| File: | build/source/clang/lib/CodeGen/CGBuiltin.cpp |
| Warning: | line 1059, column 22 Value stored to 'RetType' during its initialization is never read |
Press '?' to see keyboard shortcuts
Keyboard shortcuts:
| 1 | //===---- CGBuiltin.cpp - Emit LLVM Code for builtins ---------------------===// |
| 2 | // |
| 3 | // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. |
| 4 | // See https://llvm.org/LICENSE.txt for license information. |
| 5 | // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
| 6 | // |
| 7 | //===----------------------------------------------------------------------===// |
| 8 | // |
| 9 | // This contains code to emit Builtin calls as LLVM code. |
| 10 | // |
| 11 | //===----------------------------------------------------------------------===// |
| 12 | |
| 13 | #include "ABIInfo.h" |
| 14 | #include "CGCUDARuntime.h" |
| 15 | #include "CGCXXABI.h" |
| 16 | #include "CGObjCRuntime.h" |
| 17 | #include "CGOpenCLRuntime.h" |
| 18 | #include "CGRecordLayout.h" |
| 19 | #include "CodeGenFunction.h" |
| 20 | #include "CodeGenModule.h" |
| 21 | #include "ConstantEmitter.h" |
| 22 | #include "PatternInit.h" |
| 23 | #include "TargetInfo.h" |
| 24 | #include "clang/AST/ASTContext.h" |
| 25 | #include "clang/AST/Attr.h" |
| 26 | #include "clang/AST/Decl.h" |
| 27 | #include "clang/AST/OSLog.h" |
| 28 | #include "clang/Basic/TargetBuiltins.h" |
| 29 | #include "clang/Basic/TargetInfo.h" |
| 30 | #include "clang/CodeGen/CGFunctionInfo.h" |
| 31 | #include "clang/Frontend/FrontendDiagnostic.h" |
| 32 | #include "llvm/ADT/APFloat.h" |
| 33 | #include "llvm/ADT/APInt.h" |
| 34 | #include "llvm/ADT/SmallPtrSet.h" |
| 35 | #include "llvm/ADT/StringExtras.h" |
| 36 | #include "llvm/Analysis/ValueTracking.h" |
| 37 | #include "llvm/IR/DataLayout.h" |
| 38 | #include "llvm/IR/InlineAsm.h" |
| 39 | #include "llvm/IR/Intrinsics.h" |
| 40 | #include "llvm/IR/IntrinsicsAArch64.h" |
| 41 | #include "llvm/IR/IntrinsicsAMDGPU.h" |
| 42 | #include "llvm/IR/IntrinsicsARM.h" |
| 43 | #include "llvm/IR/IntrinsicsBPF.h" |
| 44 | #include "llvm/IR/IntrinsicsHexagon.h" |
| 45 | #include "llvm/IR/IntrinsicsLoongArch.h" |
| 46 | #include "llvm/IR/IntrinsicsNVPTX.h" |
| 47 | #include "llvm/IR/IntrinsicsPowerPC.h" |
| 48 | #include "llvm/IR/IntrinsicsR600.h" |
| 49 | #include "llvm/IR/IntrinsicsRISCV.h" |
| 50 | #include "llvm/IR/IntrinsicsS390.h" |
| 51 | #include "llvm/IR/IntrinsicsVE.h" |
| 52 | #include "llvm/IR/IntrinsicsWebAssembly.h" |
| 53 | #include "llvm/IR/IntrinsicsX86.h" |
| 54 | #include "llvm/IR/MDBuilder.h" |
| 55 | #include "llvm/IR/MatrixBuilder.h" |
| 56 | #include "llvm/Support/ConvertUTF.h" |
| 57 | #include "llvm/Support/ScopedPrinter.h" |
| 58 | #include "llvm/TargetParser/AArch64TargetParser.h" |
| 59 | #include "llvm/TargetParser/X86TargetParser.h" |
| 60 | #include <optional> |
| 61 | #include <sstream> |
| 62 | |
| 63 | using namespace clang; |
| 64 | using namespace CodeGen; |
| 65 | using namespace llvm; |
| 66 | |
| 67 | static void initializeAlloca(CodeGenFunction &CGF, AllocaInst *AI, Value *Size, |
| 68 | Align AlignmentInBytes) { |
| 69 | ConstantInt *Byte; |
| 70 | switch (CGF.getLangOpts().getTrivialAutoVarInit()) { |
| 71 | case LangOptions::TrivialAutoVarInitKind::Uninitialized: |
| 72 | // Nothing to initialize. |
| 73 | return; |
| 74 | case LangOptions::TrivialAutoVarInitKind::Zero: |
| 75 | Byte = CGF.Builder.getInt8(0x00); |
| 76 | break; |
| 77 | case LangOptions::TrivialAutoVarInitKind::Pattern: { |
| 78 | llvm::Type *Int8 = llvm::IntegerType::getInt8Ty(CGF.CGM.getLLVMContext()); |
| 79 | Byte = llvm::dyn_cast<llvm::ConstantInt>( |
| 80 | initializationPatternFor(CGF.CGM, Int8)); |
| 81 | break; |
| 82 | } |
| 83 | } |
| 84 | if (CGF.CGM.stopAutoInit()) |
| 85 | return; |
| 86 | auto *I = CGF.Builder.CreateMemSet(AI, Byte, Size, AlignmentInBytes); |
| 87 | I->addAnnotationMetadata("auto-init"); |
| 88 | } |
| 89 | |
| 90 | /// getBuiltinLibFunction - Given a builtin id for a function like |
| 91 | /// "__builtin_fabsf", return a Function* for "fabsf". |
| 92 | llvm::Constant *CodeGenModule::getBuiltinLibFunction(const FunctionDecl *FD, |
| 93 | unsigned BuiltinID) { |
| 94 | assert(Context.BuiltinInfo.isLibFunction(BuiltinID))(static_cast <bool> (Context.BuiltinInfo.isLibFunction( BuiltinID)) ? void (0) : __assert_fail ("Context.BuiltinInfo.isLibFunction(BuiltinID)" , "clang/lib/CodeGen/CGBuiltin.cpp", 94, __extension__ __PRETTY_FUNCTION__ )); |
| 95 | |
| 96 | // Get the name, skip over the __builtin_ prefix (if necessary). |
| 97 | StringRef Name; |
| 98 | GlobalDecl D(FD); |
| 99 | |
| 100 | // TODO: This list should be expanded or refactored after all GCC-compatible |
| 101 | // std libcall builtins are implemented. |
| 102 | static SmallDenseMap<unsigned, StringRef, 8> F128Builtins{ |
| 103 | {Builtin::BI__builtin_printf, "__printfieee128"}, |
| 104 | {Builtin::BI__builtin_vsnprintf, "__vsnprintfieee128"}, |
| 105 | {Builtin::BI__builtin_vsprintf, "__vsprintfieee128"}, |
| 106 | {Builtin::BI__builtin_sprintf, "__sprintfieee128"}, |
| 107 | {Builtin::BI__builtin_snprintf, "__snprintfieee128"}, |
| 108 | {Builtin::BI__builtin_fprintf, "__fprintfieee128"}, |
| 109 | {Builtin::BI__builtin_nexttowardf128, "__nexttowardieee128"}, |
| 110 | }; |
| 111 | |
| 112 | // The AIX library functions frexpl, ldexpl, and modfl are for 128-bit |
| 113 | // IBM 'long double' (i.e. __ibm128). Map to the 'double' versions |
| 114 | // if it is 64-bit 'long double' mode. |
| 115 | static SmallDenseMap<unsigned, StringRef, 4> AIXLongDouble64Builtins{ |
| 116 | {Builtin::BI__builtin_frexpl, "frexp"}, |
| 117 | {Builtin::BI__builtin_ldexpl, "ldexp"}, |
| 118 | {Builtin::BI__builtin_modfl, "modf"}, |
| 119 | }; |
| 120 | |
| 121 | // If the builtin has been declared explicitly with an assembler label, |
| 122 | // use the mangled name. This differs from the plain label on platforms |
| 123 | // that prefix labels. |
| 124 | if (FD->hasAttr<AsmLabelAttr>()) |
| 125 | Name = getMangledName(D); |
| 126 | else { |
| 127 | // TODO: This mutation should also be applied to other targets other than |
| 128 | // PPC, after backend supports IEEE 128-bit style libcalls. |
| 129 | if (getTriple().isPPC64() && |
| 130 | &getTarget().getLongDoubleFormat() == &llvm::APFloat::IEEEquad() && |
| 131 | F128Builtins.find(BuiltinID) != F128Builtins.end()) |
| 132 | Name = F128Builtins[BuiltinID]; |
| 133 | else if (getTriple().isOSAIX() && |
| 134 | &getTarget().getLongDoubleFormat() == |
| 135 | &llvm::APFloat::IEEEdouble() && |
| 136 | AIXLongDouble64Builtins.find(BuiltinID) != |
| 137 | AIXLongDouble64Builtins.end()) |
| 138 | Name = AIXLongDouble64Builtins[BuiltinID]; |
| 139 | else |
| 140 | Name = Context.BuiltinInfo.getName(BuiltinID).substr(10); |
| 141 | } |
| 142 | |
| 143 | llvm::FunctionType *Ty = |
| 144 | cast<llvm::FunctionType>(getTypes().ConvertType(FD->getType())); |
| 145 | |
| 146 | return GetOrCreateLLVMFunction(Name, Ty, D, /*ForVTable=*/false); |
| 147 | } |
| 148 | |
| 149 | /// Emit the conversions required to turn the given value into an |
| 150 | /// integer of the given size. |
| 151 | static Value *EmitToInt(CodeGenFunction &CGF, llvm::Value *V, |
| 152 | QualType T, llvm::IntegerType *IntType) { |
| 153 | V = CGF.EmitToMemory(V, T); |
| 154 | |
| 155 | if (V->getType()->isPointerTy()) |
| 156 | return CGF.Builder.CreatePtrToInt(V, IntType); |
| 157 | |
| 158 | assert(V->getType() == IntType)(static_cast <bool> (V->getType() == IntType) ? void (0) : __assert_fail ("V->getType() == IntType", "clang/lib/CodeGen/CGBuiltin.cpp" , 158, __extension__ __PRETTY_FUNCTION__)); |
| 159 | return V; |
| 160 | } |
| 161 | |
| 162 | static Value *EmitFromInt(CodeGenFunction &CGF, llvm::Value *V, |
| 163 | QualType T, llvm::Type *ResultType) { |
| 164 | V = CGF.EmitFromMemory(V, T); |
| 165 | |
| 166 | if (ResultType->isPointerTy()) |
| 167 | return CGF.Builder.CreateIntToPtr(V, ResultType); |
| 168 | |
| 169 | assert(V->getType() == ResultType)(static_cast <bool> (V->getType() == ResultType) ? void (0) : __assert_fail ("V->getType() == ResultType", "clang/lib/CodeGen/CGBuiltin.cpp" , 169, __extension__ __PRETTY_FUNCTION__)); |
| 170 | return V; |
| 171 | } |
| 172 | |
| 173 | static llvm::Value *CheckAtomicAlignment(CodeGenFunction &CGF, |
| 174 | const CallExpr *E) { |
| 175 | ASTContext &Ctx = CGF.getContext(); |
| 176 | Address Ptr = CGF.EmitPointerWithAlignment(E->getArg(0)); |
| 177 | unsigned Bytes = Ptr.getElementType()->isPointerTy() |
| 178 | ? Ctx.getTypeSizeInChars(Ctx.VoidPtrTy).getQuantity() |
| 179 | : Ptr.getElementType()->getScalarSizeInBits() / 8; |
| 180 | unsigned Align = Ptr.getAlignment().getQuantity(); |
| 181 | if (Align % Bytes != 0) { |
| 182 | DiagnosticsEngine &Diags = CGF.CGM.getDiags(); |
| 183 | Diags.Report(E->getBeginLoc(), diag::warn_sync_op_misaligned); |
| 184 | } |
| 185 | return Ptr.getPointer(); |
| 186 | } |
| 187 | |
| 188 | /// Utility to insert an atomic instruction based on Intrinsic::ID |
| 189 | /// and the expression node. |
| 190 | static Value *MakeBinaryAtomicValue( |
| 191 | CodeGenFunction &CGF, llvm::AtomicRMWInst::BinOp Kind, const CallExpr *E, |
| 192 | AtomicOrdering Ordering = AtomicOrdering::SequentiallyConsistent) { |
| 193 | |
| 194 | QualType T = E->getType(); |
| 195 | assert(E->getArg(0)->getType()->isPointerType())(static_cast <bool> (E->getArg(0)->getType()-> isPointerType()) ? void (0) : __assert_fail ("E->getArg(0)->getType()->isPointerType()" , "clang/lib/CodeGen/CGBuiltin.cpp", 195, __extension__ __PRETTY_FUNCTION__ )); |
| 196 | assert(CGF.getContext().hasSameUnqualifiedType(T,(static_cast <bool> (CGF.getContext().hasSameUnqualifiedType (T, E->getArg(0)->getType()->getPointeeType())) ? void (0) : __assert_fail ("CGF.getContext().hasSameUnqualifiedType(T, E->getArg(0)->getType()->getPointeeType())" , "clang/lib/CodeGen/CGBuiltin.cpp", 197, __extension__ __PRETTY_FUNCTION__ )) |
| 197 | E->getArg(0)->getType()->getPointeeType()))(static_cast <bool> (CGF.getContext().hasSameUnqualifiedType (T, E->getArg(0)->getType()->getPointeeType())) ? void (0) : __assert_fail ("CGF.getContext().hasSameUnqualifiedType(T, E->getArg(0)->getType()->getPointeeType())" , "clang/lib/CodeGen/CGBuiltin.cpp", 197, __extension__ __PRETTY_FUNCTION__ )); |
| 198 | assert(CGF.getContext().hasSameUnqualifiedType(T, E->getArg(1)->getType()))(static_cast <bool> (CGF.getContext().hasSameUnqualifiedType (T, E->getArg(1)->getType())) ? void (0) : __assert_fail ("CGF.getContext().hasSameUnqualifiedType(T, E->getArg(1)->getType())" , "clang/lib/CodeGen/CGBuiltin.cpp", 198, __extension__ __PRETTY_FUNCTION__ )); |
| 199 | |
| 200 | llvm::Value *DestPtr = CheckAtomicAlignment(CGF, E); |
| 201 | unsigned AddrSpace = DestPtr->getType()->getPointerAddressSpace(); |
| 202 | |
| 203 | llvm::IntegerType *IntType = |
| 204 | llvm::IntegerType::get(CGF.getLLVMContext(), |
| 205 | CGF.getContext().getTypeSize(T)); |
| 206 | llvm::Type *IntPtrType = IntType->getPointerTo(AddrSpace); |
| 207 | |
| 208 | llvm::Value *Args[2]; |
| 209 | Args[0] = CGF.Builder.CreateBitCast(DestPtr, IntPtrType); |
| 210 | Args[1] = CGF.EmitScalarExpr(E->getArg(1)); |
| 211 | llvm::Type *ValueType = Args[1]->getType(); |
| 212 | Args[1] = EmitToInt(CGF, Args[1], T, IntType); |
| 213 | |
| 214 | llvm::Value *Result = CGF.Builder.CreateAtomicRMW( |
| 215 | Kind, Args[0], Args[1], Ordering); |
| 216 | return EmitFromInt(CGF, Result, T, ValueType); |
| 217 | } |
| 218 | |
| 219 | static Value *EmitNontemporalStore(CodeGenFunction &CGF, const CallExpr *E) { |
| 220 | Value *Val = CGF.EmitScalarExpr(E->getArg(0)); |
| 221 | Value *Address = CGF.EmitScalarExpr(E->getArg(1)); |
| 222 | |
| 223 | // Convert the type of the pointer to a pointer to the stored type. |
| 224 | Val = CGF.EmitToMemory(Val, E->getArg(0)->getType()); |
| 225 | unsigned SrcAddrSpace = Address->getType()->getPointerAddressSpace(); |
| 226 | Value *BC = CGF.Builder.CreateBitCast( |
| 227 | Address, llvm::PointerType::get(Val->getType(), SrcAddrSpace), "cast"); |
| 228 | LValue LV = CGF.MakeNaturalAlignAddrLValue(BC, E->getArg(0)->getType()); |
| 229 | LV.setNontemporal(true); |
| 230 | CGF.EmitStoreOfScalar(Val, LV, false); |
| 231 | return nullptr; |
| 232 | } |
| 233 | |
| 234 | static Value *EmitNontemporalLoad(CodeGenFunction &CGF, const CallExpr *E) { |
| 235 | Value *Address = CGF.EmitScalarExpr(E->getArg(0)); |
| 236 | |
| 237 | LValue LV = CGF.MakeNaturalAlignAddrLValue(Address, E->getType()); |
| 238 | LV.setNontemporal(true); |
| 239 | return CGF.EmitLoadOfScalar(LV, E->getExprLoc()); |
| 240 | } |
| 241 | |
| 242 | static RValue EmitBinaryAtomic(CodeGenFunction &CGF, |
| 243 | llvm::AtomicRMWInst::BinOp Kind, |
| 244 | const CallExpr *E) { |
| 245 | return RValue::get(MakeBinaryAtomicValue(CGF, Kind, E)); |
| 246 | } |
| 247 | |
| 248 | /// Utility to insert an atomic instruction based Intrinsic::ID and |
| 249 | /// the expression node, where the return value is the result of the |
| 250 | /// operation. |
| 251 | static RValue EmitBinaryAtomicPost(CodeGenFunction &CGF, |
| 252 | llvm::AtomicRMWInst::BinOp Kind, |
| 253 | const CallExpr *E, |
| 254 | Instruction::BinaryOps Op, |
| 255 | bool Invert = false) { |
| 256 | QualType T = E->getType(); |
| 257 | assert(E->getArg(0)->getType()->isPointerType())(static_cast <bool> (E->getArg(0)->getType()-> isPointerType()) ? void (0) : __assert_fail ("E->getArg(0)->getType()->isPointerType()" , "clang/lib/CodeGen/CGBuiltin.cpp", 257, __extension__ __PRETTY_FUNCTION__ )); |
| 258 | assert(CGF.getContext().hasSameUnqualifiedType(T,(static_cast <bool> (CGF.getContext().hasSameUnqualifiedType (T, E->getArg(0)->getType()->getPointeeType())) ? void (0) : __assert_fail ("CGF.getContext().hasSameUnqualifiedType(T, E->getArg(0)->getType()->getPointeeType())" , "clang/lib/CodeGen/CGBuiltin.cpp", 259, __extension__ __PRETTY_FUNCTION__ )) |
| 259 | E->getArg(0)->getType()->getPointeeType()))(static_cast <bool> (CGF.getContext().hasSameUnqualifiedType (T, E->getArg(0)->getType()->getPointeeType())) ? void (0) : __assert_fail ("CGF.getContext().hasSameUnqualifiedType(T, E->getArg(0)->getType()->getPointeeType())" , "clang/lib/CodeGen/CGBuiltin.cpp", 259, __extension__ __PRETTY_FUNCTION__ )); |
| 260 | assert(CGF.getContext().hasSameUnqualifiedType(T, E->getArg(1)->getType()))(static_cast <bool> (CGF.getContext().hasSameUnqualifiedType (T, E->getArg(1)->getType())) ? void (0) : __assert_fail ("CGF.getContext().hasSameUnqualifiedType(T, E->getArg(1)->getType())" , "clang/lib/CodeGen/CGBuiltin.cpp", 260, __extension__ __PRETTY_FUNCTION__ )); |
| 261 | |
| 262 | llvm::Value *DestPtr = CheckAtomicAlignment(CGF, E); |
| 263 | unsigned AddrSpace = DestPtr->getType()->getPointerAddressSpace(); |
| 264 | |
| 265 | llvm::IntegerType *IntType = |
| 266 | llvm::IntegerType::get(CGF.getLLVMContext(), |
| 267 | CGF.getContext().getTypeSize(T)); |
| 268 | llvm::Type *IntPtrType = IntType->getPointerTo(AddrSpace); |
| 269 | |
| 270 | llvm::Value *Args[2]; |
| 271 | Args[1] = CGF.EmitScalarExpr(E->getArg(1)); |
| 272 | llvm::Type *ValueType = Args[1]->getType(); |
| 273 | Args[1] = EmitToInt(CGF, Args[1], T, IntType); |
| 274 | Args[0] = CGF.Builder.CreateBitCast(DestPtr, IntPtrType); |
| 275 | |
| 276 | llvm::Value *Result = CGF.Builder.CreateAtomicRMW( |
| 277 | Kind, Args[0], Args[1], llvm::AtomicOrdering::SequentiallyConsistent); |
| 278 | Result = CGF.Builder.CreateBinOp(Op, Result, Args[1]); |
| 279 | if (Invert) |
| 280 | Result = |
| 281 | CGF.Builder.CreateBinOp(llvm::Instruction::Xor, Result, |
| 282 | llvm::ConstantInt::getAllOnesValue(IntType)); |
| 283 | Result = EmitFromInt(CGF, Result, T, ValueType); |
| 284 | return RValue::get(Result); |
| 285 | } |
| 286 | |
| 287 | /// Utility to insert an atomic cmpxchg instruction. |
| 288 | /// |
| 289 | /// @param CGF The current codegen function. |
| 290 | /// @param E Builtin call expression to convert to cmpxchg. |
| 291 | /// arg0 - address to operate on |
| 292 | /// arg1 - value to compare with |
| 293 | /// arg2 - new value |
| 294 | /// @param ReturnBool Specifies whether to return success flag of |
| 295 | /// cmpxchg result or the old value. |
| 296 | /// |
| 297 | /// @returns result of cmpxchg, according to ReturnBool |
| 298 | /// |
| 299 | /// Note: In order to lower Microsoft's _InterlockedCompareExchange* intrinsics |
| 300 | /// invoke the function EmitAtomicCmpXchgForMSIntrin. |
| 301 | static Value *MakeAtomicCmpXchgValue(CodeGenFunction &CGF, const CallExpr *E, |
| 302 | bool ReturnBool) { |
| 303 | QualType T = ReturnBool ? E->getArg(1)->getType() : E->getType(); |
| 304 | llvm::Value *DestPtr = CheckAtomicAlignment(CGF, E); |
| 305 | unsigned AddrSpace = DestPtr->getType()->getPointerAddressSpace(); |
| 306 | |
| 307 | llvm::IntegerType *IntType = llvm::IntegerType::get( |
| 308 | CGF.getLLVMContext(), CGF.getContext().getTypeSize(T)); |
| 309 | llvm::Type *IntPtrType = IntType->getPointerTo(AddrSpace); |
| 310 | |
| 311 | Value *Args[3]; |
| 312 | Args[0] = CGF.Builder.CreateBitCast(DestPtr, IntPtrType); |
| 313 | Args[1] = CGF.EmitScalarExpr(E->getArg(1)); |
| 314 | llvm::Type *ValueType = Args[1]->getType(); |
| 315 | Args[1] = EmitToInt(CGF, Args[1], T, IntType); |
| 316 | Args[2] = EmitToInt(CGF, CGF.EmitScalarExpr(E->getArg(2)), T, IntType); |
| 317 | |
| 318 | Value *Pair = CGF.Builder.CreateAtomicCmpXchg( |
| 319 | Args[0], Args[1], Args[2], llvm::AtomicOrdering::SequentiallyConsistent, |
| 320 | llvm::AtomicOrdering::SequentiallyConsistent); |
| 321 | if (ReturnBool) |
| 322 | // Extract boolean success flag and zext it to int. |
| 323 | return CGF.Builder.CreateZExt(CGF.Builder.CreateExtractValue(Pair, 1), |
| 324 | CGF.ConvertType(E->getType())); |
| 325 | else |
| 326 | // Extract old value and emit it using the same type as compare value. |
| 327 | return EmitFromInt(CGF, CGF.Builder.CreateExtractValue(Pair, 0), T, |
| 328 | ValueType); |
| 329 | } |
| 330 | |
| 331 | /// This function should be invoked to emit atomic cmpxchg for Microsoft's |
| 332 | /// _InterlockedCompareExchange* intrinsics which have the following signature: |
| 333 | /// T _InterlockedCompareExchange(T volatile *Destination, |
| 334 | /// T Exchange, |
| 335 | /// T Comparand); |
| 336 | /// |
| 337 | /// Whereas the llvm 'cmpxchg' instruction has the following syntax: |
| 338 | /// cmpxchg *Destination, Comparand, Exchange. |
| 339 | /// So we need to swap Comparand and Exchange when invoking |
| 340 | /// CreateAtomicCmpXchg. That is the reason we could not use the above utility |
| 341 | /// function MakeAtomicCmpXchgValue since it expects the arguments to be |
| 342 | /// already swapped. |
| 343 | |
| 344 | static |
| 345 | Value *EmitAtomicCmpXchgForMSIntrin(CodeGenFunction &CGF, const CallExpr *E, |
| 346 | AtomicOrdering SuccessOrdering = AtomicOrdering::SequentiallyConsistent) { |
| 347 | assert(E->getArg(0)->getType()->isPointerType())(static_cast <bool> (E->getArg(0)->getType()-> isPointerType()) ? void (0) : __assert_fail ("E->getArg(0)->getType()->isPointerType()" , "clang/lib/CodeGen/CGBuiltin.cpp", 347, __extension__ __PRETTY_FUNCTION__ )); |
| 348 | assert(CGF.getContext().hasSameUnqualifiedType((static_cast <bool> (CGF.getContext().hasSameUnqualifiedType ( E->getType(), E->getArg(0)->getType()->getPointeeType ())) ? void (0) : __assert_fail ("CGF.getContext().hasSameUnqualifiedType( E->getType(), E->getArg(0)->getType()->getPointeeType())" , "clang/lib/CodeGen/CGBuiltin.cpp", 349, __extension__ __PRETTY_FUNCTION__ )) |
| 349 | E->getType(), E->getArg(0)->getType()->getPointeeType()))(static_cast <bool> (CGF.getContext().hasSameUnqualifiedType ( E->getType(), E->getArg(0)->getType()->getPointeeType ())) ? void (0) : __assert_fail ("CGF.getContext().hasSameUnqualifiedType( E->getType(), E->getArg(0)->getType()->getPointeeType())" , "clang/lib/CodeGen/CGBuiltin.cpp", 349, __extension__ __PRETTY_FUNCTION__ )); |
| 350 | assert(CGF.getContext().hasSameUnqualifiedType(E->getType(),(static_cast <bool> (CGF.getContext().hasSameUnqualifiedType (E->getType(), E->getArg(1)->getType())) ? void (0) : __assert_fail ("CGF.getContext().hasSameUnqualifiedType(E->getType(), E->getArg(1)->getType())" , "clang/lib/CodeGen/CGBuiltin.cpp", 351, __extension__ __PRETTY_FUNCTION__ )) |
| 351 | E->getArg(1)->getType()))(static_cast <bool> (CGF.getContext().hasSameUnqualifiedType (E->getType(), E->getArg(1)->getType())) ? void (0) : __assert_fail ("CGF.getContext().hasSameUnqualifiedType(E->getType(), E->getArg(1)->getType())" , "clang/lib/CodeGen/CGBuiltin.cpp", 351, __extension__ __PRETTY_FUNCTION__ )); |
| 352 | assert(CGF.getContext().hasSameUnqualifiedType(E->getType(),(static_cast <bool> (CGF.getContext().hasSameUnqualifiedType (E->getType(), E->getArg(2)->getType())) ? void (0) : __assert_fail ("CGF.getContext().hasSameUnqualifiedType(E->getType(), E->getArg(2)->getType())" , "clang/lib/CodeGen/CGBuiltin.cpp", 353, __extension__ __PRETTY_FUNCTION__ )) |
| 353 | E->getArg(2)->getType()))(static_cast <bool> (CGF.getContext().hasSameUnqualifiedType (E->getType(), E->getArg(2)->getType())) ? void (0) : __assert_fail ("CGF.getContext().hasSameUnqualifiedType(E->getType(), E->getArg(2)->getType())" , "clang/lib/CodeGen/CGBuiltin.cpp", 353, __extension__ __PRETTY_FUNCTION__ )); |
| 354 | |
| 355 | auto *Destination = CGF.EmitScalarExpr(E->getArg(0)); |
| 356 | auto *Comparand = CGF.EmitScalarExpr(E->getArg(2)); |
| 357 | auto *Exchange = CGF.EmitScalarExpr(E->getArg(1)); |
| 358 | |
| 359 | // For Release ordering, the failure ordering should be Monotonic. |
| 360 | auto FailureOrdering = SuccessOrdering == AtomicOrdering::Release ? |
| 361 | AtomicOrdering::Monotonic : |
| 362 | SuccessOrdering; |
| 363 | |
| 364 | // The atomic instruction is marked volatile for consistency with MSVC. This |
| 365 | // blocks the few atomics optimizations that LLVM has. If we want to optimize |
| 366 | // _Interlocked* operations in the future, we will have to remove the volatile |
| 367 | // marker. |
| 368 | auto *Result = CGF.Builder.CreateAtomicCmpXchg( |
| 369 | Destination, Comparand, Exchange, |
| 370 | SuccessOrdering, FailureOrdering); |
| 371 | Result->setVolatile(true); |
| 372 | return CGF.Builder.CreateExtractValue(Result, 0); |
| 373 | } |
| 374 | |
| 375 | // 64-bit Microsoft platforms support 128 bit cmpxchg operations. They are |
| 376 | // prototyped like this: |
| 377 | // |
| 378 | // unsigned char _InterlockedCompareExchange128...( |
| 379 | // __int64 volatile * _Destination, |
| 380 | // __int64 _ExchangeHigh, |
| 381 | // __int64 _ExchangeLow, |
| 382 | // __int64 * _ComparandResult); |
| 383 | static Value *EmitAtomicCmpXchg128ForMSIntrin(CodeGenFunction &CGF, |
| 384 | const CallExpr *E, |
| 385 | AtomicOrdering SuccessOrdering) { |
| 386 | assert(E->getNumArgs() == 4)(static_cast <bool> (E->getNumArgs() == 4) ? void (0 ) : __assert_fail ("E->getNumArgs() == 4", "clang/lib/CodeGen/CGBuiltin.cpp" , 386, __extension__ __PRETTY_FUNCTION__)); |
| 387 | llvm::Value *Destination = CGF.EmitScalarExpr(E->getArg(0)); |
| 388 | llvm::Value *ExchangeHigh = CGF.EmitScalarExpr(E->getArg(1)); |
| 389 | llvm::Value *ExchangeLow = CGF.EmitScalarExpr(E->getArg(2)); |
| 390 | llvm::Value *ComparandPtr = CGF.EmitScalarExpr(E->getArg(3)); |
| 391 | |
| 392 | assert(Destination->getType()->isPointerTy())(static_cast <bool> (Destination->getType()->isPointerTy ()) ? void (0) : __assert_fail ("Destination->getType()->isPointerTy()" , "clang/lib/CodeGen/CGBuiltin.cpp", 392, __extension__ __PRETTY_FUNCTION__ )); |
| 393 | assert(!ExchangeHigh->getType()->isPointerTy())(static_cast <bool> (!ExchangeHigh->getType()->isPointerTy ()) ? void (0) : __assert_fail ("!ExchangeHigh->getType()->isPointerTy()" , "clang/lib/CodeGen/CGBuiltin.cpp", 393, __extension__ __PRETTY_FUNCTION__ )); |
| 394 | assert(!ExchangeLow->getType()->isPointerTy())(static_cast <bool> (!ExchangeLow->getType()->isPointerTy ()) ? void (0) : __assert_fail ("!ExchangeLow->getType()->isPointerTy()" , "clang/lib/CodeGen/CGBuiltin.cpp", 394, __extension__ __PRETTY_FUNCTION__ )); |
| 395 | assert(ComparandPtr->getType()->isPointerTy())(static_cast <bool> (ComparandPtr->getType()->isPointerTy ()) ? void (0) : __assert_fail ("ComparandPtr->getType()->isPointerTy()" , "clang/lib/CodeGen/CGBuiltin.cpp", 395, __extension__ __PRETTY_FUNCTION__ )); |
| 396 | |
| 397 | // For Release ordering, the failure ordering should be Monotonic. |
| 398 | auto FailureOrdering = SuccessOrdering == AtomicOrdering::Release |
| 399 | ? AtomicOrdering::Monotonic |
| 400 | : SuccessOrdering; |
| 401 | |
| 402 | // Convert to i128 pointers and values. |
| 403 | llvm::Type *Int128Ty = llvm::IntegerType::get(CGF.getLLVMContext(), 128); |
| 404 | llvm::Type *Int128PtrTy = Int128Ty->getPointerTo(); |
| 405 | Destination = CGF.Builder.CreateBitCast(Destination, Int128PtrTy); |
| 406 | Address ComparandResult(CGF.Builder.CreateBitCast(ComparandPtr, Int128PtrTy), |
| 407 | Int128Ty, CGF.getContext().toCharUnitsFromBits(128)); |
| 408 | |
| 409 | // (((i128)hi) << 64) | ((i128)lo) |
| 410 | ExchangeHigh = CGF.Builder.CreateZExt(ExchangeHigh, Int128Ty); |
| 411 | ExchangeLow = CGF.Builder.CreateZExt(ExchangeLow, Int128Ty); |
| 412 | ExchangeHigh = |
| 413 | CGF.Builder.CreateShl(ExchangeHigh, llvm::ConstantInt::get(Int128Ty, 64)); |
| 414 | llvm::Value *Exchange = CGF.Builder.CreateOr(ExchangeHigh, ExchangeLow); |
| 415 | |
| 416 | // Load the comparand for the instruction. |
| 417 | llvm::Value *Comparand = CGF.Builder.CreateLoad(ComparandResult); |
| 418 | |
| 419 | auto *CXI = CGF.Builder.CreateAtomicCmpXchg(Destination, Comparand, Exchange, |
| 420 | SuccessOrdering, FailureOrdering); |
| 421 | |
| 422 | // The atomic instruction is marked volatile for consistency with MSVC. This |
| 423 | // blocks the few atomics optimizations that LLVM has. If we want to optimize |
| 424 | // _Interlocked* operations in the future, we will have to remove the volatile |
| 425 | // marker. |
| 426 | CXI->setVolatile(true); |
| 427 | |
| 428 | // Store the result as an outparameter. |
| 429 | CGF.Builder.CreateStore(CGF.Builder.CreateExtractValue(CXI, 0), |
| 430 | ComparandResult); |
| 431 | |
| 432 | // Get the success boolean and zero extend it to i8. |
| 433 | Value *Success = CGF.Builder.CreateExtractValue(CXI, 1); |
| 434 | return CGF.Builder.CreateZExt(Success, CGF.Int8Ty); |
| 435 | } |
| 436 | |
| 437 | static Value *EmitAtomicIncrementValue(CodeGenFunction &CGF, const CallExpr *E, |
| 438 | AtomicOrdering Ordering = AtomicOrdering::SequentiallyConsistent) { |
| 439 | assert(E->getArg(0)->getType()->isPointerType())(static_cast <bool> (E->getArg(0)->getType()-> isPointerType()) ? void (0) : __assert_fail ("E->getArg(0)->getType()->isPointerType()" , "clang/lib/CodeGen/CGBuiltin.cpp", 439, __extension__ __PRETTY_FUNCTION__ )); |
| 440 | |
| 441 | auto *IntTy = CGF.ConvertType(E->getType()); |
| 442 | auto *Result = CGF.Builder.CreateAtomicRMW( |
| 443 | AtomicRMWInst::Add, |
| 444 | CGF.EmitScalarExpr(E->getArg(0)), |
| 445 | ConstantInt::get(IntTy, 1), |
| 446 | Ordering); |
| 447 | return CGF.Builder.CreateAdd(Result, ConstantInt::get(IntTy, 1)); |
| 448 | } |
| 449 | |
| 450 | static Value *EmitAtomicDecrementValue(CodeGenFunction &CGF, const CallExpr *E, |
| 451 | AtomicOrdering Ordering = AtomicOrdering::SequentiallyConsistent) { |
| 452 | assert(E->getArg(0)->getType()->isPointerType())(static_cast <bool> (E->getArg(0)->getType()-> isPointerType()) ? void (0) : __assert_fail ("E->getArg(0)->getType()->isPointerType()" , "clang/lib/CodeGen/CGBuiltin.cpp", 452, __extension__ __PRETTY_FUNCTION__ )); |
| 453 | |
| 454 | auto *IntTy = CGF.ConvertType(E->getType()); |
| 455 | auto *Result = CGF.Builder.CreateAtomicRMW( |
| 456 | AtomicRMWInst::Sub, |
| 457 | CGF.EmitScalarExpr(E->getArg(0)), |
| 458 | ConstantInt::get(IntTy, 1), |
| 459 | Ordering); |
| 460 | return CGF.Builder.CreateSub(Result, ConstantInt::get(IntTy, 1)); |
| 461 | } |
| 462 | |
| 463 | // Build a plain volatile load. |
| 464 | static Value *EmitISOVolatileLoad(CodeGenFunction &CGF, const CallExpr *E) { |
| 465 | Value *Ptr = CGF.EmitScalarExpr(E->getArg(0)); |
| 466 | QualType ElTy = E->getArg(0)->getType()->getPointeeType(); |
| 467 | CharUnits LoadSize = CGF.getContext().getTypeSizeInChars(ElTy); |
| 468 | llvm::Type *ITy = |
| 469 | llvm::IntegerType::get(CGF.getLLVMContext(), LoadSize.getQuantity() * 8); |
| 470 | Ptr = CGF.Builder.CreateBitCast(Ptr, ITy->getPointerTo()); |
| 471 | llvm::LoadInst *Load = CGF.Builder.CreateAlignedLoad(ITy, Ptr, LoadSize); |
| 472 | Load->setVolatile(true); |
| 473 | return Load; |
| 474 | } |
| 475 | |
| 476 | // Build a plain volatile store. |
| 477 | static Value *EmitISOVolatileStore(CodeGenFunction &CGF, const CallExpr *E) { |
| 478 | Value *Ptr = CGF.EmitScalarExpr(E->getArg(0)); |
| 479 | Value *Value = CGF.EmitScalarExpr(E->getArg(1)); |
| 480 | QualType ElTy = E->getArg(0)->getType()->getPointeeType(); |
| 481 | CharUnits StoreSize = CGF.getContext().getTypeSizeInChars(ElTy); |
| 482 | llvm::Type *ITy = |
| 483 | llvm::IntegerType::get(CGF.getLLVMContext(), StoreSize.getQuantity() * 8); |
| 484 | Ptr = CGF.Builder.CreateBitCast(Ptr, ITy->getPointerTo()); |
| 485 | llvm::StoreInst *Store = |
| 486 | CGF.Builder.CreateAlignedStore(Value, Ptr, StoreSize); |
| 487 | Store->setVolatile(true); |
| 488 | return Store; |
| 489 | } |
| 490 | |
| 491 | // Emit a simple mangled intrinsic that has 1 argument and a return type |
| 492 | // matching the argument type. Depending on mode, this may be a constrained |
| 493 | // floating-point intrinsic. |
| 494 | static Value *emitUnaryMaybeConstrainedFPBuiltin(CodeGenFunction &CGF, |
| 495 | const CallExpr *E, unsigned IntrinsicID, |
| 496 | unsigned ConstrainedIntrinsicID) { |
| 497 | llvm::Value *Src0 = CGF.EmitScalarExpr(E->getArg(0)); |
| 498 | |
| 499 | if (CGF.Builder.getIsFPConstrained()) { |
| 500 | CodeGenFunction::CGFPOptionsRAII FPOptsRAII(CGF, E); |
| 501 | Function *F = CGF.CGM.getIntrinsic(ConstrainedIntrinsicID, Src0->getType()); |
| 502 | return CGF.Builder.CreateConstrainedFPCall(F, { Src0 }); |
| 503 | } else { |
| 504 | Function *F = CGF.CGM.getIntrinsic(IntrinsicID, Src0->getType()); |
| 505 | return CGF.Builder.CreateCall(F, Src0); |
| 506 | } |
| 507 | } |
| 508 | |
| 509 | // Emit an intrinsic that has 2 operands of the same type as its result. |
| 510 | // Depending on mode, this may be a constrained floating-point intrinsic. |
| 511 | static Value *emitBinaryMaybeConstrainedFPBuiltin(CodeGenFunction &CGF, |
| 512 | const CallExpr *E, unsigned IntrinsicID, |
| 513 | unsigned ConstrainedIntrinsicID) { |
| 514 | llvm::Value *Src0 = CGF.EmitScalarExpr(E->getArg(0)); |
| 515 | llvm::Value *Src1 = CGF.EmitScalarExpr(E->getArg(1)); |
| 516 | |
| 517 | if (CGF.Builder.getIsFPConstrained()) { |
| 518 | CodeGenFunction::CGFPOptionsRAII FPOptsRAII(CGF, E); |
| 519 | Function *F = CGF.CGM.getIntrinsic(ConstrainedIntrinsicID, Src0->getType()); |
| 520 | return CGF.Builder.CreateConstrainedFPCall(F, { Src0, Src1 }); |
| 521 | } else { |
| 522 | Function *F = CGF.CGM.getIntrinsic(IntrinsicID, Src0->getType()); |
| 523 | return CGF.Builder.CreateCall(F, { Src0, Src1 }); |
| 524 | } |
| 525 | } |
| 526 | |
| 527 | // Emit an intrinsic that has 3 operands of the same type as its result. |
| 528 | // Depending on mode, this may be a constrained floating-point intrinsic. |
| 529 | static Value *emitTernaryMaybeConstrainedFPBuiltin(CodeGenFunction &CGF, |
| 530 | const CallExpr *E, unsigned IntrinsicID, |
| 531 | unsigned ConstrainedIntrinsicID) { |
| 532 | llvm::Value *Src0 = CGF.EmitScalarExpr(E->getArg(0)); |
| 533 | llvm::Value *Src1 = CGF.EmitScalarExpr(E->getArg(1)); |
| 534 | llvm::Value *Src2 = CGF.EmitScalarExpr(E->getArg(2)); |
| 535 | |
| 536 | if (CGF.Builder.getIsFPConstrained()) { |
| 537 | CodeGenFunction::CGFPOptionsRAII FPOptsRAII(CGF, E); |
| 538 | Function *F = CGF.CGM.getIntrinsic(ConstrainedIntrinsicID, Src0->getType()); |
| 539 | return CGF.Builder.CreateConstrainedFPCall(F, { Src0, Src1, Src2 }); |
| 540 | } else { |
| 541 | Function *F = CGF.CGM.getIntrinsic(IntrinsicID, Src0->getType()); |
| 542 | return CGF.Builder.CreateCall(F, { Src0, Src1, Src2 }); |
| 543 | } |
| 544 | } |
| 545 | |
| 546 | // Emit an intrinsic where all operands are of the same type as the result. |
| 547 | // Depending on mode, this may be a constrained floating-point intrinsic. |
| 548 | static Value *emitCallMaybeConstrainedFPBuiltin(CodeGenFunction &CGF, |
| 549 | unsigned IntrinsicID, |
| 550 | unsigned ConstrainedIntrinsicID, |
| 551 | llvm::Type *Ty, |
| 552 | ArrayRef<Value *> Args) { |
| 553 | Function *F; |
| 554 | if (CGF.Builder.getIsFPConstrained()) |
| 555 | F = CGF.CGM.getIntrinsic(ConstrainedIntrinsicID, Ty); |
| 556 | else |
| 557 | F = CGF.CGM.getIntrinsic(IntrinsicID, Ty); |
| 558 | |
| 559 | if (CGF.Builder.getIsFPConstrained()) |
| 560 | return CGF.Builder.CreateConstrainedFPCall(F, Args); |
| 561 | else |
| 562 | return CGF.Builder.CreateCall(F, Args); |
| 563 | } |
| 564 | |
| 565 | // Emit a simple mangled intrinsic that has 1 argument and a return type |
| 566 | // matching the argument type. |
| 567 | static Value *emitUnaryBuiltin(CodeGenFunction &CGF, const CallExpr *E, |
| 568 | unsigned IntrinsicID, |
| 569 | llvm::StringRef Name = "") { |
| 570 | llvm::Value *Src0 = CGF.EmitScalarExpr(E->getArg(0)); |
| 571 | |
| 572 | Function *F = CGF.CGM.getIntrinsic(IntrinsicID, Src0->getType()); |
| 573 | return CGF.Builder.CreateCall(F, Src0, Name); |
| 574 | } |
| 575 | |
| 576 | // Emit an intrinsic that has 2 operands of the same type as its result. |
| 577 | static Value *emitBinaryBuiltin(CodeGenFunction &CGF, |
| 578 | const CallExpr *E, |
| 579 | unsigned IntrinsicID) { |
| 580 | llvm::Value *Src0 = CGF.EmitScalarExpr(E->getArg(0)); |
| 581 | llvm::Value *Src1 = CGF.EmitScalarExpr(E->getArg(1)); |
| 582 | |
| 583 | Function *F = CGF.CGM.getIntrinsic(IntrinsicID, Src0->getType()); |
| 584 | return CGF.Builder.CreateCall(F, { Src0, Src1 }); |
| 585 | } |
| 586 | |
| 587 | // Emit an intrinsic that has 3 operands of the same type as its result. |
| 588 | static Value *emitTernaryBuiltin(CodeGenFunction &CGF, |
| 589 | const CallExpr *E, |
| 590 | unsigned IntrinsicID) { |
| 591 | llvm::Value *Src0 = CGF.EmitScalarExpr(E->getArg(0)); |
| 592 | llvm::Value *Src1 = CGF.EmitScalarExpr(E->getArg(1)); |
| 593 | llvm::Value *Src2 = CGF.EmitScalarExpr(E->getArg(2)); |
| 594 | |
| 595 | Function *F = CGF.CGM.getIntrinsic(IntrinsicID, Src0->getType()); |
| 596 | return CGF.Builder.CreateCall(F, { Src0, Src1, Src2 }); |
| 597 | } |
| 598 | |
| 599 | // Emit an intrinsic that has 1 float or double operand, and 1 integer. |
| 600 | static Value *emitFPIntBuiltin(CodeGenFunction &CGF, |
| 601 | const CallExpr *E, |
| 602 | unsigned IntrinsicID) { |
| 603 | llvm::Value *Src0 = CGF.EmitScalarExpr(E->getArg(0)); |
| 604 | llvm::Value *Src1 = CGF.EmitScalarExpr(E->getArg(1)); |
| 605 | |
| 606 | Function *F = CGF.CGM.getIntrinsic(IntrinsicID, Src0->getType()); |
| 607 | return CGF.Builder.CreateCall(F, {Src0, Src1}); |
| 608 | } |
| 609 | |
| 610 | // Emit an intrinsic that has overloaded integer result and fp operand. |
| 611 | static Value * |
| 612 | emitMaybeConstrainedFPToIntRoundBuiltin(CodeGenFunction &CGF, const CallExpr *E, |
| 613 | unsigned IntrinsicID, |
| 614 | unsigned ConstrainedIntrinsicID) { |
| 615 | llvm::Type *ResultType = CGF.ConvertType(E->getType()); |
| 616 | llvm::Value *Src0 = CGF.EmitScalarExpr(E->getArg(0)); |
| 617 | |
| 618 | if (CGF.Builder.getIsFPConstrained()) { |
| 619 | CodeGenFunction::CGFPOptionsRAII FPOptsRAII(CGF, E); |
| 620 | Function *F = CGF.CGM.getIntrinsic(ConstrainedIntrinsicID, |
| 621 | {ResultType, Src0->getType()}); |
| 622 | return CGF.Builder.CreateConstrainedFPCall(F, {Src0}); |
| 623 | } else { |
| 624 | Function *F = |
| 625 | CGF.CGM.getIntrinsic(IntrinsicID, {ResultType, Src0->getType()}); |
| 626 | return CGF.Builder.CreateCall(F, Src0); |
| 627 | } |
| 628 | } |
| 629 | |
| 630 | /// EmitFAbs - Emit a call to @llvm.fabs(). |
| 631 | static Value *EmitFAbs(CodeGenFunction &CGF, Value *V) { |
| 632 | Function *F = CGF.CGM.getIntrinsic(Intrinsic::fabs, V->getType()); |
| 633 | llvm::CallInst *Call = CGF.Builder.CreateCall(F, V); |
| 634 | Call->setDoesNotAccessMemory(); |
| 635 | return Call; |
| 636 | } |
| 637 | |
| 638 | /// Emit the computation of the sign bit for a floating point value. Returns |
| 639 | /// the i1 sign bit value. |
| 640 | static Value *EmitSignBit(CodeGenFunction &CGF, Value *V) { |
| 641 | LLVMContext &C = CGF.CGM.getLLVMContext(); |
| 642 | |
| 643 | llvm::Type *Ty = V->getType(); |
| 644 | int Width = Ty->getPrimitiveSizeInBits(); |
| 645 | llvm::Type *IntTy = llvm::IntegerType::get(C, Width); |
| 646 | V = CGF.Builder.CreateBitCast(V, IntTy); |
| 647 | if (Ty->isPPC_FP128Ty()) { |
| 648 | // We want the sign bit of the higher-order double. The bitcast we just |
| 649 | // did works as if the double-double was stored to memory and then |
| 650 | // read as an i128. The "store" will put the higher-order double in the |
| 651 | // lower address in both little- and big-Endian modes, but the "load" |
| 652 | // will treat those bits as a different part of the i128: the low bits in |
| 653 | // little-Endian, the high bits in big-Endian. Therefore, on big-Endian |
| 654 | // we need to shift the high bits down to the low before truncating. |
| 655 | Width >>= 1; |
| 656 | if (CGF.getTarget().isBigEndian()) { |
| 657 | Value *ShiftCst = llvm::ConstantInt::get(IntTy, Width); |
| 658 | V = CGF.Builder.CreateLShr(V, ShiftCst); |
| 659 | } |
| 660 | // We are truncating value in order to extract the higher-order |
| 661 | // double, which we will be using to extract the sign from. |
| 662 | IntTy = llvm::IntegerType::get(C, Width); |
| 663 | V = CGF.Builder.CreateTrunc(V, IntTy); |
| 664 | } |
| 665 | Value *Zero = llvm::Constant::getNullValue(IntTy); |
| 666 | return CGF.Builder.CreateICmpSLT(V, Zero); |
| 667 | } |
| 668 | |
| 669 | static RValue emitLibraryCall(CodeGenFunction &CGF, const FunctionDecl *FD, |
| 670 | const CallExpr *E, llvm::Constant *calleeValue) { |
| 671 | CGCallee callee = CGCallee::forDirect(calleeValue, GlobalDecl(FD)); |
| 672 | return CGF.EmitCall(E->getCallee()->getType(), callee, E, ReturnValueSlot()); |
| 673 | } |
| 674 | |
| 675 | /// Emit a call to llvm.{sadd,uadd,ssub,usub,smul,umul}.with.overflow.* |
| 676 | /// depending on IntrinsicID. |
| 677 | /// |
| 678 | /// \arg CGF The current codegen function. |
| 679 | /// \arg IntrinsicID The ID for the Intrinsic we wish to generate. |
| 680 | /// \arg X The first argument to the llvm.*.with.overflow.*. |
| 681 | /// \arg Y The second argument to the llvm.*.with.overflow.*. |
| 682 | /// \arg Carry The carry returned by the llvm.*.with.overflow.*. |
| 683 | /// \returns The result (i.e. sum/product) returned by the intrinsic. |
| 684 | static llvm::Value *EmitOverflowIntrinsic(CodeGenFunction &CGF, |
| 685 | const llvm::Intrinsic::ID IntrinsicID, |
| 686 | llvm::Value *X, llvm::Value *Y, |
| 687 | llvm::Value *&Carry) { |
| 688 | // Make sure we have integers of the same width. |
| 689 | assert(X->getType() == Y->getType() &&(static_cast <bool> (X->getType() == Y->getType() && "Arguments must be the same type. (Did you forget to make sure both " "arguments have the same integer width?)") ? void (0) : __assert_fail ("X->getType() == Y->getType() && \"Arguments must be the same type. (Did you forget to make sure both \" \"arguments have the same integer width?)\"" , "clang/lib/CodeGen/CGBuiltin.cpp", 691, __extension__ __PRETTY_FUNCTION__ )) |
| 690 | "Arguments must be the same type. (Did you forget to make sure both "(static_cast <bool> (X->getType() == Y->getType() && "Arguments must be the same type. (Did you forget to make sure both " "arguments have the same integer width?)") ? void (0) : __assert_fail ("X->getType() == Y->getType() && \"Arguments must be the same type. (Did you forget to make sure both \" \"arguments have the same integer width?)\"" , "clang/lib/CodeGen/CGBuiltin.cpp", 691, __extension__ __PRETTY_FUNCTION__ )) |
| 691 | "arguments have the same integer width?)")(static_cast <bool> (X->getType() == Y->getType() && "Arguments must be the same type. (Did you forget to make sure both " "arguments have the same integer width?)") ? void (0) : __assert_fail ("X->getType() == Y->getType() && \"Arguments must be the same type. (Did you forget to make sure both \" \"arguments have the same integer width?)\"" , "clang/lib/CodeGen/CGBuiltin.cpp", 691, __extension__ __PRETTY_FUNCTION__ )); |
| 692 | |
| 693 | Function *Callee = CGF.CGM.getIntrinsic(IntrinsicID, X->getType()); |
| 694 | llvm::Value *Tmp = CGF.Builder.CreateCall(Callee, {X, Y}); |
| 695 | Carry = CGF.Builder.CreateExtractValue(Tmp, 1); |
| 696 | return CGF.Builder.CreateExtractValue(Tmp, 0); |
| 697 | } |
| 698 | |
| 699 | static Value *emitRangedBuiltin(CodeGenFunction &CGF, |
| 700 | unsigned IntrinsicID, |
| 701 | int low, int high) { |
| 702 | llvm::MDBuilder MDHelper(CGF.getLLVMContext()); |
| 703 | llvm::MDNode *RNode = MDHelper.createRange(APInt(32, low), APInt(32, high)); |
| 704 | Function *F = CGF.CGM.getIntrinsic(IntrinsicID, {}); |
| 705 | llvm::Instruction *Call = CGF.Builder.CreateCall(F); |
| 706 | Call->setMetadata(llvm::LLVMContext::MD_range, RNode); |
| 707 | Call->setMetadata(llvm::LLVMContext::MD_noundef, |
| 708 | llvm::MDNode::get(CGF.getLLVMContext(), std::nullopt)); |
| 709 | return Call; |
| 710 | } |
| 711 | |
| 712 | namespace { |
| 713 | struct WidthAndSignedness { |
| 714 | unsigned Width; |
| 715 | bool Signed; |
| 716 | }; |
| 717 | } |
| 718 | |
| 719 | static WidthAndSignedness |
| 720 | getIntegerWidthAndSignedness(const clang::ASTContext &context, |
| 721 | const clang::QualType Type) { |
| 722 | assert(Type->isIntegerType() && "Given type is not an integer.")(static_cast <bool> (Type->isIntegerType() && "Given type is not an integer.") ? void (0) : __assert_fail ( "Type->isIntegerType() && \"Given type is not an integer.\"" , "clang/lib/CodeGen/CGBuiltin.cpp", 722, __extension__ __PRETTY_FUNCTION__ )); |
| 723 | unsigned Width = Type->isBooleanType() ? 1 |
| 724 | : Type->isBitIntType() ? context.getIntWidth(Type) |
| 725 | : context.getTypeInfo(Type).Width; |
| 726 | bool Signed = Type->isSignedIntegerType(); |
| 727 | return {Width, Signed}; |
| 728 | } |
| 729 | |
| 730 | // Given one or more integer types, this function produces an integer type that |
| 731 | // encompasses them: any value in one of the given types could be expressed in |
| 732 | // the encompassing type. |
| 733 | static struct WidthAndSignedness |
| 734 | EncompassingIntegerType(ArrayRef<struct WidthAndSignedness> Types) { |
| 735 | assert(Types.size() > 0 && "Empty list of types.")(static_cast <bool> (Types.size() > 0 && "Empty list of types." ) ? void (0) : __assert_fail ("Types.size() > 0 && \"Empty list of types.\"" , "clang/lib/CodeGen/CGBuiltin.cpp", 735, __extension__ __PRETTY_FUNCTION__ )); |
| 736 | |
| 737 | // If any of the given types is signed, we must return a signed type. |
| 738 | bool Signed = false; |
| 739 | for (const auto &Type : Types) { |
| 740 | Signed |= Type.Signed; |
| 741 | } |
| 742 | |
| 743 | // The encompassing type must have a width greater than or equal to the width |
| 744 | // of the specified types. Additionally, if the encompassing type is signed, |
| 745 | // its width must be strictly greater than the width of any unsigned types |
| 746 | // given. |
| 747 | unsigned Width = 0; |
| 748 | for (const auto &Type : Types) { |
| 749 | unsigned MinWidth = Type.Width + (Signed && !Type.Signed); |
| 750 | if (Width < MinWidth) { |
| 751 | Width = MinWidth; |
| 752 | } |
| 753 | } |
| 754 | |
| 755 | return {Width, Signed}; |
| 756 | } |
| 757 | |
| 758 | Value *CodeGenFunction::EmitVAStartEnd(Value *ArgValue, bool IsStart) { |
| 759 | llvm::Type *DestType = Int8PtrTy; |
| 760 | if (ArgValue->getType() != DestType) |
| 761 | ArgValue = |
| 762 | Builder.CreateBitCast(ArgValue, DestType, ArgValue->getName().data()); |
| 763 | |
| 764 | Intrinsic::ID inst = IsStart ? Intrinsic::vastart : Intrinsic::vaend; |
| 765 | return Builder.CreateCall(CGM.getIntrinsic(inst), ArgValue); |
| 766 | } |
| 767 | |
| 768 | /// Checks if using the result of __builtin_object_size(p, @p From) in place of |
| 769 | /// __builtin_object_size(p, @p To) is correct |
| 770 | static bool areBOSTypesCompatible(int From, int To) { |
| 771 | // Note: Our __builtin_object_size implementation currently treats Type=0 and |
| 772 | // Type=2 identically. Encoding this implementation detail here may make |
| 773 | // improving __builtin_object_size difficult in the future, so it's omitted. |
| 774 | return From == To || (From == 0 && To == 1) || (From == 3 && To == 2); |
| 775 | } |
| 776 | |
| 777 | static llvm::Value * |
| 778 | getDefaultBuiltinObjectSizeResult(unsigned Type, llvm::IntegerType *ResType) { |
| 779 | return ConstantInt::get(ResType, (Type & 2) ? 0 : -1, /*isSigned=*/true); |
| 780 | } |
| 781 | |
| 782 | llvm::Value * |
| 783 | CodeGenFunction::evaluateOrEmitBuiltinObjectSize(const Expr *E, unsigned Type, |
| 784 | llvm::IntegerType *ResType, |
| 785 | llvm::Value *EmittedE, |
| 786 | bool IsDynamic) { |
| 787 | uint64_t ObjectSize; |
| 788 | if (!E->tryEvaluateObjectSize(ObjectSize, getContext(), Type)) |
| 789 | return emitBuiltinObjectSize(E, Type, ResType, EmittedE, IsDynamic); |
| 790 | return ConstantInt::get(ResType, ObjectSize, /*isSigned=*/true); |
| 791 | } |
| 792 | |
| 793 | /// Returns a Value corresponding to the size of the given expression. |
| 794 | /// This Value may be either of the following: |
| 795 | /// - A llvm::Argument (if E is a param with the pass_object_size attribute on |
| 796 | /// it) |
| 797 | /// - A call to the @llvm.objectsize intrinsic |
| 798 | /// |
| 799 | /// EmittedE is the result of emitting `E` as a scalar expr. If it's non-null |
| 800 | /// and we wouldn't otherwise try to reference a pass_object_size parameter, |
| 801 | /// we'll call @llvm.objectsize on EmittedE, rather than emitting E. |
| 802 | llvm::Value * |
| 803 | CodeGenFunction::emitBuiltinObjectSize(const Expr *E, unsigned Type, |
| 804 | llvm::IntegerType *ResType, |
| 805 | llvm::Value *EmittedE, bool IsDynamic) { |
| 806 | // We need to reference an argument if the pointer is a parameter with the |
| 807 | // pass_object_size attribute. |
| 808 | if (auto *D = dyn_cast<DeclRefExpr>(E->IgnoreParenImpCasts())) { |
| 809 | auto *Param = dyn_cast<ParmVarDecl>(D->getDecl()); |
| 810 | auto *PS = D->getDecl()->getAttr<PassObjectSizeAttr>(); |
| 811 | if (Param != nullptr && PS != nullptr && |
| 812 | areBOSTypesCompatible(PS->getType(), Type)) { |
| 813 | auto Iter = SizeArguments.find(Param); |
| 814 | assert(Iter != SizeArguments.end())(static_cast <bool> (Iter != SizeArguments.end()) ? void (0) : __assert_fail ("Iter != SizeArguments.end()", "clang/lib/CodeGen/CGBuiltin.cpp" , 814, __extension__ __PRETTY_FUNCTION__)); |
| 815 | |
| 816 | const ImplicitParamDecl *D = Iter->second; |
| 817 | auto DIter = LocalDeclMap.find(D); |
| 818 | assert(DIter != LocalDeclMap.end())(static_cast <bool> (DIter != LocalDeclMap.end()) ? void (0) : __assert_fail ("DIter != LocalDeclMap.end()", "clang/lib/CodeGen/CGBuiltin.cpp" , 818, __extension__ __PRETTY_FUNCTION__)); |
| 819 | |
| 820 | return EmitLoadOfScalar(DIter->second, /*Volatile=*/false, |
| 821 | getContext().getSizeType(), E->getBeginLoc()); |
| 822 | } |
| 823 | } |
| 824 | |
| 825 | // LLVM can't handle Type=3 appropriately, and __builtin_object_size shouldn't |
| 826 | // evaluate E for side-effects. In either case, we shouldn't lower to |
| 827 | // @llvm.objectsize. |
| 828 | if (Type == 3 || (!EmittedE && E->HasSideEffects(getContext()))) |
| 829 | return getDefaultBuiltinObjectSizeResult(Type, ResType); |
| 830 | |
| 831 | Value *Ptr = EmittedE ? EmittedE : EmitScalarExpr(E); |
| 832 | assert(Ptr->getType()->isPointerTy() &&(static_cast <bool> (Ptr->getType()->isPointerTy( ) && "Non-pointer passed to __builtin_object_size?") ? void (0) : __assert_fail ("Ptr->getType()->isPointerTy() && \"Non-pointer passed to __builtin_object_size?\"" , "clang/lib/CodeGen/CGBuiltin.cpp", 833, __extension__ __PRETTY_FUNCTION__ )) |
| 833 | "Non-pointer passed to __builtin_object_size?")(static_cast <bool> (Ptr->getType()->isPointerTy( ) && "Non-pointer passed to __builtin_object_size?") ? void (0) : __assert_fail ("Ptr->getType()->isPointerTy() && \"Non-pointer passed to __builtin_object_size?\"" , "clang/lib/CodeGen/CGBuiltin.cpp", 833, __extension__ __PRETTY_FUNCTION__ )); |
| 834 | |
| 835 | Function *F = |
| 836 | CGM.getIntrinsic(Intrinsic::objectsize, {ResType, Ptr->getType()}); |
| 837 | |
| 838 | // LLVM only supports 0 and 2, make sure that we pass along that as a boolean. |
| 839 | Value *Min = Builder.getInt1((Type & 2) != 0); |
| 840 | // For GCC compatibility, __builtin_object_size treat NULL as unknown size. |
| 841 | Value *NullIsUnknown = Builder.getTrue(); |
| 842 | Value *Dynamic = Builder.getInt1(IsDynamic); |
| 843 | return Builder.CreateCall(F, {Ptr, Min, NullIsUnknown, Dynamic}); |
| 844 | } |
| 845 | |
| 846 | namespace { |
| 847 | /// A struct to generically describe a bit test intrinsic. |
| 848 | struct BitTest { |
| 849 | enum ActionKind : uint8_t { TestOnly, Complement, Reset, Set }; |
| 850 | enum InterlockingKind : uint8_t { |
| 851 | Unlocked, |
| 852 | Sequential, |
| 853 | Acquire, |
| 854 | Release, |
| 855 | NoFence |
| 856 | }; |
| 857 | |
| 858 | ActionKind Action; |
| 859 | InterlockingKind Interlocking; |
| 860 | bool Is64Bit; |
| 861 | |
| 862 | static BitTest decodeBitTestBuiltin(unsigned BuiltinID); |
| 863 | }; |
| 864 | } // namespace |
| 865 | |
| 866 | BitTest BitTest::decodeBitTestBuiltin(unsigned BuiltinID) { |
| 867 | switch (BuiltinID) { |
| 868 | // Main portable variants. |
| 869 | case Builtin::BI_bittest: |
| 870 | return {TestOnly, Unlocked, false}; |
| 871 | case Builtin::BI_bittestandcomplement: |
| 872 | return {Complement, Unlocked, false}; |
| 873 | case Builtin::BI_bittestandreset: |
| 874 | return {Reset, Unlocked, false}; |
| 875 | case Builtin::BI_bittestandset: |
| 876 | return {Set, Unlocked, false}; |
| 877 | case Builtin::BI_interlockedbittestandreset: |
| 878 | return {Reset, Sequential, false}; |
| 879 | case Builtin::BI_interlockedbittestandset: |
| 880 | return {Set, Sequential, false}; |
| 881 | |
| 882 | // X86-specific 64-bit variants. |
| 883 | case Builtin::BI_bittest64: |
| 884 | return {TestOnly, Unlocked, true}; |
| 885 | case Builtin::BI_bittestandcomplement64: |
| 886 | return {Complement, Unlocked, true}; |
| 887 | case Builtin::BI_bittestandreset64: |
| 888 | return {Reset, Unlocked, true}; |
| 889 | case Builtin::BI_bittestandset64: |
| 890 | return {Set, Unlocked, true}; |
| 891 | case Builtin::BI_interlockedbittestandreset64: |
| 892 | return {Reset, Sequential, true}; |
| 893 | case Builtin::BI_interlockedbittestandset64: |
| 894 | return {Set, Sequential, true}; |
| 895 | |
| 896 | // ARM/AArch64-specific ordering variants. |
| 897 | case Builtin::BI_interlockedbittestandset_acq: |
| 898 | return {Set, Acquire, false}; |
| 899 | case Builtin::BI_interlockedbittestandset_rel: |
| 900 | return {Set, Release, false}; |
| 901 | case Builtin::BI_interlockedbittestandset_nf: |
| 902 | return {Set, NoFence, false}; |
| 903 | case Builtin::BI_interlockedbittestandreset_acq: |
| 904 | return {Reset, Acquire, false}; |
| 905 | case Builtin::BI_interlockedbittestandreset_rel: |
| 906 | return {Reset, Release, false}; |
| 907 | case Builtin::BI_interlockedbittestandreset_nf: |
| 908 | return {Reset, NoFence, false}; |
| 909 | } |
| 910 | llvm_unreachable("expected only bittest intrinsics")::llvm::llvm_unreachable_internal("expected only bittest intrinsics" , "clang/lib/CodeGen/CGBuiltin.cpp", 910); |
| 911 | } |
| 912 | |
| 913 | static char bitActionToX86BTCode(BitTest::ActionKind A) { |
| 914 | switch (A) { |
| 915 | case BitTest::TestOnly: return '\0'; |
| 916 | case BitTest::Complement: return 'c'; |
| 917 | case BitTest::Reset: return 'r'; |
| 918 | case BitTest::Set: return 's'; |
| 919 | } |
| 920 | llvm_unreachable("invalid action")::llvm::llvm_unreachable_internal("invalid action", "clang/lib/CodeGen/CGBuiltin.cpp" , 920); |
| 921 | } |
| 922 | |
| 923 | static llvm::Value *EmitX86BitTestIntrinsic(CodeGenFunction &CGF, |
| 924 | BitTest BT, |
| 925 | const CallExpr *E, Value *BitBase, |
| 926 | Value *BitPos) { |
| 927 | char Action = bitActionToX86BTCode(BT.Action); |
| 928 | char SizeSuffix = BT.Is64Bit ? 'q' : 'l'; |
| 929 | |
| 930 | // Build the assembly. |
| 931 | SmallString<64> Asm; |
| 932 | raw_svector_ostream AsmOS(Asm); |
| 933 | if (BT.Interlocking != BitTest::Unlocked) |
| 934 | AsmOS << "lock "; |
| 935 | AsmOS << "bt"; |
| 936 | if (Action) |
| 937 | AsmOS << Action; |
| 938 | AsmOS << SizeSuffix << " $2, ($1)"; |
| 939 | |
| 940 | // Build the constraints. FIXME: We should support immediates when possible. |
| 941 | std::string Constraints = "={@ccc},r,r,~{cc},~{memory}"; |
| 942 | std::string_view MachineClobbers = CGF.getTarget().getClobbers(); |
| 943 | if (!MachineClobbers.empty()) { |
| 944 | Constraints += ','; |
| 945 | Constraints += MachineClobbers; |
| 946 | } |
| 947 | llvm::IntegerType *IntType = llvm::IntegerType::get( |
| 948 | CGF.getLLVMContext(), |
| 949 | CGF.getContext().getTypeSize(E->getArg(1)->getType())); |
| 950 | llvm::Type *IntPtrType = IntType->getPointerTo(); |
| 951 | llvm::FunctionType *FTy = |
| 952 | llvm::FunctionType::get(CGF.Int8Ty, {IntPtrType, IntType}, false); |
| 953 | |
| 954 | llvm::InlineAsm *IA = |
| 955 | llvm::InlineAsm::get(FTy, Asm, Constraints, /*hasSideEffects=*/true); |
| 956 | return CGF.Builder.CreateCall(IA, {BitBase, BitPos}); |
| 957 | } |
| 958 | |
| 959 | static llvm::AtomicOrdering |
| 960 | getBitTestAtomicOrdering(BitTest::InterlockingKind I) { |
| 961 | switch (I) { |
| 962 | case BitTest::Unlocked: return llvm::AtomicOrdering::NotAtomic; |
| 963 | case BitTest::Sequential: return llvm::AtomicOrdering::SequentiallyConsistent; |
| 964 | case BitTest::Acquire: return llvm::AtomicOrdering::Acquire; |
| 965 | case BitTest::Release: return llvm::AtomicOrdering::Release; |
| 966 | case BitTest::NoFence: return llvm::AtomicOrdering::Monotonic; |
| 967 | } |
| 968 | llvm_unreachable("invalid interlocking")::llvm::llvm_unreachable_internal("invalid interlocking", "clang/lib/CodeGen/CGBuiltin.cpp" , 968); |
| 969 | } |
| 970 | |
| 971 | /// Emit a _bittest* intrinsic. These intrinsics take a pointer to an array of |
| 972 | /// bits and a bit position and read and optionally modify the bit at that |
| 973 | /// position. The position index can be arbitrarily large, i.e. it can be larger |
| 974 | /// than 31 or 63, so we need an indexed load in the general case. |
| 975 | static llvm::Value *EmitBitTestIntrinsic(CodeGenFunction &CGF, |
| 976 | unsigned BuiltinID, |
| 977 | const CallExpr *E) { |
| 978 | Value *BitBase = CGF.EmitScalarExpr(E->getArg(0)); |
| 979 | Value *BitPos = CGF.EmitScalarExpr(E->getArg(1)); |
| 980 | |
| 981 | BitTest BT = BitTest::decodeBitTestBuiltin(BuiltinID); |
| 982 | |
| 983 | // X86 has special BT, BTC, BTR, and BTS instructions that handle the array |
| 984 | // indexing operation internally. Use them if possible. |
| 985 | if (CGF.getTarget().getTriple().isX86()) |
| 986 | return EmitX86BitTestIntrinsic(CGF, BT, E, BitBase, BitPos); |
| 987 | |
| 988 | // Otherwise, use generic code to load one byte and test the bit. Use all but |
| 989 | // the bottom three bits as the array index, and the bottom three bits to form |
| 990 | // a mask. |
| 991 | // Bit = BitBaseI8[BitPos >> 3] & (1 << (BitPos & 0x7)) != 0; |
| 992 | Value *ByteIndex = CGF.Builder.CreateAShr( |
| 993 | BitPos, llvm::ConstantInt::get(BitPos->getType(), 3), "bittest.byteidx"); |
| 994 | Value *BitBaseI8 = CGF.Builder.CreatePointerCast(BitBase, CGF.Int8PtrTy); |
| 995 | Address ByteAddr(CGF.Builder.CreateInBoundsGEP(CGF.Int8Ty, BitBaseI8, |
| 996 | ByteIndex, "bittest.byteaddr"), |
| 997 | CGF.Int8Ty, CharUnits::One()); |
| 998 | Value *PosLow = |
| 999 | CGF.Builder.CreateAnd(CGF.Builder.CreateTrunc(BitPos, CGF.Int8Ty), |
| 1000 | llvm::ConstantInt::get(CGF.Int8Ty, 0x7)); |
| 1001 | |
| 1002 | // The updating instructions will need a mask. |
| 1003 | Value *Mask = nullptr; |
| 1004 | if (BT.Action != BitTest::TestOnly) { |
| 1005 | Mask = CGF.Builder.CreateShl(llvm::ConstantInt::get(CGF.Int8Ty, 1), PosLow, |
| 1006 | "bittest.mask"); |
| 1007 | } |
| 1008 | |
| 1009 | // Check the action and ordering of the interlocked intrinsics. |
| 1010 | llvm::AtomicOrdering Ordering = getBitTestAtomicOrdering(BT.Interlocking); |
| 1011 | |
| 1012 | Value *OldByte = nullptr; |
| 1013 | if (Ordering != llvm::AtomicOrdering::NotAtomic) { |
| 1014 | // Emit a combined atomicrmw load/store operation for the interlocked |
| 1015 | // intrinsics. |
| 1016 | llvm::AtomicRMWInst::BinOp RMWOp = llvm::AtomicRMWInst::Or; |
| 1017 | if (BT.Action == BitTest::Reset) { |
| 1018 | Mask = CGF.Builder.CreateNot(Mask); |
| 1019 | RMWOp = llvm::AtomicRMWInst::And; |
| 1020 | } |
| 1021 | OldByte = CGF.Builder.CreateAtomicRMW(RMWOp, ByteAddr.getPointer(), Mask, |
| 1022 | Ordering); |
| 1023 | } else { |
| 1024 | // Emit a plain load for the non-interlocked intrinsics. |
| 1025 | OldByte = CGF.Builder.CreateLoad(ByteAddr, "bittest.byte"); |
| 1026 | Value *NewByte = nullptr; |
| 1027 | switch (BT.Action) { |
| 1028 | case BitTest::TestOnly: |
| 1029 | // Don't store anything. |
| 1030 | break; |
| 1031 | case BitTest::Complement: |
| 1032 | NewByte = CGF.Builder.CreateXor(OldByte, Mask); |
| 1033 | break; |
| 1034 | case BitTest::Reset: |
| 1035 | NewByte = CGF.Builder.CreateAnd(OldByte, CGF.Builder.CreateNot(Mask)); |
| 1036 | break; |
| 1037 | case BitTest::Set: |
| 1038 | NewByte = CGF.Builder.CreateOr(OldByte, Mask); |
| 1039 | break; |
| 1040 | } |
| 1041 | if (NewByte) |
| 1042 | CGF.Builder.CreateStore(NewByte, ByteAddr); |
| 1043 | } |
| 1044 | |
| 1045 | // However we loaded the old byte, either by plain load or atomicrmw, shift |
| 1046 | // the bit into the low position and mask it to 0 or 1. |
| 1047 | Value *ShiftedByte = CGF.Builder.CreateLShr(OldByte, PosLow, "bittest.shr"); |
| 1048 | return CGF.Builder.CreateAnd( |
| 1049 | ShiftedByte, llvm::ConstantInt::get(CGF.Int8Ty, 1), "bittest.res"); |
| 1050 | } |
| 1051 | |
| 1052 | static llvm::Value *emitPPCLoadReserveIntrinsic(CodeGenFunction &CGF, |
| 1053 | unsigned BuiltinID, |
| 1054 | const CallExpr *E) { |
| 1055 | Value *Addr = CGF.EmitScalarExpr(E->getArg(0)); |
| 1056 | |
| 1057 | SmallString<64> Asm; |
| 1058 | raw_svector_ostream AsmOS(Asm); |
| 1059 | llvm::IntegerType *RetType = CGF.Int32Ty; |
Value stored to 'RetType' during its initialization is never read | |
| 1060 | |
| 1061 | switch (BuiltinID) { |
| 1062 | case clang::PPC::BI__builtin_ppc_ldarx: |
| 1063 | AsmOS << "ldarx "; |
| 1064 | RetType = CGF.Int64Ty; |
| 1065 | break; |
| 1066 | case clang::PPC::BI__builtin_ppc_lwarx: |
| 1067 | AsmOS << "lwarx "; |
| 1068 | RetType = CGF.Int32Ty; |
| 1069 | break; |
| 1070 | case clang::PPC::BI__builtin_ppc_lharx: |
| 1071 | AsmOS << "lharx "; |
| 1072 | RetType = CGF.Int16Ty; |
| 1073 | break; |
| 1074 | case clang::PPC::BI__builtin_ppc_lbarx: |
| 1075 | AsmOS << "lbarx "; |
| 1076 | RetType = CGF.Int8Ty; |
| 1077 | break; |
| 1078 | default: |
| 1079 | llvm_unreachable("Expected only PowerPC load reserve intrinsics")::llvm::llvm_unreachable_internal("Expected only PowerPC load reserve intrinsics" , "clang/lib/CodeGen/CGBuiltin.cpp", 1079); |
| 1080 | } |
| 1081 | |
| 1082 | AsmOS << "$0, ${1:y}"; |
| 1083 | |
| 1084 | std::string Constraints = "=r,*Z,~{memory}"; |
| 1085 | std::string_view MachineClobbers = CGF.getTarget().getClobbers(); |
| 1086 | if (!MachineClobbers.empty()) { |
| 1087 | Constraints += ','; |
| 1088 | Constraints += MachineClobbers; |
| 1089 | } |
| 1090 | |
| 1091 | llvm::Type *IntPtrType = RetType->getPointerTo(); |
| 1092 | llvm::FunctionType *FTy = |
| 1093 | llvm::FunctionType::get(RetType, {IntPtrType}, false); |
| 1094 | |
| 1095 | llvm::InlineAsm *IA = |
| 1096 | llvm::InlineAsm::get(FTy, Asm, Constraints, /*hasSideEffects=*/true); |
| 1097 | llvm::CallInst *CI = CGF.Builder.CreateCall(IA, {Addr}); |
| 1098 | CI->addParamAttr( |
| 1099 | 0, Attribute::get(CGF.getLLVMContext(), Attribute::ElementType, RetType)); |
| 1100 | return CI; |
| 1101 | } |
| 1102 | |
| 1103 | namespace { |
| 1104 | enum class MSVCSetJmpKind { |
| 1105 | _setjmpex, |
| 1106 | _setjmp3, |
| 1107 | _setjmp |
| 1108 | }; |
| 1109 | } |
| 1110 | |
| 1111 | /// MSVC handles setjmp a bit differently on different platforms. On every |
| 1112 | /// architecture except 32-bit x86, the frame address is passed. On x86, extra |
| 1113 | /// parameters can be passed as variadic arguments, but we always pass none. |
| 1114 | static RValue EmitMSVCRTSetJmp(CodeGenFunction &CGF, MSVCSetJmpKind SJKind, |
| 1115 | const CallExpr *E) { |
| 1116 | llvm::Value *Arg1 = nullptr; |
| 1117 | llvm::Type *Arg1Ty = nullptr; |
| 1118 | StringRef Name; |
| 1119 | bool IsVarArg = false; |
| 1120 | if (SJKind == MSVCSetJmpKind::_setjmp3) { |
| 1121 | Name = "_setjmp3"; |
| 1122 | Arg1Ty = CGF.Int32Ty; |
| 1123 | Arg1 = llvm::ConstantInt::get(CGF.IntTy, 0); |
| 1124 | IsVarArg = true; |
| 1125 | } else { |
| 1126 | Name = SJKind == MSVCSetJmpKind::_setjmp ? "_setjmp" : "_setjmpex"; |
| 1127 | Arg1Ty = CGF.Int8PtrTy; |
| 1128 | if (CGF.getTarget().getTriple().getArch() == llvm::Triple::aarch64) { |
| 1129 | Arg1 = CGF.Builder.CreateCall( |
| 1130 | CGF.CGM.getIntrinsic(Intrinsic::sponentry, CGF.AllocaInt8PtrTy)); |
| 1131 | } else |
| 1132 | Arg1 = CGF.Builder.CreateCall( |
| 1133 | CGF.CGM.getIntrinsic(Intrinsic::frameaddress, CGF.AllocaInt8PtrTy), |
| 1134 | llvm::ConstantInt::get(CGF.Int32Ty, 0)); |
| 1135 | } |
| 1136 | |
| 1137 | // Mark the call site and declaration with ReturnsTwice. |
| 1138 | llvm::Type *ArgTypes[2] = {CGF.Int8PtrTy, Arg1Ty}; |
| 1139 | llvm::AttributeList ReturnsTwiceAttr = llvm::AttributeList::get( |
| 1140 | CGF.getLLVMContext(), llvm::AttributeList::FunctionIndex, |
| 1141 | llvm::Attribute::ReturnsTwice); |
| 1142 | llvm::FunctionCallee SetJmpFn = CGF.CGM.CreateRuntimeFunction( |
| 1143 | llvm::FunctionType::get(CGF.IntTy, ArgTypes, IsVarArg), Name, |
| 1144 | ReturnsTwiceAttr, /*Local=*/true); |
| 1145 | |
| 1146 | llvm::Value *Buf = CGF.Builder.CreateBitOrPointerCast( |
| 1147 | CGF.EmitScalarExpr(E->getArg(0)), CGF.Int8PtrTy); |
| 1148 | llvm::Value *Args[] = {Buf, Arg1}; |
| 1149 | llvm::CallBase *CB = CGF.EmitRuntimeCallOrInvoke(SetJmpFn, Args); |
| 1150 | CB->setAttributes(ReturnsTwiceAttr); |
| 1151 | return RValue::get(CB); |
| 1152 | } |
| 1153 | |
| 1154 | // Many of MSVC builtins are on x64, ARM and AArch64; to avoid repeating code, |
| 1155 | // we handle them here. |
| 1156 | enum class CodeGenFunction::MSVCIntrin { |
| 1157 | _BitScanForward, |
| 1158 | _BitScanReverse, |
| 1159 | _InterlockedAnd, |
| 1160 | _InterlockedDecrement, |
| 1161 | _InterlockedExchange, |
| 1162 | _InterlockedExchangeAdd, |
| 1163 | _InterlockedExchangeSub, |
| 1164 | _InterlockedIncrement, |
| 1165 | _InterlockedOr, |
| 1166 | _InterlockedXor, |
| 1167 | _InterlockedExchangeAdd_acq, |
| 1168 | _InterlockedExchangeAdd_rel, |
| 1169 | _InterlockedExchangeAdd_nf, |
| 1170 | _InterlockedExchange_acq, |
| 1171 | _InterlockedExchange_rel, |
| 1172 | _InterlockedExchange_nf, |
| 1173 | _InterlockedCompareExchange_acq, |
| 1174 | _InterlockedCompareExchange_rel, |
| 1175 | _InterlockedCompareExchange_nf, |
| 1176 | _InterlockedCompareExchange128, |
| 1177 | _InterlockedCompareExchange128_acq, |
| 1178 | _InterlockedCompareExchange128_rel, |
| 1179 | _InterlockedCompareExchange128_nf, |
| 1180 | _InterlockedOr_acq, |
| 1181 | _InterlockedOr_rel, |
| 1182 | _InterlockedOr_nf, |
| 1183 | _InterlockedXor_acq, |
| 1184 | _InterlockedXor_rel, |
| 1185 | _InterlockedXor_nf, |
| 1186 | _InterlockedAnd_acq, |
| 1187 | _InterlockedAnd_rel, |
| 1188 | _InterlockedAnd_nf, |
| 1189 | _InterlockedIncrement_acq, |
| 1190 | _InterlockedIncrement_rel, |
| 1191 | _InterlockedIncrement_nf, |
| 1192 | _InterlockedDecrement_acq, |
| 1193 | _InterlockedDecrement_rel, |
| 1194 | _InterlockedDecrement_nf, |
| 1195 | __fastfail, |
| 1196 | }; |
| 1197 | |
| 1198 | static std::optional<CodeGenFunction::MSVCIntrin> |
| 1199 | translateArmToMsvcIntrin(unsigned BuiltinID) { |
| 1200 | using MSVCIntrin = CodeGenFunction::MSVCIntrin; |
| 1201 | switch (BuiltinID) { |
| 1202 | default: |
| 1203 | return std::nullopt; |
| 1204 | case clang::ARM::BI_BitScanForward: |
| 1205 | case clang::ARM::BI_BitScanForward64: |
| 1206 | return MSVCIntrin::_BitScanForward; |
| 1207 | case clang::ARM::BI_BitScanReverse: |
| 1208 | case clang::ARM::BI_BitScanReverse64: |
| 1209 | return MSVCIntrin::_BitScanReverse; |
| 1210 | case clang::ARM::BI_InterlockedAnd64: |
| 1211 | return MSVCIntrin::_InterlockedAnd; |
| 1212 | case clang::ARM::BI_InterlockedExchange64: |
| 1213 | return MSVCIntrin::_InterlockedExchange; |
| 1214 | case clang::ARM::BI_InterlockedExchangeAdd64: |
| 1215 | return MSVCIntrin::_InterlockedExchangeAdd; |
| 1216 | case clang::ARM::BI_InterlockedExchangeSub64: |
| 1217 | return MSVCIntrin::_InterlockedExchangeSub; |
| 1218 | case clang::ARM::BI_InterlockedOr64: |
| 1219 | return MSVCIntrin::_InterlockedOr; |
| 1220 | case clang::ARM::BI_InterlockedXor64: |
| 1221 | return MSVCIntrin::_InterlockedXor; |
| 1222 | case clang::ARM::BI_InterlockedDecrement64: |
| 1223 | return MSVCIntrin::_InterlockedDecrement; |
| 1224 | case clang::ARM::BI_InterlockedIncrement64: |
| 1225 | return MSVCIntrin::_InterlockedIncrement; |
| 1226 | case clang::ARM::BI_InterlockedExchangeAdd8_acq: |
| 1227 | case clang::ARM::BI_InterlockedExchangeAdd16_acq: |
| 1228 | case clang::ARM::BI_InterlockedExchangeAdd_acq: |
| 1229 | case clang::ARM::BI_InterlockedExchangeAdd64_acq: |
| 1230 | return MSVCIntrin::_InterlockedExchangeAdd_acq; |
| 1231 | case clang::ARM::BI_InterlockedExchangeAdd8_rel: |
| 1232 | case clang::ARM::BI_InterlockedExchangeAdd16_rel: |
| 1233 | case clang::ARM::BI_InterlockedExchangeAdd_rel: |
| 1234 | case clang::ARM::BI_InterlockedExchangeAdd64_rel: |
| 1235 | return MSVCIntrin::_InterlockedExchangeAdd_rel; |
| 1236 | case clang::ARM::BI_InterlockedExchangeAdd8_nf: |
| 1237 | case clang::ARM::BI_InterlockedExchangeAdd16_nf: |
| 1238 | case clang::ARM::BI_InterlockedExchangeAdd_nf: |
| 1239 | case clang::ARM::BI_InterlockedExchangeAdd64_nf: |
| 1240 | return MSVCIntrin::_InterlockedExchangeAdd_nf; |
| 1241 | case clang::ARM::BI_InterlockedExchange8_acq: |
| 1242 | case clang::ARM::BI_InterlockedExchange16_acq: |
| 1243 | case clang::ARM::BI_InterlockedExchange_acq: |
| 1244 | case clang::ARM::BI_InterlockedExchange64_acq: |
| 1245 | return MSVCIntrin::_InterlockedExchange_acq; |
| 1246 | case clang::ARM::BI_InterlockedExchange8_rel: |
| 1247 | case clang::ARM::BI_InterlockedExchange16_rel: |
| 1248 | case clang::ARM::BI_InterlockedExchange_rel: |
| 1249 | case clang::ARM::BI_InterlockedExchange64_rel: |
| 1250 | return MSVCIntrin::_InterlockedExchange_rel; |
| 1251 | case clang::ARM::BI_InterlockedExchange8_nf: |
| 1252 | case clang::ARM::BI_InterlockedExchange16_nf: |
| 1253 | case clang::ARM::BI_InterlockedExchange_nf: |
| 1254 | case clang::ARM::BI_InterlockedExchange64_nf: |
| 1255 | return MSVCIntrin::_InterlockedExchange_nf; |
| 1256 | case clang::ARM::BI_InterlockedCompareExchange8_acq: |
| 1257 | case clang::ARM::BI_InterlockedCompareExchange16_acq: |
| 1258 | case clang::ARM::BI_InterlockedCompareExchange_acq: |
| 1259 | case clang::ARM::BI_InterlockedCompareExchange64_acq: |
| 1260 | return MSVCIntrin::_InterlockedCompareExchange_acq; |
| 1261 | case clang::ARM::BI_InterlockedCompareExchange8_rel: |
| 1262 | case clang::ARM::BI_InterlockedCompareExchange16_rel: |
| 1263 | case clang::ARM::BI_InterlockedCompareExchange_rel: |
| 1264 | case clang::ARM::BI_InterlockedCompareExchange64_rel: |
| 1265 | return MSVCIntrin::_InterlockedCompareExchange_rel; |
| 1266 | case clang::ARM::BI_InterlockedCompareExchange8_nf: |
| 1267 | case clang::ARM::BI_InterlockedCompareExchange16_nf: |
| 1268 | case clang::ARM::BI_InterlockedCompareExchange_nf: |
| 1269 | case clang::ARM::BI_InterlockedCompareExchange64_nf: |
| 1270 | return MSVCIntrin::_InterlockedCompareExchange_nf; |
| 1271 | case clang::ARM::BI_InterlockedOr8_acq: |
| 1272 | case clang::ARM::BI_InterlockedOr16_acq: |
| 1273 | case clang::ARM::BI_InterlockedOr_acq: |
| 1274 | case clang::ARM::BI_InterlockedOr64_acq: |
| 1275 | return MSVCIntrin::_InterlockedOr_acq; |
| 1276 | case clang::ARM::BI_InterlockedOr8_rel: |
| 1277 | case clang::ARM::BI_InterlockedOr16_rel: |
| 1278 | case clang::ARM::BI_InterlockedOr_rel: |
| 1279 | case clang::ARM::BI_InterlockedOr64_rel: |
| 1280 | return MSVCIntrin::_InterlockedOr_rel; |
| 1281 | case clang::ARM::BI_InterlockedOr8_nf: |
| 1282 | case clang::ARM::BI_InterlockedOr16_nf: |
| 1283 | case clang::ARM::BI_InterlockedOr_nf: |
| 1284 | case clang::ARM::BI_InterlockedOr64_nf: |
| 1285 | return MSVCIntrin::_InterlockedOr_nf; |
| 1286 | case clang::ARM::BI_InterlockedXor8_acq: |
| 1287 | case clang::ARM::BI_InterlockedXor16_acq: |
| 1288 | case clang::ARM::BI_InterlockedXor_acq: |
| 1289 | case clang::ARM::BI_InterlockedXor64_acq: |
| 1290 | return MSVCIntrin::_InterlockedXor_acq; |
| 1291 | case clang::ARM::BI_InterlockedXor8_rel: |
| 1292 | case clang::ARM::BI_InterlockedXor16_rel: |
| 1293 | case clang::ARM::BI_InterlockedXor_rel: |
| 1294 | case clang::ARM::BI_InterlockedXor64_rel: |
| 1295 | return MSVCIntrin::_InterlockedXor_rel; |
| 1296 | case clang::ARM::BI_InterlockedXor8_nf: |
| 1297 | case clang::ARM::BI_InterlockedXor16_nf: |
| 1298 | case clang::ARM::BI_InterlockedXor_nf: |
| 1299 | case clang::ARM::BI_InterlockedXor64_nf: |
| 1300 | return MSVCIntrin::_InterlockedXor_nf; |
| 1301 | case clang::ARM::BI_InterlockedAnd8_acq: |
| 1302 | case clang::ARM::BI_InterlockedAnd16_acq: |
| 1303 | case clang::ARM::BI_InterlockedAnd_acq: |
| 1304 | case clang::ARM::BI_InterlockedAnd64_acq: |
| 1305 | return MSVCIntrin::_InterlockedAnd_acq; |
| 1306 | case clang::ARM::BI_InterlockedAnd8_rel: |
| 1307 | case clang::ARM::BI_InterlockedAnd16_rel: |
| 1308 | case clang::ARM::BI_InterlockedAnd_rel: |
| 1309 | case clang::ARM::BI_InterlockedAnd64_rel: |
| 1310 | return MSVCIntrin::_InterlockedAnd_rel; |
| 1311 | case clang::ARM::BI_InterlockedAnd8_nf: |
| 1312 | case clang::ARM::BI_InterlockedAnd16_nf: |
| 1313 | case clang::ARM::BI_InterlockedAnd_nf: |
| 1314 | case clang::ARM::BI_InterlockedAnd64_nf: |
| 1315 | return MSVCIntrin::_InterlockedAnd_nf; |
| 1316 | case clang::ARM::BI_InterlockedIncrement16_acq: |
| 1317 | case clang::ARM::BI_InterlockedIncrement_acq: |
| 1318 | case clang::ARM::BI_InterlockedIncrement64_acq: |
| 1319 | return MSVCIntrin::_InterlockedIncrement_acq; |
| 1320 | case clang::ARM::BI_InterlockedIncrement16_rel: |
| 1321 | case clang::ARM::BI_InterlockedIncrement_rel: |
| 1322 | case clang::ARM::BI_InterlockedIncrement64_rel: |
| 1323 | return MSVCIntrin::_InterlockedIncrement_rel; |
| 1324 | case clang::ARM::BI_InterlockedIncrement16_nf: |
| 1325 | case clang::ARM::BI_InterlockedIncrement_nf: |
| 1326 | case clang::ARM::BI_InterlockedIncrement64_nf: |
| 1327 | return MSVCIntrin::_InterlockedIncrement_nf; |
| 1328 | case clang::ARM::BI_InterlockedDecrement16_acq: |
| 1329 | case clang::ARM::BI_InterlockedDecrement_acq: |
| 1330 | case clang::ARM::BI_InterlockedDecrement64_acq: |
| 1331 | return MSVCIntrin::_InterlockedDecrement_acq; |
| 1332 | case clang::ARM::BI_InterlockedDecrement16_rel: |
| 1333 | case clang::ARM::BI_InterlockedDecrement_rel: |
| 1334 | case clang::ARM::BI_InterlockedDecrement64_rel: |
| 1335 | return MSVCIntrin::_InterlockedDecrement_rel; |
| 1336 | case clang::ARM::BI_InterlockedDecrement16_nf: |
| 1337 | case clang::ARM::BI_InterlockedDecrement_nf: |
| 1338 | case clang::ARM::BI_InterlockedDecrement64_nf: |
| 1339 | return MSVCIntrin::_InterlockedDecrement_nf; |
| 1340 | } |
| 1341 | llvm_unreachable("must return from switch")::llvm::llvm_unreachable_internal("must return from switch", "clang/lib/CodeGen/CGBuiltin.cpp" , 1341); |
| 1342 | } |
| 1343 | |
| 1344 | static std::optional<CodeGenFunction::MSVCIntrin> |
| 1345 | translateAarch64ToMsvcIntrin(unsigned BuiltinID) { |
| 1346 | using MSVCIntrin = CodeGenFunction::MSVCIntrin; |
| 1347 | switch (BuiltinID) { |
| 1348 | default: |
| 1349 | return std::nullopt; |
| 1350 | case clang::AArch64::BI_BitScanForward: |
| 1351 | case clang::AArch64::BI_BitScanForward64: |
| 1352 | return MSVCIntrin::_BitScanForward; |
| 1353 | case clang::AArch64::BI_BitScanReverse: |
| 1354 | case clang::AArch64::BI_BitScanReverse64: |
| 1355 | return MSVCIntrin::_BitScanReverse; |
| 1356 | case clang::AArch64::BI_InterlockedAnd64: |
| 1357 | return MSVCIntrin::_InterlockedAnd; |
| 1358 | case clang::AArch64::BI_InterlockedExchange64: |
| 1359 | return MSVCIntrin::_InterlockedExchange; |
| 1360 | case clang::AArch64::BI_InterlockedExchangeAdd64: |
| 1361 | return MSVCIntrin::_InterlockedExchangeAdd; |
| 1362 | case clang::AArch64::BI_InterlockedExchangeSub64: |
| 1363 | return MSVCIntrin::_InterlockedExchangeSub; |
| 1364 | case clang::AArch64::BI_InterlockedOr64: |
| 1365 | return MSVCIntrin::_InterlockedOr; |
| 1366 | case clang::AArch64::BI_InterlockedXor64: |
| 1367 | return MSVCIntrin::_InterlockedXor; |
| 1368 | case clang::AArch64::BI_InterlockedDecrement64: |
| 1369 | return MSVCIntrin::_InterlockedDecrement; |
| 1370 | case clang::AArch64::BI_InterlockedIncrement64: |
| 1371 | return MSVCIntrin::_InterlockedIncrement; |
| 1372 | case clang::AArch64::BI_InterlockedExchangeAdd8_acq: |
| 1373 | case clang::AArch64::BI_InterlockedExchangeAdd16_acq: |
| 1374 | case clang::AArch64::BI_InterlockedExchangeAdd_acq: |
| 1375 | case clang::AArch64::BI_InterlockedExchangeAdd64_acq: |
| 1376 | return MSVCIntrin::_InterlockedExchangeAdd_acq; |
| 1377 | case clang::AArch64::BI_InterlockedExchangeAdd8_rel: |
| 1378 | case clang::AArch64::BI_InterlockedExchangeAdd16_rel: |
| 1379 | case clang::AArch64::BI_InterlockedExchangeAdd_rel: |
| 1380 | case clang::AArch64::BI_InterlockedExchangeAdd64_rel: |
| 1381 | return MSVCIntrin::_InterlockedExchangeAdd_rel; |
| 1382 | case clang::AArch64::BI_InterlockedExchangeAdd8_nf: |
| 1383 | case clang::AArch64::BI_InterlockedExchangeAdd16_nf: |
| 1384 | case clang::AArch64::BI_InterlockedExchangeAdd_nf: |
| 1385 | case clang::AArch64::BI_InterlockedExchangeAdd64_nf: |
| 1386 | return MSVCIntrin::_InterlockedExchangeAdd_nf; |
| 1387 | case clang::AArch64::BI_InterlockedExchange8_acq: |
| 1388 | case clang::AArch64::BI_InterlockedExchange16_acq: |
| 1389 | case clang::AArch64::BI_InterlockedExchange_acq: |
| 1390 | case clang::AArch64::BI_InterlockedExchange64_acq: |
| 1391 | return MSVCIntrin::_InterlockedExchange_acq; |
| 1392 | case clang::AArch64::BI_InterlockedExchange8_rel: |
| 1393 | case clang::AArch64::BI_InterlockedExchange16_rel: |
| 1394 | case clang::AArch64::BI_InterlockedExchange_rel: |
| 1395 | case clang::AArch64::BI_InterlockedExchange64_rel: |
| 1396 | return MSVCIntrin::_InterlockedExchange_rel; |
| 1397 | case clang::AArch64::BI_InterlockedExchange8_nf: |
| 1398 | case clang::AArch64::BI_InterlockedExchange16_nf: |
| 1399 | case clang::AArch64::BI_InterlockedExchange_nf: |
| 1400 | case clang::AArch64::BI_InterlockedExchange64_nf: |
| 1401 | return MSVCIntrin::_InterlockedExchange_nf; |
| 1402 | case clang::AArch64::BI_InterlockedCompareExchange8_acq: |
| 1403 | case clang::AArch64::BI_InterlockedCompareExchange16_acq: |
| 1404 | case clang::AArch64::BI_InterlockedCompareExchange_acq: |
| 1405 | case clang::AArch64::BI_InterlockedCompareExchange64_acq: |
| 1406 | return MSVCIntrin::_InterlockedCompareExchange_acq; |
| 1407 | case clang::AArch64::BI_InterlockedCompareExchange8_rel: |
| 1408 | case clang::AArch64::BI_InterlockedCompareExchange16_rel: |
| 1409 | case clang::AArch64::BI_InterlockedCompareExchange_rel: |
| 1410 | case clang::AArch64::BI_InterlockedCompareExchange64_rel: |
| 1411 | return MSVCIntrin::_InterlockedCompareExchange_rel; |
| 1412 | case clang::AArch64::BI_InterlockedCompareExchange8_nf: |
| 1413 | case clang::AArch64::BI_InterlockedCompareExchange16_nf: |
| 1414 | case clang::AArch64::BI_InterlockedCompareExchange_nf: |
| 1415 | case clang::AArch64::BI_InterlockedCompareExchange64_nf: |
| 1416 | return MSVCIntrin::_InterlockedCompareExchange_nf; |
| 1417 | case clang::AArch64::BI_InterlockedCompareExchange128: |
| 1418 | return MSVCIntrin::_InterlockedCompareExchange128; |
| 1419 | case clang::AArch64::BI_InterlockedCompareExchange128_acq: |
| 1420 | return MSVCIntrin::_InterlockedCompareExchange128_acq; |
| 1421 | case clang::AArch64::BI_InterlockedCompareExchange128_nf: |
| 1422 | return MSVCIntrin::_InterlockedCompareExchange128_nf; |
| 1423 | case clang::AArch64::BI_InterlockedCompareExchange128_rel: |
| 1424 | return MSVCIntrin::_InterlockedCompareExchange128_rel; |
| 1425 | case clang::AArch64::BI_InterlockedOr8_acq: |
| 1426 | case clang::AArch64::BI_InterlockedOr16_acq: |
| 1427 | case clang::AArch64::BI_InterlockedOr_acq: |
| 1428 | case clang::AArch64::BI_InterlockedOr64_acq: |
| 1429 | return MSVCIntrin::_InterlockedOr_acq; |
| 1430 | case clang::AArch64::BI_InterlockedOr8_rel: |
| 1431 | case clang::AArch64::BI_InterlockedOr16_rel: |
| 1432 | case clang::AArch64::BI_InterlockedOr_rel: |
| 1433 | case clang::AArch64::BI_InterlockedOr64_rel: |
| 1434 | return MSVCIntrin::_InterlockedOr_rel; |
| 1435 | case clang::AArch64::BI_InterlockedOr8_nf: |
| 1436 | case clang::AArch64::BI_InterlockedOr16_nf: |
| 1437 | case clang::AArch64::BI_InterlockedOr_nf: |
| 1438 | case clang::AArch64::BI_InterlockedOr64_nf: |
| 1439 | return MSVCIntrin::_InterlockedOr_nf; |
| 1440 | case clang::AArch64::BI_InterlockedXor8_acq: |
| 1441 | case clang::AArch64::BI_InterlockedXor16_acq: |
| 1442 | case clang::AArch64::BI_InterlockedXor_acq: |
| 1443 | case clang::AArch64::BI_InterlockedXor64_acq: |
| 1444 | return MSVCIntrin::_InterlockedXor_acq; |
| 1445 | case clang::AArch64::BI_InterlockedXor8_rel: |
| 1446 | case clang::AArch64::BI_InterlockedXor16_rel: |
| 1447 | case clang::AArch64::BI_InterlockedXor_rel: |
| 1448 | case clang::AArch64::BI_InterlockedXor64_rel: |
| 1449 | return MSVCIntrin::_InterlockedXor_rel; |
| 1450 | case clang::AArch64::BI_InterlockedXor8_nf: |
| 1451 | case clang::AArch64::BI_InterlockedXor16_nf: |
| 1452 | case clang::AArch64::BI_InterlockedXor_nf: |
| 1453 | case clang::AArch64::BI_InterlockedXor64_nf: |
| 1454 | return MSVCIntrin::_InterlockedXor_nf; |
| 1455 | case clang::AArch64::BI_InterlockedAnd8_acq: |
| 1456 | case clang::AArch64::BI_InterlockedAnd16_acq: |
| 1457 | case clang::AArch64::BI_InterlockedAnd_acq: |
| 1458 | case clang::AArch64::BI_InterlockedAnd64_acq: |
| 1459 | return MSVCIntrin::_InterlockedAnd_acq; |
| 1460 | case clang::AArch64::BI_InterlockedAnd8_rel: |
| 1461 | case clang::AArch64::BI_InterlockedAnd16_rel: |
| 1462 | case clang::AArch64::BI_InterlockedAnd_rel: |
| 1463 | case clang::AArch64::BI_InterlockedAnd64_rel: |
| 1464 | return MSVCIntrin::_InterlockedAnd_rel; |
| 1465 | case clang::AArch64::BI_InterlockedAnd8_nf: |
| 1466 | case clang::AArch64::BI_InterlockedAnd16_nf: |
| 1467 | case clang::AArch64::BI_InterlockedAnd_nf: |
| 1468 | case clang::AArch64::BI_InterlockedAnd64_nf: |
| 1469 | return MSVCIntrin::_InterlockedAnd_nf; |
| 1470 | case clang::AArch64::BI_InterlockedIncrement16_acq: |
| 1471 | case clang::AArch64::BI_InterlockedIncrement_acq: |
| 1472 | case clang::AArch64::BI_InterlockedIncrement64_acq: |
| 1473 | return MSVCIntrin::_InterlockedIncrement_acq; |
| 1474 | case clang::AArch64::BI_InterlockedIncrement16_rel: |
| 1475 | case clang::AArch64::BI_InterlockedIncrement_rel: |
| 1476 | case clang::AArch64::BI_InterlockedIncrement64_rel: |
| 1477 | return MSVCIntrin::_InterlockedIncrement_rel; |
| 1478 | case clang::AArch64::BI_InterlockedIncrement16_nf: |
| 1479 | case clang::AArch64::BI_InterlockedIncrement_nf: |
| 1480 | case clang::AArch64::BI_InterlockedIncrement64_nf: |
| 1481 | return MSVCIntrin::_InterlockedIncrement_nf; |
| 1482 | case clang::AArch64::BI_InterlockedDecrement16_acq: |
| 1483 | case clang::AArch64::BI_InterlockedDecrement_acq: |
| 1484 | case clang::AArch64::BI_InterlockedDecrement64_acq: |
| 1485 | return MSVCIntrin::_InterlockedDecrement_acq; |
| 1486 | case clang::AArch64::BI_InterlockedDecrement16_rel: |
| 1487 | case clang::AArch64::BI_InterlockedDecrement_rel: |
| 1488 | case clang::AArch64::BI_InterlockedDecrement64_rel: |
| 1489 | return MSVCIntrin::_InterlockedDecrement_rel; |
| 1490 | case clang::AArch64::BI_InterlockedDecrement16_nf: |
| 1491 | case clang::AArch64::BI_InterlockedDecrement_nf: |
| 1492 | case clang::AArch64::BI_InterlockedDecrement64_nf: |
| 1493 | return MSVCIntrin::_InterlockedDecrement_nf; |
| 1494 | } |
| 1495 | llvm_unreachable("must return from switch")::llvm::llvm_unreachable_internal("must return from switch", "clang/lib/CodeGen/CGBuiltin.cpp" , 1495); |
| 1496 | } |
| 1497 | |
| 1498 | static std::optional<CodeGenFunction::MSVCIntrin> |
| 1499 | translateX86ToMsvcIntrin(unsigned BuiltinID) { |
| 1500 | using MSVCIntrin = CodeGenFunction::MSVCIntrin; |
| 1501 | switch (BuiltinID) { |
| 1502 | default: |
| 1503 | return std::nullopt; |
| 1504 | case clang::X86::BI_BitScanForward: |
| 1505 | case clang::X86::BI_BitScanForward64: |
| 1506 | return MSVCIntrin::_BitScanForward; |
| 1507 | case clang::X86::BI_BitScanReverse: |
| 1508 | case clang::X86::BI_BitScanReverse64: |
| 1509 | return MSVCIntrin::_BitScanReverse; |
| 1510 | case clang::X86::BI_InterlockedAnd64: |
| 1511 | return MSVCIntrin::_InterlockedAnd; |
| 1512 | case clang::X86::BI_InterlockedCompareExchange128: |
| 1513 | return MSVCIntrin::_InterlockedCompareExchange128; |
| 1514 | case clang::X86::BI_InterlockedExchange64: |
| 1515 | return MSVCIntrin::_InterlockedExchange; |
| 1516 | case clang::X86::BI_InterlockedExchangeAdd64: |
| 1517 | return MSVCIntrin::_InterlockedExchangeAdd; |
| 1518 | case clang::X86::BI_InterlockedExchangeSub64: |
| 1519 | return MSVCIntrin::_InterlockedExchangeSub; |
| 1520 | case clang::X86::BI_InterlockedOr64: |
| 1521 | return MSVCIntrin::_InterlockedOr; |
| 1522 | case clang::X86::BI_InterlockedXor64: |
| 1523 | return MSVCIntrin::_InterlockedXor; |
| 1524 | case clang::X86::BI_InterlockedDecrement64: |
| 1525 | return MSVCIntrin::_InterlockedDecrement; |
| 1526 | case clang::X86::BI_InterlockedIncrement64: |
| 1527 | return MSVCIntrin::_InterlockedIncrement; |
| 1528 | } |
| 1529 | llvm_unreachable("must return from switch")::llvm::llvm_unreachable_internal("must return from switch", "clang/lib/CodeGen/CGBuiltin.cpp" , 1529); |
| 1530 | } |
| 1531 | |
| 1532 | // Emit an MSVC intrinsic. Assumes that arguments have *not* been evaluated. |
| 1533 | Value *CodeGenFunction::EmitMSVCBuiltinExpr(MSVCIntrin BuiltinID, |
| 1534 | const CallExpr *E) { |
| 1535 | switch (BuiltinID) { |
| 1536 | case MSVCIntrin::_BitScanForward: |
| 1537 | case MSVCIntrin::_BitScanReverse: { |
| 1538 | Address IndexAddress(EmitPointerWithAlignment(E->getArg(0))); |
| 1539 | Value *ArgValue = EmitScalarExpr(E->getArg(1)); |
| 1540 | |
| 1541 | llvm::Type *ArgType = ArgValue->getType(); |
| 1542 | llvm::Type *IndexType = IndexAddress.getElementType(); |
| 1543 | llvm::Type *ResultType = ConvertType(E->getType()); |
| 1544 | |
| 1545 | Value *ArgZero = llvm::Constant::getNullValue(ArgType); |
| 1546 | Value *ResZero = llvm::Constant::getNullValue(ResultType); |
| 1547 | Value *ResOne = llvm::ConstantInt::get(ResultType, 1); |
| 1548 | |
| 1549 | BasicBlock *Begin = Builder.GetInsertBlock(); |
| 1550 | BasicBlock *End = createBasicBlock("bitscan_end", this->CurFn); |
| 1551 | Builder.SetInsertPoint(End); |
| 1552 | PHINode *Result = Builder.CreatePHI(ResultType, 2, "bitscan_result"); |
| 1553 | |
| 1554 | Builder.SetInsertPoint(Begin); |
| 1555 | Value *IsZero = Builder.CreateICmpEQ(ArgValue, ArgZero); |
| 1556 | BasicBlock *NotZero = createBasicBlock("bitscan_not_zero", this->CurFn); |
| 1557 | Builder.CreateCondBr(IsZero, End, NotZero); |
| 1558 | Result->addIncoming(ResZero, Begin); |
| 1559 | |
| 1560 | Builder.SetInsertPoint(NotZero); |
| 1561 | |
| 1562 | if (BuiltinID == MSVCIntrin::_BitScanForward) { |
| 1563 | Function *F = CGM.getIntrinsic(Intrinsic::cttz, ArgType); |
| 1564 | Value *ZeroCount = Builder.CreateCall(F, {ArgValue, Builder.getTrue()}); |
| 1565 | ZeroCount = Builder.CreateIntCast(ZeroCount, IndexType, false); |
| 1566 | Builder.CreateStore(ZeroCount, IndexAddress, false); |
| 1567 | } else { |
| 1568 | unsigned ArgWidth = cast<llvm::IntegerType>(ArgType)->getBitWidth(); |
| 1569 | Value *ArgTypeLastIndex = llvm::ConstantInt::get(IndexType, ArgWidth - 1); |
| 1570 | |
| 1571 | Function *F = CGM.getIntrinsic(Intrinsic::ctlz, ArgType); |
| 1572 | Value *ZeroCount = Builder.CreateCall(F, {ArgValue, Builder.getTrue()}); |
| 1573 | ZeroCount = Builder.CreateIntCast(ZeroCount, IndexType, false); |
| 1574 | Value *Index = Builder.CreateNSWSub(ArgTypeLastIndex, ZeroCount); |
| 1575 | Builder.CreateStore(Index, IndexAddress, false); |
| 1576 | } |
| 1577 | Builder.CreateBr(End); |
| 1578 | Result->addIncoming(ResOne, NotZero); |
| 1579 | |
| 1580 | Builder.SetInsertPoint(End); |
| 1581 | return Result; |
| 1582 | } |
| 1583 | case MSVCIntrin::_InterlockedAnd: |
| 1584 | return MakeBinaryAtomicValue(*this, AtomicRMWInst::And, E); |
| 1585 | case MSVCIntrin::_InterlockedExchange: |
| 1586 | return MakeBinaryAtomicValue(*this, AtomicRMWInst::Xchg, E); |
| 1587 | case MSVCIntrin::_InterlockedExchangeAdd: |
| 1588 | return MakeBinaryAtomicValue(*this, AtomicRMWInst::Add, E); |
| 1589 | case MSVCIntrin::_InterlockedExchangeSub: |
| 1590 | return MakeBinaryAtomicValue(*this, AtomicRMWInst::Sub, E); |
| 1591 | case MSVCIntrin::_InterlockedOr: |
| 1592 | return MakeBinaryAtomicValue(*this, AtomicRMWInst::Or, E); |
| 1593 | case MSVCIntrin::_InterlockedXor: |
| 1594 | return MakeBinaryAtomicValue(*this, AtomicRMWInst::Xor, E); |
| 1595 | case MSVCIntrin::_InterlockedExchangeAdd_acq: |
| 1596 | return MakeBinaryAtomicValue(*this, AtomicRMWInst::Add, E, |
| 1597 | AtomicOrdering::Acquire); |
| 1598 | case MSVCIntrin::_InterlockedExchangeAdd_rel: |
| 1599 | return MakeBinaryAtomicValue(*this, AtomicRMWInst::Add, E, |
| 1600 | AtomicOrdering::Release); |
| 1601 | case MSVCIntrin::_InterlockedExchangeAdd_nf: |
| 1602 | return MakeBinaryAtomicValue(*this, AtomicRMWInst::Add, E, |
| 1603 | AtomicOrdering::Monotonic); |
| 1604 | case MSVCIntrin::_InterlockedExchange_acq: |
| 1605 | return MakeBinaryAtomicValue(*this, AtomicRMWInst::Xchg, E, |
| 1606 | AtomicOrdering::Acquire); |
| 1607 | case MSVCIntrin::_InterlockedExchange_rel: |
| 1608 | return MakeBinaryAtomicValue(*this, AtomicRMWInst::Xchg, E, |
| 1609 | AtomicOrdering::Release); |
| 1610 | case MSVCIntrin::_InterlockedExchange_nf: |
| 1611 | return MakeBinaryAtomicValue(*this, AtomicRMWInst::Xchg, E, |
| 1612 | AtomicOrdering::Monotonic); |
| 1613 | case MSVCIntrin::_InterlockedCompareExchange_acq: |
| 1614 | return EmitAtomicCmpXchgForMSIntrin(*this, E, AtomicOrdering::Acquire); |
| 1615 | case MSVCIntrin::_InterlockedCompareExchange_rel: |
| 1616 | return EmitAtomicCmpXchgForMSIntrin(*this, E, AtomicOrdering::Release); |
| 1617 | case MSVCIntrin::_InterlockedCompareExchange_nf: |
| 1618 | return EmitAtomicCmpXchgForMSIntrin(*this, E, AtomicOrdering::Monotonic); |
| 1619 | case MSVCIntrin::_InterlockedCompareExchange128: |
| 1620 | return EmitAtomicCmpXchg128ForMSIntrin( |
| 1621 | *this, E, AtomicOrdering::SequentiallyConsistent); |
| 1622 | case MSVCIntrin::_InterlockedCompareExchange128_acq: |
| 1623 | return EmitAtomicCmpXchg128ForMSIntrin(*this, E, AtomicOrdering::Acquire); |
| 1624 | case MSVCIntrin::_InterlockedCompareExchange128_rel: |
| 1625 | return EmitAtomicCmpXchg128ForMSIntrin(*this, E, AtomicOrdering::Release); |
| 1626 | case MSVCIntrin::_InterlockedCompareExchange128_nf: |
| 1627 | return EmitAtomicCmpXchg128ForMSIntrin(*this, E, AtomicOrdering::Monotonic); |
| 1628 | case MSVCIntrin::_InterlockedOr_acq: |
| 1629 | return MakeBinaryAtomicValue(*this, AtomicRMWInst::Or, E, |
| 1630 | AtomicOrdering::Acquire); |
| 1631 | case MSVCIntrin::_InterlockedOr_rel: |
| 1632 | return MakeBinaryAtomicValue(*this, AtomicRMWInst::Or, E, |
| 1633 | AtomicOrdering::Release); |
| 1634 | case MSVCIntrin::_InterlockedOr_nf: |
| 1635 | return MakeBinaryAtomicValue(*this, AtomicRMWInst::Or, E, |
| 1636 | AtomicOrdering::Monotonic); |
| 1637 | case MSVCIntrin::_InterlockedXor_acq: |
| 1638 | return MakeBinaryAtomicValue(*this, AtomicRMWInst::Xor, E, |
| 1639 | AtomicOrdering::Acquire); |
| 1640 | case MSVCIntrin::_InterlockedXor_rel: |
| 1641 | return MakeBinaryAtomicValue(*this, AtomicRMWInst::Xor, E, |
| 1642 | AtomicOrdering::Release); |
| 1643 | case MSVCIntrin::_InterlockedXor_nf: |
| 1644 | return MakeBinaryAtomicValue(*this, AtomicRMWInst::Xor, E, |
| 1645 | AtomicOrdering::Monotonic); |
| 1646 | case MSVCIntrin::_InterlockedAnd_acq: |
| 1647 | return MakeBinaryAtomicValue(*this, AtomicRMWInst::And, E, |
| 1648 | AtomicOrdering::Acquire); |
| 1649 | case MSVCIntrin::_InterlockedAnd_rel: |
| 1650 | return MakeBinaryAtomicValue(*this, AtomicRMWInst::And, E, |
| 1651 | AtomicOrdering::Release); |
| 1652 | case MSVCIntrin::_InterlockedAnd_nf: |
| 1653 | return MakeBinaryAtomicValue(*this, AtomicRMWInst::And, E, |
| 1654 | AtomicOrdering::Monotonic); |
| 1655 | case MSVCIntrin::_InterlockedIncrement_acq: |
| 1656 | return EmitAtomicIncrementValue(*this, E, AtomicOrdering::Acquire); |
| 1657 | case MSVCIntrin::_InterlockedIncrement_rel: |
| 1658 | return EmitAtomicIncrementValue(*this, E, AtomicOrdering::Release); |
| 1659 | case MSVCIntrin::_InterlockedIncrement_nf: |
| 1660 | return EmitAtomicIncrementValue(*this, E, AtomicOrdering::Monotonic); |
| 1661 | case MSVCIntrin::_InterlockedDecrement_acq: |
| 1662 | return EmitAtomicDecrementValue(*this, E, AtomicOrdering::Acquire); |
| 1663 | case MSVCIntrin::_InterlockedDecrement_rel: |
| 1664 | return EmitAtomicDecrementValue(*this, E, AtomicOrdering::Release); |
| 1665 | case MSVCIntrin::_InterlockedDecrement_nf: |
| 1666 | return EmitAtomicDecrementValue(*this, E, AtomicOrdering::Monotonic); |
| 1667 | |
| 1668 | case MSVCIntrin::_InterlockedDecrement: |
| 1669 | return EmitAtomicDecrementValue(*this, E); |
| 1670 | case MSVCIntrin::_InterlockedIncrement: |
| 1671 | return EmitAtomicIncrementValue(*this, E); |
| 1672 | |
| 1673 | case MSVCIntrin::__fastfail: { |
| 1674 | // Request immediate process termination from the kernel. The instruction |
| 1675 | // sequences to do this are documented on MSDN: |
| 1676 | // https://msdn.microsoft.com/en-us/library/dn774154.aspx |
| 1677 | llvm::Triple::ArchType ISA = getTarget().getTriple().getArch(); |
| 1678 | StringRef Asm, Constraints; |
| 1679 | switch (ISA) { |
| 1680 | default: |
| 1681 | ErrorUnsupported(E, "__fastfail call for this architecture"); |
| 1682 | break; |
| 1683 | case llvm::Triple::x86: |
| 1684 | case llvm::Triple::x86_64: |
| 1685 | Asm = "int $$0x29"; |
| 1686 | Constraints = "{cx}"; |
| 1687 | break; |
| 1688 | case llvm::Triple::thumb: |
| 1689 | Asm = "udf #251"; |
| 1690 | Constraints = "{r0}"; |
| 1691 | break; |
| 1692 | case llvm::Triple::aarch64: |
| 1693 | Asm = "brk #0xF003"; |
| 1694 | Constraints = "{w0}"; |
| 1695 | } |
| 1696 | llvm::FunctionType *FTy = llvm::FunctionType::get(VoidTy, {Int32Ty}, false); |
| 1697 | llvm::InlineAsm *IA = |
| 1698 | llvm::InlineAsm::get(FTy, Asm, Constraints, /*hasSideEffects=*/true); |
| 1699 | llvm::AttributeList NoReturnAttr = llvm::AttributeList::get( |
| 1700 | getLLVMContext(), llvm::AttributeList::FunctionIndex, |
| 1701 | llvm::Attribute::NoReturn); |
| 1702 | llvm::CallInst *CI = Builder.CreateCall(IA, EmitScalarExpr(E->getArg(0))); |
| 1703 | CI->setAttributes(NoReturnAttr); |
| 1704 | return CI; |
| 1705 | } |
| 1706 | } |
| 1707 | llvm_unreachable("Incorrect MSVC intrinsic!")::llvm::llvm_unreachable_internal("Incorrect MSVC intrinsic!" , "clang/lib/CodeGen/CGBuiltin.cpp", 1707); |
| 1708 | } |
| 1709 | |
| 1710 | namespace { |
| 1711 | // ARC cleanup for __builtin_os_log_format |
| 1712 | struct CallObjCArcUse final : EHScopeStack::Cleanup { |
| 1713 | CallObjCArcUse(llvm::Value *object) : object(object) {} |
| 1714 | llvm::Value *object; |
| 1715 | |
| 1716 | void Emit(CodeGenFunction &CGF, Flags flags) override { |
| 1717 | CGF.EmitARCIntrinsicUse(object); |
| 1718 | } |
| 1719 | }; |
| 1720 | } |
| 1721 | |
| 1722 | Value *CodeGenFunction::EmitCheckedArgForBuiltin(const Expr *E, |
| 1723 | BuiltinCheckKind Kind) { |
| 1724 | assert((Kind == BCK_CLZPassedZero || Kind == BCK_CTZPassedZero)(static_cast <bool> ((Kind == BCK_CLZPassedZero || Kind == BCK_CTZPassedZero) && "Unsupported builtin check kind" ) ? void (0) : __assert_fail ("(Kind == BCK_CLZPassedZero || Kind == BCK_CTZPassedZero) && \"Unsupported builtin check kind\"" , "clang/lib/CodeGen/CGBuiltin.cpp", 1725, __extension__ __PRETTY_FUNCTION__ )) |
| 1725 | && "Unsupported builtin check kind")(static_cast <bool> ((Kind == BCK_CLZPassedZero || Kind == BCK_CTZPassedZero) && "Unsupported builtin check kind" ) ? void (0) : __assert_fail ("(Kind == BCK_CLZPassedZero || Kind == BCK_CTZPassedZero) && \"Unsupported builtin check kind\"" , "clang/lib/CodeGen/CGBuiltin.cpp", 1725, __extension__ __PRETTY_FUNCTION__ )); |
| 1726 | |
| 1727 | Value *ArgValue = EmitScalarExpr(E); |
| 1728 | if (!SanOpts.has(SanitizerKind::Builtin) || !getTarget().isCLZForZeroUndef()) |
| 1729 | return ArgValue; |
| 1730 | |
| 1731 | SanitizerScope SanScope(this); |
| 1732 | Value *Cond = Builder.CreateICmpNE( |
| 1733 | ArgValue, llvm::Constant::getNullValue(ArgValue->getType())); |
| 1734 | EmitCheck(std::make_pair(Cond, SanitizerKind::Builtin), |
| 1735 | SanitizerHandler::InvalidBuiltin, |
| 1736 | {EmitCheckSourceLocation(E->getExprLoc()), |
| 1737 | llvm::ConstantInt::get(Builder.getInt8Ty(), Kind)}, |
| 1738 | std::nullopt); |
| 1739 | return ArgValue; |
| 1740 | } |
| 1741 | |
| 1742 | /// Get the argument type for arguments to os_log_helper. |
| 1743 | static CanQualType getOSLogArgType(ASTContext &C, int Size) { |
| 1744 | QualType UnsignedTy = C.getIntTypeForBitwidth(Size * 8, /*Signed=*/false); |
| 1745 | return C.getCanonicalType(UnsignedTy); |
| 1746 | } |
| 1747 | |
| 1748 | llvm::Function *CodeGenFunction::generateBuiltinOSLogHelperFunction( |
| 1749 | const analyze_os_log::OSLogBufferLayout &Layout, |
| 1750 | CharUnits BufferAlignment) { |
| 1751 | ASTContext &Ctx = getContext(); |
| 1752 | |
| 1753 | llvm::SmallString<64> Name; |
| 1754 | { |
| 1755 | raw_svector_ostream OS(Name); |
| 1756 | OS << "__os_log_helper"; |
| 1757 | OS << "_" << BufferAlignment.getQuantity(); |
| 1758 | OS << "_" << int(Layout.getSummaryByte()); |
| 1759 | OS << "_" << int(Layout.getNumArgsByte()); |
| 1760 | for (const auto &Item : Layout.Items) |
| 1761 | OS << "_" << int(Item.getSizeByte()) << "_" |
| 1762 | << int(Item.getDescriptorByte()); |
| 1763 | } |
| 1764 | |
| 1765 | if (llvm::Function *F = CGM.getModule().getFunction(Name)) |
| 1766 | return F; |
| 1767 | |
| 1768 | llvm::SmallVector<QualType, 4> ArgTys; |
| 1769 | FunctionArgList Args; |
| 1770 | Args.push_back(ImplicitParamDecl::Create( |
| 1771 | Ctx, nullptr, SourceLocation(), &Ctx.Idents.get("buffer"), Ctx.VoidPtrTy, |
| 1772 | ImplicitParamDecl::Other)); |
| 1773 | ArgTys.emplace_back(Ctx.VoidPtrTy); |
| 1774 | |
| 1775 | for (unsigned int I = 0, E = Layout.Items.size(); I < E; ++I) { |
| 1776 | char Size = Layout.Items[I].getSizeByte(); |
| 1777 | if (!Size) |
| 1778 | continue; |
| 1779 | |
| 1780 | QualType ArgTy = getOSLogArgType(Ctx, Size); |
| 1781 | Args.push_back(ImplicitParamDecl::Create( |
| 1782 | Ctx, nullptr, SourceLocation(), |
| 1783 | &Ctx.Idents.get(std::string("arg") + llvm::to_string(I)), ArgTy, |
| 1784 | ImplicitParamDecl::Other)); |
| 1785 | ArgTys.emplace_back(ArgTy); |
| 1786 | } |
| 1787 | |
| 1788 | QualType ReturnTy = Ctx.VoidTy; |
| 1789 | |
| 1790 | // The helper function has linkonce_odr linkage to enable the linker to merge |
| 1791 | // identical functions. To ensure the merging always happens, 'noinline' is |
| 1792 | // attached to the function when compiling with -Oz. |
| 1793 | const CGFunctionInfo &FI = |
| 1794 | CGM.getTypes().arrangeBuiltinFunctionDeclaration(ReturnTy, Args); |
| 1795 | llvm::FunctionType *FuncTy = CGM.getTypes().GetFunctionType(FI); |
| 1796 | llvm::Function *Fn = llvm::Function::Create( |
| 1797 | FuncTy, llvm::GlobalValue::LinkOnceODRLinkage, Name, &CGM.getModule()); |
| 1798 | Fn->setVisibility(llvm::GlobalValue::HiddenVisibility); |
| 1799 | CGM.SetLLVMFunctionAttributes(GlobalDecl(), FI, Fn, /*IsThunk=*/false); |
| 1800 | CGM.SetLLVMFunctionAttributesForDefinition(nullptr, Fn); |
| 1801 | Fn->setDoesNotThrow(); |
| 1802 | |
| 1803 | // Attach 'noinline' at -Oz. |
| 1804 | if (CGM.getCodeGenOpts().OptimizeSize == 2) |
| 1805 | Fn->addFnAttr(llvm::Attribute::NoInline); |
| 1806 | |
| 1807 | auto NL = ApplyDebugLocation::CreateEmpty(*this); |
| 1808 | StartFunction(GlobalDecl(), ReturnTy, Fn, FI, Args); |
| 1809 | |
| 1810 | // Create a scope with an artificial location for the body of this function. |
| 1811 | auto AL = ApplyDebugLocation::CreateArtificial(*this); |
| 1812 | |
| 1813 | CharUnits Offset; |
| 1814 | Address BufAddr = |
| 1815 | Address(Builder.CreateLoad(GetAddrOfLocalVar(Args[0]), "buf"), Int8Ty, |
| 1816 | BufferAlignment); |
| 1817 | Builder.CreateStore(Builder.getInt8(Layout.getSummaryByte()), |
| 1818 | Builder.CreateConstByteGEP(BufAddr, Offset++, "summary")); |
| 1819 | Builder.CreateStore(Builder.getInt8(Layout.getNumArgsByte()), |
| 1820 | Builder.CreateConstByteGEP(BufAddr, Offset++, "numArgs")); |
| 1821 | |
| 1822 | unsigned I = 1; |
| 1823 | for (const auto &Item : Layout.Items) { |
| 1824 | Builder.CreateStore( |
| 1825 | Builder.getInt8(Item.getDescriptorByte()), |
| 1826 | Builder.CreateConstByteGEP(BufAddr, Offset++, "argDescriptor")); |
| 1827 | Builder.CreateStore( |
| 1828 | Builder.getInt8(Item.getSizeByte()), |
| 1829 | Builder.CreateConstByteGEP(BufAddr, Offset++, "argSize")); |
| 1830 | |
| 1831 | CharUnits Size = Item.size(); |
| 1832 | if (!Size.getQuantity()) |
| 1833 | continue; |
| 1834 | |
| 1835 | Address Arg = GetAddrOfLocalVar(Args[I]); |
| 1836 | Address Addr = Builder.CreateConstByteGEP(BufAddr, Offset, "argData"); |
| 1837 | Addr = |
| 1838 | Builder.CreateElementBitCast(Addr, Arg.getElementType(), "argDataCast"); |
| 1839 | Builder.CreateStore(Builder.CreateLoad(Arg), Addr); |
| 1840 | Offset += Size; |
| 1841 | ++I; |
| 1842 | } |
| 1843 | |
| 1844 | FinishFunction(); |
| 1845 | |
| 1846 | return Fn; |
| 1847 | } |
| 1848 | |
| 1849 | RValue CodeGenFunction::emitBuiltinOSLogFormat(const CallExpr &E) { |
| 1850 | assert(E.getNumArgs() >= 2 &&(static_cast <bool> (E.getNumArgs() >= 2 && "__builtin_os_log_format takes at least 2 arguments" ) ? void (0) : __assert_fail ("E.getNumArgs() >= 2 && \"__builtin_os_log_format takes at least 2 arguments\"" , "clang/lib/CodeGen/CGBuiltin.cpp", 1851, __extension__ __PRETTY_FUNCTION__ )) |
| 1851 | "__builtin_os_log_format takes at least 2 arguments")(static_cast <bool> (E.getNumArgs() >= 2 && "__builtin_os_log_format takes at least 2 arguments" ) ? void (0) : __assert_fail ("E.getNumArgs() >= 2 && \"__builtin_os_log_format takes at least 2 arguments\"" , "clang/lib/CodeGen/CGBuiltin.cpp", 1851, __extension__ __PRETTY_FUNCTION__ )); |
| 1852 | ASTContext &Ctx = getContext(); |
| 1853 | analyze_os_log::OSLogBufferLayout Layout; |
| 1854 | analyze_os_log::computeOSLogBufferLayout(Ctx, &E, Layout); |
| 1855 | Address BufAddr = EmitPointerWithAlignment(E.getArg(0)); |
| 1856 | llvm::SmallVector<llvm::Value *, 4> RetainableOperands; |
| 1857 | |
| 1858 | // Ignore argument 1, the format string. It is not currently used. |
| 1859 | CallArgList Args; |
| 1860 | Args.add(RValue::get(BufAddr.getPointer()), Ctx.VoidPtrTy); |
| 1861 | |
| 1862 | for (const auto &Item : Layout.Items) { |
| 1863 | int Size = Item.getSizeByte(); |
| 1864 | if (!Size) |
| 1865 | continue; |
| 1866 | |
| 1867 | llvm::Value *ArgVal; |
| 1868 | |
| 1869 | if (Item.getKind() == analyze_os_log::OSLogBufferItem::MaskKind) { |
| 1870 | uint64_t Val = 0; |
| 1871 | for (unsigned I = 0, E = Item.getMaskType().size(); I < E; ++I) |
| 1872 | Val |= ((uint64_t)Item.getMaskType()[I]) << I * 8; |
| 1873 | ArgVal = llvm::Constant::getIntegerValue(Int64Ty, llvm::APInt(64, Val)); |
| 1874 | } else if (const Expr *TheExpr = Item.getExpr()) { |
| 1875 | ArgVal = EmitScalarExpr(TheExpr, /*Ignore*/ false); |
| 1876 | |
| 1877 | // If a temporary object that requires destruction after the full |
| 1878 | // expression is passed, push a lifetime-extended cleanup to extend its |
| 1879 | // lifetime to the end of the enclosing block scope. |
| 1880 | auto LifetimeExtendObject = [&](const Expr *E) { |
| 1881 | E = E->IgnoreParenCasts(); |
| 1882 | // Extend lifetimes of objects returned by function calls and message |
| 1883 | // sends. |
| 1884 | |
| 1885 | // FIXME: We should do this in other cases in which temporaries are |
| 1886 | // created including arguments of non-ARC types (e.g., C++ |
| 1887 | // temporaries). |
| 1888 | if (isa<CallExpr>(E) || isa<ObjCMessageExpr>(E)) |
| 1889 | return true; |
| 1890 | return false; |
| 1891 | }; |
| 1892 | |
| 1893 | if (TheExpr->getType()->isObjCRetainableType() && |
| 1894 | getLangOpts().ObjCAutoRefCount && LifetimeExtendObject(TheExpr)) { |
| 1895 | assert(getEvaluationKind(TheExpr->getType()) == TEK_Scalar &&(static_cast <bool> (getEvaluationKind(TheExpr->getType ()) == TEK_Scalar && "Only scalar can be a ObjC retainable type" ) ? void (0) : __assert_fail ("getEvaluationKind(TheExpr->getType()) == TEK_Scalar && \"Only scalar can be a ObjC retainable type\"" , "clang/lib/CodeGen/CGBuiltin.cpp", 1896, __extension__ __PRETTY_FUNCTION__ )) |
| 1896 | "Only scalar can be a ObjC retainable type")(static_cast <bool> (getEvaluationKind(TheExpr->getType ()) == TEK_Scalar && "Only scalar can be a ObjC retainable type" ) ? void (0) : __assert_fail ("getEvaluationKind(TheExpr->getType()) == TEK_Scalar && \"Only scalar can be a ObjC retainable type\"" , "clang/lib/CodeGen/CGBuiltin.cpp", 1896, __extension__ __PRETTY_FUNCTION__ )); |
| 1897 | if (!isa<Constant>(ArgVal)) { |
| 1898 | CleanupKind Cleanup = getARCCleanupKind(); |
| 1899 | QualType Ty = TheExpr->getType(); |
| 1900 | Address Alloca = Address::invalid(); |
| 1901 | Address Addr = CreateMemTemp(Ty, "os.log.arg", &Alloca); |
| 1902 | ArgVal = EmitARCRetain(Ty, ArgVal); |
| 1903 | Builder.CreateStore(ArgVal, Addr); |
| 1904 | pushLifetimeExtendedDestroy(Cleanup, Alloca, Ty, |
| 1905 | CodeGenFunction::destroyARCStrongPrecise, |
| 1906 | Cleanup & EHCleanup); |
| 1907 | |
| 1908 | // Push a clang.arc.use call to ensure ARC optimizer knows that the |
| 1909 | // argument has to be alive. |
| 1910 | if (CGM.getCodeGenOpts().OptimizationLevel != 0) |
| 1911 | pushCleanupAfterFullExpr<CallObjCArcUse>(Cleanup, ArgVal); |
| 1912 | } |
| 1913 | } |
| 1914 | } else { |
| 1915 | ArgVal = Builder.getInt32(Item.getConstValue().getQuantity()); |
| 1916 | } |
| 1917 | |
| 1918 | unsigned ArgValSize = |
| 1919 | CGM.getDataLayout().getTypeSizeInBits(ArgVal->getType()); |
| 1920 | llvm::IntegerType *IntTy = llvm::Type::getIntNTy(getLLVMContext(), |
| 1921 | ArgValSize); |
| 1922 | ArgVal = Builder.CreateBitOrPointerCast(ArgVal, IntTy); |
| 1923 | CanQualType ArgTy = getOSLogArgType(Ctx, Size); |
| 1924 | // If ArgVal has type x86_fp80, zero-extend ArgVal. |
| 1925 | ArgVal = Builder.CreateZExtOrBitCast(ArgVal, ConvertType(ArgTy)); |
| 1926 | Args.add(RValue::get(ArgVal), ArgTy); |
| 1927 | } |
| 1928 | |
| 1929 | const CGFunctionInfo &FI = |
| 1930 | CGM.getTypes().arrangeBuiltinFunctionCall(Ctx.VoidTy, Args); |
| 1931 | llvm::Function *F = CodeGenFunction(CGM).generateBuiltinOSLogHelperFunction( |
| 1932 | Layout, BufAddr.getAlignment()); |
| 1933 | EmitCall(FI, CGCallee::forDirect(F), ReturnValueSlot(), Args); |
| 1934 | return RValue::get(BufAddr.getPointer()); |
| 1935 | } |
| 1936 | |
| 1937 | static bool isSpecialUnsignedMultiplySignedResult( |
| 1938 | unsigned BuiltinID, WidthAndSignedness Op1Info, WidthAndSignedness Op2Info, |
| 1939 | WidthAndSignedness ResultInfo) { |
| 1940 | return BuiltinID == Builtin::BI__builtin_mul_overflow && |
| 1941 | Op1Info.Width == Op2Info.Width && Op2Info.Width == ResultInfo.Width && |
| 1942 | !Op1Info.Signed && !Op2Info.Signed && ResultInfo.Signed; |
| 1943 | } |
| 1944 | |
| 1945 | static RValue EmitCheckedUnsignedMultiplySignedResult( |
| 1946 | CodeGenFunction &CGF, const clang::Expr *Op1, WidthAndSignedness Op1Info, |
| 1947 | const clang::Expr *Op2, WidthAndSignedness Op2Info, |
| 1948 | const clang::Expr *ResultArg, QualType ResultQTy, |
| 1949 | WidthAndSignedness ResultInfo) { |
| 1950 | assert(isSpecialUnsignedMultiplySignedResult((static_cast <bool> (isSpecialUnsignedMultiplySignedResult ( Builtin::BI__builtin_mul_overflow, Op1Info, Op2Info, ResultInfo ) && "Cannot specialize this multiply") ? void (0) : __assert_fail ("isSpecialUnsignedMultiplySignedResult( Builtin::BI__builtin_mul_overflow, Op1Info, Op2Info, ResultInfo) && \"Cannot specialize this multiply\"" , "clang/lib/CodeGen/CGBuiltin.cpp", 1952, __extension__ __PRETTY_FUNCTION__ )) |
| 1951 | Builtin::BI__builtin_mul_overflow, Op1Info, Op2Info, ResultInfo) &&(static_cast <bool> (isSpecialUnsignedMultiplySignedResult ( Builtin::BI__builtin_mul_overflow, Op1Info, Op2Info, ResultInfo ) && "Cannot specialize this multiply") ? void (0) : __assert_fail ("isSpecialUnsignedMultiplySignedResult( Builtin::BI__builtin_mul_overflow, Op1Info, Op2Info, ResultInfo) && \"Cannot specialize this multiply\"" , "clang/lib/CodeGen/CGBuiltin.cpp", 1952, __extension__ __PRETTY_FUNCTION__ )) |
| 1952 | "Cannot specialize this multiply")(static_cast <bool> (isSpecialUnsignedMultiplySignedResult ( Builtin::BI__builtin_mul_overflow, Op1Info, Op2Info, ResultInfo ) && "Cannot specialize this multiply") ? void (0) : __assert_fail ("isSpecialUnsignedMultiplySignedResult( Builtin::BI__builtin_mul_overflow, Op1Info, Op2Info, ResultInfo) && \"Cannot specialize this multiply\"" , "clang/lib/CodeGen/CGBuiltin.cpp", 1952, __extension__ __PRETTY_FUNCTION__ )); |
| 1953 | |
| 1954 | llvm::Value *V1 = CGF.EmitScalarExpr(Op1); |
| 1955 | llvm::Value *V2 = CGF.EmitScalarExpr(Op2); |
| 1956 | |
| 1957 | llvm::Value *HasOverflow; |
| 1958 | llvm::Value *Result = EmitOverflowIntrinsic( |
| 1959 | CGF, llvm::Intrinsic::umul_with_overflow, V1, V2, HasOverflow); |
| 1960 | |
| 1961 | // The intrinsic call will detect overflow when the value is > UINT_MAX, |
| 1962 | // however, since the original builtin had a signed result, we need to report |
| 1963 | // an overflow when the result is greater than INT_MAX. |
| 1964 | auto IntMax = llvm::APInt::getSignedMaxValue(ResultInfo.Width); |
| 1965 | llvm::Value *IntMaxValue = llvm::ConstantInt::get(Result->getType(), IntMax); |
| 1966 | |
| 1967 | llvm::Value *IntMaxOverflow = CGF.Builder.CreateICmpUGT(Result, IntMaxValue); |
| 1968 | HasOverflow = CGF.Builder.CreateOr(HasOverflow, IntMaxOverflow); |
| 1969 | |
| 1970 | bool isVolatile = |
| 1971 | ResultArg->getType()->getPointeeType().isVolatileQualified(); |
| 1972 | Address ResultPtr = CGF.EmitPointerWithAlignment(ResultArg); |
| 1973 | CGF.Builder.CreateStore(CGF.EmitToMemory(Result, ResultQTy), ResultPtr, |
| 1974 | isVolatile); |
| 1975 | return RValue::get(HasOverflow); |
| 1976 | } |
| 1977 | |
| 1978 | /// Determine if a binop is a checked mixed-sign multiply we can specialize. |
| 1979 | static bool isSpecialMixedSignMultiply(unsigned BuiltinID, |
| 1980 | WidthAndSignedness Op1Info, |
| 1981 | WidthAndSignedness Op2Info, |
| 1982 | WidthAndSignedness ResultInfo) { |
| 1983 | return BuiltinID == Builtin::BI__builtin_mul_overflow && |
| 1984 | std::max(Op1Info.Width, Op2Info.Width) >= ResultInfo.Width && |
| 1985 | Op1Info.Signed != Op2Info.Signed; |
| 1986 | } |
| 1987 | |
| 1988 | /// Emit a checked mixed-sign multiply. This is a cheaper specialization of |
| 1989 | /// the generic checked-binop irgen. |
| 1990 | static RValue |
| 1991 | EmitCheckedMixedSignMultiply(CodeGenFunction &CGF, const clang::Expr *Op1, |
| 1992 | WidthAndSignedness Op1Info, const clang::Expr *Op2, |
| 1993 | WidthAndSignedness Op2Info, |
| 1994 | const clang::Expr *ResultArg, QualType ResultQTy, |
| 1995 | WidthAndSignedness ResultInfo) { |
| 1996 | assert(isSpecialMixedSignMultiply(Builtin::BI__builtin_mul_overflow, Op1Info,(static_cast <bool> (isSpecialMixedSignMultiply(Builtin ::BI__builtin_mul_overflow, Op1Info, Op2Info, ResultInfo) && "Not a mixed-sign multipliction we can specialize") ? void ( 0) : __assert_fail ("isSpecialMixedSignMultiply(Builtin::BI__builtin_mul_overflow, Op1Info, Op2Info, ResultInfo) && \"Not a mixed-sign multipliction we can specialize\"" , "clang/lib/CodeGen/CGBuiltin.cpp", 1998, __extension__ __PRETTY_FUNCTION__ )) |
| 1997 | Op2Info, ResultInfo) &&(static_cast <bool> (isSpecialMixedSignMultiply(Builtin ::BI__builtin_mul_overflow, Op1Info, Op2Info, ResultInfo) && "Not a mixed-sign multipliction we can specialize") ? void ( 0) : __assert_fail ("isSpecialMixedSignMultiply(Builtin::BI__builtin_mul_overflow, Op1Info, Op2Info, ResultInfo) && \"Not a mixed-sign multipliction we can specialize\"" , "clang/lib/CodeGen/CGBuiltin.cpp", 1998, __extension__ __PRETTY_FUNCTION__ )) |
| 1998 | "Not a mixed-sign multipliction we can specialize")(static_cast <bool> (isSpecialMixedSignMultiply(Builtin ::BI__builtin_mul_overflow, Op1Info, Op2Info, ResultInfo) && "Not a mixed-sign multipliction we can specialize") ? void ( 0) : __assert_fail ("isSpecialMixedSignMultiply(Builtin::BI__builtin_mul_overflow, Op1Info, Op2Info, ResultInfo) && \"Not a mixed-sign multipliction we can specialize\"" , "clang/lib/CodeGen/CGBuiltin.cpp", 1998, __extension__ __PRETTY_FUNCTION__ )); |
| 1999 | |
| 2000 | // Emit the signed and unsigned operands. |
| 2001 | const clang::Expr *SignedOp = Op1Info.Signed ? Op1 : Op2; |
| 2002 | const clang::Expr *UnsignedOp = Op1Info.Signed ? Op2 : Op1; |
| 2003 | llvm::Value *Signed = CGF.EmitScalarExpr(SignedOp); |
| 2004 | llvm::Value *Unsigned = CGF.EmitScalarExpr(UnsignedOp); |
| 2005 | unsigned SignedOpWidth = Op1Info.Signed ? Op1Info.Width : Op2Info.Width; |
| 2006 | unsigned UnsignedOpWidth = Op1Info.Signed ? Op2Info.Width : Op1Info.Width; |
| 2007 | |
| 2008 | // One of the operands may be smaller than the other. If so, [s|z]ext it. |
| 2009 | if (SignedOpWidth < UnsignedOpWidth) |
| 2010 | Signed = CGF.Builder.CreateSExt(Signed, Unsigned->getType(), "op.sext"); |
| 2011 | if (UnsignedOpWidth < SignedOpWidth) |
| 2012 | Unsigned = CGF.Builder.CreateZExt(Unsigned, Signed->getType(), "op.zext"); |
| 2013 | |
| 2014 | llvm::Type *OpTy = Signed->getType(); |
| 2015 | llvm::Value *Zero = llvm::Constant::getNullValue(OpTy); |
| 2016 | Address ResultPtr = CGF.EmitPointerWithAlignment(ResultArg); |
| 2017 | llvm::Type *ResTy = ResultPtr.getElementType(); |
| 2018 | unsigned OpWidth = std::max(Op1Info.Width, Op2Info.Width); |
| 2019 | |
| 2020 | // Take the absolute value of the signed operand. |
| 2021 | llvm::Value *IsNegative = CGF.Builder.CreateICmpSLT(Signed, Zero); |
| 2022 | llvm::Value *AbsOfNegative = CGF.Builder.CreateSub(Zero, Signed); |
| 2023 | llvm::Value *AbsSigned = |
| 2024 | CGF.Builder.CreateSelect(IsNegative, AbsOfNegative, Signed); |
| 2025 | |
| 2026 | // Perform a checked unsigned multiplication. |
| 2027 | llvm::Value *UnsignedOverflow; |
| 2028 | llvm::Value *UnsignedResult = |
| 2029 | EmitOverflowIntrinsic(CGF, llvm::Intrinsic::umul_with_overflow, AbsSigned, |
| 2030 | Unsigned, UnsignedOverflow); |
| 2031 | |
| 2032 | llvm::Value *Overflow, *Result; |
| 2033 | if (ResultInfo.Signed) { |
| 2034 | // Signed overflow occurs if the result is greater than INT_MAX or lesser |
| 2035 | // than INT_MIN, i.e when |Result| > (INT_MAX + IsNegative). |
| 2036 | auto IntMax = |
| 2037 | llvm::APInt::getSignedMaxValue(ResultInfo.Width).zext(OpWidth); |
| 2038 | llvm::Value *MaxResult = |
| 2039 | CGF.Builder.CreateAdd(llvm::ConstantInt::get(OpTy, IntMax), |
| 2040 | CGF.Builder.CreateZExt(IsNegative, OpTy)); |
| 2041 | llvm::Value *SignedOverflow = |
| 2042 | CGF.Builder.CreateICmpUGT(UnsignedResult, MaxResult); |
| 2043 | Overflow = CGF.Builder.CreateOr(UnsignedOverflow, SignedOverflow); |
| 2044 | |
| 2045 | // Prepare the signed result (possibly by negating it). |
| 2046 | llvm::Value *NegativeResult = CGF.Builder.CreateNeg(UnsignedResult); |
| 2047 | llvm::Value *SignedResult = |
| 2048 | CGF.Builder.CreateSelect(IsNegative, NegativeResult, UnsignedResult); |
| 2049 | Result = CGF.Builder.CreateTrunc(SignedResult, ResTy); |
| 2050 | } else { |
| 2051 | // Unsigned overflow occurs if the result is < 0 or greater than UINT_MAX. |
| 2052 | llvm::Value *Underflow = CGF.Builder.CreateAnd( |
| 2053 | IsNegative, CGF.Builder.CreateIsNotNull(UnsignedResult)); |
| 2054 | Overflow = CGF.Builder.CreateOr(UnsignedOverflow, Underflow); |
| 2055 | if (ResultInfo.Width < OpWidth) { |
| 2056 | auto IntMax = |
| 2057 | llvm::APInt::getMaxValue(ResultInfo.Width).zext(OpWidth); |
| 2058 | llvm::Value *TruncOverflow = CGF.Builder.CreateICmpUGT( |
| 2059 | UnsignedResult, llvm::ConstantInt::get(OpTy, IntMax)); |
| 2060 | Overflow = CGF.Builder.CreateOr(Overflow, TruncOverflow); |
| 2061 | } |
| 2062 | |
| 2063 | // Negate the product if it would be negative in infinite precision. |
| 2064 | Result = CGF.Builder.CreateSelect( |
| 2065 | IsNegative, CGF.Builder.CreateNeg(UnsignedResult), UnsignedResult); |
| 2066 | |
| 2067 | Result = CGF.Builder.CreateTrunc(Result, ResTy); |
| 2068 | } |
| 2069 | assert(Overflow && Result && "Missing overflow or result")(static_cast <bool> (Overflow && Result && "Missing overflow or result") ? void (0) : __assert_fail ("Overflow && Result && \"Missing overflow or result\"" , "clang/lib/CodeGen/CGBuiltin.cpp", 2069, __extension__ __PRETTY_FUNCTION__ )); |
| 2070 | |
| 2071 | bool isVolatile = |
| 2072 | ResultArg->getType()->getPointeeType().isVolatileQualified(); |
| 2073 | CGF.Builder.CreateStore(CGF.EmitToMemory(Result, ResultQTy), ResultPtr, |
| 2074 | isVolatile); |
| 2075 | return RValue::get(Overflow); |
| 2076 | } |
| 2077 | |
| 2078 | static bool |
| 2079 | TypeRequiresBuiltinLaunderImp(const ASTContext &Ctx, QualType Ty, |
| 2080 | llvm::SmallPtrSetImpl<const Decl *> &Seen) { |
| 2081 | if (const auto *Arr = Ctx.getAsArrayType(Ty)) |
| 2082 | Ty = Ctx.getBaseElementType(Arr); |
| 2083 | |
| 2084 | const auto *Record = Ty->getAsCXXRecordDecl(); |
| 2085 | if (!Record) |
| 2086 | return false; |
| 2087 | |
| 2088 | // We've already checked this type, or are in the process of checking it. |
| 2089 | if (!Seen.insert(Record).second) |
| 2090 | return false; |
| 2091 | |
| 2092 | assert(Record->hasDefinition() &&(static_cast <bool> (Record->hasDefinition() && "Incomplete types should already be diagnosed") ? void (0) : __assert_fail ("Record->hasDefinition() && \"Incomplete types should already be diagnosed\"" , "clang/lib/CodeGen/CGBuiltin.cpp", 2093, __extension__ __PRETTY_FUNCTION__ )) |
| 2093 | "Incomplete types should already be diagnosed")(static_cast <bool> (Record->hasDefinition() && "Incomplete types should already be diagnosed") ? void (0) : __assert_fail ("Record->hasDefinition() && \"Incomplete types should already be diagnosed\"" , "clang/lib/CodeGen/CGBuiltin.cpp", 2093, __extension__ __PRETTY_FUNCTION__ )); |
| 2094 | |
| 2095 | if (Record->isDynamicClass()) |
| 2096 | return true; |
| 2097 | |
| 2098 | for (FieldDecl *F : Record->fields()) { |
| 2099 | if (TypeRequiresBuiltinLaunderImp(Ctx, F->getType(), Seen)) |
| 2100 | return true; |
| 2101 | } |
| 2102 | return false; |
| 2103 | } |
| 2104 | |
| 2105 | /// Determine if the specified type requires laundering by checking if it is a |
| 2106 | /// dynamic class type or contains a subobject which is a dynamic class type. |
| 2107 | static bool TypeRequiresBuiltinLaunder(CodeGenModule &CGM, QualType Ty) { |
| 2108 | if (!CGM.getCodeGenOpts().StrictVTablePointers) |
| 2109 | return false; |
| 2110 | llvm::SmallPtrSet<const Decl *, 16> Seen; |
| 2111 | return TypeRequiresBuiltinLaunderImp(CGM.getContext(), Ty, Seen); |
| 2112 | } |
| 2113 | |
| 2114 | RValue CodeGenFunction::emitRotate(const CallExpr *E, bool IsRotateRight) { |
| 2115 | llvm::Value *Src = EmitScalarExpr(E->getArg(0)); |
| 2116 | llvm::Value *ShiftAmt = EmitScalarExpr(E->getArg(1)); |
| 2117 | |
| 2118 | // The builtin's shift arg may have a different type than the source arg and |
| 2119 | // result, but the LLVM intrinsic uses the same type for all values. |
| 2120 | llvm::Type *Ty = Src->getType(); |
| 2121 | ShiftAmt = Builder.CreateIntCast(ShiftAmt, Ty, false); |
| 2122 | |
| 2123 | // Rotate is a special case of LLVM funnel shift - 1st 2 args are the same. |
| 2124 | unsigned IID = IsRotateRight ? Intrinsic::fshr : Intrinsic::fshl; |
| 2125 | Function *F = CGM.getIntrinsic(IID, Ty); |
| 2126 | return RValue::get(Builder.CreateCall(F, { Src, Src, ShiftAmt })); |
| 2127 | } |
| 2128 | |
| 2129 | // Map math builtins for long-double to f128 version. |
| 2130 | static unsigned mutateLongDoubleBuiltin(unsigned BuiltinID) { |
| 2131 | switch (BuiltinID) { |
| 2132 | #define MUTATE_LDBL(func) \ |
| 2133 | case Builtin::BI__builtin_##func##l: \ |
| 2134 | return Builtin::BI__builtin_##func##f128; |
| 2135 | MUTATE_LDBL(sqrt) |
| 2136 | MUTATE_LDBL(cbrt) |
| 2137 | MUTATE_LDBL(fabs) |
| 2138 | MUTATE_LDBL(log) |
| 2139 | MUTATE_LDBL(log2) |
| 2140 | MUTATE_LDBL(log10) |
| 2141 | MUTATE_LDBL(log1p) |
| 2142 | MUTATE_LDBL(logb) |
| 2143 | MUTATE_LDBL(exp) |
| 2144 | MUTATE_LDBL(exp2) |
| 2145 | MUTATE_LDBL(expm1) |
| 2146 | MUTATE_LDBL(fdim) |
| 2147 | MUTATE_LDBL(hypot) |
| 2148 | MUTATE_LDBL(ilogb) |
| 2149 | MUTATE_LDBL(pow) |
| 2150 | MUTATE_LDBL(fmin) |
| 2151 | MUTATE_LDBL(fmax) |
| 2152 | MUTATE_LDBL(ceil) |
| 2153 | MUTATE_LDBL(trunc) |
| 2154 | MUTATE_LDBL(rint) |
| 2155 | MUTATE_LDBL(nearbyint) |
| 2156 | MUTATE_LDBL(round) |
| 2157 | MUTATE_LDBL(floor) |
| 2158 | MUTATE_LDBL(lround) |
| 2159 | MUTATE_LDBL(llround) |
| 2160 | MUTATE_LDBL(lrint) |
| 2161 | MUTATE_LDBL(llrint) |
| 2162 | MUTATE_LDBL(fmod) |
| 2163 | MUTATE_LDBL(modf) |
| 2164 | MUTATE_LDBL(nan) |
| 2165 | MUTATE_LDBL(nans) |
| 2166 | MUTATE_LDBL(inf) |
| 2167 | MUTATE_LDBL(fma) |
| 2168 | MUTATE_LDBL(sin) |
| 2169 | MUTATE_LDBL(cos) |
| 2170 | MUTATE_LDBL(tan) |
| 2171 | MUTATE_LDBL(sinh) |
| 2172 | MUTATE_LDBL(cosh) |
| 2173 | MUTATE_LDBL(tanh) |
| 2174 | MUTATE_LDBL(asin) |
| 2175 | MUTATE_LDBL(acos) |
| 2176 | MUTATE_LDBL(atan) |
| 2177 | MUTATE_LDBL(asinh) |
| 2178 | MUTATE_LDBL(acosh) |
| 2179 | MUTATE_LDBL(atanh) |
| 2180 | MUTATE_LDBL(atan2) |
| 2181 | MUTATE_LDBL(erf) |
| 2182 | MUTATE_LDBL(erfc) |
| 2183 | MUTATE_LDBL(ldexp) |
| 2184 | MUTATE_LDBL(frexp) |
| 2185 | MUTATE_LDBL(huge_val) |
| 2186 | MUTATE_LDBL(copysign) |
| 2187 | MUTATE_LDBL(nextafter) |
| 2188 | MUTATE_LDBL(nexttoward) |
| 2189 | MUTATE_LDBL(remainder) |
| 2190 | MUTATE_LDBL(remquo) |
| 2191 | MUTATE_LDBL(scalbln) |
| 2192 | MUTATE_LDBL(scalbn) |
| 2193 | MUTATE_LDBL(tgamma) |
| 2194 | MUTATE_LDBL(lgamma) |
| 2195 | #undef MUTATE_LDBL |
| 2196 | default: |
| 2197 | return BuiltinID; |
| 2198 | } |
| 2199 | } |
| 2200 | |
| 2201 | RValue CodeGenFunction::EmitBuiltinExpr(const GlobalDecl GD, unsigned BuiltinID, |
| 2202 | const CallExpr *E, |
| 2203 | ReturnValueSlot ReturnValue) { |
| 2204 | const FunctionDecl *FD = GD.getDecl()->getAsFunction(); |
| 2205 | // See if we can constant fold this builtin. If so, don't emit it at all. |
| 2206 | // TODO: Extend this handling to all builtin calls that we can constant-fold. |
| 2207 | Expr::EvalResult Result; |
| 2208 | if (E->isPRValue() && E->EvaluateAsRValue(Result, CGM.getContext()) && |
| 2209 | !Result.hasSideEffects()) { |
| 2210 | if (Result.Val.isInt()) |
| 2211 | return RValue::get(llvm::ConstantInt::get(getLLVMContext(), |
| 2212 | Result.Val.getInt())); |
| 2213 | if (Result.Val.isFloat()) |
| 2214 | return RValue::get(llvm::ConstantFP::get(getLLVMContext(), |
| 2215 | Result.Val.getFloat())); |
| 2216 | } |
| 2217 | |
| 2218 | // If current long-double semantics is IEEE 128-bit, replace math builtins |
| 2219 | // of long-double with f128 equivalent. |
| 2220 | // TODO: This mutation should also be applied to other targets other than PPC, |
| 2221 | // after backend supports IEEE 128-bit style libcalls. |
| 2222 | if (getTarget().getTriple().isPPC64() && |
| 2223 | &getTarget().getLongDoubleFormat() == &llvm::APFloat::IEEEquad()) |
| 2224 | BuiltinID = mutateLongDoubleBuiltin(BuiltinID); |
| 2225 | |
| 2226 | // If the builtin has been declared explicitly with an assembler label, |
| 2227 | // disable the specialized emitting below. Ideally we should communicate the |
| 2228 | // rename in IR, or at least avoid generating the intrinsic calls that are |
| 2229 | // likely to get lowered to the renamed library functions. |
| 2230 | const unsigned BuiltinIDIfNoAsmLabel = |
| 2231 | FD->hasAttr<AsmLabelAttr>() ? 0 : BuiltinID; |
| 2232 | |
| 2233 | // There are LLVM math intrinsics/instructions corresponding to math library |
| 2234 | // functions except the LLVM op will never set errno while the math library |
| 2235 | // might. Also, math builtins have the same semantics as their math library |
| 2236 | // twins. Thus, we can transform math library and builtin calls to their |
| 2237 | // LLVM counterparts if the call is marked 'const' (known to never set errno). |
| 2238 | // In case FP exceptions are enabled, the experimental versions of the |
| 2239 | // intrinsics model those. |
| 2240 | bool ConstWithoutErrnoAndExceptions = |
| 2241 | getContext().BuiltinInfo.isConstWithoutErrnoAndExceptions(BuiltinID); |
| 2242 | bool ConstWithoutExceptions = |
| 2243 | getContext().BuiltinInfo.isConstWithoutExceptions(BuiltinID); |
| 2244 | if (FD->hasAttr<ConstAttr>() || |
| 2245 | ((ConstWithoutErrnoAndExceptions || ConstWithoutExceptions) && |
| 2246 | (!ConstWithoutErrnoAndExceptions || (!getLangOpts().MathErrno)))) { |
| 2247 | switch (BuiltinIDIfNoAsmLabel) { |
| 2248 | case Builtin::BIceil: |
| 2249 | case Builtin::BIceilf: |
| 2250 | case Builtin::BIceill: |
| 2251 | case Builtin::BI__builtin_ceil: |
| 2252 | case Builtin::BI__builtin_ceilf: |
| 2253 | case Builtin::BI__builtin_ceilf16: |
| 2254 | case Builtin::BI__builtin_ceill: |
| 2255 | case Builtin::BI__builtin_ceilf128: |
| 2256 | return RValue::get(emitUnaryMaybeConstrainedFPBuiltin(*this, E, |
| 2257 | Intrinsic::ceil, |
| 2258 | Intrinsic::experimental_constrained_ceil)); |
| 2259 | |
| 2260 | case Builtin::BIcopysign: |
| 2261 | case Builtin::BIcopysignf: |
| 2262 | case Builtin::BIcopysignl: |
| 2263 | case Builtin::BI__builtin_copysign: |
| 2264 | case Builtin::BI__builtin_copysignf: |
| 2265 | case Builtin::BI__builtin_copysignf16: |
| 2266 | case Builtin::BI__builtin_copysignl: |
| 2267 | case Builtin::BI__builtin_copysignf128: |
| 2268 | return RValue::get(emitBinaryBuiltin(*this, E, Intrinsic::copysign)); |
| 2269 | |
| 2270 | case Builtin::BIcos: |
| 2271 | case Builtin::BIcosf: |
| 2272 | case Builtin::BIcosl: |
| 2273 | case Builtin::BI__builtin_cos: |
| 2274 | case Builtin::BI__builtin_cosf: |
| 2275 | case Builtin::BI__builtin_cosf16: |
| 2276 | case Builtin::BI__builtin_cosl: |
| 2277 | case Builtin::BI__builtin_cosf128: |
| 2278 | return RValue::get(emitUnaryMaybeConstrainedFPBuiltin(*this, E, |
| 2279 | Intrinsic::cos, |
| 2280 | Intrinsic::experimental_constrained_cos)); |
| 2281 | |
| 2282 | case Builtin::BIexp: |
| 2283 | case Builtin::BIexpf: |
| 2284 | case Builtin::BIexpl: |
| 2285 | case Builtin::BI__builtin_exp: |
| 2286 | case Builtin::BI__builtin_expf: |
| 2287 | case Builtin::BI__builtin_expf16: |
| 2288 | case Builtin::BI__builtin_expl: |
| 2289 | case Builtin::BI__builtin_expf128: |
| 2290 | return RValue::get(emitUnaryMaybeConstrainedFPBuiltin(*this, E, |
| 2291 | Intrinsic::exp, |
| 2292 | Intrinsic::experimental_constrained_exp)); |
| 2293 | |
| 2294 | case Builtin::BIexp2: |
| 2295 | case Builtin::BIexp2f: |
| 2296 | case Builtin::BIexp2l: |
| 2297 | case Builtin::BI__builtin_exp2: |
| 2298 | case Builtin::BI__builtin_exp2f: |
| 2299 | case Builtin::BI__builtin_exp2f16: |
| 2300 | case Builtin::BI__builtin_exp2l: |
| 2301 | case Builtin::BI__builtin_exp2f128: |
| 2302 | return RValue::get(emitUnaryMaybeConstrainedFPBuiltin(*this, E, |
| 2303 | Intrinsic::exp2, |
| 2304 | Intrinsic::experimental_constrained_exp2)); |
| 2305 | |
| 2306 | case Builtin::BIfabs: |
| 2307 | case Builtin::BIfabsf: |
| 2308 | case Builtin::BIfabsl: |
| 2309 | case Builtin::BI__builtin_fabs: |
| 2310 | case Builtin::BI__builtin_fabsf: |
| 2311 | case Builtin::BI__builtin_fabsf16: |
| 2312 | case Builtin::BI__builtin_fabsl: |
| 2313 | case Builtin::BI__builtin_fabsf128: |
| 2314 | return RValue::get(emitUnaryBuiltin(*this, E, Intrinsic::fabs)); |
| 2315 | |
| 2316 | case Builtin::BIfloor: |
| 2317 | case Builtin::BIfloorf: |
| 2318 | case Builtin::BIfloorl: |
| 2319 | case Builtin::BI__builtin_floor: |
| 2320 | case Builtin::BI__builtin_floorf: |
| 2321 | case Builtin::BI__builtin_floorf16: |
| 2322 | case Builtin::BI__builtin_floorl: |
| 2323 | case Builtin::BI__builtin_floorf128: |
| 2324 | return RValue::get(emitUnaryMaybeConstrainedFPBuiltin(*this, E, |
| 2325 | Intrinsic::floor, |
| 2326 | Intrinsic::experimental_constrained_floor)); |
| 2327 | |
| 2328 | case Builtin::BIfma: |
| 2329 | case Builtin::BIfmaf: |
| 2330 | case Builtin::BIfmal: |
| 2331 | case Builtin::BI__builtin_fma: |
| 2332 | case Builtin::BI__builtin_fmaf: |
| 2333 | case Builtin::BI__builtin_fmaf16: |
| 2334 | case Builtin::BI__builtin_fmal: |
| 2335 | case Builtin::BI__builtin_fmaf128: |
| 2336 | return RValue::get(emitTernaryMaybeConstrainedFPBuiltin(*this, E, |
| 2337 | Intrinsic::fma, |
| 2338 | Intrinsic::experimental_constrained_fma)); |
| 2339 | |
| 2340 | case Builtin::BIfmax: |
| 2341 | case Builtin::BIfmaxf: |
| 2342 | case Builtin::BIfmaxl: |
| 2343 | case Builtin::BI__builtin_fmax: |
| 2344 | case Builtin::BI__builtin_fmaxf: |
| 2345 | case Builtin::BI__builtin_fmaxf16: |
| 2346 | case Builtin::BI__builtin_fmaxl: |
| 2347 | case Builtin::BI__builtin_fmaxf128: |
| 2348 | return RValue::get(emitBinaryMaybeConstrainedFPBuiltin(*this, E, |
| 2349 | Intrinsic::maxnum, |
| 2350 | Intrinsic::experimental_constrained_maxnum)); |
| 2351 | |
| 2352 | case Builtin::BIfmin: |
| 2353 | case Builtin::BIfminf: |
| 2354 | case Builtin::BIfminl: |
| 2355 | case Builtin::BI__builtin_fmin: |
| 2356 | case Builtin::BI__builtin_fminf: |
| 2357 | case Builtin::BI__builtin_fminf16: |
| 2358 | case Builtin::BI__builtin_fminl: |
| 2359 | case Builtin::BI__builtin_fminf128: |
| 2360 | return RValue::get(emitBinaryMaybeConstrainedFPBuiltin(*this, E, |
| 2361 | Intrinsic::minnum, |
| 2362 | Intrinsic::experimental_constrained_minnum)); |
| 2363 | |
| 2364 | // fmod() is a special-case. It maps to the frem instruction rather than an |
| 2365 | // LLVM intrinsic. |
| 2366 | case Builtin::BIfmod: |
| 2367 | case Builtin::BIfmodf: |
| 2368 | case Builtin::BIfmodl: |
| 2369 | case Builtin::BI__builtin_fmod: |
| 2370 | case Builtin::BI__builtin_fmodf: |
| 2371 | case Builtin::BI__builtin_fmodf16: |
| 2372 | case Builtin::BI__builtin_fmodl: |
| 2373 | case Builtin::BI__builtin_fmodf128: { |
| 2374 | CodeGenFunction::CGFPOptionsRAII FPOptsRAII(*this, E); |
| 2375 | Value *Arg1 = EmitScalarExpr(E->getArg(0)); |
| 2376 | Value *Arg2 = EmitScalarExpr(E->getArg(1)); |
| 2377 | return RValue::get(Builder.CreateFRem(Arg1, Arg2, "fmod")); |
| 2378 | } |
| 2379 | |
| 2380 | case Builtin::BIlog: |
| 2381 | case Builtin::BIlogf: |
| 2382 | case Builtin::BIlogl: |
| 2383 | case Builtin::BI__builtin_log: |
| 2384 | case Builtin::BI__builtin_logf: |
| 2385 | case Builtin::BI__builtin_logf16: |
| 2386 | case Builtin::BI__builtin_logl: |
| 2387 | case Builtin::BI__builtin_logf128: |
| 2388 | return RValue::get(emitUnaryMaybeConstrainedFPBuiltin(*this, E, |
| 2389 | Intrinsic::log, |
| 2390 | Intrinsic::experimental_constrained_log)); |
| 2391 | |
| 2392 | case Builtin::BIlog10: |
| 2393 | case Builtin::BIlog10f: |
| 2394 | case Builtin::BIlog10l: |
| 2395 | case Builtin::BI__builtin_log10: |
| 2396 | case Builtin::BI__builtin_log10f: |
| 2397 | case Builtin::BI__builtin_log10f16: |
| 2398 | case Builtin::BI__builtin_log10l: |
| 2399 | case Builtin::BI__builtin_log10f128: |
| 2400 | return RValue::get(emitUnaryMaybeConstrainedFPBuiltin(*this, E, |
| 2401 | Intrinsic::log10, |
| 2402 | Intrinsic::experimental_constrained_log10)); |
| 2403 | |
| 2404 | case Builtin::BIlog2: |
| 2405 | case Builtin::BIlog2f: |
| 2406 | case Builtin::BIlog2l: |
| 2407 | case Builtin::BI__builtin_log2: |
| 2408 | case Builtin::BI__builtin_log2f: |
| 2409 | case Builtin::BI__builtin_log2f16: |
| 2410 | case Builtin::BI__builtin_log2l: |
| 2411 | case Builtin::BI__builtin_log2f128: |
| 2412 | return RValue::get(emitUnaryMaybeConstrainedFPBuiltin(*this, E, |
| 2413 | Intrinsic::log2, |
| 2414 | Intrinsic::experimental_constrained_log2)); |
| 2415 | |
| 2416 | case Builtin::BInearbyint: |
| 2417 | case Builtin::BInearbyintf: |
| 2418 | case Builtin::BInearbyintl: |
| 2419 | case Builtin::BI__builtin_nearbyint: |
| 2420 | case Builtin::BI__builtin_nearbyintf: |
| 2421 | case Builtin::BI__builtin_nearbyintl: |
| 2422 | case Builtin::BI__builtin_nearbyintf128: |
| 2423 | return RValue::get(emitUnaryMaybeConstrainedFPBuiltin(*this, E, |
| 2424 | Intrinsic::nearbyint, |
| 2425 | Intrinsic::experimental_constrained_nearbyint)); |
| 2426 | |
| 2427 | case Builtin::BIpow: |
| 2428 | case Builtin::BIpowf: |
| 2429 | case Builtin::BIpowl: |
| 2430 | case Builtin::BI__builtin_pow: |
| 2431 | case Builtin::BI__builtin_powf: |
| 2432 | case Builtin::BI__builtin_powf16: |
| 2433 | case Builtin::BI__builtin_powl: |
| 2434 | case Builtin::BI__builtin_powf128: |
| 2435 | return RValue::get(emitBinaryMaybeConstrainedFPBuiltin(*this, E, |
| 2436 | Intrinsic::pow, |
| 2437 | Intrinsic::experimental_constrained_pow)); |
| 2438 | |
| 2439 | case Builtin::BIrint: |
| 2440 | case Builtin::BIrintf: |
| 2441 | case Builtin::BIrintl: |
| 2442 | case Builtin::BI__builtin_rint: |
| 2443 | case Builtin::BI__builtin_rintf: |
| 2444 | case Builtin::BI__builtin_rintf16: |
| 2445 | case Builtin::BI__builtin_rintl: |
| 2446 | case Builtin::BI__builtin_rintf128: |
| 2447 | return RValue::get(emitUnaryMaybeConstrainedFPBuiltin(*this, E, |
