From 73370a3e896f6550203841cfd83e34ee6ea3aa7f Mon Sep 17 00:00:00 2001 From: Michael Zuckerman Date: Wed, 2 Mar 2016 09:05:46 +0000 Subject: [PATCH] [CLANG] [AVX512] [BUILTIN] Adding PSRAV Differential Revision: http://reviews.llvm.org/D17699 git-svn-id: https://llvm.org/svn/llvm-project/cfe/trunk@262471 91177308-0d34-0410-b5e6-96231b3b80d8 --- include/clang/Basic/BuiltinsX86.def | 7 ++ lib/Headers/avx512bwintrin.h | 29 +++++++++ lib/Headers/avx512vlbwintrin.h | 60 +++++++++++++++++ lib/Headers/avx512vlintrin.h | 99 +++++++++++++++++++++++++++++ test/CodeGen/avx512bw-builtins.c | 18 ++++++ test/CodeGen/avx512vl-builtins.c | 60 +++++++++++++++++ 6 files changed, 273 insertions(+) diff --git a/include/clang/Basic/BuiltinsX86.def b/include/clang/Basic/BuiltinsX86.def index 63d4c9ca6f..e602088234 100644 --- a/include/clang/Basic/BuiltinsX86.def +++ b/include/clang/Basic/BuiltinsX86.def @@ -1678,6 +1678,13 @@ TARGET_BUILTIN(__builtin_ia32_psrlq128_mask, "V2LLiV2LLiV2LLiV2LLiUc","","avx512 TARGET_BUILTIN(__builtin_ia32_psrlq256_mask, "V4LLiV4LLiV2LLiV4LLiUc","","avx512vl") TARGET_BUILTIN(__builtin_ia32_psrlqi128_mask, "V2LLiV2LLiIiV2LLiUc","","avx512vl") TARGET_BUILTIN(__builtin_ia32_psrlqi256_mask, "V4LLiV4LLiIiV4LLiUc","","avx512vl") +TARGET_BUILTIN(__builtin_ia32_psrav32hi_mask, "V32sV32sV32sV32sUi","","avx512bw") +TARGET_BUILTIN(__builtin_ia32_psrav16hi_mask, "V16sV16sV16sV16sUs","","avx512bw,avx512vl") +TARGET_BUILTIN(__builtin_ia32_psrav8hi_mask, "V8sV8sV8sV8sUc","","avx512bw,avx512vl") +TARGET_BUILTIN(__builtin_ia32_psrav4si_mask, "V4iV4iV4iV4iUc","","avx512vl") +TARGET_BUILTIN(__builtin_ia32_psrav8si_mask, "V8iV8iV8iV8iUc","","avx512vl") +TARGET_BUILTIN(__builtin_ia32_psravq128_mask, "V2LLiV2LLiV2LLiV2LLiUc","","avx512vl") +TARGET_BUILTIN(__builtin_ia32_psravq256_mask, "V4LLiV4LLiV4LLiV4LLiUc","","avx512vl") #undef BUILTIN #undef TARGET_BUILTIN diff --git a/lib/Headers/avx512bwintrin.h b/lib/Headers/avx512bwintrin.h index c1523cfa27..8c0e7da697 100644 --- a/lib/Headers/avx512bwintrin.h +++ b/lib/Headers/avx512bwintrin.h @@ -1749,6 +1749,35 @@ _mm512_maskz_srlv_epi16 (__mmask32 __U, __m512i __A, __m512i __B) (__mmask32) __U); } +static __inline__ __m512i __DEFAULT_FN_ATTRS +_mm512_srav_epi16 (__m512i __A, __m512i __B) +{ + return (__m512i) __builtin_ia32_psrav32hi_mask ((__v32hi) __A, + (__v32hi) __B, + (__v32hi) + _mm512_setzero_hi (), + (__mmask32) -1); +} + +static __inline__ __m512i __DEFAULT_FN_ATTRS +_mm512_mask_srav_epi16 (__m512i __W, __mmask32 __U, __m512i __A, + __m512i __B) +{ + return (__m512i) __builtin_ia32_psrav32hi_mask ((__v32hi) __A, + (__v32hi) __B, + (__v32hi) __W, + (__mmask32) __U); +} + +static __inline__ __m512i __DEFAULT_FN_ATTRS +_mm512_maskz_srav_epi16 (__mmask32 __U, __m512i __A, __m512i __B) +{ + return (__m512i) __builtin_ia32_psrav32hi_mask ((__v32hi) __A, + (__v32hi) __B, + (__v32hi) + _mm512_setzero_hi (), + (__mmask32) __U); +} #undef __DEFAULT_FN_ATTRS #endif diff --git a/lib/Headers/avx512vlbwintrin.h b/lib/Headers/avx512vlbwintrin.h index 8f0a60384c..7184e1c109 100644 --- a/lib/Headers/avx512vlbwintrin.h +++ b/lib/Headers/avx512vlbwintrin.h @@ -2655,6 +2655,66 @@ _mm_maskz_srlv_epi16 (__mmask8 __U, __m128i __A, __m128i __B) (__mmask8) __U); } +static __inline__ __m256i __DEFAULT_FN_ATTRS +_mm256_srav_epi16 (__m256i __A, __m256i __B) +{ + return (__m256i) __builtin_ia32_psrav16hi_mask ((__v16hi) __A, + (__v16hi) __B, + (__v16hi) + _mm256_setzero_si256 (), + (__mmask16) -1); +} + +static __inline__ __m256i __DEFAULT_FN_ATTRS +_mm256_mask_srav_epi16 (__m256i __W, __mmask16 __U, __m256i __A, + __m256i __B) +{ + return (__m256i) __builtin_ia32_psrav16hi_mask ((__v16hi) __A, + (__v16hi) __B, + (__v16hi) __W, + (__mmask16) __U); +} + +static __inline__ __m256i __DEFAULT_FN_ATTRS +_mm256_maskz_srav_epi16 (__mmask16 __U, __m256i __A, __m256i __B) +{ + return (__m256i) __builtin_ia32_psrav16hi_mask ((__v16hi) __A, + (__v16hi) __B, + (__v16hi) + _mm256_setzero_si256 (), + (__mmask16) __U); +} + +static __inline__ __m128i __DEFAULT_FN_ATTRS +_mm_srav_epi16 (__m128i __A, __m128i __B) +{ + return (__m128i) __builtin_ia32_psrav8hi_mask ((__v8hi) __A, + (__v8hi) __B, + (__v8hi) + _mm_setzero_hi (), + (__mmask8) -1); +} + +static __inline__ __m128i __DEFAULT_FN_ATTRS +_mm_mask_srav_epi16 (__m128i __W, __mmask8 __U, __m128i __A, + __m128i __B) +{ + return (__m128i) __builtin_ia32_psrav8hi_mask ((__v8hi) __A, + (__v8hi) __B, + (__v8hi) __W, + (__mmask8) __U); +} + +static __inline__ __m128i __DEFAULT_FN_ATTRS +_mm_maskz_srav_epi16 (__mmask8 __U, __m128i __A, __m128i __B) +{ + return (__m128i) __builtin_ia32_psrav8hi_mask ((__v8hi) __A, + (__v8hi) __B, + (__v8hi) + _mm_setzero_si128 (), + (__mmask8) __U); +} + #undef __DEFAULT_FN_ATTRS #endif /* __AVX512VLBWINTRIN_H */ diff --git a/lib/Headers/avx512vlintrin.h b/lib/Headers/avx512vlintrin.h index 3f45770a32..047ec9ca39 100644 --- a/lib/Headers/avx512vlintrin.h +++ b/lib/Headers/avx512vlintrin.h @@ -5772,6 +5772,105 @@ __builtin_ia32_psrlqi256_mask ((__v4di)( __A),( __imm),\ (__mmask8)( __U));\ }) +static __inline__ __m128i __DEFAULT_FN_ATTRS +_mm_mask_srav_epi32 (__m128i __W, __mmask8 __U, __m128i __X, + __m128i __Y) +{ + return (__m128i) __builtin_ia32_psrav4si_mask ((__v4si) __X, + (__v4si) __Y, + (__v4si) __W, + (__mmask8) __U); +} + +static __inline__ __m128i __DEFAULT_FN_ATTRS +_mm_maskz_srav_epi32 (__mmask8 __U, __m128i __X, __m128i __Y) +{ + return (__m128i) __builtin_ia32_psrav4si_mask ((__v4si) __X, + (__v4si) __Y, + (__v4si) + _mm_setzero_si128 (), + (__mmask8) __U); +} + +static __inline__ __m256i __DEFAULT_FN_ATTRS +_mm256_mask_srav_epi32 (__m256i __W, __mmask8 __U, __m256i __X, + __m256i __Y) +{ + return (__m256i) __builtin_ia32_psrav8si_mask ((__v8si) __X, + (__v8si) __Y, + (__v8si) __W, + (__mmask8) __U); +} + +static __inline__ __m256i __DEFAULT_FN_ATTRS +_mm256_maskz_srav_epi32 (__mmask8 __U, __m256i __X, __m256i __Y) +{ + return (__m256i) __builtin_ia32_psrav8si_mask ((__v8si) __X, + (__v8si) __Y, + (__v8si) + _mm256_setzero_si256 (), + (__mmask8) __U); +} + +static __inline__ __m128i __DEFAULT_FN_ATTRS +_mm_srav_epi64 (__m128i __X, __m128i __Y) +{ + return (__m128i) __builtin_ia32_psravq128_mask ((__v2di) __X, + (__v2di) __Y, + (__v2di) + _mm_setzero_di (), + (__mmask8) -1); +} + +static __inline__ __m128i __DEFAULT_FN_ATTRS +_mm_mask_srav_epi64 (__m128i __W, __mmask8 __U, __m128i __X, + __m128i __Y) +{ + return (__m128i) __builtin_ia32_psravq128_mask ((__v2di) __X, + (__v2di) __Y, + (__v2di) __W, + (__mmask8) __U); +} + +static __inline__ __m128i __DEFAULT_FN_ATTRS +_mm_maskz_srav_epi64 (__mmask8 __U, __m128i __X, __m128i __Y) +{ + return (__m128i) __builtin_ia32_psravq128_mask ((__v2di) __X, + (__v2di) __Y, + (__v2di) + _mm_setzero_di (), + (__mmask8) __U); +} + +static __inline__ __m256i __DEFAULT_FN_ATTRS +_mm256_srav_epi64 (__m256i __X, __m256i __Y) +{ + return (__m256i) __builtin_ia32_psravq256_mask ((__v4di) __X, + (__v4di) __Y, + (__v4di) + _mm256_setzero_si256 (), + (__mmask8) -1); +} + +static __inline__ __m256i __DEFAULT_FN_ATTRS +_mm256_mask_srav_epi64 (__m256i __W, __mmask8 __U, __m256i __X, + __m256i __Y) +{ + return (__m256i) __builtin_ia32_psravq256_mask ((__v4di) __X, + (__v4di) __Y, + (__v4di) __W, + (__mmask8) __U); +} + +static __inline__ __m256i __DEFAULT_FN_ATTRS +_mm256_maskz_srav_epi64 (__mmask8 __U, __m256i __X, __m256i __Y) +{ + return (__m256i) __builtin_ia32_psravq256_mask ((__v4di) __X, + (__v4di) __Y, + (__v4di) + _mm256_setzero_si256 (), + (__mmask8) __U); +} #undef __DEFAULT_FN_ATTRS #undef __DEFAULT_FN_ATTRS_BOTH diff --git a/test/CodeGen/avx512bw-builtins.c b/test/CodeGen/avx512bw-builtins.c index 4fceab7058..1f10109c60 100644 --- a/test/CodeGen/avx512bw-builtins.c +++ b/test/CodeGen/avx512bw-builtins.c @@ -1185,4 +1185,22 @@ __m512i test_mm512_maskz_srlv_epi16(__mmask32 __U, __m512i __A, __m512i __B) { return _mm512_maskz_srlv_epi16(__U, __A, __B); } +__m512i test_mm512_srav_epi16(__m512i __A, __m512i __B) { + // CHECK-LABEL: @test_mm512_srav_epi16 + // CHECK: @llvm.x86.avx512.mask.psrav + return _mm512_srav_epi16(__A, __B); +} + +__m512i test_mm512_mask_srav_epi16(__m512i __W, __mmask32 __U, __m512i __A, __m512i __B) { + // CHECK-LABEL: @test_mm512_mask_srav_epi16 + // CHECK: @llvm.x86.avx512.mask.psrav + return _mm512_mask_srav_epi16(__W, __U, __A, __B); +} + +__m512i test_mm512_maskz_srav_epi16(__mmask32 __U, __m512i __A, __m512i __B) { + // CHECK-LABEL: @test_mm512_maskz_srav_epi16 + // CHECK: @llvm.x86.avx512.mask.psrav + return _mm512_maskz_srav_epi16(__U, __A, __B); +} + diff --git a/test/CodeGen/avx512vl-builtins.c b/test/CodeGen/avx512vl-builtins.c index e28653b56f..7061369940 100644 --- a/test/CodeGen/avx512vl-builtins.c +++ b/test/CodeGen/avx512vl-builtins.c @@ -3876,3 +3876,63 @@ __m256i test_mm256_maskz_srli_epi64(__mmask8 __U, __m256i __A) { return _mm256_maskz_srli_epi64(__U, __A, 5); } +__m128i test_mm_mask_srav_epi32(__m128i __W, __mmask8 __U, __m128i __X, __m128i __Y) { + // CHECK-LABEL: @test_mm_mask_srav_epi32 + // CHECK: @llvm.x86.avx512.mask.psrav + return _mm_mask_srav_epi32(__W, __U, __X, __Y); +} + +__m128i test_mm_maskz_srav_epi32(__mmask8 __U, __m128i __X, __m128i __Y) { + // CHECK-LABEL: @test_mm_maskz_srav_epi32 + // CHECK: @llvm.x86.avx512.mask.psrav + return _mm_maskz_srav_epi32(__U, __X, __Y); +} + +__m256i test_mm256_mask_srav_epi32(__m256i __W, __mmask8 __U, __m256i __X, __m256i __Y) { + // CHECK-LABEL: @test_mm256_mask_srav_epi32 + // CHECK: @llvm.x86.avx512.mask.psrav + return _mm256_mask_srav_epi32(__W, __U, __X, __Y); +} + +__m256i test_mm256_maskz_srav_epi32(__mmask8 __U, __m256i __X, __m256i __Y) { + // CHECK-LABEL: @test_mm256_maskz_srav_epi32 + // CHECK: @llvm.x86.avx512.mask.psrav + return _mm256_maskz_srav_epi32(__U, __X, __Y); +} + +__m128i test_mm_srav_epi64(__m128i __X, __m128i __Y) { + // CHECK-LABEL: @test_mm_srav_epi64 + // CHECK: @llvm.x86.avx512.mask.psrav.q.128 + return _mm_srav_epi64(__X, __Y); +} + +__m128i test_mm_mask_srav_epi64(__m128i __W, __mmask8 __U, __m128i __X, __m128i __Y) { + // CHECK-LABEL: @test_mm_mask_srav_epi64 + // CHECK: @llvm.x86.avx512.mask.psrav.q.128 + return _mm_mask_srav_epi64(__W, __U, __X, __Y); +} + +__m128i test_mm_maskz_srav_epi64(__mmask8 __U, __m128i __X, __m128i __Y) { + // CHECK-LABEL: @test_mm_maskz_srav_epi64 + // CHECK: @llvm.x86.avx512.mask.psrav.q.128 + return _mm_maskz_srav_epi64(__U, __X, __Y); +} + +__m256i test_mm256_srav_epi64(__m256i __X, __m256i __Y) { + // CHECK-LABEL: @test_mm256_srav_epi64 + // CHECK: @llvm.x86.avx512.mask.psrav.q.256 + return _mm256_srav_epi64(__X, __Y); +} + +__m256i test_mm256_mask_srav_epi64(__m256i __W, __mmask8 __U, __m256i __X, __m256i __Y) { + // CHECK-LABEL: @test_mm256_mask_srav_epi64 + // CHECK: @llvm.x86.avx512.mask.psrav.q.256 + return _mm256_mask_srav_epi64(__W, __U, __X, __Y); +} + +__m256i test_mm256_maskz_srav_epi64(__mmask8 __U, __m256i __X, __m256i __Y) { + // CHECK-LABEL: @test_mm256_maskz_srav_epi64 + // CHECK: @llvm.x86.avx512.mask.psrav.q.256 + return _mm256_maskz_srav_epi64(__U, __X, __Y); +} + -- 2.50.1