[X86] Avoid SFB - Fix inconsistent codegen with/without debug info

[cdawson]

Summary

Fixes https://bugs.llvm.org/show_bug.cgi?id=40969

The functions findPotentiallyBlockedCopies and buildCopy are currently not
accounting for the presence of debug instructions. In the former this results
in the optimization not being trigerred, and in the latter results in
inconsistent codegen.

This patch enables the optimization to be performed in a debug build and
ensures the codegen is consistent with non-debug builds.

[lebedev.ri]

Hm, can -debugify be run by llc?

[lebedev.ri]

llvm/test/CodeGen/X86/avoid-sfb-g-no-change.ll

1

Use utils/update_llc_test_checks

[gbedwell]

In D61680#1495085, @lebedev.ri wrote:

Hm, can -debugify be run by llc?

No. The debugify pass itself lives inside llvm/tools/opt.

[cdawson]

llvm/test/CodeGen/X86/avoid-sfb-g-no-change.ll

1

I tried that initially but it doesn't generate any checks for the IR.

[probinson]

The point is not that a certain set of moves comes out, but that the *same* set of moves comes out regardless of debug info.
So, the test wants to have two copies of the same function, one with debug instructions and one without; and both functions should produce the same sequence of moves.

To keep the test from being really tedious, I am pretty sure you can run FileCheck twice and have each run look at a different function, with clever use of -D on the FileCheck command line. This probably means you have to capture the llc output in a file, but that's okay.

[aprantl]

llvm/lib/Target/X86/X86AvoidStoreForwardingBlocks.cpp

410

std::prev ?

[probinson]

I wonder if this should be an MIR test? Fewer moving parts.

[cdawson]

Addressed review comments

[probinson]

llvm/test/CodeGen/X86/avoid-sfb-g-no-change.mir

247 ↗

(On Diff #198842)

Add ; DEBUG-LABEL: name: nodebug at the end, to prove the CHECK lines are all looking at the correct copy of the function.

[cdawson]

Addressed revision comments

[probinson]

LGTM

[leonardchan]

Hi, it seems that this patch results in an assertion error when building zircon:

clang: /usr/local/google/home/leonardchan/llvm-monorepo/llvm-project-2/llvm/lib/Target/X86/X86AvoidStoreForwardingBlocks.cpp:293: int getAddrOffset(llvm::MachineInstr *): Assertion `AddrOffset != -1 && "Expected Memory Operand"' failed.

Could you take a look and either fix or revert this patch? I filed a bug report with the full stack trace and reproducer you could use (https://bugs.llvm.org/show_bug.cgi?id=41838)

[cdawson]

This fix addresses https://bugs.llvm.org/show_bug.cgi?id=41838

The previous test case didn't trigger the codepath which contained the assertion. The updated test rectifies this.

[cdawson]

Reopening after updating patch

[cdawson]

Upload full diff with both sets of changes.

[cdawson]

Requesting review

[cdawson]

Ping

llvm/lib/Target/X86/X86AvoidStoreForwardingBlocks.cpp

MachineInstr *NewLoad =
           .addMemOperand(
               MBB->getParent()->getMachineMemOperand(LMMO, LMMOffset, Size));
   if (LoadBase.isReg())
     getBaseOperand(NewLoad).setIsKill(false);
   LLVM_DEBUG(NewLoad->dump());
   // If the load and store are consecutive, use the loadInst location to
   // reduce register pressure.
   MachineInstr *StInst = StoreInst;
-  if (StoreInst->getPrevNode() == LoadInst)
+  auto PrevInstrIt = skipDebugInstructionsBackward(
+      --MachineBasicBlock::instr_iterator(StoreInst), MBB->instr_begin());
+  if (PrevInstrIt.getNodePtr() == LoadInst)
     StInst = LoadInst;
   MachineInstr *NewStore =
       BuildMI(*MBB, StInst, StInst->getDebugLoc(), TII->get(NStoreOpcode))
           .add(StoreBase)
           .addImm(1)
           .addReg(X86::NoRegister)
           .addImm(StoreDisp)
           .addReg(X86::NoRegister)

