[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();
----------------
MaheshRavishankar wrote:

I dont see any benefit in not flagging in the verifier when a tensor.dim is accessed with index greater than the tensor dim (or negative dim). In dynamic index cases you cant do much, its undefined behavior. If you know this statically though its better to flag in the verifier closer to when it is created, rather than letting it plumb through compilation. In all probability this will lead to a runtime error which would be hard to track down. Flagging it in the verifier on creation is better to diagnose errors since I cant think of a valid reason one would create a `tensor.dim` operation with negative index or index out of bounds.

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