From: Michael Zuckerman Date: Sun, 10 Apr 2016 17:24:03 +0000 (+0000) Subject: [Clang][AVX512][BuiltIn] Adding avx512 ( vcvt ) builtin to clang X-Git-Url: https://granicus.if.org/sourcecode?a=commitdiff_plain;h=b41d1d15a0bf13e3d6d807c2fbdfff81845dd5b4;p=clang [Clang][AVX512][BuiltIn] Adding avx512 ( vcvt ) builtin to clang Differential Revision: http://reviews.llvm.org/D18932 git-svn-id: https://llvm.org/svn/llvm-project/cfe/trunk@265904 91177308-0d34-0410-b5e6-96231b3b80d8 --- diff --git a/include/clang/Basic/BuiltinsX86.def b/include/clang/Basic/BuiltinsX86.def index 94d1cedd0e..ba66843760 100644 --- a/include/clang/Basic/BuiltinsX86.def +++ b/include/clang/Basic/BuiltinsX86.def @@ -1832,6 +1832,22 @@ TARGET_BUILTIN(__builtin_ia32_vplzcntd_128_mask, "V4iV4iV4iUc","","avx512vl") TARGET_BUILTIN(__builtin_ia32_vplzcntd_256_mask, "V8iV8iV8iUc","","avx512vl") TARGET_BUILTIN(__builtin_ia32_vplzcntq_128_mask, "V2LLiV2LLiV2LLiUc","","avx512vl") TARGET_BUILTIN(__builtin_ia32_vplzcntq_256_mask, "V4LLiV4LLiV4LLiUc","","avx512vl") +TARGET_BUILTIN(__builtin_ia32_vcvtsd2si64, "LLiV2dIi","","avx512f") +TARGET_BUILTIN(__builtin_ia32_vcvtsd2si32, "iV2dIi","","avx512f") +TARGET_BUILTIN(__builtin_ia32_vcvtsd2usi32, "UiV2dIi","","avx512f") +TARGET_BUILTIN(__builtin_ia32_vcvtsd2usi64, "ULLiV2dIi","","avx512f") +TARGET_BUILTIN(__builtin_ia32_vcvtss2si32, "iV4fIi","","avx512f") +TARGET_BUILTIN(__builtin_ia32_vcvtss2si64, "LLiV4fIi","","avx512f") +TARGET_BUILTIN(__builtin_ia32_vcvtss2usi32, "UiV4fIi","","avx512f") +TARGET_BUILTIN(__builtin_ia32_vcvtss2usi64, "ULLiV4fIi","","avx512f") +TARGET_BUILTIN(__builtin_ia32_vcvttsd2si32, "iV2dIi","","avx512f") +TARGET_BUILTIN(__builtin_ia32_vcvttsd2si64, "LLiV2dIi","","avx512f") +TARGET_BUILTIN(__builtin_ia32_vcvttsd2usi32, "UiV2dIi","","avx512f") +TARGET_BUILTIN(__builtin_ia32_vcvttsd2usi64, "ULLiV2dIi","","avx512f") +TARGET_BUILTIN(__builtin_ia32_vcvttss2si32, "iV4fIi","","avx512f") +TARGET_BUILTIN(__builtin_ia32_vcvttss2si64, "LLiV4fIi","","avx512f") +TARGET_BUILTIN(__builtin_ia32_vcvttss2usi32, "UiV4fIi","","avx512f") +TARGET_BUILTIN(__builtin_ia32_vcvttss2usi64, "ULLiV4fIi","","avx512f") #undef BUILTIN #undef TARGET_BUILTIN diff --git a/lib/Headers/avx512fintrin.h b/lib/Headers/avx512fintrin.h index ca949769c4..23f5ac1b33 100644 --- a/lib/Headers/avx512fintrin.h +++ b/lib/Headers/avx512fintrin.h @@ -3982,6 +3982,9 @@ _mm512_mask_unpackhi_pd (__m512d __W, __mmask8 __U, __m512d __A, __m512d __B) (__v8df) __W, (__mmask8) __U); } +#define _mm_cvt_roundsd_si64( __A, __R) __extension__ ({ \ +__builtin_ia32_vcvtsd2si64 ((__v2df)( __A),( __R));\ +}) static __inline__ __m512d __DEFAULT_FN_ATTRS _mm512_maskz_unpackhi_pd (__mmask8 __U, __m512d __A, __m512d __B) @@ -4050,6 +4053,186 @@ _mm512_maskz_unpacklo_ps (__mmask16 __U, __m512 __A, __m512 __B) (__mmask16) __U); } +#define _mm_cvt_roundsd_i64( __A, __R) __extension__ ({ \ +__builtin_ia32_vcvtsd2si64 ((__v2df)( __A),( __R));\ +}) + +#define _mm_cvt_roundsd_si32( __A, __R) __extension__ ({ \ +__builtin_ia32_vcvtsd2si32 ((__v2df)( __A),( __R));\ +}) + +#define _mm_cvt_roundsd_i32( __A, __R) __extension__ ({ \ +__builtin_ia32_vcvtsd2si32 ((__v2df)( __A),( __R));\ +}) + +#define _mm_cvt_roundsd_u32( __A, __R) __extension__ ({ \ +__builtin_ia32_vcvtsd2usi32 ((__v2df)( __A),( __R));\ +}) + +static __inline__ unsigned __DEFAULT_FN_ATTRS +_mm_cvtsd_u32 (__m128d __A) +{ + return (unsigned) __builtin_ia32_vcvtsd2usi32 ((__v2df) __A, + _MM_FROUND_CUR_DIRECTION); +} + +#define _mm_cvt_roundsd_u64( __A, __R) __extension__ ({ \ +__builtin_ia32_vcvtsd2usi64 ((__v2df)( __A),( __R));\ +}) + +static __inline__ unsigned long long __DEFAULT_FN_ATTRS +_mm_cvtsd_u64 (__m128d __A) +{ + return (unsigned long long) __builtin_ia32_vcvtsd2usi64 ((__v2df) + __A, + _MM_FROUND_CUR_DIRECTION); +} + +#define _mm_cvt_roundss_si32( __A, __R) __extension__ ({ \ +__builtin_ia32_vcvtss2si32 ((__v4sf)( __A),( __R));\ +}) + +#define _mm_cvt_roundss_i32( __A, __R) __extension__ ({ \ +__builtin_ia32_vcvtss2si32 ((__v4sf)( __A),( __R));\ +}) + +#define _mm_cvt_roundss_si64( __A, __R) __extension__ ({ \ +__builtin_ia32_vcvtss2si64 ((__v4sf)( __A),( __R));\ +}) + +#define _mm_cvt_roundss_i64( __A, __R) __extension__ ({ \ +__builtin_ia32_vcvtss2si64 ((__v4sf)( __A),( __R));\ +}) + +#define _mm_cvt_roundss_u32( __A, __R) __extension__ ({ \ +__builtin_ia32_vcvtss2usi32 ((__v4sf)( __A),( __R));\ +}) + +static __inline__ unsigned __DEFAULT_FN_ATTRS +_mm_cvtss_u32 (__m128 __A) +{ + return (unsigned) __builtin_ia32_vcvtss2usi32 ((__v4sf) __A, + _MM_FROUND_CUR_DIRECTION); +} + +#define _mm_cvt_roundss_u64( __A, __R) __extension__ ({ \ +__builtin_ia32_vcvtss2usi64 ((__v4sf)( __A),( __R));\ +}) + +static __inline__ unsigned long long __DEFAULT_FN_ATTRS +_mm_cvtss_u64 (__m128 __A) +{ + return (unsigned long long) __builtin_ia32_vcvtss2usi64 ((__v4sf) + __A, + _MM_FROUND_CUR_DIRECTION); +} + +#define _mm_cvtt_roundsd_i32( __A, __R) __extension__ ({ \ +__builtin_ia32_vcvttsd2si32 ((__v2df)( __A),( __R));\ +}) + +#define _mm_cvtt_roundsd_si32( __A, __R) __extension__ ({ \ +__builtin_ia32_vcvttsd2si32 ((__v2df)( __A),( __R));\ +}) + +static __inline__ int __DEFAULT_FN_ATTRS +_mm_cvttsd_i32 (__m128d __A) +{ + return (int) __builtin_ia32_vcvttsd2si32 ((__v2df) __A, + _MM_FROUND_CUR_DIRECTION); +} + +#define _mm_cvtt_roundsd_si64( __A, __R) __extension__ ({ \ +__builtin_ia32_vcvttsd2si64 ((__v2df)( __A),( __R));\ +}) + +#define _mm_cvtt_roundsd_i64( __A, __R) __extension__ ({ \ +__builtin_ia32_vcvttsd2si64 ((__v2df)( __A),( __R));\ +}) + +static __inline__ long long __DEFAULT_FN_ATTRS +_mm_cvttsd_i64 (__m128d __A) +{ + return (long long) __builtin_ia32_vcvttsd2si64 ((__v2df) __A, + _MM_FROUND_CUR_DIRECTION); +} + +#define _mm_cvtt_roundsd_u32( __A, __R) __extension__ ({ \ +__builtin_ia32_vcvttsd2usi32 ((__v2df)( __A),( __R));\ +}) + +static __inline__ unsigned __DEFAULT_FN_ATTRS +_mm_cvttsd_u32 (__m128d __A) +{ + return (unsigned) __builtin_ia32_vcvttsd2usi32 ((__v2df) __A, + _MM_FROUND_CUR_DIRECTION); +} + +#define _mm_cvtt_roundsd_u64( __A, __R) __extension__ ({ \ +__builtin_ia32_vcvttsd2usi64 ((__v2df)( __A),( __R));\ +}) + +static __inline__ unsigned long long __DEFAULT_FN_ATTRS +_mm_cvttsd_u64 (__m128d __A) +{ + return (unsigned long long) __builtin_ia32_vcvttsd2usi64 ((__v2df) + __A, + _MM_FROUND_CUR_DIRECTION); +} + +#define _mm_cvtt_roundss_i32( __A, __R) __extension__ ({ \ +__builtin_ia32_vcvttss2si32 ((__v4sf)( __A),( __R));\ +}) + +#define _mm_cvtt_roundss_si32( __A, __R) __extension__ ({ \ +__builtin_ia32_vcvttss2si32 ((__v4sf)( __A),( __R));\ +}) + +static __inline__ int __DEFAULT_FN_ATTRS +_mm_cvttss_i32 (__m128 __A) +{ + return (int) __builtin_ia32_vcvttss2si32 ((__v4sf) __A, + _MM_FROUND_CUR_DIRECTION); +} + +#define _mm_cvtt_roundss_i64( __A, __R) __extension__ ({ \ +__builtin_ia32_vcvttss2si64 ((__v4sf)( __A),( __R));\ +}) + +#define _mm_cvtt_roundss_si64( __A, __R) __extension__ ({ \ +__builtin_ia32_vcvttss2si64 ((__v4sf)( __A),( __R));\ +}) + +static __inline__ long long __DEFAULT_FN_ATTRS +_mm_cvttss_i64 (__m128 __A) +{ + return (long long) __builtin_ia32_vcvttss2si64 ((__v4sf) __A, + _MM_FROUND_CUR_DIRECTION); +} + +#define _mm_cvtt_roundss_u32( __A, __R) __extension__ ({ \ +__builtin_ia32_vcvttss2usi32 ((__v4sf)( __A),( __R));\ +}) + +static __inline__ unsigned __DEFAULT_FN_ATTRS +_mm_cvttss_u32 (__m128 __A) +{ + return (unsigned) __builtin_ia32_vcvttss2usi32 ((__v4sf) __A, + _MM_FROUND_CUR_DIRECTION); +} + +#define _mm_cvtt_roundss_u64( __A, __R) __extension__ ({ \ +__builtin_ia32_vcvttss2usi64 ((__v4sf)( __A),( __R));\ +}) + +static __inline__ unsigned long long __DEFAULT_FN_ATTRS +_mm_cvttss_u64 (__m128 __A) +{ + return (unsigned long