From 24008b27d5f4c12c43f88e6f4037f16ec43a8bc3 Mon Sep 17 00:00:00 2001 From: Craig Topper Date: Sat, 25 Nov 2017 19:32:12 +0000 Subject: [PATCH] [X86] Use separate builtins for fma4 scalar intrinsics. Use negations to remove some of the scalar fma3 builtins. fma4 instructions zero the upper bits of the xmm register. fma3 instructions leave the bits unmodified. This requires separate builtins for the different semantics. While we're cleaning up the scalar builtins this also removes the fma3 fmsub/fnmadd/fnmsub builtins by using negates in the header file. git-svn-id: https://llvm.org/svn/llvm-project/cfe/trunk@318985 91177308-0d34-0410-b5e6-96231b3b80d8 --- include/clang/Basic/BuiltinsX86.def | 12 ++++-------- lib/Headers/fma4intrin.h | 16 ++++++++-------- lib/Headers/fmaintrin.h | 12 ++++++------ test/CodeGen/fma-builtins.c | 20 ++++++++++++++------ test/CodeGen/fma4-builtins.c | 24 ++++++++++++++++-------- 5 files changed, 48 insertions(+), 36 deletions(-) diff --git a/include/clang/Basic/BuiltinsX86.def b/include/clang/Basic/BuiltinsX86.def index 8962e7a41c..b77033a0b3 100644 --- a/include/clang/Basic/BuiltinsX86.def +++ b/include/clang/Basic/BuiltinsX86.def @@ -682,14 +682,10 @@ TARGET_BUILTIN(__builtin_ia32_sha256msg2, "V4iV4iV4i", "", "sha") // FMA TARGET_BUILTIN(__builtin_ia32_vfmaddps, "V4fV4fV4fV4f", "", "fma|fma4") TARGET_BUILTIN(__builtin_ia32_vfmaddpd, "V2dV2dV2dV2d", "", "fma|fma4") -TARGET_BUILTIN(__builtin_ia32_vfmaddss3, "V4fV4fV4fV4f", "", "fma|fma4") -TARGET_BUILTIN(__builtin_ia32_vfmaddsd3, "V2dV2dV2dV2d", "", "fma|fma4") -TARGET_BUILTIN(__builtin_ia32_vfmsubss3, "V4fV4fV4fV4f", "", "fma|fma4") -TARGET_BUILTIN(__builtin_ia32_vfmsubsd3, "V2dV2dV2dV2d", "", "fma|fma4") -TARGET_BUILTIN(__builtin_ia32_vfnmaddss3, "V4fV4fV4fV4f", "", "fma|fma4") -TARGET_BUILTIN(__builtin_ia32_vfnmaddsd3, "V2dV2dV2dV2d", "", "fma|fma4") -TARGET_BUILTIN(__builtin_ia32_vfnmsubss3, "V4fV4fV4fV4f", "", "fma|fma4") -TARGET_BUILTIN(__builtin_ia32_vfnmsubsd3, "V2dV2dV2dV2d", "", "fma|fma4") +TARGET_BUILTIN(__builtin_ia32_vfmaddss3, "V4fV4fV4fV4f", "", "fma") +TARGET_BUILTIN(__builtin_ia32_vfmaddsd3, "V2dV2dV2dV2d", "", "fma") +TARGET_BUILTIN(__builtin_ia32_vfmaddss, "V4fV4fV4fV4f", "", "fma4") +TARGET_BUILTIN(__builtin_ia32_vfmaddsd, "V2dV2dV2dV2d", "", "fma4") TARGET_BUILTIN(__builtin_ia32_vfmaddsubps, "V4fV4fV4fV4f", "", "fma|fma4") TARGET_BUILTIN(__builtin_ia32_vfmaddsubpd, "V2dV2dV2dV2d", "", "fma|fma4") TARGET_BUILTIN(__builtin_ia32_vfmaddps256, "V8fV8fV8fV8f", "", "fma|fma4") diff --git a/lib/Headers/fma4intrin.h b/lib/Headers/fma4intrin.h index cda20e7d81..962b1a60a2 100644 --- a/lib/Headers/fma4intrin.h +++ b/lib/Headers/fma4intrin.h @@ -48,13 +48,13 @@ _mm_macc_pd(__m128d __A, __m128d __B, __m128d __C) static __inline__ __m128 __DEFAULT_FN_ATTRS _mm_macc_ss(__m128 __A, __m128 __B, __m128 __C) { - return (__m128)__builtin_ia32_vfmaddss3((__v4sf)__A, (__v4sf)__B, (__v4sf)__C); + return (__m128)__builtin_ia32_vfmaddss((__v4sf)__A, (__v4sf)__B, (__v4sf)__C); } static __inline__ __m128d __DEFAULT_FN_ATTRS _mm_macc_sd(__m128d __A, __m128d __B, __m128d __C) { - return (__m128d)__builtin_ia32_vfmaddsd3((__v2df)__A, (__v2df)__B, (__v2df)__C); + return (__m128d)__builtin_ia32_vfmaddsd((__v2df)__A, (__v2df)__B, (__v2df)__C); } static __inline__ __m128 __DEFAULT_FN_ATTRS @@ -72,13 +72,13 @@ _mm_msub_pd(__m128d __A, __m128d __B, __m128d __C) static __inline__ __m128 __DEFAULT_FN_ATTRS _mm_msub_ss(__m128 __A, __m128 __B, __m128 __C) { - return (__m128)__builtin_ia32_vfmsubss3((__v4sf)__A, (__v4sf)__B, (__v4sf)__C); + return (__m128)__builtin_ia32_vfmaddss((__v4sf)__A, (__v4sf)__B, -(__v4sf)__C); } static __inline__ __m128d __DEFAULT_FN_ATTRS _mm_msub_sd(__m128d __A, __m128d __B, __m128d __C) { - return (__m128d)__builtin_ia32_vfmsubsd3((__v2df)__A, (__v2df)__B, (__v2df)__C); + return (__m128d)__builtin_ia32_vfmaddsd((__v2df)__A, (__v2df)__B, -(__v2df)__C); } static __inline__ __m128 __DEFAULT_FN_ATTRS @@ -96,13 +96,13 @@ _mm_nmacc_pd(__m128d __A, __m128d __B, __m128d __C) static __inline__ __m128 __DEFAULT_FN_ATTRS _mm_nmacc_ss(__m128 __A, __m128 __B, __m128 __C) { - return (__m128)__builtin_ia32_vfnmaddss3((__v4sf)__A, (__v4sf)__B, (__v4sf)__C); + return (__m128)__builtin_ia32_vfmaddss(-(__v4sf)__A, (__v4sf)__B, (__v4sf)__C); } static __inline__ __m128d __DEFAULT_FN_ATTRS _mm_nmacc_sd(__m128d __A, __m128d __B, __m128d __C) { - return (__m128d)__builtin_ia32_vfnmaddsd3((__v2df)__A, (__v2df)__B, (__v2df)__C); + return (__m128d)__builtin_ia32_vfmaddsd(-(__v2df)__A, (__v2df)__B, (__v2df)__C); } static __inline__ __m128 __DEFAULT_FN_ATTRS @@ -120,13 +120,13 @@ _mm_nmsub_pd(__m128d __A, __m128d __B, __m128d __C) static __inline__ __m128 __DEFAULT_FN_ATTRS _mm_nmsub_ss(__m128 __A, __m128 __B, __m128 __C) { - return (__m128)__builtin_ia32_vfnmsubss3((__v4sf)__A, (__v4sf)__B, (__v4sf)__C); + return (__m128)__builtin_ia32_vfmaddss(-(__v4sf)__A, (__v4sf)__B, -(__v4sf)__C); } static __inline__ __m128d __DEFAULT_FN_ATTRS _mm_nmsub_sd(__m128d __A, __m128d __B, __m128d __C) { - return (__m128d)__builtin_ia32_vfnmsubsd3((__v2df)__A, (__v2df)__B, (__v2df)__C); + return (__m128d)__builtin_ia32_vfmaddsd(-(__v2df)__A, (__v2df)__B, -(__v2df)__C); } static __inline__ __m128 __DEFAULT_FN_ATTRS diff --git a/lib/Headers/fmaintrin.h b/lib/Headers/fmaintrin.h index 86a1198879..478a0ac81c 100644 --- a/lib/Headers/fmaintrin.h +++ b/lib/Headers/fmaintrin.h @@ -70,13 +70,13 @@ _mm_fmsub_pd(__m128d __A, __m128d __B, __m128d __C) static __inline__ __m128 __DEFAULT_FN_ATTRS _mm_fmsub_ss(__m128 __A, __m128 __B, __m128 __C) { - return (__m128)__builtin_ia32_vfmsubss3((__v4sf)__A, (__v4sf)__B, (__v4sf)__C); + return (__m128)__builtin_ia32_vfmaddss3((__v4sf)__A, (__v4sf)__B, -(__v4sf)__C); } static __inline__ __m128d __DEFAULT_FN_ATTRS _mm_fmsub_sd(__m128d __A, __m128d __B, __m128d __C) { - return (__m128d)__builtin_ia32_vfmsubsd3((__v2df)__A, (__v2df)__B, (__v2df)__C); + return (__m128d)__builtin_ia32_vfmaddsd3((__v2df)__A, (__v2df)__B, -(__v2df)__C); } static __inline__ __m128 __DEFAULT_FN_ATTRS @@ -94,13 +94,13 @@ _mm_fnmadd_pd(__m128d __A, __m128d __B, __m128d __C) static __inline__ __m128 __DEFAULT_FN_ATTRS _mm_fnmadd_ss(__m128 __A, __m128 __B, __m128 __C) { - return (__m128)__builtin_ia32_vfnmaddss3((__v4sf)__A, (__v4sf)__B, (__v4sf)__C); + return (__m128)__builtin_ia32_vfmaddss3((__v4sf)__A, -(__v4sf)__B, (__v4sf)__C); } static __inline__ __m128d __DEFAULT_FN_ATTRS _mm_fnmadd_sd(__m128d __A, __m128d __B, __m128d __C) { - return (__m128d)__builtin_ia32_vfnmaddsd3((__v2df)__A, (__v2df)__B, (__v2df)__C); + return (__m128d)__builtin_ia32_vfmaddsd3((__v2df)__A, -(__v2df)__B, (__v2df)__C); } static __inline__ __m128 __DEFAULT_FN_ATTRS @@ -118,13 +118,13 @@ _mm_fnmsub_pd(__m128d __A, __m128d __B, __m128d __C) static __inline__ __m128 __DEFAULT_FN_ATTRS _mm_fnmsub_ss(__m128 __A, __m128 __B, __m128 __C) { - return (__m128)__builtin_ia32_vfnmsubss3((__v4sf)__A, (__v4sf)__B, (__v4sf)__C); + return (__m128)__builtin_ia32_vfmaddss3((__v4sf)__A, -(__v4sf)__B, -(__v4sf)__C); } static __inline__ __m128d __DEFAULT_FN_ATTRS _mm_fnmsub_sd(__m128d __A, __m128d __B, __m128d __C) { - return (__m128d)__builtin_ia32_vfnmsubsd3((__v2df)__A, (__v2df)__B, (__v2df)__C); + return (__m128d)__builtin_ia32_vfmaddsd3((__v2df)__A, -(__v2df)__B, -(__v2df)__C); } static __inline__ __m128 __DEFAULT_FN_ATTRS diff --git a/test/CodeGen/fma-builtins.c b/test/CodeGen/fma-builtins.c index d5905d33a9..6f792a78d8 100644 --- a/test/CodeGen/fma-builtins.c +++ b/test/CodeGen/fma-builtins.c @@ -43,13 +43,15 @@ __m128d test_mm_fmsub_pd(__m128d a, __m128d b, __m128d c) { __m128 test_mm_fmsub_ss(__m128 a, __m128 b, __m128 c) { // CHECK-LABEL: test_mm_fmsub_ss - // CHECK: @llvm.x86.fma.vfmsub.ss + // CHECK: [[NEG:%.+]] = fsub <4 x float> , %{{.+}} + // CHECK: @llvm.x86.fma.vfmadd.ss(<4 x float> %{{.+}}, <4 x float> %{{.+}}, <4 x float> [[NEG]]) return _mm_fmsub_ss(a, b, c); } __m128d test_mm_fmsub_sd(__m128d a, __m128d b, __m128d c) { // CHECK-LABEL: test_mm_fmsub_sd - // CHECK: @llvm.x86.fma.vfmsub.sd + // CHECK: [[NEG:%.+]] = fsub <2 x double> , %{{.+}} + // CHECK: @llvm.x86.fma.vfmadd.sd(<2 x double> %{{.+}}, <2 x double> %{{.+}}, <2 x double> [[NEG]]) return _mm_fmsub_sd(a, b, c); } @@ -69,13 +71,15 @@ __m128d test_mm_fnmadd_pd(__m128d a, __m128d b, __m128d c) { __m128 test_mm_fnmadd_ss(__m128 a, __m128 b, __m128 c) { // CHECK-LABEL: test_mm_fnmadd_ss - // CHECK: @llvm.x86.fma.vfnmadd.ss + // CHECK: [[NEG:%.+]] = fsub <4 x float> , %{{.+}} + // CHECK: @llvm.x86.fma.vfmadd.ss(<4 x float> %{{.+}}, <4 x float> [[NEG]], <4 x float> %{{.+}}) return _mm_fnmadd_ss(a, b, c); } __m128d test_mm_fnmadd_sd(__m128d a, __m128d b, __m128d c) { // CHECK-LABEL: test_mm_fnmadd_sd - // CHECK: @llvm.x86.fma.vfnmadd.sd + // CHECK: [[NEG:%.+]] = fsub <2 x double> , %{{.