| 2448 | Intrinsic::rint, |
| 2449 | Intrinsic::experimental_constrained_rint)); |
| 2450 | |
| 2451 | case Builtin::BIround: |
| 2452 | case Builtin::BIroundf: |
| 2453 | case Builtin::BIroundl: |
| 2454 | case Builtin::BI__builtin_round: |
| 2455 | case Builtin::BI__builtin_roundf: |
| 2456 | case Builtin::BI__builtin_roundf16: |
| 2457 | case Builtin::BI__builtin_roundl: |
| 2458 | case Builtin::BI__builtin_roundf128: |
| 2459 | return RValue::get(emitUnaryMaybeConstrainedFPBuiltin(*this, E, |
| 2460 | Intrinsic::round, |
| 2461 | Intrinsic::experimental_constrained_round)); |
| 2462 | |
| 2463 | case Builtin::BIroundeven: |
| 2464 | case Builtin::BIroundevenf: |
| 2465 | case Builtin::BIroundevenl: |
| 2466 | case Builtin::BI__builtin_roundeven: |
| 2467 | case Builtin::BI__builtin_roundevenf: |
| 2468 | case Builtin::BI__builtin_roundevenf16: |
| 2469 | case Builtin::BI__builtin_roundevenl: |
| 2470 | case Builtin::BI__builtin_roundevenf128: |
| 2471 | return RValue::get(emitUnaryMaybeConstrainedFPBuiltin(*this, E, |
| 2472 | Intrinsic::roundeven, |
| 2473 | Intrinsic::experimental_constrained_roundeven)); |
| 2474 | |
| 2475 | case Builtin::BIsin: |
| 2476 | case Builtin::BIsinf: |
| 2477 | case Builtin::BIsinl: |
| 2478 | case Builtin::BI__builtin_sin: |
| 2479 | case Builtin::BI__builtin_sinf: |
| 2480 | case Builtin::BI__builtin_sinf16: |
| 2481 | case Builtin::BI__builtin_sinl: |
| 2482 | case Builtin::BI__builtin_sinf128: |
| 2483 | return RValue::get(emitUnaryMaybeConstrainedFPBuiltin(*this, E, |
| 2484 | Intrinsic::sin, |
| 2485 | Intrinsic::experimental_constrained_sin)); |
| 2486 | |
| 2487 | case Builtin::BIsqrt: |
| 2488 | case Builtin::BIsqrtf: |
| 2489 | case Builtin::BIsqrtl: |
| 2490 | case Builtin::BI__builtin_sqrt: |
| 2491 | case Builtin::BI__builtin_sqrtf: |
| 2492 | case Builtin::BI__builtin_sqrtf16: |
| 2493 | case Builtin::BI__builtin_sqrtl: |
| 2494 | case Builtin::BI__builtin_sqrtf128: |
| 2495 | return RValue::get(emitUnaryMaybeConstrainedFPBuiltin(*this, E, |
| 2496 | Intrinsic::sqrt, |
| 2497 | Intrinsic::experimental_constrained_sqrt)); |
| 2498 | |
| 2499 | case Builtin::BItrunc: |
| 2500 | case Builtin::BItruncf: |
| 2501 | case Builtin::BItruncl: |
| 2502 | case Builtin::BI__builtin_trunc: |
| 2503 | case Builtin::BI__builtin_truncf: |
| 2504 | case Builtin::BI__builtin_truncf16: |
| 2505 | case Builtin::BI__builtin_truncl: |
| 2506 | case Builtin::BI__builtin_truncf128: |
| 2507 | return RValue::get(emitUnaryMaybeConstrainedFPBuiltin(*this, E, |
| 2508 | Intrinsic::trunc, |
| 2509 | Intrinsic::experimental_constrained_trunc)); |
| 2510 | |
| 2511 | case Builtin::BIlround: |
| 2512 | case Builtin::BIlroundf: |
| 2513 | case Builtin::BIlroundl: |
| 2514 | case Builtin::BI__builtin_lround: |
| 2515 | case Builtin::BI__builtin_lroundf: |
| 2516 | case Builtin::BI__builtin_lroundl: |
| 2517 | case Builtin::BI__builtin_lroundf128: |
| 2518 | return RValue::get(emitMaybeConstrainedFPToIntRoundBuiltin( |
| 2519 | *this, E, Intrinsic::lround, |
| 2520 | Intrinsic::experimental_constrained_lround)); |
| 2521 | |
| 2522 | case Builtin::BIllround: |
| 2523 | case Builtin::BIllroundf: |
| 2524 | case Builtin::BIllroundl: |
| 2525 | case Builtin::BI__builtin_llround: |
| 2526 | case Builtin::BI__builtin_llroundf: |
| 2527 | case Builtin::BI__builtin_llroundl: |
| 2528 | case Builtin::BI__builtin_llroundf128: |
| 2529 | return RValue::get(emitMaybeConstrainedFPToIntRoundBuiltin( |
| 2530 | *this, E, Intrinsic::llround, |
| 2531 | Intrinsic::experimental_constrained_llround)); |
| 2532 | |
| 2533 | case Builtin::BIlrint: |
| 2534 | case Builtin::BIlrintf: |
| 2535 | case Builtin::BIlrintl: |
| 2536 | case Builtin::BI__builtin_lrint: |
| 2537 | case Builtin::BI__builtin_lrintf: |
| 2538 | case Builtin::BI__builtin_lrintl: |
| 2539 | case Builtin::BI__builtin_lrintf128: |
| 2540 | return RValue::get(emitMaybeConstrainedFPToIntRoundBuiltin( |
| 2541 | *this, E, Intrinsic::lrint, |
| 2542 | Intrinsic::experimental_constrained_lrint)); |
| 2543 | |
| 2544 | case Builtin::BIllrint: |
| 2545 | case Builtin::BIllrintf: |
| 2546 | case Builtin::BIllrintl: |
| 2547 | case Builtin::BI__builtin_llrint: |
| 2548 | case Builtin::BI__builtin_llrintf: |
| 2549 | case Builtin::BI__builtin_llrintl: |
| 2550 | case Builtin::BI__builtin_llrintf128: |
| 2551 | return RValue::get(emitMaybeConstrainedFPToIntRoundBuiltin( |
| 2552 | *this, E, Intrinsic::llrint, |
| 2553 | Intrinsic::experimental_constrained_llrint)); |
| 2554 | |
| 2555 | default: |
| 2556 | break; |
| 2557 | } |
| 2558 | } |
| 2559 | |
| 2560 | switch (BuiltinIDIfNoAsmLabel) { |
| 2561 | default: break; |
| 2562 | case Builtin::BI__builtin___CFStringMakeConstantString: |
| 2563 | case Builtin::BI__builtin___NSStringMakeConstantString: |
| 2564 | return RValue::get(ConstantEmitter(*this).emitAbstract(E, E->getType())); |
| 2565 | case Builtin::BI__builtin_stdarg_start: |
| 2566 | case Builtin::BI__builtin_va_start: |
| 2567 | case Builtin::BI__va_start: |
| 2568 | case Builtin::BI__builtin_va_end: |
| 2569 | EmitVAStartEnd(BuiltinID == Builtin::BI__va_start |
| 2570 | ? EmitScalarExpr(E->getArg(0)) |
| 2571 | : EmitVAListRef(E->getArg(0)).getPointer(), |
| 2572 | BuiltinID != Builtin::BI__builtin_va_end); |
| 2573 | return RValue::get(nullptr); |
| 2574 | case Builtin::BI__builtin_va_copy: { |
| 2575 | Value *DstPtr = EmitVAListRef(E->getArg(0)).getPointer(); |
| 2576 | Value *SrcPtr = EmitVAListRef(E->getArg(1)).getPointer(); |
| 2577 | |
| 2578 | llvm::Type *Type = Int8PtrTy; |
| 2579 | |
| 2580 | DstPtr = Builder.CreateBitCast(DstPtr, Type); |
| 2581 | SrcPtr = Builder.CreateBitCast(SrcPtr, Type); |
| 2582 | Builder.CreateCall(CGM.getIntrinsic(Intrinsic::vacopy), {DstPtr, SrcPtr}); |
| 2583 | return RValue::get(nullptr); |
| 2584 | } |
| 2585 | case Builtin::BI__builtin_abs: |
| 2586 | case Builtin::BI__builtin_labs: |
| 2587 | case Builtin::BI__builtin_llabs: { |
| 2588 | // X < 0 ? -X : X |
| 2589 | // The negation has 'nsw' because abs of INT_MIN is undefined. |
| 2590 | Value *ArgValue = EmitScalarExpr(E->getArg(0)); |
| 2591 | Value *NegOp = Builder.CreateNSWNeg(ArgValue, "neg"); |
| 2592 | Constant *Zero = llvm::Constant::getNullValue(ArgValue->getType()); |
| 2593 | Value *CmpResult = Builder.CreateICmpSLT(ArgValue, Zero, "abscond"); |
| 2594 | Value *Result = Builder.CreateSelect(CmpResult, NegOp, ArgValue, "abs"); |
| 2595 | return RValue::get(Result); |
| 2596 | } |
| 2597 | case Builtin::BI__builtin_complex: { |
| 2598 | Value *Real = EmitScalarExpr(E->getArg(0)); |
| 2599 | Value *Imag = EmitScalarExpr(E->getArg(1)); |
| 2600 | return RValue::getComplex({Real, Imag}); |
| 2601 | } |
| 2602 | case Builtin::BI__builtin_conj: |
| 2603 | case Builtin::BI__builtin_conjf: |
| 2604 | case Builtin::BI__builtin_conjl: |
| 2605 | case Builtin::BIconj: |
| 2606 | case Builtin::BIconjf: |
| 2607 | case Builtin::BIconjl: { |
| 2608 | ComplexPairTy ComplexVal = EmitComplexExpr(E->getArg(0)); |
| 2609 | Value *Real = ComplexVal.first; |
| 2610 | Value *Imag = ComplexVal.second; |
| 2611 | Imag = Builder.CreateFNeg(Imag, "neg"); |
| 2612 | return RValue::getComplex(std::make_pair(Real, Imag)); |
| 2613 | } |
| 2614 | case Builtin::BI__builtin_creal: |
| 2615 | case Builtin::BI__builtin_crealf: |
| 2616 | case Builtin::BI__builtin_creall: |
| 2617 | case Builtin::BIcreal: |
| 2618 | case Builtin::BIcrealf: |
| 2619 | case Builtin::BIcreall: { |
| 2620 | ComplexPairTy ComplexVal = EmitComplexExpr(E->getArg(0)); |
| 2621 | return RValue::get(ComplexVal.first); |
| 2622 | } |
| 2623 | |
| 2624 | case Builtin::BI__builtin_preserve_access_index: { |
| 2625 | // Only enabled preserved access index region when debuginfo |
| 2626 | // is available as debuginfo is needed to preserve user-level |
| 2627 | // access pattern. |
| 2628 | if (!getDebugInfo()) { |
| 2629 | CGM.Error(E->getExprLoc(), "using builtin_preserve_access_index() without -g"); |
| 2630 | return RValue::get(EmitScalarExpr(E->getArg(0))); |
| 2631 | } |
| 2632 | |
| 2633 | // Nested builtin_preserve_access_index() not supported |
| 2634 | if (IsInPreservedAIRegion) { |
| 2635 | CGM.Error(E->getExprLoc(), "nested builtin_preserve_access_index() not supported"); |
| 2636 | return RValue::get(EmitScalarExpr(E->getArg(0))); |
| 2637 | } |
| 2638 | |
| 2639 | IsInPreservedAIRegion = true; |
| 2640 | Value *Res = EmitScalarExpr(E->getArg(0)); |
| 2641 | IsInPreservedAIRegion = false; |
| 2642 | return RValue::get(Res); |
| 2643 | } |
| 2644 | |
| 2645 | case Builtin::BI__builtin_cimag: |
| 2646 | case Builtin::BI__builtin_cimagf: |
| 2647 | case Builtin::BI__builtin_cimagl: |
| 2648 | case Builtin::BIcimag: |
| 2649 | case Builtin::BIcimagf: |
| 2650 | case Builtin::BIcimagl: { |
| 2651 | ComplexPairTy ComplexVal = EmitComplexExpr(E->getArg(0)); |
| 2652 | return RValue::get(ComplexVal.second); |
| 2653 | } |
| 2654 | |
| 2655 | case Builtin::BI__builtin_clrsb: |
| 2656 | case Builtin::BI__builtin_clrsbl: |
| 2657 | case Builtin::BI__builtin_clrsbll: { |
| 2658 | // clrsb(x) -> clz(x < 0 ? ~x : x) - 1 or |
| 2659 | Value *ArgValue = EmitScalarExpr(E->getArg(0)); |
| 2660 | |
| 2661 | llvm::Type *ArgType = ArgValue->getType(); |
| 2662 | Function *F = CGM.getIntrinsic(Intrinsic::ctlz, ArgType); |
| 2663 | |
| 2664 | llvm::Type *ResultType = ConvertType(E->getType()); |
| 2665 | Value *Zero = llvm::Constant::getNullValue(ArgType); |
| 2666 | Value *IsNeg = Builder.CreateICmpSLT(ArgValue, Zero, "isneg"); |
| 2667 | Value *Inverse = Builder.CreateNot(ArgValue, "not"); |
| 2668 | Value *Tmp = Builder.CreateSelect(IsNeg, Inverse, ArgValue); |
| 2669 | Value *Ctlz = Builder.CreateCall(F, {Tmp, Builder.getFalse()}); |
| 2670 | Value *Result = Builder.CreateSub(Ctlz, llvm::ConstantInt::get(ArgType, 1)); |
| 2671 | Result = Builder.CreateIntCast(Result, ResultType, /*isSigned*/true, |
| 2672 | "cast"); |
| 2673 | return RValue::get(Result); |
| 2674 | } |
| 2675 | case Builtin::BI__builtin_ctzs: |
| 2676 | case Builtin::BI__builtin_ctz: |
| 2677 | case Builtin::BI__builtin_ctzl: |
| 2678 | case Builtin::BI__builtin_ctzll: { |
| 2679 | Value *ArgValue = EmitCheckedArgForBuiltin(E->getArg(0), BCK_CTZPassedZero); |
| 2680 | |
| 2681 | llvm::Type *ArgType = ArgValue->getType(); |
| 2682 | Function *F = CGM.getIntrinsic(Intrinsic::cttz, ArgType); |
| 2683 | |
| 2684 | llvm::Type *ResultType = ConvertType(E->getType()); |
| 2685 | Value *ZeroUndef = Builder.getInt1(getTarget().isCLZForZeroUndef()); |
| 2686 | Value *Result = Builder.CreateCall(F, {ArgValue, ZeroUndef}); |
| 2687 | if (Result->getType() != ResultType) |
| 2688 | Result = Builder.CreateIntCast(Result, ResultType, /*isSigned*/true, |
| 2689 | "cast"); |
| 2690 | return RValue::get(Result); |
| 2691 | } |
| 2692 | case Builtin::BI__builtin_clzs: |
| 2693 | case Builtin::BI__builtin_clz: |
| 2694 | case Builtin::BI__builtin_clzl: |
| 2695 | case Builtin::BI__builtin_clzll: { |
| 2696 | Value *ArgValue = EmitCheckedArgForBuiltin(E->getArg(0), BCK_CLZPassedZero); |
| 2697 | |
| 2698 | llvm::Type *ArgType = ArgValue->getType(); |
| 2699 | Function *F = CGM.getIntrinsic(Intrinsic::ctlz, ArgType); |
| 2700 | |
| 2701 | llvm::Type *ResultType = ConvertType(E->getType()); |
| 2702 | Value *ZeroUndef = Builder.getInt1(getTarget().isCLZForZeroUndef()); |
| 2703 | Value *Result = Builder.CreateCall(F, {ArgValue, ZeroUndef}); |
| 2704 | if (Result->getType() != ResultType) |
| 2705 | Result = Builder.CreateIntCast(Result, ResultType, /*isSigned*/true, |
| 2706 | "cast"); |
| 2707 | return RValue::get(Result); |
| 2708 | } |
| 2709 | case Builtin::BI__builtin_ffs: |
| 2710 | case Builtin::BI__builtin_ffsl: |
| 2711 | case Builtin::BI__builtin_ffsll: { |
| 2712 | // ffs(x) -> x ? cttz(x) + 1 : 0 |
| 2713 | Value *ArgValue = EmitScalarExpr(E->getArg(0)); |
| 2714 | |
| 2715 | llvm::Type *ArgType = ArgValue->getType(); |
| 2716 | Function *F = CGM.getIntrinsic(Intrinsic::cttz, ArgType); |
| 2717 | |
| 2718 | llvm::Type *ResultType = ConvertType(E->getType()); |
| 2719 | Value *Tmp = |
| 2720 | Builder.CreateAdd(Builder.CreateCall(F, {ArgValue, Builder.getTrue()}), |
| 2721 | llvm::ConstantInt::get(ArgType, 1)); |
| 2722 | Value *Zero = llvm::Constant::getNullValue(ArgType); |
| 2723 | Value *IsZero = Builder.CreateICmpEQ(ArgValue, Zero, "iszero"); |
| 2724 | Value *Result = Builder.CreateSelect(IsZero, Zero, Tmp, "ffs"); |
| 2725 | if (Result->getType() != ResultType) |
| 2726 | Result = Builder.CreateIntCast(Result, ResultType, /*isSigned*/true, |
| 2727 | "cast"); |
| 2728 | return RValue::get(Result); |
| 2729 | } |
| 2730 | case Builtin::BI__builtin_parity: |
| 2731 | case Builtin::BI__builtin_parityl: |
| 2732 | case Builtin::BI__builtin_parityll: { |
| 2733 | // parity(x) -> ctpop(x) & 1 |
| 2734 | Value *ArgValue = EmitScalarExpr(E->getArg(0)); |
| 2735 | |
| 2736 | llvm::Type *ArgType = ArgValue->getType(); |
| 2737 | Function *F = CGM.getIntrinsic(Intrinsic::ctpop, ArgType); |
| 2738 | |
| 2739 | llvm::Type *ResultType = ConvertType(E->getType()); |
| 2740 | Value *Tmp = Builder.CreateCall(F, ArgValue); |
| 2741 | Value *Result = Builder.CreateAnd(Tmp, llvm::ConstantInt::get(ArgType, 1)); |
| 2742 | if (Result->getType() != ResultType) |
| 2743 | Result = Builder.CreateIntCast(Result, ResultType, /*isSigned*/true, |
| 2744 | "cast"); |
| 2745 | return RValue::get(Result); |
| 2746 | } |
| 2747 | case Builtin::BI__lzcnt16: |
| 2748 | case Builtin::BI__lzcnt: |
| 2749 | case Builtin::BI__lzcnt64: { |
| 2750 | Value *ArgValue = EmitScalarExpr(E->getArg(0)); |
| 2751 | |
| 2752 | llvm::Type *ArgType = ArgValue->getType(); |
| 2753 | Function *F = CGM.getIntrinsic(Intrinsic::ctlz, ArgType); |
| 2754 | |
| 2755 | llvm::Type *ResultType = ConvertType(E->getType()); |
| 2756 | Value *Result = Builder.CreateCall(F, {ArgValue, Builder.getFalse()}); |
| 2757 | if (Result->getType() != ResultType) |
| 2758 | Result = Builder.CreateIntCast(Result, ResultType, /*isSigned*/true, |
| 2759 | "cast"); |
| 2760 | return RValue::get(Result); |
| 2761 | } |
| 2762 | case Builtin::BI__popcnt16: |
| 2763 | case Builtin::BI__popcnt: |
| 2764 | case Builtin::BI__popcnt64: |
| 2765 | case Builtin::BI__builtin_popcount: |
| 2766 | case Builtin::BI__builtin_popcountl: |
| 2767 | case Builtin::BI__builtin_popcountll: { |
| 2768 | Value *ArgValue = EmitScalarExpr(E->getArg(0)); |
| 2769 | |
| 2770 | llvm::Type *ArgType = ArgValue->getType(); |
| 2771 | Function *F = CGM.getIntrinsic(Intrinsic::ctpop, ArgType); |
| 2772 | |
| 2773 | llvm::Type *ResultType = ConvertType(E->getType()); |
| 2774 | Value *Result = Builder.CreateCall(F, ArgValue); |
| 2775 | if (Result->getType() != ResultType) |
| 2776 | Result = Builder.CreateIntCast(Result, ResultType, /*isSigned*/true, |
| 2777 | "cast"); |
| 2778 | return RValue::get(Result); |
| 2779 | } |
| 2780 | case Builtin::BI__builtin_unpredictable: { |
| 2781 | // Always return the argument of __builtin_unpredictable. LLVM does not |
| 2782 | // handle this builtin. Metadata for this builtin should be added directly |
| 2783 | // to instructions such as branches or switches that use it. |
| 2784 | return RValue::get(EmitScalarExpr(E->getArg(0))); |
| 2785 | } |
| 2786 | case Builtin::BI__builtin_expect: { |
| 2787 | Value *ArgValue = EmitScalarExpr(E->getArg(0)); |
| 2788 | llvm::Type *ArgType = ArgValue->getType(); |
| 2789 | |
| 2790 | Value *ExpectedValue = EmitScalarExpr(E->getArg(1)); |
| 2791 | // Don't generate llvm.expect on -O0 as the backend won't use it for |
| 2792 | // anything. |
| 2793 | // Note, we still IRGen ExpectedValue because it could have side-effects. |
| 2794 | if (CGM.getCodeGenOpts().OptimizationLevel == 0) |
| 2795 | return RValue::get(ArgValue); |
| 2796 | |
| 2797 | Function *FnExpect = CGM.getIntrinsic(Intrinsic::expect, ArgType); |
| 2798 | Value *Result = |
| 2799 | Builder.CreateCall(FnExpect, {ArgValue, ExpectedValue}, "expval"); |
| 2800 | return RValue::get(Result); |
| 2801 | } |
| 2802 | case Builtin::BI__builtin_expect_with_probability: { |
| 2803 | Value *ArgValue = EmitScalarExpr(E->getArg(0)); |
| 2804 | llvm::Type *ArgType = ArgValue->getType(); |
| 2805 | |
| 2806 | Value *ExpectedValue = EmitScalarExpr(E->getArg(1)); |
| 2807 | llvm::APFloat Probability(0.0); |
| 2808 | const Expr *ProbArg = E->getArg(2); |
| 2809 | bool EvalSucceed = ProbArg->EvaluateAsFloat(Probability, CGM.getContext()); |
| 2810 | assert(EvalSucceed && "probability should be able to evaluate as float")(static_cast <bool> (EvalSucceed && "probability should be able to evaluate as float" ) ? void (0) : __assert_fail ("EvalSucceed && \"probability should be able to evaluate as float\"" , "clang/lib/CodeGen/CGBuiltin.cpp", 2810, __extension__ __PRETTY_FUNCTION__ )); |
| 2811 | (void)EvalSucceed; |
| 2812 | bool LoseInfo = false; |
| 2813 | Probability.convert(llvm::APFloat::IEEEdouble(), |
| 2814 | llvm::RoundingMode::Dynamic, &LoseInfo); |
| 2815 | llvm::Type *Ty = ConvertType(ProbArg->getType()); |
| 2816 | Constant *Confidence = ConstantFP::get(Ty, Probability); |
| 2817 | // Don't generate llvm.expect.with.probability on -O0 as the backend |
| 2818 | // won't use it for anything. |
| 2819 | // Note, we still IRGen ExpectedValue because it could have side-effects. |
| 2820 | if (CGM.getCodeGenOpts().OptimizationLevel == 0) |
| 2821 | return RValue::get(ArgValue); |
| 2822 | |
| 2823 | Function *FnExpect = |
| 2824 | CGM.getIntrinsic(Intrinsic::expect_with_probability, ArgType); |
| 2825 | Value *Result = Builder.CreateCall( |
| 2826 | FnExpect, {ArgValue, ExpectedValue, Confidence}, "expval"); |
| 2827 | return RValue::get(Result); |
| 2828 | } |
| 2829 | case Builtin::BI__builtin_assume_aligned: { |
| 2830 | const Expr *Ptr = E->getArg(0); |
| 2831 | Value *PtrValue = EmitScalarExpr(Ptr); |
| 2832 | Value *OffsetValue = |
| 2833 | (E->getNumArgs() > 2) ? EmitScalarExpr(E->getArg(2)) : nullptr; |
| 2834 | |
| 2835 | Value *AlignmentValue = EmitScalarExpr(E->getArg(1)); |
| 2836 | ConstantInt *AlignmentCI = cast<ConstantInt>(AlignmentValue); |
| 2837 | if (AlignmentCI->getValue().ugt(llvm::Value::MaximumAlignment)) |
| 2838 | AlignmentCI = ConstantInt::get(AlignmentCI->getType(), |
| 2839 | llvm::Value::MaximumAlignment); |
| 2840 | |
| 2841 | emitAlignmentAssumption(PtrValue, Ptr, |
| 2842 | /*The expr loc is sufficient.*/ SourceLocation(), |
| 2843 | AlignmentCI, OffsetValue); |
| 2844 | return RValue::get(PtrValue); |
| 2845 | } |
| 2846 | case Builtin::BI__assume: |
| 2847 | case Builtin::BI__builtin_assume: { |
| 2848 | if (E->getArg(0)->HasSideEffects(getContext())) |
| 2849 | return RValue::get(nullptr); |
| 2850 | |
| 2851 | Value *ArgValue = EmitScalarExpr(E->getArg(0)); |
| 2852 | Function *FnAssume = CGM.getIntrinsic(Intrinsic::assume); |
| 2853 | Builder.CreateCall(FnAssume, ArgValue); |
| 2854 | return RValue::get(nullptr); |
| 2855 | } |
| 2856 | case Builtin::BI__builtin_assume_separate_storage: { |
| 2857 | const Expr *Arg0 = E->getArg(0); |
| 2858 | const Expr *Arg1 = E->getArg(1); |
| 2859 | |
| 2860 | Value *Value0 = EmitScalarExpr(Arg0); |
| 2861 | Value *Value1 = EmitScalarExpr(Arg1); |
| 2862 | |
| 2863 | Value *Values[] = {Value0, Value1}; |
| 2864 | OperandBundleDefT<Value *> OBD("separate_storage", Values); |
| 2865 | Builder.CreateAssumption(ConstantInt::getTrue(getLLVMContext()), {OBD}); |
| 2866 | return RValue::get(nullptr); |
| 2867 | } |
| 2868 | case Builtin::BI__arithmetic_fence: { |
| 2869 | // Create the builtin call if FastMath is selected, and the target |
| 2870 | // supports the builtin, otherwise just return the argument. |
| 2871 | CodeGenFunction::CGFPOptionsRAII FPOptsRAII(*this, E); |
| 2872 | llvm::FastMathFlags FMF = Builder.getFastMathFlags(); |
| 2873 | bool isArithmeticFenceEnabled = |
| 2874 | FMF.allowReassoc() && |
| 2875 | getContext().getTargetInfo().checkArithmeticFenceSupported(); |
| 2876 | QualType ArgType = E->getArg(0)->getType(); |
| 2877 | if (ArgType->isComplexType()) { |
| 2878 | if (isArithmeticFenceEnabled) { |
| 2879 | QualType ElementType = ArgType->castAs<ComplexType>()->getElementType(); |
| 2880 | ComplexPairTy ComplexVal = EmitComplexExpr(E->getArg(0)); |
| 2881 | Value *Real = Builder.CreateArithmeticFence(ComplexVal.first, |
| 2882 | ConvertType(ElementType)); |
| 2883 | Value *Imag = Builder.CreateArithmeticFence(ComplexVal.second, |
| 2884 | ConvertType(ElementType)); |
| 2885 | return RValue::getComplex(std::make_pair(Real, Imag)); |
| 2886 | } |
| 2887 | ComplexPairTy ComplexVal = EmitComplexExpr(E->getArg(0)); |
| 2888 | Value *Real = ComplexVal.first; |
| 2889 | Value *Imag = ComplexVal.second; |
| 2890 | return RValue::getComplex(std::make_pair(Real, Imag)); |
| 2891 | } |
| 2892 | Value *ArgValue = EmitScalarExpr(E->getArg(0)); |
| 2893 | if (isArithmeticFenceEnabled) |
| 2894 | return RValue::get( |
| 2895 | Builder.CreateArithmeticFence(ArgValue, ConvertType(ArgType))); |
| 2896 | return RValue::get(ArgValue); |
| 2897 | } |
| 2898 | case Builtin::BI__builtin_bswap16: |
| 2899 | case Builtin::BI__builtin_bswap32: |
| 2900 | case Builtin::BI__builtin_bswap64: |
| 2901 | case Builtin::BI_byteswap_ushort: |
| 2902 | case Builtin::BI_byteswap_ulong: |
| 2903 | case Builtin::BI_byteswap_uint64: { |
| 2904 | return RValue::get(emitUnaryBuiltin(*this, E, Intrinsic::bswap)); |
| 2905 | } |
| 2906 | case Builtin::BI__builtin_bitreverse8: |
| 2907 | case Builtin::BI__builtin_bitreverse16: |
| 2908 | case Builtin::BI__builtin_bitreverse32: |
| 2909 | case Builtin::BI__builtin_bitreverse64: { |
| 2910 | return RValue::get(emitUnaryBuiltin(*this, E, Intrinsic::bitreverse)); |
| 2911 | } |
| 2912 | case Builtin::BI__builtin_rotateleft8: |
| 2913 | case Builtin::BI__builtin_rotateleft16: |
| 2914 | case Builtin::BI__builtin_rotateleft32: |
| 2915 | case Builtin::BI__builtin_rotateleft64: |
| 2916 | case Builtin::BI_rotl8: // Microsoft variants of rotate left |
| 2917 | case Builtin::BI_rotl16: |
| 2918 | case Builtin::BI_rotl: |
| 2919 | case Builtin::BI_lrotl: |
| 2920 | case Builtin::BI_rotl64: |
| 2921 | return emitRotate(E, false); |
| 2922 | |
| 2923 | case Builtin::BI__builtin_rotateright8: |
| 2924 | case Builtin::BI__builtin_rotateright16: |
| 2925 | case Builtin::BI__builtin_rotateright32: |
| 2926 | case Builtin::BI__builtin_rotateright64: |
| 2927 | case Builtin::BI_rotr8: // Microsoft variants of rotate right |
| 2928 | case Builtin::BI_rotr16: |
| 2929 | case Builtin::BI_rotr: |
| 2930 | case Builtin::BI_lrotr: |
| 2931 | case Builtin::BI_rotr64: |
| 2932 | return emitRotate(E, true); |
| 2933 | |
| 2934 | case Builtin::BI__builtin_constant_p: { |
| 2935 | llvm::Type *ResultType = ConvertType(E->getType()); |
| 2936 | |
| 2937 | const Expr *Arg = E->getArg(0); |
| 2938 | QualType ArgType = Arg->getType(); |
| 2939 | // FIXME: The allowance for Obj-C pointers and block pointers is historical |
| 2940 | // and likely a mistake. |
| 2941 | if (!ArgType->isIntegralOrEnumerationType() && !ArgType->isFloatingType() && |
| 2942 | !ArgType->isObjCObjectPointerType() && !ArgType->isBlockPointerType()) |
| 2943 | // Per the GCC documentation, only numeric constants are recognized after |
| 2944 | // inlining. |
| 2945 | return RValue::get(ConstantInt::get(ResultType, 0)); |
| 2946 | |
| 2947 | if (Arg->HasSideEffects(getContext())) |
| 2948 | // The argument is unevaluated, so be conservative if it might have |
| 2949 | // side-effects. |
| 2950 | return RValue::get(ConstantInt::get(ResultType, 0)); |
| 2951 | |
| 2952 | Value *ArgValue = EmitScalarExpr(Arg); |
| 2953 | if (ArgType->isObjCObjectPointerType()) { |
| 2954 | // Convert Objective-C objects to id because we cannot distinguish between |
| 2955 | // LLVM types for Obj-C classes as they are opaque. |
| 2956 | ArgType = CGM.getContext().getObjCIdType(); |
| 2957 | ArgValue = Builder.CreateBitCast(ArgValue, ConvertType(ArgType)); |
| 2958 | } |
| 2959 | Function *F = |
| 2960 | CGM.getIntrinsic(Intrinsic::is_constant, ConvertType(ArgType)); |
| 2961 | Value *Result = Builder.CreateCall(F, ArgValue); |
| 2962 | if (Result->getType() != ResultType) |
| 2963 | Result = Builder.CreateIntCast(Result, ResultType, /*isSigned*/false); |
| 2964 | return RValue::get(Result); |
| 2965 | } |
| 2966 | case Builtin::BI__builtin_dynamic_object_size: |
| 2967 | case Builtin::BI__builtin_object_size: { |
| 2968 | unsigned Type = |
| 2969 | E->getArg(1)->EvaluateKnownConstInt(getContext()).getZExtValue(); |
| 2970 | auto *ResType = cast<llvm::IntegerType>(ConvertType(E->getType())); |
| 2971 | |
| 2972 | // We pass this builtin onto the optimizer so that it can figure out the |
| 2973 | // object size in more complex cases. |
| 2974 | bool IsDynamic = BuiltinID == Builtin::BI__builtin_dynamic_object_size; |
| 2975 | return RValue::get(emitBuiltinObjectSize(E->getArg(0), Type, ResType, |
| 2976 | /*EmittedE=*/nullptr, IsDynamic)); |
| 2977 | } |
| 2978 | case Builtin::BI__builtin_prefetch: { |
| 2979 | Value *Locality, *RW, *Address = EmitScalarExpr(E->getArg(0)); |
| 2980 | // FIXME: Technically these constants should of type 'int', yes? |
| 2981 | RW = (E->getNumArgs() > 1) ? EmitScalarExpr(E->getArg(1)) : |
| 2982 | llvm::ConstantInt::get(Int32Ty, 0); |
| 2983 | Locality = (E->getNumArgs() > 2) ? EmitScalarExpr(E->getArg(2)) : |
| 2984 | llvm::ConstantInt::get(Int32Ty, 3); |
| 2985 | Value *Data = llvm::ConstantInt::get(Int32Ty, 1); |
| 2986 | Function *F = CGM.getIntrinsic(Intrinsic::prefetch, Address->getType()); |
| 2987 | Builder.CreateCall(F, {Address, RW, Locality, Data}); |
| 2988 | return RValue::get(nullptr); |
| 2989 | } |
| 2990 | case Builtin::BI__builtin_readcyclecounter: { |
| 2991 | Function *F = CGM.getIntrinsic(Intrinsic::readcyclecounter); |
| 2992 | return RValue::get(Builder.CreateCall(F)); |
| 2993 | } |
| 2994 | case Builtin::BI__builtin___clear_cache: { |
| 2995 | Value *Begin = EmitScalarExpr(E->getArg(0)); |
| 2996 | Value *End = EmitScalarExpr(E->getArg(1)); |
| 2997 | Function *F = CGM.getIntrinsic(Intrinsic::clear_cache); |
| 2998 | return RValue::get(Builder.CreateCall(F, {Begin, End})); |
| 2999 | } |
| 3000 | case Builtin::BI__builtin_trap: |
| 3001 | EmitTrapCall(Intrinsic::trap); |
| 3002 | return RValue::get(nullptr); |
| 3003 | case Builtin::BI__debugbreak: |
| 3004 | EmitTrapCall(Intrinsic::debugtrap); |
| 3005 | return RValue::get(nullptr); |
| 3006 | case Builtin::BI__builtin_unreachable: { |
| 3007 | EmitUnreachable(E->getExprLoc()); |
| 3008 | |
| 3009 | // We do need to preserve an insertion point. |
| 3010 | EmitBlock(createBasicBlock("unreachable.cont")); |
| 3011 | |
| 3012 | return RValue::get(nullptr); |
| 3013 | } |
| 3014 | |
| 3015 | case Builtin::BI__builtin_powi: |
| 3016 | case Builtin::BI__builtin_powif: |
| 3017 | case Builtin::BI__builtin_powil: { |
| 3018 | llvm::Value *Src0 = EmitScalarExpr(E->getArg(0)); |
| 3019 | llvm::Value *Src1 = EmitScalarExpr(E->getArg(1)); |
| 3020 | |
| 3021 | if (Builder.getIsFPConstrained()) { |
| 3022 | CodeGenFunction::CGFPOptionsRAII FPOptsRAII(*this, E); |
| 3023 | Function *F = CGM.getIntrinsic(Intrinsic::experimental_constrained_powi, |
| 3024 | Src0->getType()); |
| 3025 | return RValue::get(Builder.CreateConstrainedFPCall(F, { Src0, Src1 })); |
| 3026 | } |
| 3027 | |
| 3028 | Function *F = CGM.getIntrinsic(Intrinsic::powi, |
| 3029 | { Src0->getType(), Src1->getType() }); |
| 3030 | return RValue::get(Builder.CreateCall(F, { Src0, Src1 })); |
| 3031 | } |
| 3032 | case Builtin::BI__builtin_isgreater: |
| 3033 | case Builtin::BI__builtin_isgreaterequal: |
| 3034 | case Builtin::BI__builtin_isless: |
| 3035 | case Builtin::BI__builtin_islessequal: |
| 3036 | case Builtin::BI__builtin_islessgreater: |
| 3037 | case Builtin::BI__builtin_isunordered: { |
| 3038 | // Ordered comparisons: we know the arguments to these are matching scalar |
| 3039 | // floating point values. |
| 3040 | CodeGenFunction::CGFPOptionsRAII FPOptsRAII(*this, E); |
| 3041 | Value *LHS = EmitScalarExpr(E->getArg(0)); |
| 3042 | Value *RHS = EmitScalarExpr(E->getArg(1)); |
| 3043 | |
| 3044 | switch (BuiltinID) { |
| 3045 | default: llvm_unreachable("Unknown ordered comparison")::llvm::llvm_unreachable_internal("Unknown ordered comparison" , "clang/lib/CodeGen/CGBuiltin.cpp", 3045); |
| 3046 | case Builtin::BI__builtin_isgreater: |
| 3047 | LHS = Builder.CreateFCmpOGT(LHS, RHS, "cmp"); |
| 3048 | break; |
| 3049 | case Builtin::BI__builtin_isgreaterequal: |
| 3050 | LHS = Builder.CreateFCmpOGE(LHS, RHS, "cmp"); |
| 3051 | break; |
| 3052 | case Builtin::BI__builtin_isless: |
| 3053 | LHS = Builder.CreateFCmpOLT(LHS, RHS, "cmp"); |
| 3054 | break; |
| 3055 | case Builtin::BI__builtin_islessequal: |
| 3056 | LHS = Builder.CreateFCmpOLE(LHS, RHS, "cmp"); |
| 3057 | break; |
| 3058 | case Builtin::BI__builtin_islessgreater: |
| 3059 | LHS = Builder.CreateFCmpONE(LHS, RHS, "cmp"); |
| 3060 | break; |
| 3061 | case Builtin::BI__builtin_isunordered: |
| 3062 | LHS = Builder.CreateFCmpUNO(LHS, RHS, "cmp"); |
| 3063 | break; |
| 3064 | } |
| 3065 | // ZExt bool to int type. |
| 3066 | return RValue::get(Builder.CreateZExt(LHS, ConvertType(E->getType()))); |
| 3067 | } |
| 3068 | case Builtin::BI__builtin_isnan: { |
| 3069 | CodeGenFunction::CGFPOptionsRAII FPOptsRAII(*this, E); |
| 3070 | Value *V = EmitScalarExpr(E->getArg(0)); |
| 3071 | llvm::Type *Ty = V->getType(); |
| 3072 | const llvm::fltSemantics &Semantics = Ty->getFltSemantics(); |
| 3073 | if (!Builder.getIsFPConstrained() || |
| 3074 | Builder.getDefaultConstrainedExcept() == fp::ebIgnore || |
| 3075 | !Ty->isIEEE()) { |
| 3076 | V = Builder.CreateFCmpUNO(V, V, "cmp"); |
| 3077 | return RValue::get(Builder.CreateZExt(V, ConvertType(E->getType()))); |
| 3078 | } |
| 3079 | |
| 3080 | if (Value *Result = getTargetHooks().testFPKind(V, BuiltinID, Builder, CGM)) |
| 3081 | return RValue::get(Result); |
| 3082 | |
| 3083 | // NaN has all exp bits set and a non zero significand. Therefore: |
| 3084 | // isnan(V) == ((exp mask - (abs(V) & exp mask)) < 0) |
| 3085 | unsigned bitsize = Ty->getScalarSizeInBits(); |
| 3086 | llvm::IntegerType *IntTy = Builder.getIntNTy(bitsize); |
| 3087 | Value *IntV = Builder.CreateBitCast(V, IntTy); |
| 3088 | APInt AndMask = APInt::getSignedMaxValue(bitsize); |
| 3089 | Value *AbsV = |
| 3090 | Builder.CreateAnd(IntV, llvm::ConstantInt::get(IntTy, AndMask)); |
| 3091 | APInt ExpMask = APFloat::getInf(Semantics).bitcastToAPInt(); |
| 3092 | Value *Sub = |
| 3093 | Builder.CreateSub(llvm::ConstantInt::get(IntTy, ExpMask), AbsV); |
| 3094 | // V = sign bit (Sub) <=> V = (Sub < 0) |
| 3095 | V = Builder.CreateLShr(Sub, llvm::ConstantInt::get(IntTy, bitsize - 1)); |
| 3096 | if (bitsize > 32) |
| 3097 | V = Builder.CreateTrunc(V, ConvertType(E->getType())); |
| 3098 | return RValue::get(V); |
| 3099 | } |
| 3100 | |
| 3101 | case Builtin::BI__builtin_nondeterministic_value: { |
| 3102 | llvm::Type *Ty = ConvertType(E->getArg(0)->getType()); |
| 3103 | |
| 3104 | Value *Result = PoisonValue::get(Ty); |
| 3105 | Result = Builder.CreateFreeze(Result); |
| 3106 | |
| 3107 | return RValue::get(Result); |
| 3108 | } |
| 3109 | |
| 3110 | case Builtin::BI__builtin_elementwise_abs: { |
| 3111 | Value *Result; |
| 3112 | QualType QT = E->getArg(0)->getType(); |
| 3113 | |
| 3114 | if (auto *VecTy = QT->getAs<VectorType>()) |
| 3115 | QT = VecTy->getElementType(); |
| 3116 | if (QT->isIntegerType()) |
| 3117 | Result = Builder.CreateBinaryIntrinsic( |
| 3118 | llvm::Intrinsic::abs, EmitScalarExpr(E->getArg(0)), |
| 3119 | Builder.getFalse(), nullptr, "elt.abs"); |
| 3120 | else |
| 3121 | Result = emitUnaryBuiltin(*this, E, llvm::Intrinsic::fabs, "elt.abs"); |
| 3122 | |
| 3123 | return RValue::get(Result); |
| 3124 | } |
| 3125 | |
| 3126 | case Builtin::BI__builtin_elementwise_ceil: |
| 3127 | return RValue::get( |
| 3128 | emitUnaryBuiltin(*this, E, llvm::Intrinsic::ceil, "elt.ceil")); |
| 3129 | case Builtin::BI__builtin_elementwise_exp: |
| 3130 | return RValue::get( |
| 3131 | emitUnaryBuiltin(*this, E, llvm::Intrinsic::exp, "elt.exp")); |
| 3132 | case Builtin::BI__builtin_elementwise_exp2: |
| 3133 | return RValue::get( |
| 3134 | emitUnaryBuiltin(*this, E, llvm::Intrinsic::exp2, "elt.exp2")); |
| 3135 | case Builtin::BI__builtin_elementwise_log: |
| 3136 | return RValue::get( |
| 3137 | emitUnaryBuiltin(*this, E, llvm::Intrinsic::log, "elt.log")); |
| 3138 | case Builtin::BI__builtin_elementwise_log2: |
| 3139 | return RValue::get( |
| 3140 | emitUnaryBuiltin(*this, E, llvm::Intrinsic::log2, "elt.log2")); |
| 3141 | case Builtin::BI__builtin_elementwise_log10: |
| 3142 | return RValue::get( |
| 3143 | emitUnaryBuiltin(*this, E, llvm::Intrinsic::log10, "elt.log10")); |
| 3144 | case Builtin::BI__builtin_elementwise_cos: |
| 3145 | return RValue::get( |
| 3146 | emitUnaryBuiltin(*this, E, llvm::Intrinsic::cos, "elt.cos")); |
| 3147 | case Builtin::BI__builtin_elementwise_floor: |
| 3148 | return RValue::get( |
| 3149 | emitUnaryBuiltin(*this, E, llvm::Intrinsic::floor, "elt.floor")); |
| 3150 | case Builtin::BI__builtin_elementwise_roundeven: |
| 3151 | return RValue::get(emitUnaryBuiltin(*this, E, llvm::Intrinsic::roundeven, |
| 3152 | "elt.roundeven")); |
| 3153 | case Builtin::BI__builtin_elementwise_sin: |
| 3154 | return RValue::get( |
| 3155 | emitUnaryBuiltin(*this, E, llvm::Intrinsic::sin, "elt.sin")); |
| 3156 | |
| 3157 | case Builtin::BI__builtin_elementwise_trunc: |
| 3158 | return RValue::get( |
| 3159 | emitUnaryBuiltin(*this, E, llvm::Intrinsic::trunc, "elt.trunc")); |
| 3160 | case Builtin::BI__builtin_elementwise_canonicalize: |
| 3161 | return RValue::get( |
| 3162 | emitUnaryBuiltin(*this, E, llvm::Intrinsic::canonicalize, "elt.trunc")); |
| 3163 | case Builtin::BI__builtin_elementwise_copysign: |
| 3164 | return RValue::get(emitBinaryBuiltin(*this, E, llvm::Intrinsic::copysign)); |
| 3165 | case Builtin::BI__builtin_elementwise_fma: |
| 3166 | return RValue::get(emitTernaryBuiltin(*this, E, llvm::Intrinsic::fma)); |
| 3167 | case Builtin::BI__builtin_elementwise_add_sat: |
| 3168 | case Builtin::BI__builtin_elementwise_sub_sat: { |
| 3169 | Value *Op0 = EmitScalarExpr(E->getArg(0)); |
| 3170 | Value *Op1 = EmitScalarExpr(E->getArg(1)); |
| 3171 | Value *Result; |
| 3172 | assert(Op0->getType()->isIntOrIntVectorTy() && "integer type expected")(static_cast <bool> (Op0->getType()->isIntOrIntVectorTy () && "integer type expected") ? void (0) : __assert_fail ("Op0->getType()->isIntOrIntVectorTy() && \"integer type expected\"" , "clang/lib/CodeGen/CGBuiltin.cpp", 3172, __extension__ __PRETTY_FUNCTION__ )); |
| 3173 | QualType Ty = E->getArg(0)->getType(); |
| 3174 | if (auto *VecTy = Ty->getAs<VectorType>()) |
| 3175 | Ty = VecTy->getElementType(); |
| 3176 | bool IsSigned = Ty->isSignedIntegerType(); |
| 3177 | unsigned Opc; |
| 3178 | if (BuiltinIDIfNoAsmLabel == Builtin::BI__builtin_elementwise_add_sat) |
| 3179 | Opc = IsSigned ? llvm::Intrinsic::sadd_sat : llvm::Intrinsic::uadd_sat; |
| 3180 | else |
| 3181 | Opc = IsSigned ? llvm::Intrinsic::ssub_sat : llvm::Intrinsic::usub_sat; |
| 3182 | Result = Builder.CreateBinaryIntrinsic(Opc, Op0, Op1, nullptr, "elt.sat"); |
| 3183 | return RValue::get(Result); |
| 3184 | } |
| 3185 | |
| 3186 | case Builtin::BI__builtin_elementwise_max: { |
| 3187 | Value *Op0 = EmitScalarExpr(E->getArg(0)); |
| 3188 | Value *Op1 = EmitScalarExpr(E->getArg(1)); |
| 3189 | Value *Result; |
| 3190 | if (Op0->getType()->isIntOrIntVectorTy()) { |
| 3191 | QualType Ty = E->getArg(0)->getType(); |
| 3192 | if (auto *VecTy = Ty->getAs<VectorType>()) |
| 3193 | Ty = VecTy->getElementType(); |
| 3194 | Result = Builder.CreateBinaryIntrinsic(Ty->isSignedIntegerType() |
| 3195 | ? llvm::Intrinsic::smax |
| 3196 | : llvm::Intrinsic::umax, |
| 3197 | Op0, Op1, nullptr, "elt.max"); |
| 3198 | } else |
| 3199 | Result = Builder.CreateMaxNum(Op0, Op1, "elt.max"); |
| 3200 | return RValue::get(Result); |
| 3201 | } |
| 3202 | case Builtin::BI__builtin_elementwise_min: { |
| 3203 | Value *Op0 = EmitScalarExpr(E->getArg(0)); |
| 3204 | Value *Op1 = EmitScalarExpr(E->getArg(1)); |
| 3205 | Value *Result; |
| 3206 | if (Op0->getType()->isIntOrIntVectorTy()) { |
| 3207 | QualType Ty = E->getArg(0)->getType(); |
| 3208 | if (auto *VecTy = Ty->getAs<VectorType>()) |
| 3209 | Ty = VecTy->getElementType(); |
| 3210 | Result = Builder.CreateBinaryIntrinsic(Ty->isSignedIntegerType() |
| 3211 | ? llvm::Intrinsic::smin |
| 3212 | : llvm::Intrinsic::umin, |
| 3213 | Op0, Op1, nullptr, "elt.min"); |
| 3214 | } else |
| 3215 | Result = Builder.CreateMinNum(Op0, Op1, "elt.min"); |
| 3216 | return RValue::get(Result); |
| 3217 | } |
| 3218 | |
| 3219 | case Builtin::BI__builtin_reduce_max: { |
| 3220 | auto GetIntrinsicID = [](QualType QT) { |
| 3221 | if (auto *VecTy = QT->getAs<VectorType>()) |
| 3222 | QT = VecTy->getElementType(); |
| 3223 | if (QT->isSignedIntegerType()) |
| 3224 | return llvm::Intrinsic::vector_reduce_smax; |
| 3225 | if (QT->isUnsignedIntegerType()) |
| 3226 | return llvm::Intrinsic::vector_reduce_umax; |
| 3227 | assert(QT->isFloatingType() && "must have a float here")(static_cast <bool> (QT->isFloatingType() && "must have a float here") ? void (0) : __assert_fail ("QT->isFloatingType() && \"must have a float here\"" , "clang/lib/CodeGen/CGBuiltin.cpp", 3227, __extension__ __PRETTY_FUNCTION__ )); |
| 3228 | return llvm::Intrinsic::vector_reduce_fmax; |
| 3229 | }; |
| 3230 | return RValue::get(emitUnaryBuiltin( |
| 3231 | *this, E, GetIntrinsicID(E->getArg(0)->getType()), "rdx.min")); |
| 3232 | } |
| 3233 | |
| 3234 | case Builtin::BI__builtin_reduce_min: { |
| 3235 | auto GetIntrinsicID = [](QualType QT) { |
| 3236 | if (auto *VecTy = QT->getAs<VectorType>()) |
| 3237 | QT = VecTy->getElementType(); |
| 3238 | if (QT->isSignedIntegerType()) |
| 3239 | return llvm::Intrinsic::vector_reduce_smin; |
| 3240 | if (QT->isUnsignedIntegerType()) |
| 3241 | return llvm::Intrinsic::vector_reduce_umin; |
| 3242 | assert(QT->isFloatingType() && "must have a float here")(static_cast <bool> (QT->isFloatingType() && "must have a float here") ? void (0) : __assert_fail ("QT->isFloatingType() && \"must have a float here\"" , "clang/lib/CodeGen/CGBuiltin.cpp", 3242, __extension__ __PRETTY_FUNCTION__ )); |
| 3243 | return llvm::Intrinsic::vector_reduce_fmin; |
| 3244 | }; |
| 3245 | |
| 3246 | return RValue::get(emitUnaryBuiltin( |
| 3247 | *this, E, GetIntrinsicID(E->getArg(0)->getType()), "rdx.min")); |
| 3248 | } |
| 3249 | |
| 3250 | case Builtin::BI__builtin_reduce_add: |
| 3251 | return RValue::get(emitUnaryBuiltin( |
| 3252 | *this, E, llvm::Intrinsic::vector_reduce_add, "rdx.add")); |
| 3253 | case Builtin::BI__builtin_reduce_mul: |
| 3254 | return RValue::get(emitUnaryBuiltin( |
| 3255 | *this, E, llvm::Intrinsic::vector_reduce_mul, "rdx.mul")); |
| 3256 | case Builtin::BI__builtin_reduce_xor: |
| 3257 | return RValue::get(emitUnaryBuiltin( |
| 3258 | *this, E, llvm::Intrinsic::vector_reduce_xor, "rdx.xor")); |
| 3259 | case Builtin::BI__builtin_reduce_or: |
| 3260 | return RValue::get(emitUnaryBuiltin( |
| 3261 | *this, E, llvm::Intrinsic::vector_reduce_or, "rdx.or")); |
| 3262 | case Builtin::BI__builtin_reduce_and: |
| 3263 | return RValue::get(emitUnaryBuiltin( |
| 3264 | *this, E, llvm::Intrinsic::vector_reduce_and, "rdx.and")); |
| 3265 | |
| 3266 | case Builtin::BI__builtin_matrix_transpose: { |
| 3267 | auto *MatrixTy = E->getArg(0)->getType()->castAs<ConstantMatrixType>(); |
| 3268 | Value *MatValue = EmitScalarExpr(E->getArg(0)); |
| 3269 | MatrixBuilder MB(Builder); |
| 3270 | Value *Result = MB.CreateMatrixTranspose(MatValue, MatrixTy->getNumRows(), |
| 3271 | MatrixTy->getNumColumns()); |
| 3272 | return RValue::get(Result); |
| 3273 | } |
| 3274 | |
| 3275 | case Builtin::BI__builtin_matrix_column_major_load: { |
| 3276 | MatrixBuilder MB(Builder); |
| 3277 | // Emit everything that isn't dependent on the first parameter type |
| 3278 | Value *Stride = EmitScalarExpr(E->getArg(3)); |
| 3279 | const auto *ResultTy = E->getType()->getAs<ConstantMatrixType>(); |
| 3280 | auto *PtrTy = E->getArg(0)->getType()->getAs<PointerType>(); |
| 3281 | assert(PtrTy && "arg0 must be of pointer type")(static_cast <bool> (PtrTy && "arg0 must be of pointer type" ) ? void (0) : __assert_fail ("PtrTy && \"arg0 must be of pointer type\"" , "clang/lib/CodeGen/CGBuiltin.cpp", 3281, __extension__ __PRETTY_FUNCTION__ )); |
| 3282 | bool IsVolatile = PtrTy->getPointeeType().isVolatileQualified(); |
| 3283 | |
| 3284 | Address Src = EmitPointerWithAlignment(E->getArg(0)); |
| 3285 | EmitNonNullArgCheck(RValue::get(Src.getPointer()), E->getArg(0)->getType(), |
| 3286 | E->getArg(0)->getExprLoc(), FD, 0); |
| 3287 | Value *Result = MB.CreateColumnMajorLoad( |
| 3288 | Src.getElementType(), Src.getPointer(), |
| 3289 | Align(Src.getAlignment().getQuantity()), Stride, IsVolatile, |
| 3290 | ResultTy->getNumRows(), ResultTy->getNumColumns(), |
| 3291 | "matrix"); |
| 3292 | return RValue::get(Result); |
| 3293 | } |
| 3294 | |
| 3295 | case Builtin::BI__builtin_matrix_column_major_store: { |
| 3296 | MatrixBuilder MB(Builder); |
| 3297 | Value *Matrix = EmitScalarExpr(E->getArg(0)); |
| 3298 | Address Dst = EmitPointerWithAlignment(E->getArg(1)); |
| 3299 | Value *Stride = EmitScalarExpr(E->getArg(2)); |
| 3300 | |
| 3301 | const auto *MatrixTy = E->getArg(0)->getType()->getAs<ConstantMatrixType>(); |
| 3302 | auto *PtrTy = E->getArg(1)->getType()->getAs<PointerType>(); |
| 3303 | assert(PtrTy && "arg1 must be of pointer type")(static_cast <bool> (PtrTy && "arg1 must be of pointer type" ) ? void (0) : __assert_fail ("PtrTy && \"arg1 must be of pointer type\"" , "clang/lib/CodeGen/CGBuiltin.cpp", 3303, __extension__ __PRETTY_FUNCTION__ )); |
| 3304 | bool IsVolatile = PtrTy->getPointeeType().isVolatileQualified(); |
| 3305 | |
| 3306 | EmitNonNullArgCheck(RValue::get(Dst.getPointer()), E->getArg(1)->getType(), |
| 3307 | E->getArg(1)->getExprLoc(), FD, 0); |
| 3308 | Value *Result = MB.CreateColumnMajorStore( |
| 3309 | Matrix, Dst.getPointer(), Align(Dst.getAlignment().getQuantity()), |
| 3310 | Stride, IsVolatile, MatrixTy->getNumRows(), MatrixTy->getNumColumns()); |
| 3311 | return RValue::get(Result); |
| 3312 | } |
| 3313 | |
| 3314 | case Builtin::BIfinite: |
| 3315 | case Builtin::BI__finite: |
| 3316 | case Builtin::BIfinitef: |
| 3317 | case Builtin::BI__finitef: |
| 3318 | case Builtin::BIfinitel: |
| 3319 | case Builtin::BI__finitel: |
| 3320 | case Builtin::BI__builtin_isinf: |
| 3321 | case Builtin::BI__builtin_isfinite: { |
| 3322 | // isinf(x) --> fabs(x) == infinity |
| 3323 | // isfinite(x) --> fabs(x) != infinity |
| 3324 | // x != NaN via the ordered compare in either case. |
| 3325 | CodeGenFunction::CGFPOptionsRAII FPOptsRAII(*this, E); |
| 3326 | Value *V = EmitScalarExpr(E->getArg(0)); |
| 3327 | llvm::Type *Ty = V->getType(); |
| 3328 | if (!Builder.getIsFPConstrained() || |
| 3329 | Builder.getDefaultConstrainedExcept() == fp::ebIgnore || |
| 3330 | !Ty->isIEEE()) { |
| 3331 | Value *Fabs = EmitFAbs(*this, V); |
| 3332 | Constant *Infinity = ConstantFP::getInfinity(V->getType()); |
| 3333 | CmpInst::Predicate Pred = (BuiltinID == Builtin::BI__builtin_isinf) |
| 3334 | ? CmpInst::FCMP_OEQ |
| 3335 | : CmpInst::FCMP_ONE; |
| 3336 | Value *FCmp = Builder.CreateFCmp(Pred, Fabs, Infinity, "cmpinf"); |
| 3337 | return RValue::get(Builder.CreateZExt(FCmp, ConvertType(E->getType()))); |
| 3338 | } |
| 3339 | |
| 3340 | if (Value *Result = getTargetHooks().testFPKind(V, BuiltinID, Builder, CGM)) |
| 3341 | return RValue::get(Result); |
| 3342 | |
| 3343 | // Inf values have all exp bits set and a zero significand. Therefore: |
| 3344 | // isinf(V) == ((V << 1) == ((exp mask) << 1)) |
| 3345 | // isfinite(V) == ((V << 1) < ((exp mask) << 1)) using unsigned comparison |
| 3346 | unsigned bitsize = Ty->getScalarSizeInBits(); |
| 3347 | llvm::IntegerType *IntTy = Builder.getIntNTy(bitsize); |
| 3348 | Value *IntV = Builder.CreateBitCast(V, IntTy); |
| 3349 | Value *Shl1 = Builder.CreateShl(IntV, 1); |
| 3350 | const llvm::fltSemantics &Semantics = Ty->getFltSemantics(); |
| 3351 | APInt ExpMask = APFloat::getInf(Semantics).bitcastToAPInt(); |
| 3352 | Value *ExpMaskShl1 = llvm::ConstantInt::get(IntTy, ExpMask.shl(1)); |
| 3353 | if (BuiltinID == Builtin::BI__builtin_isinf) |
| 3354 | V = Builder.CreateICmpEQ(Shl1, ExpMaskShl1); |
| 3355 | else |
| 3356 | V = Builder.CreateICmpULT(Shl1, ExpMaskShl1); |
| 3357 | return RValue::get(Builder.CreateZExt(V, ConvertType(E->getType()))); |
| 3358 | } |
| 3359 | |
| 3360 | case Builtin::BI__builtin_isinf_sign: { |
| 3361 | // isinf_sign(x) -> fabs(x) == infinity ? (signbit(x) ? -1 : 1) : 0 |
| 3362 | CodeGenFunction::CGFPOptionsRAII FPOptsRAII(*this, E); |
| 3363 | // FIXME: for strictfp/IEEE-754 we need to not trap on SNaN here. |
| 3364 | Value *Arg = EmitScalarExpr(E->getArg(0)); |
| 3365 | Value *AbsArg = EmitFAbs(*this, Arg); |
| 3366 | Value *IsInf = Builder.CreateFCmpOEQ( |
| 3367 | AbsArg, ConstantFP::getInfinity(Arg->getType()), "isinf"); |
| 3368 | Value *IsNeg = EmitSignBit(*this, Arg); |
| 3369 | |
| 3370 | llvm::Type *IntTy = ConvertType(E->getType()); |
| 3371 | Value *Zero = Constant::getNullValue(IntTy); |
| 3372 | Value *One = ConstantInt::get(IntTy, 1); |
| 3373 | Value *NegativeOne = ConstantInt::get(IntTy, -1); |
| 3374 | Value *SignResult = Builder.CreateSelect(IsNeg, NegativeOne, One); |
| 3375 | Value *Result = Builder.CreateSelect(IsInf, SignResult, Zero); |
| 3376 | return RValue::get(Result); |
| 3377 | } |
| 3378 | |
| 3379 | case Builtin::BI__builtin_isnormal: { |
| 3380 | // isnormal(x) --> x == x && fabsf(x) < infinity && fabsf(x) >= float_min |
| 3381 | CodeGenFunction::CGFPOptionsRAII FPOptsRAII(*this, E); |
| 3382 | // FIXME: for strictfp/IEEE-754 we need to not trap on SNaN here. |
| 3383 | Value *V = EmitScalarExpr(E->getArg(0)); |
| 3384 | Value *Eq = Builder.CreateFCmpOEQ(V, V, "iseq"); |
| 3385 | |
| 3386 | Value *Abs = EmitFAbs(*this, V); |
| 3387 | Value *IsLessThanInf = |
| 3388 | Builder.CreateFCmpULT(Abs, ConstantFP::getInfinity(V->getType()),"isinf"); |
| 3389 | APFloat Smallest = APFloat::getSmallestNormalized( |
| 3390 | getContext().getFloatTypeSemantics(E->getArg(0)->getType())); |
| 3391 | Value *IsNormal = |
| 3392 | Builder.CreateFCmpUGE(Abs, ConstantFP::get(V->getContext(), Smallest), |
| 3393 | "isnormal"); |
| 3394 | V = Builder.CreateAnd(Eq, IsLessThanInf, "and"); |
| 3395 | V = Builder.CreateAnd(V, IsNormal, "and"); |
| 3396 | return RValue::get(Builder.CreateZExt(V, ConvertType(E->getType()))); |
| 3397 | } |
| 3398 | |
| 3399 | case Builtin::BI__builtin_flt_rounds: { |
| 3400 | Function *F = CGM.getIntrinsic(Intrinsic::get_rounding); |
| 3401 | |
| 3402 | llvm::Type *ResultType = ConvertType(E->getType()); |
| 3403 | Value *Result = Builder.CreateCall(F); |
| 3404 | if (Result->getType() != ResultType) |
| 3405 | Result = Builder.CreateIntCast(Result, ResultType, /*isSigned*/true, |
| 3406 | "cast"); |
| 3407 | return RValue::get(Result); |
| 3408 | } |
| 3409 | |
| 3410 | case Builtin::BI__builtin_set_flt_rounds: { |
| 3411 | Function *F = CGM.getIntrinsic(Intrinsic::set_rounding); |
| 3412 | |
| 3413 | Value *V = EmitScalarExpr(E->getArg(0)); |
| 3414 | Builder.CreateCall(F, V); |
| 3415 | return RValue::get(nullptr); |
| 3416 | } |
| 3417 | |
| 3418 | case Builtin::BI__builtin_fpclassify: { |
| 3419 | CodeGenFunction::CGFPOptionsRAII FPOptsRAII(*this, E); |
| 3420 | // FIXME: for strictfp/IEEE-754 we need to not trap on SNaN here. |
| 3421 | Value *V = EmitScalarExpr(E->getArg(5)); |
| 3422 | llvm::Type *Ty = ConvertType(E->getArg(5)->getType()); |
| 3423 | |
| 3424 | // Create Result |
| 3425 | BasicBlock *Begin = Builder.GetInsertBlock(); |
| 3426 | BasicBlock *End = createBasicBlock("fpclassify_end", this->CurFn); |
| 3427 | Builder.SetInsertPoint(End); |
| 3428 | PHINode *Result = |
| 3429 | Builder.CreatePHI(ConvertType(E->getArg(0)->getType()), 4, |
| 3430 | "fpclassify_result"); |
| 3431 | |
| 3432 | // if (V==0) return FP_ZERO |
| 3433 | Builder.SetInsertPoint(Begin); |
| 3434 | Value *IsZero = Builder.CreateFCmpOEQ(V, Constant::getNullValue(Ty), |
| 3435 | "iszero"); |
| 3436 | Value *ZeroLiteral = EmitScalarExpr(E->getArg(4)); |
| 3437 | BasicBlock *NotZero = createBasicBlock("fpclassify_not_zero", this->CurFn); |
| 3438 | Builder.CreateCondBr(IsZero, End, NotZero); |
| 3439 | Result->addIncoming(ZeroLiteral, Begin); |
| 3440 | |
| 3441 | // if (V != V) return FP_NAN |
| 3442 | Builder.SetInsertPoint(NotZero); |
| 3443 | Value *IsNan = Builder.CreateFCmpUNO(V, V, "cmp"); |
| 3444 | Value *NanLiteral = EmitScalarExpr(E->getArg(0)); |
| 3445 | BasicBlock *NotNan = createBasicBlock("fpclassify_not_nan", this->CurFn); |
| 3446 | Builder.CreateCondBr(IsNan, End, NotNan); |
| 3447 | Result->addIncoming(NanLiteral, NotZero); |
| 3448 | |
| 3449 | // if (fabs(V) == infinity) return FP_INFINITY |
| 3450 | Builder.SetInsertPoint(NotNan); |
| 3451 | Value *VAbs = EmitFAbs(*this, V); |
| 3452 | Value *IsInf = |
| 3453 | Builder.CreateFCmpOEQ(VAbs, ConstantFP::getInfinity(V->getType()), |
| 3454 | "isinf"); |
| 3455 | Value *InfLiteral = EmitScalarExpr(E->getArg(1)); |
| 3456 | BasicBlock *NotInf = createBasicBlock("fpclassify_not_inf", this->CurFn); |
| 3457 | Builder.CreateCondBr(IsInf, End, NotInf); |
| 3458 | Result->addIncoming(InfLiteral, NotNan); |
| 3459 | |
| 3460 | // if (fabs(V) >= MIN_NORMAL) return FP_NORMAL else FP_SUBNORMAL |
| 3461 | Builder.SetInsertPoint(NotInf); |
| 3462 | APFloat Smallest = APFloat::getSmallestNormalized( |
| 3463 | getContext().getFloatTypeSemantics(E->getArg(5)->getType())); |
| 3464 | Value *IsNormal = |
| 3465 | Builder.CreateFCmpUGE(VAbs, ConstantFP::get(V->getContext(), Smallest), |
| 3466 | "isnormal"); |
| 3467 | Value *NormalResult = |
| 3468 | Builder.CreateSelect(IsNormal, EmitScalarExpr(E->getArg(2)), |
| 3469 | EmitScalarExpr(E->getArg(3))); |
| 3470 | Builder.CreateBr(End); |
| 3471 | Result->addIncoming(NormalResult, NotInf); |
| 3472 | |
| 3473 | // return Result |
| 3474 | Builder.SetInsertPoint(End); |
| 3475 | return RValue::get(Result); |
| 3476 | } |
| 3477 | |
| 3478 | case Builtin::BIalloca: |
| 3479 | case Builtin::BI_alloca: |
| 3480 | case Builtin::BI__builtin_alloca_uninitialized: |
| 3481 | case Builtin::BI__builtin_alloca: { |
| 3482 | Value *Size = EmitScalarExpr(E->getArg(0)); |
| 3483 | const TargetInfo &TI = getContext().getTargetInfo(); |
| 3484 | // The alignment of the alloca should correspond to __BIGGEST_ALIGNMENT__. |
| 3485 | const Align SuitableAlignmentInBytes = |
| 3486 | CGM.getContext() |
| 3487 | .toCharUnitsFromBits(TI.getSuitableAlign()) |
| 3488 | .getAsAlign(); |
| 3489 | AllocaInst *AI = Builder.CreateAlloca(Builder.getInt8Ty(), Size); |
| 3490 | AI->setAlignment(SuitableAlignmentInBytes); |
| 3491 | if (BuiltinID != Builtin::BI__builtin_alloca_uninitialized) |
| 3492 | initializeAlloca(*this, AI, Size, SuitableAlignmentInBytes); |
| 3493 | return RValue::get(AI); |
| 3494 | } |
| 3495 | |
| 3496 | case Builtin::BI__builtin_alloca_with_align_uninitialized: |
| 3497 | case Builtin::BI__builtin_alloca_with_align: { |
| 3498 | Value *Size = EmitScalarExpr(E->getArg(0)); |
| 3499 | Value *AlignmentInBitsValue = EmitScalarExpr(E->getArg(1)); |
| 3500 | auto *AlignmentInBitsCI = cast<ConstantInt>(AlignmentInBitsValue); |
| 3501 | unsigned AlignmentInBits = AlignmentInBitsCI->getZExtValue(); |
| 3502 | const Align AlignmentInBytes = |
| 3503 | CGM.getContext().toCharUnitsFromBits(AlignmentInBits).getAsAlign(); |
| 3504 | AllocaInst *AI = Builder.CreateAlloca(Builder.getInt8Ty(), Size); |
| 3505 | AI->setAlignment(AlignmentInBytes); |
| 3506 | if (BuiltinID != Builtin::BI__builtin_alloca_with_align_uninitialized) |
| 3507 | initializeAlloca(*this, AI, Size, AlignmentInBytes); |
| 3508 | return RValue::get(AI); |
| 3509 | } |
| 3510 | |
| 3511 | case Builtin::BIbzero: |
| 3512 | case Builtin::BI__builtin_bzero: { |
| 3513 | Address Dest = EmitPointerWithAlignment(E->getArg(0)); |
| 3514 | Value *SizeVal = EmitScalarExpr(E->getArg(1)); |
| 3515 | EmitNonNullArgCheck(RValue::get(Dest.getPointer()), E->getArg(0)->getType(), |
| 3516 | E->getArg(0)->getExprLoc(), FD, 0); |
| 3517 | Builder.CreateMemSet(Dest, Builder.getInt8(0), SizeVal, false); |
| 3518 | return RValue::get(nullptr); |
| 3519 | } |
| 3520 | case Builtin::BImemcpy: |
| 3521 | case Builtin::BI__builtin_memcpy: |
| 3522 | case Builtin::BImempcpy: |
| 3523 | case Builtin::BI__builtin_mempcpy: { |
| 3524 | Address Dest = EmitPointerWithAlignment(E->getArg(0)); |
| 3525 | Address Src = EmitPointerWithAlignment(E->getArg(1)); |
| 3526 | Value *SizeVal = EmitScalarExpr(E->getArg(2)); |
| 3527 | EmitNonNullArgCheck(RValue::get(Dest.getPointer()), E->getArg(0)->getType(), |
| 3528 | E->getArg(0)->getExprLoc(), FD, 0); |
| 3529 | EmitNonNullArgCheck(RValue::get(Src.getPointer()), E->getArg(1)->getType(), |
| 3530 | E->getArg(1)->getExprLoc(), FD, 1); |
| 3531 | Builder.CreateMemCpy(Dest, Src, SizeVal, false); |
| 3532 | if (BuiltinID == Builtin::BImempcpy || |
| 3533 | BuiltinID == Builtin::BI__builtin_mempcpy) |
| 3534 | return RValue::get(Builder.CreateInBoundsGEP(Dest.getElementType(), |
| 3535 | Dest.getPointer(), SizeVal)); |
| 3536 | else |
| 3537 | return RValue::get(Dest.getPointer()); |
| 3538 | } |
| 3539 | |
| 3540 | case Builtin::BI__builtin_memcpy_inline: { |
| 3541 | Address Dest = EmitPointerWithAlignment(E->getArg(0)); |
| 3542 | Address Src = EmitPointerWithAlignment(E->getArg(1)); |
| 3543 | uint64_t Size = |
| 3544 | E->getArg(2)->EvaluateKnownConstInt(getContext()).getZExtValue(); |
| 3545 | EmitNonNullArgCheck(RValue::get(Dest.getPointer()), E->getArg(0)->getType(), |
| 3546 | E->getArg(0)->getExprLoc(), FD, 0); |
| 3547 | EmitNonNullArgCheck(RValue::get(Src.getPointer()), E->getArg(1)->getType(), |
| 3548 | E->getArg(1)->getExprLoc(), FD, 1); |
| 3549 | Builder.CreateMemCpyInline(Dest, Src, Size); |
| 3550 | return RValue::get(nullptr); |
| 3551 | } |
| 3552 | |
| 3553 | case Builtin::BI__builtin_char_memchr: |
| 3554 | BuiltinID = Builtin::BI__builtin_memchr; |
| 3555 | break; |
| 3556 | |
| 3557 | case Builtin::BI__builtin___memcpy_chk: { |
| 3558 | // fold __builtin_memcpy_chk(x, y, cst1, cst2) to memcpy iff cst1<=cst2. |
| 3559 | Expr::EvalResult SizeResult, DstSizeResult; |
| 3560 | if (!E->getArg(2)->EvaluateAsInt(SizeResult, CGM.getContext()) || |
| 3561 | !E->getArg(3)->EvaluateAsInt(DstSizeResult, CGM.getContext())) |
| 3562 | break; |
| 3563 | llvm::APSInt Size = SizeResult.Val.getInt(); |
| 3564 | llvm::APSInt DstSize = DstSizeResult.Val.getInt(); |
| 3565 | if (Size.ugt(DstSize)) |
| 3566 | break; |
| 3567 | Address Dest = EmitPointerWithAlignment(E->getArg(0)); |
| 3568 | Address Src = EmitPointerWithAlignment(E->getArg(1)); |
| 3569 | Value *SizeVal = llvm::ConstantInt::get(Builder.getContext(), Size); |
| 3570 | Builder.CreateMemCpy(Dest, Src, SizeVal, false); |
| 3571 | return RValue::get(Dest.getPointer()); |
| 3572 | } |
| 3573 | |
| 3574 | case Builtin::BI__builtin_objc_memmove_collectable: { |
| 3575 | Address DestAddr = EmitPointerWithAlignment(E->getArg(0)); |
| 3576 | Address SrcAddr = EmitPointerWithAlignment(E->getArg(1)); |
| 3577 | Value *SizeVal = EmitScalarExpr(E->getArg(2)); |
| 3578 | CGM.getObjCRuntime().EmitGCMemmoveCollectable(*this, |
| 3579 | DestAddr, SrcAddr, SizeVal); |
| 3580 | return RValue::get(DestAddr.getPointer()); |
| 3581 | } |
| 3582 | |
| 3583 | case Builtin::BI__builtin___memmove_chk: { |
| 3584 | // fold __builtin_memmove_chk(x, y, cst1, cst2) to memmove iff cst1<=cst2. |
| 3585 | Expr::EvalResult SizeResult, DstSizeResult; |
| 3586 | if (!E->getArg(2)->EvaluateAsInt(SizeResult, CGM.getContext()) || |
| 3587 | !E->getArg(3)->EvaluateAsInt(DstSizeResult, CGM.getContext())) |
| 3588 | break; |
| 3589 | llvm::APSInt Size = SizeResult.Val.getInt(); |
| 3590 | llvm::APSInt DstSize = DstSizeResult.Val.getInt(); |
| 3591 | if (Size.ugt(DstSize)) |
| 3592 | break; |
| 3593 | Address Dest = EmitPointerWithAlignment(E->getArg(0)); |
| 3594 | Address Src = EmitPointerWithAlignment(E->getArg(1)); |
| 3595 | Value *SizeVal = llvm::ConstantInt::get(Builder.getContext(), Size); |
| 3596 | Builder.CreateMemMove(Dest, Src, SizeVal, false); |
| 3597 | return RValue::get(Dest.getPointer()); |
| 3598 | } |
| 3599 | |
| 3600 | case Builtin::BImemmove: |
| 3601 | case Builtin::BI__builtin_memmove: { |
| 3602 | Address Dest = EmitPointerWithAlignment(E->getArg(0)); |
| 3603 | Address Src = EmitPointerWithAlignment(E->getArg(1)); |
| 3604 | Value *SizeVal = EmitScalarExpr(E->getArg(2)); |
| 3605 | EmitNonNullArgCheck(RValue::get(Dest.getPointer()), E->getArg(0)->getType(), |
| 3606 | E->getArg(0)->getExprLoc(), FD, 0); |
| 3607 | EmitNonNullArgCheck(RValue::get(Src.getPointer()), E->getArg(1)->getType(), |
| 3608 | E->getArg(1)->getExprLoc(), FD, 1); |
| 3609 | Builder.CreateMemMove(Dest, Src, SizeVal, false); |
| 3610 | return RValue::get(Dest.getPointer()); |
| 3611 | } |
| 3612 | case Builtin::BImemset: |
| 3613 | case Builtin::BI__builtin_memset: { |
| 3614 | Address Dest = EmitPointerWithAlignment(E->getArg(0)); |
| 3615 | Value *ByteVal = Builder.CreateTrunc(EmitScalarExpr(E->getArg(1)), |
| 3616 | Builder.getInt8Ty()); |
| 3617 | Value *SizeVal = EmitScalarExpr(E->getArg(2)); |
| 3618 | EmitNonNullArgCheck(RValue::get(Dest.getPointer()), E->getArg(0)->getType(), |
| 3619 | E->getArg(0)->getExprLoc(), FD, 0); |
| 3620 | Builder.CreateMemSet(Dest, ByteVal, SizeVal, false); |
| 3621 | return RValue::get(Dest.getPointer()); |
| 3622 | } |
| 3623 | case Builtin::BI__builtin_memset_inline: { |
| 3624 | Address Dest = EmitPointerWithAlignment(E->getArg(0)); |
| 3625 | Value *ByteVal = |
| 3626 | Builder.CreateTrunc(EmitScalarExpr(E->getArg(1)), Builder.getInt8Ty()); |
| 3627 | uint64_t Size = |
| 3628 | E->getArg(2)->EvaluateKnownConstInt(getContext()).getZExtValue(); |
| 3629 | EmitNonNullArgCheck(RValue::get(Dest.getPointer()), E->getArg(0)->getType(), |
| 3630 | E->getArg(0)->getExprLoc(), FD, 0); |
| 3631 | Builder.CreateMemSetInline(Dest, ByteVal, Size); |
| 3632 | return RValue::get(nullptr); |
| 3633 | } |
| 3634 | case Builtin::BI__builtin___memset_chk: { |
| 3635 | // fold __builtin_memset_chk(x, y, cst1, cst2) to memset iff cst1<=cst2. |
| 3636 | Expr::EvalResult SizeResult, DstSizeResult; |
| 3637 | if (!E->getArg(2)->EvaluateAsInt(SizeResult, CGM.getContext()) || |
| 3638 | !E->getArg(3)->EvaluateAsInt(DstSizeResult, CGM.getContext())) |
| 3639 | break; |
| 3640 | llvm::APSInt Size = SizeResult.Val.getInt(); |
| 3641 | llvm::APSInt DstSize = DstSizeResult.Val.getInt(); |
| 3642 | if (Size.ugt(DstSize)) |
| 3643 | break; |
| 3644 | Address Dest = EmitPointerWithAlignment(E->getArg(0)); |
| 3645 | Value *ByteVal = Builder.CreateTrunc(EmitScalarExpr(E->getArg(1)), |
| 3646 | Builder.getInt8Ty()); |
| 3647 | Value *SizeVal = llvm::ConstantInt::get(Builder.getContext(), Size); |
| 3648 | Builder.CreateMemSet(Dest, ByteVal, SizeVal, false); |
| 3649 | return RValue::get(Dest.getPointer()); |
| 3650 | } |
| 3651 | case Builtin::BI__builtin_wmemchr: { |
| 3652 | // The MSVC runtime library does not provide a definition of wmemchr, so we |
| 3653 | // need an inline implementation. |
| 3654 | if (!getTarget().getTriple().isOSMSVCRT()) |
| 3655 | break; |
| 3656 | |
| 3657 | llvm::Type *WCharTy = ConvertType(getContext().WCharTy); |
| 3658 | Value *Str = EmitScalarExpr(E->getArg(0)); |
| 3659 | Value *Chr = EmitScalarExpr(E->getArg(1)); |
| 3660 | Value *Size = EmitScalarExpr(E->getArg(2)); |
| 3661 | |
| 3662 | BasicBlock *Entry = Builder.GetInsertBlock(); |
| 3663 | BasicBlock *CmpEq = createBasicBlock("wmemchr.eq"); |
| 3664 | BasicBlock *Next = createBasicBlock("wmemchr.next"); |
| 3665 | BasicBlock *Exit = createBasicBlock("wmemchr.exit"); |
| 3666 | Value *SizeEq0 = Builder.CreateICmpEQ(Size, ConstantInt::get(SizeTy, 0)); |
| 3667 | Builder.CreateCondBr(SizeEq0, Exit, CmpEq); |
| 3668 | |
| 3669 | EmitBlock(CmpEq); |
| 3670 | PHINode *StrPhi = Builder.CreatePHI(Str->getType(), 2); |
| 3671 | StrPhi->addIncoming(Str, Entry); |
| 3672 | PHINode *SizePhi = Builder.CreatePHI(SizeTy, 2); |
| 3673 | SizePhi->addIncoming(Size, Entry); |
| 3674 | CharUnits WCharAlign = |
| 3675 | getContext().getTypeAlignInChars(getContext().WCharTy); |
| 3676 | Value *StrCh = Builder.CreateAlignedLoad(WCharTy, StrPhi, WCharAlign); |
| 3677 | Value *FoundChr = Builder.CreateConstInBoundsGEP1_32(WCharTy, StrPhi, 0); |
| 3678 | Value *StrEqChr = Builder.CreateICmpEQ(StrCh, Chr); |
| 3679 | Builder.CreateCondBr(StrEqChr, Exit, Next); |
| 3680 | |
| 3681 | EmitBlock(Next); |
| 3682 | Value *NextStr = Builder.CreateConstInBoundsGEP1_32(WCharTy, StrPhi, 1); |
| 3683 | Value *NextSize = Builder.CreateSub(SizePhi, ConstantInt::get(SizeTy, 1)); |
| 3684 | Value *NextSizeEq0 = |
| 3685 | Builder.CreateICmpEQ(NextSize, ConstantInt::get(SizeTy, 0)); |
| 3686 | Builder.CreateCondBr(NextSizeEq0, Exit, CmpEq); |
| 3687 | StrPhi->addIncoming(NextStr, Next); |
| 3688 | SizePhi->addIncoming(NextSize, Next); |
| 3689 | |
| 3690 | EmitBlock(Exit); |
| 3691 | PHINode *Ret = Builder.CreatePHI(Str->getType(), 3); |
| 3692 | Ret->addIncoming(llvm::Constant::getNullValue(Str->getType()), Entry); |
| 3693 | Ret->addIncoming(llvm::Constant::getNullValue(Str->getType()), Next); |
| 3694 | Ret->addIncoming(FoundChr, CmpEq); |
| 3695 | return RValue::get(Ret); |
| 3696 | } |
| 3697 | case Builtin::BI__builtin_wmemcmp: { |
| 3698 | // The MSVC runtime library does not provide a definition of wmemcmp, so we |
| 3699 | // need an inline implementation. |
| 3700 | if (!getTarget().getTriple().isOSMSVCRT()) |
| 3701 | break; |
| 3702 | |
| 3703 | llvm::Type *WCharTy = ConvertType(getContext().WCharTy); |
| 3704 | |
| 3705 | Value *Dst = EmitScalarExpr(E->getArg(0)); |
| 3706 | Value *Src = EmitScalarExpr(E->getArg(1)); |
| 3707 | Value *Size = EmitScalarExpr(E->getArg(2)); |
| 3708 | |
| 3709 | BasicBlock *Entry = Builder.GetInsertBlock(); |
| 3710 | BasicBlock *CmpGT = createBasicBlock("wmemcmp.gt"); |
| 3711 | BasicBlock *CmpLT = createBasicBlock("wmemcmp.lt"); |
| 3712 | BasicBlock *Next = createBasicBlock("wmemcmp.next"); |
| 3713 | BasicBlock *Exit = createBasicBlock("wmemcmp.exit"); |
| 3714 | Value *SizeEq0 = Builder.CreateICmpEQ(Size, ConstantInt::get(SizeTy, 0)); |
| 3715 | Builder.CreateCondBr(SizeEq0, Exit, CmpGT); |
| 3716 | |
| 3717 | EmitBlock(CmpGT); |
| 3718 | PHINode *DstPhi = Builder.CreatePHI(Dst->getType(), 2); |
| 3719 | DstPhi->addIncoming(Dst, Entry); |
| 3720 | PHINode *SrcPhi = Builder.CreatePHI(Src->getType(), 2); |
| 3721 | SrcPhi->addIncoming(Src, Entry); |
| 3722 | PHINode *SizePhi = Builder.CreatePHI(SizeTy, 2); |
| 3723 | SizePhi->addIncoming(Size, Entry); |
| 3724 | CharUnits WCharAlign = |
| 3725 | getContext().getTypeAlignInChars(getContext().WCharTy); |
| 3726 | Value *DstCh = Builder.CreateAlignedLoad(WCharTy, DstPhi, WCharAlign); |
| 3727 | Value *SrcCh = Builder.CreateAlignedLoad(WCharTy, SrcPhi, WCharAlign); |
| 3728 | Value *DstGtSrc = Builder.CreateICmpUGT(DstCh, SrcCh); |
| 3729 | Builder.CreateCondBr(DstGtSrc, Exit, CmpLT); |
| 3730 | |
| 3731 | EmitBlock(CmpLT); |
| 3732 | Value *DstLtSrc = Builder.CreateICmpULT(DstCh, SrcCh); |
| 3733 | Builder.CreateCondBr(DstLtSrc, Exit, Next); |
| 3734 | |
| 3735 | EmitBlock(Next); |
| 3736 | Value *NextDst = Builder.CreateConstInBoundsGEP1_32(WCharTy, DstPhi, 1); |
| 3737 | Value *NextSrc = Builder.CreateConstInBoundsGEP1_32(WCharTy, SrcPhi, 1); |
| 3738 | Value *NextSize = Builder.CreateSub(SizePhi, ConstantInt::get(SizeTy, 1)); |
| 3739 | Value *NextSizeEq0 = |
| 3740 | Builder.CreateICmpEQ(NextSize, ConstantInt::get(SizeTy, 0)); |
| 3741 | Builder.CreateCondBr(NextSizeEq0, Exit, CmpGT); |
| 3742 | DstPhi->addIncoming(NextDst, Next); |
| 3743 | SrcPhi->addIncoming(NextSrc, Next); |
| 3744 | SizePhi->addIncoming(NextSize, Next); |
| 3745 | |
| 3746 | EmitBlock(Exit); |
| 3747 | PHINode *Ret = Builder.CreatePHI(IntTy, 4); |
| 3748 | Ret->addIncoming(ConstantInt::get(IntTy, 0), Entry); |
| 3749 | Ret->addIncoming(ConstantInt::get(IntTy, 1), CmpGT); |
| 3750 | Ret->addIncoming(ConstantInt::get(IntTy, -1), CmpLT); |
| 3751 | Ret->addIncoming(ConstantInt::get(IntTy, 0), Next); |
| 3752 | return RValue::get(Ret); |
| 3753 | } |
| 3754 | case Builtin::BI__builtin_dwarf_cfa: { |
| 3755 | // The offset in bytes from the first argument to the CFA. |
| 3756 | // |
| 3757 | // Why on earth is this in the frontend? Is there any reason at |
| 3758 | // all that the backend can't reasonably determine this while |
| 3759 | // lowering llvm.eh.dwarf.cfa()? |
| 3760 | // |
| 3761 | // TODO: If there's a satisfactory reason, add a target hook for |
| 3762 | // this instead of hard-coding 0, which is correct for most targets. |
| 3763 | int32_t Offset = 0; |
| 3764 | |
| 3765 | Function *F = CGM.getIntrinsic(Intrinsic::eh_dwarf_cfa); |
| 3766 | return RValue::get(Builder.CreateCall(F, |
| 3767 | llvm::ConstantInt::get(Int32Ty, Offset))); |
| 3768 | } |
| 3769 | case Builtin::BI__builtin_return_address: { |
| 3770 | Value *Depth = ConstantEmitter(*this).emitAbstract(E->getArg(0), |
| 3771 | getContext().UnsignedIntTy); |
| 3772 | Function *F = CGM.getIntrinsic(Intrinsic::returnaddress); |
| 3773 | return RValue::get(Builder.CreateCall(F, Depth)); |
| 3774 | } |
| 3775 | case Builtin::BI_ReturnAddress: { |
| 3776 | Function *F = CGM.getIntrinsic(Intrinsic::returnaddress); |
| 3777 | return RValue::get(Builder.CreateCall(F, Builder.getInt32(0))); |
| 3778 | } |
| 3779 | case Builtin::BI__builtin_frame_address: { |
| 3780 | Value *Depth = ConstantEmitter(*this).emitAbstract(E->getArg(0), |
| 3781 | getContext().UnsignedIntTy); |
| 3782 | Function *F = CGM.getIntrinsic(Intrinsic::frameaddress, AllocaInt8PtrTy); |
| 3783 | return RValue::get(Builder.CreateCall(F, Depth)); |
| 3784 | } |
| 3785 | case Builtin::BI__builtin_extract_return_addr: { |
| 3786 | Value *Address = EmitScalarExpr(E->getArg(0)); |
| 3787 | Value *Result = getTargetHooks().decodeReturnAddress(*this, Address); |
| 3788 | return RValue::get(Result); |
| 3789 | } |
| 3790 | case Builtin::BI__builtin_frob_return_addr: { |
| 3791 | Value *Address = EmitScalarExpr(E->getArg(0)); |
| 3792 | Value *Result = getTargetHooks().encodeReturnAddress(*this, Address); |
| 3793 | return RValue::get(Result); |
| 3794 | } |
| 3795 | case Builtin::BI__builtin_dwarf_sp_column: { |
| 3796 | llvm::IntegerType *Ty |
| 3797 | = cast<llvm::IntegerType>(ConvertType(E->getType())); |
| 3798 | int Column = getTargetHooks().getDwarfEHStackPointer(CGM); |
| 3799 | if (Column == -1) { |
| 3800 | CGM.ErrorUnsupported(E, "__builtin_dwarf_sp_column"); |
| 3801 | return RValue::get(llvm::UndefValue::get(Ty)); |
| 3802 | } |
| 3803 | return RValue::get(llvm::ConstantInt::get(Ty, Column, true)); |
| 3804 | } |
| 3805 | case Builtin::BI__builtin_init_dwarf_reg_size_table: { |
| 3806 | Value *Address = EmitScalarExpr(E->getArg(0)); |
| 3807 | if (getTargetHooks().initDwarfEHRegSizeTable(*this, Address)) |
| 3808 | CGM.ErrorUnsupported(E, "__builtin_init_dwarf_reg_size_table"); |
| 3809 | return RValue::get(llvm::UndefValue::get(ConvertType(E->getType()))); |
| 3810 | } |
| 3811 | case Builtin::BI__builtin_eh_return: { |
| 3812 | Value *Int = EmitScalarExpr(E->getArg(0)); |
| 3813 | Value *Ptr = EmitScalarExpr(E->getArg(1)); |
| 3814 | |
| 3815 | llvm::IntegerType *IntTy = cast<llvm::IntegerType>(Int->getType()); |
| 3816 | assert((IntTy->getBitWidth() == 32 || IntTy->getBitWidth() == 64) &&(static_cast <bool> ((IntTy->getBitWidth() == 32 || IntTy ->getBitWidth() == 64) && "LLVM's __builtin_eh_return only supports 32- and 64-bit variants" ) ? void (0) : __assert_fail ("(IntTy->getBitWidth() == 32 || IntTy->getBitWidth() == 64) && \"LLVM's __builtin_eh_return only supports 32- and 64-bit variants\"" , "clang/lib/CodeGen/CGBuiltin.cpp", 3817, __extension__ __PRETTY_FUNCTION__ )) |
| 3817 | "LLVM's __builtin_eh_return only supports 32- and 64-bit variants")(static_cast <bool> ((IntTy->getBitWidth() == 32 || IntTy ->getBitWidth() == 64) && "LLVM's __builtin_eh_return only supports 32- and 64-bit variants" ) ? void (0) : __assert_fail ("(IntTy->getBitWidth() == 32 || IntTy->getBitWidth() == 64) && \"LLVM's __builtin_eh_return only supports 32- and 64-bit variants\"" , "clang/lib/CodeGen/CGBuiltin.cpp", 3817, __extension__ __PRETTY_FUNCTION__ )); |
| 3818 | Function *F = |
| 3819 | CGM.getIntrinsic(IntTy->getBitWidth() == 32 ? Intrinsic::eh_return_i32 |
| 3820 | : Intrinsic::eh_return_i64); |
| 3821 | Builder.CreateCall(F, {Int, Ptr}); |
| 3822 | Builder.CreateUnreachable(); |
| 3823 | |
| 3824 | // We do need to preserve an insertion point. |
| 3825 | EmitBlock(createBasicBlock("builtin_eh_return.cont")); |
| 3826 | |
| 3827 | return RValue::get(nullptr); |
| 3828 | } |
| 3829 | case Builtin::BI__builtin_unwind_init: { |
| 3830 | Function *F = CGM.getIntrinsic(Intrinsic::eh_unwind_init); |
| 3831 | Builder.CreateCall(F); |
| 3832 | return RValue::get(nullptr); |
| 3833 | } |
| 3834 | case Builtin::BI__builtin_extend_pointer: { |
| 3835 | // Extends a pointer to the size of an _Unwind_Word, which is |
| 3836 | // uint64_t on all platforms. Generally this gets poked into a |
| 3837 | // register and eventually used as an address, so if the |
| 3838 | // addressing registers are wider than pointers and the platform |
| 3839 | // doesn't implicitly ignore high-order bits when doing |
| 3840 | // addressing, we need to make sure we zext / sext based on |
| 3841 | // the platform's expectations. |
| 3842 | // |
| 3843 | // See: http://gcc.gnu.org/ml/gcc-bugs/2002-02/msg00237.html |
| 3844 | |
| 3845 | // Cast the pointer to intptr_t. |
| 3846 | Value *Ptr = EmitScalarExpr(E->getArg(0)); |
| 3847 | Value *Result = Builder.CreatePtrToInt(Ptr, IntPtrTy, "extend.cast"); |
| 3848 | |
| 3849 | // If that's 64 bits, we're done. |
| 3850 | if (IntPtrTy->getBitWidth() == 64) |
| 3851 | return RValue::get(Result); |
| 3852 | |
| 3853 | // Otherwise, ask the codegen data what to do. |
| 3854 | if (getTargetHooks().extendPointerWithSExt()) |
| 3855 | return RValue::get(Builder.CreateSExt(Result, Int64Ty, "extend.sext")); |
| 3856 | else |
| 3857 | return RValue::get(Builder.CreateZExt(Result, Int64Ty, "extend.zext")); |
| 3858 | } |
| 3859 | case Builtin::BI__builtin_setjmp: { |
| 3860 | // Buffer is a void**. |
| 3861 | Address Buf = EmitPointerWithAlignment(E->getArg(0)); |
| 3862 | |
| 3863 | // Store the frame pointer to the setjmp buffer. |
| 3864 | Value *FrameAddr = Builder.CreateCall( |
| 3865 | CGM.getIntrinsic(Intrinsic::frameaddress, AllocaInt8PtrTy), |
| 3866 | ConstantInt::get(Int32Ty, 0)); |
| 3867 | Builder.CreateStore(FrameAddr, Buf); |
| 3868 | |
| 3869 | // Store the stack pointer to the setjmp buffer. |
| 3870 | Value *StackAddr = |
| 3871 | Builder.CreateCall(CGM.getIntrinsic(Intrinsic::stacksave)); |
| 3872 | Address StackSaveSlot = Builder.CreateConstInBoundsGEP(Buf, 2); |
| 3873 | Builder.CreateStore(StackAddr, StackSaveSlot); |
| 3874 | |
| 3875 | // Call LLVM's EH setjmp, which is lightweight. |
| 3876 | Function *F = CGM.getIntrinsic(Intrinsic::eh_sjlj_setjmp); |
| 3877 | Buf = Builder.CreateElementBitCast(Buf, Int8Ty); |
| 3878 | return RValue::get(Builder.CreateCall(F, Buf.getPointer())); |
| 3879 | } |
| 3880 | case Builtin::BI__builtin_longjmp: { |
| 3881 | Value *Buf = EmitScalarExpr(E->getArg(0)); |
| 3882 | Buf = Builder.CreateBitCast(Buf, Int8PtrTy); |
| 3883 | |
| 3884 | // Call LLVM's EH longjmp, which is lightweight. |
| 3885 | Builder.CreateCall(CGM.getIntrinsic(Intrinsic::eh_sjlj_longjmp), Buf); |
| 3886 | |
| 3887 | // longjmp doesn't return; mark this as unreachable. |
| 3888 | Builder.CreateUnreachable(); |
| 3889 | |
| 3890 | // We do need to preserve an insertion point. |
| 3891 | EmitBlock(createBasicBlock("longjmp.cont")); |
| 3892 | |
| 3893 | return RValue::get(nullptr); |
| 3894 | } |
| 3895 | case Builtin::BI__builtin_launder: { |
| 3896 | const Expr *Arg = E->getArg(0); |
| 3897 | QualType ArgTy = Arg->getType()->getPointeeType(); |
| 3898 | Value *Ptr = EmitScalarExpr(Arg); |
| 3899 | if (TypeRequiresBuiltinLaunder(CGM, ArgTy)) |
| 3900 | Ptr = Builder.CreateLaunderInvariantGroup(Ptr); |
| 3901 | |
| 3902 | return RValue::get(Ptr); |
| 3903 | } |
| 3904 | case Builtin::BI__sync_fetch_and_add: |
| 3905 | case Builtin::BI__sync_fetch_and_sub: |
| 3906 | case Builtin::BI__sync_fetch_and_or: |
| 3907 | case Builtin::BI__sync_fetch_and_and: |
| 3908 | case Builtin::BI__sync_fetch_and_xor: |
| 3909 | case Builtin::BI__sync_fetch_and_nand: |
| 3910 | case Builtin::BI__sync_add_and_fetch: |
| 3911 | case Builtin::BI__sync_sub_and_fetch: |
| 3912 | case Builtin::BI__sync_and_and_fetch: |
| 3913 | case Builtin::BI__sync_or_and_fetch: |
| 3914 | case Builtin::BI__sync_xor_and_fetch: |
| 3915 | case Builtin::BI__sync_nand_and_fetch: |
| 3916 | case Builtin::BI__sync_val_compare_and_swap: |
| 3917 | case Builtin::BI__sync_bool_compare_and_swap: |
| 3918 | case Builtin::BI__sync_lock_test_and_set: |
| 3919 | case Builtin::BI__sync_lock_release: |
| 3920 | case Builtin::BI__sync_swap: |
| 3921 | llvm_unreachable("Shouldn't make it through sema")::llvm::llvm_unreachable_internal("Shouldn't make it through sema" , "clang/lib/CodeGen/CGBuiltin.cpp", 3921); |
| 3922 | case Builtin::BI__sync_fetch_and_add_1: |
| 3923 | case Builtin::BI__sync_fetch_and_add_2: |
| 3924 | case Builtin::BI__sync_fetch_and_add_4: |
| 3925 | case Builtin::BI__sync_fetch_and_add_8: |
| 3926 | case Builtin::BI__sync_fetch_and_add_16: |
| 3927 | return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Add, E); |
| 3928 | case Builtin::BI__sync_fetch_and_sub_1: |
| 3929 | case Builtin::BI__sync_fetch_and_sub_2: |
| 3930 | case Builtin::BI__sync_fetch_and_sub_4: |
| 3931 | case Builtin::BI__sync_fetch_and_sub_8: |
| 3932 | case Builtin::BI__sync_fetch_and_sub_16: |
| 3933 | return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Sub, E); |
| 3934 | case Builtin::BI__sync_fetch_and_or_1: |
| 3935 | case Builtin::BI__sync_fetch_and_or_2: |
| 3936 | case Builtin::BI__sync_fetch_and_or_4: |
| 3937 | case Builtin::BI__sync_fetch_and_or_8: |
| 3938 | case Builtin::BI__sync_fetch_and_or_16: |
| 3939 | return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Or, E); |
| 3940 | case Builtin::BI__sync_fetch_and_and_1: |
| 3941 | case Builtin::BI__sync_fetch_and_and_2: |
| 3942 | case Builtin::BI__sync_fetch_and_and_4: |
| 3943 | case Builtin::BI__sync_fetch_and_and_8: |
| 3944 | case Builtin::BI__sync_fetch_and_and_16: |
| 3945 | return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::And, E); |
| 3946 | case Builtin::BI__sync_fetch_and_xor_1: |
| 3947 | case Builtin::BI__sync_fetch_and_xor_2: |
| 3948 | case Builtin::BI__sync_fetch_and_xor_4: |
| 3949 | case Builtin::BI__sync_fetch_and_xor_8: |
| 3950 | case Builtin::BI__sync_fetch_and_xor_16: |
| 3951 | return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Xor, E); |
| 3952 | case Builtin::BI__sync_fetch_and_nand_1: |
| 3953 | case Builtin::BI__sync_fetch_and_nand_2: |
| 3954 | case Builtin::BI__sync_fetch_and_nand_4: |
| 3955 | case Builtin::BI__sync_fetch_and_nand_8: |
| 3956 | case Builtin::BI__sync_fetch_and_nand_16: |
| 3957 | return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Nand, E); |
| 3958 | |
| 3959 | // Clang extensions: not overloaded yet. |
| 3960 | case Builtin::BI__sync_fetch_and_min: |
| 3961 | return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Min, E); |
| 3962 | case Builtin::BI__sync_fetch_and_max: |
| 3963 | return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Max, E); |
| 3964 | case Builtin::BI__sync_fetch_and_umin: |
| 3965 | return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::UMin, E); |
| 3966 | case Builtin::BI__sync_fetch_and_umax: |
| 3967 | return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::UMax, E); |
| 3968 | |
| 3969 | case Builtin::BI__sync_add_and_fetch_1: |
| 3970 | case Builtin::BI__sync_add_and_fetch_2: |
| 3971 | case Builtin::BI__sync_add_and_fetch_4: |
| 3972 | case Builtin::BI__sync_add_and_fetch_8: |
| 3973 | case Builtin::BI__sync_add_and_fetch_16: |
| 3974 | return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::Add, E, |
| 3975 | llvm::Instruction::Add); |
| 3976 | case Builtin::BI__sync_sub_and_fetch_1: |
| 3977 | case Builtin::BI__sync_sub_and_fetch_2: |
| 3978 | case Builtin::BI__sync_sub_and_fetch_4: |
| 3979 | case Builtin::BI__sync_sub_and_fetch_8: |
| 3980 | case Builtin::BI__sync_sub_and_fetch_16: |
| 3981 | return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::Sub, E, |
| 3982 | llvm::Instruction::Sub); |
| 3983 | case Builtin::BI__sync_and_and_fetch_1: |
| 3984 | case Builtin::BI__sync_and_and_fetch_2: |
| 3985 | case Builtin::BI__sync_and_and_fetch_4: |
| 3986 | case Builtin::BI__sync_and_and_fetch_8: |
| 3987 | case Builtin::BI__sync_and_and_fetch_16: |
| 3988 | return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::And, E, |
| 3989 | llvm::Instruction::And); |
| 3990 | case Builtin::BI__sync_or_and_fetch_1: |
| 3991 | case Builtin::BI__sync_or_and_fetch_2: |
| 3992 | case Builtin::BI__sync_or_and_fetch_4: |
| 3993 | case Builtin::BI__sync_or_and_fetch_8: |
| 3994 | case Builtin::BI__sync_or_and_fetch_16: |
| 3995 | return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::Or, E, |
| 3996 | llvm::Instruction::Or); |
| 3997 | case Builtin::BI__sync_xor_and_fetch_1: |
| 3998 | case Builtin::BI__sync_xor_and_fetch_2: |
| 3999 | case Builtin::BI__sync_xor_and_fetch_4: |
| 4000 | case Builtin::BI__sync_xor_and_fetch_8: |
| 4001 | case Builtin::BI__sync_xor_and_fetch_16: |
| 4002 | return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::Xor, E, |
| 4003 | llvm::Instruction::Xor); |
| 4004 | case Builtin::BI__sync_nand_and_fetch_1: |
| 4005 | case Builtin::BI__sync_nand_and_fetch_2: |
| 4006 | case Builtin::BI__sync_nand_and_fetch_4: |
| 4007 | case Builtin::BI__sync_nand_and_fetch_8: |
| 4008 | case Builtin::BI__sync_nand_and_fetch_16: |
| 4009 | return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::Nand, E, |
| 4010 | llvm::Instruction::And, true); |
| 4011 | |
| 4012 | case Builtin::BI__sync_val_compare_and_swap_1: |
| 4013 | case Builtin::BI__sync_val_compare_and_swap_2: |
| 4014 | case Builtin::BI__sync_val_compare_and_swap_4: |
| 4015 | case Builtin::BI__sync_val_compare_and_swap_8: |
| 4016 | case Builtin::BI__sync_val_compare_and_swap_16: |
| 4017 | return RValue::get(MakeAtomicCmpXchgValue(*this, E, false)); |
| 4018 | |
| 4019 | case Builtin::BI__sync_bool_compare_and_swap_1: |
| 4020 | case Builtin::BI__sync_bool_compare_and_swap_2: |
| 4021 | case Builtin::BI__sync_bool_compare_and_swap_4: |
| 4022 | case Builtin::BI__sync_bool_compare_and_swap_8: |
| 4023 | case Builtin::BI__sync_bool_compare_and_swap_16: |
| 4024 | return RValue::get(MakeAtomicCmpXchgValue(*this, E, true)); |
| 4025 | |
| 4026 | case Builtin::BI__sync_swap_1: |
| 4027 | case Builtin::BI__sync_swap_2: |
| 4028 | case Builtin::BI__sync_swap_4: |
| 4029 | case Builtin::BI__sync_swap_8: |
| 4030 | case Builtin::BI__sync_swap_16: |
| 4031 | return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Xchg, E); |
| 4032 | |
| 4033 | case Builtin::BI__sync_lock_test_and_set_1: |
| 4034 | case Builtin::BI__sync_lock_test_and_set_2: |
| 4035 | case Builtin::BI__sync_lock_test_and_set_4: |
| 4036 | case Builtin::BI__sync_lock_test_and_set_8: |
| 4037 | case Builtin::BI__sync_lock_test_and_set_16: |
| 4038 | return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Xchg, E); |
| 4039 | |
| 4040 | case Builtin::BI__sync_lock_release_1: |
| 4041 | case Builtin::BI__sync_lock_release_2: |
| 4042 | case Builtin::BI__sync_lock_release_4: |
| 4043 | case Builtin::BI__sync_lock_release_8: |
| 4044 | case Builtin::BI__sync_lock_release_16: { |
| 4045 | Value *Ptr = CheckAtomicAlignment(*this, E); |
| 4046 | QualType ElTy = E->getArg(0)->getType()->getPointeeType(); |
| 4047 | CharUnits StoreSize = getContext().getTypeSizeInChars(ElTy); |
| 4048 | llvm::Type *ITy = llvm::IntegerType::get(getLLVMContext(), |
| 4049 | StoreSize.getQuantity() * 8); |
| 4050 | Ptr = Builder.CreateBitCast(Ptr, ITy->getPointerTo()); |
| 4051 | llvm::StoreInst *Store = |
| 4052 | Builder.CreateAlignedStore(llvm::Constant::getNullValue(ITy), Ptr, |
| 4053 | StoreSize); |
| 4054 | Store->setAtomic(llvm::AtomicOrdering::Release); |
| 4055 | return RValue::get(nullptr); |
| 4056 | } |
| 4057 | |
| 4058 | case Builtin::BI__sync_synchronize: { |
| 4059 | // We assume this is supposed to correspond to a C++0x-style |
| 4060 | // sequentially-consistent fence (i.e. this is only usable for |
| 4061 | // synchronization, not device I/O or anything like that). This intrinsic |
| 4062 | // is really badly designed in the sense that in theory, there isn't |
| 4063 | // any way to safely use it... but in practice, it mostly works |
| 4064 | // to use it with non-atomic loads and stores to get acquire/release |
| 4065 | // semantics. |
| 4066 | Builder.CreateFence(llvm::AtomicOrdering::SequentiallyConsistent); |
| 4067 | return RValue::get(nullptr); |
| 4068 | } |
| 4069 | |
| 4070 | case Builtin::BI__builtin_nontemporal_load: |
| 4071 | return RValue::get(EmitNontemporalLoad(*this, E)); |
| 4072 | case Builtin::BI__builtin_nontemporal_store: |
| 4073 | return RValue::get(EmitNontemporalStore(*this, E)); |
| 4074 | case Builtin::BI__c11_atomic_is_lock_free: |
| 4075 | case Builtin::BI__atomic_is_lock_free: { |
| 4076 | // Call "bool __atomic_is_lock_free(size_t size, void *ptr)". For the |
| 4077 | // __c11 builtin, ptr is 0 (indicating a properly-aligned object), since |
| 4078 | // _Atomic(T) is always properly-aligned. |
| 4079 | const char *LibCallName = "__atomic_is_lock_free"; |
| 4080 | CallArgList Args; |
| 4081 | Args.add(RValue::get(EmitScalarExpr(E->getArg(0))), |
| 4082 | getContext().getSizeType()); |
| 4083 | if (BuiltinID == Builtin::BI__atomic_is_lock_free) |
| 4084 | Args.add(RValue::get(EmitScalarExpr(E->getArg(1))), |
| 4085 | getContext().VoidPtrTy); |
| 4086 | else |
| 4087 | Args.add(RValue::get(llvm::Constant::getNullValue(VoidPtrTy)), |
| 4088 | getContext().VoidPtrTy); |
| 4089 | const CGFunctionInfo &FuncInfo = |
| 4090 | CGM.getTypes().arrangeBuiltinFunctionCall(E->getType(), Args); |
| 4091 | llvm::FunctionType *FTy = CGM.getTypes().GetFunctionType(FuncInfo); |
| 4092 | llvm::FunctionCallee Func = CGM.CreateRuntimeFunction(FTy, LibCallName); |
| 4093 | return EmitCall(FuncInfo, CGCallee::forDirect(Func), |
| 4094 | ReturnValueSlot(), Args); |
| 4095 | } |
| 4096 | |
| 4097 | case Builtin::BI__atomic_test_and_set: { |
| 4098 | // Look at the argument type to determine whether this is a volatile |
| 4099 | // operation. The parameter type is always volatile. |
| 4100 | QualType PtrTy = E->getArg(0)->IgnoreImpCasts()->getType(); |
| 4101 | bool Volatile = |
| 4102 | PtrTy->castAs<PointerType>()->getPointeeType().isVolatileQualified(); |
| 4103 | |
| 4104 | Value *Ptr = EmitScalarExpr(E->getArg(0)); |
| 4105 | unsigned AddrSpace = Ptr->getType()->getPointerAddressSpace(); |
| 4106 | Ptr = Builder.CreateBitCast(Ptr, Int8Ty->getPointerTo(AddrSpace)); |
| 4107 | Value *NewVal = Builder.getInt8(1); |
| 4108 | Value *Order = EmitScalarExpr(E->getArg(1)); |
| 4109 | if (isa<llvm::ConstantInt>(Order)) { |
| 4110 | int ord = cast<llvm::ConstantInt>(Order)->getZExtValue(); |
| 4111 | AtomicRMWInst *Result = nullptr; |
| 4112 | switch (ord) { |
| 4113 | case 0: // memory_order_relaxed |
| 4114 | default: // invalid order |
| 4115 | Result = Builder.CreateAtomicRMW(llvm::AtomicRMWInst::Xchg, Ptr, NewVal, |
| 4116 | llvm::AtomicOrdering::Monotonic); |
| 4117 | break; |
| 4118 | case 1: // memory_order_consume |
| 4119 | case 2: // memory_order_acquire |
| 4120 | Result = Builder.CreateAtomicRMW(llvm::AtomicRMWInst::Xchg, Ptr, NewVal, |
| 4121 | llvm::AtomicOrdering::Acquire); |
| 4122 | break; |
| 4123 | case 3: // memory_order_release |
| 4124 | Result = Builder.CreateAtomicRMW(llvm::AtomicRMWInst::Xchg, Ptr, NewVal, |
| 4125 | llvm::AtomicOrdering::Release); |
| 4126 | break; |
| 4127 | case 4: // memory_order_acq_rel |
| 4128 | |
| 4129 | Result = Builder.CreateAtomicRMW(llvm::AtomicRMWInst::Xchg, Ptr, NewVal, |
| 4130 | llvm::AtomicOrdering::AcquireRelease); |
| 4131 | break; |
| 4132 | case 5: // memory_order_seq_cst |
| 4133 | Result = Builder.CreateAtomicRMW( |
| 4134 | llvm::AtomicRMWInst::Xchg, Ptr, NewVal, |
| 4135 | llvm::AtomicOrdering::SequentiallyConsistent); |
| 4136 | break; |
| 4137 | } |
| 4138 | Result->setVolatile(Volatile); |
| 4139 | return RValue::get(Builder.CreateIsNotNull(Result, "tobool")); |
| 4140 | } |
| 4141 | |
| 4142 | llvm::BasicBlock *ContBB = createBasicBlock("atomic.continue", CurFn); |
| 4143 | |
| 4144 | llvm::BasicBlock *BBs[5] = { |
| 4145 | createBasicBlock("monotonic", CurFn), |
| 4146 | createBasicBlock("acquire", CurFn), |
| 4147 | createBasicBlock("release", CurFn), |
| 4148 | createBasicBlock("acqrel", CurFn), |
| 4149 | createBasicBlock("seqcst", CurFn) |
| 4150 | }; |
| 4151 | llvm::AtomicOrdering Orders[5] = { |
| 4152 | llvm::AtomicOrdering::Monotonic, llvm::AtomicOrdering::Acquire, |
| 4153 | llvm::AtomicOrdering::Release, llvm::AtomicOrdering::AcquireRelease, |
| 4154 | llvm::AtomicOrdering::SequentiallyConsistent}; |
| 4155 | |
| 4156 | Order = Builder.CreateIntCast(Order, Builder.getInt32Ty(), false); |
| 4157 | llvm::SwitchInst *SI = Builder.CreateSwitch(Order, BBs[0]); |
| 4158 | |
| 4159 | Builder.SetInsertPoint(ContBB); |
| 4160 | PHINode *Result = Builder.CreatePHI(Int8Ty, 5, "was_set"); |
| 4161 | |
| 4162 | for (unsigned i = 0; i < 5; ++i) { |
| 4163 | Builder.SetInsertPoint(BBs[i]); |
| 4164 | AtomicRMWInst *RMW = Builder.CreateAtomicRMW(llvm::AtomicRMWInst::Xchg, |
| 4165 | Ptr, NewVal, Orders[i]); |
| 4166 | RMW->setVolatile(Volatile); |
| 4167 | Result->addIncoming(RMW, BBs[i]); |
| 4168 | Builder.CreateBr(ContBB); |
| 4169 | } |
| 4170 | |
| 4171 | SI->addCase(Builder.getInt32(0), BBs[0]); |
| 4172 | SI->addCase(Builder.getInt32(1), BBs[1]); |
| 4173 | SI->addCase(Builder.getInt32(2), BBs[1]); |
| 4174 | SI->addCase(Builder.getInt32(3), BBs[2]); |
| 4175 | SI->addCase(Builder.getInt32(4), BBs[3]); |
| 4176 | SI->addCase(Builder.getInt32(5), BBs[4]); |
| 4177 | |
| 4178 | Builder.SetInsertPoint(ContBB); |
| 4179 | return RValue::get(Builder.CreateIsNotNull(Result, "tobool")); |
| 4180 | } |
| 4181 | |
| 4182 | case Builtin::BI__atomic_clear: { |
| 4183 | QualType PtrTy = E->getArg(0)->IgnoreImpCasts()->getType(); |
| 4184 | bool Volatile = |
| 4185 | PtrTy->castAs<PointerType>()->getPointeeType().isVolatileQualified(); |
| 4186 | |
| 4187 | Address Ptr = EmitPointerWithAlignment(E->getArg(0)); |
| 4188 | Ptr = Builder.CreateElementBitCast(Ptr, Int8Ty); |
| 4189 | Value *NewVal = Builder.getInt8(0); |
| 4190 | Value *Order = EmitScalarExpr(E->getArg(1)); |
| 4191 | if (isa<llvm::ConstantInt>(Order)) { |
| 4192 | int ord = cast<llvm::ConstantInt>(Order)->getZExtValue(); |
| 4193 | StoreInst *Store = Builder.CreateStore(NewVal, Ptr, Volatile); |
| 4194 | switch (ord) { |
| 4195 | case 0: // memory_order_relaxed |
| 4196 | default: // invalid order |
| 4197 | Store->setOrdering(llvm::AtomicOrdering::Monotonic); |
| 4198 | break; |
| 4199 | case 3: // memory_order_release |
| 4200 | Store->setOrdering(llvm::AtomicOrdering::Release); |
| 4201 | break; |
| 4202 | case 5: // memory_order_seq_cst |
| 4203 | Store->setOrdering(llvm::AtomicOrdering::SequentiallyConsistent); |
| 4204 | break; |
| 4205 | } |
| 4206 | return RValue::get(nullptr); |
| 4207 | } |
| 4208 | |
| 4209 | llvm::BasicBlock *ContBB = createBasicBlock("atomic.continue", CurFn); |
| 4210 | |
| 4211 | llvm::BasicBlock *BBs[3] = { |
| 4212 | createBasicBlock("monotonic", CurFn), |
| 4213 | createBasicBlock("release", CurFn), |
| 4214 | createBasicBlock("seqcst", CurFn) |
| 4215 | }; |
| 4216 | llvm::AtomicOrdering Orders[3] = { |
| 4217 | llvm::AtomicOrdering::Monotonic, llvm::AtomicOrdering::Release, |
| 4218 | llvm::AtomicOrdering::SequentiallyConsistent}; |
| 4219 | |
| 4220 | Order = Builder.CreateIntCast(Order, Builder.getInt32Ty(), false); |
| 4221 | llvm::SwitchInst *SI = Builder.CreateSwitch(Order, BBs[0]); |
| 4222 | |
| 4223 | for (unsigned i = 0; i < 3; ++i) { |
| 4224 | Builder.SetInsertPoint(BBs[i]); |
| 4225 | StoreInst *Store = Builder.CreateStore(NewVal, Ptr, Volatile); |
| 4226 | Store->setOrdering(Orders[i]); |
| 4227 | Builder.CreateBr(ContBB); |
| 4228 | } |
| 4229 | |
| 4230 | SI->addCase(Builder.getInt32(0), BBs[0]); |
| 4231 | SI->addCase(Builder.getInt32(3), BBs[1]); |
| 4232 | SI->addCase(Builder.getInt32(5), BBs[2]); |
| 4233 | |
| 4234 | Builder.SetInsertPoint(ContBB); |
| 4235 | return RValue::get(nullptr); |
| 4236 | } |
| 4237 | |
| 4238 | case Builtin::BI__atomic_thread_fence: |
| 4239 | case Builtin::BI__atomic_signal_fence: |
| 4240 | case Builtin::BI__c11_atomic_thread_fence: |
| 4241 | case Builtin::BI__c11_atomic_signal_fence: { |
| 4242 | llvm::SyncScope::ID SSID; |
| 4243 | if (BuiltinID == Builtin::BI__atomic_signal_fence || |
| 4244 | BuiltinID == Builtin::BI__c11_atomic_signal_fence) |
| 4245 | SSID = llvm::SyncScope::SingleThread; |
| 4246 | else |
| 4247 | SSID = llvm::SyncScope::System; |
| 4248 | Value *Order = EmitScalarExpr(E->getArg(0)); |
| 4249 | if (isa<llvm::ConstantInt>(Order)) { |
| 4250 | int ord = cast<llvm::ConstantInt>(Order)->getZExtValue(); |
| 4251 | switch (ord) { |
| 4252 | case 0: // memory_order_relaxed |
| 4253 | default: // invalid order |
| 4254 | break; |
| 4255 | case 1: // memory_order_consume |
| 4256 | case 2: // memory_order_acquire |
| 4257 | Builder.CreateFence(llvm::AtomicOrdering::Acquire, SSID); |
| 4258 | break; |
| 4259 | case 3: // memory_order_release |
| 4260 | Builder.CreateFence(llvm::AtomicOrdering::Release, SSID); |
| 4261 | break; |
| 4262 | case 4: // memory_order_acq_rel |
| 4263 | Builder.CreateFence(llvm::AtomicOrdering::AcquireRelease, SSID); |
| 4264 | break; |
| 4265 | case 5: // memory_order_seq_cst |
| 4266 | Builder.CreateFence(llvm::AtomicOrdering::SequentiallyConsistent, SSID); |
| 4267 | break; |
| 4268 | } |
| 4269 | return RValue::get(nullptr); |
| 4270 | } |
| 4271 | |
| 4272 | llvm::BasicBlock *AcquireBB, *ReleaseBB, *AcqRelBB, *SeqCstBB; |
| 4273 | AcquireBB = createBasicBlock("acquire", CurFn); |
| 4274 | ReleaseBB = createBasicBlock("release", CurFn); |
| 4275 | AcqRelBB = createBasicBlock("acqrel", CurFn); |
| 4276 | SeqCstBB = createBasicBlock("seqcst", CurFn); |
| 4277 | llvm::BasicBlock *ContBB = createBasicBlock("atomic.continue", CurFn); |
| 4278 | |
| 4279 | Order = Builder.CreateIntCast(Order, Builder.getInt32Ty(), false); |
| 4280 | llvm::SwitchInst *SI = Builder.CreateSwitch(Order, ContBB); |
| 4281 | |
| 4282 | Builder.SetInsertPoint(AcquireBB); |
| 4283 | Builder.CreateFence(llvm::AtomicOrdering::Acquire, SSID); |
| 4284 | Builder.CreateBr(ContBB); |
| 4285 | SI->addCase(Builder.getInt32(1), AcquireBB); |
| 4286 | SI->addCase(Builder.getInt32(2), AcquireBB); |
| 4287 | |
| 4288 | Builder.SetInsertPoint(ReleaseBB); |
| 4289 | Builder.CreateFence(llvm::AtomicOrdering::Release, SSID); |
| 4290 | Builder.CreateBr(ContBB); |
| 4291 | SI->addCase(Builder.getInt32(3), ReleaseBB); |
| 4292 | |
| 4293 | Builder.SetInsertPoint(AcqRelBB); |
| 4294 | Builder.CreateFence(llvm::AtomicOrdering::AcquireRelease, SSID); |
| 4295 | Builder.CreateBr(ContBB); |
| 4296 | SI->addCase(Builder.getInt32(4), AcqRelBB); |
| 4297 | |
| 4298 | Builder.SetInsertPoint(SeqCstBB); |
| 4299 | Builder.CreateFence(llvm::AtomicOrdering::SequentiallyConsistent, SSID); |
| 4300 | Builder.CreateBr(ContBB); |
| 4301 | SI->addCase(Builder.getInt32(5), SeqCstBB); |
| 4302 | |
| 4303 | Builder.SetInsertPoint(ContBB); |
| 4304 | return RValue::get(nullptr); |
| 4305 | } |
| 4306 | |
| 4307 | case Builtin::BI__builtin_signbit: |
| 4308 | case Builtin::BI__builtin_signbitf: |
| 4309 | case Builtin::BI__builtin_signbitl: { |
| 4310 | return RValue::get( |
| 4311 | Builder.CreateZExt(EmitSignBit(*this, EmitScalarExpr(E->getArg(0))), |
| 4312 | ConvertType(E->getType()))); |
| 4313 | } |
| 4314 | case Builtin::BI__warn_memset_zero_len: |
| 4315 | return RValue::getIgnored(); |
| 4316 | case Builtin::BI__annotation: { |
| 4317 | // Re-encode each wide string to UTF8 and make an MDString. |
| 4318 | SmallVector<Metadata *, 1> Strings; |
| 4319 | for (const Expr *Arg : E->arguments()) { |
| 4320 | const auto *Str = cast<StringLiteral>(Arg->IgnoreParenCasts()); |
| 4321 | assert(Str->getCharByteWidth() == 2)(static_cast <bool> (Str->getCharByteWidth() == 2) ? void (0) : __assert_fail ("Str->getCharByteWidth() == 2", "clang/lib/CodeGen/CGBuiltin.cpp", 4321, __extension__ __PRETTY_FUNCTION__ )); |
| 4322 | StringRef WideBytes = Str->getBytes(); |
| 4323 | std::string StrUtf8; |
| 4324 | if (!convertUTF16ToUTF8String( |
| 4325 | ArrayRef(WideBytes.data(), WideBytes.size()), StrUtf8)) { |
| 4326 | CGM.ErrorUnsupported(E, "non-UTF16 __annotation argument"); |
| 4327 | continue; |
| 4328 | } |
| 4329 | Strings.push_back(llvm::MDString::get(getLLVMContext(), StrUtf8)); |
| 4330 | } |
| 4331 | |
| 4332 | // Build and MDTuple of MDStrings and emit the intrinsic call. |
| 4333 | llvm::Function *F = |
| 4334 | CGM.getIntrinsic(llvm::Intrinsic::codeview_annotation, {}); |
| 4335 | MDTuple *StrTuple = MDTuple::get(getLLVMContext(), Strings); |
| 4336 | Builder.CreateCall(F, MetadataAsValue::get(getLLVMContext(), StrTuple)); |
| 4337 | return RValue::getIgnored(); |
| 4338 | } |
| 4339 | case Builtin::BI__builtin_annotation: { |
| 4340 | llvm::Value *AnnVal = EmitScalarExpr(E->getArg(0)); |
| 4341 | llvm::Function *F = |
| 4342 | CGM.getIntrinsic(llvm::Intrinsic::annotation, |
| 4343 | {AnnVal->getType(), CGM.ConstGlobalsPtrTy}); |
| 4344 | |
| 4345 | // Get the annotation string, go through casts. Sema requires this to be a |
| 4346 | // non-wide string literal, potentially casted, so the cast<> is safe. |
| 4347 | const Expr *AnnotationStrExpr = E->getArg(1)->IgnoreParenCasts(); |
| 4348 | StringRef Str = cast<StringLiteral>(AnnotationStrExpr)->getString(); |
| 4349 | return RValue::get( |
| 4350 | EmitAnnotationCall(F, AnnVal, Str, E->getExprLoc(), nullptr)); |
| 4351 | } |
| 4352 | case Builtin::BI__builtin_addcb: |
| 4353 | case Builtin::BI__builtin_addcs: |
| 4354 | case Builtin::BI__builtin_addc: |
| 4355 | case Builtin::BI__builtin_addcl: |
| 4356 | case Builtin::BI__builtin_addcll: |
| 4357 | case Builtin::BI__builtin_subcb: |
| 4358 | case Builtin::BI__builtin_subcs: |
| 4359 | case Builtin::BI__builtin_subc: |
| 4360 | case Builtin::BI__builtin_subcl: |
| 4361 | case Builtin::BI__builtin_subcll: { |
| 4362 | |
| 4363 | // We translate all of these builtins from expressions of the form: |
| 4364 | // int x = ..., y = ..., carryin = ..., carryout, result; |
| 4365 | // result = __builtin_addc(x, y, carryin, &carryout); |
| 4366 | // |
| 4367 | // to LLVM IR of the form: |
| 4368 | // |
| 4369 | // %tmp1 = call {i32, i1} @llvm.uadd.with.overflow.i32(i32 %x, i32 %y) |
| 4370 | // %tmpsum1 = extractvalue {i32, i1} %tmp1, 0 |
| 4371 | // %carry1 = extractvalue {i32, i1} %tmp1, 1 |
| 4372 | // %tmp2 = call {i32, i1} @llvm.uadd.with.overflow.i32(i32 %tmpsum1, |
| 4373 | // i32 %carryin) |
| 4374 | // %result = extractvalue {i32, i1} %tmp2, 0 |
| 4375 | // %carry2 = extractvalue {i32, i1} %tmp2, 1 |
| 4376 | // %tmp3 = or i1 %carry1, %carry2 |
| 4377 | // %tmp4 = zext i1 %tmp3 to i32 |
| 4378 | // store i32 %tmp4, i32* %carryout |
| 4379 | |
| 4380 | // Scalarize our inputs. |
| 4381 | llvm::Value *X = EmitScalarExpr(E->getArg(0)); |
| 4382 | llvm::Value *Y = EmitScalarExpr(E->getArg(1)); |
| 4383 | llvm::Value *Carryin = EmitScalarExpr(E->getArg(2)); |
| 4384 | Address CarryOutPtr = EmitPointerWithAlignment(E->getArg(3)); |
| 4385 | |
| 4386 | // Decide if we are lowering to a uadd.with.overflow or usub.with.overflow. |
| 4387 | llvm::Intrinsic::ID IntrinsicId; |
| 4388 | switch (BuiltinID) { |
| 4389 | default: llvm_unreachable("Unknown multiprecision builtin id.")::llvm::llvm_unreachable_internal("Unknown multiprecision builtin id." , "clang/lib/CodeGen/CGBuiltin.cpp", 4389); |
| 4390 | case Builtin::BI__builtin_addcb: |
| 4391 | case Builtin::BI__builtin_addcs: |
| 4392 | case Builtin::BI__builtin_addc: |
| 4393 | case Builtin::BI__builtin_addcl: |
| 4394 | case Builtin::BI__builtin_addcll: |
| 4395 | IntrinsicId = llvm::Intrinsic::uadd_with_overflow; |
| 4396 | break; |
| 4397 | case Builtin::BI__builtin_subcb: |
| 4398 | case Builtin::BI__builtin_subcs: |
| 4399 | case Builtin::BI__builtin_subc: |
| 4400 | case Builtin::BI__builtin_subcl: |
| 4401 | case Builtin::BI__builtin_subcll: |
| 4402 | IntrinsicId = llvm::Intrinsic::usub_with_overflow; |
| 4403 | break; |
| 4404 | } |
| 4405 | |
| 4406 | // Construct our resulting LLVM IR expression. |
| 4407 | llvm::Value *Carry1; |
| 4408 | llvm::Value *Sum1 = EmitOverflowIntrinsic(*this, IntrinsicId, |
| 4409 | X, Y, Carry1); |
| 4410 | llvm::Value *Carry2; |
| 4411 | llvm::Value *Sum2 = EmitOverflowIntrinsic(*this, IntrinsicId, |
| 4412 | Sum1, Carryin, Carry2); |
| 4413 | llvm::Value *CarryOut = Builder.CreateZExt(Builder.CreateOr(Carry1, Carry2), |
| 4414 | X->getType()); |
| 4415 | Builder.CreateStore(CarryOut, CarryOutPtr); |
| 4416 | return RValue::get(Sum2); |
| 4417 | } |
| 4418 | |
| 4419 | case Builtin::BI__builtin_add_overflow: |
| 4420 | case Builtin::BI__builtin_sub_overflow: |
| 4421 | case Builtin::BI__builtin_mul_overflow: { |
| 4422 | const clang::Expr *LeftArg = E->getArg(0); |
| 4423 | const clang::Expr *RightArg = E->getArg(1); |
| 4424 | const clang::Expr *ResultArg = E->getArg(2); |
| 4425 | |
| 4426 | clang::QualType ResultQTy = |
| 4427 | ResultArg->getType()->castAs<PointerType>()->getPointeeType(); |
| 4428 | |
| 4429 | WidthAndSignedness LeftInfo = |
| 4430 | getIntegerWidthAndSignedness(CGM.getContext(), LeftArg->getType()); |
| 4431 | WidthAndSignedness RightInfo = |
| 4432 | getIntegerWidthAndSignedness(CGM.getContext(), RightArg->getType()); |
| 4433 | WidthAndSignedness ResultInfo = |
| 4434 | getIntegerWidthAndSignedness(CGM.getContext(), ResultQTy); |
| 4435 | |
| 4436 | // Handle mixed-sign multiplication as a special case, because adding |
| 4437 | // runtime or backend support for our generic irgen would be too expensive. |
| 4438 | if (isSpecialMixedSignMultiply(BuiltinID, LeftInfo, RightInfo, ResultInfo)) |
| 4439 | return EmitCheckedMixedSignMultiply(*this, LeftArg, LeftInfo, RightArg, |
| 4440 | RightInfo, ResultArg, ResultQTy, |
| 4441 | ResultInfo); |
| 4442 | |
| 4443 | if (isSpecialUnsignedMultiplySignedResult(BuiltinID, LeftInfo, RightInfo, |
| 4444 | ResultInfo)) |
| 4445 | return EmitCheckedUnsignedMultiplySignedResult( |
| 4446 | *this, LeftArg, LeftInfo, RightArg, RightInfo, ResultArg, ResultQTy, |
| 4447 | ResultInfo); |
| 4448 | |
| 4449 | WidthAndSignedness EncompassingInfo = |
| 4450 | EncompassingIntegerType({LeftInfo, RightInfo, ResultInfo}); |
| 4451 | |
| 4452 | llvm::Type *EncompassingLLVMTy = |
| 4453 | llvm::IntegerType::get(CGM.getLLVMContext(), EncompassingInfo.Width); |
| 4454 | |
| 4455 | llvm::Type *ResultLLVMTy = CGM.getTypes().ConvertType(ResultQTy); |
| 4456 | |
| 4457 | llvm::Intrinsic::ID IntrinsicId; |
| 4458 | switch (BuiltinID) { |
| 4459 | default: |
| 4460 | llvm_unreachable("Unknown overflow builtin id.")::llvm::llvm_unreachable_internal("Unknown overflow builtin id." , "clang/lib/CodeGen/CGBuiltin.cpp", 4460); |
| 4461 | case Builtin::BI__builtin_add_overflow: |
| 4462 | IntrinsicId = EncompassingInfo.Signed |
| 4463 | ? llvm::Intrinsic::sadd_with_overflow |
| 4464 | : llvm::Intrinsic::uadd_with_overflow; |
| 4465 | break; |
| 4466 | case Builtin::BI__builtin_sub_overflow: |
| 4467 | IntrinsicId = EncompassingInfo.Signed |
| 4468 | ? llvm::Intrinsic::ssub_with_overflow |
| 4469 | : llvm::Intrinsic::usub_with_overflow; |
| 4470 | break; |
| 4471 | case Builtin::BI__builtin_mul_overflow: |
| 4472 | IntrinsicId = EncompassingInfo.Signed |
| 4473 | ? llvm::Intrinsic::smul_with_overflow |
| 4474 | : llvm::Intrinsic::umul_with_overflow; |
| 4475 | break; |
| 4476 | } |
| 4477 | |
| 4478 | llvm::Value *Left = EmitScalarExpr(LeftArg); |
| 4479 | llvm::Value *Right = EmitScalarExpr(RightArg); |
| 4480 | Address ResultPtr = EmitPointerWithAlignment(ResultArg); |
| 4481 | |
| 4482 | // Extend each operand to the encompassing type. |
| 4483 | Left = Builder.CreateIntCast(Left, EncompassingLLVMTy, LeftInfo.Signed); |
| 4484 | Right = Builder.CreateIntCast(Right, EncompassingLLVMTy, RightInfo.Signed); |
| 4485 | |
| 4486 | // Perform the operation on the extended values. |
| 4487 | llvm::Value *Overflow, *Result; |
| 4488 | Result = EmitOverflowIntrinsic(*this, IntrinsicId, Left, Right, Overflow); |
| 4489 | |
| 4490 | if (EncompassingInfo.Width > ResultInfo.Width) { |
| 4491 | // The encompassing type is wider than the result type, so we need to |
| 4492 | // truncate it. |
| 4493 | llvm::Value *ResultTrunc = Builder.CreateTrunc(Result, ResultLLVMTy); |
| 4494 | |
| 4495 | // To see if the truncation caused an overflow, we will extend |
| 4496 | // the result and then compare it to the original result. |
| 4497 | llvm::Value *ResultTruncExt = Builder.CreateIntCast( |
| 4498 | ResultTrunc, EncompassingLLVMTy, ResultInfo.Signed); |
| 4499 | llvm::Value *TruncationOverflow = |
| 4500 | Builder.CreateICmpNE(Result, ResultTruncExt); |
| 4501 | |
| 4502 | Overflow = Builder.CreateOr(Overflow, TruncationOverflow); |
| 4503 | Result = ResultTrunc; |
| 4504 | } |
| 4505 | |
| 4506 | // Finally, store the result using the pointer. |
| 4507 | bool isVolatile = |
| 4508 | ResultArg->getType()->getPointeeType().isVolatileQualified(); |
| 4509 | Builder.CreateStore(EmitToMemory(Result, ResultQTy), ResultPtr, isVolatile); |
| 4510 | |
| 4511 | return RValue::get(Overflow); |
| 4512 | } |
| 4513 | |
| 4514 | case Builtin::BI__builtin_uadd_overflow: |
| 4515 | case Builtin::BI__builtin_uaddl_overflow: |
| 4516 | case Builtin::BI__builtin_uaddll_overflow: |
| 4517 | case Builtin::BI__builtin_usub_overflow: |
| 4518 | case Builtin::BI__builtin_usubl_overflow: |
| 4519 | case Builtin::BI__builtin_usubll_overflow: |
| 4520 | case Builtin::BI__builtin_umul_overflow: |
| 4521 | case Builtin::BI__builtin_umull_overflow: |
| 4522 | case Builtin::BI__builtin_umulll_overflow: |
| 4523 | case Builtin::BI__builtin_sadd_overflow: |
| 4524 | case Builtin::BI__builtin_saddl_overflow: |
| 4525 | case Builtin::BI__builtin_saddll_overflow: |
| 4526 | case Builtin::BI__builtin_ssub_overflow: |
| 4527 | case Builtin::BI__builtin_ssubl_overflow: |
| 4528 | case Builtin::BI__builtin_ssubll_overflow: |
| 4529 | case Builtin::BI__builtin_smul_overflow: |
| 4530 | case Builtin::BI__builtin_smull_overflow: |
| 4531 | case Builtin::BI__builtin_smulll_overflow: { |
| 4532 | |
| 4533 | // We translate all of these builtins directly to the relevant llvm IR node. |
| 4534 | |
| 4535 | // Scalarize our inputs. |
| 4536 | llvm::Value *X = EmitScalarExpr(E->getArg(0)); |
| 4537 | llvm::Value *Y = EmitScalarExpr(E->getArg(1)); |
| 4538 | Address SumOutPtr = EmitPointerWithAlignment(E->getArg(2)); |
| 4539 | |
| 4540 | // Decide which of the overflow intrinsics we are lowering to: |
| 4541 | llvm::Intrinsic::ID IntrinsicId; |
| 4542 | switch (BuiltinID) { |
| 4543 | default: llvm_unreachable("Unknown overflow builtin id.")::llvm::llvm_unreachable_internal("Unknown overflow builtin id." , "clang/lib/CodeGen/CGBuiltin.cpp", 4543); |
| 4544 | case Builtin::BI__builtin_uadd_overflow: |
| 4545 | case Builtin::BI__builtin_uaddl_overflow: |
| 4546 | case Builtin::BI__builtin_uaddll_overflow: |
| 4547 | IntrinsicId = llvm::Intrinsic::uadd_with_overflow; |
| 4548 | break; |
| 4549 | case Builtin::BI__builtin_usub_overflow: |
| 4550 | case Builtin::BI__builtin_usubl_overflow: |
| 4551 | case Builtin::BI__builtin_usubll_overflow: |
| 4552 | IntrinsicId = llvm::Intrinsic::usub_with_overflow; |
| 4553 | break; |
| 4554 | case Builtin::BI__builtin_umul_overflow: |
| 4555 | case Builtin::BI__builtin_umull_overflow: |
| 4556 | case Builtin::BI__builtin_umulll_overflow: |
| 4557 | IntrinsicId = llvm::Intrinsic::umul_with_overflow; |
| 4558 | break; |
| 4559 | case Builtin::BI__builtin_sadd_overflow: |
| 4560 | case Builtin::BI__builtin_saddl_overflow: |
| 4561 | case Builtin::BI__builtin_saddll_overflow: |
| 4562 | IntrinsicId = llvm::Intrinsic::sadd_with_overflow; |
| 4563 | break; |
| 4564 | case Builtin::BI__builtin_ssub_overflow: |
| 4565 | case Builtin::BI__builtin_ssubl_overflow: |
| 4566 | case Builtin::BI__builtin_ssubll_overflow: |
| 4567 | IntrinsicId = llvm::Intrinsic::ssub_with_overflow; |
| 4568 | break; |
| 4569 | case Builtin::BI__builtin_smul_overflow: |
| 4570 | case Builtin::BI__builtin_smull_overflow: |
| 4571 | case Builtin::BI__builtin_smulll_overflow: |
| 4572 | IntrinsicId = llvm::Intrinsic::smul_with_overflow; |
| 4573 | break; |
| 4574 | } |
| 4575 | |
| 4576 | |
| 4577 | llvm::Value *Carry; |
| 4578 | llvm::Value *Sum = EmitOverflowIntrinsic(*this, IntrinsicId, X, Y, Carry); |
| 4579 | Builder.CreateStore(Sum, SumOutPtr); |
| 4580 | |
| 4581 | return RValue::get(Carry); |
| 4582 | } |
| 4583 | case Builtin::BIaddressof: |
| 4584 | case Builtin::BI__addressof: |
| 4585 | case Builtin::BI__builtin_addressof: |
| 4586 | return RValue::get(EmitLValue(E->getArg(0)).getPointer(*this)); |
| 4587 | case Builtin::BI__builtin_function_start: |
| 4588 | return RValue::get(CGM.GetFunctionStart( |
| 4589 | E->getArg(0)->getAsBuiltinConstantDeclRef(CGM.getContext()))); |
| 4590 | case Builtin::BI__builtin_operator_new: |
| 4591 | return EmitBuiltinNewDeleteCall( |
| 4592 | E->getCallee()->getType()->castAs<FunctionProtoType>(), E, false); |
| 4593 | case Builtin::BI__builtin_operator_delete: |
| 4594 | EmitBuiltinNewDeleteCall( |
| 4595 | E->getCallee()->getType()->castAs<FunctionProtoType>(), E, true); |
| 4596 | return RValue::get(nullptr); |
| 4597 | |
| 4598 | case Builtin::BI__builtin_is_aligned: |
| 4599 | return EmitBuiltinIsAligned(E); |
| 4600 | case Builtin::BI__builtin_align_up: |
| 4601 | return EmitBuiltinAlignTo(E, true); |
| 4602 | case Builtin::BI__builtin_align_down: |
| 4603 | return EmitBuiltinAlignTo(E, false); |
| 4604 | |
| 4605 | case Builtin::BI__noop: |
| 4606 | // __noop always evaluates to an integer literal zero. |
| 4607 | return RValue::get(ConstantInt::get(IntTy, 0)); |
| 4608 | case Builtin::BI__builtin_call_with_static_chain: { |
| 4609 | const CallExpr *Call = cast<CallExpr>(E->getArg(0)); |
| 4610 | const Expr *Chain = E->getArg(1); |
| 4611 | return EmitCall(Call->getCallee()->getType(), |
| 4612 | EmitCallee(Call->getCallee()), Call, ReturnValue, |
| 4613 | EmitScalarExpr(Chain)); |
| 4614 | } |
| 4615 | case Builtin::BI_InterlockedExchange8: |
| 4616 | case Builtin::BI_InterlockedExchange16: |
| 4617 | case Builtin::BI_InterlockedExchange: |
| 4618 | case Builtin::BI_InterlockedExchangePointer: |
| 4619 | return RValue::get( |
| 4620 | EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedExchange, E)); |
| 4621 | case Builtin::BI_InterlockedCompareExchangePointer: |
| 4622 | case Builtin::BI_InterlockedCompareExchangePointer_nf: { |
| 4623 | llvm::Type *RTy; |
| 4624 | llvm::IntegerType *IntType = |
| 4625 | IntegerType::get(getLLVMContext(), |
| 4626 | getContext().getTypeSize(E->getType())); |
| 4627 | llvm::Type *IntPtrType = IntType->getPointerTo(); |
| 4628 | |
| 4629 | llvm::Value *Destination = |
| 4630 | Builder.CreateBitCast(EmitScalarExpr(E->getArg(0)), IntPtrType); |
| 4631 | |
| 4632 | llvm::Value *Exchange = EmitScalarExpr(E->getArg(1)); |
| 4633 | RTy = Exchange->getType(); |
| 4634 | Exchange = Builder.CreatePtrToInt(Exchange, IntType); |
| 4635 | |
| 4636 | llvm::Value *Comparand = |
| 4637 | Builder.CreatePtrToInt(EmitScalarExpr(E->getArg(2)), IntType); |
| 4638 | |
| 4639 | auto Ordering = |
| 4640 | BuiltinID == Builtin::BI_InterlockedCompareExchangePointer_nf ? |
| 4641 | AtomicOrdering::Monotonic : AtomicOrdering::SequentiallyConsistent; |
| 4642 | |
| 4643 | auto Result = Builder.CreateAtomicCmpXchg(Destination, Comparand, Exchange, |
| 4644 | Ordering, Ordering); |
| 4645 | Result->setVolatile(true); |
| 4646 | |
| 4647 | return RValue::get(Builder.CreateIntToPtr(Builder.CreateExtractValue(Result, |
| 4648 | 0), |
| 4649 | RTy)); |
| 4650 | } |
| 4651 | case Builtin::BI_InterlockedCompareExchange8: |
| 4652 | case Builtin::BI_InterlockedCompareExchange16: |
| 4653 | case Builtin::BI_InterlockedCompareExchange: |
| 4654 | case Builtin::BI_InterlockedCompareExchange64: |
| 4655 | return RValue::get(EmitAtomicCmpXchgForMSIntrin(*this, E)); |
| 4656 | case Builtin::BI_InterlockedIncrement16: |
| 4657 | case Builtin::BI_InterlockedIncrement: |
| 4658 | return RValue::get( |
| 4659 | EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedIncrement, E)); |
| 4660 | case Builtin::BI_InterlockedDecrement16: |
| 4661 | case Builtin::BI_InterlockedDecrement: |
| 4662 | return RValue::get( |
| 4663 | EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedDecrement, E)); |
| 4664 | case Builtin::BI_InterlockedAnd8: |
| 4665 | case Builtin::BI_InterlockedAnd16: |
| 4666 | case Builtin::BI_InterlockedAnd: |
| 4667 | return RValue::get(EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedAnd, E)); |
| 4668 | case Builtin::BI_InterlockedExchangeAdd8: |
| 4669 | case Builtin::BI_InterlockedExchangeAdd16: |
| 4670 | case Builtin::BI_InterlockedExchangeAdd: |
| 4671 | return RValue::get( |
| 4672 | EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedExchangeAdd, E)); |
| 4673 | case Builtin::BI_InterlockedExchangeSub8: |
| 4674 | case Builtin::BI_InterlockedExchangeSub16: |
| 4675 | case Builtin::BI_InterlockedExchangeSub: |
| 4676 | return RValue::get( |
| 4677 | EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedExchangeSub, E)); |
| 4678 | case Builtin::BI_InterlockedOr8: |
| 4679 | case Builtin::BI_InterlockedOr16: |
| 4680 | case Builtin::BI_InterlockedOr: |
| 4681 | return RValue::get(EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedOr, E)); |
| 4682 | case Builtin::BI_InterlockedXor8: |
| 4683 | case Builtin::BI_InterlockedXor16: |
| 4684 | case Builtin::BI_InterlockedXor: |
| 4685 | return RValue::get(EmitMSVCBuiltinExpr(MSVCIntrin::_InterlockedXor, E)); |
| 4686 | |
| 4687 | case Builtin::BI_bittest64: |
| 4688 | case Builtin::BI_bittest: |
| 4689 | case Builtin::BI_bittestandcomplement64: |
| 4690 | case Builtin::BI_bittestandcomplement: |
| 4691 | case Builtin::BI_bittestandreset64: |
| 4692 | case Builtin::BI_bittestandreset: |
| 4693 | case Builtin::BI_bittestandset64: |
| 4694 | case Builtin::BI_bittestandset: |
| 4695 | case Builtin::BI_interlockedbittestandreset: |
| 4696 | case Builtin::BI_interlockedbittestandreset64: |
| 4697 | case Builtin::BI_interlockedbittestandset64: |
| 4698 | case Builtin::BI_interlockedbittestandset: |
| 4699 | case Builtin::BI_interlockedbittestandset_acq: |
| 4700 | case Builtin::BI_interlockedbittestandset_rel: |
| 4701 | case Builtin::BI_interlockedbittestandset_nf: |
| 4702 | case Builtin::BI_interlockedbittestandreset_acq: |
| 4703 | case Builtin::BI_interlockedbittestandreset_rel: |
| 4704 | case Builtin::BI_interlockedbittestandreset_nf: |
| 4705 | return RValue::get(EmitBitTestIntrinsic(*this, BuiltinID, E)); |
| 4706 | |
| 4707 | // These builtins exist to emit regular volatile loads and stores not |
| 4708 | // affected by the -fms-volatile setting. |
| 4709 | case Builtin::BI__iso_volatile_load8: |
| 4710 | case Builtin::BI__iso_volatile_load16: |
| 4711 | case Builtin::BI__iso_volatile_load32: |
| 4712 | case Builtin::BI__iso_volatile_load64: |
| 4713 | return RValue::get(EmitISOVolatileLoad(*this, E)); |
| 4714 | case Builtin::BI__iso_volatile_store8: |
| 4715 | case Builtin::BI__iso_volatile_store16: |
| 4716 | case Builtin::BI__iso_volatile_store32: |
| 4717 | case Builtin::BI__iso_volatile_store64: |
| 4718 | return RValue::get(EmitISOVolatileStore(*this, E)); |
| 4719 | |
| 4720 | case Builtin::BI__exception_code: |
| 4721 | case Builtin::BI_exception_code: |
| 4722 | return RValue::get(EmitSEHExceptionCode()); |
| 4723 | case Builtin::BI__exception_info: |
| 4724 | case Builtin::BI_exception_info: |
| 4725 | return RValue::get(EmitSEHExceptionInfo()); |
| 4726 | case Builtin::BI__abnormal_termination: |
| 4727 | case Builtin::BI_abnormal_termination: |
| 4728 | return RValue::get(EmitSEHAbnormalTermination()); |
| 4729 | case Builtin::BI_setjmpex: |
| 4730 | if (getTarget().getTriple().isOSMSVCRT() && E->getNumArgs() == 1 && |
| 4731 | E->getArg(0)->getType()->isPointerType()) |
| 4732 | return EmitMSVCRTSetJmp(*this, MSVCSetJmpKind::_setjmpex, E); |
| 4733 | break; |
| 4734 | case Builtin::BI_setjmp: |
| 4735 | if (getTarget().getTriple().isOSMSVCRT() && E->getNumArgs() == 1 && |
| 4736 | E->getArg(0)->getType()->isPointerType()) { |
| 4737 | if (getTarget().getTriple().getArch() == llvm::Triple::x86) |
| 4738 | return EmitMSVCRTSetJmp(*this, MSVCSetJmpKind::_setjmp3, E); |
| 4739 | else if (getTarget().getTriple().getArch() == llvm::Triple::aarch64) |
| 4740 | return EmitMSVCRTSetJmp(*this, MSVCSetJmpKind::_setjmpex, E); |
| 4741 | return EmitMSVCRTSetJmp(*this, MSVCSetJmpKind::_setjmp, E); |
| 4742 | } |
| 4743 | break; |
| 4744 | |
| 4745 | // C++ std:: builtins. |
| 4746 | case Builtin::BImove: |
| 4747 | case Builtin::BImove_if_noexcept: |
| 4748 | case Builtin::BIforward: |
| 4749 | case Builtin::BIforward_like: |
| 4750 | case Builtin::BIas_const: |
| 4751 | return RValue::get(EmitLValue(E->getArg(0)).getPointer(*this)); |
| 4752 | case Builtin::BI__GetExceptionInfo: { |
| 4753 | if (llvm::GlobalVariable *GV = |
| 4754 | CGM.getCXXABI().getThrowInfo(FD->getParamDecl(0)->getType())) |
| 4755 | return RValue::get(llvm::ConstantExpr::getBitCast(GV, CGM.Int8PtrTy)); |
| 4756 | break; |
| 4757 | } |
| 4758 | |
| 4759 | case Builtin::BI__fastfail: |
| 4760 | return RValue::get(EmitMSVCBuiltinExpr(MSVCIntrin::__fastfail, E)); |
| 4761 | |
| 4762 | case Builtin::BI__builtin_coro_id: |
| 4763 | return EmitCoroutineIntrinsic(E, Intrinsic::coro_id); |
| 4764 | case Builtin::BI__builtin_coro_promise: |
| 4765 | return EmitCoroutineIntrinsic(E, Intrinsic::coro_promise); |
| 4766 | case Builtin::BI__builtin_coro_resume: |
| 4767 | EmitCoroutineIntrinsic(E, Intrinsic::coro_resume); |
| 4768 | return RValue::get(nullptr); |
| 4769 | case Builtin::BI__builtin_coro_frame: |
| 4770 | return EmitCoroutineIntrinsic(E, Intrinsic::coro_frame); |
| 4771 | case Builtin::BI__builtin_coro_noop: |
| 4772 | return EmitCoroutineIntrinsic(E, Intrinsic::coro_noop); |
| 4773 | case Builtin::BI__builtin_coro_free: |
| 4774 | return EmitCoroutineIntrinsic(E, Intrinsic::coro_free); |
| 4775 | case Builtin::BI__builtin_coro_destroy: |
| 4776 | EmitCoroutineIntrinsic(E, Intrinsic::coro_destroy); |
| 4777 | return RValue::get(nullptr); |
| 4778 | case Builtin::BI__builtin_coro_done: |
| 4779 | return EmitCoroutineIntrinsic(E, Intrinsic::coro_done); |
| 4780 | case Builtin::BI__builtin_coro_alloc: |
| 4781 | return EmitCoroutineIntrinsic(E, Intrinsic::coro_alloc); |
| 4782 | case Builtin::BI__builtin_coro_begin: |
| 4783 | return EmitCoroutineIntrinsic(E, Intrinsic::coro_begin); |
| 4784 | case Builtin::BI__builtin_coro_end: |
| 4785 | return EmitCoroutineIntrinsic(E, Intrinsic::coro_end); |
| 4786 | case Builtin::BI__builtin_coro_suspend: |
| 4787 | return EmitCoroutineIntrinsic(E, Intrinsic::coro_suspend); |
| 4788 | case Builtin::BI__builtin_coro_size: |
| 4789 | return EmitCoroutineIntrinsic(E, Intrinsic::coro_size); |
| 4790 | case Builtin::BI__builtin_coro_align: |
| 4791 | return EmitCoroutineIntrinsic(E, Intrinsic::coro_align); |
| 4792 | |
| 4793 | // OpenCL v2.0 s6.13.16.2, Built-in pipe read and write functions |
| 4794 | case Builtin::BIread_pipe: |
| 4795 | case Builtin::BIwrite_pipe: { |
| 4796 | Value *Arg0 = EmitScalarExpr(E->getArg(0)), |
| 4797 | *Arg1 = EmitScalarExpr(E->getArg(1)); |
| 4798 | CGOpenCLRuntime OpenCLRT(CGM); |
| 4799 | Value *PacketSize = OpenCLRT.getPipeElemSize(E->getArg(0)); |
| 4800 | Value *PacketAlign = OpenCLRT.getPipeElemAlign(E->getArg(0)); |
| 4801 | |
| 4802 | // Type of the generic packet parameter. |
| 4803 | unsigned GenericAS = |
| 4804 | getContext().getTargetAddressSpace(LangAS::opencl_generic); |
| 4805 | llvm::Type *I8PTy = llvm::PointerType::get( |
| 4806 | llvm::Type::getInt8Ty(getLLVMContext()), GenericAS); |
| 4807 | |
| 4808 | // Testing which overloaded version we should generate the call for. |
| 4809 | if (2U == E->getNumArgs()) { |
| 4810 | const char *Name = (BuiltinID == Builtin::BIread_pipe) ? "__read_pipe_2" |
| 4811 | : "__write_pipe_2"; |
| 4812 | // Creating a generic function type to be able to call with any builtin or |
| 4813 | // user defined type. |
| 4814 | llvm::Type *ArgTys[] = {Arg0->getType(), I8PTy, Int32Ty, Int32Ty}; |
| 4815 | llvm::FunctionType *FTy = llvm::FunctionType::get( |
| 4816 | Int32Ty, llvm::ArrayRef<llvm::Type *>(ArgTys), false); |
| 4817 | Value *BCast = Builder.CreatePointerCast(Arg1, I8PTy); |
| 4818 | return RValue::get( |
| 4819 | EmitRuntimeCall(CGM.CreateRuntimeFunction(FTy, Name), |
| 4820 | {Arg0, BCast, PacketSize, PacketAlign})); |
| 4821 | } else { |
| 4822 | assert(4 == E->getNumArgs() &&(static_cast <bool> (4 == E->getNumArgs() && "Illegal number of parameters to pipe function") ? void (0) : __assert_fail ("4 == E->getNumArgs() && \"Illegal number of parameters to pipe function\"" , "clang/lib/CodeGen/CGBuiltin.cpp", 4823, __extension__ __PRETTY_FUNCTION__ )) |
| 4823 | "Illegal number of parameters to pipe function")(static_cast <bool> (4 == E->getNumArgs() && "Illegal number of parameters to pipe function") ? void (0) : __assert_fail ("4 == E->getNumArgs() && \"Illegal number of parameters to pipe function\"" , "clang/lib/CodeGen/CGBuiltin.cpp", 4823, __extension__ __PRETTY_FUNCTION__ )); |
| 4824 | const char *Name = (BuiltinID == Builtin::BIread_pipe) ? "__read_pipe_4" |
| 4825 | : "__write_pipe_4"; |
| 4826 | |
| 4827 | llvm::Type *ArgTys[] = {Arg0->getType(), Arg1->getType(), Int32Ty, I8PTy, |
| 4828 | Int32Ty, Int32Ty}; |
| 4829 | Value *Arg2 = EmitScalarExpr(E->getArg(2)), |
| 4830 | *Arg3 = EmitScalarExpr(E->getArg(3)); |
| 4831 | llvm::FunctionType *FTy = llvm::FunctionType::get( |
| 4832 | Int32Ty, llvm::ArrayRef<llvm::Type *>(ArgTys), false); |
| 4833 | Value *BCast = Builder.CreatePointerCast(Arg3, I8PTy); |
| 4834 | // We know the third argument is an integer type, but we may need to cast |
| 4835 | // it to i32. |
| 4836 | if (Arg2->getType() != Int32Ty) |
| 4837 | Arg2 = Builder.CreateZExtOrTrunc(Arg2, Int32Ty); |
| 4838 | return RValue::get( |
| 4839 | EmitRuntimeCall(CGM.CreateRuntimeFunction(FTy, Name), |
| 4840 | {Arg0, Arg1, Arg2, BCast, PacketSize, PacketAlign})); |
| 4841 | } |
| 4842 | } |
| 4843 | // OpenCL v2.0 s6.13.16 ,s9.17.3.5 - Built-in pipe reserve read and write |
| 4844 | // functions |
| 4845 | case Builtin::BIreserve_read_pipe: |
| 4846 | case Builtin::BIreserve_write_pipe: |
| 4847 | case Builtin::BIwork_group_reserve_read_pipe: |
| 4848 | case Builtin::BIwork_group_reserve_write_pipe: |
| 4849 | case Builtin::BIsub_group_reserve_read_pipe: |
| 4850 | case Builtin::BIsub_group_reserve_write_pipe: { |
| 4851 | // Composing the mangled name for the function. |
| 4852 | const char *Name; |
| 4853 | if (BuiltinID == Builtin::BIreserve_read_pipe) |
| 4854 | Name = "__reserve_read_pipe"; |
| 4855 | else if (BuiltinID == Builtin::BIreserve_write_pipe) |
| 4856 | Name = "__reserve_write_pipe"; |
| 4857 | else if (BuiltinID == Builtin::BIwork_group_reserve_read_pipe) |
| 4858 | Name = "__work_group_reserve_read_pipe"; |
| 4859 | else if (BuiltinID == Builtin::BIwork_group_reserve_write_pipe) |
| 4860 | Name = "__work_group_reserve_write_pipe"; |
| 4861 | else if (BuiltinID == Builtin::BIsub_group_reserve_read_pipe) |
| 4862 | Name = "__sub_group_reserve_read_pipe"; |
| 4863 | else |
| 4864 | Name = "__sub_group_reserve_write_pipe"; |
| 4865 | |
| 4866 | Value *Arg0 = EmitScalarExpr(E->getArg(0)), |
| 4867 | *Arg1 = EmitScalarExpr(E->getArg(1)); |
| 4868 | llvm::Type *ReservedIDTy = ConvertType(getContext().OCLReserveIDTy); |
| 4869 | CGOpenCLRuntime OpenCLRT(CGM); |
| 4870 | Value *PacketSize = OpenCLRT.getPipeElemSize(E->getArg(0)); |
| 4871 | Value *PacketAlign = OpenCLRT.getPipeElemAlign(E->getArg(0)); |
| 4872 | |
| 4873 | // Building the generic function prototype. |
| 4874 | llvm::Type *ArgTys[] = {Arg0->getType(), Int32Ty, Int32Ty, Int32Ty}; |
| 4875 | llvm::FunctionType *FTy = llvm::FunctionType::get( |
| 4876 | ReservedIDTy, llvm::ArrayRef<llvm::Type *>(ArgTys), false); |
| 4877 | // We know the second argument is an integer type, but we may need to cast |
| 4878 | // it to i32. |
| 4879 | if (Arg1->getType() != Int32Ty) |
| 4880 | Arg1 = Builder.CreateZExtOrTrunc(Arg1, Int32Ty); |
| 4881 | return RValue::get(EmitRuntimeCall(CGM.CreateRuntimeFunction(FTy, Name), |
| 4882 | {Arg0, Arg1, PacketSize, PacketAlign})); |
| 4883 | } |
| 4884 | // OpenCL v2.0 s6.13.16, s9.17.3.5 - Built-in pipe commit read and write |
| 4885 | // functions |
| 4886 | case Builtin::BIcommit_read_pipe: |
| 4887 | case Builtin::BIcommit_write_pipe: |
| 4888 | case Builtin::BIwork_group_commit_read_pipe: |
| 4889 | case Builtin::BIwork_group_commit_write_pipe: |
| 4890 | case Builtin::BIsub_group_commit_read_pipe: |
| 4891 | case Builtin::BIsub_group_commit_write_pipe: { |
| 4892 | const char *Name; |
| 4893 | if (BuiltinID == Builtin::BIcommit_read_pipe) |
| 4894 | Name = "__commit_read_pipe"; |
| 4895 | else if (BuiltinID == Builtin::BIcommit_write_pipe) |
| 4896 | Name = "__commit_write_pipe"; |
| 4897 | else if (BuiltinID == Builtin::BIwork_group_commit_read_pipe) |
| 4898 | Name = "__work_group_commit_read_pipe"; |
| 4899 | else if (BuiltinID == Builtin::BIwork_group_commit_write_pipe) |
| 4900 | Name = "__work_group_commit_write_pipe"; |
| 4901 | else if (BuiltinID == Builtin::BIsub_group_commit_read_pipe) |
| 4902 | Name = "__sub_group_commit_read_pipe"; |
| 4903 | else |
| 4904 | Name = "__sub_group_commit_write_pipe"; |
| 4905 | |
| 4906 | Value *Arg0 = EmitScalarExpr(E->getArg(0)), |
| 4907 | *Arg1 = EmitScalarExpr(E->getArg(1)); |
| 4908 | CGOpenCLRuntime OpenCLRT(CGM); |
| 4909 | Value *PacketSize = OpenCLRT.getPipeElemSize(E->getArg(0)); |
| 4910 | Value *PacketAlign = OpenCLRT.getPipeElemAlign(E->getArg(0)); |
| 4911 | |
| 4912 | // Building the generic function prototype. |
| 4913 | llvm::Type *ArgTys[] = {Arg0->getType(), Arg1->getType(), Int32Ty, Int32Ty}; |
| 4914 | llvm::FunctionType *FTy = |
| 4915 | llvm::FunctionType::get(llvm::Type::getVoidTy(getLLVMContext()), |
| 4916 | llvm::ArrayRef<llvm::Type *>(ArgTys), false); |
| 4917 | |
| 4918 | return RValue::get(EmitRuntimeCall(CGM.CreateRuntimeFunction(FTy, Name), |
| 4919 | {Arg0, Arg1, PacketSize, PacketAlign})); |
| 4920 | } |
| 4921 | // OpenCL v2.0 s6.13.16.4 Built-in pipe query functions |
| 4922 | case Builtin::BIget_pipe_num_packets: |
| 4923 | case Builtin::BIget_pipe_max_packets: { |
| 4924 | const char *BaseName; |
| 4925 | const auto *PipeTy = E->getArg(0)->getType()->castAs<PipeType>(); |
| 4926 | if (BuiltinID == Builtin::BIget_pipe_num_packets) |
| 4927 | BaseName = "__get_pipe_num_packets"; |
| 4928 | else |
| 4929 | BaseName = "__get_pipe_max_packets"; |
| 4930 | std::string Name = std::string(BaseName) + |
| 4931 | std::string(PipeTy->isReadOnly() ? "_ro" : "_wo"); |
| 4932 | |
| 4933 | // Building the generic function prototype. |
| 4934 | Value *Arg0 = EmitScalarExpr(E->getArg(0)); |
| 4935 | CGOpenCLRuntime OpenCLRT(CGM); |
| 4936 | Value *PacketSize = OpenCLRT.getPipeElemSize(E->getArg(0)); |
| 4937 | Value *PacketAlign = OpenCLRT.getPipeElemAlign(E->getArg(0)); |
| 4938 | llvm::Type *ArgTys[] = {Arg0->getType(), Int32Ty, Int32Ty}; |
| 4939 | llvm::FunctionType *FTy = llvm::FunctionType::get( |
| 4940 | Int32Ty, llvm::ArrayRef<llvm::Type *>(ArgTys), false); |
| 4941 | |
| 4942 | return RValue::get(EmitRuntimeCall(CGM.CreateRuntimeFunction(FTy, Name), |
| 4943 | {Arg0, PacketSize, PacketAlign})); |
| 4944 | } |
| 4945 | |
| 4946 | // OpenCL v2.0 s6.13.9 - Address space qualifier functions. |
| 4947 | case Builtin::BIto_global: |
| 4948 | case Builtin::BIto_local: |
| 4949 | case Builtin::BIto_private: { |
| 4950 | auto Arg0 = EmitScalarExpr(E->getArg(0)); |
| 4951 | auto NewArgT = llvm::PointerType::get(Int8Ty, |
| 4952 | CGM.getContext().getTargetAddressSpace(LangAS::opencl_generic)); |
| 4953 | auto NewRetT = llvm::PointerType::get(Int8Ty, |
| 4954 | CGM.getContext().getTargetAddressSpace( |
| 4955 | E->getType()->getPointeeType().getAddressSpace())); |
| 4956 | auto FTy = llvm::FunctionType::get(NewRetT, {NewArgT}, false); |
| 4957 | llvm::Value *NewArg; |
| 4958 | if (Arg0->getType()->getPointerAddressSpace() != |
| 4959 | NewArgT->getPointerAddressSpace()) |
| 4960 | NewArg = Builder.CreateAddrSpaceCast(Arg0, NewArgT); |
| 4961 | else |
| 4962 | NewArg = Builder.CreateBitOrPointerCast(Arg0, NewArgT); |
| 4963 | auto NewName = std::string("__") + E->getDirectCallee()->getName().str(); |
| 4964 | auto NewCall = |
| 4965 | EmitRuntimeCall(CGM.CreateRuntimeFunction(FTy, NewName), {NewArg}); |
| 4966 | return RValue::get(Builder.CreateBitOrPointerCast(NewCall, |
| 4967 | ConvertType(E->getType()))); |
| 4968 | } |
| 4969 | |
| 4970 | // OpenCL v2.0, s6.13.17 - Enqueue kernel function. |
| 4971 | // It contains four different overload formats specified in Table 6.13.17.1. |
| 4972 | case Builtin::BIenqueue_kernel: { |
| 4973 | StringRef Name; // Generated function call name |
| 4974 | unsigned NumArgs = E->getNumArgs(); |
| 4975 | |
| 4976 | llvm::Type *QueueTy = ConvertType(getContext().OCLQueueTy); |
| 4977 | llvm::Type *GenericVoidPtrTy = Builder.getInt8PtrTy( |
| 4978 | getContext().getTargetAddressSpace(LangAS::opencl_generic)); |
| 4979 | |
| 4980 | llvm::Value *Queue = EmitScalarExpr(E->getArg(0)); |
| 4981 | llvm::Value *Flags = EmitScalarExpr(E->getArg(1)); |
| 4982 | LValue NDRangeL = EmitAggExprToLValue(E->getArg(2)); |
| 4983 | llvm::Value *Range = NDRangeL.getAddress(*this).getPointer(); |
| 4984 | llvm::Type *RangeTy = NDRangeL.getAddress(*this).getType(); |
| 4985 | |
| 4986 | if (NumArgs == 4) { |
| 4987 | // The most basic form of the call with parameters: |
| 4988 | // queue_t, kernel_enqueue_flags_t, ndrange_t, block(void) |
| 4989 | Name = "__enqueue_kernel_basic"; |
| 4990 | llvm::Type *ArgTys[] = {QueueTy, Int32Ty, RangeTy, GenericVoidPtrTy, |
| 4991 | GenericVoidPtrTy}; |
| 4992 | llvm::FunctionType *FTy = llvm::FunctionType::get( |
| 4993 | Int32Ty, llvm::ArrayRef<llvm::Type *>(ArgTys), false); |
| 4994 | |
| 4995 | auto Info = |
| 4996 | CGM.getOpenCLRuntime().emitOpenCLEnqueuedBlock(*this, E->getArg(3)); |
| 4997 | llvm::Value *Kernel = |
| 4998 | Builder.CreatePointerCast(Info.KernelHandle, GenericVoidPtrTy); |
| 4999 | llvm::Value *Block = |
| 5000 | Builder.CreatePointerCast(Info.BlockArg, GenericVoidPtrTy); |
| 5001 | |
| 5002 | AttrBuilder B(Builder.getContext()); |
| 5003 | B.addByValAttr(NDRangeL.getAddress(*this).getElementType()); |
| 5004 | llvm::AttributeList ByValAttrSet = |
| 5005 | llvm::AttributeList::get(CGM.getModule().getContext(), 3U, B); |
| 5006 | |
| 5007 | auto RTCall = |
| 5008 | EmitRuntimeCall(CGM.CreateRuntimeFunction(FTy, Name, ByValAttrSet), |
| 5009 | {Queue, Flags, Range, Kernel, Block}); |
| 5010 | RTCall->setAttributes(ByValAttrSet); |
| 5011 | return RValue::get(RTCall); |
| 5012 | } |
| 5013 | assert(NumArgs >= 5 && "Invalid enqueue_kernel signature")(static_cast <bool> (NumArgs >= 5 && "Invalid enqueue_kernel signature" ) ? void (0) : __assert_fail ("NumArgs >= 5 && \"Invalid enqueue_kernel signature\"" , "clang/lib/CodeGen/CGBuiltin.cpp", 5013, __extension__ __PRETTY_FUNCTION__ )); |
| 5014 | |
| 5015 | // Create a temporary array to hold the sizes of local pointer arguments |
| 5016 | // for the block. \p First is the position of the first size argument. |
| 5017 | auto CreateArrayForSizeVar = [=](unsigned First) |
| 5018 | -> std::tuple<llvm::Value *, llvm::Value *, llvm::Value *> { |
| 5019 | llvm::APInt ArraySize(32, NumArgs - First); |
| 5020 | QualType SizeArrayTy = getContext().getConstantArrayType( |
| 5021 | getContext().getSizeType(), ArraySize, nullptr, ArrayType::Normal, |
| 5022 | /*IndexTypeQuals=*/0); |
| 5023 | auto Tmp = CreateMemTemp(SizeArrayTy, "block_sizes"); |
| 5024 | llvm::Value *TmpPtr = Tmp.getPointer(); |
| 5025 | llvm::Value *TmpSize = EmitLifetimeStart( |
| 5026 | CGM.getDataLayout().getTypeAllocSize(Tmp.getElementType()), TmpPtr); |
| 5027 | llvm::Value *ElemPtr; |
| 5028 | // Each of the following arguments specifies the size of the corresponding |
| 5029 | // argument passed to the enqueued block. |
| 5030 | auto *Zero = llvm::ConstantInt::get(IntTy, 0); |
| 5031 | for (unsigned I = First; I < NumArgs; ++I) { |
| 5032 | auto *Index = llvm::ConstantInt::get(IntTy, I - First); |
| 5033 | auto *GEP = Builder.CreateGEP(Tmp.getElementType(), TmpPtr, |
| 5034 | {Zero, Index}); |
| 5035 | if (I == First) |
| 5036 | ElemPtr = GEP; |
| 5037 | auto *V = |
| 5038 | Builder.CreateZExtOrTrunc(EmitScalarExpr(E->getArg(I)), SizeTy); |
| 5039 | Builder.CreateAlignedStore( |
| 5040 | V, GEP, CGM.getDataLayout().getPrefTypeAlign(SizeTy)); |
| 5041 | } |
| 5042 | return std::tie(ElemPtr, TmpSize, TmpPtr); |
| 5043 | }; |
| 5044 | |
| 5045 | // Could have events and/or varargs. |
| 5046 | if (E->getArg(3)->getType()->isBlockPointerType()) { |
| 5047 | // No events passed, but has variadic arguments. |
| 5048 | Name = "__enqueue_kernel_varargs"; |
| 5049 | auto Info = |
| 5050 | CGM.getOpenCLRuntime().emitOpenCLEnqueuedBlock(*this, E->getArg(3)); |
| 5051 | llvm::Value *Kernel = |
| 5052 | Builder.CreatePointerCast(Info.KernelHandle, GenericVoidPtrTy); |
| 5053 | auto *Block = Builder.CreatePointerCast(Info.BlockArg, GenericVoidPtrTy); |
| 5054 | llvm::Value *ElemPtr, *TmpSize, *TmpPtr; |
| 5055 | std::tie(ElemPtr, TmpSize, TmpPtr) = CreateArrayForSizeVar(4); |
| 5056 | |
| 5057 | // Create a vector of the arguments, as well as a constant value to |
| 5058 | // express to the runtime the number of variadic arguments. |
| 5059 | llvm::Value *const Args[] = {Queue, Flags, |
| 5060 | Range, Kernel, |
| 5061 | Block, ConstantInt::get(IntTy, NumArgs - 4), |
| 5062 | ElemPtr}; |
| 5063 | llvm::Type *const ArgTys[] = { |
| 5064 | QueueTy, IntTy, RangeTy, GenericVoidPtrTy, |
| 5065 | GenericVoidPtrTy, IntTy, ElemPtr->getType()}; |
| 5066 | |
| 5067 | llvm::FunctionType *FTy = llvm::FunctionType::get(Int32Ty, ArgTys, false); |
| 5068 | auto Call = RValue::get( |
| 5069 | EmitRuntimeCall(CGM.CreateRuntimeFunction(FTy, Name), Args)); |
| 5070 | if (TmpSize) |
| 5071 | EmitLifetimeEnd(TmpSize, TmpPtr); |
| 5072 | return Call; |
| 5073 | } |
| 5074 | // Any calls now have event arguments passed. |
| 5075 | if (NumArgs >= 7) { |
| 5076 | llvm::Type *EventTy = ConvertType(getContext().OCLClkEventTy); |
