[DebugInfo][GVNSink] Fix #77415: GVNSink fails to optimize LLVM IR with debug info #77602
Conversation
|
@llvm/pr-subscribers-debuginfo @llvm/pr-subscribers-llvm-transforms Author: Shan Huang (Apochens) ChangesThis PR fixes issue #77415 and is revised from PR #77419 . PR #77419 breaks the newly added test in the same PR on windows, because GVNSink is non-deterministic when sorting According to the report, what the CheckFile wants to match is the So I refined the regression test with a more general check rule to only detect whether there is an instruction with suffix Full diff: https://github.com/llvm/llvm-project/pull/77602.diff 2 Files Affected:
diff --git a/llvm/lib/Transforms/Scalar/GVNSink.cpp b/llvm/lib/Transforms/Scalar/GVNSink.cpp
index 2b38831139a580..9db66b3793b8dd 100644
--- a/llvm/lib/Transforms/Scalar/GVNSink.cpp
+++ b/llvm/lib/Transforms/Scalar/GVNSink.cpp
@@ -132,7 +132,7 @@ class LockstepReverseIterator {
ActiveBlocks.remove(BB);
continue;
}
- Insts.push_back(BB->getTerminator()->getPrevNode());
+ Insts.push_back(BB->getTerminator()->getPrevNonDebugInstruction());
}
if (Insts.empty())
Fail = true;
@@ -168,7 +168,7 @@ class LockstepReverseIterator {
if (Inst == &Inst->getParent()->front())
ActiveBlocks.remove(Inst->getParent());
else
- NewInsts.push_back(Inst->getPrevNode());
+ NewInsts.push_back(Inst->getPrevNonDebugInstruction());
}
if (NewInsts.empty()) {
Fail = true;
@@ -834,7 +834,7 @@ void GVNSink::sinkLastInstruction(ArrayRef<BasicBlock *> Blocks,
BasicBlock *BBEnd) {
SmallVector<Instruction *, 4> Insts;
for (BasicBlock *BB : Blocks)
- Insts.push_back(BB->getTerminator()->getPrevNode());
+ Insts.push_back(BB->getTerminator()->getPrevNonDebugInstruction());
Instruction *I0 = Insts.front();
SmallVector<Value *, 4> NewOperands;
diff --git a/llvm/test/Transforms/GVNSink/sink-ignore-dbg-intrinsics.ll b/llvm/test/Transforms/GVNSink/sink-ignore-dbg-intrinsics.ll
new file mode 100644
index 00000000000000..ee6039a0d806ab
--- /dev/null
+++ b/llvm/test/Transforms/GVNSink/sink-ignore-dbg-intrinsics.ll
@@ -0,0 +1,86 @@
+; RUN: opt < %s -passes=gvn-sink -S | FileCheck %s
+
+; Function Attrs: noinline nounwind uwtable
+define dso_local i32 @fun(i32 noundef %a, i32 noundef %b) #0 !dbg !10 {
+entry:
+ tail call void @llvm.dbg.value(metadata i32 %a, metadata !15, metadata !DIExpression()), !dbg !16
+ tail call void @llvm.dbg.value(metadata i32 %b, metadata !17, metadata !DIExpression()), !dbg !16
+ %cmp = icmp sgt i32 %b, 10, !dbg !18
+ br i1 %cmp, label %if.then, label %if.else, !dbg !20
+
+if.then: ; preds = %entry
+ %add = add nsw i32 %a, 1, !dbg !21
+ tail call void @llvm.dbg.value(metadata i32 %add, metadata !23, metadata !DIExpression()), !dbg !24
+ %xor = xor i32 %add, 1, !dbg !25
+ tail call void @llvm.dbg.value(metadata i32 %xor, metadata !26, metadata !DIExpression()), !dbg !24
+ tail call void @llvm.dbg.value(metadata i32 %xor, metadata !27, metadata !DIExpression()), !dbg !16
+ br label %if.end, !dbg !28
+
+if.else: ; preds = %entry
+ %add1 = add nsw i32 %b, 1, !dbg !29
+ tail call void @llvm.dbg.value(metadata i32 %add1, metadata !31, metadata !DIExpression()), !dbg !32
+ %xor2 = xor i32 %add1, 1, !dbg !33
+ tail call void @llvm.dbg.value(metadata i32 %xor2, metadata !34, metadata !DIExpression()), !dbg !32
+ tail call void @llvm.dbg.value(metadata i32 %xor2, metadata !27, metadata !DIExpression()), !dbg !16
+ br label %if.end
+
+; CHECK-LABEL: if.end:
+; CHECK: [[SINK:%.*.sink]] = phi i32 [ %a, %if.then ], [ %b, %if.else ]
+; CHECK: %add = add nsw i32 [[SINK]], 1
+; CHECK: %xor = xor i32 %add, 1
+if.end: ; preds = %if.else, %if.then
+ %ret.0 = phi i32 [ %xor, %if.then ], [ %xor2, %if.else ], !dbg !35
