Index: include/clang/Basic/BuiltinsX86.def =================================================================== --- include/clang/Basic/BuiltinsX86.def +++ include/clang/Basic/BuiltinsX86.def @@ -1942,6 +1942,22 @@ TARGET_BUILTIN(__builtin_ia32_pternlogq128_maskz, "V2LLiV2LLiV2LLiV2LLiIiUc","","avx512vl") TARGET_BUILTIN(__builtin_ia32_pternlogq256_mask, "V4LLiV4LLiV4LLiV4LLiIiUc","","avx512vl") TARGET_BUILTIN(__builtin_ia32_pternlogq256_maskz, "V4LLiV4LLiV4LLiV4LLiIiUc","","avx512vl") +TARGET_BUILTIN(__builtin_ia32_cvtw2mask512, "UiV32s","","avx512bw") +TARGET_BUILTIN(__builtin_ia32_cvtsd2ss_round_mask, "V4fV4fV4fV2dUcIi","","avx512f") +TARGET_BUILTIN(__builtin_ia32_cvtsi2sd64, "V2dV2dULLiIi","","avx512f") +TARGET_BUILTIN(__builtin_ia32_cvtsi2ss32, "V4fV4fiIi","","avx512f") +TARGET_BUILTIN(__builtin_ia32_cvtsi2ss64, "V4fV4fULLiIi","","avx512f") +TARGET_BUILTIN(__builtin_ia32_cvtss2sd_round_mask, "V2dV2dV2dV4fUcIi","","avx512f") +TARGET_BUILTIN(__builtin_ia32_cvtusi2sd32, "V2dV2di","","avx512f") +TARGET_BUILTIN(__builtin_ia32_cvtusi2sd64, "V2dV2ddIi","","avx512f") +TARGET_BUILTIN(__builtin_ia32_cvtusi2ss32, "V4fV4fiIi","","avx512f") +TARGET_BUILTIN(__builtin_ia32_cvtusi2ss64, "V4fV4fdIi","","avx512f") +TARGET_BUILTIN(__builtin_ia32_cvtw2mask128, "UcV8s","","avx512bw,avx512vl") +TARGET_BUILTIN(__builtin_ia32_cvtw2mask256, "UsV16s","","avx512bw,avx512vl") +TARGET_BUILTIN(__builtin_ia32_vcvtph2ps_mask, "V4fV8sV4fUc","","avx512vl") +TARGET_BUILTIN(__builtin_ia32_vcvtph2ps256_mask, "V8fV8sV8fUc","","avx512vl") +TARGET_BUILTIN(__builtin_ia32_vcvtps2ph_mask, "V8sV4fIiV8sUc","","avx512vl") +TARGET_BUILTIN(__builtin_ia32_vcvtps2ph256_mask, "V8sV8fIiV8sUc","","avx512vl") #undef BUILTIN #undef TARGET_BUILTIN Index: lib/Headers/avx512bwintrin.h =================================================================== --- lib/Headers/avx512bwintrin.h +++ lib/Headers/avx512bwintrin.h @@ -2057,6 +2057,11 @@ (__v32hi) __B, __U); } +static __inline__ __mmask32 __DEFAULT_FN_ATTRS +_mm512_movepi16_mask (__m512i __A) +{ + return (__mmask32) __builtin_ia32_cvtw2mask512 ((__v32hi) __A); +} #undef __DEFAULT_FN_ATTRS Index: lib/Headers/avx512fintrin.h =================================================================== --- lib/Headers/avx512fintrin.h +++ lib/Headers/avx512fintrin.h @@ -5306,6 +5306,133 @@ (__mmask8)( __U));\ }) +#define _mm_cvt_roundsd_ss( __A, __B, __R) __extension__ ({ \ +__builtin_ia32_cvtsd2ss_round_mask ((__v4sf)( __A),\ + (__v2df)( __B),\ + (__v4sf) _mm_setzero_ps (),\ + (__mmask8) -1,\ + ( __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_setzero_ps (),\ + (__mmask8)-1,\ + ( __R));\ +}) + +#define _mm512_cvtt_roundpd_epu32( __A, __R) __extension__ ({ \ +__builtin_ia32_cvttpd2udq512_mask ((__v8df)( __A),\ + (__v8si)\ + _mm256_undefined_si256 (),\ + (__mmask8) -1,( __R));\ +}) + +#define _mm512_mask_cvtt_roundpd_epu32( __W, __U, __A, __R) __extension__ ({ \ +__builtin_ia32_cvttpd2udq512_mask ((__v8df)( __A),\ + (__v8si)( __W),\ + (__mmask8)( __U),( __R));\ +}) + +#define _mm512_maskz_cvtt_roundpd_epu32( __U, __A, __R) __extension__ ({ \ +__builtin_ia32_cvttpd2udq512_mask ((__v8df)( __A),\ + (__v8si)\ + _mm256_setzero_si256 (),\ + (__mmask8)( __U),( __R));\ +}) + +static __inline__ __m256i __DEFAULT_FN_ATTRS +_mm512_cvttpd_epu32 (__m512d __A) +{ + return (__m256i) __builtin_ia32_cvttpd2udq512_mask ((__v8df) __A, + (__v8si) + _mm256_undefined_si256 (), + (__mmask8) -1, + _MM_FROUND_CUR_DIRECTION); +} + +static __inline__ __m256i __DEFAULT_FN_ATTRS +_mm512_mask_cvttpd_epu32 (__m256i __W, __mmask8 __U, __m512d __A) +{ + return (__m256i) __builtin_ia32_cvttpd2udq512_mask ((__v8df) __A, + (__v8si) __W, + (__mmask8) __U, + _MM_FROUND_CUR_DIRECTION); +} + +static __inline__ __m256i __DEFAULT_FN_ATTRS +_mm512_maskz_cvttpd_epu32 (__mmask8 __U, __m512d __A) +{ + return (__m256i) __builtin_ia32_cvttpd2udq512_mask ((__v8df) __A, + (__v8si) + _mm256_setzero_si256 (), + (__mmask8) __U, + _MM_FROUND_CUR_DIRECTION); +} + +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 Index: lib/Headers/avx512vlbwintrin.h =================================================================== --- lib/Headers/avx512vlbwintrin.h +++ lib/Headers/avx512vlbwintrin.h @@ -3172,7 +3172,17 @@ (__v16hi) __B, __U); } +static __inline__ __mmask8 __DEFAULT_FN_ATTRS +_mm_movepi16_mask (__m128i __A) +{ + return (__mmask8) __builtin_ia32_cvtw2mask128 ((__v8hi) __A); +} +static __inline__ __mmask16 __DEFAULT_FN_ATTRS +_mm256_movepi16_mask (__m256i __A) +{ + return (__mmask16) __builtin_ia32_cvtw2mask256 ((__v16hi) __A); +} #undef __DEFAULT_FN_ATTRS Index: lib/Headers/avx512vlintrin.h =================================================================== --- lib/Headers/avx512vlintrin.h +++ lib/Headers/avx512vlintrin.h @@ -7495,6 +7495,65 @@ (__mmask8)( __U));\ }) +static __inline__ __m128 __DEFAULT_FN_ATTRS +_mm_mask_cvtph_ps (__m128 __W, __mmask8 __U, __m128i __A) +{ + return (__m128) __builtin_ia32_vcvtph2ps_mask ((__v8hi) __A, + (__v4sf) __W, + (__mmask8) __U); +} + +static __inline__ __m128 __DEFAULT_FN_ATTRS +_mm_maskz_cvtph_ps (__mmask8 __U, __m128i __A) +{ + return (__m128) __builtin_ia32_vcvtph2ps_mask ((__v8hi) __A, + (__v4sf) + _mm_setzero_ps (), + (__mmask8) __U); +} + +static __inline__ __m256 __DEFAULT_FN_ATTRS +_mm256_mask_cvtph_ps (__m256 __W, __mmask8 __U, __m128i __A) +{ + return (__m256) __builtin_ia32_vcvtph2ps256_mask ((__v8hi) __A, + (__v8sf) __W, + (__mmask8) __U); +} + +static __inline__ __m256 __DEFAULT_FN_ATTRS +_mm256_maskz_cvtph_ps (__mmask8 __U, __m128i __A) +{ + return (__m256) __builtin_ia32_vcvtph2ps256_mask ((__v8hi) __A, + (__v8sf) + _mm256_setzero_ps (), + (__mmask8) __U); +} + +#define _mm_mask_cvtps_ph( __W, __U, __A, __I) __extension__ ({ \ +__builtin_ia32_vcvtps2ph_mask ((__v4sf)( __A),( __I),\ + (__v8hi)( __W),\ + (__mmask8)( __U));\ +}) + +#define _mm_maskz_cvtps_ph( __U, __A, __I) __extension__ ({ \ +__builtin_ia32_vcvtps2ph_mask ((__v4sf)( __A),( __I),\ + (__v8hi)\ + _mm_setzero_si128 (),\ + (__mmask8)( __U));\ +}) + +#define _mm256_mask_cvtps_ph( __W, __U, __A, __I) __extension__ ({ \ +__builtin_ia32_vcvtps2ph256_mask ((__v8sf)( __A),( __I),\ + (__v8hi)( __W),\ + (__mmask8)( __U));\ +}) + +#define _mm256_maskz_cvtps_ph( __U, __A, __I) __extension__ ({ \ +__builtin_ia32_vcvtps2ph256_mask ((__v8sf)( __A),( __I),\ + (__v8hi)\ + _mm_setzero_si128 (),\ + (__mmask8)( __U));\ +}) #undef __DEFAULT_FN_ATTRS #undef __DEFAULT_FN_ATTRS_BOTH Index: test/CodeGen/avx512bw-builtins.c =================================================================== --- test/CodeGen/avx512bw-builtins.c +++ test/CodeGen/avx512bw-builtins.c @@ -1404,3 +1404,10 @@ return _mm512_mask_testn_epi16_mask(__U, __A, __B); } +__mmask32 test_mm512_movepi16_mask(__m512i __A) { + // CHECK-LABEL: @test_mm512_movepi16_mask + // CHECK: @llvm.x86.avx512.cvtw2mask.512 + return _mm512_movepi16_mask(__A); +} + + Index: test/CodeGen/avx512f-builtins.c =================================================================== --- test/CodeGen/avx512f-builtins.c +++ test/CodeGen/avx512f-builtins.c @@ -3539,3 +3539,130 @@ // CHECK: @llvm.x86.avx512.maskz.pternlog.q.512 return _mm512_maskz_ternarylogic_epi64(__U, __A, __B, __C, 4); } + +__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); +} + +__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); +} + +__m256i test_mm512_cvtt_roundpd_epu32(__m512d __A) { + // CHECK-LABEL: @test_mm512_cvtt_roundpd_epu32 + // CHECK: @llvm.x86.avx512.mask.cvttpd2udq.512 + return _mm512_cvtt_roundpd_epu32(__A, _MM_FROUND_CUR_DIRECTION); +} + +__m256i test_mm512_mask_cvtt_roundpd_epu32(__m256i __W, __mmask8 __U, __m512d __A) { + // CHECK-LABEL: @test_mm512_mask_cvtt_roundpd_epu32 + // CHECK: @llvm.x86.avx512.mask.cvttpd2udq.512 + return _mm512_mask_cvtt_roundpd_epu32(__W, __U, __A, _MM_FROUND_CUR_DIRECTION); +} + +__m256i test_mm512_maskz_cvtt_roundpd_epu32(__mmask8 __U, __m512d __A) { + // CHECK-LABEL: @test_mm512_maskz_cvtt_roundpd_epu32 + // CHECK: @llvm.x86.avx512.mask.cvttpd2udq.512 + return _mm512_maskz_cvtt_roundpd_epu32(__U, __A, _MM_FROUND_CUR_DIRECTION); +} + +__m256i test_mm512_cvttpd_epu32(__m512d __A) { + // CHECK-LABEL: @test_mm512_cvttpd_epu32 + // CHECK: @llvm.x86.avx512.mask.cvttpd2udq.512 + return _mm512_cvttpd_epu32(__A); +} + +__m256i test_mm512_mask_cvttpd_epu32(__m256i __W, __mmask8 __U, __m512d __A) { + // CHECK-LABEL: @test_mm512_mask_cvttpd_epu32 + // CHECK: @llvm.x86.avx512.mask.cvttpd2udq.512 + return _mm512_mask_cvttpd_epu32(__W, __U, __A); +} + +__m256i test_mm512_maskz_cvttpd_epu32(__mmask8 __U, __m512d __A) { + // CHECK-LABEL: @test_mm512_maskz_cvttpd_epu32 + // CHECK: @llvm.x86.avx512.mask.cvttpd2udq.512 + return _mm512_maskz_cvttpd_epu32(__U, __A); +} + +__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); +} + Index: test/CodeGen/avx512vl-builtins.c =================================================================== --- test/CodeGen/avx512vl-builtins.c +++ test/CodeGen/avx512vl-builtins.c @@ -5087,3 +5087,52 @@ // CHECK: @llvm.x86.avx512.maskz.pternlog.q.256 return _mm256_maskz_ternarylogic_epi64(__U, __A, __B, __C, 4); } + +__m128 test_mm_mask_cvtph_ps(__m128 __W, __mmask8 __U, __m128i __A) { + // CHECK-LABEL: @test_mm_mask_cvtph_ps + // CHECK: @llvm.x86.avx512.mask.vcvtph2ps.128 + return _mm_mask_cvtph_ps(__W, __U, __A); +} + +__m128 test_mm_maskz_cvtph_ps(__mmask8 __U, __m128i __A) { + // CHECK-LABEL: @test_mm_maskz_cvtph_ps + // CHECK: @llvm.x86.avx512.mask.vcvtph2ps.128 + return _mm_maskz_cvtph_ps(__U, __A); +} + +__m256 test_mm256_mask_cvtph_ps(__m256 __W, __mmask8 __U, __m128i __A) { + // CHECK-LABEL: @test_mm256_mask_cvtph_ps + // CHECK: @llvm.x86.avx512.mask.vcvtph2ps.256 + return _mm256_mask_cvtph_ps(__W, __U, __A); +} + +__m256 test_mm256_maskz_cvtph_ps(__mmask8 __U, __m128i __A) { + // CHECK-LABEL: @test_mm256_maskz_cvtph_ps + // CHECK: @llvm.x86.avx512.mask.vcvtph2ps.256 + return _mm256_maskz_cvtph_ps(__U, __A); +} + +__m128i test_mm_mask_cvtps_ph(__m128i __W, __mmask8 __U, __m128 __A) { + // CHECK-LABEL: @test_mm_mask_cvtps_ph + // CHECK: @llvm.x86.avx512.mask.vcvtps2ph.128 + return _mm_mask_cvtps_ph(__W, __U, __A, _MM_FROUND_CUR_DIRECTION); +} + +__m128i test_mm_maskz_cvtps_ph(__mmask8 __U, __m128 __A) { + // CHECK-LABEL: @test_mm_maskz_cvtps_ph + // CHECK: @llvm.x86.avx512.mask.vcvtps2ph.128 + return _mm_maskz_cvtps_ph(__U, __A, _MM_FROUND_CUR_DIRECTION); +} + +__m128i test_mm256_mask_cvtps_ph(__m128i __W, __mmask8 __U, __m256 __A) { + // CHECK-LABEL: @test_mm256_mask_cvtps_ph + // CHECK: @llvm.x86.avx512.mask.vcvtps2ph.256 + return _mm256_mask_cvtps_ph(__W, __U, __A, _MM_FROUND_CUR_DIRECTION); +} + +__m128i test_mm256_maskz_cvtps_ph(__mmask8 __U, __m256 __A) { + // CHECK-LABEL: @test_mm256_maskz_cvtps_ph + // CHECK: @llvm.x86.avx512.mask.vcvtps2ph.256 + return _mm256_maskz_cvtps_ph(__U, __A, _MM_FROUND_CUR_DIRECTION); +} + Index: test/CodeGen/avx512vlbw-builtins.c =================================================================== --- test/CodeGen/avx512vlbw-builtins.c +++ test/CodeGen/avx512vlbw-builtins.c @@ -2172,3 +2172,16 @@ return _mm256_mask_testn_epi16_mask(__U, __A, __B); } +__mmask8 test_mm_movepi16_mask(__m128i __A) { + // CHECK-LABEL: @test_mm_movepi16_mask + // CHECK: @llvm.x86.avx512.cvtw2mask.128 + return _mm_movepi16_mask(__A); +} + +__mmask16 test_mm256_movepi16_mask(__m256i __A) { + // CHECK-LABEL: @test_mm256_movepi16_mask + // CHECK: @llvm.x86.avx512.cvtw2mask.256 + return _mm256_movepi16_mask(__A); +} + +