[SimplifyCFG] Preserve metadata when unconditionalizing branches (constant condition).

[Meinersbur]

Summary

When replacing a conditional branch by an unconditional one because the condition is a constant, transfer the metadata to the new branch instruction.

Part of fix for llvm.org/PR50060

[lebedev.ri]

The diff seems to be reverse from what you intended?

llvm/test/Transforms/SimplifyCFG/pr50060-constantfold-loopid.ll

1

Did you mean to update the test instead of deleting it

[Meinersbur]

Accidentally uploaded a reversed patch

[lebedev.ri]

llvm/lib/Transforms/Utils/Local.cpp

172

What about prof md?

[jeroen.dobbelaere]

llvm/lib/Transforms/Utils/Local.cpp

151

Shouldn't we also take care of the metadata here ?

166

Can't we just copy the metadata from the BI directly ? Or will removing predecessors modify the attached metadata ?

[Meinersbur]

llvm/lib/Transforms/Utils/Local.cpp

151

Yes, this patch is about the fixing PR50060 and I don't have a test case for this one.

166

Removing code bears the risk that instructions are invalidated and should not be accessed anymore. I am not sure enough about removePredecessor since it causes an orphaned block. However, NewBr could be created before removePredecessor.
However, BI->eraseFromParent() is also only called afterwards, so it should be fine.

172

Edge weights are not in the scope of this patch.

[Meinersbur]

• Use copyMetadata.

[Meinersbur]

llvm/lib/Transforms/Utils/Local.cpp

151

See D101226

[lebedev.ri]

llvm/lib/Transforms/Utils/Local.cpp

172

Indeed.
But my question is, this copies *all* metadata.
What happens if the cond br had prof md?
They won't really make sense on uncond br.
Will verifier be ok with that, or should this specify which md to copy?

[Meinersbur]

llvm/lib/Transforms/Utils/Local.cpp

172

Thanks for the hint. Looking at the pre-merge fails, the sanitizer check indeed complains about "Wrong number of operands". I did not think it was related, all the LLVM checks succeed. I could not find a common handling of metadata for branches such as llvm::copyMetadataForLoad for loads.

[Meinersbur]

• Transfer only selected metadata kinds.

[lebedev.ri]

This LGTM, thank you.

llvm/lib/Transforms/Utils/Local.cpp

172

I'm guessing you are intentionally not copying mustprocess?

[jeroen.dobbelaere]

llvm/lib/Transforms/Utils/Local.cpp

172

I'm guessing you are intentionally not copying mustprocess?

mustprogress is part of MD_loop.

llvm/lib/Transforms/Utils/Local.cpp

if (Dest2 == Dest1) {       // Conditional branch to same location?
       //     br bool %cond, label %Dest, label %Dest
       // and changes it into:  br label %Dest
 
       // Let the basic block know that we are letting go of one copy of it.
       assert(BI->getParent() && "Terminator not inserted in block!");
       Dest1->removePredecessor(BI->getParent());
 
       // Replace the conditional branch with an unconditional one.
       Builder.CreateBr(Dest1);
       Value *Cond = BI->getCondition();
       BI->eraseFromParent();
       if (DeleteDeadConditions)
         RecursivelyDeleteTriviallyDeadInstructions(Cond, TLI);
       return true;
     }
 
     if (auto *Cond = dyn_cast<ConstantInt>(BI->getCondition())) {
       // Are we branching on constant?
       // YES.  Change to unconditional branch...
       BasicBlock *Destination = Cond->getZExtValue() ? Dest1 : Dest2;
       BasicBlock *OldDest = Cond->getZExtValue() ? Dest2 : Dest1;
 
       // Let the basic block know that we are letting go of it.  Based on this,
       // it will adjust it's PHI nodes.
       OldDest->removePredecessor(BB);
 
       // Replace the conditional branch with an unconditional one.
-      Builder.CreateBr(Destination);
+      BranchInst *NewBI = Builder.CreateBr(Destination);
+
+      // Transfer the metadata to the new branch instruction.
+      NewBI->copyMetadata(*BI, {LLVMContext::MD_loop, LLVMContext::MD_dbg,
+                                LLVMContext::MD_annotation});
+
       BI->eraseFromParent();
       if (DTU)
         DTU->applyUpdates({{DominatorTree::Delete, BB, OldDest}});
       return true;
     }
 
