[PATCH] D124284: [SLP]Try partial store vectorization if supported by target.

[Dave Green via Phabricator via llvm-commits]

dmgreen added a comment.

I ran the same set of benchmarks again without issue this time (well, there was an issue, but it turned out that someone had changed the benchmark sources :) ). They might not be the most amazing SLP tests, but no remaining objections from me.

(The reason I suggested the truncate code the way I did was because by default the legalizing rules for smaller than legal power2 types is to promote integers to larger sizes. So under MVE where we only have 128bit vectors, a v4i8 vector will be promoted to a v4i32. We would then need a v4i32->v4i8 truncstore for it to be legal. Which it does have! So it would allow some of the smaller than legal types to be vectorized, essentially treating the v4i8 operations as v4i32's. If we are going for more conservative it might make sense to ignore that though, and float types will always widen as opposed to promote by default.)


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D124284/new/

https://reviews.llvm.org/D124284