From be4c2933a2c370d292122fc2187d75436aeaea83 Mon Sep 17 00:00:00 2001 From: Craig Topper Date: Fri, 6 Jul 2018 07:14:47 +0000 Subject: [PATCH] [X86] Implement _builtin_ia32_vfmaddss and _builtin_ia32_vfmaddsd with native IR using llvm.fma intrinsic. This generates some extra zeroing currently, but we should be able to quickly address that with some isel patterns. llvm-svn: 336417 --- clang/lib/CodeGen/CGBuiltin.cpp | 10 ++++++++ clang/test/CodeGen/fma4-builtins.c | 48 +++++++++++++++++++++++++++++++------- 2 files changed, 50 insertions(+), 8 deletions(-) diff --git a/clang/lib/CodeGen/CGBuiltin.cpp b/clang/lib/CodeGen/CGBuiltin.cpp index 56cf6a9..f2efca8c 100644 --- a/clang/lib/CodeGen/CGBuiltin.cpp +++ b/clang/lib/CodeGen/CGBuiltin.cpp @@ -9134,6 +9134,16 @@ Value *CodeGenFunction::EmitX86BuiltinExpr(unsigned BuiltinID, Value *Res = Builder.CreateCall(FMA, {A, B, C} ); return Builder.CreateInsertElement(Ops[0], Res, (uint64_t)0); } + case X86::BI__builtin_ia32_vfmaddss: + case X86::BI__builtin_ia32_vfmaddsd: { + Value *A = Builder.CreateExtractElement(Ops[0], (uint64_t)0); + Value *B = Builder.CreateExtractElement(Ops[1], (uint64_t)0); + Value *C = Builder.CreateExtractElement(Ops[2], (uint64_t)0); + Function *FMA = CGM.getIntrinsic(Intrinsic::fma, A->getType()); + Value *Res = Builder.CreateCall(FMA, {A, B, C} ); + Value *Zero = Constant::getNullValue(Ops[0]->getType()); + return Builder.CreateInsertElement(Zero, Res, (uint64_t)0); + } case X86::BI__builtin_ia32_vfmaddps: case X86::BI__builtin_ia32_vfmaddpd: case X86::BI__builtin_ia32_vfmaddps256: diff --git a/clang/test/CodeGen/fma4-builtins.c b/clang/test/CodeGen/fma4-builtins.c index 8d88a51..8078a6e 100644 --- a/clang/test/CodeGen/fma4-builtins.c +++ b/clang/test/CodeGen/fma4-builtins.c @@ -17,13 +17,21 @@ __m128d test_mm_macc_pd(__m128d a, __m128d b, __m128d c) { __m128 test_mm_macc_ss(__m128 a, __m128 b, __m128 c) { // CHECK-LABEL: test_mm_macc_ss - // CHECK: @llvm.x86.fma4.vfmadd.ss + // CHECK: extractelement <4 x float> %{{.*}}, i64 0 + // CHECK: extractelement <4 x float> %{{.*}}, i64 0 + // CHECK: extractelement <4 x float> %{{.*}}, i64 0 + // CHECK: call float @llvm.fma.f32(float %{{.*}}, float %{{.*}}, float %{{.*}}) + // CHECK: insertelement <4 x float> zeroinitializer, float %{{.*}}, i64 0 return _mm_macc_ss(a, b, c); } __m128d test_mm_macc_sd(__m128d a, __m128d b, __m128d c) { // CHECK-LABEL: test_mm_macc_sd - // CHECK: @llvm.x86.fma4.vfmadd.sd + // CHECK: extractelement <2 x double> %{{.*}}, i64 0 + // CHECK: extractelement <2 x double> %{{.*}}, i64 0 + // CHECK: extractelement <2 x double> %{{.*}}, i64 0 + // CHECK: call double @llvm.fma.f64(double %{{.*}}, double %{{.*}}, double %{{.*}}) + // CHECK: insertelement <2 x double> zeroinitializer, double %{{.*}}, i64 0 return _mm_macc_sd(a, b, c); } @@ -44,14 +52,22 @@ __m128d test_mm_msub_pd(__m128d a, __m128d b, __m128d c) { __m128 test_mm_msub_ss(__m128 a, __m128 b, __m128 c) { // CHECK-LABEL: test_mm_msub_ss // CHECK: [[NEG:%.+]] = fsub <4 x float> , %{{.+}} - // CHECK: @llvm.x86.fma4.vfmadd.ss(<4 x float> %{{.+}}, <4 x float> %{{.+}}, <4 x float> [[NEG]]) + // CHECK: extractelement <4 x float> %{{.*}}, i64 0 + // CHECK: extractelement <4 x float> %{{.*}}, i64 0 + // CHECK: [[C:%.+]] = extractelement <4 x float> [[NEG]], i64 0 + // CHECK: call float @llvm.fma.f32(float %{{.*}}, float %{{.