 }
 
 void X86AvoidSFBPass::findPotentiallylBlockedCopies(MachineFunction &MF) {
   for (auto &MBB : MF)
     for (auto &MI : MBB) {
       if (!isPotentialBlockedMemCpyLd(MI.getOpcode()))
         continue;
       int DefVR = MI.getOperand(0).getReg();
-      if (!MRI->hasOneUse(DefVR))
+      if (!MRI->hasOneNonDBGUse(DefVR))
         continue;
       for (auto UI = MRI->use_nodbg_begin(DefVR), UE = MRI->use_nodbg_end();
            UI != UE;) {
         MachineOperand &StoreMO = *UI++;
         MachineInstr &StoreMI = *StoreMO.getParent();
         // Skip cases where the memcpy may overlap.
         if (StoreMI.getParent() == MI.getParent() &&
             isPotentialBlockedMemCpyPair(MI.getOpcode(), StoreMI.getOpcode()) &&

llvm/test/CodeGen/X86/avoid-sfb-g-no-change.ll

+; RUN: llc < %s -mtriple=x86_64-unknown-linux-gnu | FileCheck %s
+
+; This was generated from:
+;
+; using alpha = float __attribute__((ext_vector_type(4)));
+;
+; void bravo(alpha charlie) {
+;   unsigned char *delta = (unsigned char *)&charlie;
+;   delta[0] = 0;
+;   volatile alpha echo = charlie;
+; }
+;
+; Using the command line:
+; clang -g -c 1.cpp -O2 -S -emit-llvm -fno-strict-aliasing --target=x86_64-unknown-unknown -o test.ll
+
+; CHECK:  movaps  %xmm0, -16(%rbp)
+; CHECK:  movb  $0, -16(%rbp)
+; CHECK:  movb  -16(%rbp), %al
+; CHECK:  movb  %al, -32(%rbp)
+; CHECK:  movq  -15(%rbp), %rax
+; CHECK:  movq  %rax, -31(%rbp)
+; CHECK:  movl  -7(%rbp), %eax
+; CHECK:  movl  %eax, -23(%rbp)
+; CHECK:  movzwl  -3(%rbp), %eax
+; CHECK:  movw  %ax, -19(%rbp)
+; CHECK:  movb  -1(%rbp), %al
+; CHECK:  movb  %al, -17(%rbp)
+
+; Function Attrs: nounwind uwtable
+define dso_local void @_Z5bravoDv4_f(<4 x float> %charlie) local_unnamed_addr #0 !dbg !10 {
+entry:
+  %charlie.addr = alloca <4 x float>, align 16
+  %echo = alloca <4 x float>, align 16
+  call void @llvm.dbg.value(metadata <4 x float> %charlie, metadata !19, metadata !DIExpression()), !dbg !23
+  store <4 x float> %charlie, <4 x float>* %charlie.addr, align 16
+  call void @llvm.dbg.value(metadata i8* undef, metadata !20, metadata !DIExpression()), !dbg !24
+  %charlie.addr.0.charlie.addr.0.arrayidx.sroa_cast = bitcast <4 x float>* %charlie.addr to i8*, !dbg !25
+  store i8 0, i8* %charlie.addr.0.charlie.addr.0.arrayidx.sroa_cast, align 16, !dbg !25
+  %echo.0.echo.0..sroa_cast = bitcast <4 x float>* %echo to i8*, !dbg !26
+  call void @llvm.lifetime.start.p0i8(i64 16, i8* nonnull %echo.0.echo.0..sroa_cast), !dbg !26
+  call void @llvm.dbg.declare(metadata <4 x float>* %echo, metadata !21, metadata !DIExpression()), !dbg !27
+  %charlie.addr.0.charlie.addr.0. = load <4 x float>, <4 x float>* %charlie.addr, align 16, !dbg !28
+  call void @llvm.dbg.value(metadata <4 x float> %charlie.addr.0.charlie.addr.0., metadata !19, metadata !DIExpression()), !dbg !23
+  store volatile <4 x float> %charlie.addr.0.charlie.addr.0., <4 x float>* %echo, align 16, !dbg !27
+  call void @llvm.lifetime.end.p0i8(i64 16, i8* nonnull %echo.0.echo.0..sroa_cast), !dbg !29
+  ret void, !dbg !29
+}
+
+; Function Attrs: nounwind readnone speculatable
+declare void @llvm.dbg.declare(metadata, metadata, metadata) #1
