https://github.com/chaitanyav updated https://github.com/llvm/llvm-project/pull/161056
>From 83f379bd8ea7b079411a68b1827bd7130cf5ec1e Mon Sep 17 00:00:00 2001 From: NagaChaitanya Vellanki <[email protected]> Date: Sun, 28 Sep 2025 00:29:57 -0700 Subject: [PATCH] [X86][Clang] VectorExprEvaluator::VisitCallExpr / InterpretBuiltin - Allow AVX/AVX512 IFMA madd52 intrinsics to be used in constexpr Resolves #160498 --- clang/include/clang/Basic/BuiltinsX86.td | 14 +- clang/lib/AST/ByteCode/InterpBuiltin.cpp | 15 ++ clang/lib/AST/ExprConstant.cpp | 46 ++++ clang/lib/Headers/avx512ifmaintrin.h | 63 +++-- clang/lib/Headers/avx512ifmavlintrin.h | 43 +++- clang/lib/Headers/avxifmaintrin.h | 18 +- clang/test/CodeGen/X86/avx512ifma-builtins.c | 224 +++++++++++++++++- .../test/CodeGen/X86/avx512ifmavl-builtins.c | 212 +++++++++++++++-- clang/test/CodeGen/X86/avxifma-builtins.c | 158 +++++++++--- 9 files changed, 667 insertions(+), 126 deletions(-) diff --git a/clang/include/clang/Basic/BuiltinsX86.td b/clang/include/clang/Basic/BuiltinsX86.td index e98bee28c15be..9a0919739d2bd 100644 --- a/clang/include/clang/Basic/BuiltinsX86.td +++ b/clang/include/clang/Basic/BuiltinsX86.td @@ -2101,24 +2101,18 @@ let Features = "avx512vl", Attributes = [NoThrow, RequiredVectorWidth<256>] in { def movdqa64store256_mask : X86Builtin<"void(_Vector<4, long long int *>, _Vector<4, long long int>, unsigned char)">; } -let Features = "avx512ifma", Attributes = [NoThrow, Const, RequiredVectorWidth<512>] in { +let Features = "avx512ifma", Attributes = [NoThrow, Const, Constexpr, RequiredVectorWidth<512>] in { def vpmadd52huq512 : X86Builtin<"_Vector<8, long long int>(_Vector<8, long long int>, _Vector<8, long long int>, _Vector<8, long long int>)">; def vpmadd52luq512 : X86Builtin<"_Vector<8, long long int>(_Vector<8, long long int>, _Vector<8, long long int>, _Vector<8, long long int>)">; } -let Features = "avx512ifma,avx512vl|avxifma", Attributes = [NoThrow, Const, RequiredVectorWidth<128>] in { +let Features = "avx512ifma,avx512vl|avxifma", Attributes = [NoThrow, Const, Constexpr, RequiredVectorWidth<128>] in { def vpmadd52huq128 : X86Builtin<"_Vector<2, long long int>(_Vector<2, long long int>, _Vector<2, long long int>, _Vector<2, long long int>)">; -} - -let Features = "avx512ifma,avx512vl|avxifma", Attributes = [NoThrow, Const, RequiredVectorWidth<256>] in { - def vpmadd52huq256 : X86Builtin<"_Vector<4, long long int>(_Vector<4, long long int>, _Vector<4, long long int>, _Vector<4, long long int>)">; -} - -let Features = "avx512ifma,avx512vl|avxifma", Attributes = [NoThrow, Const, RequiredVectorWidth<128>] in { def vpmadd52luq128 : X86Builtin<"_Vector<2, long long int>(_Vector<2, long long int>, _Vector<2, long long int>, _Vector<2, long long int>)">; } -let Features = "avx512ifma,avx512vl|avxifma", Attributes = [NoThrow, Const, RequiredVectorWidth<256>] in { +let Features = "avx512ifma,avx512vl|avxifma", Attributes = [NoThrow, Const, Constexpr, RequiredVectorWidth<256>] in { + def vpmadd52huq256 : X86Builtin<"_Vector<4, long long int>(_Vector<4, long long int>, _Vector<4, long long int>, _Vector<4, long long int>)">; def vpmadd52luq256 : X86Builtin<"_Vector<4, long long int>(_Vector<4, long long int>, _Vector<4, long long int>, _Vector<4, long long int>)">; } diff --git a/clang/lib/AST/ByteCode/InterpBuiltin.cpp b/clang/lib/AST/ByteCode/InterpBuiltin.cpp index a2e97fcafdfef..30801b6f9060f 100644 --- a/clang/lib/AST/ByteCode/InterpBuiltin.cpp +++ b/clang/lib/AST/ByteCode/InterpBuiltin.cpp @@ -3523,6 +3523,21 @@ bool InterpretBuiltin(InterpState &S, CodePtr OpPC, const CallExpr *Call, return F; }); + case X86::BI__builtin_ia32_vpmadd52luq128: + case X86::BI__builtin_ia32_vpmadd52luq256: + case X86::BI__builtin_ia32_vpmadd52luq512: + return interp__builtin_elementwise_triop( + S, OpPC, Call, [](const APSInt &A, const APSInt &B, const APSInt &C) { + return A + (B.trunc(52) * C.trunc(52)).trunc(52).zext(64); + }); + case X86::BI__builtin_ia32_vpmadd52huq128: + case X86::BI__builtin_ia32_vpmadd52huq256: + case X86::BI__builtin_ia32_vpmadd52huq512: + return interp__builtin_elementwise_triop( + S, OpPC, Call, [](const APSInt &A, const APSInt &B, const APSInt &C) { + return A + llvm::APIntOps::mulhu(B.trunc(52), C.trunc(52)).zext(64); + }); + case X86::BI__builtin_ia32_vpshldd128: case X86::BI__builtin_ia32_vpshldd256: case X86::BI__builtin_ia32_vpshldd512: diff --git a/clang/lib/AST/ExprConstant.cpp b/clang/lib/AST/ExprConstant.cpp index b706b14945b6d..93926bcf02179 100644 --- a/clang/lib/AST/ExprConstant.cpp +++ b/clang/lib/AST/ExprConstant.cpp @@ -11869,6 +11869,52 @@ bool VectorExprEvaluator::VisitCallExpr(const CallExpr *E) { return Success(APValue(ResultElements.data(), ResultElements.size()), E); } + case X86::BI__builtin_ia32_vpmadd52luq128: + case X86::BI__builtin_ia32_vpmadd52luq256: + case X86::BI__builtin_ia32_vpmadd52luq512: { + APValue A, B, C; + if (!EvaluateAsRValue(Info, E->getArg(0), A) || + !EvaluateAsRValue(Info, E->getArg(1), B) || + !EvaluateAsRValue(Info, E->getArg(2), C)) + return false; + + unsigned ALen = A.getVectorLength(); + SmallVector<APValue, 4> ResultElements; + ResultElements.reserve(ALen); + + for (unsigned EltNum = 0; EltNum < ALen; EltNum += 1) { + APInt AElt = A.getVectorElt(EltNum).getInt(); + APInt BElt = B.getVectorElt(EltNum).getInt().trunc(52); + APInt CElt = C.getVectorElt(EltNum).getInt().trunc(52); + APSInt ResElt(AElt + (BElt * CElt).trunc(52).zext(64), false); + ResultElements.push_back(APValue(ResElt)); + } + + return Success(APValue(ResultElements.data(), ResultElements.size()), E); + } + case X86::BI__builtin_ia32_vpmadd52huq128: + case X86::BI__builtin_ia32_vpmadd52huq256: + case X86::BI__builtin_ia32_vpmadd52huq512: { + APValue A, B, C; + if (!EvaluateAsRValue(Info, E->getArg(0), A) || + !EvaluateAsRValue(Info, E->getArg(1), B) || + !EvaluateAsRValue(Info, E->getArg(2), C)) + return