| 5077 | llvm::PointerType *EventPtrTy = EventTy->getPointerTo( |
| 5078 | CGM.getContext().getTargetAddressSpace(LangAS::opencl_generic)); |
| 5079 | |
| 5080 | llvm::Value *NumEvents = |
| 5081 | Builder.CreateZExtOrTrunc(EmitScalarExpr(E->getArg(3)), Int32Ty); |
| 5082 | |
| 5083 | // Since SemaOpenCLBuiltinEnqueueKernel allows fifth and sixth arguments |
| 5084 | // to be a null pointer constant (including `0` literal), we can take it |
| 5085 | // into account and emit null pointer directly. |
| 5086 | llvm::Value *EventWaitList = nullptr; |
| 5087 | if (E->getArg(4)->isNullPointerConstant( |
| 5088 | getContext(), Expr::NPC_ValueDependentIsNotNull)) { |
| 5089 | EventWaitList = llvm::ConstantPointerNull::get(EventPtrTy); |
| 5090 | } else { |
| 5091 | EventWaitList = E->getArg(4)->getType()->isArrayType() |
| 5092 | ? EmitArrayToPointerDecay(E->getArg(4)).getPointer() |
| 5093 | : EmitScalarExpr(E->getArg(4)); |
| 5094 | // Convert to generic address space. |
| 5095 | EventWaitList = Builder.CreatePointerCast(EventWaitList, EventPtrTy); |
| 5096 | } |
| 5097 | llvm::Value *EventRet = nullptr; |
| 5098 | if (E->getArg(5)->isNullPointerConstant( |
| 5099 | getContext(), Expr::NPC_ValueDependentIsNotNull)) { |
| 5100 | EventRet = llvm::ConstantPointerNull::get(EventPtrTy); |
| 5101 | } else { |
| 5102 | EventRet = |
| 5103 | Builder.CreatePointerCast(EmitScalarExpr(E->getArg(5)), EventPtrTy); |
| 5104 | } |
| 5105 | |
| 5106 | auto Info = |
| 5107 | CGM.getOpenCLRuntime().emitOpenCLEnqueuedBlock(*this, E->getArg(6)); |
| 5108 | llvm::Value *Kernel = |
| 5109 | Builder.CreatePointerCast(Info.KernelHandle, GenericVoidPtrTy); |
| 5110 | llvm::Value *Block = |
| 5111 | Builder.CreatePointerCast(Info.BlockArg, GenericVoidPtrTy); |
| 5112 | |
| 5113 | std::vector<llvm::Type *> ArgTys = { |
| 5114 | QueueTy, Int32Ty, RangeTy, Int32Ty, |
| 5115 | EventPtrTy, EventPtrTy, GenericVoidPtrTy, GenericVoidPtrTy}; |
| 5116 | |
| 5117 | std::vector<llvm::Value *> Args = {Queue, Flags, Range, |
| 5118 | NumEvents, EventWaitList, EventRet, |
| 5119 | Kernel, Block}; |
| 5120 | |
| 5121 | if (NumArgs == 7) { |
| 5122 | // Has events but no variadics. |
| 5123 | Name = "__enqueue_kernel_basic_events"; |
| 5124 | llvm::FunctionType *FTy = llvm::FunctionType::get( |
| 5125 | Int32Ty, llvm::ArrayRef<llvm::Type *>(ArgTys), false); |
| 5126 | return RValue::get( |
| 5127 | EmitRuntimeCall(CGM.CreateRuntimeFunction(FTy, Name), |
| 5128 | llvm::ArrayRef<llvm::Value *>(Args))); |
| 5129 | } |
| 5130 | // Has event info and variadics |
| 5131 | // Pass the number of variadics to the runtime function too. |
| 5132 | Args.push_back(ConstantInt::get(Int32Ty, NumArgs - 7)); |
| 5133 | ArgTys.push_back(Int32Ty); |
| 5134 | Name = "__enqueue_kernel_events_varargs"; |
| 5135 | |
| 5136 | llvm::Value *ElemPtr, *TmpSize, *TmpPtr; |
| 5137 | std::tie(ElemPtr, TmpSize, TmpPtr) = CreateArrayForSizeVar(7); |
| 5138 | Args.push_back(ElemPtr); |
| 5139 | ArgTys.push_back(ElemPtr->getType()); |
| 5140 | |
| 5141 | llvm::FunctionType *FTy = llvm::FunctionType::get( |
| 5142 | Int32Ty, llvm::ArrayRef<llvm::Type *>(ArgTys), false); |
| 5143 | auto Call = |
| 5144 | RValue::get(EmitRuntimeCall(CGM.CreateRuntimeFunction(FTy, Name), |
| 5145 | llvm::ArrayRef<llvm::Value *>(Args))); |
| 5146 | if (TmpSize) |
| 5147 | EmitLifetimeEnd(TmpSize, TmpPtr); |
| 5148 | return Call; |
| 5149 | } |
| 5150 | [[fallthrough]]; |
| 5151 | } |
| 5152 | // OpenCL v2.0 s6.13.17.6 - Kernel query functions need bitcast of block |
| 5153 | // parameter. |
| 5154 | case Builtin::BIget_kernel_work_group_size: { |
| 5155 | llvm::Type *GenericVoidPtrTy = Builder.getInt8PtrTy( |
| 5156 | getContext().getTargetAddressSpace(LangAS::opencl_generic)); |
| 5157 | auto Info = |
| 5158 | CGM.getOpenCLRuntime().emitOpenCLEnqueuedBlock(*this, E->getArg(0)); |
| 5159 | Value *Kernel = |
| 5160 | Builder.CreatePointerCast(Info.KernelHandle, GenericVoidPtrTy); |
| 5161 | Value *Arg = Builder.CreatePointerCast(Info.BlockArg, GenericVoidPtrTy); |
| 5162 | return RValue::get(EmitRuntimeCall( |
| 5163 | CGM.CreateRuntimeFunction( |
| 5164 | llvm::FunctionType::get(IntTy, {GenericVoidPtrTy, GenericVoidPtrTy}, |
| 5165 | false), |
| 5166 | "__get_kernel_work_group_size_impl"), |
| 5167 | {Kernel, Arg})); |
| 5168 | } |
| 5169 | case Builtin::BIget_kernel_preferred_work_group_size_multiple: { |
| 5170 | llvm::Type *GenericVoidPtrTy = Builder.getInt8PtrTy( |
| 5171 | getContext().getTargetAddressSpace(LangAS::opencl_generic)); |
| 5172 | auto Info = |
| 5173 | CGM.getOpenCLRuntime().emitOpenCLEnqueuedBlock(*this, E->getArg(0)); |
| 5174 | Value *Kernel = |
| 5175 | Builder.CreatePointerCast(Info.KernelHandle, GenericVoidPtrTy); |
| 5176 | Value *Arg = Builder.CreatePointerCast(Info.BlockArg, GenericVoidPtrTy); |
| 5177 | return RValue::get(EmitRuntimeCall( |
| 5178 | CGM.CreateRuntimeFunction( |
| 5179 | llvm::FunctionType::get(IntTy, {GenericVoidPtrTy, GenericVoidPtrTy}, |
| 5180 | false), |
| 5181 | "__get_kernel_preferred_work_group_size_multiple_impl"), |
| 5182 | {Kernel, Arg})); |
| 5183 | } |
| 5184 | case Builtin::BIget_kernel_max_sub_group_size_for_ndrange: |
| 5185 | case Builtin::BIget_kernel_sub_group_count_for_ndrange: { |
| 5186 | llvm::Type *GenericVoidPtrTy = Builder.getInt8PtrTy( |
| 5187 | getContext().getTargetAddressSpace(LangAS::opencl_generic)); |
| 5188 | LValue NDRangeL = EmitAggExprToLValue(E->getArg(0)); |
| 5189 | llvm::Value *NDRange = NDRangeL.getAddress(*this).getPointer(); |
| 5190 | auto Info = |
| 5191 | CGM.getOpenCLRuntime().emitOpenCLEnqueuedBlock(*this, E->getArg(1)); |
| 5192 | Value *Kernel = |
| 5193 | Builder.CreatePointerCast(Info.KernelHandle, GenericVoidPtrTy); |
| 5194 | Value *Block = Builder.CreatePointerCast(Info.BlockArg, GenericVoidPtrTy); |
| 5195 | const char *Name = |
| 5196 | BuiltinID == Builtin::BIget_kernel_max_sub_group_size_for_ndrange |
| 5197 | ? "__get_kernel_max_sub_group_size_for_ndrange_impl" |
| 5198 | : "__get_kernel_sub_group_count_for_ndrange_impl"; |
| 5199 | return RValue::get(EmitRuntimeCall( |
| 5200 | CGM.CreateRuntimeFunction( |
| 5201 | llvm::FunctionType::get( |
| 5202 | IntTy, {NDRange->getType(), GenericVoidPtrTy, GenericVoidPtrTy}, |
| 5203 | false), |
| 5204 | Name), |
| 5205 | {NDRange, Kernel, Block})); |
| 5206 | } |
| 5207 | |
| 5208 | case Builtin::BI__builtin_store_half: |
| 5209 | case Builtin::BI__builtin_store_halff: { |
| 5210 | Value *Val = EmitScalarExpr(E->getArg(0)); |
| 5211 | Address Address = EmitPointerWithAlignment(E->getArg(1)); |
| 5212 | Value *HalfVal = Builder.CreateFPTrunc(Val, Builder.getHalfTy()); |
| 5213 | Builder.CreateStore(HalfVal, Address); |
| 5214 | return RValue::get(nullptr); |
| 5215 | } |
| 5216 | case Builtin::BI__builtin_load_half: { |
| 5217 | Address Address = EmitPointerWithAlignment(E->getArg(0)); |
| 5218 | Value *HalfVal = Builder.CreateLoad(Address); |
| 5219 | return RValue::get(Builder.CreateFPExt(HalfVal, Builder.getDoubleTy())); |
| 5220 | } |
| 5221 | case Builtin::BI__builtin_load_halff: { |
| 5222 | Address Address = EmitPointerWithAlignment(E->getArg(0)); |
| 5223 | Value *HalfVal = Builder.CreateLoad(Address); |
| 5224 | return RValue::get(Builder.CreateFPExt(HalfVal, Builder.getFloatTy())); |
| 5225 | } |
| 5226 | case Builtin::BIprintf: |
| 5227 | if (getTarget().getTriple().isNVPTX() || |
| 5228 | getTarget().getTriple().isAMDGCN()) { |
| 5229 | if (getLangOpts().OpenMPIsDevice) |
| 5230 | return EmitOpenMPDevicePrintfCallExpr(E); |
| 5231 | if (getTarget().getTriple().isNVPTX()) |
| 5232 | return EmitNVPTXDevicePrintfCallExpr(E); |
| 5233 | if (getTarget().getTriple().isAMDGCN() && getLangOpts().HIP) |
| 5234 | return EmitAMDGPUDevicePrintfCallExpr(E); |
| 5235 | } |
| 5236 | |
| 5237 | break; |
| 5238 | case Builtin::BI__builtin_canonicalize: |
| 5239 | case Builtin::BI__builtin_canonicalizef: |
| 5240 | case Builtin::BI__builtin_canonicalizef16: |
| 5241 | case Builtin::BI__builtin_canonicalizel: |
| 5242 | return RValue::get(emitUnaryBuiltin(*this, E, Intrinsic::canonicalize)); |
| 5243 | |
| 5244 | case Builtin::BI__builtin_thread_pointer: { |
| 5245 | if (!getContext().getTargetInfo().isTLSSupported()) |
| 5246 | CGM.ErrorUnsupported(E, "__builtin_thread_pointer"); |
| 5247 | // Fall through - it's already mapped to the intrinsic by ClangBuiltin. |
| 5248 | break; |
| 5249 | } |
| 5250 | case Builtin::BI__builtin_os_log_format: |
| 5251 | return emitBuiltinOSLogFormat(*E); |
| 5252 | |
| 5253 | case Builtin::BI__xray_customevent: { |
| 5254 | if (!ShouldXRayInstrumentFunction()) |
| 5255 | return RValue::getIgnored(); |
| 5256 | |
| 5257 | if (!CGM.getCodeGenOpts().XRayInstrumentationBundle.has( |
| 5258 | XRayInstrKind::Custom)) |
| 5259 | return RValue::getIgnored(); |
| 5260 | |
| 5261 | if (const auto *XRayAttr = CurFuncDecl->getAttr<XRayInstrumentAttr>()) |
| 5262 | if (XRayAttr->neverXRayInstrument() && !AlwaysEmitXRayCustomEvents()) |
| 5263 | return RValue::getIgnored(); |
| 5264 | |
| 5265 | Function *F = CGM.getIntrinsic(Intrinsic::xray_customevent); |
| 5266 | auto FTy = F->getFunctionType(); |
| 5267 | auto Arg0 = E->getArg(0); |
| 5268 | auto Arg0Val = EmitScalarExpr(Arg0); |
| 5269 | auto Arg0Ty = Arg0->getType(); |
| 5270 | auto PTy0 = FTy->getParamType(0); |
| 5271 | if (PTy0 != Arg0Val->getType()) { |
| 5272 | if (Arg0Ty->isArrayType()) |
| 5273 | Arg0Val = EmitArrayToPointerDecay(Arg0).getPointer(); |
| 5274 | else |
| 5275 | Arg0Val = Builder.CreatePointerCast(Arg0Val, PTy0); |
| 5276 | } |
| 5277 | auto Arg1 = EmitScalarExpr(E->getArg(1)); |
| 5278 | auto PTy1 = FTy->getParamType(1); |
| 5279 | if (PTy1 != Arg1->getType()) |
| 5280 | Arg1 = Builder.CreateTruncOrBitCast(Arg1, PTy1); |
| 5281 | return RValue::get(Builder.CreateCall(F, {Arg0Val, Arg1})); |
| 5282 | } |
| 5283 | |
| 5284 | case Builtin::BI__xray_typedevent: { |
| 5285 | // TODO: There should be a way to always emit events even if the current |
| 5286 | // function is not instrumented. Losing events in a stream can cripple |
| 5287 | // a trace. |
| 5288 | if (!ShouldXRayInstrumentFunction()) |
| 5289 | return RValue::getIgnored(); |
| 5290 | |
| 5291 | if (!CGM.getCodeGenOpts().XRayInstrumentationBundle.has( |
| 5292 | XRayInstrKind::Typed)) |
| 5293 | return RValue::getIgnored(); |
| 5294 | |
| 5295 | if (const auto *XRayAttr = CurFuncDecl->getAttr<XRayInstrumentAttr>()) |
| 5296 | if (XRayAttr->neverXRayInstrument() && !AlwaysEmitXRayTypedEvents()) |
| 5297 | return RValue::getIgnored(); |
| 5298 | |
| 5299 | Function *F = CGM.getIntrinsic(Intrinsic::xray_typedevent); |
| 5300 | auto FTy = F->getFunctionType(); |
| 5301 | auto Arg0 = EmitScalarExpr(E->getArg(0)); |
| 5302 | auto PTy0 = FTy->getParamType(0); |
| 5303 | if (PTy0 != Arg0->getType()) |
| 5304 | Arg0 = Builder.CreateTruncOrBitCast(Arg0, PTy0); |
| 5305 | auto Arg1 = E->getArg(1); |
| 5306 | auto Arg1Val = EmitScalarExpr(Arg1); |
| 5307 | auto Arg1Ty = Arg1->getType(); |
| 5308 | auto PTy1 = FTy->getParamType(1); |
| 5309 | if (PTy1 != Arg1Val->getType()) { |
| 5310 | if (Arg1Ty->isArrayType()) |
| 5311 | Arg1Val = EmitArrayToPointerDecay(Arg1).getPointer(); |
| 5312 | else |
| 5313 | Arg1Val = Builder.CreatePointerCast(Arg1Val, PTy1); |
| 5314 | } |
| 5315 | auto Arg2 = EmitScalarExpr(E->getArg(2)); |
| 5316 | auto PTy2 = FTy->getParamType(2); |
| 5317 | if (PTy2 != Arg2->getType()) |
| 5318 | Arg2 = Builder.CreateTruncOrBitCast(Arg2, PTy2); |
| 5319 | return RValue::get(Builder.CreateCall(F, {Arg0, Arg1Val, Arg2})); |
| 5320 | } |
| 5321 | |
| 5322 | case Builtin::BI__builtin_ms_va_start: |
| 5323 | case Builtin::BI__builtin_ms_va_end: |
| 5324 | return RValue::get( |
| 5325 | EmitVAStartEnd(EmitMSVAListRef(E->getArg(0)).getPointer(), |
| 5326 | BuiltinID == Builtin::BI__builtin_ms_va_start)); |
| 5327 | |
| 5328 | case Builtin::BI__builtin_ms_va_copy: { |
| 5329 | // Lower this manually. We can't reliably determine whether or not any |
| 5330 | // given va_copy() is for a Win64 va_list from the calling convention |
| 5331 | // alone, because it's legal to do this from a System V ABI function. |
| 5332 | // With opaque pointer types, we won't have enough information in LLVM |
| 5333 | // IR to determine this from the argument types, either. Best to do it |
| 5334 | // now, while we have enough information. |
| 5335 | Address DestAddr = EmitMSVAListRef(E->getArg(0)); |
| 5336 | Address SrcAddr = EmitMSVAListRef(E->getArg(1)); |
| 5337 | |
| 5338 | llvm::Type *BPP = Int8PtrPtrTy; |
| 5339 | |
| 5340 | DestAddr = Address(Builder.CreateBitCast(DestAddr.getPointer(), BPP, "cp"), |
| 5341 | Int8PtrTy, DestAddr.getAlignment()); |
| 5342 | SrcAddr = Address(Builder.CreateBitCast(SrcAddr.getPointer(), BPP, "ap"), |
| 5343 | Int8PtrTy, SrcAddr.getAlignment()); |
| 5344 | |
| 5345 | Value *ArgPtr = Builder.CreateLoad(SrcAddr, "ap.val"); |
| 5346 | return RValue::get(Builder.CreateStore(ArgPtr, DestAddr)); |
| 5347 | } |
| 5348 | |
| 5349 | case Builtin::BI__builtin_get_device_side_mangled_name: { |
| 5350 | auto Name = CGM.getCUDARuntime().getDeviceSideName( |
| 5351 | cast<DeclRefExpr>(E->getArg(0)->IgnoreImpCasts())->getDecl()); |
| 5352 | auto Str = CGM.GetAddrOfConstantCString(Name, ""); |
| 5353 | llvm::Constant *Zeros[] = {llvm::ConstantInt::get(SizeTy, 0), |
| 5354 | llvm::ConstantInt::get(SizeTy, 0)}; |
| 5355 | auto *Ptr = llvm::ConstantExpr::getGetElementPtr(Str.getElementType(), |
| 5356 | Str.getPointer(), Zeros); |
| 5357 | return RValue::get(Ptr); |
| 5358 | } |
| 5359 | } |
| 5360 | |
| 5361 | // If this is an alias for a lib function (e.g. __builtin_sin), emit |
| 5362 | // the call using the normal call path, but using the unmangled |
| 5363 | // version of the function name. |
| 5364 | if (getContext().BuiltinInfo.isLibFunction(BuiltinID)) |
| 5365 | return emitLibraryCall(*this, FD, E, |
| 5366 | CGM.getBuiltinLibFunction(FD, BuiltinID)); |
| 5367 | |
| 5368 | // If this is a predefined lib function (e.g. malloc), emit the call |
| 5369 | // using exactly the normal call path. |
| 5370 | if (getContext().BuiltinInfo.isPredefinedLibFunction(BuiltinID)) |
| 5371 | return emitLibraryCall(*this, FD, E, |
| 5372 | cast<llvm::Constant>(EmitScalarExpr(E->getCallee()))); |
| 5373 | |
| 5374 | // Check that a call to a target specific builtin has the correct target |
| 5375 | // features. |
| 5376 | // This is down here to avoid non-target specific builtins, however, if |
| 5377 | // generic builtins start to require generic target features then we |
| 5378 | // can move this up to the beginning of the function. |
| 5379 | checkTargetFeatures(E, FD); |
| 5380 | |
| 5381 | if (unsigned VectorWidth = getContext().BuiltinInfo.getRequiredVectorWidth(BuiltinID)) |
| 5382 | LargestVectorWidth = std::max(LargestVectorWidth, VectorWidth); |
| 5383 | |
| 5384 | // See if we have a target specific intrinsic. |
| 5385 | StringRef Name = getContext().BuiltinInfo.getName(BuiltinID); |
| 5386 | Intrinsic::ID IntrinsicID = Intrinsic::not_intrinsic; |
| 5387 | StringRef Prefix = |
| 5388 | llvm::Triple::getArchTypePrefix(getTarget().getTriple().getArch()); |
| 5389 | if (!Prefix.empty()) { |
| 5390 | IntrinsicID = Intrinsic::getIntrinsicForClangBuiltin(Prefix.data(), Name); |
| 5391 | // NOTE we don't need to perform a compatibility flag check here since the |
| 5392 | // intrinsics are declared in Builtins*.def via LANGBUILTIN which filter the |
| 5393 | // MS builtins via ALL_MS_LANGUAGES and are filtered earlier. |
| 5394 | if (IntrinsicID == Intrinsic::not_intrinsic) |
| 5395 | IntrinsicID = Intrinsic::getIntrinsicForMSBuiltin(Prefix.data(), Name); |
| 5396 | } |
| 5397 | |
| 5398 | if (IntrinsicID != Intrinsic::not_intrinsic) { |
| 5399 | SmallVector<Value*, 16> Args; |
| 5400 | |
| 5401 | // Find out if any arguments are required to be integer constant |
| 5402 | // expressions. |
| 5403 | unsigned ICEArguments = 0; |
| 5404 | ASTContext::GetBuiltinTypeError Error; |
| 5405 | getContext().GetBuiltinType(BuiltinID, Error, &ICEArguments); |
| 5406 | assert(Error == ASTContext::GE_None && "Should not codegen an error")(static_cast <bool> (Error == ASTContext::GE_None && "Should not codegen an error") ? void (0) : __assert_fail ("Error == ASTContext::GE_None && \"Should not codegen an error\"" , "clang/lib/CodeGen/CGBuiltin.cpp", 5406, __extension__ __PRETTY_FUNCTION__ )); |
| 5407 | |
| 5408 | Function *F = CGM.getIntrinsic(IntrinsicID); |
| 5409 | llvm::FunctionType *FTy = F->getFunctionType(); |
| 5410 | |
| 5411 | for (unsigned i = 0, e = E->getNumArgs(); i != e; ++i) { |
| 5412 | Value *ArgValue; |
| 5413 | // If this is a normal argument, just emit it as a scalar. |
| 5414 | if ((ICEArguments & (1 << i)) == 0) { |
| 5415 | ArgValue = EmitScalarExpr(E->getArg(i)); |
| 5416 | } else { |
| 5417 | // If this is required to be a constant, constant fold it so that we |
| 5418 | // know that the generated intrinsic gets a ConstantInt. |
| 5419 | ArgValue = llvm::ConstantInt::get( |
| 5420 | getLLVMContext(), |
| 5421 | *E->getArg(i)->getIntegerConstantExpr(getContext())); |
| 5422 | } |
| 5423 | |
| 5424 | // If the intrinsic arg type is different from the builtin arg type |
| 5425 | // we need to do a bit cast. |
| 5426 | llvm::Type *PTy = FTy->getParamType(i); |
| 5427 | if (PTy != ArgValue->getType()) { |
| 5428 | // XXX - vector of pointers? |
| 5429 | if (auto *PtrTy = dyn_cast<llvm::PointerType>(PTy)) { |
| 5430 | if (PtrTy->getAddressSpace() != |
| 5431 | ArgValue->getType()->getPointerAddressSpace()) { |
| 5432 | ArgValue = Builder.CreateAddrSpaceCast( |
| 5433 | ArgValue, |
| 5434 | ArgValue->getType()->getPointerTo(PtrTy->getAddressSpace())); |
| 5435 | } |
| 5436 | } |
| 5437 | |
| 5438 | assert(PTy->canLosslesslyBitCastTo(FTy->getParamType(i)) &&(static_cast <bool> (PTy->canLosslesslyBitCastTo(FTy ->getParamType(i)) && "Must be able to losslessly bit cast to param" ) ? void (0) : __assert_fail ("PTy->canLosslesslyBitCastTo(FTy->getParamType(i)) && \"Must be able to losslessly bit cast to param\"" , "clang/lib/CodeGen/CGBuiltin.cpp", 5439, __extension__ __PRETTY_FUNCTION__ )) |
| 5439 | "Must be able to losslessly bit cast to param")(static_cast <bool> (PTy->canLosslesslyBitCastTo(FTy ->getParamType(i)) && "Must be able to losslessly bit cast to param" ) ? void (0) : __assert_fail ("PTy->canLosslesslyBitCastTo(FTy->getParamType(i)) && \"Must be able to losslessly bit cast to param\"" , "clang/lib/CodeGen/CGBuiltin.cpp", 5439, __extension__ __PRETTY_FUNCTION__ )); |
| 5440 | // Cast vector type (e.g., v256i32) to x86_amx, this only happen |
| 5441 | // in amx intrinsics. |
| 5442 | if (PTy->isX86_AMXTy()) |
| 5443 | ArgValue = Builder.CreateIntrinsic(Intrinsic::x86_cast_vector_to_tile, |
| 5444 | {ArgValue->getType()}, {ArgValue}); |
| 5445 | else |
| 5446 | ArgValue = Builder.CreateBitCast(ArgValue, PTy); |
| 5447 | } |
| 5448 | |
| 5449 | Args.push_back(ArgValue); |
| 5450 | } |
| 5451 | |
| 5452 | Value *V = Builder.CreateCall(F, Args); |
| 5453 | QualType BuiltinRetType = E->getType(); |
| 5454 | |
| 5455 | llvm::Type *RetTy = VoidTy; |
| 5456 | if (!BuiltinRetType->isVoidType()) |
| 5457 | RetTy = ConvertType(BuiltinRetType); |
| 5458 | |
| 5459 | if (RetTy != V->getType()) { |
| 5460 | // XXX - vector of pointers? |
| 5461 | if (auto *PtrTy = dyn_cast<llvm::PointerType>(RetTy)) { |
| 5462 | if (PtrTy->getAddressSpace() != V->getType()->getPointerAddressSpace()) { |
| 5463 | V = Builder.CreateAddrSpaceCast( |
| 5464 | V, V->getType()->getPointerTo(PtrTy->getAddressSpace())); |
| 5465 | } |
| 5466 | } |
| 5467 | |
| 5468 | assert(V->getType()->canLosslesslyBitCastTo(RetTy) &&(static_cast <bool> (V->getType()->canLosslesslyBitCastTo (RetTy) && "Must be able to losslessly bit cast result type" ) ? void (0) : __assert_fail ("V->getType()->canLosslesslyBitCastTo(RetTy) && \"Must be able to losslessly bit cast result type\"" , "clang/lib/CodeGen/CGBuiltin.cpp", 5469, __extension__ __PRETTY_FUNCTION__ )) |
| 5469 | "Must be able to losslessly bit cast result type")(static_cast <bool> (V->getType()->canLosslesslyBitCastTo (RetTy) && "Must be able to losslessly bit cast result type" ) ? void (0) : __assert_fail ("V->getType()->canLosslesslyBitCastTo(RetTy) && \"Must be able to losslessly bit cast result type\"" , "clang/lib/CodeGen/CGBuiltin.cpp", 5469, __extension__ __PRETTY_FUNCTION__ )); |
| 5470 | // Cast x86_amx to vector type (e.g., v256i32), this only happen |
| 5471 | // in amx intrinsics. |
| 5472 | if (V->getType()->isX86_AMXTy()) |
| 5473 | V = Builder.CreateIntrinsic(Intrinsic::x86_cast_tile_to_vector, {RetTy}, |
| 5474 | {V}); |
| 5475 | else |
| 5476 | V = Builder.CreateBitCast(V, RetTy); |
| 5477 | } |
| 5478 | |
| 5479 | if (RetTy->isVoidTy()) |
| 5480 | return RValue::get(nullptr); |
| 5481 | |
| 5482 | return RValue::get(V); |
| 5483 | } |
| 5484 | |
| 5485 | // Some target-specific builtins can have aggregate return values, e.g. |
| 5486 | // __builtin_arm_mve_vld2q_u32. So if the result is an aggregate, force |
| 5487 | // ReturnValue to be non-null, so that the target-specific emission code can |
| 5488 | // always just emit into it. |
| 5489 | TypeEvaluationKind EvalKind = getEvaluationKind(E->getType()); |
| 5490 | if (EvalKind == TEK_Aggregate && ReturnValue.isNull()) { |
| 5491 | Address DestPtr = CreateMemTemp(E->getType(), "agg.tmp"); |
| 5492 | ReturnValue = ReturnValueSlot(DestPtr, false); |
| 5493 | } |
| 5494 | |
| 5495 | // Now see if we can emit a target-specific builtin. |
| 5496 | if (Value *V = EmitTargetBuiltinExpr(BuiltinID, E, ReturnValue)) { |
| 5497 | switch (EvalKind) { |
| 5498 | case TEK_Scalar: |
| 5499 | if (V->getType()->isVoidTy()) |
| 5500 | return RValue::get(nullptr); |
| 5501 | return RValue::get(V); |
| 5502 | case TEK_Aggregate: |
| 5503 | return RValue::getAggregate(ReturnValue.getValue(), |
| 5504 | ReturnValue.isVolatile()); |
| 5505 | case TEK_Complex: |
| 5506 | llvm_unreachable("No current target builtin returns complex")::llvm::llvm_unreachable_internal("No current target builtin returns complex" , "clang/lib/CodeGen/CGBuiltin.cpp", 5506); |
| 5507 | } |
| 5508 | llvm_unreachable("Bad evaluation kind in EmitBuiltinExpr")::llvm::llvm_unreachable_internal("Bad evaluation kind in EmitBuiltinExpr" , "clang/lib/CodeGen/CGBuiltin.cpp", 5508); |
| 5509 | } |
| 5510 | |
| 5511 | ErrorUnsupported(E, "builtin function"); |
| 5512 | |
| 5513 | // Unknown builtin, for now just dump it out and return undef. |
| 5514 | return GetUndefRValue(E->getType()); |
| 5515 | } |
| 5516 | |
| 5517 | static Value *EmitTargetArchBuiltinExpr(CodeGenFunction *CGF, |
| 5518 | unsigned BuiltinID, const CallExpr *E, |
| 5519 | ReturnValueSlot ReturnValue, |
| 5520 | llvm::Triple::ArchType Arch) { |
| 5521 | switch (Arch) { |
| 5522 | case llvm::Triple::arm: |
| 5523 | case llvm::Triple::armeb: |
| 5524 | case llvm::Triple::thumb: |
| 5525 | case llvm::Triple::thumbeb: |
| 5526 | return CGF->EmitARMBuiltinExpr(BuiltinID, E, ReturnValue, Arch); |
| 5527 | case llvm::Triple::aarch64: |
| 5528 | case llvm::Triple::aarch64_32: |
| 5529 | case llvm::Triple::aarch64_be: |
| 5530 | return CGF->EmitAArch64BuiltinExpr(BuiltinID, E, Arch); |
| 5531 | case llvm::Triple::bpfeb: |
| 5532 | case llvm::Triple::bpfel: |
| 5533 | return CGF->EmitBPFBuiltinExpr(BuiltinID, E); |
| 5534 | case llvm::Triple::x86: |
| 5535 | case llvm::Triple::x86_64: |
| 5536 | return CGF->EmitX86BuiltinExpr(BuiltinID, E); |
| 5537 | case llvm::Triple::ppc: |
| 5538 | case llvm::Triple::ppcle: |
| 5539 | case llvm::Triple::ppc64: |
| 5540 | case llvm::Triple::ppc64le: |
| 5541 | return CGF->EmitPPCBuiltinExpr(BuiltinID, E); |
| 5542 | case llvm::Triple::r600: |
| 5543 | case llvm::Triple::amdgcn: |
| 5544 | return CGF->EmitAMDGPUBuiltinExpr(BuiltinID, E); |
| 5545 | case llvm::Triple::systemz: |
| 5546 | return CGF->EmitSystemZBuiltinExpr(BuiltinID, E); |
| 5547 | case llvm::Triple::nvptx: |
| 5548 | case llvm::Triple::nvptx64: |
| 5549 | return CGF->EmitNVPTXBuiltinExpr(BuiltinID, E); |
| 5550 | case llvm::Triple::wasm32: |
| 5551 | case llvm::Triple::wasm64: |
| 5552 | return CGF->EmitWebAssemblyBuiltinExpr(BuiltinID, E); |
| 5553 | case llvm::Triple::hexagon: |
| 5554 | return CGF->EmitHexagonBuiltinExpr(BuiltinID, E); |
| 5555 | case llvm::Triple::riscv32: |
| 5556 | case llvm::Triple::riscv64: |
| 5557 | return CGF->EmitRISCVBuiltinExpr(BuiltinID, E, ReturnValue); |
| 5558 | case llvm::Triple::loongarch32: |
| 5559 | case llvm::Triple::loongarch64: |
| 5560 | return CGF->EmitLoongArchBuiltinExpr(BuiltinID, E); |
| 5561 | default: |
| 5562 | return nullptr; |
| 5563 | } |
| 5564 | } |
| 5565 | |
| 5566 | Value *CodeGenFunction::EmitTargetBuiltinExpr(unsigned BuiltinID, |
| 5567 | const CallExpr *E, |
| 5568 | ReturnValueSlot ReturnValue) { |
| 5569 | if (getContext().BuiltinInfo.isAuxBuiltinID(BuiltinID)) { |
| 5570 | assert(getContext().getAuxTargetInfo() && "Missing aux target info")(static_cast <bool> (getContext().getAuxTargetInfo() && "Missing aux target info") ? void (0) : __assert_fail ("getContext().getAuxTargetInfo() && \"Missing aux target info\"" , "clang/lib/CodeGen/CGBuiltin.cpp", 5570, __extension__ __PRETTY_FUNCTION__ )); |
| 5571 | return EmitTargetArchBuiltinExpr( |
| 5572 | this, getContext().BuiltinInfo.getAuxBuiltinID(BuiltinID), E, |
| 5573 | ReturnValue, getContext().getAuxTargetInfo()->getTriple().getArch()); |
| 5574 | } |
| 5575 | |
| 5576 | return EmitTargetArchBuiltinExpr(this, BuiltinID, E, ReturnValue, |
| 5577 | getTarget().getTriple().getArch()); |
| 5578 | } |
| 5579 | |
| 5580 | static llvm::FixedVectorType *GetNeonType(CodeGenFunction *CGF, |
| 5581 | NeonTypeFlags TypeFlags, |
| 5582 | bool HasLegalHalfType = true, |
| 5583 | bool V1Ty = false, |
| 5584 | bool AllowBFloatArgsAndRet = true) { |
| 5585 | int IsQuad = TypeFlags.isQuad(); |
| 5586 | switch (TypeFlags.getEltType()) { |
| 5587 | case NeonTypeFlags::Int8: |
| 5588 | case NeonTypeFlags::Poly8: |
| 5589 | return llvm::FixedVectorType::get(CGF->Int8Ty, V1Ty ? 1 : (8 << IsQuad)); |
| 5590 | case NeonTypeFlags::Int16: |
| 5591 | case NeonTypeFlags::Poly16: |
| 5592 | return llvm::FixedVectorType::get(CGF->Int16Ty, V1Ty ? 1 : (4 << IsQuad)); |
| 5593 | case NeonTypeFlags::BFloat16: |
| 5594 | if (AllowBFloatArgsAndRet) |
| 5595 | return llvm::FixedVectorType::get(CGF->BFloatTy, V1Ty ? 1 : (4 << IsQuad)); |
| 5596 | else |
| 5597 | return llvm::FixedVectorType::get(CGF->Int16Ty, V1Ty ? 1 : (4 << IsQuad)); |
| 5598 | case NeonTypeFlags::Float16: |
| 5599 | if (HasLegalHalfType) |
| 5600 | return llvm::FixedVectorType::get(CGF->HalfTy, V1Ty ? 1 : (4 << IsQuad)); |
| 5601 | else |
| 5602 | return llvm::FixedVectorType::get(CGF->Int16Ty, V1Ty ? 1 : (4 << IsQuad)); |
| 5603 | case NeonTypeFlags::Int32: |
| 5604 | return llvm::FixedVectorType::get(CGF->Int32Ty, V1Ty ? 1 : (2 << IsQuad)); |
| 5605 | case NeonTypeFlags::Int64: |
| 5606 | case NeonTypeFlags::Poly64: |
| 5607 | return llvm::FixedVectorType::get(CGF->Int64Ty, V1Ty ? 1 : (1 << IsQuad)); |
| 5608 | case NeonTypeFlags::Poly128: |
| 5609 | // FIXME: i128 and f128 doesn't get fully support in Clang and llvm. |
| 5610 | // There is a lot of i128 and f128 API missing. |
| 5611 | // so we use v16i8 to represent poly128 and get pattern matched. |
| 5612 | return llvm::FixedVectorType::get(CGF->Int8Ty, 16); |
| 5613 | case NeonTypeFlags::Float32: |
| 5614 | return llvm::FixedVectorType::get(CGF->FloatTy, V1Ty ? 1 : (2 << IsQuad)); |
| 5615 | case NeonTypeFlags::Float64: |
| 5616 | return llvm::FixedVectorType::get(CGF->DoubleTy, V1Ty ? 1 : (1 << IsQuad)); |
| 5617 | } |
| 5618 | llvm_unreachable("Unknown vector element type!")::llvm::llvm_unreachable_internal("Unknown vector element type!" , "clang/lib/CodeGen/CGBuiltin.cpp", 5618); |
| 5619 | } |
| 5620 | |
| 5621 | static llvm::VectorType *GetFloatNeonType(CodeGenFunction *CGF, |
| 5622 | NeonTypeFlags IntTypeFlags) { |
| 5623 | int IsQuad = IntTypeFlags.isQuad(); |
| 5624 | switch (IntTypeFlags.getEltType()) { |
| 5625 | case NeonTypeFlags::Int16: |
| 5626 | return llvm::FixedVectorType::get(CGF->HalfTy, (4 << IsQuad)); |
| 5627 | case NeonTypeFlags::Int32: |
| 5628 | return llvm::FixedVectorType::get(CGF->FloatTy, (2 << IsQuad)); |
| 5629 | case NeonTypeFlags::Int64: |
| 5630 | return llvm::FixedVectorType::get(CGF->DoubleTy, (1 << IsQuad)); |
| 5631 | default: |
| 5632 | llvm_unreachable("Type can't be converted to floating-point!")::llvm::llvm_unreachable_internal("Type can't be converted to floating-point!" , "clang/lib/CodeGen/CGBuiltin.cpp", 5632); |
| 5633 | } |
| 5634 | } |
| 5635 | |
| 5636 | Value *CodeGenFunction::EmitNeonSplat(Value *V, Constant *C, |
| 5637 | const ElementCount &Count) { |
| 5638 | Value *SV = llvm::ConstantVector::getSplat(Count, C); |
| 5639 | return Builder.CreateShuffleVector(V, V, SV, "lane"); |
| 5640 | } |
| 5641 | |
| 5642 | Value *CodeGenFunction::EmitNeonSplat(Value *V, Constant *C) { |
| 5643 | ElementCount EC = cast<llvm::VectorType>(V->getType())->getElementCount(); |
| 5644 | return EmitNeonSplat(V, C, EC); |
| 5645 | } |
| 5646 | |
| 5647 | Value *CodeGenFunction::EmitNeonCall(Function *F, SmallVectorImpl<Value*> &Ops, |
| 5648 | const char *name, |
| 5649 | unsigned shift, bool rightshift) { |
| 5650 | unsigned j = 0; |
| 5651 | for (Function::const_arg_iterator ai = F->arg_begin(), ae = F->arg_end(); |
| 5652 | ai != ae; ++ai, ++j) { |
| 5653 | if (F->isConstrainedFPIntrinsic()) |
| 5654 | if (ai->getType()->isMetadataTy()) |
| 5655 | continue; |
| 5656 | if (shift > 0 && shift == j) |
| 5657 | Ops[j] = EmitNeonShiftVector(Ops[j], ai->getType(), rightshift); |
| 5658 | else |
| 5659 | Ops[j] = Builder.CreateBitCast(Ops[j], ai->getType(), name); |
| 5660 | } |
| 5661 | |
| 5662 | if (F->isConstrainedFPIntrinsic()) |
| 5663 | return Builder.CreateConstrainedFPCall(F, Ops, name); |
| 5664 | else |
| 5665 | return Builder.CreateCall(F, Ops, name); |
| 5666 | } |
| 5667 | |
| 5668 | Value *CodeGenFunction::EmitNeonShiftVector(Value *V, llvm::Type *Ty, |
| 5669 | bool neg) { |
| 5670 | int SV = cast<ConstantInt>(V)->getSExtValue(); |
| 5671 | return ConstantInt::get(Ty, neg ? -SV : SV); |
| 5672 | } |
| 5673 | |
| 5674 | // Right-shift a vector by a constant. |
| 5675 | Value *CodeGenFunction::EmitNeonRShiftImm(Value *Vec, Value *Shift, |
| 5676 | llvm::Type *Ty, bool usgn, |
| 5677 | const char *name) { |
| 5678 | llvm::VectorType *VTy = cast<llvm::VectorType>(Ty); |
| 5679 | |
| 5680 | int ShiftAmt = cast<ConstantInt>(Shift)->getSExtValue(); |
| 5681 | int EltSize = VTy->getScalarSizeInBits(); |
| 5682 | |
| 5683 | Vec = Builder.CreateBitCast(Vec, Ty); |
| 5684 | |
| 5685 | // lshr/ashr are undefined when the shift amount is equal to the vector |
| 5686 | // element size. |
| 5687 | if (ShiftAmt == EltSize) { |
| 5688 | if (usgn) { |
| 5689 | // Right-shifting an unsigned value by its size yields 0. |
| 5690 | return llvm::ConstantAggregateZero::get(VTy); |
| 5691 | } else { |
| 5692 | // Right-shifting a signed value by its size is equivalent |
| 5693 | // to a shift of size-1. |
| 5694 | --ShiftAmt; |
| 5695 | Shift = ConstantInt::get(VTy->getElementType(), ShiftAmt); |
| 5696 | } |
| 5697 | } |
| 5698 | |
| 5699 | Shift = EmitNeonShiftVector(Shift, Ty, false); |
| 5700 | if (usgn) |
| 5701 | return Builder.CreateLShr(Vec, Shift, name); |
| 5702 | else |
| 5703 | return Builder.CreateAShr(Vec, Shift, name); |
| 5704 | } |
| 5705 | |
| 5706 | enum { |
| 5707 | AddRetType = (1 << 0), |
| 5708 | Add1ArgType = (1 << 1), |
| 5709 | Add2ArgTypes = (1 << 2), |
| 5710 | |
| 5711 | VectorizeRetType = (1 << 3), |
| 5712 | VectorizeArgTypes = (1 << 4), |
| 5713 | |
| 5714 | InventFloatType = (1 << 5), |
| 5715 | UnsignedAlts = (1 << 6), |
| 5716 | |
| 5717 | Use64BitVectors = (1 << 7), |
| 5718 | Use128BitVectors = (1 << 8), |
| 5719 | |
| 5720 | Vectorize1ArgType = Add1ArgType | VectorizeArgTypes, |
| 5721 | VectorRet = AddRetType | VectorizeRetType, |
| 5722 | VectorRetGetArgs01 = |
| 5723 | AddRetType | Add2ArgTypes | VectorizeRetType | VectorizeArgTypes, |
| 5724 | FpCmpzModifiers = |
| 5725 | AddRetType | VectorizeRetType | Add1ArgType | InventFloatType |
| 5726 | }; |
| 5727 | |
| 5728 | namespace { |
| 5729 | struct ARMVectorIntrinsicInfo { |
| 5730 | const char *NameHint; |
| 5731 | unsigned BuiltinID; |
| 5732 | unsigned LLVMIntrinsic; |
| 5733 | unsigned AltLLVMIntrinsic; |
| 5734 | uint64_t TypeModifier; |
| 5735 | |
| 5736 | bool operator<(unsigned RHSBuiltinID) const { |
| 5737 | return BuiltinID < RHSBuiltinID; |
| 5738 | } |
| 5739 | bool operator<(const ARMVectorIntrinsicInfo &TE) const { |
| 5740 | return BuiltinID < TE.BuiltinID; |
| 5741 | } |
| 5742 | }; |
| 5743 | } // end anonymous namespace |
| 5744 | |
| 5745 | #define NEONMAP0(NameBase) \ |
| 5746 | { #NameBase, NEON::BI__builtin_neon_ ## NameBase, 0, 0, 0 } |
| 5747 | |
| 5748 | #define NEONMAP1(NameBase, LLVMIntrinsic, TypeModifier) \ |
| 5749 | { #NameBase, NEON:: BI__builtin_neon_ ## NameBase, \ |
| 5750 | Intrinsic::LLVMIntrinsic, 0, TypeModifier } |
| 5751 | |
| 5752 | #define NEONMAP2(NameBase, LLVMIntrinsic, AltLLVMIntrinsic, TypeModifier) \ |
| 5753 | { #NameBase, NEON:: BI__builtin_neon_ ## NameBase, \ |
| 5754 | Intrinsic::LLVMIntrinsic, Intrinsic::AltLLVMIntrinsic, \ |
| 5755 | TypeModifier } |
| 5756 | |
| 5757 | static const ARMVectorIntrinsicInfo ARMSIMDIntrinsicMap [] = { |
| 5758 | NEONMAP1(__a32_vcvt_bf16_f32, arm_neon_vcvtfp2bf, 0), |
| 5759 | NEONMAP0(splat_lane_v), |
| 5760 | NEONMAP0(splat_laneq_v), |
| 5761 | NEONMAP0(splatq_lane_v), |
| 5762 | NEONMAP0(splatq_laneq_v), |
| 5763 | NEONMAP2(vabd_v, arm_neon_vabdu, arm_neon_vabds, Add1ArgType | UnsignedAlts), |
| 5764 | NEONMAP2(vabdq_v, arm_neon_vabdu, arm_neon_vabds, Add1ArgType | UnsignedAlts), |
| 5765 | NEONMAP1(vabs_v, arm_neon_vabs, 0), |
| 5766 | NEONMAP1(vabsq_v, arm_neon_vabs, 0), |
| 5767 | NEONMAP0(vadd_v), |
| 5768 | NEONMAP0(vaddhn_v), |
| 5769 | NEONMAP0(vaddq_v), |
| 5770 | NEONMAP1(vaesdq_u8, arm_neon_aesd, 0), |
| 5771 | NEONMAP1(vaeseq_u8, arm_neon_aese, 0), |
| 5772 | NEONMAP1(vaesimcq_u8, arm_neon_aesimc, 0), |
| 5773 | NEONMAP1(vaesmcq_u8, arm_neon_aesmc, 0), |
| 5774 | NEONMAP1(vbfdot_f32, arm_neon_bfdot, 0), |
| 5775 | NEONMAP1(vbfdotq_f32, arm_neon_bfdot, 0), |
| 5776 | NEONMAP1(vbfmlalbq_f32, arm_neon_bfmlalb, 0), |
| 5777 | NEONMAP1(vbfmlaltq_f32, arm_neon_bfmlalt, 0), |
| 5778 | NEONMAP1(vbfmmlaq_f32, arm_neon_bfmmla, 0), |
| 5779 | NEONMAP1(vbsl_v, arm_neon_vbsl, AddRetType), |
| 5780 | NEONMAP1(vbslq_v, arm_neon_vbsl, AddRetType), |
| 5781 | NEONMAP1(vcadd_rot270_f16, arm_neon_vcadd_rot270, Add1ArgType), |
| 5782 | NEONMAP1(vcadd_rot270_f32, arm_neon_vcadd_rot270, Add1ArgType), |
| 5783 | NEONMAP1(vcadd_rot90_f16, arm_neon_vcadd_rot90, Add1ArgType), |
| 5784 | NEONMAP1(vcadd_rot90_f32, arm_neon_vcadd_rot90, Add1ArgType), |
| 5785 | NEONMAP1(vcaddq_rot270_f16, arm_neon_vcadd_rot270, Add1ArgType), |
| 5786 | NEONMAP1(vcaddq_rot270_f32, arm_neon_vcadd_rot270, Add1ArgType), |
| 5787 | NEONMAP1(vcaddq_rot270_f64, arm_neon_vcadd_rot270, Add1ArgType), |
| 5788 | NEONMAP1(vcaddq_rot90_f16, arm_neon_vcadd_rot90, Add1ArgType), |
| 5789 | NEONMAP1(vcaddq_rot90_f32, arm_neon_vcadd_rot90, Add1ArgType), |
| 5790 | NEONMAP1(vcaddq_rot90_f64, arm_neon_vcadd_rot90, Add1ArgType), |
| 5791 | NEONMAP1(vcage_v, arm_neon_vacge, 0), |
| 5792 | NEONMAP1(vcageq_v, arm_neon_vacge, 0), |
| 5793 | NEONMAP1(vcagt_v, arm_neon_vacgt, 0), |
| 5794 | NEONMAP1(vcagtq_v, arm_neon_vacgt, 0), |
| 5795 | NEONMAP1(vcale_v, arm_neon_vacge, 0), |
| 5796 | NEONMAP1(vcaleq_v, arm_neon_vacge, 0), |
| 5797 | NEONMAP1(vcalt_v, arm_neon_vacgt, 0), |
| 5798 | NEONMAP1(vcaltq_v, arm_neon_vacgt, 0), |
| 5799 | NEONMAP0(vceqz_v), |
| 5800 | NEONMAP0(vceqzq_v), |
| 5801 | NEONMAP0(vcgez_v), |
| 5802 | NEONMAP0(vcgezq_v), |
| 5803 | NEONMAP0(vcgtz_v), |
| 5804 | NEONMAP0(vcgtzq_v), |
| 5805 | NEONMAP0(vclez_v), |
| 5806 | NEONMAP0(vclezq_v), |
| 5807 | NEONMAP1(vcls_v, arm_neon_vcls, Add1ArgType), |
| 5808 | NEONMAP1(vclsq_v, arm_neon_vcls, Add1ArgType), |
| 5809 | NEONMAP0(vcltz_v), |
| 5810 | NEONMAP0(vcltzq_v), |
| 5811 | NEONMAP1(vclz_v, ctlz, Add1ArgType), |
| 5812 | NEONMAP1(vclzq_v, ctlz, Add1ArgType), |
| 5813 | NEONMAP1(vcnt_v, ctpop, Add1ArgType), |
| 5814 | NEONMAP1(vcntq_v, ctpop, Add1ArgType), |
| 5815 | NEONMAP1(vcvt_f16_f32, arm_neon_vcvtfp2hf, 0), |
| 5816 | NEONMAP0(vcvt_f16_s16), |
| 5817 | NEONMAP0(vcvt_f16_u16), |
| 5818 | NEONMAP1(vcvt_f32_f16, arm_neon_vcvthf2fp, 0), |
| 5819 | NEONMAP0(vcvt_f32_v), |
| 5820 | NEONMAP1(vcvt_n_f16_s16, arm_neon_vcvtfxs2fp, 0), |
| 5821 | NEONMAP1(vcvt_n_f16_u16, arm_neon_vcvtfxu2fp, 0), |
| 5822 | NEONMAP2(vcvt_n_f32_v, arm_neon_vcvtfxu2fp, arm_neon_vcvtfxs2fp, 0), |
| 5823 | NEONMAP1(vcvt_n_s16_f16, arm_neon_vcvtfp2fxs, 0), |
| 5824 | NEONMAP1(vcvt_n_s32_v, arm_neon_vcvtfp2fxs, 0), |
| 5825 | NEONMAP1(vcvt_n_s64_v, arm_neon_vcvtfp2fxs, 0), |
| 5826 | NEONMAP1(vcvt_n_u16_f16, arm_neon_vcvtfp2fxu, 0), |
| 5827 | NEONMAP1(vcvt_n_u32_v, arm_neon_vcvtfp2fxu, 0), |
| 5828 | NEONMAP1(vcvt_n_u64_v, arm_neon_vcvtfp2fxu, 0), |
| 5829 | NEONMAP0(vcvt_s16_f16), |
| 5830 | NEONMAP0(vcvt_s32_v), |
| 5831 | NEONMAP0(vcvt_s64_v), |
| 5832 | NEONMAP0(vcvt_u16_f16), |
| 5833 | NEONMAP0(vcvt_u32_v), |
| 5834 | NEONMAP0(vcvt_u64_v), |
| 5835 | NEONMAP1(vcvta_s16_f16, arm_neon_vcvtas, 0), |
| 5836 | NEONMAP1(vcvta_s32_v, arm_neon_vcvtas, 0), |
| 5837 | NEONMAP1(vcvta_s64_v, arm_neon_vcvtas, 0), |
| 5838 | NEONMAP1(vcvta_u16_f16, arm_neon_vcvtau, 0), |
| 5839 | NEONMAP1(vcvta_u32_v, arm_neon_vcvtau, 0), |
| 5840 | NEONMAP1(vcvta_u64_v, arm_neon_vcvtau, 0), |
| 5841 | NEONMAP1(vcvtaq_s16_f16, arm_neon_vcvtas, 0), |
| 5842 | NEONMAP1(vcvtaq_s32_v, arm_neon_vcvtas, 0), |
| 5843 | NEONMAP1(vcvtaq_s64_v, arm_neon_vcvtas, 0), |
| 5844 | NEONMAP1(vcvtaq_u16_f16, arm_neon_vcvtau, 0), |
| 5845 | NEONMAP1(vcvtaq_u32_v, arm_neon_vcvtau, 0), |
| 5846 | NEONMAP1(vcvtaq_u64_v, arm_neon_vcvtau, 0), |
| 5847 | NEONMAP1(vcvth_bf16_f32, arm_neon_vcvtbfp2bf, 0), |
| 5848 | NEONMAP1(vcvtm_s16_f16, arm_neon_vcvtms, 0), |
| 5849 | NEONMAP1(vcvtm_s32_v, arm_neon_vcvtms, 0), |
| 5850 | NEONMAP1(vcvtm_s64_v, arm_neon_vcvtms, 0), |
| 5851 | NEONMAP1(vcvtm_u16_f16, arm_neon_vcvtmu, 0), |
| 5852 | NEONMAP1(vcvtm_u32_v, arm_neon_vcvtmu, 0), |
| 5853 | NEONMAP1(vcvtm_u64_v, arm_neon_vcvtmu, 0), |
| 5854 | NEONMAP1(vcvtmq_s16_f16, arm_neon_vcvtms, 0), |
| 5855 | NEONMAP1(vcvtmq_s32_v, arm_neon_vcvtms, 0), |
| 5856 | NEONMAP1(vcvtmq_s64_v, arm_neon_vcvtms, 0), |
| 5857 | NEONMAP1(vcvtmq_u16_f16, arm_neon_vcvtmu, 0), |
| 5858 | NEONMAP1(vcvtmq_u32_v, arm_neon_vcvtmu, 0), |
| 5859 | NEONMAP1(vcvtmq_u64_v, arm_neon_vcvtmu, 0), |
| 5860 | NEONMAP1(vcvtn_s16_f16, arm_neon_vcvtns, 0), |
| 5861 | NEONMAP1(vcvtn_s32_v, arm_neon_vcvtns, 0), |
| 5862 | NEONMAP1(vcvtn_s64_v, arm_neon_vcvtns, 0), |
| 5863 | NEONMAP1(vcvtn_u16_f16, arm_neon_vcvtnu, 0), |
| 5864 | NEONMAP1(vcvtn_u32_v, arm_neon_vcvtnu, 0), |
| 5865 | NEONMAP1(vcvtn_u64_v, arm_neon_vcvtnu, 0), |
| 5866 | NEONMAP1(vcvtnq_s16_f16, arm_neon_vcvtns, 0), |
| 5867 | NEONMAP1(vcvtnq_s32_v, arm_neon_vcvtns, 0), |
| 5868 | NEONMAP1(vcvtnq_s64_v, arm_neon_vcvtns, 0), |
| 5869 | NEONMAP1(vcvtnq_u16_f16, arm_neon_vcvtnu, 0), |
| 5870 | NEONMAP1(vcvtnq_u32_v, arm_neon_vcvtnu, 0), |
| 5871 | NEONMAP1(vcvtnq_u64_v, arm_neon_vcvtnu, 0), |
| 5872 | NEONMAP1(vcvtp_s16_f16, arm_neon_vcvtps, 0), |
| 5873 | NEONMAP1(vcvtp_s32_v, arm_neon_vcvtps, 0), |
| 5874 | NEONMAP1(vcvtp_s64_v, arm_neon_vcvtps, 0), |
| 5875 | NEONMAP1(vcvtp_u16_f16, arm_neon_vcvtpu, 0), |
| 5876 | NEONMAP1(vcvtp_u32_v, arm_neon_vcvtpu, 0), |
| 5877 | NEONMAP1(vcvtp_u64_v, arm_neon_vcvtpu, 0), |
| 5878 | NEONMAP1(vcvtpq_s16_f16, arm_neon_vcvtps, 0), |
| 5879 | NEONMAP1(vcvtpq_s32_v, arm_neon_vcvtps, 0), |
| 5880 | NEONMAP1(vcvtpq_s64_v, arm_neon_vcvtps, 0), |
| 5881 | NEONMAP1(vcvtpq_u16_f16, arm_neon_vcvtpu, 0), |
| 5882 | NEONMAP1(vcvtpq_u32_v, arm_neon_vcvtpu, 0), |
| 5883 | NEONMAP1(vcvtpq_u64_v, arm_neon_vcvtpu, 0), |
| 5884 | NEONMAP0(vcvtq_f16_s16), |
| 5885 | NEONMAP0(vcvtq_f16_u16), |
| 5886 | NEONMAP0(vcvtq_f32_v), |
| 5887 | NEONMAP1(vcvtq_n_f16_s16, arm_neon_vcvtfxs2fp, 0), |
| 5888 | NEONMAP1(vcvtq_n_f16_u16, arm_neon_vcvtfxu2fp, 0), |
| 5889 | NEONMAP2(vcvtq_n_f32_v, arm_neon_vcvtfxu2fp, arm_neon_vcvtfxs2fp, 0), |
| 5890 | NEONMAP1(vcvtq_n_s16_f16, arm_neon_vcvtfp2fxs, 0), |
| 5891 | NEONMAP1(vcvtq_n_s32_v, arm_neon_vcvtfp2fxs, 0), |
| 5892 | NEONMAP1(vcvtq_n_s64_v, arm_neon_vcvtfp2fxs, 0), |
| 5893 | NEONMAP1(vcvtq_n_u16_f16, arm_neon_vcvtfp2fxu, 0), |
| 5894 | NEONMAP1(vcvtq_n_u32_v, arm_neon_vcvtfp2fxu, 0), |
| 5895 | NEONMAP1(vcvtq_n_u64_v, arm_neon_vcvtfp2fxu, 0), |
| 5896 | NEONMAP0(vcvtq_s16_f16), |
| 5897 | NEONMAP0(vcvtq_s32_v), |
| 5898 | NEONMAP0(vcvtq_s64_v), |
| 5899 | NEONMAP0(vcvtq_u16_f16), |
| 5900 | NEONMAP0(vcvtq_u32_v), |
| 5901 | NEONMAP0(vcvtq_u64_v), |
| 5902 | NEONMAP1(vdot_s32, arm_neon_sdot, 0), |
| 5903 | NEONMAP1(vdot_u32, arm_neon_udot, 0), |
| 5904 | NEONMAP1(vdotq_s32, arm_neon_sdot, 0), |
| 5905 | NEONMAP1(vdotq_u32, arm_neon_udot, 0), |
| 5906 | NEONMAP0(vext_v), |
| 5907 | NEONMAP0(vextq_v), |
| 5908 | NEONMAP0(vfma_v), |
| 5909 | NEONMAP0(vfmaq_v), |
| 5910 | NEONMAP2(vhadd_v, arm_neon_vhaddu, arm_neon_vhadds, Add1ArgType | UnsignedAlts), |
| 5911 | NEONMAP2(vhaddq_v, arm_neon_vhaddu, arm_neon_vhadds, Add1ArgType | UnsignedAlts), |
| 5912 | NEONMAP2(vhsub_v, arm_neon_vhsubu, arm_neon_vhsubs, Add1ArgType | UnsignedAlts), |
| 5913 | NEONMAP2(vhsubq_v, arm_neon_vhsubu, arm_neon_vhsubs, Add1ArgType | UnsignedAlts), |
| 5914 | NEONMAP0(vld1_dup_v), |
| 5915 | NEONMAP1(vld1_v, arm_neon_vld1, 0), |
| 5916 | NEONMAP1(vld1_x2_v, arm_neon_vld1x2, 0), |
| 5917 | NEONMAP1(vld1_x3_v, arm_neon_vld1x3, 0), |
| 5918 | NEONMAP1(vld1_x4_v, arm_neon_vld1x4, 0), |
| 5919 | NEONMAP0(vld1q_dup_v), |
| 5920 | NEONMAP1(vld1q_v, arm_neon_vld1, 0), |
| 5921 | NEONMAP1(vld1q_x2_v, arm_neon_vld1x2, 0), |
| 5922 | NEONMAP1(vld1q_x3_v, arm_neon_vld1x3, 0), |
| 5923 | NEONMAP1(vld1q_x4_v, arm_neon_vld1x4, 0), |
| 5924 | NEONMAP1(vld2_dup_v, arm_neon_vld2dup, 0), |
| 5925 | NEONMAP1(vld2_lane_v, arm_neon_vld2lane, 0), |
| 5926 | NEONMAP1(vld2_v, arm_neon_vld2, 0), |
| 5927 | NEONMAP1(vld2q_dup_v, arm_neon_vld2dup, 0), |
| 5928 | NEONMAP1(vld2q_lane_v, arm_neon_vld2lane, 0), |
| 5929 | NEONMAP1(vld2q_v, arm_neon_vld2, 0), |
| 5930 | NEONMAP1(vld3_dup_v, arm_neon_vld3dup, 0), |
| 5931 | NEONMAP1(vld3_lane_v, arm_neon_vld3lane, 0), |
| 5932 | NEONMAP1(vld3_v, arm_neon_vld3, 0), |
| 5933 | NEONMAP1(vld3q_dup_v, arm_neon_vld3dup, 0), |
| 5934 | NEONMAP1(vld3q_lane_v, arm_neon_vld3lane, 0), |
| 5935 | NEONMAP1(vld3q_v, arm_neon_vld3, 0), |
| 5936 | NEONMAP1(vld4_dup_v, arm_neon_vld4dup, 0), |
| 5937 | NEONMAP1(vld4_lane_v, arm_neon_vld4lane, 0), |
| 5938 | NEONMAP1(vld4_v, arm_neon_vld4, 0), |
| 5939 | NEONMAP1(vld4q_dup_v, arm_neon_vld4dup, 0), |
| 5940 | NEONMAP1(vld4q_lane_v, arm_neon_vld4lane, 0), |
| 5941 | NEONMAP1(vld4q_v, arm_neon_vld4, 0), |
| 5942 | NEONMAP2(vmax_v, arm_neon_vmaxu, arm_neon_vmaxs, Add1ArgType | UnsignedAlts), |
| 5943 | NEONMAP1(vmaxnm_v, arm_neon_vmaxnm, Add1ArgType), |
| 5944 | NEONMAP1(vmaxnmq_v, arm_neon_vmaxnm, Add1ArgType), |
| 5945 | NEONMAP2(vmaxq_v, arm_neon_vmaxu, arm_neon_vmaxs, Add1ArgType | UnsignedAlts), |
| 5946 | NEONMAP2(vmin_v, arm_neon_vminu, arm_neon_vmins, Add1ArgType | UnsignedAlts), |
| 5947 | NEONMAP1(vminnm_v, arm_neon_vminnm, Add1ArgType), |
| 5948 | NEONMAP1(vminnmq_v, arm_neon_vminnm, Add1ArgType), |
| 5949 | NEONMAP2(vminq_v, arm_neon_vminu, arm_neon_vmins, Add1ArgType | UnsignedAlts), |
| 5950 | NEONMAP1(vmmlaq_s32, arm_neon_smmla, 0), |
| 5951 | NEONMAP1(vmmlaq_u32, arm_neon_ummla, 0), |
| 5952 | NEONMAP0(vmovl_v), |
| 5953 | NEONMAP0(vmovn_v), |
| 5954 | NEONMAP1(vmul_v, arm_neon_vmulp, Add1ArgType), |
| 5955 | NEONMAP0(vmull_v), |
| 5956 | NEONMAP1(vmulq_v, arm_neon_vmulp, Add1ArgType), |
| 5957 | NEONMAP2(vpadal_v, arm_neon_vpadalu, arm_neon_vpadals, UnsignedAlts), |
| 5958 | NEONMAP2(vpadalq_v, arm_neon_vpadalu, arm_neon_vpadals, UnsignedAlts), |
| 5959 | NEONMAP1(vpadd_v, arm_neon_vpadd, Add1ArgType), |
| 5960 | NEONMAP2(vpaddl_v, arm_neon_vpaddlu, arm_neon_vpaddls, UnsignedAlts), |
| 5961 | NEONMAP2(vpaddlq_v, arm_neon_vpaddlu, arm_neon_vpaddls, UnsignedAlts), |
| 5962 | NEONMAP1(vpaddq_v, arm_neon_vpadd, Add1ArgType), |
| 5963 | NEONMAP2(vpmax_v, arm_neon_vpmaxu, arm_neon_vpmaxs, Add1ArgType | UnsignedAlts), |
| 5964 | NEONMAP2(vpmin_v, arm_neon_vpminu, arm_neon_vpmins, Add1ArgType | UnsignedAlts), |
| 5965 | NEONMAP1(vqabs_v, arm_neon_vqabs, Add1ArgType), |
| 5966 | NEONMAP1(vqabsq_v, arm_neon_vqabs, Add1ArgType), |
| 5967 | NEONMAP2(vqadd_v, uadd_sat, sadd_sat, Add1ArgType | UnsignedAlts), |
| 5968 | NEONMAP2(vqaddq_v, uadd_sat, sadd_sat, Add1ArgType | UnsignedAlts), |
| 5969 | NEONMAP2(vqdmlal_v, arm_neon_vqdmull, sadd_sat, 0), |
| 5970 | NEONMAP2(vqdmlsl_v, arm_neon_vqdmull, ssub_sat, 0), |
| 5971 | NEONMAP1(vqdmulh_v, arm_neon_vqdmulh, Add1ArgType), |
| 5972 | NEONMAP1(vqdmulhq_v, arm_neon_vqdmulh, Add1ArgType), |
| 5973 | NEONMAP1(vqdmull_v, arm_neon_vqdmull, Add1ArgType), |
| 5974 | NEONMAP2(vqmovn_v, arm_neon_vqmovnu, arm_neon_vqmovns, Add1ArgType | UnsignedAlts), |
| 5975 | NEONMAP1(vqmovun_v, arm_neon_vqmovnsu, Add1ArgType), |
| 5976 | NEONMAP1(vqneg_v, arm_neon_vqneg, Add1ArgType), |
| 5977 | NEONMAP1(vqnegq_v, arm_neon_vqneg, Add1ArgType), |
| 5978 | NEONMAP1(vqrdmlah_s16, arm_neon_vqrdmlah, Add1ArgType), |
| 5979 | NEONMAP1(vqrdmlah_s32, arm_neon_vqrdmlah, Add1ArgType), |
| 5980 | NEONMAP1(vqrdmlahq_s16, arm_neon_vqrdmlah, Add1ArgType), |
| 5981 | NEONMAP1(vqrdmlahq_s32, arm_neon_vqrdmlah, Add1ArgType), |
| 5982 | NEONMAP1(vqrdmlsh_s16, arm_neon_vqrdmlsh, Add1ArgType), |
| 5983 | NEONMAP1(vqrdmlsh_s32, arm_neon_vqrdmlsh, Add1ArgType), |
| 5984 | NEONMAP1(vqrdmlshq_s16, arm_neon_vqrdmlsh, Add1ArgType), |
| 5985 | NEONMAP1(vqrdmlshq_s32, arm_neon_vqrdmlsh, Add1ArgType), |
| 5986 | NEONMAP1(vqrdmulh_v, arm_neon_vqrdmulh, Add1ArgType), |
| 5987 | NEONMAP1(vqrdmulhq_v, arm_neon_vqrdmulh, Add1ArgType), |
| 5988 | NEONMAP2(vqrshl_v, arm_neon_vqrshiftu, arm_neon_vqrshifts, Add1ArgType | UnsignedAlts), |
| 5989 | NEONMAP2(vqrshlq_v, arm_neon_vqrshiftu, arm_neon_vqrshifts, Add1ArgType | UnsignedAlts), |
| 5990 | NEONMAP2(vqshl_n_v, arm_neon_vqshiftu, arm_neon_vqshifts, UnsignedAlts), |
| 5991 | NEONMAP2(vqshl_v, arm_neon_vqshiftu, arm_neon_vqshifts, Add1ArgType | UnsignedAlts), |
| 5992 | NEONMAP2(vqshlq_n_v, arm_neon_vqshiftu, arm_neon_vqshifts, UnsignedAlts), |
| 5993 | NEONMAP2(vqshlq_v, arm_neon_vqshiftu, arm_neon_vqshifts, Add1ArgType | UnsignedAlts), |
| 5994 | NEONMAP1(vqshlu_n_v, arm_neon_vqshiftsu, 0), |
| 5995 | NEONMAP1(vqshluq_n_v, arm_neon_vqshiftsu, 0), |
| 5996 | NEONMAP2(vqsub_v, usub_sat, ssub_sat, Add1ArgType | UnsignedAlts), |
| 5997 | NEONMAP2(vqsubq_v, usub_sat, ssub_sat, Add1ArgType | UnsignedAlts), |
| 5998 | NEONMAP1(vraddhn_v, arm_neon_vraddhn, Add1ArgType), |
| 5999 | NEONMAP2(vrecpe_v, arm_neon_vrecpe, arm_neon_vrecpe, 0), |
| 6000 | NEONMAP2(vrecpeq_v, arm_neon_vrecpe, arm_neon_vrecpe, 0), |
| 6001 | NEONMAP1(vrecps_v, arm_neon_vrecps, Add1ArgType), |
| 6002 | NEONMAP1(vrecpsq_v, arm_neon_vrecps, Add1ArgType), |
| 6003 | NEONMAP2(vrhadd_v, arm_neon_vrhaddu, arm_neon_vrhadds, Add1ArgType | UnsignedAlts), |
| 6004 | NEONMAP2(vrhaddq_v, arm_neon_vrhaddu, arm_neon_vrhadds, Add1ArgType | UnsignedAlts), |
| 6005 | NEONMAP1(vrnd_v, arm_neon_vrintz, Add1ArgType), |
| 6006 | NEONMAP1(vrnda_v, arm_neon_vrinta, Add1ArgType), |
| 6007 | NEONMAP1(vrndaq_v, arm_neon_vrinta, Add1ArgType), |
| 6008 | NEONMAP0(vrndi_v), |
| 6009 | NEONMAP0(vrndiq_v), |
| 6010 | NEONMAP1(vrndm_v, arm_neon_vrintm, Add1ArgType), |
| 6011 | NEONMAP1(vrndmq_v, arm_neon_vrintm, Add1ArgType), |
| 6012 | NEONMAP1(vrndn_v, arm_neon_vrintn, Add1ArgType), |
| 6013 | NEONMAP1(vrndnq_v, arm_neon_vrintn, Add1ArgType), |
| 6014 | NEONMAP1(vrndp_v, arm_neon_vrintp, Add1ArgType), |
| 6015 | NEONMAP1(vrndpq_v, arm_neon_vrintp, Add1ArgType), |
| 6016 | NEONMAP1(vrndq_v, arm_neon_vrintz, Add1ArgType), |
| 6017 | NEONMAP1(vrndx_v, arm_neon_vrintx, Add1ArgType), |
| 6018 | NEONMAP1(vrndxq_v, arm_neon_vrintx, Add1ArgType), |
| 6019 | NEONMAP2(vrshl_v, arm_neon_vrshiftu, arm_neon_vrshifts, Add1ArgType | UnsignedAlts), |
| 6020 | NEONMAP2(vrshlq_v, arm_neon_vrshiftu, arm_neon_vrshifts, Add1ArgType | UnsignedAlts), |
| 6021 | NEONMAP2(vrshr_n_v, arm_neon_vrshiftu, arm_neon_vrshifts, UnsignedAlts), |
| 6022 | NEONMAP2(vrshrq_n_v, arm_neon_vrshiftu, arm_neon_vrshifts, UnsignedAlts), |
| 6023 | NEONMAP2(vrsqrte_v, arm_neon_vrsqrte, arm_neon_vrsqrte, 0), |
| 6024 | NEONMAP2(vrsqrteq_v, arm_neon_vrsqrte, arm_neon_vrsqrte, 0), |
| 6025 | NEONMAP1(vrsqrts_v, arm_neon_vrsqrts, Add1ArgType), |
| 6026 | NEONMAP1(vrsqrtsq_v, arm_neon_vrsqrts, Add1ArgType), |
| 6027 | NEONMAP1(vrsubhn_v, arm_neon_vrsubhn, Add1ArgType), |
| 6028 | NEONMAP1(vsha1su0q_u32, arm_neon_sha1su0, 0), |
| 6029 | NEONMAP1(vsha1su1q_u32, arm_neon_sha1su1, 0), |
| 6030 | NEONMAP1(vsha256h2q_u32, arm_neon_sha256h2, 0), |
| 6031 | NEONMAP1(vsha256hq_u32, arm_neon_sha256h, 0), |
| 6032 | NEONMAP1(vsha256su0q_u32, arm_neon_sha256su0, 0), |
| 6033 | NEONMAP1(vsha256su1q_u32, arm_neon_sha256su1, 0), |
| 6034 | NEONMAP0(vshl_n_v), |
| 6035 | NEONMAP2(vshl_v, arm_neon_vshiftu, arm_neon_vshifts, Add1ArgType | UnsignedAlts), |
| 6036 | NEONMAP0(vshll_n_v), |
| 6037 | NEONMAP0(vshlq_n_v), |
| 6038 | NEONMAP2(vshlq_v, arm_neon_vshiftu, arm_neon_vshifts, Add1ArgType | UnsignedAlts), |
| 6039 | NEONMAP0(vshr_n_v), |
| 6040 | NEONMAP0(vshrn_n_v), |
| 6041 | NEONMAP0(vshrq_n_v), |
| 6042 | NEONMAP1(vst1_v, arm_neon_vst1, 0), |
| 6043 | NEONMAP1(vst1_x2_v, arm_neon_vst1x2, 0), |
| 6044 | NEONMAP1(vst1_x3_v, arm_neon_vst1x3, 0), |
| 6045 | NEONMAP1(vst1_x4_v, arm_neon_vst1x4, 0), |
| 6046 | NEONMAP1(vst1q_v, arm_neon_vst1, 0), |
| 6047 | NEONMAP1(vst1q_x2_v, arm_neon_vst1x2, 0), |
| 6048 | NEONMAP1(vst1q_x3_v, arm_neon_vst1x3, 0), |
| 6049 | NEONMAP1(vst1q_x4_v, arm_neon_vst1x4, 0), |
| 6050 | NEONMAP1(vst2_lane_v, arm_neon_vst2lane, 0), |
| 6051 | NEONMAP1(vst2_v, arm_neon_vst2, 0), |
| 6052 | NEONMAP1(vst2q_lane_v, arm_neon_vst2lane, 0), |
| 6053 | NEONMAP1(vst2q_v, arm_neon_vst2, 0), |
| 6054 | NEONMAP1(vst3_lane_v, arm_neon_vst3lane, 0), |
| 6055 | NEONMAP1(vst3_v, arm_neon_vst3, 0), |
| 6056 | NEONMAP1(vst3q_lane_v, arm_neon_vst3lane, 0), |
| 6057 | NEONMAP1(vst3q_v, arm_neon_vst3, 0), |
| 6058 | NEONMAP1(vst4_lane_v, arm_neon_vst4lane, 0), |
| 6059 | NEONMAP1(vst4_v, arm_neon_vst4, 0), |
| 6060 | NEONMAP1(vst4q_lane_v, arm_neon_vst4lane, 0), |
| 6061 | NEONMAP1(vst4q_v, arm_neon_vst4, 0), |
| 6062 | NEONMAP0(vsubhn_v), |
| 6063 | NEONMAP0(vtrn_v), |
| 6064 | NEONMAP0(vtrnq_v), |
| 6065 | NEONMAP0(vtst_v), |
| 6066 | NEONMAP0(vtstq_v), |
| 6067 | NEONMAP1(vusdot_s32, arm_neon_usdot, 0), |
| 6068 | NEONMAP1(vusdotq_s32, arm_neon_usdot, 0), |
| 6069 | NEONMAP1(vusmmlaq_s32, arm_neon_usmmla, 0), |
| 6070 | NEONMAP0(vuzp_v), |
| 6071 | NEONMAP0(vuzpq_v), |
| 6072 | NEONMAP0(vzip_v), |
| 6073 | NEONMAP0(vzipq_v) |
| 6074 | }; |
| 6075 | |
| 6076 | static const ARMVectorIntrinsicInfo AArch64SIMDIntrinsicMap[] = { |
| 6077 | NEONMAP1(__a64_vcvtq_low_bf16_f32, aarch64_neon_bfcvtn, 0), |
| 6078 | NEONMAP0(splat_lane_v), |
| 6079 | NEONMAP0(splat_laneq_v), |
| 6080 | NEONMAP0(splatq_lane_v), |
| 6081 | NEONMAP0(splatq_laneq_v), |
| 6082 | NEONMAP1(vabs_v, aarch64_neon_abs, 0), |
| 6083 | NEONMAP1(vabsq_v, aarch64_neon_abs, 0), |
| 6084 | NEONMAP0(vadd_v), |
| 6085 | NEONMAP0(vaddhn_v), |
| 6086 | NEONMAP0(vaddq_p128), |
| 6087 | NEONMAP0(vaddq_v), |
| 6088 | NEONMAP1(vaesdq_u8, aarch64_crypto_aesd, 0), |
| 6089 | NEONMAP1(vaeseq_u8, aarch64_crypto_aese, 0), |
| 6090 | NEONMAP1(vaesimcq_u8, aarch64_crypto_aesimc, 0), |
| 6091 | NEONMAP1(vaesmcq_u8, aarch64_crypto_aesmc, 0), |
| 6092 | NEONMAP2(vbcaxq_s16, aarch64_crypto_bcaxu, aarch64_crypto_bcaxs, Add1ArgType | UnsignedAlts), |
| 6093 | NEONMAP2(vbcaxq_s32, aarch64_crypto_bcaxu, aarch64_crypto_bcaxs, Add1ArgType | UnsignedAlts), |
| 6094 | NEONMAP2(vbcaxq_s64, aarch64_crypto_bcaxu, aarch64_crypto_bcaxs, Add1ArgType | UnsignedAlts), |
| 6095 | NEONMAP2(vbcaxq_s8, aarch64_crypto_bcaxu, aarch64_crypto_bcaxs, Add1ArgType | UnsignedAlts), |
| 6096 | NEONMAP2(vbcaxq_u16, aarch64_crypto_bcaxu, aarch64_crypto_bcaxs, Add1ArgType | UnsignedAlts), |
| 6097 | NEONMAP2(vbcaxq_u32, aarch64_crypto_bcaxu, aarch64_crypto_bcaxs, Add1ArgType | UnsignedAlts), |
| 6098 | NEONMAP2(vbcaxq_u64, aarch64_crypto_bcaxu, aarch64_crypto_bcaxs, Add1ArgType | UnsignedAlts), |
| 6099 | NEONMAP2(vbcaxq_u8, aarch64_crypto_bcaxu, aarch64_crypto_bcaxs, Add1ArgType | UnsignedAlts), |
| 6100 | NEONMAP1(vbfdot_f32, aarch64_neon_bfdot, 0), |
| 6101 | NEONMAP1(vbfdotq_f32, aarch64_neon_bfdot, 0), |
| 6102 | NEONMAP1(vbfmlalbq_f32, aarch64_neon_bfmlalb, 0), |
| 6103 | NEONMAP1(vbfmlaltq_f32, aarch64_neon_bfmlalt, 0), |
| 6104 | NEONMAP1(vbfmmlaq_f32, aarch64_neon_bfmmla, 0), |
| 6105 | NEONMAP1(vcadd_rot270_f16, aarch64_neon_vcadd_rot270, Add1ArgType), |
| 6106 | NEONMAP1(vcadd_rot270_f32, aarch64_neon_vcadd_rot270, Add1ArgType), |
| 6107 | NEONMAP1(vcadd_rot90_f16, aarch64_neon_vcadd_rot90, Add1ArgType), |
| 6108 | NEONMAP1(vcadd_rot90_f32, aarch64_neon_vcadd_rot90, Add1ArgType), |
| 6109 | NEONMAP1(vcaddq_rot270_f16, aarch64_neon_vcadd_rot270, Add1ArgType), |
| 6110 | NEONMAP1(vcaddq_rot270_f32, aarch64_neon_vcadd_rot270, Add1ArgType), |
| 6111 | NEONMAP1(vcaddq_rot270_f64, aarch64_neon_vcadd_rot270, Add1ArgType), |
| 6112 | NEONMAP1(vcaddq_rot90_f16, aarch64_neon_vcadd_rot90, Add1ArgType), |
| 6113 | NEONMAP1(vcaddq_rot90_f32, aarch64_neon_vcadd_rot90, Add1ArgType), |
| 6114 | NEONMAP1(vcaddq_rot90_f64, aarch64_neon_vcadd_rot90, Add1ArgType), |
| 6115 | NEONMAP1(vcage_v, aarch64_neon_facge, 0), |
| 6116 | NEONMAP1(vcageq_v, aarch64_neon_facge, 0), |
| 6117 | NEONMAP1(vcagt_v, aarch64_neon_facgt, 0), |
| 6118 | NEONMAP1(vcagtq_v, aarch64_neon_facgt, 0), |
| 6119 | NEONMAP1(vcale_v, aarch64_neon_facge, 0), |
| 6120 | NEONMAP1(vcaleq_v, aarch64_neon_facge, 0), |
| 6121 | NEONMAP1(vcalt_v, aarch64_neon_facgt, 0), |
| 6122 | NEONMAP1(vcaltq_v, aarch64_neon_facgt, 0), |
| 6123 | NEONMAP0(vceqz_v), |
| 6124 | NEONMAP0(vceqzq_v), |
| 6125 | NEONMAP0(vcgez_v), |
| 6126 | NEONMAP0(vcgezq_v), |
| 6127 | NEONMAP0(vcgtz_v), |
| 6128 | NEONMAP0(vcgtzq_v), |
| 6129 | NEONMAP0(vclez_v), |
| 6130 | NEONMAP0(vclezq_v), |
| 6131 | NEONMAP1(vcls_v, aarch64_neon_cls, Add1ArgType), |
| 6132 | NEONMAP1(vclsq_v, aarch64_neon_cls, Add1ArgType), |
| 6133 | NEONMAP0(vcltz_v), |
| 6134 | NEONMAP0(vcltzq_v), |
| 6135 | NEONMAP1(vclz_v, ctlz, Add1ArgType), |
| 6136 | NEONMAP1(vclzq_v, ctlz, Add1ArgType), |
| 6137 | NEONMAP1(vcmla_f16, aarch64_neon_vcmla_rot0, Add1ArgType), |
| 6138 | NEONMAP1(vcmla_f32, aarch64_neon_vcmla_rot0, Add1ArgType), |
| 6139 | NEONMAP1(vcmla_rot180_f16, aarch64_neon_vcmla_rot180, Add1ArgType), |
| 6140 | NEONMAP1(vcmla_rot180_f32, aarch64_neon_vcmla_rot180, Add1ArgType), |
| 6141 | NEONMAP1(vcmla_rot270_f16, aarch64_neon_vcmla_rot270, Add1ArgType), |
| 6142 | NEONMAP1(vcmla_rot270_f32, aarch64_neon_vcmla_rot270, Add1ArgType), |
| 6143 | NEONMAP1(vcmla_rot90_f16, aarch64_neon_vcmla_rot90, Add1ArgType), |
| 6144 | NEONMAP1(vcmla_rot90_f32, aarch64_neon_vcmla_rot90, Add1ArgType), |
| 6145 | NEONMAP1(vcmlaq_f16, aarch64_neon_vcmla_rot0, Add1ArgType), |
| 6146 | NEONMAP1(vcmlaq_f32, aarch64_neon_vcmla_rot0, Add1ArgType), |
| 6147 | NEONMAP1(vcmlaq_f64, aarch64_neon_vcmla_rot0, Add1ArgType), |
| 6148 | NEONMAP1(vcmlaq_rot180_f16, aarch64_neon_vcmla_rot180, Add1ArgType), |
| 6149 | NEONMAP1(vcmlaq_rot180_f32, aarch64_neon_vcmla_rot180, Add1ArgType), |
| 6150 | NEONMAP1(vcmlaq_rot180_f64, aarch64_neon_vcmla_rot180, Add1ArgType), |
| 6151 | NEONMAP1(vcmlaq_rot270_f16, aarch64_neon_vcmla_rot270, Add1ArgType), |
| 6152 | NEONMAP1(vcmlaq_rot270_f32, aarch64_neon_vcmla_rot270, Add1ArgType), |
| 6153 | NEONMAP1(vcmlaq_rot270_f64, aarch64_neon_vcmla_rot270, Add1ArgType), |
| 6154 | NEONMAP1(vcmlaq_rot90_f16, aarch64_neon_vcmla_rot90, Add1ArgType), |
| 6155 | NEONMAP1(vcmlaq_rot90_f32, aarch64_neon_vcmla_rot90, Add1ArgType), |
| 6156 | NEONMAP1(vcmlaq_rot90_f64, aarch64_neon_vcmla_rot90, Add1ArgType), |
| 6157 | NEONMAP1(vcnt_v, ctpop, Add1ArgType), |
| 6158 | NEONMAP1(vcntq_v, ctpop, Add1ArgType), |
| 6159 | NEONMAP1(vcvt_f16_f32, aarch64_neon_vcvtfp2hf, 0), |
| 6160 | NEONMAP0(vcvt_f16_s16), |
| 6161 | NEONMAP0(vcvt_f16_u16), |
| 6162 | NEONMAP1(vcvt_f32_f16, aarch64_neon_vcvthf2fp, 0), |
| 6163 | NEONMAP0(vcvt_f32_v), |
| 6164 | NEONMAP1(vcvt_n_f16_s16, aarch64_neon_vcvtfxs2fp, 0), |
| 6165 | NEONMAP1(vcvt_n_f16_u16, aarch64_neon_vcvtfxu2fp, 0), |
| 6166 | NEONMAP2(vcvt_n_f32_v, aarch64_neon_vcvtfxu2fp, aarch64_neon_vcvtfxs2fp, 0), |
| 6167 | NEONMAP2(vcvt_n_f64_v, aarch64_neon_vcvtfxu2fp, aarch64_neon_vcvtfxs2fp, 0), |
| 6168 | NEONMAP1(vcvt_n_s16_f16, aarch64_neon_vcvtfp2fxs, 0), |
| 6169 | NEONMAP1(vcvt_n_s32_v, aarch64_neon_vcvtfp2fxs, 0), |
| 6170 | NEONMAP1(vcvt_n_s64_v, aarch64_neon_vcvtfp2fxs, 0), |
| 6171 | NEONMAP1(vcvt_n_u16_f16, aarch64_neon_vcvtfp2fxu, 0), |
| 6172 | NEONMAP1(vcvt_n_u32_v, aarch64_neon_vcvtfp2fxu, 0), |
| 6173 | NEONMAP1(vcvt_n_u64_v, aarch64_neon_vcvtfp2fxu, 0), |
| 6174 | NEONMAP0(vcvtq_f16_s16), |
| 6175 | NEONMAP0(vcvtq_f16_u16), |
| 6176 | NEONMAP0(vcvtq_f32_v), |
| 6177 | NEONMAP1(vcvtq_high_bf16_f32, aarch64_neon_bfcvtn2, 0), |
| 6178 | NEONMAP1(vcvtq_n_f16_s16, aarch64_neon_vcvtfxs2fp, 0), |
| 6179 | NEONMAP1(vcvtq_n_f16_u16, aarch64_neon_vcvtfxu2fp, 0), |
| 6180 | NEONMAP2(vcvtq_n_f32_v, aarch64_neon_vcvtfxu2fp, aarch64_neon_vcvtfxs2fp, 0), |
| 6181 | NEONMAP2(vcvtq_n_f64_v, aarch64_neon_vcvtfxu2fp, aarch64_neon_vcvtfxs2fp, 0), |
| 6182 | NEONMAP1(vcvtq_n_s16_f16, aarch64_neon_vcvtfp2fxs, 0), |
| 6183 | NEONMAP1(vcvtq_n_s32_v, aarch64_neon_vcvtfp2fxs, 0), |
| 6184 | NEONMAP1(vcvtq_n_s64_v, aarch64_neon_vcvtfp2fxs, 0), |
| 6185 | NEONMAP1(vcvtq_n_u16_f16, aarch64_neon_vcvtfp2fxu, 0), |
| 6186 | NEONMAP1(vcvtq_n_u32_v, aarch64_neon_vcvtfp2fxu, 0), |
| 6187 | NEONMAP1(vcvtq_n_u64_v, aarch64_neon_vcvtfp2fxu, 0), |
| 6188 | NEONMAP1(vcvtx_f32_v, aarch64_neon_fcvtxn, AddRetType | Add1ArgType), |
| 6189 | NEONMAP1(vdot_s32, aarch64_neon_sdot, 0), |
| 6190 | NEONMAP1(vdot_u32, aarch64_neon_udot, 0), |
| 6191 | NEONMAP1(vdotq_s32, aarch64_neon_sdot, 0), |
| 6192 | NEONMAP1(vdotq_u32, aarch64_neon_udot, 0), |
| 6193 | NEONMAP2(veor3q_s16, aarch64_crypto_eor3u, aarch64_crypto_eor3s, Add1ArgType | UnsignedAlts), |
| 6194 | NEONMAP2(veor3q_s32, aarch64_crypto_eor3u, aarch64_crypto_eor3s, Add1ArgType | UnsignedAlts), |
| 6195 | NEONMAP2(veor3q_s64, aarch64_crypto_eor3u, aarch64_crypto_eor3s, Add1ArgType | UnsignedAlts), |
| 6196 | NEONMAP2(veor3q_s8, aarch64_crypto_eor3u, aarch64_crypto_eor3s, Add1ArgType | UnsignedAlts), |
| 6197 | NEONMAP2(veor3q_u16, aarch64_crypto_eor3u, aarch64_crypto_eor3s, Add1ArgType | UnsignedAlts), |
| 6198 | NEONMAP2(veor3q_u32, aarch64_crypto_eor3u, aarch64_crypto_eor3s, Add1ArgType | UnsignedAlts), |
| 6199 | NEONMAP2(veor3q_u64, aarch64_crypto_eor3u, aarch64_crypto_eor3s, Add1ArgType | UnsignedAlts), |
| 6200 | NEONMAP2(veor3q_u8, aarch64_crypto_eor3u, aarch64_crypto_eor3s, Add1ArgType | UnsignedAlts), |
| 6201 | NEONMAP0(vext_v), |
| 6202 | NEONMAP0(vextq_v), |
| 6203 | NEONMAP0(vfma_v), |
| 6204 | NEONMAP0(vfmaq_v), |
| 6205 | NEONMAP1(vfmlal_high_f16, aarch64_neon_fmlal2, 0), |
| 6206 | NEONMAP1(vfmlal_low_f16, aarch64_neon_fmlal, 0), |
| 6207 | NEONMAP1(vfmlalq_high_f16, aarch64_neon_fmlal2, 0), |
| 6208 | NEONMAP1(vfmlalq_low_f16, aarch64_neon_fmlal, 0), |
| 6209 | NEONMAP1(vfmlsl_high_f16, aarch64_neon_fmlsl2, 0), |
| 6210 | NEONMAP1(vfmlsl_low_f16, aarch64_neon_fmlsl, 0), |
| 6211 | NEONMAP1(vfmlslq_high_f16, aarch64_neon_fmlsl2, 0), |
| 6212 | NEONMAP1(vfmlslq_low_f16, aarch64_neon_fmlsl, 0), |
| 6213 | NEONMAP2(vhadd_v, aarch64_neon_uhadd, aarch64_neon_shadd, Add1ArgType | UnsignedAlts), |
| 6214 | NEONMAP2(vhaddq_v, aarch64_neon_uhadd, aarch64_neon_shadd, Add1ArgType | UnsignedAlts), |
| 6215 | NEONMAP2(vhsub_v, aarch64_neon_uhsub, aarch64_neon_shsub, Add1ArgType | UnsignedAlts), |
| 6216 | NEONMAP2(vhsubq_v, aarch64_neon_uhsub, aarch64_neon_shsub, Add1ArgType | UnsignedAlts), |
| 6217 | NEONMAP1(vld1_x2_v, aarch64_neon_ld1x2, 0), |
| 6218 | NEONMAP1(vld1_x3_v, aarch64_neon_ld1x3, 0), |
| 6219 | NEONMAP1(vld1_x4_v, aarch64_neon_ld1x4, 0), |
| 6220 | NEONMAP1(vld1q_x2_v, aarch64_neon_ld1x2, 0), |
| 6221 | NEONMAP1(vld1q_x3_v, aarch64_neon_ld1x3, 0), |
| 6222 | NEONMAP1(vld1q_x4_v, aarch64_neon_ld1x4, 0), |
| 6223 | NEONMAP1(vmmlaq_s32, aarch64_neon_smmla, 0), |
| 6224 | NEONMAP1(vmmlaq_u32, aarch64_neon_ummla, 0), |
| 6225 | NEONMAP0(vmovl_v), |
| 6226 | NEONMAP0(vmovn_v), |
| 6227 | NEONMAP1(vmul_v, aarch64_neon_pmul, Add1ArgType), |
| 6228 | NEONMAP1(vmulq_v, aarch64_neon_pmul, Add1ArgType), |
| 6229 | NEONMAP1(vpadd_v, aarch64_neon_addp, Add1ArgType), |
| 6230 | NEONMAP2(vpaddl_v, aarch64_neon_uaddlp, aarch64_neon_saddlp, UnsignedAlts), |
| 6231 | NEONMAP2(vpaddlq_v, aarch64_neon_uaddlp, aarch64_neon_saddlp, UnsignedAlts), |
| 6232 | NEONMAP1(vpaddq_v, aarch64_neon_addp, Add1ArgType), |
| 6233 | NEONMAP1(vqabs_v, aarch64_neon_sqabs, Add1ArgType), |
| 6234 | NEONMAP1(vqabsq_v, aarch64_neon_sqabs, Add1ArgType), |
| 6235 | NEONMAP2(vqadd_v, aarch64_neon_uqadd, aarch64_neon_sqadd, Add1ArgType | UnsignedAlts), |
| 6236 | NEONMAP2(vqaddq_v, aarch64_neon_uqadd, aarch64_neon_sqadd, Add1ArgType | UnsignedAlts), |
| 6237 | NEONMAP2(vqdmlal_v, aarch64_neon_sqdmull, aarch64_neon_sqadd, 0), |
| 6238 | NEONMAP2(vqdmlsl_v, aarch64_neon_sqdmull, aarch64_neon_sqsub, 0), |
| 6239 | NEONMAP1(vqdmulh_lane_v, aarch64_neon_sqdmulh_lane, 0), |
| 6240 | NEONMAP1(vqdmulh_laneq_v, aarch64_neon_sqdmulh_laneq, 0), |
| 6241 | NEONMAP1(vqdmulh_v, aarch64_neon_sqdmulh, Add1ArgType), |
| 6242 | NEONMAP1(vqdmulhq_lane_v, aarch64_neon_sqdmulh_lane, 0), |
| 6243 | NEONMAP1(vqdmulhq_laneq_v, aarch64_neon_sqdmulh_laneq, 0), |
| 6244 | NEONMAP1(vqdmulhq_v, aarch64_neon_sqdmulh, Add1ArgType), |
| 6245 | NEONMAP1(vqdmull_v, aarch64_neon_sqdmull, Add1ArgType), |
| 6246 | NEONMAP2(vqmovn_v, aarch64_neon_uqxtn, aarch64_neon_sqxtn, Add1ArgType | UnsignedAlts), |
| 6247 | NEONMAP1(vqmovun_v, aarch64_neon_sqxtun, Add1ArgType), |
| 6248 | NEONMAP1(vqneg_v, aarch64_neon_sqneg, Add1ArgType), |
| 6249 | NEONMAP1(vqnegq_v, aarch64_neon_sqneg, Add1ArgType), |
| 6250 | NEONMAP1(vqrdmlah_s16, aarch64_neon_sqrdmlah, Add1ArgType), |
| 6251 | NEONMAP1(vqrdmlah_s32, aarch64_neon_sqrdmlah, Add1ArgType), |
| 6252 | NEONMAP1(vqrdmlahq_s16, aarch64_neon_sqrdmlah, Add1ArgType), |
| 6253 | NEONMAP1(vqrdmlahq_s32, aarch64_neon_sqrdmlah, Add1ArgType), |
| 6254 | NEONMAP1(vqrdmlsh_s16, aarch64_neon_sqrdmlsh, Add1ArgType), |
| 6255 | NEONMAP1(vqrdmlsh_s32, aarch64_neon_sqrdmlsh, Add1ArgType), |
| 6256 | NEONMAP1(vqrdmlshq_s16, aarch64_neon_sqrdmlsh, Add1ArgType), |
| 6257 | NEONMAP1(vqrdmlshq_s32, aarch64_neon_sqrdmlsh, Add1ArgType), |
| 6258 | NEONMAP1(vqrdmulh_lane_v, aarch64_neon_sqrdmulh_lane, 0), |
| 6259 | NEONMAP1(vqrdmulh_laneq_v, aarch64_neon_sqrdmulh_laneq, 0), |
| 6260 | NEONMAP1(vqrdmulh_v, aarch64_neon_sqrdmulh, Add1ArgType), |
| 6261 | NEONMAP1(vqrdmulhq_lane_v, aarch64_neon_sqrdmulh_lane, 0), |
| 6262 | NEONMAP1(vqrdmulhq_laneq_v, aarch64_neon_sqrdmulh_laneq, 0), |
| 6263 | NEONMAP1(vqrdmulhq_v, aarch64_neon_sqrdmulh, Add1ArgType), |
| 6264 | NEONMAP2(vqrshl_v, aarch64_neon_uqrshl, aarch64_neon_sqrshl, Add1ArgType | UnsignedAlts), |
| 6265 | NEONMAP2(vqrshlq_v, aarch64_neon_uqrshl, aarch64_neon_sqrshl, Add1ArgType | UnsignedAlts), |
| 6266 | NEONMAP2(vqshl_n_v, aarch64_neon_uqshl, aarch64_neon_sqshl, UnsignedAlts), |
| 6267 | NEONMAP2(vqshl_v, aarch64_neon_uqshl, aarch64_neon_sqshl, Add1ArgType | UnsignedAlts), |
| 6268 | NEONMAP2(vqshlq_n_v, aarch64_neon_uqshl, aarch64_neon_sqshl,UnsignedAlts), |
| 6269 | NEONMAP2(vqshlq_v, aarch64_neon_uqshl, aarch64_neon_sqshl, Add1ArgType | UnsignedAlts), |
| 6270 | NEONMAP1(vqshlu_n_v, aarch64_neon_sqshlu, 0), |
| 6271 | NEONMAP1(vqshluq_n_v, aarch64_neon_sqshlu, 0), |
| 6272 | NEONMAP2(vqsub_v, aarch64_neon_uqsub, aarch64_neon_sqsub, Add1ArgType | UnsignedAlts), |
| 6273 | NEONMAP2(vqsubq_v, aarch64_neon_uqsub, aarch64_neon_sqsub, Add1ArgType | UnsignedAlts), |
| 6274 | NEONMAP1(vraddhn_v, aarch64_neon_raddhn, Add1ArgType), |
| 6275 | NEONMAP1(vrax1q_u64, aarch64_crypto_rax1, 0), |
| 6276 | NEONMAP2(vrecpe_v, aarch64_neon_frecpe, aarch64_neon_urecpe, 0), |
| 6277 | NEONMAP2(vrecpeq_v, aarch64_neon_frecpe, aarch64_neon_urecpe, 0), |
| 6278 | NEONMAP1(vrecps_v, aarch64_neon_frecps, Add1ArgType), |
| 6279 | NEONMAP1(vrecpsq_v, aarch64_neon_frecps, Add1ArgType), |
| 6280 | NEONMAP2(vrhadd_v, aarch64_neon_urhadd, aarch64_neon_srhadd, Add1ArgType | UnsignedAlts), |
| 6281 | NEONMAP2(vrhaddq_v, aarch64_neon_urhadd, aarch64_neon_srhadd, Add1ArgType | UnsignedAlts), |
| 6282 | NEONMAP1(vrnd32x_f32, aarch64_neon_frint32x, Add1ArgType), |
| 6283 | NEONMAP1(vrnd32xq_f32, aarch64_neon_frint32x, Add1ArgType), |
| 6284 | NEONMAP1(vrnd32z_f32, aarch64_neon_frint32z, Add1ArgType), |
| 6285 | NEONMAP1(vrnd32zq_f32, aarch64_neon_frint32z, Add1ArgType), |
| 6286 | NEONMAP1(vrnd64x_f32, aarch64_neon_frint64x, Add1ArgType), |
| 6287 | NEONMAP1(vrnd64xq_f32, aarch64_neon_frint64x, Add1ArgType), |
| 6288 | NEONMAP1(vrnd64z_f32, aarch64_neon_frint64z, Add1ArgType), |
| 6289 | NEONMAP1(vrnd64zq_f32, aarch64_neon_frint64z, Add1ArgType), |
| 6290 | NEONMAP0(vrndi_v), |
| 6291 | NEONMAP0(vrndiq_v), |
| 6292 | NEONMAP2(vrshl_v, aarch64_neon_urshl, aarch64_neon_srshl, Add1ArgType | UnsignedAlts), |
| 6293 | NEONMAP2(vrshlq_v, aarch64_neon_urshl, aarch64_neon_srshl, Add1ArgType | UnsignedAlts), |
| 6294 | NEONMAP2(vrshr_n_v, aarch64_neon_urshl, aarch64_neon_srshl, UnsignedAlts), |
| 6295 | NEONMAP2(vrshrq_n_v, aarch64_neon_urshl, aarch64_neon_srshl, UnsignedAlts), |
| 6296 | NEONMAP2(vrsqrte_v, aarch64_neon_frsqrte, aarch64_neon_ursqrte, 0), |
| 6297 | NEONMAP2(vrsqrteq_v, aarch64_neon_frsqrte, aarch64_neon_ursqrte, 0), |
| 6298 | NEONMAP1(vrsqrts_v, aarch64_neon_frsqrts, Add1ArgType), |
| 6299 | NEONMAP1(vrsqrtsq_v, aarch64_neon_frsqrts, Add1ArgType), |
| 6300 | NEONMAP1(vrsubhn_v, aarch64_neon_rsubhn, Add1ArgType), |
| 6301 | NEONMAP1(vsha1su0q_u32, aarch64_crypto_sha1su0, 0), |
| 6302 | NEONMAP1(vsha1su1q_u32, aarch64_crypto_sha1su1, 0), |
| 6303 | NEONMAP1(vsha256h2q_u32, aarch64_crypto_sha256h2, 0), |
| 6304 | NEONMAP1(vsha256hq_u32, aarch64_crypto_sha256h, 0), |
| 6305 | NEONMAP1(vsha256su0q_u32, aarch64_crypto_sha256su0, 0), |
| 6306 | NEONMAP1(vsha256su1q_u32, aarch64_crypto_sha256su1, 0), |
| 6307 | NEONMAP1(vsha512h2q_u64, aarch64_crypto_sha512h2, 0), |
| 6308 | NEONMAP1(vsha512hq_u64, aarch64_crypto_sha512h, 0), |
| 6309 | NEONMAP1(vsha512su0q_u64, aarch64_crypto_sha512su0, 0), |
| 6310 | NEONMAP1(vsha512su1q_u64, aarch64_crypto_sha512su1, 0), |
| 6311 | NEONMAP0(vshl_n_v), |
| 6312 | NEONMAP2(vshl_v, aarch64_neon_ushl, aarch64_neon_sshl, Add1ArgType | UnsignedAlts), |
| 6313 | NEONMAP0(vshll_n_v), |
| 6314 | NEONMAP0(vshlq_n_v), |
| 6315 | NEONMAP2(vshlq_v, aarch64_neon_ushl, aarch64_neon_sshl, Add1ArgType | UnsignedAlts), |
| 6316 | NEONMAP0(vshr_n_v), |
| 6317 | NEONMAP0(vshrn_n_v), |
| 6318 | NEONMAP0(vshrq_n_v), |
| 6319 | NEONMAP1(vsm3partw1q_u32, aarch64_crypto_sm3partw1, 0), |
| 6320 | NEONMAP1(vsm3partw2q_u32, aarch64_crypto_sm3partw2, 0), |
| 6321 | NEONMAP1(vsm3ss1q_u32, aarch64_crypto_sm3ss1, 0), |
| 6322 | NEONMAP1(vsm3tt1aq_u32, aarch64_crypto_sm3tt1a, 0), |
| 6323 | NEONMAP1(vsm3tt1bq_u32, aarch64_crypto_sm3tt1b, 0), |
| 6324 | NEONMAP1(vsm3tt2aq_u32, aarch64_crypto_sm3tt2a, 0), |
| 6325 | NEONMAP1(vsm3tt2bq_u32, aarch64_crypto_sm3tt2b, 0), |
| 6326 | NEONMAP1(vsm4ekeyq_u32, aarch64_crypto_sm4ekey, 0), |
| 6327 | NEONMAP1(vsm4eq_u32, aarch64_crypto_sm4e, 0), |
| 6328 | NEONMAP1(vst1_x2_v, aarch64_neon_st1x2, 0), |
| 6329 | NEONMAP1(vst1_x3_v, aarch64_neon_st1x3, 0), |
| 6330 | NEONMAP1(vst1_x4_v, aarch64_neon_st1x4, 0), |
| 6331 | NEONMAP1(vst1q_x2_v, aarch64_neon_st1x2, 0), |
| 6332 | NEONMAP1(vst1q_x3_v, aarch64_neon_st1x3, 0), |
| 6333 | NEONMAP1(vst1q_x4_v, aarch64_neon_st1x4, 0), |
| 6334 | NEONMAP0(vsubhn_v), |
| 6335 | NEONMAP0(vtst_v), |
| 6336 | NEONMAP0(vtstq_v), |
| 6337 | NEONMAP1(vusdot_s32, aarch64_neon_usdot, 0), |
| 6338 | NEONMAP1(vusdotq_s32, aarch64_neon_usdot, 0), |
| 6339 | NEONMAP1(vusmmlaq_s32, aarch64_neon_usmmla, 0), |
| 6340 | NEONMAP1(vxarq_u64, aarch64_crypto_xar, 0), |
| 6341 | }; |
| 6342 | |
| 6343 | static const ARMVectorIntrinsicInfo AArch64SISDIntrinsicMap[] = { |
| 6344 | NEONMAP1(vabdd_f64, aarch64_sisd_fabd, Add1ArgType), |
| 6345 | NEONMAP1(vabds_f32, aarch64_sisd_fabd, Add1ArgType), |
| 6346 | NEONMAP1(vabsd_s64, aarch64_neon_abs, Add1ArgType), |
| 6347 | NEONMAP1(vaddlv_s32, aarch64_neon_saddlv, AddRetType | Add1ArgType), |
| 6348 | NEONMAP1(vaddlv_u32, aarch64_neon_uaddlv, AddRetType | Add1ArgType), |
| 6349 | NEONMAP1(vaddlvq_s32, aarch64_neon_saddlv, AddRetType | Add1ArgType), |
| 6350 | NEONMAP1(vaddlvq_u32, aarch64_neon_uaddlv, AddRetType | Add1ArgType), |
| 6351 | NEONMAP1(vaddv_f32, aarch64_neon_faddv, AddRetType | Add1ArgType), |
| 6352 | NEONMAP1(vaddv_s32, aarch64_neon_saddv, AddRetType | Add1ArgType), |
| 6353 | NEONMAP1(vaddv_u32, aarch64_neon_uaddv, AddRetType | Add1ArgType), |
| 6354 | NEONMAP1(vaddvq_f32, aarch64_neon_faddv, AddRetType | Add1ArgType), |
| 6355 | NEONMAP1(vaddvq_f64, aarch64_neon_faddv, AddRetType | Add1ArgType), |
| 6356 | NEONMAP1(vaddvq_s32, aarch64_neon_saddv, AddRetType | Add1ArgType), |
| 6357 | NEONMAP1(vaddvq_s64, aarch64_neon_saddv, AddRetType | Add1ArgType), |
| 6358 | NEONMAP1(vaddvq_u32, aarch64_neon_uaddv, AddRetType | Add1ArgType), |
| 6359 | NEONMAP1(vaddvq_u64, aarch64_neon_uaddv, AddRetType | Add1ArgType), |
| 6360 | NEONMAP1(vcaged_f64, aarch64_neon_facge, AddRetType | Add1ArgType), |
| 6361 | NEONMAP1(vcages_f32, aarch64_neon_facge, AddRetType | Add1ArgType), |
| 6362 | NEONMAP1(vcagtd_f64, aarch64_neon_facgt, AddRetType | Add1ArgType), |
| 6363 | NEONMAP1(vcagts_f32, aarch64_neon_facgt, AddRetType | Add1ArgType), |
| 6364 | NEONMAP1(vcaled_f64, aarch64_neon_facge, AddRetType | Add1ArgType), |
| 6365 | NEONMAP1(vcales_f32, aarch64_neon_facge, AddRetType | Add1ArgType), |
| 6366 | NEONMAP1(vcaltd_f64, aarch64_neon_facgt, AddRetType | Add1ArgType), |
| 6367 | NEONMAP1(vcalts_f32, aarch64_neon_facgt, AddRetType | Add1ArgType), |
| 6368 | NEONMAP1(vcvtad_s64_f64, aarch64_neon_fcvtas, AddRetType | Add1ArgType), |
| 6369 | NEONMAP1(vcvtad_u64_f64, aarch64_neon_fcvtau, AddRetType | Add1ArgType), |
| 6370 | NEONMAP1(vcvtas_s32_f32, aarch64_neon_fcvtas, AddRetType | Add1ArgType), |
| 6371 | NEONMAP1(vcvtas_u32_f32, aarch64_neon_fcvtau, AddRetType | Add1ArgType), |
| 6372 | NEONMAP1(vcvtd_n_f64_s64, aarch64_neon_vcvtfxs2fp, AddRetType | Add1ArgType), |
| 6373 | NEONMAP1(vcvtd_n_f64_u64, aarch64_neon_vcvtfxu2fp, AddRetType | Add1ArgType), |
| 6374 | NEONMAP1(vcvtd_n_s64_f64, aarch64_neon_vcvtfp2fxs, AddRetType | Add1ArgType), |
| 6375 | NEONMAP1(vcvtd_n_u64_f64, aarch64_neon_vcvtfp2fxu, AddRetType | Add1ArgType), |
| 6376 | NEONMAP1(vcvtd_s64_f64, aarch64_neon_fcvtzs, AddRetType | Add1ArgType), |
| 6377 | NEONMAP1(vcvtd_u64_f64, aarch64_neon_fcvtzu, AddRetType | Add1ArgType), |
| 6378 | NEONMAP1(vcvth_bf16_f32, aarch64_neon_bfcvt, 0), |
| 6379 | NEONMAP1(vcvtmd_s64_f64, aarch64_neon_fcvtms, AddRetType | Add1ArgType), |
| 6380 | NEONMAP1(vcvtmd_u64_f64, aarch64_neon_fcvtmu, AddRetType | Add1ArgType), |
| 6381 | NEONMAP1(vcvtms_s32_f32, aarch64_neon_fcvtms, AddRetType | Add1ArgType), |
| 6382 | NEONMAP1(vcvtms_u32_f32, aarch64_neon_fcvtmu, AddRetType | Add1ArgType), |
| 6383 | NEONMAP1(vcvtnd_s64_f64, aarch64_neon_fcvtns, AddRetType | Add1ArgType), |
| 6384 | NEONMAP1(vcvtnd_u64_f64, aarch64_neon_fcvtnu, AddRetType | Add1ArgType), |
| 6385 | NEONMAP1(vcvtns_s32_f32, aarch64_neon_fcvtns, AddRetType | Add1ArgType), |
| 6386 | NEONMAP1(vcvtns_u32_f32, aarch64_neon_fcvtnu, AddRetType | Add1ArgType), |
| 6387 | NEONMAP1(vcvtpd_s64_f64, aarch64_neon_fcvtps, AddRetType | Add1ArgType), |
| 6388 | NEONMAP1(vcvtpd_u64_f64, aarch64_neon_fcvtpu, AddRetType | Add1ArgType), |
| 6389 | NEONMAP1(vcvtps_s32_f32, aarch64_neon_fcvtps, AddRetType | Add1ArgType), |
| 6390 | NEONMAP1(vcvtps_u32_f32, aarch64_neon_fcvtpu, AddRetType | Add1ArgType), |
| 6391 | NEONMAP1(vcvts_n_f32_s32, aarch64_neon_vcvtfxs2fp, AddRetType | Add1ArgType), |
| 6392 | NEONMAP1(vcvts_n_f32_u32, aarch64_neon_vcvtfxu2fp, AddRetType | Add1ArgType), |
| 6393 | NEONMAP1(vcvts_n_s32_f32, aarch64_neon_vcvtfp2fxs, AddRetType | Add1ArgType), |
| 6394 | NEONMAP1(vcvts_n_u32_f32, aarch64_neon_vcvtfp2fxu, AddRetType | Add1ArgType), |
| 6395 | NEONMAP1(vcvts_s32_f32, aarch64_neon_fcvtzs, AddRetType | Add1ArgType), |
| 6396 | NEONMAP1(vcvts_u32_f32, aarch64_neon_fcvtzu, AddRetType | Add1ArgType), |
| 6397 | NEONMAP1(vcvtxd_f32_f64, aarch64_sisd_fcvtxn, 0), |
| 6398 | NEONMAP1(vmaxnmv_f32, aarch64_neon_fmaxnmv, AddRetType | Add1ArgType), |
| 6399 | NEONMAP1(vmaxnmvq_f32, aarch64_neon_fmaxnmv, AddRetType | Add1ArgType), |
| 6400 | NEONMAP1(vmaxnmvq_f64, aarch64_neon_fmaxnmv, AddRetType | Add1ArgType), |
| 6401 | NEONMAP1(vmaxv_f32, aarch64_neon_fmaxv, AddRetType | Add1ArgType), |
| 6402 | NEONMAP1(vmaxv_s32, aarch64_neon_smaxv, AddRetType | Add1ArgType), |
| 6403 | NEONMAP1(vmaxv_u32, aarch64_neon_umaxv, AddRetType | Add1ArgType), |
| 6404 | NEONMAP1(vmaxvq_f32, aarch64_neon_fmaxv, AddRetType | Add1ArgType), |
| 6405 | NEONMAP1(vmaxvq_f64, aarch64_neon_fmaxv, AddRetType | Add1ArgType), |
| 6406 | NEONMAP1(vmaxvq_s32, aarch64_neon_smaxv, AddRetType | Add1ArgType), |
| 6407 | NEONMAP1(vmaxvq_u32, aarch64_neon_umaxv, AddRetType | Add1ArgType), |
| 6408 | NEONMAP1(vminnmv_f32, aarch64_neon_fminnmv, AddRetType | Add1ArgType), |
| 6409 | NEONMAP1(vminnmvq_f32, aarch64_neon_fminnmv, AddRetType | Add1ArgType), |
| 6410 | NEONMAP1(vminnmvq_f64, aarch64_neon_fminnmv, AddRetType | Add1ArgType), |
| 6411 | NEONMAP1(vminv_f32, aarch64_neon_fminv, AddRetType | Add1ArgType), |
| 6412 | NEONMAP1(vminv_s32, aarch64_neon_sminv, AddRetType | Add1ArgType), |
| 6413 | NEONMAP1(vminv_u32, aarch64_neon_uminv, AddRetType | Add1ArgType), |
| 6414 | NEONMAP1(vminvq_f32, aarch64_neon_fminv, AddRetType | Add1ArgType), |
| 6415 | NEONMAP1(vminvq_f64, aarch64_neon_fminv, AddRetType | Add1ArgType), |
| 6416 | NEONMAP1(vminvq_s32, aarch64_neon_sminv, AddRetType | Add1ArgType), |
| 6417 | NEONMAP1(vminvq_u32, aarch64_neon_uminv, AddRetType | Add1ArgType), |
| 6418 | NEONMAP1(vmull_p64, aarch64_neon_pmull64, 0), |
| 6419 | NEONMAP1(vmulxd_f64, aarch64_neon_fmulx, Add1ArgType), |
| 6420 | NEONMAP1(vmulxs_f32, aarch64_neon_fmulx, Add1ArgType), |
| 6421 | NEONMAP1(vpaddd_s64, aarch64_neon_uaddv, AddRetType | Add1ArgType), |
| 6422 | NEONMAP1(vpaddd_u64, aarch64_neon_uaddv, AddRetType | Add1ArgType), |
| 6423 | NEONMAP1(vpmaxnmqd_f64, aarch64_neon_fmaxnmv, AddRetType | Add1ArgType), |
| 6424 | NEONMAP1(vpmaxnms_f32, aarch64_neon_fmaxnmv, AddRetType | Add1ArgType), |
| 6425 | NEONMAP1(vpmaxqd_f64, aarch64_neon_fmaxv, AddRetType | Add1ArgType), |
| 6426 | NEONMAP1(vpmaxs_f32, aarch64_neon_fmaxv, AddRetType | Add1ArgType), |
| 6427 | NEONMAP1(vpminnmqd_f64, aarch64_neon_fminnmv, AddRetType | Add1ArgType), |
| 6428 | NEONMAP1(vpminnms_f32, aarch64_neon_fminnmv, AddRetType | Add1ArgType), |
| 6429 | NEONMAP1(vpminqd_f64, aarch64_neon_fminv, AddRetType | Add1ArgType), |
| 6430 | NEONMAP1(vpmins_f32, aarch64_neon_fminv, AddRetType | Add1ArgType), |
| 6431 | NEONMAP1(vqabsb_s8, aarch64_neon_sqabs, Vectorize1ArgType | Use64BitVectors), |
| 6432 | NEONMAP1(vqabsd_s64, aarch64_neon_sqabs, Add1ArgType), |
| 6433 | NEONMAP1(vqabsh_s16, aarch64_neon_sqabs, Vectorize1ArgType | Use64BitVectors), |
| 6434 | NEONMAP1(vqabss_s32, aarch64_neon_sqabs, Add1ArgType), |
| 6435 | NEONMAP1(vqaddb_s8, aarch64_neon_sqadd, Vectorize1ArgType | Use64BitVectors), |
| 6436 | NEONMAP1(vqaddb_u8, aarch64_neon_uqadd, Vectorize1ArgType | Use64BitVectors), |
| 6437 | NEONMAP1(vqaddd_s64, aarch64_neon_sqadd, Add1ArgType), |
| 6438 | NEONMAP1(vqaddd_u64, aarch64_neon_uqadd, Add1ArgType), |
| 6439 | NEONMAP1(vqaddh_s16, aarch64_neon_sqadd, Vectorize1ArgType | Use64BitVectors), |
| 6440 | NEONMAP1(vqaddh_u16, aarch64_neon_uqadd, Vectorize1ArgType | Use64BitVectors), |
| 6441 | NEONMAP1(vqadds_s32, aarch64_neon_sqadd, Add1ArgType), |
| 6442 | NEONMAP1(vqadds_u32, aarch64_neon_uqadd, Add1ArgType), |
| 6443 | NEONMAP1(vqdmulhh_s16, aarch64_neon_sqdmulh, Vectorize1ArgType | Use64BitVectors), |
| 6444 | NEONMAP1(vqdmulhs_s32, aarch64_neon_sqdmulh, Add1ArgType), |
| 6445 | NEONMAP1(vqdmullh_s16, aarch64_neon_sqdmull, VectorRet | Use128BitVectors), |
| 6446 | NEONMAP1(vqdmulls_s32, aarch64_neon_sqdmulls_scalar, 0), |
| 6447 | NEONMAP1(vqmovnd_s64, aarch64_neon_scalar_sqxtn, AddRetType | Add1ArgType), |
| 6448 | NEONMAP1(vqmovnd_u64, aarch64_neon_scalar_uqxtn, AddRetType | Add1ArgType), |
| 6449 | NEONMAP1(vqmovnh_s16, aarch64_neon_sqxtn, VectorRet | Use64BitVectors), |
| 6450 | NEONMAP1(vqmovnh_u16, aarch64_neon_uqxtn, VectorRet | Use64BitVectors), |
| 6451 | NEONMAP1(vqmovns_s32, aarch64_neon_sqxtn, VectorRet | Use64BitVectors), |
| 6452 | NEONMAP1(vqmovns_u32, aarch64_neon_uqxtn, VectorRet | Use64BitVectors), |
| 6453 | NEONMAP1(vqmovund_s64, aarch64_neon_scalar_sqxtun, AddRetType | Add1ArgType), |
| 6454 | NEONMAP1(vqmovunh_s16, aarch64_neon_sqxtun, VectorRet | Use64BitVectors), |
| 6455 | NEONMAP1(vqmovuns_s32, aarch64_neon_sqxtun, VectorRet | Use64BitVectors), |
| 6456 | NEONMAP1(vqnegb_s8, aarch64_neon_sqneg, Vectorize1ArgType | Use64BitVectors), |
| 6457 | NEONMAP1(vqnegd_s64, aarch64_neon_sqneg, Add1ArgType), |
| 6458 | NEONMAP1(vqnegh_s16, aarch64_neon_sqneg, Vectorize1ArgType | Use64BitVectors), |
| 6459 | NEONMAP1(vqnegs_s32, aarch64_neon_sqneg, Add1ArgType), |
| 6460 | NEONMAP1(vqrdmlahh_s16, aarch64_neon_sqrdmlah, Vectorize1ArgType | Use64BitVectors), |
| 6461 | NEONMAP1(vqrdmlahs_s32, aarch64_neon_sqrdmlah, Add1ArgType), |
| 6462 | NEONMAP1(vqrdmlshh_s16, aarch64_neon_sqrdmlsh, Vectorize1ArgType | Use64BitVectors), |
| 6463 | NEONMAP1(vqrdmlshs_s32, aarch64_neon_sqrdmlsh, Add1ArgType), |
| 6464 | NEONMAP1(vqrdmulhh_s16, aarch64_neon_sqrdmulh, Vectorize1ArgType | Use64BitVectors), |
| 6465 | NEONMAP1(vqrdmulhs_s32, aarch64_neon_sqrdmulh, Add1ArgType), |
| 6466 | NEONMAP1(vqrshlb_s8, aarch64_neon_sqrshl, Vectorize1ArgType | Use64BitVectors), |
| 6467 | NEONMAP1(vqrshlb_u8, aarch64_neon_uqrshl, Vectorize1ArgType | Use64BitVectors), |
| 6468 | NEONMAP1(vqrshld_s64, aarch64_neon_sqrshl, Add1ArgType), |
| 6469 | NEONMAP1(vqrshld_u64, aarch64_neon_uqrshl, Add1ArgType), |
| 6470 | NEONMAP1(vqrshlh_s16, aarch64_neon_sqrshl, Vectorize1ArgType | Use64BitVectors), |
| 6471 | NEONMAP1(vqrshlh_u16, aarch64_neon_uqrshl, Vectorize1ArgType | Use64BitVectors), |
| 6472 | NEONMAP1(vqrshls_s32, aarch64_neon_sqrshl, Add1ArgType), |
| 6473 | NEONMAP1(vqrshls_u32, aarch64_neon_uqrshl, Add1ArgType), |
| 6474 | NEONMAP1(vqrshrnd_n_s64, aarch64_neon_sqrshrn, AddRetType), |
| 6475 | NEONMAP1(vqrshrnd_n_u64, aarch64_neon_uqrshrn, AddRetType), |
| 6476 | NEONMAP1(vqrshrnh_n_s16, aarch64_neon_sqrshrn, VectorRet | Use64BitVectors), |
| 6477 | NEONMAP1(vqrshrnh_n_u16, aarch64_neon_uqrshrn, VectorRet | Use64BitVectors), |
| 6478 | NEONMAP1(vqrshrns_n_s32, aarch64_neon_sqrshrn, VectorRet | Use64BitVectors), |
| 6479 | NEONMAP1(vqrshrns_n_u32, aarch64_neon_uqrshrn, VectorRet | Use64BitVectors), |
| 6480 | NEONMAP1(vqrshrund_n_s64, aarch64_neon_sqrshrun, AddRetType), |
| 6481 | NEONMAP1(vqrshrunh_n_s16, aarch64_neon_sqrshrun, VectorRet | Use64BitVectors), |
| 6482 | NEONMAP1(vqrshruns_n_s32, aarch64_neon_sqrshrun, VectorRet | Use64BitVectors), |
| 6483 | NEONMAP1(vqshlb_n_s8, aarch64_neon_sqshl, Vectorize1ArgType | Use64BitVectors), |
| 6484 | NEONMAP1(vqshlb_n_u8, aarch64_neon_uqshl, Vectorize1ArgType | Use64BitVectors), |
| 6485 | NEONMAP1(vqshlb_s8, aarch64_neon_sqshl, Vectorize1ArgType | Use64BitVectors), |
| 6486 | NEONMAP1(vqshlb_u8, aarch64_neon_uqshl, Vectorize1ArgType | Use64BitVectors), |
| 6487 | NEONMAP1(vqshld_s64, aarch64_neon_sqshl, Add1ArgType), |
| 6488 | NEONMAP1(vqshld_u64, aarch64_neon_uqshl, Add1ArgType), |
| 6489 | NEONMAP1(vqshlh_n_s16, aarch64_neon_sqshl, Vectorize1ArgType | Use64BitVectors), |
| 6490 | NEONMAP1(vqshlh_n_u16, aarch64_neon_uqshl, Vectorize1ArgType | Use64BitVectors), |
| 6491 | NEONMAP1(vqshlh_s16, aarch64_neon_sqshl, Vectorize1ArgType | Use64BitVectors), |
| 6492 | NEONMAP1(vqshlh_u16, aarch64_neon_uqshl, Vectorize1ArgType | Use64BitVectors), |
| 6493 | NEONMAP1(vqshls_n_s32, aarch64_neon_sqshl, Add1ArgType), |
| 6494 | NEONMAP1(vqshls_n_u32, aarch64_neon_uqshl, Add1ArgType), |
| 6495 | NEONMAP1(vqshls_s32, aarch64_neon_sqshl, Add1ArgType), |
| 6496 | NEONMAP1(vqshls_u32, aarch64_neon_uqshl, Add1ArgType), |
| 6497 | NEONMAP1(vqshlub_n_s8, aarch64_neon_sqshlu, Vectorize1ArgType | Use64BitVectors), |
| 6498 | NEONMAP1(vqshluh_n_s16, aarch64_neon_sqshlu, Vectorize1ArgType | Use64BitVectors), |
| 6499 | NEONMAP1(vqshlus_n_s32, aarch64_neon_sqshlu, Add1ArgType), |
| 6500 | NEONMAP1(vqshrnd_n_s64, aarch64_neon_sqshrn, AddRetType), |
| 6501 | NEONMAP1(vqshrnd_n_u64, aarch64_neon_uqshrn, AddRetType), |
| 6502 | NEONMAP1(vqshrnh_n_s16, aarch64_neon_sqshrn, VectorRet | Use64BitVectors), |
| 6503 | NEONMAP1(vqshrnh_n_u16, aarch64_neon_uqshrn, VectorRet | Use64BitVectors), |
| 6504 | NEONMAP1(vqshrns_n_s32, aarch64_neon_sqshrn, VectorRet | Use64BitVectors), |
| 6505 | NEONMAP1(vqshrns_n_u32, aarch64_neon_uqshrn, VectorRet | Use64BitVectors), |
| 6506 | NEONMAP1(vqshrund_n_s64, aarch64_neon_sqshrun, AddRetType), |
| 6507 | NEONMAP1(vqshrunh_n_s16, aarch64_neon_sqshrun, VectorRet | Use64BitVectors), |
| 6508 | NEONMAP1(vqshruns_n_s32, aarch64_neon_sqshrun, VectorRet | Use64BitVectors), |
| 6509 | NEONMAP1(vqsubb_s8, aarch64_neon_sqsub, Vectorize1ArgType | Use64BitVectors), |
| 6510 | NEONMAP1(vqsubb_u8, aarch64_neon_uqsub, Vectorize1ArgType | Use64BitVectors), |
| 6511 | NEONMAP1(vqsubd_s64, aarch64_neon_sqsub, Add1ArgType), |
| 6512 | NEONMAP1(vqsubd_u64, aarch64_neon_uqsub, Add1ArgType), |
| 6513 | NEONMAP1(vqsubh_s16, aarch64_neon_sqsub, Vectorize1ArgType | Use64BitVectors), |
| 6514 | NEONMAP1(vqsubh_u16, aarch64_neon_uqsub, Vectorize1ArgType | Use64BitVectors), |
| 6515 | NEONMAP1(vqsubs_s32, aarch64_neon_sqsub, Add1ArgType), |
| 6516 | NEONMAP1(vqsubs_u32, aarch64_neon_uqsub, Add1ArgType), |
| 6517 | NEONMAP1(vrecped_f64, aarch64_neon_frecpe, Add1ArgType), |
| 6518 | NEONMAP1(vrecpes_f32, aarch64_neon_frecpe, Add1ArgType), |
| 6519 | NEONMAP1(vrecpxd_f64, aarch64_neon_frecpx, Add1ArgType), |
| 6520 | NEONMAP1(vrecpxs_f32, aarch64_neon_frecpx, Add1ArgType), |
| 6521 | NEONMAP1(vrshld_s64, aarch64_neon_srshl, Add1ArgType), |
| 6522 | NEONMAP1(vrshld_u64, aarch64_neon_urshl, Add1ArgType), |
| 6523 | NEONMAP1(vrsqrted_f64, aarch64_neon_frsqrte, Add1ArgType), |
| 6524 | NEONMAP1(vrsqrtes_f32, aarch64_neon_frsqrte, Add1ArgType), |
| 6525 | NEONMAP1(vrsqrtsd_f64, aarch64_neon_frsqrts, Add1ArgType), |
| 6526 | NEONMAP1(vrsqrtss_f32, aarch64_neon_frsqrts, Add1ArgType), |
| 6527 | NEONMAP1(vsha1cq_u32, aarch64_crypto_sha1c, 0), |
| 6528 | NEONMAP1(vsha1h_u32, aarch64_crypto_sha1h, 0), |
| 6529 | NEONMAP1(vsha1mq_u32, aarch64_crypto_sha1m, 0), |
| 6530 | NEONMAP1(vsha1pq_u32, aarch64_crypto_sha1p, 0), |
| 6531 | NEONMAP1(vshld_s64, aarch64_neon_sshl, Add1ArgType), |
| 6532 | NEONMAP1(vshld_u64, aarch64_neon_ushl, Add1ArgType), |
| 6533 | NEONMAP1(vslid_n_s64, aarch64_neon_vsli, Vectorize1ArgType), |
| 6534 | NEONMAP1(vslid_n_u64, aarch64_neon_vsli, Vectorize1ArgType), |
| 6535 | NEONMAP1(vsqaddb_u8, aarch64_neon_usqadd, Vectorize1ArgType | Use64BitVectors), |
| 6536 | NEONMAP1(vsqaddd_u64, aarch64_neon_usqadd, Add1ArgType), |
| 6537 | NEONMAP1(vsqaddh_u16, aarch64_neon_usqadd, Vectorize1ArgType | Use64BitVectors), |
| 6538 | NEONMAP1(vsqadds_u32, aarch64_neon_usqadd, Add1ArgType), |
| 6539 | NEONMAP1(vsrid_n_s64, aarch64_neon_vsri, Vectorize1ArgType), |
| 6540 | NEONMAP1(vsrid_n_u64, aarch64_neon_vsri, Vectorize1ArgType), |
| 6541 | NEONMAP1(vuqaddb_s8, aarch64_neon_suqadd, Vectorize1ArgType | Use64BitVectors), |
| 6542 | NEONMAP1(vuqaddd_s64, aarch64_neon_suqadd, Add1ArgType), |
| 6543 | NEONMAP1(vuqaddh_s16, aarch64_neon_suqadd, Vectorize1ArgType | Use64BitVectors), |
| 6544 | NEONMAP1(vuqadds_s32, aarch64_neon_suqadd, Add1ArgType), |
| 6545 | // FP16 scalar intrinisics go here. |
| 6546 | NEONMAP1(vabdh_f16, aarch64_sisd_fabd, Add1ArgType), |
| 6547 | NEONMAP1(vcvtah_s32_f16, aarch64_neon_fcvtas, AddRetType | Add1ArgType), |
| 6548 | NEONMAP1(vcvtah_s64_f16, aarch64_neon_fcvtas, AddRetType | Add1ArgType), |
| 6549 | NEONMAP1(vcvtah_u32_f16, aarch64_neon_fcvtau, AddRetType | Add1ArgType), |
| 6550 | NEONMAP1(vcvtah_u64_f16, aarch64_neon_fcvtau, AddRetType | Add1ArgType), |
| 6551 | NEONMAP1(vcvth_n_f16_s32, aarch64_neon_vcvtfxs2fp, AddRetType | Add1ArgType), |
| 6552 | NEONMAP1(vcvth_n_f16_s64, aarch64_neon_vcvtfxs2fp, AddRetType | Add1ArgType), |
| 6553 | NEONMAP1(vcvth_n_f16_u32, aarch64_neon_vcvtfxu2fp, AddRetType | Add1ArgType), |
| 6554 | NEONMAP1(vcvth_n_f16_u64, aarch64_neon_vcvtfxu2fp, AddRetType | Add1ArgType), |
| 6555 | NEONMAP1(vcvth_n_s32_f16, aarch64_neon_vcvtfp2fxs, AddRetType | Add1ArgType), |
| 6556 | NEONMAP1(vcvth_n_s64_f16, aarch64_neon_vcvtfp2fxs, AddRetType | Add1ArgType), |
| 6557 | NEONMAP1(vcvth_n_u32_f16, aarch64_neon_vcvtfp2fxu, AddRetType | Add1ArgType), |
| 6558 | NEONMAP1(vcvth_n_u64_f16, aarch64_neon_vcvtfp2fxu, AddRetType | Add1ArgType), |
| 6559 | NEONMAP1(vcvth_s32_f16, aarch64_neon_fcvtzs, AddRetType | Add1ArgType), |
| 6560 | NEONMAP1(vcvth_s64_f16, aarch64_neon_fcvtzs, AddRetType | Add1ArgType), |
| 6561 | NEONMAP1(vcvth_u32_f16, aarch64_neon_fcvtzu, AddRetType | Add1ArgType), |
| 6562 | NEONMAP1(vcvth_u64_f16, aarch64_neon_fcvtzu, AddRetType | Add1ArgType), |
| 6563 | NEONMAP1(vcvtmh_s32_f16, aarch64_neon_fcvtms, AddRetType | Add1ArgType), |
| 6564 | NEONMAP1(vcvtmh_s64_f16, aarch64_neon_fcvtms, AddRetType | Add1ArgType), |
| 6565 | NEONMAP1(vcvtmh_u32_f16, aarch64_neon_fcvtmu, AddRetType | Add1ArgType), |
| 6566 | NEONMAP1(vcvtmh_u64_f16, aarch64_neon_fcvtmu, AddRetType | Add1ArgType), |
| 6567 | NEONMAP1(vcvtnh_s32_f16, aarch64_neon_fcvtns, AddRetType | Add1ArgType), |
| 6568 | NEONMAP1(vcvtnh_s64_f16, aarch64_neon_fcvtns, AddRetType | Add1ArgType), |
| 6569 | NEONMAP1(vcvtnh_u32_f16, aarch64_neon_fcvtnu, AddRetType | Add1ArgType), |
| 6570 | NEONMAP1(vcvtnh_u64_f16, aarch64_neon_fcvtnu, AddRetType | Add1ArgType), |
| 6571 | NEONMAP1(vcvtph_s32_f16, aarch64_neon_fcvtps, AddRetType | Add1ArgType), |
| 6572 | NEONMAP1(vcvtph_s64_f16, aarch64_neon_fcvtps, AddRetType | Add1ArgType), |
| 6573 | NEONMAP1(vcvtph_u32_f16, aarch64_neon_fcvtpu, AddRetType | Add1ArgType), |
| 6574 | NEONMAP1(vcvtph_u64_f16, aarch64_neon_fcvtpu, AddRetType | Add1ArgType), |
| 6575 | NEONMAP1(vmulxh_f16, aarch64_neon_fmulx, Add1ArgType), |
| 6576 | NEONMAP1(vrecpeh_f16, aarch64_neon_frecpe, Add1ArgType), |
| 6577 | NEONMAP1(vrecpxh_f16, aarch64_neon_frecpx, Add1ArgType), |
| 6578 | NEONMAP1(vrsqrteh_f16, aarch64_neon_frsqrte, Add1ArgType), |
| 6579 | NEONMAP1(vrsqrtsh_f16, aarch64_neon_frsqrts, Add1ArgType), |
| 6580 | }; |
| 6581 | |
| 6582 | // Some intrinsics are equivalent for codegen. |
| 6583 | static const std::pair<unsigned, unsigned> NEONEquivalentIntrinsicMap[] = { |
| 6584 | { NEON::BI__builtin_neon_splat_lane_bf16, NEON::BI__builtin_neon_splat_lane_v, }, |
| 6585 | { NEON::BI__builtin_neon_splat_laneq_bf16, NEON::BI__builtin_neon_splat_laneq_v, }, |
| 6586 | { NEON::BI__builtin_neon_splatq_lane_bf16, NEON::BI__builtin_neon_splatq_lane_v, }, |
| 6587 | { NEON::BI__builtin_neon_splatq_laneq_bf16, NEON::BI__builtin_neon_splatq_laneq_v, }, |
| 6588 | { NEON::BI__builtin_neon_vabd_f16, NEON::BI__builtin_neon_vabd_v, }, |
| 6589 | { NEON::BI__builtin_neon_vabdq_f16, NEON::BI__builtin_neon_vabdq_v, }, |
| 6590 | { NEON::BI__builtin_neon_vabs_f16, NEON::BI__builtin_neon_vabs_v, }, |
| 6591 | { NEON::BI__builtin_neon_vabsq_f16, NEON::BI__builtin_neon_vabsq_v, }, |
| 6592 | { NEON::BI__builtin_neon_vbsl_f16, NEON::BI__builtin_neon_vbsl_v, }, |
| 6593 | { NEON::BI__builtin_neon_vbslq_f16, NEON::BI__builtin_neon_vbslq_v, }, |
| 6594 | { NEON::BI__builtin_neon_vcage_f16, NEON::BI__builtin_neon_vcage_v, }, |
| 6595 | { NEON::BI__builtin_neon_vcageq_f16, NEON::BI__builtin_neon_vcageq_v, }, |
| 6596 | { NEON::BI__builtin_neon_vcagt_f16, NEON::BI__builtin_neon_vcagt_v, }, |
| 6597 | { NEON::BI__builtin_neon_vcagtq_f16, NEON::BI__builtin_neon_vcagtq_v, }, |
| 6598 | { NEON::BI__builtin_neon_vcale_f16, NEON::BI__builtin_neon_vcale_v, }, |
| 6599 | { NEON::BI__builtin_neon_vcaleq_f16, NEON::BI__builtin_neon_vcaleq_v, }, |
| 6600 | { NEON::BI__builtin_neon_vcalt_f16, NEON::BI__builtin_neon_vcalt_v, }, |
| 6601 | { NEON::BI__builtin_neon_vcaltq_f16, NEON::BI__builtin_neon_vcaltq_v, }, |
| 6602 | { NEON::BI__builtin_neon_vceqz_f16, NEON::BI__builtin_neon_vceqz_v, }, |
| 6603 | { NEON::BI__builtin_neon_vceqzq_f16, NEON::BI__builtin_neon_vceqzq_v, }, |
| 6604 | { NEON::BI__builtin_neon_vcgez_f16, NEON::BI__builtin_neon_vcgez_v, }, |
| 6605 | { NEON::BI__builtin_neon_vcgezq_f16, NEON::BI__builtin_neon_vcgezq_v, }, |
| 6606 | { NEON::BI__builtin_neon_vcgtz_f16, NEON::BI__builtin_neon_vcgtz_v, }, |
| 6607 | { NEON::BI__builtin_neon_vcgtzq_f16, NEON::BI__builtin_neon_vcgtzq_v, }, |
| 6608 | { NEON::BI__builtin_neon_vclez_f16, NEON::BI__builtin_neon_vclez_v, }, |
| 6609 | { NEON::BI__builtin_neon_vclezq_f16, NEON::BI__builtin_neon_vclezq_v, }, |
| 6610 | { NEON::BI__builtin_neon_vcltz_f16, NEON::BI__builtin_neon_vcltz_v, }, |
| 6611 | { NEON::BI__builtin_neon_vcltzq_f16, NEON::BI__builtin_neon_vcltzq_v, }, |
| 6612 | { NEON::BI__builtin_neon_vext_f16, NEON::BI__builtin_neon_vext_v, }, |
| 6613 | { NEON::BI__builtin_neon_vextq_f16, NEON::BI__builtin_neon_vextq_v, }, |
| 6614 | { NEON::BI__builtin_neon_vfma_f16, NEON::BI__builtin_neon_vfma_v, }, |
| 6615 | { NEON::BI__builtin_neon_vfma_lane_f16, NEON::BI__builtin_neon_vfma_lane_v, }, |
| 6616 | { NEON::BI__builtin_neon_vfma_laneq_f16, NEON::BI__builtin_neon_vfma_laneq_v, }, |
| 6617 | { NEON::BI__builtin_neon_vfmaq_f16, NEON::BI__builtin_neon_vfmaq_v, }, |
| 6618 | { NEON::BI__builtin_neon_vfmaq_lane_f16, NEON::BI__builtin_neon_vfmaq_lane_v, }, |
| 6619 | { NEON::BI__builtin_neon_vfmaq_laneq_f16, NEON::BI__builtin_neon_vfmaq_laneq_v, }, |
| 6620 | { NEON::BI__builtin_neon_vld1_bf16_x2, NEON::BI__builtin_neon_vld1_x2_v }, |
| 6621 | { NEON::BI__builtin_neon_vld1_bf16_x3, NEON::BI__builtin_neon_vld1_x3_v }, |
| 6622 | { NEON::BI__builtin_neon_vld1_bf16_x4, NEON::BI__builtin_neon_vld1_x4_v }, |
| 6623 | { NEON::BI__builtin_neon_vld1_bf16, NEON::BI__builtin_neon_vld1_v }, |
| 6624 | { NEON::BI__builtin_neon_vld1_dup_bf16, NEON::BI__builtin_neon_vld1_dup_v }, |
| 6625 | { NEON::BI__builtin_neon_vld1_lane_bf16, NEON::BI__builtin_neon_vld1_lane_v }, |
| 6626 | { NEON::BI__builtin_neon_vld1q_bf16_x2, NEON::BI__builtin_neon_vld1q_x2_v }, |
| 6627 | { NEON::BI__builtin_neon_vld1q_bf16_x3, NEON::BI__builtin_neon_vld1q_x3_v }, |
| 6628 | { NEON::BI__builtin_neon_vld1q_bf16_x4, NEON::BI__builtin_neon_vld1q_x4_v }, |
| 6629 | { NEON::BI__builtin_neon_vld1q_bf16, NEON::BI__builtin_neon_vld1q_v }, |
| 6630 | { NEON::BI__builtin_neon_vld1q_dup_bf16, NEON::BI__builtin_neon_vld1q_dup_v }, |
| 6631 | { NEON::BI__builtin_neon_vld1q_lane_bf16, NEON::BI__builtin_neon_vld1q_lane_v }, |
| 6632 | { NEON::BI__builtin_neon_vld2_bf16, NEON::BI__builtin_neon_vld2_v }, |
| 6633 | { NEON::BI__builtin_neon_vld2_dup_bf16, NEON::BI__builtin_neon_vld2_dup_v }, |
| 6634 | { NEON::BI__builtin_neon_vld2_lane_bf16, NEON::BI__builtin_neon_vld2_lane_v }, |
| 6635 | { NEON::BI__builtin_neon_vld2q_bf16, NEON::BI__builtin_neon_vld2q_v }, |
| 6636 | { NEON::BI__builtin_neon_vld2q_dup_bf16, NEON::BI__builtin_neon_vld2q_dup_v }, |
| 6637 | { NEON::BI__builtin_neon_vld2q_lane_bf16, NEON::BI__builtin_neon_vld2q_lane_v }, |
| 6638 | { NEON::BI__builtin_neon_vld3_bf16, NEON::BI__builtin_neon_vld3_v }, |
| 6639 | { NEON::BI__builtin_neon_vld3_dup_bf16, NEON::BI__builtin_neon_vld3_dup_v }, |
| 6640 | { NEON::BI__builtin_neon_vld3_lane_bf16, NEON::BI__builtin_neon_vld3_lane_v }, |
| 6641 | { NEON::BI__builtin_neon_vld3q_bf16, NEON::BI__builtin_neon_vld3q_v }, |
| 6642 | { NEON::BI__builtin_neon_vld3q_dup_bf16, NEON::BI__builtin_neon_vld3q_dup_v }, |
| 6643 | { NEON::BI__builtin_neon_vld3q_lane_bf16, NEON::BI__builtin_neon_vld3q_lane_v }, |
| 6644 | { NEON::BI__builtin_neon_vld4_bf16, NEON::BI__builtin_neon_vld4_v }, |
| 6645 | { NEON::BI__builtin_neon_vld4_dup_bf16, NEON::BI__builtin_neon_vld4_dup_v }, |
| 6646 | { NEON::BI__builtin_neon_vld4_lane_bf16, NEON::BI__builtin_neon_vld4_lane_v }, |
| 6647 | { NEON::BI__builtin_neon_vld4q_bf16, NEON::BI__builtin_neon_vld4q_v }, |
| 6648 | { NEON::BI__builtin_neon_vld4q_dup_bf16, NEON::BI__builtin_neon_vld4q_dup_v }, |
| 6649 | { NEON::BI__builtin_neon_vld4q_lane_bf16, NEON::BI__builtin_neon_vld4q_lane_v }, |
| 6650 | { NEON::BI__builtin_neon_vmax_f16, NEON::BI__builtin_neon_vmax_v, }, |
| 6651 | { NEON::BI__builtin_neon_vmaxnm_f16, NEON::BI__builtin_neon_vmaxnm_v, }, |
| 6652 | { NEON::BI__builtin_neon_vmaxnmq_f16, NEON::BI__builtin_neon_vmaxnmq_v, }, |
| 6653 | { NEON::BI__builtin_neon_vmaxq_f16, NEON::BI__builtin_neon_vmaxq_v, }, |
| 6654 | { NEON::BI__builtin_neon_vmin_f16, NEON::BI__builtin_neon_vmin_v, }, |
| 6655 | { NEON::BI__builtin_neon_vminnm_f16, NEON::BI__builtin_neon_vminnm_v, }, |
| 6656 | { NEON::BI__builtin_neon_vminnmq_f16, NEON::BI__builtin_neon_vminnmq_v, }, |
| 6657 | { NEON::BI__builtin_neon_vminq_f16, NEON::BI__builtin_neon_vminq_v, }, |
| 6658 | { NEON::BI__builtin_neon_vmulx_f16, NEON::BI__builtin_neon_vmulx_v, }, |
| 6659 | { NEON::BI__builtin_neon_vmulxq_f16, NEON::BI__builtin_neon_vmulxq_v, }, |
| 6660 | { NEON::BI__builtin_neon_vpadd_f16, NEON::BI__builtin_neon_vpadd_v, }, |
| 6661 | { NEON::BI__builtin_neon_vpaddq_f16, NEON::BI__builtin_neon_vpaddq_v, }, |
| 6662 | { NEON::BI__builtin_neon_vpmax_f16, NEON::BI__builtin_neon_vpmax_v, }, |
| 6663 | { NEON::BI__builtin_neon_vpmaxnm_f16, NEON::BI__builtin_neon_vpmaxnm_v, }, |
| 6664 | { NEON::BI__builtin_neon_vpmaxnmq_f16, NEON::BI__builtin_neon_vpmaxnmq_v, }, |
| 6665 | { NEON::BI__builtin_neon_vpmaxq_f16, NEON::BI__builtin_neon_vpmaxq_v, }, |
| 6666 | { NEON::BI__builtin_neon_vpmin_f16, NEON::BI__builtin_neon_vpmin_v, }, |
| 6667 | { NEON::BI__builtin_neon_vpminnm_f16, NEON::BI__builtin_neon_vpminnm_v, }, |
| 6668 | { NEON::BI__builtin_neon_vpminnmq_f16, NEON::BI__builtin_neon_vpminnmq_v, }, |
| 6669 | { NEON::BI__builtin_neon_vpminq_f16, NEON::BI__builtin_neon_vpminq_v, }, |
| 6670 | { NEON::BI__builtin_neon_vrecpe_f16, NEON::BI__builtin_neon_vrecpe_v, }, |
| 6671 | { NEON::BI__builtin_neon_vrecpeq_f16, NEON::BI__builtin_neon_vrecpeq_v, }, |
| 6672 | { NEON::BI__builtin_neon_vrecps_f16, NEON::BI__builtin_neon_vrecps_v, }, |
| 6673 | { NEON::BI__builtin_neon_vrecpsq_f16, NEON::BI__builtin_neon_vrecpsq_v, }, |
| 6674 | { NEON::BI__builtin_neon_vrnd_f16, NEON::BI__builtin_neon_vrnd_v, }, |
| 6675 | { NEON::BI__builtin_neon_vrnda_f16, NEON::BI__builtin_neon_vrnda_v, }, |
| 6676 | { NEON::BI__builtin_neon_vrndaq_f16, NEON::BI__builtin_neon_vrndaq_v, }, |
| 6677 | { NEON::BI__builtin_neon_vrndi_f16, NEON::BI__builtin_neon_vrndi_v, }, |
| 6678 | { NEON::BI__builtin_neon_vrndiq_f16, NEON::BI__builtin_neon_vrndiq_v, }, |
| 6679 | { NEON::BI__builtin_neon_vrndm_f16, NEON::BI__builtin_neon_vrndm_v, }, |
| 6680 | { NEON::BI__builtin_neon_vrndmq_f16, NEON::BI__builtin_neon_vrndmq_v, }, |
| 6681 | { NEON::BI__builtin_neon_vrndn_f16, NEON::BI__builtin_neon_vrndn_v, }, |
| 6682 | { NEON::BI__builtin_neon_vrndnq_f16, NEON::BI__builtin_neon_vrndnq_v, }, |
| 6683 | { NEON::BI__builtin_neon_vrndp_f16, NEON::BI__builtin_neon_vrndp_v, }, |
| 6684 | { NEON::BI__builtin_neon_vrndpq_f16, NEON::BI__builtin_neon_vrndpq_v, }, |
| 6685 | { NEON::BI__builtin_neon_vrndq_f16, NEON::BI__builtin_neon_vrndq_v, }, |
| 6686 | { NEON::BI__builtin_neon_vrndx_f16, NEON::BI__builtin_neon_vrndx_v, }, |
| 6687 | { NEON::BI__builtin_neon_vrndxq_f16, NEON::BI__builtin_neon_vrndxq_v, }, |
| 6688 | { NEON::BI__builtin_neon_vrsqrte_f16, NEON::BI__builtin_neon_vrsqrte_v, }, |
| 6689 | { NEON::BI__builtin_neon_vrsqrteq_f16, NEON::BI__builtin_neon_vrsqrteq_v, }, |
| 6690 | { NEON::BI__builtin_neon_vrsqrts_f16, NEON::BI__builtin_neon_vrsqrts_v, }, |
| 6691 | { NEON::BI__builtin_neon_vrsqrtsq_f16, NEON::BI__builtin_neon_vrsqrtsq_v, }, |
| 6692 | { NEON::BI__builtin_neon_vsqrt_f16, NEON::BI__builtin_neon_vsqrt_v, }, |
| 6693 | { NEON::BI__builtin_neon_vsqrtq_f16, NEON::BI__builtin_neon_vsqrtq_v, }, |
| 6694 | { NEON::BI__builtin_neon_vst1_bf16_x2, NEON::BI__builtin_neon_vst1_x2_v }, |
| 6695 | { NEON::BI__builtin_neon_vst1_bf16_x3, NEON::BI__builtin_neon_vst1_x3_v }, |
| 6696 | { NEON::BI__builtin_neon_vst1_bf16_x4, NEON::BI__builtin_neon_vst1_x4_v }, |
| 6697 | { NEON::BI__builtin_neon_vst1_bf16, NEON::BI__builtin_neon_vst1_v }, |
| 6698 | { NEON::BI__builtin_neon_vst1_lane_bf16, NEON::BI__builtin_neon_vst1_lane_v }, |
| 6699 | { NEON::BI__builtin_neon_vst1q_bf16_x2, NEON::BI__builtin_neon_vst1q_x2_v }, |
| 6700 | { NEON::BI__builtin_neon_vst1q_bf16_x3, NEON::BI__builtin_neon_vst1q_x3_v }, |
| 6701 | { NEON::BI__builtin_neon_vst1q_bf16_x4, NEON::BI__builtin_neon_vst1q_x4_v }, |
| 6702 | { NEON::BI__builtin_neon_vst1q_bf16, NEON::BI__builtin_neon_vst1q_v }, |
| 6703 | { NEON::BI__builtin_neon_vst1q_lane_bf16, NEON::BI__builtin_neon_vst1q_lane_v }, |
| 6704 | { NEON::BI__builtin_neon_vst2_bf16, NEON::BI__builtin_neon_vst2_v }, |
| 6705 | { NEON::BI__builtin_neon_vst2_lane_bf16, NEON::BI__builtin_neon_vst2_lane_v }, |
| 6706 | { NEON::BI__builtin_neon_vst2q_bf16, NEON::BI__builtin_neon_vst2q_v }, |
| 6707 | { NEON::BI__builtin_neon_vst2q_lane_bf16, NEON::BI__builtin_neon_vst2q_lane_v }, |
| 6708 | { NEON::BI__builtin_neon_vst3_bf16, NEON::BI__builtin_neon_vst3_v }, |
| 6709 | { NEON::BI__builtin_neon_vst3_lane_bf16, NEON::BI__builtin_neon_vst3_lane_v }, |
| 6710 | { NEON::BI__builtin_neon_vst3q_bf16, NEON::BI__builtin_neon_vst3q_v }, |
| 6711 | { NEON::BI__builtin_neon_vst3q_lane_bf16, NEON::BI__builtin_neon_vst3q_lane_v }, |
| 6712 | { NEON::BI__builtin_neon_vst4_bf16, NEON::BI__builtin_neon_vst4_v }, |
| 6713 | { NEON::BI__builtin_neon_vst4_lane_bf16, NEON::BI__builtin_neon_vst4_lane_v }, |
| 6714 | { NEON::BI__builtin_neon_vst4q_bf16, NEON::BI__builtin_neon_vst4q_v }, |
| 6715 | { NEON::BI__builtin_neon_vst4q_lane_bf16, NEON::BI__builtin_neon_vst4q_lane_v }, |
| 6716 | { NEON::BI__builtin_neon_vtrn_f16, NEON::BI__builtin_neon_vtrn_v, }, |
| 6717 | { NEON::BI__builtin_neon_vtrnq_f16, NEON::BI__builtin_neon_vtrnq_v, }, |
| 6718 | { NEON::BI__builtin_neon_vuzp_f16, NEON::BI__builtin_neon_vuzp_v, }, |
| 6719 | { NEON::BI__builtin_neon_vuzpq_f16, NEON::BI__builtin_neon_vuzpq_v, }, |
| 6720 | { NEON::BI__builtin_neon_vzip_f16, NEON::BI__builtin_neon_vzip_v, }, |
| 6721 | { NEON::BI__builtin_neon_vzipq_f16, NEON::BI__builtin_neon_vzipq_v, }, |
| 6722 | }; |
| 6723 | |
| 6724 | #undef NEONMAP0 |
| 6725 | #undef NEONMAP1 |
| 6726 | #undef NEONMAP2 |
| 6727 | |
| 6728 | #define SVEMAP1(NameBase, LLVMIntrinsic, TypeModifier) \ |
| 6729 | { \ |
| 6730 | #NameBase, SVE::BI__builtin_sve_##NameBase, Intrinsic::LLVMIntrinsic, 0, \ |
| 6731 | TypeModifier \ |
| 6732 | } |
| 6733 | |
| 6734 | #define SVEMAP2(NameBase, TypeModifier) \ |
| 6735 | { #NameBase, SVE::BI__builtin_sve_##NameBase, 0, 0, TypeModifier } |
| 6736 | static const ARMVectorIntrinsicInfo AArch64SVEIntrinsicMap[] = { |
| 6737 | #define GET_SVE_LLVM_INTRINSIC_MAP |
| 6738 | #include "clang/Basic/arm_sve_builtin_cg.inc" |
| 6739 | #include "clang/Basic/BuiltinsAArch64NeonSVEBridge_cg.def" |
| 6740 | #undef GET_SVE_LLVM_INTRINSIC_MAP |
| 6741 | }; |
| 6742 | |
| 6743 | #undef SVEMAP1 |
| 6744 | #undef SVEMAP2 |
| 6745 | |
| 6746 | static bool NEONSIMDIntrinsicsProvenSorted = false; |
| 6747 | |
| 6748 | static bool AArch64SIMDIntrinsicsProvenSorted = false; |
| 6749 | static bool AArch64SISDIntrinsicsProvenSorted = false; |
| 6750 | static bool AArch64SVEIntrinsicsProvenSorted = false; |
| 6751 | |
| 6752 | static const ARMVectorIntrinsicInfo * |
| 6753 | findARMVectorIntrinsicInMap(ArrayRef<ARMVectorIntrinsicInfo> IntrinsicMap, |
| 6754 | unsigned BuiltinID, bool &MapProvenSorted) { |
| 6755 | |
| 6756 | #ifndef NDEBUG |
| 6757 | if (!MapProvenSorted) { |
| 6758 | assert(llvm::is_sorted(IntrinsicMap))(static_cast <bool> (llvm::is_sorted(IntrinsicMap)) ? void (0) : __assert_fail ("llvm::is_sorted(IntrinsicMap)", "clang/lib/CodeGen/CGBuiltin.cpp" , 6758, __extension__ __PRETTY_FUNCTION__)); |
| 6759 | MapProvenSorted = true; |
| 6760 | } |
| 6761 | #endif |
| 6762 | |
| 6763 | const ARMVectorIntrinsicInfo *Builtin = |
| 6764 | llvm::lower_bound(IntrinsicMap, BuiltinID); |
| 6765 | |
| 6766 | if (Builtin != IntrinsicMap.end() && Builtin->BuiltinID == BuiltinID) |
| 6767 | return Builtin; |
| 6768 | |
| 6769 | return nullptr; |
| 6770 | } |
| 6771 | |
| 6772 | Function *CodeGenFunction::LookupNeonLLVMIntrinsic(unsigned IntrinsicID, |
| 6773 | unsigned Modifier, |
| 6774 | llvm::Type *ArgType, |
| 6775 | const CallExpr *E) { |
| 6776 | int VectorSize = 0; |
| 6777 | if (Modifier & Use64BitVectors) |
| 6778 | VectorSize = 64; |
| 6779 | else if (Modifier & Use128BitVectors) |
| 6780 | VectorSize = 128; |
| 6781 | |
| 6782 | // Return type. |
| 6783 | SmallVector<llvm::Type *, 3> Tys; |
| 6784 | if (Modifier & AddRetType) { |
| 6785 | llvm::Type *Ty = ConvertType(E->getCallReturnType(getContext())); |
| 6786 | if (Modifier & VectorizeRetType) |
| 6787 | Ty = llvm::FixedVectorType::get( |
| 6788 | Ty, VectorSize ? VectorSize / Ty->getPrimitiveSizeInBits() : 1); |
| 6789 | |
| 6790 | Tys.push_back(Ty); |
| 6791 | } |
| 6792 | |
| 6793 | // Arguments. |
| 6794 | if (Modifier & VectorizeArgTypes) { |
| 6795 | int Elts = VectorSize ? VectorSize / ArgType->getPrimitiveSizeInBits() : 1; |
| 6796 | ArgType = llvm::FixedVectorType::get(ArgType, Elts); |
| 6797 | } |
| 6798 | |
| 6799 | if (Modifier & (Add1ArgType | Add2ArgTypes)) |
| 6800 | Tys.push_back(ArgType); |
| 6801 | |
| 6802 | if (Modifier & Add2ArgTypes) |
| 6803 | Tys.push_back(ArgType); |
| 6804 | |
| 6805 | if (Modifier & InventFloatType) |
| 6806 | Tys.push_back(FloatTy); |
| 6807 | |
| 6808 | return CGM.getIntrinsic(IntrinsicID, Tys); |
| 6809 | } |
| 6810 | |
| 6811 | static Value *EmitCommonNeonSISDBuiltinExpr( |
| 6812 | CodeGenFunction &CGF, const ARMVectorIntrinsicInfo &SISDInfo, |
| 6813 | SmallVectorImpl<Value *> &Ops, const CallExpr *E) { |
| 6814 | unsigned BuiltinID = SISDInfo.BuiltinID; |
| 6815 | unsigned int Int = SISDInfo.LLVMIntrinsic; |
| 6816 | unsigned Modifier = SISDInfo.TypeModifier; |
| 6817 | const char *s = SISDInfo.NameHint; |
| 6818 | |
| 6819 | switch (BuiltinID) { |
| 6820 | case NEON::BI__builtin_neon_vcled_s64: |
| 6821 | case NEON::BI__builtin_neon_vcled_u64: |
| 6822 | case NEON::BI__builtin_neon_vcles_f32: |
| 6823 | case NEON::BI__builtin_neon_vcled_f64: |
| 6824 | case NEON::BI__builtin_neon_vcltd_s64: |
| 6825 | case NEON::BI__builtin_neon_vcltd_u64: |
| 6826 | case NEON::BI__builtin_neon_vclts_f32: |
| 6827 | case NEON::BI__builtin_neon_vcltd_f64: |
| 6828 | case NEON::BI__builtin_neon_vcales_f32: |
| 6829 | case NEON::BI__builtin_neon_vcaled_f64: |
| 6830 | case NEON::BI__builtin_neon_vcalts_f32: |
| 6831 | case NEON::BI__builtin_neon_vcaltd_f64: |
| 6832 | // Only one direction of comparisons actually exist, cmle is actually a cmge |
| 6833 | // with swapped operands. The table gives us the right intrinsic but we |
| 6834 | // still need to do the swap. |
| 6835 | std::swap(Ops[0], Ops[1]); |
| 6836 | break; |
| 6837 | } |
| 6838 | |
| 6839 | assert(Int && "Generic code assumes a valid intrinsic")(static_cast <bool> (Int && "Generic code assumes a valid intrinsic" ) ? void (0) : __assert_fail ("Int && \"Generic code assumes a valid intrinsic\"" , "clang/lib/CodeGen/CGBuiltin.cpp", 6839, __extension__ __PRETTY_FUNCTION__ )); |
| 6840 | |
| 6841 | // Determine the type(s) of this overloaded AArch64 intrinsic. |
| 6842 | const Expr *Arg = E->getArg(0); |
| 6843 | llvm::Type *ArgTy = CGF.ConvertType(Arg->getType()); |
| 6844 | Function *F = CGF.LookupNeonLLVMIntrinsic(Int, Modifier, ArgTy, E); |
| 6845 | |
| 6846 | int j = 0; |
| 6847 | ConstantInt *C0 = ConstantInt::get(CGF.SizeTy, 0); |
| 6848 | for (Function::const_arg_iterator ai = F->arg_begin(), ae = F->arg_end(); |
| 6849 | ai != ae; ++ai, ++j) { |
| 6850 | llvm::Type *ArgTy = ai->getType(); |
| 6851 | if (Ops[j]->getType()->getPrimitiveSizeInBits() == |
| 6852 | ArgTy->getPrimitiveSizeInBits()) |
| 6853 | continue; |
| 6854 | |
| 6855 | assert(ArgTy->isVectorTy() && !Ops[j]->getType()->isVectorTy())(static_cast <bool> (ArgTy->isVectorTy() && ! Ops[j]->getType()->isVectorTy()) ? void (0) : __assert_fail ("ArgTy->isVectorTy() && !Ops[j]->getType()->isVectorTy()" , "clang/lib/CodeGen/CGBuiltin.cpp", 6855, __extension__ __PRETTY_FUNCTION__ )); |
| 6856 | // The constant argument to an _n_ intrinsic always has Int32Ty, so truncate |
| 6857 | // it before inserting. |
| 6858 | Ops[j] = CGF.Builder.CreateTruncOrBitCast( |
| 6859 | Ops[j], cast<llvm::VectorType>(ArgTy)->getElementType()); |
| 6860 | Ops[j] = |
| 6861 | CGF.Builder.CreateInsertElement(PoisonValue::get(ArgTy), Ops[j], C0); |
| 6862 | } |
| 6863 | |
| 6864 | Value *Result = CGF.EmitNeonCall(F, Ops, s); |
| 6865 | llvm::Type *ResultType = CGF.ConvertType(E->getType()); |
| 6866 | if (ResultType->getPrimitiveSizeInBits().getFixedValue() < |
| 6867 | Result->getType()->getPrimitiveSizeInBits().getFixedValue()) |
| 6868 | return CGF.Builder.CreateExtractElement(Result, C0); |
| 6869 | |
| 6870 | return CGF.Builder.CreateBitCast(Result, ResultType, s); |
| 6871 | } |
| 6872 | |
| 6873 | Value *CodeGenFunction::EmitCommonNeonBuiltinExpr( |
| 6874 | unsigned BuiltinID, unsigned LLVMIntrinsic, unsigned AltLLVMIntrinsic, |
| 6875 | const char *NameHint, unsigned Modifier, const CallExpr *E, |
| 6876 | SmallVectorImpl<llvm::Value *> &Ops, Address PtrOp0, Address PtrOp1, |
| 6877 | llvm::Triple::ArchType Arch) { |
| 6878 | // Get the last argument, which specifies the vector type. |
| 6879 | const Expr *Arg = E->getArg(E->getNumArgs() - 1); |
| 6880 | std::optional<llvm::APSInt> NeonTypeConst = |
| 6881 | Arg->getIntegerConstantExpr(getContext()); |
| 6882 | if (!NeonTypeConst) |
| 6883 | return nullptr; |
| 6884 | |
| 6885 | // Determine the type of this overloaded NEON intrinsic. |
| 6886 | NeonTypeFlags Type(NeonTypeConst->getZExtValue()); |
| 6887 | bool Usgn = Type.isUnsigned(); |
| 6888 | bool Quad = Type.isQuad(); |
| 6889 | const bool HasLegalHalfType = getTarget().hasLegalHalfType(); |
| 6890 | const bool AllowBFloatArgsAndRet = |
| 6891 | getTargetHooks().getABIInfo().allowBFloatArgsAndRet(); |
| 6892 | |
| 6893 | llvm::FixedVectorType *VTy = |
| 6894 | GetNeonType(this, Type, HasLegalHalfType, false, AllowBFloatArgsAndRet); |
| 6895 | llvm::Type *Ty = VTy; |
| 6896 | if (!Ty) |
| 6897 | return nullptr; |
| 6898 | |
| 6899 | auto getAlignmentValue32 = [&](Address addr) -> Value* { |
| 6900 | return Builder.getInt32(addr.getAlignment().getQuantity()); |
| 6901 | }; |
| 6902 | |
| 6903 | unsigned Int = LLVMIntrinsic; |
| 6904 | if ((Modifier & UnsignedAlts) && !Usgn) |
| 6905 | Int = AltLLVMIntrinsic; |
| 6906 | |
| 6907 | switch (BuiltinID) { |
| 6908 | default: break; |
| 6909 | case NEON::BI__builtin_neon_splat_lane_v: |
| 6910 | case NEON::BI__builtin_neon_splat_laneq_v: |
| 6911 | case NEON::BI__builtin_neon_splatq_lane_v: |
| 6912 | case NEON::BI__builtin_neon_splatq_laneq_v: { |
| 6913 | auto NumElements = VTy->getElementCount(); |
| 6914 | if (BuiltinID == NEON::BI__builtin_neon_splatq_lane_v) |
| 6915 | NumElements = NumElements * 2; |
| 6916 | if (BuiltinID == NEON::BI__builtin_neon_splat_laneq_v) |
| 6917 | NumElements = NumElements.divideCoefficientBy(2); |
| 6918 | |
| 6919 | Ops[0] = Builder.CreateBitCast(Ops[0], VTy); |
| 6920 | return EmitNeonSplat(Ops[0], cast<ConstantInt>(Ops[1]), NumElements); |
| 6921 | } |
| 6922 | case NEON::BI__builtin_neon_vpadd_v: |
| 6923 | case NEON::BI__builtin_neon_vpaddq_v: |
| 6924 | // We don't allow fp/int overloading of intrinsics. |
| 6925 | if (VTy->getElementType()->isFloatingPointTy() && |
| 6926 | Int == Intrinsic::aarch64_neon_addp) |
| 6927 | Int = Intrinsic::aarch64_neon_faddp; |
| 6928 | break; |
| 6929 | case NEON::BI__builtin_neon_vabs_v: |
| 6930 | case NEON::BI__builtin_neon_vabsq_v: |
| 6931 | if (VTy->getElementType()->isFloatingPointTy()) |
| 6932 | return EmitNeonCall(CGM.getIntrinsic(Intrinsic::fabs, Ty), Ops, "vabs"); |
| 6933 | return EmitNeonCall(CGM.getIntrinsic(LLVMIntrinsic, Ty), Ops, "vabs"); |
| 6934 | case NEON::BI__builtin_neon_vadd_v: |
| 6935 | case NEON::BI__builtin_neon_vaddq_v: { |
| 6936 | llvm::Type *VTy = llvm::FixedVectorType::get(Int8Ty, Quad ? 16 : 8); |
| 6937 | Ops[0] = Builder.CreateBitCast(Ops[0], VTy); |
| 6938 | Ops[1] = Builder.CreateBitCast(Ops[1], VTy); |
| 6939 | Ops[0] = Builder.CreateXor(Ops[0], Ops[1]); |
| 6940 | return Builder.CreateBitCast(Ops[0], Ty); |
| 6941 | } |
| 6942 | case NEON::BI__builtin_neon_vaddhn_v: { |
| 6943 | llvm::FixedVectorType *SrcTy = |
| 6944 | llvm::FixedVectorType::getExtendedElementVectorType(VTy); |
| 6945 | |
| 6946 | // %sum = add <4 x i32> %lhs, %rhs |
| 6947 | Ops[0] = Builder.CreateBitCast(Ops[0], SrcTy); |
| 6948 | Ops[1] = Builder.CreateBitCast(Ops[1], SrcTy); |
| 6949 | Ops[0] = Builder.CreateAdd(Ops[0], Ops[1], "vaddhn"); |
| 6950 | |
| 6951 | // %high = lshr <4 x i32> %sum, <i32 16, i32 16, i32 16, i32 16> |
| 6952 | Constant *ShiftAmt = |
| 6953 | ConstantInt::get(SrcTy, SrcTy->getScalarSizeInBits() / 2); |
| 6954 | Ops[0] = Builder.CreateLShr(Ops[0], ShiftAmt, "vaddhn"); |
| 6955 | |
| 6956 | // %res = trunc <4 x i32> %high to <4 x i16> |
| 6957 | return Builder.CreateTrunc(Ops[0], VTy, "vaddhn"); |
| 6958 | } |
| 6959 | case NEON::BI__builtin_neon_vcale_v: |
| 6960 | case NEON::BI__builtin_neon_vcaleq_v: |
| 6961 | case NEON::BI__builtin_neon_vcalt_v: |
| 6962 | case NEON::BI__builtin_neon_vcaltq_v: |
| 6963 | std::swap(Ops[0], Ops[1]); |
| 6964 | [[fallthrough]]; |
| 6965 | case NEON::BI__builtin_neon_vcage_v: |
| 6966 | case NEON::BI__builtin_neon_vcageq_v: |
| 6967 | case NEON::BI__builtin_neon_vcagt_v: |
| 6968 | case NEON::BI__builtin_neon_vcagtq_v: { |
| 6969 | llvm::Type *Ty; |
| 6970 | switch (VTy->getScalarSizeInBits()) { |
| 6971 | default: llvm_unreachable("unexpected type")::llvm::llvm_unreachable_internal("unexpected type", "clang/lib/CodeGen/CGBuiltin.cpp" , 6971); |
| 6972 | case 32: |
| 6973 | Ty = FloatTy; |
| 6974 | break; |
| 6975 | case 64: |
| 6976 | Ty = DoubleTy; |
| 6977 | break; |
| 6978 | case 16: |
| 6979 | Ty = HalfTy; |
| 6980 | break; |
| 6981 | } |
| 6982 | auto *VecFlt = llvm::FixedVectorType::get(Ty, VTy->getNumElements()); |
| 6983 | llvm::Type *Tys[] = { VTy, VecFlt }; |
| 6984 | Function *F = CGM.getIntrinsic(LLVMIntrinsic, Tys); |
| 6985 | return EmitNeonCall(F, Ops, NameHint); |
| 6986 | } |
| 6987 | case NEON::BI__builtin_neon_vceqz_v: |
| 6988 | case NEON::BI__builtin_neon_vceqzq_v: |
| 6989 | return EmitAArch64CompareBuiltinExpr(Ops[0], Ty, ICmpInst::FCMP_OEQ, |
| 6990 | ICmpInst::ICMP_EQ, "vceqz"); |
| 6991 | case NEON::BI__builtin_neon_vcgez_v: |
| 6992 | case NEON::BI__builtin_neon_vcgezq_v: |
| 6993 | return EmitAArch64CompareBuiltinExpr(Ops[0], Ty, ICmpInst::FCMP_OGE, |
| 6994 | ICmpInst::ICMP_SGE, "vcgez"); |
| 6995 | case NEON::BI__builtin_neon_vclez_v: |
| 6996 | case NEON::BI__builtin_neon_vclezq_v: |
| 6997 | return EmitAArch64CompareBuiltinExpr(Ops[0], Ty, ICmpInst::FCMP_OLE, |
| 6998 | ICmpInst::ICMP_SLE, "vclez"); |
| 6999 | case NEON::BI__builtin_neon_vcgtz_v: |
| 7000 | case NEON::BI__builtin_neon_vcgtzq_v: |
| 7001 | return EmitAArch64CompareBuiltinExpr(Ops[0], Ty, ICmpInst::FCMP_OGT, |
| 7002 | ICmpInst::ICMP_SGT, "vcgtz"); |
| 7003 | case NEON::BI__builtin_neon_vcltz_v: |
| 7004 | case NEON::BI__builtin_neon_vcltzq_v: |
| 7005 | return EmitAArch64CompareBuiltinExpr(Ops[0], Ty, ICmpInst::FCMP_OLT, |
| 7006 | ICmpInst::ICMP_SLT, "vcltz"); |
| 7007 | case NEON::BI__builtin_neon_vclz_v: |
| 7008 | case NEON::BI__builtin_neon_vclzq_v: |
| 7009 | // We generate target-independent intrinsic, which needs a second argument |
| 7010 | // for whether or not clz of zero is undefined; on ARM it isn't. |
| 7011 | Ops.push_back(Builder.getInt1(getTarget().isCLZForZeroUndef())); |
| 7012 | break; |
| 7013 | case NEON::BI__builtin_neon_vcvt_f32_v: |
| 7014 | case NEON::BI__builtin_neon_vcvtq_f32_v: |
| 7015 | Ops[0] = Builder.CreateBitCast(Ops[0], Ty); |
| 7016 | Ty = GetNeonType(this, NeonTypeFlags(NeonTypeFlags::Float32, false, Quad), |
| 7017 | HasLegalHalfType); |
| 7018 | return Usgn ? Builder.CreateUIToFP(Ops[0], Ty, "vcvt") |
| 7019 | : Builder.CreateSIToFP(Ops[0], Ty, "vcvt"); |
| 7020 | case NEON::BI__builtin_neon_vcvt_f16_s16: |
| 7021 | case NEON::BI__builtin_neon_vcvt_f16_u16: |
| 7022 | case NEON::BI__builtin_neon_vcvtq_f16_s16: |
| 7023 | case NEON::BI__builtin_neon_vcvtq_f16_u16: |
| 7024 | Ops[0] = Builder.CreateBitCast(Ops[0], Ty); |
| 7025 | Ty = GetNeonType(this, NeonTypeFlags(NeonTypeFlags::Float16, false, Quad), |
| 7026 | HasLegalHalfType); |
| 7027 | return Usgn ? Builder.CreateUIToFP(Ops[0], Ty, "vcvt") |
| 7028 | : Builder.CreateSIToFP(Ops[0], Ty, "vcvt"); |
| 7029 | case NEON::BI__builtin_neon_vcvt_n_f16_s16: |
| 7030 | case NEON::BI__builtin_neon_vcvt_n_f16_u16: |
| 7031 | case NEON::BI__builtin_neon_vcvtq_n_f16_s16: |
| 7032 | case NEON::BI__builtin_neon_vcvtq_n_f16_u16: { |
| 7033 | llvm::Type *Tys[2] = { GetFloatNeonType(this, Type), Ty }; |
| 7034 | Function *F = CGM.getIntrinsic(Int, Tys); |
| 7035 | return EmitNeonCall(F, Ops, "vcvt_n"); |
| 7036 | } |
| 7037 | case NEON::BI__builtin_neon_vcvt_n_f32_v: |
| 7038 | case NEON::BI__builtin_neon_vcvt_n_f64_v: |
| 7039 | case NEON::BI__builtin_neon_vcvtq_n_f32_v: |
| 7040 | case NEON::BI__builtin_neon_vcvtq_n_f64_v: { |
| 7041 | llvm::Type *Tys[2] = { GetFloatNeonType(this, Type), Ty }; |
| 7042 | Int = Usgn ? LLVMIntrinsic : AltLLVMIntrinsic; |
| 7043 | Function *F = CGM.getIntrinsic(Int, Tys); |
| 7044 | return EmitNeonCall(F, Ops, "vcvt_n"); |
| 7045 | } |
| 7046 | case NEON::BI__builtin_neon_vcvt_n_s16_f16: |
| 7047 | case NEON::BI__builtin_neon_vcvt_n_s32_v: |
| 7048 | case NEON::BI__builtin_neon_vcvt_n_u16_f16: |
| 7049 | case NEON::BI__builtin_neon_vcvt_n_u32_v: |
| 7050 | case NEON::BI__builtin_neon_vcvt_n_s64_v: |
| 7051 | case NEON::BI__builtin_neon_vcvt_n_u64_v: |
| 7052 | case NEON::BI__builtin_neon_vcvtq_n_s16_f16: |
| 7053 | case NEON::BI__builtin_neon_vcvtq_n_s32_v: |
| 7054 | case NEON::BI__builtin_neon_vcvtq_n_u16_f16: |
| 7055 | case NEON::BI__builtin_neon_vcvtq_n_u32_v: |
| 7056 | case NEON::BI__builtin_neon_vcvtq_n_s64_v: |
| 7057 | case NEON::BI__builtin_neon_vcvtq_n_u64_v: { |
| 7058 | llvm::Type *Tys[2] = { Ty, GetFloatNeonType(this, Type) }; |
| 7059 | Function *F = CGM.getIntrinsic(LLVMIntrinsic, Tys); |
| 7060 | return EmitNeonCall(F, Ops, "vcvt_n"); |
| 7061 | } |
| 7062 | case NEON::BI__builtin_neon_vcvt_s32_v: |
| 7063 | case NEON::BI__builtin_neon_vcvt_u32_v: |
| 7064 | case NEON::BI__builtin_neon_vcvt_s64_v: |
| 7065 | case NEON::BI__builtin_neon_vcvt_u64_v: |
| 7066 | case NEON::BI__builtin_neon_vcvt_s16_f16: |
| 7067 | case NEON::BI__builtin_neon_vcvt_u16_f16: |
| 7068 | case NEON::BI__builtin_neon_vcvtq_s32_v: |
| 7069 | case NEON::BI__builtin_neon_vcvtq_u32_v: |
| 7070 | case NEON::BI__builtin_neon_vcvtq_s64_v: |
| 7071 | case NEON::BI__builtin_neon_vcvtq_u64_v: |
| 7072 | case NEON::BI__builtin_neon_vcvtq_s16_f16: |
| 7073 | case NEON::BI__builtin_neon_vcvtq_u16_f16: { |
| 7074 | Ops[0] = Builder.CreateBitCast(Ops[0], GetFloatNeonType(this, Type)); |
| 7075 | return Usgn ? Builder.CreateFPToUI(Ops[0], Ty, "vcvt") |
| 7076 | : Builder.CreateFPToSI(Ops[0], Ty, "vcvt"); |
| 7077 | } |
| 7078 | case NEON::BI__builtin_neon_vcvta_s16_f16: |
| 7079 | case NEON::BI__builtin_neon_vcvta_s32_v: |
| 7080 | case NEON::BI__builtin_neon_vcvta_s64_v: |
| 7081 | case NEON::BI__builtin_neon_vcvta_u16_f16: |
| 7082 | case NEON::BI__builtin_neon_vcvta_u32_v: |
| 7083 | case NEON::BI__builtin_neon_vcvta_u64_v: |
| 7084 | case NEON::BI__builtin_neon_vcvtaq_s16_f16: |
| 7085 | case NEON::BI__builtin_neon_vcvtaq_s32_v: |
| 7086 | case NEON::BI__builtin_neon_vcvtaq_s64_v: |
| 7087 | case NEON::BI__builtin_neon_vcvtaq_u16_f16: |
| 7088 | case NEON::BI__builtin_neon_vcvtaq_u32_v: |
| 7089 | case NEON::BI__builtin_neon_vcvtaq_u64_v: |
| 7090 | case NEON::BI__builtin_neon_vcvtn_s16_f16: |
| 7091 | case NEON::BI__builtin_neon_vcvtn_s32_v: |
| 7092 | case NEON::BI__builtin_neon_vcvtn_s64_v: |
| 7093 | case NEON::BI__builtin_neon_vcvtn_u16_f16: |
| 7094 | case NEON::BI__builtin_neon_vcvtn_u32_v: |
| 7095 | case NEON::BI__builtin_neon_vcvtn_u64_v: |
| 7096 | case NEON::BI__builtin_neon_vcvtnq_s16_f16: |
| 7097 | case NEON::BI__builtin_neon_vcvtnq_s32_v: |
| 7098 | case NEON::BI__builtin_neon_vcvtnq_s64_v: |
| 7099 | case NEON::BI__builtin_neon_vcvtnq_u16_f16: |
| 7100 | case NEON::BI__builtin_neon_vcvtnq_u32_v: |
| 7101 | case NEON::BI__builtin_neon_vcvtnq_u64_v: |
| 7102 | case NEON::BI__builtin_neon_vcvtp_s16_f16: |
| 7103 | case NEON::BI__builtin_neon_vcvtp_s32_v: |
| 7104 | case NEON::BI__builtin_neon_vcvtp_s64_v: |
| 7105 | case NEON::BI__builtin_neon_vcvtp_u16_f16: |
| 7106 | case NEON::BI__builtin_neon_vcvtp_u32_v: |
| 7107 | case NEON::BI__builtin_neon_vcvtp_u64_v: |
| 7108 | case NEON::BI__builtin_neon_vcvtpq_s16_f16: |
| 7109 | case NEON::BI__builtin_neon_vcvtpq_s32_v: |
| 7110 | case NEON::BI__builtin_neon_vcvtpq_s64_v: |
| 7111 | case NEON::BI__builtin_neon_vcvtpq_u16_f16: |
| 7112 | case NEON::BI__builtin_neon_vcvtpq_u32_v: |
| 7113 | case NEON::BI__builtin_neon_vcvtpq_u64_v: |
| 7114 | case NEON::BI__builtin_neon_vcvtm_s16_f16: |
| 7115 | case NEON::BI__builtin_neon_vcvtm_s32_v: |
| 7116 | case NEON::BI__builtin_neon_vcvtm_s64_v: |
| 7117 | case NEON::BI__builtin_neon_vcvtm_u16_f16: |
| 7118 | case NEON::BI__builtin_neon_vcvtm_u32_v: |
| 7119 | case NEON::BI__builtin_neon_vcvtm_u64_v: |
| 7120 | case NEON::BI__builtin_neon_vcvtmq_s16_f16: |
| 7121 | case NEON::BI__builtin_neon_vcvtmq_s32_v: |
| 7122 | case NEON::BI__builtin_neon_vcvtmq_s64_v: |
| 7123 | case NEON::BI__builtin_neon_vcvtmq_u16_f16: |
| 7124 | case NEON::BI__builtin_neon_vcvtmq_u32_v: |
| 7125 | case NEON::BI__builtin_neon_vcvtmq_u64_v: { |
| 7126 | llvm::Type *Tys[2] = { Ty, GetFloatNeonType(this, Type) }; |
| 7127 | return EmitNeonCall(CGM.getIntrinsic(LLVMIntrinsic, Tys), Ops, NameHint); |
| 7128 | } |
| 7129 | case NEON::BI__builtin_neon_vcvtx_f32_v: { |
| 7130 | llvm::Type *Tys[2] = { VTy->getTruncatedElementVectorType(VTy), Ty}; |
| 7131 | return EmitNeonCall(CGM.getIntrinsic(LLVMIntrinsic, Tys), Ops, NameHint); |
| 7132 | |
| 7133 | } |
| 7134 | case NEON::BI__builtin_neon_vext_v: |
| 7135 | case NEON::BI__builtin_neon_vextq_v: { |
| 7136 | int CV = cast<ConstantInt>(Ops[2])->getSExtValue(); |
| 7137 | SmallVector<int, 16> Indices; |
| 7138 | for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) |
| 7139 | Indices.push_back(i+CV); |
| 7140 | |
| 7141 | Ops[0] = Builder.CreateBitCast(Ops[0], Ty); |
| 7142 | Ops[1] = Builder.CreateBitCast(Ops[1], Ty); |
| 7143 | return Builder.CreateShuffleVector(Ops[0], Ops[1], Indices, "vext"); |
| 7144 | } |
| 7145 | case NEON::BI__builtin_neon_vfma_v: |
| 7146 | case NEON::BI__builtin_neon_vfmaq_v: { |
| 7147 | Ops[0] = Builder.CreateBitCast(Ops[0], Ty); |
| 7148 | Ops[1] = Builder.CreateBitCast(Ops[1], Ty); |
| 7149 | Ops[2] = Builder.CreateBitCast(Ops[2], Ty); |
| 7150 | |
| 7151 | // NEON intrinsic puts accumulator first, unlike the LLVM fma. |
| 7152 | return emitCallMaybeConstrainedFPBuiltin( |
| 7153 | *this, Intrinsic::fma, Intrinsic::experimental_constrained_fma, Ty, |
| 7154 | {Ops[1], Ops[2], Ops[0]}); |
| 7155 | } |
| 7156 | case NEON::BI__builtin_neon_vld1_v: |
| 7157 | case NEON::BI__builtin_neon_vld1q_v: { |
| 7158 | llvm::Type *Tys[] = {Ty, Int8PtrTy}; |
| 7159 | Ops.push_back(getAlignmentValue32(PtrOp0)); |
| 7160 | return EmitNeonCall(CGM.getIntrinsic(LLVMIntrinsic, Tys), Ops, "vld1"); |
| 7161 | } |
| 7162 | case NEON::BI__builtin_neon_vld1_x2_v: |
| 7163 | case NEON::BI__builtin_neon_vld1q_x2_v: |
| 7164 | case NEON::BI__builtin_neon_vld1_x3_v: |
| 7165 | case NEON::BI__builtin_neon_vld1q_x3_v: |
| 7166 | case NEON::BI__builtin_neon_vld1_x4_v: |
| 7167 | case NEON::BI__builtin_neon_vld1q_x4_v: { |
| 7168 | llvm::Type *PTy = llvm::PointerType::getUnqual(VTy->getElementType()); |
| 7169 | Ops[1] = Builder.CreateBitCast(Ops[1], PTy); |
| 7170 | llvm::Type *Tys[2] = { VTy, PTy }; |
| 7171 | Function *F = CGM.getIntrinsic(LLVMIntrinsic, Tys); |
| 7172 | Ops[1] = Builder.CreateCall(F, Ops[1], "vld1xN"); |
| 7173 | Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); |
| 7174 | Ops[0] = Builder.CreateBitCast(Ops[0], Ty); |
| 7175 | return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]); |
| 7176 | } |
| 7177 | case NEON::BI__builtin_neon_vld2_v: |
| 7178 | case NEON::BI__builtin_neon_vld2q_v: |
| 7179 | case NEON::BI__builtin_neon_vld3_v: |
| 7180 | case NEON::BI__builtin_neon_vld3q_v: |
| 7181 | case NEON::BI__builtin_neon_vld4_v: |
| 7182 | case NEON::BI__builtin_neon_vld4q_v: |
| 7183 | case NEON::BI__builtin_neon_vld2_dup_v: |
| 7184 | case NEON::BI__builtin_neon_vld2q_dup_v: |
| 7185 | case NEON::BI__builtin_neon_vld3_dup_v: |
| 7186 | case NEON::BI__builtin_neon_vld3q_dup_v: |
| 7187 | case NEON::BI__builtin_neon_vld4_dup_v: |
| 7188 | case NEON::BI__builtin_neon_vld4q_dup_v: { |
| 7189 | llvm::Type *Tys[] = {Ty, Int8PtrTy}; |
| 7190 | Function *F = CGM.getIntrinsic(LLVMIntrinsic, Tys); |
| 7191 | Value *Align = getAlignmentValue32(PtrOp1); |
| 7192 | Ops[1] = Builder.CreateCall(F, {Ops[1], Align}, NameHint); |
| 7193 | Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); |
| 7194 | Ops[0] = Builder.CreateBitCast(Ops[0], Ty); |
| 7195 | return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]); |
| 7196 | } |
| 7197 | case NEON::BI__builtin_neon_vld1_dup_v: |
| 7198 | case NEON::BI__builtin_neon_vld1q_dup_v: { |
| 7199 | Value *V = PoisonValue::get(Ty); |
| 7200 | PtrOp0 = Builder.CreateElementBitCast(PtrOp0, VTy->getElementType()); |
| 7201 | LoadInst *Ld = Builder.CreateLoad(PtrOp0); |
| 7202 | llvm::Constant *CI = ConstantInt::get(SizeTy, 0); |
| 7203 | Ops[0] = Builder.CreateInsertElement(V, Ld, CI); |
| 7204 | return EmitNeonSplat(Ops[0], CI); |
| 7205 | } |
| 7206 | case NEON::BI__builtin_neon_vld2_lane_v: |
| 7207 | case NEON::BI__builtin_neon_vld2q_lane_v: |
| 7208 | case NEON::BI__builtin_neon_vld3_lane_v: |
| 7209 | case NEON::BI__builtin_neon_vld3q_lane_v: |
| 7210 | case NEON::BI__builtin_neon_vld4_lane_v: |
| 7211 | case NEON::BI__builtin_neon_vld4q_lane_v: { |
| 7212 | llvm::Type *Tys[] = {Ty, Int8PtrTy}; |
| 7213 | Function *F = CGM.getIntrinsic(LLVMIntrinsic, Tys); |
| 7214 | for (unsigned I = 2; I < Ops.size() - 1; ++I) |
| 7215 | Ops[I] = Builder.CreateBitCast(Ops[I], Ty); |
| 7216 | Ops.push_back(getAlignmentValue32(PtrOp1)); |
| 7217 | Ops[1] = Builder.CreateCall(F, ArrayRef(Ops).slice(1), NameHint); |
| 7218 | Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); |
| 7219 | Ops[0] = Builder.CreateBitCast(Ops[0], Ty); |
| 7220 | return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]); |
| 7221 | } |
| 7222 | case NEON::BI__builtin_neon_vmovl_v: { |
| 7223 | llvm::FixedVectorType *DTy = |
| 7224 | llvm::FixedVectorType::getTruncatedElementVectorType(VTy); |
| 7225 | Ops[0] = Builder.CreateBitCast(Ops[0], DTy); |
| 7226 | if (Usgn) |
| 7227 | return Builder.CreateZExt(Ops[0], Ty, "vmovl"); |
| 7228 | return Builder.CreateSExt(Ops[0], Ty, "vmovl"); |
| 7229 | } |
| 7230 | case NEON::BI__builtin_neon_vmovn_v: { |
| 7231 | llvm::FixedVectorType *QTy = |
| 7232 | llvm::FixedVectorType::getExtendedElementVectorType(VTy); |
| 7233 | Ops[0] = Builder.CreateBitCast(Ops[0], QTy); |
| 7234 | return Builder.CreateTrunc(Ops[0], Ty, "vmovn"); |
| 7235 | } |
| 7236 | case NEON::BI__builtin_neon_vmull_v: |
| 7237 | // FIXME: the integer vmull operations could be emitted in terms of pure |
| 7238 | // LLVM IR (2 exts followed by a mul). Unfortunately LLVM has a habit of |
| 7239 | // hoisting the exts outside loops. Until global ISel comes along that can |
| 7240 | // see through such movement this leads to bad CodeGen. So we need an |
| 7241 | // intrinsic for now. |
| 7242 | Int = Usgn ? Intrinsic::arm_neon_vmullu : Intrinsic::arm_neon_vmulls; |
| 7243 | Int = Type.isPoly() ? (unsigned)Intrinsic::arm_neon_vmullp : Int; |
| 7244 | return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmull"); |
| 7245 | case NEON::BI__builtin_neon_vpadal_v: |
| 7246 | case NEON::BI__builtin_neon_vpadalq_v: { |
| 7247 | // The source operand type has twice as many elements of half the size. |
| 7248 | unsigned EltBits = VTy->getElementType()->getPrimitiveSizeInBits(); |
| 7249 | llvm::Type *EltTy = |
| 7250 | llvm::IntegerType::get(getLLVMContext(), EltBits / 2); |
| 7251 | auto *NarrowTy = |
| 7252 | llvm::FixedVectorType::get(EltTy, VTy->getNumElements() * 2); |
| 7253 | llvm::Type *Tys[2] = { Ty, NarrowTy }; |
| 7254 | return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, NameHint); |
| 7255 | } |
| 7256 | case NEON::BI__builtin_neon_vpaddl_v: |
| 7257 | case NEON::BI__builtin_neon_vpaddlq_v: { |
| 7258 | // The source operand type has twice as many elements of half the size. |
| 7259 | unsigned EltBits = VTy->getElementType()->getPrimitiveSizeInBits(); |
| 7260 | llvm::Type *EltTy = llvm::IntegerType::get(getLLVMContext(), EltBits / 2); |
| 7261 | auto *NarrowTy = |
| 7262 | llvm::FixedVectorType::get(EltTy, VTy->getNumElements() * 2); |
| 7263 | llvm::Type *Tys[2] = { Ty, NarrowTy }; |
| 7264 | return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vpaddl"); |
| 7265 | } |
| 7266 | case NEON::BI__builtin_neon_vqdmlal_v: |
| 7267 | case NEON::BI__builtin_neon_vqdmlsl_v: { |
| 7268 | SmallVector<Value *, 2> MulOps(Ops.begin() + 1, Ops.end()); |
| 7269 | Ops[1] = |
| 7270 | EmitNeonCall(CGM.getIntrinsic(LLVMIntrinsic, Ty), MulOps, "vqdmlal"); |
| 7271 | Ops.resize(2); |
| 7272 | return EmitNeonCall(CGM.getIntrinsic(AltLLVMIntrinsic, Ty), Ops, NameHint); |
| 7273 | } |
| 7274 | case NEON::BI__builtin_neon_vqdmulhq_lane_v: |
| 7275 | case NEON::BI__builtin_neon_vqdmulh_lane_v: |
| 7276 | case NEON::BI__builtin_neon_vqrdmulhq_lane_v: |
| 7277 | case NEON::BI__builtin_neon_vqrdmulh_lane_v: { |
| 7278 | auto *RTy = cast<llvm::FixedVectorType>(Ty); |
| 7279 | if (BuiltinID == NEON::BI__builtin_neon_vqdmulhq_lane_v || |
| 7280 | BuiltinID == NEON::BI__builtin_neon_vqrdmulhq_lane_v) |
| 7281 | RTy = llvm::FixedVectorType::get(RTy->getElementType(), |
| 7282 | RTy->getNumElements() * 2); |
| 7283 | llvm::Type *Tys[2] = { |
| 7284 | RTy, GetNeonType(this, NeonTypeFlags(Type.getEltType(), false, |
| 7285 | /*isQuad*/ false))}; |
| 7286 | return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, NameHint); |
| 7287 | } |
| 7288 | case NEON::BI__builtin_neon_vqdmulhq_laneq_v: |
| 7289 | case NEON::BI__builtin_neon_vqdmulh_laneq_v: |
| 7290 | case NEON::BI__builtin_neon_vqrdmulhq_laneq_v: |
| 7291 | case NEON::BI__builtin_neon_vqrdmulh_laneq_v: { |
| 7292 | llvm::Type *Tys[2] = { |
| 7293 | Ty, GetNeonType(this, NeonTypeFlags(Type.getEltType(), false, |
| 7294 | /*isQuad*/ true))}; |
| 7295 | return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, NameHint); |
| 7296 | } |
| 7297 | case NEON::BI__builtin_neon_vqshl_n_v: |
| 7298 | case NEON::BI__builtin_neon_vqshlq_n_v: |
| 7299 | return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqshl_n", |
| 7300 | 1, false); |
| 7301 | case NEON::BI__builtin_neon_vqshlu_n_v: |
| 7302 | case NEON::BI__builtin_neon_vqshluq_n_v: |
| 7303 | return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqshlu_n", |
| 7304 | 1, false); |
| 7305 | case NEON::BI__builtin_neon_vrecpe_v: |
| 7306 | case NEON::BI__builtin_neon_vrecpeq_v: |
| 7307 | case NEON::BI__builtin_neon_vrsqrte_v: |
| 7308 | case NEON::BI__builtin_neon_vrsqrteq_v: |
| 7309 | Int = Ty->isFPOrFPVectorTy() ? LLVMIntrinsic : AltLLVMIntrinsic; |
| 7310 | return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, NameHint); |
| 7311 | case NEON::BI__builtin_neon_vrndi_v: |
| 7312 | case NEON::BI__builtin_neon_vrndiq_v: |
| 7313 | Int = Builder.getIsFPConstrained() |
| 7314 | ? Intrinsic::experimental_constrained_nearbyint |
| 7315 | : Intrinsic::nearbyint; |
| 7316 | return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, NameHint); |
| 7317 | case NEON::BI__builtin_neon_vrshr_n_v: |
| 7318 | case NEON::BI__builtin_neon_vrshrq_n_v: |
| 7319 | return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrshr_n", |
| 7320 | 1, true); |
| 7321 | case NEON::BI__builtin_neon_vsha512hq_u64: |
| 7322 | case NEON::BI__builtin_neon_vsha512h2q_u64: |
| 7323 | case NEON::BI__builtin_neon_vsha512su0q_u64: |
| 7324 | case NEON::BI__builtin_neon_vsha512su1q_u64: { |
| 7325 | Function *F = CGM.getIntrinsic(Int); |
| 7326 | return EmitNeonCall(F, Ops, ""); |
| 7327 | } |
| 7328 | case NEON::BI__builtin_neon_vshl_n_v: |
| 7329 | case NEON::BI__builtin_neon_vshlq_n_v: |
| 7330 | Ops[1] = EmitNeonShiftVector(Ops[1], Ty, false); |
| 7331 | return Builder.CreateShl(Builder.CreateBitCast(Ops[0],Ty), Ops[1], |
| 7332 | "vshl_n"); |
| 7333 | case NEON::BI__builtin_neon_vshll_n_v: { |
| 7334 | llvm::FixedVectorType *SrcTy = |
| 7335 | llvm::FixedVectorType::getTruncatedElementVectorType(VTy); |
| 7336 | Ops[0] = Builder.CreateBitCast(Ops[0], SrcTy); |
| 7337 | if (Usgn) |
| 7338 | Ops[0] = Builder.CreateZExt(Ops[0], VTy); |
| 7339 | else |
| 7340 | Ops[0] = Builder.CreateSExt(Ops[0], VTy); |
| 7341 | Ops[1] = EmitNeonShiftVector(Ops[1], VTy, false); |
| 7342 | return Builder.CreateShl(Ops[0], Ops[1], "vshll_n"); |
| 7343 | } |
| 7344 | case NEON::BI__builtin_neon_vshrn_n_v: { |
| 7345 | llvm::FixedVectorType *SrcTy = |
| 7346 | llvm::FixedVectorType::getExtendedElementVectorType(VTy); |
| 7347 | Ops[0] = Builder.CreateBitCast(Ops[0], SrcTy); |
| 7348 | Ops[1] = EmitNeonShiftVector(Ops[1], SrcTy, false); |
| 7349 | if (Usgn) |
| 7350 | Ops[0] = Builder.CreateLShr(Ops[0], Ops[1]); |
| 7351 | else |
| 7352 | Ops[0] = Builder.CreateAShr(Ops[0], Ops[1]); |
| 7353 | return Builder.CreateTrunc(Ops[0], Ty, "vshrn_n"); |
| 7354 | } |
| 7355 | case NEON::BI__builtin_neon_vshr_n_v: |
| 7356 | case NEON::BI__builtin_neon_vshrq_n_v: |
| 7357 | return EmitNeonRShiftImm(Ops[0], Ops[1], Ty, Usgn, "vshr_n"); |
| 7358 | case NEON::BI__builtin_neon_vst1_v: |
| 7359 | case NEON::BI__builtin_neon_vst1q_v: |
| 7360 | case NEON::BI__builtin_neon_vst2_v: |
| 7361 | case NEON::BI__builtin_neon_vst2q_v: |
| 7362 | case NEON::BI__builtin_neon_vst3_v: |