+ tail call void @llvm.dbg.value(metadata i32 %ret.0, metadata !27, metadata !DIExpression()), !dbg !16
+ ret i32 %ret.0, !dbg !36
+}
+
+; Function Attrs: nocallback nofree nosync nounwind speculatable willreturn memory(none)
+declare void @llvm.dbg.declare(metadata, metadata, metadata) #1
+
+; Function Attrs: nocallback nofree nosync nounwind speculatable willreturn memory(none)
+declare void @llvm.dbg.value(metadata, metadata, metadata) #1
+
+attributes #0 = { noinline nounwind uwtable "frame-pointer"="all" "min-legal-vector-width"="0" "no-trapping-math"="true" "stack-protector-buffer-size"="8" "target-cpu"="x86-64" "target-features"="+cmov,+cx8,+fxsr,+mmx,+sse,+sse2,+x87" "tune-cpu"="generic" }
+attributes #1 = { nocallback nofree nosync nounwind speculatable willreturn memory(none) }
+
+!llvm.dbg.cu = !{!0}
+!llvm.module.flags = !{!2, !3, !4, !5, !6, !7, !8}
+!llvm.ident = !{!9}
+
+!0 = distinct !DICompileUnit(language: DW_LANG_C11, file: !1, producer: "clang version 18.0.0git (https://github.com/llvm/llvm-project.git 5dfcb3e5d1d16bb4f8fce52b3c089119ed977e7f)", isOptimized: false, runtimeVersion: 0, emissionKind: FullDebug, splitDebugInlining: false, nameTableKind: None)
+!1 = !DIFile(filename: "main.c", directory: "/home/hs/llvm-test", checksumkind: CSK_MD5, checksum: "68c28c3d0877bed08ff43db70c573802")
+!2 = !{i32 7, !"Dwarf Version", i32 5}
+!3 = !{i32 2, !"Debug Info Version", i32 3}
+!4 = !{i32 1, !"wchar_size", i32 4}
+!5 = !{i32 8, !"PIC Level", i32 2}
+!6 = !{i32 7, !"PIE Level", i32 2}
+!7 = !{i32 7, !"uwtable", i32 2}
+!8 = !{i32 7, !"frame-pointer", i32 2}
+!9 = !{!"clang version 18.0.0git (https://github.com/llvm/llvm-project.git 5dfcb3e5d1d16bb4f8fce52b3c089119ed977e7f)"}
+!10 = distinct !DISubprogram(name: "fun", scope: !1, file: !1, line: 1, type: !11, scopeLine: 1, flags: DIFlagPrototyped, spFlags: DISPFlagDefinition, unit: !0, retainedNodes: !14)
+!11 = !DISubroutineType(types: !12)
+!12 = !{!13, !13, !13}
+!13 = !DIBasicType(name: "int", size: 32, encoding: DW_ATE_signed)
+!14 = !{}
+!15 = !DILocalVariable(name: "a", arg: 1, scope: !10, file: !1, line: 1, type: !13)
+!16 = !DILocation(line: 0, scope: !10)
+!17 = !DILocalVariable(name: "b", arg: 2, scope: !10, file: !1, line: 1, type: !13)
+!18 = !DILocation(line: 3, column: 11, scope: !19)
+!19 = distinct !DILexicalBlock(scope: !10, file: !1, line: 3, column: 9)
+!20 = !DILocation(line: 3, column: 9, scope: !10)
+!21 = !DILocation(line: 4, column: 20, scope: !22)
+!22 = distinct !DILexicalBlock(scope: !19, file: !1, line: 3, column: 17)
+!23 = !DILocalVariable(name: "a1", scope: !22, file: !1, line: 4, type: !13)
+!24 = !DILocation(line: 0, scope: !22)
+!25 = !DILocation(line: 5, column: 21, scope: !22)
+!26 = !DILocalVariable(name: "a2", scope: !22, file: !1, line: 5, type: !13)
+!27 = !DILocalVariable(name: "ret", scope: !10, file: !1, line: 2, type: !13)
+!28 = !DILocation(line: 7, column: 5, scope: !22)
+!29 = !DILocation(line: 8, column: 20, scope: !30)
+!30 = distinct !DILexicalBlock(scope: !19, file: !1, line: 7, column: 12)
+!31 = !DILocalVariable(name: "b1", scope: !30, file: !1, line: 8, type: !13)
+!32 = !DILocation(line: 0, scope: !30)
+!33 = !DILocation(line: 9, column: 21, scope: !30)
+!34 = !DILocalVariable(name: "b2", scope: !30, file: !1, line: 9, type: !13)
+!35 = !DILocation(line: 0, scope: !19)
+!36 = !DILocation(line: 12, column: 5, scope: !10)
\ No newline at end of file
|
Apochens
changed the title
|
I'm not willing to merge any changes to this pass before the non-determinism issue is fixed, so we can actually test things without fear of breaking the build. |
|
@nikic Hi, I think I have a solution for the non-determinism issue. This issue is caused by unsigned GVNSink::sinkBB(BasicBlock *BBEnd) {
...