false; + + unsigned ALen = A.getVectorLength(); + SmallVector<APValue, 4> ResultElements; + ResultElements.reserve(ALen); + + for (unsigned EltNum = 0; EltNum < ALen; EltNum += 1) { + APInt AElt = A.getVectorElt(EltNum).getInt(); + APInt BElt = B.getVectorElt(EltNum).getInt().trunc(52); + APInt CElt = C.getVectorElt(EltNum).getInt().trunc(52); + APSInt ResElt(AElt + llvm::APIntOps::mulhu(BElt, CElt).zext(64), false); + ResultElements.push_back(APValue(ResElt)); + } + + return Success(APValue(ResultElements.data(), ResultElements.size()), E); + } case clang::X86::BI__builtin_ia32_vprotbi: case clang::X86::BI__builtin_ia32_vprotdi: case clang::X86::BI__builtin_ia32_vprotqi: diff --git a/clang/lib/Headers/avx512ifmaintrin.h b/clang/lib/Headers/avx512ifmaintrin.h index f01b322ce7787..625a8ff66dc60 100644 --- a/clang/lib/Headers/avx512ifmaintrin.h +++ b/clang/lib/Headers/avx512ifmaintrin.h @@ -15,54 +15,53 @@ #define __IFMAINTRIN_H /* Define the default attributes for the functions in this file. */ +#if defined(__cplusplus) && (__cplusplus >= 201103L) +#define __DEFAULT_FN_ATTRS \ + constexpr \ + __attribute__((__always_inline__, __nodebug__, __target__("avx512ifma"), \ + __min_vector_width__(512))) +#else #define __DEFAULT_FN_ATTRS \ __attribute__((__always_inline__, __nodebug__, __target__("avx512ifma"), \ __min_vector_width__(512))) +#endif static __inline__ __m512i __DEFAULT_FN_ATTRS -_mm512_madd52hi_epu64 (__m512i __X, __m512i __Y, __m512i __Z) -{ - return (__m512i)__builtin_ia32_vpmadd52huq512((__v8di) __X, (__v8di) __Y, - (__v8di) __Z); +_mm512_madd52hi_epu64(__m512i __X, __m512i __Y, __m512i __Z) { + return (__m512i)__builtin_ia32_vpmadd52huq512((__v8di)__X, (__v8di)__Y, + (__v8di)__Z); } -static __inline__ __m512i __DEFAULT_FN_ATTRS -_mm512_mask_madd52hi_epu64 (__m512i __W, __mmask8 __M, __m512i __X, __m512i __Y) -{ - return (__m512i)__builtin_ia32_selectq_512(__M, - (__v8di)_mm512_madd52hi_epu64(__W, __X, __Y), - (__v8di)__W); +static __inline__ __m512i __DEFAULT_FN_ATTRS _mm512_mask_madd52hi_epu64( + __m512i __W, __mmask8 __M, __m512i __X, __m512i __Y) { + return (__m512i)__builtin_ia32_selectq_512( + __M, (__v8di)_mm512_madd52hi_epu64(__W, __X, __Y), (__v8di)__W); } -static __inline__ __m512i __DEFAULT_FN_ATTRS -_mm512_maskz_madd52hi_epu64 (__mmask8 __M, __m512i __X, __m512i __Y, __m512i __Z) -{ - return (__m512i)__builtin_ia32_selectq_512(__M, - (__v8di)_mm512_madd52hi_epu64(__X, __Y, __Z), - (__v8di)_mm512_setzero_si512()); +static __inline__ __m512i __DEFAULT_FN_ATTRS _mm512_maskz_madd52hi_epu64( + __mmask8 __M, __m512i __X, __m512i __Y, __m512i __Z) { + return (__m512i)__builtin_ia32_selectq_512( + __M, (__v8di)_mm512_madd52hi_epu64(__X, __Y, __Z), + (__v8di)_mm512_setzero_si512()); } static __inline__ __m512i __DEFAULT_FN_ATTRS -_mm512_madd52lo_epu64 (__m512i __X, __m512i __Y, __m512i __Z) -{ - return (__m512i)__builtin_ia32_vpmadd52luq512((__v8di) __X, (__v8di) __Y, - (__v8di) __Z); +_mm512_madd52lo_epu64(__m512i __X, __m512i __Y, __m512i __Z) { + return (__m512i)__builtin_ia32_vpmadd52luq512((__v8di)__X, (__v8di)__Y, + (__v8di)__Z); } -static __inline__ __m512i __DEFAULT_FN_ATTRS -_mm512_mask_madd52lo_epu64 (__m512i __W, __mmask8 __M, __m512i __X, __m512i __Y) -{ - return (__m512i)__builtin_ia32_selectq_512(__M, - (__v8di)_mm512_madd52lo_epu64(__W, __X, __Y), - (__v8di)__W); +static __inline__ __m512i __DEFAULT_FN_ATTRS _mm512_mask_madd52lo_epu64( + __m512i __W, __mmask8 __M, __m512i __X, __m512i __Y) { + return (__m512i)__builtin_ia32_selectq_512( + __M, (__v8di)_mm512_madd52lo_epu64(__W, __X, __Y), (__v8di)__W); } -static __inline__ __m512i __DEFAULT_FN_ATTRS -_mm512_maskz_madd52lo_epu64 (__mmask8 __M, __m512i __X, __m512i __Y, __m512i __Z) -{ - return (__m512i)__builtin_ia32_selectq_512(__M, - (__v8di)_mm512_madd52lo_epu64(__X, __Y, __Z), - (__v8di)_mm512_setzero_si512()); +static __inline__ __m512i __DEFAULT_FN_ATTRS _mm512_maskz_madd52lo_epu64( + __mmask8 __M, __m512i __X, __m512i __Y, __m512i __Z) { + return (__m512i)__builtin_ia32_selectq_512( + __M, (__v8di)_mm512_madd52lo_epu64(__X, __Y, __Z), + (__v8di)_mm512_setzero_si512()); } #undef __DEFAULT_FN_ATTRS diff --git a/clang/lib/Headers/avx512ifmavlintrin.h b/clang/lib/Headers/avx512ifmavlintrin.h index a72b56113a12b..9cfac4dd01455 100644 --- a/clang/lib/Headers/avx512ifmavlintrin.h +++ b/clang/lib/Headers/avx512ifmavlintrin.h @@ -15,6 +15,16 @@ #define __IFMAVLINTRIN_H /* Define the default attributes for the functions in this file. */ +#if defined(__cplusplus) && (__cplusplus >= 201103L) +#define __DEFAULT_FN_ATTRS128 \ + constexpr __attribute__((__always_inline__, __nodebug__, \ + __target__("avx512ifma,avx512vl"), \ + __min_vector_width__(128))) +#define __DEFAULT_FN_ATTRS256 \ + constexpr __attribute__((__always_inline__, __nodebug__, \ + __target__("avx512ifma,avx512vl"), \ + __min_vector_width__(256))) +#else #define __DEFAULT_FN_ATTRS128 \ __attribute__((__always_inline__, __nodebug__, \ __target__("avx512ifma,avx512vl"), \ @@ -23,22 +33,31 @@ __attribute__((__always_inline__, __nodebug__, \ __target__("avx512ifma,avx512vl"), \ __min_vector_width__(256))) +#endif -#define _mm_madd52hi_epu64(X, Y, Z) \ - ((__m128i)__builtin_ia32_vpmadd52huq128((__v2di)(X), (__v2di)(Y), \ - (__v2di)(Z))) +static __inline__ __m128i __DEFAULT_FN_ATTRS128 +_mm_madd52hi_epu64(__m128i __X, __m128i __Y, __m128i __Z) { + return (__m128i)__builtin_ia32_vpmadd52huq128((__v2di)__X, (__v2di)__Y, + (__v2di)__Z); +} -#define _mm256_madd52hi_epu64(X, Y, Z) \ - ((__m256i)__builtin_ia32_vpmadd52huq256((__v4di)(X), (__v4di)(Y), \ - (__v4di)(Z))) +static __inline__ __m256i __DEFAULT_FN_ATTRS256 +_mm256_madd52hi_epu64(__m256i __X, __m256i __Y, __m256i __Z) { + return (__m256i)__builtin_ia32_vpmadd52huq256((__v4di)__X, (__v4di)__Y, + (__v4di)__Z); +} -#define _mm_madd52lo_epu64(X, Y, Z) \ - ((__m128i)__builtin_ia32_vpmadd52luq128((__v2di)(X), (__v2di)(Y), \ - (__v2di)(Z))) +static __inline__ __m128i __DEFAULT_FN_ATTRS128 +_mm_madd52lo_epu64(__m128i __X, __m128i __Y, __m128i __Z) { + return (__m128i)__builtin_ia32_vpmadd52luq128((__v2di)__X, (__v2di)__Y, + (__v2di)__Z); +} -#define _mm256_madd52lo_epu64(X, Y, Z) \ - ((__m256i)__builtin_ia32_vpmadd52luq256((__v4di)(X), (__v4di)(Y), \ - (__v4di)(Z))) +static __inline__ __m256i __DEFAULT_FN_ATTRS256 +_mm256_madd52lo_epu64(__m256i __X, __m256i __Y, __m256i __Z) { + return (__m256i)__builtin_ia32_vpmadd52luq256((__v4di)__X, (__v4di)__Y, + (__v4di)__Z); +} static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_mask_madd52hi_epu64 (__m128i __W, __mmask8 __M, __m128i __X, __m128i __Y) diff --git a/clang/lib/Headers/avxifmaintrin.h b/clang/lib/Headers/avxifmaintrin.h index 5c782d2a5b865..6916be7b49b15 100644 --- a/clang/lib/Headers/avxifmaintrin.h +++ b/clang/lib/Headers/avxifmaintrin.h @@ -22,6 +22,14 @@ __attribute__((__always_inline__, __nodebug__, __target__("avxifma"), \ __min_vector_width__(256))) +#if defined(__cplusplus) && (__cplusplus >= 201103L) +#define __DEFAULT_FN_ATTRS256_CONSTEXPR constexpr __DEFAULT_FN_ATTRS256 +#define __DEFAULT_FN_ATTRS128_CONSTEXPR constexpr __DEFAULT_FN_ATTRS128 +#else +#define __DEFAULT_FN_ATTRS256_CONSTEXPR __DEFAULT_FN_ATTRS256 +#define __DEFAULT_FN_ATTRS128_CONSTEXPR __DEFAULT_FN_ATTRS128 +#endif + // must vex-encoding /// Multiply packed unsigned 52-bit integers in each 64-bit element of \a __Y @@ -55,7 +63,7 @@ /// ENDFOR /// dst[MAX:128] := 0 /// \endcode -static __inline__ __m128i __DEFAULT_FN_ATTRS128 +static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR _mm_madd52hi_avx_epu64(__m128i __X, __m128i __Y, __m128i __Z) { return (__m128i)__builtin_ia32_vpmadd52huq128((__v2di)__X, (__v2di)__Y, (__v2di)__Z); @@ -92,7 +100,7 @@ _mm_madd52hi_avx_epu64(__m128i __X, __m128i __Y, __m128i __Z) { /// ENDFOR /// dst[MAX:256] := 0 /// \endcode -static __inline__ __m256i __DEFAULT_FN_ATTRS256 +static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR _mm256_madd52hi_avx_epu64(__m256i __X, __m256i __Y, __m256i __Z) { return (__m256i)__builtin_ia32_vpmadd52huq256((__v4di)__X, (__v4di)__Y, (__v4di)__Z); @@ -129,7 +137,7 @@ _mm256_madd52hi_avx_epu64(__m256i __X, __m256i __Y, __m256i __Z) { /// ENDFOR /// dst[MAX:128] := 0 /// \endcode -static __inline__ __m128i __DEFAULT_FN_ATTRS128 +static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR _mm_madd52lo_avx_epu64(__m128i __X, __m128i __Y, __m128i __Z) { return (__m128i)__builtin_ia32_vpmadd52luq128((__v2di)__X, (__v2di)__Y, (__v2di)__Z); @@ -166,12 +174,14 @@ _mm_madd52lo_avx_epu64(__m128i __X, __m128i __Y, __m128i __Z) { /// ENDFOR /// dst[MAX:256] := 0 /// \endcode -static __inline__ __m256i __DEFAULT_FN_ATTRS256 +static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR _mm256_madd52lo_avx_epu64(__m256i __X, __m256i __Y, __m256i __Z) { return (__m256i)__builtin_ia32_vpmadd52luq256((__v4di)__X, (__v4di)__Y, (__v4di)__Z); } #undef __DEFAULT_FN_ATTRS128 #undef __DEFAULT_FN_ATTRS256 +#undef __DEFAULT_FN_ATTRS256_CONSTEXPR +#undef __DEFAULT_FN_ATTRS128_CONSTEXPR #endif // __AVXIFMAINTRIN_H diff --git a/clang/test/CodeGen/X86/avx512ifma-builtins.c b/clang/test/CodeGen/X86/avx512ifma-builtins.c index eebefb0bad4ab..9e8c5464db383 100644 --- a/clang/test/CodeGen/X86/avx512ifma-builtins.c +++ b/clang/test/CodeGen/X86/avx512ifma-builtins.c @@ -3,50 +3,250 @@ // RUN: %clang_cc1 -x c++ -flax-vector-conversions=none -ffreestanding %s -triple=x86_64-apple-darwin -target-feature +avx512ifma -emit-llvm -o - -Wall -Werror | FileCheck %s // RUN: %clang_cc1 -x c++ -flax-vector-conversions=none -ffreestanding %s -triple=i386-apple-darwin -target-feature +avx512ifma -emit-llvm -o - -Wall -Werror | FileCheck %s -// RUN: %clang_cc1 -x c -flax-vector-conversions=none -ffreestanding %s -triple=x86_64-apple-darwin -target-feature +avx512ifma -emit-llvm -o - -Wall -Werror -fexperimental-new-constant-interpreter | FileCheck %s -// RUN: %clang_cc1 -x c -flax-vector-conversions=none -ffreestanding %s -triple=i386-apple-darwin -target-feature +avx512ifma -emit-llvm -o - -Wall -Werror -fexperimental-new-constant-interpreter | FileCheck %s -// RUN: %clang_cc1 -x c++ -flax-vector-conversions=none -ffreestanding %s -triple=x86_64-apple-darwin -target-feature +avx512ifma -emit-llvm -o - -Wall -Werror -fexperimental-new-constant-interpreter | FileCheck %s -// RUN: %clang_cc1 -x c++ -flax-vector-conversions=none -ffreestanding %s -triple=i386-apple-darwin -target-feature +avx512ifma -emit-llvm -o - -Wall -Werror -fexperimental-new-constant-interpreter | FileCheck %s - - #include <immintrin.h> +#include "builtin_test_helpers.h" __m512i test_mm512_madd52hi_epu64(__m512i __X, __m512i __Y, __m512i __Z) { // CHECK-LABEL: test_mm512_madd52hi_epu64 // CHECK: call {{.*}}<8 x i64> @llvm.x86.avx512.vpmadd52h.uq.512(<8 x i64> %{{.*}}, <8 x i64> %{{.*}}, <8 x i64> %{{.*}}) - return _mm512_madd52hi_epu64(__X, __Y, __Z); + return _mm512_madd52hi_epu64(__X, __Y, __Z); } +TEST_CONSTEXPR(match_v8di(_mm512_madd52hi_epu64( + (__m512i)(__v8du){100, 0, 0, 0, 0, 0, 0, 0}, + (__m512i)(__v8du){10, 0, 0, 0, 0, 0, 0, 0}, + (__m512i)(__v8du){5, 0, 0, 0, 0, 0, 0, 0}), + 100, 0, 0, 0, 0, 0, 0, 0), + "mm512_madd52hi_epu64: high bits remain in accumulator"); + +TEST_CONSTEXPR(match_v8di(_mm512_madd52hi_epu64( + (__m512i)(__v8du){0, 0, 0, 0, 0, 0, 0, 0}, + (__m512i)(__v8du){0xFFFFFFFFFFFFFull, 0, 0, 0, + 0, 0, 0, 0}, + (__m512i)(__v8du){0xFFFFFFFFFFFFFull, 0, 0, 0, + 0, 0, 0, 0}), + 0xFFFFFFFFFFFFEull, 0, 0, 0, 0, 0, 0, 0), + "mm512_madd52hi_epu64: max value multiply high bits"); + +TEST_CONSTEXPR(match_v8di(_mm512_madd52hi_epu64( + (__m512i)(__v8du){1, 2, 3, 4, 5, 6, 7, 8}, + (__m512i)(__v8du){0xFFFFFFFFFFFFFull, + 0xFFFFFFFFFFFFFull, + 0xFFFFFFFFFFFFFull, + 0xFFFFFFFFFFFFFull, + 0xFFFFFFFFFFFFFull, + 0xFFFFFFFFFFFFFull, + 0xFFFFFFFFFFFFFull, + 0xFFFFFFFFFFFFFull}, + (__m512i)(__v8du){0xFFFFFFFFFFFFFull, + 0xFFFFFFFFFFFFFull, + 0xFFFFFFFFFFFFFull, + 0xFFFFFFFFFFFFFull, + 0xFFFFFFFFFFFFFull, + 0xFFFFFFFFFFFFFull, + 0xFFFFFFFFFFFFFull, + 0xFFFFFFFFFFFFFull}), + 4503599627370495ull, 4503599627370496ull, + 4503599627370497ull, 4503599627370498ull, + 4503599627370499ull, 4503599627370500ull, + 4503599627370501ull, 4503599627370502ull), + "mm512_madd52hi_epu64: eight-lane high bits with carry"); + __m512i test_mm512_mask_madd52hi_epu64(__m512i __W, __mmask8 __M, __m512i __X, __m512i __Y) { // CHECK-LABEL: test_mm512_mask_madd52hi_epu64 // CHECK: call {{.