From 347daedef2a65151b13d1e6c6feb813b6d9b0c93 Mon Sep 17 00:00:00 2001 From: Craig Topper Date: Thu, 16 Aug 2018 06:20:29 +0000 Subject: [PATCH] [X86] Remove masking from the 512-bit padds and psubs builtins. Use select builtin instead. git-svn-id: https://llvm.org/svn/llvm-project/cfe/trunk@339843 91177308-0d34-0410-b5e6-96231b3b80d8 --- include/clang/Basic/BuiltinsX86.def | 8 +-- lib/Headers/avx512bwintrin.h | 88 +++++++++++------------------ test/CodeGen/avx512bw-builtins.c | 32 +++++++---- 3 files changed, 56 insertions(+), 72 deletions(-) diff --git a/include/clang/Basic/BuiltinsX86.def b/include/clang/Basic/BuiltinsX86.def index e98f7d612c..2d0b96ea59 100644 --- a/include/clang/Basic/BuiltinsX86.def +++ b/include/clang/Basic/BuiltinsX86.def @@ -1038,8 +1038,8 @@ TARGET_BUILTIN(__builtin_ia32_packssdw512, "V32sV16iV16i", "ncV:512:", "avx512bw TARGET_BUILTIN(__builtin_ia32_packsswb512, "V64cV32sV32s", "ncV:512:", "avx512bw") TARGET_BUILTIN(__builtin_ia32_packusdw512, "V32sV16iV16i", "ncV:512:", "avx512bw") TARGET_BUILTIN(__builtin_ia32_packuswb512, "V64cV32sV32s", "ncV:512:", "avx512bw") -TARGET_BUILTIN(__builtin_ia32_paddsb512_mask, "V64cV64cV64cV64cULLi", "ncV:512:", "avx512bw") -TARGET_BUILTIN(__builtin_ia32_paddsw512_mask, "V32sV32sV32sV32sUi", "ncV:512:", "avx512bw") +TARGET_BUILTIN(__builtin_ia32_paddsb512, "V64cV64cV64c", "ncV:512:", "avx512bw") +TARGET_BUILTIN(__builtin_ia32_paddsw512, "V32sV32sV32s", "ncV:512:", "avx512bw") TARGET_BUILTIN(__builtin_ia32_paddusb512_mask, "V64cV64cV64cV64cULLi", "ncV:512:", "avx512bw") TARGET_BUILTIN(__builtin_ia32_paddusw512_mask, "V32sV32sV32sV32sUi", "ncV:512:", "avx512bw") TARGET_BUILTIN(__builtin_ia32_pmaxsb512, "V64cV64cV64c", "ncV:512:", "avx512bw") @@ -1051,8 +1051,8 @@ TARGET_BUILTIN(__builtin_ia32_pminsw512, "V32sV32sV32s", "ncV:512:", "avx512bw") TARGET_BUILTIN(__builtin_ia32_pminub512, "V64cV64cV64c", "ncV:512:", "avx512bw") TARGET_BUILTIN(__builtin_ia32_pminuw512, "V32sV32sV32s", "ncV:512:", "avx512bw") TARGET_BUILTIN(__builtin_ia32_pshufb512, "V64cV64cV64c", "ncV:512:", "avx512bw") -TARGET_BUILTIN(__builtin_ia32_psubsb512_mask, "V64cV64cV64cV64cULLi", "ncV:512:", "avx512bw") -TARGET_BUILTIN(__builtin_ia32_psubsw512_mask, "V32sV32sV32sV32sUi", "ncV:512:", "avx512bw") +TARGET_BUILTIN(__builtin_ia32_psubsb512, "V64cV64cV64c", "ncV:512:", "avx512bw") +TARGET_BUILTIN(__builtin_ia32_psubsw512, "V32sV32sV32s", "ncV:512:", "avx512bw") TARGET_BUILTIN(__builtin_ia32_psubusb512_mask, "V64cV64cV64cV64cULLi", "ncV:512:", "avx512bw") TARGET_BUILTIN(__builtin_ia32_psubusw512_mask, "V32sV32sV32sV32sUi", "ncV:512:", "avx512bw") diff --git a/lib/Headers/avx512bwintrin.h b/lib/Headers/avx512bwintrin.h index fc46323749..7eda8a6dde 100644 --- a/lib/Headers/avx512bwintrin.h +++ b/lib/Headers/avx512bwintrin.h @@ -422,57 +422,45 @@ _mm512_maskz_packus_epi16(__mmask64 __M, __m512i __A, __m512i __B) static __inline__ __m512i __DEFAULT_FN_ATTRS _mm512_adds_epi8 (__m512i __A, __m512i __B) { - return (__m512i) __builtin_ia32_paddsb512_mask ((__v64qi) __A, - (__v64qi) __B, - (__v64qi) _mm512_setzero_si512(), - (__mmask64) -1); + return (__m512i)__builtin_ia32_paddsb512((__v64qi)__A, (__v64qi)__B); } static __inline__ __m512i __DEFAULT_FN_ATTRS -_mm512_mask_adds_epi8 (__m512i __W, __mmask64 __U, __m512i __A, - __m512i __B) +_mm512_mask_adds_epi8 (__m512i __W, __mmask64 __U, __m512i __A, __m512i __B) { - return (__m512i) __builtin_ia32_paddsb512_mask ((__v64qi) __A, - (__v64qi) __B, - (__v64qi) __W, - (__mmask64) __U); + return (__m512i)__builtin_ia32_selectb_512((__mmask64)__U, + (__v64qi)_mm512_adds_epi8(__A, __B), + (__v64qi)__W); } static __inline__ __m512i __DEFAULT_FN_ATTRS _mm512_maskz_adds_epi8 (__mmask64 __U, __m512i __A, __m512i __B) { - return (__m512i) __builtin_ia32_paddsb512_mask ((__v64qi) __A, - (__v64qi) __B, - (__v64qi) _mm512_setzero_si512(), - (__mmask64) __U); + return (__m512i)__builtin_ia32_selectb_512((__mmask64)__U, + (__v64qi)_mm512_adds_epi8(__A, __B), + (__v64qi)_mm512_setzero_si512()); } static __inline__ __m512i __DEFAULT_FN_ATTRS _mm512_adds_epi16 (__m512i __A, __m512i __B) { - return (__m512i) __builtin_ia32_paddsw512_mask ((__v32hi) __A, - (__v32hi) __B, - (__v32hi) _mm512_setzero_si512(), - (__mmask32) -1); + return (__m512i)__builtin_ia32_paddsw512((__v32hi)__A, (__v32hi)__B); } static __inline__ __m512i __DEFAULT_FN_ATTRS -_mm512_mask_adds_epi16 (__m512i __W, __mmask32 __U, __m512i __A, - __m512i __B) +_mm512_mask_adds_epi16 (__m512i __W, __mmask32 __U, __m512i __A, __m512i __B) { - return (__m512i) __builtin_ia32_paddsw512_mask ((__v32hi) __A, - (__v32hi) __B, - (__v32hi) __W, - (__mmask32) __U); + return (__m512i)__builtin_ia32_selectw_512((__mmask32)__U, + (__v32hi)_mm512_adds_epi16(__A, __B), + (__v32hi)__W); } static __inline__ __m512i __DEFAULT_FN_ATTRS _mm512_maskz_adds_epi16 (__mmask32 __U, __m512i __A, __m512i __B) { - return (__m512i) __builtin_ia32_paddsw512_mask ((__v32hi) __A, - (__v32hi) __B, - (__v32hi) _mm512_setzero_si512(), - (__mmask32) __U); + return (__m512i)__builtin_ia32_selectw_512((__mmask32)__U, + (__v32hi)_mm512_adds_epi16(__A, __B), + (__v32hi)_mm512_setzero_si512()); } static __inline__ __m512i __DEFAULT_FN_ATTRS @@ -787,57 +775,45 @@ _mm512_maskz_shuffle_epi8(__mmask64 __U, __m512i __A, __m512i __B) static __inline__ __m512i __DEFAULT_FN_ATTRS _mm512_subs_epi8 (__m512i __A, __m512i __B) { - return (__m512i) __builtin_ia32_psubsb512_mask ((__v64qi) __A, - (__v64qi) __B, - (__v64qi) _mm512_setzero_si512(), - (__mmask64) -1); + return (__m512i)__builtin_ia32_psubsb512((__v64qi)__A, (__v64qi)__B); } static __inline__ __m512i __DEFAULT_FN_ATTRS -_mm512_mask_subs_epi8 (__m512i __W, __mmask64 __U, __m512i __A, - __m512i __B) +_mm512_mask_subs_epi8 (__m512i __W, __mmask64 __U, __m512i __A, __m512i __B) { - return (__m512i) __builtin_ia32_psubsb512_mask ((__v64qi) __A, - (__v64qi) __B, - (__v64qi) __W, - (__mmask64) __U); + return (__m512i)__builtin_ia32_selectb_512((__mmask64)__U, + (__v64qi)_mm512_subs_epi8(__A, __B), + (__v64qi)__W); } static __inline__ __m512i __DEFAULT_FN_ATTRS _mm512_maskz_subs_epi8 (__mmask64 __U, __m512i __A, __m512i __B) { - return (__m512i) __builtin_ia32_psubsb512_mask ((__v64qi) __A, - (__v64qi) __B, - (__v64qi) _mm512_setzero_si512(), - (__mmask64) __U); + return (__m512i)__builtin_ia32_selectb_512((__mmask64)__U, + (__v64qi)_mm512_subs_epi8(__A, __B), + (__v64qi)_mm512_setzero_si512()); } static __inline__ __m512i __DEFAULT_FN_ATTRS _mm512_subs_epi16 (__m512i __A, __m512i __B) { - return (__m512i) __builtin_ia32_psubsw512_mask ((__v32hi) __A, - (__v32hi) __B, - (__v32hi) _mm512_setzero_si512(), - (__mmask32) -1); + return (__m512i)__builtin_ia32_psubsw512((__v32hi)__A, (__v32hi)__B); } static __inline__ __m512i __DEFAULT_FN_ATTRS -_mm512_mask_subs_epi16 (__m512i __W, __mmask32 __U, __m512i __A, - __m512i __B) +_mm512_mask_subs_epi16 (__m512i __W, __mmask32 __U, __m512i __A, __m512i __B) { - return (__m512i) __builtin_ia32_psubsw512_mask ((__v32hi) __A, - (__v32hi) __B, - (__v32hi) __W, - (__mmask32) __U); + return (__m512i)__builtin_ia32_selectw_512((__mmask32)__U, + (__v32hi)_mm512_subs_epi16(__A, __B), + (__v32hi)__W); } static __inline__ __m512i __DEFAULT_FN_ATTRS _mm512_maskz_subs_epi16 (__mmask32 __U, __m512i __A, __m512i __B) { - return (__m512i) __builtin_ia32_psubsw512_mask ((__v32hi) __A, - (__v32hi) __B, - (__v32hi) _mm512_setzero_si512(), - (__mmask32) __U); + return (__m512i)__builtin_ia32_selectw_512((__mmask32)__U, + (__v32hi)_mm512_subs_epi16(__A, __B), + (__v32hi)_mm512_setzero_si512()); } static __inline__ __m512i __DEFAULT_FN_ATTRS diff --git a/test/CodeGen/avx512bw-builtins.c b/test/CodeGen/avx512bw-builtins.c index 2023a71194..96f142ce82 100644 --- a/test/CodeGen/avx512bw-builtins.c +++ b/test/CodeGen/avx512bw-builtins.c @@ -594,32 +594,36 @@ __m512i test_mm512_maskz_packus_epi16(__mmask64 __M, __m512i __A, __m512i __B) { } __m512i test_mm512_adds_epi8(__m512i __A, __m512i __B) { // CHECK-LABEL: @test_mm512_adds_epi8 - // CHECK: @llvm.x86.avx512.mask.padds.b.512 + // CHECK: @llvm.x86.avx512.padds.b.512 return _mm512_adds_epi8(__A,__B); } __m512i test_mm512_mask_adds_epi8(__m512i __W, __mmask64 __U, __m512i __A, __m512i __B) { // CHECK-LABEL: @test_mm512_mask_adds_epi8 - // CHECK: @llvm.x86.avx512.mask.padds.b.512 + // CHECK: @llvm.x86.avx512.padds.b.512 + // CHECK: select <64 x i1> %{{.*}}, <64 x i8> %{{.*}}, <64 x i8> %{{.*}} return _mm512_mask_adds_epi8(__W,__U,__A,__B); } __m512i test_mm512_maskz_adds_epi8(__mmask64 __U, __m512i __A, __m512i __B) { // CHECK-LABEL: @test_mm512_maskz_adds_epi8 - // CHECK: @llvm.x86.avx512.mask.padds.b.512 + // CHECK: @llvm.x86.avx512.padds.b.512 + // CHECK: select <64 x i1> %{{.*}}, <64 x i8> %{{.*}}, <64 x i8> %{{.