bb1 +; CHECK-NEXT: add w0, w2, #3 +; CHECK-NEXT: ret +; CHECK-NEXT: .LBB0_8: // %sw.bb3 +; CHECK-NEXT: add w0, w2, #4 +; CHECK-NEXT: ret +; CHECK-NEXT: .LBB0_9: // %sw.bb5 +; CHECK-NEXT: add w0, w2, #5 +; CHECK-NEXT: .LBB0_10: // %return +; CHECK-NEXT: ret +entry:

• %b = add nsw i32 %input, %n
• %cmp = icmp eq i32 %n, 0
• br i1 %cmp, label %bb.0, label %return

+bb.0:

• switch i32 %b, label %return [
• i32 0, label %sw.bb
• i32 1, label %sw.bb1
• i32 2, label %sw.bb3
• i32 4, label %sw.bb5
• i32 200, label %sw.bb7
• ]

+sw.bb:

• %add = add nsw i32 %a, 1
• br label %return

+sw.bb1:

• %add2 = add nsw i32 %a, 3
• br label %return

+sw.bb3:

• %add4 = add nsw i32 %a, 4
• br label %return

+sw.bb5:

• %add6 = add nsw i32 %a, 5
• br label %return

+sw.bb7:

• %add8 = add nsw i32 %a, 7
• br label %return

+return:

• %retval.0 = phi i32 [ %add8, %sw.bb7 ], [ %add6, %sw.bb5 ], [ %add4, %sw.bb3 ], [ %add2, %sw.bb1 ], [ %add, %sw.bb ], [ 100, %bb.0 ], [ 0, %entry ]
• ret i32 %retval.0

+} diff --git a/llvm/test/CodeGen/AArch64/redundant-orrwrs-from-zero-extend.mir b/llvm/test/CodeGen/AArch64/redundant-orrwrs-from-zero-extend.mir new file mode 100644 index 000000000000..37540dde048f --- /dev/null +++ b/llvm/test/CodeGen/AArch64/redundant-orrwrs-from-zero-extend.mir @@ -0,0 +1,69 @@ +# RUN: llc -mtriple=aarch64 -run-pass aarch64-mi-peephole-opt -verify-machineinstrs -o - %s | FileCheck %s +--- +name: test1 +# CHECK-LABEL: name: test1 +alignment: 4 +tracksRegLiveness: true +registers:

• • { id: 0, class: gpr32 }
• • { id: 1, class: gpr32 }
• • { id: 2, class: gpr32 }
• • { id: 3, class: gpr32 }
• • { id: 4, class: gpr64 }

+body: |

• bb.0:
• liveins: $w0, $w1
• %0:gpr32 = COPY $w0
• %1:gpr32 = COPY $w1
• B %bb.1
• bb.1:
• %2:gpr32 = nsw ADDWrr %0, %1
• B %bb.2
• bb.2:
• ; CHECK-LABEL: bb.2:
• ; CHECK-NOT: %3:gpr32 = ORRWrs $wzr, %2, 0
• ; The ORRWrs should be removed.
• %3:gpr32 = ORRWrs $wzr, %2, 0
• %4:gpr64 = SUBREG_TO_REG 0, %3, %subreg.sub_32
• B %bb.3
• bb.3:
• $x0 = COPY %4
• RET_ReallyLR implicit $x0

+... +--- +name: test2 +# CHECK-LABEL: name: test2 +alignment: 4 +tracksRegLiveness: true +registers:

• • { id: 0, class: gpr64 }
• • { id: 1, class: gpr32 }
• • { id: 2, class: gpr32 }
• • { id: 3, class: gpr64 }

+body: |

• bb.0:
• liveins: $x0
• %0:gpr64 = COPY $x0
• B %bb.1
• bb.1:
• %1:gpr32 = EXTRACT_SUBREG %0, %subreg.sub_32
• B %bb.2
• bb.2:
• ; CHECK-LABEL: bb.2:
• ; CHECK: %2:gpr32 = ORRWrs $wzr, %1, 0
• ; The ORRWrs should not be removed.
• %2:gpr32 = ORRWrs $wzr, %1, 0
• %3:gpr64 = SUBREG_TO_REG 0, %2, %subreg.sub_32
• B %bb.3
• bb.3:
• $x0 = COPY %3
• RET_ReallyLR implicit $x0

+...

</cut>