[CostModel][X86] Move AVX512/AVX2 uniform shift costs into the generic uniform cost...
authorSimon Pilgrim <llvm-dev@redking.me.uk>
Mon, 12 Sep 2022 11:08:42 +0000 (12:08 +0100)
committerSimon Pilgrim <llvm-dev@redking.me.uk>
Mon, 12 Sep 2022 11:08:42 +0000 (12:08 +0100)
commitbd0109f39243e12f494e53bc99e95d76a1585ef8
tree7880dc5c59a901f0e5ea30ca76fd976bd93c3eb5
parentcde2736cdf960e94b27115ef2827bb5f770e6ab7
[CostModel][X86] Move AVX512/AVX2 uniform shift costs into the generic uniform cost tables

They shouldn't be happening after XOP shift costs - AVX2 shift supports takes preference over XOP for everything but vXi8 shifts - the improvement is pretty limited as it only affects bdver4 targets but it does help clean up a fraction of the messy shift cost logic....
llvm/lib/Target/X86/X86TargetTransformInfo.cpp
llvm/test/Analysis/CostModel/X86/vshift-ashr-cost-inseltpoison.ll
llvm/test/Analysis/CostModel/X86/vshift-ashr-cost.ll