*}} return _mm512_maskz_adds_epi8(__U,__A,__B); } __m512i test_mm512_adds_epi16(__m512i __A, __m512i __B) { // CHECK-LABEL: @test_mm512_adds_epi16 - // CHECK: @llvm.x86.avx512.mask.padds.w.512 + // CHECK: @llvm.x86.avx512.padds.w.512 return _mm512_adds_epi16(__A,__B); } __m512i test_mm512_mask_adds_epi16(__m512i __W, __mmask32 __U, __m512i __A, __m512i __B) { // CHECK-LABEL: @test_mm512_mask_adds_epi16 - // CHECK: @llvm.x86.avx512.mask.padds.w.512 + // CHECK: @llvm.x86.avx512.padds.w.512 + // CHECK: select <32 x i1> %{{.*}}, <32 x i16> %{{.*}}, <32 x i16> %{{.*}} return _mm512_mask_adds_epi16(__W,__U,__A,__B); } __m512i test_mm512_maskz_adds_epi16(__mmask32 __U, __m512i __A, __m512i __B) { // CHECK-LABEL: @test_mm512_maskz_adds_epi16 - // CHECK: @llvm.x86.avx512.mask.padds.w.512 + // CHECK: @llvm.x86.avx512.padds.w.512 + // CHECK: select <32 x i1> %{{.*}}, <32 x i16> %{{.*}}, <32 x i16> %{{.*}} return _mm512_maskz_adds_epi16(__U,__A,__B); } __m512i test_mm512_adds_epu8(__m512i __A, __m512i __B) { @@ -925,32 +929,36 @@ __m512i test_mm512_maskz_shuffle_epi8(__mmask64 __U, __m512i __A, __m512i __B) { } __m512i test_mm512_subs_epi8(__m512i __A, __m512i __B) { // CHECK-LABEL: @test_mm512_subs_epi8 - // CHECK: @llvm.x86.avx512.mask.psubs.b.512 + // CHECK: @llvm.x86.avx512.psubs.b.512 return _mm512_subs_epi8(__A,__B); } __m512i test_mm512_mask_subs_epi8(__m512i __W, __mmask64 __U, __m512i __A, __m512i __B) { // CHECK-LABEL: @test_mm512_mask_subs_epi8 - // CHECK: @llvm.x86.avx512.mask.psubs.b.512 + // CHECK: @llvm.x86.avx512.psubs.b.512 + // CHECK: select <64 x i1> %{{.*}}, <64 x i8> %{{.*}}, <64 x i8> %{{.*}} return _mm512_mask_subs_epi8(__W,__U,__A,__B); } __m512i test_mm512_maskz_subs_epi8(__mmask64 __U, __m512i __A, __m512i __B) { // CHECK-LABEL: @test_mm512_maskz_subs_epi8 - // CHECK: @llvm.x86.avx512.mask.psubs.b.512 + // CHECK: @llvm.x86.avx512.psubs.b.512 + // CHECK: select <64 x i1> %{{.*}}, <64 x i8> %{{.*}}, <64 x i8> %{{.*}} return _mm512_maskz_subs_epi8(__U,__A,__B); } __m512i test_mm512_subs_epi16(__m512i __A, __m512i __B) { // CHECK-LABEL: @test_mm512_subs_epi16 - // CHECK: @llvm.x86.avx512.mask.psubs.w.512 + // CHECK: @llvm.x86.avx512.psubs.w.512 return _mm512_subs_epi16(__A,__B); } __m512i test_mm512_mask_subs_epi16(__m512i __W, __mmask32 __U, __m512i __A, __m512i __B) { // CHECK-LABEL: @test_mm512_mask_subs_epi16 - // CHECK: @llvm.x86.avx512.mask.psubs.w.512 + // CHECK: @llvm.x86.avx512.psubs.w.512 + // CHECK: select <32 x i1> %{{.*}}, <32 x i16> %{{.*}}, <32 x i16> %{{.*}} return _mm512_mask_subs_epi16(__W,__U,__A,__B); } __m512i test_mm512_maskz_subs_epi16(__mmask32 __U, __m512i __A, __m512i __B) { // CHECK-LABEL: @test_mm512_maskz_subs_epi16 - // CHECK: @llvm.x86.avx512.mask.psubs.w.512 + // CHECK: @llvm.x86.avx512.psubs.w.512 + // CHECK: select <32 x i1> %{{.*}}, <32 x i16> %{{.*}}, <32 x i16> %{{.*}} return _mm512_maskz_subs_epi16(__U,__A,__B); } __m512i test_mm512_subs_epu8(__m512i __A, __m512i __B) { -- 2.40.0