From 3006132971575978d3b6fb675ae864382dee688a Mon Sep 17 00:00:00 2001 From: Michael Zuckerman Date: Wed, 4 May 2016 08:55:11 +0000 Subject: [PATCH] [Clang][Builtin][AVX512] Adding intrinsics fot cvt{u}si2s{d|s} cvt{sd|ss}2{ss|sd} instruction set Differential Revision: http://reviews.llvm.org/D19765 git-svn-id: https://llvm.org/svn/llvm-project/cfe/trunk@268481 91177308-0d34-0410-b5e6-96231b3b80d8 --- include/clang/Basic/BuiltinsX86.def | 9 +++ lib/Headers/avx512fintrin.h | 111 +++++++++++++++++++++++++++ test/CodeGen/avx512f-builtins.c | 115 ++++++++++++++++++++++++++++ 3 files changed, 235 insertions(+) diff --git a/include/clang/Basic/BuiltinsX86.def b/include/clang/Basic/BuiltinsX86.def index 6379a118bc..d6baa71824 100644 --- a/include/clang/Basic/BuiltinsX86.def +++ b/include/clang/Basic/BuiltinsX86.def @@ -2259,6 +2259,15 @@ TARGET_BUILTIN(__builtin_ia32_vcvtps2ph256_mask, "V8sV8fIiV8sUc","","avx512vl") TARGET_BUILTIN(__builtin_ia32_cvtw2mask512, "UiV32s","","avx512bw") TARGET_BUILTIN(__builtin_ia32_cvtw2mask128, "UcV8s","","avx512bw,avx512vl") TARGET_BUILTIN(__builtin_ia32_cvtw2mask256, "UsV16s","","avx512bw,avx512vl") +TARGET_BUILTIN(__builtin_ia32_cvtsd2ss_round_mask, "V4fV4fV4fV2dUcIi","","avx512f") +TARGET_BUILTIN(__builtin_ia32_cvtsi2sd64, "V2dV2dLLiIi","","avx512f") +TARGET_BUILTIN(__builtin_ia32_cvtsi2ss32, "V4fV4fiIi","","avx512f") +TARGET_BUILTIN(__builtin_ia32_cvtsi2ss64, "V4fV4fLLiIi","","avx512f") +TARGET_BUILTIN(__builtin_ia32_cvtss2sd_round_mask, "V2dV2dV2dV4fUcIi","","avx512f") +TARGET_BUILTIN(__builtin_ia32_cvtusi2sd32, "V2dV2dUi","","avx512f") +TARGET_BUILTIN(__builtin_ia32_cvtusi2sd64, "V2dV2dULLiIi","","avx512f") +TARGET_BUILTIN(__builtin_ia32_cvtusi2ss32, "V4fV4fUiIi","","avx512f") +TARGET_BUILTIN(__builtin_ia32_cvtusi2ss64, "V4fV4fULLiIi","","avx512f") #undef BUILTIN #undef TARGET_BUILTIN diff --git a/lib/Headers/avx512fintrin.h b/lib/Headers/avx512fintrin.h index 7fa318129d..6500bb5ef5 100644 --- a/lib/Headers/avx512fintrin.h +++ b/lib/Headers/avx512fintrin.h @@ -8066,6 +8066,117 @@ _mm512_mask_compressstoreu_epi32 (void *__P, __mmask16 __U, __m512i __A) (__mmask16) __U); } +#define _mm_cvt_roundsd_ss( __A, __B, __R) __extension__ ({ \ +__builtin_ia32_cvtsd2ss_round_mask ((__v4sf)( __A),\ + (__v2df)( __B),\ + (__v4sf) _mm_undefined_ps (),\ + (__mmask8) -1,\ + ( __R));\ +}) + +#define _mm_mask_cvt_roundsd_ss( __W, __U, __A, __B, __R) __extension__ ({ \ +__builtin_ia32_cvtsd2ss_round_mask ((__v4sf)( __A),\ + (__v2df)( __B),\ + (__v4sf) __W,\ + (__mmask8) __U,\ + ( __R));\ +}) + +#define _mm_maskz_cvt_roundsd_ss( __U, __A, __B, __R) __extension__ ({ \ +__builtin_ia32_cvtsd2ss_round_mask ((__v4sf)( __A),\ + (__v2df)( __B),\ + (__v4sf) _mm_setzero_ps (),\ + (__mmask8) __U,\ + ( __R));\ +}) + +#define _mm_cvt_roundi64_sd( __A, __B, __R) __extension__ ({ \ +__builtin_ia32_cvtsi2sd64 ((__v2df)( __A),( __B),( __R));\ +}) + +#define _mm_cvt_roundsi64_sd( __A, __B, __R) __extension__ ({ \ +__builtin_ia32_cvtsi2sd64 ((__v2df)( __A),( __B),( __R));\ +}) + +#define _mm_cvt_roundsi32_ss( __A, __B, __R) __extension__ ({ \ +__builtin_ia32_cvtsi2ss32 ((__v4sf)( __A),( __B),( __R));\ +}) + +#define _mm_cvt_roundi32_ss( __A, __B, __R) __extension__ ({ \ +__builtin_ia32_cvtsi2ss32 ((__v4sf)( __A),( __B),( __R));\ +}) + +#define _mm_cvt_roundsi64_ss( __A, __B, __R) __extension__ ({ \ +__builtin_ia32_cvtsi2ss64 ((__v4sf)( __A),( __B),( __R));\ +}) + +#define _mm_cvt_roundi64_ss( __A, __B, __R) __extension__ ({ \ +__builtin_ia32_cvtsi2ss64 ((__v4sf)( __A),( __B),( __R));\ +}) + +#define _mm_cvt_roundss_sd( __A, __B, __R) __extension__ ({ \ +__builtin_ia32_cvtss2sd_round_mask ((__v2df)( __A),\ + (__v4sf)( __B),\ + (__v2df) _mm_undefined_pd (),\ + (__mmask8)-1,\ + ( __R));\ +}) + +#define _mm_mask_cvt_roundss_sd(__W, __U,__A, __B, __R) __extension__ ({ \ +__builtin_ia32_cvtss2sd_round_mask ((__v2df)( __A),\ + (__v4sf)( __B),\ + (__v2df) __W,\ + (__mmask8) __U,\ + ( __R));\ +}) + +#define _mm_maskz_cvt_roundss_sd( __U,__A, __B, __R) __extension__ ({ \ +__builtin_ia32_cvtss2sd_round_mask ((__v2df)( __A),\ + (__v4sf)( __B),\ + (__v2df) _mm_setzero_pd(),\ + (__mmask8) __U,\ + ( __R));\ +}) + +static __inline__ __m128d __DEFAULT_FN_ATTRS +_mm_cvtu32_sd (__m128d __A, unsigned __B) +{ + return (__m128d) __builtin_ia32_cvtusi2sd32 ((__v2df) __A, __B); +} + +#define _mm_cvt_roundu64_sd( __A, __B, __R) __extension__ ({ \ +__builtin_ia32_cvtusi2sd64 ((__v2df)( __A),( __B),( __R));\ +}) + +static __inline__ __m128d __DEFAULT_FN_ATTRS +_mm_cvtu64_sd (__m128d __A, unsigned long long __B) +{ + return (__m128d) __builtin_ia32_cvtusi2sd64 ((__v2df) __A, __B, + _MM_FROUND_CUR_DIRECTION); +} + +#define _mm_cvt_roundu32_ss( __A, __B, __R) __extension__ ({ \ +__builtin_ia32_cvtusi2ss32 ((__v4sf)( __A),( __B),( __R));\ +}) + +static __inline__ __m128 __DEFAULT_FN_ATTRS +_mm_cvtu32_ss (__m128 __A, unsigned __B) +{ + return (__m128) __builtin_ia32_cvtusi2ss32 ((__v4sf) __A, __B, + _MM_FROUND_CUR_DIRECTION); +} + +#define _mm_cvt_roundu64_ss( __A, __B, __R) __extension__ ({ \ +__builtin_ia32_cvtusi2ss64 ((__v4sf)( __A),( __B),( __R));\ +}) + +static __inline__ __m128 __DEFAULT_FN_ATTRS +_mm_cvtu64_ss (__m128 __A, unsigned long long __B) +{ + return (__m128) __builtin_ia32_cvtusi2ss64 ((__v4sf) __A, __B, + _MM_FROUND_CUR_DIRECTION); +} + #undef __DEFAULT_FN_ATTRS #endif // __AVX512FINTRIN_H diff --git a/test/CodeGen/avx512f-builtins.c b/test/CodeGen/avx512f-builtins.c index b21a3931af..6581ae73d4 100644 --- a/test/CodeGen/avx512f-builtins.c +++ b/test/CodeGen/avx512f-builtins.c @@ -5613,3 +5613,118 @@ __m512i test_mm512_castsi256_si512(__m256i __A) { return _mm512_castsi256_si512(__A); } + +__m128 test_mm_cvt_roundsd_ss(__m128 __A, __m128d __B) { + // CHECK-LABEL: @test_mm_cvt_roundsd_ss + // CHECK: @llvm.x86.avx512.mask.cvtsd2ss.round + return _mm_cvt_roundsd_ss(__A, __B, _MM_FROUND_CUR_DIRECTION); +} + +__m128 test_mm_mask_cvt_roundsd_ss(__m128 __W, __mmask8 __U, __m128 __A, __m128d __B) { + // CHECK-LABEL: @test_mm_mask_cvt_roundsd_ss + // CHECK: @llvm.x86.avx512.mask.cvtsd2ss.round + return _mm_mask_cvt_roundsd_ss(__W, __U, __A, __B, _MM_FROUND_CUR_DIRECTION); +} + +__m128 test_mm_maskz_cvt_roundsd_ss(__mmask8 __U, __m128 __A, __m128d __B) { + // CHECK-LABEL: @test_mm_maskz_cvt_roundsd_ss + // CHECK: @llvm.x86.avx512.mask.cvtsd2ss.round + return _mm_maskz_cvt_roundsd_ss(__U, __A, __B, _MM_FROUND_CUR_DIRECTION); +} + +__m128d test_mm_cvt_roundi64_sd(__m128d __A, long long __B) { + // CHECK-LABEL: @test_mm_cvt_roundi64_sd + // CHECK: @llvm.x86.avx512.cvtsi2sd64 + return _mm_cvt_roundi64_sd(__A, __B, _MM_FROUND_CUR_DIRECTION); +} + +__m128d test_mm_cvt_roundsi64_sd(__m128d __A, long long __B) { + // CHECK-LABEL: @test_mm_cvt_roundsi64_sd + // CHECK: @llvm.x86.avx512.cvtsi2sd64 + return _mm_cvt_roundsi64_sd(__A, __B, _MM_FROUND_CUR_DIRECTION); +} + +__m128 test_mm_cvt_roundsi32_ss(__m128 __A, int __B) { + // CHECK-LABEL: @test_mm_cvt_roundsi32_ss + // CHECK: @llvm.x86.avx512.cvtsi2ss32 + return _mm_cvt_roundsi32_ss(__A, __B, _MM_FROUND_CUR_DIRECTION); +} + +__m128 test_mm_cvt_roundi32_ss(__m128 __A, int __B) { + // CHECK-LABEL: @test_mm_cvt_roundi32_ss + // CHECK: @llvm.x86.avx512.cvtsi2ss32 + return _mm_cvt_roundi32_ss(__A, __B, _MM_FROUND_CUR_DIRECTION); +} + +__m128 test_mm_cvt_roundsi64_ss(__m128 __A, long long __B) { + // CHECK-LABEL: @test_mm_cvt_roundsi64_ss + // CHECK: @llvm.x86.avx512.cvtsi2ss64 + return _mm_cvt_roundsi64_ss(__A, __B, _MM_FROUND_CUR_DIRECTION); +} + +__m128 test_mm_cvt_roundi64_ss(__m128 __A, long long __B) { + // CHECK-LABEL: @test_mm_cvt_roundi64_ss + // CHECK: @llvm.x86.avx512.cvtsi2ss64 + return _mm_cvt_roundi64_ss(__A, __B, _MM_FROUND_CUR_DIRECTION); +} + +__m128d test_mm_cvt_roundss_sd(__m128d __A, __m128 __B) { + // CHECK-LABEL: @test_mm_cvt_roundss_sd + // CHECK: @llvm.x86.avx512.mask.cvtss2sd.round + return _mm_cvt_roundss_sd(__A, __B, _MM_FROUND_CUR_DIRECTION); +} + +__m128d test_mm_mask_cvt_roundss_sd(__m128d __W, __mmask8 __U, __m128d __A, __m128 __B) { + // CHECK-LABEL: @test_mm_mask_cvt_roundss_sd + // CHECK: @llvm.x86.avx512.mask.cvtss2sd.round + return _mm_mask_cvt_roundss_sd(__W, __U, __A, __B, _MM_FROUND_CUR_DIRECTION); +} + +__m128d test_mm_maskz_cvt_roundss_sd( __mmask8 __U, __m128d __A, __m128 __B) { + // CHECK-LABEL: @test_mm_maskz_cvt_roundss_sd + // CHECK: @llvm.x86.avx512.mask.cvtss2sd.round + return _mm_maskz_cvt_roundss_sd( __U, __A, __B, _MM_FROUND_CUR_DIRECTION); +} + +__m128d test_mm_cvtu32_sd(__m128d __A, unsigned __B) { + // CHECK-LABEL: @test_mm_cvtu32_sd + // CHECK: @llvm.x86.avx512.cvtusi2sd + return _mm_cvtu32_sd(__A, __B); +} + +__m128d test_mm_cvt_roundu64_sd(__m128d __A, unsigned long long __B) { + // CHECK-LABEL: @test_mm_cvt_roundu64_sd + // CHECK: @llvm.x86.avx512.cvtusi642sd + return _mm_cvt_roundu64_sd(__A, __B, _MM_FROUND_CUR_DIRECTION); +} + +__m128d test_mm_cvtu64_sd(__m128d __A, unsigned long long __B) { + // CHECK-LABEL: @test_mm_cvtu64_sd + // CHECK: @llvm.x86.avx512.cvtusi642sd + return _mm_cvtu64_sd(__A, __B); +} + +__m128 test_mm_cvt_roundu32_ss(__m128 __A, unsigned __B) { + // CHECK-LABEL: @test_mm_cvt_roundu32_ss + // CHECK: @llvm.x86.avx512.cvtusi2ss + return _mm_cvt_roundu32_ss(__A, __B, _MM_FROUND_CUR_DIRECTION); +} + +__m128 test_mm_cvtu32_ss(__m128 __A, unsigned __B) { + // CHECK-LABEL: @test_mm_cvtu32_ss + // CHECK: @llvm.x86.avx512.cvtusi2ss + return _mm_cvtu32_ss(__A, __B); +} + +__m128 test_mm_cvt_roundu64_ss(__m128 __A, unsigned long long __B) { + // CHECK-LABEL: @test_mm_cvt_roundu64_ss + // CHECK: @llvm.x86.avx512.cvtusi642ss + return _mm_cvt_roundu64_ss(__A, __B, _MM_FROUND_CUR_DIRECTION); +} + +__m128 test_mm_cvtu64_ss(__m128 __A, unsigned long long __B) { + // CHECK-LABEL: @test_mm_cvtu64_ss + // CHECK: @llvm.x86.avx512.cvtusi642ss + return _mm_cvtu64_ss(__A, __B); +} + -- 2.40.0