[Mlir-commits] [mlir] [mlir][VectorToLLVM] emit `inbounds|nuw` GEP flags when lowering `vector.load/store` (PR #202118)

[Andrzej Warzyński]

banach-space wrote:

> The missing flags have two concrete downstream effects:
> 
> * Index arithmetic (`llvm.mul/llvm.add`): with no `overflow<nsw,nuw>` annotation SCEV cannot prove the induction-variable computation is monotone, thus loop auto-vectorizer bails out.
> * GEP: with no `inbounds|nuw`the BasicAliasAnalysis falls back to conservative aliasing, thus further vectorization and alias-based optimizations are blocked.

Why would this matter if MLIR already generates vectorized code? Genuine question.

https://github.com/llvm/llvm-project/pull/202118