SmallVector<BasicBlock *, 4> Preds;
for (auto *B : predecessors(BBEnd)) {
auto *T = B->getTerminator();
if (isa<BranchInst>(T) || isa<SwitchInst>(T))
Preds.push_back(B);
else
return 0;
}
...
llvm::sort(Preds);
...
LockstepReverseIterator LRI(Preds);
...
while (LRI.isValid()) {
auto Cand = analyzeInstructionForSinking(LRI, InstNum, MemoryInstNum,
NeededPHIs, PHIContents);
...
}
...
for (unsigned I = 0; I < C.NumInstructions; ++I)
sinkLastInstruction(C.Blocks, InsertBB);
return C.NumInstructions;
}std::optional<SinkingInstructionCandidate>
GVNSink::analyzeInstructionForSinking(LockstepReverseIterator &LRI, ...) {
...
auto &ActivePreds = LRI.getActiveBlocks();
...
SinkingInstructionCandidate Cand;
append_range(Cand.Blocks, ActivePreds);
return Cand;
}So the order of basic blocks in function I think this issue can be solved by sorting BBs according to their names, instead of their memory addresses. |
hiraditya
force-pushed
the
#77415_gvnsink_opt_failure
branch
from
1e7de06
to
0ea7153
Compare
|
Hopefully this passes as non-determinism issues were addressed recently. Thanks for fixing the bug with dbg info. |
|
@Apochens do you mind updating the patch, there is a test failure. I think you just need to update the check directives. |
|
@Apochens OK, I'll solve this! Thanks! |
| ; CHECK: [[SINK:%.*.sink]] = phi i32 [ %a, %if.then ], [ %b, %if.else ] | ||
| ; CHECK: %add = add nsw i32 [[SINK]], 1 | ||
| ; CHECK: %xor = xor i32 %add, 1 | ||
| ; CHECK: [[SINK:%.*sink]] = phi i32 [ {{.*}} ], [ {{.*}} ] |
There was a problem hiding this comment.
nit: we should have a fixed order here.
maybe swap the PHI operands and see if that passed the check.
There was a problem hiding this comment.
or better use the llvm-project/llvm/utils/update_test_checks.py to update the test
There was a problem hiding this comment.
Thanks! Now the incoming blocks has a fixed order.
|
@hiraditya I don't have the permission to merge this PR. So, if the PR is ready, could you please help in merging it? Then, I can continue to fix the misleading debug location update. |
This PR fixes issue #77415 and is revised from PR #77419 .
PR #77419 breaks the newly added test in the same PR on windows, because GVNSink is non-deterministic when sorting
BasicBlock*pointers. This is reflected in the failure report.According to the report, what the CheckFile wants to match is the
%a.sink, however there is%b.sink. But this mismatch does not mean that this commit is wrong, since the occurrence of either%a.sinkor%b.sinkis correct. The root cause of this test failure is the strict check rule in the regression test committed.So I refined the regression test with a more general check rule to only detect whether there is an instruction with suffix
.sinkin theif.endblock. Hope this won't fail the test. If this PR still fails to build, I will close this PR and try to find another right way to fix this.