| 7363 | case NEON::BI__builtin_neon_vst3q_v: |
| 7364 | case NEON::BI__builtin_neon_vst4_v: |
| 7365 | case NEON::BI__builtin_neon_vst4q_v: |
| 7366 | case NEON::BI__builtin_neon_vst2_lane_v: |
| 7367 | case NEON::BI__builtin_neon_vst2q_lane_v: |
| 7368 | case NEON::BI__builtin_neon_vst3_lane_v: |
| 7369 | case NEON::BI__builtin_neon_vst3q_lane_v: |
| 7370 | case NEON::BI__builtin_neon_vst4_lane_v: |
| 7371 | case NEON::BI__builtin_neon_vst4q_lane_v: { |
| 7372 | llvm::Type *Tys[] = {Int8PtrTy, Ty}; |
| 7373 | Ops.push_back(getAlignmentValue32(PtrOp0)); |
| 7374 | return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, ""); |
| 7375 | } |
| 7376 | case NEON::BI__builtin_neon_vsm3partw1q_u32: |
| 7377 | case NEON::BI__builtin_neon_vsm3partw2q_u32: |
| 7378 | case NEON::BI__builtin_neon_vsm3ss1q_u32: |
| 7379 | case NEON::BI__builtin_neon_vsm4ekeyq_u32: |
| 7380 | case NEON::BI__builtin_neon_vsm4eq_u32: { |
| 7381 | Function *F = CGM.getIntrinsic(Int); |
| 7382 | return EmitNeonCall(F, Ops, ""); |
| 7383 | } |
| 7384 | case NEON::BI__builtin_neon_vsm3tt1aq_u32: |
| 7385 | case NEON::BI__builtin_neon_vsm3tt1bq_u32: |
| 7386 | case NEON::BI__builtin_neon_vsm3tt2aq_u32: |
| 7387 | case NEON::BI__builtin_neon_vsm3tt2bq_u32: { |
| 7388 | Function *F = CGM.getIntrinsic(Int); |
| 7389 | Ops[3] = Builder.CreateZExt(Ops[3], Int64Ty); |
| 7390 | return EmitNeonCall(F, Ops, ""); |
| 7391 | } |
| 7392 | case NEON::BI__builtin_neon_vst1_x2_v: |
| 7393 | case NEON::BI__builtin_neon_vst1q_x2_v: |
| 7394 | case NEON::BI__builtin_neon_vst1_x3_v: |
| 7395 | case NEON::BI__builtin_neon_vst1q_x3_v: |
| 7396 | case NEON::BI__builtin_neon_vst1_x4_v: |
| 7397 | case NEON::BI__builtin_neon_vst1q_x4_v: { |
| 7398 | llvm::Type *PTy = llvm::PointerType::getUnqual(VTy->getElementType()); |
| 7399 | // TODO: Currently in AArch32 mode the pointer operand comes first, whereas |
| 7400 | // in AArch64 it comes last. We may want to stick to one or another. |
| 7401 | if (Arch == llvm::Triple::aarch64 || Arch == llvm::Triple::aarch64_be || |
| 7402 | Arch == llvm::Triple::aarch64_32) { |
| 7403 | llvm::Type *Tys[2] = { VTy, PTy }; |
| 7404 | std::rotate(Ops.begin(), Ops.begin() + 1, Ops.end()); |
| 7405 | return EmitNeonCall(CGM.getIntrinsic(LLVMIntrinsic, Tys), Ops, ""); |
| 7406 | } |
| 7407 | llvm::Type *Tys[2] = { PTy, VTy }; |
| 7408 | return EmitNeonCall(CGM.getIntrinsic(LLVMIntrinsic, Tys), Ops, ""); |
| 7409 | } |
| 7410 | case NEON::BI__builtin_neon_vsubhn_v: { |
| 7411 | llvm::FixedVectorType *SrcTy = |
| 7412 | llvm::FixedVectorType::getExtendedElementVectorType(VTy); |
| 7413 | |
| 7414 | // %sum = add <4 x i32> %lhs, %rhs |
| 7415 | Ops[0] = Builder.CreateBitCast(Ops[0], SrcTy); |
| 7416 | Ops[1] = Builder.CreateBitCast(Ops[1], SrcTy); |
| 7417 | Ops[0] = Builder.CreateSub(Ops[0], Ops[1], "vsubhn"); |
| 7418 | |
| 7419 | // %high = lshr <4 x i32> %sum, <i32 16, i32 16, i32 16, i32 16> |
| 7420 | Constant *ShiftAmt = |
| 7421 | ConstantInt::get(SrcTy, SrcTy->getScalarSizeInBits() / 2); |
| 7422 | Ops[0] = Builder.CreateLShr(Ops[0], ShiftAmt, "vsubhn"); |
| 7423 | |
| 7424 | // %res = trunc <4 x i32> %high to <4 x i16> |
| 7425 | return Builder.CreateTrunc(Ops[0], VTy, "vsubhn"); |
| 7426 | } |
| 7427 | case NEON::BI__builtin_neon_vtrn_v: |
| 7428 | case NEON::BI__builtin_neon_vtrnq_v: { |
| 7429 | Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty)); |
| 7430 | Ops[1] = Builder.CreateBitCast(Ops[1], Ty); |
| 7431 | Ops[2] = Builder.CreateBitCast(Ops[2], Ty); |
| 7432 | Value *SV = nullptr; |
| 7433 | |
| 7434 | for (unsigned vi = 0; vi != 2; ++vi) { |
| 7435 | SmallVector<int, 16> Indices; |
| 7436 | for (unsigned i = 0, e = VTy->getNumElements(); i != e; i += 2) { |
| 7437 | Indices.push_back(i+vi); |
| 7438 | Indices.push_back(i+e+vi); |
| 7439 | } |
| 7440 | Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ty, Ops[0], vi); |
| 7441 | SV = Builder.CreateShuffleVector(Ops[1], Ops[2], Indices, "vtrn"); |
| 7442 | SV = Builder.CreateDefaultAlignedStore(SV, Addr); |
| 7443 | } |
| 7444 | return SV; |
| 7445 | } |
| 7446 | case NEON::BI__builtin_neon_vtst_v: |
| 7447 | case NEON::BI__builtin_neon_vtstq_v: { |
| 7448 | Ops[0] = Builder.CreateBitCast(Ops[0], Ty); |
| 7449 | Ops[1] = Builder.CreateBitCast(Ops[1], Ty); |
| 7450 | Ops[0] = Builder.CreateAnd(Ops[0], Ops[1]); |
| 7451 | Ops[0] = Builder.CreateICmp(ICmpInst::ICMP_NE, Ops[0], |
| 7452 | ConstantAggregateZero::get(Ty)); |
| 7453 | return Builder.CreateSExt(Ops[0], Ty, "vtst"); |
| 7454 | } |
| 7455 | case NEON::BI__builtin_neon_vuzp_v: |
| 7456 | case NEON::BI__builtin_neon_vuzpq_v: { |
| 7457 | Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty)); |
| 7458 | Ops[1] = Builder.CreateBitCast(Ops[1], Ty); |
| 7459 | Ops[2] = Builder.CreateBitCast(Ops[2], Ty); |
| 7460 | Value *SV = nullptr; |
| 7461 | |
| 7462 | for (unsigned vi = 0; vi != 2; ++vi) { |
| 7463 | SmallVector<int, 16> Indices; |
| 7464 | for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) |
| 7465 | Indices.push_back(2*i+vi); |
| 7466 | |
| 7467 | Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ty, Ops[0], vi); |
| 7468 | SV = Builder.CreateShuffleVector(Ops[1], Ops[2], Indices, "vuzp"); |
| 7469 | SV = Builder.CreateDefaultAlignedStore(SV, Addr); |
| 7470 | } |
| 7471 | return SV; |
| 7472 | } |
| 7473 | case NEON::BI__builtin_neon_vxarq_u64: { |
| 7474 | Function *F = CGM.getIntrinsic(Int); |
| 7475 | Ops[2] = Builder.CreateZExt(Ops[2], Int64Ty); |
| 7476 | return EmitNeonCall(F, Ops, ""); |
| 7477 | } |
| 7478 | case NEON::BI__builtin_neon_vzip_v: |
| 7479 | case NEON::BI__builtin_neon_vzipq_v: { |
| 7480 | Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty)); |
| 7481 | Ops[1] = Builder.CreateBitCast(Ops[1], Ty); |
| 7482 | Ops[2] = Builder.CreateBitCast(Ops[2], Ty); |
| 7483 | Value *SV = nullptr; |
| 7484 | |
| 7485 | for (unsigned vi = 0; vi != 2; ++vi) { |
| 7486 | SmallVector<int, 16> Indices; |
| 7487 | for (unsigned i = 0, e = VTy->getNumElements(); i != e; i += 2) { |
| 7488 | Indices.push_back((i + vi*e) >> 1); |
| 7489 | Indices.push_back(((i + vi*e) >> 1)+e); |
| 7490 | } |
| 7491 | Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ty, Ops[0], vi); |
| 7492 | SV = Builder.CreateShuffleVector(Ops[1], Ops[2], Indices, "vzip"); |
| 7493 | SV = Builder.CreateDefaultAlignedStore(SV, Addr); |
| 7494 | } |
| 7495 | return SV; |
| 7496 | } |
| 7497 | case NEON::BI__builtin_neon_vdot_s32: |
| 7498 | case NEON::BI__builtin_neon_vdot_u32: |
| 7499 | case NEON::BI__builtin_neon_vdotq_s32: |
| 7500 | case NEON::BI__builtin_neon_vdotq_u32: { |
| 7501 | auto *InputTy = |
| 7502 | llvm::FixedVectorType::get(Int8Ty, Ty->getPrimitiveSizeInBits() / 8); |
| 7503 | llvm::Type *Tys[2] = { Ty, InputTy }; |
| 7504 | return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vdot"); |
| 7505 | } |
| 7506 | case NEON::BI__builtin_neon_vfmlal_low_f16: |
| 7507 | case NEON::BI__builtin_neon_vfmlalq_low_f16: { |
| 7508 | auto *InputTy = |
| 7509 | llvm::FixedVectorType::get(HalfTy, Ty->getPrimitiveSizeInBits() / 16); |
| 7510 | llvm::Type *Tys[2] = { Ty, InputTy }; |
| 7511 | return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vfmlal_low"); |
| 7512 | } |
| 7513 | case NEON::BI__builtin_neon_vfmlsl_low_f16: |
| 7514 | case NEON::BI__builtin_neon_vfmlslq_low_f16: { |
| 7515 | auto *InputTy = |
| 7516 | llvm::FixedVectorType::get(HalfTy, Ty->getPrimitiveSizeInBits() / 16); |
| 7517 | llvm::Type *Tys[2] = { Ty, InputTy }; |
| 7518 | return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vfmlsl_low"); |
| 7519 | } |
| 7520 | case NEON::BI__builtin_neon_vfmlal_high_f16: |
| 7521 | case NEON::BI__builtin_neon_vfmlalq_high_f16: { |
| 7522 | auto *InputTy = |
| 7523 | llvm::FixedVectorType::get(HalfTy, Ty->getPrimitiveSizeInBits() / 16); |
| 7524 | llvm::Type *Tys[2] = { Ty, InputTy }; |
| 7525 | return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vfmlal_high"); |
| 7526 | } |
| 7527 | case NEON::BI__builtin_neon_vfmlsl_high_f16: |
| 7528 | case NEON::BI__builtin_neon_vfmlslq_high_f16: { |
| 7529 | auto *InputTy = |
| 7530 | llvm::FixedVectorType::get(HalfTy, Ty->getPrimitiveSizeInBits() / 16); |
| 7531 | llvm::Type *Tys[2] = { Ty, InputTy }; |
| 7532 | return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vfmlsl_high"); |
| 7533 | } |
| 7534 | case NEON::BI__builtin_neon_vmmlaq_s32: |
| 7535 | case NEON::BI__builtin_neon_vmmlaq_u32: { |
| 7536 | auto *InputTy = |
| 7537 | llvm::FixedVectorType::get(Int8Ty, Ty->getPrimitiveSizeInBits() / 8); |
| 7538 | llvm::Type *Tys[2] = { Ty, InputTy }; |
| 7539 | return EmitNeonCall(CGM.getIntrinsic(LLVMIntrinsic, Tys), Ops, "vmmla"); |
| 7540 | } |
| 7541 | case NEON::BI__builtin_neon_vusmmlaq_s32: { |
| 7542 | auto *InputTy = |
| 7543 | llvm::FixedVectorType::get(Int8Ty, Ty->getPrimitiveSizeInBits() / 8); |
| 7544 | llvm::Type *Tys[2] = { Ty, InputTy }; |
| 7545 | return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vusmmla"); |
| 7546 | } |
| 7547 | case NEON::BI__builtin_neon_vusdot_s32: |
| 7548 | case NEON::BI__builtin_neon_vusdotq_s32: { |
| 7549 | auto *InputTy = |
| 7550 | llvm::FixedVectorType::get(Int8Ty, Ty->getPrimitiveSizeInBits() / 8); |
| 7551 | llvm::Type *Tys[2] = { Ty, InputTy }; |
| 7552 | return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vusdot"); |
| 7553 | } |
| 7554 | case NEON::BI__builtin_neon_vbfdot_f32: |
| 7555 | case NEON::BI__builtin_neon_vbfdotq_f32: { |
| 7556 | llvm::Type *InputTy = |
| 7557 | llvm::FixedVectorType::get(BFloatTy, Ty->getPrimitiveSizeInBits() / 16); |
| 7558 | llvm::Type *Tys[2] = { Ty, InputTy }; |
| 7559 | return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vbfdot"); |
| 7560 | } |
| 7561 | case NEON::BI__builtin_neon___a32_vcvt_bf16_f32: { |
| 7562 | llvm::Type *Tys[1] = { Ty }; |
| 7563 | Function *F = CGM.getIntrinsic(Int, Tys); |
| 7564 | return EmitNeonCall(F, Ops, "vcvtfp2bf"); |
| 7565 | } |
| 7566 | |
| 7567 | } |
| 7568 | |
| 7569 | assert(Int && "Expected valid intrinsic number")(static_cast <bool> (Int && "Expected valid intrinsic number" ) ? void (0) : __assert_fail ("Int && \"Expected valid intrinsic number\"" , "clang/lib/CodeGen/CGBuiltin.cpp", 7569, __extension__ __PRETTY_FUNCTION__ )); |
| 7570 | |
| 7571 | // Determine the type(s) of this overloaded AArch64 intrinsic. |
| 7572 | Function *F = LookupNeonLLVMIntrinsic(Int, Modifier, Ty, E); |
| 7573 | |
| 7574 | Value *Result = EmitNeonCall(F, Ops, NameHint); |
| 7575 | llvm::Type *ResultType = ConvertType(E->getType()); |
| 7576 | // AArch64 intrinsic one-element vector type cast to |
| 7577 | // scalar type expected by the builtin |
| 7578 | return Builder.CreateBitCast(Result, ResultType, NameHint); |
| 7579 | } |
| 7580 | |
| 7581 | Value *CodeGenFunction::EmitAArch64CompareBuiltinExpr( |
| 7582 | Value *Op, llvm::Type *Ty, const CmpInst::Predicate Fp, |
| 7583 | const CmpInst::Predicate Ip, const Twine &Name) { |
| 7584 | llvm::Type *OTy = Op->getType(); |
| 7585 | |
| 7586 | // FIXME: this is utterly horrific. We should not be looking at previous |
| 7587 | // codegen context to find out what needs doing. Unfortunately TableGen |
| 7588 | // currently gives us exactly the same calls for vceqz_f32 and vceqz_s32 |
| 7589 | // (etc). |
| 7590 | if (BitCastInst *BI = dyn_cast<BitCastInst>(Op)) |
| 7591 | OTy = BI->getOperand(0)->getType(); |
| 7592 | |
| 7593 | Op = Builder.CreateBitCast(Op, OTy); |
| 7594 | if (OTy->getScalarType()->isFloatingPointTy()) { |
| 7595 | if (Fp == CmpInst::FCMP_OEQ) |
| 7596 | Op = Builder.CreateFCmp(Fp, Op, Constant::getNullValue(OTy)); |
| 7597 | else |
| 7598 | Op = Builder.CreateFCmpS(Fp, Op, Constant::getNullValue(OTy)); |
| 7599 | } else { |
| 7600 | Op = Builder.CreateICmp(Ip, Op, Constant::getNullValue(OTy)); |
| 7601 | } |
| 7602 | return Builder.CreateSExt(Op, Ty, Name); |
| 7603 | } |
| 7604 | |
| 7605 | static Value *packTBLDVectorList(CodeGenFunction &CGF, ArrayRef<Value *> Ops, |
| 7606 | Value *ExtOp, Value *IndexOp, |
| 7607 | llvm::Type *ResTy, unsigned IntID, |
| 7608 | const char *Name) { |
| 7609 | SmallVector<Value *, 2> TblOps; |
| 7610 | if (ExtOp) |
| 7611 | TblOps.push_back(ExtOp); |
| 7612 | |
| 7613 | // Build a vector containing sequential number like (0, 1, 2, ..., 15) |
| 7614 | SmallVector<int, 16> Indices; |
| 7615 | auto *TblTy = cast<llvm::FixedVectorType>(Ops[0]->getType()); |
| 7616 | for (unsigned i = 0, e = TblTy->getNumElements(); i != e; ++i) { |
| 7617 | Indices.push_back(2*i); |
| 7618 | Indices.push_back(2*i+1); |
| 7619 | } |
| 7620 | |
| 7621 | int PairPos = 0, End = Ops.size() - 1; |
| 7622 | while (PairPos < End) { |
| 7623 | TblOps.push_back(CGF.Builder.CreateShuffleVector(Ops[PairPos], |
| 7624 | Ops[PairPos+1], Indices, |
| 7625 | Name)); |
| 7626 | PairPos += 2; |
| 7627 | } |
| 7628 | |
| 7629 | // If there's an odd number of 64-bit lookup table, fill the high 64-bit |
| 7630 | // of the 128-bit lookup table with zero. |
| 7631 | if (PairPos == End) { |
| 7632 | Value *ZeroTbl = ConstantAggregateZero::get(TblTy); |
| 7633 | TblOps.push_back(CGF.Builder.CreateShuffleVector(Ops[PairPos], |
| 7634 | ZeroTbl, Indices, Name)); |
| 7635 | } |
| 7636 | |
| 7637 | Function *TblF; |
| 7638 | TblOps.push_back(IndexOp); |
| 7639 | TblF = CGF.CGM.getIntrinsic(IntID, ResTy); |
| 7640 | |
| 7641 | return CGF.EmitNeonCall(TblF, TblOps, Name); |
| 7642 | } |
| 7643 | |
| 7644 | Value *CodeGenFunction::GetValueForARMHint(unsigned BuiltinID) { |
| 7645 | unsigned Value; |
| 7646 | switch (BuiltinID) { |
| 7647 | default: |
| 7648 | return nullptr; |
| 7649 | case clang::ARM::BI__builtin_arm_nop: |
| 7650 | Value = 0; |
| 7651 | break; |
| 7652 | case clang::ARM::BI__builtin_arm_yield: |
| 7653 | case clang::ARM::BI__yield: |
| 7654 | Value = 1; |
| 7655 | break; |
| 7656 | case clang::ARM::BI__builtin_arm_wfe: |
| 7657 | case clang::ARM::BI__wfe: |
| 7658 | Value = 2; |
| 7659 | break; |
| 7660 | case clang::ARM::BI__builtin_arm_wfi: |
| 7661 | case clang::ARM::BI__wfi: |
| 7662 | Value = 3; |
| 7663 | break; |
| 7664 | case clang::ARM::BI__builtin_arm_sev: |
| 7665 | case clang::ARM::BI__sev: |
| 7666 | Value = 4; |
| 7667 | break; |
| 7668 | case clang::ARM::BI__builtin_arm_sevl: |
| 7669 | case clang::ARM::BI__sevl: |
| 7670 | Value = 5; |
| 7671 | break; |
| 7672 | } |
| 7673 | |
| 7674 | return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::arm_hint), |
| 7675 | llvm::ConstantInt::get(Int32Ty, Value)); |
| 7676 | } |
| 7677 | |
| 7678 | enum SpecialRegisterAccessKind { |
| 7679 | NormalRead, |
| 7680 | VolatileRead, |
| 7681 | Write, |
| 7682 | }; |
| 7683 | |
| 7684 | // Generates the IR for the read/write special register builtin, |
| 7685 | // ValueType is the type of the value that is to be written or read, |
| 7686 | // RegisterType is the type of the register being written to or read from. |
| 7687 | static Value *EmitSpecialRegisterBuiltin(CodeGenFunction &CGF, |
| 7688 | const CallExpr *E, |
| 7689 | llvm::Type *RegisterType, |
| 7690 | llvm::Type *ValueType, |
| 7691 | SpecialRegisterAccessKind AccessKind, |
| 7692 | StringRef SysReg = "") { |
| 7693 | // write and register intrinsics only support 32, 64 and 128 bit operations. |
| 7694 | assert((RegisterType->isIntegerTy(32) || RegisterType->isIntegerTy(64) ||(static_cast <bool> ((RegisterType->isIntegerTy(32) || RegisterType->isIntegerTy(64) || RegisterType->isIntegerTy (128)) && "Unsupported size for register.") ? void (0 ) : __assert_fail ("(RegisterType->isIntegerTy(32) || RegisterType->isIntegerTy(64) || RegisterType->isIntegerTy(128)) && \"Unsupported size for register.\"" , "clang/lib/CodeGen/CGBuiltin.cpp", 7696, __extension__ __PRETTY_FUNCTION__ )) |
| 7695 | RegisterType->isIntegerTy(128)) &&(static_cast <bool> ((RegisterType->isIntegerTy(32) || RegisterType->isIntegerTy(64) || RegisterType->isIntegerTy (128)) && "Unsupported size for register.") ? void (0 ) : __assert_fail ("(RegisterType->isIntegerTy(32) || RegisterType->isIntegerTy(64) || RegisterType->isIntegerTy(128)) && \"Unsupported size for register.\"" , "clang/lib/CodeGen/CGBuiltin.cpp", 7696, __extension__ __PRETTY_FUNCTION__ )) |
| 7696 | "Unsupported size for register.")(static_cast <bool> ((RegisterType->isIntegerTy(32) || RegisterType->isIntegerTy(64) || RegisterType->isIntegerTy (128)) && "Unsupported size for register.") ? void (0 ) : __assert_fail ("(RegisterType->isIntegerTy(32) || RegisterType->isIntegerTy(64) || RegisterType->isIntegerTy(128)) && \"Unsupported size for register.\"" , "clang/lib/CodeGen/CGBuiltin.cpp", 7696, __extension__ __PRETTY_FUNCTION__ )); |
| 7697 | |
| 7698 | CodeGen::CGBuilderTy &Builder = CGF.Builder; |
| 7699 | CodeGen::CodeGenModule &CGM = CGF.CGM; |
| 7700 | LLVMContext &Context = CGM.getLLVMContext(); |
| 7701 | |
| 7702 | if (SysReg.empty()) { |
| 7703 | const Expr *SysRegStrExpr = E->getArg(0)->IgnoreParenCasts(); |
| 7704 | SysReg = cast<clang::StringLiteral>(SysRegStrExpr)->getString(); |
| 7705 | } |
| 7706 | |
| 7707 | llvm::Metadata *Ops[] = { llvm::MDString::get(Context, SysReg) }; |
| 7708 | llvm::MDNode *RegName = llvm::MDNode::get(Context, Ops); |
| 7709 | llvm::Value *Metadata = llvm::MetadataAsValue::get(Context, RegName); |
| 7710 | |
| 7711 | llvm::Type *Types[] = { RegisterType }; |
| 7712 | |
| 7713 | bool MixedTypes = RegisterType->isIntegerTy(64) && ValueType->isIntegerTy(32); |
| 7714 | assert(!(RegisterType->isIntegerTy(32) && ValueType->isIntegerTy(64))(static_cast <bool> (!(RegisterType->isIntegerTy(32) && ValueType->isIntegerTy(64)) && "Can't fit 64-bit value in 32-bit register" ) ? void (0) : __assert_fail ("!(RegisterType->isIntegerTy(32) && ValueType->isIntegerTy(64)) && \"Can't fit 64-bit value in 32-bit register\"" , "clang/lib/CodeGen/CGBuiltin.cpp", 7715, __extension__ __PRETTY_FUNCTION__ )) |
| 7715 | && "Can't fit 64-bit value in 32-bit register")(static_cast <bool> (!(RegisterType->isIntegerTy(32) && ValueType->isIntegerTy(64)) && "Can't fit 64-bit value in 32-bit register" ) ? void (0) : __assert_fail ("!(RegisterType->isIntegerTy(32) && ValueType->isIntegerTy(64)) && \"Can't fit 64-bit value in 32-bit register\"" , "clang/lib/CodeGen/CGBuiltin.cpp", 7715, __extension__ __PRETTY_FUNCTION__ )); |
| 7716 | |
| 7717 | if (AccessKind != Write) { |
| 7718 | assert(AccessKind == NormalRead || AccessKind == VolatileRead)(static_cast <bool> (AccessKind == NormalRead || AccessKind == VolatileRead) ? void (0) : __assert_fail ("AccessKind == NormalRead || AccessKind == VolatileRead" , "clang/lib/CodeGen/CGBuiltin.cpp", 7718, __extension__ __PRETTY_FUNCTION__ )); |
| 7719 | llvm::Function *F = CGM.getIntrinsic( |
| 7720 | AccessKind == VolatileRead ? llvm::Intrinsic::read_volatile_register |
| 7721 | : llvm::Intrinsic::read_register, |
| 7722 | Types); |
| 7723 | llvm::Value *Call = Builder.CreateCall(F, Metadata); |
| 7724 | |
| 7725 | if (MixedTypes) |
| 7726 | // Read into 64 bit register and then truncate result to 32 bit. |
| 7727 | return Builder.CreateTrunc(Call, ValueType); |
| 7728 | |
| 7729 | if (ValueType->isPointerTy()) |
| 7730 | // Have i32/i64 result (Call) but want to return a VoidPtrTy (i8*). |
| 7731 | return Builder.CreateIntToPtr(Call, ValueType); |
| 7732 | |
| 7733 | return Call; |
| 7734 | } |
| 7735 | |
| 7736 | llvm::Function *F = CGM.getIntrinsic(llvm::Intrinsic::write_register, Types); |
| 7737 | llvm::Value *ArgValue = CGF.EmitScalarExpr(E->getArg(1)); |
| 7738 | if (MixedTypes) { |
| 7739 | // Extend 32 bit write value to 64 bit to pass to write. |
| 7740 | ArgValue = Builder.CreateZExt(ArgValue, RegisterType); |
| 7741 | return Builder.CreateCall(F, { Metadata, ArgValue }); |
| 7742 | } |
| 7743 | |
| 7744 | if (ValueType->isPointerTy()) { |
| 7745 | // Have VoidPtrTy ArgValue but want to return an i32/i64. |
| 7746 | ArgValue = Builder.CreatePtrToInt(ArgValue, RegisterType); |
| 7747 | return Builder.CreateCall(F, { Metadata, ArgValue }); |
| 7748 | } |
| 7749 | |
| 7750 | return Builder.CreateCall(F, { Metadata, ArgValue }); |
| 7751 | } |
| 7752 | |
| 7753 | /// Return true if BuiltinID is an overloaded Neon intrinsic with an extra |
| 7754 | /// argument that specifies the vector type. |
| 7755 | static bool HasExtraNeonArgument(unsigned BuiltinID) { |
| 7756 | switch (BuiltinID) { |
| 7757 | default: break; |
| 7758 | case NEON::BI__builtin_neon_vget_lane_i8: |
| 7759 | case NEON::BI__builtin_neon_vget_lane_i16: |
| 7760 | case NEON::BI__builtin_neon_vget_lane_bf16: |
| 7761 | case NEON::BI__builtin_neon_vget_lane_i32: |
| 7762 | case NEON::BI__builtin_neon_vget_lane_i64: |
| 7763 | case NEON::BI__builtin_neon_vget_lane_f32: |
| 7764 | case NEON::BI__builtin_neon_vgetq_lane_i8: |
| 7765 | case NEON::BI__builtin_neon_vgetq_lane_i16: |
| 7766 | case NEON::BI__builtin_neon_vgetq_lane_bf16: |
| 7767 | case NEON::BI__builtin_neon_vgetq_lane_i32: |
| 7768 | case NEON::BI__builtin_neon_vgetq_lane_i64: |
| 7769 | case NEON::BI__builtin_neon_vgetq_lane_f32: |
| 7770 | case NEON::BI__builtin_neon_vduph_lane_bf16: |
| 7771 | case NEON::BI__builtin_neon_vduph_laneq_bf16: |
| 7772 | case NEON::BI__builtin_neon_vset_lane_i8: |
| 7773 | case NEON::BI__builtin_neon_vset_lane_i16: |
| 7774 | case NEON::BI__builtin_neon_vset_lane_bf16: |
| 7775 | case NEON::BI__builtin_neon_vset_lane_i32: |
| 7776 | case NEON::BI__builtin_neon_vset_lane_i64: |
| 7777 | case NEON::BI__builtin_neon_vset_lane_f32: |
| 7778 | case NEON::BI__builtin_neon_vsetq_lane_i8: |
| 7779 | case NEON::BI__builtin_neon_vsetq_lane_i16: |
| 7780 | case NEON::BI__builtin_neon_vsetq_lane_bf16: |
| 7781 | case NEON::BI__builtin_neon_vsetq_lane_i32: |
| 7782 | case NEON::BI__builtin_neon_vsetq_lane_i64: |
| 7783 | case NEON::BI__builtin_neon_vsetq_lane_f32: |
| 7784 | case NEON::BI__builtin_neon_vsha1h_u32: |
| 7785 | case NEON::BI__builtin_neon_vsha1cq_u32: |
| 7786 | case NEON::BI__builtin_neon_vsha1pq_u32: |
| 7787 | case NEON::BI__builtin_neon_vsha1mq_u32: |
| 7788 | case NEON::BI__builtin_neon_vcvth_bf16_f32: |
| 7789 | case clang::ARM::BI_MoveToCoprocessor: |
| 7790 | case clang::ARM::BI_MoveToCoprocessor2: |
| 7791 | return false; |
| 7792 | } |
| 7793 | return true; |
| 7794 | } |
| 7795 | |
| 7796 | Value *CodeGenFunction::EmitARMBuiltinExpr(unsigned BuiltinID, |
| 7797 | const CallExpr *E, |
| 7798 | ReturnValueSlot ReturnValue, |
| 7799 | llvm::Triple::ArchType Arch) { |
| 7800 | if (auto Hint = GetValueForARMHint(BuiltinID)) |
| 7801 | return Hint; |
| 7802 | |
| 7803 | if (BuiltinID == clang::ARM::BI__emit) { |
| 7804 | bool IsThumb = getTarget().getTriple().getArch() == llvm::Triple::thumb; |
| 7805 | llvm::FunctionType *FTy = |
| 7806 | llvm::FunctionType::get(VoidTy, /*Variadic=*/false); |
| 7807 | |
| 7808 | Expr::EvalResult Result; |
| 7809 | if (!E->getArg(0)->EvaluateAsInt(Result, CGM.getContext())) |
| 7810 | llvm_unreachable("Sema will ensure that the parameter is constant")::llvm::llvm_unreachable_internal("Sema will ensure that the parameter is constant" , "clang/lib/CodeGen/CGBuiltin.cpp", 7810); |
| 7811 | |
| 7812 | llvm::APSInt Value = Result.Val.getInt(); |
| 7813 | uint64_t ZExtValue = Value.zextOrTrunc(IsThumb ? 16 : 32).getZExtValue(); |
| 7814 | |
| 7815 | llvm::InlineAsm *Emit = |
| 7816 | IsThumb ? InlineAsm::get(FTy, ".inst.n 0x" + utohexstr(ZExtValue), "", |
| 7817 | /*hasSideEffects=*/true) |
| 7818 | : InlineAsm::get(FTy, ".inst 0x" + utohexstr(ZExtValue), "", |
| 7819 | /*hasSideEffects=*/true); |
| 7820 | |
| 7821 | return Builder.CreateCall(Emit); |
| 7822 | } |
| 7823 | |
| 7824 | if (BuiltinID == clang::ARM::BI__builtin_arm_dbg) { |
| 7825 | Value *Option = EmitScalarExpr(E->getArg(0)); |
| 7826 | return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::arm_dbg), Option); |
| 7827 | } |
| 7828 | |
| 7829 | if (BuiltinID == clang::ARM::BI__builtin_arm_prefetch) { |
| 7830 | Value *Address = EmitScalarExpr(E->getArg(0)); |
| 7831 | Value *RW = EmitScalarExpr(E->getArg(1)); |
| 7832 | Value *IsData = EmitScalarExpr(E->getArg(2)); |
| 7833 | |
| 7834 | // Locality is not supported on ARM target |
| 7835 | Value *Locality = llvm::ConstantInt::get(Int32Ty, 3); |
| 7836 | |
| 7837 | Function *F = CGM.getIntrinsic(Intrinsic::prefetch, Address->getType()); |
| 7838 | return Builder.CreateCall(F, {Address, RW, Locality, IsData}); |
| 7839 | } |
| 7840 | |
| 7841 | if (BuiltinID == clang::ARM::BI__builtin_arm_rbit) { |
| 7842 | llvm::Value *Arg = EmitScalarExpr(E->getArg(0)); |
| 7843 | return Builder.CreateCall( |
| 7844 | CGM.getIntrinsic(Intrinsic::bitreverse, Arg->getType()), Arg, "rbit"); |
| 7845 | } |
| 7846 | |
| 7847 | if (BuiltinID == clang::ARM::BI__builtin_arm_cls) { |
| 7848 | llvm::Value *Arg = EmitScalarExpr(E->getArg(0)); |
| 7849 | return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::arm_cls), Arg, "cls"); |
| 7850 | } |
| 7851 | if (BuiltinID == clang::ARM::BI__builtin_arm_cls64) { |
| 7852 | llvm::Value *Arg = EmitScalarExpr(E->getArg(0)); |
| 7853 | return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::arm_cls64), Arg, |
| 7854 | "cls"); |
| 7855 | } |
| 7856 | |
| 7857 | if (BuiltinID == clang::ARM::BI__clear_cache) { |
| 7858 | assert(E->getNumArgs() == 2 && "__clear_cache takes 2 arguments")(static_cast <bool> (E->getNumArgs() == 2 && "__clear_cache takes 2 arguments") ? void (0) : __assert_fail ("E->getNumArgs() == 2 && \"__clear_cache takes 2 arguments\"" , "clang/lib/CodeGen/CGBuiltin.cpp", 7858, __extension__ __PRETTY_FUNCTION__ )); |
| 7859 | const FunctionDecl *FD = E->getDirectCallee(); |
| 7860 | Value *Ops[2]; |
| 7861 | for (unsigned i = 0; i < 2; i++) |
| 7862 | Ops[i] = EmitScalarExpr(E->getArg(i)); |
| 7863 | llvm::Type *Ty = CGM.getTypes().ConvertType(FD->getType()); |
| 7864 | llvm::FunctionType *FTy = cast<llvm::FunctionType>(Ty); |
| 7865 | StringRef Name = FD->getName(); |
| 7866 | return EmitNounwindRuntimeCall(CGM.CreateRuntimeFunction(FTy, Name), Ops); |
| 7867 | } |
| 7868 | |
| 7869 | if (BuiltinID == clang::ARM::BI__builtin_arm_mcrr || |
| 7870 | BuiltinID == clang::ARM::BI__builtin_arm_mcrr2) { |
| 7871 | Function *F; |
| 7872 | |
| 7873 | switch (BuiltinID) { |
| 7874 | default: llvm_unreachable("unexpected builtin")::llvm::llvm_unreachable_internal("unexpected builtin", "clang/lib/CodeGen/CGBuiltin.cpp" , 7874); |
| 7875 | case clang::ARM::BI__builtin_arm_mcrr: |
| 7876 | F = CGM.getIntrinsic(Intrinsic::arm_mcrr); |
| 7877 | break; |
| 7878 | case clang::ARM::BI__builtin_arm_mcrr2: |
| 7879 | F = CGM.getIntrinsic(Intrinsic::arm_mcrr2); |
| 7880 | break; |
| 7881 | } |
| 7882 | |
| 7883 | // MCRR{2} instruction has 5 operands but |
| 7884 | // the intrinsic has 4 because Rt and Rt2 |
| 7885 | // are represented as a single unsigned 64 |
| 7886 | // bit integer in the intrinsic definition |
| 7887 | // but internally it's represented as 2 32 |
| 7888 | // bit integers. |
| 7889 | |
| 7890 | Value *Coproc = EmitScalarExpr(E->getArg(0)); |
| 7891 | Value *Opc1 = EmitScalarExpr(E->getArg(1)); |
| 7892 | Value *RtAndRt2 = EmitScalarExpr(E->getArg(2)); |
| 7893 | Value *CRm = EmitScalarExpr(E->getArg(3)); |
| 7894 | |
| 7895 | Value *C1 = llvm::ConstantInt::get(Int64Ty, 32); |
| 7896 | Value *Rt = Builder.CreateTruncOrBitCast(RtAndRt2, Int32Ty); |
| 7897 | Value *Rt2 = Builder.CreateLShr(RtAndRt2, C1); |
| 7898 | Rt2 = Builder.CreateTruncOrBitCast(Rt2, Int32Ty); |
| 7899 | |
| 7900 | return Builder.CreateCall(F, {Coproc, Opc1, Rt, Rt2, CRm}); |
| 7901 | } |
| 7902 | |
| 7903 | if (BuiltinID == clang::ARM::BI__builtin_arm_mrrc || |
| 7904 | BuiltinID == clang::ARM::BI__builtin_arm_mrrc2) { |
| 7905 | Function *F; |
| 7906 | |
| 7907 | switch (BuiltinID) { |
| 7908 | default: llvm_unreachable("unexpected builtin")::llvm::llvm_unreachable_internal("unexpected builtin", "clang/lib/CodeGen/CGBuiltin.cpp" , 7908); |
| 7909 | case clang::ARM::BI__builtin_arm_mrrc: |
| 7910 | F = CGM.getIntrinsic(Intrinsic::arm_mrrc); |
| 7911 | break; |
| 7912 | case clang::ARM::BI__builtin_arm_mrrc2: |
| 7913 | F = CGM.getIntrinsic(Intrinsic::arm_mrrc2); |
| 7914 | break; |
| 7915 | } |
| 7916 | |
| 7917 | Value *Coproc = EmitScalarExpr(E->getArg(0)); |
| 7918 | Value *Opc1 = EmitScalarExpr(E->getArg(1)); |
| 7919 | Value *CRm = EmitScalarExpr(E->getArg(2)); |
| 7920 | Value *RtAndRt2 = Builder.CreateCall(F, {Coproc, Opc1, CRm}); |
| 7921 | |
| 7922 | // Returns an unsigned 64 bit integer, represented |
| 7923 | // as two 32 bit integers. |
| 7924 | |
| 7925 | Value *Rt = Builder.CreateExtractValue(RtAndRt2, 1); |
| 7926 | Value *Rt1 = Builder.CreateExtractValue(RtAndRt2, 0); |
| 7927 | Rt = Builder.CreateZExt(Rt, Int64Ty); |
| 7928 | Rt1 = Builder.CreateZExt(Rt1, Int64Ty); |
| 7929 | |
| 7930 | Value *ShiftCast = llvm::ConstantInt::get(Int64Ty, 32); |
| 7931 | RtAndRt2 = Builder.CreateShl(Rt, ShiftCast, "shl", true); |
| 7932 | RtAndRt2 = Builder.CreateOr(RtAndRt2, Rt1); |
| 7933 | |
| 7934 | return Builder.CreateBitCast(RtAndRt2, ConvertType(E->getType())); |
| 7935 | } |
| 7936 | |
| 7937 | if (BuiltinID == clang::ARM::BI__builtin_arm_ldrexd || |
| 7938 | ((BuiltinID == clang::ARM::BI__builtin_arm_ldrex || |
| 7939 | BuiltinID == clang::ARM::BI__builtin_arm_ldaex) && |
| 7940 | getContext().getTypeSize(E->getType()) == 64) || |
| 7941 | BuiltinID == clang::ARM::BI__ldrexd) { |
| 7942 | Function *F; |
| 7943 | |
| 7944 | switch (BuiltinID) { |
| 7945 | default: llvm_unreachable("unexpected builtin")::llvm::llvm_unreachable_internal("unexpected builtin", "clang/lib/CodeGen/CGBuiltin.cpp" , 7945); |
| 7946 | case clang::ARM::BI__builtin_arm_ldaex: |
| 7947 | F = CGM.getIntrinsic(Intrinsic::arm_ldaexd); |
| 7948 | break; |
| 7949 | case clang::ARM::BI__builtin_arm_ldrexd: |
| 7950 | case clang::ARM::BI__builtin_arm_ldrex: |
| 7951 | case clang::ARM::BI__ldrexd: |
| 7952 | F = CGM.getIntrinsic(Intrinsic::arm_ldrexd); |
| 7953 | break; |
| 7954 | } |
| 7955 | |
| 7956 | Value *LdPtr = EmitScalarExpr(E->getArg(0)); |
| 7957 | Value *Val = Builder.CreateCall(F, Builder.CreateBitCast(LdPtr, Int8PtrTy), |
| 7958 | "ldrexd"); |
| 7959 | |
| 7960 | Value *Val0 = Builder.CreateExtractValue(Val, 1); |
| 7961 | Value *Val1 = Builder.CreateExtractValue(Val, 0); |
| 7962 | Val0 = Builder.CreateZExt(Val0, Int64Ty); |
| 7963 | Val1 = Builder.CreateZExt(Val1, Int64Ty); |
| 7964 | |
| 7965 | Value *ShiftCst = llvm::ConstantInt::get(Int64Ty, 32); |
| 7966 | Val = Builder.CreateShl(Val0, ShiftCst, "shl", true /* nuw */); |
| 7967 | Val = Builder.CreateOr(Val, Val1); |
| 7968 | return Builder.CreateBitCast(Val, ConvertType(E->getType())); |
| 7969 | } |
| 7970 | |
| 7971 | if (BuiltinID == clang::ARM::BI__builtin_arm_ldrex || |
| 7972 | BuiltinID == clang::ARM::BI__builtin_arm_ldaex) { |
| 7973 | Value *LoadAddr = EmitScalarExpr(E->getArg(0)); |
| 7974 | |
| 7975 | QualType Ty = E->getType(); |
| 7976 | llvm::Type *RealResTy = ConvertType(Ty); |
| 7977 | llvm::Type *IntTy = |
| 7978 | llvm::IntegerType::get(getLLVMContext(), getContext().getTypeSize(Ty)); |
| 7979 | llvm::Type *PtrTy = IntTy->getPointerTo(); |
| 7980 | LoadAddr = Builder.CreateBitCast(LoadAddr, PtrTy); |
| 7981 | |
| 7982 | Function *F = CGM.getIntrinsic( |
| 7983 | BuiltinID == clang::ARM::BI__builtin_arm_ldaex ? Intrinsic::arm_ldaex |
| 7984 | : Intrinsic::arm_ldrex, |
| 7985 | PtrTy); |
| 7986 | CallInst *Val = Builder.CreateCall(F, LoadAddr, "ldrex"); |
| 7987 | Val->addParamAttr( |
| 7988 | 0, Attribute::get(getLLVMContext(), Attribute::ElementType, IntTy)); |
| 7989 | |
| 7990 | if (RealResTy->isPointerTy()) |
| 7991 | return Builder.CreateIntToPtr(Val, RealResTy); |
| 7992 | else { |
| 7993 | llvm::Type *IntResTy = llvm::IntegerType::get( |
| 7994 | getLLVMContext(), CGM.getDataLayout().getTypeSizeInBits(RealResTy)); |
| 7995 | return Builder.CreateBitCast(Builder.CreateTruncOrBitCast(Val, IntResTy), |
| 7996 | RealResTy); |
| 7997 | } |
| 7998 | } |
| 7999 | |
| 8000 | if (BuiltinID == clang::ARM::BI__builtin_arm_strexd || |
| 8001 | ((BuiltinID == clang::ARM::BI__builtin_arm_stlex || |
| 8002 | BuiltinID == clang::ARM::BI__builtin_arm_strex) && |
| 8003 | getContext().getTypeSize(E->getArg(0)->getType()) == 64)) { |
| 8004 | Function *F = CGM.getIntrinsic( |
| 8005 | BuiltinID == clang::ARM::BI__builtin_arm_stlex ? Intrinsic::arm_stlexd |
| 8006 | : Intrinsic::arm_strexd); |
| 8007 | llvm::Type *STy = llvm::StructType::get(Int32Ty, Int32Ty); |
| 8008 | |
| 8009 | Address Tmp = CreateMemTemp(E->getArg(0)->getType()); |
| 8010 | Value *Val = EmitScalarExpr(E->getArg(0)); |
| 8011 | Builder.CreateStore(Val, Tmp); |
| 8012 | |
| 8013 | Address LdPtr = Builder.CreateElementBitCast(Tmp, STy); |
| 8014 | Val = Builder.CreateLoad(LdPtr); |
| 8015 | |
| 8016 | Value *Arg0 = Builder.CreateExtractValue(Val, 0); |
| 8017 | Value *Arg1 = Builder.CreateExtractValue(Val, 1); |
| 8018 | Value *StPtr = Builder.CreateBitCast(EmitScalarExpr(E->getArg(1)), Int8PtrTy); |
| 8019 | return Builder.CreateCall(F, {Arg0, Arg1, StPtr}, "strexd"); |
| 8020 | } |
| 8021 | |
| 8022 | if (BuiltinID == clang::ARM::BI__builtin_arm_strex || |
| 8023 | BuiltinID == clang::ARM::BI__builtin_arm_stlex) { |
| 8024 | Value *StoreVal = EmitScalarExpr(E->getArg(0)); |
| 8025 | Value *StoreAddr = EmitScalarExpr(E->getArg(1)); |
| 8026 | |
| 8027 | QualType Ty = E->getArg(0)->getType(); |
| 8028 | llvm::Type *StoreTy = llvm::IntegerType::get(getLLVMContext(), |
| 8029 | getContext().getTypeSize(Ty)); |
| 8030 | StoreAddr = Builder.CreateBitCast(StoreAddr, StoreTy->getPointerTo()); |
| 8031 | |
| 8032 | if (StoreVal->getType()->isPointerTy()) |
| 8033 | StoreVal = Builder.CreatePtrToInt(StoreVal, Int32Ty); |
| 8034 | else { |
| 8035 | llvm::Type *IntTy = llvm::IntegerType::get( |
| 8036 | getLLVMContext(), |
| 8037 | CGM.getDataLayout().getTypeSizeInBits(StoreVal->getType())); |
| 8038 | StoreVal = Builder.CreateBitCast(StoreVal, IntTy); |
| 8039 | StoreVal = Builder.CreateZExtOrBitCast(StoreVal, Int32Ty); |
| 8040 | } |
| 8041 | |
| 8042 | Function *F = CGM.getIntrinsic( |
| 8043 | BuiltinID == clang::ARM::BI__builtin_arm_stlex ? Intrinsic::arm_stlex |
| 8044 | : Intrinsic::arm_strex, |
| 8045 | StoreAddr->getType()); |
| 8046 | |
| 8047 | CallInst *CI = Builder.CreateCall(F, {StoreVal, StoreAddr}, "strex"); |
| 8048 | CI->addParamAttr( |
| 8049 | 1, Attribute::get(getLLVMContext(), Attribute::ElementType, StoreTy)); |
| 8050 | return CI; |
| 8051 | } |
| 8052 | |
| 8053 | if (BuiltinID == clang::ARM::BI__builtin_arm_clrex) { |
| 8054 | Function *F = CGM.getIntrinsic(Intrinsic::arm_clrex); |
| 8055 | return Builder.CreateCall(F); |
| 8056 | } |
| 8057 | |
| 8058 | // CRC32 |
| 8059 | Intrinsic::ID CRCIntrinsicID = Intrinsic::not_intrinsic; |
| 8060 | switch (BuiltinID) { |
| 8061 | case clang::ARM::BI__builtin_arm_crc32b: |
| 8062 | CRCIntrinsicID = Intrinsic::arm_crc32b; break; |
| 8063 | case clang::ARM::BI__builtin_arm_crc32cb: |
| 8064 | CRCIntrinsicID = Intrinsic::arm_crc32cb; break; |
| 8065 | case clang::ARM::BI__builtin_arm_crc32h: |
| 8066 | CRCIntrinsicID = Intrinsic::arm_crc32h; break; |
| 8067 | case clang::ARM::BI__builtin_arm_crc32ch: |
| 8068 | CRCIntrinsicID = Intrinsic::arm_crc32ch; break; |
| 8069 | case clang::ARM::BI__builtin_arm_crc32w: |
| 8070 | case clang::ARM::BI__builtin_arm_crc32d: |
| 8071 | CRCIntrinsicID = Intrinsic::arm_crc32w; break; |
| 8072 | case clang::ARM::BI__builtin_arm_crc32cw: |
| 8073 | case clang::ARM::BI__builtin_arm_crc32cd: |
| 8074 | CRCIntrinsicID = Intrinsic::arm_crc32cw; break; |
| 8075 | } |
| 8076 | |
| 8077 | if (CRCIntrinsicID != Intrinsic::not_intrinsic) { |
| 8078 | Value *Arg0 = EmitScalarExpr(E->getArg(0)); |
| 8079 | Value *Arg1 = EmitScalarExpr(E->getArg(1)); |
| 8080 | |
| 8081 | // crc32{c,}d intrinsics are implemented as two calls to crc32{c,}w |
| 8082 | // intrinsics, hence we need different codegen for these cases. |
| 8083 | if (BuiltinID == clang::ARM::BI__builtin_arm_crc32d || |
| 8084 | BuiltinID == clang::ARM::BI__builtin_arm_crc32cd) { |
| 8085 | Value *C1 = llvm::ConstantInt::get(Int64Ty, 32); |
| 8086 | Value *Arg1a = Builder.CreateTruncOrBitCast(Arg1, Int32Ty); |
| 8087 | Value *Arg1b = Builder.CreateLShr(Arg1, C1); |
| 8088 | Arg1b = Builder.CreateTruncOrBitCast(Arg1b, Int32Ty); |
| 8089 | |
| 8090 | Function *F = CGM.getIntrinsic(CRCIntrinsicID); |
| 8091 | Value *Res = Builder.CreateCall(F, {Arg0, Arg1a}); |
| 8092 | return Builder.CreateCall(F, {Res, Arg1b}); |
| 8093 | } else { |
| 8094 | Arg1 = Builder.CreateZExtOrBitCast(Arg1, Int32Ty); |
| 8095 | |
| 8096 | Function *F = CGM.getIntrinsic(CRCIntrinsicID); |
| 8097 | return Builder.CreateCall(F, {Arg0, Arg1}); |
| 8098 | } |
| 8099 | } |
| 8100 | |
| 8101 | if (BuiltinID == clang::ARM::BI__builtin_arm_rsr || |
| 8102 | BuiltinID == clang::ARM::BI__builtin_arm_rsr64 || |
| 8103 | BuiltinID == clang::ARM::BI__builtin_arm_rsrp || |
| 8104 | BuiltinID == clang::ARM::BI__builtin_arm_wsr || |
| 8105 | BuiltinID == clang::ARM::BI__builtin_arm_wsr64 || |
| 8106 | BuiltinID == clang::ARM::BI__builtin_arm_wsrp) { |
| 8107 | |
| 8108 | SpecialRegisterAccessKind AccessKind = Write; |
| 8109 | if (BuiltinID == clang::ARM::BI__builtin_arm_rsr || |
| 8110 | BuiltinID == clang::ARM::BI__builtin_arm_rsr64 || |
| 8111 | BuiltinID == clang::ARM::BI__builtin_arm_rsrp) |
| 8112 | AccessKind = VolatileRead; |
| 8113 | |
| 8114 | bool IsPointerBuiltin = BuiltinID == clang::ARM::BI__builtin_arm_rsrp || |
| 8115 | BuiltinID == clang::ARM::BI__builtin_arm_wsrp; |
| 8116 | |
| 8117 | bool Is64Bit = BuiltinID == clang::ARM::BI__builtin_arm_rsr64 || |
| 8118 | BuiltinID == clang::ARM::BI__builtin_arm_wsr64; |
| 8119 | |
| 8120 | llvm::Type *ValueType; |
| 8121 | llvm::Type *RegisterType; |
| 8122 | if (IsPointerBuiltin) { |
| 8123 | ValueType = VoidPtrTy; |
| 8124 | RegisterType = Int32Ty; |
| 8125 | } else if (Is64Bit) { |
| 8126 | ValueType = RegisterType = Int64Ty; |
| 8127 | } else { |
| 8128 | ValueType = RegisterType = Int32Ty; |
| 8129 | } |
| 8130 | |
| 8131 | return EmitSpecialRegisterBuiltin(*this, E, RegisterType, ValueType, |
| 8132 | AccessKind); |
| 8133 | } |
| 8134 | |
| 8135 | if (BuiltinID == ARM::BI__builtin_sponentry) { |
| 8136 | llvm::Function *F = CGM.getIntrinsic(Intrinsic::sponentry, AllocaInt8PtrTy); |
| 8137 | return Builder.CreateCall(F); |
| 8138 | } |
| 8139 | |
| 8140 | // Handle MSVC intrinsics before argument evaluation to prevent double |
| 8141 | // evaluation. |
| 8142 | if (std::optional<MSVCIntrin> MsvcIntId = translateArmToMsvcIntrin(BuiltinID)) |
| 8143 | return EmitMSVCBuiltinExpr(*MsvcIntId, E); |
| 8144 | |
| 8145 | // Deal with MVE builtins |
| 8146 | if (Value *Result = EmitARMMVEBuiltinExpr(BuiltinID, E, ReturnValue, Arch)) |
| 8147 | return Result; |
| 8148 | // Handle CDE builtins |
| 8149 | if (Value *Result = EmitARMCDEBuiltinExpr(BuiltinID, E, ReturnValue, Arch)) |
| 8150 | return Result; |
| 8151 | |
| 8152 | // Some intrinsics are equivalent - if they are use the base intrinsic ID. |
| 8153 | auto It = llvm::find_if(NEONEquivalentIntrinsicMap, [BuiltinID](auto &P) { |
| 8154 | return P.first == BuiltinID; |
| 8155 | }); |
| 8156 | if (It != end(NEONEquivalentIntrinsicMap)) |
| 8157 | BuiltinID = It->second; |
| 8158 | |
| 8159 | // Find out if any arguments are required to be integer constant |
| 8160 | // expressions. |
| 8161 | unsigned ICEArguments = 0; |
| 8162 | ASTContext::GetBuiltinTypeError Error; |
| 8163 | getContext().GetBuiltinType(BuiltinID, Error, &ICEArguments); |
| 8164 | assert(Error == ASTContext::GE_None && "Should not codegen an error")(static_cast <bool> (Error == ASTContext::GE_None && "Should not codegen an error") ? void (0) : __assert_fail ("Error == ASTContext::GE_None && \"Should not codegen an error\"" , "clang/lib/CodeGen/CGBuiltin.cpp", 8164, __extension__ __PRETTY_FUNCTION__ )); |
| 8165 | |
| 8166 | auto getAlignmentValue32 = [&](Address addr) -> Value* { |
| 8167 | return Builder.getInt32(addr.getAlignment().getQuantity()); |
| 8168 | }; |
| 8169 | |
| 8170 | Address PtrOp0 = Address::invalid(); |
| 8171 | Address PtrOp1 = Address::invalid(); |
| 8172 | SmallVector<Value*, 4> Ops; |
| 8173 | bool HasExtraArg = HasExtraNeonArgument(BuiltinID); |
| 8174 | unsigned NumArgs = E->getNumArgs() - (HasExtraArg ? 1 : 0); |
| 8175 | for (unsigned i = 0, e = NumArgs; i != e; i++) { |
| 8176 | if (i == 0) { |
| 8177 | switch (BuiltinID) { |
| 8178 | case NEON::BI__builtin_neon_vld1_v: |
| 8179 | case NEON::BI__builtin_neon_vld1q_v: |
| 8180 | case NEON::BI__builtin_neon_vld1q_lane_v: |
| 8181 | case NEON::BI__builtin_neon_vld1_lane_v: |
| 8182 | case NEON::BI__builtin_neon_vld1_dup_v: |
| 8183 | case NEON::BI__builtin_neon_vld1q_dup_v: |
| 8184 | case NEON::BI__builtin_neon_vst1_v: |
| 8185 | case NEON::BI__builtin_neon_vst1q_v: |
| 8186 | case NEON::BI__builtin_neon_vst1q_lane_v: |
| 8187 | case NEON::BI__builtin_neon_vst1_lane_v: |
| 8188 | case NEON::BI__builtin_neon_vst2_v: |
| 8189 | case NEON::BI__builtin_neon_vst2q_v: |
| 8190 | case NEON::BI__builtin_neon_vst2_lane_v: |
| 8191 | case NEON::BI__builtin_neon_vst2q_lane_v: |
| 8192 | case NEON::BI__builtin_neon_vst3_v: |
| 8193 | case NEON::BI__builtin_neon_vst3q_v: |
| 8194 | case NEON::BI__builtin_neon_vst3_lane_v: |
| 8195 | case NEON::BI__builtin_neon_vst3q_lane_v: |
| 8196 | case NEON::BI__builtin_neon_vst4_v: |
| 8197 | case NEON::BI__builtin_neon_vst4q_v: |
| 8198 | case NEON::BI__builtin_neon_vst4_lane_v: |
| 8199 | case NEON::BI__builtin_neon_vst4q_lane_v: |
| 8200 | // Get the alignment for the argument in addition to the value; |
| 8201 | // we'll use it later. |
| 8202 | PtrOp0 = EmitPointerWithAlignment(E->getArg(0)); |
| 8203 | Ops.push_back(PtrOp0.getPointer()); |
| 8204 | continue; |
| 8205 | } |
| 8206 | } |
| 8207 | if (i == 1) { |
| 8208 | switch (BuiltinID) { |
| 8209 | case NEON::BI__builtin_neon_vld2_v: |
| 8210 | case NEON::BI__builtin_neon_vld2q_v: |
| 8211 | case NEON::BI__builtin_neon_vld3_v: |
| 8212 | case NEON::BI__builtin_neon_vld3q_v: |
| 8213 | case NEON::BI__builtin_neon_vld4_v: |
| 8214 | case NEON::BI__builtin_neon_vld4q_v: |
| 8215 | case NEON::BI__builtin_neon_vld2_lane_v: |
| 8216 | case NEON::BI__builtin_neon_vld2q_lane_v: |
| 8217 | case NEON::BI__builtin_neon_vld3_lane_v: |
| 8218 | case NEON::BI__builtin_neon_vld3q_lane_v: |
| 8219 | case NEON::BI__builtin_neon_vld4_lane_v: |
| 8220 | case NEON::BI__builtin_neon_vld4q_lane_v: |
| 8221 | case NEON::BI__builtin_neon_vld2_dup_v: |
| 8222 | case NEON::BI__builtin_neon_vld2q_dup_v: |
| 8223 | case NEON::BI__builtin_neon_vld3_dup_v: |
| 8224 | case NEON::BI__builtin_neon_vld3q_dup_v: |
| 8225 | case NEON::BI__builtin_neon_vld4_dup_v: |
| 8226 | case NEON::BI__builtin_neon_vld4q_dup_v: |
| 8227 | // Get the alignment for the argument in addition to the value; |
| 8228 | // we'll use it later. |
| 8229 | PtrOp1 = EmitPointerWithAlignment(E->getArg(1)); |
| 8230 | Ops.push_back(PtrOp1.getPointer()); |
| 8231 | continue; |
| 8232 | } |
| 8233 | } |
| 8234 | |
| 8235 | if ((ICEArguments & (1 << i)) == 0) { |
| 8236 | Ops.push_back(EmitScalarExpr(E->getArg(i))); |
| 8237 | } else { |
| 8238 | // If this is required to be a constant, constant fold it so that we know |
| 8239 | // that the generated intrinsic gets a ConstantInt. |
| 8240 | Ops.push_back(llvm::ConstantInt::get( |
| 8241 | getLLVMContext(), |
| 8242 | *E->getArg(i)->getIntegerConstantExpr(getContext()))); |
| 8243 | } |
| 8244 | } |
| 8245 | |
| 8246 | switch (BuiltinID) { |
| 8247 | default: break; |
| 8248 | |
| 8249 | case NEON::BI__builtin_neon_vget_lane_i8: |
| 8250 | case NEON::BI__builtin_neon_vget_lane_i16: |
| 8251 | case NEON::BI__builtin_neon_vget_lane_i32: |
| 8252 | case NEON::BI__builtin_neon_vget_lane_i64: |
| 8253 | case NEON::BI__builtin_neon_vget_lane_bf16: |
| 8254 | case NEON::BI__builtin_neon_vget_lane_f32: |
| 8255 | case NEON::BI__builtin_neon_vgetq_lane_i8: |
| 8256 | case NEON::BI__builtin_neon_vgetq_lane_i16: |
| 8257 | case NEON::BI__builtin_neon_vgetq_lane_i32: |
| 8258 | case NEON::BI__builtin_neon_vgetq_lane_i64: |
| 8259 | case NEON::BI__builtin_neon_vgetq_lane_bf16: |
| 8260 | case NEON::BI__builtin_neon_vgetq_lane_f32: |
| 8261 | case NEON::BI__builtin_neon_vduph_lane_bf16: |
| 8262 | case NEON::BI__builtin_neon_vduph_laneq_bf16: |
| 8263 | return Builder.CreateExtractElement(Ops[0], Ops[1], "vget_lane"); |
| 8264 | |
| 8265 | case NEON::BI__builtin_neon_vrndns_f32: { |
| 8266 | Value *Arg = EmitScalarExpr(E->getArg(0)); |
| 8267 | llvm::Type *Tys[] = {Arg->getType()}; |
| 8268 | Function *F = CGM.getIntrinsic(Intrinsic::arm_neon_vrintn, Tys); |
| 8269 | return Builder.CreateCall(F, {Arg}, "vrndn"); } |
| 8270 | |
| 8271 | case NEON::BI__builtin_neon_vset_lane_i8: |
| 8272 | case NEON::BI__builtin_neon_vset_lane_i16: |
| 8273 | case NEON::BI__builtin_neon_vset_lane_i32: |
| 8274 | case NEON::BI__builtin_neon_vset_lane_i64: |
| 8275 | case NEON::BI__builtin_neon_vset_lane_bf16: |
| 8276 | case NEON::BI__builtin_neon_vset_lane_f32: |
| 8277 | case NEON::BI__builtin_neon_vsetq_lane_i8: |
| 8278 | case NEON::BI__builtin_neon_vsetq_lane_i16: |
| 8279 | case NEON::BI__builtin_neon_vsetq_lane_i32: |
| 8280 | case NEON::BI__builtin_neon_vsetq_lane_i64: |
| 8281 | case NEON::BI__builtin_neon_vsetq_lane_bf16: |
| 8282 | case NEON::BI__builtin_neon_vsetq_lane_f32: |
| 8283 | return Builder.CreateInsertElement(Ops[1], Ops[0], Ops[2], "vset_lane"); |
| 8284 | |
| 8285 | case NEON::BI__builtin_neon_vsha1h_u32: |
| 8286 | return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_sha1h), Ops, |
| 8287 | "vsha1h"); |
| 8288 | case NEON::BI__builtin_neon_vsha1cq_u32: |
| 8289 | return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_sha1c), Ops, |
| 8290 | "vsha1h"); |
| 8291 | case NEON::BI__builtin_neon_vsha1pq_u32: |
| 8292 | return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_sha1p), Ops, |
| 8293 | "vsha1h"); |
| 8294 | case NEON::BI__builtin_neon_vsha1mq_u32: |
| 8295 | return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_sha1m), Ops, |
| 8296 | "vsha1h"); |
| 8297 | |
| 8298 | case NEON::BI__builtin_neon_vcvth_bf16_f32: { |
| 8299 | return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vcvtbfp2bf), Ops, |
| 8300 | "vcvtbfp2bf"); |
| 8301 | } |
| 8302 | |
| 8303 | // The ARM _MoveToCoprocessor builtins put the input register value as |
| 8304 | // the first argument, but the LLVM intrinsic expects it as the third one. |
| 8305 | case clang::ARM::BI_MoveToCoprocessor: |
| 8306 | case clang::ARM::BI_MoveToCoprocessor2: { |
| 8307 | Function *F = CGM.getIntrinsic(BuiltinID == clang::ARM::BI_MoveToCoprocessor |
| 8308 | ? Intrinsic::arm_mcr |
| 8309 | : Intrinsic::arm_mcr2); |
| 8310 | return Builder.CreateCall(F, {Ops[1], Ops[2], Ops[0], |
| 8311 | Ops[3], Ops[4], Ops[5]}); |
| 8312 | } |
| 8313 | } |
| 8314 | |
| 8315 | // Get the last argument, which specifies the vector type. |
| 8316 | assert(HasExtraArg)(static_cast <bool> (HasExtraArg) ? void (0) : __assert_fail ("HasExtraArg", "clang/lib/CodeGen/CGBuiltin.cpp", 8316, __extension__ __PRETTY_FUNCTION__)); |
| 8317 | const Expr *Arg = E->getArg(E->getNumArgs()-1); |
| 8318 | std::optional<llvm::APSInt> Result = |
| 8319 | Arg->getIntegerConstantExpr(getContext()); |
| 8320 | if (!Result) |
| 8321 | return nullptr; |
| 8322 | |
| 8323 | if (BuiltinID == clang::ARM::BI__builtin_arm_vcvtr_f || |
| 8324 | BuiltinID == clang::ARM::BI__builtin_arm_vcvtr_d) { |
| 8325 | // Determine the overloaded type of this builtin. |
| 8326 | llvm::Type *Ty; |
| 8327 | if (BuiltinID == clang::ARM::BI__builtin_arm_vcvtr_f) |
| 8328 | Ty = FloatTy; |
| 8329 | else |
| 8330 | Ty = DoubleTy; |
| 8331 | |
| 8332 | // Determine whether this is an unsigned conversion or not. |
| 8333 | bool usgn = Result->getZExtValue() == 1; |
| 8334 | unsigned Int = usgn ? Intrinsic::arm_vcvtru : Intrinsic::arm_vcvtr; |
| 8335 | |
| 8336 | // Call the appropriate intrinsic. |
| 8337 | Function *F = CGM.getIntrinsic(Int, Ty); |
| 8338 | return Builder.CreateCall(F, Ops, "vcvtr"); |
| 8339 | } |
| 8340 | |
| 8341 | // Determine the type of this overloaded NEON intrinsic. |
| 8342 | NeonTypeFlags Type = Result->getZExtValue(); |
| 8343 | bool usgn = Type.isUnsigned(); |
| 8344 | bool rightShift = false; |
| 8345 | |
| 8346 | llvm::FixedVectorType *VTy = |
| 8347 | GetNeonType(this, Type, getTarget().hasLegalHalfType(), false, |
| 8348 | getTarget().hasBFloat16Type()); |
| 8349 | llvm::Type *Ty = VTy; |
| 8350 | if (!Ty) |
| 8351 | return nullptr; |
| 8352 | |
| 8353 | // Many NEON builtins have identical semantics and uses in ARM and |
| 8354 | // AArch64. Emit these in a single function. |
| 8355 | auto IntrinsicMap = ArrayRef(ARMSIMDIntrinsicMap); |
| 8356 | const ARMVectorIntrinsicInfo *Builtin = findARMVectorIntrinsicInMap( |
| 8357 | IntrinsicMap, BuiltinID, NEONSIMDIntrinsicsProvenSorted); |
| 8358 | if (Builtin) |
| 8359 | return EmitCommonNeonBuiltinExpr( |
| 8360 | Builtin->BuiltinID, Builtin->LLVMIntrinsic, Builtin->AltLLVMIntrinsic, |
| 8361 | Builtin->NameHint, Builtin->TypeModifier, E, Ops, PtrOp0, PtrOp1, Arch); |
| 8362 | |
| 8363 | unsigned Int; |
| 8364 | switch (BuiltinID) { |
| 8365 | default: return nullptr; |
| 8366 | case NEON::BI__builtin_neon_vld1q_lane_v: |
| 8367 | // Handle 64-bit integer elements as a special case. Use shuffles of |
| 8368 | // one-element vectors to avoid poor code for i64 in the backend. |
| 8369 | if (VTy->getElementType()->isIntegerTy(64)) { |
| 8370 | // Extract the other lane. |
| 8371 | Ops[1] = Builder.CreateBitCast(Ops[1], Ty); |
| 8372 | int Lane = cast<ConstantInt>(Ops[2])->getZExtValue(); |
| 8373 | Value *SV = llvm::ConstantVector::get(ConstantInt::get(Int32Ty, 1-Lane)); |
| 8374 | Ops[1] = Builder.CreateShuffleVector(Ops[1], Ops[1], SV); |
| 8375 | // Load the value as a one-element vector. |
| 8376 | Ty = llvm::FixedVectorType::get(VTy->getElementType(), 1); |
| 8377 | llvm::Type *Tys[] = {Ty, Int8PtrTy}; |
| 8378 | Function *F = CGM.getIntrinsic(Intrinsic::arm_neon_vld1, Tys); |
| 8379 | Value *Align = getAlignmentValue32(PtrOp0); |
| 8380 | Value *Ld = Builder.CreateCall(F, {Ops[0], Align}); |
| 8381 | // Combine them. |
| 8382 | int Indices[] = {1 - Lane, Lane}; |
| 8383 | return Builder.CreateShuffleVector(Ops[1], Ld, Indices, "vld1q_lane"); |
| 8384 | } |
| 8385 | [[fallthrough]]; |
| 8386 | case NEON::BI__builtin_neon_vld1_lane_v: { |
| 8387 | Ops[1] = Builder.CreateBitCast(Ops[1], Ty); |
| 8388 | PtrOp0 = Builder.CreateElementBitCast(PtrOp0, VTy->getElementType()); |
| 8389 | Value *Ld = Builder.CreateLoad(PtrOp0); |
| 8390 | return Builder.CreateInsertElement(Ops[1], Ld, Ops[2], "vld1_lane"); |
| 8391 | } |
| 8392 | case NEON::BI__builtin_neon_vqrshrn_n_v: |
| 8393 | Int = |
| 8394 | usgn ? Intrinsic::arm_neon_vqrshiftnu : Intrinsic::arm_neon_vqrshiftns; |
| 8395 | return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqrshrn_n", |
| 8396 | 1, true); |
| 8397 | case NEON::BI__builtin_neon_vqrshrun_n_v: |
| 8398 | return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vqrshiftnsu, Ty), |
| 8399 | Ops, "vqrshrun_n", 1, true); |
| 8400 | case NEON::BI__builtin_neon_vqshrn_n_v: |
| 8401 | Int = usgn ? Intrinsic::arm_neon_vqshiftnu : Intrinsic::arm_neon_vqshiftns; |
| 8402 | return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqshrn_n", |
| 8403 | 1, true); |
| 8404 | case NEON::BI__builtin_neon_vqshrun_n_v: |
| 8405 | return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vqshiftnsu, Ty), |
| 8406 | Ops, "vqshrun_n", 1, true); |
| 8407 | case NEON::BI__builtin_neon_vrecpe_v: |
| 8408 | case NEON::BI__builtin_neon_vrecpeq_v: |
| 8409 | return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vrecpe, Ty), |
| 8410 | Ops, "vrecpe"); |
| 8411 | case NEON::BI__builtin_neon_vrshrn_n_v: |
| 8412 | return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vrshiftn, Ty), |
| 8413 | Ops, "vrshrn_n", 1, true); |
| 8414 | case NEON::BI__builtin_neon_vrsra_n_v: |
| 8415 | case NEON::BI__builtin_neon_vrsraq_n_v: |
| 8416 | Ops[0] = Builder.CreateBitCast(Ops[0], Ty); |
| 8417 | Ops[1] = Builder.CreateBitCast(Ops[1], Ty); |
| 8418 | Ops[2] = EmitNeonShiftVector(Ops[2], Ty, true); |
| 8419 | Int = usgn ? Intrinsic::arm_neon_vrshiftu : Intrinsic::arm_neon_vrshifts; |
| 8420 | Ops[1] = Builder.CreateCall(CGM.getIntrinsic(Int, Ty), {Ops[1], Ops[2]}); |
| 8421 | return Builder.CreateAdd(Ops[0], Ops[1], "vrsra_n"); |
| 8422 | case NEON::BI__builtin_neon_vsri_n_v: |
| 8423 | case NEON::BI__builtin_neon_vsriq_n_v: |
| 8424 | rightShift = true; |
| 8425 | [[fallthrough]]; |
| 8426 | case NEON::BI__builtin_neon_vsli_n_v: |
| 8427 | case NEON::BI__builtin_neon_vsliq_n_v: |
| 8428 | Ops[2] = EmitNeonShiftVector(Ops[2], Ty, rightShift); |
| 8429 | return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vshiftins, Ty), |
| 8430 | Ops, "vsli_n"); |
| 8431 | case NEON::BI__builtin_neon_vsra_n_v: |
| 8432 | case NEON::BI__builtin_neon_vsraq_n_v: |
| 8433 | Ops[0] = Builder.CreateBitCast(Ops[0], Ty); |
| 8434 | Ops[1] = EmitNeonRShiftImm(Ops[1], Ops[2], Ty, usgn, "vsra_n"); |
| 8435 | return Builder.CreateAdd(Ops[0], Ops[1]); |
| 8436 | case NEON::BI__builtin_neon_vst1q_lane_v: |
| 8437 | // Handle 64-bit integer elements as a special case. Use a shuffle to get |
| 8438 | // a one-element vector and avoid poor code for i64 in the backend. |
| 8439 | if (VTy->getElementType()->isIntegerTy(64)) { |
| 8440 | Ops[1] = Builder.CreateBitCast(Ops[1], Ty); |
| 8441 | Value *SV = llvm::ConstantVector::get(cast<llvm::Constant>(Ops[2])); |
| 8442 | Ops[1] = Builder.CreateShuffleVector(Ops[1], Ops[1], SV); |
| 8443 | Ops[2] = getAlignmentValue32(PtrOp0); |
| 8444 | llvm::Type *Tys[] = {Int8PtrTy, Ops[1]->getType()}; |
| 8445 | return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::arm_neon_vst1, |
| 8446 | Tys), Ops); |
| 8447 | } |
| 8448 | [[fallthrough]]; |
| 8449 | case NEON::BI__builtin_neon_vst1_lane_v: { |
| 8450 | Ops[1] = Builder.CreateBitCast(Ops[1], Ty); |
| 8451 | Ops[1] = Builder.CreateExtractElement(Ops[1], Ops[2]); |
| 8452 | auto St = Builder.CreateStore( |
| 8453 | Ops[1], Builder.CreateElementBitCast(PtrOp0, Ops[1]->getType())); |
| 8454 | return St; |
| 8455 | } |
| 8456 | case NEON::BI__builtin_neon_vtbl1_v: |
| 8457 | return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbl1), |
| 8458 | Ops, "vtbl1"); |
| 8459 | case NEON::BI__builtin_neon_vtbl2_v: |
| 8460 | return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbl2), |
| 8461 | Ops, "vtbl2"); |
| 8462 | case NEON::BI__builtin_neon_vtbl3_v: |
| 8463 | return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbl3), |
| 8464 | Ops, "vtbl3"); |
| 8465 | case NEON::BI__builtin_neon_vtbl4_v: |
| 8466 | return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbl4), |
| 8467 | Ops, "vtbl4"); |
| 8468 | case NEON::BI__builtin_neon_vtbx1_v: |
| 8469 | return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbx1), |
| 8470 | Ops, "vtbx1"); |
| 8471 | case NEON::BI__builtin_neon_vtbx2_v: |
| 8472 | return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbx2), |
| 8473 | Ops, "vtbx2"); |
| 8474 | case NEON::BI__builtin_neon_vtbx3_v: |
| 8475 | return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbx3), |
| 8476 | Ops, "vtbx3"); |
| 8477 | case NEON::BI__builtin_neon_vtbx4_v: |
| 8478 | return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbx4), |
| 8479 | Ops, "vtbx4"); |
| 8480 | } |
| 8481 | } |
| 8482 | |
| 8483 | template<typename Integer> |
| 8484 | static Integer GetIntegerConstantValue(const Expr *E, ASTContext &Context) { |
| 8485 | return E->getIntegerConstantExpr(Context)->getExtValue(); |
| 8486 | } |
| 8487 | |
| 8488 | static llvm::Value *SignOrZeroExtend(CGBuilderTy &Builder, llvm::Value *V, |
| 8489 | llvm::Type *T, bool Unsigned) { |
| 8490 | // Helper function called by Tablegen-constructed ARM MVE builtin codegen, |
| 8491 | // which finds it convenient to specify signed/unsigned as a boolean flag. |
| 8492 | return Unsigned ? Builder.CreateZExt(V, T) : Builder.CreateSExt(V, T); |
| 8493 | } |
| 8494 | |
| 8495 | static llvm::Value *MVEImmediateShr(CGBuilderTy &Builder, llvm::Value *V, |
| 8496 | uint32_t Shift, bool Unsigned) { |
| 8497 | // MVE helper function for integer shift right. This must handle signed vs |
| 8498 | // unsigned, and also deal specially with the case where the shift count is |
| 8499 | // equal to the lane size. In LLVM IR, an LShr with that parameter would be |
| 8500 | // undefined behavior, but in MVE it's legal, so we must convert it to code |
| 8501 | // that is not undefined in IR. |
| 8502 | unsigned LaneBits = cast<llvm::VectorType>(V->getType()) |
| 8503 | ->getElementType() |
| 8504 | ->getPrimitiveSizeInBits(); |
| 8505 | if (Shift == LaneBits) { |
| 8506 | // An unsigned shift of the full lane size always generates zero, so we can |
| 8507 | // simply emit a zero vector. A signed shift of the full lane size does the |
| 8508 | // same thing as shifting by one bit fewer. |
| 8509 | if (Unsigned) |
| 8510 | return llvm::Constant::getNullValue(V->getType()); |
| 8511 | else |
| 8512 | --Shift; |
| 8513 | } |
| 8514 | return Unsigned ? Builder.CreateLShr(V, Shift) : Builder.CreateAShr(V, Shift); |
| 8515 | } |
| 8516 | |
| 8517 | static llvm::Value *ARMMVEVectorSplat(CGBuilderTy &Builder, llvm::Value *V) { |
| 8518 | // MVE-specific helper function for a vector splat, which infers the element |
| 8519 | // count of the output vector by knowing that MVE vectors are all 128 bits |
| 8520 | // wide. |
| 8521 | unsigned Elements = 128 / V->getType()->getPrimitiveSizeInBits(); |
| 8522 | return Builder.CreateVectorSplat(Elements, V); |
| 8523 | } |
| 8524 | |
| 8525 | static llvm::Value *ARMMVEVectorReinterpret(CGBuilderTy &Builder, |
| 8526 | CodeGenFunction *CGF, |
| 8527 | llvm::Value *V, |
| 8528 | llvm::Type *DestType) { |
| 8529 | // Convert one MVE vector type into another by reinterpreting its in-register |
| 8530 | // format. |
| 8531 | // |
| 8532 | // Little-endian, this is identical to a bitcast (which reinterprets the |
| 8533 | // memory format). But big-endian, they're not necessarily the same, because |
| 8534 | // the register and memory formats map to each other differently depending on |
| 8535 | // the lane size. |
| 8536 | // |
| 8537 | // We generate a bitcast whenever we can (if we're little-endian, or if the |
| 8538 | // lane sizes are the same anyway). Otherwise we fall back to an IR intrinsic |
| 8539 | // that performs the different kind of reinterpretation. |
| 8540 | if (CGF->getTarget().isBigEndian() && |
| 8541 | V->getType()->getScalarSizeInBits() != DestType->getScalarSizeInBits()) { |
| 8542 | return Builder.CreateCall( |
| 8543 | CGF->CGM.getIntrinsic(Intrinsic::arm_mve_vreinterpretq, |
| 8544 | {DestType, V->getType()}), |
| 8545 | V); |
| 8546 | } else { |
| 8547 | return Builder.CreateBitCast(V, DestType); |
| 8548 | } |
| 8549 | } |
| 8550 | |
| 8551 | static llvm::Value *VectorUnzip(CGBuilderTy &Builder, llvm::Value *V, bool Odd) { |
| 8552 | // Make a shufflevector that extracts every other element of a vector (evens |
| 8553 | // or odds, as desired). |
| 8554 | SmallVector<int, 16> Indices; |
| 8555 | unsigned InputElements = |
| 8556 | cast<llvm::FixedVectorType>(V->getType())->getNumElements(); |
| 8557 | for (unsigned i = 0; i < InputElements; i += 2) |
| 8558 | Indices.push_back(i + Odd); |
| 8559 | return Builder.CreateShuffleVector(V, Indices); |
| 8560 | } |
| 8561 | |
| 8562 | static llvm::Value *VectorZip(CGBuilderTy &Builder, llvm::Value *V0, |
| 8563 | llvm::Value *V1) { |
| 8564 | // Make a shufflevector that interleaves two vectors element by element. |
| 8565 | assert(V0->getType() == V1->getType() && "Can't zip different vector types")(static_cast <bool> (V0->getType() == V1->getType () && "Can't zip different vector types") ? void (0) : __assert_fail ("V0->getType() == V1->getType() && \"Can't zip different vector types\"" , "clang/lib/CodeGen/CGBuiltin.cpp", 8565, __extension__ __PRETTY_FUNCTION__ )); |
| 8566 | SmallVector<int, 16> Indices; |
| 8567 | unsigned InputElements = |
| 8568 | cast<llvm::FixedVectorType>(V0->getType())->getNumElements(); |
| 8569 | for (unsigned i = 0; i < InputElements; i++) { |
| 8570 | Indices.push_back(i); |
| 8571 | Indices.push_back(i + InputElements); |
| 8572 | } |
| 8573 | return Builder.CreateShuffleVector(V0, V1, Indices); |
| 8574 | } |
| 8575 | |
| 8576 | template<unsigned HighBit, unsigned OtherBits> |
| 8577 | static llvm::Value *ARMMVEConstantSplat(CGBuilderTy &Builder, llvm::Type *VT) { |
| 8578 | // MVE-specific helper function to make a vector splat of a constant such as |
| 8579 | // UINT_MAX or INT_MIN, in which all bits below the highest one are equal. |
| 8580 | llvm::Type *T = cast<llvm::VectorType>(VT)->getElementType(); |
| 8581 | unsigned LaneBits = T->getPrimitiveSizeInBits(); |
| 8582 | uint32_t Value = HighBit << (LaneBits - 1); |
| 8583 | if (OtherBits) |
| 8584 | Value |= (1UL << (LaneBits - 1)) - 1; |
| 8585 | llvm::Value *Lane = llvm::ConstantInt::get(T, Value); |
| 8586 | return ARMMVEVectorSplat(Builder, Lane); |
| 8587 | } |
| 8588 | |
| 8589 | static llvm::Value *ARMMVEVectorElementReverse(CGBuilderTy &Builder, |
| 8590 | llvm::Value *V, |
| 8591 | unsigned ReverseWidth) { |
| 8592 | // MVE-specific helper function which reverses the elements of a |
| 8593 | // vector within every (ReverseWidth)-bit collection of lanes. |
| 8594 | SmallVector<int, 16> Indices; |
| 8595 | unsigned LaneSize = V->getType()->getScalarSizeInBits(); |
| 8596 | unsigned Elements = 128 / LaneSize; |
| 8597 | unsigned Mask = ReverseWidth / LaneSize - 1; |
| 8598 | for (unsigned i = 0; i < Elements; i++) |
| 8599 | Indices.push_back(i ^ Mask); |
| 8600 | return Builder.CreateShuffleVector(V, Indices); |
| 8601 | } |
| 8602 | |
| 8603 | Value *CodeGenFunction::EmitARMMVEBuiltinExpr(unsigned BuiltinID, |
| 8604 | const CallExpr *E, |
| 8605 | ReturnValueSlot ReturnValue, |
| 8606 | llvm::Triple::ArchType Arch) { |
| 8607 | enum class CustomCodeGen { VLD24, VST24 } CustomCodeGenType; |
| 8608 | Intrinsic::ID IRIntr; |
| 8609 | unsigned NumVectors; |
| 8610 | |
| 8611 | // Code autogenerated by Tablegen will handle all the simple builtins. |
| 8612 | switch (BuiltinID) { |
| 8613 | #include "clang/Basic/arm_mve_builtin_cg.inc" |
| 8614 | |
| 8615 | // If we didn't match an MVE builtin id at all, go back to the |
| 8616 | // main EmitARMBuiltinExpr. |
| 8617 | default: |
| 8618 | return nullptr; |
| 8619 | } |
| 8620 | |
| 8621 | // Anything that breaks from that switch is an MVE builtin that |
| 8622 | // needs handwritten code to generate. |
| 8623 | |
| 8624 | switch (CustomCodeGenType) { |
| 8625 | |
| 8626 | case CustomCodeGen::VLD24: { |
| 8627 | llvm::SmallVector<Value *, 4> Ops; |
| 8628 | llvm::SmallVector<llvm::Type *, 4> Tys; |
| 8629 | |
| 8630 | auto MvecCType = E->getType(); |
| 8631 | auto MvecLType = ConvertType(MvecCType); |
| 8632 | assert(MvecLType->isStructTy() &&(static_cast <bool> (MvecLType->isStructTy() && "Return type for vld[24]q should be a struct") ? void (0) : __assert_fail ("MvecLType->isStructTy() && \"Return type for vld[24]q should be a struct\"" , "clang/lib/CodeGen/CGBuiltin.cpp", 8633, __extension__ __PRETTY_FUNCTION__ )) |
| 8633 | "Return type for vld[24]q should be a struct")(static_cast <bool> (MvecLType->isStructTy() && "Return type for vld[24]q should be a struct") ? void (0) : __assert_fail ("MvecLType->isStructTy() && \"Return type for vld[24]q should be a struct\"" , "clang/lib/CodeGen/CGBuiltin.cpp", 8633, __extension__ __PRETTY_FUNCTION__ )); |
| 8634 | assert(MvecLType->getStructNumElements() == 1 &&(static_cast <bool> (MvecLType->getStructNumElements () == 1 && "Return-type struct for vld[24]q should have one element" ) ? void (0) : __assert_fail ("MvecLType->getStructNumElements() == 1 && \"Return-type struct for vld[24]q should have one element\"" , "clang/lib/CodeGen/CGBuiltin.cpp", 8635, __extension__ __PRETTY_FUNCTION__ )) |
| 8635 | "Return-type struct for vld[24]q should have one element")(static_cast <bool> (MvecLType->getStructNumElements () == 1 && "Return-type struct for vld[24]q should have one element" ) ? void (0) : __assert_fail ("MvecLType->getStructNumElements() == 1 && \"Return-type struct for vld[24]q should have one element\"" , "clang/lib/CodeGen/CGBuiltin.cpp", 8635, __extension__ __PRETTY_FUNCTION__ )); |
| 8636 | auto MvecLTypeInner = MvecLType->getStructElementType(0); |
| 8637 | assert(MvecLTypeInner->isArrayTy() &&(static_cast <bool> (MvecLTypeInner->isArrayTy() && "Return-type struct for vld[24]q should contain an array") ? void (0) : __assert_fail ("MvecLTypeInner->isArrayTy() && \"Return-type struct for vld[24]q should contain an array\"" , "clang/lib/CodeGen/CGBuiltin.cpp", 8638, __extension__ __PRETTY_FUNCTION__ )) |
| 8638 | "Return-type struct for vld[24]q should contain an array")(static_cast <bool> (MvecLTypeInner->isArrayTy() && "Return-type struct for vld[24]q should contain an array") ? void (0) : __assert_fail ("MvecLTypeInner->isArrayTy() && \"Return-type struct for vld[24]q should contain an array\"" , "clang/lib/CodeGen/CGBuiltin.cpp", 8638, __extension__ __PRETTY_FUNCTION__ )); |
| 8639 | assert(MvecLTypeInner->getArrayNumElements() == NumVectors &&(static_cast <bool> (MvecLTypeInner->getArrayNumElements () == NumVectors && "Array member of return-type struct vld[24]q has wrong length" ) ? void (0) : __assert_fail ("MvecLTypeInner->getArrayNumElements() == NumVectors && \"Array member of return-type struct vld[24]q has wrong length\"" , "clang/lib/CodeGen/CGBuiltin.cpp", 8640, __extension__ __PRETTY_FUNCTION__ )) |
| 8640 | "Array member of return-type struct vld[24]q has wrong length")(static_cast <bool> (MvecLTypeInner->getArrayNumElements () == NumVectors && "Array member of return-type struct vld[24]q has wrong length" ) ? void (0) : __assert_fail ("MvecLTypeInner->getArrayNumElements() == NumVectors && \"Array member of return-type struct vld[24]q has wrong length\"" , "clang/lib/CodeGen/CGBuiltin.cpp", 8640, __extension__ __PRETTY_FUNCTION__ )); |
| 8641 | auto VecLType = MvecLTypeInner->getArrayElementType(); |
| 8642 | |
| 8643 | Tys.push_back(VecLType); |
| 8644 | |
| 8645 | auto Addr = E->getArg(0); |
| 8646 | Ops.push_back(EmitScalarExpr(Addr)); |
| 8647 | Tys.push_back(ConvertType(Addr->getType())); |
| 8648 | |
| 8649 | Function *F = CGM.getIntrinsic(IRIntr, ArrayRef(Tys)); |
| 8650 | Value *LoadResult = Builder.CreateCall(F, Ops); |
| 8651 | Value *MvecOut = PoisonValue::get(MvecLType); |
| 8652 | for (unsigned i = 0; i < NumVectors; ++i) { |
| 8653 | Value *Vec = Builder.CreateExtractValue(LoadResult, i); |
| 8654 | MvecOut = Builder.CreateInsertValue(MvecOut, Vec, {0, i}); |
| 8655 | } |
| 8656 | |
| 8657 | if (ReturnValue.isNull()) |
| 8658 | return MvecOut; |
| 8659 | else |
| 8660 | return Builder.CreateStore(MvecOut, ReturnValue.getValue()); |
| 8661 | } |
| 8662 | |
| 8663 | case CustomCodeGen::VST24: { |
| 8664 | llvm::SmallVector<Value *, 4> Ops; |
| 8665 | llvm::SmallVector<llvm::Type *, 4> Tys; |
| 8666 | |
| 8667 | auto Addr = E->getArg(0); |
| 8668 | Ops.push_back(EmitScalarExpr(Addr)); |
| 8669 | Tys.push_back(ConvertType(Addr->getType())); |
| 8670 | |
| 8671 | auto MvecCType = E->getArg(1)->getType(); |
| 8672 | auto MvecLType = ConvertType(MvecCType); |
| 8673 | assert(MvecLType->isStructTy() && "Data type for vst2q should be a struct")(static_cast <bool> (MvecLType->isStructTy() && "Data type for vst2q should be a struct") ? void (0) : __assert_fail ("MvecLType->isStructTy() && \"Data type for vst2q should be a struct\"" , "clang/lib/CodeGen/CGBuiltin.cpp", 8673, __extension__ __PRETTY_FUNCTION__ )); |
| 8674 | assert(MvecLType->getStructNumElements() == 1 &&(static_cast <bool> (MvecLType->getStructNumElements () == 1 && "Data-type struct for vst2q should have one element" ) ? void (0) : __assert_fail ("MvecLType->getStructNumElements() == 1 && \"Data-type struct for vst2q should have one element\"" , "clang/lib/CodeGen/CGBuiltin.cpp", 8675, __extension__ __PRETTY_FUNCTION__ )) |
| 8675 | "Data-type struct for vst2q should have one element")(static_cast <bool> (MvecLType->getStructNumElements () == 1 && "Data-type struct for vst2q should have one element" ) ? void (0) : __assert_fail ("MvecLType->getStructNumElements() == 1 && \"Data-type struct for vst2q should have one element\"" , "clang/lib/CodeGen/CGBuiltin.cpp", 8675, __extension__ __PRETTY_FUNCTION__ )); |
| 8676 | auto MvecLTypeInner = MvecLType->getStructElementType(0); |
| 8677 | assert(MvecLTypeInner->isArrayTy() &&(static_cast <bool> (MvecLTypeInner->isArrayTy() && "Data-type struct for vst2q should contain an array") ? void (0) : __assert_fail ("MvecLTypeInner->isArrayTy() && \"Data-type struct for vst2q should contain an array\"" , "clang/lib/CodeGen/CGBuiltin.cpp", 8678, __extension__ __PRETTY_FUNCTION__ )) |
| 8678 | "Data-type struct for vst2q should contain an array")(static_cast <bool> (MvecLTypeInner->isArrayTy() && "Data-type struct for vst2q should contain an array") ? void (0) : __assert_fail ("MvecLTypeInner->isArrayTy() && \"Data-type struct for vst2q should contain an array\"" , "clang/lib/CodeGen/CGBuiltin.cpp", 8678, __extension__ __PRETTY_FUNCTION__ )); |
| 8679 | assert(MvecLTypeInner->getArrayNumElements() == NumVectors &&(static_cast <bool> (MvecLTypeInner->getArrayNumElements () == NumVectors && "Array member of return-type struct vld[24]q has wrong length" ) ? void (0) : __assert_fail ("MvecLTypeInner->getArrayNumElements() == NumVectors && \"Array member of return-type struct vld[24]q has wrong length\"" , "clang/lib/CodeGen/CGBuiltin.cpp", 8680, __extension__ __PRETTY_FUNCTION__ )) |
| 8680 | "Array member of return-type struct vld[24]q has wrong length")(static_cast <bool> (MvecLTypeInner->getArrayNumElements () == NumVectors && "Array member of return-type struct vld[24]q has wrong length" ) ? void (0) : __assert_fail ("MvecLTypeInner->getArrayNumElements() == NumVectors && \"Array member of return-type struct vld[24]q has wrong length\"" , "clang/lib/CodeGen/CGBuiltin.cpp", 8680, __extension__ __PRETTY_FUNCTION__ )); |
| 8681 | auto VecLType = MvecLTypeInner->getArrayElementType(); |
| 8682 | |
| 8683 | Tys.push_back(VecLType); |
| 8684 | |
| 8685 | AggValueSlot MvecSlot = CreateAggTemp(MvecCType); |
| 8686 | EmitAggExpr(E->getArg(1), MvecSlot); |
| 8687 | auto Mvec = Builder.CreateLoad(MvecSlot.getAddress()); |
| 8688 | for (unsigned i = 0; i < NumVectors; i++) |
| 8689 | Ops.push_back(Builder.CreateExtractValue(Mvec, {0, i})); |
| 8690 | |
| 8691 | Function *F = CGM.getIntrinsic(IRIntr, ArrayRef(Tys)); |
| 8692 | Value *ToReturn = nullptr; |
| 8693 | for (unsigned i = 0; i < NumVectors; i++) { |
| 8694 | Ops.push_back(llvm::ConstantInt::get(Int32Ty, i)); |
| 8695 | ToReturn = Builder.CreateCall(F, Ops); |
| 8696 | Ops.pop_back(); |
| 8697 | } |
| 8698 | return ToReturn; |
| 8699 | } |
| 8700 | } |
| 8701 | llvm_unreachable("unknown custom codegen type.")::llvm::llvm_unreachable_internal("unknown custom codegen type." , "clang/lib/CodeGen/CGBuiltin.cpp", 8701); |
| 8702 | } |
| 8703 | |
| 8704 | Value *CodeGenFunction::EmitARMCDEBuiltinExpr(unsigned BuiltinID, |
| 8705 | const CallExpr *E, |
| 8706 | ReturnValueSlot ReturnValue, |
| 8707 | llvm::Triple::ArchType Arch) { |
| 8708 | switch (BuiltinID) { |
| 8709 | default: |
| 8710 | return nullptr; |
| 8711 | #include "clang/Basic/arm_cde_builtin_cg.inc" |
| 8712 | } |
| 8713 | } |
| 8714 | |
| 8715 | static Value *EmitAArch64TblBuiltinExpr(CodeGenFunction &CGF, unsigned BuiltinID, |
| 8716 | const CallExpr *E, |
| 8717 | SmallVectorImpl<Value *> &Ops, |
| 8718 | llvm::Triple::ArchType Arch) { |
| 8719 | unsigned int Int = 0; |
| 8720 | const char *s = nullptr; |
| 8721 | |
| 8722 | switch (BuiltinID) { |
| 8723 | default: |
| 8724 | return nullptr; |
| 8725 | case NEON::BI__builtin_neon_vtbl1_v: |
| 8726 | case NEON::BI__builtin_neon_vqtbl1_v: |
| 8727 | case NEON::BI__builtin_neon_vqtbl1q_v: |
| 8728 | case NEON::BI__builtin_neon_vtbl2_v: |
| 8729 | case NEON::BI__builtin_neon_vqtbl2_v: |
| 8730 | case NEON::BI__builtin_neon_vqtbl2q_v: |
| 8731 | case NEON::BI__builtin_neon_vtbl3_v: |
| 8732 | case NEON::BI__builtin_neon_vqtbl3_v: |
| 8733 | case NEON::BI__builtin_neon_vqtbl3q_v: |
| 8734 | case NEON::BI__builtin_neon_vtbl4_v: |
| 8735 | case NEON::BI__builtin_neon_vqtbl4_v: |
| 8736 | case NEON::BI__builtin_neon_vqtbl4q_v: |
| 8737 | break; |
| 8738 | case NEON::BI__builtin_neon_vtbx1_v: |
| 8739 | case NEON::BI__builtin_neon_vqtbx1_v: |
| 8740 | case NEON::BI__builtin_neon_vqtbx1q_v: |
| 8741 | case NEON::BI__builtin_neon_vtbx2_v: |
| 8742 | case NEON::BI__builtin_neon_vqtbx2_v: |
| 8743 | case NEON::BI__builtin_neon_vqtbx2q_v: |
| 8744 | case NEON::BI__builtin_neon_vtbx3_v: |
| 8745 | case NEON::BI__builtin_neon_vqtbx3_v: |
| 8746 | case NEON::BI__builtin_neon_vqtbx3q_v: |
| 8747 | case NEON::BI__builtin_neon_vtbx4_v: |
| 8748 | case NEON::BI__builtin_neon_vqtbx4_v: |
| 8749 | case NEON::BI__builtin_neon_vqtbx4q_v: |
| 8750 | break; |
| 8751 | } |
| 8752 | |
| 8753 | assert(E->getNumArgs() >= 3)(static_cast <bool> (E->getNumArgs() >= 3) ? void (0) : __assert_fail ("E->getNumArgs() >= 3", "clang/lib/CodeGen/CGBuiltin.cpp" , 8753, __extension__ __PRETTY_FUNCTION__)); |
| 8754 | |
| 8755 | // Get the last argument, which specifies the vector type. |
| 8756 | const Expr *Arg = E->getArg(E->getNumArgs() - 1); |
| 8757 | std::optional<llvm::APSInt> Result = |
| 8758 | Arg->getIntegerConstantExpr(CGF.getContext()); |
| 8759 | if (!Result) |
| 8760 | return nullptr; |
| 8761 | |
| 8762 | // Determine the type of this overloaded NEON intrinsic. |
| 8763 | NeonTypeFlags Type = Result->getZExtValue(); |
| 8764 | llvm::FixedVectorType *Ty = GetNeonType(&CGF, Type); |
| 8765 | if (!Ty) |
| 8766 | return nullptr; |
| 8767 | |
| 8768 | CodeGen::CGBuilderTy &Builder = CGF.Builder; |
| 8769 | |
| 8770 | // AArch64 scalar builtins are not overloaded, they do not have an extra |
| 8771 | // argument that specifies the vector type, need to handle each case. |
| 8772 | switch (BuiltinID) { |
| 8773 | case NEON::BI__builtin_neon_vtbl1_v: { |
| 8774 | return packTBLDVectorList(CGF, ArrayRef(Ops).slice(0, 1), nullptr, Ops[1], |
| 8775 | Ty, Intrinsic::aarch64_neon_tbl1, "vtbl1"); |
| 8776 | } |
| 8777 | case NEON::BI__builtin_neon_vtbl2_v: { |
| 8778 | return packTBLDVectorList(CGF, ArrayRef(Ops).slice(0, 2), nullptr, Ops[2], |
| 8779 | Ty, Intrinsic::aarch64_neon_tbl1, "vtbl1"); |
| 8780 | } |
| 8781 | case NEON::BI__builtin_neon_vtbl3_v: { |
| 8782 | return packTBLDVectorList(CGF, ArrayRef(Ops).slice(0, 3), nullptr, Ops[3], |
| 8783 | Ty, Intrinsic::aarch64_neon_tbl2, "vtbl2"); |
| 8784 | } |
| 8785 | case NEON::BI__builtin_neon_vtbl4_v: { |
| 8786 | return packTBLDVectorList(CGF, ArrayRef(Ops).slice(0, 4), nullptr, Ops[4], |
| 8787 | Ty, Intrinsic::aarch64_neon_tbl2, "vtbl2"); |
| 8788 | } |
| 8789 | case NEON::BI__builtin_neon_vtbx1_v: { |
| 8790 | Value *TblRes = |
| 8791 | packTBLDVectorList(CGF, ArrayRef(Ops).slice(1, 1), nullptr, Ops[2], Ty, |
| 8792 | Intrinsic::aarch64_neon_tbl1, "vtbl1"); |
| 8793 | |
| 8794 | llvm::Constant *EightV = ConstantInt::get(Ty, 8); |
| 8795 | Value *CmpRes = Builder.CreateICmp(ICmpInst::ICMP_UGE, Ops[2], EightV); |
| 8796 | CmpRes = Builder.CreateSExt(CmpRes, Ty); |
| 8797 | |
| 8798 | Value *EltsFromInput = Builder.CreateAnd(CmpRes, Ops[0]); |
| 8799 | Value *EltsFromTbl = Builder.CreateAnd(Builder.CreateNot(CmpRes), TblRes); |
| 8800 | return Builder.CreateOr(EltsFromInput, EltsFromTbl, "vtbx"); |
| 8801 | } |
| 8802 | case NEON::BI__builtin_neon_vtbx2_v: { |
| 8803 | return packTBLDVectorList(CGF, ArrayRef(Ops).slice(1, 2), Ops[0], Ops[3], |
| 8804 | Ty, Intrinsic::aarch64_neon_tbx1, "vtbx1"); |
| 8805 | } |
| 8806 | case NEON::BI__builtin_neon_vtbx3_v: { |
| 8807 | Value *TblRes = |
| 8808 | packTBLDVectorList(CGF, ArrayRef(Ops).slice(1, 3), nullptr, Ops[4], Ty, |
| 8809 | Intrinsic::aarch64_neon_tbl2, "vtbl2"); |
| 8810 | |
| 8811 | llvm::Constant *TwentyFourV = ConstantInt::get(Ty, 24); |
| 8812 | Value *CmpRes = Builder.CreateICmp(ICmpInst::ICMP_UGE, Ops[4], |
| 8813 | TwentyFourV); |
| 8814 | CmpRes = Builder.CreateSExt(CmpRes, Ty); |
| 8815 | |
| 8816 | Value *EltsFromInput = Builder.CreateAnd(CmpRes, Ops[0]); |
| 8817 | Value *EltsFromTbl = Builder.CreateAnd(Builder.CreateNot(CmpRes), TblRes); |
| 8818 | return Builder.CreateOr(EltsFromInput, EltsFromTbl, "vtbx"); |
| 8819 | } |
| 8820 | case NEON::BI__builtin_neon_vtbx4_v: { |
| 8821 | return packTBLDVectorList(CGF, ArrayRef(Ops).slice(1, 4), Ops[0], Ops[5], |
| 8822 | Ty, Intrinsic::aarch64_neon_tbx2, "vtbx2"); |
| 8823 | } |
| 8824 | case NEON::BI__builtin_neon_vqtbl1_v: |
| 8825 | case NEON::BI__builtin_neon_vqtbl1q_v: |
| 8826 | Int = Intrinsic::aarch64_neon_tbl1; s = "vtbl1"; break; |
| 8827 | case NEON::BI__builtin_neon_vqtbl2_v: |
| 8828 | case NEON::BI__builtin_neon_vqtbl2q_v: { |
| 8829 | Int = Intrinsic::aarch64_neon_tbl2; s = "vtbl2"; break; |
| 8830 | case NEON::BI__builtin_neon_vqtbl3_v: |
| 8831 | case NEON::BI__builtin_neon_vqtbl3q_v: |
| 8832 | Int = Intrinsic::aarch64_neon_tbl3; s = "vtbl3"; break; |
| 8833 | case NEON::BI__builtin_neon_vqtbl4_v: |
| 8834 | case NEON::BI__builtin_neon_vqtbl4q_v: |
| 8835 | Int = Intrinsic::aarch64_neon_tbl4; s = "vtbl4"; break; |
| 8836 | case NEON::BI__builtin_neon_vqtbx1_v: |
| 8837 | case NEON::BI__builtin_neon_vqtbx1q_v: |
| 8838 | Int = Intrinsic::aarch64_neon_tbx1; s = "vtbx1"; break; |
| 8839 | case NEON::BI__builtin_neon_vqtbx2_v: |
| 8840 | case NEON::BI__builtin_neon_vqtbx2q_v: |
| 8841 | Int = Intrinsic::aarch64_neon_tbx2; s = "vtbx2"; break; |
| 8842 | case NEON::BI__builtin_neon_vqtbx3_v: |
| 8843 | case NEON::BI__builtin_neon_vqtbx3q_v: |
| 8844 | Int = Intrinsic::aarch64_neon_tbx3; s = "vtbx3"; break; |
| 8845 | case NEON::BI__builtin_neon_vqtbx4_v: |
| 8846 | case NEON::BI__builtin_neon_vqtbx4q_v: |
| 8847 | Int = Intrinsic::aarch64_neon_tbx4; s = "vtbx4"; break; |
| 8848 | } |
| 8849 | } |
| 8850 | |
| 8851 | if (!Int) |
| 8852 | return nullptr; |
| 8853 | |
| 8854 | Function *F = CGF.CGM.getIntrinsic(Int, Ty); |
| 8855 | return CGF.EmitNeonCall(F, Ops, s); |
| 8856 | } |
| 8857 | |
| 8858 | Value *CodeGenFunction::vectorWrapScalar16(Value *Op) { |
| 8859 | auto *VTy = llvm::FixedVectorType::get(Int16Ty, 4); |
| 8860 | Op = Builder.CreateBitCast(Op, Int16Ty); |
| 8861 | Value *V = PoisonValue::get(VTy); |
| 8862 | llvm::Constant *CI = ConstantInt::get(SizeTy, 0); |
| 8863 | Op = Builder.CreateInsertElement(V, Op, CI); |
| 8864 | return Op; |
| 8865 | } |
| 8866 | |
| 8867 | /// SVEBuiltinMemEltTy - Returns the memory element type for this memory |
| 8868 | /// access builtin. Only required if it can't be inferred from the base pointer |
| 8869 | /// operand. |
| 8870 | llvm::Type *CodeGenFunction::SVEBuiltinMemEltTy(const SVETypeFlags &TypeFlags) { |
| 8871 | switch (TypeFlags.getMemEltType()) { |
| 8872 | case SVETypeFlags::MemEltTyDefault: |
| 8873 | return getEltType(TypeFlags); |
| 8874 | case SVETypeFlags::MemEltTyInt8: |
| 8875 | return Builder.getInt8Ty(); |
| 8876 | case SVETypeFlags::MemEltTyInt16: |
| 8877 | return Builder.getInt16Ty(); |
| 8878 | case SVETypeFlags::MemEltTyInt32: |
| 8879 | return Builder.getInt32Ty(); |
| 8880 | case SVETypeFlags::MemEltTyInt64: |
| 8881 | return Builder.getInt64Ty(); |
| 8882 | } |
| 8883 | llvm_unreachable("Unknown MemEltType")::llvm::llvm_unreachable_internal("Unknown MemEltType", "clang/lib/CodeGen/CGBuiltin.cpp" , 8883); |
| 8884 | } |
| 8885 | |
| 8886 | llvm::Type *CodeGenFunction::getEltType(const SVETypeFlags &TypeFlags) { |
| 8887 | switch (TypeFlags.getEltType()) { |
| 8888 | default: |
| 8889 | llvm_unreachable("Invalid SVETypeFlag!")::llvm::llvm_unreachable_internal("Invalid SVETypeFlag!", "clang/lib/CodeGen/CGBuiltin.cpp" , 8889); |
| 8890 | |
| 8891 | case SVETypeFlags::EltTyInt8: |
| 8892 | return Builder.getInt8Ty(); |
| 8893 | case SVETypeFlags::EltTyInt16: |
| 8894 | return Builder.getInt16Ty(); |
| 8895 | case SVETypeFlags::EltTyInt32: |
| 8896 | return Builder.getInt32Ty(); |
| 8897 | case SVETypeFlags::EltTyInt64: |
| 8898 | return Builder.getInt64Ty(); |
| 8899 | |
| 8900 | case SVETypeFlags::EltTyFloat16: |
| 8901 | return Builder.getHalfTy(); |
| 8902 | case SVETypeFlags::EltTyFloat32: |
| 8903 | return Builder.getFloatTy(); |
| 8904 | case SVETypeFlags::EltTyFloat64: |
| 8905 | return Builder.getDoubleTy(); |
| 8906 | |
| 8907 | case SVETypeFlags::EltTyBFloat16: |
| 8908 | return Builder.getBFloatTy(); |
| 8909 | |
| 8910 | case SVETypeFlags::EltTyBool8: |
| 8911 | case SVETypeFlags::EltTyBool16: |
| 8912 | case SVETypeFlags::EltTyBool32: |
| 8913 | case SVETypeFlags::EltTyBool64: |
| 8914 | return Builder.getInt1Ty(); |
| 8915 | } |
| 8916 | } |
| 8917 | |
| 8918 | // Return the llvm predicate vector type corresponding to the specified element |
| 8919 | // TypeFlags. |
| 8920 | llvm::ScalableVectorType * |
| 8921 | CodeGenFunction::getSVEPredType(const SVETypeFlags &TypeFlags) { |
| 8922 | switch (TypeFlags.getEltType()) { |
| 8923 | default: llvm_unreachable("Unhandled SVETypeFlag!")::llvm::llvm_unreachable_internal("Unhandled SVETypeFlag!", "clang/lib/CodeGen/CGBuiltin.cpp" , 8923); |
| 8924 | |
| 8925 | case SVETypeFlags::EltTyInt8: |
| 8926 | return llvm::ScalableVectorType::get(Builder.getInt1Ty(), 16); |
| 8927 | case SVETypeFlags::EltTyInt16: |
| 8928 | return llvm::ScalableVectorType::get(Builder.getInt1Ty(), 8); |
| 8929 | case SVETypeFlags::EltTyInt32: |
| 8930 | return llvm::ScalableVectorType::get(Builder.getInt1Ty(), 4); |
| 8931 | case SVETypeFlags::EltTyInt64: |
| 8932 | return llvm::ScalableVectorType::get(Builder.getInt1Ty(), 2); |
| 8933 | |
| 8934 | case SVETypeFlags::EltTyBFloat16: |
| 8935 | return llvm::ScalableVectorType::get(Builder.getInt1Ty(), 8); |
| 8936 | case SVETypeFlags::EltTyFloat16: |
| 8937 | return llvm::ScalableVectorType::get(Builder.getInt1Ty(), 8); |
| 8938 | case SVETypeFlags::EltTyFloat32: |
| 8939 | return llvm::ScalableVectorType::get(Builder.getInt1Ty(), 4); |
| 8940 | case SVETypeFlags::EltTyFloat64: |
| 8941 | return llvm::ScalableVectorType::get(Builder.getInt1Ty(), 2); |
| 8942 | |
| 8943 | case SVETypeFlags::EltTyBool8: |
| 8944 | return llvm::ScalableVectorType::get(Builder.getInt1Ty(), 16); |
| 8945 | case SVETypeFlags::EltTyBool16: |
| 8946 | return llvm::ScalableVectorType::get(Builder.getInt1Ty(), 8); |
| 8947 | case SVETypeFlags::EltTyBool32: |
| 8948 | return llvm::ScalableVectorType::get(Builder.getInt1Ty(), 4); |
| 8949 | case SVETypeFlags::EltTyBool64: |
| 8950 | return llvm::ScalableVectorType::get(Builder.getInt1Ty(), 2); |
| 8951 | } |
| 8952 | } |
| 8953 | |
| 8954 | // Return the llvm vector type corresponding to the specified element TypeFlags. |
| 8955 | llvm::ScalableVectorType * |
| 8956 | CodeGenFunction::getSVEType(const SVETypeFlags &TypeFlags) { |
| 8957 | switch (TypeFlags.getEltType()) { |
| 8958 | default: |
| 8959 | llvm_unreachable("Invalid SVETypeFlag!")::llvm::llvm_unreachable_internal("Invalid SVETypeFlag!", "clang/lib/CodeGen/CGBuiltin.cpp" , 8959); |
| 8960 | |
| 8961 | case SVETypeFlags::EltTyInt8: |
| 8962 | return llvm::ScalableVectorType::get(Builder.getInt8Ty(), 16); |
| 8963 | case SVETypeFlags::EltTyInt16: |
| 8964 | return llvm::ScalableVectorType::get(Builder.getInt16Ty(), 8); |
| 8965 | case SVETypeFlags::EltTyInt32: |
| 8966 | return llvm::ScalableVectorType::get(Builder.getInt32Ty(), 4); |
| 8967 | case SVETypeFlags::EltTyInt64: |
| 8968 | return llvm::ScalableVectorType::get(Builder.getInt64Ty(), 2); |
| 8969 | |
| 8970 | case SVETypeFlags::EltTyFloat16: |
| 8971 | return llvm::ScalableVectorType::get(Builder.getHalfTy(), 8); |
| 8972 | case SVETypeFlags::EltTyBFloat16: |
| 8973 | return llvm::ScalableVectorType::get(Builder.getBFloatTy(), 8); |
| 8974 | case SVETypeFlags::EltTyFloat32: |
| 8975 | return llvm::ScalableVectorType::get(Builder.getFloatTy(), 4); |
| 8976 | case SVETypeFlags::EltTyFloat64: |
| 8977 | return llvm::ScalableVectorType::get(Builder.getDoubleTy(), 2); |
| 8978 | |
| 8979 | case SVETypeFlags::EltTyBool8: |
| 8980 | return llvm::ScalableVectorType::get(Builder.getInt1Ty(), 16); |
| 8981 | case SVETypeFlags::EltTyBool16: |
| 8982 | return llvm::ScalableVectorType::get(Builder.getInt1Ty(), 8); |
| 8983 | case SVETypeFlags::EltTyBool32: |
| 8984 | return llvm::ScalableVectorType::get(Builder.getInt1Ty(), 4); |
| 8985 | case SVETypeFlags::EltTyBool64: |
| 8986 | return llvm::ScalableVectorType::get(Builder.getInt1Ty(), 2); |
| 8987 | } |
| 8988 | } |
| 8989 | |
| 8990 | llvm::Value * |
| 8991 | CodeGenFunction::EmitSVEAllTruePred(const SVETypeFlags &TypeFlags) { |
| 8992 | Function *Ptrue = |
| 8993 | CGM.getIntrinsic(Intrinsic::aarch64_sve_ptrue, getSVEPredType(TypeFlags)); |
| 8994 | return Builder.CreateCall(Ptrue, {Builder.getInt32(/*SV_ALL*/ 31)}); |
| 8995 | } |
| 8996 | |
| 8997 | constexpr unsigned SVEBitsPerBlock = 128; |
| 8998 | |
| 8999 | static llvm::ScalableVectorType *getSVEVectorForElementType(llvm::Type *EltTy) { |
| 9000 | unsigned NumElts = SVEBitsPerBlock / EltTy->getScalarSizeInBits(); |
| 9001 | return llvm::ScalableVectorType::get(EltTy, NumElts); |
| 9002 | } |
| 9003 | |
| 9004 | // Reinterpret the input predicate so that it can be used to correctly isolate |
| 9005 | // the elements of the specified datatype. |
| 9006 | Value *CodeGenFunction::EmitSVEPredicateCast(Value *Pred, |
| 9007 | llvm::ScalableVectorType *VTy) { |
| 9008 | auto *RTy = llvm::VectorType::get(IntegerType::get(getLLVMContext(), 1), VTy); |
| 9009 | if (Pred->getType() == RTy) |
| 9010 | return Pred; |
| 9011 | |
| 9012 | unsigned IntID; |
| 9013 | llvm::Type *IntrinsicTy; |
| 9014 | switch (VTy->getMinNumElements()) { |
| 9015 | default: |
| 9016 | llvm_unreachable("unsupported element count!")::llvm::llvm_unreachable_internal("unsupported element count!" , "clang/lib/CodeGen/CGBuiltin.cpp", 9016); |
| 9017 | case 2: |
| 9018 | case 4: |
| 9019 | case 8: |
| 9020 | IntID = Intrinsic::aarch64_sve_convert_from_svbool; |
| 9021 | IntrinsicTy = RTy; |
| 9022 | break; |
| 9023 | case 16: |
| 9024 | IntID = Intrinsic::aarch64_sve_convert_to_svbool; |
| 9025 | IntrinsicTy = Pred->getType(); |
| 9026 | break; |
| 9027 | } |
| 9028 | |
| 9029 | Function *F = CGM.getIntrinsic(IntID, IntrinsicTy); |
| 9030 | Value *C = Builder.CreateCall(F, Pred); |
| 9031 | assert(C->getType() == RTy && "Unexpected return type!")(static_cast <bool> (C->getType() == RTy && "Unexpected return type!" ) ? void (0) : __assert_fail ("C->getType() == RTy && \"Unexpected return type!\"" , "clang/lib/CodeGen/CGBuiltin.cpp", 9031, __extension__ __PRETTY_FUNCTION__ )); |
| 9032 | return C; |
| 9033 | } |
| 9034 | |
| 9035 | Value *CodeGenFunction::EmitSVEGatherLoad(const SVETypeFlags &TypeFlags, |
| 9036 | SmallVectorImpl<Value *> &Ops, |
| 9037 | unsigned IntID) { |
| 9038 | auto *ResultTy = getSVEType(TypeFlags); |
| 9039 | auto *OverloadedTy = |
| 9040 | llvm::ScalableVectorType::get(SVEBuiltinMemEltTy(TypeFlags), ResultTy); |
| 9041 | |
| 9042 | // At the ACLE level there's only one predicate type, svbool_t, which is |
| 9043 | // mapped to <n x 16 x i1>. However, this might be incompatible with the |
| 9044 | // actual type being loaded. For example, when loading doubles (i64) the |
| 9045 | // predicated should be <n x 2 x i1> instead. At the IR level the type of |
| 9046 | // the predicate and the data being loaded must match. Cast accordingly. |
| 9047 | Ops[0] = EmitSVEPredicateCast(Ops[0], OverloadedTy); |
| 9048 | |
| 9049 | Function *F = nullptr; |
| 9050 | if (Ops[1]->getType()->isVectorTy()) |
| 9051 | // This is the "vector base, scalar offset" case. In order to uniquely |
| 9052 | // map this built-in to an LLVM IR intrinsic, we need both the return type |
| 9053 | // and the type of the vector base. |
| 9054 | F = CGM.getIntrinsic(IntID, {OverloadedTy, Ops[1]->getType()}); |
| 9055 | else |
| 9056 | // This is the "scalar base, vector offset case". The type of the offset |
| 9057 | // is encoded in the name of the intrinsic. We only need to specify the |
| 9058 | // return type in order to uniquely map this built-in to an LLVM IR |
| 9059 | // intrinsic. |
| 9060 | F = CGM.getIntrinsic(IntID, OverloadedTy); |
| 9061 | |
| 9062 | // Pass 0 when the offset is missing. This can only be applied when using |
| 9063 | // the "vector base" addressing mode for which ACLE allows no offset. The |
| 9064 | // corresponding LLVM IR always requires an offset. |
| 9065 | if (Ops.size() == 2) { |
| 9066 | assert(Ops[1]->getType()->isVectorTy() && "Scalar base requires an offset")(static_cast <bool> (Ops[1]->getType()->isVectorTy () && "Scalar base requires an offset") ? void (0) : __assert_fail ("Ops[1]->getType()->isVectorTy() && \"Scalar base requires an offset\"" , "clang/lib/CodeGen/CGBuiltin.cpp", 9066, __extension__ __PRETTY_FUNCTION__ )); |
| 9067 | Ops.push_back(ConstantInt::get(Int64Ty, 0)); |
| 9068 | } |
| 9069 | |
| 9070 | // For "vector base, scalar index" scale the index so that it becomes a |
| 9071 | // scalar offset. |
| 9072 | if (!TypeFlags.isByteIndexed() && Ops[1]->getType()->isVectorTy()) { |
| 9073 | unsigned BytesPerElt = |
| 9074 | OverloadedTy->getElementType()->getScalarSizeInBits() / 8; |
| 9075 | Ops[2] = Builder.CreateShl(Ops[2], Log2_32(BytesPerElt)); |
| 9076 | } |
| 9077 | |
| 9078 | Value *Call = Builder.CreateCall(F, Ops); |
| 9079 | |
| 9080 | // The following sext/zext is only needed when ResultTy != OverloadedTy. In |
| 9081 | // other cases it's folded into a nop. |
| 9082 | return TypeFlags.isZExtReturn() ? Builder.CreateZExt(Call, ResultTy) |
| 9083 | : Builder.CreateSExt(Call, ResultTy); |
| 9084 | } |
| 9085 | |
| 9086 | Value *CodeGenFunction::EmitSVEScatterStore(const SVETypeFlags &TypeFlags, |
| 9087 | SmallVectorImpl<Value *> &Ops, |
| 9088 | unsigned IntID) { |
| 9089 | auto *SrcDataTy = getSVEType(TypeFlags); |
| 9090 | auto *OverloadedTy = |
| 9091 | llvm::ScalableVectorType::get(SVEBuiltinMemEltTy(TypeFlags), SrcDataTy); |
| 9092 | |
| 9093 | // In ACLE the source data is passed in the last argument, whereas in LLVM IR |
| 9094 | // it's the first argument. Move it accordingly. |
| 9095 | Ops.insert(Ops.begin(), Ops.pop_back_val()); |
| 9096 | |
| 9097 | Function *F = nullptr; |
| 9098 | if (Ops[2]->getType()->isVectorTy()) |
| 9099 | // This is the "vector base, scalar offset" case. In order to uniquely |
| 9100 | // map this built-in to an LLVM IR intrinsic, we need both the return type |
| 9101 | // and the type of the vector base. |
| 9102 | F = CGM.getIntrinsic(IntID, {OverloadedTy, Ops[2]->getType()}); |
| 9103 | else |
| 9104 | // This is the "scalar base, vector offset case". The type of the offset |
| 9105 | // is encoded in the name of the intrinsic. We only need to specify the |
| 9106 | // return type in order to uniquely map this built-in to an LLVM IR |
| 9107 | // intrinsic. |
| 9108 | F = CGM.getIntrinsic(IntID, OverloadedTy); |
| 9109 | |
| 9110 | // Pass 0 when the offset is missing. This can only be applied when using |
| 9111 | // the "vector base" addressing mode for which ACLE allows no offset. The |
| 9112 | // corresponding LLVM IR always requires an offset. |
| 9113 | if (Ops.size() == 3) { |
| 9114 | assert(Ops[1]->getType()->isVectorTy() && "Scalar base requires an offset")(static_cast <bool> (Ops[1]->getType()->isVectorTy () && "Scalar base requires an offset") ? void (0) : __assert_fail ("Ops[1]->getType()->isVectorTy() && \"Scalar base requires an offset\"" , "clang/lib/CodeGen/CGBuiltin.cpp", 9114, __extension__ __PRETTY_FUNCTION__ )); |
| 9115 | Ops.push_back(ConstantInt::get(Int64Ty, 0)); |
| 9116 | } |
| 9117 | |
| 9118 | // Truncation is needed when SrcDataTy != OverloadedTy. In other cases it's |
| 9119 | // folded into a nop. |
| 9120 | Ops[0] = Builder.CreateTrunc(Ops[0], OverloadedTy); |
| 9121 | |
| 9122 | // At the ACLE level there's only one predicate type, svbool_t, which is |
| 9123 | // mapped to <n x 16 x i1>. However, this might be incompatible with the |
| 9124 | // actual type being stored. For example, when storing doubles (i64) the |
| 9125 | // predicated should be <n x 2 x i1> instead. At the IR level the type of |
| 9126 | // the predicate and the data being stored must match. Cast accordingly. |
| 9127 | Ops[1] = EmitSVEPredicateCast(Ops[1], OverloadedTy); |
| 9128 | |
| 9129 | // For "vector base, scalar index" scale the index so that it becomes a |
| 9130 | // scalar offset. |
| 9131 | if (!TypeFlags.isByteIndexed() && Ops[2]->getType()->isVectorTy()) { |
| 9132 | unsigned BytesPerElt = |
| 9133 | OverloadedTy->getElementType()->getScalarSizeInBits() / 8; |
| 9134 | Ops[3] = Builder.CreateShl(Ops[3], Log2_32(BytesPerElt)); |
| 9135 | } |
| 9136 | |
| 9137 | return Builder.CreateCall(F, Ops); |
| 9138 | } |
| 9139 | |
| 9140 | Value *CodeGenFunction::EmitSVEGatherPrefetch(const SVETypeFlags &TypeFlags, |
| 9141 | SmallVectorImpl<Value *> &Ops, |
| 9142 | unsigned IntID) { |
| 9143 | // The gather prefetches are overloaded on the vector input - this can either |
| 9144 | // be the vector of base addresses or vector of offsets. |
| 9145 | auto *OverloadedTy = dyn_cast<llvm::ScalableVectorType>(Ops[1]->getType()); |
| 9146 | if (!OverloadedTy) |
| 9147 | OverloadedTy = cast<llvm::ScalableVectorType>(Ops[2]->getType()); |
| 9148 | |
| 9149 | // Cast the predicate from svbool_t to the right number of elements. |
| 9150 | Ops[0] = EmitSVEPredicateCast(Ops[0], OverloadedTy); |
| 9151 | |
| 9152 | // vector + imm addressing modes |
| 9153 | if (Ops[1]->getType()->isVectorTy()) { |
| 9154 | if (Ops.size() == 3) { |
| 9155 | // Pass 0 for 'vector+imm' when the index is omitted. |
| 9156 | Ops.push_back(ConstantInt::get(Int64Ty, 0)); |
| 9157 | |
| 9158 | // The sv_prfop is the last operand in the builtin and IR intrinsic. |
| 9159 | std::swap(Ops[2], Ops[3]); |
| 9160 | } else { |
| 9161 | // Index needs to be passed as scaled offset. |
| 9162 | llvm::Type *MemEltTy = SVEBuiltinMemEltTy(TypeFlags); |
| 9163 | unsigned BytesPerElt = MemEltTy->getPrimitiveSizeInBits() / 8; |
| 9164 | if (BytesPerElt > 1) |
| 9165 | Ops[2] = Builder.CreateShl(Ops[2], Log2_32(BytesPerElt)); |
| 9166 | } |
| 9167 | } |
| 9168 | |
| 9169 | Function *F = CGM.getIntrinsic(IntID, OverloadedTy); |
| 9170 | return Builder.CreateCall(F, Ops); |
| 9171 | } |
| 9172 | |
| 9173 | Value *CodeGenFunction::EmitSVEStructLoad(const SVETypeFlags &TypeFlags, |
| 9174 | SmallVectorImpl<Value*> &Ops, |
| 9175 | unsigned IntID) { |
| 9176 | llvm::ScalableVectorType *VTy = getSVEType(TypeFlags); |
| 9177 | auto VecPtrTy = llvm::PointerType::getUnqual(VTy); |
| 9178 | auto EltPtrTy = llvm::PointerType::getUnqual(VTy->getElementType()); |
| 9179 | |
| 9180 | unsigned N; |
| 9181 | switch (IntID) { |
| 9182 | case Intrinsic::aarch64_sve_ld2_sret: |
| 9183 | N = 2; |
| 9184 | break; |
| 9185 | case Intrinsic::aarch64_sve_ld3_sret: |
| 9186 | N = 3; |
| 9187 | break; |
| 9188 | case Intrinsic::aarch64_sve_ld4_sret: |
| 9189 | N = 4; |
| 9190 | break; |
| 9191 | default: |
| 9192 | llvm_unreachable("unknown intrinsic!")::llvm::llvm_unreachable_internal("unknown intrinsic!", "clang/lib/CodeGen/CGBuiltin.cpp" , 9192); |
| 9193 | } |
| 9194 | auto RetTy = llvm::VectorType::get(VTy->getElementType(), |
| 9195 | VTy->getElementCount() * N); |
| 9196 | |
| 9197 | Value *Predicate = EmitSVEPredicateCast(Ops[0], VTy); |
| 9198 | Value *BasePtr= Builder.CreateBitCast(Ops[1], VecPtrTy); |
| 9199 | |
| 9200 | // Does the load have an offset? |
| 9201 | if (Ops.size() > 2) |
| 9202 | BasePtr = Builder.CreateGEP(VTy, BasePtr, Ops[2]); |
| 9203 | |
| 9204 | BasePtr = Builder.CreateBitCast(BasePtr, EltPtrTy); |
| 9205 | Function *F = CGM.getIntrinsic(IntID, {VTy}); |
| 9206 | Value *Call = Builder.CreateCall(F, {Predicate, BasePtr}); |
| 9207 | unsigned MinElts = VTy->getMinNumElements(); |
| 9208 | Value *Ret = llvm::PoisonValue::get(RetTy); |
| 9209 | for (unsigned I = 0; I < N; I++) { |
| 9210 | Value *Idx = ConstantInt::get(CGM.Int64Ty, I * MinElts); |
| 9211 | Value *SRet = Builder.CreateExtractValue(Call, I); |
| 9212 | Ret = Builder.CreateInsertVector(RetTy, Ret, SRet, Idx); |
| 9213 | } |
| 9214 | return Ret; |
| 9215 | } |
| 9216 | |
| 9217 | Value *CodeGenFunction::EmitSVEStructStore(const SVETypeFlags &TypeFlags, |
| 9218 | SmallVectorImpl<Value*> &Ops, |
| 9219 | unsigned IntID) { |
| 9220 | llvm::ScalableVectorType *VTy = getSVEType(TypeFlags); |
| 9221 | auto VecPtrTy = llvm::PointerType::getUnqual(VTy); |
| 9222 | auto EltPtrTy = llvm::PointerType::getUnqual(VTy->getElementType()); |
| 9223 | |
| 9224 | unsigned N; |
| 9225 | switch (IntID) { |
| 9226 | case Intrinsic::aarch64_sve_st2: |
| 9227 | N = 2; |
| 9228 | break; |
| 9229 | case Intrinsic::aarch64_sve_st3: |
| 9230 | N = 3; |
| 9231 | break; |
| 9232 | case Intrinsic::aarch64_sve_st4: |
| 9233 | N = 4; |
| 9234 | break; |
| 9235 | default: |
| 9236 | llvm_unreachable("unknown intrinsic!")::llvm::llvm_unreachable_internal("unknown intrinsic!", "clang/lib/CodeGen/CGBuiltin.cpp" , 9236); |
| 9237 | } |
| 9238 | |
| 9239 | Value *Predicate = EmitSVEPredicateCast(Ops[0], VTy); |
| 9240 | Value *BasePtr = Builder.CreateBitCast(Ops[1], VecPtrTy); |
| 9241 | |
| 9242 | // Does the store have an offset? |
| 9243 | if (Ops.size() > 3) |
| 9244 | BasePtr = Builder.CreateGEP(VTy, BasePtr, Ops[2]); |
| 9245 | |
| 9246 | BasePtr = Builder.CreateBitCast(BasePtr, EltPtrTy); |
| 9247 | Value *Val = Ops.back(); |
| 9248 | |
| 9249 | // The llvm.aarch64.sve.st2/3/4 intrinsics take legal part vectors, so we |
| 9250 | // need to break up the tuple vector. |
| 9251 | SmallVector<llvm::Value*, 5> Operands; |
| 9252 | unsigned MinElts = VTy->getElementCount().getKnownMinValue(); |
| 9253 | for (unsigned I = 0; I < N; ++I) { |
| 9254 | Value *Idx = ConstantInt::get(CGM.Int64Ty, I * MinElts); |
| 9255 | Operands.push_back(Builder.CreateExtractVector(VTy, Val, Idx)); |
| 9256 | } |
| 9257 | Operands.append({Predicate, BasePtr}); |
| 9258 | |
| 9259 | Function *F = CGM.getIntrinsic(IntID, { VTy }); |
| 9260 | return Builder.CreateCall(F, Operands); |
| 9261 | } |
| 9262 | |
| 9263 | // SVE2's svpmullb and svpmullt builtins are similar to the svpmullb_pair and |
| 9264 | // svpmullt_pair intrinsics, with the exception that their results are bitcast |
| 9265 | // to a wider type. |
| 9266 | Value *CodeGenFunction::EmitSVEPMull(const SVETypeFlags &TypeFlags, |
| 9267 | SmallVectorImpl<Value *> &Ops, |
| 9268 | unsigned BuiltinID) { |
| 9269 | // Splat scalar operand to vector (intrinsics with _n infix) |
| 9270 | if (TypeFlags.hasSplatOperand()) { |
| 9271 | unsigned OpNo = TypeFlags.getSplatOperand(); |
| 9272 | Ops[OpNo] = EmitSVEDupX(Ops[OpNo]); |
| 9273 | } |
| 9274 | |
| 9275 | // The pair-wise function has a narrower overloaded type. |
| 9276 | Function *F = CGM.getIntrinsic(BuiltinID, Ops[0]->getType()); |
| 9277 | Value *Call = Builder.CreateCall(F, {Ops[0], Ops[1]}); |
| 9278 | |
| 9279 | // Now bitcast to the wider result type. |
| 9280 | llvm::ScalableVectorType *Ty = getSVEType(TypeFlags); |
| 9281 | return EmitSVEReinterpret(Call, Ty); |
| 9282 | } |
| 9283 | |
| 9284 | Value *CodeGenFunction::EmitSVEMovl(const SVETypeFlags &TypeFlags, |
| 9285 | ArrayRef<Value *> Ops, unsigned BuiltinID) { |
| 9286 | llvm::Type *OverloadedTy = getSVEType(TypeFlags); |
| 9287 | Function *F = CGM.getIntrinsic(BuiltinID, OverloadedTy); |
| 9288 | return Builder.CreateCall(F, {Ops[0], Builder.getInt32(0)}); |
| 9289 | } |
| 9290 | |
| 9291 | Value *CodeGenFunction::EmitSVEPrefetchLoad(const SVETypeFlags &TypeFlags, |
| 9292 | SmallVectorImpl<Value *> &Ops, |
| 9293 | unsigned BuiltinID) { |
| 9294 | auto *MemEltTy = SVEBuiltinMemEltTy(TypeFlags); |
| 9295 | auto *VectorTy = getSVEVectorForElementType(MemEltTy); |
| 9296 | auto *MemoryTy = llvm::ScalableVectorType::get(MemEltTy, VectorTy); |
| 9297 | |
| 9298 | Value *Predicate = EmitSVEPredicateCast(Ops[0], MemoryTy); |
| 9299 | Value *BasePtr = Ops[1]; |
| 9300 | |
| 9301 | // Implement the index operand if not omitted. |
| 9302 | if (Ops.size() > 3) { |
| 9303 | BasePtr = Builder.CreateBitCast(BasePtr, MemoryTy->getPointerTo()); |
| 9304 | BasePtr = Builder.CreateGEP(MemoryTy, BasePtr, Ops[2]); |
| 9305 | } |
| 9306 | |
| 9307 | // Prefetch intriniscs always expect an i8* |
| 9308 | BasePtr = Builder.CreateBitCast(BasePtr, llvm::PointerType::getUnqual(Int8Ty)); |
| 9309 | Value *PrfOp = Ops.back(); |
| 9310 | |
| 9311 | Function *F = CGM.getIntrinsic(BuiltinID, Predicate->getType()); |
| 9312 | return Builder.CreateCall(F, {Predicate, BasePtr, PrfOp}); |
| 9313 | } |
| 9314 | |
| 9315 | Value *CodeGenFunction::EmitSVEMaskedLoad(const CallExpr *E, |
| 9316 | llvm::Type *ReturnTy, |
| 9317 | SmallVectorImpl<Value *> &Ops, |
| 9318 | unsigned BuiltinID, |
| 9319 | bool IsZExtReturn) { |
| 9320 | QualType LangPTy = E->getArg(1)->getType(); |
| 9321 | llvm::Type *MemEltTy = CGM.getTypes().ConvertType( |
| 9322 | LangPTy->castAs<PointerType>()->getPointeeType()); |
| 9323 | |
| 9324 | // The vector type that is returned may be different from the |
| 9325 | // eventual type loaded from memory. |
| 9326 | auto VectorTy = cast<llvm::ScalableVectorType>(ReturnTy); |
| 9327 | auto MemoryTy = llvm::ScalableVectorType::get(MemEltTy, VectorTy); |
| 9328 | |
| 9329 | Value *Predicate = EmitSVEPredicateCast(Ops[0], MemoryTy); |
| 9330 | Value *BasePtr = Builder.CreateBitCast(Ops[1], MemoryTy->getPointerTo()); |
| 9331 | |
| 9332 | // Does the load have an offset? |
| 9333 | if (Ops.size() > 2) |
| 9334 | BasePtr = Builder.CreateGEP(MemoryTy, BasePtr, Ops[2]); |
| 9335 | |
| 9336 | BasePtr = Builder.CreateBitCast(BasePtr, MemEltTy->getPointerTo()); |
| 9337 | Function *F = CGM.getIntrinsic(BuiltinID, MemoryTy); |
| 9338 | auto *Load = |
| 9339 | cast<llvm::Instruction>(Builder.CreateCall(F, {Predicate, BasePtr})); |
| 9340 | auto TBAAInfo = CGM.getTBAAAccessInfo(LangPTy->getPointeeType()); |
| 9341 | CGM.DecorateInstructionWithTBAA(Load, TBAAInfo); |
| 9342 | |
| 9343 | return IsZExtReturn ? Builder.CreateZExt(Load, VectorTy) |
| 9344 | : Builder.CreateSExt(Load, VectorTy); |
| 9345 | } |
| 9346 | |
| 9347 | Value *CodeGenFunction::EmitSVEMaskedStore(const CallExpr *E, |
| 9348 | SmallVectorImpl<Value *> &Ops, |
| 9349 | unsigned BuiltinID) { |
| 9350 | QualType LangPTy = E->getArg(1)->getType(); |
| 9351 | llvm::Type *MemEltTy = CGM.getTypes().ConvertType( |
| 9352 | LangPTy->castAs<PointerType>()->getPointeeType()); |
| 9353 | |
| 9354 | // The vector type that is stored may be different from the |
| 9355 | // eventual type stored to memory. |
| 9356 | auto VectorTy = cast<llvm::ScalableVectorType>(Ops.back()->getType()); |
| 9357 | auto MemoryTy = llvm::ScalableVectorType::get(MemEltTy, VectorTy); |
| 9358 | |
| 9359 | Value *Predicate = EmitSVEPredicateCast(Ops[0], MemoryTy); |
| 9360 | Value *BasePtr = Builder.CreateBitCast(Ops[1], MemoryTy->getPointerTo()); |
| 9361 | |
| 9362 | // Does the store have an offset? |
| 9363 | if (Ops.size() == 4) |
| 9364 | BasePtr = Builder.CreateGEP(MemoryTy, BasePtr, Ops[2]); |
| 9365 | |
| 9366 | // Last value is always the data |
| 9367 | llvm::Value *Val = Builder.CreateTrunc(Ops.back(), MemoryTy); |
| 9368 | |
| 9369 | BasePtr = Builder.CreateBitCast(BasePtr, MemEltTy->getPointerTo()); |
| 9370 | Function *F = CGM.getIntrinsic(BuiltinID, MemoryTy); |
| 9371 | auto *Store = |
| 9372 | cast<llvm::Instruction>(Builder.CreateCall(F, {Val, Predicate, BasePtr})); |
| 9373 | auto TBAAInfo = CGM.getTBAAAccessInfo(LangPTy->getPointeeType()); |
| 9374 | CGM.DecorateInstructionWithTBAA(Store, TBAAInfo); |
| 9375 | return Store; |
| 9376 | } |
| 9377 | |
| 9378 | // Limit the usage of scalable llvm IR generated by the ACLE by using the |
| 9379 | // sve dup.x intrinsic instead of IRBuilder::CreateVectorSplat. |
| 9380 | Value *CodeGenFunction::EmitSVEDupX(Value *Scalar, llvm::Type *Ty) { |
| 9381 | return Builder.CreateVectorSplat( |
| 9382 | cast<llvm::VectorType>(Ty)->getElementCount(), Scalar); |
| 9383 | } |
| 9384 | |
| 9385 | Value *CodeGenFunction::EmitSVEDupX(Value* Scalar) { |
| 9386 | return EmitSVEDupX(Scalar, getSVEVectorForElementType(Scalar->getType())); |
| 9387 | } |
| 9388 | |
| 9389 | Value *CodeGenFunction::EmitSVEReinterpret(Value *Val, llvm::Type *Ty) { |
| 9390 | // FIXME: For big endian this needs an additional REV, or needs a separate |
| 9391 | // intrinsic that is code-generated as a no-op, because the LLVM bitcast |
| 9392 | // instruction is defined as 'bitwise' equivalent from memory point of |
| 9393 | // view (when storing/reloading), whereas the svreinterpret builtin |
| 9394 | // implements bitwise equivalent cast from register point of view. |
| 9395 | // LLVM CodeGen for a bitcast must add an explicit REV for big-endian. |
| 9396 | return Builder.CreateBitCast(Val, Ty); |
| 9397 | } |
| 9398 | |
| 9399 | static void InsertExplicitZeroOperand(CGBuilderTy &Builder, llvm::Type *Ty, |
| 9400 | SmallVectorImpl<Value *> &Ops) { |
| 9401 | auto *SplatZero = Constant::getNullValue(Ty); |
| 9402 | Ops.insert(Ops.begin(), SplatZero); |
| 9403 | } |
| 9404 | |
| 9405 | static void InsertExplicitUndefOperand(CGBuilderTy &Builder, llvm::Type *Ty, |
| 9406 | SmallVectorImpl<Value *> &Ops) { |
| 9407 | auto *SplatUndef = UndefValue::get(Ty); |
| 9408 | Ops.insert(Ops.begin(), SplatUndef); |
| 9409 | } |
| 9410 | |
| 9411 | SmallVector<llvm::Type *, 2> |
| 9412 | CodeGenFunction::getSVEOverloadTypes(const SVETypeFlags &TypeFlags, |
| 9413 | llvm::Type *ResultType, |
| 9414 | ArrayRef<Value *> Ops) { |
| 9415 | if (TypeFlags.isOverloadNone()) |
| 9416 | return {}; |
| 9417 | |
| 9418 | llvm::Type *DefaultType = getSVEType(TypeFlags); |
| 9419 | |
| 9420 | if (TypeFlags.isOverloadWhile()) |
| 9421 | return {DefaultType, Ops[1]->getType()}; |
| 9422 | |
| 9423 | if (TypeFlags.isOverloadWhileRW()) |
| 9424 | return {getSVEPredType(TypeFlags), Ops[0]->getType()}; |
| 9425 | |
| 9426 | if (TypeFlags.isOverloadCvt()) |
| 9427 | return {Ops[0]->getType(), Ops.back()->getType()}; |
| 9428 | |
| 9429 | assert(TypeFlags.isOverloadDefault() && "Unexpected value for overloads")(static_cast <bool> (TypeFlags.isOverloadDefault() && "Unexpected value for overloads") ? void (0) : __assert_fail ("TypeFlags.isOverloadDefault() && \"Unexpected value for overloads\"" , "clang/lib/CodeGen/CGBuiltin.cpp", 9429, __extension__ __PRETTY_FUNCTION__ )); |
| 9430 | return {DefaultType}; |
| 9431 | } |
| 9432 | |
| 9433 | Value *CodeGenFunction::EmitSVETupleSetOrGet(const SVETypeFlags &TypeFlags, |
| 9434 | llvm::Type *Ty, |
| 9435 | ArrayRef<Value *> Ops) { |
| 9436 | assert((TypeFlags.isTupleSet() || TypeFlags.isTupleGet()) &&(static_cast <bool> ((TypeFlags.isTupleSet() || TypeFlags .isTupleGet()) && "Expects TypleFlag isTupleSet or TypeFlags.isTupleSet()" ) ? void (0) : __assert_fail ("(TypeFlags.isTupleSet() || TypeFlags.isTupleGet()) && \"Expects TypleFlag isTupleSet or TypeFlags.isTupleSet()\"" , "clang/lib/CodeGen/CGBuiltin.cpp", 9437, __extension__ __PRETTY_FUNCTION__ )) |
| 9437 | "Expects TypleFlag isTupleSet or TypeFlags.isTupleSet()")(static_cast <bool> ((TypeFlags.isTupleSet() || TypeFlags .isTupleGet()) && "Expects TypleFlag isTupleSet or TypeFlags.isTupleSet()" ) ? void (0) : __assert_fail ("(TypeFlags.isTupleSet() || TypeFlags.isTupleGet()) && \"Expects TypleFlag isTupleSet or TypeFlags.isTupleSet()\"" , "clang/lib/CodeGen/CGBuiltin.cpp", 9437, __extension__ __PRETTY_FUNCTION__ )); |
| 9438 | |
| 9439 | unsigned I = cast<ConstantInt>(Ops[1])->getSExtValue(); |
| 9440 | auto *SingleVecTy = dyn_cast<llvm::ScalableVectorType>( |
| 9441 | TypeFlags.isTupleSet() ? Ops[2]->getType() : Ty); |
| 9442 | Value *Idx = ConstantInt::get(CGM.Int64Ty, |
| 9443 | I * SingleVecTy->getMinNumElements()); |
| 9444 | |
| 9445 | if (TypeFlags.isTupleSet()) |
| 9446 | return Builder.CreateInsertVector(Ty, Ops[0], Ops[2], Idx); |
| 9447 | return Builder.CreateExtractVector(Ty, Ops[0], Idx); |
| 9448 | } |
| 9449 | |
| 9450 | Value *CodeGenFunction::EmitSVETupleCreate(const SVETypeFlags &TypeFlags, |
| 9451 | llvm::Type *Ty, |
| 9452 | ArrayRef<Value *> Ops) { |
| 9453 | assert(TypeFlags.isTupleCreate() && "Expects TypleFlag isTupleCreate")(static_cast <bool> (TypeFlags.isTupleCreate() && "Expects TypleFlag isTupleCreate") ? void (0) : __assert_fail ("TypeFlags.isTupleCreate() && \"Expects TypleFlag isTupleCreate\"" , "clang/lib/CodeGen/CGBuiltin.cpp", 9453, __extension__ __PRETTY_FUNCTION__ )); |
| 9454 | |
| 9455 | auto *SrcTy = dyn_cast<llvm::ScalableVectorType>(Ops[0]->getType()); |
| 9456 | unsigned MinElts = SrcTy->getMinNumElements(); |
| 9457 | Value *Call = llvm::PoisonValue::get(Ty); |
| 9458 | for (unsigned I = 0; I < Ops.size(); I++) { |
| 9459 | Value *Idx = ConstantInt::get(CGM.Int64Ty, I * MinElts); |
| 9460 | Call = Builder.CreateInsertVector(Ty, Call, Ops[I], Idx); |
| 9461 | } |
| 9462 | |
| 9463 | return Call; |
| 9464 | } |
| 9465 | |
| 9466 | Value *CodeGenFunction::EmitAArch64SVEBuiltinExpr(unsigned BuiltinID, |
| 9467 | const CallExpr *E) { |
| 9468 | // Find out if any arguments are required to be integer constant expressions. |
| 9469 | unsigned ICEArguments = 0; |
| 9470 | ASTContext::GetBuiltinTypeError Error; |
| 9471 | getContext().GetBuiltinType(BuiltinID, Error, &ICEArguments); |
| 9472 | assert(Error == ASTContext::GE_None && "Should not codegen an error")(static_cast <bool> (Error == ASTContext::GE_None && "Should not codegen an error") ? void (0) : __assert_fail ("Error == ASTContext::GE_None && \"Should not codegen an error\"" , "clang/lib/CodeGen/CGBuiltin.cpp", 9472, __extension__ __PRETTY_FUNCTION__ )); |
| 9473 | |
| 9474 | llvm::Type *Ty = ConvertType(E->getType()); |
| 9475 | if (BuiltinID >= SVE::BI__builtin_sve_reinterpret_s8_s8 && |
| 9476 | BuiltinID <= SVE::BI__builtin_sve_reinterpret_f64_f64) { |
| 9477 | Value *Val = EmitScalarExpr(E->getArg(0)); |
| 9478 | return EmitSVEReinterpret(Val, Ty); |
| 9479 | } |
| 9480 | |
| 9481 | llvm::SmallVector<Value *, 4> Ops; |
| 9482 | for (unsigned i = 0, e = E->getNumArgs(); i != e; i++) { |
| 9483 | if ((ICEArguments & (1 << i)) == 0) |
| 9484 | Ops.push_back(EmitScalarExpr(E->getArg(i))); |
| 9485 | else { |
| 9486 | // If this is required to be a constant, constant fold it so that we know |
| 9487 | // that the generated intrinsic gets a ConstantInt. |
| 9488 | std::optional<llvm::APSInt> Result = |
| 9489 | E->getArg(i)->getIntegerConstantExpr(getContext()); |
| 9490 | assert(Result && "Expected argument to be a constant")(static_cast <bool> (Result && "Expected argument to be a constant" ) ? void (0) : __assert_fail ("Result && \"Expected argument to be a constant\"" , "clang/lib/CodeGen/CGBuiltin.cpp", 9490, __extension__ __PRETTY_FUNCTION__ )); |
| 9491 | |
| 9492 | // Immediates for SVE llvm intrinsics are always 32bit. We can safely |
| 9493 | // truncate because the immediate has been range checked and no valid |
| 9494 | // immediate requires more than a handful of bits. |
| 9495 | *Result = Result->extOrTrunc(32); |
| 9496 | Ops.push_back(llvm::ConstantInt::get(getLLVMContext(), *Result)); |
| 9497 | } |
| 9498 | } |
| 9499 | |
| 9500 | auto *Builtin = findARMVectorIntrinsicInMap(AArch64SVEIntrinsicMap, BuiltinID, |
| 9501 | AArch64SVEIntrinsicsProvenSorted); |
| 9502 | SVETypeFlags TypeFlags(Builtin->TypeModifier); |
| 9503 | if (TypeFlags.isLoad()) |
| 9504 | return EmitSVEMaskedLoad(E, Ty, Ops, Builtin->LLVMIntrinsic, |
| 9505 | TypeFlags.isZExtReturn()); |
| 9506 | else if (TypeFlags.isStore()) |
| 9507 | return EmitSVEMaskedStore(E, Ops, Builtin->LLVMIntrinsic); |
| 9508 | else if (TypeFlags.isGatherLoad()) |
| 9509 | return EmitSVEGatherLoad(TypeFlags, Ops, Builtin->LLVMIntrinsic); |
| 9510 | else if (TypeFlags.isScatterStore()) |
| 9511 | return EmitSVEScatterStore(TypeFlags, Ops, Builtin->LLVMIntrinsic); |
| 9512 | else if (TypeFlags.isPrefetch()) |
| 9513 | return EmitSVEPrefetchLoad(TypeFlags, Ops, Builtin->LLVMIntrinsic); |
| 9514 | else if (TypeFlags.isGatherPrefetch()) |
| 9515 | return EmitSVEGatherPrefetch(TypeFlags, Ops, Builtin->LLVMIntrinsic); |
| 9516 | else if (TypeFlags.isStructLoad()) |
| 9517 | return EmitSVEStructLoad(TypeFlags, Ops, Builtin->LLVMIntrinsic); |
| 9518 | else if (TypeFlags.isStructStore()) |
| 9519 | return EmitSVEStructStore(TypeFlags, Ops, Builtin->LLVMIntrinsic); |
| 9520 | else if (TypeFlags.isTupleSet() || TypeFlags.isTupleGet()) |
| 9521 | return EmitSVETupleSetOrGet(TypeFlags, Ty, Ops); |
| 9522 | else if (TypeFlags.isTupleCreate()) |
| 9523 | return EmitSVETupleCreate(TypeFlags, Ty, Ops); |
| 9524 | else if (TypeFlags.isUndef()) |
| 9525 | return UndefValue::get(Ty); |
| 9526 | else if (Builtin->LLVMIntrinsic != 0) { |
| 9527 | if (TypeFlags.getMergeType() == SVETypeFlags::MergeZeroExp) |
| 9528 | InsertExplicitZeroOperand(Builder, Ty, Ops); |
| 9529 | |
| 9530 | if (TypeFlags.getMergeType() == SVETypeFlags::MergeAnyExp) |
| 9531 | InsertExplicitUndefOperand(Builder, Ty, Ops); |
| 9532 | |
| 9533 | // Some ACLE builtins leave out the argument to specify the predicate |
| 9534 | // pattern, which is expected to be expanded to an SV_ALL pattern. |
| 9535 | if (TypeFlags.isAppendSVALL()) |
| 9536 | Ops.push_back(Builder.getInt32(/*SV_ALL*/ 31)); |
| 9537 | if (TypeFlags.isInsertOp1SVALL()) |
| 9538 | Ops.insert(&Ops[1], Builder.getInt32(/*SV_ALL*/ 31)); |
| 9539 | |
| 9540 | // Predicates must match the main datatype. |
| 9541 | for (unsigned i = 0, e = Ops.size(); i != e; ++i) |
| 9542 | if (auto PredTy = dyn_cast<llvm::VectorType>(Ops[i]->getType())) |
| 9543 | if (PredTy->getElementType()->isIntegerTy(1)) |
| 9544 | Ops[i] = EmitSVEPredicateCast(Ops[i], getSVEType(TypeFlags)); |
| 9545 | |
| 9546 | // Splat scalar operand to vector (intrinsics with _n infix) |
| 9547 | if (TypeFlags.hasSplatOperand()) { |
| 9548 | unsigned OpNo = TypeFlags.getSplatOperand(); |
| 9549 | Ops[OpNo] = EmitSVEDupX(Ops[OpNo]); |
| 9550 | } |
| 9551 | |
| 9552 | if (TypeFlags.isReverseCompare()) |
| 9553 | std::swap(Ops[1], Ops[2]); |
| 9554 | else if (TypeFlags.isReverseUSDOT()) |
| 9555 | std::swap(Ops[1], Ops[2]); |
| 9556 | else if (TypeFlags.isReverseMergeAnyBinOp() && |
| 9557 | TypeFlags.getMergeType() == SVETypeFlags::MergeAny) |
| 9558 | std::swap(Ops[1], Ops[2]); |
| 9559 | else if (TypeFlags.isReverseMergeAnyAccOp() && |
| 9560 | TypeFlags.getMergeType() == SVETypeFlags::MergeAny) |
| 9561 | std::swap(Ops[1], Ops[3]); |
| 9562 | |
| 9563 | // Predicated intrinsics with _z suffix need a select w/ zeroinitializer. |
| 9564 | if (TypeFlags.getMergeType() == SVETypeFlags::MergeZero) { |
| 9565 | llvm::Type *OpndTy = Ops[1]->getType(); |
| 9566 | auto *SplatZero = Constant::getNullValue(OpndTy); |
| 9567 | Ops[1] = Builder.CreateSelect(Ops[0], Ops[1], SplatZero); |
| 9568 | } |
| 9569 | |
| 9570 | Function *F = CGM.getIntrinsic(Builtin->LLVMIntrinsic, |
| 9571 | getSVEOverloadTypes(TypeFlags, Ty, Ops)); |
| 9572 | Value *Call = Builder.CreateCall(F, Ops); |
| 9573 | |
| 9574 | // Predicate results must be converted to svbool_t. |
| 9575 | if (auto PredTy = dyn_cast<llvm::VectorType>(Call->getType())) |
| 9576 | if (PredTy->getScalarType()->isIntegerTy(1)) |
| 9577 | Call = EmitSVEPredicateCast(Call, cast<llvm::ScalableVectorType>(Ty)); |
| 9578 | |
| 9579 | return Call; |
| 9580 | } |
| 9581 | |
| 9582 | switch (BuiltinID) { |
| 9583 | default: |
| 9584 | return nullptr; |
| 9585 | |
| 9586 | case SVE::BI__builtin_sve_svmov_b_z: { |
| 9587 | // svmov_b_z(pg, op) <=> svand_b_z(pg, op, op) |
| 9588 | SVETypeFlags TypeFlags(Builtin->TypeModifier); |
| 9589 | llvm::Type* OverloadedTy = getSVEType(TypeFlags); |
| 9590 | Function *F = CGM.getIntrinsic(Intrinsic::aarch64_sve_and_z, OverloadedTy); |
| 9591 | return Builder.CreateCall(F, {Ops[0], Ops[1], Ops[1]}); |
| 9592 | } |
| 9593 | |
| 9594 | case SVE::BI__builtin_sve_svnot_b_z: { |
| 9595 | // svnot_b_z(pg, op) <=> sveor_b_z(pg, op, pg) |
| 9596 | SVETypeFlags TypeFlags(Builtin->TypeModifier); |
| 9597 | llvm::Type* OverloadedTy = getSVEType(TypeFlags); |
| 9598 | Function *F = CGM.getIntrinsic(Intrinsic::aarch64_sve_eor_z, OverloadedTy); |
| 9599 | return Builder.CreateCall(F, {Ops[0], Ops[1], Ops[0]}); |
| 9600 | } |
| 9601 | |
| 9602 | case SVE::BI__builtin_sve_svmovlb_u16: |
| 9603 | case SVE::BI__builtin_sve_svmovlb_u32: |
| 9604 | case SVE::BI__builtin_sve_svmovlb_u64: |
| 9605 | return EmitSVEMovl(TypeFlags, Ops, Intrinsic::aarch64_sve_ushllb); |
| 9606 | |
| 9607 | case SVE::BI__builtin_sve_svmovlb_s16: |
| 9608 | case SVE::BI__builtin_sve_svmovlb_s32: |
| 9609 | case SVE::BI__builtin_sve_svmovlb_s64: |
| 9610 | return EmitSVEMovl(TypeFlags, Ops, Intrinsic::aarch64_sve_sshllb); |
| 9611 | |
| 9612 | case SVE::BI__builtin_sve_svmovlt_u16: |
| 9613 | case SVE::BI__builtin_sve_svmovlt_u32: |
| 9614 | case SVE::BI__builtin_sve_svmovlt_u64: |
| 9615 | return EmitSVEMovl(TypeFlags, Ops, Intrinsic::aarch64_sve_ushllt); |
| 9616 | |
| 9617 | case SVE::BI__builtin_sve_svmovlt_s16: |
| 9618 | case SVE::BI__builtin_sve_svmovlt_s32: |
| 9619 | case SVE::BI__builtin_sve_svmovlt_s64: |
| 9620 | return EmitSVEMovl(TypeFlags, Ops, Intrinsic::aarch64_sve_sshllt); |
| 9621 | |
| 9622 | case SVE::BI__builtin_sve_svpmullt_u16: |
| 9623 | case SVE::BI__builtin_sve_svpmullt_u64: |
| 9624 | case SVE::BI__builtin_sve_svpmullt_n_u16: |
| 9625 | case SVE::BI__builtin_sve_svpmullt_n_u64: |
| 9626 | return EmitSVEPMull(TypeFlags, Ops, Intrinsic::aarch64_sve_pmullt_pair); |
| 9627 | |
| 9628 | case SVE::BI__builtin_sve_svpmullb_u16: |
| 9629 | case SVE::BI__builtin_sve_svpmullb_u64: |
| 9630 | case SVE::BI__builtin_sve_svpmullb_n_u16: |
| 9631 | case SVE::BI__builtin_sve_svpmullb_n_u64: |
| 9632 | return EmitSVEPMull(TypeFlags, Ops, Intrinsic::aarch64_sve_pmullb_pair); |
| 9633 | |
| 9634 | case SVE::BI__builtin_sve_svdup_n_b8: |
| 9635 | case SVE::BI__builtin_sve_svdup_n_b16: |
| 9636 | case SVE::BI__builtin_sve_svdup_n_b32: |
| 9637 | case SVE::BI__builtin_sve_svdup_n_b64: { |
| 9638 | Value *CmpNE = |
| 9639 | Builder.CreateICmpNE(Ops[0], Constant::getNullValue(Ops[0]->getType())); |
| 9640 | llvm::ScalableVectorType *OverloadedTy = getSVEType(TypeFlags); |
| 9641 | Value *Dup = EmitSVEDupX(CmpNE, OverloadedTy); |
| 9642 | return EmitSVEPredicateCast(Dup, cast<llvm::ScalableVectorType>(Ty)); |
| 9643 | } |
| 9644 | |
| 9645 | case SVE::BI__builtin_sve_svdupq_n_b8: |
| 9646 | case SVE::BI__builtin_sve_svdupq_n_b16: |
| 9647 | case SVE::BI__builtin_sve_svdupq_n_b32: |
| 9648 | case SVE::BI__builtin_sve_svdupq_n_b64: |
| 9649 | case SVE::BI__builtin_sve_svdupq_n_u8: |
| 9650 | case SVE::BI__builtin_sve_svdupq_n_s8: |
| 9651 | case SVE::BI__builtin_sve_svdupq_n_u64: |
| 9652 | case SVE::BI__builtin_sve_svdupq_n_f64: |
| 9653 | case SVE::BI__builtin_sve_svdupq_n_s64: |
| 9654 | case SVE::BI__builtin_sve_svdupq_n_u16: |
| 9655 | case SVE::BI__builtin_sve_svdupq_n_f16: |
| 9656 | case SVE::BI__builtin_sve_svdupq_n_bf16: |
| 9657 | case SVE::BI__builtin_sve_svdupq_n_s16: |
| 9658 | case SVE::BI__builtin_sve_svdupq_n_u32: |
| 9659 | case SVE::BI__builtin_sve_svdupq_n_f32: |
| 9660 | case SVE::BI__builtin_sve_svdupq_n_s32: { |
| 9661 | // These builtins are implemented by storing each element to an array and using |
| 9662 | // ld1rq to materialize a vector. |
| 9663 | unsigned NumOpnds = Ops.size(); |
| 9664 | |
| 9665 | bool IsBoolTy = |
| 9666 | cast<llvm::VectorType>(Ty)->getElementType()->isIntegerTy(1); |
| 9667 | |
| 9668 | // For svdupq_n_b* the element type of is an integer of type 128/numelts, |
| 9669 | // so that the compare can use the width that is natural for the expected |
| 9670 | // number of predicate lanes. |
| 9671 | llvm::Type *EltTy = Ops[0]->getType(); |
| 9672 | if (IsBoolTy) |
| 9673 | EltTy = IntegerType::get(getLLVMContext(), SVEBitsPerBlock / NumOpnds); |
| 9674 | |
| 9675 | SmallVector<llvm::Value *, 16> VecOps; |
| 9676 | for (unsigned I = 0; I < NumOpnds; ++I) |
| 9677 | VecOps.push_back(Builder.CreateZExt(Ops[I], EltTy)); |
| 9678 | Value *Vec = BuildVector(VecOps); |
| 9679 | |
| 9680 | llvm::Type *OverloadedTy = getSVEVectorForElementType(EltTy); |
| 9681 | Value *InsertSubVec = Builder.CreateInsertVector( |
| 9682 | OverloadedTy, PoisonValue::get(OverloadedTy), Vec, Builder.getInt64(0)); |
| 9683 | |
| 9684 | Function *F = |
| 9685 | CGM.getIntrinsic(Intrinsic::aarch64_sve_dupq_lane, OverloadedTy); |
| 9686 | Value *DupQLane = |
| 9687 | Builder.CreateCall(F, {InsertSubVec, Builder.getInt64(0)}); |
| 9688 | |
| 9689 | if (!IsBoolTy) |
| 9690 | return DupQLane; |
| 9691 | |
| 9692 | SVETypeFlags TypeFlags(Builtin->TypeModifier); |
| 9693 | Value *Pred = EmitSVEAllTruePred(TypeFlags); |
| 9694 | |
| 9695 | // For svdupq_n_b* we need to add an additional 'cmpne' with '0'. |
| 9696 | F = CGM.getIntrinsic(NumOpnds == 2 ? Intrinsic::aarch64_sve_cmpne |
| 9697 | : Intrinsic::aarch64_sve_cmpne_wide, |
| 9698 | OverloadedTy); |
| 9699 | Value *Call = Builder.CreateCall( |
| 9700 | F, {Pred, DupQLane, EmitSVEDupX(Builder.getInt64(0))}); |
| 9701 | return EmitSVEPredicateCast(Call, cast<llvm::ScalableVectorType>(Ty)); |
| 9702 | } |
| 9703 | |
| 9704 | case SVE::BI__builtin_sve_svpfalse_b: |
| 9705 | return ConstantInt::getFalse(Ty); |
| 9706 | |
| 9707 | case SVE::BI__builtin_sve_svlen_bf16: |
| 9708 | case SVE::BI__builtin_sve_svlen_f16: |
| 9709 | case SVE::BI__builtin_sve_svlen_f32: |
| 9710 | case SVE::BI__builtin_sve_svlen_f64: |
| 9711 | case SVE::BI__builtin_sve_svlen_s8: |
| 9712 | case SVE::BI__builtin_sve_svlen_s16: |
| 9713 | case SVE::BI__builtin_sve_svlen_s32: |
| 9714 | case SVE::BI__builtin_sve_svlen_s64: |
| 9715 | case SVE::BI__builtin_sve_svlen_u8: |
| 9716 | case SVE::BI__builtin_sve_svlen_u16: |
| 9717 | case SVE::BI__builtin_sve_svlen_u32: |
| 9718 | case SVE::BI__builtin_sve_svlen_u64: { |
| 9719 | SVETypeFlags TF(Builtin->TypeModifier); |
| 9720 | auto VTy = cast<llvm::VectorType>(getSVEType(TF)); |
| 9721 | auto *NumEls = |
| 9722 | llvm::ConstantInt::get(Ty, VTy->getElementCount().getKnownMinValue()); |
| 9723 | |
| 9724 | Function *F = CGM.getIntrinsic(Intrinsic::vscale, Ty); |
| 9725 | return Builder.CreateMul(NumEls, Builder.CreateCall(F)); |
| 9726 | } |
| 9727 | |
| 9728 | case SVE::BI__builtin_sve_svtbl2_u8: |
| 9729 | case SVE::BI__builtin_sve_svtbl2_s8: |
| 9730 | case SVE::BI__builtin_sve_svtbl2_u16: |
| 9731 | case SVE::BI__builtin_sve_svtbl2_s16: |
| 9732 | case SVE::BI__builtin_sve_svtbl2_u32: |
| 9733 | case SVE::BI__builtin_sve_svtbl2_s32: |
| 9734 | case SVE::BI__builtin_sve_svtbl2_u64: |
| 9735 | case SVE::BI__builtin_sve_svtbl2_s64: |
| 9736 | case SVE::BI__builtin_sve_svtbl2_f16: |
| 9737 | case SVE::BI__builtin_sve_svtbl2_bf16: |
| 9738 | case SVE::BI__builtin_sve_svtbl2_f32: |
| 9739 | case SVE::BI__builtin_sve_svtbl2_f64: { |
| 9740 | SVETypeFlags TF(Builtin->TypeModifier); |
| 9741 | auto VTy = cast<llvm::ScalableVectorType>(getSVEType(TF)); |
| 9742 | Value *V0 = Builder.CreateExtractVector(VTy, Ops[0], |
| 9743 | ConstantInt::get(CGM.Int64Ty, 0)); |
| 9744 | unsigned MinElts = VTy->getMinNumElements(); |
| 9745 | Value *V1 = Builder.CreateExtractVector( |
| 9746 | VTy, Ops[0], ConstantInt::get(CGM.Int64Ty, MinElts)); |
| 9747 | Function *F = CGM.getIntrinsic(Intrinsic::aarch64_sve_tbl2, VTy); |
| 9748 | return Builder.CreateCall(F, {V0, V1, Ops[1]}); |
| 9749 | } |
| 9750 | |
| 9751 | case SVE::BI__builtin_sve_svset_neonq_s8: |
| 9752 | case SVE::BI__builtin_sve_svset_neonq_s16: |
| 9753 | case SVE::BI__builtin_sve_svset_neonq_s32: |
| 9754 | case SVE::BI__builtin_sve_svset_neonq_s64: |
| 9755 | case SVE::BI__builtin_sve_svset_neonq_u8: |
| 9756 | case SVE::BI__builtin_sve_svset_neonq_u16: |
| 9757 | case SVE::BI__builtin_sve_svset_neonq_u32: |
| 9758 | case SVE::BI__builtin_sve_svset_neonq_u64: |
| 9759 | case SVE::BI__builtin_sve_svset_neonq_f16: |
| 9760 | case SVE::BI__builtin_sve_svset_neonq_f32: |
| 9761 | case SVE::BI__builtin_sve_svset_neonq_f64: |
| 9762 | case SVE::BI__builtin_sve_svset_neonq_bf16: { |
| 9763 | return Builder.CreateInsertVector(Ty, Ops[0], Ops[1], Builder.getInt64(0)); |
| 9764 | } |
| 9765 | |
| 9766 | case SVE::BI__builtin_sve_svget_neonq_s8: |
| 9767 | case SVE::BI__builtin_sve_svget_neonq_s16: |
| 9768 | case SVE::BI__builtin_sve_svget_neonq_s32: |
| 9769 | case SVE::BI__builtin_sve_svget_neonq_s64: |
| 9770 | case SVE::BI__builtin_sve_svget_neonq_u8: |
| 9771 | case SVE::BI__builtin_sve_svget_neonq_u16: |
| 9772 | case SVE::BI__builtin_sve_svget_neonq_u32: |
| 9773 | case SVE::BI__builtin_sve_svget_neonq_u64: |
| 9774 | case SVE::BI__builtin_sve_svget_neonq_f16: |
| 9775 | case SVE::BI__builtin_sve_svget_neonq_f32: |
| 9776 | case SVE::BI__builtin_sve_svget_neonq_f64: |
| 9777 | case SVE::BI__builtin_sve_svget_neonq_bf16: { |
| 9778 | return Builder.CreateExtractVector(Ty, Ops[0], Builder.getInt64(0)); |
| 9779 | } |
| 9780 | |
| 9781 | case SVE::BI__builtin_sve_svdup_neonq_s8: |
| 9782 | case SVE::BI__builtin_sve_svdup_neonq_s16: |
| 9783 | case SVE::BI__builtin_sve_svdup_neonq_s32: |
| 9784 | case SVE::BI__builtin_sve_svdup_neonq_s64: |
| 9785 | case SVE::BI__builtin_sve_svdup_neonq_u8: |
| 9786 | case SVE::BI__builtin_sve_svdup_neonq_u16: |
| 9787 | case SVE::BI__builtin_sve_svdup_neonq_u32: |
| 9788 | case SVE::BI__builtin_sve_svdup_neonq_u64: |
| 9789 | case SVE::BI__builtin_sve_svdup_neonq_f16: |
| 9790 | case SVE::BI__builtin_sve_svdup_neonq_f32: |
| 9791 | case SVE::BI__builtin_sve_svdup_neonq_f64: |
| 9792 | case SVE::BI__builtin_sve_svdup_neonq_bf16: { |
| 9793 | Value *Insert = Builder.CreateInsertVector(Ty, PoisonValue::get(Ty), Ops[0], |
| 9794 | Builder.getInt64(0)); |
| 9795 | return Builder.CreateIntrinsic(Intrinsic::aarch64_sve_dupq_lane, {Ty}, |
| 9796 | {Insert, Builder.getInt64(0)}); |
| 9797 | } |
| 9798 | } |
| 9799 | |
| 9800 | /// Should not happen |
| 9801 | return nullptr; |
| 9802 | } |
| 9803 | |
| 9804 | Value *CodeGenFunction::EmitAArch64BuiltinExpr(unsigned BuiltinID, |
| 9805 | const CallExpr *E, |
| 9806 | llvm::Triple::ArchType Arch) { |
| 9807 | if (BuiltinID >= clang::AArch64::FirstSVEBuiltin && |
| 9808 | BuiltinID <= clang::AArch64::LastSVEBuiltin) |
| 9809 | return EmitAArch64SVEBuiltinExpr(BuiltinID, E); |
| 9810 | |
| 9811 | unsigned HintID = static_cast<unsigned>(-1); |
| 9812 | switch (BuiltinID) { |
| 9813 | default: break; |
| 9814 | case clang::AArch64::BI__builtin_arm_nop: |
| 9815 | HintID = 0; |
| 9816 | break; |
| 9817 | case clang::AArch64::BI__builtin_arm_yield: |
| 9818 | case clang::AArch64::BI__yield: |
| 9819 | HintID = 1; |
| 9820 | break; |
| 9821 | case clang::AArch64::BI__builtin_arm_wfe: |
| 9822 | case clang::AArch64::BI__wfe: |
| 9823 | HintID = 2; |
| 9824 | break; |
| 9825 | case clang::AArch64::BI__builtin_arm_wfi: |
| 9826 | case clang::AArch64::BI__wfi: |
| 9827 | HintID = 3; |
| 9828 | break; |
| 9829 | case clang::AArch64::BI__builtin_arm_sev: |
| 9830 | case clang::AArch64::BI__sev: |
| 9831 | HintID = 4; |
| 9832 | break; |
| 9833 | case clang::AArch64::BI__builtin_arm_sevl: |
| 9834 | case clang::AArch64::BI__sevl: |
| 9835 | HintID = 5; |
| 9836 | break; |
| 9837 | } |
| 9838 | |
| 9839 | if (HintID != static_cast<unsigned>(-1)) { |
| 9840 | Function *F = CGM.getIntrinsic(Intrinsic::aarch64_hint); |
| 9841 | return Builder.CreateCall(F, llvm::ConstantInt::get(Int32Ty, HintID)); |
| 9842 | } |
| 9843 | |
| 9844 | if (BuiltinID == clang::AArch64::BI__builtin_arm_rbit) { |
| 9845 | assert((getContext().getTypeSize(E->getType()) == 32) &&(static_cast <bool> ((getContext().getTypeSize(E->getType ()) == 32) && "rbit of unusual size!") ? void (0) : __assert_fail ("(getContext().getTypeSize(E->getType()) == 32) && \"rbit of unusual size!\"" , "clang/lib/CodeGen/CGBuiltin.cpp", 9846, __extension__ __PRETTY_FUNCTION__ )) |
| 9846 | "rbit of unusual size!")(static_cast <bool> ((getContext().getTypeSize(E->getType ()) == 32) && "rbit of unusual size!") ? void (0) : __assert_fail ("(getContext().getTypeSize(E->getType()) == 32) && \"rbit of unusual size!\"" , "clang/lib/CodeGen/CGBuiltin.cpp", 9846, __extension__ __PRETTY_FUNCTION__ )); |
| 9847 | llvm::Value *Arg = EmitScalarExpr(E->getArg(0)); |
| 9848 | return Builder.CreateCall( |
| 9849 | CGM.getIntrinsic(Intrinsic::bitreverse, Arg->getType()), Arg, "rbit"); |
| 9850 | } |
| 9851 | if (BuiltinID == clang::AArch64::BI__builtin_arm_rbit64) { |
| 9852 | assert((getContext().getTypeSize(E->getType()) == 64) &&(static_cast <bool> ((getContext().getTypeSize(E->getType ()) == 64) && "rbit of unusual size!") ? void (0) : __assert_fail ("(getContext().getTypeSize(E->getType()) == 64) && \"rbit of unusual size!\"" , "clang/lib/CodeGen/CGBuiltin.cpp", 9853, __extension__ __PRETTY_FUNCTION__ )) |
| 9853 | "rbit of unusual size!")(static_cast <bool> ((getContext().getTypeSize(E->getType ()) == 64) && "rbit of unusual size!") ? void (0) : __assert_fail ("(getContext().getTypeSize(E->getType()) == 64) && \"rbit of unusual size!\"" , "clang/lib/CodeGen/CGBuiltin.cpp", 9853, __extension__ __PRETTY_FUNCTION__ )); |
| 9854 | llvm::Value *Arg = EmitScalarExpr(E->getArg(0)); |
| 9855 | return Builder.CreateCall( |
| 9856 | CGM.getIntrinsic(Intrinsic::bitreverse, Arg->getType()), Arg, "rbit"); |
| 9857 | } |
| 9858 | |
| 9859 | if (BuiltinID == clang::AArch64::BI__builtin_arm_cls) { |
| 9860 | llvm::Value *Arg = EmitScalarExpr(E->getArg(0)); |
| 9861 | return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::aarch64_cls), Arg, |
| 9862 | "cls"); |
| 9863 | } |
| 9864 | if (BuiltinID == clang::AArch64::BI__builtin_arm_cls64) { |
| 9865 | llvm::Value *Arg = EmitScalarExpr(E->getArg(0)); |
| 9866 | return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::aarch64_cls64), Arg, |
| 9867 | "cls"); |
| 9868 | } |
| 9869 | |
| 9870 | if (BuiltinID == clang::AArch64::BI__builtin_arm_rint32zf || |
| 9871 | BuiltinID == clang::AArch64::BI__builtin_arm_rint32z) { |
| 9872 | llvm::Value *Arg = EmitScalarExpr(E->getArg(0)); |
| 9873 | llvm::Type *Ty = Arg->getType(); |
| 9874 | return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::aarch64_frint32z, Ty), |
| 9875 | Arg, "frint32z"); |
| 9876 | } |
| 9877 | |
| 9878 | if (BuiltinID == clang::AArch64::BI__builtin_arm_rint64zf || |
| 9879 | BuiltinID == clang::AArch64::BI__builtin_arm_rint64z) { |
| 9880 | llvm::Value *Arg = EmitScalarExpr(E->getArg(0)); |
| 9881 | llvm::Type *Ty = Arg->getType(); |
| 9882 | return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::aarch64_frint64z, Ty), |
| 9883 | Arg, "frint64z"); |
| 9884 | } |
| 9885 | |
| 9886 | if (BuiltinID == clang::AArch64::BI__builtin_arm_rint32xf || |
| 9887 | BuiltinID == clang::AArch64::BI__builtin_arm_rint32x) { |
| 9888 | llvm::Value *Arg = EmitScalarExpr(E->getArg(0)); |
| 9889 | llvm::Type *Ty = Arg->getType(); |
| 9890 | return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::aarch64_frint32x, Ty), |
| 9891 | Arg, "frint32x"); |
| 9892 | } |
| 9893 | |
| 9894 | if (BuiltinID == clang::AArch64::BI__builtin_arm_rint64xf || |
| 9895 | BuiltinID == clang::AArch64::BI__builtin_arm_rint64x) { |
| 9896 | llvm::Value *Arg = EmitScalarExpr(E->getArg(0)); |
| 9897 | llvm::Type *Ty = Arg->getType(); |
