[Clang] VectorExprEvaluator::VisitCallExpr / InterpretBuiltin - Allow AVX512 VPMULTISHIFTQB intrinsics to be used in constexpr #168995

chaitanyav · 2025-11-21T03:56:26Z

Resolves: #167477

github-actions · 2025-11-21T04:20:04Z

🐧 Linux x64 Test Results

111425 tests passed
4451 tests skipped

llvmbot · 2025-11-21T05:11:42Z

@llvm/pr-subscribers-backend-x86

Author: NagaChaitanya Vellanki (chaitanyav)

Changes

Resolves: #167477

Patch is 32.87 KiB, truncated to 20.00 KiB below, full version: https://github.com/llvm/llvm-project/pull/168995.diff

7 Files Affected:

(modified) clang/include/clang/Basic/BuiltinsX86.td (+3-3)
(modified) clang/lib/AST/ByteCode/InterpBuiltin.cpp (+63)
(modified) clang/lib/AST/ExprConstant.cpp (+45)
(modified) clang/lib/Headers/avx512vbmiintrin.h (+17-21)
(modified) clang/lib/Headers/avx512vbmivlintrin.h (+32-40)
(modified) clang/test/CodeGen/X86/avx512vbmi-builtins.c (+130-3)
(modified) clang/test/CodeGen/X86/avx512vbmivl-builtin.c (+88-6)