+
+; Function Attrs: argmemonly nounwind
+declare void @llvm.lifetime.start.p0i8(i64 immarg, i8* nocapture) #2
+
+; Function Attrs: argmemonly nounwind
+declare void @llvm.lifetime.end.p0i8(i64 immarg, i8* nocapture) #2
+
+; Function Attrs: nounwind readnone speculatable
+declare void @llvm.dbg.value(metadata, metadata, metadata) #1
+
+attributes #0 = { nounwind uwtable "correctly-rounded-divide-sqrt-fp-math"="false" "disable-tail-calls"="false" "less-precise-fpmad"="false" "min-legal-vector-width"="128" "no-frame-pointer-elim"="true" "no-frame-pointer-elim-non-leaf" "no-infs-fp-math"="false" "no-jump-tables"="false" "no-nans-fp-math"="false" "no-signed-zeros-fp-math"="false" "no-trapping-math"="false" "stack-protector-buffer-size"="8" "target-cpu"="x86-64" "target-features"="+cx8,+fxsr,+mmx,+sse,+sse2,+x87" "unsafe-fp-math"="false" "use-soft-float"="false" }
+attributes #1 = { nounwind readnone speculatable }
+attributes #2 = { argmemonly nounwind }
+
+!llvm.dbg.cu = !{!0}
+!llvm.module.flags = !{!6, !7, !8}
+!llvm.ident = !{!9}
+
+!0 = distinct !DICompileUnit(language: DW_LANG_C_plus_plus, file: !1, producer: "clang version 9.0.0 (https://github.com/llvm/llvm-project.git c4018508182effbdf77905533365ee9ee5ae9a12)", isOptimized: true, runtimeVersion: 0, emissionKind: FullDebug, enums: !2, retainedTypes: !3, nameTableKind: None)
+!1 = !DIFile(filename: "1.cpp", directory: "C:\5CUsers\5Cgbdawsoc\5CDocuments\5Cllvm\5Cbg40969")
+!2 = !{}
+!3 = !{!4}
+!4 = !DIDerivedType(tag: DW_TAG_pointer_type, baseType: !5, size: 64)
+!5 = !DIBasicType(name: "unsigned char", size: 8, encoding: DW_ATE_unsigned_char)
+!6 = !{i32 2, !"Dwarf Version", i32 4}
+!7 = !{i32 2, !"Debug Info Version", i32 3}
+!8 = !{i32 1, !"wchar_size", i32 4}
+!9 = !{!"clang version 9.0.0 (https://github.com/llvm/llvm-project.git c4018508182effbdf77905533365ee9ee5ae9a12)"}
+!10 = distinct !DISubprogram(name: "bravo", linkageName: "_Z5bravoDv4_f", scope: !1, file: !1, line: 3, type: !11, scopeLine: 3, flags: DIFlagPrototyped, spFlags: DISPFlagDefinition | DISPFlagOptimized, unit: !0, retainedNodes: !18)
+!11 = !DISubroutineType(types: !12)
+!12 = !{null, !13}
+!13 = !DIDerivedType(tag: DW_TAG_typedef, name: "alpha", file: !1, line: 1, baseType: !14)
+!14 = !DICompositeType(tag: DW_TAG_array_type, baseType: !15, size: 128, flags: DIFlagVector, elements: !16)
+!15 = !DIBasicType(name: "float", size: 32, encoding: DW_ATE_float)
+!16 = !{!17}
+!17 = !DISubrange(count: 4)
+!18 = !{!19, !20, !21}
+!19 = !DILocalVariable(name: "charlie", arg: 1, scope: !10, file: !1, line: 3, type: !13)
+!20 = !DILocalVariable(name: "delta", scope: !10, file: !1, line: 4, type: !4)
+!21 = !DILocalVariable(name: "echo", scope: !10, file: !1, line: 6, type: !22)
+!22 = !DIDerivedType(tag: DW_TAG_volatile_type, baseType: !13)
+!23 = !DILocation(line: 3, column: 18, scope: !10)
+!24 = !DILocation(line: 4, column: 18, scope: !10)
+!25 = !DILocation(line: 5, column: 12, scope: !10)
+!26 = !DILocation(line: 6, column: 3, scope: !10)
+!27 = !DILocation(line: 6, column: 18, scope: !10)
+!28 = !DILocation(line: 6, column: 25, scope: !10)
+!29 = !DILocation(line: 7, column: 1, scope: !10)