*}}<8 x i64> @llvm.x86.avx512.vpmadd52h.uq.512(<8 x i64> %{{.*}}, <8 x i64> %{{.*}}, <8 x i64> %{{.*}}) // CHECK: select <8 x i1> %{{.*}}, <8 x i64> %{{.*}}, <8 x i64> %{{.*}} - return _mm512_mask_madd52hi_epu64(__W, __M, __X, __Y); + return _mm512_mask_madd52hi_epu64(__W, __M, __X, __Y); } +TEST_CONSTEXPR(match_v8di(_mm512_mask_madd52hi_epu64( + (__m512i)(__v8du){111, 222, 333, 444, 555, 666, + 777, 888}, + 0x00, + (__m512i)(__v8du){1, 2, 3, 4, 5, 6, 7, 8}, + (__m512i)(__v8du){10, 20, 30, 40, 50, 60, 70, + 80}), + 111, 222, 333, 444, 555, 666, 777, 888), + "mm512_mask_madd52hi_epu64: zero mask preserves __W"); + +TEST_CONSTEXPR(match_v8di(_mm512_mask_madd52hi_epu64( + (__m512i)(__v8du){10, 20, 30, 40, 50, 60, 70, + 80}, + 0xFF, + (__m512i)(__v8du){100, 200, 300, 400, 500, 600, + 700, 800}, + (__m512i)(__v8du){10, 20, 30, 40, 50, 60, 70, + 80}), + 10, 20, 30, 40, 50, 60, 70, 80), + "mm512_mask_madd52hi_epu64: all mask bits set"); + __m512i test_mm512_maskz_madd52hi_epu64(__mmask8 __M, __m512i __X, __m512i __Y, __m512i __Z) { // CHECK-LABEL: test_mm512_maskz_madd52hi_epu64 // CHECK: call {{.*}}<8 x i64> @llvm.x86.avx512.vpmadd52h.uq.512(<8 x i64> %{{.*}}, <8 x i64> %{{.*}}, <8 x i64> %{{.*}}) // CHECK: select <8 x i1> %{{.*}}, <8 x i64> %{{.*}}, <8 x i64> %{{.*}} - return _mm512_maskz_madd52hi_epu64(__M, __X, __Y, __Z); + return _mm512_maskz_madd52hi_epu64(__M, __X, __Y, __Z); } +TEST_CONSTEXPR(match_v8di(_mm512_maskz_madd52hi_epu64( + 0x00, + (__m512i)(__v8du){1, 2, 3, 4, 5, 6, 7, 8}, + (__m512i)(__v8du){10, 20, 30, 40, 50, 60, 70, + 80}, + (__m512i)(__v8du){100, 200, 300, 400, 500, 600, + 700, 800}), + 0, 0, 0, 0, 0, 0, 0, 0), + "mm512_maskz_madd52hi_epu64: zero mask"); + +TEST_CONSTEXPR(match_v8di(_mm512_maskz_madd52hi_epu64( + 0xFF, + (__m512i)(__v8du){1, 2, 3, 4, 5, 6, 7, 8}, + (__m512i)(__v8du){10, 20, 30, 40, 50, 60, 70, + 80}, + (__m512i)(__v8du){100, 200, 300, 400, 500, 600, + 700, 800}), + 1, 2, 3, 4, 5, 6, 7, 8), + "mm512_maskz_madd52hi_epu64: all mask bits set"); + __m512i test_mm512_madd52lo_epu64(__m512i __X, __m512i __Y, __m512i __Z) { // CHECK-LABEL: test_mm512_madd52lo_epu64 // CHECK: call {{.*}}<8 x i64> @llvm.x86.avx512.vpmadd52l.uq.512(<8 x i64> %{{.*}}, <8 x i64> %{{.*}}, <8 x i64> %{{.*}}) - return _mm512_madd52lo_epu64(__X, __Y, __Z); + return _mm512_madd52lo_epu64(__X, __Y, __Z); } +TEST_CONSTEXPR(match_v8di(_mm512_madd52lo_epu64( + (__m512i)(__v8du){0, 0, 0, 0, 0, 0, 0, 0}, + (__m512i)(__v8du){10, 0, 0, 0, 0, 0, 0, 0}, + (__m512i)(__v8du){5, 0, 0, 0, 0, 0, 0, 0}), + 50, 0, 0, 0, 0, 0, 0, 0), + "mm512_madd52lo_epu64: basic multiply-add low bits"); + +TEST_CONSTEXPR(match_v8di(_mm512_madd52lo_epu64( + (__m512i)(__v8du){100, 0, 0, 0, 0, 0, 0, 0}, + (__m512i)(__v8du){20, 0, 0, 0, 0, 0, 0, 0}, + (__m512i)(__v8du){30, 0, 0, 0, 0, 0, 0, 0}), + 700, 0, 0, 0, 0, 0, 0, 0), + "mm512_madd52lo_epu64: accumulator test"); + +TEST_CONSTEXPR(match_v8di(_mm512_madd52lo_epu64( + (__m512i)(__v8du){0, 0, 0, 0, 0, 0, 0, 0}, + (__m512i)(__v8du){0xFFFFFFFFFFFFFull, 0, 0, 0, + 0, 0, 0, 0}, + (__m512i)(__v8du){1, 0, 0, 0, 0, 0, 0, 0}), + 0xFFFFFFFFFFFFFull, 0, 0, 0, 0, 0, 0, 0), + "mm512_madd52lo_epu64: max 52-bit value"); + +TEST_CONSTEXPR(match_v8di(_mm512_madd52lo_epu64( + (__m512i)(__v8du){0, 0, 0, 0, 0, 0, 0, 0}, + (__m512i)(__v8du){0x1F000000000000ull, 0, 0, 0, + 0, 0, 0, 0}, + (__m512i)(__v8du){2, 0, 0, 0, 0, 0, 0, 0}), + 0xE000000000000ull, 0, 0, 0, 0, 0, 0, 0), + "mm512_madd52lo_epu64: 52-bit truncation test"); + +TEST_CONSTEXPR(match_v8di(_mm512_madd52lo_epu64( + (__m512i)(__v8du){1, 2, 3, 4, 5, 6, 7, 8}, + (__m512i)(__v8du){10, 20, 30, 40, 50, 60, 70, + 80}, + (__m512i)(__v8du){2, 3, 4, 5, 6, 7, 8, 9}), + 21, 62, 123, 204, 305, 426, 567, 728), + "mm512_madd52lo_epu64: eight-lane computation"); + +TEST_CONSTEXPR(match_v8di(_mm512_madd52lo_epu64( + (__m512i)(__v8du){0xFFFFFFFFFFFFFull, 0, 0, 0, + 0, 0, 0, 0}, + (__m512i)(__v8du){10, 0, 0, 0, 0, 0, 0, 0}, + (__m512i)(__v8du){5, 0, 0, 0, 0, 0, 0, 0}), + 4503599627370545ull, 0, 0, 0, 0, 0, 0, 0), + "mm512_madd52lo_epu64: accumulator with large value"); + +TEST_CONSTEXPR(match_v8di(_mm512_madd52lo_epu64( + (__m512i)(__v8du){10, 20, 30, 40, 50, 60, 70, + 80}, + (__m512i)(__v8du){100, 200, 300, 400, 500, 600, + 700, 800}, + (__m512i)(__v8du){2, 3, 4, 5, 6, 7, 8, 9}), + 210, 620, 1230, 2040, 3050, 4260, 5670, 7280), + "mm512_madd52lo_epu64: eight-lane with larger values"); + +TEST_CONSTEXPR(match_v8di(_mm512_madd52lo_epu64( + (__m512i)(__v8du){0, 0, 0, 0, 0, 0, 0, 0}, + (__m512i)(__v8du){0x1F000000000000ull, + 0x1F000000000000ull, 0, 0, 0, + 0, 0, 0}, + (__m512i)(__v8du){2, 3, 0, 0, 0, 0, 0, 0}), + 0xE000000000000ull, 0xD000000000000ull, 0, 0, 0, 0, + 0, 0), + "mm512_madd52lo_epu64: two-lane truncation test"); + +TEST_CONSTEXPR(match_v8di(_mm512_madd52lo_epu64( + (__m512i)(__v8du){0, 0, 0, 0, 0, 0, 0, 0}, + (__m512i)(__v8du){0xFFFFFFFFFFFFFull, + 0xFFFFFFFFFFFFFull, + 0xFFFFFFFFFFFFFull, + 0xFFFFFFFFFFFFFull, + 0xFFFFFFFFFFFFFull, + 0xFFFFFFFFFFFFFull, + 0xFFFFFFFFFFFFFull, + 0xFFFFFFFFFFFFFull}, + (__m512i)(__v8du){1, 1, 1, 1, 1, 1, 1, 1}), + 0xFFFFFFFFFFFFFull, 0xFFFFFFFFFFFFFull, + 0xFFFFFFFFFFFFFull, 0xFFFFFFFFFFFFFull, + 0xFFFFFFFFFFFFFull, 0xFFFFFFFFFFFFFull, + 0xFFFFFFFFFFFFFull, 0xFFFFFFFFFFFFFull), + "mm512_madd52lo_epu64: all lanes max value"); + __m512i test_mm512_mask_madd52lo_epu64(__m512i __W, __mmask8 __M, __m512i __X, __m512i __Y) { // CHECK-LABEL: test_mm512_mask_madd52lo_epu64 // CHECK: call {{.