[X86][AVX] Fixed v16i16/v32i8 ADD/SUB costs on AVX1 subtargets
authorSimon Pilgrim <llvm-dev@redking.me.uk>
Mon, 14 Nov 2016 14:45:16 +0000 (14:45 +0000)
committerSimon Pilgrim <llvm-dev@redking.me.uk>
Mon, 14 Nov 2016 14:45:16 +0000 (14:45 +0000)
commit27fed8e5d636d67ed5e2dff77705dcae1fcd0b15
tree13c14878cb5ff358c251334e2ec8c047ddb6f258
parente784395c99db265d8e7d358b1e03b1240abdad72
[X86][AVX] Fixed v16i16/v32i8 ADD/SUB costs on AVX1 subtargets

Add explicit v16i16/v32i8 ADD/SUB costs, matching the costs of v4i64/v8i32 - they were missing for some reason.

This has side effects on the LV max bandwidth tests (AVX1 now prefers 128-bit vectors vs AVX2 which still prefers 256-bit)

llvm-svn: 286832
llvm/lib/Target/X86/X86TargetTransformInfo.cpp
llvm/test/Analysis/CostModel/X86/arith.ll
llvm/test/Transforms/LoopVectorize/X86/vector_max_bandwidth.ll