+}} + // CHECK: @llvm.x86.fma.vfmadd.sd(<2 x double> %{{.+}}, <2 x double> [[NEG]], <2 x double> %{{.+}}) return _mm_fnmadd_sd(a, b, c); } @@ -97,13 +101,17 @@ __m128d test_mm_fnmsub_pd(__m128d a, __m128d b, __m128d c) { __m128 test_mm_fnmsub_ss(__m128 a, __m128 b, __m128 c) { // CHECK-LABEL: test_mm_fnmsub_ss - // CHECK: @llvm.x86.fma.vfnmsub.ss + // CHECK: [[NEG:%.+]] = fsub <4 x float> , %{{.+}} + // CHECK: [[NEG2:%.+]] = fsub <4 x float> , %{{.+}} + // CHECK: @llvm.x86.fma.vfmadd.ss(<4 x float> %{{.+}}, <4 x float> [[NEG]], <4 x float> [[NEG2]]) return _mm_fnmsub_ss(a, b, c); } __m128d test_mm_fnmsub_sd(__m128d a, __m128d b, __m128d c) { // CHECK-LABEL: test_mm_fnmsub_sd - // CHECK: @llvm.x86.fma.vfnmsub.sd + // CHECK: [[NEG:%.+]] = fsub <2 x double> , %{{.+}} + // CHECK: [[NEG2:%.+]] = fsub <2 x double> , %{{.+}} + // CHECK: @llvm.x86.fma.vfmadd.sd(<2 x double> %{{.+}}, <2 x double> [[NEG]], <2 x double> [[NEG2]]) return _mm_fnmsub_sd(a, b, c); } diff --git a/test/CodeGen/fma4-builtins.c b/test/CodeGen/fma4-builtins.c index 68a2ec6e3a..c848d4a751 100644 --- a/test/CodeGen/fma4-builtins.c +++ b/test/CodeGen/fma4-builtins.c @@ -17,13 +17,13 @@ __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.fma.vfmadd.ss + // CHECK: @llvm.x86.fma4.vfmadd.ss 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.fma.vfmadd.sd + // CHECK: @llvm.x86.fma4.vfmadd.sd return _mm_macc_sd(a, b, c); } @@ -43,13 +43,15 @@ __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: @llvm.x86.fma.vfmsub.ss + // CHECK: [[NEG:%.+]] = fsub <4 x float> , %{{.+}} + // CHECK: @llvm.x86.fma4.vfmadd.ss(<4 x float> %{{.+}}, <4 x float> %{{.+}}, <4 x float> [[NEG]]) 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: @llvm.x86.fma.vfmsub.sd + // CHECK: [[NEG:%.+]] = fsub <2 x double> , %{{.+}} + // CHECK: @llvm.x86.fma4.vfmadd.sd(<2 x double> %{{.+}}, <2 x double> %{{.+}}, <2 x double> [[NEG]]) return _mm_msub_sd(a, b, c); } @@ -69,13 +71,15 @@ __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: @llvm.x86.fma.vfnmadd.ss + // CHECK: [[NEG:%.+]] = fsub <4 x float> , %{{.+}} + // CHECK: @llvm.x86.fma4.vfmadd.ss(<4 x float> [[NEG]], <4 x float> %{{.+}}, <4 x float> %{{.+}}) 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: @llvm.x86.fma.vfnmadd.sd + // CHECK: [[NEG:%.+]] = fsub <2 x double> , %{{.+}} + // CHECK: @llvm.x86.fma4.vfmadd.sd(<2 x double> [[NEG]], <2 x double> %{{.+}}, <2 x double> %{{.+}}) return _mm_nmacc_sd(a, b, c); } @@ -97,13 +101,17 @@ __m128d test_mm_nmsub_pd(__m128d a, __m128d b, __m128d c) { __m128 test_mm_nmsub_ss(__m128 a, __m128 b, __m128 c) { // CHECK-LABEL: test_mm_nmsub_ss - // CHECK: @llvm.x86.fma.vfnmsub.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]]) return _mm_nmsub_ss(a, b, c); } __m128d test_mm_nmsub_sd(__m128d a, __m128d b, __m128d c) { // CHECK-LABEL: test_mm_nmsub_sd - // CHECK: @llvm.x86.fma.vfnmsub.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]]) return _mm_nmsub_sd(a, b, c); } -- 2.50.1