*}}<8 x i64> @llvm.x86.avx512.vpmadd52l.uq.512(<8 x i64> %{{.*}}, <8 x i64> %{{.*}}, <8 x i64> %{{.*}}) // CHECK: select <8 x i1> %{{.*}}, <8 x i64> %{{.*}}, <8 x i64> %{{.*}} - return _mm512_mask_madd52lo_epu64(__W, __M, __X, __Y); + return _mm512_mask_madd52lo_epu64(__W, __M, __X, __Y); } +TEST_CONSTEXPR(match_v8di(_mm512_mask_madd52lo_epu64( + (__m512i)(__v8du){111, 222, 333, 444, 555, 666, + 777, 888}, + 0x00, + (__m512i)(__v8du){1, 2, 3, 4, 5, 6, 7, 8}, + (__m512i)(__v8du){10, 20, 30, 40, 50, 60, 70, + 80}), + 111, 222, 333, 444, 555, 666, 777, 888), + "mm512_mask_madd52lo_epu64: zero mask preserves __W"); + +TEST_CONSTEXPR(match_v8di(_mm512_mask_madd52lo_epu64( + (__m512i)(__v8du){1000, 2000, 3000, 4000, 5000, + 6000, 7000, 8000}, + 0xFF, + (__m512i)(__v8du){100, 200, 300, 400, 500, 600, + 700, 800}, + (__m512i)(__v8du){20, 30, 40, 50, 60, 70, 80, + 90}), + 3000, 8000, 15000, 24000, 35000, 48000, 63000, + 80000), + "mm512_mask_madd52lo_epu64: all mask bits set"); + __m512i test_mm512_maskz_madd52lo_epu64(__mmask8 __M, __m512i __X, __m512i __Y, __m512i __Z) { // CHECK-LABEL: test_mm512_maskz_madd52lo_epu64 // CHECK: call {{.*}}<8 x i64> @llvm.x86.avx512.vpmadd52l.uq.512(<8 x i64> %{{.*}}, <8 x i64> %{{.*}}, <8 x i64> %{{.*}}) // CHECK: select <8 x i1> %{{.*}}, <8 x i64> %{{.*}}, <8 x i64> %{{.*}} - return _mm512_maskz_madd52lo_epu64(__M, __X, __Y, __Z); + return _mm512_maskz_madd52lo_epu64(__M, __X, __Y, __Z); } + +TEST_CONSTEXPR(match_v8di(_mm512_maskz_madd52lo_epu64( + 0x00, + (__m512i)(__v8du){1, 2, 3, 4, 5, 6, 7, 8}, + (__m512i)(__v8du){10, 20, 30, 40, 50, 60, 70, + 80}, + (__m512i)(__v8du){2, 3, 4, 5, 6, 7, 8, 9}), + 0, 0, 0, 0, 0, 0, 0, 0), + "mm512_maskz_madd52lo_epu64: zero mask"); + +TEST_CONSTEXPR(match_v8di(_mm512_maskz_madd52lo_epu64( + 0xFF, + (__m512i)(__v8du){100, 200, 300, 400, 500, 600, + 700, 800}, + (__m512i)(__v8du){20, 30, 40, 50, 60, 70, 80, + 90}, + (__m512i)(__v8du){30, 40, 50, 60, 70, 80, 90, + 100}), + 700, 1400, 2300, 3400, 4700, 6200, 7900, 9800), + "mm512_maskz_madd52lo_epu64: all mask bits set"); diff --git a/clang/test/CodeGen/X86/avx512ifmavl-builtins.c b/clang/test/CodeGen/X86/avx512ifmavl-builtins.c index 89108fc037520..b41c821a9049b 100644 --- a/clang/test/CodeGen/X86/avx512ifmavl-builtins.c +++ b/clang/test/CodeGen/X86/avx512ifmavl-builtins.c @@ -3,90 +3,266 @@ // RUN: %clang_cc1 -x c++ %s -flax-vector-conversions=none -ffreestanding -triple=x86_64-apple-darwin -target-feature +avx512ifma -target-feature +avx512vl -emit-llvm -o - -Wall -Werror | FileCheck %s // RUN: %clang_cc1 -x c++ %s -flax-vector-conversions=none -ffreestanding -triple=i386-apple-darwin -target-feature +avx512ifma -target-feature +avx512vl -emit-llvm -o - -Wall -Werror | FileCheck %s -// RUN: %clang_cc1 -x c %s -flax-vector-conversions=none -ffreestanding -triple=x86_64-apple-darwin -target-feature +avx512ifma -target-feature +avx512vl -emit-llvm -o - -Wall -Werror -fexperimental-new-constant-interpreter | FileCheck %s -// RUN: %clang_cc1 -x c %s -flax-vector-conversions=none -ffreestanding -triple=i386-apple-darwin -target-feature +avx512ifma -target-feature +avx512vl -emit-llvm -o - -Wall -Werror -fexperimental-new-constant-interpreter | FileCheck %s -// RUN: %clang_cc1 -x c++ %s -flax-vector-conversions=none -ffreestanding -triple=x86_64-apple-darwin -target-feature +avx512ifma -target-feature +avx512vl -emit-llvm -o - -Wall -Werror -fexperimental-new-constant-interpreter | FileCheck %s -// RUN: %clang_cc1 -x c++ %s -flax-vector-conversions=none -ffreestanding -triple=i386-apple-darwin -target-feature +avx512ifma -target-feature +avx512vl -emit-llvm -o - -Wall -Werror -fexperimental-new-constant-interpreter | FileCheck %s - - #include <immintrin.h> +#include "builtin_test_helpers.h" __m128i test_mm_madd52hi_epu64(__m128i __X, __m128i __Y, __m128i __Z) { // CHECK-LABEL: test_mm_madd52hi_epu64 // CHECK: call {{.*}}<2 x i64> @llvm.x86.avx512.vpmadd52h.uq.128(<2 x i64> %{{.*}}, <2 x i64> %{{.*}}, <2 x i64> %{{.*}}) - return _mm_madd52hi_epu64(__X, __Y, __Z); + return _mm_madd52hi_epu64(__X, __Y, __Z); } +TEST_CONSTEXPR(match_v2di(_mm_madd52hi_epu64( + (__m128i)((__v2du){100, 0}), + (__m128i)((__v2du){10, 0}), + (__m128i)((__v2du){5, 0})), + 100, 0), + "mm_madd52hi_epu64: high bits remain in accumulator"); + +TEST_CONSTEXPR(match_v2di(_mm_madd52hi_epu64( + (__m128i)((__v2du){0, 0}), + (__m128i)((__v2du){0xFFFFFFFFFFFFFull, 0}), + (__m128i)((__v2du){0xFFFFFFFFFFFFFull, 0})), + 0xFFFFFFFFFFFFEull, 0), + "mm_madd52hi_epu64: max value multiply high bits"); + __m128i test_mm_mask_madd52hi_epu64(__m128i __W, __mmask8 __M, __m128i __X, __m128i __Y) { // CHECK-LABEL: test_mm_mask_madd52hi_epu64 // CHECK: call {{.*}}<2 x i64> @llvm.x86.avx512.vpmadd52h.uq.128(<2 x i64> %{{.*}}, <2 x i64> %{{.*}}, <2 x i64> %{{.*}}) // CHECK: select <2 x i1> %{{.*}}, <2 x i64> %{{.*}}, <2 x i64> %{{.*}} - return _mm_mask_madd52hi_epu64(__W, __M, __X, __Y); + return _mm_mask_madd52hi_epu64(__W, __M, __X, __Y); } +TEST_CONSTEXPR(match_v2di(_mm_mask_madd52hi_epu64((__m128i)((__v2du){111, 222}), + 0x0, + (__m128i)((__v2du){1, 2}), + (__m128i)((__v2du){10, 20})), + 111, 222), + "mm_mask_madd52hi_epu64: zero mask preserves __W"); + +TEST_CONSTEXPR(match_v2di(_mm_mask_madd52hi_epu64((__m128i)((__v2du){10, 20}), + 0x2, + (__m128i)((__v2du){0x1000000000000ULL, 0x1000000000000ULL}), + (__m128i)((__v2du){0x1000000000000ULL, 0x1000000000000ULL})), + 10, 0x100000000014ULL), + "mm_mask_madd52hi_epu64: mask 0x2"); + __m128i test_mm_maskz_madd52hi_epu64(__mmask8 __M, __m128i __X, __m128i __Y, __m128i __Z) { // CHECK-LABEL: test_mm_maskz_madd52hi_epu64 // CHECK: call {{.