From 6127cce465f74ceca607e405d83aa5951766bc3f Mon Sep 17 00:00:00 2001 From: Craig Topper Date: Sun, 23 Oct 2016 07:35:39 +0000 Subject: [PATCH] [AVX-512] Remove masked 128/256-bit packss/packus builtins and replace with selects and the older unmasked builtins. git-svn-id: https://llvm.org/svn/llvm-project/cfe/trunk@284935 91177308-0d34-0410-b5e6-96231b3b80d8 --- include/clang/Basic/BuiltinsX86.def | 8 -- lib/Headers/avx512vlbwintrin.h | 149 ++++++++++++---------------- test/CodeGen/avx512vlbw-builtins.c | 48 ++++++--- 3 files changed, 96 insertions(+), 109 deletions(-) diff --git a/include/clang/Basic/BuiltinsX86.def b/include/clang/Basic/BuiltinsX86.def index 803923e225..551ab3e095 100644 --- a/include/clang/Basic/BuiltinsX86.def +++ b/include/clang/Basic/BuiltinsX86.def @@ -1120,14 +1120,6 @@ TARGET_BUILTIN(__builtin_ia32_vpconflictsi_512_mask, "V16iV16iV16iUs", "", "avx5 TARGET_BUILTIN(__builtin_ia32_vplzcntd_512_mask, "V16iV16iV16iUs", "", "avx512cd") TARGET_BUILTIN(__builtin_ia32_vplzcntq_512_mask, "V8LLiV8LLiV8LLiUc", "", "avx512cd") -TARGET_BUILTIN(__builtin_ia32_packssdw128_mask, "V8sV4iV4iV8sUc", "", "avx512vl,avx512bw") -TARGET_BUILTIN(__builtin_ia32_packssdw256_mask, "V16sV8iV8iV16sUs", "", "avx512vl,avx512bw") -TARGET_BUILTIN(__builtin_ia32_packsswb128_mask, "V16cV8sV8sV16cUs", "", "avx512vl,avx512bw") -TARGET_BUILTIN(__builtin_ia32_packsswb256_mask, "V32cV16sV16sV32cUi", "", "avx512vl,avx512bw") -TARGET_BUILTIN(__builtin_ia32_packusdw128_mask, "V8sV4iV4iV8sUc", "", "avx512vl,avx512bw") -TARGET_BUILTIN(__builtin_ia32_packusdw256_mask, "V16sV8iV8iV16sUs", "", "avx512vl,avx512bw") -TARGET_BUILTIN(__builtin_ia32_packuswb128_mask, "V16cV8sV8sV16cUs", "", "avx512vl,avx512bw") -TARGET_BUILTIN(__builtin_ia32_packuswb256_mask, "V32cV16sV16sV32cUi", "", "avx512vl,avx512bw") TARGET_BUILTIN(__builtin_ia32_pmaxsb128_mask, "V16cV16cV16cV16cUs", "", "avx512vl,avx512bw") TARGET_BUILTIN(__builtin_ia32_pmaxsb256_mask, "V32cV32cV32cV32cUi", "", "avx512vl,avx512bw") TARGET_BUILTIN(__builtin_ia32_pmaxsw128_mask, "V8sV8sV8sV8sUc", "", "avx512vl,avx512bw") diff --git a/lib/Headers/avx512vlbwintrin.h b/lib/Headers/avx512vlbwintrin.h index c65e01dbe4..ebd72302ed 100644 --- a/lib/Headers/avx512vlbwintrin.h +++ b/lib/Headers/avx512vlbwintrin.h @@ -851,151 +851,130 @@ _mm256_maskz_abs_epi16(__mmask16 __U, __m256i __A) } static __inline__ __m128i __DEFAULT_FN_ATTRS -_mm_maskz_packs_epi32 (__mmask8 __M, __m128i __A, __m128i __B) -{ - return (__m128i) __builtin_ia32_packssdw128_mask ((__v4si) __A, - (__v4si) __B, - (__v8hi) _mm_setzero_si128 (), __M); +_mm_maskz_packs_epi32(__mmask8 __M, __m128i __A, __m128i __B) { + return (__m128i)__builtin_ia32_selectw_128((__mmask8)__M, + (__v8hi)_mm_packs_epi32(__A, __B), + (__v8hi)_mm_setzero_si128()); } static __inline__ __m128i __DEFAULT_FN_ATTRS -_mm_mask_packs_epi32 (__m128i __W, __mmask16 __M, __m128i __A, - __m128i __B) +_mm_mask_packs_epi32(__m128i __W, __mmask16 __M, __m128i __A, __m128i __B) { - return (__m128i) __builtin_ia32_packssdw128_mask ((__v4si) __A, - (__v4si) __B, - (__v8hi) __W, __M); + return (__m128i)__builtin_ia32_selectw_128((__mmask8)__M, + (__v8hi)_mm_packs_epi32(__A, __B), + (__v8hi)__W); } static __inline__ __m256i __DEFAULT_FN_ATTRS -_mm256_maskz_packs_epi32 (__mmask16 __M, __m256i __A, __m256i __B) +_mm256_maskz_packs_epi32(__mmask16 __M, __m256i __A, __m256i __B) { - return (__m256i) __builtin_ia32_packssdw256_mask ((__v8si) __A, - (__v8si) __B, - (__v16hi) _mm256_setzero_si256 (), - __M); + return (__m256i)__builtin_ia32_selectw_256((__mmask16)__M, + (__v16hi)_mm256_packs_epi32(__A, __B), + (__v16hi)_mm256_setzero_si256()); } static __inline__ __m256i __DEFAULT_FN_ATTRS -_mm256_mask_packs_epi32 (__m256i __W, __mmask16 __M, __m256i __A, - __m256i __B) +_mm256_mask_packs_epi32(__m256i __W, __mmask16 __M, __m256i __A, __m256i __B) { - return (__m256i) __builtin_ia32_packssdw256_mask ((__v8si) __A, - (__v8si) __B, - (__v16hi) __W, __M); + return (__m256i)__builtin_ia32_selectw_256((__mmask16)__M, + (__v16hi)_mm256_packs_epi32(__A, __B), + (__v16hi)__W); } static __inline__ __m128i __DEFAULT_FN_ATTRS -_mm_maskz_packs_epi16 (__mmask16 __M, __m128i __A, __m128i __B) +_mm_maskz_packs_epi16(__mmask16 __M, __m128i __A, __m128i __B) { - return (__m128i) __builtin_ia32_packsswb128_mask ((__v8hi) __A, - (__v8hi) __B, - (__v16qi) _mm_setzero_si128 (), - __M); + return (__m128i)__builtin_ia32_selectb_128((__mmask16)__M, + (__v16qi)_mm_packs_epi16(__A, __B), + (__v16qi)_mm_setzero_si128()); } static __inline__ __m128i __DEFAULT_FN_ATTRS -_mm_mask_packs_epi16 (__m128i __W, __mmask16 __M, __m128i __A, - __m128i __B) +_mm_mask_packs_epi16(__m128i __W, __mmask16 __M, __m128i __A, __m128i __B) { - return (__m128i) __builtin_ia32_packsswb128_mask ((__v8hi) __A, - (__v8hi) __B, - (__v16qi) __W, - __M); + return (__m128i)__builtin_ia32_selectb_128((__mmask16)__M, + (__v16qi)_mm_packs_epi16(__A, __B), + (__v16qi)__W); } static __inline__ __m256i __DEFAULT_FN_ATTRS -_mm256_maskz_packs_epi16 (__mmask32 __M, __m256i __A, __m256i __B) +_mm256_maskz_packs_epi16(__mmask32 __M, __m256i __A, __m256i __B) { - return (__m256i) __builtin_ia32_packsswb256_mask ((__v16hi) __A, - (__v16hi) __B, - (__v32qi) _mm256_setzero_si256 (), - __M); + return (__m256i)__builtin_ia32_selectb_256((__mmask32)__M, + (__v32qi)_mm256_packs_epi16(__A, __B), + (__v32qi)_mm256_setzero_si256()); } static __inline__ __m256i __DEFAULT_FN_ATTRS -_mm256_mask_packs_epi16 (__m256i __W, __mmask32 __M, __m256i __A, - __m256i __B) +_mm256_mask_packs_epi16(__m256i __W, __mmask32 __M, __m256i __A, __m256i __B) { - return (__m256i) __builtin_ia32_packsswb256_mask ((__v16hi) __A, - (__v16hi) __B, - (__v32qi) __W, - __M); + return (__m256i)__builtin_ia32_selectb_256((__mmask32)__M, + (__v32qi)_mm256_packs_epi16(__A, __B), + (__v32qi)__W); } static __inline__ __m128i __DEFAULT_FN_ATTRS -_mm_maskz_packus_epi32 (__mmask8 __M, __m128i __A, __m128i __B) +_mm_maskz_packus_epi32(__mmask8 __M, __m128i __A, __m128i __B) { - return (__m128i) __builtin_ia32_packusdw128_mask ((__v4si) __A, - (__v4si) __B, - (__v8hi) _mm_setzero_si128 (), - __M); + return (__m128i)__builtin_ia32_selectw_128((__mmask8)__M, + (__v8hi)_mm_packus_epi32(__A, __B), + (__v8hi)_mm_setzero_si128()); } static __inline__ __m128i __DEFAULT_FN_ATTRS -_mm_mask_packus_epi32 (__m128i __W, __mmask16 __M, __m128i __A, - __m128i __B) +_mm_mask_packus_epi32(__m128i __W, __mmask16 __M, __m128i __A, __m128i __B) { - return (__m128i) __builtin_ia32_packusdw128_mask ((__v4si) __A, - (__v4si) __B, - (__v8hi) __W, __M); + return (__m128i)__builtin_ia32_selectw_128((__mmask8)__M, + (__v8hi)_mm_packus_epi32(__A, __B), + (__v8hi)__W); } static __inline__ __m256i __DEFAULT_FN_ATTRS -_mm256_maskz_packus_epi32 (__mmask16 __M, __m256i __A, __m256i __B) +_mm256_maskz_packus_epi32(__mmask16 __M, __m256i __A, __m256i __B) { - return (__m256i) __builtin_ia32_packusdw256_mask ((__v8si) __A, - (__v8si) __B, - (__v16hi) _mm256_setzero_si256 (), - __M); + return (__m256i)__builtin_ia32_selectw_256((__mmask16)__M, + (__v16hi)_mm256_packus_epi32(__A, __B), + (__v16hi)_mm256_setzero_si256()); } static __inline__ __m256i __DEFAULT_FN_ATTRS -_mm256_mask_packus_epi32 (__m256i __W, __mmask16 __M, __m256i __A, - __m256i __B) +_mm256_mask_packus_epi32(__m256i __W, __mmask16 __M, __m256i __A, __m256i __B) { - return (__m256i) __builtin_ia32_packusdw256_mask ((__v8si) __A, - (__v8si) __B, - (__v16hi) __W, - __M); + return (__m256i)__builtin_ia32_selectw_256((__mmask16)__M, + (__v16hi)_mm256_packus_epi32(__A, __B), + (__v16hi)__W); } static __inline__ __m128i __DEFAULT_FN_ATTRS -_mm_maskz_packus_epi16 (__mmask16 __M, __m128i __A, __m128i __B) +_mm_maskz_packus_epi16(__mmask16 __M, __m128i __A, __m128i __B) { - return (__m128i) __builtin_ia32_packuswb128_mask ((__v8hi) __A, - (__v8hi) __B, - (__v16qi) _mm_setzero_si128 (), - __M); + return (__m128i)__builtin_ia32_selectb_128((__mmask16)__M, + (__v16qi)_mm_packus_epi16(__A, __B), + (__v16qi)_mm_setzero_si128()); } static __inline__ __m128i __DEFAULT_FN_ATTRS -_mm_mask_packus_epi16 (__m128i __W, __mmask16 __M, __m128i __A, - __m128i __B) +_mm_mask_packus_epi16(__m128i __W, __mmask16 __M, __m128i __A, __m128i __B) { - return (__m128i) __builtin_ia32_packuswb128_mask ((__v8hi) __A, - (__v8hi) __B, - (__v16qi) __W, - __M); + return (__m128i)__builtin_ia32_selectb_128((__mmask16)__M, + (__v16qi)_mm_packus_epi16(__A, __B), + (__v16qi)__W); } static __inline__ __m256i __DEFAULT_FN_ATTRS -_mm256_maskz_packus_epi16 (__mmask32 __M, __m256i __A, __m256i __B) +_mm256_maskz_packus_epi16(__mmask32 __M, __m256i __A, __m256i __B) { - return (__m256i) __builtin_ia32_packuswb256_mask ((__v16hi) __A, - (__v16hi) __B, - (__v32qi) _mm256_setzero_si256 (), - __M); + return (__m256i)__builtin_ia32_selectb_256((__mmask32)__M, + (__v32qi)_mm256_packus_epi16(__A, __B), + (__v32qi)_mm256_setzero_si256()); } static __inline__ __m256i __DEFAULT_FN_ATTRS -_mm256_mask_packus_epi16 (__m256i __W, __mmask32 __M, __m256i __A, - __m256i __B) +_mm256_mask_packus_epi16(__m256i __W, __mmask32 __M, __m256i __A, __m256i __B) { - return (__m256i) __builtin_ia32_packuswb256_mask ((__v16hi) __A, - (__v16hi) __B, - (__v32qi) __W, - __M); + return (__m256i)__builtin_ia32_selectb_256((__mmask32)__M, + (__v32qi)_mm256_packus_epi16(__A, __B), + (__v32qi)__W); } static __inline__ __m128i __DEFAULT_FN_ATTRS diff --git a/test/CodeGen/avx512vlbw-builtins.c b/test/CodeGen/avx512vlbw-builtins.c index 1923350c2e..0eaa004e0b 100644 --- a/test/CodeGen/avx512vlbw-builtins.c +++ b/test/CodeGen/avx512vlbw-builtins.c @@ -954,90 +954,106 @@ __m256i test_mm256_maskz_abs_epi16(__mmask16 __U, __m256i __A) { __m128i test_mm_maskz_packs_epi32(__mmask8 __M, __m128i __A, __m128i __B) { // CHECK-LABEL: @test_mm_maskz_packs_epi32 - // CHECK: @llvm.x86.avx512.mask.packssdw.128 + // CHECK: @llvm.x86.sse2.packssdw + // CHECK: select <8 x i1> %{{.*}}, <8 x i16> %{{.*}}, <8 x i16> %{{.*}} return _mm_maskz_packs_epi32(__M,__A,__B); } __m128i test_mm_mask_packs_epi32(__m128i __W, __mmask16 __M, __m128i __A, __m128i __B) { // CHECK-LABEL: @test_mm_mask_packs_epi32 - // CHECK: @llvm.x86.avx512.mask.packssdw.128 + // CHECK: @llvm.x86.sse2.packssdw + // CHECK: select <8 x i1> %{{.*}}, <8 x i16> %{{.*}}, <8 x i16> %{{.*}} return _mm_mask_packs_epi32(__W,__M,__A,__B); } __m256i test_mm256_maskz_packs_epi32(__mmask16 __M, __m256i __A, __m256i __B) { // CHECK-LABEL: @test_mm256_maskz_packs_epi32 - // CHECK: @llvm.x86.avx512.mask.packssdw.256 + // CHECK: @llvm.x86.avx2.packssdw + // CHECK: select <16 x i1> %{{.*}}, <16 x i16> %{{.*}}, <16 x i16> %{{.*}} return _mm256_maskz_packs_epi32(__M,__A,__B); } __m256i test_mm256_mask_packs_epi32(__m256i __W, __mmask16 __M, __m256i __A, __m256i __B) { // CHECK-LABEL: @test_mm256_mask_packs_epi32 - // CHECK: @llvm.x86.avx512.mask.packssdw.256 + // CHECK: @llvm.x86.avx2.packssdw + // CHECK: select <16 x i1> %{{.*}}, <16 x i16> %{{.*}}, <16 x i16> %{{.*}} return _mm256_mask_packs_epi32(__W,__M,__A,__B); } __m128i test_mm_maskz_packs_epi16(__mmask16 __M, __m128i __A, __m128i __B) { // CHECK-LABEL: @test_mm_maskz_packs_epi16 - // CHECK: @llvm.x86.avx512.mask.packsswb.128 + // CHECK: @llvm.x86.sse2.packsswb + // CHECK: select <16 x i1> %{{.*}}, <16 x i8> %{{.*}}, <16 x i8> %{{.*}} return _mm_maskz_packs_epi16(__M,__A,__B); } __m128i test_mm_mask_packs_epi16(__m128i __W, __mmask16 __M, __m128i __A, __m128i __B) { // CHECK-LABEL: @test_mm_mask_packs_epi16 - // CHECK: @llvm.x86.avx512.mask.packsswb.128 + // CHECK: @llvm.x86.sse2.packsswb + // CHECK: select <16 x i1> %{{.*}}, <16 x i8> %{{.*}}, <16 x i8> %{{.*}} return _mm_mask_packs_epi16(__W,__M,__A,__B); } __m256i test_mm256_maskz_packs_epi16(__mmask32 __M, __m256i __A, __m256i __B) { // CHECK-LABEL: @test_mm256_maskz_packs_epi16 - // CHECK: @llvm.x86.avx512.mask.packsswb.256 + // CHECK: @llvm.x86.avx2.packsswb + // CHECK: select <32 x i1> %{{.*}}, <32 x i8> %{{.*}}, <32 x i8> %{{.