     return false;
   }

llvm/test/Transforms/SimplifyCFG/pr50060-constantfold-loopid.ll

+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
+; RUN: opt -S -simplifycfg < %s | FileCheck %s
+;
+; The branch in the latch do.cond is conditional by a constant and the
+; conditional branch replaced by an unconditional one.
+; Ensure that the llvm.loop metadata is transferred to the new branch
+; (In for.cond.cleanup after further simplifications).
+;
+; llvm.org/PR50060
+;
+
+@n = dso_local global i32 0, align 4
+@C = dso_local global i32 0, align 4
+
+; Function Attrs: nounwind
+define dso_local void @_Z6test01v() addrspace(1) #0 {
+; CHECK-LABEL: @_Z6test01v(
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    [[J:%.*]] = alloca i32, align 4
+; CHECK-NEXT:    [[I:%.*]] = alloca i32, align 4
+; CHECK-NEXT:    br label [[DO_BODY:%.*]]
+; CHECK:       do.body:
+; CHECK-NEXT:    [[TMP0:%.*]] = load i32, i32* @C, align 4, !tbaa [[TBAA2:![0-9]+]]
+; CHECK-NEXT:    [[INC:%.*]] = add nsw i32 [[TMP0]], 1
+; CHECK-NEXT:    [[TMP1:%.*]] = bitcast i32* [[J]] to i8*
+; CHECK-NEXT:    call addrspace(1) void @llvm.lifetime.start.p0i8(i64 4, i8* [[TMP1]]) #[[ATTR2:[0-9]+]]
+; CHECK-NEXT:    store i32 0, i32* [[J]], align 4, !tbaa [[TBAA2]]
+; CHECK-NEXT:    br label [[FOR_COND:%.*]]
+; CHECK:       for.cond:
+; CHECK-NEXT:    [[TMP2:%.*]] = load i32, i32* [[J]], align 4, !tbaa [[TBAA2]]
+; CHECK-NEXT:    [[CMP:%.*]] = icmp slt i32 [[TMP2]], 3
+; CHECK-NEXT:    br i1 [[CMP]], label [[FOR_BODY:%.*]], label [[FOR_COND_CLEANUP:%.*]]
+; CHECK:       for.cond.cleanup:
+; CHECK-NEXT:    [[TMP3:%.*]] = bitcast i32* [[J]] to i8*
+; CHECK-NEXT:    call addrspace(1) void @llvm.lifetime.end.p0i8(i64 4, i8* [[TMP3]]) #[[ATTR2]]
+; CHECK-NEXT:    br label [[DO_BODY]], !llvm.loop [[LOOP6:![0-9]+]]
+; CHECK:       for.body:
+; CHECK-NEXT:    store i32 undef, i32* [[I]], align 4
+; CHECK-NEXT:    [[TMP4:%.*]] = bitcast i32* [[I]] to i8*
+; CHECK-NEXT:    call addrspace(1) void @llvm.lifetime.start.p0i8(i64 4, i8* [[TMP4]]) #[[ATTR2]]
+; CHECK-NEXT:    store i32 0, i32* [[I]], align 4, !tbaa [[TBAA2]]
+; CHECK-NEXT:    br label [[FOR_COND1:%.*]]
+; CHECK:       for.cond1:
+; CHECK-NEXT:    [[TMP5:%.*]] = load i32, i32* [[I]], align 4, !tbaa [[TBAA2]]
+; CHECK-NEXT:    [[TMP6:%.*]] = load i32, i32* @n, align 4, !tbaa [[TBAA2]]
+; CHECK-NEXT:    [[CMP2:%.*]] = icmp slt i32 [[TMP5]], [[TMP6]]
+; CHECK-NEXT:    br i1 [[CMP2]], label [[FOR_BODY4:%.*]], label [[FOR_COND_CLEANUP3:%.*]]
+; CHECK:       for.cond.cleanup3:
+; CHECK-NEXT:    [[TMP7:%.*]] = bitcast i32* [[I]] to i8*
+; CHECK-NEXT:    call addrspace(1) void @llvm.lifetime.end.p0i8(i64 4, i8* [[TMP7]]) #[[ATTR2]]
+; CHECK-NEXT:    [[TMP8:%.*]] = load i32, i32* [[J]], align 4, !tbaa [[TBAA2]]
+; CHECK-NEXT:    [[INC7:%.*]] = add nsw i32 [[TMP8]], 1
+; CHECK-NEXT:    store i32 [[INC7]], i32* [[J]], align 4, !tbaa [[TBAA2]]
+; CHECK-NEXT:    br label [[FOR_COND]], !llvm.loop [[LOOP8:![0-9]+]]
+; CHECK:       for.body4:
+; CHECK-NEXT:    [[TMP9:%.*]] = load i32, i32* [[I]], align 4, !tbaa [[TBAA2]]
+; CHECK-NEXT:    store volatile i32 [[TMP9]], i32* @C, align 4, !tbaa [[TBAA2]]
+; CHECK-NEXT:    [[TMP10:%.*]] = load i32, i32* [[I]], align 4, !tbaa [[TBAA2]]
+; CHECK-NEXT:    [[INC5:%.*]] = add nsw i32 [[TMP10]], 1
+; CHECK-NEXT:    store i32 [[INC5]], i32* [[I]], align 4, !tbaa [[TBAA2]]
+; CHECK-NEXT:    br label [[FOR_COND1]], !llvm.loop [[LOOP11:![0-9]+]]
+;
+entry:
+  %j = alloca i32, align 4
+  %i = alloca i32, align 4
+  br label %do.body
+
+do.body:                                          ; preds = %do.cond, %entry