| 9898 | return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::aarch64_frint64x, Ty), |
| 9899 | Arg, "frint64x"); |
| 9900 | } |
| 9901 | |
| 9902 | if (BuiltinID == clang::AArch64::BI__builtin_arm_jcvt) { |
| 9903 | assert((getContext().getTypeSize(E->getType()) == 32) &&(static_cast <bool> ((getContext().getTypeSize(E->getType ()) == 32) && "__jcvt of unusual size!") ? void (0) : __assert_fail ("(getContext().getTypeSize(E->getType()) == 32) && \"__jcvt of unusual size!\"" , "clang/lib/CodeGen/CGBuiltin.cpp", 9904, __extension__ __PRETTY_FUNCTION__ )) |
| 9904 | "__jcvt of unusual size!")(static_cast <bool> ((getContext().getTypeSize(E->getType ()) == 32) && "__jcvt of unusual size!") ? void (0) : __assert_fail ("(getContext().getTypeSize(E->getType()) == 32) && \"__jcvt of unusual size!\"" , "clang/lib/CodeGen/CGBuiltin.cpp", 9904, __extension__ __PRETTY_FUNCTION__ )); |
| 9905 | llvm::Value *Arg = EmitScalarExpr(E->getArg(0)); |
| 9906 | return Builder.CreateCall( |
| 9907 | CGM.getIntrinsic(Intrinsic::aarch64_fjcvtzs), Arg); |
| 9908 | } |
| 9909 | |
| 9910 | if (BuiltinID == clang::AArch64::BI__builtin_arm_ld64b || |
| 9911 | BuiltinID == clang::AArch64::BI__builtin_arm_st64b || |
| 9912 | BuiltinID == clang::AArch64::BI__builtin_arm_st64bv || |
| 9913 | BuiltinID == clang::AArch64::BI__builtin_arm_st64bv0) { |
| 9914 | llvm::Value *MemAddr = EmitScalarExpr(E->getArg(0)); |
| 9915 | llvm::Value *ValPtr = EmitScalarExpr(E->getArg(1)); |
| 9916 | |
| 9917 | if (BuiltinID == clang::AArch64::BI__builtin_arm_ld64b) { |
| 9918 | // Load from the address via an LLVM intrinsic, receiving a |
| 9919 | // tuple of 8 i64 words, and store each one to ValPtr. |
| 9920 | Function *F = CGM.getIntrinsic(Intrinsic::aarch64_ld64b); |
| 9921 | llvm::Value *Val = Builder.CreateCall(F, MemAddr); |
| 9922 | llvm::Value *ToRet; |
| 9923 | for (size_t i = 0; i < 8; i++) { |
| 9924 | llvm::Value *ValOffsetPtr = |
| 9925 | Builder.CreateGEP(Int64Ty, ValPtr, Builder.getInt32(i)); |
| 9926 | Address Addr = |
| 9927 | Address(ValOffsetPtr, Int64Ty, CharUnits::fromQuantity(8)); |
| 9928 | ToRet = Builder.CreateStore(Builder.CreateExtractValue(Val, i), Addr); |
| 9929 | } |
| 9930 | return ToRet; |
| 9931 | } else { |
| 9932 | // Load 8 i64 words from ValPtr, and store them to the address |
| 9933 | // via an LLVM intrinsic. |
| 9934 | SmallVector<llvm::Value *, 9> Args; |
| 9935 | Args.push_back(MemAddr); |
| 9936 | for (size_t i = 0; i < 8; i++) { |
| 9937 | llvm::Value *ValOffsetPtr = |
| 9938 | Builder.CreateGEP(Int64Ty, ValPtr, Builder.getInt32(i)); |
| 9939 | Address Addr = |
| 9940 | Address(ValOffsetPtr, Int64Ty, CharUnits::fromQuantity(8)); |
| 9941 | Args.push_back(Builder.CreateLoad(Addr)); |
| 9942 | } |
| 9943 | |
| 9944 | auto Intr = (BuiltinID == clang::AArch64::BI__builtin_arm_st64b |
| 9945 | ? Intrinsic::aarch64_st64b |
| 9946 | : BuiltinID == clang::AArch64::BI__builtin_arm_st64bv |
| 9947 | ? Intrinsic::aarch64_st64bv |
| 9948 | : Intrinsic::aarch64_st64bv0); |
| 9949 | Function *F = CGM.getIntrinsic(Intr); |
| 9950 | return Builder.CreateCall(F, Args); |
| 9951 | } |
| 9952 | } |
| 9953 | |
| 9954 | if (BuiltinID == clang::AArch64::BI__builtin_arm_rndr || |
| 9955 | BuiltinID == clang::AArch64::BI__builtin_arm_rndrrs) { |
| 9956 | |
| 9957 | auto Intr = (BuiltinID == clang::AArch64::BI__builtin_arm_rndr |
| 9958 | ? Intrinsic::aarch64_rndr |
| 9959 | : Intrinsic::aarch64_rndrrs); |
| 9960 | Function *F = CGM.getIntrinsic(Intr); |
| 9961 | llvm::Value *Val = Builder.CreateCall(F); |
| 9962 | Value *RandomValue = Builder.CreateExtractValue(Val, 0); |
| 9963 | Value *Status = Builder.CreateExtractValue(Val, 1); |
| 9964 | |
| 9965 | Address MemAddress = EmitPointerWithAlignment(E->getArg(0)); |
| 9966 | Builder.CreateStore(RandomValue, MemAddress); |
| 9967 | Status = Builder.CreateZExt(Status, Int32Ty); |
| 9968 | return Status; |
| 9969 | } |
| 9970 | |
| 9971 | if (BuiltinID == clang::AArch64::BI__clear_cache) { |
| 9972 | assert(E->getNumArgs() == 2 && "__clear_cache takes 2 arguments")(static_cast <bool> (E->getNumArgs() == 2 && "__clear_cache takes 2 arguments") ? void (0) : __assert_fail ("E->getNumArgs() == 2 && \"__clear_cache takes 2 arguments\"" , "clang/lib/CodeGen/CGBuiltin.cpp", 9972, __extension__ __PRETTY_FUNCTION__ )); |
| 9973 | const FunctionDecl *FD = E->getDirectCallee(); |
| 9974 | Value *Ops[2]; |
| 9975 | for (unsigned i = 0; i < 2; i++) |
| 9976 | Ops[i] = EmitScalarExpr(E->getArg(i)); |
| 9977 | llvm::Type *Ty = CGM.getTypes().ConvertType(FD->getType()); |
| 9978 | llvm::FunctionType *FTy = cast<llvm::FunctionType>(Ty); |
| 9979 | StringRef Name = FD->getName(); |
| 9980 | return EmitNounwindRuntimeCall(CGM.CreateRuntimeFunction(FTy, Name), Ops); |
| 9981 | } |
| 9982 | |
| 9983 | if ((BuiltinID == clang::AArch64::BI__builtin_arm_ldrex || |
| 9984 | BuiltinID == clang::AArch64::BI__builtin_arm_ldaex) && |
| 9985 | getContext().getTypeSize(E->getType()) == 128) { |
| 9986 | Function *F = |
| 9987 | CGM.getIntrinsic(BuiltinID == clang::AArch64::BI__builtin_arm_ldaex |
| 9988 | ? Intrinsic::aarch64_ldaxp |
| 9989 | : Intrinsic::aarch64_ldxp); |
| 9990 | |
| 9991 | Value *LdPtr = EmitScalarExpr(E->getArg(0)); |
| 9992 | Value *Val = Builder.CreateCall(F, Builder.CreateBitCast(LdPtr, Int8PtrTy), |
| 9993 | "ldxp"); |
| 9994 | |
| 9995 | Value *Val0 = Builder.CreateExtractValue(Val, 1); |
| 9996 | Value *Val1 = Builder.CreateExtractValue(Val, 0); |
| 9997 | llvm::Type *Int128Ty = llvm::IntegerType::get(getLLVMContext(), 128); |
| 9998 | Val0 = Builder.CreateZExt(Val0, Int128Ty); |
| 9999 | Val1 = Builder.CreateZExt(Val1, Int128Ty); |
| 10000 | |
| 10001 | Value *ShiftCst = llvm::ConstantInt::get(Int128Ty, 64); |
| 10002 | Val = Builder.CreateShl(Val0, ShiftCst, "shl", true /* nuw */); |
| 10003 | Val = Builder.CreateOr(Val, Val1); |
| 10004 | return Builder.CreateBitCast(Val, ConvertType(E->getType())); |
| 10005 | } else if (BuiltinID == clang::AArch64::BI__builtin_arm_ldrex || |
| 10006 | BuiltinID == clang::AArch64::BI__builtin_arm_ldaex) { |
| 10007 | Value *LoadAddr = EmitScalarExpr(E->getArg(0)); |
| 10008 | |
| 10009 | QualType Ty = E->getType(); |
| 10010 | llvm::Type *RealResTy = ConvertType(Ty); |
| 10011 | llvm::Type *IntTy = |
| 10012 | llvm::IntegerType::get(getLLVMContext(), getContext().getTypeSize(Ty)); |
| 10013 | llvm::Type *PtrTy = IntTy->getPointerTo(); |
| 10014 | LoadAddr = Builder.CreateBitCast(LoadAddr, PtrTy); |
| 10015 | |
| 10016 | Function *F = |
| 10017 | CGM.getIntrinsic(BuiltinID == clang::AArch64::BI__builtin_arm_ldaex |
| 10018 | ? Intrinsic::aarch64_ldaxr |
| 10019 | : Intrinsic::aarch64_ldxr, |
| 10020 | PtrTy); |
| 10021 | CallInst *Val = Builder.CreateCall(F, LoadAddr, "ldxr"); |
| 10022 | Val->addParamAttr( |
| 10023 | 0, Attribute::get(getLLVMContext(), Attribute::ElementType, IntTy)); |
| 10024 | |
| 10025 | if (RealResTy->isPointerTy()) |
| 10026 | return Builder.CreateIntToPtr(Val, RealResTy); |
| 10027 | |
| 10028 | llvm::Type *IntResTy = llvm::IntegerType::get( |
| 10029 | getLLVMContext(), CGM.getDataLayout().getTypeSizeInBits(RealResTy)); |
| 10030 | return Builder.CreateBitCast(Builder.CreateTruncOrBitCast(Val, IntResTy), |
| 10031 | RealResTy); |
| 10032 | } |
| 10033 | |
| 10034 | if ((BuiltinID == clang::AArch64::BI__builtin_arm_strex || |
| 10035 | BuiltinID == clang::AArch64::BI__builtin_arm_stlex) && |
| 10036 | getContext().getTypeSize(E->getArg(0)->getType()) == 128) { |
| 10037 | Function *F = |
| 10038 | CGM.getIntrinsic(BuiltinID == clang::AArch64::BI__builtin_arm_stlex |
| 10039 | ? Intrinsic::aarch64_stlxp |
| 10040 | : Intrinsic::aarch64_stxp); |
| 10041 | llvm::Type *STy = llvm::StructType::get(Int64Ty, Int64Ty); |
| 10042 | |
| 10043 | Address Tmp = CreateMemTemp(E->getArg(0)->getType()); |
| 10044 | EmitAnyExprToMem(E->getArg(0), Tmp, Qualifiers(), /*init*/ true); |
| 10045 | |
| 10046 | Tmp = Builder.CreateElementBitCast(Tmp, STy); |
| 10047 | llvm::Value *Val = Builder.CreateLoad(Tmp); |
| 10048 | |
| 10049 | Value *Arg0 = Builder.CreateExtractValue(Val, 0); |
| 10050 | Value *Arg1 = Builder.CreateExtractValue(Val, 1); |
| 10051 | Value *StPtr = Builder.CreateBitCast(EmitScalarExpr(E->getArg(1)), |
| 10052 | Int8PtrTy); |
| 10053 | return Builder.CreateCall(F, {Arg0, Arg1, StPtr}, "stxp"); |
| 10054 | } |
| 10055 | |
| 10056 | if (BuiltinID == clang::AArch64::BI__builtin_arm_strex || |
| 10057 | BuiltinID == clang::AArch64::BI__builtin_arm_stlex) { |
| 10058 | Value *StoreVal = EmitScalarExpr(E->getArg(0)); |
| 10059 | Value *StoreAddr = EmitScalarExpr(E->getArg(1)); |
| 10060 | |
| 10061 | QualType Ty = E->getArg(0)->getType(); |
| 10062 | llvm::Type *StoreTy = llvm::IntegerType::get(getLLVMContext(), |
| 10063 | getContext().getTypeSize(Ty)); |
| 10064 | StoreAddr = Builder.CreateBitCast(StoreAddr, StoreTy->getPointerTo()); |
| 10065 | |
| 10066 | if (StoreVal->getType()->isPointerTy()) |
| 10067 | StoreVal = Builder.CreatePtrToInt(StoreVal, Int64Ty); |
| 10068 | else { |
| 10069 | llvm::Type *IntTy = llvm::IntegerType::get( |
| 10070 | getLLVMContext(), |
| 10071 | CGM.getDataLayout().getTypeSizeInBits(StoreVal->getType())); |
| 10072 | StoreVal = Builder.CreateBitCast(StoreVal, IntTy); |
| 10073 | StoreVal = Builder.CreateZExtOrBitCast(StoreVal, Int64Ty); |
| 10074 | } |
| 10075 | |
| 10076 | Function *F = |
| 10077 | CGM.getIntrinsic(BuiltinID == clang::AArch64::BI__builtin_arm_stlex |
| 10078 | ? Intrinsic::aarch64_stlxr |
| 10079 | : Intrinsic::aarch64_stxr, |
| 10080 | StoreAddr->getType()); |
| 10081 | CallInst *CI = Builder.CreateCall(F, {StoreVal, StoreAddr}, "stxr"); |
| 10082 | CI->addParamAttr( |
| 10083 | 1, Attribute::get(getLLVMContext(), Attribute::ElementType, StoreTy)); |
| 10084 | return CI; |
| 10085 | } |
| 10086 | |
| 10087 | if (BuiltinID == clang::AArch64::BI__getReg) { |
| 10088 | Expr::EvalResult Result; |
| 10089 | if (!E->getArg(0)->EvaluateAsInt(Result, CGM.getContext())) |
| 10090 | llvm_unreachable("Sema will ensure that the parameter is constant")::llvm::llvm_unreachable_internal("Sema will ensure that the parameter is constant" , "clang/lib/CodeGen/CGBuiltin.cpp", 10090); |
| 10091 | |
| 10092 | llvm::APSInt Value = Result.Val.getInt(); |
| 10093 | LLVMContext &Context = CGM.getLLVMContext(); |
| 10094 | std::string Reg = Value == 31 ? "sp" : "x" + toString(Value, 10); |
| 10095 | |
| 10096 | llvm::Metadata *Ops[] = {llvm::MDString::get(Context, Reg)}; |
| 10097 | llvm::MDNode *RegName = llvm::MDNode::get(Context, Ops); |
| 10098 | llvm::Value *Metadata = llvm::MetadataAsValue::get(Context, RegName); |
| 10099 | |
| 10100 | llvm::Function *F = |
| 10101 | CGM.getIntrinsic(llvm::Intrinsic::read_register, {Int64Ty}); |
| 10102 | return Builder.CreateCall(F, Metadata); |
| 10103 | } |
| 10104 | |
| 10105 | if (BuiltinID == clang::AArch64::BI__break) { |
| 10106 | Expr::EvalResult Result; |
| 10107 | if (!E->getArg(0)->EvaluateAsInt(Result, CGM.getContext())) |
| 10108 | llvm_unreachable("Sema will ensure that the parameter is constant")::llvm::llvm_unreachable_internal("Sema will ensure that the parameter is constant" , "clang/lib/CodeGen/CGBuiltin.cpp", 10108); |
| 10109 | |
| 10110 | llvm::Function *F = CGM.getIntrinsic(llvm::Intrinsic::aarch64_break); |
| 10111 | return Builder.CreateCall(F, {EmitScalarExpr(E->getArg(0))}); |
| 10112 | } |
| 10113 | |
| 10114 | if (BuiltinID == clang::AArch64::BI__builtin_arm_clrex) { |
| 10115 | Function *F = CGM.getIntrinsic(Intrinsic::aarch64_clrex); |
| 10116 | return Builder.CreateCall(F); |
| 10117 | } |
| 10118 | |
| 10119 | if (BuiltinID == clang::AArch64::BI_ReadWriteBarrier) |
| 10120 | return Builder.CreateFence(llvm::AtomicOrdering::SequentiallyConsistent, |
| 10121 | llvm::SyncScope::SingleThread); |
| 10122 | |
| 10123 | // CRC32 |
| 10124 | Intrinsic::ID CRCIntrinsicID = Intrinsic::not_intrinsic; |
| 10125 | switch (BuiltinID) { |
| 10126 | case clang::AArch64::BI__builtin_arm_crc32b: |
| 10127 | CRCIntrinsicID = Intrinsic::aarch64_crc32b; break; |
| 10128 | case clang::AArch64::BI__builtin_arm_crc32cb: |
| 10129 | CRCIntrinsicID = Intrinsic::aarch64_crc32cb; break; |
| 10130 | case clang::AArch64::BI__builtin_arm_crc32h: |
| 10131 | CRCIntrinsicID = Intrinsic::aarch64_crc32h; break; |
| 10132 | case clang::AArch64::BI__builtin_arm_crc32ch: |
| 10133 | CRCIntrinsicID = Intrinsic::aarch64_crc32ch; break; |
| 10134 | case clang::AArch64::BI__builtin_arm_crc32w: |
| 10135 | CRCIntrinsicID = Intrinsic::aarch64_crc32w; break; |
| 10136 | case clang::AArch64::BI__builtin_arm_crc32cw: |
| 10137 | CRCIntrinsicID = Intrinsic::aarch64_crc32cw; break; |
| 10138 | case clang::AArch64::BI__builtin_arm_crc32d: |
| 10139 | CRCIntrinsicID = Intrinsic::aarch64_crc32x; break; |
| 10140 | case clang::AArch64::BI__builtin_arm_crc32cd: |
| 10141 | CRCIntrinsicID = Intrinsic::aarch64_crc32cx; break; |
| 10142 | } |
| 10143 | |
| 10144 | if (CRCIntrinsicID != Intrinsic::not_intrinsic) { |
| 10145 | Value *Arg0 = EmitScalarExpr(E->getArg(0)); |
| 10146 | Value *Arg1 = EmitScalarExpr(E->getArg(1)); |
| 10147 | Function *F = CGM.getIntrinsic(CRCIntrinsicID); |
| 10148 | |
| 10149 | llvm::Type *DataTy = F->getFunctionType()->getParamType(1); |
| 10150 | Arg1 = Builder.CreateZExtOrBitCast(Arg1, DataTy); |
| 10151 | |
| 10152 | return Builder.CreateCall(F, {Arg0, Arg1}); |
| 10153 | } |
| 10154 | |
| 10155 | // Memory Operations (MOPS) |
| 10156 | if (BuiltinID == AArch64::BI__builtin_arm_mops_memset_tag) { |
| 10157 | Value *Dst = EmitScalarExpr(E->getArg(0)); |
| 10158 | Value *Val = EmitScalarExpr(E->getArg(1)); |
| 10159 | Value *Size = EmitScalarExpr(E->getArg(2)); |
| 10160 | Dst = Builder.CreatePointerCast(Dst, Int8PtrTy); |
| 10161 | Val = Builder.CreateTrunc(Val, Int8Ty); |
| 10162 | Size = Builder.CreateIntCast(Size, Int64Ty, false); |
| 10163 | return Builder.CreateCall( |
| 10164 | CGM.getIntrinsic(Intrinsic::aarch64_mops_memset_tag), {Dst, Val, Size}); |
| 10165 | } |
| 10166 | |
| 10167 | // Memory Tagging Extensions (MTE) Intrinsics |
| 10168 | Intrinsic::ID MTEIntrinsicID = Intrinsic::not_intrinsic; |
| 10169 | switch (BuiltinID) { |
| 10170 | case clang::AArch64::BI__builtin_arm_irg: |
| 10171 | MTEIntrinsicID = Intrinsic::aarch64_irg; break; |
| 10172 | case clang::AArch64::BI__builtin_arm_addg: |
| 10173 | MTEIntrinsicID = Intrinsic::aarch64_addg; break; |
| 10174 | case clang::AArch64::BI__builtin_arm_gmi: |
| 10175 | MTEIntrinsicID = Intrinsic::aarch64_gmi; break; |
| 10176 | case clang::AArch64::BI__builtin_arm_ldg: |
| 10177 | MTEIntrinsicID = Intrinsic::aarch64_ldg; break; |
| 10178 | case clang::AArch64::BI__builtin_arm_stg: |
| 10179 | MTEIntrinsicID = Intrinsic::aarch64_stg; break; |
| 10180 | case clang::AArch64::BI__builtin_arm_subp: |
| 10181 | MTEIntrinsicID = Intrinsic::aarch64_subp; break; |
| 10182 | } |
| 10183 | |
| 10184 | if (MTEIntrinsicID != Intrinsic::not_intrinsic) { |
| 10185 | llvm::Type *T = ConvertType(E->getType()); |
| 10186 | |
| 10187 | if (MTEIntrinsicID == Intrinsic::aarch64_irg) { |
| 10188 | Value *Pointer = EmitScalarExpr(E->getArg(0)); |
| 10189 | Value *Mask = EmitScalarExpr(E->getArg(1)); |
| 10190 | |
| 10191 | Pointer = Builder.CreatePointerCast(Pointer, Int8PtrTy); |
| 10192 | Mask = Builder.CreateZExt(Mask, Int64Ty); |
| 10193 | Value *RV = Builder.CreateCall( |
| 10194 | CGM.getIntrinsic(MTEIntrinsicID), {Pointer, Mask}); |
| 10195 | return Builder.CreatePointerCast(RV, T); |
| 10196 | } |
| 10197 | if (MTEIntrinsicID == Intrinsic::aarch64_addg) { |
| 10198 | Value *Pointer = EmitScalarExpr(E->getArg(0)); |
| 10199 | Value *TagOffset = EmitScalarExpr(E->getArg(1)); |
| 10200 | |
| 10201 | Pointer = Builder.CreatePointerCast(Pointer, Int8PtrTy); |
| 10202 | TagOffset = Builder.CreateZExt(TagOffset, Int64Ty); |
| 10203 | Value *RV = Builder.CreateCall( |
| 10204 | CGM.getIntrinsic(MTEIntrinsicID), {Pointer, TagOffset}); |
| 10205 | return Builder.CreatePointerCast(RV, T); |
| 10206 | } |
| 10207 | if (MTEIntrinsicID == Intrinsic::aarch64_gmi) { |
| 10208 | Value *Pointer = EmitScalarExpr(E->getArg(0)); |
| 10209 | Value *ExcludedMask = EmitScalarExpr(E->getArg(1)); |
| 10210 | |
| 10211 | ExcludedMask = Builder.CreateZExt(ExcludedMask, Int64Ty); |
| 10212 | Pointer = Builder.CreatePointerCast(Pointer, Int8PtrTy); |
| 10213 | return Builder.CreateCall( |
| 10214 | CGM.getIntrinsic(MTEIntrinsicID), {Pointer, ExcludedMask}); |
| 10215 | } |
| 10216 | // Although it is possible to supply a different return |
| 10217 | // address (first arg) to this intrinsic, for now we set |
| 10218 | // return address same as input address. |
| 10219 | if (MTEIntrinsicID == Intrinsic::aarch64_ldg) { |
| 10220 | Value *TagAddress = EmitScalarExpr(E->getArg(0)); |
| 10221 | TagAddress = Builder.CreatePointerCast(TagAddress, Int8PtrTy); |
| 10222 | Value *RV = Builder.CreateCall( |
| 10223 | CGM.getIntrinsic(MTEIntrinsicID), {TagAddress, TagAddress}); |
| 10224 | return Builder.CreatePointerCast(RV, T); |
| 10225 | } |
| 10226 | // Although it is possible to supply a different tag (to set) |
| 10227 | // to this intrinsic (as first arg), for now we supply |
| 10228 | // the tag that is in input address arg (common use case). |
| 10229 | if (MTEIntrinsicID == Intrinsic::aarch64_stg) { |
| 10230 | Value *TagAddress = EmitScalarExpr(E->getArg(0)); |
| 10231 | TagAddress = Builder.CreatePointerCast(TagAddress, Int8PtrTy); |
| 10232 | return Builder.CreateCall( |
| 10233 | CGM.getIntrinsic(MTEIntrinsicID), {TagAddress, TagAddress}); |
| 10234 | } |
| 10235 | if (MTEIntrinsicID == Intrinsic::aarch64_subp) { |
| 10236 | Value *PointerA = EmitScalarExpr(E->getArg(0)); |
| 10237 | Value *PointerB = EmitScalarExpr(E->getArg(1)); |
| 10238 | PointerA = Builder.CreatePointerCast(PointerA, Int8PtrTy); |
| 10239 | PointerB = Builder.CreatePointerCast(PointerB, Int8PtrTy); |
| 10240 | return Builder.CreateCall( |
| 10241 | CGM.getIntrinsic(MTEIntrinsicID), {PointerA, PointerB}); |
| 10242 | } |
| 10243 | } |
| 10244 | |
| 10245 | if (BuiltinID == clang::AArch64::BI__builtin_arm_rsr || |
| 10246 | BuiltinID == clang::AArch64::BI__builtin_arm_rsr64 || |
| 10247 | BuiltinID == clang::AArch64::BI__builtin_arm_rsr128 || |
| 10248 | BuiltinID == clang::AArch64::BI__builtin_arm_rsrp || |
| 10249 | BuiltinID == clang::AArch64::BI__builtin_arm_wsr || |
| 10250 | BuiltinID == clang::AArch64::BI__builtin_arm_wsr64 || |
| 10251 | BuiltinID == clang::AArch64::BI__builtin_arm_wsr128 || |
| 10252 | BuiltinID == clang::AArch64::BI__builtin_arm_wsrp) { |
| 10253 | |
| 10254 | SpecialRegisterAccessKind AccessKind = Write; |
| 10255 | if (BuiltinID == clang::AArch64::BI__builtin_arm_rsr || |
| 10256 | BuiltinID == clang::AArch64::BI__builtin_arm_rsr64 || |
| 10257 | BuiltinID == clang::AArch64::BI__builtin_arm_rsr128 || |
| 10258 | BuiltinID == clang::AArch64::BI__builtin_arm_rsrp) |
| 10259 | AccessKind = VolatileRead; |
| 10260 | |
| 10261 | bool IsPointerBuiltin = BuiltinID == clang::AArch64::BI__builtin_arm_rsrp || |
| 10262 | BuiltinID == clang::AArch64::BI__builtin_arm_wsrp; |
| 10263 | |
| 10264 | bool Is32Bit = BuiltinID == clang::AArch64::BI__builtin_arm_rsr || |
| 10265 | BuiltinID == clang::AArch64::BI__builtin_arm_wsr; |
| 10266 | |
| 10267 | bool Is128Bit = BuiltinID == clang::AArch64::BI__builtin_arm_rsr128 || |
| 10268 | BuiltinID == clang::AArch64::BI__builtin_arm_wsr128; |
| 10269 | |
| 10270 | llvm::Type *ValueType; |
| 10271 | llvm::Type *RegisterType = Int64Ty; |
| 10272 | if (Is32Bit) { |
| 10273 | ValueType = Int32Ty; |
| 10274 | } else if (Is128Bit) { |
| 10275 | llvm::Type *Int128Ty = |
| 10276 | llvm::IntegerType::getInt128Ty(CGM.getLLVMContext()); |
| 10277 | ValueType = Int128Ty; |
| 10278 | RegisterType = Int128Ty; |
| 10279 | } else if (IsPointerBuiltin) { |
| 10280 | ValueType = VoidPtrTy; |
| 10281 | } else { |
| 10282 | ValueType = Int64Ty; |
| 10283 | }; |
| 10284 | |
| 10285 | return EmitSpecialRegisterBuiltin(*this, E, RegisterType, ValueType, |
| 10286 | AccessKind); |
| 10287 | } |
| 10288 | |
| 10289 | if (BuiltinID == clang::AArch64::BI_ReadStatusReg || |
| 10290 | BuiltinID == clang::AArch64::BI_WriteStatusReg) { |
| 10291 | LLVMContext &Context = CGM.getLLVMContext(); |
| 10292 | |
| 10293 | unsigned SysReg = |
| 10294 | E->getArg(0)->EvaluateKnownConstInt(getContext()).getZExtValue(); |
| 10295 | |
| 10296 | std::string SysRegStr; |
| 10297 | llvm::raw_string_ostream(SysRegStr) << |
| 10298 | ((1 << 1) | ((SysReg >> 14) & 1)) << ":" << |
| 10299 | ((SysReg >> 11) & 7) << ":" << |
| 10300 | ((SysReg >> 7) & 15) << ":" << |
| 10301 | ((SysReg >> 3) & 15) << ":" << |
| 10302 | ( SysReg & 7); |
| 10303 | |
| 10304 | llvm::Metadata *Ops[] = { llvm::MDString::get(Context, SysRegStr) }; |
| 10305 | llvm::MDNode *RegName = llvm::MDNode::get(Context, Ops); |
| 10306 | llvm::Value *Metadata = llvm::MetadataAsValue::get(Context, RegName); |
| 10307 | |
| 10308 | llvm::Type *RegisterType = Int64Ty; |
| 10309 | llvm::Type *Types[] = { RegisterType }; |
| 10310 | |
| 10311 | if (BuiltinID == clang::AArch64::BI_ReadStatusReg) { |
| 10312 | llvm::Function *F = CGM.getIntrinsic(llvm::Intrinsic::read_register, Types); |
| 10313 | |
| 10314 | return Builder.CreateCall(F, Metadata); |
| 10315 | } |
| 10316 | |
| 10317 | llvm::Function *F = CGM.getIntrinsic(llvm::Intrinsic::write_register, Types); |
| 10318 | llvm::Value *ArgValue = EmitScalarExpr(E->getArg(1)); |
| 10319 | |
| 10320 | return Builder.CreateCall(F, { Metadata, ArgValue }); |
| 10321 | } |
| 10322 | |
| 10323 | if (BuiltinID == clang::AArch64::BI_AddressOfReturnAddress) { |
| 10324 | llvm::Function *F = |
| 10325 | CGM.getIntrinsic(Intrinsic::addressofreturnaddress, AllocaInt8PtrTy); |
| 10326 | return Builder.CreateCall(F); |
| 10327 | } |
| 10328 | |
| 10329 | if (BuiltinID == clang::AArch64::BI__builtin_sponentry) { |
| 10330 | llvm::Function *F = CGM.getIntrinsic(Intrinsic::sponentry, AllocaInt8PtrTy); |
| 10331 | return Builder.CreateCall(F); |
| 10332 | } |
| 10333 | |
| 10334 | if (BuiltinID == clang::AArch64::BI__mulh || |
| 10335 | BuiltinID == clang::AArch64::BI__umulh) { |
| 10336 | llvm::Type *ResType = ConvertType(E->getType()); |
| 10337 | llvm::Type *Int128Ty = llvm::IntegerType::get(getLLVMContext(), 128); |
| 10338 | |
| 10339 | bool IsSigned = BuiltinID == clang::AArch64::BI__mulh; |
| 10340 | Value *LHS = |
| 10341 | Builder.CreateIntCast(EmitScalarExpr(E->getArg(0)), Int128Ty, IsSigned); |
| 10342 | Value *RHS = |
| 10343 | Builder.CreateIntCast(EmitScalarExpr(E->getArg(1)), Int128Ty, IsSigned); |
| 10344 | |
| 10345 | Value *MulResult, *HigherBits; |
| 10346 | if (IsSigned) { |
| 10347 | MulResult = Builder.CreateNSWMul(LHS, RHS); |
| 10348 | HigherBits = Builder.CreateAShr(MulResult, 64); |
| 10349 | } else { |
| 10350 | MulResult = Builder.CreateNUWMul(LHS, RHS); |
| 10351 | HigherBits = Builder.CreateLShr(MulResult, 64); |
| 10352 | } |
| 10353 | HigherBits = Builder.CreateIntCast(HigherBits, ResType, IsSigned); |
| 10354 | |
| 10355 | return HigherBits; |
| 10356 | } |
| 10357 | |
| 10358 | if (BuiltinID == AArch64::BI__writex18byte || |
| 10359 | BuiltinID == AArch64::BI__writex18word || |
| 10360 | BuiltinID == AArch64::BI__writex18dword || |
| 10361 | BuiltinID == AArch64::BI__writex18qword) { |
| 10362 | llvm::Type *IntTy = ConvertType(E->getArg(1)->getType()); |
| 10363 | |
| 10364 | // Read x18 as i8* |
| 10365 | LLVMContext &Context = CGM.getLLVMContext(); |
| 10366 | llvm::Metadata *Ops[] = {llvm::MDString::get(Context, "x18")}; |
| 10367 | llvm::MDNode *RegName = llvm::MDNode::get(Context, Ops); |
| 10368 | llvm::Value *Metadata = llvm::MetadataAsValue::get(Context, RegName); |
| 10369 | llvm::Function *F = |
| 10370 | CGM.getIntrinsic(llvm::Intrinsic::read_register, {Int64Ty}); |
| 10371 | llvm::Value *X18 = Builder.CreateCall(F, Metadata); |
| 10372 | X18 = Builder.CreateIntToPtr(X18, llvm::PointerType::get(Int8Ty, 0)); |
| 10373 | |
| 10374 | // Store val at x18 + offset |
| 10375 | Value *Offset = Builder.CreateZExt(EmitScalarExpr(E->getArg(0)), Int64Ty); |
| 10376 | Value *Ptr = Builder.CreateGEP(Int8Ty, X18, Offset); |
| 10377 | Ptr = Builder.CreatePointerCast(Ptr, llvm::PointerType::get(IntTy, 0)); |
| 10378 | Value *Val = EmitScalarExpr(E->getArg(1)); |
| 10379 | StoreInst *Store = Builder.CreateAlignedStore(Val, Ptr, CharUnits::One()); |
| 10380 | return Store; |
| 10381 | } |
| 10382 | |
| 10383 | if (BuiltinID == AArch64::BI__readx18byte || |
| 10384 | BuiltinID == AArch64::BI__readx18word || |
| 10385 | BuiltinID == AArch64::BI__readx18dword || |
| 10386 | BuiltinID == AArch64::BI__readx18qword) { |
| 10387 | llvm::Type *IntTy = ConvertType(E->getType()); |
| 10388 | |
| 10389 | // Read x18 as i8* |
| 10390 | LLVMContext &Context = CGM.getLLVMContext(); |
| 10391 | llvm::Metadata *Ops[] = {llvm::MDString::get(Context, "x18")}; |
| 10392 | llvm::MDNode *RegName = llvm::MDNode::get(Context, Ops); |
| 10393 | llvm::Value *Metadata = llvm::MetadataAsValue::get(Context, RegName); |
| 10394 | llvm::Function *F = |
| 10395 | CGM.getIntrinsic(llvm::Intrinsic::read_register, {Int64Ty}); |
| 10396 | llvm::Value *X18 = Builder.CreateCall(F, Metadata); |
| 10397 | X18 = Builder.CreateIntToPtr(X18, llvm::PointerType::get(Int8Ty, 0)); |
| 10398 | |
| 10399 | // Load x18 + offset |
| 10400 | Value *Offset = Builder.CreateZExt(EmitScalarExpr(E->getArg(0)), Int64Ty); |
| 10401 | Value *Ptr = Builder.CreateGEP(Int8Ty, X18, Offset); |
| 10402 | Ptr = Builder.CreatePointerCast(Ptr, llvm::PointerType::get(IntTy, 0)); |
| 10403 | LoadInst *Load = Builder.CreateAlignedLoad(IntTy, Ptr, CharUnits::One()); |
| 10404 | return Load; |
| 10405 | } |
| 10406 | |
| 10407 | // Handle MSVC intrinsics before argument evaluation to prevent double |
| 10408 | // evaluation. |
| 10409 | if (std::optional<MSVCIntrin> MsvcIntId = |
| 10410 | translateAarch64ToMsvcIntrin(BuiltinID)) |
| 10411 | return EmitMSVCBuiltinExpr(*MsvcIntId, E); |
| 10412 | |
| 10413 | // Some intrinsics are equivalent - if they are use the base intrinsic ID. |
| 10414 | auto It = llvm::find_if(NEONEquivalentIntrinsicMap, [BuiltinID](auto &P) { |
| 10415 | return P.first == BuiltinID; |
| 10416 | }); |
| 10417 | if (It != end(NEONEquivalentIntrinsicMap)) |
| 10418 | BuiltinID = It->second; |
| 10419 | |
| 10420 | // Find out if any arguments are required to be integer constant |
| 10421 | // expressions. |
| 10422 | unsigned ICEArguments = 0; |
| 10423 | ASTContext::GetBuiltinTypeError Error; |
| 10424 | getContext().GetBuiltinType(BuiltinID, Error, &ICEArguments); |
| 10425 | assert(Error == ASTContext::GE_None && "Should not codegen an error")(static_cast <bool> (Error == ASTContext::GE_None && "Should not codegen an error") ? void (0) : __assert_fail ("Error == ASTContext::GE_None && \"Should not codegen an error\"" , "clang/lib/CodeGen/CGBuiltin.cpp", 10425, __extension__ __PRETTY_FUNCTION__ )); |
| 10426 | |
| 10427 | llvm::SmallVector<Value*, 4> Ops; |
| 10428 | Address PtrOp0 = Address::invalid(); |
| 10429 | for (unsigned i = 0, e = E->getNumArgs() - 1; i != e; i++) { |
| 10430 | if (i == 0) { |
| 10431 | switch (BuiltinID) { |
| 10432 | case NEON::BI__builtin_neon_vld1_v: |
| 10433 | case NEON::BI__builtin_neon_vld1q_v: |
| 10434 | case NEON::BI__builtin_neon_vld1_dup_v: |
| 10435 | case NEON::BI__builtin_neon_vld1q_dup_v: |
| 10436 | case NEON::BI__builtin_neon_vld1_lane_v: |
| 10437 | case NEON::BI__builtin_neon_vld1q_lane_v: |
| 10438 | case NEON::BI__builtin_neon_vst1_v: |
| 10439 | case NEON::BI__builtin_neon_vst1q_v: |
| 10440 | case NEON::BI__builtin_neon_vst1_lane_v: |
| 10441 | case NEON::BI__builtin_neon_vst1q_lane_v: |
| 10442 | // Get the alignment for the argument in addition to the value; |
| 10443 | // we'll use it later. |
| 10444 | PtrOp0 = EmitPointerWithAlignment(E->getArg(0)); |
| 10445 | Ops.push_back(PtrOp0.getPointer()); |
| 10446 | continue; |
| 10447 | } |
| 10448 | } |
| 10449 | if ((ICEArguments & (1 << i)) == 0) { |
| 10450 | Ops.push_back(EmitScalarExpr(E->getArg(i))); |
| 10451 | } else { |
| 10452 | // If this is required to be a constant, constant fold it so that we know |
| 10453 | // that the generated intrinsic gets a ConstantInt. |
| 10454 | Ops.push_back(llvm::ConstantInt::get( |
| 10455 | getLLVMContext(), |
| 10456 | *E->getArg(i)->getIntegerConstantExpr(getContext()))); |
| 10457 | } |
| 10458 | } |
| 10459 | |
| 10460 | auto SISDMap = ArrayRef(AArch64SISDIntrinsicMap); |
| 10461 | const ARMVectorIntrinsicInfo *Builtin = findARMVectorIntrinsicInMap( |
| 10462 | SISDMap, BuiltinID, AArch64SISDIntrinsicsProvenSorted); |
| 10463 | |
| 10464 | if (Builtin) { |
| 10465 | Ops.push_back(EmitScalarExpr(E->getArg(E->getNumArgs() - 1))); |
| 10466 | Value *Result = EmitCommonNeonSISDBuiltinExpr(*this, *Builtin, Ops, E); |
| 10467 | assert(Result && "SISD intrinsic should have been handled")(static_cast <bool> (Result && "SISD intrinsic should have been handled" ) ? void (0) : __assert_fail ("Result && \"SISD intrinsic should have been handled\"" , "clang/lib/CodeGen/CGBuiltin.cpp", 10467, __extension__ __PRETTY_FUNCTION__ )); |
| 10468 | return Result; |
| 10469 | } |
| 10470 | |
| 10471 | const Expr *Arg = E->getArg(E->getNumArgs()-1); |
| 10472 | NeonTypeFlags Type(0); |
| 10473 | if (std::optional<llvm::APSInt> Result = |
| 10474 | Arg->getIntegerConstantExpr(getContext())) |
| 10475 | // Determine the type of this overloaded NEON intrinsic. |
| 10476 | Type = NeonTypeFlags(Result->getZExtValue()); |
| 10477 | |
| 10478 | bool usgn = Type.isUnsigned(); |
| 10479 | bool quad = Type.isQuad(); |
| 10480 | |
| 10481 | // Handle non-overloaded intrinsics first. |
| 10482 | switch (BuiltinID) { |
| 10483 | default: break; |
| 10484 | case NEON::BI__builtin_neon_vabsh_f16: |
| 10485 | Ops.push_back(EmitScalarExpr(E->getArg(0))); |
| 10486 | return EmitNeonCall(CGM.getIntrinsic(Intrinsic::fabs, HalfTy), Ops, "vabs"); |
| 10487 | case NEON::BI__builtin_neon_vaddq_p128: { |
| 10488 | llvm::Type *Ty = GetNeonType(this, NeonTypeFlags::Poly128); |
| 10489 | Ops.push_back(EmitScalarExpr(E->getArg(1))); |
| 10490 | Ops[0] = Builder.CreateBitCast(Ops[0], Ty); |
| 10491 | Ops[1] = Builder.CreateBitCast(Ops[1], Ty); |
| 10492 | Ops[0] = Builder.CreateXor(Ops[0], Ops[1]); |
| 10493 | llvm::Type *Int128Ty = llvm::Type::getIntNTy(getLLVMContext(), 128); |
| 10494 | return Builder.CreateBitCast(Ops[0], Int128Ty); |
| 10495 | } |
| 10496 | case NEON::BI__builtin_neon_vldrq_p128: { |
| 10497 | llvm::Type *Int128Ty = llvm::Type::getIntNTy(getLLVMContext(), 128); |
| 10498 | llvm::Type *Int128PTy = llvm::PointerType::get(Int128Ty, 0); |
| 10499 | Value *Ptr = Builder.CreateBitCast(EmitScalarExpr(E->getArg(0)), Int128PTy); |
| 10500 | return Builder.CreateAlignedLoad(Int128Ty, Ptr, |
| 10501 | CharUnits::fromQuantity(16)); |
| 10502 | } |
| 10503 | case NEON::BI__builtin_neon_vstrq_p128: { |
| 10504 | llvm::Type *Int128PTy = llvm::Type::getIntNPtrTy(getLLVMContext(), 128); |
| 10505 | Value *Ptr = Builder.CreateBitCast(Ops[0], Int128PTy); |
| 10506 | return Builder.CreateDefaultAlignedStore(EmitScalarExpr(E->getArg(1)), Ptr); |
| 10507 | } |
| 10508 | case NEON::BI__builtin_neon_vcvts_f32_u32: |
| 10509 | case NEON::BI__builtin_neon_vcvtd_f64_u64: |
| 10510 | usgn = true; |
| 10511 | [[fallthrough]]; |
| 10512 | case NEON::BI__builtin_neon_vcvts_f32_s32: |
| 10513 | case NEON::BI__builtin_neon_vcvtd_f64_s64: { |
| 10514 | Ops.push_back(EmitScalarExpr(E->getArg(0))); |
| 10515 | bool Is64 = Ops[0]->getType()->getPrimitiveSizeInBits() == 64; |
| 10516 | llvm::Type *InTy = Is64 ? Int64Ty : Int32Ty; |
| 10517 | llvm::Type *FTy = Is64 ? DoubleTy : FloatTy; |
| 10518 | Ops[0] = Builder.CreateBitCast(Ops[0], InTy); |
| 10519 | if (usgn) |
| 10520 | return Builder.CreateUIToFP(Ops[0], FTy); |
| 10521 | return Builder.CreateSIToFP(Ops[0], FTy); |
| 10522 | } |
| 10523 | case NEON::BI__builtin_neon_vcvth_f16_u16: |
| 10524 | case NEON::BI__builtin_neon_vcvth_f16_u32: |
| 10525 | case NEON::BI__builtin_neon_vcvth_f16_u64: |
| 10526 | usgn = true; |
| 10527 | [[fallthrough]]; |
| 10528 | case NEON::BI__builtin_neon_vcvth_f16_s16: |
| 10529 | case NEON::BI__builtin_neon_vcvth_f16_s32: |
| 10530 | case NEON::BI__builtin_neon_vcvth_f16_s64: { |
| 10531 | Ops.push_back(EmitScalarExpr(E->getArg(0))); |
| 10532 | llvm::Type *FTy = HalfTy; |
| 10533 | llvm::Type *InTy; |
| 10534 | if (Ops[0]->getType()->getPrimitiveSizeInBits() == 64) |
| 10535 | InTy = Int64Ty; |
| 10536 | else if (Ops[0]->getType()->getPrimitiveSizeInBits() == 32) |
| 10537 | InTy = Int32Ty; |
| 10538 | else |
| 10539 | InTy = Int16Ty; |
| 10540 | Ops[0] = Builder.CreateBitCast(Ops[0], InTy); |
| 10541 | if (usgn) |
| 10542 | return Builder.CreateUIToFP(Ops[0], FTy); |
| 10543 | return Builder.CreateSIToFP(Ops[0], FTy); |
| 10544 | } |
| 10545 | case NEON::BI__builtin_neon_vcvtah_u16_f16: |
| 10546 | case NEON::BI__builtin_neon_vcvtmh_u16_f16: |
| 10547 | case NEON::BI__builtin_neon_vcvtnh_u16_f16: |
| 10548 | case NEON::BI__builtin_neon_vcvtph_u16_f16: |
| 10549 | case NEON::BI__builtin_neon_vcvth_u16_f16: |
| 10550 | case NEON::BI__builtin_neon_vcvtah_s16_f16: |
| 10551 | case NEON::BI__builtin_neon_vcvtmh_s16_f16: |
| 10552 | case NEON::BI__builtin_neon_vcvtnh_s16_f16: |
| 10553 | case NEON::BI__builtin_neon_vcvtph_s16_f16: |
| 10554 | case NEON::BI__builtin_neon_vcvth_s16_f16: { |
| 10555 | unsigned Int; |
| 10556 | llvm::Type* InTy = Int32Ty; |
| 10557 | llvm::Type* FTy = HalfTy; |
| 10558 | llvm::Type *Tys[2] = {InTy, FTy}; |
| 10559 | Ops.push_back(EmitScalarExpr(E->getArg(0))); |
| 10560 | switch (BuiltinID) { |
| 10561 | default: llvm_unreachable("missing builtin ID in switch!")::llvm::llvm_unreachable_internal("missing builtin ID in switch!" , "clang/lib/CodeGen/CGBuiltin.cpp", 10561); |
| 10562 | case NEON::BI__builtin_neon_vcvtah_u16_f16: |
| 10563 | Int = Intrinsic::aarch64_neon_fcvtau; break; |
| 10564 | case NEON::BI__builtin_neon_vcvtmh_u16_f16: |
| 10565 | Int = Intrinsic::aarch64_neon_fcvtmu; break; |
| 10566 | case NEON::BI__builtin_neon_vcvtnh_u16_f16: |
| 10567 | Int = Intrinsic::aarch64_neon_fcvtnu; break; |
| 10568 | case NEON::BI__builtin_neon_vcvtph_u16_f16: |
| 10569 | Int = Intrinsic::aarch64_neon_fcvtpu; break; |
| 10570 | case NEON::BI__builtin_neon_vcvth_u16_f16: |
| 10571 | Int = Intrinsic::aarch64_neon_fcvtzu; break; |
| 10572 | case NEON::BI__builtin_neon_vcvtah_s16_f16: |
| 10573 | Int = Intrinsic::aarch64_neon_fcvtas; break; |
| 10574 | case NEON::BI__builtin_neon_vcvtmh_s16_f16: |
| 10575 | Int = Intrinsic::aarch64_neon_fcvtms; break; |
| 10576 | case NEON::BI__builtin_neon_vcvtnh_s16_f16: |
| 10577 | Int = Intrinsic::aarch64_neon_fcvtns; break; |
| 10578 | case NEON::BI__builtin_neon_vcvtph_s16_f16: |
| 10579 | Int = Intrinsic::aarch64_neon_fcvtps; break; |
| 10580 | case NEON::BI__builtin_neon_vcvth_s16_f16: |
| 10581 | Int = Intrinsic::aarch64_neon_fcvtzs; break; |
| 10582 | } |
| 10583 | Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "fcvt"); |
| 10584 | return Builder.CreateTrunc(Ops[0], Int16Ty); |
| 10585 | } |
| 10586 | case NEON::BI__builtin_neon_vcaleh_f16: |
| 10587 | case NEON::BI__builtin_neon_vcalth_f16: |
| 10588 | case NEON::BI__builtin_neon_vcageh_f16: |
| 10589 | case NEON::BI__builtin_neon_vcagth_f16: { |
| 10590 | unsigned Int; |
| 10591 | llvm::Type* InTy = Int32Ty; |
| 10592 | llvm::Type* FTy = HalfTy; |
| 10593 | llvm::Type *Tys[2] = {InTy, FTy}; |
| 10594 | Ops.push_back(EmitScalarExpr(E->getArg(1))); |
| 10595 | switch (BuiltinID) { |
| 10596 | default: llvm_unreachable("missing builtin ID in switch!")::llvm::llvm_unreachable_internal("missing builtin ID in switch!" , "clang/lib/CodeGen/CGBuiltin.cpp", 10596); |
| 10597 | case NEON::BI__builtin_neon_vcageh_f16: |
| 10598 | Int = Intrinsic::aarch64_neon_facge; break; |
| 10599 | case NEON::BI__builtin_neon_vcagth_f16: |
| 10600 | Int = Intrinsic::aarch64_neon_facgt; break; |
| 10601 | case NEON::BI__builtin_neon_vcaleh_f16: |
| 10602 | Int = Intrinsic::aarch64_neon_facge; std::swap(Ops[0], Ops[1]); break; |
| 10603 | case NEON::BI__builtin_neon_vcalth_f16: |
| 10604 | Int = Intrinsic::aarch64_neon_facgt; std::swap(Ops[0], Ops[1]); break; |
| 10605 | } |
| 10606 | Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "facg"); |
| 10607 | return Builder.CreateTrunc(Ops[0], Int16Ty); |
| 10608 | } |
| 10609 | case NEON::BI__builtin_neon_vcvth_n_s16_f16: |
| 10610 | case NEON::BI__builtin_neon_vcvth_n_u16_f16: { |
| 10611 | unsigned Int; |
| 10612 | llvm::Type* InTy = Int32Ty; |
| 10613 | llvm::Type* FTy = HalfTy; |
| 10614 | llvm::Type *Tys[2] = {InTy, FTy}; |
| 10615 | Ops.push_back(EmitScalarExpr(E->getArg(1))); |
| 10616 | switch (BuiltinID) { |
| 10617 | default: llvm_unreachable("missing builtin ID in switch!")::llvm::llvm_unreachable_internal("missing builtin ID in switch!" , "clang/lib/CodeGen/CGBuiltin.cpp", 10617); |
| 10618 | case NEON::BI__builtin_neon_vcvth_n_s16_f16: |
| 10619 | Int = Intrinsic::aarch64_neon_vcvtfp2fxs; break; |
| 10620 | case NEON::BI__builtin_neon_vcvth_n_u16_f16: |
| 10621 | Int = Intrinsic::aarch64_neon_vcvtfp2fxu; break; |
| 10622 | } |
| 10623 | Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "fcvth_n"); |
| 10624 | return Builder.CreateTrunc(Ops[0], Int16Ty); |
| 10625 | } |
| 10626 | case NEON::BI__builtin_neon_vcvth_n_f16_s16: |
| 10627 | case NEON::BI__builtin_neon_vcvth_n_f16_u16: { |
| 10628 | unsigned Int; |
| 10629 | llvm::Type* FTy = HalfTy; |
| 10630 | llvm::Type* InTy = Int32Ty; |
| 10631 | llvm::Type *Tys[2] = {FTy, InTy}; |
| 10632 | Ops.push_back(EmitScalarExpr(E->getArg(1))); |
| 10633 | switch (BuiltinID) { |
| 10634 | default: llvm_unreachable("missing builtin ID in switch!")::llvm::llvm_unreachable_internal("missing builtin ID in switch!" , "clang/lib/CodeGen/CGBuiltin.cpp", 10634); |
| 10635 | case NEON::BI__builtin_neon_vcvth_n_f16_s16: |
| 10636 | Int = Intrinsic::aarch64_neon_vcvtfxs2fp; |
| 10637 | Ops[0] = Builder.CreateSExt(Ops[0], InTy, "sext"); |
| 10638 | break; |
| 10639 | case NEON::BI__builtin_neon_vcvth_n_f16_u16: |
| 10640 | Int = Intrinsic::aarch64_neon_vcvtfxu2fp; |
| 10641 | Ops[0] = Builder.CreateZExt(Ops[0], InTy); |
| 10642 | break; |
| 10643 | } |
| 10644 | return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "fcvth_n"); |
| 10645 | } |
| 10646 | case NEON::BI__builtin_neon_vpaddd_s64: { |
| 10647 | auto *Ty = llvm::FixedVectorType::get(Int64Ty, 2); |
| 10648 | Value *Vec = EmitScalarExpr(E->getArg(0)); |
| 10649 | // The vector is v2f64, so make sure it's bitcast to that. |
| 10650 | Vec = Builder.CreateBitCast(Vec, Ty, "v2i64"); |
| 10651 | llvm::Value *Idx0 = llvm::ConstantInt::get(SizeTy, 0); |
| 10652 | llvm::Value *Idx1 = llvm::ConstantInt::get(SizeTy, 1); |
| 10653 | Value *Op0 = Builder.CreateExtractElement(Vec, Idx0, "lane0"); |
| 10654 | Value *Op1 = Builder.CreateExtractElement(Vec, Idx1, "lane1"); |
| 10655 | // Pairwise addition of a v2f64 into a scalar f64. |
| 10656 | return Builder.CreateAdd(Op0, Op1, "vpaddd"); |
| 10657 | } |
| 10658 | case NEON::BI__builtin_neon_vpaddd_f64: { |
| 10659 | auto *Ty = llvm::FixedVectorType::get(DoubleTy, 2); |
| 10660 | Value *Vec = EmitScalarExpr(E->getArg(0)); |
| 10661 | // The vector is v2f64, so make sure it's bitcast to that. |
| 10662 | Vec = Builder.CreateBitCast(Vec, Ty, "v2f64"); |
| 10663 | llvm::Value *Idx0 = llvm::ConstantInt::get(SizeTy, 0); |
| 10664 | llvm::Value *Idx1 = llvm::ConstantInt::get(SizeTy, 1); |
| 10665 | Value *Op0 = Builder.CreateExtractElement(Vec, Idx0, "lane0"); |
| 10666 | Value *Op1 = Builder.CreateExtractElement(Vec, Idx1, "lane1"); |
| 10667 | // Pairwise addition of a v2f64 into a scalar f64. |
| 10668 | return Builder.CreateFAdd(Op0, Op1, "vpaddd"); |
| 10669 | } |
| 10670 | case NEON::BI__builtin_neon_vpadds_f32: { |
| 10671 | auto *Ty = llvm::FixedVectorType::get(FloatTy, 2); |
| 10672 | Value *Vec = EmitScalarExpr(E->getArg(0)); |
| 10673 | // The vector is v2f32, so make sure it's bitcast to that. |
| 10674 | Vec = Builder.CreateBitCast(Vec, Ty, "v2f32"); |
| 10675 | llvm::Value *Idx0 = llvm::ConstantInt::get(SizeTy, 0); |
| 10676 | llvm::Value *Idx1 = llvm::ConstantInt::get(SizeTy, 1); |
| 10677 | Value *Op0 = Builder.CreateExtractElement(Vec, Idx0, "lane0"); |
| 10678 | Value *Op1 = Builder.CreateExtractElement(Vec, Idx1, "lane1"); |
| 10679 | // Pairwise addition of a v2f32 into a scalar f32. |
| 10680 | return Builder.CreateFAdd(Op0, Op1, "vpaddd"); |
| 10681 | } |
| 10682 | case NEON::BI__builtin_neon_vceqzd_s64: |
| 10683 | case NEON::BI__builtin_neon_vceqzd_f64: |
| 10684 | case NEON::BI__builtin_neon_vceqzs_f32: |
| 10685 | case NEON::BI__builtin_neon_vceqzh_f16: |
| 10686 | Ops.push_back(EmitScalarExpr(E->getArg(0))); |
| 10687 | return EmitAArch64CompareBuiltinExpr( |
| 10688 | Ops[0], ConvertType(E->getCallReturnType(getContext())), |
| 10689 | ICmpInst::FCMP_OEQ, ICmpInst::ICMP_EQ, "vceqz"); |
| 10690 | case NEON::BI__builtin_neon_vcgezd_s64: |
| 10691 | case NEON::BI__builtin_neon_vcgezd_f64: |
| 10692 | case NEON::BI__builtin_neon_vcgezs_f32: |
| 10693 | case NEON::BI__builtin_neon_vcgezh_f16: |
| 10694 | Ops.push_back(EmitScalarExpr(E->getArg(0))); |
| 10695 | return EmitAArch64CompareBuiltinExpr( |
| 10696 | Ops[0], ConvertType(E->getCallReturnType(getContext())), |
| 10697 | ICmpInst::FCMP_OGE, ICmpInst::ICMP_SGE, "vcgez"); |
| 10698 | case NEON::BI__builtin_neon_vclezd_s64: |
| 10699 | case NEON::BI__builtin_neon_vclezd_f64: |
| 10700 | case NEON::BI__builtin_neon_vclezs_f32: |
| 10701 | case NEON::BI__builtin_neon_vclezh_f16: |
| 10702 | Ops.push_back(EmitScalarExpr(E->getArg(0))); |
| 10703 | return EmitAArch64CompareBuiltinExpr( |
| 10704 | Ops[0], ConvertType(E->getCallReturnType(getContext())), |
| 10705 | ICmpInst::FCMP_OLE, ICmpInst::ICMP_SLE, "vclez"); |
| 10706 | case NEON::BI__builtin_neon_vcgtzd_s64: |
| 10707 | case NEON::BI__builtin_neon_vcgtzd_f64: |
| 10708 | case NEON::BI__builtin_neon_vcgtzs_f32: |
| 10709 | case NEON::BI__builtin_neon_vcgtzh_f16: |
| 10710 | Ops.push_back(EmitScalarExpr(E->getArg(0))); |
| 10711 | return EmitAArch64CompareBuiltinExpr( |
| 10712 | Ops[0], ConvertType(E->getCallReturnType(getContext())), |
| 10713 | ICmpInst::FCMP_OGT, ICmpInst::ICMP_SGT, "vcgtz"); |
| 10714 | case NEON::BI__builtin_neon_vcltzd_s64: |
| 10715 | case NEON::BI__builtin_neon_vcltzd_f64: |
| 10716 | case NEON::BI__builtin_neon_vcltzs_f32: |
| 10717 | case NEON::BI__builtin_neon_vcltzh_f16: |
| 10718 | Ops.push_back(EmitScalarExpr(E->getArg(0))); |
| 10719 | return EmitAArch64CompareBuiltinExpr( |
| 10720 | Ops[0], ConvertType(E->getCallReturnType(getContext())), |
| 10721 | ICmpInst::FCMP_OLT, ICmpInst::ICMP_SLT, "vcltz"); |
| 10722 | |
| 10723 | case NEON::BI__builtin_neon_vceqzd_u64: { |
| 10724 | Ops.push_back(EmitScalarExpr(E->getArg(0))); |
| 10725 | Ops[0] = Builder.CreateBitCast(Ops[0], Int64Ty); |
| 10726 | Ops[0] = |
| 10727 | Builder.CreateICmpEQ(Ops[0], llvm::Constant::getNullValue(Int64Ty)); |
| 10728 | return Builder.CreateSExt(Ops[0], Int64Ty, "vceqzd"); |
| 10729 | } |
| 10730 | case NEON::BI__builtin_neon_vceqd_f64: |
| 10731 | case NEON::BI__builtin_neon_vcled_f64: |
| 10732 | case NEON::BI__builtin_neon_vcltd_f64: |
| 10733 | case NEON::BI__builtin_neon_vcged_f64: |
| 10734 | case NEON::BI__builtin_neon_vcgtd_f64: { |
| 10735 | llvm::CmpInst::Predicate P; |
| 10736 | switch (BuiltinID) { |
| 10737 | default: llvm_unreachable("missing builtin ID in switch!")::llvm::llvm_unreachable_internal("missing builtin ID in switch!" , "clang/lib/CodeGen/CGBuiltin.cpp", 10737); |
| 10738 | case NEON::BI__builtin_neon_vceqd_f64: P = llvm::FCmpInst::FCMP_OEQ; break; |
| 10739 | case NEON::BI__builtin_neon_vcled_f64: P = llvm::FCmpInst::FCMP_OLE; break; |
| 10740 | case NEON::BI__builtin_neon_vcltd_f64: P = llvm::FCmpInst::FCMP_OLT; break; |
| 10741 | case NEON::BI__builtin_neon_vcged_f64: P = llvm::FCmpInst::FCMP_OGE; break; |
| 10742 | case NEON::BI__builtin_neon_vcgtd_f64: P = llvm::FCmpInst::FCMP_OGT; break; |
| 10743 | } |
| 10744 | Ops.push_back(EmitScalarExpr(E->getArg(1))); |
| 10745 | Ops[0] = Builder.CreateBitCast(Ops[0], DoubleTy); |
| 10746 | Ops[1] = Builder.CreateBitCast(Ops[1], DoubleTy); |
| 10747 | if (P == llvm::FCmpInst::FCMP_OEQ) |
| 10748 | Ops[0] = Builder.CreateFCmp(P, Ops[0], Ops[1]); |
| 10749 | else |
| 10750 | Ops[0] = Builder.CreateFCmpS(P, Ops[0], Ops[1]); |
| 10751 | return Builder.CreateSExt(Ops[0], Int64Ty, "vcmpd"); |
| 10752 | } |
| 10753 | case NEON::BI__builtin_neon_vceqs_f32: |
| 10754 | case NEON::BI__builtin_neon_vcles_f32: |
| 10755 | case NEON::BI__builtin_neon_vclts_f32: |
| 10756 | case NEON::BI__builtin_neon_vcges_f32: |
| 10757 | case NEON::BI__builtin_neon_vcgts_f32: { |
| 10758 | llvm::CmpInst::Predicate P; |
| 10759 | switch (BuiltinID) { |
| 10760 | default: llvm_unreachable("missing builtin ID in switch!")::llvm::llvm_unreachable_internal("missing builtin ID in switch!" , "clang/lib/CodeGen/CGBuiltin.cpp", 10760); |
| 10761 | case NEON::BI__builtin_neon_vceqs_f32: P = llvm::FCmpInst::FCMP_OEQ; break; |
| 10762 | case NEON::BI__builtin_neon_vcles_f32: P = llvm::FCmpInst::FCMP_OLE; break; |
| 10763 | case NEON::BI__builtin_neon_vclts_f32: P = llvm::FCmpInst::FCMP_OLT; break; |
| 10764 | case NEON::BI__builtin_neon_vcges_f32: P = llvm::FCmpInst::FCMP_OGE; break; |
| 10765 | case NEON::BI__builtin_neon_vcgts_f32: P = llvm::FCmpInst::FCMP_OGT; break; |
| 10766 | } |
| 10767 | Ops.push_back(EmitScalarExpr(E->getArg(1))); |
| 10768 | Ops[0] = Builder.CreateBitCast(Ops[0], FloatTy); |
| 10769 | Ops[1] = Builder.CreateBitCast(Ops[1], FloatTy); |
| 10770 | if (P == llvm::FCmpInst::FCMP_OEQ) |
| 10771 | Ops[0] = Builder.CreateFCmp(P, Ops[0], Ops[1]); |
| 10772 | else |
| 10773 | Ops[0] = Builder.CreateFCmpS(P, Ops[0], Ops[1]); |
| 10774 | return Builder.CreateSExt(Ops[0], Int32Ty, "vcmpd"); |
| 10775 | } |
| 10776 | case NEON::BI__builtin_neon_vceqh_f16: |
| 10777 | case NEON::BI__builtin_neon_vcleh_f16: |
| 10778 | case NEON::BI__builtin_neon_vclth_f16: |
| 10779 | case NEON::BI__builtin_neon_vcgeh_f16: |
| 10780 | case NEON::BI__builtin_neon_vcgth_f16: { |
| 10781 | llvm::CmpInst::Predicate P; |
| 10782 | switch (BuiltinID) { |
| 10783 | default: llvm_unreachable("missing builtin ID in switch!")::llvm::llvm_unreachable_internal("missing builtin ID in switch!" , "clang/lib/CodeGen/CGBuiltin.cpp", 10783); |
| 10784 | case NEON::BI__builtin_neon_vceqh_f16: P = llvm::FCmpInst::FCMP_OEQ; break; |
| 10785 | case NEON::BI__builtin_neon_vcleh_f16: P = llvm::FCmpInst::FCMP_OLE; break; |
| 10786 | case NEON::BI__builtin_neon_vclth_f16: P = llvm::FCmpInst::FCMP_OLT; break; |
| 10787 | case NEON::BI__builtin_neon_vcgeh_f16: P = llvm::FCmpInst::FCMP_OGE; break; |
| 10788 | case NEON::BI__builtin_neon_vcgth_f16: P = llvm::FCmpInst::FCMP_OGT; break; |
| 10789 | } |
| 10790 | Ops.push_back(EmitScalarExpr(E->getArg(1))); |
| 10791 | Ops[0] = Builder.CreateBitCast(Ops[0], HalfTy); |
| 10792 | Ops[1] = Builder.CreateBitCast(Ops[1], HalfTy); |
| 10793 | if (P == llvm::FCmpInst::FCMP_OEQ) |
| 10794 | Ops[0] = Builder.CreateFCmp(P, Ops[0], Ops[1]); |
| 10795 | else |
| 10796 | Ops[0] = Builder.CreateFCmpS(P, Ops[0], Ops[1]); |
| 10797 | return Builder.CreateSExt(Ops[0], Int16Ty, "vcmpd"); |
| 10798 | } |
| 10799 | case NEON::BI__builtin_neon_vceqd_s64: |
| 10800 | case NEON::BI__builtin_neon_vceqd_u64: |
| 10801 | case NEON::BI__builtin_neon_vcgtd_s64: |
| 10802 | case NEON::BI__builtin_neon_vcgtd_u64: |
| 10803 | case NEON::BI__builtin_neon_vcltd_s64: |
| 10804 | case NEON::BI__builtin_neon_vcltd_u64: |
| 10805 | case NEON::BI__builtin_neon_vcged_u64: |
| 10806 | case NEON::BI__builtin_neon_vcged_s64: |
| 10807 | case NEON::BI__builtin_neon_vcled_u64: |
| 10808 | case NEON::BI__builtin_neon_vcled_s64: { |
| 10809 | llvm::CmpInst::Predicate P; |
| 10810 | switch (BuiltinID) { |
| 10811 | default: llvm_unreachable("missing builtin ID in switch!")::llvm::llvm_unreachable_internal("missing builtin ID in switch!" , "clang/lib/CodeGen/CGBuiltin.cpp", 10811); |
| 10812 | case NEON::BI__builtin_neon_vceqd_s64: |
| 10813 | case NEON::BI__builtin_neon_vceqd_u64:P = llvm::ICmpInst::ICMP_EQ;break; |
| 10814 | case NEON::BI__builtin_neon_vcgtd_s64:P = llvm::ICmpInst::ICMP_SGT;break; |
| 10815 | case NEON::BI__builtin_neon_vcgtd_u64:P = llvm::ICmpInst::ICMP_UGT;break; |
| 10816 | case NEON::BI__builtin_neon_vcltd_s64:P = llvm::ICmpInst::ICMP_SLT;break; |
| 10817 | case NEON::BI__builtin_neon_vcltd_u64:P = llvm::ICmpInst::ICMP_ULT;break; |
| 10818 | case NEON::BI__builtin_neon_vcged_u64:P = llvm::ICmpInst::ICMP_UGE;break; |
| 10819 | case NEON::BI__builtin_neon_vcged_s64:P = llvm::ICmpInst::ICMP_SGE;break; |
| 10820 | case NEON::BI__builtin_neon_vcled_u64:P = llvm::ICmpInst::ICMP_ULE;break; |
| 10821 | case NEON::BI__builtin_neon_vcled_s64:P = llvm::ICmpInst::ICMP_SLE;break; |
| 10822 | } |
| 10823 | Ops.push_back(EmitScalarExpr(E->getArg(1))); |
| 10824 | Ops[0] = Builder.CreateBitCast(Ops[0], Int64Ty); |
| 10825 | Ops[1] = Builder.CreateBitCast(Ops[1], Int64Ty); |
| 10826 | Ops[0] = Builder.CreateICmp(P, Ops[0], Ops[1]); |
| 10827 | return Builder.CreateSExt(Ops[0], Int64Ty, "vceqd"); |
| 10828 | } |
| 10829 | case NEON::BI__builtin_neon_vtstd_s64: |
| 10830 | case NEON::BI__builtin_neon_vtstd_u64: { |
| 10831 | Ops.push_back(EmitScalarExpr(E->getArg(1))); |
| 10832 | Ops[0] = Builder.CreateBitCast(Ops[0], Int64Ty); |
| 10833 | Ops[1] = Builder.CreateBitCast(Ops[1], Int64Ty); |
| 10834 | Ops[0] = Builder.CreateAnd(Ops[0], Ops[1]); |
| 10835 | Ops[0] = Builder.CreateICmp(ICmpInst::ICMP_NE, Ops[0], |
| 10836 | llvm::Constant::getNullValue(Int64Ty)); |
| 10837 | return Builder.CreateSExt(Ops[0], Int64Ty, "vtstd"); |
| 10838 | } |
| 10839 | case NEON::BI__builtin_neon_vset_lane_i8: |
| 10840 | case NEON::BI__builtin_neon_vset_lane_i16: |
| 10841 | case NEON::BI__builtin_neon_vset_lane_i32: |
| 10842 | case NEON::BI__builtin_neon_vset_lane_i64: |
| 10843 | case NEON::BI__builtin_neon_vset_lane_bf16: |
| 10844 | case NEON::BI__builtin_neon_vset_lane_f32: |
| 10845 | case NEON::BI__builtin_neon_vsetq_lane_i8: |
| 10846 | case NEON::BI__builtin_neon_vsetq_lane_i16: |
| 10847 | case NEON::BI__builtin_neon_vsetq_lane_i32: |
| 10848 | case NEON::BI__builtin_neon_vsetq_lane_i64: |
| 10849 | case NEON::BI__builtin_neon_vsetq_lane_bf16: |
| 10850 | case NEON::BI__builtin_neon_vsetq_lane_f32: |
| 10851 | Ops.push_back(EmitScalarExpr(E->getArg(2))); |
| 10852 | return Builder.CreateInsertElement(Ops[1], Ops[0], Ops[2], "vset_lane"); |
| 10853 | case NEON::BI__builtin_neon_vset_lane_f64: |
| 10854 | // The vector type needs a cast for the v1f64 variant. |
| 10855 | Ops[1] = |
| 10856 | Builder.CreateBitCast(Ops[1], llvm::FixedVectorType::get(DoubleTy, 1)); |
| 10857 | Ops.push_back(EmitScalarExpr(E->getArg(2))); |
| 10858 | return Builder.CreateInsertElement(Ops[1], Ops[0], Ops[2], "vset_lane"); |
| 10859 | case NEON::BI__builtin_neon_vsetq_lane_f64: |
| 10860 | // The vector type needs a cast for the v2f64 variant. |
| 10861 | Ops[1] = |
| 10862 | Builder.CreateBitCast(Ops[1], llvm::FixedVectorType::get(DoubleTy, 2)); |
| 10863 | Ops.push_back(EmitScalarExpr(E->getArg(2))); |
| 10864 | return Builder.CreateInsertElement(Ops[1], Ops[0], Ops[2], "vset_lane"); |
| 10865 | |
| 10866 | case NEON::BI__builtin_neon_vget_lane_i8: |
| 10867 | case NEON::BI__builtin_neon_vdupb_lane_i8: |
| 10868 | Ops[0] = |
| 10869 | Builder.CreateBitCast(Ops[0], llvm::FixedVectorType::get(Int8Ty, 8)); |
| 10870 | return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), |
| 10871 | "vget_lane"); |
| 10872 | case NEON::BI__builtin_neon_vgetq_lane_i8: |
| 10873 | case NEON::BI__builtin_neon_vdupb_laneq_i8: |
| 10874 | Ops[0] = |
| 10875 | Builder.CreateBitCast(Ops[0], llvm::FixedVectorType::get(Int8Ty, 16)); |
| 10876 | return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), |
| 10877 | "vgetq_lane"); |
| 10878 | case NEON::BI__builtin_neon_vget_lane_i16: |
| 10879 | case NEON::BI__builtin_neon_vduph_lane_i16: |
| 10880 | Ops[0] = |
| 10881 | Builder.CreateBitCast(Ops[0], llvm::FixedVectorType::get(Int16Ty, 4)); |
| 10882 | return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), |
| 10883 | "vget_lane"); |
| 10884 | case NEON::BI__builtin_neon_vgetq_lane_i16: |
| 10885 | case NEON::BI__builtin_neon_vduph_laneq_i16: |
| 10886 | Ops[0] = |
| 10887 | Builder.CreateBitCast(Ops[0], llvm::FixedVectorType::get(Int16Ty, 8)); |
| 10888 | return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), |
| 10889 | "vgetq_lane"); |
| 10890 | case NEON::BI__builtin_neon_vget_lane_i32: |
| 10891 | case NEON::BI__builtin_neon_vdups_lane_i32: |
| 10892 | Ops[0] = |
| 10893 | Builder.CreateBitCast(Ops[0], llvm::FixedVectorType::get(Int32Ty, 2)); |
| 10894 | return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), |
| 10895 | "vget_lane"); |
| 10896 | case NEON::BI__builtin_neon_vdups_lane_f32: |
| 10897 | Ops[0] = |
| 10898 | Builder.CreateBitCast(Ops[0], llvm::FixedVectorType::get(FloatTy, 2)); |
| 10899 | return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), |
| 10900 | "vdups_lane"); |
| 10901 | case NEON::BI__builtin_neon_vgetq_lane_i32: |
| 10902 | case NEON::BI__builtin_neon_vdups_laneq_i32: |
| 10903 | Ops[0] = |
| 10904 | Builder.CreateBitCast(Ops[0], llvm::FixedVectorType::get(Int32Ty, 4)); |
| 10905 | return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), |
| 10906 | "vgetq_lane"); |
| 10907 | case NEON::BI__builtin_neon_vget_lane_i64: |
| 10908 | case NEON::BI__builtin_neon_vdupd_lane_i64: |
| 10909 | Ops[0] = |
| 10910 | Builder.CreateBitCast(Ops[0], llvm::FixedVectorType::get(Int64Ty, 1)); |
| 10911 | return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), |
| 10912 | "vget_lane"); |
| 10913 | case NEON::BI__builtin_neon_vdupd_lane_f64: |
| 10914 | Ops[0] = |
| 10915 | Builder.CreateBitCast(Ops[0], llvm::FixedVectorType::get(DoubleTy, 1)); |
| 10916 | return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), |
| 10917 | "vdupd_lane"); |
| 10918 | case NEON::BI__builtin_neon_vgetq_lane_i64: |
| 10919 | case NEON::BI__builtin_neon_vdupd_laneq_i64: |
| 10920 | Ops[0] = |
| 10921 | Builder.CreateBitCast(Ops[0], llvm::FixedVectorType::get(Int64Ty, 2)); |
| 10922 | return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), |
| 10923 | "vgetq_lane"); |
| 10924 | case NEON::BI__builtin_neon_vget_lane_f32: |
| 10925 | Ops[0] = |
| 10926 | Builder.CreateBitCast(Ops[0], llvm::FixedVectorType::get(FloatTy, 2)); |
| 10927 | return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), |
| 10928 | "vget_lane"); |
| 10929 | case NEON::BI__builtin_neon_vget_lane_f64: |
| 10930 | Ops[0] = |
| 10931 | Builder.CreateBitCast(Ops[0], llvm::FixedVectorType::get(DoubleTy, 1)); |
| 10932 | return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), |
| 10933 | "vget_lane"); |
| 10934 | case NEON::BI__builtin_neon_vgetq_lane_f32: |
| 10935 | case NEON::BI__builtin_neon_vdups_laneq_f32: |
| 10936 | Ops[0] = |
| 10937 | Builder.CreateBitCast(Ops[0], llvm::FixedVectorType::get(FloatTy, 4)); |
| 10938 | return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), |
| 10939 | "vgetq_lane"); |
| 10940 | case NEON::BI__builtin_neon_vgetq_lane_f64: |
| 10941 | case NEON::BI__builtin_neon_vdupd_laneq_f64: |
| 10942 | Ops[0] = |
| 10943 | Builder.CreateBitCast(Ops[0], llvm::FixedVectorType::get(DoubleTy, 2)); |
| 10944 | return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), |
| 10945 | "vgetq_lane"); |
| 10946 | case NEON::BI__builtin_neon_vaddh_f16: |
| 10947 | Ops.push_back(EmitScalarExpr(E->getArg(1))); |
| 10948 | return Builder.CreateFAdd(Ops[0], Ops[1], "vaddh"); |
| 10949 | case NEON::BI__builtin_neon_vsubh_f16: |
| 10950 | Ops.push_back(EmitScalarExpr(E->getArg(1))); |
| 10951 | return Builder.CreateFSub(Ops[0], Ops[1], "vsubh"); |
| 10952 | case NEON::BI__builtin_neon_vmulh_f16: |
| 10953 | Ops.push_back(EmitScalarExpr(E->getArg(1))); |
| 10954 | return Builder.CreateFMul(Ops[0], Ops[1], "vmulh"); |
| 10955 | case NEON::BI__builtin_neon_vdivh_f16: |
| 10956 | Ops.push_back(EmitScalarExpr(E->getArg(1))); |
| 10957 | return Builder.CreateFDiv(Ops[0], Ops[1], "vdivh"); |
| 10958 | case NEON::BI__builtin_neon_vfmah_f16: |
| 10959 | // NEON intrinsic puts accumulator first, unlike the LLVM fma. |
| 10960 | return emitCallMaybeConstrainedFPBuiltin( |
| 10961 | *this, Intrinsic::fma, Intrinsic::experimental_constrained_fma, HalfTy, |
| 10962 | {EmitScalarExpr(E->getArg(1)), EmitScalarExpr(E->getArg(2)), Ops[0]}); |
| 10963 | case NEON::BI__builtin_neon_vfmsh_f16: { |
| 10964 | Value* Neg = Builder.CreateFNeg(EmitScalarExpr(E->getArg(1)), "vsubh"); |
| 10965 | |
| 10966 | // NEON intrinsic puts accumulator first, unlike the LLVM fma. |
| 10967 | return emitCallMaybeConstrainedFPBuiltin( |
| 10968 | *this, Intrinsic::fma, Intrinsic::experimental_constrained_fma, HalfTy, |
| 10969 | {Neg, EmitScalarExpr(E->getArg(2)), Ops[0]}); |
| 10970 | } |
| 10971 | case NEON::BI__builtin_neon_vaddd_s64: |
| 10972 | case NEON::BI__builtin_neon_vaddd_u64: |
| 10973 | return Builder.CreateAdd(Ops[0], EmitScalarExpr(E->getArg(1)), "vaddd"); |
| 10974 | case NEON::BI__builtin_neon_vsubd_s64: |
| 10975 | case NEON::BI__builtin_neon_vsubd_u64: |
| 10976 | return Builder.CreateSub(Ops[0], EmitScalarExpr(E->getArg(1)), "vsubd"); |
| 10977 | case NEON::BI__builtin_neon_vqdmlalh_s16: |
| 10978 | case NEON::BI__builtin_neon_vqdmlslh_s16: { |
| 10979 | SmallVector<Value *, 2> ProductOps; |
| 10980 | ProductOps.push_back(vectorWrapScalar16(Ops[1])); |
| 10981 | ProductOps.push_back(vectorWrapScalar16(EmitScalarExpr(E->getArg(2)))); |
| 10982 | auto *VTy = llvm::FixedVectorType::get(Int32Ty, 4); |
| 10983 | Ops[1] = EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_sqdmull, VTy), |
| 10984 | ProductOps, "vqdmlXl"); |
| 10985 | Constant *CI = ConstantInt::get(SizeTy, 0); |
| 10986 | Ops[1] = Builder.CreateExtractElement(Ops[1], CI, "lane0"); |
| 10987 | |
| 10988 | unsigned AccumInt = BuiltinID == NEON::BI__builtin_neon_vqdmlalh_s16 |
| 10989 | ? Intrinsic::aarch64_neon_sqadd |
| 10990 | : Intrinsic::aarch64_neon_sqsub; |
| 10991 | return EmitNeonCall(CGM.getIntrinsic(AccumInt, Int32Ty), Ops, "vqdmlXl"); |
| 10992 | } |
| 10993 | case NEON::BI__builtin_neon_vqshlud_n_s64: { |
| 10994 | Ops.push_back(EmitScalarExpr(E->getArg(1))); |
| 10995 | Ops[1] = Builder.CreateZExt(Ops[1], Int64Ty); |
| 10996 | return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_sqshlu, Int64Ty), |
| 10997 | Ops, "vqshlu_n"); |
| 10998 | } |
| 10999 | case NEON::BI__builtin_neon_vqshld_n_u64: |
| 11000 | case NEON::BI__builtin_neon_vqshld_n_s64: { |
| 11001 | unsigned Int = BuiltinID == NEON::BI__builtin_neon_vqshld_n_u64 |
| 11002 | ? Intrinsic::aarch64_neon_uqshl |
| 11003 | : Intrinsic::aarch64_neon_sqshl; |
| 11004 | Ops.push_back(EmitScalarExpr(E->getArg(1))); |
| 11005 | Ops[1] = Builder.CreateZExt(Ops[1], Int64Ty); |
| 11006 | return EmitNeonCall(CGM.getIntrinsic(Int, Int64Ty), Ops, "vqshl_n"); |
| 11007 | } |
| 11008 | case NEON::BI__builtin_neon_vrshrd_n_u64: |
| 11009 | case NEON::BI__builtin_neon_vrshrd_n_s64: { |
| 11010 | unsigned Int = BuiltinID == NEON::BI__builtin_neon_vrshrd_n_u64 |
| 11011 | ? Intrinsic::aarch64_neon_urshl |
| 11012 | : Intrinsic::aarch64_neon_srshl; |
| 11013 | Ops.push_back(EmitScalarExpr(E->getArg(1))); |
| 11014 | int SV = cast<ConstantInt>(Ops[1])->getSExtValue(); |
| 11015 | Ops[1] = ConstantInt::get(Int64Ty, -SV); |
| 11016 | return EmitNeonCall(CGM.getIntrinsic(Int, Int64Ty), Ops, "vrshr_n"); |
| 11017 | } |
| 11018 | case NEON::BI__builtin_neon_vrsrad_n_u64: |
| 11019 | case NEON::BI__builtin_neon_vrsrad_n_s64: { |
| 11020 | unsigned Int = BuiltinID == NEON::BI__builtin_neon_vrsrad_n_u64 |
| 11021 | ? Intrinsic::aarch64_neon_urshl |
| 11022 | : Intrinsic::aarch64_neon_srshl; |
| 11023 | Ops[1] = Builder.CreateBitCast(Ops[1], Int64Ty); |
| 11024 | Ops.push_back(Builder.CreateNeg(EmitScalarExpr(E->getArg(2)))); |
| 11025 | Ops[1] = Builder.CreateCall(CGM.getIntrinsic(Int, Int64Ty), |
| 11026 | {Ops[1], Builder.CreateSExt(Ops[2], Int64Ty)}); |
| 11027 | return Builder.CreateAdd(Ops[0], Builder.CreateBitCast(Ops[1], Int64Ty)); |
| 11028 | } |
| 11029 | case NEON::BI__builtin_neon_vshld_n_s64: |
| 11030 | case NEON::BI__builtin_neon_vshld_n_u64: { |
| 11031 | llvm::ConstantInt *Amt = cast<ConstantInt>(EmitScalarExpr(E->getArg(1))); |
| 11032 | return Builder.CreateShl( |
| 11033 | Ops[0], ConstantInt::get(Int64Ty, Amt->getZExtValue()), "shld_n"); |
| 11034 | } |
| 11035 | case NEON::BI__builtin_neon_vshrd_n_s64: { |
| 11036 | llvm::ConstantInt *Amt = cast<ConstantInt>(EmitScalarExpr(E->getArg(1))); |
| 11037 | return Builder.CreateAShr( |
| 11038 | Ops[0], ConstantInt::get(Int64Ty, std::min(static_cast<uint64_t>(63), |
| 11039 | Amt->getZExtValue())), |
| 11040 | "shrd_n"); |
| 11041 | } |
| 11042 | case NEON::BI__builtin_neon_vshrd_n_u64: { |
| 11043 | llvm::ConstantInt *Amt = cast<ConstantInt>(EmitScalarExpr(E->getArg(1))); |
| 11044 | uint64_t ShiftAmt = Amt->getZExtValue(); |
| 11045 | // Right-shifting an unsigned value by its size yields 0. |
| 11046 | if (ShiftAmt == 64) |
| 11047 | return ConstantInt::get(Int64Ty, 0); |
| 11048 | return Builder.CreateLShr(Ops[0], ConstantInt::get(Int64Ty, ShiftAmt), |
| 11049 | "shrd_n"); |
| 11050 | } |
| 11051 | case NEON::BI__builtin_neon_vsrad_n_s64: { |
| 11052 | llvm::ConstantInt *Amt = cast<ConstantInt>(EmitScalarExpr(E->getArg(2))); |
| 11053 | Ops[1] = Builder.CreateAShr( |
| 11054 | Ops[1], ConstantInt::get(Int64Ty, std::min(static_cast<uint64_t>(63), |
| 11055 | Amt->getZExtValue())), |
| 11056 | "shrd_n"); |
| 11057 | return Builder.CreateAdd(Ops[0], Ops[1]); |
| 11058 | } |
| 11059 | case NEON::BI__builtin_neon_vsrad_n_u64: { |
| 11060 | llvm::ConstantInt *Amt = cast<ConstantInt>(EmitScalarExpr(E->getArg(2))); |
| 11061 | uint64_t ShiftAmt = Amt->getZExtValue(); |
| 11062 | // Right-shifting an unsigned value by its size yields 0. |
| 11063 | // As Op + 0 = Op, return Ops[0] directly. |
| 11064 | if (ShiftAmt == 64) |
| 11065 | return Ops[0]; |
| 11066 | Ops[1] = Builder.CreateLShr(Ops[1], ConstantInt::get(Int64Ty, ShiftAmt), |
| 11067 | "shrd_n"); |
| 11068 | return Builder.CreateAdd(Ops[0], Ops[1]); |
| 11069 | } |
| 11070 | case NEON::BI__builtin_neon_vqdmlalh_lane_s16: |
| 11071 | case NEON::BI__builtin_neon_vqdmlalh_laneq_s16: |
| 11072 | case NEON::BI__builtin_neon_vqdmlslh_lane_s16: |
| 11073 | case NEON::BI__builtin_neon_vqdmlslh_laneq_s16: { |
| 11074 | Ops[2] = Builder.CreateExtractElement(Ops[2], EmitScalarExpr(E->getArg(3)), |
| 11075 | "lane"); |
| 11076 | SmallVector<Value *, 2> ProductOps; |
| 11077 | ProductOps.push_back(vectorWrapScalar16(Ops[1])); |
| 11078 | ProductOps.push_back(vectorWrapScalar16(Ops[2])); |
| 11079 | auto *VTy = llvm::FixedVectorType::get(Int32Ty, 4); |
| 11080 | Ops[1] = EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_sqdmull, VTy), |
| 11081 | ProductOps, "vqdmlXl"); |
| 11082 | Constant *CI = ConstantInt::get(SizeTy, 0); |
| 11083 | Ops[1] = Builder.CreateExtractElement(Ops[1], CI, "lane0"); |
| 11084 | Ops.pop_back(); |
| 11085 | |
| 11086 | unsigned AccInt = (BuiltinID == NEON::BI__builtin_neon_vqdmlalh_lane_s16 || |
| 11087 | BuiltinID == NEON::BI__builtin_neon_vqdmlalh_laneq_s16) |
| 11088 | ? Intrinsic::aarch64_neon_sqadd |
| 11089 | : Intrinsic::aarch64_neon_sqsub; |
| 11090 | return EmitNeonCall(CGM.getIntrinsic(AccInt, Int32Ty), Ops, "vqdmlXl"); |
| 11091 | } |
| 11092 | case NEON::BI__builtin_neon_vqdmlals_s32: |
| 11093 | case NEON::BI__builtin_neon_vqdmlsls_s32: { |
| 11094 | SmallVector<Value *, 2> ProductOps; |
| 11095 | ProductOps.push_back(Ops[1]); |
| 11096 | ProductOps.push_back(EmitScalarExpr(E->getArg(2))); |
| 11097 | Ops[1] = |
| 11098 | EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_sqdmulls_scalar), |
| 11099 | ProductOps, "vqdmlXl"); |
| 11100 | |
| 11101 | unsigned AccumInt = BuiltinID == NEON::BI__builtin_neon_vqdmlals_s32 |
| 11102 | ? Intrinsic::aarch64_neon_sqadd |
| 11103 | : Intrinsic::aarch64_neon_sqsub; |
| 11104 | return EmitNeonCall(CGM.getIntrinsic(AccumInt, Int64Ty), Ops, "vqdmlXl"); |
| 11105 | } |
| 11106 | case NEON::BI__builtin_neon_vqdmlals_lane_s32: |
| 11107 | case NEON::BI__builtin_neon_vqdmlals_laneq_s32: |
| 11108 | case NEON::BI__builtin_neon_vqdmlsls_lane_s32: |
| 11109 | case NEON::BI__builtin_neon_vqdmlsls_laneq_s32: { |
| 11110 | Ops[2] = Builder.CreateExtractElement(Ops[2], EmitScalarExpr(E->getArg(3)), |
| 11111 | "lane"); |
| 11112 | SmallVector<Value *, 2> ProductOps; |
| 11113 | ProductOps.push_back(Ops[1]); |
| 11114 | ProductOps.push_back(Ops[2]); |
| 11115 | Ops[1] = |
| 11116 | EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_sqdmulls_scalar), |
| 11117 | ProductOps, "vqdmlXl"); |
| 11118 | Ops.pop_back(); |
| 11119 | |
| 11120 | unsigned AccInt = (BuiltinID == NEON::BI__builtin_neon_vqdmlals_lane_s32 || |
| 11121 | BuiltinID == NEON::BI__builtin_neon_vqdmlals_laneq_s32) |
| 11122 | ? Intrinsic::aarch64_neon_sqadd |
| 11123 | : Intrinsic::aarch64_neon_sqsub; |
| 11124 | return EmitNeonCall(CGM.getIntrinsic(AccInt, Int64Ty), Ops, "vqdmlXl"); |
| 11125 | } |
| 11126 | case NEON::BI__builtin_neon_vget_lane_bf16: |
| 11127 | case NEON::BI__builtin_neon_vduph_lane_bf16: |
| 11128 | case NEON::BI__builtin_neon_vduph_lane_f16: { |
| 11129 | return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), |
| 11130 | "vget_lane"); |
| 11131 | } |
| 11132 | case NEON::BI__builtin_neon_vgetq_lane_bf16: |
| 11133 | case NEON::BI__builtin_neon_vduph_laneq_bf16: |
| 11134 | case NEON::BI__builtin_neon_vduph_laneq_f16: { |
| 11135 | return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), |
| 11136 | "vgetq_lane"); |
| 11137 | } |
| 11138 | |
| 11139 | case clang::AArch64::BI_InterlockedAdd: { |
| 11140 | Value *Arg0 = EmitScalarExpr(E->getArg(0)); |
| 11141 | Value *Arg1 = EmitScalarExpr(E->getArg(1)); |
| 11142 | AtomicRMWInst *RMWI = Builder.CreateAtomicRMW( |
| 11143 | AtomicRMWInst::Add, Arg0, Arg1, |
| 11144 | llvm::AtomicOrdering::SequentiallyConsistent); |
| 11145 | return Builder.CreateAdd(RMWI, Arg1); |
| 11146 | } |
| 11147 | } |
| 11148 | |
| 11149 | llvm::FixedVectorType *VTy = GetNeonType(this, Type); |
| 11150 | llvm::Type *Ty = VTy; |
| 11151 | if (!Ty) |
| 11152 | return nullptr; |
| 11153 | |
| 11154 | // Not all intrinsics handled by the common case work for AArch64 yet, so only |
| 11155 | // defer to common code if it's been added to our special map. |
| 11156 | Builtin = findARMVectorIntrinsicInMap(AArch64SIMDIntrinsicMap, BuiltinID, |
| 11157 | AArch64SIMDIntrinsicsProvenSorted); |
| 11158 | |
| 11159 | if (Builtin) |
| 11160 | return EmitCommonNeonBuiltinExpr( |
| 11161 | Builtin->BuiltinID, Builtin->LLVMIntrinsic, Builtin->AltLLVMIntrinsic, |
| 11162 | Builtin->NameHint, Builtin->TypeModifier, E, Ops, |
| 11163 | /*never use addresses*/ Address::invalid(), Address::invalid(), Arch); |
| 11164 | |
| 11165 | if (Value *V = EmitAArch64TblBuiltinExpr(*this, BuiltinID, E, Ops, Arch)) |
| 11166 | return V; |
| 11167 | |
| 11168 | unsigned Int; |
| 11169 | switch (BuiltinID) { |
| 11170 | default: return nullptr; |
| 11171 | case NEON::BI__builtin_neon_vbsl_v: |
| 11172 | case NEON::BI__builtin_neon_vbslq_v: { |
| 11173 | llvm::Type *BitTy = llvm::VectorType::getInteger(VTy); |
| 11174 | Ops[0] = Builder.CreateBitCast(Ops[0], BitTy, "vbsl"); |
| 11175 | Ops[1] = Builder.CreateBitCast(Ops[1], BitTy, "vbsl"); |
| 11176 | Ops[2] = Builder.CreateBitCast(Ops[2], BitTy, "vbsl"); |
| 11177 | |
| 11178 | Ops[1] = Builder.CreateAnd(Ops[0], Ops[1], "vbsl"); |
| 11179 | Ops[2] = Builder.CreateAnd(Builder.CreateNot(Ops[0]), Ops[2], "vbsl"); |
| 11180 | Ops[0] = Builder.CreateOr(Ops[1], Ops[2], "vbsl"); |
| 11181 | return Builder.CreateBitCast(Ops[0], Ty); |
| 11182 | } |
| 11183 | case NEON::BI__builtin_neon_vfma_lane_v: |
| 11184 | case NEON::BI__builtin_neon_vfmaq_lane_v: { // Only used for FP types |
| 11185 | // The ARM builtins (and instructions) have the addend as the first |
| 11186 | // operand, but the 'fma' intrinsics have it last. Swap it around here. |
| 11187 | Value *Addend = Ops[0]; |
| 11188 | Value *Multiplicand = Ops[1]; |
| 11189 | Value *LaneSource = Ops[2]; |
| 11190 | Ops[0] = Multiplicand; |
| 11191 | Ops[1] = LaneSource; |
| 11192 | Ops[2] = Addend; |
| 11193 | |
| 11194 | // Now adjust things to handle the lane access. |
| 11195 | auto *SourceTy = BuiltinID == NEON::BI__builtin_neon_vfmaq_lane_v |
| 11196 | ? llvm::FixedVectorType::get(VTy->getElementType(), |
| 11197 | VTy->getNumElements() / 2) |
| 11198 | : VTy; |
| 11199 | llvm::Constant *cst = cast<Constant>(Ops[3]); |
| 11200 | Value *SV = llvm::ConstantVector::getSplat(VTy->getElementCount(), cst); |
| 11201 | Ops[1] = Builder.CreateBitCast(Ops[1], SourceTy); |
| 11202 | Ops[1] = Builder.CreateShuffleVector(Ops[1], Ops[1], SV, "lane"); |
| 11203 | |
| 11204 | Ops.pop_back(); |
| 11205 | Int = Builder.getIsFPConstrained() ? Intrinsic::experimental_constrained_fma |
| 11206 | : Intrinsic::fma; |
| 11207 | return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "fmla"); |
| 11208 | } |
| 11209 | case NEON::BI__builtin_neon_vfma_laneq_v: { |
| 11210 | auto *VTy = cast<llvm::FixedVectorType>(Ty); |
| 11211 | // v1f64 fma should be mapped to Neon scalar f64 fma |
| 11212 | if (VTy && VTy->getElementType() == DoubleTy) { |
| 11213 | Ops[0] = Builder.CreateBitCast(Ops[0], DoubleTy); |
| 11214 | Ops[1] = Builder.CreateBitCast(Ops[1], DoubleTy); |
| 11215 | llvm::FixedVectorType *VTy = |
| 11216 | GetNeonType(this, NeonTypeFlags(NeonTypeFlags::Float64, false, true)); |
| 11217 | Ops[2] = Builder.CreateBitCast(Ops[2], VTy); |
| 11218 | Ops[2] = Builder.CreateExtractElement(Ops[2], Ops[3], "extract"); |
| 11219 | Value *Result; |
| 11220 | Result = emitCallMaybeConstrainedFPBuiltin( |
| 11221 | *this, Intrinsic::fma, Intrinsic::experimental_constrained_fma, |
| 11222 | DoubleTy, {Ops[1], Ops[2], Ops[0]}); |
| 11223 | return Builder.CreateBitCast(Result, Ty); |
| 11224 | } |
| 11225 | Ops[0] = Builder.CreateBitCast(Ops[0], Ty); |
| 11226 | Ops[1] = Builder.CreateBitCast(Ops[1], Ty); |
| 11227 | |
| 11228 | auto *STy = llvm::FixedVectorType::get(VTy->getElementType(), |
| 11229 | VTy->getNumElements() * 2); |
| 11230 | Ops[2] = Builder.CreateBitCast(Ops[2], STy); |
| 11231 | Value *SV = llvm::ConstantVector::getSplat(VTy->getElementCount(), |
| 11232 | cast<ConstantInt>(Ops[3])); |
| 11233 | Ops[2] = Builder.CreateShuffleVector(Ops[2], Ops[2], SV, "lane"); |
| 11234 | |
| 11235 | return emitCallMaybeConstrainedFPBuiltin( |
| 11236 | *this, Intrinsic::fma, Intrinsic::experimental_constrained_fma, Ty, |
| 11237 | {Ops[2], Ops[1], Ops[0]}); |
| 11238 | } |
| 11239 | case NEON::BI__builtin_neon_vfmaq_laneq_v: { |
| 11240 | Ops[0] = Builder.CreateBitCast(Ops[0], Ty); |
| 11241 | Ops[1] = Builder.CreateBitCast(Ops[1], Ty); |
| 11242 | |
| 11243 | Ops[2] = Builder.CreateBitCast(Ops[2], Ty); |
| 11244 | Ops[2] = EmitNeonSplat(Ops[2], cast<ConstantInt>(Ops[3])); |
| 11245 | return emitCallMaybeConstrainedFPBuiltin( |
| 11246 | *this, Intrinsic::fma, Intrinsic::experimental_constrained_fma, Ty, |
| 11247 | {Ops[2], Ops[1], Ops[0]}); |
| 11248 | } |
| 11249 | case NEON::BI__builtin_neon_vfmah_lane_f16: |
| 11250 | case NEON::BI__builtin_neon_vfmas_lane_f32: |
| 11251 | case NEON::BI__builtin_neon_vfmah_laneq_f16: |
| 11252 | case NEON::BI__builtin_neon_vfmas_laneq_f32: |
| 11253 | case NEON::BI__builtin_neon_vfmad_lane_f64: |
| 11254 | case NEON::BI__builtin_neon_vfmad_laneq_f64: { |
| 11255 | Ops.push_back(EmitScalarExpr(E->getArg(3))); |
| 11256 | llvm::Type *Ty = ConvertType(E->getCallReturnType(getContext())); |
| 11257 | Ops[2] = Builder.CreateExtractElement(Ops[2], Ops[3], "extract"); |
| 11258 | return emitCallMaybeConstrainedFPBuiltin( |
| 11259 | *this, Intrinsic::fma, Intrinsic::experimental_constrained_fma, Ty, |
| 11260 | {Ops[1], Ops[2], Ops[0]}); |
| 11261 | } |
| 11262 | case NEON::BI__builtin_neon_vmull_v: |
| 11263 | // FIXME: improve sharing scheme to cope with 3 alternative LLVM intrinsics. |
| 11264 | Int = usgn ? Intrinsic::aarch64_neon_umull : Intrinsic::aarch64_neon_smull; |
| 11265 | if (Type.isPoly()) Int = Intrinsic::aarch64_neon_pmull; |
| 11266 | return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmull"); |
| 11267 | case NEON::BI__builtin_neon_vmax_v: |
| 11268 | case NEON::BI__builtin_neon_vmaxq_v: |
| 11269 | // FIXME: improve sharing scheme to cope with 3 alternative LLVM intrinsics. |
| 11270 | Int = usgn ? Intrinsic::aarch64_neon_umax : Intrinsic::aarch64_neon_smax; |
| 11271 | if (Ty->isFPOrFPVectorTy()) Int = Intrinsic::aarch64_neon_fmax; |
| 11272 | return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmax"); |
| 11273 | case NEON::BI__builtin_neon_vmaxh_f16: { |
| 11274 | Ops.push_back(EmitScalarExpr(E->getArg(1))); |
| 11275 | Int = Intrinsic::aarch64_neon_fmax; |
| 11276 | return EmitNeonCall(CGM.getIntrinsic(Int, HalfTy), Ops, "vmax"); |
| 11277 | } |
| 11278 | case NEON::BI__builtin_neon_vmin_v: |
| 11279 | case NEON::BI__builtin_neon_vminq_v: |
| 11280 | // FIXME: improve sharing scheme to cope with 3 alternative LLVM intrinsics. |
| 11281 | Int = usgn ? Intrinsic::aarch64_neon_umin : Intrinsic::aarch64_neon_smin; |
| 11282 | if (Ty->isFPOrFPVectorTy()) Int = Intrinsic::aarch64_neon_fmin; |
| 11283 | return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmin"); |
| 11284 | case NEON::BI__builtin_neon_vminh_f16: { |
| 11285 | Ops.push_back(EmitScalarExpr(E->getArg(1))); |
| 11286 | Int = Intrinsic::aarch64_neon_fmin; |
| 11287 | return EmitNeonCall(CGM.getIntrinsic(Int, HalfTy), Ops, "vmin"); |
| 11288 | } |
| 11289 | case NEON::BI__builtin_neon_vabd_v: |
| 11290 | case NEON::BI__builtin_neon_vabdq_v: |
| 11291 | // FIXME: improve sharing scheme to cope with 3 alternative LLVM intrinsics. |
| 11292 | Int = usgn ? Intrinsic::aarch64_neon_uabd : Intrinsic::aarch64_neon_sabd; |
| 11293 | if (Ty->isFPOrFPVectorTy()) Int = Intrinsic::aarch64_neon_fabd; |
| 11294 | return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vabd"); |
| 11295 | case NEON::BI__builtin_neon_vpadal_v: |
| 11296 | case NEON::BI__builtin_neon_vpadalq_v: { |
| 11297 | unsigned ArgElts = VTy->getNumElements(); |
| 11298 | llvm::IntegerType *EltTy = cast<IntegerType>(VTy->getElementType()); |
| 11299 | unsigned BitWidth = EltTy->getBitWidth(); |
| 11300 | auto *ArgTy = llvm::FixedVectorType::get( |
| 11301 | llvm::IntegerType::get(getLLVMContext(), BitWidth / 2), 2 * ArgElts); |
| 11302 | llvm::Type* Tys[2] = { VTy, ArgTy }; |
| 11303 | Int = usgn ? Intrinsic::aarch64_neon_uaddlp : Intrinsic::aarch64_neon_saddlp; |
| 11304 | SmallVector<llvm::Value*, 1> TmpOps; |
| 11305 | TmpOps.push_back(Ops[1]); |
| 11306 | Function *F = CGM.getIntrinsic(Int, Tys); |
| 11307 | llvm::Value *tmp = EmitNeonCall(F, TmpOps, "vpadal"); |
| 11308 | llvm::Value *addend = Builder.CreateBitCast(Ops[0], tmp->getType()); |
| 11309 | return Builder.CreateAdd(tmp, addend); |
| 11310 | } |
| 11311 | case NEON::BI__builtin_neon_vpmin_v: |
| 11312 | case NEON::BI__builtin_neon_vpminq_v: |
| 11313 | // FIXME: improve sharing scheme to cope with 3 alternative LLVM intrinsics. |
| 11314 | Int = usgn ? Intrinsic::aarch64_neon_uminp : Intrinsic::aarch64_neon_sminp; |
| 11315 | if (Ty->isFPOrFPVectorTy()) Int = Intrinsic::aarch64_neon_fminp; |
| 11316 | return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vpmin"); |
| 11317 | case NEON::BI__builtin_neon_vpmax_v: |
| 11318 | case NEON::BI__builtin_neon_vpmaxq_v: |
| 11319 | // FIXME: improve sharing scheme to cope with 3 alternative LLVM intrinsics. |
| 11320 | Int = usgn ? Intrinsic::aarch64_neon_umaxp : Intrinsic::aarch64_neon_smaxp; |
| 11321 | if (Ty->isFPOrFPVectorTy()) Int = Intrinsic::aarch64_neon_fmaxp; |
| 11322 | return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vpmax"); |
| 11323 | case NEON::BI__builtin_neon_vminnm_v: |
| 11324 | case NEON::BI__builtin_neon_vminnmq_v: |
| 11325 | Int = Intrinsic::aarch64_neon_fminnm; |
| 11326 | return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vminnm"); |
| 11327 | case NEON::BI__builtin_neon_vminnmh_f16: |
| 11328 | Ops.push_back(EmitScalarExpr(E->getArg(1))); |
| 11329 | Int = Intrinsic::aarch64_neon_fminnm; |
| 11330 | return EmitNeonCall(CGM.getIntrinsic(Int, HalfTy), Ops, "vminnm"); |
| 11331 | case NEON::BI__builtin_neon_vmaxnm_v: |
| 11332 | case NEON::BI__builtin_neon_vmaxnmq_v: |
| 11333 | Int = Intrinsic::aarch64_neon_fmaxnm; |
| 11334 | return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmaxnm"); |
| 11335 | case NEON::BI__builtin_neon_vmaxnmh_f16: |
| 11336 | Ops.push_back(EmitScalarExpr(E->getArg(1))); |
| 11337 | Int = Intrinsic::aarch64_neon_fmaxnm; |
| 11338 | return EmitNeonCall(CGM.getIntrinsic(Int, HalfTy), Ops, "vmaxnm"); |
| 11339 | case NEON::BI__builtin_neon_vrecpss_f32: { |
| 11340 | Ops.push_back(EmitScalarExpr(E->getArg(1))); |
| 11341 | return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_frecps, FloatTy), |
| 11342 | Ops, "vrecps"); |
| 11343 | } |
| 11344 | case NEON::BI__builtin_neon_vrecpsd_f64: |
| 11345 | Ops.push_back(EmitScalarExpr(E->getArg(1))); |
| 11346 | return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_frecps, DoubleTy), |
| 11347 | Ops, "vrecps"); |
| 11348 | case NEON::BI__builtin_neon_vrecpsh_f16: |
| 11349 | Ops.push_back(EmitScalarExpr(E->getArg(1))); |
| 11350 | return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_frecps, HalfTy), |
| 11351 | Ops, "vrecps"); |
| 11352 | case NEON::BI__builtin_neon_vqshrun_n_v: |
| 11353 | Int = Intrinsic::aarch64_neon_sqshrun; |
| 11354 | return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqshrun_n"); |
| 11355 | case NEON::BI__builtin_neon_vqrshrun_n_v: |
| 11356 | Int = Intrinsic::aarch64_neon_sqrshrun; |
| 11357 | return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqrshrun_n"); |
| 11358 | case NEON::BI__builtin_neon_vqshrn_n_v: |
| 11359 | Int = usgn ? Intrinsic::aarch64_neon_uqshrn : Intrinsic::aarch64_neon_sqshrn; |
| 11360 | return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqshrn_n"); |
| 11361 | case NEON::BI__builtin_neon_vrshrn_n_v: |
| 11362 | Int = Intrinsic::aarch64_neon_rshrn; |
| 11363 | return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrshrn_n"); |
| 11364 | case NEON::BI__builtin_neon_vqrshrn_n_v: |
| 11365 | Int = usgn ? Intrinsic::aarch64_neon_uqrshrn : Intrinsic::aarch64_neon_sqrshrn; |
| 11366 | return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqrshrn_n"); |
| 11367 | case NEON::BI__builtin_neon_vrndah_f16: { |
| 11368 | Ops.push_back(EmitScalarExpr(E->getArg(0))); |
| 11369 | Int = Builder.getIsFPConstrained() |
| 11370 | ? Intrinsic::experimental_constrained_round |
| 11371 | : Intrinsic::round; |
| 11372 | return EmitNeonCall(CGM.getIntrinsic(Int, HalfTy), Ops, "vrnda"); |
| 11373 | } |
| 11374 | case NEON::BI__builtin_neon_vrnda_v: |
| 11375 | case NEON::BI__builtin_neon_vrndaq_v: { |
| 11376 | Int = Builder.getIsFPConstrained() |
| 11377 | ? Intrinsic::experimental_constrained_round |
| 11378 | : Intrinsic::round; |
| 11379 | return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrnda"); |
| 11380 | } |
| 11381 | case NEON::BI__builtin_neon_vrndih_f16: { |
| 11382 | Ops.push_back(EmitScalarExpr(E->getArg(0))); |
| 11383 | Int = Builder.getIsFPConstrained() |
| 11384 | ? Intrinsic::experimental_constrained_nearbyint |
| 11385 | : Intrinsic::nearbyint; |
| 11386 | return EmitNeonCall(CGM.getIntrinsic(Int, HalfTy), Ops, "vrndi"); |
| 11387 | } |
| 11388 | case NEON::BI__builtin_neon_vrndmh_f16: { |
| 11389 | Ops.push_back(EmitScalarExpr(E->getArg(0))); |
| 11390 | Int = Builder.getIsFPConstrained() |
| 11391 | ? Intrinsic::experimental_constrained_floor |
| 11392 | : Intrinsic::floor; |
| 11393 | return EmitNeonCall(CGM.getIntrinsic(Int, HalfTy), Ops, "vrndm"); |
| 11394 | } |
| 11395 | case NEON::BI__builtin_neon_vrndm_v: |
| 11396 | case NEON::BI__builtin_neon_vrndmq_v: { |
| 11397 | Int = Builder.getIsFPConstrained() |
| 11398 | ? Intrinsic::experimental_constrained_floor |
| 11399 | : Intrinsic::floor; |
| 11400 | return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrndm"); |
| 11401 | } |
| 11402 | case NEON::BI__builtin_neon_vrndnh_f16: { |
| 11403 | Ops.push_back(EmitScalarExpr(E->getArg(0))); |
| 11404 | Int = Builder.getIsFPConstrained() |
| 11405 | ? Intrinsic::experimental_constrained_roundeven |
| 11406 | : Intrinsic::roundeven; |
| 11407 | return EmitNeonCall(CGM.getIntrinsic(Int, HalfTy), Ops, "vrndn"); |
| 11408 | } |
| 11409 | case NEON::BI__builtin_neon_vrndn_v: |
| 11410 | case NEON::BI__builtin_neon_vrndnq_v: { |
| 11411 | Int = Builder.getIsFPConstrained() |
| 11412 | ? Intrinsic::experimental_constrained_roundeven |
| 11413 | : Intrinsic::roundeven; |
| 11414 | return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrndn"); |
| 11415 | } |
| 11416 | case NEON::BI__builtin_neon_vrndns_f32: { |
| 11417 | Ops.push_back(EmitScalarExpr(E->getArg(0))); |
| 11418 | Int = Builder.getIsFPConstrained() |
| 11419 | ? Intrinsic::experimental_constrained_roundeven |
| 11420 | : Intrinsic::roundeven; |
| 11421 | return EmitNeonCall(CGM.getIntrinsic(Int, FloatTy), Ops, "vrndn"); |
| 11422 | } |
| 11423 | case NEON::BI__builtin_neon_vrndph_f16: { |
| 11424 | Ops.push_back(EmitScalarExpr(E->getArg(0))); |
| 11425 | Int = Builder.getIsFPConstrained() |
| 11426 | ? Intrinsic::experimental_constrained_ceil |
| 11427 | : Intrinsic::ceil; |
| 11428 | return EmitNeonCall(CGM.getIntrinsic(Int, HalfTy), Ops, "vrndp"); |
| 11429 | } |
| 11430 | case NEON::BI__builtin_neon_vrndp_v: |
| 11431 | case NEON::BI__builtin_neon_vrndpq_v: { |
| 11432 | Int = Builder.getIsFPConstrained() |
| 11433 | ? Intrinsic::experimental_constrained_ceil |
| 11434 | : Intrinsic::ceil; |
| 11435 | return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrndp"); |
| 11436 | } |
| 11437 | case NEON::BI__builtin_neon_vrndxh_f16: { |
| 11438 | Ops.push_back(EmitScalarExpr(E->getArg(0))); |
| 11439 | Int = Builder.getIsFPConstrained() |
| 11440 | ? Intrinsic::experimental_constrained_rint |
| 11441 | : Intrinsic::rint; |
| 11442 | return EmitNeonCall(CGM.getIntrinsic(Int, HalfTy), Ops, "vrndx"); |
| 11443 | } |
| 11444 | case NEON::BI__builtin_neon_vrndx_v: |
| 11445 | case NEON::BI__builtin_neon_vrndxq_v: { |
| 11446 | Int = Builder.getIsFPConstrained() |
| 11447 | ? Intrinsic::experimental_constrained_rint |
| 11448 | : Intrinsic::rint; |
| 11449 | return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrndx"); |
| 11450 | } |
| 11451 | case NEON::BI__builtin_neon_vrndh_f16: { |
| 11452 | Ops.push_back(EmitScalarExpr(E->getArg(0))); |
| 11453 | Int = Builder.getIsFPConstrained() |
| 11454 | ? Intrinsic::experimental_constrained_trunc |
| 11455 | : Intrinsic::trunc; |
| 11456 | return EmitNeonCall(CGM.getIntrinsic(Int, HalfTy), Ops, "vrndz"); |
| 11457 | } |
| 11458 | case NEON::BI__builtin_neon_vrnd32x_f32: |
| 11459 | case NEON::BI__builtin_neon_vrnd32xq_f32: { |
| 11460 | Ops.push_back(EmitScalarExpr(E->getArg(0))); |
| 11461 | Int = Intrinsic::aarch64_neon_frint32x; |
| 11462 | return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrnd32x"); |
| 11463 | } |
| 11464 | case NEON::BI__builtin_neon_vrnd32z_f32: |
| 11465 | case NEON::BI__builtin_neon_vrnd32zq_f32: { |
| 11466 | Ops.push_back(EmitScalarExpr(E->getArg(0))); |
| 11467 | Int = Intrinsic::aarch64_neon_frint32z; |
| 11468 | return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrnd32z"); |
| 11469 | } |
| 11470 | case NEON::BI__builtin_neon_vrnd64x_f32: |
| 11471 | case NEON::BI__builtin_neon_vrnd64xq_f32: { |
| 11472 | Ops.push_back(EmitScalarExpr(E->getArg(0))); |
| 11473 | Int = Intrinsic::aarch64_neon_frint64x; |
| 11474 | return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrnd64x"); |
| 11475 | } |
| 11476 | case NEON::BI__builtin_neon_vrnd64z_f32: |
| 11477 | case NEON::BI__builtin_neon_vrnd64zq_f32: { |
| 11478 | Ops.push_back(EmitScalarExpr(E->getArg(0))); |
| 11479 | Int = Intrinsic::aarch64_neon_frint64z; |
| 11480 | return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrnd64z"); |
| 11481 | } |
| 11482 | case NEON::BI__builtin_neon_vrnd_v: |
| 11483 | case NEON::BI__builtin_neon_vrndq_v: { |
| 11484 | Int = Builder.getIsFPConstrained() |
| 11485 | ? Intrinsic::experimental_constrained_trunc |
| 11486 | : Intrinsic::trunc; |
| 11487 | return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrndz"); |
| 11488 | } |
| 11489 | case NEON::BI__builtin_neon_vcvt_f64_v: |
| 11490 | case NEON::BI__builtin_neon_vcvtq_f64_v: |
| 11491 | Ops[0] = Builder.CreateBitCast(Ops[0], Ty); |
| 11492 | Ty = GetNeonType(this, NeonTypeFlags(NeonTypeFlags::Float64, false, quad)); |
| 11493 | return usgn ? Builder.CreateUIToFP(Ops[0], Ty, "vcvt") |
| 11494 | : Builder.CreateSIToFP(Ops[0], Ty, "vcvt"); |
| 11495 | case NEON::BI__builtin_neon_vcvt_f64_f32: { |
| 11496 | assert(Type.getEltType() == NeonTypeFlags::Float64 && quad &&(static_cast <bool> (Type.getEltType() == NeonTypeFlags ::Float64 && quad && "unexpected vcvt_f64_f32 builtin" ) ? void (0) : __assert_fail ("Type.getEltType() == NeonTypeFlags::Float64 && quad && \"unexpected vcvt_f64_f32 builtin\"" , "clang/lib/CodeGen/CGBuiltin.cpp", 11497, __extension__ __PRETTY_FUNCTION__ )) |
| 11497 | "unexpected vcvt_f64_f32 builtin")(static_cast <bool> (Type.getEltType() == NeonTypeFlags ::Float64 && quad && "unexpected vcvt_f64_f32 builtin" ) ? void (0) : __assert_fail ("Type.getEltType() == NeonTypeFlags::Float64 && quad && \"unexpected vcvt_f64_f32 builtin\"" , "clang/lib/CodeGen/CGBuiltin.cpp", 11497, __extension__ __PRETTY_FUNCTION__ )); |
| 11498 | NeonTypeFlags SrcFlag = NeonTypeFlags(NeonTypeFlags::Float32, false, false); |
| 11499 | Ops[0] = Builder.CreateBitCast(Ops[0], GetNeonType(this, SrcFlag)); |
| 11500 | |
| 11501 | return Builder.CreateFPExt(Ops[0], Ty, "vcvt"); |
| 11502 | } |
| 11503 | case NEON::BI__builtin_neon_vcvt_f32_f64: { |
| 11504 | assert(Type.getEltType() == NeonTypeFlags::Float32 &&(static_cast <bool> (Type.getEltType() == NeonTypeFlags ::Float32 && "unexpected vcvt_f32_f64 builtin") ? void (0) : __assert_fail ("Type.getEltType() == NeonTypeFlags::Float32 && \"unexpected vcvt_f32_f64 builtin\"" , "clang/lib/CodeGen/CGBuiltin.cpp", 11505, __extension__ __PRETTY_FUNCTION__ )) |
| 11505 | "unexpected vcvt_f32_f64 builtin")(static_cast <bool> (Type.getEltType() == NeonTypeFlags ::Float32 && "unexpected vcvt_f32_f64 builtin") ? void (0) : __assert_fail ("Type.getEltType() == NeonTypeFlags::Float32 && \"unexpected vcvt_f32_f64 builtin\"" , "clang/lib/CodeGen/CGBuiltin.cpp", 11505, __extension__ __PRETTY_FUNCTION__ )); |
| 11506 | NeonTypeFlags SrcFlag = NeonTypeFlags(NeonTypeFlags::Float64, false, true); |
| 11507 | Ops[0] = Builder.CreateBitCast(Ops[0], GetNeonType(this, SrcFlag)); |
| 11508 | |
| 11509 | return Builder.CreateFPTrunc(Ops[0], Ty, "vcvt"); |
| 11510 | } |
| 11511 | case NEON::BI__builtin_neon_vcvt_s32_v: |
| 11512 | case NEON::BI__builtin_neon_vcvt_u32_v: |
| 11513 | case NEON::BI__builtin_neon_vcvt_s64_v: |
| 11514 | case NEON::BI__builtin_neon_vcvt_u64_v: |
| 11515 | case NEON::BI__builtin_neon_vcvt_s16_f16: |
| 11516 | case NEON::BI__builtin_neon_vcvt_u16_f16: |
| 11517 | case NEON::BI__builtin_neon_vcvtq_s32_v: |
| 11518 | case NEON::BI__builtin_neon_vcvtq_u32_v: |
| 11519 | case NEON::BI__builtin_neon_vcvtq_s64_v: |
| 11520 | case NEON::BI__builtin_neon_vcvtq_u64_v: |
| 11521 | case NEON::BI__builtin_neon_vcvtq_s16_f16: |
| 11522 | case NEON::BI__builtin_neon_vcvtq_u16_f16: { |
| 11523 | Int = |
| 11524 | usgn ? Intrinsic::aarch64_neon_fcvtzu : Intrinsic::aarch64_neon_fcvtzs; |
| 11525 | llvm::Type *Tys[2] = {Ty, GetFloatNeonType(this, Type)}; |
| 11526 | return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vcvtz"); |
| 11527 | } |
| 11528 | case NEON::BI__builtin_neon_vcvta_s16_f16: |
| 11529 | case NEON::BI__builtin_neon_vcvta_u16_f16: |
| 11530 | case NEON::BI__builtin_neon_vcvta_s32_v: |
| 11531 | case NEON::BI__builtin_neon_vcvtaq_s16_f16: |
| 11532 | case NEON::BI__builtin_neon_vcvtaq_s32_v: |
| 11533 | case NEON::BI__builtin_neon_vcvta_u32_v: |
| 11534 | case NEON::BI__builtin_neon_vcvtaq_u16_f16: |
| 11535 | case NEON::BI__builtin_neon_vcvtaq_u32_v: |
| 11536 | case NEON::BI__builtin_neon_vcvta_s64_v: |
| 11537 | case NEON::BI__builtin_neon_vcvtaq_s64_v: |
| 11538 | case NEON::BI__builtin_neon_vcvta_u64_v: |
| 11539 | case NEON::BI__builtin_neon_vcvtaq_u64_v: { |
| 11540 | Int = usgn ? Intrinsic::aarch64_neon_fcvtau : Intrinsic::aarch64_neon_fcvtas; |
| 11541 | llvm::Type *Tys[2] = { Ty, GetFloatNeonType(this, Type) }; |
| 11542 | return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vcvta"); |
| 11543 | } |
| 11544 | case NEON::BI__builtin_neon_vcvtm_s16_f16: |
| 11545 | case NEON::BI__builtin_neon_vcvtm_s32_v: |
| 11546 | case NEON::BI__builtin_neon_vcvtmq_s16_f16: |
| 11547 | case NEON::BI__builtin_neon_vcvtmq_s32_v: |
| 11548 | case NEON::BI__builtin_neon_vcvtm_u16_f16: |
| 11549 | case NEON::BI__builtin_neon_vcvtm_u32_v: |
| 11550 | case NEON::BI__builtin_neon_vcvtmq_u16_f16: |
| 11551 | case NEON::BI__builtin_neon_vcvtmq_u32_v: |
| 11552 | case NEON::BI__builtin_neon_vcvtm_s64_v: |
| 11553 | case NEON::BI__builtin_neon_vcvtmq_s64_v: |
| 11554 | case NEON::BI__builtin_neon_vcvtm_u64_v: |
| 11555 | case NEON::BI__builtin_neon_vcvtmq_u64_v: { |
| 11556 | Int = usgn ? Intrinsic::aarch64_neon_fcvtmu : Intrinsic::aarch64_neon_fcvtms; |
| 11557 | llvm::Type *Tys[2] = { Ty, GetFloatNeonType(this, Type) }; |
| 11558 | return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vcvtm"); |
| 11559 | } |
| 11560 | case NEON::BI__builtin_neon_vcvtn_s16_f16: |
| 11561 | case NEON::BI__builtin_neon_vcvtn_s32_v: |
| 11562 | case NEON::BI__builtin_neon_vcvtnq_s16_f16: |
| 11563 | case NEON::BI__builtin_neon_vcvtnq_s32_v: |
| 11564 | case NEON::BI__builtin_neon_vcvtn_u16_f16: |
| 11565 | case NEON::BI__builtin_neon_vcvtn_u32_v: |
| 11566 | case NEON::BI__builtin_neon_vcvtnq_u16_f16: |
| 11567 | case NEON::BI__builtin_neon_vcvtnq_u32_v: |
| 11568 | case NEON::BI__builtin_neon_vcvtn_s64_v: |
| 11569 | case NEON::BI__builtin_neon_vcvtnq_s64_v: |
| 11570 | case NEON::BI__builtin_neon_vcvtn_u64_v: |
| 11571 | case NEON::BI__builtin_neon_vcvtnq_u64_v: { |
| 11572 | Int = usgn ? Intrinsic::aarch64_neon_fcvtnu : Intrinsic::aarch64_neon_fcvtns; |
| 11573 | llvm::Type *Tys[2] = { Ty, GetFloatNeonType(this, Type) }; |
| 11574 | return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vcvtn"); |
| 11575 | } |
| 11576 | case NEON::BI__builtin_neon_vcvtp_s16_f16: |
| 11577 | case NEON::BI__builtin_neon_vcvtp_s32_v: |
| 11578 | case NEON::BI__builtin_neon_vcvtpq_s16_f16: |
| 11579 | case NEON::BI__builtin_neon_vcvtpq_s32_v: |
| 11580 | case NEON::BI__builtin_neon_vcvtp_u16_f16: |
| 11581 | case NEON::BI__builtin_neon_vcvtp_u32_v: |
| 11582 | case NEON::BI__builtin_neon_vcvtpq_u16_f16: |
| 11583 | case NEON::BI__builtin_neon_vcvtpq_u32_v: |
| 11584 | case NEON::BI__builtin_neon_vcvtp_s64_v: |
| 11585 | case NEON::BI__builtin_neon_vcvtpq_s64_v: |
| 11586 | case NEON::BI__builtin_neon_vcvtp_u64_v: |
| 11587 | case NEON::BI__builtin_neon_vcvtpq_u64_v: { |
| 11588 | Int = usgn ? Intrinsic::aarch64_neon_fcvtpu : Intrinsic::aarch64_neon_fcvtps; |
| 11589 | llvm::Type *Tys[2] = { Ty, GetFloatNeonType(this, Type) }; |
| 11590 | return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vcvtp"); |
| 11591 | } |
| 11592 | case NEON::BI__builtin_neon_vmulx_v: |
| 11593 | case NEON::BI__builtin_neon_vmulxq_v: { |
| 11594 | Int = Intrinsic::aarch64_neon_fmulx; |
| 11595 | return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmulx"); |
| 11596 | } |
| 11597 | case NEON::BI__builtin_neon_vmulxh_lane_f16: |
| 11598 | case NEON::BI__builtin_neon_vmulxh_laneq_f16: { |
| 11599 | // vmulx_lane should be mapped to Neon scalar mulx after |
| 11600 | // extracting the scalar element |
| 11601 | Ops.push_back(EmitScalarExpr(E->getArg(2))); |
| 11602 | Ops[1] = Builder.CreateExtractElement(Ops[1], Ops[2], "extract"); |
| 11603 | Ops.pop_back(); |
| 11604 | Int = Intrinsic::aarch64_neon_fmulx; |
| 11605 | return EmitNeonCall(CGM.getIntrinsic(Int, HalfTy), Ops, "vmulx"); |
| 11606 | } |
| 11607 | case NEON::BI__builtin_neon_vmul_lane_v: |
| 11608 | case NEON::BI__builtin_neon_vmul_laneq_v: { |
| 11609 | // v1f64 vmul_lane should be mapped to Neon scalar mul lane |
| 11610 | bool Quad = false; |
| 11611 | if (BuiltinID == NEON::BI__builtin_neon_vmul_laneq_v) |
| 11612 | Quad = true; |
| 11613 | Ops[0] = Builder.CreateBitCast(Ops[0], DoubleTy); |
| 11614 | llvm::FixedVectorType *VTy = |
| 11615 | GetNeonType(this, NeonTypeFlags(NeonTypeFlags::Float64, false, Quad)); |
| 11616 | Ops[1] = Builder.CreateBitCast(Ops[1], VTy); |
| 11617 | Ops[1] = Builder.CreateExtractElement(Ops[1], Ops[2], "extract"); |
| 11618 | Value *Result = Builder.CreateFMul(Ops[0], Ops[1]); |
| 11619 | return Builder.CreateBitCast(Result, Ty); |
| 11620 | } |
| 11621 | case NEON::BI__builtin_neon_vnegd_s64: |
| 11622 | return Builder.CreateNeg(EmitScalarExpr(E->getArg(0)), "vnegd"); |
| 11623 | case NEON::BI__builtin_neon_vnegh_f16: |
| 11624 | return Builder.CreateFNeg(EmitScalarExpr(E->getArg(0)), "vnegh"); |
| 11625 | case NEON::BI__builtin_neon_vpmaxnm_v: |
| 11626 | case NEON::BI__builtin_neon_vpmaxnmq_v: { |
| 11627 | Int = Intrinsic::aarch64_neon_fmaxnmp; |
| 11628 | return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vpmaxnm"); |
| 11629 | } |
| 11630 | case NEON::BI__builtin_neon_vpminnm_v: |
| 11631 | case NEON::BI__builtin_neon_vpminnmq_v: { |
| 11632 | Int = Intrinsic::aarch64_neon_fminnmp; |
| 11633 | return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vpminnm"); |
| 11634 | } |
| 11635 | case NEON::BI__builtin_neon_vsqrth_f16: { |
| 11636 | Ops.push_back(EmitScalarExpr(E->getArg(0))); |
| 11637 | Int = Builder.getIsFPConstrained() |
| 11638 | ? Intrinsic::experimental_constrained_sqrt |
| 11639 | : Intrinsic::sqrt; |
| 11640 | return EmitNeonCall(CGM.getIntrinsic(Int, HalfTy), Ops, "vsqrt"); |
| 11641 | } |
| 11642 | case NEON::BI__builtin_neon_vsqrt_v: |
| 11643 | case NEON::BI__builtin_neon_vsqrtq_v: { |
| 11644 | Int = Builder.getIsFPConstrained() |
| 11645 | ? Intrinsic::experimental_constrained_sqrt |
| 11646 | : Intrinsic::sqrt; |
| 11647 | Ops[0] = Builder.CreateBitCast(Ops[0], Ty); |
| 11648 | return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vsqrt"); |
| 11649 | } |
| 11650 | case NEON::BI__builtin_neon_vrbit_v: |
| 11651 | case NEON::BI__builtin_neon_vrbitq_v: { |
| 11652 | Int = Intrinsic::bitreverse; |
| 11653 | return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrbit"); |
| 11654 | } |
| 11655 | case NEON::BI__builtin_neon_vaddv_u8: |
| 11656 | // FIXME: These are handled by the AArch64 scalar code. |
| 11657 | usgn = true; |
| 11658 | [[fallthrough]]; |
| 11659 | case NEON::BI__builtin_neon_vaddv_s8: { |
| 11660 | Int = usgn ? Intrinsic::aarch64_neon_uaddv : Intrinsic::aarch64_neon_saddv; |
| 11661 | Ty = Int32Ty; |
| 11662 | VTy = llvm::FixedVectorType::get(Int8Ty, 8); |
| 11663 | llvm::Type *Tys[2] = { Ty, VTy }; |
| 11664 | Ops.push_back(EmitScalarExpr(E->getArg(0))); |
| 11665 | Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddv"); |
| 11666 | return Builder.CreateTrunc(Ops[0], Int8Ty); |
| 11667 | } |
| 11668 | case NEON::BI__builtin_neon_vaddv_u16: |
| 11669 | usgn = true; |
| 11670 | [[fallthrough]]; |
| 11671 | case NEON::BI__builtin_neon_vaddv_s16: { |
| 11672 | Int = usgn ? Intrinsic::aarch64_neon_uaddv : Intrinsic::aarch64_neon_saddv; |
| 11673 | Ty = Int32Ty; |
| 11674 | VTy = llvm::FixedVectorType::get(Int16Ty, 4); |
| 11675 | llvm::Type *Tys[2] = { Ty, VTy }; |
| 11676 | Ops.push_back(EmitScalarExpr(E->getArg(0))); |
| 11677 | Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddv"); |
| 11678 | return Builder.CreateTrunc(Ops[0], Int16Ty); |
| 11679 | } |
| 11680 | case NEON::BI__builtin_neon_vaddvq_u8: |
| 11681 | usgn = true; |
| 11682 | [[fallthrough]]; |
| 11683 | case NEON::BI__builtin_neon_vaddvq_s8: { |
| 11684 | Int = usgn ? Intrinsic::aarch64_neon_uaddv : Intrinsic::aarch64_neon_saddv; |
| 11685 | Ty = Int32Ty; |
| 11686 | VTy = llvm::FixedVectorType::get(Int8Ty, 16); |
| 11687 | llvm::Type *Tys[2] = { Ty, VTy }; |
| 11688 | Ops.push_back(EmitScalarExpr(E->getArg(0))); |
| 11689 | Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddv"); |
| 11690 | return Builder.CreateTrunc(Ops[0], Int8Ty); |
| 11691 | } |
| 11692 | case NEON::BI__builtin_neon_vaddvq_u16: |
| 11693 | usgn = true; |
| 11694 | [[fallthrough]]; |
| 11695 | case NEON::BI__builtin_neon_vaddvq_s16: { |
| 11696 | Int = usgn ? Intrinsic::aarch64_neon_uaddv : Intrinsic::aarch64_neon_saddv; |
| 11697 | Ty = Int32Ty; |
| 11698 | VTy = llvm::FixedVectorType::get(Int16Ty, 8); |
| 11699 | llvm::Type *Tys[2] = { Ty, VTy }; |
| 11700 | Ops.push_back(EmitScalarExpr(E->getArg(0))); |
| 11701 | Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddv"); |
| 11702 | return Builder.CreateTrunc(Ops[0], Int16Ty); |
| 11703 | } |
| 11704 | case NEON::BI__builtin_neon_vmaxv_u8: { |
| 11705 | Int = Intrinsic::aarch64_neon_umaxv; |
| 11706 | Ty = Int32Ty; |
| 11707 | VTy = llvm::FixedVectorType::get(Int8Ty, 8); |
| 11708 | llvm::Type *Tys[2] = { Ty, VTy }; |
| 11709 | Ops.push_back(EmitScalarExpr(E->getArg(0))); |
| 11710 | Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv"); |
| 11711 | return Builder.CreateTrunc(Ops[0], Int8Ty); |
| 11712 | } |
| 11713 | case NEON::BI__builtin_neon_vmaxv_u16: { |
| 11714 | Int = Intrinsic::aarch64_neon_umaxv; |
| 11715 | Ty = Int32Ty; |
| 11716 | VTy = llvm::FixedVectorType::get(Int16Ty, 4); |
| 11717 | llvm::Type *Tys[2] = { Ty, VTy }; |
| 11718 | Ops.push_back(EmitScalarExpr(E->getArg(0))); |
| 11719 | Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv"); |
| 11720 | return Builder.CreateTrunc(Ops[0], Int16Ty); |
| 11721 | } |
| 11722 | case NEON::BI__builtin_neon_vmaxvq_u8: { |
| 11723 | Int = Intrinsic::aarch64_neon_umaxv; |
| 11724 | Ty = Int32Ty; |
| 11725 | VTy = llvm::FixedVectorType::get(Int8Ty, 16); |
| 11726 | llvm::Type *Tys[2] = { Ty, VTy }; |
| 11727 | Ops.push_back(EmitScalarExpr(E->getArg(0))); |
| 11728 | Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv"); |
| 11729 | return Builder.CreateTrunc(Ops[0], Int8Ty); |
| 11730 | } |
| 11731 | case NEON::BI__builtin_neon_vmaxvq_u16: { |
| 11732 | Int = Intrinsic::aarch64_neon_umaxv; |
| 11733 | Ty = Int32Ty; |
| 11734 | VTy = llvm::FixedVectorType::get(Int16Ty, 8); |
| 11735 | llvm::Type *Tys[2] = { Ty, VTy }; |
| 11736 | Ops.push_back(EmitScalarExpr(E->getArg(0))); |
| 11737 | Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv"); |
| 11738 | return Builder.CreateTrunc(Ops[0], Int16Ty); |
| 11739 | } |
| 11740 | case NEON::BI__builtin_neon_vmaxv_s8: { |
| 11741 | Int = Intrinsic::aarch64_neon_smaxv; |
| 11742 | Ty = Int32Ty; |
| 11743 | VTy = llvm::FixedVectorType::get(Int8Ty, 8); |
| 11744 | llvm::Type *Tys[2] = { Ty, VTy }; |
| 11745 | Ops.push_back(EmitScalarExpr(E->getArg(0))); |
| 11746 | Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv"); |
| 11747 | return Builder.CreateTrunc(Ops[0], Int8Ty); |
| 11748 | } |
| 11749 | case NEON::BI__builtin_neon_vmaxv_s16: { |
| 11750 | Int = Intrinsic::aarch64_neon_smaxv; |
| 11751 | Ty = Int32Ty; |
| 11752 | VTy = llvm::FixedVectorType::get(Int16Ty, 4); |
| 11753 | llvm::Type *Tys[2] = { Ty, VTy }; |
| 11754 | Ops.push_back(EmitScalarExpr(E->getArg(0))); |
| 11755 | Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv"); |
| 11756 | return Builder.CreateTrunc(Ops[0], Int16Ty); |
| 11757 | } |
| 11758 | case NEON::BI__builtin_neon_vmaxvq_s8: { |
| 11759 | Int = Intrinsic::aarch64_neon_smaxv; |
| 11760 | Ty = Int32Ty; |
| 11761 | VTy = llvm::FixedVectorType::get(Int8Ty, 16); |
| 11762 | llvm::Type *Tys[2] = { Ty, VTy }; |
| 11763 | Ops.push_back(EmitScalarExpr(E->getArg(0))); |
| 11764 | Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv"); |
| 11765 | return Builder.CreateTrunc(Ops[0], Int8Ty); |
| 11766 | } |
| 11767 | case NEON::BI__builtin_neon_vmaxvq_s16: { |
| 11768 | Int = Intrinsic::aarch64_neon_smaxv; |
| 11769 | Ty = Int32Ty; |
| 11770 | VTy = llvm::FixedVectorType::get(Int16Ty, 8); |
| 11771 | llvm::Type *Tys[2] = { Ty, VTy }; |
| 11772 | Ops.push_back(EmitScalarExpr(E->getArg(0))); |
| 11773 | Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv"); |
| 11774 | return Builder.CreateTrunc(Ops[0], Int16Ty); |
| 11775 | } |
| 11776 | case NEON::BI__builtin_neon_vmaxv_f16: { |
| 11777 | Int = Intrinsic::aarch64_neon_fmaxv; |
| 11778 | Ty = HalfTy; |
| 11779 | VTy = llvm::FixedVectorType::get(HalfTy, 4); |
| 11780 | llvm::Type *Tys[2] = { Ty, VTy }; |
| 11781 | Ops.push_back(EmitScalarExpr(E->getArg(0))); |
| 11782 | Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv"); |
| 11783 | return Builder.CreateTrunc(Ops[0], HalfTy); |
| 11784 | } |
| 11785 | case NEON::BI__builtin_neon_vmaxvq_f16: { |
| 11786 | Int = Intrinsic::aarch64_neon_fmaxv; |
| 11787 | Ty = HalfTy; |
| 11788 | VTy = llvm::FixedVectorType::get(HalfTy, 8); |
| 11789 | llvm::Type *Tys[2] = { Ty, VTy }; |
| 11790 | Ops.push_back(EmitScalarExpr(E->getArg(0))); |
| 11791 | Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv"); |
| 11792 | return Builder.CreateTrunc(Ops[0], HalfTy); |
| 11793 | } |
| 11794 | case NEON::BI__builtin_neon_vminv_u8: { |
| 11795 | Int = Intrinsic::aarch64_neon_uminv; |
| 11796 | Ty = Int32Ty; |
| 11797 | VTy = llvm::FixedVectorType::get(Int8Ty, 8); |
| 11798 | llvm::Type *Tys[2] = { Ty, VTy }; |
| 11799 | Ops.push_back(EmitScalarExpr(E->getArg(0))); |
| 11800 | Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv"); |
| 11801 | return Builder.CreateTrunc(Ops[0], Int8Ty); |
| 11802 | } |
| 11803 | case NEON::BI__builtin_neon_vminv_u16: { |
| 11804 | Int = Intrinsic::aarch64_neon_uminv; |
| 11805 | Ty = Int32Ty; |
| 11806 | VTy = llvm::FixedVectorType::get(Int16Ty, 4); |
| 11807 | llvm::Type *Tys[2] = { Ty, VTy }; |
| 11808 | Ops.push_back(EmitScalarExpr(E->getArg(0))); |
| 11809 | Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv"); |
| 11810 | return Builder.CreateTrunc(Ops[0], Int16Ty); |
| 11811 | } |
| 11812 | case NEON::BI__builtin_neon_vminvq_u8: { |
| 11813 | Int = Intrinsic::aarch64_neon_uminv; |
| 11814 | Ty = Int32Ty; |
| 11815 | VTy = llvm::FixedVectorType::get(Int8Ty, 16); |
| 11816 | llvm::Type *Tys[2] = { Ty, VTy }; |
| 11817 | Ops.push_back(EmitScalarExpr(E->getArg(0))); |
| 11818 | Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv"); |
| 11819 | return Builder.CreateTrunc(Ops[0], Int8Ty); |
| 11820 | } |
| 11821 | case NEON::BI__builtin_neon_vminvq_u16: { |
| 11822 | Int = Intrinsic::aarch64_neon_uminv; |
| 11823 | Ty = Int32Ty; |
| 11824 | VTy = llvm::FixedVectorType::get(Int16Ty, 8); |
| 11825 | llvm::Type *Tys[2] = { Ty, VTy }; |
| 11826 | Ops.push_back(EmitScalarExpr(E->getArg(0))); |
| 11827 | Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv"); |
| 11828 | return Builder.CreateTrunc(Ops[0], Int16Ty); |
| 11829 | } |
| 11830 | case NEON::BI__builtin_neon_vminv_s8: { |
| 11831 | Int = Intrinsic::aarch64_neon_sminv; |
| 11832 | Ty = Int32Ty; |
| 11833 | VTy = llvm::FixedVectorType::get(Int8Ty, 8); |
| 11834 | llvm::Type *Tys[2] = { Ty, VTy }; |
| 11835 | Ops.push_back(EmitScalarExpr(E->getArg(0))); |
| 11836 | Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv"); |
| 11837 | return Builder.CreateTrunc(Ops[0], Int8Ty); |
| 11838 | } |
| 11839 | case NEON::BI__builtin_neon_vminv_s16: { |
| 11840 | Int = Intrinsic::aarch64_neon_sminv; |
| 11841 | Ty = Int32Ty; |
| 11842 | VTy = llvm::FixedVectorType::get(Int16Ty, 4); |
| 11843 | llvm::Type *Tys[2] = { Ty, VTy }; |
| 11844 | Ops.push_back(EmitScalarExpr(E->getArg(0))); |
| 11845 | Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv"); |
| 11846 | return Builder.CreateTrunc(Ops[0], Int16Ty); |
| 11847 | } |
| 11848 | case NEON::BI__builtin_neon_vminvq_s8: { |
| 11849 | Int = Intrinsic::aarch64_neon_sminv; |
| 11850 | Ty = Int32Ty; |
| 11851 | VTy = llvm::FixedVectorType::get(Int8Ty, 16); |
| 11852 | llvm::Type *Tys[2] = { Ty, VTy }; |
| 11853 | Ops.push_back(EmitScalarExpr(E->getArg(0))); |
| 11854 | Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv"); |
| 11855 | return Builder.CreateTrunc(Ops[0], Int8Ty); |
| 11856 | } |
| 11857 | case NEON::BI__builtin_neon_vminvq_s16: { |
| 11858 | Int = Intrinsic::aarch64_neon_sminv; |
| 11859 | Ty = Int32Ty; |
| 11860 | VTy = llvm::FixedVectorType::get(Int16Ty, 8); |
| 11861 | llvm::Type *Tys[2] = { Ty, VTy }; |
| 11862 | Ops.push_back(EmitScalarExpr(E->getArg(0))); |
| 11863 | Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv"); |
| 11864 | return Builder.CreateTrunc(Ops[0], Int16Ty); |
| 11865 | } |
| 11866 | case NEON::BI__builtin_neon_vminv_f16: { |
| 11867 | Int = Intrinsic::aarch64_neon_fminv; |
| 11868 | Ty = HalfTy; |
| 11869 | VTy = llvm::FixedVectorType::get(HalfTy, 4); |
| 11870 | llvm::Type *Tys[2] = { Ty, VTy }; |
| 11871 | Ops.push_back(EmitScalarExpr(E->getArg(0))); |
| 11872 | Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv"); |
| 11873 | return Builder.CreateTrunc(Ops[0], HalfTy); |
| 11874 | } |
| 11875 | case NEON::BI__builtin_neon_vminvq_f16: { |
| 11876 | Int = Intrinsic::aarch64_neon_fminv; |
| 11877 | Ty = HalfTy; |
| 11878 | VTy = llvm::FixedVectorType::get(HalfTy, 8); |
| 11879 | llvm::Type *Tys[2] = { Ty, VTy }; |
| 11880 | Ops.push_back(EmitScalarExpr(E->getArg(0))); |
| 11881 | Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv"); |
| 11882 | return Builder.CreateTrunc(Ops[0], HalfTy); |
| 11883 | } |
| 11884 | case NEON::BI__builtin_neon_vmaxnmv_f16: { |
| 11885 | Int = Intrinsic::aarch64_neon_fmaxnmv; |
| 11886 | Ty = HalfTy; |
| 11887 | VTy = llvm::FixedVectorType::get(HalfTy, 4); |
| 11888 | llvm::Type *Tys[2] = { Ty, VTy }; |
| 11889 | Ops.push_back(EmitScalarExpr(E->getArg(0))); |
| 11890 | Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxnmv"); |
| 11891 | return Builder.CreateTrunc(Ops[0], HalfTy); |
| 11892 | } |
| 11893 | case NEON::BI__builtin_neon_vmaxnmvq_f16: { |
| 11894 | Int = Intrinsic::aarch64_neon_fmaxnmv; |
| 11895 | Ty = HalfTy; |
| 11896 | VTy = llvm::FixedVectorType::get(HalfTy, 8); |
| 11897 | llvm::Type *Tys[2] = { Ty, VTy }; |
| 11898 | Ops.push_back(EmitScalarExpr(E->getArg(0))); |
| 11899 | Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxnmv"); |
| 11900 | return Builder.CreateTrunc(Ops[0], HalfTy); |
| 11901 | } |
| 11902 | case NEON::BI__builtin_neon_vminnmv_f16: { |
| 11903 | Int = Intrinsic::aarch64_neon_fminnmv; |
| 11904 | Ty = HalfTy; |
| 11905 | VTy = llvm::FixedVectorType::get(HalfTy, 4); |
| 11906 | llvm::Type *Tys[2] = { Ty, VTy }; |
| 11907 | Ops.push_back(EmitScalarExpr(E->getArg(0))); |
| 11908 | Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminnmv"); |
| 11909 | return Builder.CreateTrunc(Ops[0], HalfTy); |
| 11910 | } |
| 11911 | case NEON::BI__builtin_neon_vminnmvq_f16: { |
| 11912 | Int = Intrinsic::aarch64_neon_fminnmv; |
| 11913 | Ty = HalfTy; |
| 11914 | VTy = llvm::FixedVectorType::get(HalfTy, 8); |
| 11915 | llvm::Type *Tys[2] = { Ty, VTy }; |
| 11916 | Ops.push_back(EmitScalarExpr(E->getArg(0))); |
| 11917 | Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminnmv"); |
| 11918 | return Builder.CreateTrunc(Ops[0], HalfTy); |
| 11919 | } |
| 11920 | case NEON::BI__builtin_neon_vmul_n_f64: { |
| 11921 | Ops[0] = Builder.CreateBitCast(Ops[0], DoubleTy); |
| 11922 | Value *RHS = Builder.CreateBitCast(EmitScalarExpr(E->getArg(1)), DoubleTy); |
| 11923 | return Builder.CreateFMul(Ops[0], RHS); |
| 11924 | } |
| 11925 | case NEON::BI__builtin_neon_vaddlv_u8: { |
| 11926 | Int = Intrinsic::aarch64_neon_uaddlv; |
| 11927 | Ty = Int32Ty; |
| 11928 | VTy = llvm::FixedVectorType::get(Int8Ty, 8); |
| 11929 | llvm::Type *Tys[2] = { Ty, VTy }; |
| 11930 | Ops.push_back(EmitScalarExpr(E->getArg(0))); |
| 11931 | Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv"); |
| 11932 | return Builder.CreateTrunc(Ops[0], Int16Ty); |
| 11933 | } |
| 11934 | case NEON::BI__builtin_neon_vaddlv_u16: { |
| 11935 | Int = Intrinsic::aarch64_neon_uaddlv; |
| 11936 | Ty = Int32Ty; |
| 11937 | VTy = llvm::FixedVectorType::get(Int16Ty, 4); |
| 11938 | llvm::Type *Tys[2] = { Ty, VTy }; |
| 11939 | Ops.push_back(EmitScalarExpr(E->getArg(0))); |
| 11940 | return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv"); |
| 11941 | } |
| 11942 | case NEON::BI__builtin_neon_vaddlvq_u8: { |
| 11943 | Int = Intrinsic::aarch64_neon_uaddlv; |
| 11944 | Ty = Int32Ty; |
| 11945 | VTy = llvm::FixedVectorType::get(Int8Ty, 16); |
| 11946 | llvm::Type *Tys[2] = { Ty, VTy }; |
| 11947 | Ops.push_back(EmitScalarExpr(E->getArg(0))); |
| 11948 | Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv"); |
| 11949 | return Builder.CreateTrunc(Ops[0], Int16Ty); |
| 11950 | } |
| 11951 | case NEON::BI__builtin_neon_vaddlvq_u16: { |
| 11952 | Int = Intrinsic::aarch64_neon_uaddlv; |
| 11953 | Ty = Int32Ty; |
| 11954 | VTy = llvm::FixedVectorType::get(Int16Ty, 8); |
| 11955 | llvm::Type *Tys[2] = { Ty, VTy }; |
| 11956 | Ops.push_back(EmitScalarExpr(E->getArg(0))); |
| 11957 | return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv"); |
| 11958 | } |
| 11959 | case NEON::BI__builtin_neon_vaddlv_s8: { |
| 11960 | Int = Intrinsic::aarch64_neon_saddlv; |
| 11961 | Ty = Int32Ty; |
| 11962 | VTy = llvm::FixedVectorType::get(Int8Ty, 8); |
| 11963 | llvm::Type *Tys[2] = { Ty, VTy }; |
| 11964 | Ops.push_back(EmitScalarExpr(E->getArg(0))); |
| 11965 | Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv"); |
| 11966 | return Builder.CreateTrunc(Ops[0], Int16Ty); |
| 11967 | } |
| 11968 | case NEON::BI__builtin_neon_vaddlv_s16: { |
| 11969 | Int = Intrinsic::aarch64_neon_saddlv; |
| 11970 | Ty = Int32Ty; |
| 11971 | VTy = llvm::FixedVectorType::get(Int16Ty, 4); |
| 11972 | llvm::Type *Tys[2] = { Ty, VTy }; |
| 11973 | Ops.push_back(EmitScalarExpr(E->getArg(0))); |
| 11974 | return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv"); |
| 11975 | } |
| 11976 | case NEON::BI__builtin_neon_vaddlvq_s8: { |
| 11977 | Int = Intrinsic::aarch64_neon_saddlv; |
| 11978 | Ty = Int32Ty; |
| 11979 | VTy = llvm::FixedVectorType::get(Int8Ty, 16); |
| 11980 | llvm::Type *Tys[2] = { Ty, VTy }; |
| 11981 | Ops.push_back(EmitScalarExpr(E->getArg(0))); |
| 11982 | Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv"); |
| 11983 | return Builder.CreateTrunc(Ops[0], Int16Ty); |
| 11984 | } |
| 11985 | case NEON::BI__builtin_neon_vaddlvq_s16: { |
| 11986 | Int = Intrinsic::aarch64_neon_saddlv; |
| 11987 | Ty = Int32Ty; |
| 11988 | VTy = llvm::FixedVectorType::get(Int16Ty, 8); |
| 11989 | llvm::Type *Tys[2] = { Ty, VTy }; |
| 11990 | Ops.push_back(EmitScalarExpr(E->getArg(0))); |
| 11991 | return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv"); |
| 11992 | } |
| 11993 | case NEON::BI__builtin_neon_vsri_n_v: |
| 11994 | case NEON::BI__builtin_neon_vsriq_n_v: { |
| 11995 | Int = Intrinsic::aarch64_neon_vsri; |
| 11996 | llvm::Function *Intrin = CGM.getIntrinsic(Int, Ty); |
| 11997 | return EmitNeonCall(Intrin, Ops, "vsri_n"); |
| 11998 | } |
| 11999 | case NEON::BI__builtin_neon_vsli_n_v: |
| 12000 | case NEON::BI__builtin_neon_vsliq_n_v: { |
| 12001 | Int = Intrinsic::aarch64_neon_vsli; |
| 12002 | llvm::Function *Intrin = CGM.getIntrinsic(Int, Ty); |
| 12003 | return EmitNeonCall(Intrin, Ops, "vsli_n"); |
| 12004 | } |
| 12005 | case NEON::BI__builtin_neon_vsra_n_v: |
| 12006 | case NEON::BI__builtin_neon_vsraq_n_v: |
| 12007 | Ops[0] = Builder.CreateBitCast(Ops[0], Ty); |
| 12008 | Ops[1] = EmitNeonRShiftImm(Ops[1], Ops[2], Ty, usgn, "vsra_n"); |
| 12009 | return Builder.CreateAdd(Ops[0], Ops[1]); |
| 12010 | case NEON::BI__builtin_neon_vrsra_n_v: |
| 12011 | case NEON::BI__builtin_neon_vrsraq_n_v: { |
| 12012 | Int = usgn ? Intrinsic::aarch64_neon_urshl : Intrinsic::aarch64_neon_srshl; |
| 12013 | SmallVector<llvm::Value*,2> TmpOps; |
| 12014 | TmpOps.push_back(Ops[1]); |
| 12015 | TmpOps.push_back(Ops[2]); |
| 12016 | Function* F = CGM.getIntrinsic(Int, Ty); |
| 12017 | llvm::Value *tmp = EmitNeonCall(F, TmpOps, "vrshr_n", 1, true); |
| 12018 | Ops[0] = Builder.CreateBitCast(Ops[0], VTy); |
| 12019 | return Builder.CreateAdd(Ops[0], tmp); |
| 12020 | } |
| 12021 | case NEON::BI__builtin_neon_vld1_v: |
| 12022 | case NEON::BI__builtin_neon_vld1q_v: { |
| 12023 | Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(VTy)); |
| 12024 | return Builder.CreateAlignedLoad(VTy, Ops[0], PtrOp0.getAlignment()); |
| 12025 | } |
| 12026 | case NEON::BI__builtin_neon_vst1_v: |
| 12027 | case NEON::BI__builtin_neon_vst1q_v: |
| 12028 | Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(VTy)); |
| 12029 | Ops[1] = Builder.CreateBitCast(Ops[1], VTy); |
| 12030 | return Builder.CreateAlignedStore(Ops[1], Ops[0], PtrOp0.getAlignment()); |
| 12031 | case NEON::BI__builtin_neon_vld1_lane_v: |
| 12032 | case NEON::BI__builtin_neon_vld1q_lane_v: { |
| 12033 | Ops[1] = Builder.CreateBitCast(Ops[1], Ty); |
| 12034 | Ty = llvm::PointerType::getUnqual(VTy->getElementType()); |
| 12035 | Ops[0] = Builder.CreateBitCast(Ops[0], Ty); |
| 12036 | Ops[0] = Builder.CreateAlignedLoad(VTy->getElementType(), Ops[0], |
| 12037 | PtrOp0.getAlignment()); |
| 12038 | return Builder.CreateInsertElement(Ops[1], Ops[0], Ops[2], "vld1_lane"); |
| 12039 | } |
| 12040 | case NEON::BI__builtin_neon_vld1_dup_v: |
| 12041 | case NEON::BI__builtin_neon_vld1q_dup_v: { |
| 12042 | Value *V = PoisonValue::get(Ty); |
| 12043 | Ty = llvm::PointerType::getUnqual(VTy->getElementType()); |
| 12044 | Ops[0] = Builder.CreateBitCast(Ops[0], Ty); |
| 12045 | Ops[0] = Builder.CreateAlignedLoad(VTy->getElementType(), Ops[0], |
| 12046 | PtrOp0.getAlignment()); |
| 12047 | llvm::Constant *CI = ConstantInt::get(Int32Ty, 0); |
| 12048 | Ops[0] = Builder.CreateInsertElement(V, Ops[0], CI); |
| 12049 | return EmitNeonSplat(Ops[0], CI); |
| 12050 | } |
| 12051 | case NEON::BI__builtin_neon_vst1_lane_v: |
| 12052 | case NEON::BI__builtin_neon_vst1q_lane_v: |
| 12053 | Ops[1] = Builder.CreateBitCast(Ops[1], Ty); |
| 12054 | Ops[1] = Builder.CreateExtractElement(Ops[1], Ops[2]); |
| 12055 | Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); |
| 12056 | return Builder.CreateAlignedStore(Ops[1], Builder.CreateBitCast(Ops[0], Ty), |
| 12057 | PtrOp0.getAlignment()); |
| 12058 | case NEON::BI__builtin_neon_vld2_v: |
| 12059 | case NEON::BI__builtin_neon_vld2q_v: { |
| 12060 | llvm::Type *PTy = llvm::PointerType::getUnqual(VTy); |
| 12061 | Ops[1] = Builder.CreateBitCast(Ops[1], PTy); |
| 12062 | llvm::Type *Tys[2] = { VTy, PTy }; |
| 12063 | Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld2, Tys); |
| 12064 | Ops[1] = Builder.CreateCall(F, Ops[1], "vld2"); |
| 12065 | Ops[0] = Builder.CreateBitCast(Ops[0], |
| 12066 | llvm::PointerType::getUnqual(Ops[1]->getType())); |
| 12067 | return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]); |
| 12068 | } |
| 12069 | case NEON::BI__builtin_neon_vld3_v: |
| 12070 | case NEON::BI__builtin_neon_vld3q_v: { |
| 12071 | llvm::Type *PTy = llvm::PointerType::getUnqual(VTy); |
| 12072 | Ops[1] = Builder.CreateBitCast(Ops[1], PTy); |
| 12073 | llvm::Type *Tys[2] = { VTy, PTy }; |
| 12074 | Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld3, Tys); |
| 12075 | Ops[1] = Builder.CreateCall(F, Ops[1], "vld3"); |
| 12076 | Ops[0] = Builder.CreateBitCast(Ops[0], |
| 12077 | llvm::PointerType::getUnqual(Ops[1]->getType())); |
| 12078 | return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]); |
| 12079 | } |
| 12080 | case NEON::BI__builtin_neon_vld4_v: |
| 12081 | case NEON::BI__builtin_neon_vld4q_v: { |
| 12082 | llvm::Type *PTy = llvm::PointerType::getUnqual(VTy); |
| 12083 | Ops[1] = Builder.CreateBitCast(Ops[1], PTy); |
| 12084 | llvm::Type *Tys[2] = { VTy, PTy }; |
| 12085 | Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld4, Tys); |
| 12086 | Ops[1] = Builder.CreateCall(F, Ops[1], "vld4"); |
| 12087 | Ops[0] = Builder.CreateBitCast(Ops[0], |
| 12088 | llvm::PointerType::getUnqual(Ops[1]->getType())); |
| 12089 | return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]); |
| 12090 | } |
| 12091 | case NEON::BI__builtin_neon_vld2_dup_v: |
| 12092 | case NEON::BI__builtin_neon_vld2q_dup_v: { |
| 12093 | llvm::Type *PTy = |
| 12094 | llvm::PointerType::getUnqual(VTy->getElementType()); |
| 12095 | Ops[1] = Builder.CreateBitCast(Ops[1], PTy); |
| 12096 | llvm::Type *Tys[2] = { VTy, PTy }; |
| 12097 | Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld2r, Tys); |
| 12098 | Ops[1] = Builder.CreateCall(F, Ops[1], "vld2"); |
| 12099 | Ops[0] = Builder.CreateBitCast(Ops[0], |
| 12100 | llvm::PointerType::getUnqual(Ops[1]->getType())); |
| 12101 | return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]); |
| 12102 | } |
| 12103 | case NEON::BI__builtin_neon_vld3_dup_v: |
| 12104 | case NEON::BI__builtin_neon_vld3q_dup_v: { |
| 12105 | llvm::Type *PTy = |
| 12106 | llvm::PointerType::getUnqual(VTy->getElementType()); |
| 12107 | Ops[1] = Builder.CreateBitCast(Ops[1], PTy); |
| 12108 | llvm::Type *Tys[2] = { VTy, PTy }; |
| 12109 | Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld3r, Tys); |
| 12110 | Ops[1] = Builder.CreateCall(F, Ops[1], "vld3"); |
| 12111 | Ops[0] = Builder.CreateBitCast(Ops[0], |
| 12112 | llvm::PointerType::getUnqual(Ops[1]->getType())); |
| 12113 | return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]); |
| 12114 | } |
| 12115 | case NEON::BI__builtin_neon_vld4_dup_v: |
| 12116 | case NEON::BI__builtin_neon_vld4q_dup_v: { |
| 12117 | llvm::Type *PTy = |
| 12118 | llvm::PointerType::getUnqual(VTy->getElementType()); |
| 12119 | Ops[1] = Builder.CreateBitCast(Ops[1], PTy); |
| 12120 | llvm::Type *Tys[2] = { VTy, PTy }; |
| 12121 | Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld4r, Tys); |
| 12122 | Ops[1] = Builder.CreateCall(F, Ops[1], "vld4"); |
| 12123 | Ops[0] = Builder.CreateBitCast(Ops[0], |
| 12124 | llvm::PointerType::getUnqual(Ops[1]->getType())); |
| 12125 | return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]); |
| 12126 | } |
| 12127 | case NEON::BI__builtin_neon_vld2_lane_v: |
| 12128 | case NEON::BI__builtin_neon_vld2q_lane_v: { |
| 12129 | llvm::Type *Tys[2] = { VTy, Ops[1]->getType() }; |
| 12130 | Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld2lane, Tys); |
| 12131 | std::rotate(Ops.begin() + 1, Ops.begin() + 2, Ops.end()); |
| 12132 | Ops[1] = Builder.CreateBitCast(Ops[1], Ty); |
| 12133 | Ops[2] = Builder.CreateBitCast(Ops[2], Ty); |
| 12134 | Ops[3] = Builder.CreateZExt(Ops[3], Int64Ty); |
| 12135 | Ops[1] = Builder.CreateCall(F, ArrayRef(Ops).slice(1), "vld2_lane"); |
| 12136 | Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); |
| 12137 | Ops[0] = Builder.CreateBitCast(Ops[0], Ty); |
| 12138 | return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]); |
| 12139 | } |
| 12140 | case NEON::BI__builtin_neon_vld3_lane_v: |
| 12141 | case NEON::BI__builtin_neon_vld3q_lane_v: { |
| 12142 | llvm::Type *Tys[2] = { VTy, Ops[1]->getType() }; |
| 12143 | Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld3lane, Tys); |
| 12144 | std::rotate(Ops.begin() + 1, Ops.begin() + 2, Ops.end()); |
| 12145 | Ops[1] = Builder.CreateBitCast(Ops[1], Ty); |
| 12146 | Ops[2] = Builder.CreateBitCast(Ops[2], Ty); |
| 12147 | Ops[3] = Builder.CreateBitCast(Ops[3], Ty); |
| 12148 | Ops[4] = Builder.CreateZExt(Ops[4], Int64Ty); |
| 12149 | Ops[1] = Builder.CreateCall(F, ArrayRef(Ops).slice(1), "vld3_lane"); |
| 12150 | Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); |
| 12151 | Ops[0] = Builder.CreateBitCast(Ops[0], Ty); |
| 12152 | return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]); |
| 12153 | } |
| 12154 | case NEON::BI__builtin_neon_vld4_lane_v: |
| 12155 | case NEON::BI__builtin_neon_vld4q_lane_v: { |
| 12156 | llvm::Type *Tys[2] = { VTy, Ops[1]->getType() }; |
| 12157 | Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld4lane, Tys); |
| 12158 | std::rotate(Ops.begin() + 1, Ops.begin() + 2, Ops.end()); |
| 12159 | Ops[1] = Builder.CreateBitCast(Ops[1], Ty); |
| 12160 | Ops[2] = Builder.CreateBitCast(Ops[2], Ty); |
| 12161 | Ops[3] = Builder.CreateBitCast(Ops[3], Ty); |
| 12162 | Ops[4] = Builder.CreateBitCast(Ops[4], Ty); |
| 12163 | Ops[5] = Builder.CreateZExt(Ops[5], Int64Ty); |
| 12164 | Ops[1] = Builder.CreateCall(F, ArrayRef(Ops).slice(1), "vld4_lane"); |
| 12165 | Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); |
| 12166 | Ops[0] = Builder.CreateBitCast(Ops[0], Ty); |
| 12167 | return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]); |
| 12168 | } |
| 12169 | case NEON::BI__builtin_neon_vst2_v: |
| 12170 | case NEON::BI__builtin_neon_vst2q_v: { |
| 12171 | std::rotate(Ops.begin(), Ops.begin() + 1, Ops.end()); |
| 12172 | llvm::Type *Tys[2] = { VTy, Ops[2]->getType() }; |
| 12173 | return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_st2, Tys), |
| 12174 | Ops, ""); |
| 12175 | } |
| 12176 | case NEON::BI__builtin_neon_vst2_lane_v: |
| 12177 | case NEON::BI__builtin_neon_vst2q_lane_v: { |
| 12178 | std::rotate(Ops.begin(), Ops.begin() + 1, Ops.end()); |
| 12179 | Ops[2] = Builder.CreateZExt(Ops[2], Int64Ty); |
| 12180 | llvm::Type *Tys[2] = { VTy, Ops[3]->getType() }; |
| 12181 | return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_st2lane, Tys), |
| 12182 | Ops, ""); |
| 12183 | } |
| 12184 | case NEON::BI__builtin_neon_vst3_v: |
| 12185 | case NEON::BI__builtin_neon_vst3q_v: { |
| 12186 | std::rotate(Ops.begin(), Ops.begin() + 1, Ops.end()); |
| 12187 | llvm::Type *Tys[2] = { VTy, Ops[3]->getType() }; |
| 12188 | return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_st3, Tys), |
| 12189 | Ops, ""); |
| 12190 | } |
| 12191 | case NEON::BI__builtin_neon_vst3_lane_v: |
| 12192 | case NEON::BI__builtin_neon_vst3q_lane_v: { |
| 12193 | std::rotate(Ops.begin(), Ops.begin() + 1, Ops.end()); |
| 12194 | Ops[3] = Builder.CreateZExt(Ops[3], Int64Ty); |
| 12195 | llvm::Type *Tys[2] = { VTy, Ops[4]->getType() }; |
| 12196 | return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_st3lane, Tys), |
| 12197 | Ops, ""); |
| 12198 | } |
| 12199 | case NEON::BI__builtin_neon_vst4_v: |
| 12200 | case NEON::BI__builtin_neon_vst4q_v: { |
| 12201 | std::rotate(Ops.begin(), Ops.begin() + 1, Ops.end()); |
| 12202 | llvm::Type *Tys[2] = { VTy, Ops[4]->getType() }; |
| 12203 | return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_st4, Tys), |
| 12204 | Ops, ""); |
| 12205 | } |
| 12206 | case NEON::BI__builtin_neon_vst4_lane_v: |
| 12207 | case NEON::BI__builtin_neon_vst4q_lane_v: { |
| 12208 | std::rotate(Ops.begin(), Ops.begin() + 1, Ops.end()); |
| 12209 | Ops[4] = Builder.CreateZExt(Ops[4], Int64Ty); |
| 12210 | llvm::Type *Tys[2] = { VTy, Ops[5]->getType() }; |
| 12211 | return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_st4lane, Tys), |
| 12212 | Ops, ""); |
| 12213 | } |
| 12214 | case NEON::BI__builtin_neon_vtrn_v: |
| 12215 | case NEON::BI__builtin_neon_vtrnq_v: { |
| 12216 | Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty)); |
| 12217 | Ops[1] = Builder.CreateBitCast(Ops[1], Ty); |
| 12218 | Ops[2] = Builder.CreateBitCast(Ops[2], Ty); |
| 12219 | Value *SV = nullptr; |
| 12220 | |
| 12221 | for (unsigned vi = 0; vi != 2; ++vi) { |
| 12222 | SmallVector<int, 16> Indices; |
| 12223 | for (unsigned i = 0, e = VTy->getNumElements(); i != e; i += 2) { |
| 12224 | Indices.push_back(i+vi); |
| 12225 | Indices.push_back(i+e+vi); |
| 12226 | } |
| 12227 | Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ty, Ops[0], vi); |
| 12228 | SV = Builder.CreateShuffleVector(Ops[1], Ops[2], Indices, "vtrn"); |
| 12229 | SV = Builder.CreateDefaultAlignedStore(SV, Addr); |
| 12230 | } |
| 12231 | return SV; |
| 12232 | } |
| 12233 | case NEON::BI__builtin_neon_vuzp_v: |
| 12234 | case NEON::BI__builtin_neon_vuzpq_v: { |
| 12235 | Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty)); |
| 12236 | Ops[1] = Builder.CreateBitCast(Ops[1], Ty); |
| 12237 | Ops[2] = Builder.CreateBitCast(Ops[2], Ty); |
| 12238 | Value *SV = nullptr; |
| 12239 | |
| 12240 | for (unsigned vi = 0; vi != 2; ++vi) { |
| 12241 | SmallVector<int, 16> Indices; |
| 12242 | for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) |
| 12243 | Indices.push_back(2*i+vi); |
| 12244 | |
| 12245 | Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ty, Ops[0], vi); |
| 12246 | SV = Builder.CreateShuffleVector(Ops[1], Ops[2], Indices, "vuzp"); |
| 12247 | SV = Builder.CreateDefaultAlignedStore(SV, Addr); |
| 12248 | } |
| 12249 | return SV; |
| 12250 | } |
| 12251 | case NEON::BI__builtin_neon_vzip_v: |
| 12252 | case NEON::BI__builtin_neon_vzipq_v: { |
| 12253 | Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty)); |
| 12254 | Ops[1] = Builder.CreateBitCast(Ops[1], Ty); |
| 12255 | Ops[2] = Builder.CreateBitCast(Ops[2], Ty); |
| 12256 | Value *SV = nullptr; |
| 12257 | |
| 12258 | for (unsigned vi = 0; vi != 2; ++vi) { |
| 12259 | SmallVector<int, 16> Indices; |
| 12260 | for (unsigned i = 0, e = VTy->getNumElements(); i != e; i += 2) { |
| 12261 | Indices.push_back((i + vi*e) >> 1); |
| 12262 | Indices.push_back(((i + vi*e) >> 1)+e); |
| 12263 | } |
| 12264 | Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ty, Ops[0], vi); |
| 12265 | SV = Builder.CreateShuffleVector(Ops[1], Ops[2], Indices, "vzip"); |
| 12266 | SV = Builder.CreateDefaultAlignedStore(SV, Addr); |
| 12267 | } |
| 12268 | return SV; |
| 12269 | } |
| 12270 | case NEON::BI__builtin_neon_vqtbl1q_v: { |
| 12271 | return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbl1, Ty), |
| 12272 | Ops, "vtbl1"); |
| 12273 | } |
| 12274 | case NEON::BI__builtin_neon_vqtbl2q_v: { |
| 12275 | return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbl2, Ty), |
| 12276 | Ops, "vtbl2"); |
| 12277 | } |
| 12278 | case NEON::BI__builtin_neon_vqtbl3q_v: { |
| 12279 | return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbl3, Ty), |
| 12280 | Ops, "vtbl3"); |
| 12281 | } |
| 12282 | case NEON::BI__builtin_neon_vqtbl4q_v: { |
| 12283 | return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbl4, Ty), |
| 12284 | Ops, "vtbl4"); |
| 12285 | } |
| 12286 | case NEON::BI__builtin_neon_vqtbx1q_v: { |
| 12287 | return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbx1, Ty), |
| 12288 | Ops, "vtbx1"); |
| 12289 | } |
| 12290 | case NEON::BI__builtin_neon_vqtbx2q_v: { |
| 12291 | return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbx2, Ty), |
| 12292 | Ops, "vtbx2"); |
| 12293 | } |
| 12294 | case NEON::BI__builtin_neon_vqtbx3q_v: { |
| 12295 | return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbx3, Ty), |
| 12296 | Ops, "vtbx3"); |
| 12297 | } |
| 12298 | case NEON::BI__builtin_neon_vqtbx4q_v: { |
| 12299 | return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbx4, Ty), |
| 12300 | Ops, "vtbx4"); |
| 12301 | } |
| 12302 | case NEON::BI__builtin_neon_vsqadd_v: |
| 12303 | case NEON::BI__builtin_neon_vsqaddq_v: { |
| 12304 | Int = Intrinsic::aarch64_neon_usqadd; |
| 12305 | return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vsqadd"); |
| 12306 | } |
| 12307 | case NEON::BI__builtin_neon_vuqadd_v: |
| 12308 | case NEON::BI__builtin_neon_vuqaddq_v: { |
| 12309 | Int = Intrinsic::aarch64_neon_suqadd; |
| 12310 | return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vuqadd"); |
| 12311 | } |
| 12312 | } |
| 12313 | } |
| 12314 | |
| 12315 | Value *CodeGenFunction::EmitBPFBuiltinExpr(unsigned BuiltinID, |
| 12316 | const CallExpr *E) { |
| 12317 | assert((BuiltinID == BPF::BI__builtin_preserve_field_info ||(static_cast <bool> ((BuiltinID == BPF::BI__builtin_preserve_field_info || BuiltinID == BPF::BI__builtin_btf_type_id || BuiltinID == BPF::BI__builtin_preserve_type_info || BuiltinID == BPF::BI__builtin_preserve_enum_value ) && "unexpected BPF builtin") ? void (0) : __assert_fail ("(BuiltinID == BPF::BI__builtin_preserve_field_info || BuiltinID == BPF::BI__builtin_btf_type_id || BuiltinID == BPF::BI__builtin_preserve_type_info || BuiltinID == BPF::BI__builtin_preserve_enum_value) && \"unexpected BPF builtin\"" , "clang/lib/CodeGen/CGBuiltin.cpp", 12321, __extension__ __PRETTY_FUNCTION__ )) |
| 12318 | BuiltinID == BPF::BI__builtin_btf_type_id ||(static_cast <bool> ((BuiltinID == BPF::BI__builtin_preserve_field_info || BuiltinID == BPF::BI__builtin_btf_type_id || BuiltinID == BPF::BI__builtin_preserve_type_info || BuiltinID == BPF::BI__builtin_preserve_enum_value ) && "unexpected BPF builtin") ? void (0) : __assert_fail ("(BuiltinID == BPF::BI__builtin_preserve_field_info || BuiltinID == BPF::BI__builtin_btf_type_id || BuiltinID == BPF::BI__builtin_preserve_type_info || BuiltinID == BPF::BI__builtin_preserve_enum_value) && \"unexpected BPF builtin\"" , "clang/lib/CodeGen/CGBuiltin.cpp", 12321, __extension__ __PRETTY_FUNCTION__ )) |
| 12319 | BuiltinID == BPF::BI__builtin_preserve_type_info ||(static_cast <bool> ((BuiltinID == BPF::BI__builtin_preserve_field_info || BuiltinID == BPF::BI__builtin_btf_type_id || BuiltinID == BPF::BI__builtin_preserve_type_info || BuiltinID == BPF::BI__builtin_preserve_enum_value ) && "unexpected BPF builtin") ? void (0) : __assert_fail ("(BuiltinID == BPF::BI__builtin_preserve_field_info || BuiltinID == BPF::BI__builtin_btf_type_id || BuiltinID == BPF::BI__builtin_preserve_type_info || BuiltinID == BPF::BI__builtin_preserve_enum_value) && \"unexpected BPF builtin\"" , "clang/lib/CodeGen/CGBuiltin.cpp", 12321, __extension__ __PRETTY_FUNCTION__ )) |
| 12320 | BuiltinID == BPF::BI__builtin_preserve_enum_value) &&(static_cast <bool> ((BuiltinID == BPF::BI__builtin_preserve_field_info || BuiltinID == BPF::BI__builtin_btf_type_id || BuiltinID == BPF::BI__builtin_preserve_type_info || BuiltinID == BPF::BI__builtin_preserve_enum_value ) && "unexpected BPF builtin") ? void (0) : __assert_fail ("(BuiltinID == BPF::BI__builtin_preserve_field_info || BuiltinID == BPF::BI__builtin_btf_type_id || BuiltinID == BPF::BI__builtin_preserve_type_info || BuiltinID == BPF::BI__builtin_preserve_enum_value) && \"unexpected BPF builtin\"" , "clang/lib/CodeGen/CGBuiltin.cpp", 12321, __extension__ __PRETTY_FUNCTION__ )) |
| 12321 | "unexpected BPF builtin")(static_cast <bool> ((BuiltinID == BPF::BI__builtin_preserve_field_info || BuiltinID == BPF::BI__builtin_btf_type_id || BuiltinID == BPF::BI__builtin_preserve_type_info || BuiltinID == BPF::BI__builtin_preserve_enum_value ) && "unexpected BPF builtin") ? void (0) : __assert_fail ("(BuiltinID == BPF::BI__builtin_preserve_field_info || BuiltinID == BPF::BI__builtin_btf_type_id || BuiltinID == BPF::BI__builtin_preserve_type_info || BuiltinID == BPF::BI__builtin_preserve_enum_value) && \"unexpected BPF builtin\"" , "clang/lib/CodeGen/CGBuiltin.cpp", 12321, __extension__ __PRETTY_FUNCTION__ )); |
| 12322 | |
| 12323 | // A sequence number, injected into IR builtin functions, to |
| 12324 | // prevent CSE given the only difference of the function |
| 12325 | // may just be the debuginfo metadata. |
| 12326 | static uint32_t BuiltinSeqNum; |
| 12327 | |
| 12328 | switch (BuiltinID) { |
| 12329 | default: |
