[Mlir-commits] [mlir] [mlir][scf] Fix `for-loop-peeling` crash (PR #77697)

[Ivan Butygin]

Hardcode84 wrote:

> Good point, how could this be done with the PoisonAttr instead?

The easiest solution is just to remove corresponding check from verifier and check `scf.for` canonicalizers doesn't crash or do something unexpected in this case.

We can potentially drop the loop entirely and replace results with poisons, but it will introduce dependence on poison dialect and need some discussion, I think.

I've created an issue to track https://github.com/llvm/llvm-project/issues/77755

https://github.com/llvm/llvm-project/pull/77697