*}} return _mm256_maskz_packs_epi16(__M,__A,__B); } __m256i test_mm256_mask_packs_epi16(__m256i __W, __mmask32 __M, __m256i __A, __m256i __B) { // CHECK-LABEL: @test_mm256_mask_packs_epi16 - // CHECK: @llvm.x86.avx512.mask.packsswb.256 + // CHECK: @llvm.x86.avx2.packsswb + // CHECK: select <32 x i1> %{{.*}}, <32 x i8> %{{.*}}, <32 x i8> %{{.*}} return _mm256_mask_packs_epi16(__W,__M,__A,__B); } __m128i test_mm_mask_packus_epi32(__m128i __W, __mmask16 __M, __m128i __A, __m128i __B) { // CHECK-LABEL: @test_mm_mask_packus_epi32 - // CHECK: @llvm.x86.avx512.mask.packusdw.128 + // CHECK: @llvm.x86.sse41.packusdw + // CHECK: select <8 x i1> %{{.*}}, <8 x i16> %{{.*}}, <8 x i16> %{{.*}} return _mm_mask_packus_epi32(__W,__M,__A,__B); } __m128i test_mm_maskz_packus_epi32(__mmask8 __M, __m128i __A, __m128i __B) { // CHECK-LABEL: @test_mm_maskz_packus_epi32 - // CHECK: @llvm.x86.avx512.mask.packusdw.128 + // CHECK: @llvm.x86.sse41.packusdw + // CHECK: select <8 x i1> %{{.*}}, <8 x i16> %{{.*}}, <8 x i16> %{{.*}} return _mm_maskz_packus_epi32(__M,__A,__B); } __m256i test_mm256_maskz_packus_epi32(__mmask16 __M, __m256i __A, __m256i __B) { // CHECK-LABEL: @test_mm256_maskz_packus_epi32 - // CHECK: @llvm.x86.avx512.mask.packusdw.256 + // CHECK: @llvm.x86.avx2.packusdw + // CHECK: select <16 x i1> %{{.*}}, <16 x i16> %{{.*}}, <16 x i16> %{{.*}} return _mm256_maskz_packus_epi32(__M,__A,__B); } __m256i test_mm256_mask_packus_epi32(__m256i __W, __mmask16 __M, __m256i __A, __m256i __B) { // CHECK-LABEL: @test_mm256_mask_packus_epi32 - // CHECK: @llvm.x86.avx512.mask.packusdw.256 + // CHECK: @llvm.x86.avx2.packusdw + // CHECK: select <16 x i1> %{{.*}}, <16 x i16> %{{.*}}, <16 x i16> %{{.*}} return _mm256_mask_packus_epi32(__W,__M,__A,__B); } __m128i test_mm_maskz_packus_epi16(__mmask16 __M, __m128i __A, __m128i __B) { // CHECK-LABEL: @test_mm_maskz_packus_epi16 - // CHECK: @llvm.x86.avx512.mask.packuswb.128 + // CHECK: @llvm.x86.sse2.packuswb + // CHECK: select <16 x i1> %{{.*}}, <16 x i8> %{{.*}}, <16 x i8> %{{.*}} return _mm_maskz_packus_epi16(__M,__A,__B); } __m128i test_mm_mask_packus_epi16(__m128i __W, __mmask16 __M, __m128i __A, __m128i __B) { // CHECK-LABEL: @test_mm_mask_packus_epi16 - // CHECK: @llvm.x86.avx512.mask.packuswb.128 + // CHECK: @llvm.x86.sse2.packuswb + // CHECK: select <16 x i1> %{{.*}}, <16 x i8> %{{.*}}, <16 x i8> %{{.*}} return _mm_mask_packus_epi16(__W,__M,__A,__B); } __m256i test_mm256_maskz_packus_epi16(__mmask32 __M, __m256i __A, __m256i __B) { // CHECK-LABEL: @test_mm256_maskz_packus_epi16 - // CHECK: @llvm.x86.avx512.mask.packuswb.256 + // CHECK: @llvm.x86.avx2.packuswb + // CHECK: select <32 x i1> %{{.*}}, <32 x i8> %{{.*}}, <32 x i8> %{{.*}} return _mm256_maskz_packus_epi16(__M,__A,__B); } __m256i test_mm256_mask_packus_epi16(__m256i __W, __mmask32 __M, __m256i __A, __m256i __B) { // CHECK-LABEL: @test_mm256_mask_packus_epi16 - // CHECK: @llvm.x86.avx512.mask.packuswb.256 + // CHECK: @llvm.x86.avx2.packuswb + // CHECK: select <32 x i1> %{{.*}}, <32 x i8> %{{.*}}, <32 x i8> %{{.*}} return _mm256_mask_packus_epi16(__W,__M,__A,__B); } -- 2.50.1