[llvm] b4df0da - [AMDGPU] Fix a potential wrong return value indicating whether a pass modifies a function (#88197)

[via llvm-commits]

Author: Shilei Tian
Date: 2024-04-12T09:34:46-04:00
New Revision: b4df0da9e8421c8026bd09980b8a6acd7a6ce8c9

URL: https://github.com/llvm/llvm-project/commit/b4df0da9e8421c8026bd09980b8a6acd7a6ce8c9
DIFF: https://github.com/llvm/llvm-project/commit/b4df0da9e8421c8026bd09980b8a6acd7a6ce8c9.diff

LOG: [AMDGPU] Fix a potential wrong return value indicating whether a pass modifies a function (#88197)

When the alloca is too big for vectorization, the function could have
already
been modified in previous iteration of the `for` loop.

Added: 
    llvm/test/CodeGen/AMDGPU/half-alloca-promotion.ll

Modified: 
    llvm/lib/Target/AMDGPU/AMDGPUPromoteAlloca.cpp

Removed: 
    


################################################################################
diff  --git a/llvm/lib/Target/AMDGPU/AMDGPUPromoteAlloca.cpp b/llvm/lib/Target/AMDGPU/AMDGPUPromoteAlloca.cpp
index 6f3cdf54dceec7..c0846b123d1870 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPUPromoteAlloca.cpp
+++ b/llvm/lib/Target/AMDGPU/AMDGPUPromoteAlloca.cpp
@@ -336,7 +336,7 @@ bool AMDGPUPromoteAllocaImpl::run(Function &F, bool PromoteToLDS) {
     if (AllocaCost > VectorizationBudget) {
       LLVM_DEBUG(dbgs() << "  Alloca too big for vectorization: " << *AI
                         << "\n");
-      return false;
+      return Changed;
     }
 
     if (tryPromoteAllocaToVector(*AI)) {

diff  --git a/llvm/test/CodeGen/AMDGPU/half-alloca-promotion.ll b/llvm/test/CodeGen/AMDGPU/half-alloca-promotion.ll
new file mode 100644
index 00000000000000..cfec49f3652fbd
--- /dev/null
+++ b/llvm/test/CodeGen/AMDGPU/half-alloca-promotion.ll
@@ -0,0 +1,11 @@
+; RUN: opt -S -mtriple=amdgcn-amd-amdhsa -passes="amdgpu-promote-alloca-to-vector" -o - %s
+; We don't really need to check anything here because with expensive check, this
+; test case crashes. The correctness of the pass is beyond the scope.
+
+define fastcc void @foo() {
+entry:
+  %det = alloca [4 x i32], align 16, addrspace(5)
+  %trkltPosTmpYZ = alloca [2 x float], align 4, addrspace(5)
+  %trkltCovTmp = alloca [3 x float], align 4, addrspace(5)
+  ret void
+}