| 12330 | llvm_unreachable("Unexpected BPF builtin")::llvm::llvm_unreachable_internal("Unexpected BPF builtin", "clang/lib/CodeGen/CGBuiltin.cpp" , 12330); |
| 12331 | case BPF::BI__builtin_preserve_field_info: { |
| 12332 | const Expr *Arg = E->getArg(0); |
| 12333 | bool IsBitField = Arg->IgnoreParens()->getObjectKind() == OK_BitField; |
| 12334 | |
| 12335 | if (!getDebugInfo()) { |
| 12336 | CGM.Error(E->getExprLoc(), |
| 12337 | "using __builtin_preserve_field_info() without -g"); |
| 12338 | return IsBitField ? EmitLValue(Arg).getBitFieldPointer() |
| 12339 | : EmitLValue(Arg).getPointer(*this); |
| 12340 | } |
| 12341 | |
| 12342 | // Enable underlying preserve_*_access_index() generation. |
| 12343 | bool OldIsInPreservedAIRegion = IsInPreservedAIRegion; |
| 12344 | IsInPreservedAIRegion = true; |
| 12345 | Value *FieldAddr = IsBitField ? EmitLValue(Arg).getBitFieldPointer() |
| 12346 | : EmitLValue(Arg).getPointer(*this); |
| 12347 | IsInPreservedAIRegion = OldIsInPreservedAIRegion; |
| 12348 | |
| 12349 | ConstantInt *C = cast<ConstantInt>(EmitScalarExpr(E->getArg(1))); |
| 12350 | Value *InfoKind = ConstantInt::get(Int64Ty, C->getSExtValue()); |
| 12351 | |
| 12352 | // Built the IR for the preserve_field_info intrinsic. |
| 12353 | llvm::Function *FnGetFieldInfo = llvm::Intrinsic::getDeclaration( |
| 12354 | &CGM.getModule(), llvm::Intrinsic::bpf_preserve_field_info, |
| 12355 | {FieldAddr->getType()}); |
| 12356 | return Builder.CreateCall(FnGetFieldInfo, {FieldAddr, InfoKind}); |
| 12357 | } |
| 12358 | case BPF::BI__builtin_btf_type_id: |
| 12359 | case BPF::BI__builtin_preserve_type_info: { |
| 12360 | if (!getDebugInfo()) { |
| 12361 | CGM.Error(E->getExprLoc(), "using builtin function without -g"); |
| 12362 | return nullptr; |
| 12363 | } |
| 12364 | |
| 12365 | const Expr *Arg0 = E->getArg(0); |
| 12366 | llvm::DIType *DbgInfo = getDebugInfo()->getOrCreateStandaloneType( |
| 12367 | Arg0->getType(), Arg0->getExprLoc()); |
| 12368 | |
| 12369 | ConstantInt *Flag = cast<ConstantInt>(EmitScalarExpr(E->getArg(1))); |
| 12370 | Value *FlagValue = ConstantInt::get(Int64Ty, Flag->getSExtValue()); |
| 12371 | Value *SeqNumVal = ConstantInt::get(Int32Ty, BuiltinSeqNum++); |
| 12372 | |
| 12373 | llvm::Function *FnDecl; |
| 12374 | if (BuiltinID == BPF::BI__builtin_btf_type_id) |
| 12375 | FnDecl = llvm::Intrinsic::getDeclaration( |
| 12376 | &CGM.getModule(), llvm::Intrinsic::bpf_btf_type_id, {}); |
| 12377 | else |
| 12378 | FnDecl = llvm::Intrinsic::getDeclaration( |
| 12379 | &CGM.getModule(), llvm::Intrinsic::bpf_preserve_type_info, {}); |
| 12380 | CallInst *Fn = Builder.CreateCall(FnDecl, {SeqNumVal, FlagValue}); |
| 12381 | Fn->setMetadata(LLVMContext::MD_preserve_access_index, DbgInfo); |
| 12382 | return Fn; |
| 12383 | } |
| 12384 | case BPF::BI__builtin_preserve_enum_value: { |
| 12385 | if (!getDebugInfo()) { |
| 12386 | CGM.Error(E->getExprLoc(), "using builtin function without -g"); |
| 12387 | return nullptr; |
| 12388 | } |
| 12389 | |
| 12390 | const Expr *Arg0 = E->getArg(0); |
| 12391 | llvm::DIType *DbgInfo = getDebugInfo()->getOrCreateStandaloneType( |
| 12392 | Arg0->getType(), Arg0->getExprLoc()); |
| 12393 | |
| 12394 | // Find enumerator |
| 12395 | const auto *UO = cast<UnaryOperator>(Arg0->IgnoreParens()); |
| 12396 | const auto *CE = cast<CStyleCastExpr>(UO->getSubExpr()); |
| 12397 | const auto *DR = cast<DeclRefExpr>(CE->getSubExpr()); |
| 12398 | const auto *Enumerator = cast<EnumConstantDecl>(DR->getDecl()); |
| 12399 | |
| 12400 | auto &InitVal = Enumerator->getInitVal(); |
| 12401 | std::string InitValStr; |
| 12402 | if (InitVal.isNegative() || InitVal > uint64_t(INT64_MAX(9223372036854775807L))) |
| 12403 | InitValStr = std::to_string(InitVal.getSExtValue()); |
| 12404 | else |
| 12405 | InitValStr = std::to_string(InitVal.getZExtValue()); |
| 12406 | std::string EnumStr = Enumerator->getNameAsString() + ":" + InitValStr; |
| 12407 | Value *EnumStrVal = Builder.CreateGlobalStringPtr(EnumStr); |
| 12408 | |
| 12409 | ConstantInt *Flag = cast<ConstantInt>(EmitScalarExpr(E->getArg(1))); |
| 12410 | Value *FlagValue = ConstantInt::get(Int64Ty, Flag->getSExtValue()); |
| 12411 | Value *SeqNumVal = ConstantInt::get(Int32Ty, BuiltinSeqNum++); |
| 12412 | |
| 12413 | llvm::Function *IntrinsicFn = llvm::Intrinsic::getDeclaration( |
| 12414 | &CGM.getModule(), llvm::Intrinsic::bpf_preserve_enum_value, {}); |
| 12415 | CallInst *Fn = |
| 12416 | Builder.CreateCall(IntrinsicFn, {SeqNumVal, EnumStrVal, FlagValue}); |
| 12417 | Fn->setMetadata(LLVMContext::MD_preserve_access_index, DbgInfo); |
| 12418 | return Fn; |
| 12419 | } |
| 12420 | } |
| 12421 | } |
| 12422 | |
| 12423 | llvm::Value *CodeGenFunction:: |
| 12424 | BuildVector(ArrayRef<llvm::Value*> Ops) { |
| 12425 | assert((Ops.size() & (Ops.size() - 1)) == 0 &&(static_cast <bool> ((Ops.size() & (Ops.size() - 1) ) == 0 && "Not a power-of-two sized vector!") ? void ( 0) : __assert_fail ("(Ops.size() & (Ops.size() - 1)) == 0 && \"Not a power-of-two sized vector!\"" , "clang/lib/CodeGen/CGBuiltin.cpp", 12426, __extension__ __PRETTY_FUNCTION__ )) |
| 12426 | "Not a power-of-two sized vector!")(static_cast <bool> ((Ops.size() & (Ops.size() - 1) ) == 0 && "Not a power-of-two sized vector!") ? void ( 0) : __assert_fail ("(Ops.size() & (Ops.size() - 1)) == 0 && \"Not a power-of-two sized vector!\"" , "clang/lib/CodeGen/CGBuiltin.cpp", 12426, __extension__ __PRETTY_FUNCTION__ )); |
| 12427 | bool AllConstants = true; |
| 12428 | for (unsigned i = 0, e = Ops.size(); i != e && AllConstants; ++i) |
| 12429 | AllConstants &= isa<Constant>(Ops[i]); |
| 12430 | |
| 12431 | // If this is a constant vector, create a ConstantVector. |
| 12432 | if (AllConstants) { |
| 12433 | SmallVector<llvm::Constant*, 16> CstOps; |
| 12434 | for (unsigned i = 0, e = Ops.size(); i != e; ++i) |
| 12435 | CstOps.push_back(cast<Constant>(Ops[i])); |
| 12436 | return llvm::ConstantVector::get(CstOps); |
| 12437 | } |
| 12438 | |
| 12439 | // Otherwise, insertelement the values to build the vector. |
| 12440 | Value *Result = llvm::PoisonValue::get( |
| 12441 | llvm::FixedVectorType::get(Ops[0]->getType(), Ops.size())); |
| 12442 | |
| 12443 | for (unsigned i = 0, e = Ops.size(); i != e; ++i) |
| 12444 | Result = Builder.CreateInsertElement(Result, Ops[i], Builder.getInt64(i)); |
| 12445 | |
| 12446 | return Result; |
| 12447 | } |
| 12448 | |
| 12449 | // Convert the mask from an integer type to a vector of i1. |
| 12450 | static Value *getMaskVecValue(CodeGenFunction &CGF, Value *Mask, |
| 12451 | unsigned NumElts) { |
| 12452 | |
| 12453 | auto *MaskTy = llvm::FixedVectorType::get( |
| 12454 | CGF.Builder.getInt1Ty(), |
| 12455 | cast<IntegerType>(Mask->getType())->getBitWidth()); |
| 12456 | Value *MaskVec = CGF.Builder.CreateBitCast(Mask, MaskTy); |
| 12457 | |
| 12458 | // If we have less than 8 elements, then the starting mask was an i8 and |
| 12459 | // we need to extract down to the right number of elements. |
| 12460 | if (NumElts < 8) { |
| 12461 | int Indices[4]; |
| 12462 | for (unsigned i = 0; i != NumElts; ++i) |
| 12463 | Indices[i] = i; |
| 12464 | MaskVec = CGF.Builder.CreateShuffleVector( |
| 12465 | MaskVec, MaskVec, ArrayRef(Indices, NumElts), "extract"); |
| 12466 | } |
| 12467 | return MaskVec; |
| 12468 | } |
| 12469 | |
| 12470 | static Value *EmitX86MaskedStore(CodeGenFunction &CGF, ArrayRef<Value *> Ops, |
| 12471 | Align Alignment) { |
| 12472 | // Cast the pointer to right type. |
| 12473 | Value *Ptr = CGF.Builder.CreateBitCast(Ops[0], |
| 12474 | llvm::PointerType::getUnqual(Ops[1]->getType())); |
| 12475 | |
| 12476 | Value *MaskVec = getMaskVecValue( |
| 12477 | CGF, Ops[2], |
| 12478 | cast<llvm::FixedVectorType>(Ops[1]->getType())->getNumElements()); |
| 12479 | |
| 12480 | return CGF.Builder.CreateMaskedStore(Ops[1], Ptr, Alignment, MaskVec); |
| 12481 | } |
| 12482 | |
| 12483 | static Value *EmitX86MaskedLoad(CodeGenFunction &CGF, ArrayRef<Value *> Ops, |
| 12484 | Align Alignment) { |
| 12485 | // Cast the pointer to right type. |
| 12486 | llvm::Type *Ty = Ops[1]->getType(); |
| 12487 | Value *Ptr = |
| 12488 | CGF.Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty)); |
| 12489 | |
| 12490 | Value *MaskVec = getMaskVecValue( |
| 12491 | CGF, Ops[2], cast<llvm::FixedVectorType>(Ty)->getNumElements()); |
| 12492 | |
| 12493 | return CGF.Builder.CreateMaskedLoad(Ty, Ptr, Alignment, MaskVec, Ops[1]); |
| 12494 | } |
| 12495 | |
| 12496 | static Value *EmitX86ExpandLoad(CodeGenFunction &CGF, |
| 12497 | ArrayRef<Value *> Ops) { |
| 12498 | auto *ResultTy = cast<llvm::VectorType>(Ops[1]->getType()); |
| 12499 | llvm::Type *PtrTy = ResultTy->getElementType(); |
| 12500 | |
| 12501 | // Cast the pointer to element type. |
| 12502 | Value *Ptr = CGF.Builder.CreateBitCast(Ops[0], |
| 12503 | llvm::PointerType::getUnqual(PtrTy)); |
| 12504 | |
| 12505 | Value *MaskVec = getMaskVecValue( |
| 12506 | CGF, Ops[2], cast<FixedVectorType>(ResultTy)->getNumElements()); |
| 12507 | |
| 12508 | llvm::Function *F = CGF.CGM.getIntrinsic(Intrinsic::masked_expandload, |
| 12509 | ResultTy); |
| 12510 | return CGF.Builder.CreateCall(F, { Ptr, MaskVec, Ops[1] }); |
| 12511 | } |
| 12512 | |
| 12513 | static Value *EmitX86CompressExpand(CodeGenFunction &CGF, |
| 12514 | ArrayRef<Value *> Ops, |
| 12515 | bool IsCompress) { |
| 12516 | auto *ResultTy = cast<llvm::FixedVectorType>(Ops[1]->getType()); |
| 12517 | |
| 12518 | Value *MaskVec = getMaskVecValue(CGF, Ops[2], ResultTy->getNumElements()); |
| 12519 | |
| 12520 | Intrinsic::ID IID = IsCompress ? Intrinsic::x86_avx512_mask_compress |
| 12521 | : Intrinsic::x86_avx512_mask_expand; |
| 12522 | llvm::Function *F = CGF.CGM.getIntrinsic(IID, ResultTy); |
| 12523 | return CGF.Builder.CreateCall(F, { Ops[0], Ops[1], MaskVec }); |
| 12524 | } |
| 12525 | |
| 12526 | static Value *EmitX86CompressStore(CodeGenFunction &CGF, |
| 12527 | ArrayRef<Value *> Ops) { |
| 12528 | auto *ResultTy = cast<llvm::FixedVectorType>(Ops[1]->getType()); |
| 12529 | llvm::Type *PtrTy = ResultTy->getElementType(); |
| 12530 | |
| 12531 | // Cast the pointer to element type. |
| 12532 | Value *Ptr = CGF.Builder.CreateBitCast(Ops[0], |
| 12533 | llvm::PointerType::getUnqual(PtrTy)); |
| 12534 | |
| 12535 | Value *MaskVec = getMaskVecValue(CGF, Ops[2], ResultTy->getNumElements()); |
| 12536 | |
| 12537 | llvm::Function *F = CGF.CGM.getIntrinsic(Intrinsic::masked_compressstore, |
| 12538 | ResultTy); |
| 12539 | return CGF.Builder.CreateCall(F, { Ops[1], Ptr, MaskVec }); |
| 12540 | } |
| 12541 | |
| 12542 | static Value *EmitX86MaskLogic(CodeGenFunction &CGF, Instruction::BinaryOps Opc, |
| 12543 | ArrayRef<Value *> Ops, |
| 12544 | bool InvertLHS = false) { |
| 12545 | unsigned NumElts = Ops[0]->getType()->getIntegerBitWidth(); |
| 12546 | Value *LHS = getMaskVecValue(CGF, Ops[0], NumElts); |
| 12547 | Value *RHS = getMaskVecValue(CGF, Ops[1], NumElts); |
| 12548 | |
| 12549 | if (InvertLHS) |
| 12550 | LHS = CGF.Builder.CreateNot(LHS); |
| 12551 | |
| 12552 | return CGF.Builder.CreateBitCast(CGF.Builder.CreateBinOp(Opc, LHS, RHS), |
| 12553 | Ops[0]->getType()); |
| 12554 | } |
| 12555 | |
| 12556 | static Value *EmitX86FunnelShift(CodeGenFunction &CGF, Value *Op0, Value *Op1, |
| 12557 | Value *Amt, bool IsRight) { |
| 12558 | llvm::Type *Ty = Op0->getType(); |
| 12559 | |
| 12560 | // Amount may be scalar immediate, in which case create a splat vector. |
| 12561 | // Funnel shifts amounts are treated as modulo and types are all power-of-2 so |
| 12562 | // we only care about the lowest log2 bits anyway. |
| 12563 | if (Amt->getType() != Ty) { |
| 12564 | unsigned NumElts = cast<llvm::FixedVectorType>(Ty)->getNumElements(); |
| 12565 | Amt = CGF.Builder.CreateIntCast(Amt, Ty->getScalarType(), false); |
| 12566 | Amt = CGF.Builder.CreateVectorSplat(NumElts, Amt); |
| 12567 | } |
| 12568 | |
| 12569 | unsigned IID = IsRight ? Intrinsic::fshr : Intrinsic::fshl; |
| 12570 | Function *F = CGF.CGM.getIntrinsic(IID, Ty); |
| 12571 | return CGF.Builder.CreateCall(F, {Op0, Op1, Amt}); |
| 12572 | } |
| 12573 | |
| 12574 | static Value *EmitX86vpcom(CodeGenFunction &CGF, ArrayRef<Value *> Ops, |
| 12575 | bool IsSigned) { |
| 12576 | Value *Op0 = Ops[0]; |
| 12577 | Value *Op1 = Ops[1]; |
| 12578 | llvm::Type *Ty = Op0->getType(); |
| 12579 | uint64_t Imm = cast<llvm::ConstantInt>(Ops[2])->getZExtValue() & 0x7; |
| 12580 | |
| 12581 | CmpInst::Predicate Pred; |
| 12582 | switch (Imm) { |
| 12583 | case 0x0: |
| 12584 | Pred = IsSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT; |
| 12585 | break; |
| 12586 | case 0x1: |
| 12587 | Pred = IsSigned ? ICmpInst::ICMP_SLE : ICmpInst::ICMP_ULE; |
| 12588 | break; |
| 12589 | case 0x2: |
| 12590 | Pred = IsSigned ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT; |
| 12591 | break; |
| 12592 | case 0x3: |
| 12593 | Pred = IsSigned ? ICmpInst::ICMP_SGE : ICmpInst::ICMP_UGE; |
| 12594 | break; |
| 12595 | case 0x4: |
| 12596 | Pred = ICmpInst::ICMP_EQ; |
| 12597 | break; |
| 12598 | case 0x5: |
| 12599 | Pred = ICmpInst::ICMP_NE; |
| 12600 | break; |
| 12601 | case 0x6: |
| 12602 | return llvm::Constant::getNullValue(Ty); // FALSE |
| 12603 | case 0x7: |
| 12604 | return llvm::Constant::getAllOnesValue(Ty); // TRUE |
| 12605 | default: |
| 12606 | llvm_unreachable("Unexpected XOP vpcom/vpcomu predicate")::llvm::llvm_unreachable_internal("Unexpected XOP vpcom/vpcomu predicate" , "clang/lib/CodeGen/CGBuiltin.cpp", 12606); |
| 12607 | } |
| 12608 | |
| 12609 | Value *Cmp = CGF.Builder.CreateICmp(Pred, Op0, Op1); |
| 12610 | Value *Res = CGF.Builder.CreateSExt(Cmp, Ty); |
| 12611 | return Res; |
| 12612 | } |
| 12613 | |
| 12614 | static Value *EmitX86Select(CodeGenFunction &CGF, |
| 12615 | Value *Mask, Value *Op0, Value *Op1) { |
| 12616 | |
| 12617 | // If the mask is all ones just return first argument. |
| 12618 | if (const auto *C = dyn_cast<Constant>(Mask)) |
| 12619 | if (C->isAllOnesValue()) |
| 12620 | return Op0; |
| 12621 | |
| 12622 | Mask = getMaskVecValue( |
| 12623 | CGF, Mask, cast<llvm::FixedVectorType>(Op0->getType())->getNumElements()); |
| 12624 | |
| 12625 | return CGF.Builder.CreateSelect(Mask, Op0, Op1); |
| 12626 | } |
| 12627 | |
| 12628 | static Value *EmitX86ScalarSelect(CodeGenFunction &CGF, |
| 12629 | Value *Mask, Value *Op0, Value *Op1) { |
| 12630 | // If the mask is all ones just return first argument. |
| 12631 | if (const auto *C = dyn_cast<Constant>(Mask)) |
| 12632 | if (C->isAllOnesValue()) |
| 12633 | return Op0; |
| 12634 | |
| 12635 | auto *MaskTy = llvm::FixedVectorType::get( |
| 12636 | CGF.Builder.getInt1Ty(), Mask->getType()->getIntegerBitWidth()); |
| 12637 | Mask = CGF.Builder.CreateBitCast(Mask, MaskTy); |
| 12638 | Mask = CGF.Builder.CreateExtractElement(Mask, (uint64_t)0); |
| 12639 | return CGF.Builder.CreateSelect(Mask, Op0, Op1); |
| 12640 | } |
| 12641 | |
| 12642 | static Value *EmitX86MaskedCompareResult(CodeGenFunction &CGF, Value *Cmp, |
| 12643 | unsigned NumElts, Value *MaskIn) { |
| 12644 | if (MaskIn) { |
| 12645 | const auto *C = dyn_cast<Constant>(MaskIn); |
| 12646 | if (!C || !C->isAllOnesValue()) |
| 12647 | Cmp = CGF.Builder.CreateAnd(Cmp, getMaskVecValue(CGF, MaskIn, NumElts)); |
| 12648 | } |
| 12649 | |
| 12650 | if (NumElts < 8) { |
| 12651 | int Indices[8]; |
| 12652 | for (unsigned i = 0; i != NumElts; ++i) |
| 12653 | Indices[i] = i; |
| 12654 | for (unsigned i = NumElts; i != 8; ++i) |
| 12655 | Indices[i] = i % NumElts + NumElts; |
| 12656 | Cmp = CGF.Builder.CreateShuffleVector( |
| 12657 | Cmp, llvm::Constant::getNullValue(Cmp->getType()), Indices); |
| 12658 | } |
| 12659 | |
| 12660 | return CGF.Builder.CreateBitCast(Cmp, |
| 12661 | IntegerType::get(CGF.getLLVMContext(), |
| 12662 | std::max(NumElts, 8U))); |
| 12663 | } |
| 12664 | |
| 12665 | static Value *EmitX86MaskedCompare(CodeGenFunction &CGF, unsigned CC, |
| 12666 | bool Signed, ArrayRef<Value *> Ops) { |
| 12667 | assert((Ops.size() == 2 || Ops.size() == 4) &&(static_cast <bool> ((Ops.size() == 2 || Ops.size() == 4 ) && "Unexpected number of arguments") ? void (0) : __assert_fail ("(Ops.size() == 2 || Ops.size() == 4) && \"Unexpected number of arguments\"" , "clang/lib/CodeGen/CGBuiltin.cpp", 12668, __extension__ __PRETTY_FUNCTION__ )) |
| 12668 | "Unexpected number of arguments")(static_cast <bool> ((Ops.size() == 2 || Ops.size() == 4 ) && "Unexpected number of arguments") ? void (0) : __assert_fail ("(Ops.size() == 2 || Ops.size() == 4) && \"Unexpected number of arguments\"" , "clang/lib/CodeGen/CGBuiltin.cpp", 12668, __extension__ __PRETTY_FUNCTION__ )); |
| 12669 | unsigned NumElts = |
| 12670 | cast<llvm::FixedVectorType>(Ops[0]->getType())->getNumElements(); |
| 12671 | Value *Cmp; |
| 12672 | |
| 12673 | if (CC == 3) { |
| 12674 | Cmp = Constant::getNullValue( |
| 12675 | llvm::FixedVectorType::get(CGF.Builder.getInt1Ty(), NumElts)); |
| 12676 | } else if (CC == 7) { |
| 12677 | Cmp = Constant::getAllOnesValue( |
| 12678 | llvm::FixedVectorType::get(CGF.Builder.getInt1Ty(), NumElts)); |
| 12679 | } else { |
| 12680 | ICmpInst::Predicate Pred; |
| 12681 | switch (CC) { |
| 12682 | default: llvm_unreachable("Unknown condition code")::llvm::llvm_unreachable_internal("Unknown condition code", "clang/lib/CodeGen/CGBuiltin.cpp" , 12682); |
| 12683 | case 0: Pred = ICmpInst::ICMP_EQ; break; |
| 12684 | case 1: Pred = Signed ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT; break; |
| 12685 | case 2: Pred = Signed ? ICmpInst::ICMP_SLE : ICmpInst::ICMP_ULE; break; |
| 12686 | case 4: Pred = ICmpInst::ICMP_NE; break; |
| 12687 | case 5: Pred = Signed ? ICmpInst::ICMP_SGE : ICmpInst::ICMP_UGE; break; |
| 12688 | case 6: Pred = Signed ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT; break; |
| 12689 | } |
| 12690 | Cmp = CGF.Builder.CreateICmp(Pred, Ops[0], Ops[1]); |
| 12691 | } |
| 12692 | |
| 12693 | Value *MaskIn = nullptr; |
| 12694 | if (Ops.size() == 4) |
| 12695 | MaskIn = Ops[3]; |
| 12696 | |
| 12697 | return EmitX86MaskedCompareResult(CGF, Cmp, NumElts, MaskIn); |
| 12698 | } |
| 12699 | |
| 12700 | static Value *EmitX86ConvertToMask(CodeGenFunction &CGF, Value *In) { |
| 12701 | Value *Zero = Constant::getNullValue(In->getType()); |
| 12702 | return EmitX86MaskedCompare(CGF, 1, true, { In, Zero }); |
| 12703 | } |
| 12704 | |
| 12705 | static Value *EmitX86ConvertIntToFp(CodeGenFunction &CGF, const CallExpr *E, |
| 12706 | ArrayRef<Value *> Ops, bool IsSigned) { |
| 12707 | unsigned Rnd = cast<llvm::ConstantInt>(Ops[3])->getZExtValue(); |
| 12708 | llvm::Type *Ty = Ops[1]->getType(); |
| 12709 | |
| 12710 | Value *Res; |
| 12711 | if (Rnd != 4) { |
| 12712 | Intrinsic::ID IID = IsSigned ? Intrinsic::x86_avx512_sitofp_round |
| 12713 | : Intrinsic::x86_avx512_uitofp_round; |
| 12714 | Function *F = CGF.CGM.getIntrinsic(IID, { Ty, Ops[0]->getType() }); |
| 12715 | Res = CGF.Builder.CreateCall(F, { Ops[0], Ops[3] }); |
| 12716 | } else { |
| 12717 | CodeGenFunction::CGFPOptionsRAII FPOptsRAII(CGF, E); |
| 12718 | Res = IsSigned ? CGF.Builder.CreateSIToFP(Ops[0], Ty) |
| 12719 | : CGF.Builder.CreateUIToFP(Ops[0], Ty); |
| 12720 | } |
| 12721 | |
| 12722 | return EmitX86Select(CGF, Ops[2], Res, Ops[1]); |
| 12723 | } |
| 12724 | |
| 12725 | // Lowers X86 FMA intrinsics to IR. |
| 12726 | static Value *EmitX86FMAExpr(CodeGenFunction &CGF, const CallExpr *E, |
| 12727 | ArrayRef<Value *> Ops, unsigned BuiltinID, |
| 12728 | bool IsAddSub) { |
| 12729 | |
| 12730 | bool Subtract = false; |
| 12731 | Intrinsic::ID IID = Intrinsic::not_intrinsic; |
| 12732 | switch (BuiltinID) { |
| 12733 | default: break; |
| 12734 | case clang::X86::BI__builtin_ia32_vfmsubph512_mask3: |
| 12735 | Subtract = true; |
| 12736 | [[fallthrough]]; |
| 12737 | case clang::X86::BI__builtin_ia32_vfmaddph512_mask: |
| 12738 | case clang::X86::BI__builtin_ia32_vfmaddph512_maskz: |
| 12739 | case clang::X86::BI__builtin_ia32_vfmaddph512_mask3: |
| 12740 | IID = llvm::Intrinsic::x86_avx512fp16_vfmadd_ph_512; |
| 12741 | break; |
| 12742 | case clang::X86::BI__builtin_ia32_vfmsubaddph512_mask3: |
| 12743 | Subtract = true; |
| 12744 | [[fallthrough]]; |
| 12745 | case clang::X86::BI__builtin_ia32_vfmaddsubph512_mask: |
| 12746 | case clang::X86::BI__builtin_ia32_vfmaddsubph512_maskz: |
| 12747 | case clang::X86::BI__builtin_ia32_vfmaddsubph512_mask3: |
| 12748 | IID = llvm::Intrinsic::x86_avx512fp16_vfmaddsub_ph_512; |
| 12749 | break; |
| 12750 | case clang::X86::BI__builtin_ia32_vfmsubps512_mask3: |
| 12751 | Subtract = true; |
| 12752 | [[fallthrough]]; |
| 12753 | case clang::X86::BI__builtin_ia32_vfmaddps512_mask: |
| 12754 | case clang::X86::BI__builtin_ia32_vfmaddps512_maskz: |
| 12755 | case clang::X86::BI__builtin_ia32_vfmaddps512_mask3: |
| 12756 | IID = llvm::Intrinsic::x86_avx512_vfmadd_ps_512; break; |
| 12757 | case clang::X86::BI__builtin_ia32_vfmsubpd512_mask3: |
| 12758 | Subtract = true; |
| 12759 | [[fallthrough]]; |
| 12760 | case clang::X86::BI__builtin_ia32_vfmaddpd512_mask: |
| 12761 | case clang::X86::BI__builtin_ia32_vfmaddpd512_maskz: |
| 12762 | case clang::X86::BI__builtin_ia32_vfmaddpd512_mask3: |
| 12763 | IID = llvm::Intrinsic::x86_avx512_vfmadd_pd_512; break; |
| 12764 | case clang::X86::BI__builtin_ia32_vfmsubaddps512_mask3: |
| 12765 | Subtract = true; |
| 12766 | [[fallthrough]]; |
| 12767 | case clang::X86::BI__builtin_ia32_vfmaddsubps512_mask: |
| 12768 | case clang::X86::BI__builtin_ia32_vfmaddsubps512_maskz: |
| 12769 | case clang::X86::BI__builtin_ia32_vfmaddsubps512_mask3: |
| 12770 | IID = llvm::Intrinsic::x86_avx512_vfmaddsub_ps_512; |
| 12771 | break; |
| 12772 | case clang::X86::BI__builtin_ia32_vfmsubaddpd512_mask3: |
| 12773 | Subtract = true; |
| 12774 | [[fallthrough]]; |
| 12775 | case clang::X86::BI__builtin_ia32_vfmaddsubpd512_mask: |
| 12776 | case clang::X86::BI__builtin_ia32_vfmaddsubpd512_maskz: |
| 12777 | case clang::X86::BI__builtin_ia32_vfmaddsubpd512_mask3: |
| 12778 | IID = llvm::Intrinsic::x86_avx512_vfmaddsub_pd_512; |
| 12779 | break; |
| 12780 | } |
| 12781 | |
| 12782 | Value *A = Ops[0]; |
| 12783 | Value *B = Ops[1]; |
| 12784 | Value *C = Ops[2]; |
| 12785 | |
| 12786 | if (Subtract) |
| 12787 | C = CGF.Builder.CreateFNeg(C); |
| 12788 | |
| 12789 | Value *Res; |
| 12790 | |
| 12791 | // Only handle in case of _MM_FROUND_CUR_DIRECTION/4 (no rounding). |
| 12792 | if (IID != Intrinsic::not_intrinsic && |
| 12793 | (cast<llvm::ConstantInt>(Ops.back())->getZExtValue() != (uint64_t)4 || |
| 12794 | IsAddSub)) { |
| 12795 | Function *Intr = CGF.CGM.getIntrinsic(IID); |
| 12796 | Res = CGF.Builder.CreateCall(Intr, {A, B, C, Ops.back() }); |
| 12797 | } else { |
| 12798 | llvm::Type *Ty = A->getType(); |
| 12799 | Function *FMA; |
| 12800 | if (CGF.Builder.getIsFPConstrained()) { |
| 12801 | CodeGenFunction::CGFPOptionsRAII FPOptsRAII(CGF, E); |
| 12802 | FMA = CGF.CGM.getIntrinsic(Intrinsic::experimental_constrained_fma, Ty); |
| 12803 | Res = CGF.Builder.CreateConstrainedFPCall(FMA, {A, B, C}); |
| 12804 | } else { |
| 12805 | FMA = CGF.CGM.getIntrinsic(Intrinsic::fma, Ty); |
| 12806 | Res = CGF.Builder.CreateCall(FMA, {A, B, C}); |
| 12807 | } |
| 12808 | } |
| 12809 | |
| 12810 | // Handle any required masking. |
| 12811 | Value *MaskFalseVal = nullptr; |
| 12812 | switch (BuiltinID) { |
| 12813 | case clang::X86::BI__builtin_ia32_vfmaddph512_mask: |
| 12814 | case clang::X86::BI__builtin_ia32_vfmaddps512_mask: |
| 12815 | case clang::X86::BI__builtin_ia32_vfmaddpd512_mask: |
| 12816 | case clang::X86::BI__builtin_ia32_vfmaddsubph512_mask: |
| 12817 | case clang::X86::BI__builtin_ia32_vfmaddsubps512_mask: |
| 12818 | case clang::X86::BI__builtin_ia32_vfmaddsubpd512_mask: |
| 12819 | MaskFalseVal = Ops[0]; |
| 12820 | break; |
| 12821 | case clang::X86::BI__builtin_ia32_vfmaddph512_maskz: |
| 12822 | case clang::X86::BI__builtin_ia32_vfmaddps512_maskz: |
| 12823 | case clang::X86::BI__builtin_ia32_vfmaddpd512_maskz: |
| 12824 | case clang::X86::BI__builtin_ia32_vfmaddsubph512_maskz: |
| 12825 | case clang::X86::BI__builtin_ia32_vfmaddsubps512_maskz: |
| 12826 | case clang::X86::BI__builtin_ia32_vfmaddsubpd512_maskz: |
| 12827 | MaskFalseVal = Constant::getNullValue(Ops[0]->getType()); |
| 12828 | break; |
| 12829 | case clang::X86::BI__builtin_ia32_vfmsubph512_mask3: |
| 12830 | case clang::X86::BI__builtin_ia32_vfmaddph512_mask3: |
| 12831 | case clang::X86::BI__builtin_ia32_vfmsubps512_mask3: |
| 12832 | case clang::X86::BI__builtin_ia32_vfmaddps512_mask3: |
| 12833 | case clang::X86::BI__builtin_ia32_vfmsubpd512_mask3: |
| 12834 | case clang::X86::BI__builtin_ia32_vfmaddpd512_mask3: |
| 12835 | case clang::X86::BI__builtin_ia32_vfmsubaddph512_mask3: |
| 12836 | case clang::X86::BI__builtin_ia32_vfmaddsubph512_mask3: |
| 12837 | case clang::X86::BI__builtin_ia32_vfmsubaddps512_mask3: |
| 12838 | case clang::X86::BI__builtin_ia32_vfmaddsubps512_mask3: |
| 12839 | case clang::X86::BI__builtin_ia32_vfmsubaddpd512_mask3: |
| 12840 | case clang::X86::BI__builtin_ia32_vfmaddsubpd512_mask3: |
| 12841 | MaskFalseVal = Ops[2]; |
| 12842 | break; |
| 12843 | } |
| 12844 | |
| 12845 | if (MaskFalseVal) |
| 12846 | return EmitX86Select(CGF, Ops[3], Res, MaskFalseVal); |
| 12847 | |
| 12848 | return Res; |
| 12849 | } |
| 12850 | |
| 12851 | static Value *EmitScalarFMAExpr(CodeGenFunction &CGF, const CallExpr *E, |
| 12852 | MutableArrayRef<Value *> Ops, Value *Upper, |
| 12853 | bool ZeroMask = false, unsigned PTIdx = 0, |
| 12854 | bool NegAcc = false) { |
| 12855 | unsigned Rnd = 4; |
| 12856 | if (Ops.size() > 4) |
| 12857 | Rnd = cast<llvm::ConstantInt>(Ops[4])->getZExtValue(); |
| 12858 | |
| 12859 | if (NegAcc) |
| 12860 | Ops[2] = CGF.Builder.CreateFNeg(Ops[2]); |
| 12861 | |
| 12862 | Ops[0] = CGF.Builder.CreateExtractElement(Ops[0], (uint64_t)0); |
| 12863 | Ops[1] = CGF.Builder.CreateExtractElement(Ops[1], (uint64_t)0); |
| 12864 | Ops[2] = CGF.Builder.CreateExtractElement(Ops[2], (uint64_t)0); |
| 12865 | Value *Res; |
| 12866 | if (Rnd != 4) { |
| 12867 | Intrinsic::ID IID; |
| 12868 | |
| 12869 | switch (Ops[0]->getType()->getPrimitiveSizeInBits()) { |
| 12870 | case 16: |
| 12871 | IID = Intrinsic::x86_avx512fp16_vfmadd_f16; |
| 12872 | break; |
| 12873 | case 32: |
| 12874 | IID = Intrinsic::x86_avx512_vfmadd_f32; |
| 12875 | break; |
| 12876 | case 64: |
| 12877 | IID = Intrinsic::x86_avx512_vfmadd_f64; |
| 12878 | break; |
| 12879 | default: |
| 12880 | llvm_unreachable("Unexpected size")::llvm::llvm_unreachable_internal("Unexpected size", "clang/lib/CodeGen/CGBuiltin.cpp" , 12880); |
| 12881 | } |
| 12882 | Res = CGF.Builder.CreateCall(CGF.CGM.getIntrinsic(IID), |
| 12883 | {Ops[0], Ops[1], Ops[2], Ops[4]}); |
| 12884 | } else if (CGF.Builder.getIsFPConstrained()) { |
| 12885 | CodeGenFunction::CGFPOptionsRAII FPOptsRAII(CGF, E); |
| 12886 | Function *FMA = CGF.CGM.getIntrinsic( |
| 12887 | Intrinsic::experimental_constrained_fma, Ops[0]->getType()); |
| 12888 | Res = CGF.Builder.CreateConstrainedFPCall(FMA, Ops.slice(0, 3)); |
| 12889 | } else { |
| 12890 | Function *FMA = CGF.CGM.getIntrinsic(Intrinsic::fma, Ops[0]->getType()); |
| 12891 | Res = CGF.Builder.CreateCall(FMA, Ops.slice(0, 3)); |
| 12892 | } |
| 12893 | // If we have more than 3 arguments, we need to do masking. |
| 12894 | if (Ops.size() > 3) { |
| 12895 | Value *PassThru = ZeroMask ? Constant::getNullValue(Res->getType()) |
| 12896 | : Ops[PTIdx]; |
| 12897 | |
| 12898 | // If we negated the accumulator and the its the PassThru value we need to |
| 12899 | // bypass the negate. Conveniently Upper should be the same thing in this |
| 12900 | // case. |
| 12901 | if (NegAcc && PTIdx == 2) |
| 12902 | PassThru = CGF.Builder.CreateExtractElement(Upper, (uint64_t)0); |
| 12903 | |
| 12904 | Res = EmitX86ScalarSelect(CGF, Ops[3], Res, PassThru); |
| 12905 | } |
| 12906 | return CGF.Builder.CreateInsertElement(Upper, Res, (uint64_t)0); |
| 12907 | } |
| 12908 | |
| 12909 | static Value *EmitX86Muldq(CodeGenFunction &CGF, bool IsSigned, |
| 12910 | ArrayRef<Value *> Ops) { |
| 12911 | llvm::Type *Ty = Ops[0]->getType(); |
| 12912 | // Arguments have a vXi32 type so cast to vXi64. |
| 12913 | Ty = llvm::FixedVectorType::get(CGF.Int64Ty, |
| 12914 | Ty->getPrimitiveSizeInBits() / 64); |
| 12915 | Value *LHS = CGF.Builder.CreateBitCast(Ops[0], Ty); |
| 12916 | Value *RHS = CGF.Builder.CreateBitCast(Ops[1], Ty); |
| 12917 | |
| 12918 | if (IsSigned) { |
| 12919 | // Shift left then arithmetic shift right. |
| 12920 | Constant *ShiftAmt = ConstantInt::get(Ty, 32); |
| 12921 | LHS = CGF.Builder.CreateShl(LHS, ShiftAmt); |
| 12922 | LHS = CGF.Builder.CreateAShr(LHS, ShiftAmt); |
| 12923 | RHS = CGF.Builder.CreateShl(RHS, ShiftAmt); |
| 12924 | RHS = CGF.Builder.CreateAShr(RHS, ShiftAmt); |
| 12925 | } else { |
| 12926 | // Clear the upper bits. |
| 12927 | Constant *Mask = ConstantInt::get(Ty, 0xffffffff); |
| 12928 | LHS = CGF.Builder.CreateAnd(LHS, Mask); |
| 12929 | RHS = CGF.Builder.CreateAnd(RHS, Mask); |
| 12930 | } |
| 12931 | |
| 12932 | return CGF.Builder.CreateMul(LHS, RHS); |
| 12933 | } |
| 12934 | |
| 12935 | // Emit a masked pternlog intrinsic. This only exists because the header has to |
| 12936 | // use a macro and we aren't able to pass the input argument to a pternlog |
| 12937 | // builtin and a select builtin without evaluating it twice. |
| 12938 | static Value *EmitX86Ternlog(CodeGenFunction &CGF, bool ZeroMask, |
| 12939 | ArrayRef<Value *> Ops) { |
| 12940 | llvm::Type *Ty = Ops[0]->getType(); |
| 12941 | |
| 12942 | unsigned VecWidth = Ty->getPrimitiveSizeInBits(); |
| 12943 | unsigned EltWidth = Ty->getScalarSizeInBits(); |
| 12944 | Intrinsic::ID IID; |
| 12945 | if (VecWidth == 128 && EltWidth == 32) |
| 12946 | IID = Intrinsic::x86_avx512_pternlog_d_128; |
| 12947 | else if (VecWidth == 256 && EltWidth == 32) |
| 12948 | IID = Intrinsic::x86_avx512_pternlog_d_256; |
| 12949 | else if (VecWidth == 512 && EltWidth == 32) |
| 12950 | IID = Intrinsic::x86_avx512_pternlog_d_512; |
| 12951 | else if (VecWidth == 128 && EltWidth == 64) |
| 12952 | IID = Intrinsic::x86_avx512_pternlog_q_128; |
| 12953 | else if (VecWidth == 256 && EltWidth == 64) |
| 12954 | IID = Intrinsic::x86_avx512_pternlog_q_256; |
| 12955 | else if (VecWidth == 512 && EltWidth == 64) |
| 12956 | IID = Intrinsic::x86_avx512_pternlog_q_512; |
| 12957 | else |
| 12958 | llvm_unreachable("Unexpected intrinsic")::llvm::llvm_unreachable_internal("Unexpected intrinsic", "clang/lib/CodeGen/CGBuiltin.cpp" , 12958); |
| 12959 | |
| 12960 | Value *Ternlog = CGF.Builder.CreateCall(CGF.CGM.getIntrinsic(IID), |
| 12961 | Ops.drop_back()); |
| 12962 | Value *PassThru = ZeroMask ? ConstantAggregateZero::get(Ty) : Ops[0]; |
| 12963 | return EmitX86Select(CGF, Ops[4], Ternlog, PassThru); |
| 12964 | } |
| 12965 | |
| 12966 | static Value *EmitX86SExtMask(CodeGenFunction &CGF, Value *Op, |
| 12967 | llvm::Type *DstTy) { |
| 12968 | unsigned NumberOfElements = |
| 12969 | cast<llvm::FixedVectorType>(DstTy)->getNumElements(); |
| 12970 | Value *Mask = getMaskVecValue(CGF, Op, NumberOfElements); |
| 12971 | return CGF.Builder.CreateSExt(Mask, DstTy, "vpmovm2"); |
| 12972 | } |
| 12973 | |
| 12974 | Value *CodeGenFunction::EmitX86CpuIs(const CallExpr *E) { |
| 12975 | const Expr *CPUExpr = E->getArg(0)->IgnoreParenCasts(); |
| 12976 | StringRef CPUStr = cast<clang::StringLiteral>(CPUExpr)->getString(); |
| 12977 | return EmitX86CpuIs(CPUStr); |
| 12978 | } |
| 12979 | |
| 12980 | // Convert F16 halfs to floats. |
| 12981 | static Value *EmitX86CvtF16ToFloatExpr(CodeGenFunction &CGF, |
| 12982 | ArrayRef<Value *> Ops, |
| 12983 | llvm::Type *DstTy) { |
| 12984 | assert((Ops.size() == 1 || Ops.size() == 3 || Ops.size() == 4) &&(static_cast <bool> ((Ops.size() == 1 || Ops.size() == 3 || Ops.size() == 4) && "Unknown cvtph2ps intrinsic") ? void (0) : __assert_fail ("(Ops.size() == 1 || Ops.size() == 3 || Ops.size() == 4) && \"Unknown cvtph2ps intrinsic\"" , "clang/lib/CodeGen/CGBuiltin.cpp", 12985, __extension__ __PRETTY_FUNCTION__ )) |
| 12985 | "Unknown cvtph2ps intrinsic")(static_cast <bool> ((Ops.size() == 1 || Ops.size() == 3 || Ops.size() == 4) && "Unknown cvtph2ps intrinsic") ? void (0) : __assert_fail ("(Ops.size() == 1 || Ops.size() == 3 || Ops.size() == 4) && \"Unknown cvtph2ps intrinsic\"" , "clang/lib/CodeGen/CGBuiltin.cpp", 12985, __extension__ __PRETTY_FUNCTION__ )); |
| 12986 | |
| 12987 | // If the SAE intrinsic doesn't use default rounding then we can't upgrade. |
| 12988 | if (Ops.size() == 4 && cast<llvm::ConstantInt>(Ops[3])->getZExtValue() != 4) { |
| 12989 | Function *F = |
| 12990 | CGF.CGM.getIntrinsic(Intrinsic::x86_avx512_mask_vcvtph2ps_512); |
| 12991 | return CGF.Builder.CreateCall(F, {Ops[0], Ops[1], Ops[2], Ops[3]}); |
| 12992 | } |
| 12993 | |
| 12994 | unsigned NumDstElts = cast<llvm::FixedVectorType>(DstTy)->getNumElements(); |
| 12995 | Value *Src = Ops[0]; |
| 12996 | |
| 12997 | // Extract the subvector. |
| 12998 | if (NumDstElts != |
| 12999 | cast<llvm::FixedVectorType>(Src->getType())->getNumElements()) { |
| 13000 | assert(NumDstElts == 4 && "Unexpected vector size")(static_cast <bool> (NumDstElts == 4 && "Unexpected vector size" ) ? void (0) : __assert_fail ("NumDstElts == 4 && \"Unexpected vector size\"" , "clang/lib/CodeGen/CGBuiltin.cpp", 13000, __extension__ __PRETTY_FUNCTION__ )); |
| 13001 | Src = CGF.Builder.CreateShuffleVector(Src, ArrayRef<int>{0, 1, 2, 3}); |
| 13002 | } |
| 13003 | |
| 13004 | // Bitcast from vXi16 to vXf16. |
| 13005 | auto *HalfTy = llvm::FixedVectorType::get( |
| 13006 | llvm::Type::getHalfTy(CGF.getLLVMContext()), NumDstElts); |
| 13007 | Src = CGF.Builder.CreateBitCast(Src, HalfTy); |
| 13008 | |
| 13009 | // Perform the fp-extension. |
| 13010 | Value *Res = CGF.Builder.CreateFPExt(Src, DstTy, "cvtph2ps"); |
| 13011 | |
| 13012 | if (Ops.size() >= 3) |
| 13013 | Res = EmitX86Select(CGF, Ops[2], Res, Ops[1]); |
| 13014 | return Res; |
| 13015 | } |
| 13016 | |
| 13017 | Value *CodeGenFunction::EmitX86CpuIs(StringRef CPUStr) { |
| 13018 | |
| 13019 | llvm::Type *Int32Ty = Builder.getInt32Ty(); |
| 13020 | |
| 13021 | // Matching the struct layout from the compiler-rt/libgcc structure that is |
| 13022 | // filled in: |
| 13023 | // unsigned int __cpu_vendor; |
| 13024 | // unsigned int __cpu_type; |
| 13025 | // unsigned int __cpu_subtype; |
| 13026 | // unsigned int __cpu_features[1]; |
| 13027 | llvm::Type *STy = llvm::StructType::get(Int32Ty, Int32Ty, Int32Ty, |
| 13028 | llvm::ArrayType::get(Int32Ty, 1)); |
| 13029 | |
| 13030 | // Grab the global __cpu_model. |
| 13031 | llvm::Constant *CpuModel = CGM.CreateRuntimeVariable(STy, "__cpu_model"); |
| 13032 | cast<llvm::GlobalValue>(CpuModel)->setDSOLocal(true); |
| 13033 | |
| 13034 | // Calculate the index needed to access the correct field based on the |
| 13035 | // range. Also adjust the expected value. |
| 13036 | unsigned Index; |
| 13037 | unsigned Value; |
| 13038 | std::tie(Index, Value) = StringSwitch<std::pair<unsigned, unsigned>>(CPUStr) |
| 13039 | #define X86_VENDOR(ENUM, STRING) \ |
| 13040 | .Case(STRING, {0u, static_cast<unsigned>(llvm::X86::ENUM)}) |
| 13041 | #define X86_CPU_TYPE_ALIAS(ENUM, ALIAS) \ |
| 13042 | .Case(ALIAS, {1u, static_cast<unsigned>(llvm::X86::ENUM)}) |
| 13043 | #define X86_CPU_TYPE(ENUM, STR) \ |
| 13044 | .Case(STR, {1u, static_cast<unsigned>(llvm::X86::ENUM)}) |
| 13045 | #define X86_CPU_SUBTYPE_ALIAS(ENUM, ALIAS) \ |
| 13046 | .Case(ALIAS, {2u, static_cast<unsigned>(llvm::X86::ENUM)}) |
| 13047 | #define X86_CPU_SUBTYPE(ENUM, STR) \ |
| 13048 | .Case(STR, {2u, static_cast<unsigned>(llvm::X86::ENUM)}) |
| 13049 | #include "llvm/TargetParser/X86TargetParser.def" |
| 13050 | .Default({0, 0}); |
| 13051 | assert(Value != 0 && "Invalid CPUStr passed to CpuIs")(static_cast <bool> (Value != 0 && "Invalid CPUStr passed to CpuIs" ) ? void (0) : __assert_fail ("Value != 0 && \"Invalid CPUStr passed to CpuIs\"" , "clang/lib/CodeGen/CGBuiltin.cpp", 13051, __extension__ __PRETTY_FUNCTION__ )); |
| 13052 | |
| 13053 | // Grab the appropriate field from __cpu_model. |
| 13054 | llvm::Value *Idxs[] = {ConstantInt::get(Int32Ty, 0), |
| 13055 | ConstantInt::get(Int32Ty, Index)}; |
| 13056 | llvm::Value *CpuValue = Builder.CreateGEP(STy, CpuModel, Idxs); |
| 13057 | CpuValue = Builder.CreateAlignedLoad(Int32Ty, CpuValue, |
| 13058 | CharUnits::fromQuantity(4)); |
| 13059 | |
| 13060 | // Check the value of the field against the requested value. |
| 13061 | return Builder.CreateICmpEQ(CpuValue, |
| 13062 | llvm::ConstantInt::get(Int32Ty, Value)); |
| 13063 | } |
| 13064 | |
| 13065 | Value *CodeGenFunction::EmitX86CpuSupports(const CallExpr *E) { |
| 13066 | const Expr *FeatureExpr = E->getArg(0)->IgnoreParenCasts(); |
| 13067 | StringRef FeatureStr = cast<StringLiteral>(FeatureExpr)->getString(); |
| 13068 | return EmitX86CpuSupports(FeatureStr); |
| 13069 | } |
| 13070 | |
| 13071 | Value *CodeGenFunction::EmitX86CpuSupports(ArrayRef<StringRef> FeatureStrs) { |
| 13072 | return EmitX86CpuSupports(llvm::X86::getCpuSupportsMask(FeatureStrs)); |
| 13073 | } |
| 13074 | |
| 13075 | llvm::Value *CodeGenFunction::EmitX86CpuSupports(uint64_t FeaturesMask) { |
| 13076 | uint32_t Features1 = Lo_32(FeaturesMask); |
| 13077 | uint32_t Features2 = Hi_32(FeaturesMask); |
| 13078 | |
| 13079 | Value *Result = Builder.getTrue(); |
| 13080 | |
| 13081 | if (Features1 != 0) { |
| 13082 | // Matching the struct layout from the compiler-rt/libgcc structure that is |
| 13083 | // filled in: |
| 13084 | // unsigned int __cpu_vendor; |
| 13085 | // unsigned int __cpu_type; |
| 13086 | // unsigned int __cpu_subtype; |
| 13087 | // unsigned int __cpu_features[1]; |
| 13088 | llvm::Type *STy = llvm::StructType::get(Int32Ty, Int32Ty, Int32Ty, |
| 13089 | llvm::ArrayType::get(Int32Ty, 1)); |
| 13090 | |
| 13091 | // Grab the global __cpu_model. |
| 13092 | llvm::Constant *CpuModel = CGM.CreateRuntimeVariable(STy, "__cpu_model"); |
| 13093 | cast<llvm::GlobalValue>(CpuModel)->setDSOLocal(true); |
| 13094 | |
| 13095 | // Grab the first (0th) element from the field __cpu_features off of the |
| 13096 | // global in the struct STy. |
| 13097 | Value *Idxs[] = {Builder.getInt32(0), Builder.getInt32(3), |
| 13098 | Builder.getInt32(0)}; |
| 13099 | Value *CpuFeatures = Builder.CreateGEP(STy, CpuModel, Idxs); |
| 13100 | Value *Features = Builder.CreateAlignedLoad(Int32Ty, CpuFeatures, |
| 13101 | CharUnits::fromQuantity(4)); |
| 13102 | |
| 13103 | // Check the value of the bit corresponding to the feature requested. |
| 13104 | Value *Mask = Builder.getInt32(Features1); |
| 13105 | Value *Bitset = Builder.CreateAnd(Features, Mask); |
| 13106 | Value *Cmp = Builder.CreateICmpEQ(Bitset, Mask); |
| 13107 | Result = Builder.CreateAnd(Result, Cmp); |
| 13108 | } |
| 13109 | |
| 13110 | if (Features2 != 0) { |
| 13111 | llvm::Constant *CpuFeatures2 = CGM.CreateRuntimeVariable(Int32Ty, |
| 13112 | "__cpu_features2"); |
| 13113 | cast<llvm::GlobalValue>(CpuFeatures2)->setDSOLocal(true); |
| 13114 | |
| 13115 | Value *Features = Builder.CreateAlignedLoad(Int32Ty, CpuFeatures2, |
| 13116 | CharUnits::fromQuantity(4)); |
| 13117 | |
| 13118 | // Check the value of the bit corresponding to the feature requested. |
| 13119 | Value *Mask = Builder.getInt32(Features2); |
| 13120 | Value *Bitset = Builder.CreateAnd(Features, Mask); |
| 13121 | Value *Cmp = Builder.CreateICmpEQ(Bitset, Mask); |
| 13122 | Result = Builder.CreateAnd(Result, Cmp); |
| 13123 | } |
| 13124 | |
| 13125 | return Result; |
| 13126 | } |
| 13127 | |
| 13128 | Value *CodeGenFunction::EmitAArch64CpuInit() { |
| 13129 | llvm::FunctionType *FTy = llvm::FunctionType::get(VoidTy, false); |
| 13130 | llvm::FunctionCallee Func = |
| 13131 | CGM.CreateRuntimeFunction(FTy, "init_cpu_features_resolver"); |
| 13132 | cast<llvm::GlobalValue>(Func.getCallee())->setDSOLocal(true); |
| 13133 | cast<llvm::GlobalValue>(Func.getCallee()) |
| 13134 | ->setDLLStorageClass(llvm::GlobalValue::DefaultStorageClass); |
| 13135 | return Builder.CreateCall(Func); |
| 13136 | } |
| 13137 | |
| 13138 | Value *CodeGenFunction::EmitX86CpuInit() { |
| 13139 | llvm::FunctionType *FTy = llvm::FunctionType::get(VoidTy, |
| 13140 | /*Variadic*/ false); |
| 13141 | llvm::FunctionCallee Func = |
| 13142 | CGM.CreateRuntimeFunction(FTy, "__cpu_indicator_init"); |
| 13143 | cast<llvm::GlobalValue>(Func.getCallee())->setDSOLocal(true); |
| 13144 | cast<llvm::GlobalValue>(Func.getCallee()) |
| 13145 | ->setDLLStorageClass(llvm::GlobalValue::DefaultStorageClass); |
| 13146 | return Builder.CreateCall(Func); |
| 13147 | } |
| 13148 | |
| 13149 | llvm::Value * |
| 13150 | CodeGenFunction::EmitAArch64CpuSupports(ArrayRef<StringRef> FeaturesStrs) { |
| 13151 | uint64_t FeaturesMask = llvm::AArch64::getCpuSupportsMask(FeaturesStrs); |
| 13152 | Value *Result = Builder.getTrue(); |
| 13153 | if (FeaturesMask != 0) { |
| 13154 | // Get features from structure in runtime library |
| 13155 | // struct { |
| 13156 | // unsigned long long features; |
| 13157 | // } __aarch64_cpu_features; |
| 13158 | llvm::Type *STy = llvm::StructType::get(Int64Ty); |
| 13159 | llvm::Constant *AArch64CPUFeatures = |
| 13160 | CGM.CreateRuntimeVariable(STy, "__aarch64_cpu_features"); |
| 13161 | cast<llvm::GlobalValue>(AArch64CPUFeatures)->setDSOLocal(true); |
| 13162 | llvm::Value *CpuFeatures = Builder.CreateGEP( |
| 13163 | STy, AArch64CPUFeatures, |
| 13164 | {ConstantInt::get(Int32Ty, 0), ConstantInt::get(Int32Ty, 0)}); |
| 13165 | Value *Features = Builder.CreateAlignedLoad(Int64Ty, CpuFeatures, |
| 13166 | CharUnits::fromQuantity(8)); |
| 13167 | Value *Mask = Builder.getInt64(FeaturesMask); |
| 13168 | Value *Bitset = Builder.CreateAnd(Features, Mask); |
| 13169 | Value *Cmp = Builder.CreateICmpEQ(Bitset, Mask); |
| 13170 | Result = Builder.CreateAnd(Result, Cmp); |
| 13171 | } |
| 13172 | return Result; |
| 13173 | } |
| 13174 | |
| 13175 | Value *CodeGenFunction::EmitX86BuiltinExpr(unsigned BuiltinID, |
| 13176 | const CallExpr *E) { |
| 13177 | if (BuiltinID == X86::BI__builtin_cpu_is) |
| 13178 | return EmitX86CpuIs(E); |
| 13179 | if (BuiltinID == X86::BI__builtin_cpu_supports) |
| 13180 | return EmitX86CpuSupports(E); |
| 13181 | if (BuiltinID == X86::BI__builtin_cpu_init) |
| 13182 | return EmitX86CpuInit(); |
| 13183 | |
| 13184 | // Handle MSVC intrinsics before argument evaluation to prevent double |
| 13185 | // evaluation. |
| 13186 | if (std::optional<MSVCIntrin> MsvcIntId = translateX86ToMsvcIntrin(BuiltinID)) |
| 13187 | return EmitMSVCBuiltinExpr(*MsvcIntId, E); |
| 13188 | |
| 13189 | SmallVector<Value*, 4> Ops; |
| 13190 | bool IsMaskFCmp = false; |
| 13191 | bool IsConjFMA = false; |
| 13192 | |
| 13193 | // Find out if any arguments are required to be integer constant expressions. |
| 13194 | unsigned ICEArguments = 0; |
| 13195 | ASTContext::GetBuiltinTypeError Error; |
| 13196 | getContext().GetBuiltinType(BuiltinID, Error, &ICEArguments); |
| 13197 | assert(Error == ASTContext::GE_None && "Should not codegen an error")(static_cast <bool> (Error == ASTContext::GE_None && "Should not codegen an error") ? void (0) : __assert_fail ("Error == ASTContext::GE_None && \"Should not codegen an error\"" , "clang/lib/CodeGen/CGBuiltin.cpp", 13197, __extension__ __PRETTY_FUNCTION__ )); |
| 13198 | |
| 13199 | for (unsigned i = 0, e = E->getNumArgs(); i != e; i++) { |
| 13200 | // If this is a normal argument, just emit it as a scalar. |
| 13201 | if ((ICEArguments & (1 << i)) == 0) { |
| 13202 | Ops.push_back(EmitScalarExpr(E->getArg(i))); |
| 13203 | continue; |
| 13204 | } |
| 13205 | |
| 13206 | // If this is required to be a constant, constant fold it so that we know |
| 13207 | // that the generated intrinsic gets a ConstantInt. |
| 13208 | Ops.push_back(llvm::ConstantInt::get( |
| 13209 | getLLVMContext(), *E->getArg(i)->getIntegerConstantExpr(getContext()))); |
| 13210 | } |
| 13211 | |
| 13212 | // These exist so that the builtin that takes an immediate can be bounds |
| 13213 | // checked by clang to avoid passing bad immediates to the backend. Since |
| 13214 | // AVX has a larger immediate than SSE we would need separate builtins to |
| 13215 | // do the different bounds checking. Rather than create a clang specific |
| 13216 | // SSE only builtin, this implements eight separate builtins to match gcc |
| 13217 | // implementation. |
| 13218 | auto getCmpIntrinsicCall = [this, &Ops](Intrinsic::ID ID, unsigned Imm) { |
| 13219 | Ops.push_back(llvm::ConstantInt::get(Int8Ty, Imm)); |
| 13220 | llvm::Function *F = CGM.getIntrinsic(ID); |
| 13221 | return Builder.CreateCall(F, Ops); |
| 13222 | }; |
| 13223 | |
| 13224 | // For the vector forms of FP comparisons, translate the builtins directly to |
| 13225 | // IR. |
| 13226 | // TODO: The builtins could be removed if the SSE header files used vector |
| 13227 | // extension comparisons directly (vector ordered/unordered may need |
| 13228 | // additional support via __builtin_isnan()). |
| 13229 | auto getVectorFCmpIR = [this, &Ops, E](CmpInst::Predicate Pred, |
| 13230 | bool IsSignaling) { |
| 13231 | CodeGenFunction::CGFPOptionsRAII FPOptsRAII(*this, E); |
| 13232 | Value *Cmp; |
| 13233 | if (IsSignaling) |
| 13234 | Cmp = Builder.CreateFCmpS(Pred, Ops[0], Ops[1]); |
| 13235 | else |
| 13236 | Cmp = Builder.CreateFCmp(Pred, Ops[0], Ops[1]); |
| 13237 | llvm::VectorType *FPVecTy = cast<llvm::VectorType>(Ops[0]->getType()); |
| 13238 | llvm::VectorType *IntVecTy = llvm::VectorType::getInteger(FPVecTy); |
| 13239 | Value *Sext = Builder.CreateSExt(Cmp, IntVecTy); |
| 13240 | return Builder.CreateBitCast(Sext, FPVecTy); |
| 13241 | }; |
| 13242 | |
| 13243 | switch (BuiltinID) { |
| 13244 | default: return nullptr; |
| 13245 | case X86::BI_mm_prefetch: { |
| 13246 | Value *Address = Ops[0]; |
| 13247 | ConstantInt *C = cast<ConstantInt>(Ops[1]); |
| 13248 | Value *RW = ConstantInt::get(Int32Ty, (C->getZExtValue() >> 2) & 0x1); |
| 13249 | Value *Locality = ConstantInt::get(Int32Ty, C->getZExtValue() & 0x3); |
| 13250 | Value *Data = ConstantInt::get(Int32Ty, 1); |
| 13251 | Function *F = CGM.getIntrinsic(Intrinsic::prefetch, Address->getType()); |
| 13252 | return Builder.CreateCall(F, {Address, RW, Locality, Data}); |
| 13253 | } |
| 13254 | case X86::BI_mm_clflush: { |
| 13255 | return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::x86_sse2_clflush), |
| 13256 | Ops[0]); |
| 13257 | } |
| 13258 | case X86::BI_mm_lfence: { |
| 13259 | return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::x86_sse2_lfence)); |
| 13260 | } |
| 13261 | case X86::BI_mm_mfence: { |
| 13262 | return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::x86_sse2_mfence)); |
| 13263 | } |
| 13264 | case X86::BI_mm_sfence: { |
| 13265 | return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::x86_sse_sfence)); |
| 13266 | } |
| 13267 | case X86::BI_mm_pause: { |
| 13268 | return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::x86_sse2_pause)); |
| 13269 | } |
| 13270 | case X86::BI__rdtsc: { |
| 13271 | return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::x86_rdtsc)); |
| 13272 | } |
| 13273 | case X86::BI__builtin_ia32_rdtscp: { |
| 13274 | Value *Call = Builder.CreateCall(CGM.getIntrinsic(Intrinsic::x86_rdtscp)); |
| 13275 | Builder.CreateDefaultAlignedStore(Builder.CreateExtractValue(Call, 1), |
| 13276 | Ops[0]); |
| 13277 | return Builder.CreateExtractValue(Call, 0); |
| 13278 | } |
| 13279 | case X86::BI__builtin_ia32_lzcnt_u16: |
| 13280 | case X86::BI__builtin_ia32_lzcnt_u32: |
| 13281 | case X86::BI__builtin_ia32_lzcnt_u64: { |
| 13282 | Function *F = CGM.getIntrinsic(Intrinsic::ctlz, Ops[0]->getType()); |
| 13283 | return Builder.CreateCall(F, {Ops[0], Builder.getInt1(false)}); |
| 13284 | } |
| 13285 | case X86::BI__builtin_ia32_tzcnt_u16: |
| 13286 | case X86::BI__builtin_ia32_tzcnt_u32: |
| 13287 | case X86::BI__builtin_ia32_tzcnt_u64: { |
| 13288 | Function *F = CGM.getIntrinsic(Intrinsic::cttz, Ops[0]->getType()); |
| 13289 | return Builder.CreateCall(F, {Ops[0], Builder.getInt1(false)}); |
| 13290 | } |
| 13291 | case X86::BI__builtin_ia32_undef128: |
| 13292 | case X86::BI__builtin_ia32_undef256: |
| 13293 | case X86::BI__builtin_ia32_undef512: |
| 13294 | // The x86 definition of "undef" is not the same as the LLVM definition |
| 13295 | // (PR32176). We leave optimizing away an unnecessary zero constant to the |
| 13296 | // IR optimizer and backend. |
| 13297 | // TODO: If we had a "freeze" IR instruction to generate a fixed undef |
| 13298 | // value, we should use that here instead of a zero. |
| 13299 | return llvm::Constant::getNullValue(ConvertType(E->getType())); |
| 13300 | case X86::BI__builtin_ia32_vec_init_v8qi: |
| 13301 | case X86::BI__builtin_ia32_vec_init_v4hi: |
| 13302 | case X86::BI__builtin_ia32_vec_init_v2si: |
| 13303 | return Builder.CreateBitCast(BuildVector(Ops), |
| 13304 | llvm::Type::getX86_MMXTy(getLLVMContext())); |
| 13305 | case X86::BI__builtin_ia32_vec_ext_v2si: |
| 13306 | case X86::BI__builtin_ia32_vec_ext_v16qi: |
| 13307 | case X86::BI__builtin_ia32_vec_ext_v8hi: |
| 13308 | case X86::BI__builtin_ia32_vec_ext_v4si: |
| 13309 | case X86::BI__builtin_ia32_vec_ext_v4sf: |
| 13310 | case X86::BI__builtin_ia32_vec_ext_v2di: |
| 13311 | case X86::BI__builtin_ia32_vec_ext_v32qi: |
| 13312 | case X86::BI__builtin_ia32_vec_ext_v16hi: |
| 13313 | case X86::BI__builtin_ia32_vec_ext_v8si: |
| 13314 | case X86::BI__builtin_ia32_vec_ext_v4di: { |
| 13315 | unsigned NumElts = |
| 13316 | cast<llvm::FixedVectorType>(Ops[0]->getType())->getNumElements(); |
| 13317 | uint64_t Index = cast<ConstantInt>(Ops[1])->getZExtValue(); |
| 13318 | Index &= NumElts - 1; |
| 13319 | // These builtins exist so we can ensure the index is an ICE and in range. |
| 13320 | // Otherwise we could just do this in the header file. |
| 13321 | return Builder.CreateExtractElement(Ops[0], Index); |
| 13322 | } |
| 13323 | case X86::BI__builtin_ia32_vec_set_v16qi: |
| 13324 | case X86::BI__builtin_ia32_vec_set_v8hi: |
| 13325 | case X86::BI__builtin_ia32_vec_set_v4si: |
| 13326 | case X86::BI__builtin_ia32_vec_set_v2di: |
| 13327 | case X86::BI__builtin_ia32_vec_set_v32qi: |
| 13328 | case X86::BI__builtin_ia32_vec_set_v16hi: |
| 13329 | case X86::BI__builtin_ia32_vec_set_v8si: |
| 13330 | case X86::BI__builtin_ia32_vec_set_v4di: { |
| 13331 | unsigned NumElts = |
| 13332 | cast<llvm::FixedVectorType>(Ops[0]->getType())->getNumElements(); |
| 13333 | unsigned Index = cast<ConstantInt>(Ops[2])->getZExtValue(); |
| 13334 | Index &= NumElts - 1; |
| 13335 | // These builtins exist so we can ensure the index is an ICE and in range. |
| 13336 | // Otherwise we could just do this in the header file. |
| 13337 | return Builder.CreateInsertElement(Ops[0], Ops[1], Index); |
| 13338 | } |
| 13339 | case X86::BI_mm_setcsr: |
| 13340 | case X86::BI__builtin_ia32_ldmxcsr: { |
| 13341 | Address Tmp = CreateMemTemp(E->getArg(0)->getType()); |
| 13342 | Builder.CreateStore(Ops[0], Tmp); |
| 13343 | return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::x86_sse_ldmxcsr), |
| 13344 | Builder.CreateBitCast(Tmp.getPointer(), Int8PtrTy)); |
| 13345 | } |
| 13346 | case X86::BI_mm_getcsr: |
| 13347 | case X86::BI__builtin_ia32_stmxcsr: { |
| 13348 | Address Tmp = CreateMemTemp(E->getType()); |
| 13349 | Builder.CreateCall(CGM.getIntrinsic(Intrinsic::x86_sse_stmxcsr), |
| 13350 | Builder.CreateBitCast(Tmp.getPointer(), Int8PtrTy)); |
| 13351 | return Builder.CreateLoad(Tmp, "stmxcsr"); |
| 13352 | } |
| 13353 | case X86::BI__builtin_ia32_xsave: |
| 13354 | case X86::BI__builtin_ia32_xsave64: |
| 13355 | case X86::BI__builtin_ia32_xrstor: |
| 13356 | case X86::BI__builtin_ia32_xrstor64: |
| 13357 | case X86::BI__builtin_ia32_xsaveopt: |
| 13358 | case X86::BI__builtin_ia32_xsaveopt64: |
| 13359 | case X86::BI__builtin_ia32_xrstors: |
| 13360 | case X86::BI__builtin_ia32_xrstors64: |
| 13361 | case X86::BI__builtin_ia32_xsavec: |
| 13362 | case X86::BI__builtin_ia32_xsavec64: |
| 13363 | case X86::BI__builtin_ia32_xsaves: |
| 13364 | case X86::BI__builtin_ia32_xsaves64: |
| 13365 | case X86::BI__builtin_ia32_xsetbv: |
| 13366 | case X86::BI_xsetbv: { |
| 13367 | Intrinsic::ID ID; |
| 13368 | #define INTRINSIC_X86_XSAVE_ID(NAME) \ |
| 13369 | case X86::BI__builtin_ia32_##NAME: \ |
| 13370 | ID = Intrinsic::x86_##NAME; \ |
| 13371 | break |
| 13372 | switch (BuiltinID) { |
| 13373 | default: llvm_unreachable("Unsupported intrinsic!")::llvm::llvm_unreachable_internal("Unsupported intrinsic!", "clang/lib/CodeGen/CGBuiltin.cpp" , 13373); |
| 13374 | INTRINSIC_X86_XSAVE_ID(xsave); |
| 13375 | INTRINSIC_X86_XSAVE_ID(xsave64); |
| 13376 | INTRINSIC_X86_XSAVE_ID(xrstor); |
| 13377 | INTRINSIC_X86_XSAVE_ID(xrstor64); |
| 13378 | INTRINSIC_X86_XSAVE_ID(xsaveopt); |
| 13379 | INTRINSIC_X86_XSAVE_ID(xsaveopt64); |
| 13380 | INTRINSIC_X86_XSAVE_ID(xrstors); |
| 13381 | INTRINSIC_X86_XSAVE_ID(xrstors64); |
| 13382 | INTRINSIC_X86_XSAVE_ID(xsavec); |
| 13383 | INTRINSIC_X86_XSAVE_ID(xsavec64); |
| 13384 | INTRINSIC_X86_XSAVE_ID(xsaves); |
| 13385 | INTRINSIC_X86_XSAVE_ID(xsaves64); |
| 13386 | INTRINSIC_X86_XSAVE_ID(xsetbv); |
| 13387 | case X86::BI_xsetbv: |
| 13388 | ID = Intrinsic::x86_xsetbv; |
| 13389 | break; |
| 13390 | } |
| 13391 | #undef INTRINSIC_X86_XSAVE_ID |
| 13392 | Value *Mhi = Builder.CreateTrunc( |
| 13393 | Builder.CreateLShr(Ops[1], ConstantInt::get(Int64Ty, 32)), Int32Ty); |
| 13394 | Value *Mlo = Builder.CreateTrunc(Ops[1], Int32Ty); |
| 13395 | Ops[1] = Mhi; |
| 13396 | Ops.push_back(Mlo); |
| 13397 | return Builder.CreateCall(CGM.getIntrinsic(ID), Ops); |
| 13398 | } |
| 13399 | case X86::BI__builtin_ia32_xgetbv: |
| 13400 | case X86::BI_xgetbv: |
| 13401 | return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::x86_xgetbv), Ops); |
| 13402 | case X86::BI__builtin_ia32_storedqudi128_mask: |
| 13403 | case X86::BI__builtin_ia32_storedqusi128_mask: |
| 13404 | case X86::BI__builtin_ia32_storedquhi128_mask: |
| 13405 | case X86::BI__builtin_ia32_storedquqi128_mask: |
| 13406 | case X86::BI__builtin_ia32_storeupd128_mask: |
| 13407 | case X86::BI__builtin_ia32_storeups128_mask: |
| 13408 | case X86::BI__builtin_ia32_storedqudi256_mask: |
| 13409 | case X86::BI__builtin_ia32_storedqusi256_mask: |
| 13410 | case X86::BI__builtin_ia32_storedquhi256_mask: |
| 13411 | case X86::BI__builtin_ia32_storedquqi256_mask: |
| 13412 | case X86::BI__builtin_ia32_storeupd256_mask: |
| 13413 | case X86::BI__builtin_ia32_storeups256_mask: |
| 13414 | case X86::BI__builtin_ia32_storedqudi512_mask: |
| 13415 | case X86::BI__builtin_ia32_storedqusi512_mask: |
| 13416 | case X86::BI__builtin_ia32_storedquhi512_mask: |
| 13417 | case X86::BI__builtin_ia32_storedquqi512_mask: |
| 13418 | case X86::BI__builtin_ia32_storeupd512_mask: |
| 13419 | case X86::BI__builtin_ia32_storeups512_mask: |
| 13420 | return EmitX86MaskedStore(*this, Ops, Align(1)); |
| 13421 | |
| 13422 | case X86::BI__builtin_ia32_storesh128_mask: |
| 13423 | case X86::BI__builtin_ia32_storess128_mask: |
| 13424 | case X86::BI__builtin_ia32_storesd128_mask: |
| 13425 | return EmitX86MaskedStore(*this, Ops, Align(1)); |
| 13426 | |
| 13427 | case X86::BI__builtin_ia32_vpopcntb_128: |
| 13428 | case X86::BI__builtin_ia32_vpopcntd_128: |
| 13429 | case X86::BI__builtin_ia32_vpopcntq_128: |
| 13430 | case X86::BI__builtin_ia32_vpopcntw_128: |
| 13431 | case X86::BI__builtin_ia32_vpopcntb_256: |
| 13432 | case X86::BI__builtin_ia32_vpopcntd_256: |
| 13433 | case X86::BI__builtin_ia32_vpopcntq_256: |
| 13434 | case X86::BI__builtin_ia32_vpopcntw_256: |
| 13435 | case X86::BI__builtin_ia32_vpopcntb_512: |
| 13436 | case X86::BI__builtin_ia32_vpopcntd_512: |
| 13437 | case X86::BI__builtin_ia32_vpopcntq_512: |
| 13438 | case X86::BI__builtin_ia32_vpopcntw_512: { |
| 13439 | llvm::Type *ResultType = ConvertType(E->getType()); |
| 13440 | llvm::Function *F = CGM.getIntrinsic(Intrinsic::ctpop, ResultType); |
| 13441 | return Builder.CreateCall(F, Ops); |
| 13442 | } |
| 13443 | case X86::BI__builtin_ia32_cvtmask2b128: |
| 13444 | case X86::BI__builtin_ia32_cvtmask2b256: |
| 13445 | case X86::BI__builtin_ia32_cvtmask2b512: |
| 13446 | case X86::BI__builtin_ia32_cvtmask2w128: |
| 13447 | case X86::BI__builtin_ia32_cvtmask2w256: |
| 13448 | case X86::BI__builtin_ia32_cvtmask2w512: |
| 13449 | case X86::BI__builtin_ia32_cvtmask2d128: |
| 13450 | case X86::BI__builtin_ia32_cvtmask2d256: |
| 13451 | case X86::BI__builtin_ia32_cvtmask2d512: |
| 13452 | case X86::BI__builtin_ia32_cvtmask2q128: |
| 13453 | case X86::BI__builtin_ia32_cvtmask2q256: |
| 13454 | case X86::BI__builtin_ia32_cvtmask2q512: |
| 13455 | return EmitX86SExtMask(*this, Ops[0], ConvertType(E->getType())); |
| 13456 | |
| 13457 | case X86::BI__builtin_ia32_cvtb2mask128: |
| 13458 | case X86::BI__builtin_ia32_cvtb2mask256: |
| 13459 | case X86::BI__builtin_ia32_cvtb2mask512: |
| 13460 | case X86::BI__builtin_ia32_cvtw2mask128: |
| 13461 | case X86::BI__builtin_ia32_cvtw2mask256: |
| 13462 | case X86::BI__builtin_ia32_cvtw2mask512: |
| 13463 | case X86::BI__builtin_ia32_cvtd2mask128: |
| 13464 | case X86::BI__builtin_ia32_cvtd2mask256: |
| 13465 | case X86::BI__builtin_ia32_cvtd2mask512: |
| 13466 | case X86::BI__builtin_ia32_cvtq2mask128: |
| 13467 | case X86::BI__builtin_ia32_cvtq2mask256: |
| 13468 | case X86::BI__builtin_ia32_cvtq2mask512: |
| 13469 | return EmitX86ConvertToMask(*this, Ops[0]); |
| 13470 | |
| 13471 | case X86::BI__builtin_ia32_cvtdq2ps512_mask: |
| 13472 | case X86::BI__builtin_ia32_cvtqq2ps512_mask: |
| 13473 | case X86::BI__builtin_ia32_cvtqq2pd512_mask: |
| 13474 | case X86::BI__builtin_ia32_vcvtw2ph512_mask: |
| 13475 | case X86::BI__builtin_ia32_vcvtdq2ph512_mask: |
| 13476 | case X86::BI__builtin_ia32_vcvtqq2ph512_mask: |
| 13477 | return EmitX86ConvertIntToFp(*this, E, Ops, /*IsSigned*/ true); |
| 13478 | case X86::BI__builtin_ia32_cvtudq2ps512_mask: |
| 13479 | case X86::BI__builtin_ia32_cvtuqq2ps512_mask: |
| 13480 | case X86::BI__builtin_ia32_cvtuqq2pd512_mask: |
| 13481 | case X86::BI__builtin_ia32_vcvtuw2ph512_mask: |
| 13482 | case X86::BI__builtin_ia32_vcvtudq2ph512_mask: |
| 13483 | case X86::BI__builtin_ia32_vcvtuqq2ph512_mask: |
| 13484 | return EmitX86ConvertIntToFp(*this, E, Ops, /*IsSigned*/ false); |
| 13485 | |
| 13486 | case X86::BI__builtin_ia32_vfmaddss3: |
| 13487 | case X86::BI__builtin_ia32_vfmaddsd3: |
| 13488 | case X86::BI__builtin_ia32_vfmaddsh3_mask: |
| 13489 | case X86::BI__builtin_ia32_vfmaddss3_mask: |
| 13490 | case X86::BI__builtin_ia32_vfmaddsd3_mask: |
| 13491 | return EmitScalarFMAExpr(*this, E, Ops, Ops[0]); |
| 13492 | case X86::BI__builtin_ia32_vfmaddss: |
| 13493 | case X86::BI__builtin_ia32_vfmaddsd: |
| 13494 | return EmitScalarFMAExpr(*this, E, Ops, |
| 13495 | Constant::getNullValue(Ops[0]->getType())); |
| 13496 | case X86::BI__builtin_ia32_vfmaddsh3_maskz: |
| 13497 | case X86::BI__builtin_ia32_vfmaddss3_maskz: |
| 13498 | case X86::BI__builtin_ia32_vfmaddsd3_maskz: |
| 13499 | return EmitScalarFMAExpr(*this, E, Ops, Ops[0], /*ZeroMask*/ true); |
| 13500 | case X86::BI__builtin_ia32_vfmaddsh3_mask3: |
| 13501 | case X86::BI__builtin_ia32_vfmaddss3_mask3: |
| 13502 | case X86::BI__builtin_ia32_vfmaddsd3_mask3: |
| 13503 | return EmitScalarFMAExpr(*this, E, Ops, Ops[2], /*ZeroMask*/ false, 2); |
| 13504 | case X86::BI__builtin_ia32_vfmsubsh3_mask3: |
| 13505 | case X86::BI__builtin_ia32_vfmsubss3_mask3: |
| 13506 | case X86::BI__builtin_ia32_vfmsubsd3_mask3: |
| 13507 | return EmitScalarFMAExpr(*this, E, Ops, Ops[2], /*ZeroMask*/ false, 2, |
| 13508 | /*NegAcc*/ true); |
| 13509 | case X86::BI__builtin_ia32_vfmaddph: |
| 13510 | case X86::BI__builtin_ia32_vfmaddps: |
| 13511 | case X86::BI__builtin_ia32_vfmaddpd: |
| 13512 | case X86::BI__builtin_ia32_vfmaddph256: |
| 13513 | case X86::BI__builtin_ia32_vfmaddps256: |
| 13514 | case X86::BI__builtin_ia32_vfmaddpd256: |
| 13515 | case X86::BI__builtin_ia32_vfmaddph512_mask: |
| 13516 | case X86::BI__builtin_ia32_vfmaddph512_maskz: |
| 13517 | case X86::BI__builtin_ia32_vfmaddph512_mask3: |
| 13518 | case X86::BI__builtin_ia32_vfmaddps512_mask: |
| 13519 | case X86::BI__builtin_ia32_vfmaddps512_maskz: |
| 13520 | case X86::BI__builtin_ia32_vfmaddps512_mask3: |
| 13521 | case X86::BI__builtin_ia32_vfmsubps512_mask3: |
| 13522 | case X86::BI__builtin_ia32_vfmaddpd512_mask: |
| 13523 | case X86::BI__builtin_ia32_vfmaddpd512_maskz: |
| 13524 | case X86::BI__builtin_ia32_vfmaddpd512_mask3: |
| 13525 | case X86::BI__builtin_ia32_vfmsubpd512_mask3: |
| 13526 | case X86::BI__builtin_ia32_vfmsubph512_mask3: |
| 13527 | return EmitX86FMAExpr(*this, E, Ops, BuiltinID, /*IsAddSub*/ false); |
| 13528 | case X86::BI__builtin_ia32_vfmaddsubph512_mask: |
| 13529 | case X86::BI__builtin_ia32_vfmaddsubph512_maskz: |
| 13530 | case X86::BI__builtin_ia32_vfmaddsubph512_mask3: |
| 13531 | case X86::BI__builtin_ia32_vfmsubaddph512_mask3: |
| 13532 | case X86::BI__builtin_ia32_vfmaddsubps512_mask: |
| 13533 | case X86::BI__builtin_ia32_vfmaddsubps512_maskz: |
| 13534 | case X86::BI__builtin_ia32_vfmaddsubps512_mask3: |
| 13535 | case X86::BI__builtin_ia32_vfmsubaddps512_mask3: |
| 13536 | case X86::BI__builtin_ia32_vfmaddsubpd512_mask: |
| 13537 | case X86::BI__builtin_ia32_vfmaddsubpd512_maskz: |
| 13538 | case X86::BI__builtin_ia32_vfmaddsubpd512_mask3: |
| 13539 | case X86::BI__builtin_ia32_vfmsubaddpd512_mask3: |
| 13540 | return EmitX86FMAExpr(*this, E, Ops, BuiltinID, /*IsAddSub*/ true); |
| 13541 | |
| 13542 | case X86::BI__builtin_ia32_movdqa32store128_mask: |
| 13543 | case X86::BI__builtin_ia32_movdqa64store128_mask: |
| 13544 | case X86::BI__builtin_ia32_storeaps128_mask: |
| 13545 | case X86::BI__builtin_ia32_storeapd128_mask: |
| 13546 | case X86::BI__builtin_ia32_movdqa32store256_mask: |
| 13547 | case X86::BI__builtin_ia32_movdqa64store256_mask: |
| 13548 | case X86::BI__builtin_ia32_storeaps256_mask: |
| 13549 | case X86::BI__builtin_ia32_storeapd256_mask: |
| 13550 | case X86::BI__builtin_ia32_movdqa32store512_mask: |
| 13551 | case X86::BI__builtin_ia32_movdqa64store512_mask: |
| 13552 | case X86::BI__builtin_ia32_storeaps512_mask: |
| 13553 | case X86::BI__builtin_ia32_storeapd512_mask: |
| 13554 | return EmitX86MaskedStore( |
| 13555 | *this, Ops, |
| 13556 | getContext().getTypeAlignInChars(E->getArg(1)->getType()).getAsAlign()); |
| 13557 | |
| 13558 | case X86::BI__builtin_ia32_loadups128_mask: |
| 13559 | case X86::BI__builtin_ia32_loadups256_mask: |
| 13560 | case X86::BI__builtin_ia32_loadups512_mask: |
| 13561 | case X86::BI__builtin_ia32_loadupd128_mask: |
| 13562 | case X86::BI__builtin_ia32_loadupd256_mask: |
| 13563 | case X86::BI__builtin_ia32_loadupd512_mask: |
| 13564 | case X86::BI__builtin_ia32_loaddquqi128_mask: |
| 13565 | case X86::BI__builtin_ia32_loaddquqi256_mask: |
| 13566 | case X86::BI__builtin_ia32_loaddquqi512_mask: |
| 13567 | case X86::BI__builtin_ia32_loaddquhi128_mask: |
| 13568 | case X86::BI__builtin_ia32_loaddquhi256_mask: |
| 13569 | case X86::BI__builtin_ia32_loaddquhi512_mask: |
| 13570 | case X86::BI__builtin_ia32_loaddqusi128_mask: |
| 13571 | case X86::BI__builtin_ia32_loaddqusi256_mask: |
| 13572 | case X86::BI__builtin_ia32_loaddqusi512_mask: |
| 13573 | case X86::BI__builtin_ia32_loaddqudi128_mask: |
| 13574 | case X86::BI__builtin_ia32_loaddqudi256_mask: |
| 13575 | case X86::BI__builtin_ia32_loaddqudi512_mask: |
| 13576 | return EmitX86MaskedLoad(*this, Ops, Align(1)); |
| 13577 | |
| 13578 | case X86::BI__builtin_ia32_loadsh128_mask: |
| 13579 | case X86::BI__builtin_ia32_loadss128_mask: |
| 13580 | case X86::BI__builtin_ia32_loadsd128_mask: |
| 13581 | return EmitX86MaskedLoad(*this, Ops, Align(1)); |
| 13582 | |
| 13583 | case X86::BI__builtin_ia32_loadaps128_mask: |
| 13584 | case X86::BI__builtin_ia32_loadaps256_mask: |
| 13585 | case X86::BI__builtin_ia32_loadaps512_mask: |
| 13586 | case X86::BI__builtin_ia32_loadapd128_mask: |
| 13587 | case X86::BI__builtin_ia32_loadapd256_mask: |
| 13588 | case X86::BI__builtin_ia32_loadapd512_mask: |
| 13589 | case X86::BI__builtin_ia32_movdqa32load128_mask: |
| 13590 | case X86::BI__builtin_ia32_movdqa32load256_mask: |
| 13591 | case X86::BI__builtin_ia32_movdqa32load512_mask: |
| 13592 | case X86::BI__builtin_ia32_movdqa64load128_mask: |
| 13593 | case X86::BI__builtin_ia32_movdqa64load256_mask: |
| 13594 | case X86::BI__builtin_ia32_movdqa64load512_mask: |
| 13595 | return EmitX86MaskedLoad( |
| 13596 | *this, Ops, |
| 13597 | getContext().getTypeAlignInChars(E->getArg(1)->getType()).getAsAlign()); |
| 13598 | |
| 13599 | case X86::BI__builtin_ia32_expandloaddf128_mask: |
| 13600 | case X86::BI__builtin_ia32_expandloaddf256_mask: |
| 13601 | case X86::BI__builtin_ia32_expandloaddf512_mask: |
| 13602 | case X86::BI__builtin_ia32_expandloadsf128_mask: |
| 13603 | case X86::BI__builtin_ia32_expandloadsf256_mask: |
| 13604 | case X86::BI__builtin_ia32_expandloadsf512_mask: |
| 13605 | case X86::BI__builtin_ia32_expandloaddi128_mask: |
| 13606 | case X86::BI__builtin_ia32_expandloaddi256_mask: |
| 13607 | case X86::BI__builtin_ia32_expandloaddi512_mask: |
| 13608 | case X86::BI__builtin_ia32_expandloadsi128_mask: |
| 13609 | case X86::BI__builtin_ia32_expandloadsi256_mask: |
| 13610 | case X86::BI__builtin_ia32_expandloadsi512_mask: |
| 13611 | case X86::BI__builtin_ia32_expandloadhi128_mask: |
| 13612 | case X86::BI__builtin_ia32_expandloadhi256_mask: |
| 13613 | case X86::BI__builtin_ia32_expandloadhi512_mask: |
| 13614 | case X86::BI__builtin_ia32_expandloadqi128_mask: |
| 13615 | case X86::BI__builtin_ia32_expandloadqi256_mask: |
| 13616 | case X86::BI__builtin_ia32_expandloadqi512_mask: |
| 13617 | return EmitX86ExpandLoad(*this, Ops); |
| 13618 | |
| 13619 | case X86::BI__builtin_ia32_compressstoredf128_mask: |
| 13620 | case X86::BI__builtin_ia32_compressstoredf256_mask: |
| 13621 | case X86::BI__builtin_ia32_compressstoredf512_mask: |
| 13622 | case X86::BI__builtin_ia32_compressstoresf128_mask: |
| 13623 | case X86::BI__builtin_ia32_compressstoresf256_mask: |
| 13624 | case X86::BI__builtin_ia32_compressstoresf512_mask: |
| 13625 | case X86::BI__builtin_ia32_compressstoredi128_mask: |
| 13626 | case X86::BI__builtin_ia32_compressstoredi256_mask: |
| 13627 | case X86::BI__builtin_ia32_compressstoredi512_mask: |
| 13628 | case X86::BI__builtin_ia32_compressstoresi128_mask: |
| 13629 | case X86::BI__builtin_ia32_compressstoresi256_mask: |
| 13630 | case X86::BI__builtin_ia32_compressstoresi512_mask: |
| 13631 | case X86::BI__builtin_ia32_compressstorehi128_mask: |
| 13632 | case X86::BI__builtin_ia32_compressstorehi256_mask: |
| 13633 | case X86::BI__builtin_ia32_compressstorehi512_mask: |
| 13634 | case X86::BI__builtin_ia32_compressstoreqi128_mask: |
| 13635 | case X86::BI__builtin_ia32_compressstoreqi256_mask: |
| 13636 | case X86::BI__builtin_ia32_compressstoreqi512_mask: |
| 13637 | return EmitX86CompressStore(*this, Ops); |
| 13638 | |
| 13639 | case X86::BI__builtin_ia32_expanddf128_mask: |
| 13640 | case X86::BI__builtin_ia32_expanddf256_mask: |
| 13641 | case X86::BI__builtin_ia32_expanddf512_mask: |
| 13642 | case X86::BI__builtin_ia32_expandsf128_mask: |
| 13643 | case X86::BI__builtin_ia32_expandsf256_mask: |
| 13644 | case X86::BI__builtin_ia32_expandsf512_mask: |
| 13645 | case X86::BI__builtin_ia32_expanddi128_mask: |
| 13646 | case X86::BI__builtin_ia32_expanddi256_mask: |
| 13647 | case X86::BI__builtin_ia32_expanddi512_mask: |
| 13648 | case X86::BI__builtin_ia32_expandsi128_mask: |
| 13649 | case X86::BI__builtin_ia32_expandsi256_mask: |
| 13650 | case X86::BI__builtin_ia32_expandsi512_mask: |
| 13651 | case X86::BI__builtin_ia32_expandhi128_mask: |
| 13652 | case X86::BI__builtin_ia32_expandhi256_mask: |
| 13653 | case X86::BI__builtin_ia32_expandhi512_mask: |
| 13654 | case X86::BI__builtin_ia32_expandqi128_mask: |
| 13655 | case X86::BI__builtin_ia32_expandqi256_mask: |
| 13656 | case X86::BI__builtin_ia32_expandqi512_mask: |
| 13657 | return EmitX86CompressExpand(*this, Ops, /*IsCompress*/false); |
| 13658 | |
| 13659 | case X86::BI__builtin_ia32_compressdf128_mask: |
| 13660 | case X86::BI__builtin_ia32_compressdf256_mask: |
| 13661 | case X86::BI__builtin_ia32_compressdf512_mask: |
| 13662 | case X86::BI__builtin_ia32_compresssf128_mask: |
| 13663 | case X86::BI__builtin_ia32_compresssf256_mask: |
| 13664 | case X86::BI__builtin_ia32_compresssf512_mask: |
| 13665 | case X86::BI__builtin_ia32_compressdi128_mask: |
| 13666 | case X86::BI__builtin_ia32_compressdi256_mask: |
| 13667 | case X86::BI__builtin_ia32_compressdi512_mask: |
| 13668 | case X86::BI__builtin_ia32_compresssi128_mask: |
| 13669 | case X86::BI__builtin_ia32_compresssi256_mask: |
| 13670 | case X86::BI__builtin_ia32_compresssi512_mask: |
| 13671 | case X86::BI__builtin_ia32_compresshi128_mask: |
| 13672 | case X86::BI__builtin_ia32_compresshi256_mask: |
| 13673 | case X86::BI__builtin_ia32_compresshi512_mask: |
| 13674 | case X86::BI__builtin_ia32_compressqi128_mask: |
| 13675 | case X86::BI__builtin_ia32_compressqi256_mask: |
| 13676 | case X86::BI__builtin_ia32_compressqi512_mask: |
| 13677 | return EmitX86CompressExpand(*this, Ops, /*IsCompress*/true); |
| 13678 | |
| 13679 | case X86::BI__builtin_ia32_gather3div2df: |
| 13680 | case X86::BI__builtin_ia32_gather3div2di: |
| 13681 | case X86::BI__builtin_ia32_gather3div4df: |
| 13682 | case X86::BI__builtin_ia32_gather3div4di: |
| 13683 | case X86::BI__builtin_ia32_gather3div4sf: |
| 13684 | case X86::BI__builtin_ia32_gather3div4si: |
| 13685 | case X86::BI__builtin_ia32_gather3div8sf: |
| 13686 | case X86::BI__builtin_ia32_gather3div8si: |
| 13687 | case X86::BI__builtin_ia32_gather3siv2df: |
| 13688 | case X86::BI__builtin_ia32_gather3siv2di: |
| 13689 | case X86::BI__builtin_ia32_gather3siv4df: |
| 13690 | case X86::BI__builtin_ia32_gather3siv4di: |
| 13691 | case X86::BI__builtin_ia32_gather3siv4sf: |
| 13692 | case X86::BI__builtin_ia32_gather3siv4si: |
| 13693 | case X86::BI__builtin_ia32_gather3siv8sf: |
| 13694 | case X86::BI__builtin_ia32_gather3siv8si: |
| 13695 | case X86::BI__builtin_ia32_gathersiv8df: |
| 13696 | case X86::BI__builtin_ia32_gathersiv16sf: |
| 13697 | case X86::BI__builtin_ia32_gatherdiv8df: |
| 13698 | case X86::BI__builtin_ia32_gatherdiv16sf: |
| 13699 | case X86::BI__builtin_ia32_gathersiv8di: |
| 13700 | case X86::BI__builtin_ia32_gathersiv16si: |
| 13701 | case X86::BI__builtin_ia32_gatherdiv8di: |
| 13702 | case X86::BI__builtin_ia32_gatherdiv16si: { |
| 13703 | Intrinsic::ID IID; |
| 13704 | switch (BuiltinID) { |
| 13705 | default: llvm_unreachable("Unexpected builtin")::llvm::llvm_unreachable_internal("Unexpected builtin", "clang/lib/CodeGen/CGBuiltin.cpp" , 13705); |
| 13706 | case X86::BI__builtin_ia32_gather3div2df: |
| 13707 | IID = Intrinsic::x86_avx512_mask_gather3div2_df; |
| 13708 | break; |
| 13709 | case X86::BI__builtin_ia32_gather3div2di: |
| 13710 | IID = Intrinsic::x86_avx512_mask_gather3div2_di; |
| 13711 | break; |
| 13712 | case X86::BI__builtin_ia32_gather3div4df: |
| 13713 | IID = Intrinsic::x86_avx512_mask_gather3div4_df; |
| 13714 | break; |
| 13715 | case X86::BI__builtin_ia32_gather3div4di: |
| 13716 | IID = Intrinsic::x86_avx512_mask_gather3div4_di; |
| 13717 | break; |
| 13718 | case X86::BI__builtin_ia32_gather3div4sf: |
| 13719 | IID = Intrinsic::x86_avx512_mask_gather3div4_sf; |
| 13720 | break; |
| 13721 | case X86::BI__builtin_ia32_gather3div4si: |
| 13722 | IID = Intrinsic::x86_avx512_mask_gather3div4_si; |
| 13723 | break; |
| 13724 | case X86::BI__builtin_ia32_gather3div8sf: |
| 13725 | IID = Intrinsic::x86_avx512_mask_gather3div8_sf; |
| 13726 | break; |
| 13727 | case X86::BI__builtin_ia32_gather3div8si: |
| 13728 | IID = Intrinsic::x86_avx512_mask_gather3div8_si; |
| 13729 | break; |
| 13730 | case X86::BI__builtin_ia32_gather3siv2df: |
| 13731 | IID = Intrinsic::x86_avx512_mask_gather3siv2_df; |
| 13732 | break; |
| 13733 | case X86::BI__builtin_ia32_gather3siv2di: |
| 13734 | IID = Intrinsic::x86_avx512_mask_gather3siv2_di; |
| 13735 | break; |
| 13736 | case X86::BI__builtin_ia32_gather3siv4df: |
| 13737 | IID = Intrinsic::x86_avx512_mask_gather3siv4_df; |
| 13738 | break; |
| 13739 | case X86::BI__builtin_ia32_gather3siv4di: |
| 13740 | IID = Intrinsic::x86_avx512_mask_gather3siv4_di; |
| 13741 | break; |
| 13742 | case X86::BI__builtin_ia32_gather3siv4sf: |
| 13743 | IID = Intrinsic::x86_avx512_mask_gather3siv4_sf; |
| 13744 | break; |
| 13745 | case X86::BI__builtin_ia32_gather3siv4si: |
| 13746 | IID = Intrinsic::x86_avx512_mask_gather3siv4_si; |
| 13747 | break; |
| 13748 | case X86::BI__builtin_ia32_gather3siv8sf: |
| 13749 | IID = Intrinsic::x86_avx512_mask_gather3siv8_sf; |
| 13750 | break; |
| 13751 | case X86::BI__builtin_ia32_gather3siv8si: |
| 13752 | IID = Intrinsic::x86_avx512_mask_gather3siv8_si; |
| 13753 | break; |
| 13754 | case X86::BI__builtin_ia32_gathersiv8df: |
| 13755 | IID = Intrinsic::x86_avx512_mask_gather_dpd_512; |
| 13756 | break; |
| 13757 | case X86::BI__builtin_ia32_gathersiv16sf: |
| 13758 | IID = Intrinsic::x86_avx512_mask_gather_dps_512; |
| 13759 | break; |
| 13760 | case X86::BI__builtin_ia32_gatherdiv8df: |
| 13761 | IID = Intrinsic::x86_avx512_mask_gather_qpd_512; |
| 13762 | break; |
| 13763 | case X86::BI__builtin_ia32_gatherdiv16sf: |
| 13764 | IID = Intrinsic::x86_avx512_mask_gather_qps_512; |
| 13765 | break; |
| 13766 | case X86::BI__builtin_ia32_gathersiv8di: |
| 13767 | IID = Intrinsic::x86_avx512_mask_gather_dpq_512; |
| 13768 | break; |
| 13769 | case X86::BI__builtin_ia32_gathersiv16si: |
| 13770 | IID = Intrinsic::x86_avx512_mask_gather_dpi_512; |
| 13771 | break; |
| 13772 | case X86::BI__builtin_ia32_gatherdiv8di: |
| 13773 | IID = Intrinsic::x86_avx512_mask_gather_qpq_512; |
| 13774 | break; |
| 13775 | case X86::BI__builtin_ia32_gatherdiv16si: |
| 13776 | IID = Intrinsic::x86_avx512_mask_gather_qpi_512; |
| 13777 | break; |
| 13778 | } |
| 13779 | |
| 13780 | unsigned MinElts = std::min( |
| 13781 | cast<llvm::FixedVectorType>(Ops[0]->getType())->getNumElements(), |
| 13782 | cast<llvm::FixedVectorType>(Ops[2]->getType())->getNumElements()); |
| 13783 | Ops[3] = getMaskVecValue(*this, Ops[3], MinElts); |
| 13784 | Function *Intr = CGM.getIntrinsic(IID); |
| 13785 | return Builder.CreateCall(Intr, Ops); |
| 13786 | } |
| 13787 | |
| 13788 | case X86::BI__builtin_ia32_scattersiv8df: |
| 13789 | case X86::BI__builtin_ia32_scattersiv16sf: |
| 13790 | case X86::BI__builtin_ia32_scatterdiv8df: |
| 13791 | case X86::BI__builtin_ia32_scatterdiv16sf: |
| 13792 | case X86::BI__builtin_ia32_scattersiv8di: |
| 13793 | case X86::BI__builtin_ia32_scattersiv16si: |
| 13794 | case X86::BI__builtin_ia32_scatterdiv8di: |
| 13795 | case X86::BI__builtin_ia32_scatterdiv16si: |
| 13796 | case X86::BI__builtin_ia32_scatterdiv2df: |
| 13797 | case X86::BI__builtin_ia32_scatterdiv2di: |
| 13798 | case X86::BI__builtin_ia32_scatterdiv4df: |
| 13799 | case X86::BI__builtin_ia32_scatterdiv4di: |
| 13800 | case X86::BI__builtin_ia32_scatterdiv4sf: |
| 13801 | case X86::BI__builtin_ia32_scatterdiv4si: |
| 13802 | case X86::BI__builtin_ia32_scatterdiv8sf: |
| 13803 | case X86::BI__builtin_ia32_scatterdiv8si: |
| 13804 | case X86::BI__builtin_ia32_scattersiv2df: |
| 13805 | case X86::BI__builtin_ia32_scattersiv2di: |
| 13806 | case X86::BI__builtin_ia32_scattersiv4df: |
| 13807 | case X86::BI__builtin_ia32_scattersiv4di: |
| 13808 | case X86::BI__builtin_ia32_scattersiv4sf: |
| 13809 | case X86::BI__builtin_ia32_scattersiv4si: |
| 13810 | case X86::BI__builtin_ia32_scattersiv8sf: |
| 13811 | case X86::BI__builtin_ia32_scattersiv8si: { |
| 13812 | Intrinsic::ID IID; |
| 13813 | switch (BuiltinID) { |
| 13814 | default: llvm_unreachable("Unexpected builtin")::llvm::llvm_unreachable_internal("Unexpected builtin", "clang/lib/CodeGen/CGBuiltin.cpp" , 13814); |
| 13815 | case X86::BI__builtin_ia32_scattersiv8df: |
| 13816 | IID = Intrinsic::x86_avx512_mask_scatter_dpd_512; |
| 13817 | break; |
| 13818 | case X86::BI__builtin_ia32_scattersiv16sf: |
| 13819 | IID = Intrinsic::x86_avx512_mask_scatter_dps_512; |
| 13820 | break; |
| 13821 | case X86::BI__builtin_ia32_scatterdiv8df: |
| 13822 | IID = Intrinsic::x86_avx512_mask_scatter_qpd_512; |
| 13823 | break; |
| 13824 | case X86::BI__builtin_ia32_scatterdiv16sf: |
| 13825 | IID = Intrinsic::x86_avx512_mask_scatter_qps_512; |
| 13826 | break; |
| 13827 | case X86::BI__builtin_ia32_scattersiv8di: |
| 13828 | IID = Intrinsic::x86_avx512_mask_scatter_dpq_512; |
| 13829 | break; |
| 13830 | case X86::BI__builtin_ia32_scattersiv16si: |
| 13831 | IID = Intrinsic::x86_avx512_mask_scatter_dpi_512; |
| 13832 | break; |
| 13833 | case X86::BI__builtin_ia32_scatterdiv8di: |
| 13834 | IID = Intrinsic::x86_avx512_mask_scatter_qpq_512; |
| 13835 | break; |
| 13836 | case X86::BI__builtin_ia32_scatterdiv16si: |
| 13837 | IID = Intrinsic::x86_avx512_mask_scatter_qpi_512; |
| 13838 | break; |
| 13839 | case X86::BI__builtin_ia32_scatterdiv2df: |
| 13840 | IID = Intrinsic::x86_avx512_mask_scatterdiv2_df; |
| 13841 | break; |
| 13842 | case X86::BI__builtin_ia32_scatterdiv2di: |
| 13843 | IID = Intrinsic::x86_avx512_mask_scatterdiv2_di; |
| 13844 | break; |
| 13845 | case X86::BI__builtin_ia32_scatterdiv4df: |
| 13846 | IID = Intrinsic::x86_avx512_mask_scatterdiv4_df; |
| 13847 | break; |
| 13848 | case X86::BI__builtin_ia32_scatterdiv4di: |
| 13849 | IID = Intrinsic::x86_avx512_mask_scatterdiv4_di; |
| 13850 | break; |
| 13851 | case X86::BI__builtin_ia32_scatterdiv4sf: |
| 13852 | IID = Intrinsic::x86_avx512_mask_scatterdiv4_sf; |
| 13853 | break; |
| 13854 | case X86::BI__builtin_ia32_scatterdiv4si: |
| 13855 | IID = Intrinsic::x86_avx512_mask_scatterdiv4_si; |
| 13856 | break; |
| 13857 | case X86::BI__builtin_ia32_scatterdiv8sf: |
| 13858 | IID = Intrinsic::x86_avx512_mask_scatterdiv8_sf; |
| 13859 | break; |
| 13860 | case X86::BI__builtin_ia32_scatterdiv8si: |
| 13861 | IID = Intrinsic::x86_avx512_mask_scatterdiv8_si; |
| 13862 | break; |
| 13863 | case X86::BI__builtin_ia32_scattersiv2df: |
| 13864 | IID = Intrinsic::x86_avx512_mask_scattersiv2_df; |
| 13865 | break; |
| 13866 | case X86::BI__builtin_ia32_scattersiv2di: |
| 13867 | IID = Intrinsic::x86_avx512_mask_scattersiv2_di; |
| 13868 | break; |
| 13869 | case X86::BI__builtin_ia32_scattersiv4df: |
| 13870 | IID = Intrinsic::x86_avx512_mask_scattersiv4_df; |
| 13871 | break; |
| 13872 | case X86::BI__builtin_ia32_scattersiv4di: |
| 13873 | IID = Intrinsic::x86_avx512_mask_scattersiv4_di; |
| 13874 | break; |
| 13875 | case X86::BI__builtin_ia32_scattersiv4sf: |
| 13876 | IID = Intrinsic::x86_avx512_mask_scattersiv4_sf; |
| 13877 | break; |
| 13878 | case X86::BI__builtin_ia32_scattersiv4si: |
| 13879 | IID = Intrinsic::x86_avx512_mask_scattersiv4_si; |
| 13880 | break; |
| 13881 | case X86::BI__builtin_ia32_scattersiv8sf: |
| 13882 | IID = Intrinsic::x86_avx512_mask_scattersiv8_sf; |
| 13883 | break; |
| 13884 | case X86::BI__builtin_ia32_scattersiv8si: |
| 13885 | IID = Intrinsic::x86_avx512_mask_scattersiv8_si; |
| 13886 | break; |
| 13887 | } |
| 13888 | |
| 13889 | unsigned MinElts = std::min( |
| 13890 | cast<llvm::FixedVectorType>(Ops[2]->getType())->getNumElements(), |
| 13891 | cast<llvm::FixedVectorType>(Ops[3]->getType())->getNumElements()); |
| 13892 | Ops[1] = getMaskVecValue(*this, Ops[1], MinElts); |
| 13893 | Function *Intr = CGM.getIntrinsic(IID); |
| 13894 | return Builder.CreateCall(Intr, Ops); |
| 13895 | } |
| 13896 | |
| 13897 | case X86::BI__builtin_ia32_vextractf128_pd256: |
| 13898 | case X86::BI__builtin_ia32_vextractf128_ps256: |
| 13899 | case X86::BI__builtin_ia32_vextractf128_si256: |
| 13900 | case X86::BI__builtin_ia32_extract128i256: |
| 13901 | case X86::BI__builtin_ia32_extractf64x4_mask: |
| 13902 | case X86::BI__builtin_ia32_extractf32x4_mask: |
| 13903 | case X86::BI__builtin_ia32_extracti64x4_mask: |
| 13904 | case X86::BI__builtin_ia32_extracti32x4_mask: |
| 13905 | case X86::BI__builtin_ia32_extractf32x8_mask: |
| 13906 | case X86::BI__builtin_ia32_extracti32x8_mask: |
| 13907 | case X86::BI__builtin_ia32_extractf32x4_256_mask: |
| 13908 | case X86::BI__builtin_ia32_extracti32x4_256_mask: |
| 13909 | case X86::BI__builtin_ia32_extractf64x2_256_mask: |
| 13910 | case X86::BI__builtin_ia32_extracti64x2_256_mask: |
| 13911 | case X86::BI__builtin_ia32_extractf64x2_512_mask: |
| 13912 | case X86::BI__builtin_ia32_extracti64x2_512_mask: { |
| 13913 | auto *DstTy = cast<llvm::FixedVectorType>(ConvertType(E->getType())); |
| 13914 | unsigned NumElts = DstTy->getNumElements(); |
| 13915 | unsigned SrcNumElts = |
| 13916 | cast<llvm::FixedVectorType>(Ops[0]->getType())->getNumElements(); |
| 13917 | unsigned SubVectors = SrcNumElts / NumElts; |
| 13918 | unsigned Index = cast<ConstantInt>(Ops[1])->getZExtValue(); |
| 13919 | assert(llvm::isPowerOf2_32(SubVectors) && "Expected power of 2 subvectors")(static_cast <bool> (llvm::isPowerOf2_32(SubVectors) && "Expected power of 2 subvectors") ? void (0) : __assert_fail ("llvm::isPowerOf2_32(SubVectors) && \"Expected power of 2 subvectors\"" , "clang/lib/CodeGen/CGBuiltin.cpp", 13919, __extension__ __PRETTY_FUNCTION__ )); |
| 13920 | Index &= SubVectors - 1; // Remove any extra bits. |
| 13921 | Index *= NumElts; |
| 13922 | |
| 13923 | int Indices[16]; |
| 13924 | for (unsigned i = 0; i != NumElts; ++i) |
| 13925 | Indices[i] = i + Index; |
| 13926 | |
| 13927 | Value *Res = Builder.CreateShuffleVector(Ops[0], ArrayRef(Indices, NumElts), |
| 13928 | "extract"); |
| 13929 | |
| 13930 | if (Ops.size() == 4) |
| 13931 | Res = EmitX86Select(*this, Ops[3], Res, Ops[2]); |
| 13932 | |
| 13933 | return Res; |
| 13934 | } |
| 13935 | case X86::BI__builtin_ia32_vinsertf128_pd256: |
| 13936 | case X86::BI__builtin_ia32_vinsertf128_ps256: |
| 13937 | case X86::BI__builtin_ia32_vinsertf128_si256: |
| 13938 | case X86::BI__builtin_ia32_insert128i256: |
| 13939 | case X86::BI__builtin_ia32_insertf64x4: |
| 13940 | case X86::BI__builtin_ia32_insertf32x4: |
| 13941 | case X86::BI__builtin_ia32_inserti64x4: |
| 13942 | case X86::BI__builtin_ia32_inserti32x4: |
| 13943 | case X86::BI__builtin_ia32_insertf32x8: |
| 13944 | case X86::BI__builtin_ia32_inserti32x8: |
| 13945 | case X86::BI__builtin_ia32_insertf32x4_256: |
| 13946 | case X86::BI__builtin_ia32_inserti32x4_256: |
| 13947 | case X86::BI__builtin_ia32_insertf64x2_256: |
| 13948 | case X86::BI__builtin_ia32_inserti64x2_256: |
| 13949 | case X86::BI__builtin_ia32_insertf64x2_512: |
| 13950 | case X86::BI__builtin_ia32_inserti64x2_512: { |
| 13951 | unsigned DstNumElts = |
| 13952 | cast<llvm::FixedVectorType>(Ops[0]->getType())->getNumElements(); |
| 13953 | unsigned SrcNumElts = |
| 13954 | cast<llvm::FixedVectorType>(Ops[1]->getType())->getNumElements(); |
| 13955 | unsigned SubVectors = DstNumElts / SrcNumElts; |
| 13956 | unsigned Index = cast<ConstantInt>(Ops[2])->getZExtValue(); |
| 13957 | assert(llvm::isPowerOf2_32(SubVectors) && "Expected power of 2 subvectors")(static_cast <bool> (llvm::isPowerOf2_32(SubVectors) && "Expected power of 2 subvectors") ? void (0) : __assert_fail ("llvm::isPowerOf2_32(SubVectors) && \"Expected power of 2 subvectors\"" , "clang/lib/CodeGen/CGBuiltin.cpp", 13957, __extension__ __PRETTY_FUNCTION__ )); |
| 13958 | Index &= SubVectors - 1; // Remove any extra bits. |
| 13959 | Index *= SrcNumElts; |
| 13960 | |
| 13961 | int Indices[16]; |
| 13962 | for (unsigned i = 0; i != DstNumElts; ++i) |
| 13963 | Indices[i] = (i >= SrcNumElts) ? SrcNumElts + (i % SrcNumElts) : i; |
| 13964 | |
| 13965 | Value *Op1 = Builder.CreateShuffleVector( |
| 13966 | Ops[1], ArrayRef(Indices, DstNumElts), "widen"); |
| 13967 | |
| 13968 | for (unsigned i = 0; i != DstNumElts; ++i) { |
| 13969 | if (i >= Index && i < (Index + SrcNumElts)) |
| 13970 | Indices[i] = (i - Index) + DstNumElts; |
| 13971 | else |
| 13972 | Indices[i] = i; |
| 13973 | } |
| 13974 | |
| 13975 | return Builder.CreateShuffleVector(Ops[0], Op1, |
| 13976 | ArrayRef(Indices, DstNumElts), "insert"); |
| 13977 | } |
| 13978 | case X86::BI__builtin_ia32_pmovqd512_mask: |
| 13979 | case X86::BI__builtin_ia32_pmovwb512_mask: { |
| 13980 | Value *Res = Builder.CreateTrunc(Ops[0], Ops[1]->getType()); |
| 13981 | return EmitX86Select(*this, Ops[2], Res, Ops[1]); |
| 13982 | } |
| 13983 | case X86::BI__builtin_ia32_pmovdb512_mask: |
| 13984 | case X86::BI__builtin_ia32_pmovdw512_mask: |
| 13985 | case X86::BI__builtin_ia32_pmovqw512_mask: { |
| 13986 | if (const auto *C = dyn_cast<Constant>(Ops[2])) |
| 13987 | if (C->isAllOnesValue()) |
| 13988 | return Builder.CreateTrunc(Ops[0], Ops[1]->getType()); |
| 13989 | |
| 13990 | Intrinsic::ID IID; |
| 13991 | switch (BuiltinID) { |
| 13992 | default: llvm_unreachable("Unsupported intrinsic!")::llvm::llvm_unreachable_internal("Unsupported intrinsic!", "clang/lib/CodeGen/CGBuiltin.cpp" , 13992); |
| 13993 | case X86::BI__builtin_ia32_pmovdb512_mask: |
| 13994 | IID = Intrinsic::x86_avx512_mask_pmov_db_512; |
| 13995 | break; |
| 13996 | case X86::BI__builtin_ia32_pmovdw512_mask: |
| 13997 | IID = Intrinsic::x86_avx512_mask_pmov_dw_512; |
| 13998 | break; |
| 13999 | case X86::BI__builtin_ia32_pmovqw512_mask: |
| 14000 | IID = Intrinsic::x86_avx512_mask_pmov_qw_512; |
| 14001 | break; |
| 14002 | } |
| 14003 | |
| 14004 | Function *Intr = CGM.getIntrinsic(IID); |
| 14005 | return Builder.CreateCall(Intr, Ops); |
| 14006 | } |
| 14007 | case X86::BI__builtin_ia32_pblendw128: |
| 14008 | case X86::BI__builtin_ia32_blendpd: |
| 14009 | case X86::BI__builtin_ia32_blendps: |
| 14010 | case X86::BI__builtin_ia32_blendpd256: |
| 14011 | case X86::BI__builtin_ia32_blendps256: |
| 14012 | case X86::BI__builtin_ia32_pblendw256: |
| 14013 | case X86::BI__builtin_ia32_pblendd128: |
| 14014 | case X86::BI__builtin_ia32_pblendd256: { |
| 14015 | unsigned NumElts = |
| 14016 | cast<llvm::FixedVectorType>(Ops[0]->getType())->getNumElements(); |
| 14017 | unsigned Imm = cast<llvm::ConstantInt>(Ops[2])->getZExtValue(); |
| 14018 | |
| 14019 | int Indices[16]; |
| 14020 | // If there are more than 8 elements, the immediate is used twice so make |
| 14021 | // sure we handle that. |
| 14022 | for (unsigned i = 0; i != NumElts; ++i) |
| 14023 | Indices[i] = ((Imm >> (i % 8)) & 0x1) ? NumElts + i : i; |
| 14024 | |
| 14025 | return Builder.CreateShuffleVector(Ops[0], Ops[1], |
| 14026 | ArrayRef(Indices, NumElts), "blend"); |
| 14027 | } |
| 14028 | case X86::BI__builtin_ia32_pshuflw: |
| 14029 | case X86::BI__builtin_ia32_pshuflw256: |
| 14030 | case X86::BI__builtin_ia32_pshuflw512: { |
| 14031 | uint32_t Imm = cast<llvm::ConstantInt>(Ops[1])->getZExtValue(); |
| 14032 | auto *Ty = cast<llvm::FixedVectorType>(Ops[0]->getType()); |
| 14033 | unsigned NumElts = Ty->getNumElements(); |
| 14034 | |
| 14035 | // Splat the 8-bits of immediate 4 times to help the loop wrap around. |
| 14036 | Imm = (Imm & 0xff) * 0x01010101; |
| 14037 | |
| 14038 | int Indices[32]; |
| 14039 | for (unsigned l = 0; l != NumElts; l += 8) { |
| 14040 | for (unsigned i = 0; i != 4; ++i) { |
| 14041 | Indices[l + i] = l + (Imm & 3); |
| 14042 | Imm >>= 2; |
| 14043 | } |
| 14044 | for (unsigned i = 4; i != 8; ++i) |
| 14045 | Indices[l + i] = l + i; |
| 14046 | } |
| 14047 | |
| 14048 | return Builder.CreateShuffleVector(Ops[0], ArrayRef(Indices, NumElts), |
| 14049 | "pshuflw"); |
| 14050 | } |
| 14051 | case X86::BI__builtin_ia32_pshufhw: |
| 14052 | case X86::BI__builtin_ia32_pshufhw256: |
| 14053 | case X86::BI__builtin_ia32_pshufhw512: { |
| 14054 | uint32_t Imm = cast<llvm::ConstantInt>(Ops[1])->getZExtValue(); |
| 14055 | auto *Ty = cast<llvm::FixedVectorType>(Ops[0]->getType()); |
| 14056 | unsigned NumElts = Ty->getNumElements(); |
| 14057 | |
| 14058 | // Splat the 8-bits of immediate 4 times to help the loop wrap around. |
| 14059 | Imm = (Imm & 0xff) * 0x01010101; |
| 14060 | |
| 14061 | int Indices[32]; |
| 14062 | for (unsigned l = 0; l != NumElts; l += 8) { |
| 14063 | for (unsigned i = 0; i != 4; ++i) |
| 14064 | Indices[l + i] = l + i; |
| 14065 | for (unsigned i = 4; i != 8; ++i) { |
| 14066 | Indices[l + i] = l + 4 + (Imm & 3); |
| 14067 | Imm >>= 2; |
| 14068 | } |
| 14069 | } |
| 14070 | |
| 14071 | return Builder.CreateShuffleVector(Ops[0], ArrayRef(Indices, NumElts), |
| 14072 | "pshufhw"); |
| 14073 | } |
| 14074 | case X86::BI__builtin_ia32_pshufd: |
| 14075 | case X86::BI__builtin_ia32_pshufd256: |
| 14076 | case X86::BI__builtin_ia32_pshufd512: |
| 14077 | case X86::BI__builtin_ia32_vpermilpd: |
| 14078 | case X86::BI__builtin_ia32_vpermilps: |
| 14079 | case X86::BI__builtin_ia32_vpermilpd256: |
| 14080 | case X86::BI__builtin_ia32_vpermilps256: |
| 14081 | case X86::BI__builtin_ia32_vpermilpd512: |
| 14082 | case X86::BI__builtin_ia32_vpermilps512: { |
| 14083 | uint32_t Imm = cast<llvm::ConstantInt>(Ops[1])->getZExtValue(); |
| 14084 | auto *Ty = cast<llvm::FixedVectorType>(Ops[0]->getType()); |
| 14085 | unsigned NumElts = Ty->getNumElements(); |
| 14086 | unsigned NumLanes = Ty->getPrimitiveSizeInBits() / 128; |
| 14087 | unsigned NumLaneElts = NumElts / NumLanes; |
| 14088 | |
| 14089 | // Splat the 8-bits of immediate 4 times to help the loop wrap around. |
| 14090 | Imm = (Imm & 0xff) * 0x01010101; |
| 14091 | |
| 14092 | int Indices[16]; |
| 14093 | for (unsigned l = 0; l != NumElts; l += NumLaneElts) { |
| 14094 | for (unsigned i = 0; i != NumLaneElts; ++i) { |
| 14095 | Indices[i + l] = (Imm % NumLaneElts) + l; |
| 14096 | Imm /= NumLaneElts; |
| 14097 | } |
| 14098 | } |
| 14099 | |
| 14100 | return Builder.CreateShuffleVector(Ops[0], ArrayRef(Indices, NumElts), |
| 14101 | "permil"); |
| 14102 | } |
| 14103 | case X86::BI__builtin_ia32_shufpd: |
| 14104 | case X86::BI__builtin_ia32_shufpd256: |
| 14105 | case X86::BI__builtin_ia32_shufpd512: |
| 14106 | case X86::BI__builtin_ia32_shufps: |
| 14107 | case X86::BI__builtin_ia32_shufps256: |
| 14108 | case X86::BI__builtin_ia32_shufps512: { |
| 14109 | uint32_t Imm = cast<llvm::ConstantInt>(Ops[2])->getZExtValue(); |
| 14110 | auto *Ty = cast<llvm::FixedVectorType>(Ops[0]->getType()); |
| 14111 | unsigned NumElts = Ty->getNumElements(); |
| 14112 | unsigned NumLanes = Ty->getPrimitiveSizeInBits() / 128; |
| 14113 | unsigned NumLaneElts = NumElts / NumLanes; |
| 14114 | |
| 14115 | // Splat the 8-bits of immediate 4 times to help the loop wrap around. |
| 14116 | Imm = (Imm & 0xff) * 0x01010101; |
| 14117 | |
| 14118 | int Indices[16]; |
| 14119 | for (unsigned l = 0; l != NumElts; l += NumLaneElts) { |
| 14120 | for (unsigned i = 0; i != NumLaneElts; ++i) { |
| 14121 | unsigned Index = Imm % NumLaneElts; |
| 14122 | Imm /= NumLaneElts; |
| 14123 | if (i >= (NumLaneElts / 2)) |
| 14124 | Index += NumElts; |
| 14125 | Indices[l + i] = l + Index; |
| 14126 | } |
| 14127 | } |
| 14128 | |
| 14129 | return Builder.CreateShuffleVector(Ops[0], Ops[1], |
| 14130 | ArrayRef(Indices, NumElts), "shufp"); |
| 14131 | } |
| 14132 | case X86::BI__builtin_ia32_permdi256: |
| 14133 | case X86::BI__builtin_ia32_permdf256: |
| 14134 | case X86::BI__builtin_ia32_permdi512: |
| 14135 | case X86::BI__builtin_ia32_permdf512: { |
| 14136 | unsigned Imm = cast<llvm::ConstantInt>(Ops[1])->getZExtValue(); |
| 14137 | auto *Ty = cast<llvm::FixedVectorType>(Ops[0]->getType()); |
| 14138 | unsigned NumElts = Ty->getNumElements(); |
| 14139 | |
| 14140 | // These intrinsics operate on 256-bit lanes of four 64-bit elements. |
| 14141 | int Indices[8]; |
| 14142 | for (unsigned l = 0; l != NumElts; l += 4) |
| 14143 | for (unsigned i = 0; i != 4; ++i) |
| 14144 | Indices[l + i] = l + ((Imm >> (2 * i)) & 0x3); |
| 14145 | |
| 14146 | return Builder.CreateShuffleVector(Ops[0], ArrayRef(Indices, NumElts), |
| 14147 | "perm"); |
| 14148 | } |
| 14149 | case X86::BI__builtin_ia32_palignr128: |
| 14150 | case X86::BI__builtin_ia32_palignr256: |
| 14151 | case X86::BI__builtin_ia32_palignr512: { |
| 14152 | unsigned ShiftVal = cast<llvm::ConstantInt>(Ops[2])->getZExtValue() & 0xff; |
| 14153 | |
| 14154 | unsigned NumElts = |
| 14155 | cast<llvm::FixedVectorType>(Ops[0]->getType())->getNumElements(); |
| 14156 | assert(NumElts % 16 == 0)(static_cast <bool> (NumElts % 16 == 0) ? void (0) : __assert_fail ("NumElts % 16 == 0", "clang/lib/CodeGen/CGBuiltin.cpp", 14156 , __extension__ __PRETTY_FUNCTION__)); |
| 14157 | |
| 14158 | // If palignr is shifting the pair of vectors more than the size of two |
| 14159 | // lanes, emit zero. |
| 14160 | if (ShiftVal >= 32) |
| 14161 | return llvm::Constant::getNullValue(ConvertType(E->getType())); |
| 14162 | |
| 14163 | // If palignr is shifting the pair of input vectors more than one lane, |
| 14164 | // but less than two lanes, convert to shifting in zeroes. |
| 14165 | if (ShiftVal > 16) { |
| 14166 | ShiftVal -= 16; |
| 14167 | Ops[1] = Ops[0]; |
| 14168 | Ops[0] = llvm::Constant::getNullValue(Ops[0]->getType()); |
| 14169 | } |
| 14170 | |
| 14171 | int Indices[64]; |
| 14172 | // 256-bit palignr operates on 128-bit lanes so we need to handle that |
| 14173 | for (unsigned l = 0; l != NumElts; l += 16) { |
| 14174 | for (unsigned i = 0; i != 16; ++i) { |
| 14175 | unsigned Idx = ShiftVal + i; |
| 14176 | if (Idx >= 16) |
| 14177 | Idx += NumElts - 16; // End of lane, switch operand. |
| 14178 | Indices[l + i] = Idx + l; |
| 14179 | } |
| 14180 | } |
| 14181 | |
| 14182 | return Builder.CreateShuffleVector(Ops[1], Ops[0], |
| 14183 | ArrayRef(Indices, NumElts), "palignr"); |
| 14184 | } |
| 14185 | case X86::BI__builtin_ia32_alignd128: |
| 14186 | case X86::BI__builtin_ia32_alignd256: |
| 14187 | case X86::BI__builtin_ia32_alignd512: |
| 14188 | case X86::BI__builtin_ia32_alignq128: |
| 14189 | case X86::BI__builtin_ia32_alignq256: |
| 14190 | case X86::BI__builtin_ia32_alignq512: { |
| 14191 | unsigned NumElts = |
| 14192 | cast<llvm::FixedVectorType>(Ops[0]->getType())->getNumElements(); |
| 14193 | unsigned ShiftVal = cast<llvm::ConstantInt>(Ops[2])->getZExtValue() & 0xff; |
| 14194 | |
| 14195 | // Mask the shift amount to width of a vector. |
| 14196 | ShiftVal &= NumElts - 1; |
| 14197 | |
| 14198 | int Indices[16]; |
| 14199 | for (unsigned i = 0; i != NumElts; ++i) |
| 14200 | Indices[i] = i + ShiftVal; |
| 14201 | |
| 14202 | return Builder.CreateShuffleVector(Ops[1], Ops[0], |
| 14203 | ArrayRef(Indices, NumElts), "valign"); |
| 14204 | } |
| 14205 | case X86::BI__builtin_ia32_shuf_f32x4_256: |
| 14206 | case X86::BI__builtin_ia32_shuf_f64x2_256: |
| 14207 | case X86::BI__builtin_ia32_shuf_i32x4_256: |
| 14208 | case X86::BI__builtin_ia32_shuf_i64x2_256: |
| 14209 | case X86::BI__builtin_ia32_shuf_f32x4: |
| 14210 | case X86::BI__builtin_ia32_shuf_f64x2: |
| 14211 | case X86::BI__builtin_ia32_shuf_i32x4: |
| 14212 | case X86::BI__builtin_ia32_shuf_i64x2: { |
| 14213 | unsigned Imm = cast<llvm::ConstantInt>(Ops[2])->getZExtValue(); |
| 14214 | auto *Ty = cast<llvm::FixedVectorType>(Ops[0]->getType()); |
| 14215 | unsigned NumElts = Ty->getNumElements(); |
| 14216 | unsigned NumLanes = Ty->getPrimitiveSizeInBits() == 512 ? 4 : 2; |
| 14217 | unsigned NumLaneElts = NumElts / NumLanes; |
| 14218 | |
| 14219 | int Indices[16]; |
| 14220 | for (unsigned l = 0; l != NumElts; l += NumLaneElts) { |
| 14221 | unsigned Index = (Imm % NumLanes) * NumLaneElts; |
| 14222 | Imm /= NumLanes; // Discard the bits we just used. |
| 14223 | if (l >= (NumElts / 2)) |
| 14224 | Index += NumElts; // Switch to other source. |
| 14225 | for (unsigned i = 0; i != NumLaneElts; ++i) { |
| 14226 | Indices[l + i] = Index + i; |
| 14227 | } |
| 14228 | } |
| 14229 | |
| 14230 | return Builder.CreateShuffleVector(Ops[0], Ops[1], |
| 14231 | ArrayRef(Indices, NumElts), "shuf"); |
| 14232 | } |
| 14233 | |
| 14234 | case X86::BI__builtin_ia32_vperm2f128_pd256: |
| 14235 | case X86::BI__builtin_ia32_vperm2f128_ps256: |
| 14236 | case X86::BI__builtin_ia32_vperm2f128_si256: |
| 14237 | case X86::BI__builtin_ia32_permti256: { |
| 14238 | unsigned Imm = cast<llvm::ConstantInt>(Ops[2])->getZExtValue(); |
| 14239 | unsigned NumElts = |
| 14240 | cast<llvm::FixedVectorType>(Ops[0]->getType())->getNumElements(); |
| 14241 | |
| 14242 | // This takes a very simple approach since there are two lanes and a |
| 14243 | // shuffle can have 2 inputs. So we reserve the first input for the first |
| 14244 | // lane and the second input for the second lane. This may result in |
| 14245 | // duplicate sources, but this can be dealt with in the backend. |
| 14246 | |
| 14247 | Value *OutOps[2]; |
| 14248 | int Indices[8]; |
| 14249 | for (unsigned l = 0; l != 2; ++l) { |
| 14250 | // Determine the source for this lane. |
| 14251 | if (Imm & (1 << ((l * 4) + 3))) |
| 14252 | OutOps[l] = llvm::ConstantAggregateZero::get(Ops[0]->getType()); |
| 14253 | else if (Imm & (1 << ((l * 4) + 1))) |
| 14254 | OutOps[l] = Ops[1]; |
| 14255 | else |
| 14256 | OutOps[l] = Ops[0]; |
| 14257 | |
| 14258 | for (unsigned i = 0; i != NumElts/2; ++i) { |
| 14259 | // Start with ith element of the source for this lane. |
| 14260 | unsigned Idx = (l * NumElts) + i; |
| 14261 | // If bit 0 of the immediate half is set, switch to the high half of |
| 14262 | // the source. |
| 14263 | if (Imm & (1 << (l * 4))) |
| 14264 | Idx += NumElts/2; |
| 14265 | Indices[(l * (NumElts/2)) + i] = Idx; |
| 14266 | } |
| 14267 | } |
| 14268 | |
| 14269 | return Builder.CreateShuffleVector(OutOps[0], OutOps[1], |
| 14270 | ArrayRef(Indices, NumElts), "vperm"); |
| 14271 | } |
| 14272 | |
| 14273 | case X86::BI__builtin_ia32_pslldqi128_byteshift: |
| 14274 | case X86::BI__builtin_ia32_pslldqi256_byteshift: |
| 14275 | case X86::BI__builtin_ia32_pslldqi512_byteshift: { |
| 14276 | unsigned ShiftVal = cast<llvm::ConstantInt>(Ops[1])->getZExtValue() & 0xff; |
| 14277 | auto *ResultType = cast<llvm::FixedVectorType>(Ops[0]->getType()); |
| 14278 | // Builtin type is vXi64 so multiply by 8 to get bytes. |
| 14279 | unsigned NumElts = ResultType->getNumElements() * 8; |
| 14280 | |
| 14281 | // If pslldq is shifting the vector more than 15 bytes, emit zero. |
| 14282 | if (ShiftVal >= 16) |
| 14283 | return llvm::Constant::getNullValue(ResultType); |
| 14284 | |
| 14285 | int Indices[64]; |
| 14286 | // 256/512-bit pslldq operates on 128-bit lanes so we need to handle that |
| 14287 | for (unsigned l = 0; l != NumElts; l += 16) { |
| 14288 | for (unsigned i = 0; i != 16; ++i) { |
| 14289 | unsigned Idx = NumElts + i - ShiftVal; |
| 14290 | if (Idx < NumElts) Idx -= NumElts - 16; // end of lane, switch operand. |
| 14291 | Indices[l + i] = Idx + l; |
| 14292 | } |
| 14293 | } |
| 14294 | |
| 14295 | auto *VecTy = llvm::FixedVectorType::get(Int8Ty, NumElts); |
| 14296 | Value *Cast = Builder.CreateBitCast(Ops[0], VecTy, "cast"); |
| 14297 | Value *Zero = llvm::Constant::getNullValue(VecTy); |
| 14298 | Value *SV = Builder.CreateShuffleVector( |
| 14299 | Zero, Cast, ArrayRef(Indices, NumElts), "pslldq"); |
| 14300 | return Builder.CreateBitCast(SV, Ops[0]->getType(), "cast"); |
| 14301 | } |
| 14302 | case X86::BI__builtin_ia32_psrldqi128_byteshift: |
| 14303 | case X86::BI__builtin_ia32_psrldqi256_byteshift: |
| 14304 | case X86::BI__builtin_ia32_psrldqi512_byteshift: { |
| 14305 | unsigned ShiftVal = cast<llvm::ConstantInt>(Ops[1])->getZExtValue() & 0xff; |
| 14306 | auto *ResultType = cast<llvm::FixedVectorType>(Ops[0]->getType()); |
| 14307 | // Builtin type is vXi64 so multiply by 8 to get bytes. |
| 14308 | unsigned NumElts = ResultType->getNumElements() * 8; |
| 14309 | |
| 14310 | // If psrldq is shifting the vector more than 15 bytes, emit zero. |
| 14311 | if (ShiftVal >= 16) |
| 14312 | return llvm::Constant::getNullValue(ResultType); |
| 14313 | |
| 14314 | int Indices[64]; |
| 14315 | // 256/512-bit psrldq operates on 128-bit lanes so we need to handle that |
| 14316 | for (unsigned l = 0; l != NumElts; l += 16) { |
| 14317 | for (unsigned i = 0; i != 16; ++i) { |
| 14318 | unsigned Idx = i + ShiftVal; |
| 14319 | if (Idx >= 16) Idx += NumElts - 16; // end of lane, switch operand. |
| 14320 | Indices[l + i] = Idx + l; |
| 14321 | } |
| 14322 | } |
| 14323 | |
| 14324 | auto *VecTy = llvm::FixedVectorType::get(Int8Ty, NumElts); |
| 14325 | Value *Cast = Builder.CreateBitCast(Ops[0], VecTy, "cast"); |
| 14326 | Value *Zero = llvm::Constant::getNullValue(VecTy); |
| 14327 | Value *SV = Builder.CreateShuffleVector( |
| 14328 | Cast, Zero, ArrayRef(Indices, NumElts), "psrldq"); |
| 14329 | return Builder.CreateBitCast(SV, ResultType, "cast"); |
| 14330 | } |
| 14331 | case X86::BI__builtin_ia32_kshiftliqi: |
| 14332 | case X86::BI__builtin_ia32_kshiftlihi: |
| 14333 | case X86::BI__builtin_ia32_kshiftlisi: |
| 14334 | case X86::BI__builtin_ia32_kshiftlidi: { |
| 14335 | unsigned ShiftVal = cast<llvm::ConstantInt>(Ops[1])->getZExtValue() & 0xff; |
| 14336 | unsigned NumElts = Ops[0]->getType()->getIntegerBitWidth(); |
| 14337 | |
| 14338 | if (ShiftVal >= NumElts) |
| 14339 | return llvm::Constant::getNullValue(Ops[0]->getType()); |
| 14340 | |
| 14341 | Value *In = getMaskVecValue(*this, Ops[0], NumElts); |
| 14342 | |
| 14343 | int Indices[64]; |
| 14344 | for (unsigned i = 0; i != NumElts; ++i) |
| 14345 | Indices[i] = NumElts + i - ShiftVal; |
| 14346 | |
| 14347 | Value *Zero = llvm::Constant::getNullValue(In->getType()); |
| 14348 | Value *SV = Builder.CreateShuffleVector( |
| 14349 | Zero, In, ArrayRef(Indices, NumElts), "kshiftl"); |
| 14350 | return Builder.CreateBitCast(SV, Ops[0]->getType()); |
| 14351 | } |
| 14352 | case X86::BI__builtin_ia32_kshiftriqi: |
| 14353 | case X86::BI__builtin_ia32_kshiftrihi: |
| 14354 | case X86::BI__builtin_ia32_kshiftrisi: |
| 14355 | case X86::BI__builtin_ia32_kshiftridi: { |
| 14356 | unsigned ShiftVal = cast<llvm::ConstantInt>(Ops[1])->getZExtValue() & 0xff; |
| 14357 | unsigned NumElts = Ops[0]->getType()->getIntegerBitWidth(); |
| 14358 | |
| 14359 | if (ShiftVal >= NumElts) |
| 14360 | return llvm::Constant::getNullValue(Ops[0]->getType()); |
| 14361 | |
| 14362 | Value *In = getMaskVecValue(*this, Ops[0], NumElts); |
| 14363 | |
| 14364 | int Indices[64]; |
| 14365 | for (unsigned i = 0; i != NumElts; ++i) |
| 14366 | Indices[i] = i + ShiftVal; |
| 14367 | |
| 14368 | Value *Zero = llvm::Constant::getNullValue(In->getType()); |
| 14369 | Value *SV = Builder.CreateShuffleVector( |
| 14370 | In, Zero, ArrayRef(Indices, NumElts), "kshiftr"); |
| 14371 | return Builder.CreateBitCast(SV, Ops[0]->getType()); |
| 14372 | } |
| 14373 | case X86::BI__builtin_ia32_movnti: |
| 14374 | case X86::BI__builtin_ia32_movnti64: |
| 14375 | case X86::BI__builtin_ia32_movntsd: |
| 14376 | case X86::BI__builtin_ia32_movntss: { |
| 14377 | llvm::MDNode *Node = llvm::MDNode::get( |
| 14378 | getLLVMContext(), llvm::ConstantAsMetadata::get(Builder.getInt32(1))); |
| 14379 | |
| 14380 | Value *Ptr = Ops[0]; |
| 14381 | Value *Src = Ops[1]; |
| 14382 | |
| 14383 | // Extract the 0'th element of the source vector. |
| 14384 | if (BuiltinID == X86::BI__builtin_ia32_movntsd || |
| 14385 | BuiltinID == X86::BI__builtin_ia32_movntss) |
| 14386 | Src = Builder.CreateExtractElement(Src, (uint64_t)0, "extract"); |
| 14387 | |
| 14388 | // Convert the type of the pointer to a pointer to the stored type. |
| 14389 | Value *BC = Builder.CreateBitCast( |
| 14390 | Ptr, llvm::PointerType::getUnqual(Src->getType()), "cast"); |
| 14391 | |
| 14392 | // Unaligned nontemporal store of the scalar value. |
| 14393 | StoreInst *SI = Builder.CreateDefaultAlignedStore(Src, BC); |
| 14394 | SI->setMetadata(CGM.getModule().getMDKindID("nontemporal"), Node); |
| 14395 | SI->setAlignment(llvm::Align(1)); |
| 14396 | return SI; |
| 14397 | } |
| 14398 | // Rotate is a special case of funnel shift - 1st 2 args are the same. |
| 14399 | case X86::BI__builtin_ia32_vprotb: |
| 14400 | case X86::BI__builtin_ia32_vprotw: |
| 14401 | case X86::BI__builtin_ia32_vprotd: |
| 14402 | case X86::BI__builtin_ia32_vprotq: |
| 14403 | case X86::BI__builtin_ia32_vprotbi: |
| 14404 | case X86::BI__builtin_ia32_vprotwi: |
| 14405 | case X86::BI__builtin_ia32_vprotdi: |
| 14406 | case X86::BI__builtin_ia32_vprotqi: |
| 14407 | case X86::BI__builtin_ia32_prold128: |
| 14408 | case X86::BI__builtin_ia32_prold256: |
| 14409 | case X86::BI__builtin_ia32_prold512: |
| 14410 | case X86::BI__builtin_ia32_prolq128: |
| 14411 | case X86::BI__builtin_ia32_prolq256: |
| 14412 | case X86::BI__builtin_ia32_prolq512: |
| 14413 | case X86::BI__builtin_ia32_prolvd128: |
| 14414 | case X86::BI__builtin_ia32_prolvd256: |
| 14415 | case X86::BI__builtin_ia32_prolvd512: |
| 14416 | case X86::BI__builtin_ia32_prolvq128: |
| 14417 | case X86::BI__builtin_ia32_prolvq256: |
| 14418 | case X86::BI__builtin_ia32_prolvq512: |
| 14419 | return EmitX86FunnelShift(*this, Ops[0], Ops[0], Ops[1], false); |
| 14420 | case X86::BI__builtin_ia32_prord128: |
| 14421 | case X86::BI__builtin_ia32_prord256: |
| 14422 | case X86::BI__builtin_ia32_prord512: |
| 14423 | case X86::BI__builtin_ia32_prorq128: |
| 14424 | case X86::BI__builtin_ia32_prorq256: |
| 14425 | case X86::BI__builtin_ia32_prorq512: |
| 14426 | case X86::BI__builtin_ia32_prorvd128: |
| 14427 | case X86::BI__builtin_ia32_prorvd256: |
| 14428 | case X86::BI__builtin_ia32_prorvd512: |
| 14429 | case X86::BI__builtin_ia32_prorvq128: |
| 14430 | case X86::BI__builtin_ia32_prorvq256: |
| 14431 | case X86::BI__builtin_ia32_prorvq512: |
| 14432 | return EmitX86FunnelShift(*this, Ops[0], Ops[0], Ops[1], true); |
| 14433 | case X86::BI__builtin_ia32_selectb_128: |
| 14434 | case X86::BI__builtin_ia32_selectb_256: |
| 14435 | case X86::BI__builtin_ia32_selectb_512: |
| 14436 | case X86::BI__builtin_ia32_selectw_128: |
| 14437 | case X86::BI__builtin_ia32_selectw_256: |
| 14438 | case X86::BI__builtin_ia32_selectw_512: |
| 14439 | case X86::BI__builtin_ia32_selectd_128: |
| 14440 | case X86::BI__builtin_ia32_selectd_256: |
| 14441 | case X86::BI__builtin_ia32_selectd_512: |
| 14442 | case X86::BI__builtin_ia32_selectq_128: |
| 14443 | case X86::BI__builtin_ia32_selectq_256: |
| 14444 | case X86::BI__builtin_ia32_selectq_512: |
| 14445 | case X86::BI__builtin_ia32_selectph_128: |
| 14446 | case X86::BI__builtin_ia32_selectph_256: |
| 14447 | case X86::BI__builtin_ia32_selectph_512: |
| 14448 | case X86::BI__builtin_ia32_selectpbf_128: |
| 14449 | case X86::BI__builtin_ia32_selectpbf_256: |
| 14450 | case X86::BI__builtin_ia32_selectpbf_512: |
| 14451 | case X86::BI__builtin_ia32_selectps_128: |
| 14452 | case X86::BI__builtin_ia32_selectps_256: |
| 14453 | case X86::BI__builtin_ia32_selectps_512: |
| 14454 | case X86::BI__builtin_ia32_selectpd_128: |
| 14455 | case X86::BI__builtin_ia32_selectpd_256: |
| 14456 | case X86::BI__builtin_ia32_selectpd_512: |
| 14457 | return EmitX86Select(*this, Ops[0], Ops[1], Ops[2]); |
| 14458 | case X86::BI__builtin_ia32_selectsh_128: |
| 14459 | case X86::BI__builtin_ia32_selectsbf_128: |
| 14460 | case X86::BI__builtin_ia32_selectss_128: |
| 14461 | case X86::BI__builtin_ia32_selectsd_128: { |
| 14462 | Value *A = Builder.CreateExtractElement(Ops[1], (uint64_t)0); |
| 14463 | Value *B = Builder.CreateExtractElement(Ops[2], (uint64_t)0); |
| 14464 | A = EmitX86ScalarSelect(*this, Ops[0], A, B); |
| 14465 | return Builder.CreateInsertElement(Ops[1], A, (uint64_t)0); |
| 14466 | } |
| 14467 | case X86::BI__builtin_ia32_cmpb128_mask: |
| 14468 | case X86::BI__builtin_ia32_cmpb256_mask: |
| 14469 | case X86::BI__builtin_ia32_cmpb512_mask: |
| 14470 | case X86::BI__builtin_ia32_cmpw128_mask: |
| 14471 | case X86::BI__builtin_ia32_cmpw256_mask: |
| 14472 | case X86::BI__builtin_ia32_cmpw512_mask: |
| 14473 | case X86::BI__builtin_ia32_cmpd128_mask: |
| 14474 | case X86::BI__builtin_ia32_cmpd256_mask: |
| 14475 | case X86::BI__builtin_ia32_cmpd512_mask: |
| 14476 | case X86::BI__builtin_ia32_cmpq128_mask: |
| 14477 | case X86::BI__builtin_ia32_cmpq256_mask: |
| 14478 | case X86::BI__builtin_ia32_cmpq512_mask: { |
| 14479 | unsigned CC = cast<llvm::ConstantInt>(Ops[2])->getZExtValue() & 0x7; |
| 14480 | return EmitX86MaskedCompare(*this, CC, true, Ops); |
| 14481 | } |
| 14482 | case X86::BI__builtin_ia32_ucmpb128_mask: |
| 14483 | case X86::BI__builtin_ia32_ucmpb256_mask: |
| 14484 | case X86::BI__builtin_ia32_ucmpb512_mask: |
| 14485 | case X86::BI__builtin_ia32_ucmpw128_mask: |
| 14486 | case X86::BI__builtin_ia32_ucmpw256_mask: |
| 14487 | case X86::BI__builtin_ia32_ucmpw512_mask: |
| 14488 | case X86::BI__builtin_ia32_ucmpd128_mask: |
| 14489 | case X86::BI__builtin_ia32_ucmpd256_mask: |
| 14490 | case X86::BI__builtin_ia32_ucmpd512_mask: |
| 14491 | case X86::BI__builtin_ia32_ucmpq128_mask: |
| 14492 | case X86::BI__builtin_ia32_ucmpq256_mask: |
| 14493 | case X86::BI__builtin_ia32_ucmpq512_mask: { |
| 14494 | unsigned CC = cast<llvm::ConstantInt>(Ops[2])->getZExtValue() & 0x7; |
| 14495 | return EmitX86MaskedCompare(*this, CC, false, Ops); |
| 14496 | } |
| 14497 | case X86::BI__builtin_ia32_vpcomb: |
| 14498 | case X86::BI__builtin_ia32_vpcomw: |
| 14499 | case X86::BI__builtin_ia32_vpcomd: |
| 14500 | case X86::BI__builtin_ia32_vpcomq: |
| 14501 | return EmitX86vpcom(*this, Ops, true); |
| 14502 | case X86::BI__builtin_ia32_vpcomub: |
| 14503 | case X86::BI__builtin_ia32_vpcomuw: |
| 14504 | case X86::BI__builtin_ia32_vpcomud: |
| 14505 | case X86::BI__builtin_ia32_vpcomuq: |
| 14506 | return EmitX86vpcom(*this, Ops, false); |
| 14507 | |
| 14508 | case X86::BI__builtin_ia32_kortestcqi: |
| 14509 | case X86::BI__builtin_ia32_kortestchi: |
| 14510 | case X86::BI__builtin_ia32_kortestcsi: |
| 14511 | case X86::BI__builtin_ia32_kortestcdi: { |
| 14512 | Value *Or = EmitX86MaskLogic(*this, Instruction::Or, Ops); |
| 14513 | Value *C = llvm::Constant::getAllOnesValue(Ops[0]->getType()); |
| 14514 | Value *Cmp = Builder.CreateICmpEQ(Or, C); |
| 14515 | return Builder.CreateZExt(Cmp, ConvertType(E->getType())); |
| 14516 | } |
| 14517 | case X86::BI__builtin_ia32_kortestzqi: |
| 14518 | case X86::BI__builtin_ia32_kortestzhi: |
| 14519 | case X86::BI__builtin_ia32_kortestzsi: |
| 14520 | case X86::BI__builtin_ia32_kortestzdi: { |
| 14521 | Value *Or = EmitX86MaskLogic(*this, Instruction::Or, Ops); |
| 14522 | Value *C = llvm::Constant::getNullValue(Ops[0]->getType()); |
| 14523 | Value *Cmp = Builder.CreateICmpEQ(Or, C); |
| 14524 | return Builder.CreateZExt(Cmp, ConvertType(E->getType())); |
| 14525 | } |
| 14526 | |
| 14527 | case X86::BI__builtin_ia32_ktestcqi: |
| 14528 | case X86::BI__builtin_ia32_ktestzqi: |
| 14529 | case X86::BI__builtin_ia32_ktestchi: |
| 14530 | case X86::BI__builtin_ia32_ktestzhi: |
| 14531 | case X86::BI__builtin_ia32_ktestcsi: |
| 14532 | case X86::BI__builtin_ia32_ktestzsi: |
| 14533 | case X86::BI__builtin_ia32_ktestcdi: |
| 14534 | case X86::BI__builtin_ia32_ktestzdi: { |
| 14535 | Intrinsic::ID IID; |
| 14536 | switch (BuiltinID) { |
| 14537 | default: llvm_unreachable("Unsupported intrinsic!")::llvm::llvm_unreachable_internal("Unsupported intrinsic!", "clang/lib/CodeGen/CGBuiltin.cpp" , 14537); |
| 14538 | case X86::BI__builtin_ia32_ktestcqi: |
| 14539 | IID = Intrinsic::x86_avx512_ktestc_b; |
| 14540 | break; |
| 14541 | case X86::BI__builtin_ia32_ktestzqi: |
| 14542 | IID = Intrinsic::x86_avx512_ktestz_b; |
| 14543 | break; |
| 14544 | case X86::BI__builtin_ia32_ktestchi: |
| 14545 | IID = Intrinsic::x86_avx512_ktestc_w; |
| 14546 | break; |
| 14547 | case X86::BI__builtin_ia32_ktestzhi: |
| 14548 | IID = Intrinsic::x86_avx512_ktestz_w; |
| 14549 | break; |
| 14550 | case X86::BI__builtin_ia32_ktestcsi: |
| 14551 | IID = Intrinsic::x86_avx512_ktestc_d; |
| 14552 | break; |
| 14553 | case X86::BI__builtin_ia32_ktestzsi: |
| 14554 | IID = Intrinsic::x86_avx512_ktestz_d; |
| 14555 | break; |
| 14556 | case X86::BI__builtin_ia32_ktestcdi: |
| 14557 | IID = Intrinsic::x86_avx512_ktestc_q; |
| 14558 | break; |
| 14559 | case X86::BI__builtin_ia32_ktestzdi: |
| 14560 | IID = Intrinsic::x86_avx512_ktestz_q; |
| 14561 | break; |
| 14562 | } |
| 14563 | |
| 14564 | unsigned NumElts = Ops[0]->getType()->getIntegerBitWidth(); |
| 14565 | Value *LHS = getMaskVecValue(*this, Ops[0], NumElts); |
| 14566 | Value *RHS = getMaskVecValue(*this, Ops[1], NumElts); |
| 14567 | Function *Intr = CGM.getIntrinsic(IID); |
| 14568 | return Builder.CreateCall(Intr, {LHS, RHS}); |
| 14569 | } |
| 14570 | |
| 14571 | case X86::BI__builtin_ia32_kaddqi: |
| 14572 | case X86::BI__builtin_ia32_kaddhi: |
| 14573 | case X86::BI__builtin_ia32_kaddsi: |
| 14574 | case X86::BI__builtin_ia32_kadddi: { |
| 14575 | Intrinsic::ID IID; |
| 14576 | switch (BuiltinID) { |
| 14577 | default: llvm_unreachable("Unsupported intrinsic!")::llvm::llvm_unreachable_internal("Unsupported intrinsic!", "clang/lib/CodeGen/CGBuiltin.cpp" , 14577); |
| 14578 | case X86::BI__builtin_ia32_kaddqi: |
| 14579 | IID = Intrinsic::x86_avx512_kadd_b; |
| 14580 | break; |
| 14581 | case X86::BI__builtin_ia32_kaddhi: |
| 14582 | IID = Intrinsic::x86_avx512_kadd_w; |
| 14583 | break; |
| 14584 | case X86::BI__builtin_ia32_kaddsi: |
| 14585 | IID = Intrinsic::x86_avx512_kadd_d; |
| 14586 | break; |
| 14587 | case X86::BI__builtin_ia32_kadddi: |
| 14588 | IID = Intrinsic::x86_avx512_kadd_q; |
| 14589 | break; |
| 14590 | } |
| 14591 | |
| 14592 | unsigned NumElts = Ops[0]->getType()->getIntegerBitWidth(); |
| 14593 | Value *LHS = getMaskVecValue(*this, Ops[0], NumElts); |
| 14594 | Value *RHS = getMaskVecValue(*this, Ops[1], NumElts); |
| 14595 | Function *Intr = CGM.getIntrinsic(IID); |
| 14596 | Value *Res = Builder.CreateCall(Intr, {LHS, RHS}); |
| 14597 | return Builder.CreateBitCast(Res, Ops[0]->getType()); |
| 14598 | } |
| 14599 | case X86::BI__builtin_ia32_kandqi: |
| 14600 | case X86::BI__builtin_ia32_kandhi: |
| 14601 | case X86::BI__builtin_ia32_kandsi: |
| 14602 | case X86::BI__builtin_ia32_kanddi: |
| 14603 | return EmitX86MaskLogic(*this, Instruction::And, Ops); |
| 14604 | case X86::BI__builtin_ia32_kandnqi: |
| 14605 | case X86::BI__builtin_ia32_kandnhi: |
| 14606 | case X86::BI__builtin_ia32_kandnsi: |
| 14607 | case X86::BI__builtin_ia32_kandndi: |
| 14608 | return EmitX86MaskLogic(*this, Instruction::And, Ops, true); |
| 14609 | case X86::BI__builtin_ia32_korqi: |
| 14610 | case X86::BI__builtin_ia32_korhi: |
| 14611 | case X86::BI__builtin_ia32_korsi: |
| 14612 | case X86::BI__builtin_ia32_kordi: |
| 14613 | return EmitX86MaskLogic(*this, Instruction::Or, Ops); |
| 14614 | case X86::BI__builtin_ia32_kxnorqi: |
| 14615 | case X86::BI__builtin_ia32_kxnorhi: |
| 14616 | case X86::BI__builtin_ia32_kxnorsi: |
| 14617 | case X86::BI__builtin_ia32_kxnordi: |
| 14618 | return EmitX86MaskLogic(*this, Instruction::Xor, Ops, true); |
| 14619 | case X86::BI__builtin_ia32_kxorqi: |
| 14620 | case X86::BI__builtin_ia32_kxorhi: |
| 14621 | case X86::BI__builtin_ia32_kxorsi: |
| 14622 | case X86::BI__builtin_ia32_kxordi: |
| 14623 | return EmitX86MaskLogic(*this, Instruction::Xor, Ops); |
| 14624 | case X86::BI__builtin_ia32_knotqi: |
| 14625 | case X86::BI__builtin_ia32_knothi: |
| 14626 | case X86::BI__builtin_ia32_knotsi: |
| 14627 | case X86::BI__builtin_ia32_knotdi: { |
| 14628 | unsigned NumElts = Ops[0]->getType()->getIntegerBitWidth(); |
| 14629 | Value *Res = getMaskVecValue(*this, Ops[0], NumElts); |
| 14630 | return Builder.CreateBitCast(Builder.CreateNot(Res), |
| 14631 | Ops[0]->getType()); |
| 14632 | } |
| 14633 | case X86::BI__builtin_ia32_kmovb: |
| 14634 | case X86::BI__builtin_ia32_kmovw: |
| 14635 | case X86::BI__builtin_ia32_kmovd: |
| 14636 | case X86::BI__builtin_ia32_kmovq: { |
| 14637 | // Bitcast to vXi1 type and then back to integer. This gets the mask |
| 14638 | // register type into the IR, but might be optimized out depending on |
| 14639 | // what's around it. |
| 14640 | unsigned NumElts = Ops[0]->getType()->getIntegerBitWidth(); |
| 14641 | Value *Res = getMaskVecValue(*this, Ops[0], NumElts); |
| 14642 | return Builder.CreateBitCast(Res, Ops[0]->getType()); |
| 14643 | } |
| 14644 | |
| 14645 | case X86::BI__builtin_ia32_kunpckdi: |
| 14646 | case X86::BI__builtin_ia32_kunpcksi: |
| 14647 | case X86::BI__builtin_ia32_kunpckhi: { |
| 14648 | unsigned NumElts = Ops[0]->getType()->getIntegerBitWidth(); |
| 14649 | Value *LHS = getMaskVecValue(*this, Ops[0], NumElts); |
| 14650 | Value *RHS = getMaskVecValue(*this, Ops[1], NumElts); |
| 14651 | int Indices[64]; |
| 14652 | for (unsigned i = 0; i != NumElts; ++i) |
| 14653 | Indices[i] = i; |
| 14654 | |
| 14655 | // First extract half of each vector. This gives better codegen than |
| 14656 | // doing it in a single shuffle. |
| 14657 | LHS = Builder.CreateShuffleVector(LHS, LHS, ArrayRef(Indices, NumElts / 2)); |
| 14658 | RHS = Builder.CreateShuffleVector(RHS, RHS, ArrayRef(Indices, NumElts / 2)); |
| 14659 | // Concat the vectors. |
| 14660 | // NOTE: Operands are swapped to match the intrinsic definition. |
| 14661 | Value *Res = |
| 14662 | Builder.CreateShuffleVector(RHS, LHS, ArrayRef(Indices, NumElts)); |
| 14663 | return Builder.CreateBitCast(Res, Ops[0]->getType()); |
| 14664 | } |
| 14665 | |
| 14666 | case X86::BI__builtin_ia32_vplzcntd_128: |
| 14667 | case X86::BI__builtin_ia32_vplzcntd_256: |
| 14668 | case X86::BI__builtin_ia32_vplzcntd_512: |
| 14669 | case X86::BI__builtin_ia32_vplzcntq_128: |
| 14670 | case X86::BI__builtin_ia32_vplzcntq_256: |
| 14671 | case X86::BI__builtin_ia32_vplzcntq_512: { |
| 14672 | Function *F = CGM.getIntrinsic(Intrinsic::ctlz, Ops[0]->getType()); |
| 14673 | return Builder.CreateCall(F, {Ops[0],Builder.getInt1(false)}); |
| 14674 | } |
| 14675 | case X86::BI__builtin_ia32_sqrtss: |
| 14676 | case X86::BI__builtin_ia32_sqrtsd: { |
| 14677 | Value *A = Builder.CreateExtractElement(Ops[0], (uint64_t)0); |
| 14678 | Function *F; |
| 14679 | if (Builder.getIsFPConstrained()) { |
| 14680 | CodeGenFunction::CGFPOptionsRAII FPOptsRAII(*this, E); |
| 14681 | F = CGM.getIntrinsic(Intrinsic::experimental_constrained_sqrt, |
| 14682 | A->getType()); |
| 14683 | A = Builder.CreateConstrainedFPCall(F, {A}); |
| 14684 | } else { |
| 14685 | F = CGM.getIntrinsic(Intrinsic::sqrt, A->getType()); |
| 14686 | A = Builder.CreateCall(F, {A}); |
| 14687 | } |
| 14688 | return Builder.CreateInsertElement(Ops[0], A, (uint64_t)0); |
| 14689 | } |
| 14690 | case X86::BI__builtin_ia32_sqrtsh_round_mask: |
| 14691 | case X86::BI__builtin_ia32_sqrtsd_round_mask: |
| 14692 | case X86::BI__builtin_ia32_sqrtss_round_mask: { |
| 14693 | unsigned CC = cast<llvm::ConstantInt>(Ops[4])->getZExtValue(); |
| 14694 | // Support only if the rounding mode is 4 (AKA CUR_DIRECTION), |
| 14695 | // otherwise keep the intrinsic. |
| 14696 | if (CC != 4) { |
| 14697 | Intrinsic::ID IID; |
| 14698 | |
| 14699 | switch (BuiltinID) { |
| 14700 | default: |
| 14701 | llvm_unreachable("Unsupported intrinsic!")::llvm::llvm_unreachable_internal("Unsupported intrinsic!", "clang/lib/CodeGen/CGBuiltin.cpp" , 14701); |
| 14702 | case X86::BI__builtin_ia32_sqrtsh_round_mask: |
| 14703 | IID = Intrinsic::x86_avx512fp16_mask_sqrt_sh; |
| 14704 | break; |
| 14705 | case X86::BI__builtin_ia32_sqrtsd_round_mask: |
| 14706 | IID = Intrinsic::x86_avx512_mask_sqrt_sd; |
| 14707 | break; |
| 14708 | case X86::BI__builtin_ia32_sqrtss_round_mask: |
| 14709 | IID = Intrinsic::x86_avx512_mask_sqrt_ss; |
| 14710 | break; |
| 14711 | } |
| 14712 | return Builder.CreateCall(CGM.getIntrinsic(IID), Ops); |
| 14713 | } |
| 14714 | Value *A = Builder.CreateExtractElement(Ops[1], (uint64_t)0); |
| 14715 | Function *F; |
| 14716 | if (Builder.getIsFPConstrained()) { |
| 14717 | CodeGenFunction::CGFPOptionsRAII FPOptsRAII(*this, E); |
| 14718 | F = CGM.getIntrinsic(Intrinsic::experimental_constrained_sqrt, |
| 14719 | A->getType()); |
| 14720 | A = Builder.CreateConstrainedFPCall(F, A); |
| 14721 | } else { |
| 14722 | F = CGM.getIntrinsic(Intrinsic::sqrt, A->getType()); |
| 14723 | A = Builder.CreateCall(F, A); |
| 14724 | } |
| 14725 | Value *Src = Builder.CreateExtractElement(Ops[2], (uint64_t)0); |
| 14726 | A = EmitX86ScalarSelect(*this, Ops[3], A, Src); |
| 14727 | return Builder.CreateInsertElement(Ops[0], A, (uint64_t)0); |
| 14728 | } |
| 14729 | case X86::BI__builtin_ia32_sqrtpd256: |
| 14730 | case X86::BI__builtin_ia32_sqrtpd: |
| 14731 | case X86::BI__builtin_ia32_sqrtps256: |
| 14732 | case X86::BI__builtin_ia32_sqrtps: |
| 14733 | case X86::BI__builtin_ia32_sqrtph256: |
| 14734 | case X86::BI__builtin_ia32_sqrtph: |
| 14735 | case X86::BI__builtin_ia32_sqrtph512: |
| 14736 | case X86::BI__builtin_ia32_sqrtps512: |
| 14737 | case X86::BI__builtin_ia32_sqrtpd512: { |
| 14738 | if (Ops.size() == 2) { |
| 14739 | unsigned CC = cast<llvm::ConstantInt>(Ops[1])->getZExtValue(); |
| 14740 | // Support only if the rounding mode is 4 (AKA CUR_DIRECTION), |
| 14741 | // otherwise keep the intrinsic. |
| 14742 | if (CC != 4) { |
| 14743 | Intrinsic::ID IID; |
| 14744 | |
| 14745 | switch (BuiltinID) { |
| 14746 | default: |
| 14747 | llvm_unreachable("Unsupported intrinsic!")::llvm::llvm_unreachable_internal("Unsupported intrinsic!", "clang/lib/CodeGen/CGBuiltin.cpp" , 14747); |
| 14748 | case X86::BI__builtin_ia32_sqrtph512: |
| 14749 | IID = Intrinsic::x86_avx512fp16_sqrt_ph_512; |
| 14750 | break; |
| 14751 | case X86::BI__builtin_ia32_sqrtps512: |
| 14752 | IID = Intrinsic::x86_avx512_sqrt_ps_512; |
| 14753 | break; |
| 14754 | case X86::BI__builtin_ia32_sqrtpd512: |
| 14755 | IID = Intrinsic::x86_avx512_sqrt_pd_512; |
| 14756 | break; |
| 14757 | } |
| 14758 | return Builder.CreateCall(CGM.getIntrinsic(IID), Ops); |
| 14759 | } |
| 14760 | } |
| 14761 | if (Builder.getIsFPConstrained()) { |
| 14762 | CodeGenFunction::CGFPOptionsRAII FPOptsRAII(*this, E); |
| 14763 | Function *F = CGM.getIntrinsic(Intrinsic::experimental_constrained_sqrt, |
| 14764 | Ops[0]->getType()); |
| 14765 | return Builder.CreateConstrainedFPCall(F, Ops[0]); |
| 14766 | } else { |
| 14767 | Function *F = CGM.getIntrinsic(Intrinsic::sqrt, Ops[0]->getType()); |
| 14768 | return Builder.CreateCall(F, Ops[0]); |
| 14769 | } |
| 14770 | } |
| 14771 | |
| 14772 | case X86::BI__builtin_ia32_pmuludq128: |
| 14773 | case X86::BI__builtin_ia32_pmuludq256: |
| 14774 | case X86::BI__builtin_ia32_pmuludq512: |
| 14775 | return EmitX86Muldq(*this, /*IsSigned*/false, Ops); |
| 14776 | |
| 14777 | case X86::BI__builtin_ia32_pmuldq128: |
| 14778 | case X86::BI__builtin_ia32_pmuldq256: |
| 14779 | case X86::BI__builtin_ia32_pmuldq512: |
| 14780 | return EmitX86Muldq(*this, /*IsSigned*/true, Ops); |
| 14781 | |
| 14782 | case X86::BI__builtin_ia32_pternlogd512_mask: |
| 14783 | case X86::BI__builtin_ia32_pternlogq512_mask: |
| 14784 | case X86::BI__builtin_ia32_pternlogd128_mask: |
| 14785 | case X86::BI__builtin_ia32_pternlogd256_mask: |
| 14786 | case X86::BI__builtin_ia32_pternlogq128_mask: |
| 14787 | case X86::BI__builtin_ia32_pternlogq256_mask: |
| 14788 | return EmitX86Ternlog(*this, /*ZeroMask*/false, Ops); |
| 14789 | |
| 14790 | case X86::BI__builtin_ia32_pternlogd512_maskz: |
| 14791 | case X86::BI__builtin_ia32_pternlogq512_maskz: |
| 14792 | case X86::BI__builtin_ia32_pternlogd128_maskz: |
| 14793 | case X86::BI__builtin_ia32_pternlogd256_maskz: |
| 14794 | case X86::BI__builtin_ia32_pternlogq128_maskz: |
| 14795 | case X86::BI__builtin_ia32_pternlogq256_maskz: |
| 14796 | return EmitX86Ternlog(*this, /*ZeroMask*/true, Ops); |
| 14797 | |
| 14798 | case X86::BI__builtin_ia32_vpshldd128: |
| 14799 | case X86::BI__builtin_ia32_vpshldd256: |
| 14800 | case X86::BI__builtin_ia32_vpshldd512: |
| 14801 | case X86::BI__builtin_ia32_vpshldq128: |
| 14802 | case X86::BI__builtin_ia32_vpshldq256: |
| 14803 | case X86::BI__builtin_ia32_vpshldq512: |
| 14804 | case X86::BI__builtin_ia32_vpshldw128: |
| 14805 | case X86::BI__builtin_ia32_vpshldw256: |
| 14806 | case X86::BI__builtin_ia32_vpshldw512: |
| 14807 | return EmitX86FunnelShift(*this, Ops[0], Ops[1], Ops[2], false); |
| 14808 | |
| 14809 | case X86::BI__builtin_ia32_vpshrdd128: |
| 14810 | case X86::BI__builtin_ia32_vpshrdd256: |
| 14811 | case X86::BI__builtin_ia32_vpshrdd512: |
| 14812 | case X86::BI__builtin_ia32_vpshrdq128: |
| 14813 | case X86::BI__builtin_ia32_vpshrdq256: |
| 14814 | case X86::BI__builtin_ia32_vpshrdq512: |
| 14815 | case X86::BI__builtin_ia32_vpshrdw128: |
| 14816 | case X86::BI__builtin_ia32_vpshrdw256: |
| 14817 | case X86::BI__builtin_ia32_vpshrdw512: |
| 14818 | // Ops 0 and 1 are swapped. |
| 14819 | return EmitX86FunnelShift(*this, Ops[1], Ops[0], Ops[2], true); |
| 14820 | |
| 14821 | case X86::BI__builtin_ia32_vpshldvd128: |
| 14822 | case X86::BI__builtin_ia32_vpshldvd256: |
| 14823 | case X86::BI__builtin_ia32_vpshldvd512: |
| 14824 | case X86::BI__builtin_ia32_vpshldvq128: |
| 14825 | case X86::BI__builtin_ia32_vpshldvq256: |
| 14826 | case X86::BI__builtin_ia32_vpshldvq512: |
| 14827 | case X86::BI__builtin_ia32_vpshldvw128: |
| 14828 | case X86::BI__builtin_ia32_vpshldvw256: |
| 14829 | case X86::BI__builtin_ia32_vpshldvw512: |
| 14830 | return EmitX86FunnelShift(*this, Ops[0], Ops[1], Ops[2], false); |
| 14831 | |
| 14832 | case X86::BI__builtin_ia32_vpshrdvd128: |
| 14833 | case X86::BI__builtin_ia32_vpshrdvd256: |
| 14834 | case X86::BI__builtin_ia32_vpshrdvd512: |
| 14835 | case X86::BI__builtin_ia32_vpshrdvq128: |
| 14836 | case X86::BI__builtin_ia32_vpshrdvq256: |
| 14837 | case X86::BI__builtin_ia32_vpshrdvq512: |
| 14838 | case X86::BI__builtin_ia32_vpshrdvw128: |
| 14839 | case X86::BI__builtin_ia32_vpshrdvw256: |
| 14840 | case X86::BI__builtin_ia32_vpshrdvw512: |
| 14841 | // Ops 0 and 1 are swapped. |
| 14842 | return EmitX86FunnelShift(*this, Ops[1], Ops[0], Ops[2], true); |
| 14843 | |
| 14844 | // Reductions |
| 14845 | case X86::BI__builtin_ia32_reduce_fadd_pd512: |
| 14846 | case X86::BI__builtin_ia32_reduce_fadd_ps512: |
| 14847 | case X86::BI__builtin_ia32_reduce_fadd_ph512: |
| 14848 | case X86::BI__builtin_ia32_reduce_fadd_ph256: |
| 14849 | case X86::BI__builtin_ia32_reduce_fadd_ph128: { |
| 14850 | Function *F = |
| 14851 | CGM.getIntrinsic(Intrinsic::vector_reduce_fadd, Ops[1]->getType()); |
| 14852 | IRBuilder<>::FastMathFlagGuard FMFGuard(Builder); |
| 14853 | Builder.getFastMathFlags().setAllowReassoc(); |
| 14854 | return Builder.CreateCall(F, {Ops[0], Ops[1]}); |
| 14855 | } |
| 14856 | case X86::BI__builtin_ia32_reduce_fmul_pd512: |
| 14857 | case X86::BI__builtin_ia32_reduce_fmul_ps512: |
| 14858 | case X86::BI__builtin_ia32_reduce_fmul_ph512: |
| 14859 | case X86::BI__builtin_ia32_reduce_fmul_ph256: |
| 14860 | case X86::BI__builtin_ia32_reduce_fmul_ph128: { |
| 14861 | Function *F = |
| 14862 | CGM.getIntrinsic(Intrinsic::vector_reduce_fmul, Ops[1]->getType()); |
| 14863 | IRBuilder<>::FastMathFlagGuard FMFGuard(Builder); |
| 14864 | Builder.getFastMathFlags().setAllowReassoc(); |
| 14865 | return Builder.CreateCall(F, {Ops[0], Ops[1]}); |
| 14866 | } |
| 14867 | case X86::BI__builtin_ia32_reduce_fmax_pd512: |
| 14868 | case X86::BI__builtin_ia32_reduce_fmax_ps512: |
| 14869 | case X86::BI__builtin_ia32_reduce_fmax_ph512: |
| 14870 | case X86::BI__builtin_ia32_reduce_fmax_ph256: |
| 14871 | case X86::BI__builtin_ia32_reduce_fmax_ph128: { |
| 14872 | Function *F = |
| 14873 | CGM.getIntrinsic(Intrinsic::vector_reduce_fmax, Ops[0]->getType()); |
| 14874 | IRBuilder<>::FastMathFlagGuard FMFGuard(Builder); |
| 14875 | Builder.getFastMathFlags().setNoNaNs(); |
| 14876 | return Builder.CreateCall(F, {Ops[0]}); |
| 14877 | } |
| 14878 | case X86::BI__builtin_ia32_reduce_fmin_pd512: |
| 14879 | case X86::BI__builtin_ia32_reduce_fmin_ps512: |
| 14880 | case X86::BI__builtin_ia32_reduce_fmin_ph512: |
| 14881 | case X86::BI__builtin_ia32_reduce_fmin_ph256: |
| 14882 | case X86::BI__builtin_ia32_reduce_fmin_ph128: { |
| 14883 | Function *F = |
| 14884 | CGM.getIntrinsic(Intrinsic::vector_reduce_fmin, Ops[0]->getType()); |
| 14885 | IRBuilder<>::FastMathFlagGuard FMFGuard(Builder); |
| 14886 | Builder.getFastMathFlags().setNoNaNs(); |
| 14887 | return Builder.CreateCall(F, {Ops[0]}); |
| 14888 | } |
| 14889 | |
| 14890 | // 3DNow! |
| 14891 | case X86::BI__builtin_ia32_pswapdsf: |
| 14892 | case X86::BI__builtin_ia32_pswapdsi: { |
| 14893 | llvm::Type *MMXTy = llvm::Type::getX86_MMXTy(getLLVMContext()); |
| 14894 | Ops[0] = Builder.CreateBitCast(Ops[0], MMXTy, "cast"); |
| 14895 | llvm::Function *F = CGM.getIntrinsic(Intrinsic::x86_3dnowa_pswapd); |
| 14896 | return Builder.CreateCall(F, Ops, "pswapd"); |
| 14897 | } |
| 14898 | case X86::BI__builtin_ia32_rdrand16_step: |
| 14899 | case X86::BI__builtin_ia32_rdrand32_step: |
| 14900 | case X86::BI__builtin_ia32_rdrand64_step: |
| 14901 | case X86::BI__builtin_ia32_rdseed16_step: |
| 14902 | case X86::BI__builtin_ia32_rdseed32_step: |
| 14903 | case X86::BI__builtin_ia32_rdseed64_step: { |
| 14904 | Intrinsic::ID ID; |
| 14905 | switch (BuiltinID) { |
| 14906 | default: llvm_unreachable("Unsupported intrinsic!")::llvm::llvm_unreachable_internal("Unsupported intrinsic!", "clang/lib/CodeGen/CGBuiltin.cpp" , 14906); |
| 14907 | case X86::BI__builtin_ia32_rdrand16_step: |
| 14908 | ID = Intrinsic::x86_rdrand_16; |
| 14909 | break; |
| 14910 | case X86::BI__builtin_ia32_rdrand32_step: |
| 14911 | ID = Intrinsic::x86_rdrand_32; |
| 14912 | break; |
| 14913 | case X86::BI__builtin_ia32_rdrand64_step: |
| 14914 | ID = Intrinsic::x86_rdrand_64; |
| 14915 | break; |
| 14916 | case X86::BI__builtin_ia32_rdseed16_step: |
| 14917 | ID = Intrinsic::x86_rdseed_16; |
| 14918 | break; |
| 14919 | case X86::BI__builtin_ia32_rdseed32_step: |
| 14920 | ID = Intrinsic::x86_rdseed_32; |
| 14921 | break; |
| 14922 | case X86::BI__builtin_ia32_rdseed64_step: |
| 14923 | ID = Intrinsic::x86_rdseed_64; |
| 14924 | break; |
| 14925 | } |
| 14926 | |
| 14927 | Value *Call = Builder.CreateCall(CGM.getIntrinsic(ID)); |
| 14928 | Builder.CreateDefaultAlignedStore(Builder.CreateExtractValue(Call, 0), |
| 14929 | Ops[0]); |
| 14930 | return Builder.CreateExtractValue(Call, 1); |
| 14931 | } |
| 14932 | case X86::BI__builtin_ia32_addcarryx_u32: |
| 14933 | case X86::BI__builtin_ia32_addcarryx_u64: |
| 14934 | case X86::BI__builtin_ia32_subborrow_u32: |
| 14935 | case X86::BI__builtin_ia32_subborrow_u64: { |
| 14936 | Intrinsic::ID IID; |
| 14937 | switch (BuiltinID) { |
| 14938 | default: llvm_unreachable("Unsupported intrinsic!")::llvm::llvm_unreachable_internal("Unsupported intrinsic!", "clang/lib/CodeGen/CGBuiltin.cpp" , 14938); |
| 14939 | case X86::BI__builtin_ia32_addcarryx_u32: |
| 14940 | IID = Intrinsic::x86_addcarry_32; |
| 14941 | break; |
| 14942 | case X86::BI__builtin_ia32_addcarryx_u64: |
| 14943 | IID = Intrinsic::x86_addcarry_64; |
| 14944 | break; |
| 14945 | case X86::BI__builtin_ia32_subborrow_u32: |
| 14946 | IID = Intrinsic::x86_subborrow_32; |
| 14947 | break; |
| 14948 | case X86::BI__builtin_ia32_subborrow_u64: |
| 14949 | IID = Intrinsic::x86_subborrow_64; |
| 14950 | break; |
| 14951 | } |
| 14952 | |
| 14953 | Value *Call = Builder.CreateCall(CGM.getIntrinsic(IID), |
| 14954 | { Ops[0], Ops[1], Ops[2] }); |
| 14955 | Builder.CreateDefaultAlignedStore(Builder.CreateExtractValue(Call, 1), |
| 14956 | Ops[3]); |
| 14957 | return Builder.CreateExtractValue(Call, 0); |