*}}, float [[C]]) + // CHECK: insertelement <4 x float> zeroinitializer, float %{{.*}}, i64 0 return _mm_msub_ss(a, b, c); } __m128d test_mm_msub_sd(__m128d a, __m128d b, __m128d c) { // CHECK-LABEL: test_mm_msub_sd // CHECK: [[NEG:%.+]] = fsub <2 x double> , %{{.+}} - // CHECK: @llvm.x86.fma4.vfmadd.sd(<2 x double> %{{.+}}, <2 x double> %{{.+}}, <2 x double> [[NEG]]) + // CHECK: extractelement <2 x double> %{{.*}}, i64 0 + // CHECK: extractelement <2 x double> %{{.*}}, i64 0 + // CHECK: [[C:%.+]] = extractelement <2 x double> [[NEG]], i64 0 + // CHECK: call double @llvm.fma.f64(double %{{.*}}, double %{{.*}}, double [[C]]) + // CHECK: insertelement <2 x double> zeroinitializer, double %{{.*}}, i64 0 return _mm_msub_sd(a, b, c); } @@ -72,14 +88,22 @@ __m128d test_mm_nmacc_pd(__m128d a, __m128d b, __m128d c) { __m128 test_mm_nmacc_ss(__m128 a, __m128 b, __m128 c) { // CHECK-LABEL: test_mm_nmacc_ss // CHECK: [[NEG:%.+]] = fsub <4 x float> , %{{.+}} - // CHECK: @llvm.x86.fma4.vfmadd.ss(<4 x float> [[NEG]], <4 x float> %{{.+}}, <4 x float> %{{.+}}) + // CHECK: [[A:%.+]] = extractelement <4 x float> [[NEG]], i64 0 + // CHECK: extractelement <4 x float> %{{.*}}, i64 0 + // CHECK: extractelement <4 x float> %{{.*}}, i64 0 + // CHECK: call float @llvm.fma.f32(float [[A]], float %{{.*}}, float %{{.*}}) + // CHECK: insertelement <4 x float> zeroinitializer, float %{{.*}}, i64 0 return _mm_nmacc_ss(a, b, c); } __m128d test_mm_nmacc_sd(__m128d a, __m128d b, __m128d c) { // CHECK-LABEL: test_mm_nmacc_sd // CHECK: [[NEG:%.+]] = fsub <2 x double> , %{{.+}} - // CHECK: @llvm.x86.fma4.vfmadd.sd(<2 x double> [[NEG]], <2 x double> %{{.+}}, <2 x double> %{{.+}}) + // CHECK: [[A:%.+]] = extractelement <2 x double> [[NEG]], i64 0 + // CHECK: extractelement <2 x double> %{{.*}}, i64 0 + // CHECK: extractelement <2 x double> %{{.*}}, i64 0 + // CHECK: call double @llvm.fma.f64(double [[A]], double %{{.*}}, double %{{.*}}) + // CHECK: insertelement <2 x double> zeroinitializer, double %{{.*}}, i64 0 return _mm_nmacc_sd(a, b, c); } @@ -103,7 +127,11 @@ __m128 test_mm_nmsub_ss(__m128 a, __m128 b, __m128 c) { // CHECK-LABEL: test_mm_nmsub_ss // CHECK: [[NEG:%.+]] = fsub <4 x float> , %{{.+}} // CHECK: [[NEG2:%.+]] = fsub <4 x float> , %{{.+}} - // CHECK: @llvm.x86.fma4.vfmadd.ss(<4 x float> [[NEG]], <4 x float> %{{.+}}, <4 x float> [[NEG2]]) + // CHECK: [[A:%.+]] = extractelement <4 x float> [[NEG]], i64 0 + // CHECK: extractelement <4 x float> %{{.*}}, i64 0 + // CHECK: [[C:%.+]] = extractelement <4 x float> [[NEG2]], i64 0 + // CHECK: call float @llvm.fma.f32(float [[A]], float %{{.*}}, float [[C]]) + // CHECK: insertelement <4 x float> zeroinitializer, float %{{.*}}, i64 0 return _mm_nmsub_ss(a, b, c); } @@ -111,7 +139,11 @@ __m128d test_mm_nmsub_sd(__m128d a, __m128d b, __m128d c) { // CHECK-LABEL: test_mm_nmsub_sd // CHECK: [[NEG:%.+]] = fsub <2 x double> , %{{.+}} // CHECK: [[NEG2:%.+]] = fsub <2 x double> , %{{.+}} - // CHECK: @llvm.x86.fma4.vfmadd.sd(<2 x double> [[NEG]], <2 x double> %{{.+}}, <2 x double> [[NEG2]]) + // CHECK: [[A:%.+]] = extractelement <2 x double> [[NEG]], i64 0 + // CHECK: extractelement <2 x double> %{{.*}}, i64 0 + // CHECK: [[C:%.+]] = extractelement <2 x double> [[NEG2]], i64 0 + // CHECK: call double @llvm.fma.f64(double [[A]], double %{{.*}}, double [[C]]) + // CHECK: insertelement <2 x double> zeroinitializer, double %{{.*}}, i64 0 return _mm_nmsub_sd(a, b, c); } -- 2.7.4