long) __builtin_ia32_vcvttss2usi64 ((__v4sf) + __A, + _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 5719f77a36..20b9733c0e 100644 --- a/test/CodeGen/avx512f-builtins.c +++ b/test/CodeGen/avx512f-builtins.c @@ -2581,12 +2581,22 @@ __m512d test_mm512_mask_unpackhi_pd(__m512d __W, __mmask8 __U, __m512d __A, __m5 // CHECK: @llvm.x86.avx512.mask.unpckh.pd.512 return _mm512_mask_unpackhi_pd(__W, __U, __A, __B); } +unsigned long long test_mm_cvt_roundsd_si64(__m128d __A) { + // CHECK-LABEL: @test_mm_cvt_roundsd_si64 + // CHECK: @llvm.x86.avx512.vcvtsd2si64 + return _mm_cvt_roundsd_si64(__A, _MM_FROUND_CUR_DIRECTION); +} __m512d test_mm512_maskz_unpackhi_pd(__mmask8 __U, __m512d __A, __m512d __B) { // CHECK-LABEL: @test_mm512_maskz_unpackhi_pd // CHECK: @llvm.x86.avx512.mask.unpckh.pd.512 return _mm512_maskz_unpackhi_pd(__U, __A, __B); } +long long test_mm_cvt_roundsd_i64(__m128d __A) { + // CHECK-LABEL: @test_mm_cvt_roundsd_i64 + // CHECK: @llvm.x86.avx512.vcvtsd2si64 + return _mm_cvt_roundsd_i64(__A, _MM_FROUND_CUR_DIRECTION); +} __m512 test_mm512_mask_unpackhi_ps(__m512 __W, __mmask16 __U, __m512 __A, __m512 __B) { // CHECK-LABEL: @test_mm512_mask_unpackhi_ps @@ -2623,3 +2633,206 @@ __m512 test_mm512_maskz_unpacklo_ps(__mmask16 __U, __m512 __A, __m512 __B) { // CHECK: @llvm.x86.avx512.mask.unpckl.ps.512 return _mm512_maskz_unpacklo_ps(__U, __A, __B); } +int test_mm_cvt_roundsd_si32(__m128d __A) { + // CHECK-LABEL: @test_mm_cvt_roundsd_si32 + // CHECK: @llvm.x86.avx512.vcvtsd2si32 + return _mm_cvt_roundsd_si32(__A, _MM_FROUND_CUR_DIRECTION); +} + +int test_mm_cvt_roundsd_i32(__m128d __A) { + // CHECK-LABEL: @test_mm_cvt_roundsd_i32 + // CHECK: @llvm.x86.avx512.vcvtsd2si32 + return _mm_cvt_roundsd_i32(__A, _MM_FROUND_CUR_DIRECTION); +} + +unsigned test_mm_cvt_roundsd_u32(__m128d __A) { + // CHECK-LABEL: @test_mm_cvt_roundsd_u32 + // CHECK: @llvm.x86.avx512.vcvtsd2usi32 + return _mm_cvt_roundsd_u32(__A, _MM_FROUND_CUR_DIRECTION); +} + +unsigned test_mm_cvtsd_u32(__m128d __A) { + // CHECK-LABEL: @test_mm_cvtsd_u32 + // CHECK: @llvm.x86.avx512.vcvtsd2usi32 + return _mm_cvtsd_u32(__A); +} + +unsigned long long test_mm_cvt_roundsd_u64(__m128d __A) { + // CHECK-LABEL: @test_mm_cvt_roundsd_u64 + // CHECK: @llvm.x86.avx512.vcvtsd2usi64 + return _mm_cvt_roundsd_u64(__A, _MM_FROUND_CUR_DIRECTION); +} + +unsigned long long test_mm_cvtsd_u64(__m128d __A) { + // CHECK-LABEL: @test_mm_cvtsd_u64 + // CHECK: @llvm.x86.avx512.vcvtsd2usi64 + return _mm_cvtsd_u64(__A); +} + +int test_mm_cvt_roundss_si32(__m128 __A) { + // CHECK-LABEL: @test_mm_cvt_roundss_si32 + // CHECK: @llvm.x86.avx512.vcvtss2si32 + return _mm_cvt_roundss_si32(__A, _MM_FROUND_CUR_DIRECTION); +} + +int test_mm_cvt_roundss_i32(__m128 __A) { + // CHECK-LABEL: @test_mm_cvt_roundss_i32 + // CHECK: @llvm.x86.avx512.vcvtss2si32 + return _mm_cvt_roundss_i32(__A, _MM_FROUND_CUR_DIRECTION); +} + +int test_mm_cvt_roundss_si64(__m128 __A) { + // CHECK-LABEL: @test_mm_cvt_roundss_si64 + // CHECK: @llvm.x86.avx512.vcvtss2si64 + return _mm_cvt_roundss_si64(__A, _MM_FROUND_CUR_DIRECTION); +} + +long long test_mm_cvt_roundss_i64(__m128 __A) { + // CHECK-LABEL: @test_mm_cvt_roundss_i64 + // CHECK: @llvm.x86.avx512.vcvtss2si64 + return _mm_cvt_roundss_i64(__A, _MM_FROUND_CUR_DIRECTION); +} + +unsigned test_mm_cvt_roundss_u32(__m128 __A) { + // CHECK-LABEL: @test_mm_cvt_roundss_u32 + // CHECK: @llvm.x86.avx512.vcvtss2usi32 + return _mm_cvt_roundss_u32(__A, _MM_FROUND_CUR_DIRECTION); +} + +unsigned test_mm_cvtss_u32(__m128 __A) { + // CHECK-LABEL: @test_mm_cvtss_u32 + // CHECK: @llvm.x86.avx512.vcvtss2usi32 + return _mm_cvtss_u32(__A); +} + +unsigned long long test_mm_cvt_roundss_u64(__m128 __A) { + // CHECK-LABEL: @test_mm_cvt_roundss_u64 + // CHECK: @llvm.x86.avx512.vcvtss2usi64 + return _mm_cvt_roundss_u64(__A, _MM_FROUND_CUR_DIRECTION); +} + +unsigned long long test_mm_cvtss_u64(__m128 __A) { + // CHECK-LABEL: @test_mm_cvtss_u64 + // CHECK: @llvm.x86.avx512.vcvtss2usi64 + return _mm_cvtss_u64(__A); +} + +int test_mm_cvtt_roundsd_i32(__m128d __A) { + // CHECK-LABEL: @test_mm_cvtt_roundsd_i32 + // CHECK: @llvm.x86.avx512.cvttsd2si + return _mm_cvtt_roundsd_i32(__A, _MM_FROUND_CUR_DIRECTION); +} + +int test_mm_cvtt_roundsd_si32(__m128d __A) { + // CHECK-LABEL: @test_mm_cvtt_roundsd_si32 + // CHECK: @llvm.x86.avx512.cvttsd2si + return _mm_cvtt_roundsd_si32(__A, _MM_FROUND_CUR_DIRECTION); +} + +int test_mm_cvttsd_i32(__m128d __A) { + // CHECK-LABEL: @test_mm_cvttsd_i32 + // CHECK: @llvm.x86.avx512.cvttsd2si + return _mm_cvttsd_i32(__A); +} + +unsigned long long test_mm_cvtt_roundsd_si64(__m128d __A) { + // CHECK-LABEL: @test_mm_cvtt_roundsd_si64 + // CHECK: @llvm.x86.avx512.cvttsd2si64 + return _mm_cvtt_roundsd_si64(__A, _MM_FROUND_CUR_DIRECTION); +} + +long long test_mm_cvtt_roundsd_i64(__m128d __A) { + // CHECK-LABEL: @test_mm_cvtt_roundsd_i64 + // CHECK: @llvm.x86.avx512.cvttsd2si64 + return _mm_cvtt_roundsd_i64(__A, _MM_FROUND_CUR_DIRECTION); +} + +long long test_mm_cvttsd_i64(__m128d __A) { + // CHECK-LABEL: @test_mm_cvttsd_i64 + // CHECK: @llvm.x86.avx512.cvttsd2si64 + return _mm_cvttsd_i64(__A); +} + +unsigned