[Mlir-commits] [mlir] [mlir][affine] Fix crash when rolling back vectorization of reduction loops (PR #226912)

[Mattéo Rizza Murgier]

matteo-rm wrote:

> _This is unrelated_: I probably need to study the super vectorizer in greater depth. But wouldn't it be possible to avoid the rollback entirely by performing a (low-overhead) check beforehand?

There is already a small check with `isVectorizableLoopBody`, but expanding it would probably require some refactoring to separate vectorizability logic from the actual transformation. This would solve leftover ops not being cleaned (though they can be removed by CSE).

https://github.com/llvm/llvm-project/pull/226912