diff --git a/clang/include/clang/Basic/BuiltinsX86.td b/clang/include/clang/Basic/BuiltinsX86.td index b760c3e06b8f7..93fb511a508f3 100644 --- a/clang/include/clang/Basic/BuiltinsX86.td +++ b/clang/include/clang/Basic/BuiltinsX86.td @@ -3358,15 +3358,15 @@ let Features = "avx512f", Attributes = [NoThrow, Const, RequiredVectorWidth<128> def cvtusi2ss32 : X86Builtin<"_Vector<4, float>(_Vector<4, float>, unsigned int, _Constant int)">; } -let Features = "avx512vbmi", Attributes = [NoThrow, Const, RequiredVectorWidth<512>] in { +let Features = "avx512vbmi", Attributes = [NoThrow, Const, Constexpr, RequiredVectorWidth<512>] in { def vpmultishiftqb512 : X86Builtin<"_Vector<64, char>(_Vector<64, char>, _Vector<64, char>)">; } -let Features = "avx512vbmi,avx512vl", Attributes = [NoThrow, Const, RequiredVectorWidth<128>] in { +let Features = "avx512vbmi,avx512vl", Attributes = [NoThrow, Const, Constexpr, RequiredVectorWidth<128>] in { def vpmultishiftqb128 : X86Builtin<"_Vector<16, char>(_Vector<16, char>, _Vector<16, char>)">; } -let Features = "avx512vbmi,avx512vl", Attributes = [NoThrow, Const, RequiredVectorWidth<256>] in { +let Features = "avx512vbmi,avx512vl", Attributes = [NoThrow, Const, Constexpr, RequiredVectorWidth<256>] in { def vpmultishiftqb256 : X86Builtin<"_Vector<32, char>(_Vector<32, char>, _Vector<32, char>)">; } diff --git a/clang/lib/AST/ByteCode/InterpBuiltin.cpp b/clang/lib/AST/ByteCode/InterpBuiltin.cpp index eba71d66bc4d6..5be31239fd597 100644 --- a/clang/lib/AST/ByteCode/InterpBuiltin.cpp +++ b/clang/lib/AST/ByteCode/InterpBuiltin.cpp @@ -3468,6 +3468,65 @@ static bool interp__builtin_ia32_shuffle_generic( return true; } +static bool interp__builtin_ia32_multishiftqb(InterpState &S, CodePtr OpPC, + const CallExpr *Call) { + assert(Call->getNumArgs() == 2); + + QualType ATy = Call->getArg(0)->getType(); + QualType BTy = Call->getArg(1)->getType(); + if (!ATy->isVectorType() || !BTy->isVectorType()) { + return false; + } + + const Pointer &BPtr = S.Stk.pop<Pointer>(); + const Pointer &APtr = S.Stk.pop<Pointer>(); + const auto *AVecT = ATy->castAs<VectorType>(); + const auto *BVecT = BTy->castAs<VectorType>(); + assert(AVecT->getNumElements() == BVecT->getNumElements()); + + PrimType ElemT = *S.getContext().classify(AVecT->getElementType()); + + unsigned NumBytesInQWord = 8; + unsigned NumBitsInByte = 8; + unsigned NumBytes = AVecT->getNumElements(); + unsigned NumQWords = NumBytes / NumBytesInQWord; + const Pointer &Dst = S.Stk.peek<Pointer>(); + + for (unsigned QWordId = 0; QWordId != NumQWords; ++QWordId) { + APInt AQWord(64, 0); + APInt BQWord(64, 0); + for (unsigned ByteIdx = 0; ByteIdx != NumBytesInQWord; ++ByteIdx) { + unsigned Idx = QWordId * NumBytesInQWord + ByteIdx; + uint64_t Byte = 0; + INT_TYPE_SWITCH(ElemT, { + Byte = static_cast<uint64_t>(APtr.elem<T>(Idx)); + AQWord.insertBits(APInt(8, Byte & 0xFF), ByteIdx * NumBitsInByte); + + Byte = static_cast<uint64_t>(BPtr.elem<T>(Idx)); + BQWord.insertBits(APInt(8, Byte & 0xFF), ByteIdx * NumBitsInByte); + }); + } + + for (unsigned ByteIdx = 0; ByteIdx != NumBytesInQWord; ++ByteIdx) { + uint64_t Ctrl = + AQWord.extractBits(8, ByteIdx * NumBitsInByte).getZExtValue() & 0x3F; + + APInt Byte(8, 0); + for (unsigned BitIdx = 0; BitIdx != NumBitsInByte; ++BitIdx) { + Byte.insertBits(BQWord.extractBits(1, (Ctrl + BitIdx) & 0x3F), BitIdx); + } + INT_TYPE_SWITCH(ElemT, { + Dst.elem<T>(QWordId * NumBytesInQWord + ByteIdx) = + T::from(Byte.getZExtValue()); + }); + } + } + + Dst.initializeAllElements(); + + return true; +} + bool InterpretBuiltin(InterpState &S, CodePtr OpPC, const CallExpr *Call, uint32_t BuiltinID) { if (!S.getASTContext().BuiltinInfo.isConstantEvaluated(BuiltinID)) @@ -4669,6 +4728,10 @@ bool InterpretBuiltin(InterpState &S, CodePtr OpPC, const CallExpr *Call, return std::make_pair(0, static_cast<int>(LaneOffset + Index)); }); + case X86::BI__builtin_ia32_vpmultishiftqb128: + case X86::BI__builtin_ia32_vpmultishiftqb256: + case X86::BI__builtin_ia32_vpmultishiftqb512: + return interp__builtin_ia32_multishiftqb(S, OpPC, Call); case X86::BI__builtin_ia32_kandqi: case X86::BI__builtin_ia32_kandhi: case X86::BI__builtin_ia32_kandsi: diff --git a/clang/lib/AST/ExprConstant.cpp b/clang/lib/AST/ExprConstant.cpp index ce5301f17b3e7..21a664fefde49 100644 --- a/clang/lib/AST/ExprConstant.cpp +++ b/clang/lib/AST/ExprConstant.cpp @@ -13062,6 +13062,51 @@ bool VectorExprEvaluator::VisitCallExpr(const CallExpr *E) { return Success(R, E); } + case X86::BI__builtin_ia32_vpmultishiftqb128: + case X86::BI__builtin_ia32_vpmultishiftqb256: + case X86::BI__builtin_ia32_vpmultishiftqb512: { + assert(E->getNumArgs() == 2); + + APValue A, B; + if (!Evaluate(A, Info, E->getArg(0)) || !Evaluate(B, Info, E->getArg(1))) + return false; + + assert(A.getVectorLength() == B.getVectorLength()); + unsigned NumBytesInQWord = 8; + unsigned NumBitsInByte = 8; + unsigned NumBytes = A.getVectorLength(); + unsigned NumQWords = NumBytes / NumBytesInQWord; + SmallVector<APValue, 64> Result; + Result.reserve(NumBytes); + + for (unsigned QWordId = 0; QWordId != NumQWords; ++QWordId) { + APInt AQWord(64, 0); + APInt BQWord(64, 0); + for (unsigned ByteIdx = 0; ByteIdx != NumBytesInQWord; ++ByteIdx) { + unsigned Idx = QWordId * NumBytesInQWord + ByteIdx; + uint64_t Byte = A.getVectorElt(Idx).getInt().getZExtValue(); + AQWord.insertBits(APInt(8, Byte & 0xFF), ByteIdx * NumBitsInByte); + + Byte = B.getVectorElt(Idx).getInt().getZExtValue(); + BQWord.insertBits(APInt(8, Byte & 0xFF), ByteIdx * NumBitsInByte); + } + + for (unsigned ByteIdx = 0; ByteIdx != NumBytesInQWord; ++ByteIdx) { + uint64_t Ctrl = + AQWord.extractBits(8, ByteIdx * NumBitsInByte).getZExtValue() & + 0x3F; + + APInt Byte(8, 0); + for (unsigned BitIdx = 0; BitIdx != NumBitsInByte; ++BitIdx) { + Byte.insertBits(BQWord.extractBits(1, (Ctrl + BitIdx) & 0x3F), + BitIdx); + } + Result.push_back(APValue(APSInt(Byte, /*isUnsigned*/ true))); + } + } + return Success(APValue(Result.data(), Result.size()), E); + } + case X86::BI__builtin_ia32_phminposuw128: { APValue Source; if (!Evaluate(Source, Info, E->getArg(0))) diff --git a/clang/lib/Headers/avx512vbmiintrin.h b/clang/lib/Headers/avx512vbmiintrin.h index 84fda5c5849e8..5ac78f0849c26 100644 --- a/clang/lib/Headers/avx512vbmiintrin.h +++ b/clang/lib/Headers/avx512vbmiintrin.h @@ -15,61 +15,57 @@ #define __VBMIINTRIN_H /* Define the default attributes for the functions in this file. */ +#if defined(__cplusplus) && (__cplusplus >= 201103L) #define __DEFAULT_FN_ATTRS \ __attribute__((__always_inline__, __nodebug__, __target__("avx512vbmi"), \ - __min_vector_width__(512))) - -#if defined(__cplusplus) && (__cplusplus >= 201103L) -#define __DEFAULT_FN_ATTRS_CONSTEXPR __DEFAULT_FN_ATTRS constexpr + __min_vector_width__(512))) constexpr #else -#define __DEFAULT_FN_ATTRS_CONSTEXPR __DEFAULT_FN_ATTRS +#define __DEFAULT_FN_ATTRS \ + __attribute__((__always_inline__, __nodebug__, __target__("avx512vbmi"), \ + __min_vector_width__(512))) #endif -static __inline__ __m512i __DEFAULT_FN_ATTRS_CONSTEXPR +static __inline__ __m512i __DEFAULT_FN_ATTRS _mm512_permutex2var_epi8(__m512i __A, __m512i __I, __m512i __B) { return (__m512i)__builtin_ia32_vpermi2varqi512((__v64qi)__A, (__v64qi)__I, (__v64qi) __B); } -static __inline__ __m512i __DEFAULT_FN_ATTRS_CONSTEXPR -_mm512_mask_permutex2var_epi8(__m512i __A, __mmask64 __U, __m512i __I, - __m512i __B) { +static __inline__ __m512i __DEFAULT_FN_ATTRS _mm512_mask_permutex2var_epi8( + __m512i __A, __mmask64 __U, __m512i __I, __m512i __B) { return (__m512i)__builtin_ia32_selectb_512(__U, (__v64qi)_mm512_permutex2var_epi8(__A, __I, __B), (__v64qi)__A); } -static __inline__ __m512i __DEFAULT_FN_ATTRS_CONSTEXPR -_mm512_mask2_permutex2var_epi8(__m512i __A, __m512i __I, __mmask64 __U, - __m512i __B) { +static __inline__ __m512i __DEFAULT_FN_ATTRS _mm512_mask2_permutex2var_epi8( + __m512i __A, __m512i __I, __mmask64 __U, __m512i __B) { return (__m512i)__builtin_ia32_selectb_512(__U, (__v64qi)_mm512_permutex2var_epi8(__A, __I, __B), (__v64qi)__I); } -static __inline__ __m512i __DEFAULT_FN_ATTRS_CONSTEXPR -_mm512_maskz_permutex2var_epi8(__mmask64 __U, __m512i __A, __m512i __I, - __m512i __B) { +static __inline__ __m512i __DEFAULT_FN_ATTRS _mm512_maskz_permutex2var_epi8( + __mmask64 __U, __m512i __A, __m512i __I, __m512i __B) { return (__m512i)__builtin_ia32_selectb_512(__U, (__v64qi)_mm512_permutex2var_epi8(__A, __I, __B), (__v64qi)_mm512_setzero_si512()); } -static __inline__ __m512i __DEFAULT_FN_ATTRS_CONSTEXPR +static __inline__ __m512i __DEFAULT_FN_ATTRS _mm512_permutexvar_epi8(__m512i __A, __m512i __B) { return (__m512i)__builtin_ia32_permvarqi512((__v64qi) __B, (__v64qi) __A); } -static __inline__ __m512i __DEFAULT_FN_ATTRS_CONSTEXPR +static __inline__ __m512i __DEFAULT_FN_ATTRS _mm512_maskz_permutexvar_epi8(__mmask64 __M, __m512i __A, __m512i __B) { return (__m512i)__builtin_ia32_selectb_512((__mmask64)__M, (__v64qi)_mm512_permutexvar_epi8(__A, __B), (__v64qi)_mm512_setzero_si512()); } -static __inline__ __m512i __DEFAULT_FN_ATTRS_CONSTEXPR -_mm512_mask_permutexvar_epi8(__m512i __W, __mmask64 __M, __m512i __A, - __m512i __B) { +static __inline__ __m512i __DEFAULT_FN_ATTRS _mm512_mask_permutexvar_epi8( + __m512i __W, __mmask64 __M, __m512i __A, __m512i __B) { return (__m512i)__builtin_ia32_selectb_512((__mmask64)__M, (__v64qi)_mm512_permutexvar_epi8(__A, __B), (__v64qi)__W); @@ -97,6 +93,6 @@ _mm512_maskz_multishift_epi64_epi8(__mmask64 __M, __m512i __X, __m512i __Y) (__v64qi)_mm512_multishift_epi64_epi8(__X, __Y), (__v64qi)_mm512_setzero_si512()); } -#undef __DEFAULT_FN_ATTRS_CONSTEXPR + #undef __DEFAULT_FN_ATTRS #endif diff --git a/clang/lib/Headers/avx512vbmivlintrin.h b/clang/lib/Headers/avx512vbmivlintrin.h index 58a48dadff863..40a67bd63ca49 100644 --- a/clang/lib/Headers/avx512vbmivlintrin.h +++ b/clang/lib/Headers/avx512vbmivlintrin.h @@ -15,6 +15,16 @@ #define __VBMIVLINTRIN_H /* Define the default attributes for the functions in this file. */ +#if defined(__cplusplus) && (__cplusplus >= 201103L) +#define __DEFAULT_FN_ATTRS128 \ + __attribute__((__always_inline__, __nodebug__, \ + __target__("avx512vbmi,avx512vl"), \ + __min_vector_width__(128))) constexpr +#define __DEFAULT_FN_ATTRS256 \ + __attribute__((__always_inline__, __nodebug__, \ + __target__("avx512vbmi,avx512vl"), \ + __min_vector_width__(256))) constexpr +#else #define __DEFAULT_FN_ATTRS128 \ __attribute__((__always_inline__, __nodebug__, \ __target__("avx512vbmi,avx512vl"), \ @@ -23,111 +33,96 @@ __attribute__((__always_inline__, __nodebug__, \ __target__("avx512vbmi,avx512vl"), \ __min_vector_width__(256))) - -#if defined(__cplusplus) && (__cplusplus >= 201103L) -#define __DEFAULT_FN_ATTRS128_CONSTEXPR __DEFAULT_FN_ATTRS128 constexpr -#define __DEFAULT_FN_ATTRS256_CONSTEXPR __DEFAULT_FN_ATTRS256 constexpr -#else -#define __DEFAULT_FN_ATTRS128_CONSTEXPR __DEFAULT_FN_ATTRS128 -#define __DEFAULT_FN_ATTRS256_CONSTEXPR __DEFAULT_FN_ATTRS256 #endif -static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR +static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_permutex2var_epi8(__m128i __A, __m128i __I, __m128i __B) { return (__m128i)__builtin_ia32_vpermi2varqi128((__v16qi)__A, (__v16qi)__I, (__v16qi)__B); } -static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR -_mm_mask_permutex2var_epi8(__m128i __A, __mmask16 __U, __m128i __I, - __m128i __B) { +static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_mask_permutex2var_epi8( + __m128i __A, __mmask16 __U, __m128i __I, __m128i __B) { return (__m128i)__builtin_ia32_selectb_128(__U, (__v16qi)_mm_permutex2var_epi8(__A, __I, __B), (__v16qi)__A); } -static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR -_mm_mask2_permutex2var_epi8(__m128i __A, __m128i __I, __mmask16 __U, - __m128i __B) { +static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_mask2_permutex2var_epi8( + __m128i __A, __m128i __I, __mmask16 __U, __m128i __B) { return (__m128i)__builtin_ia32_selectb_128(__U, (__v16qi)_mm_permutex2var_epi8(__A, __I, __B), (__v16qi)__I); } -static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR -_mm_maskz_permutex2var_epi8(__mmask16 __U, __m128i __A, __m128i __I, - __m128i __B) { +static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_maskz_permutex2var_epi8( + __mmask16 __U, __m128i __A, __m128i __I, __m128i __B) { return (__m128i)__builtin_ia32_selectb_128(__U, (__v16qi)_mm_permutex2var_epi8(__A, __I, __B), (__v16qi)_mm_setzero_si128()); } -static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR +static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_permutex2var_epi8(__m256i __A, __m256i __I, __m256i __B) { return (__m256i)__builtin_ia32_vpermi2varqi256((__v32qi)__A, (__v32qi)__I, (__v32qi)__B); } -static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR -_mm256_mask_permutex2var_epi8(__m256i __A, __mmask32 __U, __m256i __I, - __m256i __B) { +static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_mask_permutex2var_epi8( + __m256i __A, __mmask32 __U, __m256i __I, __m256i __B) { return (__m256i)__builtin_ia32_selectb_256(__U, (__v32qi)_mm256_permutex2var_epi8(__A, __I, __B), (__v32qi)__A); } -static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR -_mm256_mask2_permutex2var_epi8(__m256i __A, __m256i __I, __mmask32 __U, - __m256i __B) { +static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_mask2_permutex2var_epi8( + __m256i __A, __m256i __I, __mmask32 __U, __m256i __B) { return (__m256i)__builtin_ia32_selectb_256(__U, (__v32qi)_mm256_permutex2var_epi8(__A, __I, __B), (__v32qi)__I); } -static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR -_mm256_maskz_permutex2var_epi8(__mmask32 __U, __m256i __A, __m256i __I, - __m256i __B) { +static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_maskz_permutex2var_epi8( + __mmask32 __U, __m256i __A, __m256i __I, __m256i __B) { return (__m256i)__builtin_ia32_selectb_256(__U, (__v32qi)_mm256_permutex2var_epi8(__A, __I, __B), (__v32qi)_mm256_setzero_si256()); } -static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR +static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_permutexvar_epi8(__m128i __A, __m128i __B) { return (__m128i)__builtin_ia32_permvarqi128((__v16qi)__B, (__v16qi)__A); } -static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR +static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_maskz_permutexvar_epi8(__mmask16 __M, __m128i __A, __m128i __B) { return (__m128i)__builtin_ia32_selectb_128((__mmask16)__M, (__v16qi)_mm_permutexvar_epi8(__A, __B), (__v16qi)_mm_setzero_si128()); } -static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR -_mm_mask_permutexvar_epi8(__m128i __W, __mmask16 __M, __m128i __A, - __m128i __B) { +static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_mask_permutexvar_epi8( + __m128i __W, __mmask16 __M, __m128i __A, __m128i __B) { return (__m128i)__builtin_ia32_selectb_128((__mmask16)__M, (__v16qi)_mm_permutexvar_epi8(__A, __B), (__v16qi)__W); } -static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR +static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_permutexvar_epi8(__m256i __A, __m256i __B) { return (__m256i)__builtin_ia32_permvarqi256((__v32qi) __B, (__v32qi) __A); } -static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR +static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_maskz_permutexvar_epi8(__mmask32 __M, __m256i __A, __m256i __B) { return (__m256i)__builtin_ia32_selectb_256((__mmask32)__M, (__v32qi)_mm256_permutexvar_epi8(__A, __B), (__v32qi)_mm256_setzero_si256()); } -static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR -_mm256_mask_permutexvar_epi8(__m256i __W, __mmask32 __M, __m256i __A, - __m256i __B) { +static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_mask_permutexvar_epi8( + __m256i __W, __mmask32 __M, __m256i __A, __m256i __B) { return (__m256i)__builtin_ia32_selectb_256((__mmask32)__M, (__v32qi)_mm256_permutexvar_epi8(__A, __B), (__v32qi)__W); @@ -179,9 +174,6 @@ _mm256_maskz_multishift_epi64_epi8(__mmask32 __M, __m256i __X, __m256i __Y) (__v32qi)_mm256_setzero_si256()); } -#undef __DEFAULT_FN_ATTRS128_CONSTEXPR -#undef __DEFAULT_FN_ATTRS256_CONSTEXPR #undef __DEFAULT_FN_ATTRS128 #undef __DEFAULT_FN_ATTRS256 - #endif diff --git a/clang/test/CodeGen/X86/avx512vbmi-builtins.c b/clang/test/CodeGen/X86/avx512vbmi-builtins.c index 7d506db92faeb..fcce58b63737b 100644 --- a/clang/test/CodeGen/X86/avx512vbmi-builtins.c +++ b/clang/test/CodeGen/X86/avx512vbmi-builtins.c @@ -211,18 +211,145 @@ __m512i test_mm512_mask_multishift_epi64_epi8(__m512i __W, __mmask64 __M, __m512 // CHECK-LABEL: test_mm512_mask_multishift_epi64_epi8 // CHECK: call <64 x i8> @llvm.x86.avx512.pmultishift.qb.512(<64 x i8> %{{.*}}, <64 x i8> %{{.*}}) // CHECK: select <64 x i1> %{{.*}}, <64 x i8> %{{.*}}, <64 x i8> %{{.*}} - return _mm512_mask_multishift_epi64_epi8(__W, __M, __X, __Y);  + return _mm512_mask_multishift_epi64_epi8(__W, __M, __X, __Y); } +TEST_CONSTEXPR(match_v64qu( + _mm512_mask_multishift_epi64_epi8( + (__m512i)(__v64qu){ + 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, + 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, + 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, + 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, + 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, + 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, + 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, + 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0... [truncated]

