[llvm] [LV] Add support for widening loads/stores to a VF multiple (PR #217670)
Paul Walker via llvm-commits
llvm-commits at lists.llvm.org
Wed Sep 16 03:19:41 PDT 2026
paulwalker-arm wrote:
What does this PR give us that is not already possible? or achievable elsewhere? Perhaps I'm missing a nuance but this just looks like a very complex way of grouping a set of N consecutive loads into one? If true then isn't that a combine rather than something LoopVectorize should be concern with? Even when expanded to masked_loads the problem does not changes much. Sure for our exact use case we'll need to look through a PHI to ensure the mask is generated as we need but that still feels like a post-vectorisation problem.
For non-tail-folded cases picking are larger VF just works today and code generator changes is all that's required to support the interleave case. The latter part is likely required anyway for non-loop-vectorize use cases.
I guess I'm just asking if the cost is worth the gain?
https://github.com/llvm/llvm-project/pull/217670
More information about the llvm-commits
mailing list