[PATCH] D132892: [LV] Explicily lower uniform load as single load

[Florian Hahn via Phabricator via llvm-commits]

fhahn added reviewers: Ayal, gilr.
fhahn added a comment.

> This replaces the "scalarize then clone-as-uniform" lowering for loads of uniform values with a dedicated mode of memory widening for this case. This avoids needing to relying on instcombine/GVN to clean up after redundant loads.

IIUC the main issue is that `VPReplicateRecipe::isUniform` means 'uniform-per-part/uniform-per-VF', not 'uniform-across-all-parts'.  Unfortunately the terminology is not really used very consistently across different parts of LV. We should be able to catch some cases through improving lowering for uniform VPReplicateRecipes directly:  D133019 <https://reviews.llvm.org/D133019>.

If we want to use `isUniformMemOp` to determine the `uniform-across-all-parts` property, I think it would make sense to extend `VPReplicateRecipe` to distinguish between our 2 versions of uniformity. AFAIK `VPWidenXXX` naming scheme is meant to indicate that the recipe will be widened (i.e. a wide vector instruction will be generated), but for uniform loads we only generate a single scalar instruction).


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D132892/new/

https://reviews.llvm.org/D132892