llvmbot · 2025-11-21T05:11:43Z

@llvm/pr-subscribers-clang

Author: NagaChaitanya Vellanki (chaitanyav)

Changes

Resolves: #167477

Patch is 32.87 KiB, truncated to 20.00 KiB below, full version: https://github.com/llvm/llvm-project/pull/168995.diff

7 Files Affected:

(modified) clang/include/clang/Basic/BuiltinsX86.td (+3-3)
(modified) clang/lib/AST/ByteCode/InterpBuiltin.cpp (+63)
(modified) clang/lib/AST/ExprConstant.cpp (+45)
(modified) clang/lib/Headers/avx512vbmiintrin.h (+17-21)
(modified) clang/lib/Headers/avx512vbmivlintrin.h (+32-40)
(modified) clang/test/CodeGen/X86/avx512vbmi-builtins.c (+130-3)
(modified) clang/test/CodeGen/X86/avx512vbmivl-builtin.c (+88-6)

diff --git a/clang/include/clang/Basic/BuiltinsX86.td b/clang/include/clang/Basic/BuiltinsX86.td index b760c3e06b8f7..93fb511a508f3 100644 --- a/clang/include/clang/Basic/BuiltinsX86.td +++ b/clang/include/clang/Basic/BuiltinsX86.td @@ -3358,15 +3358,15 @@ let Features = "avx512f", Attributes = [NoThrow, Const, RequiredVectorWidth<128> def cvtusi2ss32 : X86Builtin<"_Vector<4, float>(_Vector<4, float>, unsigned int, _Constant int)">; } -let Features = "avx512vbmi", Attributes = [NoThrow, Const, RequiredVectorWidth<512>] in { +let Features = "avx512vbmi", Attributes = [NoThrow, Const, Constexpr, RequiredVectorWidth<512>] in { def vpmultishiftqb512 : X86Builtin<"_Vector<64, char>(_Vector<64, char>, _Vector<64, char>)">; } -let Features = "avx512vbmi,avx512vl", Attributes = [NoThrow, Const, RequiredVectorWidth<128>] in { +let Features = "avx512vbmi,avx512vl", Attributes = [NoThrow, Const, Constexpr, RequiredVectorWidth<128>] in { def vpmultishiftqb128 : X86Builtin<"_Vector<16, char>(_Vector<16, char>, _Vector<16, char>)">; } -let Features = "avx512vbmi,avx512vl", Attributes = [NoThrow, Const, RequiredVectorWidth<256>] in { +let Features = "avx512vbmi,avx512vl", Attributes = [NoThrow, Const, Constexpr, RequiredVectorWidth<256>] in { def vpmultishiftqb256 : X86Builtin<"_Vector<32, char>(_Vector<32, char>, _Vector<32, char>)">; } diff --git a/clang/lib/AST/ByteCode/InterpBuiltin.cpp b/clang/lib/AST/ByteCode/InterpBuiltin.cpp index eba71d66bc4d6..5be31239fd597 100644 --- a/clang/lib/AST/ByteCode/InterpBuiltin.cpp +++ b/clang/lib/AST/ByteCode/InterpBuiltin.cpp @@ -3468,6 +3468,65 @@ static bool interp__builtin_ia32_shuffle_generic( return true; } +static bool interp__builtin_ia32_multishiftqb(InterpState &S, CodePtr OpPC, + const CallExpr *Call) { + assert(Call->getNumArgs() == 2); + + QualType ATy = Call->getArg(0)->getType(); + QualType BTy = Call->getArg(1)->getType(); + if (!ATy->isVectorType() || !BTy->isVectorType()) { + return false; + } + + const Pointer &BPtr = S.Stk.pop<Pointer>(); + const Pointer &APtr = S.Stk.pop<Pointer>(); + const auto *AVecT = ATy->castAs<VectorType>(); + const auto *BVecT = BTy->castAs<VectorType>(); + assert(AVecT->getNumElements() == BVecT->getNumElements()); + + PrimType ElemT = *S.getContext().classify(AVecT->getElementType()); + + unsigned NumBytesInQWord = 8; + unsigned NumBitsInByte = 8; + unsigned NumBytes = AVecT->getNumElements(); + unsigned NumQWords = NumBytes / NumBytesInQWord; + const Pointer &Dst = S.Stk.peek<Pointer>(); + + for (unsigned QWordId = 0; QWordId != NumQWords; ++QWordId) { + APInt AQWord(64, 0); + APInt BQWord(64, 0); + for (unsigned ByteIdx = 0; ByteIdx != NumBytesInQWord; ++ByteIdx) { + unsigned Idx = QWordId * NumBytesInQWord + ByteIdx; + uint64_t Byte = 0; + INT_TYPE_SWITCH(ElemT, { + Byte = static_cast<uint64_t>(APtr.elem<T>(Idx)); + AQWord.insertBits(APInt(8, Byte & 0xFF), ByteIdx * NumBitsInByte); + + Byte = static_cast<uint64_t>(BPtr.elem<T>(Idx)); + BQWord.insertBits(APInt(8, Byte & 0xFF), ByteIdx * NumBitsInByte); + }); + } + + for (unsigned ByteIdx = 0; ByteIdx != NumBytesInQWord; ++ByteIdx) { + uint64_t Ctrl = + AQWord.extractBits(8, ByteIdx * NumBitsInByte).getZExtValue() & 0x3F; + + APInt Byte(8, 0); + for (unsigned BitIdx = 0; BitIdx != NumBitsInByte; ++BitIdx) { + Byte.insertBits(BQWord.extractBits(1, (Ctrl + BitIdx) & 0x3F), BitIdx); + } + INT_TYPE_SWITCH(ElemT, { + Dst.elem<T>(QWordId * NumBytesInQWord + ByteIdx) = + T::from(Byte.getZExtValue()); + }); + } + } + + Dst.initializeAllElements(); + + return true; +} + bool InterpretBuiltin(InterpState &S, CodePtr OpPC, const CallExpr *Call, uint32_t BuiltinID) { if (!S.getASTContext().BuiltinInfo.isConstantEvaluated(BuiltinID)) @@ -4669,6 +4728,10 @@ bool InterpretBuiltin(InterpState &S, CodePtr OpPC, const CallExpr *Call, return std::make_pair(0, static_cast<int>(LaneOffset + Index)); }); + case X86::BI__builtin_ia32_vpmultishiftqb128: + case X86::BI__builtin_ia32_vpmultishiftqb256: + case X86::BI__builtin_ia32_vpmultishiftqb512: + return interp__builtin_ia32_multishiftqb(S, OpPC, Call); case X86::BI__builtin_ia32_kandqi: case X86::BI__builtin_ia32_kandhi: case X86::BI__builtin_ia32_kandsi: diff --git a/clang/lib/AST/ExprConstant.cpp b/clang/lib/AST/ExprConstant.cpp index ce5301f17b3e7..21a664fefde49 100644 --- a/clang/lib/AST/ExprConstant.cpp +++ b/clang/lib/AST/ExprConstant.cpp @@ -13062,6 +13062,51 @@ bool VectorExprEvaluator::VisitCallExpr(const CallExpr *E) { return Success(R, E); } + case X86::BI__builtin_ia32_vpmultishiftqb128: + case X86::BI__builtin_ia32_vpmultishiftqb256: + case X86::BI__builtin_ia32_vpmultishiftqb512: { + assert(E->getNumArgs() == 2); + + APValue A, B; + if (!Evaluate(A, Info, E->getArg(0)) || !Evaluate(B, Info, E->getArg(1))) + return false; + + assert(A.getVectorLength() == B.getVectorLength()); + unsigned NumBytesInQWord = 8; + unsigned NumBitsInByte = 8; + unsigned NumBytes = A.getVectorLength(); + unsigned NumQWords = NumBytes / NumBytesInQWord; + SmallVector<APValue, 64> Result; + Result.reserve(NumBytes); + + for (unsigned QWordId = 0; QWordId != NumQWords; ++QWordId) { + APInt AQWord(64, 0); + APInt BQWord(64, 0); + for (unsigned ByteIdx = 0; ByteIdx != NumBytesInQWord; ++ByteIdx) { + unsigned Idx = QWordId * NumBytesInQWord + ByteIdx; + uint64_t Byte = A.getVectorElt(Idx).getInt().getZExtValue(); + AQWord.insertBits(APInt(8, Byte & 0xFF), ByteIdx * NumBitsInByte); + + Byte = B.getVectorElt(Idx).getInt().getZExtValue(); + BQWord.insertBits(APInt(8, Byte & 0xFF), ByteIdx * NumBitsInByte); + } + + for (unsigned ByteIdx = 0; ByteIdx != NumBytesInQWord; ++ByteIdx) { + uint64_t Ctrl = + AQWord.extractBits(8, ByteIdx * NumBitsInByte).getZExtValue() & + 0x3F; + + APInt Byte(8, 0); + for (unsigned BitIdx = 0; BitIdx != NumBitsInByte; ++BitIdx) { + Byte.insertBits(BQWord.extractBits(1, (Ctrl + BitIdx) & 0x3F), + BitIdx); + } + Result.push_back(APValue(APSInt(Byte, /*isUnsigned*/ true))); + } + } + return Success(APValue(Result.data(), Result.size()), E); + } + case X86::BI__builtin_ia32_phminposuw128: { APValue Source; if (!Evaluate(Source, Info, E->getArg(0))) diff --git a/clang/lib/Headers/avx512vbmiintrin.h b/clang/lib/Headers/avx512vbmiintrin.h index 84fda5c5849e8..5ac78f0849c26 100644 --- a/clang/lib/Headers/avx512vbmiintrin.h +++ b/clang/lib/Headers/avx512vbmiintrin.h @@ -15,61 +15,57 @@ #define __VBMIINTRIN_H /* Define the default attributes for the functions in this file. */ +#if defined(__cplusplus) && (__cplusplus >= 201103L) #define __DEFAULT_FN_ATTRS \ __attribute__((__always_inline__, __nodebug__, __target__("avx512vbmi"), \ - __min_vector_width__(512))) - -#if defined(__cplusplus) && (__cplusplus >= 201103L) -#define __DEFAULT_FN_ATTRS_CONSTEXPR __DEFAULT_FN_ATTRS constexpr + __min_vector_width__(512))) constexpr #else -#define __DEFAULT_FN_ATTRS_CONSTEXPR __DEFAULT_FN_ATTRS +#define __DEFAULT_FN_ATTRS \ + __attribute__((__always_inline__, __nodebug__, __target__("avx512vbmi"), \ + __min_vector_width__(512))) #endif -static __inline__ __m512i __DEFAULT_FN_ATTRS_CONSTEXPR +static __inline__ __m512i __DEFAULT_FN_ATTRS _mm512_permutex2var_epi8(__m512i __A, __m512i __I, __m512i __B) { return (__m512i)__builtin_ia32_vpermi2varqi512((__v64qi)__A, (__v64qi)__I, (__v64qi) __B); } -static __inline__ __m512i __DEFAULT_FN_ATTRS_CONSTEXPR -_mm512_mask_permutex2var_epi8(__m512i __A, __mmask64 __U, __m512i __I, - __m512i __B) { +static __inline__ __m512i __DEFAULT_FN_ATTRS _mm512_mask_permutex2var_epi8( + __m512i __A, __mmask64 __U, __m512i __I, __m512i __B) { return (__m512i)__builtin_ia32_selectb_512(__U, (__v64qi)_mm512_permutex2var_epi8(__A, __I, __B), (__v64qi)__A); } -static __inline__ __m512i __DEFAULT_FN_ATTRS_CONSTEXPR -_mm512_mask2_permutex2var_epi8(__m512i __A, __m512i __I, __mmask64 __U, - __m512i __B) { +static __inline__ __m512i __DEFAULT_FN_ATTRS _mm512_mask2_permutex2var_epi8( + __m512i __A, __m512i __I, __mmask64 __U, __m512i __B) { return (__m512i)__builtin_ia32_selectb_512(__U, (__v64qi)_mm512_permutex2var_epi8(__A, __I, __B), (__v64qi)__I); } -static __inline__ __m512i __DEFAULT_FN_ATTRS_CONSTEXPR -_mm512_maskz_permutex2var_epi8(__mmask64 __U, __m512i __A, __m512i __I, - __m512i __B) { +static __inline__ __m512i __DEFAULT_FN_ATTRS _mm512_maskz_permutex2var_epi8( + __mmask64 __U, __m512i __A, __m512i __I, __m512i __B) { return (__m512i)__builtin_ia32_selectb_512(__U, (__v64qi)_mm512_permutex2var_epi8(__A, __I, __B), (__v64qi)_mm512_setzero_si512()); } -static __inline__ __m512i __DEFAULT_FN_ATTRS_CONSTEXPR +static __inline__ __m512i __DEFAULT_FN_ATTRS _mm512_permutexvar_epi8(__m512i __A, __m512i __B) { return (__m512i)__builtin_ia32_permvarqi512((__v64qi) __B, (__v64qi) __A); } -static __inline__ __m512i __DEFAULT_FN_ATTRS_CONSTEXPR +static __inline__ __m512i __DEFAULT_FN_ATTRS _mm512_maskz_permutexvar_epi8(__mmask64 __M, __m512i __A, __m512i __B) { return (__m512i)__builtin_ia32_selectb_512((__mmask64)__M, (__v64qi)_mm512_permutexvar_epi8(__A, __B), (__v64qi)_mm512_setzero_si512()); } -static __inline__ __m512i __DEFAULT_FN_ATTRS_CONSTEXPR -_mm512_mask_permutexvar_epi8(__m512i __W, __mmask64 __M, __m512i __A, - __m512i __B) { +static __inline__ __m512i __DEFAULT_FN_ATTRS _mm512_mask_permutexvar_epi8( + __m512i __W, __mmask64 __M, __m512i __A, __m512i __B) { return (__m512i)__builtin_ia32_selectb_512((__mmask64)__M, (__v64qi)_mm512_permutexvar_epi8(__A, __B), (__v64qi)__W); @@ -97,6 +93,6 @@ _mm512_maskz_multishift_epi64_epi8(__mmask64 __M, __m512i __X, __m512i __Y) (__v64qi)_mm512_multishift_epi64_epi8(__X, __Y), (__v64qi)_mm512_setzero_si512()); } -#undef __DEFAULT_FN_ATTRS_CONSTEXPR + #undef __DEFAULT_FN_ATTRS #endif diff --git a/clang/lib/Headers/avx512vbmivlintrin.h b/clang/lib/Headers/avx512vbmivlintrin.h index 58a48dadff863..40a67bd63ca49 100644 --- a/clang/lib/Headers/avx512vbmivlintrin.h +++ b/clang/lib/Headers/avx512vbmivlintrin.h @@ -15,6 +15,16 @@ #define __VBMIVLINTRIN_H /* Define the default attributes for the functions in this file. */ +#if defined(__cplusplus) && (__cplusplus >= 201103L) +#define __DEFAULT_FN_ATTRS128 \ + __attribute__((__always_inline__, __nodebug__, \ + __target__("avx512vbmi,avx512vl"), \ + __min_vector_width__(128))) constexpr +#define __DEFAULT_FN_ATTRS256 \ + __attribute__((__always_inline__, __nodebug__, \ + __target__("avx512vbmi,avx512vl"), \ + __min_vector_width__(256))) constexpr +#else #define __DEFAULT_FN_ATTRS128 \ __attribute__((__always_inline__, __nodebug__, \ __target__("avx512vbmi,avx512vl"), \ @@ -23,111 +33,96 @@ __attribute__((__always_inline__, __nodebug__, \ __target__("avx512vbmi,avx512vl"), \ __min_vector_width__(256))) - -#if defined(__cplusplus) && (__cplusplus >= 201103L) -#define __DEFAULT_FN_ATTRS128_CONSTEXPR __DEFAULT_FN_ATTRS128 constexpr -#define __DEFAULT_FN_ATTRS256_CONSTEXPR __DEFAULT_FN_ATTRS256 constexpr -#else -#define __DEFAULT_FN_ATTRS128_CONSTEXPR __DEFAULT_FN_ATTRS128 -#define __DEFAULT_FN_ATTRS256_CONSTEXPR __DEFAULT_FN_ATTRS256 #endif -static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR +static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_permutex2var_epi8(__m128i __A, __m128i __I, __m128i __B) { return (__m128i)__builtin_ia32_vpermi2varqi128((__v16qi)__A, (__v16qi)__I, (__v16qi)__B); } -static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR -_mm_mask_permutex2var_epi8(__m128i __A, __mmask16 __U, __m128i __I, - __m128i __B) { +static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_mask_permutex2var_epi8( + __m128i __A, __mmask16 __U, __m128i __I, __m128i __B) { return (__m128i)__builtin_ia32_selectb_128(__U, (__v16qi)_mm_permutex2var_epi8(__A, __I, __B), (__v16qi)__A); } -static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR -_mm_mask2_permutex2var_epi8(__m128i __A, __m128i __I, __mmask16 __U, - __m128i __B) { +static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_mask2_permutex2var_epi8( + __m128i __A, __m128i __I, __mmask16 __U, __m128i __B) { return (__m128i)__builtin_ia32_selectb_128(__U, (__v16qi)_mm_permutex2var_epi8(__A, __I, __B), (__v16qi)__I); } -static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR -_mm_maskz_permutex2var_epi8(__mmask16 __U, __m128i __A, __m128i __I, - __m128i __B) { +static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_maskz_permutex2var_epi8( + __mmask16 __U, __m128i __A, __m128i __I, __m128i __B) { return (__m128i)__builtin_ia32_selectb_128(__U, (__v16qi)_mm_permutex2var_epi8(__A, __I, __B), (__v16qi)_mm_setzero_si128()); } -static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR +static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_permutex2var_epi8(__m256i __A, __m256i __I, __m256i __B) { return (__m256i)__builtin_ia32_vpermi2varqi256((__v32qi)__A, (__v32qi)__I, (__v32qi)__B); } -static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR -_mm256_mask_permutex2var_epi8(__m256i __A, __mmask32 __U, __m256i __I, - __m256i __B) { +static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_mask_permutex2var_epi8( + __m256i __A, __mmask32 __U, __m256i __I, __m256i __B) { return (__m256i)__builtin_ia32_selectb_256(__U, (__v32qi)_mm256_permutex2var_epi8(__A, __I, __B), (__v32qi)__A); } -static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR -_mm256_mask2_permutex2var_epi8(__m256i __A, __m256i __I, __mmask32 __U, - __m256i __B) { +static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_mask2_permutex2var_epi8( + __m256i __A, __m256i __I, __mmask32 __U, __m256i __B) { return (__m256i)__builtin_ia32_selectb_256(__U, (__v32qi)_mm256_permutex2var_epi8(__A, __I, __B), (__v32qi)__I); } -static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR -_mm256_maskz_permutex2var_epi8(__mmask32 __U, __m256i __A, __m256i __I, - __m256i __B) { +static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_maskz_permutex2var_epi8( + __mmask32 __U, __m256i __A, __m256i __I, __m256i __B) { return (__m256i)__builtin_ia32_selectb_256(__U, (__v32qi)_mm256_permutex2var_epi8(__A, __I, __B), (__v32qi)_mm256_setzero_si256()); } -static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR +static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_permutexvar_epi8(__m128i __A, __m128i __B) { return (__m128i)__builtin_ia32_permvarqi128((__v16qi)__B, (__v16qi)__A); } -static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR +static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_maskz_permutexvar_epi8(__mmask16 __M, __m128i __A, __m128i __B) { return (__m128i)__builtin_ia32_selectb_128((__mmask16)__M, (__v16qi)_mm_permutexvar_epi8(__A, __B), (__v16qi)_mm_setzero_si128()); } -static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR -_mm_mask_permutexvar_epi8(__m128i __W, __mmask16 __M, __m128i __A, - __m128i __B) { +static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_mask_permutexvar_epi8( + __m128i __W, __mmask16 __M, __m128i __A, __m128i __B) { return (__m128i)__builtin_ia32_selectb_128((__mmask16)__M, (__v16qi)_mm_permutexvar_epi8(__A, __B), (__v16qi)__W); } -static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR +static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_permutexvar_epi8(__m256i __A, __m256i __B) { return (__m256i)__builtin_ia32_permvarqi256((__v32qi) __B, (__v32qi) __A); } -static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR +static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_maskz_permutexvar_epi8(__mmask32 __M, __m256i __A, __m256i __B) { return (__m256i)__builtin_ia32_selectb_256((__mmask32)__M, (__v32qi)_mm256_permutexvar_epi8(__A, __B), (__v32qi)_mm256_setzero_si256()); } -static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR -_mm256_mask_permutexvar_epi8(__m256i __W, __mmask32 __M, __m256i __A, - __m256i __B) { +static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_mask_permutexvar_epi8( + __m256i __W, __mmask32 __M, __m256i __A, __m256i __B) { return (__m256i)__builtin_ia32_selectb_256((__mmask32)__M, (__v32qi)_mm256_permutexvar_epi8(__A, __B), (__v32qi)__W); @@ -179,9 +174,6 @@ _mm256_maskz_multishift_epi64_epi8(__mmask32 __M, __m256i __X, __m256i __Y) (__v32qi)_mm256_setzero_si256()); } -#undef __DEFAULT_FN_ATTRS128_CONSTEXPR -#undef __DEFAULT_FN_ATTRS256_CONSTEXPR #undef __DEFAULT_FN_ATTRS128 #undef __DEFAULT_FN_ATTRS256 - #endif diff --git a/clang/test/CodeGen/X86/avx512vbmi-builtins.c b/clang/test/CodeGen/X86/avx512vbmi-builtins.c index 7d506db92faeb..fcce58b63737b 100644 --- a/clang/test/CodeGen/X86/avx512vbmi-builtins.c +++ b/clang/test/CodeGen/X86/avx512vbmi-builtins.c @@ -211,18 +211,145 @@ __m512i test_mm512_mask_multishift_epi64_epi8(__m512i __W, __mmask64 __M, __m512 // CHECK-LABEL: test_mm512_mask_multishift_epi64_epi8 // CHECK: call <64 x i8> @llvm.x86.avx512.pmultishift.qb.512(<64 x i8> %{{.*}}, <64 x i8> %{{.*}}) // CHECK: select <64 x i1> %{{.*}}, <64 x i8> %{{.*}}, <64 x i8> %{{.*}} - return _mm512_mask_multishift_epi64_epi8(__W, __M, __X, __Y);  + return _mm512_mask_multishift_epi64_epi8(__W, __M, __X, __Y); } +TEST_CONSTEXPR(match_v64qu( + _mm512_mask_multishift_epi64_epi8( + (__m512i)(__v64qu){ + 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, + 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, + 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, + 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, + 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, + 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, + 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, + 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0... [truncated]