test_mm_cvtt_roundsd_u32(__m128d __A) { + // CHECK-LABEL: @test_mm_cvtt_roundsd_u32 + // CHECK: @llvm.x86.avx512.cvttsd2usi + return _mm_cvtt_roundsd_u32(__A, _MM_FROUND_CUR_DIRECTION); +} + +unsigned test_mm_cvttsd_u32(__m128d __A) { + // CHECK-LABEL: @test_mm_cvttsd_u32 + // CHECK: @llvm.x86.avx512.cvttsd2usi + return _mm_cvttsd_u32(__A); +} + +unsigned long long test_mm_cvtt_roundsd_u64(__m128d __A) { + // CHECK-LABEL: @test_mm_cvtt_roundsd_u64 + // CHECK: @llvm.x86.avx512.cvttsd2usi64 + return _mm_cvtt_roundsd_u64(__A, _MM_FROUND_CUR_DIRECTION); +} + +unsigned long long test_mm_cvttsd_u64(__m128d __A) { + // CHECK-LABEL: @test_mm_cvttsd_u64 + // CHECK: @llvm.x86.avx512.cvttsd2usi64 + return _mm_cvttsd_u64(__A); +} + +int test_mm_cvtt_roundss_i32(__m128 __A) { + // CHECK-LABEL: @test_mm_cvtt_roundss_i32 + // CHECK: @llvm.x86.avx512.cvttss2si + return _mm_cvtt_roundss_i32(__A, _MM_FROUND_CUR_DIRECTION); +} + +int test_mm_cvtt_roundss_si32(__m128 __A) { + // CHECK-LABEL: @test_mm_cvtt_roundss_si32 + // CHECK: @llvm.x86.avx512.cvttss2si + return _mm_cvtt_roundss_si32(__A, _MM_FROUND_CUR_DIRECTION); +} + +int test_mm_cvttss_i32(__m128 __A) { + // CHECK-LABEL: @test_mm_cvttss_i32 + // CHECK: @llvm.x86.avx512.cvttss2si + return _mm_cvttss_i32(__A); +} + +float test_mm_cvtt_roundss_i64(__m128 __A) { + // CHECK-LABEL: @test_mm_cvtt_roundss_i64 + // CHECK: @llvm.x86.avx512.cvttss2si64 + return _mm_cvtt_roundss_i64(__A, _MM_FROUND_CUR_DIRECTION); +} + +long long test_mm_cvtt_roundss_si64(__m128 __A) { + // CHECK-LABEL: @test_mm_cvtt_roundss_si64 + // CHECK: @llvm.x86.avx512.cvttss2si64 + return _mm_cvtt_roundss_si64(__A, _MM_FROUND_CUR_DIRECTION); +} + +long long test_mm_cvttss_i64(__m128 __A) { + // CHECK-LABEL: @test_mm_cvttss_i64 + // CHECK: @llvm.x86.avx512.cvttss2si64 + return _mm_cvttss_i64(__A); +} + +unsigned test_mm_cvtt_roundss_u32(__m128 __A) { + // CHECK-LABEL: @test_mm_cvtt_roundss_u32 + // CHECK: @llvm.x86.avx512.cvttss2usi + return _mm_cvtt_roundss_u32(__A, _MM_FROUND_CUR_DIRECTION); +} + +unsigned test_mm_cvttss_u32(__m128 __A) { + // CHECK-LABEL: @test_mm_cvttss_u32 + // CHECK: @llvm.x86.avx512.cvttss2usi + return _mm_cvttss_u32(__A); +} + +unsigned long long test_mm_cvtt_roundss_u64(__m128 __A) { + // CHECK-LABEL: @test_mm_cvtt_roundss_u64 + // CHECK: @llvm.x86.avx512.cvttss2usi64 + return _mm_cvtt_roundss_u64(__A, _MM_FROUND_CUR_DIRECTION); +} + +unsigned long long test_mm_cvttss_u64(__m128 __A) { + // CHECK-LABEL: @test_mm_cvttss_u64 + // CHECK: @llvm.x86.avx512.cvttss2usi64 + return _mm_cvttss_u64(__A); +}