[Mlir-commits] [mlir] [MLIR] Fix use-after-scope when interchanging ploops (PR #196076)

[llvmlistbot at llvm.org]

llvmorg-github-actions[bot] wrote:


<!--LLVM PR SUMMARY COMMENT-->

@llvm/pr-subscribers-mlir-scf

Author: Benjamin Kramer (d0k)

<details>
<summary>Changes</summary>

getInductionVars returns a SmallVector, so going through zip+reverse gets us a dangling reference. Quite a footgun.

Found by asan.

---
Full diff: https://github.com/llvm/llvm-project/pull/196076.diff


1 Files Affected:

- (modified) mlir/lib/Dialect/SCF/Transforms/ParallelLoopFusion.cpp (+3-2) 


``````````diff
diff --git a/mlir/lib/Dialect/SCF/Transforms/ParallelLoopFusion.cpp b/mlir/lib/Dialect/SCF/Transforms/ParallelLoopFusion.cpp
index cbbe45a00163f..bdaade59e99df 100644
--- a/mlir/lib/Dialect/SCF/Transforms/ParallelLoopFusion.cpp
+++ b/mlir/lib/Dialect/SCF/Transforms/ParallelLoopFusion.cpp
@@ -748,8 +748,9 @@ static std::optional<ParallelOp> interchangeLoops(OpBuilder &builder,
                                   loop.getUpperBound(), loop.getStep(),
                                   loop.getInitVals(), nullptr);
   IRMapping mapping;
-  for (auto [iv, riv] : llvm::zip(loop.getInductionVars(),
-                                  llvm::reverse(newOp.getInductionVars()))) {
+  auto ivs = loop.getInductionVars();
+  auto newIvs = newOp.getInductionVars();
+  for (auto [iv, riv] : llvm::zip(ivs, llvm::reverse(newIvs))) {
     mapping.map(iv, riv);
   }
   // Copy parallel loop body

``````````

</details>


https://github.com/llvm/llvm-project/pull/196076