*}}<2 x i64> @llvm.x86.avx512.vpmadd52h.uq.128(<2 x i64> %{{.*}}, <2 x i64> %{{.*}}, <2 x i64> %{{.*}}) // CHECK: select <2 x i1> %{{.*}}, <2 x i64> %{{.*}}, <2 x i64> %{{.*}} - return _mm_maskz_madd52hi_epu64(__M, __X, __Y, __Z); + return _mm_maskz_madd52hi_epu64(__M, __X, __Y, __Z); } +TEST_CONSTEXPR(match_v2di(_mm_maskz_madd52hi_epu64(0x3, + (__m128i)((__v2du){1, 2}), + (__m128i)((__v2du){10, 20}), + (__m128i)((__v2du){100, 200})), + 1, 2), + "mm_maskz_madd52hi_epu64: all mask bits set"); + +TEST_CONSTEXPR(match_v2di(_mm_maskz_madd52hi_epu64(0x1, + (__m128i)((__v2du){0x1000000000000ULL, 0x1000000000000ULL}), + (__m128i)((__v2du){0x1000000000000ULL, 0x1000000000000ULL}), + (__m128i)((__v2du){0, 0})), + 0x1000000000000ULL, 0), + "mm_maskz_madd52hi_epu64: mask 0x1"); + __m256i test_mm256_madd52hi_epu64(__m256i __X, __m256i __Y, __m256i __Z) { // CHECK-LABEL: test_mm256_madd52hi_epu64 // CHECK: call {{.*}}<4 x i64> @llvm.x86.avx512.vpmadd52h.uq.256(<4 x i64> %{{.*}}, <4 x i64> %{{.*}}, <4 x i64> %{{.*}}) - return _mm256_madd52hi_epu64(__X, __Y, __Z); + return _mm256_madd52hi_epu64(__X, __Y, __Z); } +TEST_CONSTEXPR(match_v4di(_mm256_madd52hi_epu64( + (__m256i)((__v4du){100, 200, 300, 400}), + (__m256i)((__v4du){10, 20, 30, 40}), + (__m256i)((__v4du){5, 6, 7, 8})), + 100, 200, 300, 400), + "mm256_madd52hi_epu64: accumulator preserved"); + +TEST_CONSTEXPR(match_v4di(_mm256_madd52hi_epu64( + (__m256i)((__v4du){0, 0, 0, 0}), + (__m256i)((__v4du){0xFFFFFFFFFFFFFull, 0, 0, + 0}), + (__m256i)((__v4du){0xFFFFFFFFFFFFFull, 0, 0, + 0})), + 0xFFFFFFFFFFFFEull, 0, 0, 0), + "mm256_madd52hi_epu64: single lane max value"); + __m256i test_mm256_mask_madd52hi_epu64(__m256i __W, __mmask8 __M, __m256i __X, __m256i __Y) { // CHECK-LABEL: test_mm256_mask_madd52hi_epu64 // CHECK: call {{.*}}<4 x i64> @llvm.x86.avx512.vpmadd52h.uq.256(<4 x i64> %{{.*}}, <4 x i64> %{{.*}}, <4 x i64> %{{.*}}) // CHECK: select <4 x i1> %{{.*}}, <4 x i64> %{{.*}}, <4 x i64> %{{.*}} - return _mm256_mask_madd52hi_epu64(__W, __M, __X, __Y); + return _mm256_mask_madd52hi_epu64(__W, __M, __X, __Y); } +TEST_CONSTEXPR(match_v4di(_mm256_mask_madd52hi_epu64((__m256i)((__v4du){111, 222, 333, 444}), + 0x0, + (__m256i)((__v4du){1, 2, 3, 4}), + (__m256i)((__v4du){10, 20, 30, 40})), + 111, 222, 333, 444), + "mm256_mask_madd52hi_epu64: zero mask preserves __W"); + +TEST_CONSTEXPR(match_v4di(_mm256_mask_madd52hi_epu64((__m256i)((__v4du){10, 20, 30, 40}), + 0xA, + (__m256i)((__v4du){0x1000000000000ULL, 0x1000000000000ULL, + 0x1000000000000ULL, 0x1000000000000ULL}), + (__m256i)((__v4du){0x1000000000000ULL, 0x1000000000000ULL, + 0x1000000000000ULL, 0x1000000000000ULL})), + 10, 0x100000000014ULL, 30, 0x100000000028ULL), + "mm256_mask_madd52hi_epu64: mask 0xA"); + __m256i test_mm256_maskz_madd52hi_epu64(__mmask8 __M, __m256i __X, __m256i __Y, __m256i __Z) { // CHECK-LABEL: test_mm256_maskz_madd52hi_epu64 // CHECK: call {{.*}}<4 x i64> @llvm.x86.avx512.vpmadd52h.uq.256(<4 x i64> %{{.*}}, <4 x i64> %{{.*}}, <4 x i64> %{{.*}}) // CHECK: select <4 x i1> %{{.*}}, <4 x i64> %{{.*}}, <4 x i64> %{{.*}} - return _mm256_maskz_madd52hi_epu64(__M, __X, __Y, __Z); + return _mm256_maskz_madd52hi_epu64(__M, __X, __Y, __Z); } +TEST_CONSTEXPR(match_v4di(_mm256_maskz_madd52hi_epu64(0xF, + (__m256i)((__v4du){1, 2, 3, 4}), + (__m256i)((__v4du){10, 20, 30, 40}), + (__m256i)((__v4du){100, 200, 300, 400})), + 1, 2, 3, 4), + "mm256_maskz_madd52hi_epu64: all mask bits set"); + +TEST_CONSTEXPR(match_v4di(_mm256_maskz_madd52hi_epu64(0x5, + (__m256i)((__v4du){0x1000000000000ULL, 0x1000000000000ULL, + 0x1000000000000ULL, 0x1000000000000ULL}), + (__m256i)((__v4du){0x1000000000000ULL, 0x1000000000000ULL, + 0x1000000000000ULL, 0x1000000000000ULL}), + (__m256i)((__v4du){0, 0, 0, 0})), + 0x1000000000000ULL, 0, 0x1000000000000ULL, 0), + "mm256_maskz_madd52hi_epu64: mask 0x5"); + __m128i test_mm_madd52lo_epu64(__m128i __X, __m128i __Y, __m128i __Z) { // CHECK-LABEL: test_mm_madd52lo_epu64 // CHECK: call {{.*}}<2 x i64> @llvm.x86.avx512.vpmadd52l.uq.128(<2 x i64> %{{.*}}, <2 x i64> %{{.*}}, <2 x i64> %{{.*}}) - return _mm_madd52lo_epu64(__X, __Y, __Z); + return _mm_madd52lo_epu64(__X, __Y, __Z); } +TEST_CONSTEXPR(match_v2di(_mm_madd52lo_epu64( + (__m128i)((__v2du){0, 0}), + (__m128i)((__v2du){10, 0}), + (__m128i)((__v2du){5, 0})), + 50, 0), + "mm_madd52lo_epu64: basic multiply-add low bits"); + +TEST_CONSTEXPR(match_v2di(_mm_madd52lo_epu64( + (__m128i)((__v2du){100, 0}), + (__m128i)((__v2du){20, 0}), + (__m128i)((__v2du){30, 0})), + 700, 0), + "mm_madd52lo_epu64: accumulator test"); + +TEST_CONSTEXPR(match_v2di(_mm_madd52lo_epu64( + (__m128i)((__v2du){1, 2}), + (__m128i)((__v2du){10, 20}), + (__m128i)((__v2du){2, 3})), + 21, 62), + "mm_madd52lo_epu64: two-lane computation"); + __m128i test_mm_mask_madd52lo_epu64(__m128i __W, __mmask8 __M, __m128i __X, __m128i __Y) { // CHECK-LABEL: test_mm_mask_madd52lo_epu64 // CHECK: call {{.*}}<2 x i64> @llvm.x86.avx512.vpmadd52l.uq.128(<2 x i64> %{{.*}}, <2 x i64> %{{.*}}, <2 x i64> %{{.*}}) // CHECK: select <2 x i1> %{{.*}}, <2 x i64> %{{.*}}, <2 x i64> %{{.*}} - return _mm_mask_madd52lo_epu64(__W, __M, __X, __Y); + return _mm_mask_madd52lo_epu64(__W, __M, __X, __Y); } +TEST_CONSTEXPR(match_v2di(_mm_mask_madd52lo_epu64((__m128i)((__v2du){1000, 2000}), + 0x3, + (__m128i)((__v2du){100, 200}), + (__m128i)((__v2du){20, 30})), + 3000, 8000), + "mm_mask_madd52lo_epu64: all mask bits set"); + +TEST_CONSTEXPR(match_v2di(_mm_mask_madd52lo_epu64((__m128i)((__v2du){111, 222}), + 0x0, + (__m128i)((__v2du){1, 2}), + (__m128i)((__v2du){10, 20})), + 111, 222), + "mm_mask_madd52lo_epu64: zero mask preserves __W"); + __m128i test_mm_maskz_madd52lo_epu64(__mmask8 __M, __m128i __X, __m128i __Y, __m128i __Z) { // CHECK-LABEL: test_mm_maskz_madd52lo_epu64 // CHECK: call {{.