[PATCH] D40865: X86 AVX2: Prefer one VPERMV over ShuffleAsRepeatedMaskAndLanePermute

[Dibyendu Das via Phabricator via llvm-commits]

ddibyend added a comment.

vpermps/vpermpd/vpermd have high cost in Ryzen. I see that this patch creates cases where a vperm* is introduced where one did not exist earlier.  That may cause slowdowns in Ryzen.


https://reviews.llvm.org/D40865