+  %0 = load i32, i32* @C, align 4, !tbaa !2
+  %inc = add nsw i32 %0, 1
+  %1 = bitcast i32* %j to i8*
+  call addrspace(1) void @llvm.lifetime.start.p0i8(i64 4, i8* %1) #2
+  store i32 0, i32* %j, align 4, !tbaa !2
+  br label %for.cond
+
+for.cond:                                         ; preds = %for.inc6, %do.body
+  %2 = load i32, i32* %j, align 4, !tbaa !2
+  %cmp = icmp slt i32 %2, 3
+  br i1 %cmp, label %for.body, label %for.cond.cleanup
+
+for.cond.cleanup:                                 ; preds = %for.cond
+  %3 = bitcast i32* %j to i8*
+  call addrspace(1) void @llvm.lifetime.end.p0i8(i64 4, i8* %3) #2
+  br label %for.end8
+
+for.body:                                         ; preds = %for.cond
+  store i32 undef, i32* %i, align 4
+  %4 = bitcast i32* %i to i8*
+  call addrspace(1) void @llvm.lifetime.start.p0i8(i64 4, i8* %4) #2
+  store i32 0, i32* %i, align 4, !tbaa !2
+  br label %for.cond1
+
+for.cond1:                                        ; preds = %for.inc, %for.body
+  %5 = load i32, i32* %i, align 4, !tbaa !2
+  %6 = load i32, i32* @n, align 4, !tbaa !2
+  %cmp2 = icmp slt i32 %5, %6
+  br i1 %cmp2, label %for.body4, label %for.cond.cleanup3
+
+for.cond.cleanup3:                                ; preds = %for.cond1
+  %7 = bitcast i32* %i to i8*
+  call addrspace(1) void @llvm.lifetime.end.p0i8(i64 4, i8* %7) #2
+  br label %for.end
+
+for.body4:                                        ; preds = %for.cond1
+  %8 = load i32, i32* %i, align 4, !tbaa !2
+  store volatile i32 %8, i32* @C, align 4, !tbaa !2
+  br label %for.inc
+
+for.inc:                                          ; preds = %for.body4
+  %9 = load i32, i32* %i, align 4, !tbaa !2
+  %inc5 = add nsw i32 %9, 1
+  store i32 %inc5, i32* %i, align 4, !tbaa !2
+  br label %for.cond1, !llvm.loop !6
+
+for.end:                                          ; preds = %for.cond.cleanup3
+  br label %for.inc6
+
+for.inc6:                                         ; preds = %for.end
+  %10 = load i32, i32* %j, align 4, !tbaa !2
+  %inc7 = add nsw i32 %10, 1
+  store i32 %inc7, i32* %j, align 4, !tbaa !2
+  br label %for.cond, !llvm.loop !8
+
+for.end8:                                         ; preds = %for.cond.cleanup
+  br label %do.cond
+
+do.cond:                                          ; preds = %for.end8
+  br i1 true, label %do.body, label %do.end, !llvm.loop !10
+
+do.end:                                           ; preds = %do.cond
+  ret void
+}
+
+; Function Attrs: argmemonly nofree nosync nounwind willreturn
+declare void @llvm.lifetime.start.p0i8(i64 immarg, i8* nocapture) addrspace(1) #1
+
+; Function Attrs: argmemonly nofree nosync nounwind willreturn
+declare void @llvm.lifetime.end.p0i8(i64 immarg, i8* nocapture) addrspace(1) #1
+
+attributes #0 = { nounwind "disable-tail-calls"="false" "frame-pointer"="all" "less-precise-fpmad"="false" "min-legal-vector-width"="0" "no-infs-fp-math"="false" "no-jump-tables"="false" "no-nans-fp-math"="false" "no-signed-zeros-fp-math"="false" "no-trapping-math"="true" "stack-protector-buffer-size"="8" "unsafe-fp-math"="false" "use-soft-float"="false" }
+attributes #1 = { argmemonly nofree nosync nounwind willreturn }
+attributes #2 = { nounwind }
+
+!llvm.module.flags = !{!0}
+!llvm.ident = !{!1}
+
+!0 = !{i32 1, !"wchar_size", i32 4}
+!1 = !{!"clang)"}
+!2 = !{!3, !3, i64 0, i64 4}
+!3 = !{!4, i64 4, !"int"}
+!4 = !{!5, i64 1, !"omnipotent char"}
+!5 = !{!"Simple C++ TBAA"}
+!6 = distinct !{!6, !7}
+!7 = !{!"llvm.loop.mustprogress"}
+!8 = distinct !{!8, !7, !9}
+!9 = !{!"llvm.loop.unroll.disable"}
+!10 = distinct !{!10, !11}
+!11 = !{!"llvm.loop.unroll.count", i32 2}