[PATCH] D82227: SLP: honor requested max vector size merging PHIs

[Alexey Bataev via Phabricator via llvm-commits]

ABataev added inline comments.


================
Comment at: llvm/test/Transforms/SLPVectorizer/slp-max-phi-size.ll:167-168
+; MAX128-NEXT:    [[I37:%.*]] = fadd float 0.000000e+00, [[I36]]
+; MAX128-NEXT:    [[I38:%.*]] = fmul float [[I3]], [[FVAL]]
+; MAX128-NEXT:    [[I39:%.*]] = fadd float 0.000000e+00, [[I38]]
+; MAX128-NEXT:    [[I40:%.*]] = fmul float [[I6]], [[FVAL]]
----------------
It really makes the vectorization worse, in general. Most of these inserts/extracts will be transformed into the simple shuffles by the instcombiner. And if there is really a problem with target-specific limitations, it is better to adapt the cost model rather than introduce changes that may affect all targets. Maybe need to fix `TTI::getTypeLegalizationCost`?


CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D82227/new/

https://reviews.llvm.org/D82227