release/20.x: [VectorCombine] scalarizeLoadExtract - don't create scalar loads if any extract is waiting to be erased (#129375)

[llvmbot]

Backport 5ddf40f

Requested by: @RKSimon

[llvmbot]

@RKSimon What do you think about merging this PR to the release branch?

[llvmbot]

@llvm/pr-subscribers-vectorizers

@llvm/pr-subscribers-llvm-transforms

Author: None (llvmbot)

Changes

Backport 5ddf40f

Requested by: @RKSimon

---

Full diff: https://github.com/llvm/llvm-project/pull/129595.diff

2 Files Affected:

• (modified) llvm/lib/Transforms/Vectorize/VectorCombine.cpp (+5)
• (modified) llvm/test/Transforms/VectorCombine/X86/load-extractelement-scalarization.ll (+12)

diff --git a/llvm/lib/Transforms/Vectorize/VectorCombine.cpp b/llvm/lib/Transforms/Vectorize/VectorCombine.cpp
index 59920b5a4dd20..1deaaca05d98f 100644
--- a/llvm/lib/Transforms/Vectorize/VectorCombine.cpp
+++ b/llvm/lib/Transforms/Vectorize/VectorCombine.cpp
@@ -1403,6 +1403,11 @@ bool VectorCombine::scalarizeLoadExtract(Instruction &I) {
     if (!UI || UI->getParent() != LI->getParent())
       return false;
 
+    // If any extract is waiting to be erased, then bail out as this will
+    // distort the cost calculation and possibly lead to infinite loops.
+    if (UI->use_empty())
+      return false;
+
     // Check if any instruction between the load and the extract may modify
     // memory.
     if (LastCheckedInst->comesBefore(UI)) {
diff --git a/llvm/test/Transforms/VectorCombine/X86/load-extractelement-scalarization.ll b/llvm/test/Transforms/VectorCombine/X86/load-extractelement-scalarization.ll
index 0acfeccb92ef7..d46c8c0de4037 100644
--- a/llvm/test/Transforms/VectorCombine/X86/load-extractelement-scalarization.ll
+++ b/llvm/test/Transforms/VectorCombine/X86/load-extractelement-scalarization.ll
@@ -24,3 +24,15 @@ define void @multiple_extract(ptr %p) {
   store i32 %e1, ptr %p1, align 4
   ret void
 }
+
+; infinite loop if we fold an extract that is waiting to be erased
+define void @unused_extract(ptr %p) {
+; CHECK-LABEL: @unused_extract(
+; CHECK-NEXT:    ret void
+;
+  %load = load <4 x float>, ptr %p, align 8
+  %shuffle0 = shufflevector <4 x float> zeroinitializer, <4 x float> %load, <4 x i32> <i32 0, i32 4, i32 1, i32 5>
+  %shuffle1 = shufflevector <4 x float> %shuffle0, <4 x float> zeroinitializer, <4 x i32> <i32 0, i32 4, i32 poison, i32 poison>
+  %extract = extractelement <4 x float> %load, i64 1
+  ret void
+}

[github-actions]

@RKSimon (or anyone else). If you would like to add a note about this fix in the release notes (completely optional). Please reply to this comment with a one or two sentence description of the fix. When you are done, please add the release:note label to this PR.