[PATCH] D101555: [SLP]Improve handling of compensate external uses cost.

[Alexey Bataev via Phabricator via llvm-commits]

ABataev added inline comments.


================
Comment at: llvm/test/Transforms/SLPVectorizer/X86/hsub.ll:174
+; SSE-NEXT:    [[R2:%.*]] = fsub double [[A2]], [[A3]]
+; SSE-NEXT:    [[R3:%.*]] = fsub double [[B2]], [[B3]]
+; SSE-NEXT:    [[TMP4:%.*]] = extractelement <2 x double> [[TMP3]], i32 0
----------------
RKSimon wrote:
> These regressions looks like we need to do more in the shuffle costs to recognise when the shuffles don't cross subvector boundaries? Either for illegal types like this or across 128-bit subvector boundaries on AVX.
Yes, need to subtract scalarization overhead for insertelement instruction, trying to handle it correctly in vectorization of InsertElement instructions patch. I'm going to abandon this patch when vectorization of InsertElements is landed. Keeping it just in case.


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D101555/new/

https://reviews.llvm.org/D101555