*}}<2 x i64> @llvm.x86.avx512.vpmadd52l.uq.128(<2 x i64> %{{.*}}, <2 x i64> %{{.*}}, <2 x i64> %{{.*}}) // CHECK: select <2 x i1> %{{.*}}, <2 x i64> %{{.*}}, <2 x i64> %{{.*}} - return _mm_maskz_madd52lo_epu64(__M, __X, __Y, __Z); + return _mm_maskz_madd52lo_epu64(__M, __X, __Y, __Z); } +TEST_CONSTEXPR(match_v2di(_mm_maskz_madd52lo_epu64(0x3, + (__m128i)((__v2du){100, 200}), + (__m128i)((__v2du){20, 30}), + (__m128i)((__v2du){30, 40})), + 700, 1400), + "mm_maskz_madd52lo_epu64: all mask bits set"); + +TEST_CONSTEXPR(match_v2di(_mm_maskz_madd52lo_epu64(0x1, + (__m128i)((__v2du){100, 0}), + (__m128i)((__v2du){20, 0}), + (__m128i)((__v2du){30, 0})), + 700, 0), + "mm_maskz_madd52lo_epu64: first lane only"); + __m256i test_mm256_madd52lo_epu64(__m256i __X, __m256i __Y, __m256i __Z) { // CHECK-LABEL: test_mm256_madd52lo_epu64 // CHECK: call {{.*}}<4 x i64> @llvm.x86.avx512.vpmadd52l.uq.256(<4 x i64> %{{.*}}, <4 x i64> %{{.*}}, <4 x i64> %{{.*}}) - return _mm256_madd52lo_epu64(__X, __Y, __Z); + return _mm256_madd52lo_epu64(__X, __Y, __Z); } +TEST_CONSTEXPR(match_v4di(_mm256_madd52lo_epu64( + (__m256i)((__v4du){1, 2, 3, 4}), + (__m256i)((__v4du){10, 20, 30, 40}), + (__m256i)((__v4du){2, 3, 4, 5})), + 21, 62, 123, 204), + "mm256_madd52lo_epu64: four-lane computation"); + __m256i test_mm256_mask_madd52lo_epu64(__m256i __W, __mmask8 __M, __m256i __X, __m256i __Y) { // CHECK-LABEL: test_mm256_mask_madd52lo_epu64 // CHECK: call {{.*}}<4 x i64> @llvm.x86.avx512.vpmadd52l.uq.256(<4 x i64> %{{.*}}, <4 x i64> %{{.*}}, <4 x i64> %{{.*}}) // CHECK: select <4 x i1> %{{.*}}, <4 x i64> %{{.*}}, <4 x i64> %{{.*}} - return _mm256_mask_madd52lo_epu64(__W, __M, __X, __Y); + return _mm256_mask_madd52lo_epu64(__W, __M, __X, __Y); } +TEST_CONSTEXPR(match_v4di(_mm256_mask_madd52lo_epu64((__m256i)((__v4du){1000, 2000, 3000, 4000}), + 0xF, + (__m256i)((__v4du){100, 200, 300, 400}), + (__m256i)((__v4du){20, 30, 40, 50})), + 3000, 8000, 15000, 24000), + "mm256_mask_madd52lo_epu64: all mask bits set"); + +TEST_CONSTEXPR(match_v4di(_mm256_mask_madd52lo_epu64((__m256i)((__v4du){111, 222, 333, 444}), + 0x0, + (__m256i)((__v4du){1, 2, 3, 4}), + (__m256i)((__v4du){10, 20, 30, 40})), + 111, 222, 333, 444), + "mm256_mask_madd52lo_epu64: zero mask preserves __W"); + +TEST_CONSTEXPR(match_v4di(_mm256_mask_madd52lo_epu64((__m256i)((__v4du){11, 22, 33, 44}), + 0x5, + (__m256i)((__v4du){100, 200, 300, 400}), + (__m256i)((__v4du){10, 20, 30, 40})), + 1011, 22, 9033, 44), + "mm256_mask_madd52lo_epu64: mask 0x5"); + __m256i test_mm256_maskz_madd52lo_epu64(__mmask8 __M, __m256i __X, __m256i __Y, __m256i __Z) { // CHECK-LABEL: test_mm256_maskz_madd52lo_epu64 // CHECK: call {{.*}}<4 x i64> @llvm.x86.avx512.vpmadd52l.uq.256(<4 x i64> %{{.*}}, <4 x i64> %{{.*}}, <4 x i64> %{{.*}}) // CHECK: select <4 x i1> %{{.*}}, <4 x i64> %{{.*}}, <4 x i64> %{{.*}} - return _mm256_maskz_madd52lo_epu64(__M, __X, __Y, __Z); + return _mm256_maskz_madd52lo_epu64(__M, __X, __Y, __Z); } + +TEST_CONSTEXPR(match_v4di(_mm256_maskz_madd52lo_epu64(0xF, + (__m256i)((__v4du){100, 200, 300, 400}), + (__m256i)((__v4du){20, 30, 40, 50}), + (__m256i)((__v4du){30, 40, 50, 60})), + 700, 1400, 2300, 3400), + "mm256_maskz_madd52lo_epu64: all mask bits set"); + +TEST_CONSTEXPR(match_v4di(_mm256_maskz_madd52lo_epu64(0x9, + (__m256i)((__v4du){100, 200, 300, 400}), + (__m256i)((__v4du){10, 20, 30, 40}), + (__m256i)((__v4du){5, 10, 15, 20})), + 150, 0, 0, 1200), + "mm256_maskz_madd52lo_epu64: mask 0x9"); diff --git a/clang/test/CodeGen/X86/avxifma-builtins.c b/clang/test/CodeGen/X86/avxifma-builtins.c index aa151591ed143..6150f87f039fc 100644 --- a/clang/test/CodeGen/X86/avxifma-builtins.c +++ b/clang/test/CodeGen/X86/avxifma-builtins.c @@ -3,58 +3,140 @@ // RUN: %clang_cc1 -x c++ -ffreestanding %s -triple=x86_64-apple-darwin -target-feature +avxifma -emit-llvm -o - -Wall -Werror | FileCheck %s // RUN: %clang_cc1 -x c++ -ffreestanding %s -triple=i386-apple-darwin -target-feature +avxifma -emit-llvm -o - -Wall -Werror | FileCheck %s -// RUN: %clang_cc1 -x c -ffreestanding %s -triple=x86_64-apple-darwin -target-feature +avxifma -emit-llvm -o - -Wall -Werror -fexperimental-new-constant-interpreter | FileCheck %s -// RUN: %clang_cc1 -x c -ffreestanding %s -triple=i386-apple-darwin -target-feature +avxifma -emit-llvm -o - -Wall -Werror -fexperimental-new-constant-interpreter | FileCheck %s -// RUN: %clang_cc1 -x c++ -ffreestanding %s -triple=x86_64-apple-darwin -target-feature +avxifma -emit-llvm -o - -Wall -Werror -fexperimental-new-constant-interpreter | FileCheck %s -// RUN: %clang_cc1 -x c++ -ffreestanding %s -triple=i386-apple-darwin -target-feature +avxifma -emit-llvm -o - -Wall -Werror -fexperimental-new-constant-interpreter | FileCheck %s - - #include <immintrin.h> - -__m128i test_mm_madd52hi_epu64(__m128i __X, __m128i __Y, __m128i __Z) { -// CHECK-LABEL: test_mm_madd52hi_epu64 -// CHECK: call {{.*}}<2 x i64> @llvm.x86.avx512.vpmadd52h.uq.128(<2 x i64> %{{.*}}, <2 x i64> %{{.*}}, <2 x i64> %{{.*}}) - return _mm_madd52hi_epu64(__X, __Y, __Z); -} - -__m256i test_mm256_madd52hi_epu64(__m256i __X, __m256i __Y, __m256i __Z) { -// CHECK-LABEL: test_mm256_madd52hi_epu64 -// CHECK: call {{.*}}<4 x i64> @llvm.x86.avx512.vpmadd52h.uq.256(<4 x i64> %{{.*}}, <4 x i64> %{{.*}}, <4 x i64> %{{.*}}) - return _mm256_madd52hi_epu64(__X, __Y, __Z); -} - -__m128i test_mm_madd52lo_epu64(__m128i __X, __m128i __Y, __m128i __Z) { -// CHECK-LABEL: test_mm_madd52lo_epu64 -// CHECK: call {{.*}}<2 x i64> @llvm.x86.avx512.vpmadd52l.uq.128(<2 x i64> %{{.*}}, <2 x i64> %{{.*}}, <2 x i64> %{{.*}}) - return _mm_madd52lo_epu64(__X, __Y, __Z); -} - -__m256i test_mm256_madd52lo_epu64(__m256i __X, __m256i __Y, __m256i __Z) { -// CHECK-LABEL: test_mm256_madd52lo_epu64 -// CHECK: call {{.