[Mlir-commits] [mlir] [MLIR][Tensor] Fix out-of-bounds FoldEmptyTensorWithDimOp crash #111270 (PR #112196)

[llvmlistbot at llvm.org]

================
@@ -979,7 +979,9 @@ struct FoldEmptyTensorWithDimOp : public OpRewritePattern<DimOp> {
     auto emptyTensorOp = dimOp.getSource().getDefiningOp<EmptyOp>();
     if (!emptyTensorOp || !maybeConstantIndex)
       return failure();
-    if (!emptyTensorOp.getType().isDynamicDim(*maybeConstantIndex))
+    auto emptyTensorType = emptyTensorOp.getType();
----------------
brod4910 wrote:

Hi @MaheshRavishankar thanks for your review!

Here is the original repro.

`mlir-opt --linalg-fold-unit-extent-dims a.mlir`

a.mlir:

```mlir
func.func @func1() -> index {
  %3 = tensor.empty() : tensor<19x15x1xf16>
  %idx10 = index.constant 10
  %dim_162 = tensor.dim %3, %idx10 : tensor<19x15x1xf16>  // index out-of-bounds
  return %dim_162 : index
}
```

>From my understanding, OOBs `tensor.dim` operations result in UB but that doesn't mean that the IR is invalid (or does it?). The crash here is unexpected due to the assertion check OOBs accesses in`isDynamicDim`.

https://github.com/llvm/llvm-project/pull/112196