| 14958 | } |
| 14959 | |
| 14960 | case X86::BI__builtin_ia32_fpclassps128_mask: |
| 14961 | case X86::BI__builtin_ia32_fpclassps256_mask: |
| 14962 | case X86::BI__builtin_ia32_fpclassps512_mask: |
| 14963 | case X86::BI__builtin_ia32_fpclassph128_mask: |
| 14964 | case X86::BI__builtin_ia32_fpclassph256_mask: |
| 14965 | case X86::BI__builtin_ia32_fpclassph512_mask: |
| 14966 | case X86::BI__builtin_ia32_fpclasspd128_mask: |
| 14967 | case X86::BI__builtin_ia32_fpclasspd256_mask: |
| 14968 | case X86::BI__builtin_ia32_fpclasspd512_mask: { |
| 14969 | unsigned NumElts = |
| 14970 | cast<llvm::FixedVectorType>(Ops[0]->getType())->getNumElements(); |
| 14971 | Value *MaskIn = Ops[2]; |
| 14972 | Ops.erase(&Ops[2]); |
| 14973 | |
| 14974 | Intrinsic::ID ID; |
| 14975 | switch (BuiltinID) { |
| 14976 | default: llvm_unreachable("Unsupported intrinsic!")::llvm::llvm_unreachable_internal("Unsupported intrinsic!", "clang/lib/CodeGen/CGBuiltin.cpp" , 14976); |
| 14977 | case X86::BI__builtin_ia32_fpclassph128_mask: |
| 14978 | ID = Intrinsic::x86_avx512fp16_fpclass_ph_128; |
| 14979 | break; |
| 14980 | case X86::BI__builtin_ia32_fpclassph256_mask: |
| 14981 | ID = Intrinsic::x86_avx512fp16_fpclass_ph_256; |
| 14982 | break; |
| 14983 | case X86::BI__builtin_ia32_fpclassph512_mask: |
| 14984 | ID = Intrinsic::x86_avx512fp16_fpclass_ph_512; |
| 14985 | break; |
| 14986 | case X86::BI__builtin_ia32_fpclassps128_mask: |
| 14987 | ID = Intrinsic::x86_avx512_fpclass_ps_128; |
| 14988 | break; |
| 14989 | case X86::BI__builtin_ia32_fpclassps256_mask: |
| 14990 | ID = Intrinsic::x86_avx512_fpclass_ps_256; |
| 14991 | break; |
| 14992 | case X86::BI__builtin_ia32_fpclassps512_mask: |
| 14993 | ID = Intrinsic::x86_avx512_fpclass_ps_512; |
| 14994 | break; |
| 14995 | case X86::BI__builtin_ia32_fpclasspd128_mask: |
| 14996 | ID = Intrinsic::x86_avx512_fpclass_pd_128; |
| 14997 | break; |
| 14998 | case X86::BI__builtin_ia32_fpclasspd256_mask: |
| 14999 | ID = Intrinsic::x86_avx512_fpclass_pd_256; |
| 15000 | break; |
| 15001 | case X86::BI__builtin_ia32_fpclasspd512_mask: |
| 15002 | ID = Intrinsic::x86_avx512_fpclass_pd_512; |
| 15003 | break; |
| 15004 | } |
| 15005 | |
| 15006 | Value *Fpclass = Builder.CreateCall(CGM.getIntrinsic(ID), Ops); |
| 15007 | return EmitX86MaskedCompareResult(*this, Fpclass, NumElts, MaskIn); |
| 15008 | } |
| 15009 | |
| 15010 | case X86::BI__builtin_ia32_vp2intersect_q_512: |
| 15011 | case X86::BI__builtin_ia32_vp2intersect_q_256: |
| 15012 | case X86::BI__builtin_ia32_vp2intersect_q_128: |
| 15013 | case X86::BI__builtin_ia32_vp2intersect_d_512: |
| 15014 | case X86::BI__builtin_ia32_vp2intersect_d_256: |
| 15015 | case X86::BI__builtin_ia32_vp2intersect_d_128: { |
| 15016 | unsigned NumElts = |
| 15017 | cast<llvm::FixedVectorType>(Ops[0]->getType())->getNumElements(); |
| 15018 | Intrinsic::ID ID; |
| 15019 | |
| 15020 | switch (BuiltinID) { |
| 15021 | default: llvm_unreachable("Unsupported intrinsic!")::llvm::llvm_unreachable_internal("Unsupported intrinsic!", "clang/lib/CodeGen/CGBuiltin.cpp" , 15021); |
| 15022 | case X86::BI__builtin_ia32_vp2intersect_q_512: |
| 15023 | ID = Intrinsic::x86_avx512_vp2intersect_q_512; |
| 15024 | break; |
| 15025 | case X86::BI__builtin_ia32_vp2intersect_q_256: |
| 15026 | ID = Intrinsic::x86_avx512_vp2intersect_q_256; |
| 15027 | break; |
| 15028 | case X86::BI__builtin_ia32_vp2intersect_q_128: |
| 15029 | ID = Intrinsic::x86_avx512_vp2intersect_q_128; |
| 15030 | break; |
| 15031 | case X86::BI__builtin_ia32_vp2intersect_d_512: |
| 15032 | ID = Intrinsic::x86_avx512_vp2intersect_d_512; |
| 15033 | break; |
| 15034 | case X86::BI__builtin_ia32_vp2intersect_d_256: |
| 15035 | ID = Intrinsic::x86_avx512_vp2intersect_d_256; |
| 15036 | break; |
| 15037 | case X86::BI__builtin_ia32_vp2intersect_d_128: |
| 15038 | ID = Intrinsic::x86_avx512_vp2intersect_d_128; |
| 15039 | break; |
| 15040 | } |
| 15041 | |
| 15042 | Value *Call = Builder.CreateCall(CGM.getIntrinsic(ID), {Ops[0], Ops[1]}); |
| 15043 | Value *Result = Builder.CreateExtractValue(Call, 0); |
| 15044 | Result = EmitX86MaskedCompareResult(*this, Result, NumElts, nullptr); |
| 15045 | Builder.CreateDefaultAlignedStore(Result, Ops[2]); |
| 15046 | |
| 15047 | Result = Builder.CreateExtractValue(Call, 1); |
| 15048 | Result = EmitX86MaskedCompareResult(*this, Result, NumElts, nullptr); |
| 15049 | return Builder.CreateDefaultAlignedStore(Result, Ops[3]); |
| 15050 | } |
| 15051 | |
| 15052 | case X86::BI__builtin_ia32_vpmultishiftqb128: |
| 15053 | case X86::BI__builtin_ia32_vpmultishiftqb256: |
| 15054 | case X86::BI__builtin_ia32_vpmultishiftqb512: { |
| 15055 | Intrinsic::ID ID; |
| 15056 | switch (BuiltinID) { |
| 15057 | default: llvm_unreachable("Unsupported intrinsic!")::llvm::llvm_unreachable_internal("Unsupported intrinsic!", "clang/lib/CodeGen/CGBuiltin.cpp" , 15057); |
| 15058 | case X86::BI__builtin_ia32_vpmultishiftqb128: |
| 15059 | ID = Intrinsic::x86_avx512_pmultishift_qb_128; |
| 15060 | break; |
| 15061 | case X86::BI__builtin_ia32_vpmultishiftqb256: |
| 15062 | ID = Intrinsic::x86_avx512_pmultishift_qb_256; |
| 15063 | break; |
| 15064 | case X86::BI__builtin_ia32_vpmultishiftqb512: |
| 15065 | ID = Intrinsic::x86_avx512_pmultishift_qb_512; |
| 15066 | break; |
| 15067 | } |
| 15068 | |
| 15069 | return Builder.CreateCall(CGM.getIntrinsic(ID), Ops); |
| 15070 | } |
| 15071 | |
| 15072 | case X86::BI__builtin_ia32_vpshufbitqmb128_mask: |
| 15073 | case X86::BI__builtin_ia32_vpshufbitqmb256_mask: |
| 15074 | case X86::BI__builtin_ia32_vpshufbitqmb512_mask: { |
| 15075 | unsigned NumElts = |
| 15076 | cast<llvm::FixedVectorType>(Ops[0]->getType())->getNumElements(); |
| 15077 | Value *MaskIn = Ops[2]; |
| 15078 | Ops.erase(&Ops[2]); |
| 15079 | |
| 15080 | Intrinsic::ID ID; |
| 15081 | switch (BuiltinID) { |
| 15082 | default: llvm_unreachable("Unsupported intrinsic!")::llvm::llvm_unreachable_internal("Unsupported intrinsic!", "clang/lib/CodeGen/CGBuiltin.cpp" , 15082); |
| 15083 | case X86::BI__builtin_ia32_vpshufbitqmb128_mask: |
| 15084 | ID = Intrinsic::x86_avx512_vpshufbitqmb_128; |
| 15085 | break; |
| 15086 | case X86::BI__builtin_ia32_vpshufbitqmb256_mask: |
| 15087 | ID = Intrinsic::x86_avx512_vpshufbitqmb_256; |
| 15088 | break; |
| 15089 | case X86::BI__builtin_ia32_vpshufbitqmb512_mask: |
| 15090 | ID = Intrinsic::x86_avx512_vpshufbitqmb_512; |
| 15091 | break; |
| 15092 | } |
| 15093 | |
| 15094 | Value *Shufbit = Builder.CreateCall(CGM.getIntrinsic(ID), Ops); |
| 15095 | return EmitX86MaskedCompareResult(*this, Shufbit, NumElts, MaskIn); |
| 15096 | } |
| 15097 | |
| 15098 | // packed comparison intrinsics |
| 15099 | case X86::BI__builtin_ia32_cmpeqps: |
| 15100 | case X86::BI__builtin_ia32_cmpeqpd: |
| 15101 | return getVectorFCmpIR(CmpInst::FCMP_OEQ, /*IsSignaling*/false); |
| 15102 | case X86::BI__builtin_ia32_cmpltps: |
| 15103 | case X86::BI__builtin_ia32_cmpltpd: |
| 15104 | return getVectorFCmpIR(CmpInst::FCMP_OLT, /*IsSignaling*/true); |
| 15105 | case X86::BI__builtin_ia32_cmpleps: |
| 15106 | case X86::BI__builtin_ia32_cmplepd: |
| 15107 | return getVectorFCmpIR(CmpInst::FCMP_OLE, /*IsSignaling*/true); |
| 15108 | case X86::BI__builtin_ia32_cmpunordps: |
| 15109 | case X86::BI__builtin_ia32_cmpunordpd: |
| 15110 | return getVectorFCmpIR(CmpInst::FCMP_UNO, /*IsSignaling*/false); |
| 15111 | case X86::BI__builtin_ia32_cmpneqps: |
| 15112 | case X86::BI__builtin_ia32_cmpneqpd: |
| 15113 | return getVectorFCmpIR(CmpInst::FCMP_UNE, /*IsSignaling*/false); |
| 15114 | case X86::BI__builtin_ia32_cmpnltps: |
| 15115 | case X86::BI__builtin_ia32_cmpnltpd: |
| 15116 | return getVectorFCmpIR(CmpInst::FCMP_UGE, /*IsSignaling*/true); |
| 15117 | case X86::BI__builtin_ia32_cmpnleps: |
| 15118 | case X86::BI__builtin_ia32_cmpnlepd: |
| 15119 | return getVectorFCmpIR(CmpInst::FCMP_UGT, /*IsSignaling*/true); |
| 15120 | case X86::BI__builtin_ia32_cmpordps: |
| 15121 | case X86::BI__builtin_ia32_cmpordpd: |
| 15122 | return getVectorFCmpIR(CmpInst::FCMP_ORD, /*IsSignaling*/false); |
| 15123 | case X86::BI__builtin_ia32_cmpph128_mask: |
| 15124 | case X86::BI__builtin_ia32_cmpph256_mask: |
| 15125 | case X86::BI__builtin_ia32_cmpph512_mask: |
| 15126 | case X86::BI__builtin_ia32_cmpps128_mask: |
| 15127 | case X86::BI__builtin_ia32_cmpps256_mask: |
| 15128 | case X86::BI__builtin_ia32_cmpps512_mask: |
| 15129 | case X86::BI__builtin_ia32_cmppd128_mask: |
| 15130 | case X86::BI__builtin_ia32_cmppd256_mask: |
| 15131 | case X86::BI__builtin_ia32_cmppd512_mask: |
| 15132 | IsMaskFCmp = true; |
| 15133 | [[fallthrough]]; |
| 15134 | case X86::BI__builtin_ia32_cmpps: |
| 15135 | case X86::BI__builtin_ia32_cmpps256: |
| 15136 | case X86::BI__builtin_ia32_cmppd: |
| 15137 | case X86::BI__builtin_ia32_cmppd256: { |
| 15138 | // Lowering vector comparisons to fcmp instructions, while |
| 15139 | // ignoring signalling behaviour requested |
| 15140 | // ignoring rounding mode requested |
| 15141 | // This is only possible if fp-model is not strict and FENV_ACCESS is off. |
| 15142 | |
| 15143 | // The third argument is the comparison condition, and integer in the |
| 15144 | // range [0, 31] |
| 15145 | unsigned CC = cast<llvm::ConstantInt>(Ops[2])->getZExtValue() & 0x1f; |
| 15146 | |
| 15147 | // Lowering to IR fcmp instruction. |
| 15148 | // Ignoring requested signaling behaviour, |
| 15149 | // e.g. both _CMP_GT_OS & _CMP_GT_OQ are translated to FCMP_OGT. |
| 15150 | FCmpInst::Predicate Pred; |
| 15151 | bool IsSignaling; |
| 15152 | // Predicates for 16-31 repeat the 0-15 predicates. Only the signalling |
| 15153 | // behavior is inverted. We'll handle that after the switch. |
| 15154 | switch (CC & 0xf) { |
| 15155 | case 0x00: Pred = FCmpInst::FCMP_OEQ; IsSignaling = false; break; |
| 15156 | case 0x01: Pred = FCmpInst::FCMP_OLT; IsSignaling = true; break; |
| 15157 | case 0x02: Pred = FCmpInst::FCMP_OLE; IsSignaling = true; break; |
| 15158 | case 0x03: Pred = FCmpInst::FCMP_UNO; IsSignaling = false; break; |
| 15159 | case 0x04: Pred = FCmpInst::FCMP_UNE; IsSignaling = false; break; |
| 15160 | case 0x05: Pred = FCmpInst::FCMP_UGE; IsSignaling = true; break; |
| 15161 | case 0x06: Pred = FCmpInst::FCMP_UGT; IsSignaling = true; break; |
| 15162 | case 0x07: Pred = FCmpInst::FCMP_ORD; IsSignaling = false; break; |
| 15163 | case 0x08: Pred = FCmpInst::FCMP_UEQ; IsSignaling = false; break; |
| 15164 | case 0x09: Pred = FCmpInst::FCMP_ULT; IsSignaling = true; break; |
| 15165 | case 0x0a: Pred = FCmpInst::FCMP_ULE; IsSignaling = true; break; |
| 15166 | case 0x0b: Pred = FCmpInst::FCMP_FALSE; IsSignaling = false; break; |
| 15167 | case 0x0c: Pred = FCmpInst::FCMP_ONE; IsSignaling = false; break; |
| 15168 | case 0x0d: Pred = FCmpInst::FCMP_OGE; IsSignaling = true; break; |
| 15169 | case 0x0e: Pred = FCmpInst::FCMP_OGT; IsSignaling = true; break; |
| 15170 | case 0x0f: Pred = FCmpInst::FCMP_TRUE; IsSignaling = false; break; |
| 15171 | default: llvm_unreachable("Unhandled CC")::llvm::llvm_unreachable_internal("Unhandled CC", "clang/lib/CodeGen/CGBuiltin.cpp" , 15171); |
| 15172 | } |
| 15173 | |
| 15174 | // Invert the signalling behavior for 16-31. |
| 15175 | if (CC & 0x10) |
| 15176 | IsSignaling = !IsSignaling; |
| 15177 | |
| 15178 | // If the predicate is true or false and we're using constrained intrinsics, |
| 15179 | // we don't have a compare intrinsic we can use. Just use the legacy X86 |
| 15180 | // specific intrinsic. |
| 15181 | // If the intrinsic is mask enabled and we're using constrained intrinsics, |
| 15182 | // use the legacy X86 specific intrinsic. |
| 15183 | if (Builder.getIsFPConstrained() && |
| 15184 | (Pred == FCmpInst::FCMP_TRUE || Pred == FCmpInst::FCMP_FALSE || |
| 15185 | IsMaskFCmp)) { |
| 15186 | |
| 15187 | Intrinsic::ID IID; |
| 15188 | switch (BuiltinID) { |
| 15189 | default: llvm_unreachable("Unexpected builtin")::llvm::llvm_unreachable_internal("Unexpected builtin", "clang/lib/CodeGen/CGBuiltin.cpp" , 15189); |
| 15190 | case X86::BI__builtin_ia32_cmpps: |
| 15191 | IID = Intrinsic::x86_sse_cmp_ps; |
| 15192 | break; |
| 15193 | case X86::BI__builtin_ia32_cmpps256: |
| 15194 | IID = Intrinsic::x86_avx_cmp_ps_256; |
| 15195 | break; |
| 15196 | case X86::BI__builtin_ia32_cmppd: |
| 15197 | IID = Intrinsic::x86_sse2_cmp_pd; |
| 15198 | break; |
| 15199 | case X86::BI__builtin_ia32_cmppd256: |
| 15200 | IID = Intrinsic::x86_avx_cmp_pd_256; |
| 15201 | break; |
| 15202 | case X86::BI__builtin_ia32_cmpps512_mask: |
| 15203 | IID = Intrinsic::x86_avx512_mask_cmp_ps_512; |
| 15204 | break; |
| 15205 | case X86::BI__builtin_ia32_cmppd512_mask: |
| 15206 | IID = Intrinsic::x86_avx512_mask_cmp_pd_512; |
| 15207 | break; |
| 15208 | case X86::BI__builtin_ia32_cmpps128_mask: |
| 15209 | IID = Intrinsic::x86_avx512_mask_cmp_ps_128; |
| 15210 | break; |
| 15211 | case X86::BI__builtin_ia32_cmpps256_mask: |
| 15212 | IID = Intrinsic::x86_avx512_mask_cmp_ps_256; |
| 15213 | break; |
| 15214 | case X86::BI__builtin_ia32_cmppd128_mask: |
| 15215 | IID = Intrinsic::x86_avx512_mask_cmp_pd_128; |
| 15216 | break; |
| 15217 | case X86::BI__builtin_ia32_cmppd256_mask: |
| 15218 | IID = Intrinsic::x86_avx512_mask_cmp_pd_256; |
| 15219 | break; |
| 15220 | } |
| 15221 | |
| 15222 | Function *Intr = CGM.getIntrinsic(IID); |
| 15223 | if (IsMaskFCmp) { |
| 15224 | unsigned NumElts = |
| 15225 | cast<llvm::FixedVectorType>(Ops[0]->getType())->getNumElements(); |
| 15226 | Ops[3] = getMaskVecValue(*this, Ops[3], NumElts); |
| 15227 | Value *Cmp = Builder.CreateCall(Intr, Ops); |
| 15228 | return EmitX86MaskedCompareResult(*this, Cmp, NumElts, nullptr); |
| 15229 | } |
| 15230 | |
| 15231 | return Builder.CreateCall(Intr, Ops); |
| 15232 | } |
| 15233 | |
| 15234 | // Builtins without the _mask suffix return a vector of integers |
| 15235 | // of the same width as the input vectors |
| 15236 | if (IsMaskFCmp) { |
| 15237 | // We ignore SAE if strict FP is disabled. We only keep precise |
| 15238 | // exception behavior under strict FP. |
| 15239 | // NOTE: If strict FP does ever go through here a CGFPOptionsRAII |
| 15240 | // object will be required. |
| 15241 | unsigned NumElts = |
| 15242 | cast<llvm::FixedVectorType>(Ops[0]->getType())->getNumElements(); |
| 15243 | Value *Cmp; |
| 15244 | if (IsSignaling) |
| 15245 | Cmp = Builder.CreateFCmpS(Pred, Ops[0], Ops[1]); |
| 15246 | else |
| 15247 | Cmp = Builder.CreateFCmp(Pred, Ops[0], Ops[1]); |
| 15248 | return EmitX86MaskedCompareResult(*this, Cmp, NumElts, Ops[3]); |
| 15249 | } |
| 15250 | |
| 15251 | return getVectorFCmpIR(Pred, IsSignaling); |
| 15252 | } |
| 15253 | |
| 15254 | // SSE scalar comparison intrinsics |
| 15255 | case X86::BI__builtin_ia32_cmpeqss: |
| 15256 | return getCmpIntrinsicCall(Intrinsic::x86_sse_cmp_ss, 0); |
| 15257 | case X86::BI__builtin_ia32_cmpltss: |
| 15258 | return getCmpIntrinsicCall(Intrinsic::x86_sse_cmp_ss, 1); |
| 15259 | case X86::BI__builtin_ia32_cmpless: |
| 15260 | return getCmpIntrinsicCall(Intrinsic::x86_sse_cmp_ss, 2); |
| 15261 | case X86::BI__builtin_ia32_cmpunordss: |
| 15262 | return getCmpIntrinsicCall(Intrinsic::x86_sse_cmp_ss, 3); |
| 15263 | case X86::BI__builtin_ia32_cmpneqss: |
| 15264 | return getCmpIntrinsicCall(Intrinsic::x86_sse_cmp_ss, 4); |
| 15265 | case X86::BI__builtin_ia32_cmpnltss: |
| 15266 | return getCmpIntrinsicCall(Intrinsic::x86_sse_cmp_ss, 5); |
| 15267 | case X86::BI__builtin_ia32_cmpnless: |
| 15268 | return getCmpIntrinsicCall(Intrinsic::x86_sse_cmp_ss, 6); |
| 15269 | case X86::BI__builtin_ia32_cmpordss: |
| 15270 | return getCmpIntrinsicCall(Intrinsic::x86_sse_cmp_ss, 7); |
| 15271 | case X86::BI__builtin_ia32_cmpeqsd: |
| 15272 | return getCmpIntrinsicCall(Intrinsic::x86_sse2_cmp_sd, 0); |
| 15273 | case X86::BI__builtin_ia32_cmpltsd: |
| 15274 | return getCmpIntrinsicCall(Intrinsic::x86_sse2_cmp_sd, 1); |
| 15275 | case X86::BI__builtin_ia32_cmplesd: |
| 15276 | return getCmpIntrinsicCall(Intrinsic::x86_sse2_cmp_sd, 2); |
| 15277 | case X86::BI__builtin_ia32_cmpunordsd: |
| 15278 | return getCmpIntrinsicCall(Intrinsic::x86_sse2_cmp_sd, 3); |
| 15279 | case X86::BI__builtin_ia32_cmpneqsd: |
| 15280 | return getCmpIntrinsicCall(Intrinsic::x86_sse2_cmp_sd, 4); |
| 15281 | case X86::BI__builtin_ia32_cmpnltsd: |
| 15282 | return getCmpIntrinsicCall(Intrinsic::x86_sse2_cmp_sd, 5); |
| 15283 | case X86::BI__builtin_ia32_cmpnlesd: |
| 15284 | return getCmpIntrinsicCall(Intrinsic::x86_sse2_cmp_sd, 6); |
| 15285 | case X86::BI__builtin_ia32_cmpordsd: |
| 15286 | return getCmpIntrinsicCall(Intrinsic::x86_sse2_cmp_sd, 7); |
| 15287 | |
| 15288 | // f16c half2float intrinsics |
| 15289 | case X86::BI__builtin_ia32_vcvtph2ps: |
| 15290 | case X86::BI__builtin_ia32_vcvtph2ps256: |
| 15291 | case X86::BI__builtin_ia32_vcvtph2ps_mask: |
| 15292 | case X86::BI__builtin_ia32_vcvtph2ps256_mask: |
| 15293 | case X86::BI__builtin_ia32_vcvtph2ps512_mask: { |
| 15294 | CodeGenFunction::CGFPOptionsRAII FPOptsRAII(*this, E); |
| 15295 | return EmitX86CvtF16ToFloatExpr(*this, Ops, ConvertType(E->getType())); |
| 15296 | } |
| 15297 | |
| 15298 | // AVX512 bf16 intrinsics |
| 15299 | case X86::BI__builtin_ia32_cvtneps2bf16_128_mask: { |
| 15300 | Ops[2] = getMaskVecValue( |
| 15301 | *this, Ops[2], |
| 15302 | cast<llvm::FixedVectorType>(Ops[0]->getType())->getNumElements()); |
| 15303 | Intrinsic::ID IID = Intrinsic::x86_avx512bf16_mask_cvtneps2bf16_128; |
| 15304 | return Builder.CreateCall(CGM.getIntrinsic(IID), Ops); |
| 15305 | } |
| 15306 | case X86::BI__builtin_ia32_cvtsbf162ss_32: |
| 15307 | return Builder.CreateFPExt(Ops[0], Builder.getFloatTy()); |
| 15308 | |
| 15309 | case X86::BI__builtin_ia32_cvtneps2bf16_256_mask: |
| 15310 | case X86::BI__builtin_ia32_cvtneps2bf16_512_mask: { |
| 15311 | Intrinsic::ID IID; |
| 15312 | switch (BuiltinID) { |
| 15313 | default: llvm_unreachable("Unsupported intrinsic!")::llvm::llvm_unreachable_internal("Unsupported intrinsic!", "clang/lib/CodeGen/CGBuiltin.cpp" , 15313); |
| 15314 | case X86::BI__builtin_ia32_cvtneps2bf16_256_mask: |
| 15315 | IID = Intrinsic::x86_avx512bf16_cvtneps2bf16_256; |
| 15316 | break; |
| 15317 | case X86::BI__builtin_ia32_cvtneps2bf16_512_mask: |
| 15318 | IID = Intrinsic::x86_avx512bf16_cvtneps2bf16_512; |
| 15319 | break; |
| 15320 | } |
| 15321 | Value *Res = Builder.CreateCall(CGM.getIntrinsic(IID), Ops[0]); |
| 15322 | return EmitX86Select(*this, Ops[2], Res, Ops[1]); |
| 15323 | } |
| 15324 | |
| 15325 | case X86::BI__cpuid: |
| 15326 | case X86::BI__cpuidex: { |
| 15327 | Value *FuncId = EmitScalarExpr(E->getArg(1)); |
| 15328 | Value *SubFuncId = BuiltinID == X86::BI__cpuidex |
| 15329 | ? EmitScalarExpr(E->getArg(2)) |
| 15330 | : llvm::ConstantInt::get(Int32Ty, 0); |
| 15331 | |
| 15332 | llvm::StructType *CpuidRetTy = |
| 15333 | llvm::StructType::get(Int32Ty, Int32Ty, Int32Ty, Int32Ty); |
| 15334 | llvm::FunctionType *FTy = |
| 15335 | llvm::FunctionType::get(CpuidRetTy, {Int32Ty, Int32Ty}, false); |
| 15336 | |
| 15337 | StringRef Asm, Constraints; |
| 15338 | if (getTarget().getTriple().getArch() == llvm::Triple::x86) { |
| 15339 | Asm = "cpuid"; |
| 15340 | Constraints = "={ax},={bx},={cx},={dx},{ax},{cx}"; |
| 15341 | } else { |
| 15342 | // x86-64 uses %rbx as the base register, so preserve it. |
| 15343 | Asm = "xchgq %rbx, ${1:q}\n" |
| 15344 | "cpuid\n" |
| 15345 | "xchgq %rbx, ${1:q}"; |
| 15346 | Constraints = "={ax},=r,={cx},={dx},0,2"; |
| 15347 | } |
| 15348 | |
| 15349 | llvm::InlineAsm *IA = llvm::InlineAsm::get(FTy, Asm, Constraints, |
| 15350 | /*hasSideEffects=*/false); |
| 15351 | Value *IACall = Builder.CreateCall(IA, {FuncId, SubFuncId}); |
| 15352 | Value *BasePtr = EmitScalarExpr(E->getArg(0)); |
| 15353 | Value *Store = nullptr; |
| 15354 | for (unsigned i = 0; i < 4; i++) { |
| 15355 | Value *Extracted = Builder.CreateExtractValue(IACall, i); |
| 15356 | Value *StorePtr = Builder.CreateConstInBoundsGEP1_32(Int32Ty, BasePtr, i); |
| 15357 | Store = Builder.CreateAlignedStore(Extracted, StorePtr, getIntAlign()); |
| 15358 | } |
| 15359 | |
| 15360 | // Return the last store instruction to signal that we have emitted the |
| 15361 | // the intrinsic. |
| 15362 | return Store; |
| 15363 | } |
| 15364 | |
| 15365 | case X86::BI__emul: |
| 15366 | case X86::BI__emulu: { |
| 15367 | llvm::Type *Int64Ty = llvm::IntegerType::get(getLLVMContext(), 64); |
| 15368 | bool isSigned = (BuiltinID == X86::BI__emul); |
| 15369 | Value *LHS = Builder.CreateIntCast(Ops[0], Int64Ty, isSigned); |
| 15370 | Value *RHS = Builder.CreateIntCast(Ops[1], Int64Ty, isSigned); |
| 15371 | return Builder.CreateMul(LHS, RHS, "", !isSigned, isSigned); |
| 15372 | } |
| 15373 | case X86::BI__mulh: |
| 15374 | case X86::BI__umulh: |
| 15375 | case X86::BI_mul128: |
| 15376 | case X86::BI_umul128: { |
| 15377 | llvm::Type *ResType = ConvertType(E->getType()); |
| 15378 | llvm::Type *Int128Ty = llvm::IntegerType::get(getLLVMContext(), 128); |
| 15379 | |
| 15380 | bool IsSigned = (BuiltinID == X86::BI__mulh || BuiltinID == X86::BI_mul128); |
| 15381 | Value *LHS = Builder.CreateIntCast(Ops[0], Int128Ty, IsSigned); |
| 15382 | Value *RHS = Builder.CreateIntCast(Ops[1], Int128Ty, IsSigned); |
| 15383 | |
| 15384 | Value *MulResult, *HigherBits; |
| 15385 | if (IsSigned) { |
| 15386 | MulResult = Builder.CreateNSWMul(LHS, RHS); |
| 15387 | HigherBits = Builder.CreateAShr(MulResult, 64); |
| 15388 | } else { |
| 15389 | MulResult = Builder.CreateNUWMul(LHS, RHS); |
| 15390 | HigherBits = Builder.CreateLShr(MulResult, 64); |
| 15391 | } |
| 15392 | HigherBits = Builder.CreateIntCast(HigherBits, ResType, IsSigned); |
| 15393 | |
| 15394 | if (BuiltinID == X86::BI__mulh || BuiltinID == X86::BI__umulh) |
| 15395 | return HigherBits; |
| 15396 | |
| 15397 | Address HighBitsAddress = EmitPointerWithAlignment(E->getArg(2)); |
| 15398 | Builder.CreateStore(HigherBits, HighBitsAddress); |
| 15399 | return Builder.CreateIntCast(MulResult, ResType, IsSigned); |
| 15400 | } |
| 15401 | |
| 15402 | case X86::BI__faststorefence: { |
| 15403 | return Builder.CreateFence(llvm::AtomicOrdering::SequentiallyConsistent, |
| 15404 | llvm::SyncScope::System); |
| 15405 | } |
| 15406 | case X86::BI__shiftleft128: |
| 15407 | case X86::BI__shiftright128: { |
| 15408 | llvm::Function *F = CGM.getIntrinsic( |
| 15409 | BuiltinID == X86::BI__shiftleft128 ? Intrinsic::fshl : Intrinsic::fshr, |
| 15410 | Int64Ty); |
| 15411 | // Flip low/high ops and zero-extend amount to matching type. |
| 15412 | // shiftleft128(Low, High, Amt) -> fshl(High, Low, Amt) |
| 15413 | // shiftright128(Low, High, Amt) -> fshr(High, Low, Amt) |
| 15414 | std::swap(Ops[0], Ops[1]); |
| 15415 | Ops[2] = Builder.CreateZExt(Ops[2], Int64Ty); |
| 15416 | return Builder.CreateCall(F, Ops); |
| 15417 | } |
| 15418 | case X86::BI_ReadWriteBarrier: |
| 15419 | case X86::BI_ReadBarrier: |
| 15420 | case X86::BI_WriteBarrier: { |
| 15421 | return Builder.CreateFence(llvm::AtomicOrdering::SequentiallyConsistent, |
| 15422 | llvm::SyncScope::SingleThread); |
| 15423 | } |
| 15424 | |
| 15425 | case X86::BI_AddressOfReturnAddress: { |
| 15426 | Function *F = |
| 15427 | CGM.getIntrinsic(Intrinsic::addressofreturnaddress, AllocaInt8PtrTy); |
| 15428 | return Builder.CreateCall(F); |
| 15429 | } |
| 15430 | case X86::BI__stosb: { |
| 15431 | // We treat __stosb as a volatile memset - it may not generate "rep stosb" |
| 15432 | // instruction, but it will create a memset that won't be optimized away. |
| 15433 | return Builder.CreateMemSet(Ops[0], Ops[1], Ops[2], Align(1), true); |
| 15434 | } |
| 15435 | case X86::BI__ud2: |
| 15436 | // llvm.trap makes a ud2a instruction on x86. |
| 15437 | return EmitTrapCall(Intrinsic::trap); |
| 15438 | case X86::BI__int2c: { |
| 15439 | // This syscall signals a driver assertion failure in x86 NT kernels. |
| 15440 | llvm::FunctionType *FTy = llvm::FunctionType::get(VoidTy, false); |
| 15441 | llvm::InlineAsm *IA = |
| 15442 | llvm::InlineAsm::get(FTy, "int $$0x2c", "", /*hasSideEffects=*/true); |
| 15443 | llvm::AttributeList NoReturnAttr = llvm::AttributeList::get( |
| 15444 | getLLVMContext(), llvm::AttributeList::FunctionIndex, |
| 15445 | llvm::Attribute::NoReturn); |
| 15446 | llvm::CallInst *CI = Builder.CreateCall(IA); |
| 15447 | CI->setAttributes(NoReturnAttr); |
| 15448 | return CI; |
| 15449 | } |
| 15450 | case X86::BI__readfsbyte: |
| 15451 | case X86::BI__readfsword: |
| 15452 | case X86::BI__readfsdword: |
| 15453 | case X86::BI__readfsqword: { |
| 15454 | llvm::Type *IntTy = ConvertType(E->getType()); |
| 15455 | Value *Ptr = |
| 15456 | Builder.CreateIntToPtr(Ops[0], llvm::PointerType::get(IntTy, 257)); |
| 15457 | LoadInst *Load = Builder.CreateAlignedLoad( |
| 15458 | IntTy, Ptr, getContext().getTypeAlignInChars(E->getType())); |
| 15459 | Load->setVolatile(true); |
| 15460 | return Load; |
| 15461 | } |
| 15462 | case X86::BI__readgsbyte: |
| 15463 | case X86::BI__readgsword: |
| 15464 | case X86::BI__readgsdword: |
| 15465 | case X86::BI__readgsqword: { |
| 15466 | llvm::Type *IntTy = ConvertType(E->getType()); |
| 15467 | Value *Ptr = |
| 15468 | Builder.CreateIntToPtr(Ops[0], llvm::PointerType::get(IntTy, 256)); |
| 15469 | LoadInst *Load = Builder.CreateAlignedLoad( |
| 15470 | IntTy, Ptr, getContext().getTypeAlignInChars(E->getType())); |
| 15471 | Load->setVolatile(true); |
| 15472 | return Load; |
| 15473 | } |
| 15474 | case X86::BI__builtin_ia32_encodekey128_u32: { |
| 15475 | Intrinsic::ID IID = Intrinsic::x86_encodekey128; |
| 15476 | |
| 15477 | Value *Call = Builder.CreateCall(CGM.getIntrinsic(IID), {Ops[0], Ops[1]}); |
| 15478 | |
| 15479 | for (int i = 0; i < 3; ++i) { |
| 15480 | Value *Extract = Builder.CreateExtractValue(Call, i + 1); |
| 15481 | Value *Ptr = Builder.CreateConstGEP1_32(Int8Ty, Ops[2], i * 16); |
| 15482 | Ptr = Builder.CreateBitCast( |
| 15483 | Ptr, llvm::PointerType::getUnqual(Extract->getType())); |
| 15484 | Builder.CreateAlignedStore(Extract, Ptr, Align(1)); |
| 15485 | } |
| 15486 | |
| 15487 | return Builder.CreateExtractValue(Call, 0); |
| 15488 | } |
| 15489 | case X86::BI__builtin_ia32_encodekey256_u32: { |
| 15490 | Intrinsic::ID IID = Intrinsic::x86_encodekey256; |
| 15491 | |
| 15492 | Value *Call = |
| 15493 | Builder.CreateCall(CGM.getIntrinsic(IID), {Ops[0], Ops[1], Ops[2]}); |
| 15494 | |
| 15495 | for (int i = 0; i < 4; ++i) { |
| 15496 | Value *Extract = Builder.CreateExtractValue(Call, i + 1); |
| 15497 | Value *Ptr = Builder.CreateConstGEP1_32(Int8Ty, Ops[3], i * 16); |
| 15498 | Ptr = Builder.CreateBitCast( |
| 15499 | Ptr, llvm::PointerType::getUnqual(Extract->getType())); |
| 15500 | Builder.CreateAlignedStore(Extract, Ptr, Align(1)); |
| 15501 | } |
| 15502 | |
| 15503 | return Builder.CreateExtractValue(Call, 0); |
| 15504 | } |
| 15505 | case X86::BI__builtin_ia32_aesenc128kl_u8: |
| 15506 | case X86::BI__builtin_ia32_aesdec128kl_u8: |
| 15507 | case X86::BI__builtin_ia32_aesenc256kl_u8: |
| 15508 | case X86::BI__builtin_ia32_aesdec256kl_u8: { |
| 15509 | Intrinsic::ID IID; |
| 15510 | StringRef BlockName; |
| 15511 | switch (BuiltinID) { |
| 15512 | default: |
| 15513 | llvm_unreachable("Unexpected builtin")::llvm::llvm_unreachable_internal("Unexpected builtin", "clang/lib/CodeGen/CGBuiltin.cpp" , 15513); |
| 15514 | case X86::BI__builtin_ia32_aesenc128kl_u8: |
| 15515 | IID = Intrinsic::x86_aesenc128kl; |
| 15516 | BlockName = "aesenc128kl"; |
| 15517 | break; |
| 15518 | case X86::BI__builtin_ia32_aesdec128kl_u8: |
| 15519 | IID = Intrinsic::x86_aesdec128kl; |
| 15520 | BlockName = "aesdec128kl"; |
| 15521 | break; |
| 15522 | case X86::BI__builtin_ia32_aesenc256kl_u8: |
| 15523 | IID = Intrinsic::x86_aesenc256kl; |
| 15524 | BlockName = "aesenc256kl"; |
| 15525 | break; |
| 15526 | case X86::BI__builtin_ia32_aesdec256kl_u8: |
| 15527 | IID = Intrinsic::x86_aesdec256kl; |
| 15528 | BlockName = "aesdec256kl"; |
| 15529 | break; |
| 15530 | } |
| 15531 | |
| 15532 | Value *Call = Builder.CreateCall(CGM.getIntrinsic(IID), {Ops[1], Ops[2]}); |
| 15533 | |
| 15534 | BasicBlock *NoError = |
| 15535 | createBasicBlock(BlockName + "_no_error", this->CurFn); |
| 15536 | BasicBlock *Error = createBasicBlock(BlockName + "_error", this->CurFn); |
| 15537 | BasicBlock *End = createBasicBlock(BlockName + "_end", this->CurFn); |
| 15538 | |
| 15539 | Value *Ret = Builder.CreateExtractValue(Call, 0); |
| 15540 | Value *Succ = Builder.CreateTrunc(Ret, Builder.getInt1Ty()); |
| 15541 | Value *Out = Builder.CreateExtractValue(Call, 1); |
| 15542 | Builder.CreateCondBr(Succ, NoError, Error); |
| 15543 | |
| 15544 | Builder.SetInsertPoint(NoError); |
| 15545 | Builder.CreateDefaultAlignedStore(Out, Ops[0]); |
| 15546 | Builder.CreateBr(End); |
| 15547 | |
| 15548 | Builder.SetInsertPoint(Error); |
| 15549 | Constant *Zero = llvm::Constant::getNullValue(Out->getType()); |
| 15550 | Builder.CreateDefaultAlignedStore(Zero, Ops[0]); |
| 15551 | Builder.CreateBr(End); |
| 15552 | |
| 15553 | Builder.SetInsertPoint(End); |
| 15554 | return Builder.CreateExtractValue(Call, 0); |
| 15555 | } |
| 15556 | case X86::BI__builtin_ia32_aesencwide128kl_u8: |
| 15557 | case X86::BI__builtin_ia32_aesdecwide128kl_u8: |
| 15558 | case X86::BI__builtin_ia32_aesencwide256kl_u8: |
| 15559 | case X86::BI__builtin_ia32_aesdecwide256kl_u8: { |
| 15560 | Intrinsic::ID IID; |
| 15561 | StringRef BlockName; |
| 15562 | switch (BuiltinID) { |
| 15563 | case X86::BI__builtin_ia32_aesencwide128kl_u8: |
| 15564 | IID = Intrinsic::x86_aesencwide128kl; |
| 15565 | BlockName = "aesencwide128kl"; |
| 15566 | break; |
| 15567 | case X86::BI__builtin_ia32_aesdecwide128kl_u8: |
| 15568 | IID = Intrinsic::x86_aesdecwide128kl; |
| 15569 | BlockName = "aesdecwide128kl"; |
| 15570 | break; |
| 15571 | case X86::BI__builtin_ia32_aesencwide256kl_u8: |
| 15572 | IID = Intrinsic::x86_aesencwide256kl; |
| 15573 | BlockName = "aesencwide256kl"; |
| 15574 | break; |
| 15575 | case X86::BI__builtin_ia32_aesdecwide256kl_u8: |
| 15576 | IID = Intrinsic::x86_aesdecwide256kl; |
| 15577 | BlockName = "aesdecwide256kl"; |
| 15578 | break; |
| 15579 | } |
| 15580 | |
| 15581 | llvm::Type *Ty = FixedVectorType::get(Builder.getInt64Ty(), 2); |
| 15582 | Value *InOps[9]; |
| 15583 | InOps[0] = Ops[2]; |
| 15584 | for (int i = 0; i != 8; ++i) { |
| 15585 | Value *Ptr = Builder.CreateConstGEP1_32(Ty, Ops[1], i); |
| 15586 | InOps[i + 1] = Builder.CreateAlignedLoad(Ty, Ptr, Align(16)); |
| 15587 | } |
| 15588 | |
| 15589 | Value *Call = Builder.CreateCall(CGM.getIntrinsic(IID), InOps); |
| 15590 | |
| 15591 | BasicBlock *NoError = |
| 15592 | createBasicBlock(BlockName + "_no_error", this->CurFn); |
| 15593 | BasicBlock *Error = createBasicBlock(BlockName + "_error", this->CurFn); |
| 15594 | BasicBlock *End = createBasicBlock(BlockName + "_end", this->CurFn); |
| 15595 | |
| 15596 | Value *Ret = Builder.CreateExtractValue(Call, 0); |
| 15597 | Value *Succ = Builder.CreateTrunc(Ret, Builder.getInt1Ty()); |
| 15598 | Builder.CreateCondBr(Succ, NoError, Error); |
| 15599 | |
| 15600 | Builder.SetInsertPoint(NoError); |
| 15601 | for (int i = 0; i != 8; ++i) { |
| 15602 | Value *Extract = Builder.CreateExtractValue(Call, i + 1); |
| 15603 | Value *Ptr = Builder.CreateConstGEP1_32(Extract->getType(), Ops[0], i); |
| 15604 | Builder.CreateAlignedStore(Extract, Ptr, Align(16)); |
| 15605 | } |
| 15606 | Builder.CreateBr(End); |
| 15607 | |
| 15608 | Builder.SetInsertPoint(Error); |
| 15609 | for (int i = 0; i != 8; ++i) { |
| 15610 | Value *Out = Builder.CreateExtractValue(Call, i + 1); |
| 15611 | Constant *Zero = llvm::Constant::getNullValue(Out->getType()); |
| 15612 | Value *Ptr = Builder.CreateConstGEP1_32(Out->getType(), Ops[0], i); |
| 15613 | Builder.CreateAlignedStore(Zero, Ptr, Align(16)); |
| 15614 | } |
| 15615 | Builder.CreateBr(End); |
| 15616 | |
| 15617 | Builder.SetInsertPoint(End); |
| 15618 | return Builder.CreateExtractValue(Call, 0); |
| 15619 | } |
| 15620 | case X86::BI__builtin_ia32_vfcmaddcph512_mask: |
| 15621 | IsConjFMA = true; |
| 15622 | [[fallthrough]]; |
| 15623 | case X86::BI__builtin_ia32_vfmaddcph512_mask: { |
| 15624 | Intrinsic::ID IID = IsConjFMA |
| 15625 | ? Intrinsic::x86_avx512fp16_mask_vfcmadd_cph_512 |
| 15626 | : Intrinsic::x86_avx512fp16_mask_vfmadd_cph_512; |
| 15627 | Value *Call = Builder.CreateCall(CGM.getIntrinsic(IID), Ops); |
| 15628 | return EmitX86Select(*this, Ops[3], Call, Ops[0]); |
| 15629 | } |
| 15630 | case X86::BI__builtin_ia32_vfcmaddcsh_round_mask: |
| 15631 | IsConjFMA = true; |
| 15632 | [[fallthrough]]; |
| 15633 | case X86::BI__builtin_ia32_vfmaddcsh_round_mask: { |
| 15634 | Intrinsic::ID IID = IsConjFMA ? Intrinsic::x86_avx512fp16_mask_vfcmadd_csh |
| 15635 | : Intrinsic::x86_avx512fp16_mask_vfmadd_csh; |
| 15636 | Value *Call = Builder.CreateCall(CGM.getIntrinsic(IID), Ops); |
| 15637 | Value *And = Builder.CreateAnd(Ops[3], llvm::ConstantInt::get(Int8Ty, 1)); |
| 15638 | return EmitX86Select(*this, And, Call, Ops[0]); |
| 15639 | } |
| 15640 | case X86::BI__builtin_ia32_vfcmaddcsh_round_mask3: |
| 15641 | IsConjFMA = true; |
| 15642 | [[fallthrough]]; |
| 15643 | case X86::BI__builtin_ia32_vfmaddcsh_round_mask3: { |
| 15644 | Intrinsic::ID IID = IsConjFMA ? Intrinsic::x86_avx512fp16_mask_vfcmadd_csh |
| 15645 | : Intrinsic::x86_avx512fp16_mask_vfmadd_csh; |
| 15646 | Value *Call = Builder.CreateCall(CGM.getIntrinsic(IID), Ops); |
| 15647 | static constexpr int Mask[] = {0, 5, 6, 7}; |
| 15648 | return Builder.CreateShuffleVector(Call, Ops[2], Mask); |
| 15649 | } |
| 15650 | case X86::BI__builtin_ia32_prefetchi: |
| 15651 | return Builder.CreateCall( |
| 15652 | CGM.getIntrinsic(Intrinsic::prefetch, Ops[0]->getType()), |
| 15653 | {Ops[0], llvm::ConstantInt::get(Int32Ty, 0), Ops[1], |
| 15654 | llvm::ConstantInt::get(Int32Ty, 0)}); |
| 15655 | } |
| 15656 | } |
| 15657 | |
| 15658 | Value *CodeGenFunction::EmitPPCBuiltinExpr(unsigned BuiltinID, |
| 15659 | const CallExpr *E) { |
| 15660 | // Do not emit the builtin arguments in the arguments of a function call, |
| 15661 | // because the evaluation order of function arguments is not specified in C++. |
| 15662 | // This is important when testing to ensure the arguments are emitted in the |
| 15663 | // same order every time. Eg: |
| 15664 | // Instead of: |
| 15665 | // return Builder.CreateFDiv(EmitScalarExpr(E->getArg(0)), |
| 15666 | // EmitScalarExpr(E->getArg(1)), "swdiv"); |
| 15667 | // Use: |
| 15668 | // Value *Op0 = EmitScalarExpr(E->getArg(0)); |
| 15669 | // Value *Op1 = EmitScalarExpr(E->getArg(1)); |
| 15670 | // return Builder.CreateFDiv(Op0, Op1, "swdiv") |
| 15671 | |
| 15672 | Intrinsic::ID ID = Intrinsic::not_intrinsic; |
| 15673 | |
| 15674 | switch (BuiltinID) { |
| 15675 | default: return nullptr; |
| 15676 | |
| 15677 | // __builtin_ppc_get_timebase is GCC 4.8+'s PowerPC-specific name for what we |
| 15678 | // call __builtin_readcyclecounter. |
| 15679 | case PPC::BI__builtin_ppc_get_timebase: |
| 15680 | return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::readcyclecounter)); |
| 15681 | |
| 15682 | // vec_ld, vec_xl_be, vec_lvsl, vec_lvsr |
| 15683 | case PPC::BI__builtin_altivec_lvx: |
| 15684 | case PPC::BI__builtin_altivec_lvxl: |
| 15685 | case PPC::BI__builtin_altivec_lvebx: |
| 15686 | case PPC::BI__builtin_altivec_lvehx: |
| 15687 | case PPC::BI__builtin_altivec_lvewx: |
| 15688 | case PPC::BI__builtin_altivec_lvsl: |
| 15689 | case PPC::BI__builtin_altivec_lvsr: |
| 15690 | case PPC::BI__builtin_vsx_lxvd2x: |
| 15691 | case PPC::BI__builtin_vsx_lxvw4x: |
| 15692 | case PPC::BI__builtin_vsx_lxvd2x_be: |
| 15693 | case PPC::BI__builtin_vsx_lxvw4x_be: |
| 15694 | case PPC::BI__builtin_vsx_lxvl: |
| 15695 | case PPC::BI__builtin_vsx_lxvll: |
| 15696 | { |
| 15697 | SmallVector<Value *, 2> Ops; |
| 15698 | Ops.push_back(EmitScalarExpr(E->getArg(0))); |
| 15699 | Ops.push_back(EmitScalarExpr(E->getArg(1))); |
| 15700 | if(BuiltinID == PPC::BI__builtin_vsx_lxvl || |
| 15701 | BuiltinID == PPC::BI__builtin_vsx_lxvll){ |
| 15702 | Ops[0] = Builder.CreateBitCast(Ops[0], Int8PtrTy); |
| 15703 | }else { |
| 15704 | Ops[1] = Builder.CreateBitCast(Ops[1], Int8PtrTy); |
| 15705 | Ops[0] = Builder.CreateGEP(Int8Ty, Ops[1], Ops[0]); |
| 15706 | Ops.pop_back(); |
| 15707 | } |
| 15708 | |
| 15709 | switch (BuiltinID) { |
| 15710 | default: llvm_unreachable("Unsupported ld/lvsl/lvsr intrinsic!")::llvm::llvm_unreachable_internal("Unsupported ld/lvsl/lvsr intrinsic!" , "clang/lib/CodeGen/CGBuiltin.cpp", 15710); |
| 15711 | case PPC::BI__builtin_altivec_lvx: |
| 15712 | ID = Intrinsic::ppc_altivec_lvx; |
| 15713 | break; |
| 15714 | case PPC::BI__builtin_altivec_lvxl: |
| 15715 | ID = Intrinsic::ppc_altivec_lvxl; |
| 15716 | break; |
| 15717 | case PPC::BI__builtin_altivec_lvebx: |
| 15718 | ID = Intrinsic::ppc_altivec_lvebx; |
| 15719 | break; |
| 15720 | case PPC::BI__builtin_altivec_lvehx: |
| 15721 | ID = Intrinsic::ppc_altivec_lvehx; |
| 15722 | break; |
| 15723 | case PPC::BI__builtin_altivec_lvewx: |
| 15724 | ID = Intrinsic::ppc_altivec_lvewx; |
| 15725 | break; |
| 15726 | case PPC::BI__builtin_altivec_lvsl: |
| 15727 | ID = Intrinsic::ppc_altivec_lvsl; |
| 15728 | break; |
| 15729 | case PPC::BI__builtin_altivec_lvsr: |
| 15730 | ID = Intrinsic::ppc_altivec_lvsr; |
| 15731 | break; |
| 15732 | case PPC::BI__builtin_vsx_lxvd2x: |
| 15733 | ID = Intrinsic::ppc_vsx_lxvd2x; |
| 15734 | break; |
| 15735 | case PPC::BI__builtin_vsx_lxvw4x: |
| 15736 | ID = Intrinsic::ppc_vsx_lxvw4x; |
| 15737 | break; |
| 15738 | case PPC::BI__builtin_vsx_lxvd2x_be: |
| 15739 | ID = Intrinsic::ppc_vsx_lxvd2x_be; |
| 15740 | break; |
| 15741 | case PPC::BI__builtin_vsx_lxvw4x_be: |
| 15742 | ID = Intrinsic::ppc_vsx_lxvw4x_be; |
| 15743 | break; |
| 15744 | case PPC::BI__builtin_vsx_lxvl: |
| 15745 | ID = Intrinsic::ppc_vsx_lxvl; |
| 15746 | break; |
| 15747 | case PPC::BI__builtin_vsx_lxvll: |
| 15748 | ID = Intrinsic::ppc_vsx_lxvll; |
| 15749 | break; |
| 15750 | } |
| 15751 | llvm::Function *F = CGM.getIntrinsic(ID); |
| 15752 | return Builder.CreateCall(F, Ops, ""); |
| 15753 | } |
| 15754 | |
| 15755 | // vec_st, vec_xst_be |
| 15756 | case PPC::BI__builtin_altivec_stvx: |
| 15757 | case PPC::BI__builtin_altivec_stvxl: |
| 15758 | case PPC::BI__builtin_altivec_stvebx: |
| 15759 | case PPC::BI__builtin_altivec_stvehx: |
| 15760 | case PPC::BI__builtin_altivec_stvewx: |
| 15761 | case PPC::BI__builtin_vsx_stxvd2x: |
| 15762 | case PPC::BI__builtin_vsx_stxvw4x: |
| 15763 | case PPC::BI__builtin_vsx_stxvd2x_be: |
| 15764 | case PPC::BI__builtin_vsx_stxvw4x_be: |
| 15765 | case PPC::BI__builtin_vsx_stxvl: |
| 15766 | case PPC::BI__builtin_vsx_stxvll: |
| 15767 | { |
| 15768 | SmallVector<Value *, 3> Ops; |
| 15769 | Ops.push_back(EmitScalarExpr(E->getArg(0))); |
| 15770 | Ops.push_back(EmitScalarExpr(E->getArg(1))); |
| 15771 | Ops.push_back(EmitScalarExpr(E->getArg(2))); |
| 15772 | if(BuiltinID == PPC::BI__builtin_vsx_stxvl || |
| 15773 | BuiltinID == PPC::BI__builtin_vsx_stxvll ){ |
| 15774 | Ops[1] = Builder.CreateBitCast(Ops[1], Int8PtrTy); |
| 15775 | }else { |
| 15776 | Ops[2] = Builder.CreateBitCast(Ops[2], Int8PtrTy); |
| 15777 | Ops[1] = Builder.CreateGEP(Int8Ty, Ops[2], Ops[1]); |
| 15778 | Ops.pop_back(); |
| 15779 | } |
| 15780 | |
| 15781 | switch (BuiltinID) { |
| 15782 | default: llvm_unreachable("Unsupported st intrinsic!")::llvm::llvm_unreachable_internal("Unsupported st intrinsic!" , "clang/lib/CodeGen/CGBuiltin.cpp", 15782); |
| 15783 | case PPC::BI__builtin_altivec_stvx: |
| 15784 | ID = Intrinsic::ppc_altivec_stvx; |
| 15785 | break; |
| 15786 | case PPC::BI__builtin_altivec_stvxl: |
| 15787 | ID = Intrinsic::ppc_altivec_stvxl; |
| 15788 | break; |
| 15789 | case PPC::BI__builtin_altivec_stvebx: |
| 15790 | ID = Intrinsic::ppc_altivec_stvebx; |
| 15791 | break; |
| 15792 | case PPC::BI__builtin_altivec_stvehx: |
| 15793 | ID = Intrinsic::ppc_altivec_stvehx; |
| 15794 | break; |
| 15795 | case PPC::BI__builtin_altivec_stvewx: |
| 15796 | ID = Intrinsic::ppc_altivec_stvewx; |
| 15797 | break; |
| 15798 | case PPC::BI__builtin_vsx_stxvd2x: |
| 15799 | ID = Intrinsic::ppc_vsx_stxvd2x; |
| 15800 | break; |
| 15801 | case PPC::BI__builtin_vsx_stxvw4x: |
| 15802 | ID = Intrinsic::ppc_vsx_stxvw4x; |
| 15803 | break; |
| 15804 | case PPC::BI__builtin_vsx_stxvd2x_be: |
| 15805 | ID = Intrinsic::ppc_vsx_stxvd2x_be; |
| 15806 | break; |
| 15807 | case PPC::BI__builtin_vsx_stxvw4x_be: |
| 15808 | ID = Intrinsic::ppc_vsx_stxvw4x_be; |
| 15809 | break; |
| 15810 | case PPC::BI__builtin_vsx_stxvl: |
| 15811 | ID = Intrinsic::ppc_vsx_stxvl; |
| 15812 | break; |
| 15813 | case PPC::BI__builtin_vsx_stxvll: |
| 15814 | ID = Intrinsic::ppc_vsx_stxvll; |
| 15815 | break; |
| 15816 | } |
| 15817 | llvm::Function *F = CGM.getIntrinsic(ID); |
| 15818 | return Builder.CreateCall(F, Ops, ""); |
| 15819 | } |
| 15820 | case PPC::BI__builtin_vsx_ldrmb: { |
| 15821 | // Essentially boils down to performing an unaligned VMX load sequence so |
| 15822 | // as to avoid crossing a page boundary and then shuffling the elements |
| 15823 | // into the right side of the vector register. |
| 15824 | Value *Op0 = EmitScalarExpr(E->getArg(0)); |
| 15825 | Value *Op1 = EmitScalarExpr(E->getArg(1)); |
| 15826 | int64_t NumBytes = cast<ConstantInt>(Op1)->getZExtValue(); |
| 15827 | llvm::Type *ResTy = ConvertType(E->getType()); |
| 15828 | bool IsLE = getTarget().isLittleEndian(); |
| 15829 | |
| 15830 | // If the user wants the entire vector, just load the entire vector. |
| 15831 | if (NumBytes == 16) { |
| 15832 | Value *BC = Builder.CreateBitCast(Op0, ResTy->getPointerTo()); |
| 15833 | Value *LD = |
| 15834 | Builder.CreateLoad(Address(BC, ResTy, CharUnits::fromQuantity(1))); |
| 15835 | if (!IsLE) |
| 15836 | return LD; |
| 15837 | |
| 15838 | // Reverse the bytes on LE. |
| 15839 | SmallVector<int, 16> RevMask; |
| 15840 | for (int Idx = 0; Idx < 16; Idx++) |
| 15841 | RevMask.push_back(15 - Idx); |
| 15842 | return Builder.CreateShuffleVector(LD, LD, RevMask); |
| 15843 | } |
| 15844 | |
| 15845 | llvm::Function *Lvx = CGM.getIntrinsic(Intrinsic::ppc_altivec_lvx); |
| 15846 | llvm::Function *Lvs = CGM.getIntrinsic(IsLE ? Intrinsic::ppc_altivec_lvsr |
| 15847 | : Intrinsic::ppc_altivec_lvsl); |
| 15848 | llvm::Function *Vperm = CGM.getIntrinsic(Intrinsic::ppc_altivec_vperm); |
| 15849 | Value *HiMem = Builder.CreateGEP( |
| 15850 | Int8Ty, Op0, ConstantInt::get(Op1->getType(), NumBytes - 1)); |
| 15851 | Value *LoLd = Builder.CreateCall(Lvx, Op0, "ld.lo"); |
| 15852 | Value *HiLd = Builder.CreateCall(Lvx, HiMem, "ld.hi"); |
| 15853 | Value *Mask1 = Builder.CreateCall(Lvs, Op0, "mask1"); |
| 15854 | |
| 15855 | Op0 = IsLE ? HiLd : LoLd; |
| 15856 | Op1 = IsLE ? LoLd : HiLd; |
| 15857 | Value *AllElts = Builder.CreateCall(Vperm, {Op0, Op1, Mask1}, "shuffle1"); |
| 15858 | Constant *Zero = llvm::Constant::getNullValue(IsLE ? ResTy : AllElts->getType()); |
| 15859 | |
| 15860 | if (IsLE) { |
| 15861 | SmallVector<int, 16> Consts; |
| 15862 | for (int Idx = 0; Idx < 16; Idx++) { |
| 15863 | int Val = (NumBytes - Idx - 1 >= 0) ? (NumBytes - Idx - 1) |
| 15864 | : 16 - (NumBytes - Idx); |
| 15865 | Consts.push_back(Val); |
| 15866 | } |
| 15867 | return Builder.CreateShuffleVector(Builder.CreateBitCast(AllElts, ResTy), |
| 15868 | Zero, Consts); |
| 15869 | } |
| 15870 | SmallVector<Constant *, 16> Consts; |
| 15871 | for (int Idx = 0; Idx < 16; Idx++) |
| 15872 | Consts.push_back(Builder.getInt8(NumBytes + Idx)); |
| 15873 | Value *Mask2 = ConstantVector::get(Consts); |
| 15874 | return Builder.CreateBitCast( |
| 15875 | Builder.CreateCall(Vperm, {Zero, AllElts, Mask2}, "shuffle2"), ResTy); |
| 15876 | } |
| 15877 | case PPC::BI__builtin_vsx_strmb: { |
| 15878 | Value *Op0 = EmitScalarExpr(E->getArg(0)); |
| 15879 | Value *Op1 = EmitScalarExpr(E->getArg(1)); |
| 15880 | Value *Op2 = EmitScalarExpr(E->getArg(2)); |
| 15881 | int64_t NumBytes = cast<ConstantInt>(Op1)->getZExtValue(); |
| 15882 | bool IsLE = getTarget().isLittleEndian(); |
| 15883 | auto StoreSubVec = [&](unsigned Width, unsigned Offset, unsigned EltNo) { |
| 15884 | // Storing the whole vector, simply store it on BE and reverse bytes and |
| 15885 | // store on LE. |
| 15886 | if (Width == 16) { |
| 15887 | Value *BC = Builder.CreateBitCast(Op0, Op2->getType()->getPointerTo()); |
| 15888 | Value *StVec = Op2; |
| 15889 | if (IsLE) { |
| 15890 | SmallVector<int, 16> RevMask; |
| 15891 | for (int Idx = 0; Idx < 16; Idx++) |
| 15892 | RevMask.push_back(15 - Idx); |
| 15893 | StVec = Builder.CreateShuffleVector(Op2, Op2, RevMask); |
| 15894 | } |
| 15895 | return Builder.CreateStore( |
| 15896 | StVec, Address(BC, Op2->getType(), CharUnits::fromQuantity(1))); |
| 15897 | } |
| 15898 | auto *ConvTy = Int64Ty; |
| 15899 | unsigned NumElts = 0; |
| 15900 | switch (Width) { |
| 15901 | default: |
| 15902 | llvm_unreachable("width for stores must be a power of 2")::llvm::llvm_unreachable_internal("width for stores must be a power of 2" , "clang/lib/CodeGen/CGBuiltin.cpp", 15902); |
| 15903 | case 8: |
| 15904 | ConvTy = Int64Ty; |
| 15905 | NumElts = 2; |
| 15906 | break; |
| 15907 | case 4: |
| 15908 | ConvTy = Int32Ty; |
| 15909 | NumElts = 4; |
| 15910 | break; |
| 15911 | case 2: |
| 15912 | ConvTy = Int16Ty; |
| 15913 | NumElts = 8; |
| 15914 | break; |
| 15915 | case 1: |
| 15916 | ConvTy = Int8Ty; |
| 15917 | NumElts = 16; |
| 15918 | break; |
| 15919 | } |
| 15920 | Value *Vec = Builder.CreateBitCast( |
| 15921 | Op2, llvm::FixedVectorType::get(ConvTy, NumElts)); |
| 15922 | Value *Ptr = |
| 15923 | Builder.CreateGEP(Int8Ty, Op0, ConstantInt::get(Int64Ty, Offset)); |
| 15924 | Value *PtrBC = Builder.CreateBitCast(Ptr, ConvTy->getPointerTo()); |
| 15925 | Value *Elt = Builder.CreateExtractElement(Vec, EltNo); |
| 15926 | if (IsLE && Width > 1) { |
| 15927 | Function *F = CGM.getIntrinsic(Intrinsic::bswap, ConvTy); |
| 15928 | Elt = Builder.CreateCall(F, Elt); |
| 15929 | } |
| 15930 | return Builder.CreateStore( |
| 15931 | Elt, Address(PtrBC, ConvTy, CharUnits::fromQuantity(1))); |
| 15932 | }; |
| 15933 | unsigned Stored = 0; |
| 15934 | unsigned RemainingBytes = NumBytes; |
| 15935 | Value *Result; |
| 15936 | if (NumBytes == 16) |
| 15937 | return StoreSubVec(16, 0, 0); |
| 15938 | if (NumBytes >= 8) { |
| 15939 | Result = StoreSubVec(8, NumBytes - 8, IsLE ? 0 : 1); |
| 15940 | RemainingBytes -= 8; |
| 15941 | Stored += 8; |
| 15942 | } |
| 15943 | if (RemainingBytes >= 4) { |
| 15944 | Result = StoreSubVec(4, NumBytes - Stored - 4, |
| 15945 | IsLE ? (Stored >> 2) : 3 - (Stored >> 2)); |
| 15946 | RemainingBytes -= 4; |
| 15947 | Stored += 4; |
| 15948 | } |
| 15949 | if (RemainingBytes >= 2) { |
| 15950 | Result = StoreSubVec(2, NumBytes - Stored - 2, |
| 15951 | IsLE ? (Stored >> 1) : 7 - (Stored >> 1)); |
| 15952 | RemainingBytes -= 2; |
| 15953 | Stored += 2; |
| 15954 | } |
| 15955 | if (RemainingBytes) |
| 15956 | Result = |
| 15957 | StoreSubVec(1, NumBytes - Stored - 1, IsLE ? Stored : 15 - Stored); |
| 15958 | return Result; |
| 15959 | } |
| 15960 | // Square root |
| 15961 | case PPC::BI__builtin_vsx_xvsqrtsp: |
| 15962 | case PPC::BI__builtin_vsx_xvsqrtdp: { |
| 15963 | llvm::Type *ResultType = ConvertType(E->getType()); |
| 15964 | Value *X = EmitScalarExpr(E->getArg(0)); |
| 15965 | if (Builder.getIsFPConstrained()) { |
| 15966 | llvm::Function *F = CGM.getIntrinsic( |
| 15967 | Intrinsic::experimental_constrained_sqrt, ResultType); |
| 15968 | return Builder.CreateConstrainedFPCall(F, X); |
| 15969 | } else { |
| 15970 | llvm::Function *F = CGM.getIntrinsic(Intrinsic::sqrt, ResultType); |
| 15971 | return Builder.CreateCall(F, X); |
| 15972 | } |
| 15973 | } |
| 15974 | // Count leading zeros |
| 15975 | case PPC::BI__builtin_altivec_vclzb: |
| 15976 | case PPC::BI__builtin_altivec_vclzh: |
| 15977 | case PPC::BI__builtin_altivec_vclzw: |
| 15978 | case PPC::BI__builtin_altivec_vclzd: { |
| 15979 | llvm::Type *ResultType = ConvertType(E->getType()); |
| 15980 | Value *X = EmitScalarExpr(E->getArg(0)); |
| 15981 | Value *Undef = ConstantInt::get(Builder.getInt1Ty(), false); |
| 15982 | Function *F = CGM.getIntrinsic(Intrinsic::ctlz, ResultType); |
| 15983 | return Builder.CreateCall(F, {X, Undef}); |
| 15984 | } |
| 15985 | case PPC::BI__builtin_altivec_vctzb: |
| 15986 | case PPC::BI__builtin_altivec_vctzh: |
| 15987 | case PPC::BI__builtin_altivec_vctzw: |
| 15988 | case PPC::BI__builtin_altivec_vctzd: { |
| 15989 | llvm::Type *ResultType = ConvertType(E->getType()); |
| 15990 | Value *X = EmitScalarExpr(E->getArg(0)); |
| 15991 | Value *Undef = ConstantInt::get(Builder.getInt1Ty(), false); |
| 15992 | Function *F = CGM.getIntrinsic(Intrinsic::cttz, ResultType); |
| 15993 | return Builder.CreateCall(F, {X, Undef}); |
| 15994 | } |
| 15995 | case PPC::BI__builtin_altivec_vinsd: |
| 15996 | case PPC::BI__builtin_altivec_vinsw: |
| 15997 | case PPC::BI__builtin_altivec_vinsd_elt: |
| 15998 | case PPC::BI__builtin_altivec_vinsw_elt: { |
| 15999 | llvm::Type *ResultType = ConvertType(E->getType()); |
| 16000 | Value *Op0 = EmitScalarExpr(E->getArg(0)); |
| 16001 | Value *Op1 = EmitScalarExpr(E->getArg(1)); |
| 16002 | Value *Op2 = EmitScalarExpr(E->getArg(2)); |
| 16003 | |
| 16004 | bool IsUnaligned = (BuiltinID == PPC::BI__builtin_altivec_vinsw || |
| 16005 | BuiltinID == PPC::BI__builtin_altivec_vinsd); |
| 16006 | |
| 16007 | bool Is32bit = (BuiltinID == PPC::BI__builtin_altivec_vinsw || |
| 16008 | BuiltinID == PPC::BI__builtin_altivec_vinsw_elt); |
| 16009 | |
| 16010 | // The third argument must be a compile time constant. |
| 16011 | ConstantInt *ArgCI = dyn_cast<ConstantInt>(Op2); |
| 16012 | assert(ArgCI &&(static_cast <bool> (ArgCI && "Third Arg to vinsw/vinsd intrinsic must be a constant integer!" ) ? void (0) : __assert_fail ("ArgCI && \"Third Arg to vinsw/vinsd intrinsic must be a constant integer!\"" , "clang/lib/CodeGen/CGBuiltin.cpp", 16013, __extension__ __PRETTY_FUNCTION__ )) |
| 16013 | "Third Arg to vinsw/vinsd intrinsic must be a constant integer!")(static_cast <bool> (ArgCI && "Third Arg to vinsw/vinsd intrinsic must be a constant integer!" ) ? void (0) : __assert_fail ("ArgCI && \"Third Arg to vinsw/vinsd intrinsic must be a constant integer!\"" , "clang/lib/CodeGen/CGBuiltin.cpp", 16013, __extension__ __PRETTY_FUNCTION__ )); |
| 16014 | |
| 16015 | // Valid value for the third argument is dependent on the input type and |
| 16016 | // builtin called. |
| 16017 | int ValidMaxValue = 0; |
| 16018 | if (IsUnaligned) |
| 16019 | ValidMaxValue = (Is32bit) ? 12 : 8; |
| 16020 | else |
| 16021 | ValidMaxValue = (Is32bit) ? 3 : 1; |
| 16022 | |
| 16023 | // Get value of third argument. |
| 16024 | int64_t ConstArg = ArgCI->getSExtValue(); |
| 16025 | |
| 16026 | // Compose range checking error message. |
| 16027 | std::string RangeErrMsg = IsUnaligned ? "byte" : "element"; |
| 16028 | RangeErrMsg += " number " + llvm::to_string(ConstArg); |
| 16029 | RangeErrMsg += " is outside of the valid range [0, "; |
| 16030 | RangeErrMsg += llvm::to_string(ValidMaxValue) + "]"; |
| 16031 | |
| 16032 | // Issue error if third argument is not within the valid range. |
| 16033 | if (ConstArg < 0 || ConstArg > ValidMaxValue) |
| 16034 | CGM.Error(E->getExprLoc(), RangeErrMsg); |
| 16035 | |
| 16036 | // Input to vec_replace_elt is an element index, convert to byte index. |
| 16037 | if (!IsUnaligned) { |
| 16038 | ConstArg *= Is32bit ? 4 : 8; |
| 16039 | // Fix the constant according to endianess. |
| 16040 | if (getTarget().isLittleEndian()) |
| 16041 | ConstArg = (Is32bit ? 12 : 8) - ConstArg; |
| 16042 | } |
| 16043 | |
| 16044 | ID = Is32bit ? Intrinsic::ppc_altivec_vinsw : Intrinsic::ppc_altivec_vinsd; |
| 16045 | Op2 = ConstantInt::getSigned(Int32Ty, ConstArg); |
| 16046 | // Casting input to vector int as per intrinsic definition. |
| 16047 | Op0 = |
| 16048 | Is32bit |
| 16049 | ? Builder.CreateBitCast(Op0, llvm::FixedVectorType::get(Int32Ty, 4)) |
| 16050 | : Builder.CreateBitCast(Op0, |
| 16051 | llvm::FixedVectorType::get(Int64Ty, 2)); |
| 16052 | return Builder.CreateBitCast( |
| 16053 | Builder.CreateCall(CGM.getIntrinsic(ID), {Op0, Op1, Op2}), ResultType); |
| 16054 | } |
| 16055 | case PPC::BI__builtin_altivec_vpopcntb: |
| 16056 | case PPC::BI__builtin_altivec_vpopcnth: |
| 16057 | case PPC::BI__builtin_altivec_vpopcntw: |
| 16058 | case PPC::BI__builtin_altivec_vpopcntd: { |
| 16059 | llvm::Type *ResultType = ConvertType(E->getType()); |
| 16060 | Value *X = EmitScalarExpr(E->getArg(0)); |
| 16061 | llvm::Function *F = CGM.getIntrinsic(Intrinsic::ctpop, ResultType); |
| 16062 | return Builder.CreateCall(F, X); |
| 16063 | } |
| 16064 | case PPC::BI__builtin_altivec_vadduqm: |
| 16065 | case PPC::BI__builtin_altivec_vsubuqm: { |
| 16066 | Value *Op0 = EmitScalarExpr(E->getArg(0)); |
| 16067 | Value *Op1 = EmitScalarExpr(E->getArg(1)); |
| 16068 | llvm::Type *Int128Ty = llvm::IntegerType::get(getLLVMContext(), 128); |
| 16069 | Op0 = Builder.CreateBitCast(Op0, llvm::FixedVectorType::get(Int128Ty, 1)); |
| 16070 | Op1 = Builder.CreateBitCast(Op1, llvm::FixedVectorType::get(Int128Ty, 1)); |
| 16071 | if (BuiltinID == PPC::BI__builtin_altivec_vadduqm) |
| 16072 | return Builder.CreateAdd(Op0, Op1, "vadduqm"); |
| 16073 | else |
| 16074 | return Builder.CreateSub(Op0, Op1, "vsubuqm"); |
| 16075 | } |
| 16076 | case PPC::BI__builtin_altivec_vaddcuq_c: |
| 16077 | case PPC::BI__builtin_altivec_vsubcuq_c: { |
| 16078 | SmallVector<Value *, 2> Ops; |
| 16079 | Value *Op0 = EmitScalarExpr(E->getArg(0)); |
| 16080 | Value *Op1 = EmitScalarExpr(E->getArg(1)); |
| 16081 | llvm::Type *V1I128Ty = llvm::FixedVectorType::get( |
| 16082 | llvm::IntegerType::get(getLLVMContext(), 128), 1); |
| 16083 | Ops.push_back(Builder.CreateBitCast(Op0, V1I128Ty)); |
| 16084 | Ops.push_back(Builder.CreateBitCast(Op1, V1I128Ty)); |
| 16085 | ID = (BuiltinID == PPC::BI__builtin_altivec_vaddcuq_c) |
| 16086 | ? Intrinsic::ppc_altivec_vaddcuq |
| 16087 | : Intrinsic::ppc_altivec_vsubcuq; |
| 16088 | return Builder.CreateCall(CGM.getIntrinsic(ID), Ops, ""); |
| 16089 | } |
| 16090 | case PPC::BI__builtin_altivec_vaddeuqm_c: |
| 16091 | case PPC::BI__builtin_altivec_vaddecuq_c: |
| 16092 | case PPC::BI__builtin_altivec_vsubeuqm_c: |
| 16093 | case PPC::BI__builtin_altivec_vsubecuq_c: { |
| 16094 | SmallVector<Value *, 3> Ops; |
| 16095 | Value *Op0 = EmitScalarExpr(E->getArg(0)); |
| 16096 | Value *Op1 = EmitScalarExpr(E->getArg(1)); |
| 16097 | Value *Op2 = EmitScalarExpr(E->getArg(2)); |
| 16098 | llvm::Type *V1I128Ty = llvm::FixedVectorType::get( |
| 16099 | llvm::IntegerType::get(getLLVMContext(), 128), 1); |
| 16100 | Ops.push_back(Builder.CreateBitCast(Op0, V1I128Ty)); |
| 16101 | Ops.push_back(Builder.CreateBitCast(Op1, V1I128Ty)); |
| 16102 | Ops.push_back(Builder.CreateBitCast(Op2, V1I128Ty)); |
| 16103 | switch (BuiltinID) { |
| 16104 | default: |
| 16105 | llvm_unreachable("Unsupported intrinsic!")::llvm::llvm_unreachable_internal("Unsupported intrinsic!", "clang/lib/CodeGen/CGBuiltin.cpp" , 16105); |
| 16106 | case PPC::BI__builtin_altivec_vaddeuqm_c: |
| 16107 | ID = Intrinsic::ppc_altivec_vaddeuqm; |
| 16108 | break; |
| 16109 | case PPC::BI__builtin_altivec_vaddecuq_c: |
| 16110 | ID = Intrinsic::ppc_altivec_vaddecuq; |
| 16111 | break; |
| 16112 | case PPC::BI__builtin_altivec_vsubeuqm_c: |
| 16113 | ID = Intrinsic::ppc_altivec_vsubeuqm; |
| 16114 | break; |
| 16115 | case PPC::BI__builtin_altivec_vsubecuq_c: |
| 16116 | ID = Intrinsic::ppc_altivec_vsubecuq; |
| 16117 | break; |
| 16118 | } |
| 16119 | return Builder.CreateCall(CGM.getIntrinsic(ID), Ops, ""); |
| 16120 | } |
| 16121 | // Rotate and insert under mask operation. |
| 16122 | // __rldimi(rs, is, shift, mask) |
| 16123 | // (rotl64(rs, shift) & mask) | (is & ~mask) |
| 16124 | // __rlwimi(rs, is, shift, mask) |
| 16125 | // (rotl(rs, shift) & mask) | (is & ~mask) |
| 16126 | case PPC::BI__builtin_ppc_rldimi: |
| 16127 | case PPC::BI__builtin_ppc_rlwimi: { |
| 16128 | Value *Op0 = EmitScalarExpr(E->getArg(0)); |
| 16129 | Value *Op1 = EmitScalarExpr(E->getArg(1)); |
| 16130 | Value *Op2 = EmitScalarExpr(E->getArg(2)); |
| 16131 | Value *Op3 = EmitScalarExpr(E->getArg(3)); |
| 16132 | llvm::Type *Ty = Op0->getType(); |
| 16133 | Function *F = CGM.getIntrinsic(Intrinsic::fshl, Ty); |
| 16134 | if (BuiltinID == PPC::BI__builtin_ppc_rldimi) |
| 16135 | Op2 = Builder.CreateZExt(Op2, Int64Ty); |
| 16136 | Value *Shift = Builder.CreateCall(F, {Op0, Op0, Op2}); |
| 16137 | Value *X = Builder.CreateAnd(Shift, Op3); |
| 16138 | Value *Y = Builder.CreateAnd(Op1, Builder.CreateNot(Op3)); |
| 16139 | return Builder.CreateOr(X, Y); |
| 16140 | } |
| 16141 | // Rotate and insert under mask operation. |
| 16142 | // __rlwnm(rs, shift, mask) |
| 16143 | // rotl(rs, shift) & mask |
| 16144 | case PPC::BI__builtin_ppc_rlwnm: { |
| 16145 | Value *Op0 = EmitScalarExpr(E->getArg(0)); |
| 16146 | Value *Op1 = EmitScalarExpr(E->getArg(1)); |
| 16147 | Value *Op2 = EmitScalarExpr(E->getArg(2)); |
| 16148 | llvm::Type *Ty = Op0->getType(); |
| 16149 | Function *F = CGM.getIntrinsic(Intrinsic::fshl, Ty); |
| 16150 | Value *Shift = Builder.CreateCall(F, {Op0, Op0, Op1}); |
| 16151 | return Builder.CreateAnd(Shift, Op2); |
| 16152 | } |
| 16153 | case PPC::BI__builtin_ppc_poppar4: |
| 16154 | case PPC::BI__builtin_ppc_poppar8: { |
| 16155 | Value *Op0 = EmitScalarExpr(E->getArg(0)); |
| 16156 | llvm::Type *ArgType = Op0->getType(); |
| 16157 | Function *F = CGM.getIntrinsic(Intrinsic::ctpop, ArgType); |
| 16158 | Value *Tmp = Builder.CreateCall(F, Op0); |
| 16159 | |
| 16160 | llvm::Type *ResultType = ConvertType(E->getType()); |
| 16161 | Value *Result = Builder.CreateAnd(Tmp, llvm::ConstantInt::get(ArgType, 1)); |
| 16162 | if (Result->getType() != ResultType) |
| 16163 | Result = Builder.CreateIntCast(Result, ResultType, /*isSigned*/true, |
| 16164 | "cast"); |
| 16165 | return Result; |
| 16166 | } |
| 16167 | case PPC::BI__builtin_ppc_cmpb: { |
| 16168 | Value *Op0 = EmitScalarExpr(E->getArg(0)); |
| 16169 | Value *Op1 = EmitScalarExpr(E->getArg(1)); |
| 16170 | if (getTarget().getTriple().isPPC64()) { |
| 16171 | Function *F = |
| 16172 | CGM.getIntrinsic(Intrinsic::ppc_cmpb, {Int64Ty, Int64Ty, Int64Ty}); |
| 16173 | return Builder.CreateCall(F, {Op0, Op1}, "cmpb"); |
| 16174 | } |
| 16175 | // For 32 bit, emit the code as below: |
| 16176 | // %conv = trunc i64 %a to i32 |
| 16177 | // %conv1 = trunc i64 %b to i32 |
| 16178 | // %shr = lshr i64 %a, 32 |
| 16179 | // %conv2 = trunc i64 %shr to i32 |
| 16180 | // %shr3 = lshr i64 %b, 32 |
| 16181 | // %conv4 = trunc i64 %shr3 to i32 |
| 16182 | // %0 = tail call i32 @llvm.ppc.cmpb32(i32 %conv, i32 %conv1) |
| 16183 | // %conv5 = zext i32 %0 to i64 |
| 16184 | // %1 = tail call i32 @llvm.ppc.cmpb32(i32 %conv2, i32 %conv4) |
| 16185 | // %conv614 = zext i32 %1 to i64 |
| 16186 | // %shl = shl nuw i64 %conv614, 32 |
| 16187 | // %or = or i64 %shl, %conv5 |
| 16188 | // ret i64 %or |
| 16189 | Function *F = |
| 16190 | CGM.getIntrinsic(Intrinsic::ppc_cmpb, {Int32Ty, Int32Ty, Int32Ty}); |
| 16191 | Value *ArgOneLo = Builder.CreateTrunc(Op0, Int32Ty); |
| 16192 | Value *ArgTwoLo = Builder.CreateTrunc(Op1, Int32Ty); |
| 16193 | Constant *ShiftAmt = ConstantInt::get(Int64Ty, 32); |
| 16194 | Value *ArgOneHi = |
| 16195 | Builder.CreateTrunc(Builder.CreateLShr(Op0, ShiftAmt), Int32Ty); |
| 16196 | Value *ArgTwoHi = |
| 16197 | Builder.CreateTrunc(Builder.CreateLShr(Op1, ShiftAmt), Int32Ty); |
| 16198 | Value *ResLo = Builder.CreateZExt( |
| 16199 | Builder.CreateCall(F, {ArgOneLo, ArgTwoLo}, "cmpb"), Int64Ty); |
| 16200 | Value *ResHiShift = Builder.CreateZExt( |
| 16201 | Builder.CreateCall(F, {ArgOneHi, ArgTwoHi}, "cmpb"), Int64Ty); |
| 16202 | Value *ResHi = Builder.CreateShl(ResHiShift, ShiftAmt); |
| 16203 | return Builder.CreateOr(ResLo, ResHi); |
| 16204 | } |
| 16205 | // Copy sign |
| 16206 | case PPC::BI__builtin_vsx_xvcpsgnsp: |
| 16207 | case PPC::BI__builtin_vsx_xvcpsgndp: { |
| 16208 | llvm::Type *ResultType = ConvertType(E->getType()); |
| 16209 | Value *X = EmitScalarExpr(E->getArg(0)); |
| 16210 | Value *Y = EmitScalarExpr(E->getArg(1)); |
| 16211 | ID = Intrinsic::copysign; |
| 16212 | llvm::Function *F = CGM.getIntrinsic(ID, ResultType); |
| 16213 | return Builder.CreateCall(F, {X, Y}); |
| 16214 | } |
| 16215 | // Rounding/truncation |
| 16216 | case PPC::BI__builtin_vsx_xvrspip: |
| 16217 | case PPC::BI__builtin_vsx_xvrdpip: |
| 16218 | case PPC::BI__builtin_vsx_xvrdpim: |
| 16219 | case PPC::BI__builtin_vsx_xvrspim: |
| 16220 | case PPC::BI__builtin_vsx_xvrdpi: |
| 16221 | case PPC::BI__builtin_vsx_xvrspi: |
| 16222 | case PPC::BI__builtin_vsx_xvrdpic: |
| 16223 | case PPC::BI__builtin_vsx_xvrspic: |
| 16224 | case PPC::BI__builtin_vsx_xvrdpiz: |
| 16225 | case PPC::BI__builtin_vsx_xvrspiz: { |
| 16226 | llvm::Type *ResultType = ConvertType(E->getType()); |
| 16227 | Value *X = EmitScalarExpr(E->getArg(0)); |
| 16228 | if (BuiltinID == PPC::BI__builtin_vsx_xvrdpim || |
| 16229 | BuiltinID == PPC::BI__builtin_vsx_xvrspim) |
| 16230 | ID = Builder.getIsFPConstrained() |
| 16231 | ? Intrinsic::experimental_constrained_floor |
| 16232 | : Intrinsic::floor; |
| 16233 | else if (BuiltinID == PPC::BI__builtin_vsx_xvrdpi || |
| 16234 | BuiltinID == PPC::BI__builtin_vsx_xvrspi) |
| 16235 | ID = Builder.getIsFPConstrained() |
| 16236 | ? Intrinsic::experimental_constrained_round |
| 16237 | : Intrinsic::round; |
| 16238 | else if (BuiltinID == PPC::BI__builtin_vsx_xvrdpic || |
| 16239 | BuiltinID == PPC::BI__builtin_vsx_xvrspic) |
| 16240 | ID = Builder.getIsFPConstrained() |
| 16241 | ? Intrinsic::experimental_constrained_rint |
| 16242 | : Intrinsic::rint; |
| 16243 | else if (BuiltinID == PPC::BI__builtin_vsx_xvrdpip || |
| 16244 | BuiltinID == PPC::BI__builtin_vsx_xvrspip) |
| 16245 | ID = Builder.getIsFPConstrained() |
| 16246 | ? Intrinsic::experimental_constrained_ceil |
| 16247 | : Intrinsic::ceil; |
| 16248 | else if (BuiltinID == PPC::BI__builtin_vsx_xvrdpiz || |
| 16249 | BuiltinID == PPC::BI__builtin_vsx_xvrspiz) |
| 16250 | ID = Builder.getIsFPConstrained() |
| 16251 | ? Intrinsic::experimental_constrained_trunc |
| 16252 | : Intrinsic::trunc; |
| 16253 | llvm::Function *F = CGM.getIntrinsic(ID, ResultType); |
| 16254 | return Builder.getIsFPConstrained() ? Builder.CreateConstrainedFPCall(F, X) |
| 16255 | : Builder.CreateCall(F, X); |
| 16256 | } |
| 16257 | |
| 16258 | // Absolute value |
| 16259 | case PPC::BI__builtin_vsx_xvabsdp: |
| 16260 | case PPC::BI__builtin_vsx_xvabssp: { |
| 16261 | llvm::Type *ResultType = ConvertType(E->getType()); |
| 16262 | Value *X = EmitScalarExpr(E->getArg(0)); |
| 16263 | llvm::Function *F = CGM.getIntrinsic(Intrinsic::fabs, ResultType); |
| 16264 | return Builder.CreateCall(F, X); |
| 16265 | } |
| 16266 | |
| 16267 | // Fastmath by default |
| 16268 | case PPC::BI__builtin_ppc_recipdivf: |
| 16269 | case PPC::BI__builtin_ppc_recipdivd: |
| 16270 | case PPC::BI__builtin_ppc_rsqrtf: |
| 16271 | case PPC::BI__builtin_ppc_rsqrtd: { |
| 16272 | FastMathFlags FMF = Builder.getFastMathFlags(); |
| 16273 | Builder.getFastMathFlags().setFast(); |
| 16274 | llvm::Type *ResultType = ConvertType(E->getType()); |
| 16275 | Value *X = EmitScalarExpr(E->getArg(0)); |
| 16276 | |
| 16277 | if (BuiltinID == PPC::BI__builtin_ppc_recipdivf || |
| 16278 | BuiltinID == PPC::BI__builtin_ppc_recipdivd) { |
| 16279 | Value *Y = EmitScalarExpr(E->getArg(1)); |
| 16280 | Value *FDiv = Builder.CreateFDiv(X, Y, "recipdiv"); |
| 16281 | Builder.getFastMathFlags() &= (FMF); |
| 16282 | return FDiv; |
| 16283 | } |
| 16284 | auto *One = ConstantFP::get(ResultType, 1.0); |
| 16285 | llvm::Function *F = CGM.getIntrinsic(Intrinsic::sqrt, ResultType); |
| 16286 | Value *FDiv = Builder.CreateFDiv(One, Builder.CreateCall(F, X), "rsqrt"); |
| 16287 | Builder.getFastMathFlags() &= (FMF); |
| 16288 | return FDiv; |
| 16289 | } |
| 16290 | case PPC::BI__builtin_ppc_alignx: { |
| 16291 | Value *Op0 = EmitScalarExpr(E->getArg(0)); |
| 16292 | Value *Op1 = EmitScalarExpr(E->getArg(1)); |
| 16293 | ConstantInt *AlignmentCI = cast<ConstantInt>(Op0); |
| 16294 | if (AlignmentCI->getValue().ugt(llvm::Value::MaximumAlignment)) |
| 16295 | AlignmentCI = ConstantInt::get(AlignmentCI->getType(), |
| 16296 | llvm::Value::MaximumAlignment); |
| 16297 | |
| 16298 | emitAlignmentAssumption(Op1, E->getArg(1), |
| 16299 | /*The expr loc is sufficient.*/ SourceLocation(), |
| 16300 | AlignmentCI, nullptr); |
| 16301 | return Op1; |
| 16302 | } |
| 16303 | case PPC::BI__builtin_ppc_rdlam: { |
| 16304 | Value *Op0 = EmitScalarExpr(E->getArg(0)); |
| 16305 | Value *Op1 = EmitScalarExpr(E->getArg(1)); |
| 16306 | Value *Op2 = EmitScalarExpr(E->getArg(2)); |
| 16307 | llvm::Type *Ty = Op0->getType(); |
| 16308 | Value *ShiftAmt = Builder.CreateIntCast(Op1, Ty, false); |
| 16309 | Function *F = CGM.getIntrinsic(Intrinsic::fshl, Ty); |
| 16310 | Value *Rotate = Builder.CreateCall(F, {Op0, Op0, ShiftAmt}); |
| 16311 | return Builder.CreateAnd(Rotate, Op2); |
| 16312 | } |
| 16313 | case PPC::BI__builtin_ppc_load2r: { |
| 16314 | Function *F = CGM.getIntrinsic(Intrinsic::ppc_load2r); |
| 16315 | Value *Op0 = Builder.CreateBitCast(EmitScalarExpr(E->getArg(0)), Int8PtrTy); |
| 16316 | Value *LoadIntrinsic = Builder.CreateCall(F, {Op0}); |
| 16317 | return Builder.CreateTrunc(LoadIntrinsic, Int16Ty); |
| 16318 | } |
| 16319 | // FMA variations |
| 16320 | case PPC::BI__builtin_ppc_fnmsub: |
| 16321 | case PPC::BI__builtin_ppc_fnmsubs: |
| 16322 | case PPC::BI__builtin_vsx_xvmaddadp: |
| 16323 | case PPC::BI__builtin_vsx_xvmaddasp: |
| 16324 | case PPC::BI__builtin_vsx_xvnmaddadp: |
| 16325 | case PPC::BI__builtin_vsx_xvnmaddasp: |
| 16326 | case PPC::BI__builtin_vsx_xvmsubadp: |
| 16327 | case PPC::BI__builtin_vsx_xvmsubasp: |
| 16328 | case PPC::BI__builtin_vsx_xvnmsubadp: |
| 16329 | case PPC::BI__builtin_vsx_xvnmsubasp: { |
| 16330 | llvm::Type *ResultType = ConvertType(E->getType()); |
| 16331 | Value *X = EmitScalarExpr(E->getArg(0)); |
| 16332 | Value *Y = EmitScalarExpr(E->getArg(1)); |
| 16333 | Value *Z = EmitScalarExpr(E->getArg(2)); |
| 16334 | llvm::Function *F; |
| 16335 | if (Builder.getIsFPConstrained()) |
| 16336 | F = CGM.getIntrinsic(Intrinsic::experimental_constrained_fma, ResultType); |
| 16337 | else |
| 16338 | F = CGM.getIntrinsic(Intrinsic::fma, ResultType); |
| 16339 | switch (BuiltinID) { |
| 16340 | case PPC::BI__builtin_vsx_xvmaddadp: |
| 16341 | case PPC::BI__builtin_vsx_xvmaddasp: |
| 16342 | if (Builder.getIsFPConstrained()) |
| 16343 | return Builder.CreateConstrainedFPCall(F, {X, Y, Z}); |
| 16344 | else |
| 16345 | return Builder.CreateCall(F, {X, Y, Z}); |
| 16346 | case PPC::BI__builtin_vsx_xvnmaddadp: |
| 16347 | case PPC::BI__builtin_vsx_xvnmaddasp: |
| 16348 | if (Builder.getIsFPConstrained()) |
| 16349 | return Builder.CreateFNeg( |
| 16350 | Builder.CreateConstrainedFPCall(F, {X, Y, Z}), "neg"); |
| 16351 | else |
| 16352 | return Builder.CreateFNeg(Builder.CreateCall(F, {X, Y, Z}), "neg"); |
| 16353 | case PPC::BI__builtin_vsx_xvmsubadp: |
| 16354 | case PPC::BI__builtin_vsx_xvmsubasp: |
| 16355 | if (Builder.getIsFPConstrained()) |
| 16356 | return Builder.CreateConstrainedFPCall( |
| 16357 | F, {X, Y, Builder.CreateFNeg(Z, "neg")}); |
| 16358 | else |
| 16359 | return Builder.CreateCall(F, {X, Y, Builder.CreateFNeg(Z, "neg")}); |
| 16360 | case PPC::BI__builtin_ppc_fnmsub: |
| 16361 | case PPC::BI__builtin_ppc_fnmsubs: |
| 16362 | case PPC::BI__builtin_vsx_xvnmsubadp: |
| 16363 | case PPC::BI__builtin_vsx_xvnmsubasp: |
| 16364 | if (Builder.getIsFPConstrained()) |
| 16365 | return Builder.CreateFNeg( |
| 16366 | Builder.CreateConstrainedFPCall( |
| 16367 | F, {X, Y, Builder.CreateFNeg(Z, "neg")}), |
| 16368 | "neg"); |
| 16369 | else |
| 16370 | return Builder.CreateCall( |
| 16371 | CGM.getIntrinsic(Intrinsic::ppc_fnmsub, ResultType), {X, Y, Z}); |
| 16372 | } |
| 16373 | llvm_unreachable("Unknown FMA operation")::llvm::llvm_unreachable_internal("Unknown FMA operation", "clang/lib/CodeGen/CGBuiltin.cpp" , 16373); |
| 16374 | return nullptr; // Suppress no-return warning |
| 16375 | } |
| 16376 | |
| 16377 | case PPC::BI__builtin_vsx_insertword: { |
| 16378 | Value *Op0 = EmitScalarExpr(E->getArg(0)); |
| 16379 | Value *Op1 = EmitScalarExpr(E->getArg(1)); |
| 16380 | Value *Op2 = EmitScalarExpr(E->getArg(2)); |
| 16381 | llvm::Function *F = CGM.getIntrinsic(Intrinsic::ppc_vsx_xxinsertw); |
| 16382 | |
| 16383 | // Third argument is a compile time constant int. It must be clamped to |
| 16384 | // to the range [0, 12]. |
| 16385 | ConstantInt *ArgCI = dyn_cast<ConstantInt>(Op2); |
| 16386 | assert(ArgCI &&(static_cast <bool> (ArgCI && "Third arg to xxinsertw intrinsic must be constant integer" ) ? void (0) : __assert_fail ("ArgCI && \"Third arg to xxinsertw intrinsic must be constant integer\"" , "clang/lib/CodeGen/CGBuiltin.cpp", 16387, __extension__ __PRETTY_FUNCTION__ )) |
| 16387 | "Third arg to xxinsertw intrinsic must be constant integer")(static_cast <bool> (ArgCI && "Third arg to xxinsertw intrinsic must be constant integer" ) ? void (0) : __assert_fail ("ArgCI && \"Third arg to xxinsertw intrinsic must be constant integer\"" , "clang/lib/CodeGen/CGBuiltin.cpp", 16387, __extension__ __PRETTY_FUNCTION__ )); |
| 16388 | const int64_t MaxIndex = 12; |
| 16389 | int64_t Index = std::clamp(ArgCI->getSExtValue(), (int64_t)0, MaxIndex); |
| 16390 | |
| 16391 | // The builtin semantics don't exactly match the xxinsertw instructions |
| 16392 | // semantics (which ppc_vsx_xxinsertw follows). The builtin extracts the |
| 16393 | // word from the first argument, and inserts it in the second argument. The |
| 16394 | // instruction extracts the word from its second input register and inserts |
| 16395 | // it into its first input register, so swap the first and second arguments. |
| 16396 | std::swap(Op0, Op1); |
| 16397 | |
| 16398 | // Need to cast the second argument from a vector of unsigned int to a |
| 16399 | // vector of long long. |
| 16400 | Op1 = Builder.CreateBitCast(Op1, llvm::FixedVectorType::get(Int64Ty, 2)); |
| 16401 | |
| 16402 | if (getTarget().isLittleEndian()) { |
| 16403 | // Reverse the double words in the vector we will extract from. |
| 16404 | Op0 = Builder.CreateBitCast(Op0, llvm::FixedVectorType::get(Int64Ty, 2)); |
| 16405 | Op0 = Builder.CreateShuffleVector(Op0, Op0, ArrayRef<int>{1, 0}); |
| 16406 | |
| 16407 | // Reverse the index. |
| 16408 | Index = MaxIndex - Index; |
| 16409 | } |
| 16410 | |
| 16411 | // Intrinsic expects the first arg to be a vector of int. |
| 16412 | Op0 = Builder.CreateBitCast(Op0, llvm::FixedVectorType::get(Int32Ty, 4)); |
| 16413 | Op2 = ConstantInt::getSigned(Int32Ty, Index); |
| 16414 | return Builder.CreateCall(F, {Op0, Op1, Op2}); |
| 16415 | } |
| 16416 | |
| 16417 | case PPC::BI__builtin_vsx_extractuword: { |
| 16418 | Value *Op0 = EmitScalarExpr(E->getArg(0)); |
| 16419 | Value *Op1 = EmitScalarExpr(E->getArg(1)); |
| 16420 | llvm::Function *F = CGM.getIntrinsic(Intrinsic::ppc_vsx_xxextractuw); |
| 16421 | |
| 16422 | // Intrinsic expects the first argument to be a vector of doublewords. |
| 16423 | Op0 = Builder.CreateBitCast(Op0, llvm::FixedVectorType::get(Int64Ty, 2)); |
| 16424 | |
| 16425 | // The second argument is a compile time constant int that needs to |
| 16426 | // be clamped to the range [0, 12]. |
| 16427 | ConstantInt *ArgCI = dyn_cast<ConstantInt>(Op1); |
| 16428 | assert(ArgCI &&(static_cast <bool> (ArgCI && "Second Arg to xxextractuw intrinsic must be a constant integer!" ) ? void (0) : __assert_fail ("ArgCI && \"Second Arg to xxextractuw intrinsic must be a constant integer!\"" , "clang/lib/CodeGen/CGBuiltin.cpp", 16429, __extension__ __PRETTY_FUNCTION__ )) |
| 16429 | "Second Arg to xxextractuw intrinsic must be a constant integer!")(static_cast <bool> (ArgCI && "Second Arg to xxextractuw intrinsic must be a constant integer!" ) ? void (0) : __assert_fail ("ArgCI && \"Second Arg to xxextractuw intrinsic must be a constant integer!\"" , "clang/lib/CodeGen/CGBuiltin.cpp", 16429, __extension__ __PRETTY_FUNCTION__ )); |
| 16430 | const int64_t MaxIndex = 12; |
| 16431 | int64_t Index = std::clamp(ArgCI->getSExtValue(), (int64_t)0, MaxIndex); |
| 16432 | |
| 16433 | if (getTarget().isLittleEndian()) { |
| 16434 | // Reverse the index. |
| 16435 | Index = MaxIndex - Index; |
| 16436 | Op1 = ConstantInt::getSigned(Int32Ty, Index); |
| 16437 | |
| 16438 | // Emit the call, then reverse the double words of the results vector. |
| 16439 | Value *Call = Builder.CreateCall(F, {Op0, Op1}); |
| 16440 | |
| 16441 | Value *ShuffleCall = |
| 16442 | Builder.CreateShuffleVector(Call, Call, ArrayRef<int>{1, 0}); |
| 16443 | return ShuffleCall; |
| 16444 | } else { |
| 16445 | Op1 = ConstantInt::getSigned(Int32Ty, Index); |
| 16446 | return Builder.CreateCall(F, {Op0, Op1}); |
| 16447 | } |
| 16448 | } |
| 16449 | |
| 16450 | case PPC::BI__builtin_vsx_xxpermdi: { |
| 16451 | Value *Op0 = EmitScalarExpr(E->getArg(0)); |
| 16452 | Value *Op1 = EmitScalarExpr(E->getArg(1)); |
| 16453 | Value *Op2 = EmitScalarExpr(E->getArg(2)); |
| 16454 | ConstantInt *ArgCI = dyn_cast<ConstantInt>(Op2); |
| 16455 | assert(ArgCI && "Third arg must be constant integer!")(static_cast <bool> (ArgCI && "Third arg must be constant integer!" ) ? void (0) : __assert_fail ("ArgCI && \"Third arg must be constant integer!\"" , "clang/lib/CodeGen/CGBuiltin.cpp", 16455, __extension__ __PRETTY_FUNCTION__ )); |
| 16456 | |
| 16457 | unsigned Index = ArgCI->getZExtValue(); |
| 16458 | Op0 = Builder.CreateBitCast(Op0, llvm::FixedVectorType::get(Int64Ty, 2)); |
| 16459 | Op1 = Builder.CreateBitCast(Op1, llvm::FixedVectorType::get(Int64Ty, 2)); |
| 16460 | |
| 16461 | // Account for endianness by treating this as just a shuffle. So we use the |
| 16462 | // same indices for both LE and BE in order to produce expected results in |
| 16463 | // both cases. |
| 16464 | int ElemIdx0 = (Index & 2) >> 1; |
| 16465 | int ElemIdx1 = 2 + (Index & 1); |
| 16466 | |
| 16467 | int ShuffleElts[2] = {ElemIdx0, ElemIdx1}; |
| 16468 | Value *ShuffleCall = Builder.CreateShuffleVector(Op0, Op1, ShuffleElts); |
| 16469 | QualType BIRetType = E->getType(); |
| 16470 | auto RetTy = ConvertType(BIRetType); |
| 16471 | return Builder.CreateBitCast(ShuffleCall, RetTy); |
| 16472 | } |
| 16473 | |
| 16474 | case PPC::BI__builtin_vsx_xxsldwi: { |
| 16475 | Value *Op0 = EmitScalarExpr(E->getArg(0)); |
| 16476 | Value *Op1 = EmitScalarExpr(E->getArg(1)); |
| 16477 | Value *Op2 = EmitScalarExpr(E->getArg(2)); |
| 16478 | ConstantInt *ArgCI = dyn_cast<ConstantInt>(Op2); |
| 16479 | assert(ArgCI && "Third argument must be a compile time constant")(static_cast <bool> (ArgCI && "Third argument must be a compile time constant" ) ? void (0) : __assert_fail ("ArgCI && \"Third argument must be a compile time constant\"" , "clang/lib/CodeGen/CGBuiltin.cpp", 16479, __extension__ __PRETTY_FUNCTION__ )); |
| 16480 | unsigned Index = ArgCI->getZExtValue() & 0x3; |
| 16481 | Op0 = Builder.CreateBitCast(Op0, llvm::FixedVectorType::get(Int32Ty, 4)); |
| 16482 | Op1 = Builder.CreateBitCast(Op1, llvm::FixedVectorType::get(Int32Ty, 4)); |
| 16483 | |
| 16484 | // Create a shuffle mask |
| 16485 | int ElemIdx0; |
| 16486 | int ElemIdx1; |
| 16487 | int ElemIdx2; |
| 16488 | int ElemIdx3; |
| 16489 | if (getTarget().isLittleEndian()) { |
| 16490 | // Little endian element N comes from element 8+N-Index of the |
| 16491 | // concatenated wide vector (of course, using modulo arithmetic on |
| 16492 | // the total number of elements). |
| 16493 | ElemIdx0 = (8 - Index) % 8; |
| 16494 | ElemIdx1 = (9 - Index) % 8; |
| 16495 | ElemIdx2 = (10 - Index) % 8; |
| 16496 | ElemIdx3 = (11 - Index) % 8; |
| 16497 | } else { |
| 16498 | // Big endian ElemIdx<N> = Index + N |
| 16499 | ElemIdx0 = Index; |
| 16500 | ElemIdx1 = Index + 1; |
| 16501 | ElemIdx2 = Index + 2; |
| 16502 | ElemIdx3 = Index + 3; |
| 16503 | } |
| 16504 | |
| 16505 | int ShuffleElts[4] = {ElemIdx0, ElemIdx1, ElemIdx2, ElemIdx3}; |
| 16506 | Value *ShuffleCall = Builder.CreateShuffleVector(Op0, Op1, ShuffleElts); |
| 16507 | QualType BIRetType = E->getType(); |
| 16508 | auto RetTy = ConvertType(BIRetType); |
| 16509 | return Builder.CreateBitCast(ShuffleCall, RetTy); |
| 16510 | } |
| 16511 | |
| 16512 | case PPC::BI__builtin_pack_vector_int128: { |
| 16513 | Value *Op0 = EmitScalarExpr(E->getArg(0)); |
| 16514 | Value *Op1 = EmitScalarExpr(E->getArg(1)); |
| 16515 | bool isLittleEndian = getTarget().isLittleEndian(); |
| 16516 | Value *PoisonValue = |
| 16517 | llvm::PoisonValue::get(llvm::FixedVectorType::get(Op0->getType(), 2)); |
| 16518 | Value *Res = Builder.CreateInsertElement( |
| 16519 | PoisonValue, Op0, (uint64_t)(isLittleEndian ? 1 : 0)); |
| 16520 | Res = Builder.CreateInsertElement(Res, Op1, |
| 16521 | (uint64_t)(isLittleEndian ? 0 : 1)); |
| 16522 | return Builder.CreateBitCast(Res, ConvertType(E->getType())); |
| 16523 | } |
| 16524 | |
| 16525 | case PPC::BI__builtin_unpack_vector_int128: { |
| 16526 | Value *Op0 = EmitScalarExpr(E->getArg(0)); |
| 16527 | Value *Op1 = EmitScalarExpr(E->getArg(1)); |
| 16528 | ConstantInt *Index = cast<ConstantInt>(Op1); |
| 16529 | Value *Unpacked = Builder.CreateBitCast( |
| 16530 | Op0, llvm::FixedVectorType::get(ConvertType(E->getType()), 2)); |
| 16531 | |
| 16532 | if (getTarget().isLittleEndian()) |
| 16533 | Index = ConstantInt::get(Index->getType(), 1 - Index->getZExtValue()); |
| 16534 | |
| 16535 | return Builder.CreateExtractElement(Unpacked, Index); |
| 16536 | } |
| 16537 | |
| 16538 | case PPC::BI__builtin_ppc_sthcx: { |
| 16539 | llvm::Function *F = CGM.getIntrinsic(Intrinsic::ppc_sthcx); |
| 16540 | Value *Op0 = Builder.CreateBitCast(EmitScalarExpr(E->getArg(0)), Int8PtrTy); |
| 16541 | Value *Op1 = Builder.CreateSExt(EmitScalarExpr(E->getArg(1)), Int32Ty); |
| 16542 | return Builder.CreateCall(F, {Op0, Op1}); |
| 16543 | } |
| 16544 | |
| 16545 | // The PPC MMA builtins take a pointer to a __vector_quad as an argument. |
| 16546 | // Some of the MMA instructions accumulate their result into an existing |
| 16547 | // accumulator whereas the others generate a new accumulator. So we need to |
| 16548 | // use custom code generation to expand a builtin call with a pointer to a |
| 16549 | // load (if the corresponding instruction accumulates its result) followed by |
| 16550 | // the call to the intrinsic and a store of the result. |
| 16551 | #define CUSTOM_BUILTIN(Name, Intr, Types, Accumulate, Feature) \ |
| 16552 | case PPC::BI__builtin_##Name: |
| 16553 | #include "clang/Basic/BuiltinsPPC.def" |
| 16554 | { |
| 16555 | SmallVector<Value *, 4> Ops; |
| 16556 | for (unsigned i = 0, e = E->getNumArgs(); i != e; i++) |
| 16557 | if (E->getArg(i)->getType()->isArrayType()) |
| 16558 | Ops.push_back(EmitArrayToPointerDecay(E->getArg(i)).getPointer()); |
| 16559 | else |
| 16560 | Ops.push_back(EmitScalarExpr(E->getArg(i))); |
| 16561 | // The first argument of these two builtins is a pointer used to store their |
| 16562 | // result. However, the llvm intrinsics return their result in multiple |
| 16563 | // return values. So, here we emit code extracting these values from the |
| 16564 | // intrinsic results and storing them using that pointer. |
| 16565 | if (BuiltinID == PPC::BI__builtin_mma_disassemble_acc || |
| 16566 | BuiltinID == PPC::BI__builtin_vsx_disassemble_pair || |
| 16567 | BuiltinID == PPC::BI__builtin_mma_disassemble_pair) { |
| 16568 | unsigned NumVecs = 2; |
| 16569 | auto Intrinsic = Intrinsic::ppc_vsx_disassemble_pair; |
| 16570 | if (BuiltinID == PPC::BI__builtin_mma_disassemble_acc) { |
| 16571 | NumVecs = 4; |
| 16572 | Intrinsic = Intrinsic::ppc_mma_disassemble_acc; |
| 16573 | } |
| 16574 | llvm::Function *F = CGM.getIntrinsic(Intrinsic); |
| 16575 | Address Addr = EmitPointerWithAlignment(E->getArg(1)); |
| 16576 | Value *Vec = Builder.CreateLoad(Addr); |
| 16577 | Value *Call = Builder.CreateCall(F, {Vec}); |
| 16578 | llvm::Type *VTy = llvm::FixedVectorType::get(Int8Ty, 16); |
| 16579 | Value *Ptr = Builder.CreateBitCast(Ops[0], VTy->getPointerTo()); |
| 16580 | for (unsigned i=0; i<NumVecs; i++) { |
| 16581 | Value *Vec = Builder.CreateExtractValue(Call, i); |
| 16582 | llvm::ConstantInt* Index = llvm::ConstantInt::get(IntTy, i); |
| 16583 | Value *GEP = Builder.CreateInBoundsGEP(VTy, Ptr, Index); |
| 16584 | Builder.CreateAlignedStore(Vec, GEP, MaybeAlign(16)); |
| 16585 | } |
| 16586 | return Call; |
| 16587 | } |
| 16588 | if (BuiltinID == PPC::BI__builtin_vsx_build_pair || |
| 16589 | BuiltinID == PPC::BI__builtin_mma_build_acc) { |
| 16590 | // Reverse the order of the operands for LE, so the |
| 16591 | // same builtin call can be used on both LE and BE |
| 16592 | // without the need for the programmer to swap operands. |
| 16593 | // The operands are reversed starting from the second argument, |
| 16594 | // the first operand is the pointer to the pair/accumulator |
| 16595 | // that is being built. |
| 16596 | if (getTarget().isLittleEndian()) |
| 16597 | std::reverse(Ops.begin() + 1, Ops.end()); |
| 16598 | } |
| 16599 | bool Accumulate; |
| 16600 | switch (BuiltinID) { |
| 16601 | #define CUSTOM_BUILTIN(Name, Intr, Types, Acc, Feature) \ |
| 16602 | case PPC::BI__builtin_##Name: \ |
| 16603 | ID = Intrinsic::ppc_##Intr; \ |
| 16604 | Accumulate = Acc; \ |
| 16605 | break; |
| 16606 | #include "clang/Basic/BuiltinsPPC.def" |
| 16607 | } |
| 16608 | if (BuiltinID == PPC::BI__builtin_vsx_lxvp || |
| 16609 | BuiltinID == PPC::BI__builtin_vsx_stxvp || |
| 16610 | BuiltinID == PPC::BI__builtin_mma_lxvp || |
| 16611 | BuiltinID == PPC::BI__builtin_mma_stxvp) { |
| 16612 | if (BuiltinID == PPC::BI__builtin_vsx_lxvp || |
| 16613 | BuiltinID == PPC::BI__builtin_mma_lxvp) { |
| 16614 | Ops[1] = Builder.CreateBitCast(Ops[1], Int8PtrTy); |
| 16615 | Ops[0] = Builder.CreateGEP(Int8Ty, Ops[1], Ops[0]); |
| 16616 | } else { |
| 16617 | Ops[2] = Builder.CreateBitCast(Ops[2], Int8PtrTy); |
| 16618 | Ops[1] = Builder.CreateGEP(Int8Ty, Ops[2], Ops[1]); |
| 16619 | } |
| 16620 | Ops.pop_back(); |
| 16621 | llvm::Function *F = CGM.getIntrinsic(ID); |
| 16622 | return Builder.CreateCall(F, Ops, ""); |
| 16623 | } |
| 16624 | SmallVector<Value*, 4> CallOps; |
| 16625 | if (Accumulate) { |
| 16626 | Address Addr = EmitPointerWithAlignment(E->getArg(0)); |
| 16627 | Value *Acc = Builder.CreateLoad(Addr); |
| 16628 | CallOps.push_back(Acc); |
| 16629 | } |
| 16630 | for (unsigned i=1; i<Ops.size(); i++) |
| 16631 | CallOps.push_back(Ops[i]); |
| 16632 | llvm::Function *F = CGM.getIntrinsic(ID); |
| 16633 | Value *Call = Builder.CreateCall(F, CallOps); |
| 16634 | return Builder.CreateAlignedStore(Call, Ops[0], MaybeAlign(64)); |
| 16635 | } |
| 16636 | |
| 16637 | case PPC::BI__builtin_ppc_compare_and_swap: |
| 16638 | case PPC::BI__builtin_ppc_compare_and_swaplp: { |
| 16639 | Address Addr = EmitPointerWithAlignment(E->getArg(0)); |
| 16640 | Address OldValAddr = EmitPointerWithAlignment(E->getArg(1)); |
| 16641 | Value *OldVal = Builder.CreateLoad(OldValAddr); |
| 16642 | QualType AtomicTy = E->getArg(0)->getType()->getPointeeType(); |
| 16643 | LValue LV = MakeAddrLValue(Addr, AtomicTy); |
| 16644 | Value *Op2 = EmitScalarExpr(E->getArg(2)); |
| 16645 | auto Pair = EmitAtomicCompareExchange( |
| 16646 | LV, RValue::get(OldVal), RValue::get(Op2), E->getExprLoc(), |
| 16647 | llvm::AtomicOrdering::Monotonic, llvm::AtomicOrdering::Monotonic, true); |
| 16648 | // Unlike c11's atomic_compare_exchange, according to |
| 16649 | // https://www.ibm.com/docs/en/xl-c-and-cpp-aix/16.1?topic=functions-compare-swap-compare-swaplp |
| 16650 | // > In either case, the contents of the memory location specified by addr |
| 16651 | // > are copied into the memory location specified by old_val_addr. |
| 16652 | // But it hasn't specified storing to OldValAddr is atomic or not and |
| 16653 | // which order to use. Now following XL's codegen, treat it as a normal |
| 16654 | // store. |
| 16655 | Value *LoadedVal = Pair.first.getScalarVal(); |
| 16656 | Builder.CreateStore(LoadedVal, OldValAddr); |
| 16657 | return Builder.CreateZExt(Pair.second, Builder.getInt32Ty()); |
| 16658 | } |
| 16659 | case PPC::BI__builtin_ppc_fetch_and_add: |
| 16660 | case PPC::BI__builtin_ppc_fetch_and_addlp: { |
| 16661 | return MakeBinaryAtomicValue(*this, AtomicRMWInst::Add, E, |
| 16662 | llvm::AtomicOrdering::Monotonic); |
| 16663 | } |
| 16664 | case PPC::BI__builtin_ppc_fetch_and_and: |
| 16665 | case PPC::BI__builtin_ppc_fetch_and_andlp: { |
| 16666 | return MakeBinaryAtomicValue(*this, AtomicRMWInst::And, E, |
| 16667 | llvm::AtomicOrdering::Monotonic); |
| 16668 | } |
| 16669 | |
| 16670 | case PPC::BI__builtin_ppc_fetch_and_or: |
| 16671 | case PPC::BI__builtin_ppc_fetch_and_orlp: { |
| 16672 | return MakeBinaryAtomicValue(*this, AtomicRMWInst::Or, E, |
| 16673 | llvm::AtomicOrdering::Monotonic); |
| 16674 | } |
| 16675 | case PPC::BI__builtin_ppc_fetch_and_swap: |
| 16676 | case PPC::BI__builtin_ppc_fetch_and_swaplp: { |
| 16677 | return MakeBinaryAtomicValue(*this, AtomicRMWInst::Xchg, E, |
| 16678 | llvm::AtomicOrdering::Monotonic); |
| 16679 | } |
| 16680 | case PPC::BI__builtin_ppc_ldarx: |
| 16681 | case PPC::BI__builtin_ppc_lwarx: |
| 16682 | case PPC::BI__builtin_ppc_lharx: |
| 16683 | case PPC::BI__builtin_ppc_lbarx: |
| 16684 | return emitPPCLoadReserveIntrinsic(*this, BuiltinID, E); |
| 16685 | case PPC::BI__builtin_ppc_mfspr: { |
| 16686 | Value *Op0 = EmitScalarExpr(E->getArg(0)); |
| 16687 | llvm::Type *RetType = CGM.getDataLayout().getTypeSizeInBits(VoidPtrTy) == 32 |
| 16688 | ? Int32Ty |
| 16689 | : Int64Ty; |
| 16690 | Function *F = CGM.getIntrinsic(Intrinsic::ppc_mfspr, RetType); |
| 16691 | return Builder.CreateCall(F, {Op0}); |
| 16692 | } |
| 16693 | case PPC::BI__builtin_ppc_mtspr: { |
| 16694 | Value *Op0 = EmitScalarExpr(E->getArg(0)); |
| 16695 | Value *Op1 = EmitScalarExpr(E->getArg(1)); |
| 16696 | llvm::Type *RetType = CGM.getDataLayout().getTypeSizeInBits(VoidPtrTy) == 32 |
| 16697 | ? Int32Ty |
| 16698 | : Int64Ty; |
| 16699 | Function *F = CGM.getIntrinsic(Intrinsic::ppc_mtspr, RetType); |
| 16700 | return Builder.CreateCall(F, {Op0, Op1}); |
| 16701 | } |
| 16702 | case PPC::BI__builtin_ppc_popcntb: { |
| 16703 | Value *ArgValue = EmitScalarExpr(E->getArg(0)); |
| 16704 | llvm::Type *ArgType = ArgValue->getType(); |
| 16705 | Function *F = CGM.getIntrinsic(Intrinsic::ppc_popcntb, {ArgType, ArgType}); |
| 16706 | return Builder.CreateCall(F, {ArgValue}, "popcntb"); |
| 16707 | } |
| 16708 | case PPC::BI__builtin_ppc_mtfsf: { |
| 16709 | // The builtin takes a uint32 that needs to be cast to an |
| 16710 | // f64 to be passed to the intrinsic. |
| 16711 | Value *Op0 = EmitScalarExpr(E->getArg(0)); |
| 16712 | Value *Op1 = EmitScalarExpr(E->getArg(1)); |
| 16713 | Value *Cast = Builder.CreateUIToFP(Op1, DoubleTy); |
| 16714 | llvm::Function *F = CGM.getIntrinsic(Intrinsic::ppc_mtfsf); |
| 16715 | return Builder.CreateCall(F, {Op0, Cast}, ""); |
| 16716 | } |
| 16717 | |
| 16718 | case PPC::BI__builtin_ppc_swdiv_nochk: |
| 16719 | case PPC::BI__builtin_ppc_swdivs_nochk: { |
| 16720 | Value *Op0 = EmitScalarExpr(E->getArg(0)); |
| 16721 | Value *Op1 = EmitScalarExpr(E->getArg(1)); |
| 16722 | FastMathFlags FMF = Builder.getFastMathFlags(); |
| 16723 | Builder.getFastMathFlags().setFast(); |
| 16724 | Value *FDiv = Builder.CreateFDiv(Op0, Op1, "swdiv_nochk"); |
| 16725 | Builder.getFastMathFlags() &= (FMF); |
| 16726 | return FDiv; |
| 16727 | } |
| 16728 | case PPC::BI__builtin_ppc_fric: |
| 16729 | return RValue::get(emitUnaryMaybeConstrainedFPBuiltin( |
| 16730 | *this, E, Intrinsic::rint, |
| 16731 | Intrinsic::experimental_constrained_rint)) |
| 16732 | .getScalarVal(); |
| 16733 | case PPC::BI__builtin_ppc_frim: |
| 16734 | case PPC::BI__builtin_ppc_frims: |
| 16735 | return RValue::get(emitUnaryMaybeConstrainedFPBuiltin( |
| 16736 | *this, E, Intrinsic::floor, |
| 16737 | Intrinsic::experimental_constrained_floor)) |
| 16738 | .getScalarVal(); |
| 16739 | case PPC::BI__builtin_ppc_frin: |
| 16740 | case PPC::BI__builtin_ppc_frins: |
| 16741 | return RValue::get(emitUnaryMaybeConstrainedFPBuiltin( |
| 16742 | *this, E, Intrinsic::round, |
| 16743 | Intrinsic::experimental_constrained_round)) |
| 16744 | .getScalarVal(); |
| 16745 | case PPC::BI__builtin_ppc_frip: |
| 16746 | case PPC::BI__builtin_ppc_frips: |
| 16747 | return RValue::get(emitUnaryMaybeConstrainedFPBuiltin( |
| 16748 | *this, E, Intrinsic::ceil, |
| 16749 | Intrinsic::experimental_constrained_ceil)) |
| 16750 | .getScalarVal(); |
| 16751 | case PPC::BI__builtin_ppc_friz: |
| 16752 | case PPC::BI__builtin_ppc_frizs: |
| 16753 | return RValue::get(emitUnaryMaybeConstrainedFPBuiltin( |
| 16754 | *this, E, Intrinsic::trunc, |
| 16755 | Intrinsic::experimental_constrained_trunc)) |
| 16756 | .getScalarVal(); |
| 16757 | case PPC::BI__builtin_ppc_fsqrt: |
| 16758 | case PPC::BI__builtin_ppc_fsqrts: |
| 16759 | return RValue::get(emitUnaryMaybeConstrainedFPBuiltin( |
| 16760 | *this, E, Intrinsic::sqrt, |
| 16761 | Intrinsic::experimental_constrained_sqrt)) |
| 16762 | .getScalarVal(); |
| 16763 | case PPC::BI__builtin_ppc_test_data_class: { |
| 16764 | Value *Op0 = EmitScalarExpr(E->getArg(0)); |
| 16765 | Value *Op1 = EmitScalarExpr(E->getArg(1)); |
| 16766 | return Builder.CreateCall( |
| 16767 | CGM.getIntrinsic(Intrinsic::ppc_test_data_class, Op0->getType()), |
| 16768 | {Op0, Op1}, "test_data_class"); |
| 16769 | } |
| 16770 | case PPC::BI__builtin_ppc_maxfe: { |
| 16771 | Value *Op0 = EmitScalarExpr(E->getArg(0)); |
| 16772 | Value *Op1 = EmitScalarExpr(E->getArg(1)); |
| 16773 | Value *Op2 = EmitScalarExpr(E->getArg(2)); |
| 16774 | Value *Op3 = EmitScalarExpr(E->getArg(3)); |
| 16775 | return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::ppc_maxfe), |
| 16776 | {Op0, Op1, Op2, Op3}); |
| 16777 | } |
| 16778 | case PPC::BI__builtin_ppc_maxfl: { |
| 16779 | Value *Op0 = EmitScalarExpr(E->getArg(0)); |
| 16780 | Value *Op1 = EmitScalarExpr(E->getArg(1)); |
| 16781 | Value *Op2 = EmitScalarExpr(E->getArg(2)); |
| 16782 | Value *Op3 = EmitScalarExpr(E->getArg(3)); |
| 16783 | return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::ppc_maxfl), |
| 16784 | {Op0, Op1, Op2, Op3}); |
| 16785 | } |
| 16786 | case PPC::BI__builtin_ppc_maxfs: { |
| 16787 | Value *Op0 = EmitScalarExpr(E->getArg(0)); |
| 16788 | Value *Op1 = EmitScalarExpr(E->getArg(1)); |
| 16789 | Value *Op2 = EmitScalarExpr(E->getArg(2)); |
| 16790 | Value *Op3 = EmitScalarExpr(E->getArg(3)); |
| 16791 | return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::ppc_maxfs), |
| 16792 | {Op0, Op1, Op2, Op3}); |
| 16793 | } |
| 16794 | case PPC::BI__builtin_ppc_minfe: { |
| 16795 | Value *Op0 = EmitScalarExpr(E->getArg(0)); |
| 16796 | Value *Op1 = EmitScalarExpr(E->getArg(1)); |
| 16797 | Value *Op2 = EmitScalarExpr(E->getArg(2)); |
| 16798 | Value *Op3 = EmitScalarExpr(E->getArg(3)); |
| 16799 | return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::ppc_minfe), |
| 16800 | {Op0, Op1, Op2, Op3}); |
| 16801 | } |
| 16802 | case PPC::BI__builtin_ppc_minfl: { |
| 16803 | Value *Op0 = EmitScalarExpr(E->getArg(0)); |
| 16804 | Value *Op1 = EmitScalarExpr(E->getArg(1)); |
| 16805 | Value *Op2 = EmitScalarExpr(E->getArg(2)); |
| 16806 | Value *Op3 = EmitScalarExpr(E->getArg(3)); |
| 16807 | return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::ppc_minfl), |
| 16808 | {Op0, Op1, Op2, Op3}); |
| 16809 | } |
| 16810 | case PPC::BI__builtin_ppc_minfs: { |
| 16811 | Value *Op0 = EmitScalarExpr(E->getArg(0)); |
| 16812 | Value *Op1 = EmitScalarExpr(E->getArg(1)); |
| 16813 | Value *Op2 = EmitScalarExpr(E->getArg(2)); |
| 16814 | Value *Op3 = EmitScalarExpr(E->getArg(3)); |
| 16815 | return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::ppc_minfs), |
| 16816 | {Op0, Op1, Op2, Op3}); |
| 16817 | } |
| 16818 | case PPC::BI__builtin_ppc_swdiv: |
| 16819 | case PPC::BI__builtin_ppc_swdivs: { |
| 16820 | Value *Op0 = EmitScalarExpr(E->getArg(0)); |
| 16821 | Value *Op1 = EmitScalarExpr(E->getArg(1)); |
| 16822 | return Builder.CreateFDiv(Op0, Op1, "swdiv"); |
| 16823 | } |
| 16824 | } |
| 16825 | } |
| 16826 | |
| 16827 | namespace { |
| 16828 | // If \p E is not null pointer, insert address space cast to match return |
| 16829 | // type of \p E if necessary. |
| 16830 | Value *EmitAMDGPUDispatchPtr(CodeGenFunction &CGF, |
| 16831 | const CallExpr *E = nullptr) { |
| 16832 | auto *F = CGF.CGM.getIntrinsic(Intrinsic::amdgcn_dispatch_ptr); |
| 16833 | auto *Call = CGF.Builder.CreateCall(F); |
| 16834 | Call->addRetAttr( |
| 16835 | Attribute::getWithDereferenceableBytes(Call->getContext(), 64)); |
| 16836 | Call->addRetAttr(Attribute::getWithAlignment(Call->getContext(), Align(4))); |
| 16837 | if (!E) |
| 16838 | return Call; |
| 16839 | QualType BuiltinRetType = E->getType(); |
| 16840 | auto *RetTy = cast<llvm::PointerType>(CGF.ConvertType(BuiltinRetType)); |
| 16841 | if (RetTy == Call->getType()) |
| 16842 | return Call; |
| 16843 | return CGF.Builder.CreateAddrSpaceCast(Call, RetTy); |
| 16844 | } |
| 16845 | |
| 16846 | Value *EmitAMDGPUImplicitArgPtr(CodeGenFunction &CGF) { |
| 16847 | auto *F = CGF.CGM.getIntrinsic(Intrinsic::amdgcn_implicitarg_ptr); |
| 16848 | auto *Call = CGF.Builder.CreateCall(F); |
| 16849 | Call->addRetAttr( |
| 16850 | Attribute::getWithDereferenceableBytes(Call->getContext(), 256)); |
| 16851 | Call->addRetAttr(Attribute::getWithAlignment(Call->getContext(), Align(8))); |
| 16852 | return Call; |
| 16853 | } |
| 16854 | |
| 16855 | // \p Index is 0, 1, and 2 for x, y, and z dimension, respectively. |
| 16856 | Value *EmitAMDGPUWorkGroupSize(CodeGenFunction &CGF, unsigned Index) { |
| 16857 | bool IsCOV_5 = CGF.getTarget().getTargetOpts().CodeObjectVersion == |
| 16858 | clang::TargetOptions::COV_5; |
| 16859 | Constant *Offset; |
| 16860 | Value *DP; |
| 16861 | if (IsCOV_5) { |
| 16862 | // Indexing the implicit kernarg segment. |
| 16863 | Offset = llvm::ConstantInt::get(CGF.Int32Ty, 12 + Index * 2); |
| 16864 | DP = EmitAMDGPUImplicitArgPtr(CGF); |
| 16865 | } else { |
| 16866 | // Indexing the HSA kernel_dispatch_packet struct. |
| 16867 | Offset = llvm::ConstantInt::get(CGF.Int32Ty, 4 + Index * 2); |
| 16868 | DP = EmitAMDGPUDispatchPtr(CGF); |
| 16869 | } |
| 16870 | |
| 16871 | auto *GEP = CGF.Builder.CreateGEP(CGF.Int8Ty, DP, Offset); |
| 16872 | auto *DstTy = |
| 16873 | CGF.Int16Ty->getPointerTo(GEP->getType()->getPointerAddressSpace()); |
| 16874 | auto *Cast = CGF.Builder.CreateBitCast(GEP, DstTy); |
| 16875 | auto *LD = CGF.Builder.CreateLoad( |
| 16876 | Address(Cast, CGF.Int16Ty, CharUnits::fromQuantity(2))); |
| 16877 | llvm::MDBuilder MDHelper(CGF.getLLVMContext()); |
| 16878 | llvm::MDNode *RNode = MDHelper.createRange(APInt(16, 1), |
| 16879 | APInt(16, CGF.getTarget().getMaxOpenCLWorkGroupSize() + 1)); |
| 16880 | LD->setMetadata(llvm::LLVMContext::MD_range, RNode); |
| 16881 | LD->setMetadata(llvm::LLVMContext::MD_noundef, |
| 16882 | llvm::MDNode::get(CGF.getLLVMContext(), std::nullopt)); |
| 16883 | LD->setMetadata(llvm::LLVMContext::MD_invariant_load, |
| 16884 | llvm::MDNode::get(CGF.getLLVMContext(), std::nullopt)); |
| 16885 | return LD; |
| 16886 | } |
| 16887 | |
| 16888 | // \p Index is 0, 1, and 2 for x, y, and z dimension, respectively. |
| 16889 | Value *EmitAMDGPUGridSize(CodeGenFunction &CGF, unsigned Index) { |
| 16890 | const unsigned XOffset = 12; |
| 16891 | auto *DP = EmitAMDGPUDispatchPtr(CGF); |
| 16892 | // Indexing the HSA kernel_dispatch_packet struct. |
| 16893 | auto *Offset = llvm::ConstantInt::get(CGF.Int32Ty, XOffset + Index * 4); |
| 16894 | auto *GEP = CGF.Builder.CreateGEP(CGF.Int8Ty, DP, Offset); |
| 16895 | auto *DstTy = |
| 16896 | CGF.Int32Ty->getPointerTo(GEP->getType()->getPointerAddressSpace()); |
| 16897 | auto *Cast = CGF.Builder.CreateBitCast(GEP, DstTy); |
| 16898 | auto *LD = CGF.Builder.CreateLoad( |
| 16899 | Address(Cast, CGF.Int32Ty, CharUnits::fromQuantity(4))); |
| 16900 | LD->setMetadata(llvm::LLVMContext::MD_invariant_load, |
| 16901 | llvm::MDNode::get(CGF.getLLVMContext(), std::nullopt)); |
| 16902 | return LD; |
| 16903 | } |
| 16904 | } // namespace |
| 16905 | |
| 16906 | // For processing memory ordering and memory scope arguments of various |
| 16907 | // amdgcn builtins. |
| 16908 | // \p Order takes a C++11 comptabile memory-ordering specifier and converts |
| 16909 | // it into LLVM's memory ordering specifier using atomic C ABI, and writes |
| 16910 | // to \p AO. \p Scope takes a const char * and converts it into AMDGCN |
| 16911 | // specific SyncScopeID and writes it to \p SSID. |
| 16912 | void CodeGenFunction::ProcessOrderScopeAMDGCN(Value *Order, Value *Scope, |
| 16913 | llvm::AtomicOrdering &AO, |
| 16914 | llvm::SyncScope::ID &SSID) { |
| 16915 | int ord = cast<llvm::ConstantInt>(Order)->getZExtValue(); |
| 16916 | |
| 16917 | // Map C11/C++11 memory ordering to LLVM memory ordering |
| 16918 | assert(llvm::isValidAtomicOrderingCABI(ord))(static_cast <bool> (llvm::isValidAtomicOrderingCABI(ord )) ? void (0) : __assert_fail ("llvm::isValidAtomicOrderingCABI(ord)" , "clang/lib/CodeGen/CGBuiltin.cpp", 16918, __extension__ __PRETTY_FUNCTION__ )); |
| 16919 | switch (static_cast<llvm::AtomicOrderingCABI>(ord)) { |
| 16920 | case llvm::AtomicOrderingCABI::acquire: |
| 16921 | case llvm::AtomicOrderingCABI::consume: |
| 16922 | AO = llvm::AtomicOrdering::Acquire; |
| 16923 | break; |
| 16924 | case llvm::AtomicOrderingCABI::release: |
| 16925 | AO = llvm::AtomicOrdering::Release; |
| 16926 | break; |
| 16927 | case llvm::AtomicOrderingCABI::acq_rel: |
| 16928 | AO = llvm::AtomicOrdering::AcquireRelease; |
| 16929 | break; |
| 16930 | case llvm::AtomicOrderingCABI::seq_cst: |
| 16931 | AO = llvm::AtomicOrdering::SequentiallyConsistent; |
| 16932 | break; |
| 16933 | case llvm::AtomicOrderingCABI::relaxed: |
| 16934 | AO = llvm::AtomicOrdering::Monotonic; |
| 16935 | break; |
| 16936 | } |
| 16937 | |
| 16938 | StringRef scp; |
| 16939 | llvm::getConstantStringInfo(Scope, scp); |
| 16940 | SSID = getLLVMContext().getOrInsertSyncScopeID(scp); |
| 16941 | } |
| 16942 | |
| 16943 | Value *CodeGenFunction::EmitAMDGPUBuiltinExpr(unsigned BuiltinID, |
| 16944 | const CallExpr *E) { |
| 16945 | llvm::AtomicOrdering AO = llvm::AtomicOrdering::SequentiallyConsistent; |
| 16946 | llvm::SyncScope::ID SSID; |
| 16947 | switch (BuiltinID) { |
| 16948 | case AMDGPU::BI__builtin_amdgcn_div_scale: |
| 16949 | case AMDGPU::BI__builtin_amdgcn_div_scalef: { |
| 16950 | // Translate from the intrinsics's struct return to the builtin's out |
| 16951 | // argument. |
| 16952 | |
| 16953 | Address FlagOutPtr = EmitPointerWithAlignment(E->getArg(3)); |
| 16954 | |
| 16955 | llvm::Value *X = EmitScalarExpr(E->getArg(0)); |
| 16956 | llvm::Value *Y = EmitScalarExpr(E->getArg(1)); |
| 16957 | llvm::Value *Z = EmitScalarExpr(E->getArg(2)); |
| 16958 | |
| 16959 | llvm::Function *Callee = CGM.getIntrinsic(Intrinsic::amdgcn_div_scale, |
| 16960 | X->getType()); |
| 16961 | |
| 16962 | llvm::Value *Tmp = Builder.CreateCall(Callee, {X, Y, Z}); |
| 16963 | |
| 16964 | llvm::Value *Result = Builder.CreateExtractValue(Tmp, 0); |
| 16965 | llvm::Value *Flag = Builder.CreateExtractValue(Tmp, 1); |
| 16966 | |
| 16967 | llvm::Type *RealFlagType = FlagOutPtr.getElementType(); |
| 16968 | |
| 16969 | llvm::Value *FlagExt = Builder.CreateZExt(Flag, RealFlagType); |
| 16970 | Builder.CreateStore(FlagExt, FlagOutPtr); |
| 16971 | return Result; |
| 16972 | } |
| 16973 | case AMDGPU::BI__builtin_amdgcn_div_fmas: |
| 16974 | case AMDGPU::BI__builtin_amdgcn_div_fmasf: { |
| 16975 | llvm::Value *Src0 = EmitScalarExpr(E->getArg(0)); |
| 16976 | llvm::Value *Src1 = EmitScalarExpr(E->getArg(1)); |
| 16977 | llvm::Value *Src2 = EmitScalarExpr(E->getArg(2)); |
| 16978 | llvm::Value *Src3 = EmitScalarExpr(E->getArg(3)); |
| 16979 | |
| 16980 | llvm::Function *F = CGM.getIntrinsic(Intrinsic::amdgcn_div_fmas, |
| 16981 | Src0->getType()); |
| 16982 | llvm::Value *Src3ToBool = Builder.CreateIsNotNull(Src3); |
| 16983 | return Builder.CreateCall(F, {Src0, Src1, Src2, Src3ToBool}); |
| 16984 | } |
| 16985 | |
| 16986 | case AMDGPU::BI__builtin_amdgcn_ds_swizzle: |
| 16987 | return emitBinaryBuiltin(*this, E, Intrinsic::amdgcn_ds_swizzle); |
| 16988 | case AMDGPU::BI__builtin_amdgcn_mov_dpp8: |
| 16989 | return emitBinaryBuiltin(*this, E, Intrinsic::amdgcn_mov_dpp8); |
| 16990 | case AMDGPU::BI__builtin_amdgcn_mov_dpp: |
| 16991 | case AMDGPU::BI__builtin_amdgcn_update_dpp: { |
| 16992 | llvm::SmallVector<llvm::Value *, 6> Args; |
| 16993 | for (unsigned I = 0; I != E->getNumArgs(); ++I) |
| 16994 | Args.push_back(EmitScalarExpr(E->getArg(I))); |
| 16995 | assert(Args.size() == 5 || Args.size() == 6)(static_cast <bool> (Args.size() == 5 || Args.size() == 6) ? void (0) : __assert_fail ("Args.size() == 5 || Args.size() == 6" , "clang/lib/CodeGen/CGBuiltin.cpp", 16995, __extension__ __PRETTY_FUNCTION__ )); |
| 16996 | if (Args.size() == 5) |
| 16997 | Args.insert(Args.begin(), llvm::PoisonValue::get(Args[0]->getType())); |
| 16998 | Function *F = |
| 16999 | CGM.getIntrinsic(Intrinsic::amdgcn_update_dpp, Args[0]->getType()); |
| 17000 | return Builder.CreateCall(F, Args); |
| 17001 | } |
| 17002 | case AMDGPU::BI__builtin_amdgcn_div_fixup: |
| 17003 | case AMDGPU::BI__builtin_amdgcn_div_fixupf: |
| 17004 | case AMDGPU::BI__builtin_amdgcn_div_fixuph: |
| 17005 | return emitTernaryBuiltin(*this, E, Intrinsic::amdgcn_div_fixup); |
| 17006 | case AMDGPU::BI__builtin_amdgcn_trig_preop: |
| 17007 | case AMDGPU::BI__builtin_amdgcn_trig_preopf: |
| 17008 | return emitFPIntBuiltin(*this, E, Intrinsic::amdgcn_trig_preop); |
| 17009 | case AMDGPU::BI__builtin_amdgcn_rcp: |
| 17010 | case AMDGPU::BI__builtin_amdgcn_rcpf: |
| 17011 | case AMDGPU::BI__builtin_amdgcn_rcph: |
| 17012 | return emitUnaryBuiltin(*this, E, Intrinsic::amdgcn_rcp); |
| 17013 | case AMDGPU::BI__builtin_amdgcn_sqrt: |
| 17014 | case AMDGPU::BI__builtin_amdgcn_sqrtf: |
| 17015 | case AMDGPU::BI__builtin_amdgcn_sqrth: |
| 17016 | return emitUnaryBuiltin(*this, E, Intrinsic::amdgcn_sqrt); |
| 17017 | case AMDGPU::BI__builtin_amdgcn_rsq: |
| 17018 | case AMDGPU::BI__builtin_amdgcn_rsqf: |
| 17019 | case AMDGPU::BI__builtin_amdgcn_rsqh: |
| 17020 | return emitUnaryBuiltin(*this, E, Intrinsic::amdgcn_rsq); |
| 17021 | case AMDGPU::BI__builtin_amdgcn_rsq_clamp: |
| 17022 | case AMDGPU::BI__builtin_amdgcn_rsq_clampf: |
| 17023 | return emitUnaryBuiltin(*this, E, Intrinsic::amdgcn_rsq_clamp); |
| 17024 | case AMDGPU::BI__builtin_amdgcn_sinf: |
| 17025 | case AMDGPU::BI__builtin_amdgcn_sinh: |
| 17026 | return emitUnaryBuiltin(*this, E, Intrinsic::amdgcn_sin); |
| 17027 | case AMDGPU::BI__builtin_amdgcn_cosf: |
| 17028 | case AMDGPU::BI__builtin_amdgcn_cosh: |
| 17029 | return emitUnaryBuiltin(*this, E, Intrinsic::amdgcn_cos); |
| 17030 | case AMDGPU::BI__builtin_amdgcn_dispatch_ptr: |
| 17031 | return EmitAMDGPUDispatchPtr(*this, E); |
| 17032 | case AMDGPU::BI__builtin_amdgcn_log_clampf: |
| 17033 | return emitUnaryBuiltin(*this, E, Intrinsic::amdgcn_log_clamp); |
| 17034 | case AMDGPU::BI__builtin_amdgcn_ldexp: |
| 17035 | case AMDGPU::BI__builtin_amdgcn_ldexpf: |
| 17036 | case AMDGPU::BI__builtin_amdgcn_ldexph: |
| 17037 | return emitFPIntBuiltin(*this, E, Intrinsic::amdgcn_ldexp); |
| 17038 | case AMDGPU::BI__builtin_amdgcn_frexp_mant: |
| 17039 | case AMDGPU::BI__builtin_amdgcn_frexp_mantf: |
| 17040 | case AMDGPU::BI__builtin_amdgcn_frexp_manth: |
| 17041 | return emitUnaryBuiltin(*this, E, Intrinsic::amdgcn_frexp_mant); |
| 17042 | case AMDGPU::BI__builtin_amdgcn_frexp_exp: |
| 17043 | case AMDGPU::BI__builtin_amdgcn_frexp_expf: { |
| 17044 | Value *Src0 = EmitScalarExpr(E->getArg(0)); |
| 17045 | Function *F = CGM.getIntrinsic(Intrinsic::amdgcn_frexp_exp, |
| 17046 | { Builder.getInt32Ty(), Src0->getType() }); |
| 17047 | return Builder.CreateCall(F, Src0); |
| 17048 | } |
| 17049 | case AMDGPU::BI__builtin_amdgcn_frexp_exph: { |
| 17050 | Value *Src0 = EmitScalarExpr(E->getArg(0)); |
| 17051 | Function *F = CGM.getIntrinsic(Intrinsic::amdgcn_frexp_exp, |
| 17052 | { Builder.getInt16Ty(), Src0->getType() }); |
| 17053 | return Builder.CreateCall(F, Src0); |
| 17054 | } |
| 17055 | case AMDGPU::BI__builtin_amdgcn_fract: |
| 17056 | case AMDGPU::BI__builtin_amdgcn_fractf: |
| 17057 | case AMDGPU::BI__builtin_amdgcn_fracth: |
| 17058 | return emitUnaryBuiltin(*this, E, Intrinsic::amdgcn_fract); |
| 17059 | case AMDGPU::BI__builtin_amdgcn_lerp: |
| 17060 | return emitTernaryBuiltin(*this, E, Intrinsic::amdgcn_lerp); |
| 17061 | case AMDGPU::BI__builtin_amdgcn_ubfe: |
| 17062 | return emitTernaryBuiltin(*this, E, Intrinsic::amdgcn_ubfe); |
| 17063 | case AMDGPU::BI__builtin_amdgcn_sbfe: |
| 17064 | return emitTernaryBuiltin(*this, E, Intrinsic::amdgcn_sbfe); |
| 17065 | case AMDGPU::BI__builtin_amdgcn_ballot_w32: |
| 17066 | case AMDGPU::BI__builtin_amdgcn_ballot_w64: { |
| 17067 | llvm::Type *ResultType = ConvertType(E->getType()); |
| 17068 | llvm::Value *Src = EmitScalarExpr(E->getArg(0)); |
| 17069 | Function *F = CGM.getIntrinsic(Intrinsic::amdgcn_ballot, { ResultType }); |
| 17070 | return Builder.CreateCall(F, { Src }); |
| 17071 | } |
| 17072 | case AMDGPU::BI__builtin_amdgcn_uicmp: |
| 17073 | case AMDGPU::BI__builtin_amdgcn_uicmpl: |
| 17074 | case AMDGPU::BI__builtin_amdgcn_sicmp: |
| 17075 | case AMDGPU::BI__builtin_amdgcn_sicmpl: { |
| 17076 | llvm::Value *Src0 = EmitScalarExpr(E->getArg(0)); |
| 17077 | llvm::Value *Src1 = EmitScalarExpr(E->getArg(1)); |
| 17078 | llvm::Value *Src2 = EmitScalarExpr(E->getArg(2)); |
| 17079 | |
| 17080 | // FIXME-GFX10: How should 32 bit mask be handled? |
| 17081 | Function *F = CGM.getIntrinsic(Intrinsic::amdgcn_icmp, |
| 17082 | { Builder.getInt64Ty(), Src0->getType() }); |
| 17083 | return Builder.CreateCall(F, { Src0, Src1, Src2 }); |
| 17084 | } |
| 17085 | case AMDGPU::BI__builtin_amdgcn_fcmp: |
| 17086 | case AMDGPU::BI__builtin_amdgcn_fcmpf: { |
| 17087 | llvm::Value *Src0 = EmitScalarExpr(E->getArg(0)); |
| 17088 | llvm::Value *Src1 = EmitScalarExpr(E->getArg(1)); |
| 17089 | llvm::Value *Src2 = EmitScalarExpr(E->getArg(2)); |
| 17090 | |
| 17091 | // FIXME-GFX10: How should 32 bit mask be handled? |
| 17092 | Function *F = CGM.getIntrinsic(Intrinsic::amdgcn_fcmp, |
| 17093 | { Builder.getInt64Ty(), Src0->getType() }); |
| 17094 | return Builder.CreateCall(F, { Src0, Src1, Src2 }); |
| 17095 | } |
| 17096 | case AMDGPU::BI__builtin_amdgcn_class: |
| 17097 | case AMDGPU::BI__builtin_amdgcn_classf: |
| 17098 | case AMDGPU::BI__builtin_amdgcn_classh: |
| 17099 | return emitFPIntBuiltin(*this, E, Intrinsic::amdgcn_class); |
| 17100 | case AMDGPU::BI__builtin_amdgcn_fmed3f: |
| 17101 | case AMDGPU::BI__builtin_amdgcn_fmed3h: |
| 17102 | return emitTernaryBuiltin(*this, E, Intrinsic::amdgcn_fmed3); |
| 17103 | case AMDGPU::BI__builtin_amdgcn_ds_append: |
| 17104 | case AMDGPU::BI__builtin_amdgcn_ds_consume: { |
| 17105 | Intrinsic::ID Intrin = BuiltinID == AMDGPU::BI__builtin_amdgcn_ds_append ? |
| 17106 | Intrinsic::amdgcn_ds_append : Intrinsic::amdgcn_ds_consume; |
| 17107 | Value *Src0 = EmitScalarExpr(E->getArg(0)); |
| 17108 | Function *F = CGM.getIntrinsic(Intrin, { Src0->getType() }); |
| 17109 | return Builder.CreateCall(F, { Src0, Builder.getFalse() }); |
| 17110 | } |
| 17111 | case AMDGPU::BI__builtin_amdgcn_ds_faddf: |
| 17112 | case AMDGPU::BI__builtin_amdgcn_ds_fminf: |
| 17113 | case AMDGPU::BI__builtin_amdgcn_ds_fmaxf: { |
| 17114 | Intrinsic::ID Intrin; |
| 17115 | switch (BuiltinID) { |
| 17116 | case AMDGPU::BI__builtin_amdgcn_ds_faddf: |
| 17117 | Intrin = Intrinsic::amdgcn_ds_fadd; |
| 17118 | break; |
| 17119 | case AMDGPU::BI__builtin_amdgcn_ds_fminf: |
| 17120 | Intrin = Intrinsic::amdgcn_ds_fmin; |
| 17121 | break; |
| 17122 | case AMDGPU::BI__builtin_amdgcn_ds_fmaxf: |
| 17123 | Intrin = Intrinsic::amdgcn_ds_fmax; |
| 17124 | break; |
| 17125 | } |
| 17126 | llvm::Value *Src0 = EmitScalarExpr(E->getArg(0)); |
| 17127 | llvm::Value *Src1 = EmitScalarExpr(E->getArg(1)); |
| 17128 | llvm::Value *Src2 = EmitScalarExpr(E->getArg(2)); |
| 17129 | llvm::Value *Src3 = EmitScalarExpr(E->getArg(3)); |
| 17130 | llvm::Value *Src4 = EmitScalarExpr(E->getArg(4)); |
| 17131 | llvm::Function *F = CGM.getIntrinsic(Intrin, { Src1->getType() }); |
| 17132 | llvm::FunctionType *FTy = F->getFunctionType(); |
| 17133 | llvm::Type *PTy = FTy->getParamType(0); |
| 17134 | Src0 = Builder.CreatePointerBitCastOrAddrSpaceCast(Src0, PTy); |
| 17135 | return Builder.CreateCall(F, { Src0, Src1, Src2, Src3, Src4 }); |
| 17136 | } |
| 17137 | case AMDGPU::BI__builtin_amdgcn_global_atomic_fadd_f64: |
| 17138 | case AMDGPU::BI__builtin_amdgcn_global_atomic_fadd_f32: |
| 17139 | case AMDGPU::BI__builtin_amdgcn_global_atomic_fadd_v2f16: |
| 17140 | case AMDGPU::BI__builtin_amdgcn_global_atomic_fmin_f64: |
| 17141 | case AMDGPU::BI__builtin_amdgcn_global_atomic_fmax_f64: |
| 17142 | case AMDGPU::BI__builtin_amdgcn_flat_atomic_fadd_f64: |
| 17143 | case AMDGPU::BI__builtin_amdgcn_flat_atomic_fmin_f64: |
| 17144 | case AMDGPU::BI__builtin_amdgcn_flat_atomic_fmax_f64: |
| 17145 | case AMDGPU::BI__builtin_amdgcn_flat_atomic_fadd_f32: |
| 17146 | case AMDGPU::BI__builtin_amdgcn_flat_atomic_fadd_v2f16: { |
| 17147 | Intrinsic::ID IID; |
| 17148 | llvm::Type *ArgTy = llvm::Type::getDoubleTy(getLLVMContext()); |
| 17149 | switch (BuiltinID) { |
| 17150 | case AMDGPU::BI__builtin_amdgcn_global_atomic_fadd_f32: |
| 17151 | ArgTy = llvm::Type::getFloatTy(getLLVMContext()); |
| 17152 | IID = Intrinsic::amdgcn_global_atomic_fadd; |
| 17153 | break; |
| 17154 | case AMDGPU::BI__builtin_amdgcn_global_atomic_fadd_v2f16: |
| 17155 | ArgTy = llvm::FixedVectorType::get( |
| 17156 | llvm::Type::getHalfTy(getLLVMContext()), 2); |
| 17157 | IID = Intrinsic::amdgcn_global_atomic_fadd; |
| 17158 | break; |
| 17159 | case AMDGPU::BI__builtin_amdgcn_global_atomic_fadd_f64: |
| 17160 | IID = Intrinsic::amdgcn_global_atomic_fadd; |
| 17161 | break; |
| 17162 | case AMDGPU::BI__builtin_amdgcn_global_atomic_fmin_f64: |
| 17163 | IID = Intrinsic::amdgcn_global_atomic_fmin; |
| 17164 | break; |
| 17165 | case AMDGPU::BI__builtin_amdgcn_global_atomic_fmax_f64: |
| 17166 | IID = Intrinsic::amdgcn_global_atomic_fmax; |
| 17167 | break; |
| 17168 | case AMDGPU::BI__builtin_amdgcn_flat_atomic_fadd_f64: |
| 17169 | IID = Intrinsic::amdgcn_flat_atomic_fadd; |
| 17170 | break; |
| 17171 | case AMDGPU::BI__builtin_amdgcn_flat_atomic_fmin_f64: |
| 17172 | IID = Intrinsic::amdgcn_flat_atomic_fmin; |
| 17173 | break; |
| 17174 | case AMDGPU::BI__builtin_amdgcn_flat_atomic_fmax_f64: |
| 17175 | IID = Intrinsic::amdgcn_flat_atomic_fmax; |
| 17176 | break; |
| 17177 | case AMDGPU::BI__builtin_amdgcn_flat_atomic_fadd_f32: |
| 17178 | ArgTy = llvm::Type::getFloatTy(getLLVMContext()); |
| 17179 | IID = Intrinsic::amdgcn_flat_atomic_fadd; |
| 17180 | break; |
| 17181 | case AMDGPU::BI__builtin_amdgcn_flat_atomic_fadd_v2f16: |
| 17182 | ArgTy = llvm::FixedVectorType::get( |
| 17183 | llvm::Type::getHalfTy(getLLVMContext()), 2); |
| 17184 | IID = Intrinsic::amdgcn_flat_atomic_fadd; |
| 17185 | break; |
| 17186 | } |
| 17187 | llvm::Value *Addr = EmitScalarExpr(E->getArg(0)); |
| 17188 | llvm::Value *Val = EmitScalarExpr(E->getArg(1)); |
| 17189 | llvm::Function *F = |
| 17190 | CGM.getIntrinsic(IID, {ArgTy, Addr->getType(), Val->getType()}); |
| 17191 | return Builder.CreateCall(F, {Addr, Val}); |
| 17192 | } |
| 17193 | case AMDGPU::BI__builtin_amdgcn_global_atomic_fadd_v2bf16: |
| 17194 | case AMDGPU::BI__builtin_amdgcn_flat_atomic_fadd_v2bf16: { |
| 17195 | Intrinsic::ID IID; |
| 17196 | switch (BuiltinID) { |
| 17197 | case AMDGPU::BI__builtin_amdgcn_global_atomic_fadd_v2bf16: |
| 17198 | IID = Intrinsic::amdgcn_global_atomic_fadd_v2bf16; |
| 17199 | break; |
| 17200 | case AMDGPU::BI__builtin_amdgcn_flat_atomic_fadd_v2bf16: |
| 17201 | IID = Intrinsic::amdgcn_flat_atomic_fadd_v2bf16; |
| 17202 | break; |
| 17203 | } |
| 17204 | llvm::Value *Addr = EmitScalarExpr(E->getArg(0)); |
| 17205 | llvm::Value *Val = EmitScalarExpr(E->getArg(1)); |
| 17206 | llvm::Function *F = CGM.getIntrinsic(IID, {Addr->getType()}); |
| 17207 | return Builder.CreateCall(F, {Addr, Val}); |
| 17208 | } |
| 17209 | case AMDGPU::BI__builtin_amdgcn_ds_atomic_fadd_f64: |
| 17210 | case AMDGPU::BI__builtin_amdgcn_ds_atomic_fadd_f32: |
| 17211 | case AMDGPU::BI__builtin_amdgcn_ds_atomic_fadd_v2f16: { |
| 17212 | Intrinsic::ID IID; |
| 17213 | llvm::Type *ArgTy; |
| 17214 | switch (BuiltinID) { |
| 17215 | case AMDGPU::BI__builtin_amdgcn_ds_atomic_fadd_f32: |
| 17216 | ArgTy = llvm::Type::getFloatTy(getLLVMContext()); |
| 17217 | IID = Intrinsic::amdgcn_ds_fadd; |
| 17218 | break; |
| 17219 | case AMDGPU::BI__builtin_amdgcn_ds_atomic_fadd_f64: |
| 17220 | ArgTy = llvm::Type::getDoubleTy(getLLVMContext()); |
| 17221 | IID = Intrinsic::amdgcn_ds_fadd; |
| 17222 | break; |
| 17223 | case AMDGPU::BI__builtin_amdgcn_ds_atomic_fadd_v2f16: |
| 17224 | ArgTy = llvm::FixedVectorType::get( |
| 17225 | llvm::Type::getHalfTy(getLLVMContext()), 2); |
| 17226 | IID = Intrinsic::amdgcn_ds_fadd; |
| 17227 | break; |
| 17228 | } |
| 17229 | llvm::Value *Addr = EmitScalarExpr(E->getArg(0)); |
| 17230 | llvm::Value *Val = EmitScalarExpr(E->getArg(1)); |
| 17231 | llvm::Constant *ZeroI32 = llvm::ConstantInt::getIntegerValue( |
| 17232 | llvm::Type::getInt32Ty(getLLVMContext()), APInt(32, 0, true)); |
| 17233 | llvm::Constant *ZeroI1 = llvm::ConstantInt::getIntegerValue( |
| 17234 | llvm::Type::getInt1Ty(getLLVMContext()), APInt(1, 0)); |
| 17235 | llvm::Function *F = CGM.getIntrinsic(IID, {ArgTy}); |
| 17236 | return Builder.CreateCall(F, {Addr, Val, ZeroI32, ZeroI32, ZeroI1}); |
| 17237 | } |
| 17238 | case AMDGPU::BI__builtin_amdgcn_read_exec: { |
| 17239 | CallInst *CI = cast<CallInst>( |
| 17240 | EmitSpecialRegisterBuiltin(*this, E, Int64Ty, Int64Ty, NormalRead, "exec")); |
| 17241 | CI->setConvergent(); |
| 17242 | return CI; |
| 17243 | } |
| 17244 | case AMDGPU::BI__builtin_amdgcn_read_exec_lo: |
| 17245 | case AMDGPU::BI__builtin_amdgcn_read_exec_hi: { |
| 17246 | StringRef RegName = BuiltinID == AMDGPU::BI__builtin_amdgcn_read_exec_lo ? |
| 17247 | "exec_lo" : "exec_hi"; |
| 17248 | CallInst *CI = cast<CallInst>( |
| 17249 | EmitSpecialRegisterBuiltin(*this, E, Int32Ty, Int32Ty, NormalRead, RegName)); |
| 17250 | CI->setConvergent(); |
| 17251 | return CI; |
| 17252 | } |
| 17253 | case AMDGPU::BI__builtin_amdgcn_image_bvh_intersect_ray: |
| 17254 | case AMDGPU::BI__builtin_amdgcn_image_bvh_intersect_ray_h: |
| 17255 | case AMDGPU::BI__builtin_amdgcn_image_bvh_intersect_ray_l: |
| 17256 | case AMDGPU::BI__builtin_amdgcn_image_bvh_intersect_ray_lh: { |
| 17257 | llvm::Value *NodePtr = EmitScalarExpr(E->getArg(0)); |
| 17258 | llvm::Value *RayExtent = EmitScalarExpr(E->getArg(1)); |
| 17259 | llvm::Value *RayOrigin = EmitScalarExpr(E->getArg(2)); |
| 17260 | llvm::Value *RayDir = EmitScalarExpr(E->getArg(3)); |
| 17261 | llvm::Value *RayInverseDir = EmitScalarExpr(E->getArg(4)); |
| 17262 | llvm::Value *TextureDescr = EmitScalarExpr(E->getArg(5)); |
| 17263 | |
| 17264 | // The builtins take these arguments as vec4 where the last element is |
| 17265 | // ignored. The intrinsic takes them as vec3. |
| 17266 | RayOrigin = Builder.CreateShuffleVector(RayOrigin, RayOrigin, |
| 17267 | ArrayRef<int>{0, 1, 2}); |
| 17268 | RayDir = |
| 17269 | Builder.CreateShuffleVector(RayDir, RayDir, ArrayRef<int>{0, 1, 2}); |
| 17270 | RayInverseDir = Builder.CreateShuffleVector(RayInverseDir, RayInverseDir, |
| 17271 | ArrayRef<int>{0, 1, 2}); |
| 17272 | |
| 17273 | Function *F = CGM.getIntrinsic(Intrinsic::amdgcn_image_bvh_intersect_ray, |
| 17274 | {NodePtr->getType(), RayDir->getType()}); |
| 17275 | return Builder.CreateCall(F, {NodePtr, RayExtent, RayOrigin, RayDir, |
| 17276 | RayInverseDir, TextureDescr}); |
| 17277 | } |
| 17278 | |
| 17279 | case AMDGPU::BI__builtin_amdgcn_ds_bvh_stack_rtn: { |
| 17280 | SmallVector<Value *, 4> Args; |
| 17281 | for (int i = 0, e = E->getNumArgs(); i != e; ++i) |
| 17282 | Args.push_back(EmitScalarExpr(E->getArg(i))); |
| 17283 | |
| 17284 | Function *F = CGM.getIntrinsic(Intrinsic::amdgcn_ds_bvh_stack_rtn); |
| 17285 | Value *Call = Builder.CreateCall(F, Args); |
| 17286 | Value *Rtn = Builder.CreateExtractValue(Call, 0); |
| 17287 | Value *A = Builder.CreateExtractValue(Call, 1); |
| 17288 | llvm::Type *RetTy = ConvertType(E->getType()); |
| 17289 | Value *I0 = Builder.CreateInsertElement(PoisonValue::get(RetTy), Rtn, |
| 17290 | (uint64_t)0); |
| 17291 | return Builder.CreateInsertElement(I0, A, 1); |
| 17292 | } |
| 17293 | |
| 17294 | case AMDGPU::BI__builtin_amdgcn_wmma_bf16_16x16x16_bf16_w32: |
| 17295 | case AMDGPU::BI__builtin_amdgcn_wmma_bf16_16x16x16_bf16_w64: |
| 17296 | case AMDGPU::BI__builtin_amdgcn_wmma_f16_16x16x16_f16_w32: |
| 17297 | case AMDGPU::BI__builtin_amdgcn_wmma_f16_16x16x16_f16_w64: |
| 17298 | case AMDGPU::BI__builtin_amdgcn_wmma_f32_16x16x16_bf16_w32: |
| 17299 | case AMDGPU::BI__builtin_amdgcn_wmma_f32_16x16x16_bf16_w64: |
| 17300 | case AMDGPU::BI__builtin_amdgcn_wmma_f32_16x16x16_f16_w32: |
| 17301 | case AMDGPU::BI__builtin_amdgcn_wmma_f32_16x16x16_f16_w64: |
| 17302 | case AMDGPU::BI__builtin_amdgcn_wmma_i32_16x16x16_iu4_w32: |
| 17303 | case AMDGPU::BI__builtin_amdgcn_wmma_i32_16x16x16_iu4_w64: |
| 17304 | case AMDGPU::BI__builtin_amdgcn_wmma_i32_16x16x16_iu8_w32: |
| 17305 | case AMDGPU::BI__builtin_amdgcn_wmma_i32_16x16x16_iu8_w64: { |
| 17306 | |
| 17307 | // These operations perform a matrix multiplication and accumulation of |
| 17308 | // the form: |
| 17309 | // D = A * B + C |
| 17310 | // The return type always matches the type of matrix C. |
| 17311 | unsigned ArgForMatchingRetType; |
| 17312 | unsigned BuiltinWMMAOp; |
| 17313 | |
| 17314 | switch (BuiltinID) { |
| 17315 | case AMDGPU::BI__builtin_amdgcn_wmma_f32_16x16x16_f16_w32: |
| 17316 | case AMDGPU::BI__builtin_amdgcn_wmma_f32_16x16x16_f16_w64: |
| 17317 | ArgForMatchingRetType = 2; |
| 17318 | BuiltinWMMAOp = Intrinsic::amdgcn_wmma_f32_16x16x16_f16; |
| 17319 | break; |
| 17320 | case AMDGPU::BI__builtin_amdgcn_wmma_f32_16x16x16_bf16_w32: |
| 17321 | case AMDGPU::BI__builtin_amdgcn_wmma_f32_16x16x16_bf16_w64: |
| 17322 | ArgForMatchingRetType = 2; |
| 17323 | BuiltinWMMAOp = Intrinsic::amdgcn_wmma_f32_16x16x16_bf16; |
| 17324 | break; |
| 17325 | case AMDGPU::BI__builtin_amdgcn_wmma_f16_16x16x16_f16_w32: |
| 17326 | case AMDGPU::BI__builtin_amdgcn_wmma_f16_16x16x16_f16_w64: |
| 17327 | ArgForMatchingRetType = 2; |
| 17328 | BuiltinWMMAOp = Intrinsic::amdgcn_wmma_f16_16x16x16_f16; |
| 17329 | break; |
| 17330 | case AMDGPU::BI__builtin_amdgcn_wmma_bf16_16x16x16_bf16_w32: |
| 17331 | case AMDGPU::BI__builtin_amdgcn_wmma_bf16_16x16x16_bf16_w64: |
| 17332 | ArgForMatchingRetType = 2; |
| 17333 | BuiltinWMMAOp = Intrinsic::amdgcn_wmma_bf16_16x16x16_bf16; |
| 17334 | break; |
| 17335 | case AMDGPU::BI__builtin_amdgcn_wmma_i32_16x16x16_iu8_w32: |
| 17336 | case AMDGPU::BI__builtin_amdgcn_wmma_i32_16x16x16_iu8_w64: |
| 17337 | ArgForMatchingRetType = 4; |
| 17338 | BuiltinWMMAOp = Intrinsic::amdgcn_wmma_i32_16x16x16_iu8; |
| 17339 | break; |
| 17340 | case AMDGPU::BI__builtin_amdgcn_wmma_i32_16x16x16_iu4_w32: |
| 17341 | case AMDGPU::BI__builtin_amdgcn_wmma_i32_16x16x16_iu4_w64: |
| 17342 | ArgForMatchingRetType = 4; |
| 17343 | BuiltinWMMAOp = Intrinsic::amdgcn_wmma_i32_16x16x16_iu4; |
| 17344 | break; |
| 17345 | } |
| 17346 | |
| 17347 | SmallVector<Value *, 6> Args; |
| 17348 | for (int i = 0, e = E->getNumArgs(); i != e; ++i) |
| 17349 | Args.push_back(EmitScalarExpr(E->getArg(i))); |
| 17350 | |
| 17351 | Function *F = CGM.getIntrinsic(BuiltinWMMAOp, |
| 17352 | {Args[ArgForMatchingRetType]->getType()}); |
| 17353 | |
| 17354 | return Builder.CreateCall(F, Args); |
| 17355 | } |
| 17356 | |
| 17357 | // amdgcn workitem |
| 17358 | case AMDGPU::BI__builtin_amdgcn_workitem_id_x: |
| 17359 | return emitRangedBuiltin(*this, Intrinsic::amdgcn_workitem_id_x, 0, 1024); |
| 17360 | case AMDGPU::BI__builtin_amdgcn_workitem_id_y: |
| 17361 | return emitRangedBuiltin(*this, Intrinsic::amdgcn_workitem_id_y, 0, 1024); |
| 17362 | case AMDGPU::BI__builtin_amdgcn_workitem_id_z: |
| 17363 | return emitRangedBuiltin(*this, Intrinsic::amdgcn_workitem_id_z, 0, 1024); |
| 17364 | |
| 17365 | // amdgcn workgroup size |
| 17366 | case AMDGPU::BI__builtin_amdgcn_workgroup_size_x: |
| 17367 | return EmitAMDGPUWorkGroupSize(*this, 0); |
| 17368 | case AMDGPU::BI__builtin_amdgcn_workgroup_size_y: |
| 17369 | return EmitAMDGPUWorkGroupSize(*this, 1); |
| 17370 | case AMDGPU::BI__builtin_amdgcn_workgroup_size_z: |
| 17371 | return EmitAMDGPUWorkGroupSize(*this, 2); |
| 17372 | |
| 17373 | // amdgcn grid size |
| 17374 | case AMDGPU::BI__builtin_amdgcn_grid_size_x: |
| 17375 | return EmitAMDGPUGridSize(*this, 0); |
| 17376 | case AMDGPU::BI__builtin_amdgcn_grid_size_y: |
| 17377 | return EmitAMDGPUGridSize(*this, 1); |
| 17378 | case AMDGPU::BI__builtin_amdgcn_grid_size_z: |
| 17379 | return EmitAMDGPUGridSize(*this, 2); |
| 17380 | |
| 17381 | // r600 intrinsics |
| 17382 | case AMDGPU::BI__builtin_r600_recipsqrt_ieee: |
| 17383 | case AMDGPU::BI__builtin_r600_recipsqrt_ieeef: |
| 17384 | return emitUnaryBuiltin(*this, E, Intrinsic::r600_recipsqrt_ieee); |
| 17385 | case AMDGPU::BI__builtin_r600_read_tidig_x: |
| 17386 | return emitRangedBuiltin(*this, Intrinsic::r600_read_tidig_x, 0, 1024); |
| 17387 | case AMDGPU::BI__builtin_r600_read_tidig_y: |
| 17388 | return emitRangedBuiltin(*this, Intrinsic::r600_read_tidig_y, 0, 1024); |
| 17389 | case AMDGPU::BI__builtin_r600_read_tidig_z: |
| 17390 | return emitRangedBuiltin(*this, Intrinsic::r600_read_tidig_z, 0, 1024); |
| 17391 | case AMDGPU::BI__builtin_amdgcn_alignbit: { |
| 17392 | llvm::Value *Src0 = EmitScalarExpr(E->getArg(0)); |
| 17393 | llvm::Value *Src1 = EmitScalarExpr(E->getArg(1)); |
| 17394 | llvm::Value *Src2 = EmitScalarExpr(E->getArg(2)); |
| 17395 | Function *F = CGM.getIntrinsic(Intrinsic::fshr, Src0->getType()); |
| 17396 | return Builder.CreateCall(F, { Src0, Src1, Src2 }); |
| 17397 | } |
| 17398 | case AMDGPU::BI__builtin_amdgcn_fence: { |
| 17399 | ProcessOrderScopeAMDGCN(EmitScalarExpr(E->getArg(0)), |
| 17400 | EmitScalarExpr(E->getArg(1)), AO, SSID); |
| 17401 | return Builder.CreateFence(AO, SSID); |
| 17402 | } |
| 17403 | case AMDGPU::BI__builtin_amdgcn_atomic_inc32: |
| 17404 | case AMDGPU::BI__builtin_amdgcn_atomic_inc64: |
| 17405 | case AMDGPU::BI__builtin_amdgcn_atomic_dec32: |
| 17406 | case AMDGPU::BI__builtin_amdgcn_atomic_dec64: { |
| 17407 | unsigned BuiltinAtomicOp; |
| 17408 | llvm::Type *ResultType = ConvertType(E->getType()); |
| 17409 | |
| 17410 | switch (BuiltinID) { |
| 17411 | case AMDGPU::BI__builtin_amdgcn_atomic_inc32: |
| 17412 | case AMDGPU::BI__builtin_amdgcn_atomic_inc64: |
| 17413 | BuiltinAtomicOp = Intrinsic::amdgcn_atomic_inc; |
| 17414 | break; |
| 17415 | case AMDGPU::BI__builtin_amdgcn_atomic_dec32: |
| 17416 | case AMDGPU::BI__builtin_amdgcn_atomic_dec64: |
| 17417 | BuiltinAtomicOp = Intrinsic::amdgcn_atomic_dec; |
| 17418 | break; |
| 17419 | } |
| 17420 | |
| 17421 | Value *Ptr = EmitScalarExpr(E->getArg(0)); |
| 17422 | Value *Val = EmitScalarExpr(E->getArg(1)); |
| 17423 | |
| 17424 | llvm::Function *F = |
| 17425 | CGM.getIntrinsic(BuiltinAtomicOp, {ResultType, Ptr->getType()}); |
| 17426 | |
| 17427 | ProcessOrderScopeAMDGCN(EmitScalarExpr(E->getArg(2)), |
| 17428 | EmitScalarExpr(E->getArg(3)), AO, SSID); |
| 17429 | |
| 17430 | // llvm.amdgcn.atomic.inc and llvm.amdgcn.atomic.dec expects ordering and |
| 17431 | // scope as unsigned values |
| 17432 | Value *MemOrder = Builder.getInt32(static_cast<int>(AO)); |
| 17433 | Value *MemScope = Builder.getInt32(static_cast<int>(SSID)); |
| 17434 | |
| 17435 | QualType PtrTy = E->getArg(0)->IgnoreImpCasts()->getType(); |
| 17436 | bool Volatile = |
| 17437 | PtrTy->castAs<PointerType>()->getPointeeType().isVolatileQualified(); |
| 17438 | Value *IsVolatile = Builder.getInt1(static_cast<bool>(Volatile)); |
| 17439 | |
| 17440 | return Builder.CreateCall(F, {Ptr, Val, MemOrder, MemScope, IsVolatile}); |
| 17441 | } |
| 17442 | case AMDGPU::BI__builtin_amdgcn_s_sendmsg_rtn: |
| 17443 | case AMDGPU::BI__builtin_amdgcn_s_sendmsg_rtnl: { |
| 17444 | llvm::Value *Arg = EmitScalarExpr(E->getArg(0)); |
| 17445 | llvm::Type *ResultType = ConvertType(E->getType()); |
| 17446 | // s_sendmsg_rtn is mangled using return type only. |
| 17447 | Function *F = |
| 17448 | CGM.getIntrinsic(Intrinsic::amdgcn_s_sendmsg_rtn, {ResultType}); |
| 17449 | return Builder.CreateCall(F, {Arg}); |
| 17450 | } |
| 17451 | default: |
| 17452 | return nullptr; |
| 17453 | } |
| 17454 | } |
| 17455 | |
| 17456 | /// Handle a SystemZ function in which the final argument is a pointer |
| 17457 | /// to an int that receives the post-instruction CC value. At the LLVM level |
| 17458 | /// this is represented as a function that returns a {result, cc} pair. |
| 17459 | static Value *EmitSystemZIntrinsicWithCC(CodeGenFunction &CGF, |
| 17460 | unsigned IntrinsicID, |
| 17461 | const CallExpr *E) { |
| 17462 | unsigned NumArgs = E->getNumArgs() - 1; |
| 17463 | SmallVector<Value *, 8> Args(NumArgs); |
| 17464 | for (unsigned I = 0; I < NumArgs; ++I) |
| 17465 | Args[I] = CGF.EmitScalarExpr(E->getArg(I)); |
| 17466 | Address CCPtr = CGF.EmitPointerWithAlignment(E->getArg(NumArgs)); |
| 17467 | Function *F = CGF.CGM.getIntrinsic(IntrinsicID); |
| 17468 | Value *Call = CGF.Builder.CreateCall(F, Args); |
| 17469 | Value *CC = CGF.Builder.CreateExtractValue(Call, 1); |
| 17470 | CGF.Builder.CreateStore(CC, CCPtr); |
| 17471 | return CGF.Builder.CreateExtractValue(Call, 0); |
| 17472 | } |
| 17473 | |
| 17474 | Value *CodeGenFunction::EmitSystemZBuiltinExpr(unsigned BuiltinID, |
| 17475 | const CallExpr *E) { |
| 17476 | switch (BuiltinID) { |
| 17477 | case SystemZ::BI__builtin_tbegin: { |
| 17478 | Value *TDB = EmitScalarExpr(E->getArg(0)); |
| 17479 | Value *Control = llvm::ConstantInt::get(Int32Ty, 0xff0c); |
| 17480 | Function *F = CGM.getIntrinsic(Intrinsic::s390_tbegin); |
| 17481 | return Builder.CreateCall(F, {TDB, Control}); |
| 17482 | } |
| 17483 | case SystemZ::BI__builtin_tbegin_nofloat: { |
| 17484 | Value *TDB = EmitScalarExpr(E->getArg(0)); |
| 17485 | Value *Control = llvm::ConstantInt::get(Int32Ty, 0xff0c); |
| 17486 | Function *F = CGM.getIntrinsic(Intrinsic::s390_tbegin_nofloat); |
| 17487 | return Builder.CreateCall(F, {TDB, Control}); |
| 17488 | } |
| 17489 | case SystemZ::BI__builtin_tbeginc: { |
| 17490 | Value *TDB = llvm::ConstantPointerNull::get(Int8PtrTy); |
| 17491 | Value *Control = llvm::ConstantInt::get(Int32Ty, 0xff08); |
| 17492 | Function *F = CGM.getIntrinsic(Intrinsic::s390_tbeginc); |
| 17493 | return Builder.CreateCall(F, {TDB, Control}); |
| 17494 | } |
| 17495 | case SystemZ::BI__builtin_tabort: { |
| 17496 | Value *Data = EmitScalarExpr(E->getArg(0)); |
| 17497 | Function *F = CGM.getIntrinsic(Intrinsic::s390_tabort); |
| 17498 | return Builder.CreateCall(F, Builder.CreateSExt(Data, Int64Ty, "tabort")); |
| 17499 | } |
| 17500 | case SystemZ::BI__builtin_non_tx_store: { |
| 17501 | Value *Address = EmitScalarExpr(E->getArg(0)); |
| 17502 | Value *Data = EmitScalarExpr(E->getArg(1)); |
| 17503 | Function *F = CGM.getIntrinsic(Intrinsic::s390_ntstg); |
| 17504 | return Builder.CreateCall(F, {Data, Address}); |
| 17505 | } |
| 17506 | |
| 17507 | // Vector builtins. Note that most vector builtins are mapped automatically |
| 17508 | // to target-specific LLVM intrinsics. The ones handled specially here can |
| 17509 | // be represented via standard LLVM IR, which is preferable to enable common |
| 17510 | // LLVM optimizations. |
| 17511 | |
| 17512 | case SystemZ::BI__builtin_s390_vpopctb: |
| 17513 | case SystemZ::BI__builtin_s390_vpopcth: |
| 17514 | case SystemZ::BI__builtin_s390_vpopctf: |
| 17515 | case SystemZ::BI__builtin_s390_vpopctg: { |
| 17516 | llvm::Type *ResultType = ConvertType(E->getType()); |
| 17517 | Value *X = EmitScalarExpr(E->getArg(0)); |
| 17518 | Function *F = CGM.getIntrinsic(Intrinsic::ctpop, ResultType); |
| 17519 | return Builder.CreateCall(F, X); |
| 17520 | } |
| 17521 | |
| 17522 | case SystemZ::BI__builtin_s390_vclzb: |
| 17523 | case SystemZ::BI__builtin_s390_vclzh: |
| 17524 | case SystemZ::BI__builtin_s390_vclzf: |
| 17525 | case SystemZ::BI__builtin_s390_vclzg: { |
| 17526 | llvm::Type *ResultType = ConvertType(E->getType()); |
| 17527 | Value *X = EmitScalarExpr(E->getArg(0)); |
| 17528 | Value *Undef = ConstantInt::get(Builder.getInt1Ty(), false); |
| 17529 | Function *F = CGM.getIntrinsic(Intrinsic::ctlz, ResultType); |
| 17530 | return Builder.CreateCall(F, {X, Undef}); |
| 17531 | } |
| 17532 | |
| 17533 | case SystemZ::BI__builtin_s390_vctzb: |
| 17534 | case SystemZ::BI__builtin_s390_vctzh: |
| 17535 | case SystemZ::BI__builtin_s390_vctzf: |
| 17536 | case SystemZ::BI__builtin_s390_vctzg: { |
| 17537 | llvm::Type *ResultType = ConvertType(E->getType()); |
| 17538 | Value *X = EmitScalarExpr(E->getArg(0)); |
| 17539 | Value *Undef = ConstantInt::get(Builder.getInt1Ty(), false); |
| 17540 | Function *F = CGM.getIntrinsic(Intrinsic::cttz, ResultType); |
| 17541 | return Builder.CreateCall(F, {X, Undef}); |
| 17542 | } |
| 17543 | |
| 17544 | case SystemZ::BI__builtin_s390_vfsqsb: |
| 17545 | case SystemZ::BI__builtin_s390_vfsqdb: { |
| 17546 | llvm::Type *ResultType = ConvertType(E->getType()); |
| 17547 | Value *X = EmitScalarExpr(E->getArg(0)); |
| 17548 | if (Builder.getIsFPConstrained()) { |
| 17549 | Function *F = CGM.getIntrinsic(Intrinsic::experimental_constrained_sqrt, ResultType); |
| 17550 | return Builder.CreateConstrainedFPCall(F, { X }); |
| 17551 | } else { |
| 17552 | Function *F = CGM.getIntrinsic(Intrinsic::sqrt, ResultType); |
| 17553 | return Builder.CreateCall(F, X); |
| 17554 | } |
| 17555 | } |
| 17556 | case SystemZ::BI__builtin_s390_vfmasb: |
| 17557 | case SystemZ::BI__builtin_s390_vfmadb: { |
| 17558 | llvm::Type *ResultType = ConvertType(E->getType()); |
| 17559 | Value *X = EmitScalarExpr(E->getArg(0)); |
| 17560 | Value *Y = EmitScalarExpr(E->getArg(1)); |
| 17561 | Value *Z = EmitScalarExpr(E->getArg(2)); |
| 17562 | if (Builder.getIsFPConstrained()) { |
| 17563 | Function *F = CGM.getIntrinsic(Intrinsic::experimental_constrained_fma, ResultType); |
| 17564 | return Builder.CreateConstrainedFPCall(F, {X, Y, Z}); |
| 17565 | } else { |
| 17566 | Function *F = CGM.getIntrinsic(Intrinsic::fma, ResultType); |
| 17567 | return Builder.CreateCall(F, {X, Y, Z}); |
| 17568 | } |
| 17569 | } |
| 17570 | case SystemZ::BI__builtin_s390_vfmssb: |
| 17571 | case SystemZ::BI__builtin_s390_vfmsdb: { |
| 17572 | llvm::Type *ResultType = ConvertType(E->getType()); |
| 17573 | Value *X = EmitScalarExpr(E->getArg(0)); |
