From d8333372bc764ea7575addf3dc9b2caaa39f178c Mon Sep 17 00:00:00 2001 From: Simon Pilgrim Date: Fri, 6 Jan 2017 11:12:53 +0000 Subject: [CostModel][X86] Fix 512-bit SDIV/UDIV 'big' costs. Set the costs on the lowest target that supports the type. llvm-svn: 291229 --- llvm/lib/Target/X86/X86TargetTransformInfo.cpp | 34 ++++++++++++++------------ 1 file changed, 18 insertions(+), 16 deletions(-) (limited to 'llvm/lib/Target/X86/X86TargetTransformInfo.cpp') diff --git a/llvm/lib/Target/X86/X86TargetTransformInfo.cpp b/llvm/lib/Target/X86/X86TargetTransformInfo.cpp index 80187a9fcc9..de4839432b9 100644 --- a/llvm/lib/Target/X86/X86TargetTransformInfo.cpp +++ b/llvm/lib/Target/X86/X86TargetTransformInfo.cpp @@ -226,12 +226,8 @@ int X86TTIImpl::getArithmeticInstrCost( // Vectorizing division is a bad idea. See the SSE2 table for more comments. { ISD::SDIV, MVT::v64i8, 64*20 }, { ISD::SDIV, MVT::v32i16, 32*20 }, - { ISD::SDIV, MVT::v16i32, 16*20 }, - { ISD::SDIV, MVT::v8i64, 8*20 }, { ISD::UDIV, MVT::v64i8, 64*20 }, - { ISD::UDIV, MVT::v32i16, 32*20 }, - { ISD::UDIV, MVT::v16i32, 16*20 }, - { ISD::UDIV, MVT::v8i64, 8*20 }, + { ISD::UDIV, MVT::v32i16, 32*20 } }; // Look for AVX512BW lowering tricks for custom cases. @@ -240,17 +236,23 @@ int X86TTIImpl::getArithmeticInstrCost( return LT.first * Entry->Cost; static const CostTblEntry AVX512CostTable[] = { - { ISD::SHL, MVT::v16i32, 1 }, - { ISD::SRL, MVT::v16i32, 1 }, - { ISD::SRA, MVT::v16i32, 1 }, - { ISD::SHL, MVT::v8i64, 1 }, - { ISD::SRL, MVT::v8i64, 1 }, - { ISD::SRA, MVT::v8i64, 1 }, - - { ISD::MUL, MVT::v32i8, 13 }, // extend/pmullw/trunc sequence. - { ISD::MUL, MVT::v16i8, 5 }, // extend/pmullw/trunc sequence. - { ISD::MUL, MVT::v16i32, 1 }, // pmulld - { ISD::MUL, MVT::v8i64, 8 } // 3*pmuludq/3*shift/2*add + { ISD::SHL, MVT::v16i32, 1 }, + { ISD::SRL, MVT::v16i32, 1 }, + { ISD::SRA, MVT::v16i32, 1 }, + { ISD::SHL, MVT::v8i64, 1 }, + { ISD::SRL, MVT::v8i64, 1 }, + { ISD::SRA, MVT::v8i64, 1 }, + + { ISD::MUL, MVT::v32i8, 13 }, // extend/pmullw/trunc sequence. + { ISD::MUL, MVT::v16i8, 5 }, // extend/pmullw/trunc sequence. + { ISD::MUL, MVT::v16i32, 1 }, // pmulld + { ISD::MUL, MVT::v8i64, 8 }, // 3*pmuludq/3*shift/2*add + + // Vectorizing division is a bad idea. See the SSE2 table for more comments. + { ISD::SDIV, MVT::v16i32, 16*20 }, + { ISD::SDIV, MVT::v8i64, 8*20 }, + { ISD::UDIV, MVT::v16i32, 16*20 }, + { ISD::UDIV, MVT::v8i64, 8*20 } }; if (ST->hasAVX512()) -- cgit v1.2.3