[Mlir-commits] [mlir] [mlir][cf] Do not access erased operation in `cf.cond_br` lowering (PR #148358)

llvmlistbot at llvm.org llvmlistbot at llvm.org
Sat Jul 12 04:08:09 PDT 2025


llvmbot wrote:


<!--LLVM PR SUMMARY COMMENT-->

@llvm/pr-subscribers-mlir

Author: Matthias Springer (matthias-springer)

<details>
<summary>Changes</summary>

Do not access the erased `cf.cond_br` operation in the lowering pattern. That won't work anymore in a One-Shot Dialect Conversion and triggers a use-after-free sanitizer error.

After the One-Shot Dialect Conversion refactoring, a `ConversionPatternRewriter` will behave more like a normal `PatternRewriter`.


---
Full diff: https://github.com/llvm/llvm-project/pull/148358.diff


1 Files Affected:

- (modified) mlir/lib/Conversion/ControlFlowToLLVM/ControlFlowToLLVM.cpp (+7-10) 


``````````diff
diff --git a/mlir/lib/Conversion/ControlFlowToLLVM/ControlFlowToLLVM.cpp b/mlir/lib/Conversion/ControlFlowToLLVM/ControlFlowToLLVM.cpp
index 88a8b7fb185c5..716ab140ad6c7 100644
--- a/mlir/lib/Conversion/ControlFlowToLLVM/ControlFlowToLLVM.cpp
+++ b/mlir/lib/Conversion/ControlFlowToLLVM/ControlFlowToLLVM.cpp
@@ -138,11 +138,12 @@ struct BranchOpLowering : public ConvertOpToLLVMPattern<cf::BranchOp> {
                           TypeRange(adaptor.getOperands()));
     if (failed(convertedBlock))
       return failure();
+    auto attrs = op->getAttrDictionary();
     Operation *newOp = rewriter.replaceOpWithNewOp<LLVM::BrOp>(
         op, adaptor.getOperands(), *convertedBlock);
     // TODO: We should not just forward all attributes like that. But there are
     // existing Flang tests that depend on this behavior.
-    newOp->setAttrs(op->getAttrDictionary());
+    newOp->setAttrs(attrs);
     return success();
   }
 };
@@ -166,18 +167,14 @@ struct CondBranchOpLowering : public ConvertOpToLLVMPattern<cf::CondBranchOp> {
                           TypeRange(adaptor.getFalseDestOperands()));
     if (failed(convertedFalseBlock))
       return failure();
+    auto attrs = op->getAttrDictionary();
     auto newOp = rewriter.replaceOpWithNewOp<LLVM::CondBrOp>(
-        op, adaptor.getCondition(), *convertedTrueBlock,
-        adaptor.getTrueDestOperands(), *convertedFalseBlock,
-        adaptor.getFalseDestOperands());
-    ArrayRef<int32_t> weights = op.getWeights();
-    if (!weights.empty()) {
-      newOp.setWeights(weights);
-      op.removeBranchWeightsAttr();
-    }
+        op, adaptor.getCondition(), adaptor.getTrueDestOperands(),
+        adaptor.getFalseDestOperands(), op.getBranchWeightsAttr(),
+        *convertedTrueBlock, *convertedFalseBlock);
     // TODO: We should not just forward all attributes like that. But there are
     // existing Flang tests that depend on this behavior.
-    newOp->setAttrs(op->getAttrDictionary());
+    newOp->setAttrs(attrs);
     return success();
   }
 };

``````````

</details>


https://github.com/llvm/llvm-project/pull/148358


More information about the Mlir-commits mailing list