| 17574 | Value *Y = EmitScalarExpr(E->getArg(1)); |
| 17575 | Value *Z = EmitScalarExpr(E->getArg(2)); |
| 17576 | if (Builder.getIsFPConstrained()) { |
| 17577 | Function *F = CGM.getIntrinsic(Intrinsic::experimental_constrained_fma, ResultType); |
| 17578 | return Builder.CreateConstrainedFPCall(F, {X, Y, Builder.CreateFNeg(Z, "neg")}); |
| 17579 | } else { |
| 17580 | Function *F = CGM.getIntrinsic(Intrinsic::fma, ResultType); |
| 17581 | return Builder.CreateCall(F, {X, Y, Builder.CreateFNeg(Z, "neg")}); |
| 17582 | } |
| 17583 | } |
| 17584 | case SystemZ::BI__builtin_s390_vfnmasb: |
| 17585 | case SystemZ::BI__builtin_s390_vfnmadb: { |
| 17586 | llvm::Type *ResultType = ConvertType(E->getType()); |
| 17587 | Value *X = EmitScalarExpr(E->getArg(0)); |
| 17588 | Value *Y = EmitScalarExpr(E->getArg(1)); |
| 17589 | Value *Z = EmitScalarExpr(E->getArg(2)); |
| 17590 | if (Builder.getIsFPConstrained()) { |
| 17591 | Function *F = CGM.getIntrinsic(Intrinsic::experimental_constrained_fma, ResultType); |
| 17592 | return Builder.CreateFNeg(Builder.CreateConstrainedFPCall(F, {X, Y, Z}), "neg"); |
| 17593 | } else { |
| 17594 | Function *F = CGM.getIntrinsic(Intrinsic::fma, ResultType); |
| 17595 | return Builder.CreateFNeg(Builder.CreateCall(F, {X, Y, Z}), "neg"); |
| 17596 | } |
| 17597 | } |
| 17598 | case SystemZ::BI__builtin_s390_vfnmssb: |
| 17599 | case SystemZ::BI__builtin_s390_vfnmsdb: { |
| 17600 | llvm::Type *ResultType = ConvertType(E->getType()); |
| 17601 | Value *X = EmitScalarExpr(E->getArg(0)); |
| 17602 | Value *Y = EmitScalarExpr(E->getArg(1)); |
| 17603 | Value *Z = EmitScalarExpr(E->getArg(2)); |
| 17604 | if (Builder.getIsFPConstrained()) { |
| 17605 | Function *F = CGM.getIntrinsic(Intrinsic::experimental_constrained_fma, ResultType); |
| 17606 | Value *NegZ = Builder.CreateFNeg(Z, "sub"); |
| 17607 | return Builder.CreateFNeg(Builder.CreateConstrainedFPCall(F, {X, Y, NegZ})); |
| 17608 | } else { |
| 17609 | Function *F = CGM.getIntrinsic(Intrinsic::fma, ResultType); |
| 17610 | Value *NegZ = Builder.CreateFNeg(Z, "neg"); |
| 17611 | return Builder.CreateFNeg(Builder.CreateCall(F, {X, Y, NegZ})); |
| 17612 | } |
| 17613 | } |
| 17614 | case SystemZ::BI__builtin_s390_vflpsb: |
| 17615 | case SystemZ::BI__builtin_s390_vflpdb: { |
| 17616 | llvm::Type *ResultType = ConvertType(E->getType()); |
| 17617 | Value *X = EmitScalarExpr(E->getArg(0)); |
| 17618 | Function *F = CGM.getIntrinsic(Intrinsic::fabs, ResultType); |
| 17619 | return Builder.CreateCall(F, X); |
| 17620 | } |
| 17621 | case SystemZ::BI__builtin_s390_vflnsb: |
| 17622 | case SystemZ::BI__builtin_s390_vflndb: { |
| 17623 | llvm::Type *ResultType = ConvertType(E->getType()); |
| 17624 | Value *X = EmitScalarExpr(E->getArg(0)); |
| 17625 | Function *F = CGM.getIntrinsic(Intrinsic::fabs, ResultType); |
| 17626 | return Builder.CreateFNeg(Builder.CreateCall(F, X), "neg"); |
| 17627 | } |
| 17628 | case SystemZ::BI__builtin_s390_vfisb: |
| 17629 | case SystemZ::BI__builtin_s390_vfidb: { |
| 17630 | llvm::Type *ResultType = ConvertType(E->getType()); |
| 17631 | Value *X = EmitScalarExpr(E->getArg(0)); |
| 17632 | // Constant-fold the M4 and M5 mask arguments. |
| 17633 | llvm::APSInt M4 = *E->getArg(1)->getIntegerConstantExpr(getContext()); |
| 17634 | llvm::APSInt M5 = *E->getArg(2)->getIntegerConstantExpr(getContext()); |
| 17635 | // Check whether this instance can be represented via a LLVM standard |
| 17636 | // intrinsic. We only support some combinations of M4 and M5. |
| 17637 | Intrinsic::ID ID = Intrinsic::not_intrinsic; |
| 17638 | Intrinsic::ID CI; |
| 17639 | switch (M4.getZExtValue()) { |
| 17640 | default: break; |
| 17641 | case 0: // IEEE-inexact exception allowed |
| 17642 | switch (M5.getZExtValue()) { |
| 17643 | default: break; |
| 17644 | case 0: ID = Intrinsic::rint; |
| 17645 | CI = Intrinsic::experimental_constrained_rint; break; |
| 17646 | } |
| 17647 | break; |
| 17648 | case 4: // IEEE-inexact exception suppressed |
| 17649 | switch (M5.getZExtValue()) { |
| 17650 | default: break; |
| 17651 | case 0: ID = Intrinsic::nearbyint; |
| 17652 | CI = Intrinsic::experimental_constrained_nearbyint; break; |
| 17653 | case 1: ID = Intrinsic::round; |
| 17654 | CI = Intrinsic::experimental_constrained_round; break; |
| 17655 | case 5: ID = Intrinsic::trunc; |
| 17656 | CI = Intrinsic::experimental_constrained_trunc; break; |
| 17657 | case 6: ID = Intrinsic::ceil; |
| 17658 | CI = Intrinsic::experimental_constrained_ceil; break; |
| 17659 | case 7: ID = Intrinsic::floor; |
| 17660 | CI = Intrinsic::experimental_constrained_floor; break; |
| 17661 | } |
| 17662 | break; |
| 17663 | } |
| 17664 | if (ID != Intrinsic::not_intrinsic) { |
| 17665 | if (Builder.getIsFPConstrained()) { |
| 17666 | Function *F = CGM.getIntrinsic(CI, ResultType); |
| 17667 | return Builder.CreateConstrainedFPCall(F, X); |
| 17668 | } else { |
| 17669 | Function *F = CGM.getIntrinsic(ID, ResultType); |
| 17670 | return Builder.CreateCall(F, X); |
| 17671 | } |
| 17672 | } |
| 17673 | switch (BuiltinID) { // FIXME: constrained version? |
| 17674 | case SystemZ::BI__builtin_s390_vfisb: ID = Intrinsic::s390_vfisb; break; |
| 17675 | case SystemZ::BI__builtin_s390_vfidb: ID = Intrinsic::s390_vfidb; break; |
| 17676 | default: llvm_unreachable("Unknown BuiltinID")::llvm::llvm_unreachable_internal("Unknown BuiltinID", "clang/lib/CodeGen/CGBuiltin.cpp" , 17676); |
| 17677 | } |
| 17678 | Function *F = CGM.getIntrinsic(ID); |
| 17679 | Value *M4Value = llvm::ConstantInt::get(getLLVMContext(), M4); |
| 17680 | Value *M5Value = llvm::ConstantInt::get(getLLVMContext(), M5); |
| 17681 | return Builder.CreateCall(F, {X, M4Value, M5Value}); |
| 17682 | } |
| 17683 | case SystemZ::BI__builtin_s390_vfmaxsb: |
| 17684 | case SystemZ::BI__builtin_s390_vfmaxdb: { |
| 17685 | llvm::Type *ResultType = ConvertType(E->getType()); |
| 17686 | Value *X = EmitScalarExpr(E->getArg(0)); |
| 17687 | Value *Y = EmitScalarExpr(E->getArg(1)); |
| 17688 | // Constant-fold the M4 mask argument. |
| 17689 | llvm::APSInt M4 = *E->getArg(2)->getIntegerConstantExpr(getContext()); |
| 17690 | // Check whether this instance can be represented via a LLVM standard |
| 17691 | // intrinsic. We only support some values of M4. |
| 17692 | Intrinsic::ID ID = Intrinsic::not_intrinsic; |
| 17693 | Intrinsic::ID CI; |
| 17694 | switch (M4.getZExtValue()) { |
| 17695 | default: break; |
| 17696 | case 4: ID = Intrinsic::maxnum; |
| 17697 | CI = Intrinsic::experimental_constrained_maxnum; break; |
| 17698 | } |
| 17699 | if (ID != Intrinsic::not_intrinsic) { |
| 17700 | if (Builder.getIsFPConstrained()) { |
| 17701 | Function *F = CGM.getIntrinsic(CI, ResultType); |
| 17702 | return Builder.CreateConstrainedFPCall(F, {X, Y}); |
| 17703 | } else { |
| 17704 | Function *F = CGM.getIntrinsic(ID, ResultType); |
| 17705 | return Builder.CreateCall(F, {X, Y}); |
| 17706 | } |
| 17707 | } |
| 17708 | switch (BuiltinID) { |
| 17709 | case SystemZ::BI__builtin_s390_vfmaxsb: ID = Intrinsic::s390_vfmaxsb; break; |
| 17710 | case SystemZ::BI__builtin_s390_vfmaxdb: ID = Intrinsic::s390_vfmaxdb; break; |
| 17711 | default: llvm_unreachable("Unknown BuiltinID")::llvm::llvm_unreachable_internal("Unknown BuiltinID", "clang/lib/CodeGen/CGBuiltin.cpp" , 17711); |
| 17712 | } |
| 17713 | Function *F = CGM.getIntrinsic(ID); |
| 17714 | Value *M4Value = llvm::ConstantInt::get(getLLVMContext(), M4); |
| 17715 | return Builder.CreateCall(F, {X, Y, M4Value}); |
| 17716 | } |
| 17717 | case SystemZ::BI__builtin_s390_vfminsb: |
| 17718 | case SystemZ::BI__builtin_s390_vfmindb: { |
| 17719 | llvm::Type *ResultType = ConvertType(E->getType()); |
| 17720 | Value *X = EmitScalarExpr(E->getArg(0)); |
| 17721 | Value *Y = EmitScalarExpr(E->getArg(1)); |
| 17722 | // Constant-fold the M4 mask argument. |
| 17723 | llvm::APSInt M4 = *E->getArg(2)->getIntegerConstantExpr(getContext()); |
| 17724 | // Check whether this instance can be represented via a LLVM standard |
| 17725 | // intrinsic. We only support some values of M4. |
| 17726 | Intrinsic::ID ID = Intrinsic::not_intrinsic; |
| 17727 | Intrinsic::ID CI; |
| 17728 | switch (M4.getZExtValue()) { |
| 17729 | default: break; |
| 17730 | case 4: ID = Intrinsic::minnum; |
| 17731 | CI = Intrinsic::experimental_constrained_minnum; break; |
| 17732 | } |
| 17733 | if (ID != Intrinsic::not_intrinsic) { |
| 17734 | if (Builder.getIsFPConstrained()) { |
| 17735 | Function *F = CGM.getIntrinsic(CI, ResultType); |
| 17736 | return Builder.CreateConstrainedFPCall(F, {X, Y}); |
| 17737 | } else { |
| 17738 | Function *F = CGM.getIntrinsic(ID, ResultType); |
| 17739 | return Builder.CreateCall(F, {X, Y}); |
| 17740 | } |
| 17741 | } |
| 17742 | switch (BuiltinID) { |
| 17743 | case SystemZ::BI__builtin_s390_vfminsb: ID = Intrinsic::s390_vfminsb; break; |
| 17744 | case SystemZ::BI__builtin_s390_vfmindb: ID = Intrinsic::s390_vfmindb; break; |
| 17745 | default: llvm_unreachable("Unknown BuiltinID")::llvm::llvm_unreachable_internal("Unknown BuiltinID", "clang/lib/CodeGen/CGBuiltin.cpp" , 17745); |
| 17746 | } |
| 17747 | Function *F = CGM.getIntrinsic(ID); |
| 17748 | Value *M4Value = llvm::ConstantInt::get(getLLVMContext(), M4); |
| 17749 | return Builder.CreateCall(F, {X, Y, M4Value}); |
| 17750 | } |
| 17751 | |
| 17752 | case SystemZ::BI__builtin_s390_vlbrh: |
| 17753 | case SystemZ::BI__builtin_s390_vlbrf: |
| 17754 | case SystemZ::BI__builtin_s390_vlbrg: { |
| 17755 | llvm::Type *ResultType = ConvertType(E->getType()); |
| 17756 | Value *X = EmitScalarExpr(E->getArg(0)); |
| 17757 | Function *F = CGM.getIntrinsic(Intrinsic::bswap, ResultType); |
| 17758 | return Builder.CreateCall(F, X); |
| 17759 | } |
| 17760 | |
| 17761 | // Vector intrinsics that output the post-instruction CC value. |
| 17762 | |
| 17763 | #define INTRINSIC_WITH_CC(NAME) \ |
| 17764 | case SystemZ::BI__builtin_##NAME: \ |
| 17765 | return EmitSystemZIntrinsicWithCC(*this, Intrinsic::NAME, E) |
| 17766 | |
| 17767 | INTRINSIC_WITH_CC(s390_vpkshs); |
| 17768 | INTRINSIC_WITH_CC(s390_vpksfs); |
| 17769 | INTRINSIC_WITH_CC(s390_vpksgs); |
| 17770 | |
| 17771 | INTRINSIC_WITH_CC(s390_vpklshs); |
| 17772 | INTRINSIC_WITH_CC(s390_vpklsfs); |
| 17773 | INTRINSIC_WITH_CC(s390_vpklsgs); |
| 17774 | |
| 17775 | INTRINSIC_WITH_CC(s390_vceqbs); |
| 17776 | INTRINSIC_WITH_CC(s390_vceqhs); |
| 17777 | INTRINSIC_WITH_CC(s390_vceqfs); |
| 17778 | INTRINSIC_WITH_CC(s390_vceqgs); |
| 17779 | |
| 17780 | INTRINSIC_WITH_CC(s390_vchbs); |
| 17781 | INTRINSIC_WITH_CC(s390_vchhs); |
| 17782 | INTRINSIC_WITH_CC(s390_vchfs); |
| 17783 | INTRINSIC_WITH_CC(s390_vchgs); |
| 17784 | |
| 17785 | INTRINSIC_WITH_CC(s390_vchlbs); |
| 17786 | INTRINSIC_WITH_CC(s390_vchlhs); |
| 17787 | INTRINSIC_WITH_CC(s390_vchlfs); |
| 17788 | INTRINSIC_WITH_CC(s390_vchlgs); |
| 17789 | |
| 17790 | INTRINSIC_WITH_CC(s390_vfaebs); |
| 17791 | INTRINSIC_WITH_CC(s390_vfaehs); |
| 17792 | INTRINSIC_WITH_CC(s390_vfaefs); |
| 17793 | |
| 17794 | INTRINSIC_WITH_CC(s390_vfaezbs); |
| 17795 | INTRINSIC_WITH_CC(s390_vfaezhs); |
| 17796 | INTRINSIC_WITH_CC(s390_vfaezfs); |
| 17797 | |
| 17798 | INTRINSIC_WITH_CC(s390_vfeebs); |
| 17799 | INTRINSIC_WITH_CC(s390_vfeehs); |
| 17800 | INTRINSIC_WITH_CC(s390_vfeefs); |
| 17801 | |
| 17802 | INTRINSIC_WITH_CC(s390_vfeezbs); |
| 17803 | INTRINSIC_WITH_CC(s390_vfeezhs); |
| 17804 | INTRINSIC_WITH_CC(s390_vfeezfs); |
| 17805 | |
| 17806 | INTRINSIC_WITH_CC(s390_vfenebs); |
| 17807 | INTRINSIC_WITH_CC(s390_vfenehs); |
| 17808 | INTRINSIC_WITH_CC(s390_vfenefs); |
| 17809 | |
| 17810 | INTRINSIC_WITH_CC(s390_vfenezbs); |
| 17811 | INTRINSIC_WITH_CC(s390_vfenezhs); |
| 17812 | INTRINSIC_WITH_CC(s390_vfenezfs); |
| 17813 | |
| 17814 | INTRINSIC_WITH_CC(s390_vistrbs); |
| 17815 | INTRINSIC_WITH_CC(s390_vistrhs); |
| 17816 | INTRINSIC_WITH_CC(s390_vistrfs); |
| 17817 | |
| 17818 | INTRINSIC_WITH_CC(s390_vstrcbs); |
| 17819 | INTRINSIC_WITH_CC(s390_vstrchs); |
| 17820 | INTRINSIC_WITH_CC(s390_vstrcfs); |
| 17821 | |
| 17822 | INTRINSIC_WITH_CC(s390_vstrczbs); |
| 17823 | INTRINSIC_WITH_CC(s390_vstrczhs); |
| 17824 | INTRINSIC_WITH_CC(s390_vstrczfs); |
| 17825 | |
| 17826 | INTRINSIC_WITH_CC(s390_vfcesbs); |
| 17827 | INTRINSIC_WITH_CC(s390_vfcedbs); |
| 17828 | INTRINSIC_WITH_CC(s390_vfchsbs); |
| 17829 | INTRINSIC_WITH_CC(s390_vfchdbs); |
| 17830 | INTRINSIC_WITH_CC(s390_vfchesbs); |
| 17831 | INTRINSIC_WITH_CC(s390_vfchedbs); |
| 17832 | |
| 17833 | INTRINSIC_WITH_CC(s390_vftcisb); |
| 17834 | INTRINSIC_WITH_CC(s390_vftcidb); |
| 17835 | |
| 17836 | INTRINSIC_WITH_CC(s390_vstrsb); |
| 17837 | INTRINSIC_WITH_CC(s390_vstrsh); |
| 17838 | INTRINSIC_WITH_CC(s390_vstrsf); |
| 17839 | |
| 17840 | INTRINSIC_WITH_CC(s390_vstrszb); |
| 17841 | INTRINSIC_WITH_CC(s390_vstrszh); |
| 17842 | INTRINSIC_WITH_CC(s390_vstrszf); |
| 17843 | |
| 17844 | #undef INTRINSIC_WITH_CC |
| 17845 | |
| 17846 | default: |
| 17847 | return nullptr; |
| 17848 | } |
| 17849 | } |
| 17850 | |
| 17851 | namespace { |
| 17852 | // Helper classes for mapping MMA builtins to particular LLVM intrinsic variant. |
| 17853 | struct NVPTXMmaLdstInfo { |
| 17854 | unsigned NumResults; // Number of elements to load/store |
| 17855 | // Intrinsic IDs for row/col variants. 0 if particular layout is unsupported. |
| 17856 | unsigned IID_col; |
| 17857 | unsigned IID_row; |
| 17858 | }; |
| 17859 | |
| 17860 | #define MMA_INTR(geom_op_type, layout) \ |
| 17861 | Intrinsic::nvvm_wmma_##geom_op_type##_##layout##_stride |
| 17862 | #define MMA_LDST(n, geom_op_type) \ |
| 17863 | { n, MMA_INTR(geom_op_type, col), MMA_INTR(geom_op_type, row) } |
| 17864 | |
| 17865 | static NVPTXMmaLdstInfo getNVPTXMmaLdstInfo(unsigned BuiltinID) { |
| 17866 | switch (BuiltinID) { |
| 17867 | // FP MMA loads |
| 17868 | case NVPTX::BI__hmma_m16n16k16_ld_a: |
| 17869 | return MMA_LDST(8, m16n16k16_load_a_f16); |
| 17870 | case NVPTX::BI__hmma_m16n16k16_ld_b: |
| 17871 | return MMA_LDST(8, m16n16k16_load_b_f16); |
| 17872 | case NVPTX::BI__hmma_m16n16k16_ld_c_f16: |
| 17873 | return MMA_LDST(4, m16n16k16_load_c_f16); |
| 17874 | case NVPTX::BI__hmma_m16n16k16_ld_c_f32: |
| 17875 | return MMA_LDST(8, m16n16k16_load_c_f32); |
| 17876 | case NVPTX::BI__hmma_m32n8k16_ld_a: |
| 17877 | return MMA_LDST(8, m32n8k16_load_a_f16); |
| 17878 | case NVPTX::BI__hmma_m32n8k16_ld_b: |
| 17879 | return MMA_LDST(8, m32n8k16_load_b_f16); |
| 17880 | case NVPTX::BI__hmma_m32n8k16_ld_c_f16: |
| 17881 | return MMA_LDST(4, m32n8k16_load_c_f16); |
| 17882 | case NVPTX::BI__hmma_m32n8k16_ld_c_f32: |
| 17883 | return MMA_LDST(8, m32n8k16_load_c_f32); |
| 17884 | case NVPTX::BI__hmma_m8n32k16_ld_a: |
| 17885 | return MMA_LDST(8, m8n32k16_load_a_f16); |
| 17886 | case NVPTX::BI__hmma_m8n32k16_ld_b: |
| 17887 | return MMA_LDST(8, m8n32k16_load_b_f16); |
| 17888 | case NVPTX::BI__hmma_m8n32k16_ld_c_f16: |
| 17889 | return MMA_LDST(4, m8n32k16_load_c_f16); |
| 17890 | case NVPTX::BI__hmma_m8n32k16_ld_c_f32: |
| 17891 | return MMA_LDST(8, m8n32k16_load_c_f32); |
| 17892 | |
| 17893 | // Integer MMA loads |
| 17894 | case NVPTX::BI__imma_m16n16k16_ld_a_s8: |
| 17895 | return MMA_LDST(2, m16n16k16_load_a_s8); |
| 17896 | case NVPTX::BI__imma_m16n16k16_ld_a_u8: |
| 17897 | return MMA_LDST(2, m16n16k16_load_a_u8); |
| 17898 | case NVPTX::BI__imma_m16n16k16_ld_b_s8: |
| 17899 | return MMA_LDST(2, m16n16k16_load_b_s8); |
| 17900 | case NVPTX::BI__imma_m16n16k16_ld_b_u8: |
| 17901 | return MMA_LDST(2, m16n16k16_load_b_u8); |
| 17902 | case NVPTX::BI__imma_m16n16k16_ld_c: |
| 17903 | return MMA_LDST(8, m16n16k16_load_c_s32); |
| 17904 | case NVPTX::BI__imma_m32n8k16_ld_a_s8: |
| 17905 | return MMA_LDST(4, m32n8k16_load_a_s8); |
| 17906 | case NVPTX::BI__imma_m32n8k16_ld_a_u8: |
| 17907 | return MMA_LDST(4, m32n8k16_load_a_u8); |
| 17908 | case NVPTX::BI__imma_m32n8k16_ld_b_s8: |
| 17909 | return MMA_LDST(1, m32n8k16_load_b_s8); |
| 17910 | case NVPTX::BI__imma_m32n8k16_ld_b_u8: |
| 17911 | return MMA_LDST(1, m32n8k16_load_b_u8); |
| 17912 | case NVPTX::BI__imma_m32n8k16_ld_c: |
| 17913 | return MMA_LDST(8, m32n8k16_load_c_s32); |
| 17914 | case NVPTX::BI__imma_m8n32k16_ld_a_s8: |
| 17915 | return MMA_LDST(1, m8n32k16_load_a_s8); |
| 17916 | case NVPTX::BI__imma_m8n32k16_ld_a_u8: |
| 17917 | return MMA_LDST(1, m8n32k16_load_a_u8); |
| 17918 | case NVPTX::BI__imma_m8n32k16_ld_b_s8: |
| 17919 | return MMA_LDST(4, m8n32k16_load_b_s8); |
| 17920 | case NVPTX::BI__imma_m8n32k16_ld_b_u8: |
| 17921 | return MMA_LDST(4, m8n32k16_load_b_u8); |
| 17922 | case NVPTX::BI__imma_m8n32k16_ld_c: |
| 17923 | return MMA_LDST(8, m8n32k16_load_c_s32); |
| 17924 | |
| 17925 | // Sub-integer MMA loads. |
| 17926 | // Only row/col layout is supported by A/B fragments. |
| 17927 | case NVPTX::BI__imma_m8n8k32_ld_a_s4: |
| 17928 | return {1, 0, MMA_INTR(m8n8k32_load_a_s4, row)}; |
| 17929 | case NVPTX::BI__imma_m8n8k32_ld_a_u4: |
| 17930 | return {1, 0, MMA_INTR(m8n8k32_load_a_u4, row)}; |
| 17931 | case NVPTX::BI__imma_m8n8k32_ld_b_s4: |
| 17932 | return {1, MMA_INTR(m8n8k32_load_b_s4, col), 0}; |
| 17933 | case NVPTX::BI__imma_m8n8k32_ld_b_u4: |
| 17934 | return {1, MMA_INTR(m8n8k32_load_b_u4, col), 0}; |
| 17935 | case NVPTX::BI__imma_m8n8k32_ld_c: |
| 17936 | return MMA_LDST(2, m8n8k32_load_c_s32); |
| 17937 | case NVPTX::BI__bmma_m8n8k128_ld_a_b1: |
| 17938 | return {1, 0, MMA_INTR(m8n8k128_load_a_b1, row)}; |
| 17939 | case NVPTX::BI__bmma_m8n8k128_ld_b_b1: |
| 17940 | return {1, MMA_INTR(m8n8k128_load_b_b1, col), 0}; |
| 17941 | case NVPTX::BI__bmma_m8n8k128_ld_c: |
| 17942 | return MMA_LDST(2, m8n8k128_load_c_s32); |
| 17943 | |
| 17944 | // Double MMA loads |
| 17945 | case NVPTX::BI__dmma_m8n8k4_ld_a: |
| 17946 | return MMA_LDST(1, m8n8k4_load_a_f64); |
| 17947 | case NVPTX::BI__dmma_m8n8k4_ld_b: |
| 17948 | return MMA_LDST(1, m8n8k4_load_b_f64); |
| 17949 | case NVPTX::BI__dmma_m8n8k4_ld_c: |
| 17950 | return MMA_LDST(2, m8n8k4_load_c_f64); |
| 17951 | |
| 17952 | // Alternate float MMA loads |
| 17953 | case NVPTX::BI__mma_bf16_m16n16k16_ld_a: |
| 17954 | return MMA_LDST(4, m16n16k16_load_a_bf16); |
| 17955 | case NVPTX::BI__mma_bf16_m16n16k16_ld_b: |
| 17956 | return MMA_LDST(4, m16n16k16_load_b_bf16); |
| 17957 | case NVPTX::BI__mma_bf16_m8n32k16_ld_a: |
| 17958 | return MMA_LDST(2, m8n32k16_load_a_bf16); |
| 17959 | case NVPTX::BI__mma_bf16_m8n32k16_ld_b: |
| 17960 | return MMA_LDST(8, m8n32k16_load_b_bf16); |
| 17961 | case NVPTX::BI__mma_bf16_m32n8k16_ld_a: |
| 17962 | return MMA_LDST(8, m32n8k16_load_a_bf16); |
| 17963 | case NVPTX::BI__mma_bf16_m32n8k16_ld_b: |
| 17964 | return MMA_LDST(2, m32n8k16_load_b_bf16); |
| 17965 | case NVPTX::BI__mma_tf32_m16n16k8_ld_a: |
| 17966 | return MMA_LDST(4, m16n16k8_load_a_tf32); |
| 17967 | case NVPTX::BI__mma_tf32_m16n16k8_ld_b: |
| 17968 | return MMA_LDST(4, m16n16k8_load_b_tf32); |
| 17969 | case NVPTX::BI__mma_tf32_m16n16k8_ld_c: |
| 17970 | return MMA_LDST(8, m16n16k8_load_c_f32); |
| 17971 | |
| 17972 | // NOTE: We need to follow inconsitent naming scheme used by NVCC. Unlike |
| 17973 | // PTX and LLVM IR where stores always use fragment D, NVCC builtins always |
| 17974 | // use fragment C for both loads and stores. |
| 17975 | // FP MMA stores. |
| 17976 | case NVPTX::BI__hmma_m16n16k16_st_c_f16: |
| 17977 | return MMA_LDST(4, m16n16k16_store_d_f16); |
| 17978 | case NVPTX::BI__hmma_m16n16k16_st_c_f32: |
| 17979 | return MMA_LDST(8, m16n16k16_store_d_f32); |
| 17980 | case NVPTX::BI__hmma_m32n8k16_st_c_f16: |
| 17981 | return MMA_LDST(4, m32n8k16_store_d_f16); |
| 17982 | case NVPTX::BI__hmma_m32n8k16_st_c_f32: |
| 17983 | return MMA_LDST(8, m32n8k16_store_d_f32); |
| 17984 | case NVPTX::BI__hmma_m8n32k16_st_c_f16: |
| 17985 | return MMA_LDST(4, m8n32k16_store_d_f16); |
| 17986 | case NVPTX::BI__hmma_m8n32k16_st_c_f32: |
| 17987 | return MMA_LDST(8, m8n32k16_store_d_f32); |
| 17988 | |
| 17989 | // Integer and sub-integer MMA stores. |
| 17990 | // Another naming quirk. Unlike other MMA builtins that use PTX types in the |
| 17991 | // name, integer loads/stores use LLVM's i32. |
| 17992 | case NVPTX::BI__imma_m16n16k16_st_c_i32: |
| 17993 | return MMA_LDST(8, m16n16k16_store_d_s32); |
| 17994 | case NVPTX::BI__imma_m32n8k16_st_c_i32: |
| 17995 | return MMA_LDST(8, m32n8k16_store_d_s32); |
| 17996 | case NVPTX::BI__imma_m8n32k16_st_c_i32: |
| 17997 | return MMA_LDST(8, m8n32k16_store_d_s32); |
| 17998 | case NVPTX::BI__imma_m8n8k32_st_c_i32: |
| 17999 | return MMA_LDST(2, m8n8k32_store_d_s32); |
| 18000 | case NVPTX::BI__bmma_m8n8k128_st_c_i32: |
| 18001 | return MMA_LDST(2, m8n8k128_store_d_s32); |
| 18002 | |
| 18003 | // Double MMA store |
| 18004 | case NVPTX::BI__dmma_m8n8k4_st_c_f64: |
| 18005 | return MMA_LDST(2, m8n8k4_store_d_f64); |
| 18006 | |
| 18007 | // Alternate float MMA store |
| 18008 | case NVPTX::BI__mma_m16n16k8_st_c_f32: |
| 18009 | return MMA_LDST(8, m16n16k8_store_d_f32); |
| 18010 | |
| 18011 | default: |
| 18012 | llvm_unreachable("Unknown MMA builtin")::llvm::llvm_unreachable_internal("Unknown MMA builtin", "clang/lib/CodeGen/CGBuiltin.cpp" , 18012); |
| 18013 | } |
| 18014 | } |
| 18015 | #undef MMA_LDST |
| 18016 | #undef MMA_INTR |
| 18017 | |
| 18018 | |
| 18019 | struct NVPTXMmaInfo { |
| 18020 | unsigned NumEltsA; |
| 18021 | unsigned NumEltsB; |
| 18022 | unsigned NumEltsC; |
| 18023 | unsigned NumEltsD; |
| 18024 | |
| 18025 | // Variants are ordered by layout-A/layout-B/satf, where 'row' has priority |
| 18026 | // over 'col' for layout. The index of non-satf variants is expected to match |
| 18027 | // the undocumented layout constants used by CUDA's mma.hpp. |
| 18028 | std::array<unsigned, 8> Variants; |
| 18029 | |
| 18030 | unsigned getMMAIntrinsic(int Layout, bool Satf) { |
| 18031 | unsigned Index = Layout + 4 * Satf; |
| 18032 | if (Index >= Variants.size()) |
| 18033 | return 0; |
| 18034 | return Variants[Index]; |
| 18035 | } |
| 18036 | }; |
| 18037 | |
| 18038 | // Returns an intrinsic that matches Layout and Satf for valid combinations of |
| 18039 | // Layout and Satf, 0 otherwise. |
| 18040 | static NVPTXMmaInfo getNVPTXMmaInfo(unsigned BuiltinID) { |
| 18041 | // clang-format off |
| 18042 | #define MMA_VARIANTS(geom, type) \ |
| 18043 | Intrinsic::nvvm_wmma_##geom##_mma_row_row_##type, \ |
| 18044 | Intrinsic::nvvm_wmma_##geom##_mma_row_col_##type, \ |
| 18045 | Intrinsic::nvvm_wmma_##geom##_mma_col_row_##type, \ |
| 18046 | Intrinsic::nvvm_wmma_##geom##_mma_col_col_##type |
| 18047 | #define MMA_SATF_VARIANTS(geom, type) \ |
| 18048 | MMA_VARIANTS(geom, type), \ |
| 18049 | Intrinsic::nvvm_wmma_##geom##_mma_row_row_##type##_satfinite, \ |
| 18050 | Intrinsic::nvvm_wmma_##geom##_mma_row_col_##type##_satfinite, \ |
| 18051 | Intrinsic::nvvm_wmma_##geom##_mma_col_row_##type##_satfinite, \ |
| 18052 | Intrinsic::nvvm_wmma_##geom##_mma_col_col_##type##_satfinite |
| 18053 | // Sub-integer MMA only supports row.col layout. |
| 18054 | #define MMA_VARIANTS_I4(geom, type) \ |
| 18055 | 0, \ |
| 18056 | Intrinsic::nvvm_wmma_##geom##_mma_row_col_##type, \ |
| 18057 | 0, \ |
| 18058 | 0, \ |
| 18059 | 0, \ |
| 18060 | Intrinsic::nvvm_wmma_##geom##_mma_row_col_##type##_satfinite, \ |
| 18061 | 0, \ |
| 18062 | 0 |
| 18063 | // b1 MMA does not support .satfinite. |
| 18064 | #define MMA_VARIANTS_B1_XOR(geom, type) \ |
| 18065 | 0, \ |
| 18066 | Intrinsic::nvvm_wmma_##geom##_mma_xor_popc_row_col_##type, \ |
| 18067 | 0, \ |
| 18068 | 0, \ |
| 18069 | 0, \ |
| 18070 | 0, \ |
| 18071 | 0, \ |
| 18072 | 0 |
| 18073 | #define MMA_VARIANTS_B1_AND(geom, type) \ |
| 18074 | 0, \ |
| 18075 | Intrinsic::nvvm_wmma_##geom##_mma_and_popc_row_col_##type, \ |
| 18076 | 0, \ |
| 18077 | 0, \ |
| 18078 | 0, \ |
| 18079 | 0, \ |
| 18080 | 0, \ |
| 18081 | 0 |
| 18082 | // clang-format on |
| 18083 | switch (BuiltinID) { |
| 18084 | // FP MMA |
| 18085 | // Note that 'type' argument of MMA_SATF_VARIANTS uses D_C notation, while |
| 18086 | // NumEltsN of return value are ordered as A,B,C,D. |
| 18087 | case NVPTX::BI__hmma_m16n16k16_mma_f16f16: |
| 18088 | return {8, 8, 4, 4, {{MMA_SATF_VARIANTS(m16n16k16, f16_f16)}}}; |
| 18089 | case NVPTX::BI__hmma_m16n16k16_mma_f32f16: |
| 18090 | return {8, 8, 4, 8, {{MMA_SATF_VARIANTS(m16n16k16, f32_f16)}}}; |
| 18091 | case NVPTX::BI__hmma_m16n16k16_mma_f16f32: |
| 18092 | return {8, 8, 8, 4, {{MMA_SATF_VARIANTS(m16n16k16, f16_f32)}}}; |
| 18093 | case NVPTX::BI__hmma_m16n16k16_mma_f32f32: |
| 18094 | return {8, 8, 8, 8, {{MMA_SATF_VARIANTS(m16n16k16, f32_f32)}}}; |
| 18095 | case NVPTX::BI__hmma_m32n8k16_mma_f16f16: |
| 18096 | return {8, 8, 4, 4, {{MMA_SATF_VARIANTS(m32n8k16, f16_f16)}}}; |
| 18097 | case NVPTX::BI__hmma_m32n8k16_mma_f32f16: |
| 18098 | return {8, 8, 4, 8, {{MMA_SATF_VARIANTS(m32n8k16, f32_f16)}}}; |
| 18099 | case NVPTX::BI__hmma_m32n8k16_mma_f16f32: |
| 18100 | return {8, 8, 8, 4, {{MMA_SATF_VARIANTS(m32n8k16, f16_f32)}}}; |
| 18101 | case NVPTX::BI__hmma_m32n8k16_mma_f32f32: |
| 18102 | return {8, 8, 8, 8, {{MMA_SATF_VARIANTS(m32n8k16, f32_f32)}}}; |
| 18103 | case NVPTX::BI__hmma_m8n32k16_mma_f16f16: |
| 18104 | return {8, 8, 4, 4, {{MMA_SATF_VARIANTS(m8n32k16, f16_f16)}}}; |
| 18105 | case NVPTX::BI__hmma_m8n32k16_mma_f32f16: |
| 18106 | return {8, 8, 4, 8, {{MMA_SATF_VARIANTS(m8n32k16, f32_f16)}}}; |
| 18107 | case NVPTX::BI__hmma_m8n32k16_mma_f16f32: |
| 18108 | return {8, 8, 8, 4, {{MMA_SATF_VARIANTS(m8n32k16, f16_f32)}}}; |
| 18109 | case NVPTX::BI__hmma_m8n32k16_mma_f32f32: |
| 18110 | return {8, 8, 8, 8, {{MMA_SATF_VARIANTS(m8n32k16, f32_f32)}}}; |
| 18111 | |
| 18112 | // Integer MMA |
| 18113 | case NVPTX::BI__imma_m16n16k16_mma_s8: |
| 18114 | return {2, 2, 8, 8, {{MMA_SATF_VARIANTS(m16n16k16, s8)}}}; |
| 18115 | case NVPTX::BI__imma_m16n16k16_mma_u8: |
| 18116 | return {2, 2, 8, 8, {{MMA_SATF_VARIANTS(m16n16k16, u8)}}}; |
| 18117 | case NVPTX::BI__imma_m32n8k16_mma_s8: |
| 18118 | return {4, 1, 8, 8, {{MMA_SATF_VARIANTS(m32n8k16, s8)}}}; |
| 18119 | case NVPTX::BI__imma_m32n8k16_mma_u8: |
| 18120 | return {4, 1, 8, 8, {{MMA_SATF_VARIANTS(m32n8k16, u8)}}}; |
| 18121 | case NVPTX::BI__imma_m8n32k16_mma_s8: |
| 18122 | return {1, 4, 8, 8, {{MMA_SATF_VARIANTS(m8n32k16, s8)}}}; |
| 18123 | case NVPTX::BI__imma_m8n32k16_mma_u8: |
| 18124 | return {1, 4, 8, 8, {{MMA_SATF_VARIANTS(m8n32k16, u8)}}}; |
| 18125 | |
| 18126 | // Sub-integer MMA |
| 18127 | case NVPTX::BI__imma_m8n8k32_mma_s4: |
| 18128 | return {1, 1, 2, 2, {{MMA_VARIANTS_I4(m8n8k32, s4)}}}; |
| 18129 | case NVPTX::BI__imma_m8n8k32_mma_u4: |
| 18130 | return {1, 1, 2, 2, {{MMA_VARIANTS_I4(m8n8k32, u4)}}}; |
| 18131 | case NVPTX::BI__bmma_m8n8k128_mma_xor_popc_b1: |
| 18132 | return {1, 1, 2, 2, {{MMA_VARIANTS_B1_XOR(m8n8k128, b1)}}}; |
| 18133 | case NVPTX::BI__bmma_m8n8k128_mma_and_popc_b1: |
| 18134 | return {1, 1, 2, 2, {{MMA_VARIANTS_B1_AND(m8n8k128, b1)}}}; |
| 18135 | |
| 18136 | // Double MMA |
| 18137 | case NVPTX::BI__dmma_m8n8k4_mma_f64: |
| 18138 | return {1, 1, 2, 2, {{MMA_VARIANTS(m8n8k4, f64)}}}; |
| 18139 | |
| 18140 | // Alternate FP MMA |
| 18141 | case NVPTX::BI__mma_bf16_m16n16k16_mma_f32: |
| 18142 | return {4, 4, 8, 8, {{MMA_VARIANTS(m16n16k16, bf16)}}}; |
| 18143 | case NVPTX::BI__mma_bf16_m8n32k16_mma_f32: |
| 18144 | return {2, 8, 8, 8, {{MMA_VARIANTS(m8n32k16, bf16)}}}; |
| 18145 | case NVPTX::BI__mma_bf16_m32n8k16_mma_f32: |
| 18146 | return {8, 2, 8, 8, {{MMA_VARIANTS(m32n8k16, bf16)}}}; |
| 18147 | case NVPTX::BI__mma_tf32_m16n16k8_mma_f32: |
| 18148 | return {4, 4, 8, 8, {{MMA_VARIANTS(m16n16k8, tf32)}}}; |
| 18149 | default: |
| 18150 | llvm_unreachable("Unexpected builtin ID.")::llvm::llvm_unreachable_internal("Unexpected builtin ID.", "clang/lib/CodeGen/CGBuiltin.cpp" , 18150); |
| 18151 | } |
| 18152 | #undef MMA_VARIANTS |
| 18153 | #undef MMA_SATF_VARIANTS |
| 18154 | #undef MMA_VARIANTS_I4 |
| 18155 | #undef MMA_VARIANTS_B1_AND |
| 18156 | #undef MMA_VARIANTS_B1_XOR |
| 18157 | } |
| 18158 | |
| 18159 | static Value *MakeLdgLdu(unsigned IntrinsicID, CodeGenFunction &CGF, |
| 18160 | const CallExpr *E) { |
| 18161 | Value *Ptr = CGF.EmitScalarExpr(E->getArg(0)); |
| 18162 | QualType ArgType = E->getArg(0)->getType(); |
| 18163 | clang::CharUnits Align = CGF.CGM.getNaturalPointeeTypeAlignment(ArgType); |
| 18164 | llvm::Type *ElemTy = CGF.ConvertTypeForMem(ArgType->getPointeeType()); |
| 18165 | return CGF.Builder.CreateCall( |
| 18166 | CGF.CGM.getIntrinsic(IntrinsicID, {ElemTy, Ptr->getType()}), |
| 18167 | {Ptr, ConstantInt::get(CGF.Builder.getInt32Ty(), Align.getQuantity())}); |
| 18168 | } |
| 18169 | |
| 18170 | static Value *MakeScopedAtomic(unsigned IntrinsicID, CodeGenFunction &CGF, |
| 18171 | const CallExpr *E) { |
| 18172 | Value *Ptr = CGF.EmitScalarExpr(E->getArg(0)); |
| 18173 | llvm::Type *ElemTy = |
| 18174 | CGF.ConvertTypeForMem(E->getArg(0)->getType()->getPointeeType()); |
| 18175 | return CGF.Builder.CreateCall( |
| 18176 | CGF.CGM.getIntrinsic(IntrinsicID, {ElemTy, Ptr->getType()}), |
| 18177 | {Ptr, CGF.EmitScalarExpr(E->getArg(1))}); |
| 18178 | } |
| 18179 | |
| 18180 | static Value *MakeHalfType(unsigned IntrinsicID, unsigned BuiltinID, |
| 18181 | const CallExpr *E, CodeGenFunction &CGF) { |
| 18182 | auto &C = CGF.CGM.getContext(); |
| 18183 | if (!(C.getLangOpts().NativeHalfType || |
| 18184 | !C.getTargetInfo().useFP16ConversionIntrinsics())) { |
| 18185 | CGF.CGM.Error(E->getExprLoc(), C.BuiltinInfo.getName(BuiltinID).str() + |
| 18186 | " requires native half type support."); |
| 18187 | return nullptr; |
| 18188 | } |
| 18189 | |
| 18190 | if (IntrinsicID == Intrinsic::nvvm_ldg_global_f || |
| 18191 | IntrinsicID == Intrinsic::nvvm_ldu_global_f) |
| 18192 | return MakeLdgLdu(IntrinsicID, CGF, E); |
| 18193 | |
| 18194 | SmallVector<Value *, 16> Args; |
| 18195 | auto *F = CGF.CGM.getIntrinsic(IntrinsicID); |
| 18196 | auto *FTy = F->getFunctionType(); |
| 18197 | unsigned ICEArguments = 0; |
| 18198 | ASTContext::GetBuiltinTypeError Error; |
| 18199 | C.GetBuiltinType(BuiltinID, Error, &ICEArguments); |
| 18200 | assert(Error == ASTContext::GE_None && "Should not codegen an error")(static_cast <bool> (Error == ASTContext::GE_None && "Should not codegen an error") ? void (0) : __assert_fail ("Error == ASTContext::GE_None && \"Should not codegen an error\"" , "clang/lib/CodeGen/CGBuiltin.cpp", 18200, __extension__ __PRETTY_FUNCTION__ )); |
| 18201 | for (unsigned i = 0, e = E->getNumArgs(); i != e; ++i) { |
| 18202 | assert((ICEArguments & (1 << i)) == 0)(static_cast <bool> ((ICEArguments & (1 << i) ) == 0) ? void (0) : __assert_fail ("(ICEArguments & (1 << i)) == 0" , "clang/lib/CodeGen/CGBuiltin.cpp", 18202, __extension__ __PRETTY_FUNCTION__ )); |
| 18203 | auto *ArgValue = CGF.EmitScalarExpr(E->getArg(i)); |
| 18204 | auto *PTy = FTy->getParamType(i); |
| 18205 | if (PTy != ArgValue->getType()) |
| 18206 | ArgValue = CGF.Builder.CreateBitCast(ArgValue, PTy); |
| 18207 | Args.push_back(ArgValue); |
| 18208 | } |
| 18209 | |
| 18210 | return CGF.Builder.CreateCall(F, Args); |
| 18211 | } |
| 18212 | } // namespace |
| 18213 | |
| 18214 | Value *CodeGenFunction::EmitNVPTXBuiltinExpr(unsigned BuiltinID, |
| 18215 | const CallExpr *E) { |
| 18216 | switch (BuiltinID) { |
| 18217 | case NVPTX::BI__nvvm_atom_add_gen_i: |
| 18218 | case NVPTX::BI__nvvm_atom_add_gen_l: |
| 18219 | case NVPTX::BI__nvvm_atom_add_gen_ll: |
| 18220 | return MakeBinaryAtomicValue(*this, llvm::AtomicRMWInst::Add, E); |
| 18221 | |
| 18222 | case NVPTX::BI__nvvm_atom_sub_gen_i: |
| 18223 | case NVPTX::BI__nvvm_atom_sub_gen_l: |
| 18224 | case NVPTX::BI__nvvm_atom_sub_gen_ll: |
| 18225 | return MakeBinaryAtomicValue(*this, llvm::AtomicRMWInst::Sub, E); |
| 18226 | |
| 18227 | case NVPTX::BI__nvvm_atom_and_gen_i: |
| 18228 | case NVPTX::BI__nvvm_atom_and_gen_l: |
| 18229 | case NVPTX::BI__nvvm_atom_and_gen_ll: |
| 18230 | return MakeBinaryAtomicValue(*this, llvm::AtomicRMWInst::And, E); |
| 18231 | |
| 18232 | case NVPTX::BI__nvvm_atom_or_gen_i: |
| 18233 | case NVPTX::BI__nvvm_atom_or_gen_l: |
| 18234 | case NVPTX::BI__nvvm_atom_or_gen_ll: |
| 18235 | return MakeBinaryAtomicValue(*this, llvm::AtomicRMWInst::Or, E); |
| 18236 | |
| 18237 | case NVPTX::BI__nvvm_atom_xor_gen_i: |
| 18238 | case NVPTX::BI__nvvm_atom_xor_gen_l: |
| 18239 | case NVPTX::BI__nvvm_atom_xor_gen_ll: |
| 18240 | return MakeBinaryAtomicValue(*this, llvm::AtomicRMWInst::Xor, E); |
| 18241 | |
| 18242 | case NVPTX::BI__nvvm_atom_xchg_gen_i: |
| 18243 | case NVPTX::BI__nvvm_atom_xchg_gen_l: |
| 18244 | case NVPTX::BI__nvvm_atom_xchg_gen_ll: |
| 18245 | return MakeBinaryAtomicValue(*this, llvm::AtomicRMWInst::Xchg, E); |
| 18246 | |
| 18247 | case NVPTX::BI__nvvm_atom_max_gen_i: |
| 18248 | case NVPTX::BI__nvvm_atom_max_gen_l: |
| 18249 | case NVPTX::BI__nvvm_atom_max_gen_ll: |
| 18250 | return MakeBinaryAtomicValue(*this, llvm::AtomicRMWInst::Max, E); |
| 18251 | |
| 18252 | case NVPTX::BI__nvvm_atom_max_gen_ui: |
| 18253 | case NVPTX::BI__nvvm_atom_max_gen_ul: |
| 18254 | case NVPTX::BI__nvvm_atom_max_gen_ull: |
| 18255 | return MakeBinaryAtomicValue(*this, llvm::AtomicRMWInst::UMax, E); |
| 18256 | |
| 18257 | case NVPTX::BI__nvvm_atom_min_gen_i: |
| 18258 | case NVPTX::BI__nvvm_atom_min_gen_l: |
| 18259 | case NVPTX::BI__nvvm_atom_min_gen_ll: |
| 18260 | return MakeBinaryAtomicValue(*this, llvm::AtomicRMWInst::Min, E); |
| 18261 | |
| 18262 | case NVPTX::BI__nvvm_atom_min_gen_ui: |
| 18263 | case NVPTX::BI__nvvm_atom_min_gen_ul: |
| 18264 | case NVPTX::BI__nvvm_atom_min_gen_ull: |
| 18265 | return MakeBinaryAtomicValue(*this, llvm::AtomicRMWInst::UMin, E); |
| 18266 | |
| 18267 | case NVPTX::BI__nvvm_atom_cas_gen_i: |
| 18268 | case NVPTX::BI__nvvm_atom_cas_gen_l: |
| 18269 | case NVPTX::BI__nvvm_atom_cas_gen_ll: |
| 18270 | // __nvvm_atom_cas_gen_* should return the old value rather than the |
| 18271 | // success flag. |
| 18272 | return MakeAtomicCmpXchgValue(*this, E, /*ReturnBool=*/false); |
| 18273 | |
| 18274 | case NVPTX::BI__nvvm_atom_add_gen_f: |
| 18275 | case NVPTX::BI__nvvm_atom_add_gen_d: { |
| 18276 | Value *Ptr = EmitScalarExpr(E->getArg(0)); |
| 18277 | Value *Val = EmitScalarExpr(E->getArg(1)); |
| 18278 | return Builder.CreateAtomicRMW(llvm::AtomicRMWInst::FAdd, Ptr, Val, |
| 18279 | AtomicOrdering::SequentiallyConsistent); |
| 18280 | } |
| 18281 | |
| 18282 | case NVPTX::BI__nvvm_atom_inc_gen_ui: { |
| 18283 | Value *Ptr = EmitScalarExpr(E->getArg(0)); |
| 18284 | Value *Val = EmitScalarExpr(E->getArg(1)); |
| 18285 | Function *FnALI32 = |
| 18286 | CGM.getIntrinsic(Intrinsic::nvvm_atomic_load_inc_32, Ptr->getType()); |
| 18287 | return Builder.CreateCall(FnALI32, {Ptr, Val}); |
| 18288 | } |
| 18289 | |
| 18290 | case NVPTX::BI__nvvm_atom_dec_gen_ui: { |
| 18291 | Value *Ptr = EmitScalarExpr(E->getArg(0)); |
| 18292 | Value *Val = EmitScalarExpr(E->getArg(1)); |
| 18293 | Function *FnALD32 = |
| 18294 | CGM.getIntrinsic(Intrinsic::nvvm_atomic_load_dec_32, Ptr->getType()); |
| 18295 | return Builder.CreateCall(FnALD32, {Ptr, Val}); |
| 18296 | } |
| 18297 | |
| 18298 | case NVPTX::BI__nvvm_ldg_c: |
| 18299 | case NVPTX::BI__nvvm_ldg_c2: |
| 18300 | case NVPTX::BI__nvvm_ldg_c4: |
| 18301 | case NVPTX::BI__nvvm_ldg_s: |
| 18302 | case NVPTX::BI__nvvm_ldg_s2: |
| 18303 | case NVPTX::BI__nvvm_ldg_s4: |
| 18304 | case NVPTX::BI__nvvm_ldg_i: |
| 18305 | case NVPTX::BI__nvvm_ldg_i2: |
| 18306 | case NVPTX::BI__nvvm_ldg_i4: |
| 18307 | case NVPTX::BI__nvvm_ldg_l: |
| 18308 | case NVPTX::BI__nvvm_ldg_ll: |
| 18309 | case NVPTX::BI__nvvm_ldg_ll2: |
| 18310 | case NVPTX::BI__nvvm_ldg_uc: |
| 18311 | case NVPTX::BI__nvvm_ldg_uc2: |
| 18312 | case NVPTX::BI__nvvm_ldg_uc4: |
| 18313 | case NVPTX::BI__nvvm_ldg_us: |
| 18314 | case NVPTX::BI__nvvm_ldg_us2: |
| 18315 | case NVPTX::BI__nvvm_ldg_us4: |
| 18316 | case NVPTX::BI__nvvm_ldg_ui: |
| 18317 | case NVPTX::BI__nvvm_ldg_ui2: |
| 18318 | case NVPTX::BI__nvvm_ldg_ui4: |
| 18319 | case NVPTX::BI__nvvm_ldg_ul: |
| 18320 | case NVPTX::BI__nvvm_ldg_ull: |
| 18321 | case NVPTX::BI__nvvm_ldg_ull2: |
| 18322 | // PTX Interoperability section 2.2: "For a vector with an even number of |
| 18323 | // elements, its alignment is set to number of elements times the alignment |
| 18324 | // of its member: n*alignof(t)." |
| 18325 | return MakeLdgLdu(Intrinsic::nvvm_ldg_global_i, *this, E); |
| 18326 | case NVPTX::BI__nvvm_ldg_f: |
| 18327 | case NVPTX::BI__nvvm_ldg_f2: |
| 18328 | case NVPTX::BI__nvvm_ldg_f4: |
| 18329 | case NVPTX::BI__nvvm_ldg_d: |
| 18330 | case NVPTX::BI__nvvm_ldg_d2: |
| 18331 | return MakeLdgLdu(Intrinsic::nvvm_ldg_global_f, *this, E); |
| 18332 | |
| 18333 | case NVPTX::BI__nvvm_ldu_c: |
| 18334 | case NVPTX::BI__nvvm_ldu_c2: |
| 18335 | case NVPTX::BI__nvvm_ldu_c4: |
| 18336 | case NVPTX::BI__nvvm_ldu_s: |
| 18337 | case NVPTX::BI__nvvm_ldu_s2: |
| 18338 | case NVPTX::BI__nvvm_ldu_s4: |
| 18339 | case NVPTX::BI__nvvm_ldu_i: |
| 18340 | case NVPTX::BI__nvvm_ldu_i2: |
| 18341 | case NVPTX::BI__nvvm_ldu_i4: |
| 18342 | case NVPTX::BI__nvvm_ldu_l: |
| 18343 | case NVPTX::BI__nvvm_ldu_ll: |
| 18344 | case NVPTX::BI__nvvm_ldu_ll2: |
| 18345 | case NVPTX::BI__nvvm_ldu_uc: |
| 18346 | case NVPTX::BI__nvvm_ldu_uc2: |
| 18347 | case NVPTX::BI__nvvm_ldu_uc4: |
| 18348 | case NVPTX::BI__nvvm_ldu_us: |
| 18349 | case NVPTX::BI__nvvm_ldu_us2: |
| 18350 | case NVPTX::BI__nvvm_ldu_us4: |
| 18351 | case NVPTX::BI__nvvm_ldu_ui: |
| 18352 | case NVPTX::BI__nvvm_ldu_ui2: |
| 18353 | case NVPTX::BI__nvvm_ldu_ui4: |
| 18354 | case NVPTX::BI__nvvm_ldu_ul: |
| 18355 | case NVPTX::BI__nvvm_ldu_ull: |
| 18356 | case NVPTX::BI__nvvm_ldu_ull2: |
| 18357 | return MakeLdgLdu(Intrinsic::nvvm_ldu_global_i, *this, E); |
| 18358 | case NVPTX::BI__nvvm_ldu_f: |
| 18359 | case NVPTX::BI__nvvm_ldu_f2: |
| 18360 | case NVPTX::BI__nvvm_ldu_f4: |
| 18361 | case NVPTX::BI__nvvm_ldu_d: |
| 18362 | case NVPTX::BI__nvvm_ldu_d2: |
| 18363 | return MakeLdgLdu(Intrinsic::nvvm_ldu_global_f, *this, E); |
| 18364 | |
| 18365 | case NVPTX::BI__nvvm_atom_cta_add_gen_i: |
| 18366 | case NVPTX::BI__nvvm_atom_cta_add_gen_l: |
| 18367 | case NVPTX::BI__nvvm_atom_cta_add_gen_ll: |
| 18368 | return MakeScopedAtomic(Intrinsic::nvvm_atomic_add_gen_i_cta, *this, E); |
| 18369 | case NVPTX::BI__nvvm_atom_sys_add_gen_i: |
| 18370 | case NVPTX::BI__nvvm_atom_sys_add_gen_l: |
| 18371 | case NVPTX::BI__nvvm_atom_sys_add_gen_ll: |
| 18372 | return MakeScopedAtomic(Intrinsic::nvvm_atomic_add_gen_i_sys, *this, E); |
| 18373 | case NVPTX::BI__nvvm_atom_cta_add_gen_f: |
| 18374 | case NVPTX::BI__nvvm_atom_cta_add_gen_d: |
| 18375 | return MakeScopedAtomic(Intrinsic::nvvm_atomic_add_gen_f_cta, *this, E); |
| 18376 | case NVPTX::BI__nvvm_atom_sys_add_gen_f: |
| 18377 | case NVPTX::BI__nvvm_atom_sys_add_gen_d: |
| 18378 | return MakeScopedAtomic(Intrinsic::nvvm_atomic_add_gen_f_sys, *this, E); |
| 18379 | case NVPTX::BI__nvvm_atom_cta_xchg_gen_i: |
| 18380 | case NVPTX::BI__nvvm_atom_cta_xchg_gen_l: |
| 18381 | case NVPTX::BI__nvvm_atom_cta_xchg_gen_ll: |
| 18382 | return MakeScopedAtomic(Intrinsic::nvvm_atomic_exch_gen_i_cta, *this, E); |
| 18383 | case NVPTX::BI__nvvm_atom_sys_xchg_gen_i: |
| 18384 | case NVPTX::BI__nvvm_atom_sys_xchg_gen_l: |
| 18385 | case NVPTX::BI__nvvm_atom_sys_xchg_gen_ll: |
| 18386 | return MakeScopedAtomic(Intrinsic::nvvm_atomic_exch_gen_i_sys, *this, E); |
| 18387 | case NVPTX::BI__nvvm_atom_cta_max_gen_i: |
| 18388 | case NVPTX::BI__nvvm_atom_cta_max_gen_ui: |
| 18389 | case NVPTX::BI__nvvm_atom_cta_max_gen_l: |
| 18390 | case NVPTX::BI__nvvm_atom_cta_max_gen_ul: |
| 18391 | case NVPTX::BI__nvvm_atom_cta_max_gen_ll: |
| 18392 | case NVPTX::BI__nvvm_atom_cta_max_gen_ull: |
| 18393 | return MakeScopedAtomic(Intrinsic::nvvm_atomic_max_gen_i_cta, *this, E); |
| 18394 | case NVPTX::BI__nvvm_atom_sys_max_gen_i: |
| 18395 | case NVPTX::BI__nvvm_atom_sys_max_gen_ui: |
| 18396 | case NVPTX::BI__nvvm_atom_sys_max_gen_l: |
| 18397 | case NVPTX::BI__nvvm_atom_sys_max_gen_ul: |
| 18398 | case NVPTX::BI__nvvm_atom_sys_max_gen_ll: |
| 18399 | case NVPTX::BI__nvvm_atom_sys_max_gen_ull: |
| 18400 | return MakeScopedAtomic(Intrinsic::nvvm_atomic_max_gen_i_sys, *this, E); |
| 18401 | case NVPTX::BI__nvvm_atom_cta_min_gen_i: |
| 18402 | case NVPTX::BI__nvvm_atom_cta_min_gen_ui: |
| 18403 | case NVPTX::BI__nvvm_atom_cta_min_gen_l: |
| 18404 | case NVPTX::BI__nvvm_atom_cta_min_gen_ul: |
| 18405 | case NVPTX::BI__nvvm_atom_cta_min_gen_ll: |
| 18406 | case NVPTX::BI__nvvm_atom_cta_min_gen_ull: |
| 18407 | return MakeScopedAtomic(Intrinsic::nvvm_atomic_min_gen_i_cta, *this, E); |
| 18408 | case NVPTX::BI__nvvm_atom_sys_min_gen_i: |
| 18409 | case NVPTX::BI__nvvm_atom_sys_min_gen_ui: |
| 18410 | case NVPTX::BI__nvvm_atom_sys_min_gen_l: |
| 18411 | case NVPTX::BI__nvvm_atom_sys_min_gen_ul: |
| 18412 | case NVPTX::BI__nvvm_atom_sys_min_gen_ll: |
| 18413 | case NVPTX::BI__nvvm_atom_sys_min_gen_ull: |
| 18414 | return MakeScopedAtomic(Intrinsic::nvvm_atomic_min_gen_i_sys, *this, E); |
| 18415 | case NVPTX::BI__nvvm_atom_cta_inc_gen_ui: |
| 18416 | return MakeScopedAtomic(Intrinsic::nvvm_atomic_inc_gen_i_cta, *this, E); |
| 18417 | case NVPTX::BI__nvvm_atom_cta_dec_gen_ui: |
| 18418 | return MakeScopedAtomic(Intrinsic::nvvm_atomic_dec_gen_i_cta, *this, E); |
| 18419 | case NVPTX::BI__nvvm_atom_sys_inc_gen_ui: |
| 18420 | return MakeScopedAtomic(Intrinsic::nvvm_atomic_inc_gen_i_sys, *this, E); |
| 18421 | case NVPTX::BI__nvvm_atom_sys_dec_gen_ui: |
| 18422 | return MakeScopedAtomic(Intrinsic::nvvm_atomic_dec_gen_i_sys, *this, E); |
| 18423 | case NVPTX::BI__nvvm_atom_cta_and_gen_i: |
| 18424 | case NVPTX::BI__nvvm_atom_cta_and_gen_l: |
| 18425 | case NVPTX::BI__nvvm_atom_cta_and_gen_ll: |
| 18426 | return MakeScopedAtomic(Intrinsic::nvvm_atomic_and_gen_i_cta, *this, E); |
| 18427 | case NVPTX::BI__nvvm_atom_sys_and_gen_i: |
| 18428 | case NVPTX::BI__nvvm_atom_sys_and_gen_l: |
| 18429 | case NVPTX::BI__nvvm_atom_sys_and_gen_ll: |
| 18430 | return MakeScopedAtomic(Intrinsic::nvvm_atomic_and_gen_i_sys, *this, E); |
| 18431 | case NVPTX::BI__nvvm_atom_cta_or_gen_i: |
| 18432 | case NVPTX::BI__nvvm_atom_cta_or_gen_l: |
| 18433 | case NVPTX::BI__nvvm_atom_cta_or_gen_ll: |
| 18434 | return MakeScopedAtomic(Intrinsic::nvvm_atomic_or_gen_i_cta, *this, E); |
| 18435 | case NVPTX::BI__nvvm_atom_sys_or_gen_i: |
| 18436 | case NVPTX::BI__nvvm_atom_sys_or_gen_l: |
| 18437 | case NVPTX::BI__nvvm_atom_sys_or_gen_ll: |
| 18438 | return MakeScopedAtomic(Intrinsic::nvvm_atomic_or_gen_i_sys, *this, E); |
| 18439 | case NVPTX::BI__nvvm_atom_cta_xor_gen_i: |
| 18440 | case NVPTX::BI__nvvm_atom_cta_xor_gen_l: |
| 18441 | case NVPTX::BI__nvvm_atom_cta_xor_gen_ll: |
| 18442 | return MakeScopedAtomic(Intrinsic::nvvm_atomic_xor_gen_i_cta, *this, E); |
| 18443 | case NVPTX::BI__nvvm_atom_sys_xor_gen_i: |
| 18444 | case NVPTX::BI__nvvm_atom_sys_xor_gen_l: |
| 18445 | case NVPTX::BI__nvvm_atom_sys_xor_gen_ll: |
| 18446 | return MakeScopedAtomic(Intrinsic::nvvm_atomic_xor_gen_i_sys, *this, E); |
| 18447 | case NVPTX::BI__nvvm_atom_cta_cas_gen_i: |
| 18448 | case NVPTX::BI__nvvm_atom_cta_cas_gen_l: |
| 18449 | case NVPTX::BI__nvvm_atom_cta_cas_gen_ll: { |
| 18450 | Value *Ptr = EmitScalarExpr(E->getArg(0)); |
| 18451 | llvm::Type *ElemTy = |
| 18452 | ConvertTypeForMem(E->getArg(0)->getType()->getPointeeType()); |
| 18453 | return Builder.CreateCall( |
| 18454 | CGM.getIntrinsic( |
| 18455 | Intrinsic::nvvm_atomic_cas_gen_i_cta, {ElemTy, Ptr->getType()}), |
| 18456 | {Ptr, EmitScalarExpr(E->getArg(1)), EmitScalarExpr(E->getArg(2))}); |
| 18457 | } |
| 18458 | case NVPTX::BI__nvvm_atom_sys_cas_gen_i: |
| 18459 | case NVPTX::BI__nvvm_atom_sys_cas_gen_l: |
| 18460 | case NVPTX::BI__nvvm_atom_sys_cas_gen_ll: { |
| 18461 | Value *Ptr = EmitScalarExpr(E->getArg(0)); |
| 18462 | llvm::Type *ElemTy = |
| 18463 | ConvertTypeForMem(E->getArg(0)->getType()->getPointeeType()); |
| 18464 | return Builder.CreateCall( |
| 18465 | CGM.getIntrinsic( |
| 18466 | Intrinsic::nvvm_atomic_cas_gen_i_sys, {ElemTy, Ptr->getType()}), |
| 18467 | {Ptr, EmitScalarExpr(E->getArg(1)), EmitScalarExpr(E->getArg(2))}); |
| 18468 | } |
| 18469 | case NVPTX::BI__nvvm_match_all_sync_i32p: |
| 18470 | case NVPTX::BI__nvvm_match_all_sync_i64p: { |
| 18471 | Value *Mask = EmitScalarExpr(E->getArg(0)); |
| 18472 | Value *Val = EmitScalarExpr(E->getArg(1)); |
| 18473 | Address PredOutPtr = EmitPointerWithAlignment(E->getArg(2)); |
| 18474 | Value *ResultPair = Builder.CreateCall( |
| 18475 | CGM.getIntrinsic(BuiltinID == NVPTX::BI__nvvm_match_all_sync_i32p |
| 18476 | ? Intrinsic::nvvm_match_all_sync_i32p |
| 18477 | : Intrinsic::nvvm_match_all_sync_i64p), |
| 18478 | {Mask, Val}); |
| 18479 | Value *Pred = Builder.CreateZExt(Builder.CreateExtractValue(ResultPair, 1), |
| 18480 | PredOutPtr.getElementType()); |
| 18481 | Builder.CreateStore(Pred, PredOutPtr); |
| 18482 | return Builder.CreateExtractValue(ResultPair, 0); |
| 18483 | } |
| 18484 | |
| 18485 | // FP MMA loads |
| 18486 | case NVPTX::BI__hmma_m16n16k16_ld_a: |
| 18487 | case NVPTX::BI__hmma_m16n16k16_ld_b: |
| 18488 | case NVPTX::BI__hmma_m16n16k16_ld_c_f16: |
| 18489 | case NVPTX::BI__hmma_m16n16k16_ld_c_f32: |
| 18490 | case NVPTX::BI__hmma_m32n8k16_ld_a: |
| 18491 | case NVPTX::BI__hmma_m32n8k16_ld_b: |
| 18492 | case NVPTX::BI__hmma_m32n8k16_ld_c_f16: |
| 18493 | case NVPTX::BI__hmma_m32n8k16_ld_c_f32: |
| 18494 | case NVPTX::BI__hmma_m8n32k16_ld_a: |
| 18495 | case NVPTX::BI__hmma_m8n32k16_ld_b: |
| 18496 | case NVPTX::BI__hmma_m8n32k16_ld_c_f16: |
| 18497 | case NVPTX::BI__hmma_m8n32k16_ld_c_f32: |
| 18498 | // Integer MMA loads. |
| 18499 | case NVPTX::BI__imma_m16n16k16_ld_a_s8: |
| 18500 | case NVPTX::BI__imma_m16n16k16_ld_a_u8: |
| 18501 | case NVPTX::BI__imma_m16n16k16_ld_b_s8: |
| 18502 | case NVPTX::BI__imma_m16n16k16_ld_b_u8: |
| 18503 | case NVPTX::BI__imma_m16n16k16_ld_c: |
| 18504 | case NVPTX::BI__imma_m32n8k16_ld_a_s8: |
| 18505 | case NVPTX::BI__imma_m32n8k16_ld_a_u8: |
| 18506 | case NVPTX::BI__imma_m32n8k16_ld_b_s8: |
| 18507 | case NVPTX::BI__imma_m32n8k16_ld_b_u8: |
| 18508 | case NVPTX::BI__imma_m32n8k16_ld_c: |
| 18509 | case NVPTX::BI__imma_m8n32k16_ld_a_s8: |
| 18510 | case NVPTX::BI__imma_m8n32k16_ld_a_u8: |
| 18511 | case NVPTX::BI__imma_m8n32k16_ld_b_s8: |
| 18512 | case NVPTX::BI__imma_m8n32k16_ld_b_u8: |
| 18513 | case NVPTX::BI__imma_m8n32k16_ld_c: |
| 18514 | // Sub-integer MMA loads. |
| 18515 | case NVPTX::BI__imma_m8n8k32_ld_a_s4: |
| 18516 | case NVPTX::BI__imma_m8n8k32_ld_a_u4: |
| 18517 | case NVPTX::BI__imma_m8n8k32_ld_b_s4: |
| 18518 | case NVPTX::BI__imma_m8n8k32_ld_b_u4: |
| 18519 | case NVPTX::BI__imma_m8n8k32_ld_c: |
| 18520 | case NVPTX::BI__bmma_m8n8k128_ld_a_b1: |
| 18521 | case NVPTX::BI__bmma_m8n8k128_ld_b_b1: |
| 18522 | case NVPTX::BI__bmma_m8n8k128_ld_c: |
| 18523 | // Double MMA loads. |
| 18524 | case NVPTX::BI__dmma_m8n8k4_ld_a: |
| 18525 | case NVPTX::BI__dmma_m8n8k4_ld_b: |
| 18526 | case NVPTX::BI__dmma_m8n8k4_ld_c: |
| 18527 | // Alternate float MMA loads. |
| 18528 | case NVPTX::BI__mma_bf16_m16n16k16_ld_a: |
| 18529 | case NVPTX::BI__mma_bf16_m16n16k16_ld_b: |
| 18530 | case NVPTX::BI__mma_bf16_m8n32k16_ld_a: |
| 18531 | case NVPTX::BI__mma_bf16_m8n32k16_ld_b: |
| 18532 | case NVPTX::BI__mma_bf16_m32n8k16_ld_a: |
| 18533 | case NVPTX::BI__mma_bf16_m32n8k16_ld_b: |
| 18534 | case NVPTX::BI__mma_tf32_m16n16k8_ld_a: |
| 18535 | case NVPTX::BI__mma_tf32_m16n16k8_ld_b: |
| 18536 | case NVPTX::BI__mma_tf32_m16n16k8_ld_c: { |
| 18537 | Address Dst = EmitPointerWithAlignment(E->getArg(0)); |
| 18538 | Value *Src = EmitScalarExpr(E->getArg(1)); |
| 18539 | Value *Ldm = EmitScalarExpr(E->getArg(2)); |
| 18540 | std::optional<llvm::APSInt> isColMajorArg = |
| 18541 | E->getArg(3)->getIntegerConstantExpr(getContext()); |
| 18542 | if (!isColMajorArg) |
| 18543 | return nullptr; |
| 18544 | bool isColMajor = isColMajorArg->getSExtValue(); |
| 18545 | NVPTXMmaLdstInfo II = getNVPTXMmaLdstInfo(BuiltinID); |
| 18546 | unsigned IID = isColMajor ? II.IID_col : II.IID_row; |
| 18547 | if (IID == 0) |
| 18548 | return nullptr; |
| 18549 | |
| 18550 | Value *Result = |
| 18551 | Builder.CreateCall(CGM.getIntrinsic(IID, Src->getType()), {Src, Ldm}); |
| 18552 | |
| 18553 | // Save returned values. |
| 18554 | assert(II.NumResults)(static_cast <bool> (II.NumResults) ? void (0) : __assert_fail ("II.NumResults", "clang/lib/CodeGen/CGBuiltin.cpp", 18554, __extension__ __PRETTY_FUNCTION__)); |
| 18555 | if (II.NumResults == 1) { |
| 18556 | Builder.CreateAlignedStore(Result, Dst.getPointer(), |
| 18557 | CharUnits::fromQuantity(4)); |
| 18558 | } else { |
| 18559 | for (unsigned i = 0; i < II.NumResults; ++i) { |
| 18560 | Builder.CreateAlignedStore( |
| 18561 | Builder.CreateBitCast(Builder.CreateExtractValue(Result, i), |
| 18562 | Dst.getElementType()), |
| 18563 | Builder.CreateGEP(Dst.getElementType(), Dst.getPointer(), |
| 18564 | llvm::ConstantInt::get(IntTy, i)), |
| 18565 | CharUnits::fromQuantity(4)); |
| 18566 | } |
| 18567 | } |
| 18568 | return Result; |
| 18569 | } |
| 18570 | |
| 18571 | case NVPTX::BI__hmma_m16n16k16_st_c_f16: |
| 18572 | case NVPTX::BI__hmma_m16n16k16_st_c_f32: |
| 18573 | case NVPTX::BI__hmma_m32n8k16_st_c_f16: |
| 18574 | case NVPTX::BI__hmma_m32n8k16_st_c_f32: |
| 18575 | case NVPTX::BI__hmma_m8n32k16_st_c_f16: |
| 18576 | case NVPTX::BI__hmma_m8n32k16_st_c_f32: |
| 18577 | case NVPTX::BI__imma_m16n16k16_st_c_i32: |
| 18578 | case NVPTX::BI__imma_m32n8k16_st_c_i32: |
| 18579 | case NVPTX::BI__imma_m8n32k16_st_c_i32: |
| 18580 | case NVPTX::BI__imma_m8n8k32_st_c_i32: |
| 18581 | case NVPTX::BI__bmma_m8n8k128_st_c_i32: |
| 18582 | case NVPTX::BI__dmma_m8n8k4_st_c_f64: |
| 18583 | case NVPTX::BI__mma_m16n16k8_st_c_f32: { |
| 18584 | Value *Dst = EmitScalarExpr(E->getArg(0)); |
| 18585 | Address Src = EmitPointerWithAlignment(E->getArg(1)); |
| 18586 | Value *Ldm = EmitScalarExpr(E->getArg(2)); |
| 18587 | std::optional<llvm::APSInt> isColMajorArg = |
| 18588 | E->getArg(3)->getIntegerConstantExpr(getContext()); |
| 18589 | if (!isColMajorArg) |
| 18590 | return nullptr; |
| 18591 | bool isColMajor = isColMajorArg->getSExtValue(); |
| 18592 | NVPTXMmaLdstInfo II = getNVPTXMmaLdstInfo(BuiltinID); |
| 18593 | unsigned IID = isColMajor ? II.IID_col : II.IID_row; |
| 18594 | if (IID == 0) |
| 18595 | return nullptr; |
| 18596 | Function *Intrinsic = |
| 18597 | CGM.getIntrinsic(IID, Dst->getType()); |
| 18598 | llvm::Type *ParamType = Intrinsic->getFunctionType()->getParamType(1); |
| 18599 | SmallVector<Value *, 10> Values = {Dst}; |
| 18600 | for (unsigned i = 0; i < II.NumResults; ++i) { |
| 18601 | Value *V = Builder.CreateAlignedLoad( |
| 18602 | Src.getElementType(), |
| 18603 | Builder.CreateGEP(Src.getElementType(), Src.getPointer(), |
| 18604 | llvm::ConstantInt::get(IntTy, i)), |
| 18605 | CharUnits::fromQuantity(4)); |
| 18606 | Values.push_back(Builder.CreateBitCast(V, ParamType)); |
| 18607 | } |
| 18608 | Values.push_back(Ldm); |
| 18609 | Value *Result = Builder.CreateCall(Intrinsic, Values); |
| 18610 | return Result; |
| 18611 | } |
| 18612 | |
| 18613 | // BI__hmma_m16n16k16_mma_<Dtype><CType>(d, a, b, c, layout, satf) --> |
| 18614 | // Intrinsic::nvvm_wmma_m16n16k16_mma_sync<layout A,B><DType><CType><Satf> |
| 18615 | case NVPTX::BI__hmma_m16n16k16_mma_f16f16: |
| 18616 | case NVPTX::BI__hmma_m16n16k16_mma_f32f16: |
| 18617 | case NVPTX::BI__hmma_m16n16k16_mma_f32f32: |
| 18618 | case NVPTX::BI__hmma_m16n16k16_mma_f16f32: |
| 18619 | case NVPTX::BI__hmma_m32n8k16_mma_f16f16: |
| 18620 | case NVPTX::BI__hmma_m32n8k16_mma_f32f16: |
| 18621 | case NVPTX::BI__hmma_m32n8k16_mma_f32f32: |
| 18622 | case NVPTX::BI__hmma_m32n8k16_mma_f16f32: |
| 18623 | case NVPTX::BI__hmma_m8n32k16_mma_f16f16: |
| 18624 | case NVPTX::BI__hmma_m8n32k16_mma_f32f16: |
| 18625 | case NVPTX::BI__hmma_m8n32k16_mma_f32f32: |
| 18626 | case NVPTX::BI__hmma_m8n32k16_mma_f16f32: |
| 18627 | case NVPTX::BI__imma_m16n16k16_mma_s8: |
| 18628 | case NVPTX::BI__imma_m16n16k16_mma_u8: |
| 18629 | case NVPTX::BI__imma_m32n8k16_mma_s8: |
| 18630 | case NVPTX::BI__imma_m32n8k16_mma_u8: |
| 18631 | case NVPTX::BI__imma_m8n32k16_mma_s8: |
| 18632 | case NVPTX::BI__imma_m8n32k16_mma_u8: |
| 18633 | case NVPTX::BI__imma_m8n8k32_mma_s4: |
| 18634 | case NVPTX::BI__imma_m8n8k32_mma_u4: |
| 18635 | case NVPTX::BI__bmma_m8n8k128_mma_xor_popc_b1: |
| 18636 | case NVPTX::BI__bmma_m8n8k128_mma_and_popc_b1: |
| 18637 | case NVPTX::BI__dmma_m8n8k4_mma_f64: |
| 18638 | case NVPTX::BI__mma_bf16_m16n16k16_mma_f32: |
| 18639 | case NVPTX::BI__mma_bf16_m8n32k16_mma_f32: |
| 18640 | case NVPTX::BI__mma_bf16_m32n8k16_mma_f32: |
| 18641 | case NVPTX::BI__mma_tf32_m16n16k8_mma_f32: { |
| 18642 | Address Dst = EmitPointerWithAlignment(E->getArg(0)); |
| 18643 | Address SrcA = EmitPointerWithAlignment(E->getArg(1)); |
| 18644 | Address SrcB = EmitPointerWithAlignment(E->getArg(2)); |
| 18645 | Address SrcC = EmitPointerWithAlignment(E->getArg(3)); |
| 18646 | std::optional<llvm::APSInt> LayoutArg = |
| 18647 | E->getArg(4)->getIntegerConstantExpr(getContext()); |
| 18648 | if (!LayoutArg) |
| 18649 | return nullptr; |
| 18650 | int Layout = LayoutArg->getSExtValue(); |
| 18651 | if (Layout < 0 || Layout > 3) |
| 18652 | return nullptr; |
| 18653 | llvm::APSInt SatfArg; |
| 18654 | if (BuiltinID == NVPTX::BI__bmma_m8n8k128_mma_xor_popc_b1 || |
| 18655 | BuiltinID == NVPTX::BI__bmma_m8n8k128_mma_and_popc_b1) |
| 18656 | SatfArg = 0; // .b1 does not have satf argument. |
| 18657 | else if (std::optional<llvm::APSInt> OptSatfArg = |
| 18658 | E->getArg(5)->getIntegerConstantExpr(getContext())) |
| 18659 | SatfArg = *OptSatfArg; |
| 18660 | else |
| 18661 | return nullptr; |
| 18662 | bool Satf = SatfArg.getSExtValue(); |
| 18663 | NVPTXMmaInfo MI = getNVPTXMmaInfo(BuiltinID); |
| 18664 | unsigned IID = MI.getMMAIntrinsic(Layout, Satf); |
| 18665 | if (IID == 0) // Unsupported combination of Layout/Satf. |
| 18666 | return nullptr; |
| 18667 | |
| 18668 | SmallVector<Value *, 24> Values; |
| 18669 | Function *Intrinsic = CGM.getIntrinsic(IID); |
| 18670 | llvm::Type *AType = Intrinsic->getFunctionType()->getParamType(0); |
| 18671 | // Load A |
| 18672 | for (unsigned i = 0; i < MI.NumEltsA; ++i) { |
| 18673 | Value *V = Builder.CreateAlignedLoad( |
| 18674 | SrcA.getElementType(), |
| 18675 | Builder.CreateGEP(SrcA.getElementType(), SrcA.getPointer(), |
| 18676 | llvm::ConstantInt::get(IntTy, i)), |
| 18677 | CharUnits::fromQuantity(4)); |
| 18678 | Values.push_back(Builder.CreateBitCast(V, AType)); |
| 18679 | } |
| 18680 | // Load B |
| 18681 | llvm::Type *BType = Intrinsic->getFunctionType()->getParamType(MI.NumEltsA); |
| 18682 | for (unsigned i = 0; i < MI.NumEltsB; ++i) { |
| 18683 | Value *V = Builder.CreateAlignedLoad( |
| 18684 | SrcB.getElementType(), |
| 18685 | Builder.CreateGEP(SrcB.getElementType(), SrcB.getPointer(), |
| 18686 | llvm::ConstantInt::get(IntTy, i)), |
| 18687 | CharUnits::fromQuantity(4)); |
| 18688 | Values.push_back(Builder.CreateBitCast(V, BType)); |
| 18689 | } |
| 18690 | // Load C |
| 18691 | llvm::Type *CType = |
| 18692 | Intrinsic->getFunctionType()->getParamType(MI.NumEltsA + MI.NumEltsB); |
| 18693 | for (unsigned i = 0; i < MI.NumEltsC; ++i) { |
| 18694 | Value *V = Builder.CreateAlignedLoad( |
| 18695 | SrcC.getElementType(), |
| 18696 | Builder.CreateGEP(SrcC.getElementType(), SrcC.getPointer(), |
| 18697 | llvm::ConstantInt::get(IntTy, i)), |
| 18698 | CharUnits::fromQuantity(4)); |
| 18699 | Values.push_back(Builder.CreateBitCast(V, CType)); |
| 18700 | } |
| 18701 | Value *Result = Builder.CreateCall(Intrinsic, Values); |
| 18702 | llvm::Type *DType = Dst.getElementType(); |
| 18703 | for (unsigned i = 0; i < MI.NumEltsD; ++i) |
| 18704 | Builder.CreateAlignedStore( |
| 18705 | Builder.CreateBitCast(Builder.CreateExtractValue(Result, i), DType), |
| 18706 | Builder.CreateGEP(Dst.getElementType(), Dst.getPointer(), |
| 18707 | llvm::ConstantInt::get(IntTy, i)), |
| 18708 | CharUnits::fromQuantity(4)); |
| 18709 | return Result; |
| 18710 | } |
| 18711 | // The following builtins require half type support |
| 18712 | case NVPTX::BI__nvvm_ex2_approx_f16: |
| 18713 | return MakeHalfType(Intrinsic::nvvm_ex2_approx_f16, BuiltinID, E, *this); |
| 18714 | case NVPTX::BI__nvvm_ex2_approx_f16x2: |
| 18715 | return MakeHalfType(Intrinsic::nvvm_ex2_approx_f16x2, BuiltinID, E, *this); |
| 18716 | case NVPTX::BI__nvvm_ff2f16x2_rn: |
| 18717 | return MakeHalfType(Intrinsic::nvvm_ff2f16x2_rn, BuiltinID, E, *this); |
| 18718 | case NVPTX::BI__nvvm_ff2f16x2_rn_relu: |
| 18719 | return MakeHalfType(Intrinsic::nvvm_ff2f16x2_rn_relu, BuiltinID, E, *this); |
| 18720 | case NVPTX::BI__nvvm_ff2f16x2_rz: |
| 18721 | return MakeHalfType(Intrinsic::nvvm_ff2f16x2_rz, BuiltinID, E, *this); |
| 18722 | case NVPTX::BI__nvvm_ff2f16x2_rz_relu: |
| 18723 | return MakeHalfType(Intrinsic::nvvm_ff2f16x2_rz_relu, BuiltinID, E, *this); |
| 18724 | case NVPTX::BI__nvvm_fma_rn_f16: |
| 18725 | return MakeHalfType(Intrinsic::nvvm_fma_rn_f16, BuiltinID, E, *this); |
| 18726 | case NVPTX::BI__nvvm_fma_rn_f16x2: |
| 18727 | return MakeHalfType(Intrinsic::nvvm_fma_rn_f16x2, BuiltinID, E, *this); |
| 18728 | case NVPTX::BI__nvvm_fma_rn_ftz_f16: |
| 18729 | return MakeHalfType(Intrinsic::nvvm_fma_rn_ftz_f16, BuiltinID, E, *this); |
| 18730 | case NVPTX::BI__nvvm_fma_rn_ftz_f16x2: |
| 18731 | return MakeHalfType(Intrinsic::nvvm_fma_rn_ftz_f16x2, BuiltinID, E, *this); |
| 18732 | case NVPTX::BI__nvvm_fma_rn_ftz_relu_f16: |
| 18733 | return MakeHalfType(Intrinsic::nvvm_fma_rn_ftz_relu_f16, BuiltinID, E, |
| 18734 | *this); |
| 18735 | case NVPTX::BI__nvvm_fma_rn_ftz_relu_f16x2: |
| 18736 | return MakeHalfType(Intrinsic::nvvm_fma_rn_ftz_relu_f16x2, BuiltinID, E, |
| 18737 | *this); |
| 18738 | case NVPTX::BI__nvvm_fma_rn_ftz_sat_f16: |
| 18739 | return MakeHalfType(Intrinsic::nvvm_fma_rn_ftz_sat_f16, BuiltinID, E, |
| 18740 | *this); |
| 18741 | case NVPTX::BI__nvvm_fma_rn_ftz_sat_f16x2: |
| 18742 | return MakeHalfType(Intrinsic::nvvm_fma_rn_ftz_sat_f16x2, BuiltinID, E, |
| 18743 | *this); |
| 18744 | case NVPTX::BI__nvvm_fma_rn_relu_f16: |
| 18745 | return MakeHalfType(Intrinsic::nvvm_fma_rn_relu_f16, BuiltinID, E, *this); |
| 18746 | case NVPTX::BI__nvvm_fma_rn_relu_f16x2: |
| 18747 | return MakeHalfType(Intrinsic::nvvm_fma_rn_relu_f16x2, BuiltinID, E, *this); |
| 18748 | case NVPTX::BI__nvvm_fma_rn_sat_f16: |
| 18749 | return MakeHalfType(Intrinsic::nvvm_fma_rn_sat_f16, BuiltinID, E, *this); |
| 18750 | case NVPTX::BI__nvvm_fma_rn_sat_f16x2: |
| 18751 | return MakeHalfType(Intrinsic::nvvm_fma_rn_sat_f16x2, BuiltinID, E, *this); |
| 18752 | case NVPTX::BI__nvvm_fmax_f16: |
| 18753 | return MakeHalfType(Intrinsic::nvvm_fmax_f16, BuiltinID, E, *this); |
| 18754 | case NVPTX::BI__nvvm_fmax_f16x2: |
| 18755 | return MakeHalfType(Intrinsic::nvvm_fmax_f16x2, BuiltinID, E, *this); |
| 18756 | case NVPTX::BI__nvvm_fmax_ftz_f16: |
| 18757 | return MakeHalfType(Intrinsic::nvvm_fmax_ftz_f16, BuiltinID, E, *this); |
| 18758 | case NVPTX::BI__nvvm_fmax_ftz_f16x2: |
| 18759 | return MakeHalfType(Intrinsic::nvvm_fmax_ftz_f16x2, BuiltinID, E, *this); |
| 18760 | case NVPTX::BI__nvvm_fmax_ftz_nan_f16: |
| 18761 | return MakeHalfType(Intrinsic::nvvm_fmax_ftz_nan_f16, BuiltinID, E, *this); |
| 18762 | case NVPTX::BI__nvvm_fmax_ftz_nan_f16x2: |
| 18763 | return MakeHalfType(Intrinsic::nvvm_fmax_ftz_nan_f16x2, BuiltinID, E, |
| 18764 | *this); |
| 18765 | case NVPTX::BI__nvvm_fmax_ftz_nan_xorsign_abs_f16: |
| 18766 | return MakeHalfType(Intrinsic::nvvm_fmax_ftz_nan_xorsign_abs_f16, BuiltinID, |
| 18767 | E, *this); |
| 18768 | case NVPTX::BI__nvvm_fmax_ftz_nan_xorsign_abs_f16x2: |
| 18769 | return MakeHalfType(Intrinsic::nvvm_fmax_ftz_nan_xorsign_abs_f16x2, |
| 18770 | BuiltinID, E, *this); |
| 18771 | case NVPTX::BI__nvvm_fmax_ftz_xorsign_abs_f16: |
| 18772 | return MakeHalfType(Intrinsic::nvvm_fmax_ftz_xorsign_abs_f16, BuiltinID, E, |
| 18773 | *this); |
| 18774 | case NVPTX::BI__nvvm_fmax_ftz_xorsign_abs_f16x2: |
| 18775 | return MakeHalfType(Intrinsic::nvvm_fmax_ftz_xorsign_abs_f16x2, BuiltinID, |
| 18776 | E, *this); |
| 18777 | case NVPTX::BI__nvvm_fmax_nan_f16: |
| 18778 | return MakeHalfType(Intrinsic::nvvm_fmax_nan_f16, BuiltinID, E, *this); |
| 18779 | case NVPTX::BI__nvvm_fmax_nan_f16x2: |
| 18780 | return MakeHalfType(Intrinsic::nvvm_fmax_nan_f16x2, BuiltinID, E, *this); |
| 18781 | case NVPTX::BI__nvvm_fmax_nan_xorsign_abs_f16: |
| 18782 | return MakeHalfType(Intrinsic::nvvm_fmax_nan_xorsign_abs_f16, BuiltinID, E, |
| 18783 | *this); |
| 18784 | case NVPTX::BI__nvvm_fmax_nan_xorsign_abs_f16x2: |
| 18785 | return MakeHalfType(Intrinsic::nvvm_fmax_nan_xorsign_abs_f16x2, BuiltinID, |
| 18786 | E, *this); |
| 18787 | case NVPTX::BI__nvvm_fmax_xorsign_abs_f16: |
| 18788 | return MakeHalfType(Intrinsic::nvvm_fmax_xorsign_abs_f16, BuiltinID, E, |
| 18789 | *this); |
| 18790 | case NVPTX::BI__nvvm_fmax_xorsign_abs_f16x2: |
| 18791 | return MakeHalfType(Intrinsic::nvvm_fmax_xorsign_abs_f16x2, BuiltinID, E, |
| 18792 | *this); |
| 18793 | case NVPTX::BI__nvvm_fmin_f16: |
| 18794 | return MakeHalfType(Intrinsic::nvvm_fmin_f16, BuiltinID, E, *this); |
| 18795 | case NVPTX::BI__nvvm_fmin_f16x2: |
| 18796 | return MakeHalfType(Intrinsic::nvvm_fmin_f16x2, BuiltinID, E, *this); |
| 18797 | case NVPTX::BI__nvvm_fmin_ftz_f16: |
| 18798 | return MakeHalfType(Intrinsic::nvvm_fmin_ftz_f16, BuiltinID, E, *this); |
| 18799 | case NVPTX::BI__nvvm_fmin_ftz_f16x2: |
| 18800 | return MakeHalfType(Intrinsic::nvvm_fmin_ftz_f16x2, BuiltinID, E, *this); |
| 18801 | case NVPTX::BI__nvvm_fmin_ftz_nan_f16: |
| 18802 | return MakeHalfType(Intrinsic::nvvm_fmin_ftz_nan_f16, BuiltinID, E, *this); |
| 18803 | case NVPTX::BI__nvvm_fmin_ftz_nan_f16x2: |
| 18804 | return MakeHalfType(Intrinsic::nvvm_fmin_ftz_nan_f16x2, BuiltinID, E, |
| 18805 | *this); |
| 18806 | case NVPTX::BI__nvvm_fmin_ftz_nan_xorsign_abs_f16: |
| 18807 | return MakeHalfType(Intrinsic::nvvm_fmin_ftz_nan_xorsign_abs_f16, BuiltinID, |
| 18808 | E, *this); |
| 18809 | case NVPTX::BI__nvvm_fmin_ftz_nan_xorsign_abs_f16x2: |
| 18810 | return MakeHalfType(Intrinsic::nvvm_fmin_ftz_nan_xorsign_abs_f16x2, |
| 18811 | BuiltinID, E, *this); |
| 18812 | case NVPTX::BI__nvvm_fmin_ftz_xorsign_abs_f16: |
| 18813 | return MakeHalfType(Intrinsic::nvvm_fmin_ftz_xorsign_abs_f16, BuiltinID, E, |
| 18814 | *this); |
| 18815 | case NVPTX::BI__nvvm_fmin_ftz_xorsign_abs_f16x2: |
| 18816 | return MakeHalfType(Intrinsic::nvvm_fmin_ftz_xorsign_abs_f16x2, BuiltinID, |
| 18817 | E, *this); |
| 18818 | case NVPTX::BI__nvvm_fmin_nan_f16: |
| 18819 | return MakeHalfType(Intrinsic::nvvm_fmin_nan_f16, BuiltinID, E, *this); |
| 18820 | case NVPTX::BI__nvvm_fmin_nan_f16x2: |
| 18821 | return MakeHalfType(Intrinsic::nvvm_fmin_nan_f16x2, BuiltinID, E, *this); |
| 18822 | case NVPTX::BI__nvvm_fmin_nan_xorsign_abs_f16: |
| 18823 | return MakeHalfType(Intrinsic::nvvm_fmin_nan_xorsign_abs_f16, BuiltinID, E, |
| 18824 | *this); |
| 18825 | case NVPTX::BI__nvvm_fmin_nan_xorsign_abs_f16x2: |
| 18826 | return MakeHalfType(Intrinsic::nvvm_fmin_nan_xorsign_abs_f16x2, BuiltinID, |
| 18827 | E, *this); |
| 18828 | case NVPTX::BI__nvvm_fmin_xorsign_abs_f16: |
| 18829 | return MakeHalfType(Intrinsic::nvvm_fmin_xorsign_abs_f16, BuiltinID, E, |
| 18830 | *this); |
| 18831 | case NVPTX::BI__nvvm_fmin_xorsign_abs_f16x2: |
| 18832 | return MakeHalfType(Intrinsic::nvvm_fmin_xorsign_abs_f16x2, BuiltinID, E, |
| 18833 | *this); |
| 18834 | case NVPTX::BI__nvvm_ldg_h: |
| 18835 | return MakeHalfType(Intrinsic::nvvm_ldg_global_f, BuiltinID, E, *this); |
| 18836 | case NVPTX::BI__nvvm_ldg_h2: |
| 18837 | return MakeHalfType(Intrinsic::nvvm_ldg_global_f, BuiltinID, E, *this); |
| 18838 | case NVPTX::BI__nvvm_ldu_h: |
| 18839 | return MakeHalfType(Intrinsic::nvvm_ldu_global_f, BuiltinID, E, *this); |
| 18840 | case NVPTX::BI__nvvm_ldu_h2: { |
| 18841 | return MakeHalfType(Intrinsic::nvvm_ldu_global_f, BuiltinID, E, *this); |
| 18842 | } |
| 18843 | default: |
| 18844 | return nullptr; |
| 18845 | } |
| 18846 | } |
| 18847 | |
| 18848 | namespace { |
| 18849 | struct BuiltinAlignArgs { |
| 18850 | llvm::Value *Src = nullptr; |
| 18851 | llvm::Type *SrcType = nullptr; |
| 18852 | llvm::Value *Alignment = nullptr; |
| 18853 | llvm::Value *Mask = nullptr; |
| 18854 | llvm::IntegerType *IntType = nullptr; |
| 18855 | |
| 18856 | BuiltinAlignArgs(const CallExpr *E, CodeGenFunction &CGF) { |
| 18857 | QualType AstType = E->getArg(0)->getType(); |
| 18858 | if (AstType->isArrayType()) |
| 18859 | Src = CGF.EmitArrayToPointerDecay(E->getArg(0)).getPointer(); |
| 18860 | else |
| 18861 | Src = CGF.EmitScalarExpr(E->getArg(0)); |
| 18862 | SrcType = Src->getType(); |
| 18863 | if (SrcType->isPointerTy()) { |
| 18864 | IntType = IntegerType::get( |
| 18865 | CGF.getLLVMContext(), |
| 18866 | CGF.CGM.getDataLayout().getIndexTypeSizeInBits(SrcType)); |
| 18867 | } else { |
| 18868 | assert(SrcType->isIntegerTy())(static_cast <bool> (SrcType->isIntegerTy()) ? void ( 0) : __assert_fail ("SrcType->isIntegerTy()", "clang/lib/CodeGen/CGBuiltin.cpp" , 18868, __extension__ __PRETTY_FUNCTION__)); |
| 18869 | IntType = cast<llvm::IntegerType>(SrcType); |
| 18870 | } |
| 18871 | Alignment = CGF.EmitScalarExpr(E->getArg(1)); |
| 18872 | Alignment = CGF.Builder.CreateZExtOrTrunc(Alignment, IntType, "alignment"); |
| 18873 | auto *One = llvm::ConstantInt::get(IntType, 1); |
| 18874 | Mask = CGF.Builder.CreateSub(Alignment, One, "mask"); |
| 18875 | } |
| 18876 | }; |
| 18877 | } // namespace |
| 18878 | |
| 18879 | /// Generate (x & (y-1)) == 0. |
| 18880 | RValue CodeGenFunction::EmitBuiltinIsAligned(const CallExpr *E) { |
| 18881 | BuiltinAlignArgs Args(E, *this); |
| 18882 | llvm::Value *SrcAddress = Args.Src; |
| 18883 | if (Args.SrcType->isPointerTy()) |
| 18884 | SrcAddress = |
| 18885 | Builder.CreateBitOrPointerCast(Args.Src, Args.IntType, "src_addr"); |
| 18886 | return RValue::get(Builder.CreateICmpEQ( |
| 18887 | Builder.CreateAnd(SrcAddress, Args.Mask, "set_bits"), |
| 18888 | llvm::Constant::getNullValue(Args.IntType), "is_aligned")); |
| 18889 | } |
| 18890 | |
| 18891 | /// Generate (x & ~(y-1)) to align down or ((x+(y-1)) & ~(y-1)) to align up. |
| 18892 | /// Note: For pointer types we can avoid ptrtoint/inttoptr pairs by using the |
| 18893 | /// llvm.ptrmask intrinsic (with a GEP before in the align_up case). |
| 18894 | /// TODO: actually use ptrmask once most optimization passes know about it. |
| 18895 | RValue CodeGenFunction::EmitBuiltinAlignTo(const CallExpr *E, bool AlignUp) { |
| 18896 | BuiltinAlignArgs Args(E, *this); |
| 18897 | llvm::Value *SrcAddr = Args.Src; |
| 18898 | if (Args.Src->getType()->isPointerTy()) |
| 18899 | SrcAddr = Builder.CreatePtrToInt(Args.Src, Args.IntType, "intptr"); |
| 18900 | llvm::Value *SrcForMask = SrcAddr; |
| 18901 | if (AlignUp) { |
| 18902 | // When aligning up we have to first add the mask to ensure we go over the |
| 18903 | // next alignment value and then align down to the next valid multiple. |
| 18904 | // By adding the mask, we ensure that align_up on an already aligned |
| 18905 | // value will not change the value. |
| 18906 | SrcForMask = Builder.CreateAdd(SrcForMask, Args.Mask, "over_boundary"); |
| 18907 | } |
| 18908 | // Invert the mask to only clear the lower bits. |
| 18909 | llvm::Value *InvertedMask = Builder.CreateNot(Args.Mask, "inverted_mask"); |
| 18910 | llvm::Value *Result = |
| 18911 | Builder.CreateAnd(SrcForMask, InvertedMask, "aligned_result"); |
| 18912 | if (Args.Src->getType()->isPointerTy()) { |
| 18913 | /// TODO: Use ptrmask instead of ptrtoint+gep once it is optimized well. |
| 18914 | // Result = Builder.CreateIntrinsic( |
| 18915 | // Intrinsic::ptrmask, {Args.SrcType, SrcForMask->getType(), Args.IntType}, |
| 18916 | // {SrcForMask, NegatedMask}, nullptr, "aligned_result"); |
| 18917 | Result->setName("aligned_intptr"); |
| 18918 | llvm::Value *Difference = Builder.CreateSub(Result, SrcAddr, "diff"); |
| 18919 | // The result must point to the same underlying allocation. This means we |
| 18920 | // can use an inbounds GEP to enable better optimization. |
| 18921 | Value *Base = EmitCastToVoidPtr(Args.Src); |
| 18922 | if (getLangOpts().isSignedOverflowDefined()) |
| 18923 | Result = Builder.CreateGEP(Int8Ty, Base, Difference, "aligned_result"); |
| 18924 | else |
| 18925 | Result = EmitCheckedInBoundsGEP(Int8Ty, Base, Difference, |
| 18926 | /*SignedIndices=*/true, |
| 18927 | /*isSubtraction=*/!AlignUp, |
| 18928 | E->getExprLoc(), "aligned_result"); |
| 18929 | Result = Builder.CreatePointerCast(Result, Args.SrcType); |
| 18930 | // Emit an alignment assumption to ensure that the new alignment is |
| 18931 | // propagated to loads/stores, etc. |
| 18932 | emitAlignmentAssumption(Result, E, E->getExprLoc(), Args.Alignment); |
| 18933 | } |
| 18934 | assert(Result->getType() == Args.SrcType)(static_cast <bool> (Result->getType() == Args.SrcType ) ? void (0) : __assert_fail ("Result->getType() == Args.SrcType" , "clang/lib/CodeGen/CGBuiltin.cpp", 18934, __extension__ __PRETTY_FUNCTION__ )); |
| 18935 | return RValue::get(Result); |
| 18936 | } |
| 18937 | |
| 18938 | Value *CodeGenFunction::EmitWebAssemblyBuiltinExpr(unsigned BuiltinID, |
| 18939 | const CallExpr *E) { |
| 18940 | switch (BuiltinID) { |
| 18941 | case WebAssembly::BI__builtin_wasm_memory_size: { |
| 18942 | llvm::Type *ResultType = ConvertType(E->getType()); |
| 18943 | Value *I = EmitScalarExpr(E->getArg(0)); |
| 18944 | Function *Callee = |
| 18945 | CGM.getIntrinsic(Intrinsic::wasm_memory_size, ResultType); |
| 18946 | return Builder.CreateCall(Callee, I); |
| 18947 | } |
| 18948 | case WebAssembly::BI__builtin_wasm_memory_grow: { |
| 18949 | llvm::Type *ResultType = ConvertType(E->getType()); |
| 18950 | Value *Args[] = {EmitScalarExpr(E->getArg(0)), |
| 18951 | EmitScalarExpr(E->getArg(1))}; |
| 18952 | Function *Callee = |
| 18953 | CGM.getIntrinsic(Intrinsic::wasm_memory_grow, ResultType); |
| 18954 | return Builder.CreateCall(Callee, Args); |
| 18955 | } |
| 18956 | case WebAssembly::BI__builtin_wasm_tls_size: { |
| 18957 | llvm::Type *ResultType = ConvertType(E->getType()); |
| 18958 | Function *Callee = CGM.getIntrinsic(Intrinsic::wasm_tls_size, ResultType); |
| 18959 | return Builder.CreateCall(Callee); |
| 18960 | } |
| 18961 | case WebAssembly::BI__builtin_wasm_tls_align: { |
| 18962 | llvm::Type *ResultType = ConvertType(E->getType()); |
| 18963 | Function *Callee = CGM.getIntrinsic(Intrinsic::wasm_tls_align, ResultType); |
| 18964 | return Builder.CreateCall(Callee); |
| 18965 | } |
| 18966 | case WebAssembly::BI__builtin_wasm_tls_base: { |
| 18967 | Function *Callee = CGM.getIntrinsic(Intrinsic::wasm_tls_base); |
| 18968 | return Builder.CreateCall(Callee); |
| 18969 | } |
| 18970 | case WebAssembly::BI__builtin_wasm_throw: { |
| 18971 | Value *Tag = EmitScalarExpr(E->getArg(0)); |
| 18972 | Value *Obj = EmitScalarExpr(E->getArg(1)); |
| 18973 | Function *Callee = CGM.getIntrinsic(Intrinsic::wasm_throw); |
| 18974 | return Builder.CreateCall(Callee, {Tag, Obj}); |
| 18975 | } |
| 18976 | case WebAssembly::BI__builtin_wasm_rethrow: { |
| 18977 | Function *Callee = CGM.getIntrinsic(Intrinsic::wasm_rethrow); |
| 18978 | return Builder.CreateCall(Callee); |
| 18979 | } |
| 18980 | case WebAssembly::BI__builtin_wasm_memory_atomic_wait32: { |
| 18981 | Value *Addr = EmitScalarExpr(E->getArg(0)); |
| 18982 | Value *Expected = EmitScalarExpr(E->getArg(1)); |
| 18983 | Value *Timeout = EmitScalarExpr(E->getArg(2)); |
| 18984 | Function *Callee = CGM.getIntrinsic(Intrinsic::wasm_memory_atomic_wait32); |
| 18985 | return Builder.CreateCall(Callee, {Addr, Expected, Timeout}); |
| 18986 | } |
| 18987 | case WebAssembly::BI__builtin_wasm_memory_atomic_wait64: { |
| 18988 | Value *Addr = EmitScalarExpr(E->getArg(0)); |
| 18989 | Value *Expected = EmitScalarExpr(E->getArg(1)); |
| 18990 | Value *Timeout = EmitScalarExpr(E->getArg(2)); |
| 18991 | Function *Callee = CGM.getIntrinsic(Intrinsic::wasm_memory_atomic_wait64); |
| 18992 | return Builder.CreateCall(Callee, {Addr, Expected, Timeout}); |
| 18993 | } |
| 18994 | case WebAssembly::BI__builtin_wasm_memory_atomic_notify: { |
| 18995 | Value *Addr = EmitScalarExpr(E->getArg(0)); |
| 18996 | Value *Count = EmitScalarExpr(E->getArg(1)); |
| 18997 | Function *Callee = CGM.getIntrinsic(Intrinsic::wasm_memory_atomic_notify); |
| 18998 | return Builder.CreateCall(Callee, {Addr, Count}); |
| 18999 | } |
| 19000 | case WebAssembly::BI__builtin_wasm_trunc_s_i32_f32: |
| 19001 | case WebAssembly::BI__builtin_wasm_trunc_s_i32_f64: |
| 19002 | case WebAssembly::BI__builtin_wasm_trunc_s_i64_f32: |
| 19003 | case WebAssembly::BI__builtin_wasm_trunc_s_i64_f64: { |
| 19004 | Value *Src = EmitScalarExpr(E->getArg(0)); |
| 19005 | llvm::Type *ResT = ConvertType(E->getType()); |
| 19006 | Function *Callee = |
| 19007 | CGM.getIntrinsic(Intrinsic::wasm_trunc_signed, {ResT, Src->getType()}); |
| 19008 | return Builder.CreateCall(Callee, {Src}); |
| 19009 | } |
| 19010 | case WebAssembly::BI__builtin_wasm_trunc_u_i32_f32: |
| 19011 | case WebAssembly::BI__builtin_wasm_trunc_u_i32_f64: |
| 19012 | case WebAssembly::BI__builtin_wasm_trunc_u_i64_f32: |
| 19013 | case WebAssembly::BI__builtin_wasm_trunc_u_i64_f64: { |
| 19014 | Value *Src = EmitScalarExpr(E->getArg(0)); |
| 19015 | llvm::Type *ResT = ConvertType(E->getType()); |
| 19016 | Function *Callee = CGM.getIntrinsic(Intrinsic::wasm_trunc_unsigned, |
| 19017 | {ResT, Src->getType()}); |
| 19018 | return Builder.CreateCall(Callee, {Src}); |
| 19019 | } |
| 19020 | case WebAssembly::BI__builtin_wasm_trunc_saturate_s_i32_f32: |
| 19021 | case WebAssembly::BI__builtin_wasm_trunc_saturate_s_i32_f64: |
| 19022 | case WebAssembly::BI__builtin_wasm_trunc_saturate_s_i64_f32: |
| 19023 | case WebAssembly::BI__builtin_wasm_trunc_saturate_s_i64_f64: |
| 19024 | case WebAssembly::BI__builtin_wasm_trunc_saturate_s_i32x4_f32x4: { |
| 19025 | Value *Src = EmitScalarExpr(E->getArg(0)); |
| 19026 | llvm::Type *ResT = ConvertType(E->getType()); |
| 19027 | Function *Callee = |
| 19028 | CGM.getIntrinsic(Intrinsic::fptosi_sat, {ResT, Src->getType()}); |
| 19029 | return Builder.CreateCall(Callee, {Src}); |
| 19030 | } |
| 19031 | case WebAssembly::BI__builtin_wasm_trunc_saturate_u_i32_f32: |
| 19032 | case WebAssembly::BI__builtin_wasm_trunc_saturate_u_i32_f64: |
| 19033 | case WebAssembly::BI__builtin_wasm_trunc_saturate_u_i64_f32: |
| 19034 | case WebAssembly::BI__builtin_wasm_trunc_saturate_u_i64_f64: |
| 19035 | case WebAssembly::BI__builtin_wasm_trunc_saturate_u_i32x4_f32x4: { |
| 19036 | Value *Src = EmitScalarExpr(E->getArg(0)); |
| 19037 | llvm::Type *ResT = ConvertType(E->getType()); |
| 19038 | Function *Callee = |
| 19039 | CGM.getIntrinsic(Intrinsic::fptoui_sat, {ResT, Src->getType()}); |
| 19040 | return Builder.CreateCall(Callee, {Src}); |
| 19041 | } |
| 19042 | case WebAssembly::BI__builtin_wasm_min_f32: |
| 19043 | case WebAssembly::BI__builtin_wasm_min_f64: |
| 19044 | case WebAssembly::BI__builtin_wasm_min_f32x4: |
| 19045 | case WebAssembly::BI__builtin_wasm_min_f64x2: { |
| 19046 | Value *LHS = EmitScalarExpr(E->getArg(0)); |
| 19047 | Value *RHS = EmitScalarExpr(E->getArg(1)); |
| 19048 | Function *Callee = |
| 19049 | CGM.getIntrinsic(Intrinsic::minimum, ConvertType(E->getType())); |
| 19050 | return Builder.CreateCall(Callee, {LHS, RHS}); |
| 19051 | } |
| 19052 | case WebAssembly::BI__builtin_wasm_max_f32: |
| 19053 | case WebAssembly::BI__builtin_wasm_max_f64: |
| 19054 | case WebAssembly::BI__builtin_wasm_max_f32x4: |
| 19055 | case WebAssembly::BI__builtin_wasm_max_f64x2: { |
| 19056 | Value *LHS = EmitScalarExpr(E->getArg(0)); |
| 19057 | Value *RHS = EmitScalarExpr(E->getArg(1)); |
| 19058 | Function *Callee = |
| 19059 | CGM.getIntrinsic(Intrinsic::maximum, ConvertType(E->getType())); |
| 19060 | return Builder.CreateCall(Callee, {LHS, RHS}); |
| 19061 | } |
| 19062 | case WebAssembly::BI__builtin_wasm_pmin_f32x4: |
| 19063 | case WebAssembly::BI__builtin_wasm_pmin_f64x2: { |
| 19064 | Value *LHS = EmitScalarExpr(E->getArg(0)); |
| 19065 | Value *RHS = EmitScalarExpr(E->getArg(1)); |
| 19066 | Function *Callee = |
| 19067 | CGM.getIntrinsic(Intrinsic::wasm_pmin, ConvertType(E->getType())); |
| 19068 | return Builder.CreateCall(Callee, {LHS, RHS}); |
| 19069 | } |
| 19070 | case WebAssembly::BI__builtin_wasm_pmax_f32x4: |
| 19071 | case WebAssembly::BI__builtin_wasm_pmax_f64x2: { |
| 19072 | Value *LHS = EmitScalarExpr(E->getArg(0)); |
| 19073 | Value *RHS = EmitScalarExpr(E->getArg(1)); |
| 19074 | Function *Callee = |
| 19075 | CGM.getIntrinsic(Intrinsic::wasm_pmax, ConvertType(E->getType())); |
| 19076 | return Builder.CreateCall(Callee, {LHS, RHS}); |
| 19077 | } |
| 19078 | case WebAssembly::BI__builtin_wasm_ceil_f32x4: |
| 19079 | case WebAssembly::BI__builtin_wasm_floor_f32x4: |
| 19080 | case WebAssembly::BI__builtin_wasm_trunc_f32x4: |
| 19081 | case WebAssembly::BI__builtin_wasm_nearest_f32x4: |
| 19082 | case WebAssembly::BI__builtin_wasm_ceil_f64x2: |
| 19083 | case WebAssembly::BI__builtin_wasm_floor_f64x2: |
| 19084 | case WebAssembly::BI__builtin_wasm_trunc_f64x2: |
| 19085 | case WebAssembly::BI__builtin_wasm_nearest_f64x2: { |
| 19086 | unsigned IntNo; |
| 19087 | switch (BuiltinID) { |
| 19088 | case WebAssembly::BI__builtin_wasm_ceil_f32x4: |
| 19089 | case WebAssembly::BI__builtin_wasm_ceil_f64x2: |
| 19090 | IntNo = Intrinsic::ceil; |
| 19091 | break; |
| 19092 | case WebAssembly::BI__builtin_wasm_floor_f32x4: |
| 19093 | case WebAssembly::BI__builtin_wasm_floor_f64x2: |
| 19094 | IntNo = Intrinsic::floor; |
| 19095 | break; |
| 19096 | case WebAssembly::BI__builtin_wasm_trunc_f32x4: |
| 19097 | case WebAssembly::BI__builtin_wasm_trunc_f64x2: |
| 19098 | IntNo = Intrinsic::trunc; |
| 19099 | break; |
| 19100 | case WebAssembly::BI__builtin_wasm_nearest_f32x4: |
| 19101 | case WebAssembly::BI__builtin_wasm_nearest_f64x2: |
| 19102 | IntNo = Intrinsic::nearbyint; |
| 19103 | break; |
| 19104 | default: |
| 19105 | llvm_unreachable("unexpected builtin ID")::llvm::llvm_unreachable_internal("unexpected builtin ID", "clang/lib/CodeGen/CGBuiltin.cpp" , 19105); |
| 19106 | } |
| 19107 | Value *Value = EmitScalarExpr(E->getArg(0)); |
| 19108 | Function *Callee = CGM.getIntrinsic(IntNo, ConvertType(E->getType())); |
| 19109 | return Builder.CreateCall(Callee, Value); |
| 19110 | } |
| 19111 | case WebAssembly::BI__builtin_wasm_ref_null_extern: { |
| 19112 | Function *Callee = CGM.getIntrinsic(Intrinsic::wasm_ref_null_extern); |
| 19113 | return Builder.CreateCall(Callee); |
| 19114 | } |
| 19115 | case WebAssembly::BI__builtin_wasm_ref_null_func: { |
| 19116 | Function *Callee = CGM.getIntrinsic(Intrinsic::wasm_ref_null_func); |
| 19117 | return Builder.CreateCall(Callee); |
| 19118 | } |
| 19119 | case WebAssembly::BI__builtin_wasm_swizzle_i8x16: { |
| 19120 | Value *Src = EmitScalarExpr(E->getArg(0)); |
| 19121 | Value *Indices = EmitScalarExpr(E->getArg(1)); |
| 19122 | Function *Callee = CGM.getIntrinsic(Intrinsic::wasm_swizzle); |
| 19123 | return Builder.CreateCall(Callee, {Src, Indices}); |
| 19124 | } |
| 19125 | case WebAssembly::BI__builtin_wasm_add_sat_s_i8x16: |
| 19126 | case WebAssembly::BI__builtin_wasm_add_sat_u_i8x16: |
| 19127 | case WebAssembly::BI__builtin_wasm_add_sat_s_i16x8: |
| 19128 | case WebAssembly::BI__builtin_wasm_add_sat_u_i16x8: |
| 19129 | case WebAssembly::BI__builtin_wasm_sub_sat_s_i8x16: |
| 19130 | case WebAssembly::BI__builtin_wasm_sub_sat_u_i8x16: |
| 19131 | case WebAssembly::BI__builtin_wasm_sub_sat_s_i16x8: |
| 19132 | case WebAssembly::BI__builtin_wasm_sub_sat_u_i16x8: { |
| 19133 | unsigned IntNo; |
| 19134 | switch (BuiltinID) { |
| 19135 | case WebAssembly::BI__builtin_wasm_add_sat_s_i8x16: |
| 19136 | case WebAssembly::BI__builtin_wasm_add_sat_s_i16x8: |
| 19137 | IntNo = Intrinsic::sadd_sat; |
| 19138 | break; |
| 19139 | case WebAssembly::BI__builtin_wasm_add_sat_u_i8x16: |
| 19140 | case WebAssembly::BI__builtin_wasm_add_sat_u_i16x8: |
| 19141 | IntNo = Intrinsic::uadd_sat; |
| 19142 | break; |
| 19143 | case WebAssembly::BI__builtin_wasm_sub_sat_s_i8x16: |
| 19144 | case WebAssembly::BI__builtin_wasm_sub_sat_s_i16x8: |
| 19145 | IntNo = Intrinsic::wasm_sub_sat_signed; |
| 19146 | break; |
| 19147 | case WebAssembly::BI__builtin_wasm_sub_sat_u_i8x16: |
| 19148 | case WebAssembly::BI__builtin_wasm_sub_sat_u_i16x8: |
| 19149 | IntNo = Intrinsic::wasm_sub_sat_unsigned; |
| 19150 | break; |
| 19151 | default: |
| 19152 | llvm_unreachable("unexpected builtin ID")::llvm::llvm_unreachable_internal("unexpected builtin ID", "clang/lib/CodeGen/CGBuiltin.cpp" , 19152); |
| 19153 | } |
| 19154 | Value *LHS = EmitScalarExpr(E->getArg(0)); |
| 19155 | Value *RHS = EmitScalarExpr(E->getArg(1)); |
| 19156 | Function *Callee = CGM.getIntrinsic(IntNo, ConvertType(E->getType())); |
| 19157 | return Builder.CreateCall(Callee, {LHS, RHS}); |
| 19158 | } |
| 19159 | case WebAssembly::BI__builtin_wasm_abs_i8x16: |
| 19160 | case WebAssembly::BI__builtin_wasm_abs_i16x8: |
| 19161 | case WebAssembly::BI__builtin_wasm_abs_i32x4: |
| 19162 | case WebAssembly::BI__builtin_wasm_abs_i64x2: { |
| 19163 | Value *Vec = EmitScalarExpr(E->getArg(0)); |
| 19164 | Value *Neg = Builder.CreateNeg(Vec, "neg"); |
| 19165 | Constant *Zero = llvm::Constant::getNullValue(Vec->getType()); |
| 19166 | Value *ICmp = Builder.CreateICmpSLT(Vec, Zero, "abscond"); |
| 19167 | return Builder.CreateSelect(ICmp, Neg, Vec, "abs"); |
| 19168 | } |
| 19169 | case WebAssembly::BI__builtin_wasm_min_s_i8x16: |
| 19170 | case WebAssembly::BI__builtin_wasm_min_u_i8x16: |
| 19171 | case WebAssembly::BI__builtin_wasm_max_s_i8x16: |
| 19172 | case WebAssembly::BI__builtin_wasm_max_u_i8x16: |
| 19173 | case WebAssembly::BI__builtin_wasm_min_s_i16x8: |
| 19174 | case WebAssembly::BI__builtin_wasm_min_u_i16x8: |
| 19175 | case WebAssembly::BI__builtin_wasm_max_s_i16x8: |
| 19176 | case WebAssembly::BI__builtin_wasm_max_u_i16x8: |
| 19177 | case WebAssembly::BI__builtin_wasm_min_s_i32x4: |
| 19178 | case WebAssembly::BI__builtin_wasm_min_u_i32x4: |
| 19179 | case WebAssembly::BI__builtin_wasm_max_s_i32x4: |
| 19180 | case WebAssembly::BI__builtin_wasm_max_u_i32x4: { |
| 19181 | Value *LHS = EmitScalarExpr(E->getArg(0)); |
| 19182 | Value *RHS = EmitScalarExpr(E->getArg(1)); |
| 19183 | Value *ICmp; |
| 19184 | switch (BuiltinID) { |
| 19185 | case WebAssembly::BI__builtin_wasm_min_s_i8x16: |
| 19186 | case WebAssembly::BI__builtin_wasm_min_s_i16x8: |
| 19187 | case WebAssembly::BI__builtin_wasm_min_s_i32x4: |
| 19188 | ICmp = Builder.CreateICmpSLT(LHS, RHS); |
| 19189 | break; |
| 19190 | case WebAssembly::BI__builtin_wasm_min_u_i8x16: |
| 19191 | case WebAssembly::BI__builtin_wasm_min_u_i16x8: |
| 19192 | case WebAssembly::BI__builtin_wasm_min_u_i32x4: |
| 19193 | ICmp = Builder.CreateICmpULT(LHS, RHS); |
| 19194 | break; |
| 19195 | case WebAssembly::BI__builtin_wasm_max_s_i8x16: |
| 19196 | case WebAssembly::BI__builtin_wasm_max_s_i16x8: |
| 19197 | case WebAssembly::BI__builtin_wasm_max_s_i32x4: |
| 19198 | ICmp = Builder.CreateICmpSGT(LHS, RHS); |
| 19199 | break; |
| 19200 | case WebAssembly::BI__builtin_wasm_max_u_i8x16: |
| 19201 | case WebAssembly::BI__builtin_wasm_max_u_i16x8: |
| 19202 | case WebAssembly::BI__builtin_wasm_max_u_i32x4: |
| 19203 | ICmp = Builder.CreateICmpUGT(LHS, RHS); |
| 19204 | break; |
| 19205 | default: |
| 19206 | llvm_unreachable("unexpected builtin ID")::llvm::llvm_unreachable_internal("unexpected builtin ID", "clang/lib/CodeGen/CGBuiltin.cpp" , 19206); |
| 19207 | } |
| 19208 | return Builder.CreateSelect(ICmp, LHS, RHS); |
| 19209 | } |
| 19210 | case WebAssembly::BI__builtin_wasm_avgr_u_i8x16: |
| 19211 | case WebAssembly::BI__builtin_wasm_avgr_u_i16x8: { |
| 19212 | Value *LHS = EmitScalarExpr(E->getArg(0)); |
| 19213 | Value *RHS = EmitScalarExpr(E->getArg(1)); |
| 19214 | Function *Callee = CGM.getIntrinsic(Intrinsic::wasm_avgr_unsigned, |
| 19215 | ConvertType(E->getType())); |
| 19216 | return Builder.CreateCall(Callee, {LHS, RHS}); |
| 19217 | } |
| 19218 | case WebAssembly::BI__builtin_wasm_q15mulr_sat_s_i16x8: { |
| 19219 | Value *LHS = EmitScalarExpr(E->getArg(0)); |
| 19220 | Value *RHS = EmitScalarExpr(E->getArg(1)); |
| 19221 | Function *Callee = CGM.getIntrinsic(Intrinsic::wasm_q15mulr_sat_signed); |
| 19222 | return Builder.CreateCall(Callee, {LHS, RHS}); |
| 19223 | } |
| 19224 | case WebAssembly::BI__builtin_wasm_extadd_pairwise_i8x16_s_i16x8: |
| 19225 | case WebAssembly::BI__builtin_wasm_extadd_pairwise_i8x16_u_i16x8: |
| 19226 | case WebAssembly::BI__builtin_wasm_extadd_pairwise_i16x8_s_i32x4: |
| 19227 | case WebAssembly::BI__builtin_wasm_extadd_pairwise_i16x8_u_i32x4: { |
| 19228 | Value *Vec = EmitScalarExpr(E->getArg(0)); |
| 19229 | unsigned IntNo; |
| 19230 | switch (BuiltinID) { |
| 19231 | case WebAssembly::BI__builtin_wasm_extadd_pairwise_i8x16_s_i16x8: |
| 19232 | case WebAssembly::BI__builtin_wasm_extadd_pairwise_i16x8_s_i32x4: |
| 19233 | IntNo = Intrinsic::wasm_extadd_pairwise_signed; |
| 19234 | break; |
| 19235 | case WebAssembly::BI__builtin_wasm_extadd_pairwise_i8x16_u_i16x8: |
| 19236 | case WebAssembly::BI__builtin_wasm_extadd_pairwise_i16x8_u_i32x4: |
| 19237 | IntNo = Intrinsic::wasm_extadd_pairwise_unsigned; |
| 19238 | break; |
| 19239 | default: |
| 19240 | llvm_unreachable("unexpected builtin ID")::llvm::llvm_unreachable_internal("unexpected builtin ID", "clang/lib/CodeGen/CGBuiltin.cpp" , 19240); |
| 19241 | } |
| 19242 | |
| 19243 | Function *Callee = CGM.getIntrinsic(IntNo, ConvertType(E->getType())); |
| 19244 | return Builder.CreateCall(Callee, Vec); |
| 19245 | } |
| 19246 | case WebAssembly::BI__builtin_wasm_bitselect: { |
| 19247 | Value *V1 = EmitScalarExpr(E->getArg(0)); |
| 19248 | Value *V2 = EmitScalarExpr(E->getArg(1)); |
| 19249 | Value *C = EmitScalarExpr(E->getArg(2)); |
| 19250 | Function *Callee = |
| 19251 | CGM.getIntrinsic(Intrinsic::wasm_bitselect, ConvertType(E->getType())); |
| 19252 | return Builder.CreateCall(Callee, {V1, V2, C}); |
| 19253 | } |
| 19254 | case WebAssembly::BI__builtin_wasm_dot_s_i32x4_i16x8: { |
| 19255 | Value *LHS = EmitScalarExpr(E->getArg(0)); |
| 19256 | Value *RHS = EmitScalarExpr(E->getArg(1)); |
| 19257 | Function *Callee = CGM.getIntrinsic(Intrinsic::wasm_dot); |
| 19258 | return Builder.CreateCall(Callee, {LHS, RHS}); |
| 19259 | } |
| 19260 | case WebAssembly::BI__builtin_wasm_popcnt_i8x16: { |
| 19261 | Value *Vec = EmitScalarExpr(E->getArg(0)); |
| 19262 | Function *Callee = |
| 19263 | CGM.getIntrinsic(Intrinsic::ctpop, ConvertType(E->getType())); |
| 19264 | return Builder.CreateCall(Callee, {Vec}); |
| 19265 | } |
| 19266 | case WebAssembly::BI__builtin_wasm_any_true_v128: |
| 19267 | case WebAssembly::BI__builtin_wasm_all_true_i8x16: |
| 19268 | case WebAssembly::BI__builtin_wasm_all_true_i16x8: |
| 19269 | case WebAssembly::BI__builtin_wasm_all_true_i32x4: |
| 19270 | case WebAssembly::BI__builtin_wasm_all_true_i64x2: { |
| 19271 | unsigned IntNo; |
| 19272 | switch (BuiltinID) { |
| 19273 | case WebAssembly::BI__builtin_wasm_any_true_v128: |
| 19274 | IntNo = Intrinsic::wasm_anytrue; |
| 19275 | break; |
| 19276 | case WebAssembly::BI__builtin_wasm_all_true_i8x16: |
| 19277 | case WebAssembly::BI__builtin_wasm_all_true_i16x8: |
| 19278 | case WebAssembly::BI__builtin_wasm_all_true_i32x4: |
| 19279 | case WebAssembly::BI__builtin_wasm_all_true_i64x2: |
| 19280 | IntNo = Intrinsic::wasm_alltrue; |
| 19281 | break; |
| 19282 | default: |
| 19283 | llvm_unreachable("unexpected builtin ID")::llvm::llvm_unreachable_internal("unexpected builtin ID", "clang/lib/CodeGen/CGBuiltin.cpp" , 19283); |
| 19284 | } |
| 19285 | Value *Vec = EmitScalarExpr(E->getArg(0)); |
| 19286 | Function *Callee = CGM.getIntrinsic(IntNo, Vec->getType()); |
| 19287 | return Builder.CreateCall(Callee, {Vec}); |
| 19288 | } |
| 19289 | case WebAssembly::BI__builtin_wasm_bitmask_i8x16: |
| 19290 | case WebAssembly::BI__builtin_wasm_bitmask_i16x8: |
| 19291 | case WebAssembly::BI__builtin_wasm_bitmask_i32x4: |
| 19292 | case WebAssembly::BI__builtin_wasm_bitmask_i64x2: { |
| 19293 | Value *Vec = EmitScalarExpr(E->getArg(0)); |
| 19294 | Function *Callee = |
| 19295 | CGM.getIntrinsic(Intrinsic::wasm_bitmask, Vec->getType()); |
| 19296 | return Builder.CreateCall(Callee, {Vec}); |
| 19297 | } |
| 19298 | case WebAssembly::BI__builtin_wasm_abs_f32x4: |
| 19299 | case WebAssembly::BI__builtin_wasm_abs_f64x2: { |
| 19300 | Value *Vec = EmitScalarExpr(E->getArg(0)); |
| 19301 | Function *Callee = CGM.getIntrinsic(Intrinsic::fabs, Vec->getType()); |
| 19302 | return Builder.CreateCall(Callee, {Vec}); |
| 19303 | } |
| 19304 | case WebAssembly::BI__builtin_wasm_sqrt_f32x4: |
| 19305 | case WebAssembly::BI__builtin_wasm_sqrt_f64x2: { |
| 19306 | Value *Vec = EmitScalarExpr(E->getArg(0)); |
| 19307 | Function *Callee = CGM.getIntrinsic(Intrinsic::sqrt, Vec->getType()); |
| 19308 | return Builder.CreateCall(Callee, {Vec}); |
| 19309 | } |
| 19310 | case WebAssembly::BI__builtin_wasm_narrow_s_i8x16_i16x8: |
| 19311 | case WebAssembly::BI__builtin_wasm_narrow_u_i8x16_i16x8: |
| 19312 | case WebAssembly::BI__builtin_wasm_narrow_s_i16x8_i32x4: |
| 19313 | case WebAssembly::BI__builtin_wasm_narrow_u_i16x8_i32x4: { |
| 19314 | Value *Low = EmitScalarExpr(E->getArg(0)); |
| 19315 | Value *High = EmitScalarExpr(E->getArg(1)); |
| 19316 | unsigned IntNo; |
| 19317 | switch (BuiltinID) { |
| 19318 | case WebAssembly::BI__builtin_wasm_narrow_s_i8x16_i16x8: |
| 19319 | case WebAssembly::BI__builtin_wasm_narrow_s_i16x8_i32x4: |
| 19320 | IntNo = Intrinsic::wasm_narrow_signed; |
| 19321 | break; |
| 19322 | case WebAssembly::BI__builtin_wasm_narrow_u_i8x16_i16x8: |
| 19323 | case WebAssembly::BI__builtin_wasm_narrow_u_i16x8_i32x4: |
| 19324 | IntNo = Intrinsic::wasm_narrow_unsigned; |
| 19325 | break; |
| 19326 | default: |
| 19327 | llvm_unreachable("unexpected builtin ID")::llvm::llvm_unreachable_internal("unexpected builtin ID", "clang/lib/CodeGen/CGBuiltin.cpp" , 19327); |
| 19328 | } |
| 19329 | Function *Callee = |
| 19330 | CGM.getIntrinsic(IntNo, {ConvertType(E->getType()), Low->getType()}); |
| 19331 | return Builder.CreateCall(Callee, {Low, High}); |
| 19332 | } |
| 19333 | case WebAssembly::BI__builtin_wasm_trunc_sat_s_zero_f64x2_i32x4: |
| 19334 | case WebAssembly::BI__builtin_wasm_trunc_sat_u_zero_f64x2_i32x4: { |
| 19335 | Value *Vec = EmitScalarExpr(E->getArg(0)); |
| 19336 | unsigned IntNo; |
| 19337 | switch (BuiltinID) { |
| 19338 | case WebAssembly::BI__builtin_wasm_trunc_sat_s_zero_f64x2_i32x4: |
| 19339 | IntNo = Intrinsic::fptosi_sat; |
| 19340 | break; |
| 19341 | case WebAssembly::BI__builtin_wasm_trunc_sat_u_zero_f64x2_i32x4: |
| 19342 | IntNo = Intrinsic::fptoui_sat; |
| 19343 | break; |
| 19344 | default: |
| 19345 | llvm_unreachable("unexpected builtin ID")::llvm::llvm_unreachable_internal("unexpected builtin ID", "clang/lib/CodeGen/CGBuiltin.cpp" , 19345); |
| 19346 | } |
| 19347 | llvm::Type *SrcT = Vec->getType(); |
| 19348 | llvm::Type *TruncT = SrcT->getWithNewType(Builder.getInt32Ty()); |
| 19349 | Function *Callee = CGM.getIntrinsic(IntNo, {TruncT, SrcT}); |
| 19350 | Value *Trunc = Builder.CreateCall(Callee, Vec); |
| 19351 | Value *Splat = Constant::getNullValue(TruncT); |
| 19352 | return Builder.CreateShuffleVector(Trunc, Splat, ArrayRef<int>{0, 1, 2, 3}); |
| 19353 | } |
| 19354 | case WebAssembly::BI__builtin_wasm_shuffle_i8x16: { |
| 19355 | Value *Ops[18]; |
| 19356 | size_t OpIdx = 0; |
| 19357 | Ops[OpIdx++] = EmitScalarExpr(E->getArg(0)); |
| 19358 | Ops[OpIdx++] = EmitScalarExpr(E->getArg(1)); |
| 19359 | while (OpIdx < 18) { |
| 19360 | std::optional<llvm::APSInt> LaneConst = |
| 19361 | E->getArg(OpIdx)->getIntegerConstantExpr(getContext()); |
| 19362 | assert(LaneConst && "Constant arg isn't actually constant?")(static_cast <bool> (LaneConst && "Constant arg isn't actually constant?" ) ? void (0) : __assert_fail ("LaneConst && \"Constant arg isn't actually constant?\"" , "clang/lib/CodeGen/CGBuiltin.cpp", 19362, __extension__ __PRETTY_FUNCTION__ )); |
| 19363 | Ops[OpIdx++] = llvm::ConstantInt::get(getLLVMContext(), *LaneConst); |
| 19364 | } |
| 19365 | Function *Callee = CGM.getIntrinsic(Intrinsic::wasm_shuffle); |
| 19366 | return Builder.CreateCall(Callee, Ops); |
| 19367 | } |
| 19368 | case WebAssembly::BI__builtin_wasm_relaxed_madd_f32x4: |
| 19369 | case WebAssembly::BI__builtin_wasm_relaxed_nmadd_f32x4: |
| 19370 | case WebAssembly::BI__builtin_wasm_relaxed_madd_f64x2: |
| 19371 | case WebAssembly::BI__builtin_wasm_relaxed_nmadd_f64x2: { |
| 19372 | Value *A = EmitScalarExpr(E->getArg(0)); |
| 19373 | Value *B = EmitScalarExpr(E->getArg(1)); |
| 19374 | Value *C = EmitScalarExpr(E->getArg(2)); |
| 19375 | unsigned IntNo; |
| 19376 | switch (BuiltinID) { |
| 19377 | case WebAssembly::BI__builtin_wasm_relaxed_madd_f32x4: |
| 19378 | case WebAssembly::BI__builtin_wasm_relaxed_madd_f64x2: |
| 19379 | IntNo = Intrinsic::wasm_relaxed_madd; |
| 19380 | break; |
| 19381 | case WebAssembly::BI__builtin_wasm_relaxed_nmadd_f32x4: |
| 19382 | case WebAssembly::BI__builtin_wasm_relaxed_nmadd_f64x2: |
| 19383 | IntNo = Intrinsic::wasm_relaxed_nmadd; |
| 19384 | break; |
| 19385 | default: |
| 19386 | llvm_unreachable("unexpected builtin ID")::llvm::llvm_unreachable_internal("unexpected builtin ID", "clang/lib/CodeGen/CGBuiltin.cpp" , 19386); |
| 19387 | } |
| 19388 | Function *Callee = CGM.getIntrinsic(IntNo, A->getType()); |
| 19389 | return Builder.CreateCall(Callee, {A, B, C}); |
| 19390 | } |
| 19391 | case WebAssembly::BI__builtin_wasm_relaxed_laneselect_i8x16: |
| 19392 | case WebAssembly::BI__builtin_wasm_relaxed_laneselect_i16x8: |
| 19393 | case WebAssembly::BI__builtin_wasm_relaxed_laneselect_i32x4: |
| 19394 | case WebAssembly::BI__builtin_wasm_relaxed_laneselect_i64x2: { |
| 19395 | Value *A = EmitScalarExpr(E->getArg(0)); |
| 19396 | Value *B = EmitScalarExpr(E->getArg(1)); |
| 19397 | Value *C = EmitScalarExpr(E->getArg(2)); |
| 19398 | Function *Callee = |
| 19399 | CGM.getIntrinsic(Intrinsic::wasm_relaxed_laneselect, A->getType()); |
| 19400 | return Builder.CreateCall(Callee, {A, B, C}); |
| 19401 | } |
| 19402 | case WebAssembly::BI__builtin_wasm_relaxed_swizzle_i8x16: { |
| 19403 | Value *Src = EmitScalarExpr(E->getArg(0)); |
| 19404 | Value *Indices = EmitScalarExpr(E->getArg(1)); |
| 19405 | Function *Callee = CGM.getIntrinsic(Intrinsic::wasm_relaxed_swizzle); |
| 19406 | return Builder.CreateCall(Callee, {Src, Indices}); |
| 19407 | } |
| 19408 | case WebAssembly::BI__builtin_wasm_relaxed_min_f32x4: |
| 19409 | case WebAssembly::BI__builtin_wasm_relaxed_max_f32x4: |
| 19410 | case WebAssembly::BI__builtin_wasm_relaxed_min_f64x2: |
| 19411 | case WebAssembly::BI__builtin_wasm_relaxed_max_f64x2: { |
| 19412 | Value *LHS = EmitScalarExpr(E->getArg(0)); |
| 19413 | Value *RHS = EmitScalarExpr(E->getArg(1)); |
| 19414 | unsigned IntNo; |
| 19415 | switch (BuiltinID) { |
| 19416 | case WebAssembly::BI__builtin_wasm_relaxed_min_f32x4: |
| 19417 | case WebAssembly::BI__builtin_wasm_relaxed_min_f64x2: |
| 19418 | IntNo = Intrinsic::wasm_relaxed_min; |
| 19419 | break; |
| 19420 | case WebAssembly::BI__builtin_wasm_relaxed_max_f32x4: |
| 19421 | case WebAssembly::BI__builtin_wasm_relaxed_max_f64x2: |
| 19422 | IntNo = Intrinsic::wasm_relaxed_max; |
| 19423 | break; |
| 19424 | default: |
| 19425 | llvm_unreachable("unexpected builtin ID")::llvm::llvm_unreachable_internal("unexpected builtin ID", "clang/lib/CodeGen/CGBuiltin.cpp" , 19425); |
| 19426 | } |
| 19427 | Function *Callee = CGM.getIntrinsic(IntNo, LHS->getType()); |
| 19428 | return Builder.CreateCall(Callee, {LHS, RHS}); |
| 19429 | } |
| 19430 | case WebAssembly::BI__builtin_wasm_relaxed_trunc_s_i32x4_f32x4: |
| 19431 | case WebAssembly::BI__builtin_wasm_relaxed_trunc_u_i32x4_f32x4: |
| 19432 | case WebAssembly::BI__builtin_wasm_relaxed_trunc_s_zero_i32x4_f64x2: |
| 19433 | case WebAssembly::BI__builtin_wasm_relaxed_trunc_u_zero_i32x4_f64x2: { |
| 19434 | Value *Vec = EmitScalarExpr(E->getArg(0)); |
| 19435 | unsigned IntNo; |
| 19436 | switch (BuiltinID) { |
| 19437 | case WebAssembly::BI__builtin_wasm_relaxed_trunc_s_i32x4_f32x4: |
| 19438 | IntNo = Intrinsic::wasm_relaxed_trunc_signed; |
| 19439 | break; |
| 19440 | case WebAssembly::BI__builtin_wasm_relaxed_trunc_u_i32x4_f32x4: |
| 19441 | IntNo = Intrinsic::wasm_relaxed_trunc_unsigned; |
| 19442 | break; |
| 19443 | case WebAssembly::BI__builtin_wasm_relaxed_trunc_s_zero_i32x4_f64x2: |
| 19444 | IntNo = Intrinsic::wasm_relaxed_trunc_signed_zero; |
| 19445 | break; |
| 19446 | case WebAssembly::BI__builtin_wasm_relaxed_trunc_u_zero_i32x4_f64x2: |
| 19447 | IntNo = Intrinsic::wasm_relaxed_trunc_unsigned_zero; |
| 19448 | break; |
| 19449 | default: |
| 19450 | llvm_unreachable("unexpected builtin ID")::llvm::llvm_unreachable_internal("unexpected builtin ID", "clang/lib/CodeGen/CGBuiltin.cpp" , 19450); |
| 19451 | } |
| 19452 | Function *Callee = CGM.getIntrinsic(IntNo); |
| 19453 | return Builder.CreateCall(Callee, {Vec}); |
| 19454 | } |
| 19455 | case WebAssembly::BI__builtin_wasm_relaxed_q15mulr_s_i16x8: { |
| 19456 | Value *LHS = EmitScalarExpr(E->getArg(0)); |
| 19457 | Value *RHS = EmitScalarExpr(E->getArg(1)); |
| 19458 | Function *Callee = CGM.getIntrinsic(Intrinsic::wasm_relaxed_q15mulr_signed); |
| 19459 | return Builder.CreateCall(Callee, {LHS, RHS}); |
| 19460 | } |
| 19461 | case WebAssembly::BI__builtin_wasm_relaxed_dot_i8x16_i7x16_s_i16x8: { |
| 19462 | Value *LHS = EmitScalarExpr(E->getArg(0)); |
| 19463 | Value *RHS = EmitScalarExpr(E->getArg(1)); |
| 19464 | Function *Callee = |
| 19465 | CGM.getIntrinsic(Intrinsic::wasm_relaxed_dot_i8x16_i7x16_signed); |
| 19466 | return Builder.CreateCall(Callee, {LHS, RHS}); |
| 19467 | } |
| 19468 | case WebAssembly::BI__builtin_wasm_relaxed_dot_i8x16_i7x16_add_s_i32x4: { |
| 19469 | Value *LHS = EmitScalarExpr(E->getArg(0)); |
| 19470 | Value *RHS = EmitScalarExpr(E->getArg(1)); |
| 19471 | Value *Acc = EmitScalarExpr(E->getArg(2)); |
| 19472 | Function *Callee = |
| 19473 | CGM.getIntrinsic(Intrinsic::wasm_relaxed_dot_i8x16_i7x16_add_signed); |
| 19474 | return Builder.CreateCall(Callee, {LHS, RHS, Acc}); |
| 19475 | } |
| 19476 | case WebAssembly::BI__builtin_wasm_relaxed_dot_bf16x8_add_f32_f32x4: { |
| 19477 | Value *LHS = EmitScalarExpr(E->getArg(0)); |
| 19478 | Value *RHS = EmitScalarExpr(E->getArg(1)); |
| 19479 | Value *Acc = EmitScalarExpr(E->getArg(2)); |
| 19480 | Function *Callee = |
| 19481 | CGM.getIntrinsic(Intrinsic::wasm_relaxed_dot_bf16x8_add_f32); |
| 19482 | return Builder.CreateCall(Callee, {LHS, RHS, Acc}); |
| 19483 | } |
| 19484 | default: |
| 19485 | return nullptr; |
| 19486 | } |
| 19487 | } |
| 19488 | |
| 19489 | static std::pair<Intrinsic::ID, unsigned> |
| 19490 | getIntrinsicForHexagonNonClangBuiltin(unsigned BuiltinID) { |
| 19491 | struct Info { |
| 19492 | unsigned BuiltinID; |
| 19493 | Intrinsic::ID IntrinsicID; |
| 19494 | unsigned VecLen; |
| 19495 | }; |
| 19496 | static Info Infos[] = { |
| 19497 | #define CUSTOM_BUILTIN_MAPPING(x,s) \ |
| 19498 | { Hexagon::BI__builtin_HEXAGON_##x, Intrinsic::hexagon_##x, s }, |
| 19499 | CUSTOM_BUILTIN_MAPPING(L2_loadrub_pci, 0) |
| 19500 | CUSTOM_BUILTIN_MAPPING(L2_loadrb_pci, 0) |
| 19501 | CUSTOM_BUILTIN_MAPPING(L2_loadruh_pci, 0) |
| 19502 | CUSTOM_BUILTIN_MAPPING(L2_loadrh_pci, 0) |
| 19503 | CUSTOM_BUILTIN_MAPPING(L2_loadri_pci, 0) |
| 19504 | CUSTOM_BUILTIN_MAPPING(L2_loadrd_pci, 0) |
| 19505 | CUSTOM_BUILTIN_MAPPING(L2_loadrub_pcr, 0) |
| 19506 | CUSTOM_BUILTIN_MAPPING(L2_loadrb_pcr, 0) |
| 19507 | CUSTOM_BUILTIN_MAPPING(L2_loadruh_pcr, 0) |
| 19508 | CUSTOM_BUILTIN_MAPPING(L2_loadrh_pcr, 0) |
| 19509 | CUSTOM_BUILTIN_MAPPING(L2_loadri_pcr, 0) |
| 19510 | CUSTOM_BUILTIN_MAPPING(L2_loadrd_pcr, 0) |
| 19511 | CUSTOM_BUILTIN_MAPPING(S2_storerb_pci, 0) |
| 19512 | CUSTOM_BUILTIN_MAPPING(S2_storerh_pci, 0) |
| 19513 | CUSTOM_BUILTIN_MAPPING(S2_storerf_pci, 0) |
| 19514 | CUSTOM_BUILTIN_MAPPING(S2_storeri_pci, 0) |
| 19515 | CUSTOM_BUILTIN_MAPPING(S2_storerd_pci, 0) |
| 19516 | CUSTOM_BUILTIN_MAPPING(S2_storerb_pcr, 0) |
| 19517 | CUSTOM_BUILTIN_MAPPING(S2_storerh_pcr, 0) |
| 19518 | CUSTOM_BUILTIN_MAPPING(S2_storerf_pcr, 0) |
| 19519 | CUSTOM_BUILTIN_MAPPING(S2_storeri_pcr, 0) |
| 19520 | CUSTOM_BUILTIN_MAPPING(S2_storerd_pcr, 0) |
| 19521 | // Legacy builtins that take a vector in place of a vector predicate. |
| 19522 | CUSTOM_BUILTIN_MAPPING(V6_vmaskedstoreq, 64) |
| 19523 | CUSTOM_BUILTIN_MAPPING(V6_vmaskedstorenq, 64) |
| 19524 | CUSTOM_BUILTIN_MAPPING(V6_vmaskedstorentq, 64) |
| 19525 | CUSTOM_BUILTIN_MAPPING(V6_vmaskedstorentnq, 64) |
| 19526 | CUSTOM_BUILTIN_MAPPING(V6_vmaskedstoreq_128B, 128) |
| 19527 | CUSTOM_BUILTIN_MAPPING(V6_vmaskedstorenq_128B, 128) |
| 19528 | CUSTOM_BUILTIN_MAPPING(V6_vmaskedstorentq_128B, 128) |
| 19529 | CUSTOM_BUILTIN_MAPPING(V6_vmaskedstorentnq_128B, 128) |
| 19530 | #include "clang/Basic/BuiltinsHexagonMapCustomDep.def" |
| 19531 | #undef CUSTOM_BUILTIN_MAPPING |
| 19532 | }; |
| 19533 | |
| 19534 | auto CmpInfo = [] (Info A, Info B) { return A.BuiltinID < B.BuiltinID; }; |
| 19535 | static const bool SortOnce = (llvm::sort(Infos, CmpInfo), true); |
| 19536 | (void)SortOnce; |
| 19537 | |
| 19538 | const Info *F = llvm::lower_bound(Infos, Info{BuiltinID, 0, 0}, CmpInfo); |
| 19539 | if (F == std::end(Infos) || F->BuiltinID != BuiltinID) |
| 19540 | return {Intrinsic::not_intrinsic, 0}; |
| 19541 | |
| 19542 | return {F->IntrinsicID, F->VecLen}; |
| 19543 | } |
| 19544 | |
| 19545 | Value *CodeGenFunction::EmitHexagonBuiltinExpr(unsigned BuiltinID, |
| 19546 | const CallExpr *E) { |
| 19547 | Intrinsic::ID ID; |
| 19548 | unsigned VecLen; |
| 19549 | std::tie(ID, VecLen) = getIntrinsicForHexagonNonClangBuiltin(BuiltinID); |
| 19550 | |
| 19551 | auto MakeCircOp = [this, E](unsigned IntID, bool IsLoad) { |
| 19552 | // The base pointer is passed by address, so it needs to be loaded. |
| 19553 | Address A = EmitPointerWithAlignment(E->getArg(0)); |
| 19554 | Address BP = Address(Builder.CreateBitCast( |
| 19555 | A.getPointer(), Int8PtrPtrTy), Int8PtrTy, A.getAlignment()); |
| 19556 | llvm::Value *Base = Builder.CreateLoad(BP); |
| 19557 | // The treatment of both loads and stores is the same: the arguments for |
| 19558 | // the builtin are the same as the arguments for the intrinsic. |
| 19559 | // Load: |
| 19560 | // builtin(Base, Inc, Mod, Start) -> intr(Base, Inc, Mod, Start) |
| 19561 | // builtin(Base, Mod, Start) -> intr(Base, Mod, Start) |
| 19562 | // Store: |
| 19563 | // builtin(Base, Inc, Mod, Val, Start) -> intr(Base, Inc, Mod, Val, Start) |
| 19564 | // builtin(Base, Mod, Val, Start) -> intr(Base, Mod, Val, Start) |
| 19565 | SmallVector<llvm::Value*,5> Ops = { Base }; |
| 19566 | for (unsigned i = 1, e = E->getNumArgs(); i != e; ++i) |
| 19567 | Ops.push_back(EmitScalarExpr(E->getArg(i))); |
| 19568 | |
| 19569 | llvm::Value *Result = Builder.CreateCall(CGM.getIntrinsic(IntID), Ops); |
| 19570 | // The load intrinsics generate two results (Value, NewBase), stores |
| 19571 | // generate one (NewBase). The new base address needs to be stored. |
| 19572 | llvm::Value *NewBase = IsLoad ? Builder.CreateExtractValue(Result, 1) |
| 19573 | : Result; |
| 19574 | llvm::Value *LV = Builder.CreateBitCast( |
| 19575 | EmitScalarExpr(E->getArg(0)), NewBase->getType()->getPointerTo()); |
| 19576 | Address Dest = EmitPointerWithAlignment(E->getArg(0)); |
| 19577 | llvm::Value *RetVal = |
| 19578 | Builder.CreateAlignedStore(NewBase, LV, Dest.getAlignment()); |
| 19579 | if (IsLoad) |
| 19580 | RetVal = Builder.CreateExtractValue(Result, 0); |
| 19581 | return RetVal; |
| 19582 | }; |
| 19583 | |
| 19584 | // Handle the conversion of bit-reverse load intrinsics to bit code. |
| 19585 | // The intrinsic call after this function only reads from memory and the |
| 19586 | // write to memory is dealt by the store instruction. |
| 19587 | auto MakeBrevLd = [this, E](unsigned IntID, llvm::Type *DestTy) { |
| 19588 | // The intrinsic generates one result, which is the new value for the base |
| 19589 | // pointer. It needs to be returned. The result of the load instruction is |
| 19590 | // passed to intrinsic by address, so the value needs to be stored. |
| 19591 | llvm::Value *BaseAddress = |
| 19592 | Builder.CreateBitCast(EmitScalarExpr(E->getArg(0)), Int8PtrTy); |
| 19593 | |
| 19594 | // Expressions like &(*pt++) will be incremented per evaluation. |
| 19595 | // EmitPointerWithAlignment and EmitScalarExpr evaluates the expression |
| 19596 | // per call. |
| 19597 | Address DestAddr = EmitPointerWithAlignment(E->getArg(1)); |
| 19598 | DestAddr = Address(Builder.CreateBitCast(DestAddr.getPointer(), Int8PtrTy), |
| 19599 | Int8Ty, DestAddr.getAlignment()); |
| 19600 | llvm::Value *DestAddress = DestAddr.getPointer(); |
| 19601 | |
| 19602 | // Operands are Base, Dest, Modifier. |
| 19603 | // The intrinsic format in LLVM IR is defined as |
| 19604 | // { ValueType, i8* } (i8*, i32). |
| 19605 | llvm::Value *Result = Builder.CreateCall( |
| 19606 | CGM.getIntrinsic(IntID), {BaseAddress, EmitScalarExpr(E->getArg(2))}); |
| 19607 | |
| 19608 | // The value needs to be stored as the variable is passed by reference. |
| 19609 | llvm::Value *DestVal = Builder.CreateExtractValue(Result, 0); |
| 19610 | |
| 19611 | // The store needs to be truncated to fit the destination type. |
| 19612 | // While i32 and i64 are natively supported on Hexagon, i8 and i16 needs |
| 19613 | // to be handled with stores of respective destination type. |
| 19614 | DestVal = Builder.CreateTrunc(DestVal, DestTy); |
| 19615 | |
| 19616 | llvm::Value *DestForStore = |
| 19617 | Builder.CreateBitCast(DestAddress, DestVal->getType()->getPointerTo()); |
| 19618 | Builder.CreateAlignedStore(DestVal, DestForStore, DestAddr.getAlignment()); |
| 19619 | // The updated value of the base pointer is returned. |
| 19620 | return Builder.CreateExtractValue(Result, 1); |
| 19621 | }; |
| 19622 | |
| 19623 | auto V2Q = [this, VecLen] (llvm::Value *Vec) { |
| 19624 | Intrinsic::ID ID = VecLen == 128 ? Intrinsic::hexagon_V6_vandvrt_128B |
| 19625 | : Intrinsic::hexagon_V6_vandvrt; |
| 19626 | return Builder.CreateCall(CGM.getIntrinsic(ID), |
| 19627 | {Vec, Builder.getInt32(-1)}); |
| 19628 | }; |
| 19629 | auto Q2V = [this, VecLen] (llvm::Value *Pred) { |
| 19630 | Intrinsic::ID ID = VecLen == 128 ? Intrinsic::hexagon_V6_vandqrt_128B |
| 19631 | : Intrinsic::hexagon_V6_vandqrt; |
| 19632 | return Builder.CreateCall(CGM.getIntrinsic(ID), |
| 19633 | {Pred, Builder.getInt32(-1)}); |
| 19634 | }; |
| 19635 | |
| 19636 | switch (BuiltinID) { |
| 19637 | // These intrinsics return a tuple {Vector, VectorPred} in LLVM IR, |
| 19638 | // and the corresponding C/C++ builtins use loads/stores to update |
| 19639 | // the predicate. |
| 19640 | case Hexagon::BI__builtin_HEXAGON_V6_vaddcarry: |
| 19641 | case Hexagon::BI__builtin_HEXAGON_V6_vaddcarry_128B: |
| 19642 | case Hexagon::BI__builtin_HEXAGON_V6_vsubcarry: |
| 19643 | case Hexagon::BI__builtin_HEXAGON_V6_vsubcarry_128B: { |
| 19644 | // Get the type from the 0-th argument. |
| 19645 | llvm::Type *VecType = ConvertType(E->getArg(0)->getType()); |
| 19646 | Address PredAddr = Builder.CreateElementBitCast( |
| 19647 | EmitPointerWithAlignment(E->getArg(2)), VecType); |
| 19648 | llvm::Value *PredIn = V2Q(Builder.CreateLoad(PredAddr)); |
| 19649 | llvm::Value *Result = Builder.CreateCall(CGM.getIntrinsic(ID), |
| 19650 | {EmitScalarExpr(E->getArg(0)), EmitScalarExpr(E->getArg(1)), PredIn}); |
| 19651 | |
| 19652 | llvm::Value *PredOut = Builder.CreateExtractValue(Result, 1); |
| 19653 | Builder.CreateAlignedStore(Q2V(PredOut), PredAddr.getPointer(), |
| 19654 | PredAddr.getAlignment()); |
| 19655 | return Builder.CreateExtractValue(Result, 0); |
| 19656 | } |
| 19657 | // These are identical to the builtins above, except they don't consume |
| 19658 | // input carry, only generate carry-out. Since they still produce two |
| 19659 | // outputs, generate the store of the predicate, but no load. |
| 19660 | case Hexagon::BI__builtin_HEXAGON_V6_vaddcarryo: |
| 19661 | case Hexagon::BI__builtin_HEXAGON_V6_vaddcarryo_128B: |
| 19662 | case Hexagon::BI__builtin_HEXAGON_V6_vsubcarryo: |
| 19663 | case Hexagon::BI__builtin_HEXAGON_V6_vsubcarryo_128B: { |
| 19664 | // Get the type from the 0-th argument. |
| 19665 | llvm::Type *VecType = ConvertType(E->getArg(0)->getType()); |
| 19666 | Address PredAddr = Builder.CreateElementBitCast( |
| 19667 | EmitPointerWithAlignment(E->getArg(2)), VecType); |
| 19668 | llvm::Value *Result = Builder.CreateCall(CGM.getIntrinsic(ID), |
| 19669 | {EmitScalarExpr(E->getArg(0)), EmitScalarExpr(E->getArg(1))}); |
| 19670 | |
| 19671 | llvm::Value *PredOut = Builder.CreateExtractValue(Result, 1); |
| 19672 | Builder.CreateAlignedStore(Q2V(PredOut), PredAddr.getPointer(), |
| 19673 | PredAddr.getAlignment()); |
| 19674 | return Builder.CreateExtractValue(Result, 0); |
| 19675 | } |
| 19676 | |
| 19677 | case Hexagon::BI__builtin_HEXAGON_V6_vmaskedstoreq: |
| 19678 | case Hexagon::BI__builtin_HEXAGON_V6_vmaskedstorenq: |
| 19679 | case Hexagon::BI__builtin_HEXAGON_V6_vmaskedstorentq: |
| 19680 | case Hexagon::BI__builtin_HEXAGON_V6_vmaskedstorentnq: |
| 19681 | case Hexagon::BI__builtin_HEXAGON_V6_vmaskedstoreq_128B: |
| 19682 | case Hexagon::BI__builtin_HEXAGON_V6_vmaskedstorenq_128B: |
| 19683 | case Hexagon::BI__builtin_HEXAGON_V6_vmaskedstorentq_128B: |
| 19684 | case Hexagon::BI__builtin_HEXAGON_V6_vmaskedstorentnq_128B: { |
| 19685 | SmallVector<llvm::Value*,4> Ops; |
| 19686 | const Expr *PredOp = E->getArg(0); |
| 19687 | // There will be an implicit cast to a boolean vector. Strip it. |
| 19688 | if (auto *Cast = dyn_cast<ImplicitCastExpr>(PredOp)) { |
| 19689 | if (Cast->getCastKind() == CK_BitCast) |
| 19690 | PredOp = Cast->getSubExpr(); |
| 19691 | Ops.push_back(V2Q(EmitScalarExpr(PredOp))); |
| 19692 | } |
| 19693 | for (int i = 1, e = E->getNumArgs(); i != e; ++i) |
| 19694 | Ops.push_back(EmitScalarExpr(E->getArg(i))); |
| 19695 | return Builder.CreateCall(CGM.getIntrinsic(ID), Ops); |
| 19696 | } |
| 19697 | |
| 19698 | case Hexagon::BI__builtin_HEXAGON_L2_loadrub_pci: |
| 19699 | case Hexagon::BI__builtin_HEXAGON_L2_loadrb_pci: |
| 19700 | case Hexagon::BI__builtin_HEXAGON_L2_loadruh_pci: |
| 19701 | case Hexagon::BI__builtin_HEXAGON_L2_loadrh_pci: |
| 19702 | case Hexagon::BI__builtin_HEXAGON_L2_loadri_pci: |
| 19703 | case Hexagon::BI__builtin_HEXAGON_L2_loadrd_pci: |
| 19704 | case Hexagon::BI__builtin_HEXAGON_L2_loadrub_pcr: |
| 19705 | case Hexagon::BI__builtin_HEXAGON_L2_loadrb_pcr: |
| 19706 | case Hexagon::BI__builtin_HEXAGON_L2_loadruh_pcr: |
| 19707 | case Hexagon::BI__builtin_HEXAGON_L2_loadrh_pcr: |
| 19708 | case Hexagon::BI__builtin_HEXAGON_L2_loadri_pcr: |
| 19709 | case Hexagon::BI__builtin_HEXAGON_L2_loadrd_pcr: |
| 19710 | return MakeCircOp(ID, /*IsLoad=*/true); |
| 19711 | case Hexagon::BI__builtin_HEXAGON_S2_storerb_pci: |
| 19712 | case Hexagon::BI__builtin_HEXAGON_S2_storerh_pci: |
| 19713 | case Hexagon::BI__builtin_HEXAGON_S2_storerf_pci: |
| 19714 | case Hexagon::BI__builtin_HEXAGON_S2_storeri_pci: |
| 19715 | case Hexagon::BI__builtin_HEXAGON_S2_storerd_pci: |
| 19716 | case Hexagon::BI__builtin_HEXAGON_S2_storerb_pcr: |
| 19717 | case Hexagon::BI__builtin_HEXAGON_S2_storerh_pcr: |
| 19718 | case Hexagon::BI__builtin_HEXAGON_S2_storerf_pcr: |
| 19719 | case Hexagon::BI__builtin_HEXAGON_S2_storeri_pcr: |
| 19720 | case Hexagon::BI__builtin_HEXAGON_S2_storerd_pcr: |
| 19721 | return MakeCircOp(ID, /*IsLoad=*/false); |
| 19722 | case Hexagon::BI__builtin_brev_ldub: |
| 19723 | return MakeBrevLd(Intrinsic::hexagon_L2_loadrub_pbr, Int8Ty); |
| 19724 | case Hexagon::BI__builtin_brev_ldb: |
| 19725 | return MakeBrevLd(Intrinsic::hexagon_L2_loadrb_pbr, Int8Ty); |
| 19726 | case Hexagon::BI__builtin_brev_lduh: |
| 19727 | return MakeBrevLd(Intrinsic::hexagon_L2_loadruh_pbr, Int16Ty); |
| 19728 | case Hexagon::BI__builtin_brev_ldh: |
| 19729 | return MakeBrevLd(Intrinsic::hexagon_L2_loadrh_pbr, Int16Ty); |
| 19730 | case Hexagon::BI__builtin_brev_ldw: |
| 19731 | return MakeBrevLd(Intrinsic::hexagon_L2_loadri_pbr, Int32Ty); |
| 19732 | case Hexagon::BI__builtin_brev_ldd: |
| 19733 | return MakeBrevLd(Intrinsic::hexagon_L2_loadrd_pbr, Int64Ty); |
| 19734 | } // switch |
| 19735 | |
| 19736 | return nullptr; |
| 19737 | } |
| 19738 | |
| 19739 | Value *CodeGenFunction::EmitRISCVBuiltinExpr(unsigned BuiltinID, |
| 19740 | const CallExpr *E, |
| 19741 | ReturnValueSlot ReturnValue) { |
| 19742 | SmallVector<Value *, 4> Ops; |
| 19743 | llvm::Type *ResultType = ConvertType(E->getType()); |
| 19744 | |
| 19745 | // Find out if any arguments are required to be integer constant expressions. |
| 19746 | unsigned ICEArguments = 0; |
| 19747 | ASTContext::GetBuiltinTypeError Error; |
| 19748 | getContext().GetBuiltinType(BuiltinID, Error, &ICEArguments); |
| 19749 | if (Error == ASTContext::GE_Missing_type) { |
| 19750 | // Vector intrinsics don't have a type string. |
| 19751 | assert(BuiltinID >= clang::RISCV::FirstRVVBuiltin &&(static_cast <bool> (BuiltinID >= clang::RISCV::FirstRVVBuiltin && BuiltinID <= clang::RISCV::LastRVVBuiltin) ? void (0) : __assert_fail ("BuiltinID >= clang::RISCV::FirstRVVBuiltin && BuiltinID <= clang::RISCV::LastRVVBuiltin" , "clang/lib/CodeGen/CGBuiltin.cpp", 19752, __extension__ __PRETTY_FUNCTION__ )) |
| 19752 | BuiltinID <= clang::RISCV::LastRVVBuiltin)(static_cast <bool> (BuiltinID >= clang::RISCV::FirstRVVBuiltin && BuiltinID <= clang::RISCV::LastRVVBuiltin) ? void (0) : __assert_fail ("BuiltinID >= clang::RISCV::FirstRVVBuiltin && BuiltinID <= clang::RISCV::LastRVVBuiltin" , "clang/lib/CodeGen/CGBuiltin.cpp", 19752, __extension__ __PRETTY_FUNCTION__ )); |
| 19753 | ICEArguments = 0; |
| 19754 | if (BuiltinID == RISCVVector::BI__builtin_rvv_vget_v || |
| 19755 | BuiltinID == RISCVVector::BI__builtin_rvv_vset_v) |
| 19756 | ICEArguments = 1 << 1; |
| 19757 | } else { |
| 19758 | assert(Error == ASTContext::GE_None && "Unexpected error")(static_cast <bool> (Error == ASTContext::GE_None && "Unexpected error") ? void (0) : __assert_fail ("Error == ASTContext::GE_None && \"Unexpected error\"" , "clang/lib/CodeGen/CGBuiltin.cpp", 19758, __extension__ __PRETTY_FUNCTION__ )); |
| 19759 | } |
| 19760 | |
| 19761 | if (BuiltinID == RISCV::BI__builtin_riscv_ntl_load) |
| 19762 | ICEArguments |= (1 << 1); |
| 19763 | if (BuiltinID == RISCV::BI__builtin_riscv_ntl_store) |
| 19764 | ICEArguments |= (1 << 2); |
| 19765 | |
| 19766 | for (unsigned i = 0, e = E->getNumArgs(); i != e; i++) { |
| 19767 | // If this is a normal argument, just emit it as a scalar. |
| 19768 | if ((ICEArguments & (1 << i)) == 0) { |
| 19769 | Ops.push_back(EmitScalarExpr(E->getArg(i))); |
| 19770 | continue; |
| 19771 | } |
| 19772 | |
| 19773 | // If this is required to be a constant, constant fold it so that we know |
| 19774 | // that the generated intrinsic gets a ConstantInt. |
| 19775 | Ops.push_back(llvm::ConstantInt::get( |
| 19776 | getLLVMContext(), *E->getArg(i)->getIntegerConstantExpr(getContext()))); |
| 19777 | } |
| 19778 | |
| 19779 | Intrinsic::ID ID = Intrinsic::not_intrinsic; |
| 19780 | unsigned NF = 1; |
| 19781 | // The 0th bit simulates the `vta` of RVV |
| 19782 | // The 1st bit simulates the `vma` of RVV |
| 19783 | constexpr unsigned RVV_VTA = 0x1; |
| 19784 | constexpr unsigned RVV_VMA = 0x2; |
| 19785 | int PolicyAttrs = 0; |
| 19786 | bool IsMasked = false; |
| 19787 | |
| 19788 | // Required for overloaded intrinsics. |
| 19789 | llvm::SmallVector<llvm::Type *, 2> IntrinsicTypes; |
| 19790 | switch (BuiltinID) { |
| 19791 | default: llvm_unreachable("unexpected builtin ID")::llvm::llvm_unreachable_internal("unexpected builtin ID", "clang/lib/CodeGen/CGBuiltin.cpp" , 19791); |
| 19792 | case RISCV::BI__builtin_riscv_orc_b_32: |
| 19793 | case RISCV::BI__builtin_riscv_orc_b_64: |
| 19794 | case RISCV::BI__builtin_riscv_clz_32: |
| 19795 | case RISCV::BI__builtin_riscv_clz_64: |
| 19796 | case RISCV::BI__builtin_riscv_ctz_32: |
| 19797 | case RISCV::BI__builtin_riscv_ctz_64: |
| 19798 | case RISCV::BI__builtin_riscv_clmul: |
| 19799 | case RISCV::BI__builtin_riscv_clmulh: |
| 19800 | case RISCV::BI__builtin_riscv_clmulr: |
| 19801 | case RISCV::BI__builtin_riscv_xperm4: |
| 19802 | case RISCV::BI__builtin_riscv_xperm8: |
| 19803 | case RISCV::BI__builtin_riscv_brev8: |
| 19804 | case RISCV::BI__builtin_riscv_zip_32: |
| 19805 | case RISCV::BI__builtin_riscv_unzip_32: { |
| 19806 | switch (BuiltinID) { |
| 19807 | default: llvm_unreachable("unexpected builtin ID")::llvm::llvm_unreachable_internal("unexpected builtin ID", "clang/lib/CodeGen/CGBuiltin.cpp" , 19807); |
| 19808 | // Zbb |
| 19809 | case RISCV::BI__builtin_riscv_orc_b_32: |
| 19810 | case RISCV::BI__builtin_riscv_orc_b_64: |
| 19811 | ID = Intrinsic::riscv_orc_b; |
| 19812 | break; |
| 19813 | case RISCV::BI__builtin_riscv_clz_32: |
| 19814 | case RISCV::BI__builtin_riscv_clz_64: { |
| 19815 | Function *F = CGM.getIntrinsic(Intrinsic::ctlz, Ops[0]->getType()); |
| 19816 | return Builder.CreateCall(F, {Ops[0], Builder.getInt1(false)}); |
| 19817 | } |
| 19818 | case RISCV::BI__builtin_riscv_ctz_32: |
| 19819 | case RISCV::BI__builtin_riscv_ctz_64: { |
| 19820 | Function *F = CGM.getIntrinsic(Intrinsic::cttz, Ops[0]->getType()); |
| 19821 | return Builder.CreateCall(F, {Ops[0], Builder.getInt1(false)}); |
| 19822 | } |
| 19823 | |
| 19824 | // Zbc |
| 19825 | case RISCV::BI__builtin_riscv_clmul: |
| 19826 | ID = Intrinsic::riscv_clmul; |
| 19827 | break; |
| 19828 | case RISCV::BI__builtin_riscv_clmulh: |
| 19829 | ID = Intrinsic::riscv_clmulh; |
| 19830 | break; |
| 19831 | case RISCV::BI__builtin_riscv_clmulr: |
| 19832 | ID = Intrinsic::riscv_clmulr; |
| 19833 | break; |
| 19834 | |
| 19835 | // Zbkx |
| 19836 | case RISCV::BI__builtin_riscv_xperm8: |
| 19837 | ID = Intrinsic::riscv_xperm8; |
| 19838 | break; |
| 19839 | case RISCV::BI__builtin_riscv_xperm4: |
| 19840 | ID = Intrinsic::riscv_xperm4; |
| 19841 | break; |
| 19842 | |
| 19843 | // Zbkb |
| 19844 | case RISCV::BI__builtin_riscv_brev8: |
| 19845 | ID = Intrinsic::riscv_brev8; |
| 19846 | break; |
| 19847 | case RISCV::BI__builtin_riscv_zip_32: |
| 19848 | ID = Intrinsic::riscv_zip; |
| 19849 | break; |
| 19850 | case RISCV::BI__builtin_riscv_unzip_32: |
| 19851 | ID = Intrinsic::riscv_unzip; |
| 19852 | break; |
| 19853 | } |
| 19854 | |
| 19855 | IntrinsicTypes = {ResultType}; |
| 19856 | break; |
| 19857 | } |
| 19858 | |
| 19859 | // Zk builtins |
| 19860 | |
| 19861 | // Zknd |
| 19862 | case RISCV::BI__builtin_riscv_aes32dsi_32: |
| 19863 | ID = Intrinsic::riscv_aes32dsi; |
| 19864 | break; |
| 19865 | case RISCV::BI__builtin_riscv_aes32dsmi_32: |
| 19866 | ID = Intrinsic::riscv_aes32dsmi; |
| 19867 | break; |
| 19868 | case RISCV::BI__builtin_riscv_aes64ds_64: |
| 19869 | ID = Intrinsic::riscv_aes64ds; |
| 19870 | break; |
| 19871 | case RISCV::BI__builtin_riscv_aes64dsm_64: |
| 19872 | ID = Intrinsic::riscv_aes64dsm; |
| 19873 | break; |
| 19874 | case RISCV::BI__builtin_riscv_aes64im_64: |
| 19875 | ID = Intrinsic::riscv_aes64im; |
| 19876 | break; |
| 19877 | |
| 19878 | // Zkne |
| 19879 | case RISCV::BI__builtin_riscv_aes32esi_32: |
| 19880 | ID = Intrinsic::riscv_aes32esi; |
| 19881 | break; |
| 19882 | case RISCV::BI__builtin_riscv_aes32esmi_32: |
| 19883 | ID = Intrinsic::riscv_aes32esmi; |
| 19884 | break; |
| 19885 | case RISCV::BI__builtin_riscv_aes64es_64: |
| 19886 | ID = Intrinsic::riscv_aes64es; |
| 19887 | break; |
| 19888 | case RISCV::BI__builtin_riscv_aes64esm_64: |
| 19889 | ID = Intrinsic::riscv_aes64esm; |
| 19890 | break; |
| 19891 | |
| 19892 | // Zknd & Zkne |
| 19893 | case RISCV::BI__builtin_riscv_aes64ks1i_64: |
| 19894 | ID = Intrinsic::riscv_aes64ks1i; |
| 19895 | break; |
| 19896 | case RISCV::BI__builtin_riscv_aes64ks2_64: |
| 19897 | ID = Intrinsic::riscv_aes64ks2; |
| 19898 | break; |
| 19899 | |
| 19900 | // Zknh |
| 19901 | case RISCV::BI__builtin_riscv_sha256sig0: |
| 19902 | ID = Intrinsic::riscv_sha256sig0; |
| 19903 | IntrinsicTypes = {ResultType}; |
| 19904 | break; |
| 19905 | case RISCV::BI__builtin_riscv_sha256sig1: |
| 19906 | ID = Intrinsic::riscv_sha256sig1; |
| 19907 | IntrinsicTypes = {ResultType}; |
| 19908 | break; |
| 19909 | case RISCV::BI__builtin_riscv_sha256sum0: |
| 19910 | ID = Intrinsic::riscv_sha256sum0; |
| 19911 | IntrinsicTypes = {ResultType}; |
| 19912 | break; |
| 19913 | case RISCV::BI__builtin_riscv_sha256sum1: |
| 19914 | ID = Intrinsic::riscv_sha256sum1; |
| 19915 | IntrinsicTypes = {ResultType}; |
| 19916 | break; |
| 19917 | case RISCV::BI__builtin_riscv_sha512sig0_64: |
| 19918 | ID = Intrinsic::riscv_sha512sig0; |
| 19919 | break; |
| 19920 | case RISCV::BI__builtin_riscv_sha512sig0h_32: |
| 19921 | ID = Intrinsic::riscv_sha512sig0h; |
| 19922 | break; |
| 19923 | case RISCV::BI__builtin_riscv_sha512sig0l_32: |
| 19924 | ID = Intrinsic::riscv_sha512sig0l; |
| 19925 | break; |
| 19926 | case RISCV::BI__builtin_riscv_sha512sig1_64: |
| 19927 | ID = Intrinsic::riscv_sha512sig1; |
| 19928 | break; |
| 19929 | case RISCV::BI__builtin_riscv_sha512sig1h_32: |
| 19930 | ID = Intrinsic::riscv_sha512sig1h; |
| 19931 | break; |
| 19932 | case RISCV::BI__builtin_riscv_sha512sig1l_32: |
| 19933 | ID = Intrinsic::riscv_sha512sig1l; |
| 19934 | break; |
| 19935 | case RISCV::BI__builtin_riscv_sha512sum0_64: |
| 19936 | ID = Intrinsic::riscv_sha512sum0; |
| 19937 | break; |
| 19938 | case RISCV::BI__builtin_riscv_sha512sum0r_32: |
| 19939 | ID = Intrinsic::riscv_sha512sum0r; |
| 19940 | break; |
| 19941 | case RISCV::BI__builtin_riscv_sha512sum1_64: |
| 19942 | ID = Intrinsic::riscv_sha512sum1; |
| 19943 | break; |
| 19944 | case RISCV::BI__builtin_riscv_sha512sum1r_32: |
| 19945 | ID = Intrinsic::riscv_sha512sum1r; |
| 19946 | break; |
| 19947 | |
| 19948 | // Zksed |
| 19949 | case RISCV::BI__builtin_riscv_sm4ks: |
| 19950 | ID = Intrinsic::riscv_sm4ks; |
| 19951 | IntrinsicTypes = {ResultType}; |
| 19952 | break; |
| 19953 | case RISCV::BI__builtin_riscv_sm4ed: |
| 19954 | ID = Intrinsic::riscv_sm4ed; |
| 19955 | IntrinsicTypes = {ResultType}; |
| 19956 | break; |
| 19957 | |
| 19958 | // Zksh |
| 19959 | case RISCV::BI__builtin_riscv_sm3p0: |
| 19960 | ID = Intrinsic::riscv_sm3p0; |
| 19961 | IntrinsicTypes = {ResultType}; |
| 19962 | break; |
| 19963 | case RISCV::BI__builtin_riscv_sm3p1: |
| 19964 | ID = Intrinsic::riscv_sm3p1; |
| 19965 | IntrinsicTypes = {ResultType}; |
| 19966 | break; |
| 19967 | |
| 19968 | // Zihintntl |
| 19969 | case RISCV::BI__builtin_riscv_ntl_load: { |
| 19970 | llvm::Type *ResTy = ConvertType(E->getType()); |
| 19971 | ConstantInt *Mode = cast<ConstantInt>(Ops[1]); |
| 19972 | |
| 19973 | llvm::MDNode *RISCVDomainNode = llvm::MDNode::get( |
| 19974 | getLLVMContext(), |
| 19975 | llvm::ConstantAsMetadata::get(Builder.getInt32(Mode->getZExtValue()))); |
| 19976 | llvm::MDNode *NontemporalNode = llvm::MDNode::get( |
| 19977 | getLLVMContext(), llvm::ConstantAsMetadata::get(Builder.getInt32(1))); |
| 19978 | |
| 19979 | int Width; |
| 19980 | if(ResTy->isScalableTy()) { |
| 19981 | const ScalableVectorType *SVTy = cast<ScalableVectorType>(ResTy); |
| 19982 | llvm::Type *ScalarTy = ResTy->getScalarType(); |
| 19983 | Width = ScalarTy->getPrimitiveSizeInBits() * |
| 19984 | SVTy->getElementCount().getKnownMinValue(); |
| 19985 | } else |
| 19986 | Width = ResTy->getPrimitiveSizeInBits(); |
| 19987 | LoadInst *Load = Builder.CreateLoad( |
| 19988 | Address(Ops[0], ResTy, CharUnits::fromQuantity(Width / 8))); |
| 19989 | |
| 19990 | Load->setMetadata(CGM.getModule().getMDKindID("nontemporal"), |
| 19991 | NontemporalNode); |
| 19992 | Load->setMetadata(CGM.getModule().getMDKindID("riscv-nontemporal-domain"), |
| 19993 | RISCVDomainNode); |
| 19994 | |
| 19995 | return Load; |
| 19996 | } |
| 19997 | case RISCV::BI__builtin_riscv_ntl_store: { |
| 19998 | ConstantInt *Mode = cast<ConstantInt>(Ops[2]); |
| 19999 | |
| 20000 | llvm::MDNode *RISCVDomainNode = llvm::MDNode::get( |
| 20001 | getLLVMContext(), |
| 20002 | llvm::ConstantAsMetadata::get(Builder.getInt32(Mode->getZExtValue()))); |
| 20003 | llvm::MDNode *NontemporalNode = llvm::MDNode::get( |
| 20004 | getLLVMContext(), llvm::ConstantAsMetadata::get(Builder.getInt32(1))); |
| 20005 | |
| 20006 | Value *BC = Builder.CreateBitCast( |
| 20007 | Ops[0], llvm::PointerType::getUnqual(Ops[1]->getType()), "cast"); |
| 20008 | |
| 20009 | StoreInst *Store = Builder.CreateDefaultAlignedStore(Ops[1], BC); |
| 20010 | Store->setMetadata(CGM.getModule().getMDKindID("nontemporal"), |
| 20011 | NontemporalNode); |
| 20012 | Store->setMetadata(CGM.getModule().getMDKindID("riscv-nontemporal-domain"), |
| 20013 | RISCVDomainNode); |
| 20014 | |
| 20015 | return Store; |
| 20016 | } |
| 20017 | |
| 20018 | // Vector builtins are handled from here. |
| 20019 | #include "clang/Basic/riscv_vector_builtin_cg.inc" |
| 20020 | // SiFive Vector builtins are handled from here. |
| 20021 | #include "clang/Basic/riscv_sifive_vector_builtin_cg.inc" |
| 20022 | } |
| 20023 | |
| 20024 | assert(ID != Intrinsic::not_intrinsic)(static_cast <bool> (ID != Intrinsic::not_intrinsic) ? void (0) : __assert_fail ("ID != Intrinsic::not_intrinsic", "clang/lib/CodeGen/CGBuiltin.cpp" , 20024, __extension__ __PRETTY_FUNCTION__)); |
| 20025 | |
| 20026 | llvm::Function *F = CGM.getIntrinsic(ID, IntrinsicTypes); |
| 20027 | return Builder.CreateCall(F, Ops, ""); |
| 20028 | } |
| 20029 | |
| 20030 | Value *CodeGenFunction::EmitLoongArchBuiltinExpr(unsigned BuiltinID, |
| 20031 | const CallExpr *E) { |
| 20032 | SmallVector<Value *, 4> Ops; |
| 20033 | |
| 20034 | for (unsigned i = 0, e = E->getNumArgs(); i != e; i++) |
| 20035 | Ops.push_back(EmitScalarExpr(E->getArg(i))); |
| 20036 | |
| 20037 | Intrinsic::ID ID = Intrinsic::not_intrinsic; |
| 20038 | |
| 20039 | switch (BuiltinID) { |
| 20040 | default: |
| 20041 | llvm_unreachable("unexpected builtin ID.")::llvm::llvm_unreachable_internal("unexpected builtin ID.", "clang/lib/CodeGen/CGBuiltin.cpp" , 20041); |
| 20042 | case LoongArch::BI__builtin_loongarch_cacop_d: |
| 20043 | ID = Intrinsic::loongarch_cacop_d; |
| 20044 | break; |
| 20045 | case LoongArch::BI__builtin_loongarch_cacop_w: |
| 20046 | ID = Intrinsic::loongarch_cacop_w; |
| 20047 | break; |
| 20048 | case LoongArch::BI__builtin_loongarch_dbar: |
| 20049 | ID = Intrinsic::loongarch_dbar; |
| 20050 | break; |
| 20051 | case LoongArch::BI__builtin_loongarch_break: |
| 20052 | ID = Intrinsic::loongarch_break; |
| 20053 | break; |
| 20054 | case LoongArch::BI__builtin_loongarch_ibar: |
| 20055 | ID = Intrinsic::loongarch_ibar; |
| 20056 | break; |
| 20057 | case LoongArch::BI__builtin_loongarch_movfcsr2gr: |
| 20058 | ID = Intrinsic::loongarch_movfcsr2gr; |
| 20059 | break; |
| 20060 | case LoongArch::BI__builtin_loongarch_movgr2fcsr: |
| 20061 | ID = Intrinsic::loongarch_movgr2fcsr; |
| 20062 | break; |
| 20063 | case LoongArch::BI__builtin_loongarch_syscall: |
| 20064 | ID = Intrinsic::loongarch_syscall; |
| 20065 | break; |
| 20066 | case LoongArch::BI__builtin_loongarch_crc_w_b_w: |
| 20067 | ID = Intrinsic::loongarch_crc_w_b_w; |
| 20068 | break; |
| 20069 | case LoongArch::BI__builtin_loongarch_crc_w_h_w: |
| 20070 | ID = Intrinsic::loongarch_crc_w_h_w; |
| 20071 | break; |
| 20072 | case LoongArch::BI__builtin_loongarch_crc_w_w_w: |
| 20073 | ID = Intrinsic::loongarch_crc_w_w_w; |
| 20074 | break; |
| 20075 | case LoongArch::BI__builtin_loongarch_crc_w_d_w: |
| 20076 | ID = Intrinsic::loongarch_crc_w_d_w; |
| 20077 | break; |
| 20078 | case LoongArch::BI__builtin_loongarch_crcc_w_b_w: |
| 20079 | ID = Intrinsic::loongarch_crcc_w_b_w; |
| 20080 | break; |
| 20081 | case LoongArch::BI__builtin_loongarch_crcc_w_h_w: |
| 20082 | ID = Intrinsic::loongarch_crcc_w_h_w; |
| 20083 | break; |
| 20084 | case LoongArch::BI__builtin_loongarch_crcc_w_w_w: |
| 20085 | ID = Intrinsic::loongarch_crcc_w_w_w; |
| 20086 | break; |
| 20087 | case LoongArch::BI__builtin_loongarch_crcc_w_d_w: |
| 20088 | ID = Intrinsic::loongarch_crcc_w_d_w; |
| 20089 | break; |
| 20090 | case LoongArch::BI__builtin_loongarch_csrrd_w: |
| 20091 | ID = Intrinsic::loongarch_csrrd_w; |
| 20092 | break; |
| 20093 | case LoongArch::BI__builtin_loongarch_csrwr_w: |
| 20094 | ID = Intrinsic::loongarch_csrwr_w; |
| 20095 | break; |
| 20096 | case LoongArch::BI__builtin_loongarch_csrxchg_w: |
| 20097 | ID = Intrinsic::loongarch_csrxchg_w; |
| 20098 | break; |
| 20099 | case LoongArch::BI__builtin_loongarch_csrrd_d: |
| 20100 | ID = Intrinsic::loongarch_csrrd_d; |
| 20101 | break; |
| 20102 | case LoongArch::BI__builtin_loongarch_csrwr_d: |
| 20103 | ID = Intrinsic::loongarch_csrwr_d; |
| 20104 | break; |
| 20105 | case LoongArch::BI__builtin_loongarch_csrxchg_d: |
| 20106 | ID = Intrinsic::loongarch_csrxchg_d; |
| 20107 | break; |
| 20108 | case LoongArch::BI__builtin_loongarch_iocsrrd_b: |
| 20109 | ID = Intrinsic::loongarch_iocsrrd_b; |
| 20110 | break; |
| 20111 | case LoongArch::BI__builtin_loongarch_iocsrrd_h: |
| 20112 | ID = Intrinsic::loongarch_iocsrrd_h; |
| 20113 | break; |
| 20114 | case LoongArch::BI__builtin_loongarch_iocsrrd_w: |
| 20115 | ID = Intrinsic::loongarch_iocsrrd_w; |
| 20116 | break; |
| 20117 | case LoongArch::BI__builtin_loongarch_iocsrrd_d: |
| 20118 | ID = Intrinsic::loongarch_iocsrrd_d; |
| 20119 | break; |
| 20120 | case LoongArch::BI__builtin_loongarch_iocsrwr_b: |
| 20121 | ID = Intrinsic::loongarch_iocsrwr_b; |
| 20122 | break; |
| 20123 | case LoongArch::BI__builtin_loongarch_iocsrwr_h: |
| 20124 | ID = Intrinsic::loongarch_iocsrwr_h; |
| 20125 | break; |
| 20126 | case LoongArch::BI__builtin_loongarch_iocsrwr_w: |
| 20127 | ID = Intrinsic::loongarch_iocsrwr_w; |
| 20128 | break; |
| 20129 | case LoongArch::BI__builtin_loongarch_iocsrwr_d: |
| 20130 | ID = Intrinsic::loongarch_iocsrwr_d; |
| 20131 | break; |
| 20132 | case LoongArch::BI__builtin_loongarch_cpucfg: |
| 20133 | ID = Intrinsic::loongarch_cpucfg; |
| 20134 | break; |
| 20135 | case LoongArch::BI__builtin_loongarch_asrtle_d: |
| 20136 | ID = Intrinsic::loongarch_asrtle_d; |
| 20137 | break; |
| 20138 | case LoongArch::BI__builtin_loongarch_asrtgt_d: |
| 20139 | ID = Intrinsic::loongarch_asrtgt_d; |
| 20140 | break; |
| 20141 | case LoongArch::BI__builtin_loongarch_lddir_d: |
| 20142 | ID = Intrinsic::loongarch_lddir_d; |
| 20143 | break; |
| 20144 | case LoongArch::BI__builtin_loongarch_ldpte_d: |
| 20145 | ID = Intrinsic::loongarch_ldpte_d; |
| 20146 | break; |
| 20147 | // TODO: Support more Intrinsics. |
| 20148 | } |
| 20149 | |
| 20150 | assert(ID != Intrinsic::not_intrinsic)(static_cast <bool> (ID != Intrinsic::not_intrinsic) ? void (0) : __assert_fail ("ID != Intrinsic::not_intrinsic", "clang/lib/CodeGen/CGBuiltin.cpp" , 20150, __extension__ __PRETTY_FUNCTION__)); |
| 20151 | |
| 20152 | llvm::Function *F = CGM.getIntrinsic(ID); |
| 20153 | return Builder.CreateCall(F, Ops); |
| 20154 | } |