*}}<4 x i64> @llvm.x86.avx512.vpmadd52l.uq.256(<4 x i64> %{{.*}}, <4 x i64> %{{.*}}, <4 x i64> %{{.*}}) - return _mm256_madd52lo_epu64(__X, __Y, __Z); -} +#include "builtin_test_helpers.h" __m128i test_mm_madd52hi_avx_epu64(__m128i __X, __m128i __Y, __m128i __Z) { -// CHECK-LABEL: test_mm_madd52hi_avx_epu64 -// CHECK: call {{.*}}<2 x i64> @llvm.x86.avx512.vpmadd52h.uq.128(<2 x i64> %{{.*}}, <2 x i64> %{{.*}}, <2 x i64> %{{.*}}) + // CHECK-LABEL: test_mm_madd52hi_avx_epu64 + // CHECK: call {{.*}}<2 x i64> @llvm.x86.avx512.vpmadd52h.uq.128(<2 x i64> %{{.*}}, <2 x i64> %{{.*}}, <2 x i64> %{{.*}}) return _mm_madd52hi_avx_epu64(__X, __Y, __Z); } __m256i test_mm256_madd52hi_avx_epu64(__m256i __X, __m256i __Y, __m256i __Z) { -// CHECK-LABEL: test_mm256_madd52hi_avx_epu64 -// CHECK: call {{.*}}<4 x i64> @llvm.x86.avx512.vpmadd52h.uq.256(<4 x i64> %{{.*}}, <4 x i64> %{{.*}}, <4 x i64> %{{.*}}) + // CHECK-LABEL: test_mm256_madd52hi_avx_epu64 + // CHECK: call {{.*}}<4 x i64> @llvm.x86.avx512.vpmadd52h.uq.256(<4 x i64> %{{.*}}, <4 x i64> %{{.*}}, <4 x i64> %{{.*}}) return _mm256_madd52hi_avx_epu64(__X, __Y, __Z); } __m128i test_mm_madd52lo_avx_epu64(__m128i __X, __m128i __Y, __m128i __Z) { -// CHECK-LABEL: test_mm_madd52lo_avx_epu64 -// CHECK: call {{.*}}<2 x i64> @llvm.x86.avx512.vpmadd52l.uq.128(<2 x i64> %{{.*}}, <2 x i64> %{{.*}}, <2 x i64> %{{.*}}) + // CHECK-LABEL: test_mm_madd52lo_avx_epu64 + // CHECK: call {{.*}}<2 x i64> @llvm.x86.avx512.vpmadd52l.uq.128(<2 x i64> %{{.*}}, <2 x i64> %{{.*}}, <2 x i64> %{{.*}}) return _mm_madd52lo_avx_epu64(__X, __Y, __Z); } __m256i test_mm256_madd52lo_avx_epu64(__m256i __X, __m256i __Y, __m256i __Z) { -// CHECK-LABEL: test_mm256_madd52lo_avx_epu64 -// CHECK: call {{.*}}<4 x i64> @llvm.x86.avx512.vpmadd52l.uq.256(<4 x i64> %{{.*}}, <4 x i64> %{{.*}}, <4 x i64> %{{.*}}) + // CHECK-LABEL: test_mm256_madd52lo_avx_epu64 + // CHECK: call {{.*}}<4 x i64> @llvm.x86.avx512.vpmadd52l.uq.256(<4 x i64> %{{.*}}, <4 x i64> %{{.*}}, <4 x i64> %{{.*}}) return _mm256_madd52lo_avx_epu64(__X, __Y, __Z); } + +TEST_CONSTEXPR(match_v2di(_mm_madd52lo_avx_epu64( + (__m128i)((__v2du){0, 0}), + (__m128i)((__v2du){10, 0}), + (__m128i)((__v2du){5, 0})), + 50, 0), + "mm_madd52lo_avx_epu64: basic multiply-add low bits"); + +TEST_CONSTEXPR(match_v2di(_mm_madd52lo_avx_epu64( + (__m128i)((__v2du){100, 0}), + (__m128i)((__v2du){20, 0}), + (__m128i)((__v2du){30, 0})), + 700, 0), + "mm_madd52lo_avx_epu64: accumulator test"); + +TEST_CONSTEXPR(match_v2di(_mm_madd52lo_avx_epu64( + (__m128i)((__v2du){1, 2}), + (__m128i)((__v2du){10, 20}), + (__m128i)((__v2du){2, 3})), + 21, 62), + "mm_madd52lo_avx_epu64: two-lane computation"); + +TEST_CONSTEXPR(match_v2di(_mm_madd52lo_avx_epu64( + (__m128i)((__v2du){0, 0}), + (__m128i)((__v2du){0xFFFFFFFFFFFFFull, 0}), + (__m128i)((__v2du){1, 0})), + 0xFFFFFFFFFFFFFull, 0), + "mm_madd52lo_avx_epu64: max 52-bit value"); + +TEST_CONSTEXPR(match_v4di(_mm256_madd52lo_avx_epu64( + (__m256i)((__v4du){1, 2, 3, 4}), + (__m256i)((__v4du){10, 20, 30, 40}), + (__m256i)((__v4du){2, 3, 4, 5})), + 21, 62, 123, 204), + "mm256_madd52lo_avx_epu64: four-lane computation"); + +TEST_CONSTEXPR(match_v2di(_mm_madd52hi_avx_epu64( + (__m128i)((__v2du){100, 0}), + (__m128i)((__v2du){10, 0}), + (__m128i)((__v2du){5, 0})), + 100, 0), + "mm_madd52hi_avx_epu64: high bits remain in accumulator"); + +TEST_CONSTEXPR(match_v2di(_mm_madd52hi_avx_epu64( + (__m128i)((__v2du){0, 0}), + (__m128i)((__v2du){0xFFFFFFFFFFFFFull, 0}), + (__m128i)((__v2du){0xFFFFFFFFFFFFFull, 0})), + 0xFFFFFFFFFFFFEull, 0), + "mm_madd52hi_avx_epu64: max value multiply high bits"); + +TEST_CONSTEXPR(match_v4di(_mm256_madd52hi_avx_epu64( + (__m256i)((__v4du){100, 200, 300, 400}), + (__m256i)((__v4du){10, 20, 30, 40}), + (__m256i)((__v4du){5, 6, 7, 8})), + 100, 200, 300, 400), + "mm256_madd52hi_avx_epu64: accumulator preserved"); + +TEST_CONSTEXPR(match_v4di(_mm256_madd52hi_avx_epu64( + (__m256i)((__v4du){0, 0, 0, 0}), + (__m256i)((__v4du){0xFFFFFFFFFFFFFull, 0, 0, + 0}), + (__m256i)((__v4du){0xFFFFFFFFFFFFFull, 0, 0, + 0})), + 0xFFFFFFFFFFFFEull, 0, 0, 0), + "mm256_madd52hi_avx_epu64: single lane max value"); + +TEST_CONSTEXPR(match_v4di(_mm256_madd52lo_avx_epu64( + (__m256i)((__v4du){0, 0, 0, 0}), + (__m256i)((__v4du){0xFFFFFFFFFFFFFull, 0, 0, + 0}), + (__m256i)((__v4du){1, 0, 0, 0})), + 0xFFFFFFFFFFFFFull, 0, 0, 0), + "mm256_madd52lo_avx_epu64: single lane max value"); + +TEST_CONSTEXPR(match_v4di(_mm256_madd52lo_avx_epu64( + (__m256i)((__v4du){0, 0, 0, 0}), + (__m256i)((__v4du){0x1F000000000000ull, 0, 0, + 0}), + (__m256i)((__v4du){2, 0, 0, 0})), + 0xE000000000000ull, 0, 0, 0), + "mm256_madd52lo_avx_epu64: 52-bit truncation test"); + +TEST_CONSTEXPR(match_v2di(_mm_madd52lo_avx_epu64( + (__m128i)((__v2du){5, 10}), + (__m128i)((__v2du){100, 200}), + (__m128i)((__v2du){7, 8})), + 705, 1610), + "mm_madd52lo_avx_epu64: two-lane accumulator"); + +TEST_CONSTEXPR(match_v4di(_mm256_madd52lo_avx_epu64( + (__m256i)((__v4du){1, 2, 3, 4}), + (__m256i)((__v4du){10, 20, 30, 40}), + (__m256i)((__v4du){2, 3, 4, 5})), + 21, 62, 123, 204), + "mm256_madd52lo_avx_epu64: four-lane computation"); + +TEST_CONSTEXPR(match_v2di(_mm_madd52hi_avx_epu64( + (__m128i)((__v2du){50, 100}), + (__m128i)((__v2du){10, 20}), + (__m128i)((__v2du){5, 6})), + 50, 100), + "mm_madd52hi_avx_epu64: accumulator preserved"); + +TEST_CONSTEXPR(match_v4di(_mm256_madd52hi_avx_epu64( + (__m256i)((__v4du){0, 0, 0, 0}), + (__m256i)((__v4du){0xFFFFFFFFFFFFFull, + 0xFFFFFFFFFFFFFull, 0, 0}), + (__m256i)((__v4du){0xFFFFFFFFFFFFFull, + 0xFFFFFFFFFFFFFull, 0, 0})), + 0xFFFFFFFFFFFFEull, 0xFFFFFFFFFFFFEull, 0, 0), + "mm256_madd52hi_avx_epu64: two-lane max value"); _______________________________________________ cfe-commits mailing list [email protected] https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