chaitanyav · 2025-11-25T05:33:58Z

@RKSimon attaching the fuzz test results file
vpmultishiftqb_comprehensive_report.md

RKSimon · 2025-11-26T14:59:23Z

clang/lib/AST/ByteCode/InterpBuiltin.cpp

+ const Pointer &APtr = S.Stk.pop<Pointer>();
+ const auto *AVecT = ATy->castAs<VectorType>();
+ const auto *BVecT = BTy->castAs<VectorType>();
+ assert(AVecT->getNumElements() == BVecT->getNumElements());


Stop unused variable warnings on non-assert builds:

Suggested change

assert(AVecT->getNumElements() == BVecT->getNumElements());

assert(AVecT->getNumElements() == BTy->castAs<VectorType>()->getNumElements());

understood, i will update the assertion.

📦 Installing… ninja: Entering directory `/home/naga/llvm-project/build' [1205/4534] Building CXX object tools/clang/lib/AST/CMakeFiles/obj.clangAST.dir/ByteCode/InterpBuiltin.cpp.o /home/naga/llvm-project/clang/lib/AST/ByteCode/InterpBuiltin.cpp:3547:15: warning: unused variable 'BVecT' [-Wunused-variable] 3547 | const auto *BVecT = BTy->castAs<VectorType>(); | ^~~~~

RKSimon · 2025-11-26T17:15:30Z

clang/lib/AST/ByteCode/InterpBuiltin.cpp

+
+ APInt Byte(8, 0);
+ for (unsigned BitIdx = 0; BitIdx != NumBitsInByte; ++BitIdx) {
+ Byte.insertBits(BQWord.extractBits(1, (Ctrl + BitIdx) & 0x3F), BitIdx);


Should this be:

Byte.setBitVal(BitIdx, BQWord[(Ctrl + BitIdx) & 0x3F]);

RKSimon · 2025-11-26T17:18:11Z

clang/lib/AST/ByteCode/InterpBuiltin.cpp

+
+ for (unsigned ByteIdx = 0; ByteIdx != NumBytesInQWord; ++ByteIdx) {
+ uint64_t Ctrl =
+ AQWord.extractBits(8, ByteIdx * NumBitsInByte).getZExtValue() & 0x3F;


Do we gain anything from creating/extraction from AQWord instead of just doing this directly?

uint64_t Ctrl = static_cast<uint64_t>(APtr.elem<T>(QWordId * NumBytesInQWord + ByteIdx)) & 0x3F;

That's closer to what we do for interp__builtin_ia32_shufbitqmb_mask

made the changes to both the files

chaitanyav · 2025-11-26T18:26:18Z

verified that everything is passing in the fuzz testing also.

RKSimon · 2025-11-26T19:12:05Z

clang/lib/AST/ByteCode/InterpBuiltin.cpp

+ Byte.setBitVal(BitIdx, BQWord[(Ctrl + BitIdx) & 0x3F]);
+ }
+ INT_TYPE_SWITCH(ElemT, {
+ Dst.elem<T>(QWordId * NumBytesInQWord + ByteIdx) =


hoist this unsigned Idx = QWordId * NumBytesInQWord + ByteIdx; - we now repeat it.

RKSimon · 2025-11-26T19:13:40Z

clang/lib/AST/ExprConstant.cpp

+ }
+
+ for (unsigned ByteIdx = 0; ByteIdx != NumBytesInQWord; ++ByteIdx) {
+ uint64_t Ctrl = A.getVectorElt(QWordId * NumBytesInQWord + ByteIdx)


Hoist unsigned Idx = QWordId * NumBytesInQWord + ByteIdx; - it should help with the clang-format line wrapping

… AVX512 VPMULTISHIFTQB intrinsics to be used in constexpr Resolves:llvm#167477

* Unnecessary AQWord extraction when we can directly extract ctrl from A. * Use setBitVal instead of insertBits(extractBits)

* Use tmp variable to hold the inedx of the byte to avoid duplication and help with clang-format

RKSimon

LGTM - cheers

chaitanyav force-pushed the issue_167477 branch from 2808e74 to 4a5b92c Compare November 21, 2025 04:21

chaitanyav marked this pull request as ready for review November 21, 2025 05:11

llvmbot added clang Clang issues not falling into any other category backend:X86 clang:frontend Language frontend issues, e.g. anything involving "Sema" clang:headers Headers provided by Clang, e.g. for intrinsics clang:bytecode Issues for the clang bytecode constexpr interpreter labels Nov 21, 2025

chaitanyav requested a review from RKSimon November 21, 2025 05:12

chaitanyav self-assigned this Nov 21, 2025

chaitanyav requested a review from tbaederr November 22, 2025 04:39

chaitanyav force-pushed the issue_167477 branch 2 times, most recently from 4831cf9 to 3e26e39 Compare November 24, 2025 04:20

RKSimon requested changes Nov 26, 2025

View reviewed changes

chaitanyav force-pushed the issue_167477 branch from 3e26e39 to 439aa6a Compare November 26, 2025 16:37

RKSimon reviewed Nov 26, 2025

View reviewed changes

chaitanyav force-pushed the issue_167477 branch from 439aa6a to d5d96ec Compare November 26, 2025 18:07

RKSimon reviewed Nov 26, 2025

View reviewed changes

chaitanyav added 5 commits November 26, 2025 12:10

[Clang] VectorExprEvaluator::VisitCallExpr / InterpretBuiltin - Allow…

8ebcf0a

… AVX512 VPMULTISHIFTQB intrinsics to be used in constexpr Resolves:llvm#167477

Put the TEST_CONSTEXPR tests after their corresponding CHECK tests

ba2575e

Fix unused variable warning on non-assert builds

26ee054

Address feedback from code review

b646720

* Unnecessary AQWord extraction when we can directly extract ctrl from A. * Use setBitVal instead of insertBits(extractBits)

Address feedback from code review

2c67a38

* Use tmp variable to hold the inedx of the byte to avoid duplication and help with clang-format

chaitanyav force-pushed the issue_167477 branch from d5d96ec to 2c67a38 Compare November 26, 2025 20:10

RKSimon self-requested a review November 27, 2025 08:49

RKSimon approved these changes Nov 27, 2025

View reviewed changes

Merge branch 'main' into issue_167477

09a65cf

RKSimon enabled auto-merge (squash) November 27, 2025 10:01

RKSimon merged commit c3c3d16 into llvm:main Nov 27, 2025
9 of 10 checks passed

chaitanyav deleted the issue_167477 branch November 30, 2025 20:59

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

[Clang] VectorExprEvaluator::VisitCallExpr / InterpretBuiltin - Allow AVX512 VPMULTISHIFTQB intrinsics to be used in constexpr #168995

[Clang] VectorExprEvaluator::VisitCallExpr / InterpretBuiltin - Allow AVX512 VPMULTISHIFTQB intrinsics to be used in constexpr #168995

Uh oh!

chaitanyav commented Nov 21, 2025

github-actions bot commented Nov 21, 2025 •

edited

Loading

llvmbot commented Nov 21, 2025

llvmbot commented Nov 21, 2025

chaitanyav commented Nov 25, 2025

RKSimon Nov 26, 2025

chaitanyav Nov 26, 2025

RKSimon Nov 26, 2025

RKSimon Nov 26, 2025

chaitanyav Nov 26, 2025

chaitanyav commented Nov 26, 2025

RKSimon Nov 26, 2025

RKSimon Nov 26, 2025

RKSimon left a comment

Uh oh!

Labels

3 participants

	assert(AVecT->getNumElements() == BVecT->getNumElements());
	assert(AVecT->getNumElements() == BTy->castAs<VectorType>()->getNumElements());

[Clang] VectorExprEvaluator::VisitCallExpr / InterpretBuiltin - Allow AVX512 VPMULTISHIFTQB intrinsics to be used in constexpr #168995

[Clang] VectorExprEvaluator::VisitCallExpr / InterpretBuiltin - Allow AVX512 VPMULTISHIFTQB intrinsics to be used in constexpr #168995

Uh oh!

Conversation

chaitanyav commented Nov 21, 2025

github-actions bot commented Nov 21, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

🐧 Linux x64 Test Results

llvmbot commented Nov 21, 2025

llvmbot commented Nov 21, 2025

chaitanyav commented Nov 25, 2025

RKSimon Nov 26, 2025

Choose a reason for hiding this comment

chaitanyav Nov 26, 2025

Choose a reason for hiding this comment

RKSimon Nov 26, 2025

Choose a reason for hiding this comment

RKSimon Nov 26, 2025

Choose a reason for hiding this comment

chaitanyav Nov 26, 2025

Choose a reason for hiding this comment

chaitanyav commented Nov 26, 2025

RKSimon Nov 26, 2025

Choose a reason for hiding this comment

RKSimon Nov 26, 2025

Choose a reason for hiding this comment

RKSimon left a comment

Choose a reason for hiding this comment

Uh oh!

Labels

3 participants

github-actions bot commented Nov 21, 2025 •

edited

Loading