1 //===---- CGBuiltin.cpp - Emit LLVM Code for builtins ---------------------===// 2 // 3 // The LLVM Compiler Infrastructure 4 // 5 // This file is distributed under the University of Illinois Open Source 6 // License. See LICENSE.TXT for details. 7 // 8 //===----------------------------------------------------------------------===// 9 // 10 // This contains code to emit Builtin calls as LLVM code. 11 // 12 //===----------------------------------------------------------------------===// 13 14 #include "CodeGenFunction.h" 15 #include "CGCXXABI.h" 16 #include "CGObjCRuntime.h" 17 #include "CodeGenModule.h" 18 #include "TargetInfo.h" 19 #include "clang/AST/ASTContext.h" 20 #include "clang/AST/Decl.h" 21 #include "clang/Basic/TargetBuiltins.h" 22 #include "clang/Basic/TargetInfo.h" 23 #include "clang/CodeGen/CGFunctionInfo.h" 24 #include "llvm/ADT/StringExtras.h" 25 #include "llvm/IR/CallSite.h" 26 #include "llvm/IR/DataLayout.h" 27 #include "llvm/IR/InlineAsm.h" 28 #include "llvm/IR/Intrinsics.h" 29 #include <sstream> 30 31 using namespace clang; 32 using namespace CodeGen; 33 using namespace llvm; 34 35 /// getBuiltinLibFunction - Given a builtin id for a function like 36 /// "__builtin_fabsf", return a Function* for "fabsf". 37 llvm::Value *CodeGenModule::getBuiltinLibFunction(const FunctionDecl *FD, 38 unsigned BuiltinID) { 39 assert(Context.BuiltinInfo.isLibFunction(BuiltinID)); 40 41 // Get the name, skip over the __builtin_ prefix (if necessary). 42 StringRef Name; 43 GlobalDecl D(FD); 44 45 // If the builtin has been declared explicitly with an assembler label, 46 // use the mangled name. This differs from the plain label on platforms 47 // that prefix labels. 48 if (FD->hasAttr<AsmLabelAttr>()) 49 Name = getMangledName(D); 50 else 51 Name = Context.BuiltinInfo.getName(BuiltinID) + 10; 52 53 llvm::FunctionType *Ty = 54 cast<llvm::FunctionType>(getTypes().ConvertType(FD->getType())); 55 56 return GetOrCreateLLVMFunction(Name, Ty, D, /*ForVTable=*/false); 57 } 58 59 /// Emit the conversions required to turn the given value into an 60 /// integer of the given size. 61 static Value *EmitToInt(CodeGenFunction &CGF, llvm::Value *V, 62 QualType T, llvm::IntegerType *IntType) { 63 V = CGF.EmitToMemory(V, T); 64 65 if (V->getType()->isPointerTy()) 66 return CGF.Builder.CreatePtrToInt(V, IntType); 67 68 assert(V->getType() == IntType); 69 return V; 70 } 71 72 static Value *EmitFromInt(CodeGenFunction &CGF, llvm::Value *V, 73 QualType T, llvm::Type *ResultType) { 74 V = CGF.EmitFromMemory(V, T); 75 76 if (ResultType->isPointerTy()) 77 return CGF.Builder.CreateIntToPtr(V, ResultType); 78 79 assert(V->getType() == ResultType); 80 return V; 81 } 82 83 /// Utility to insert an atomic instruction based on Instrinsic::ID 84 /// and the expression node. 85 static Value *MakeBinaryAtomicValue(CodeGenFunction &CGF, 86 llvm::AtomicRMWInst::BinOp Kind, 87 const CallExpr *E) { 88 QualType T = E->getType(); 89 assert(E->getArg(0)->getType()->isPointerType()); 90 assert(CGF.getContext().hasSameUnqualifiedType(T, 91 E->getArg(0)->getType()->getPointeeType())); 92 assert(CGF.getContext().hasSameUnqualifiedType(T, E->getArg(1)->getType())); 93 94 llvm::Value *DestPtr = CGF.EmitScalarExpr(E->getArg(0)); 95 unsigned AddrSpace = DestPtr->getType()->getPointerAddressSpace(); 96 97 llvm::IntegerType *IntType = 98 llvm::IntegerType::get(CGF.getLLVMContext(), 99 CGF.getContext().getTypeSize(T)); 100 llvm::Type *IntPtrType = IntType->getPointerTo(AddrSpace); 101 102 llvm::Value *Args[2]; 103 Args[0] = CGF.Builder.CreateBitCast(DestPtr, IntPtrType); 104 Args[1] = CGF.EmitScalarExpr(E->getArg(1)); 105 llvm::Type *ValueType = Args[1]->getType(); 106 Args[1] = EmitToInt(CGF, Args[1], T, IntType); 107 108 llvm::Value *Result = 109 CGF.Builder.CreateAtomicRMW(Kind, Args[0], Args[1], 110 llvm::SequentiallyConsistent); 111 return EmitFromInt(CGF, Result, T, ValueType); 112 } 113 114 static Value *EmitNontemporalStore(CodeGenFunction &CGF, const CallExpr *E) { 115 Value *Val = CGF.EmitScalarExpr(E->getArg(0)); 116 Value *Address = CGF.EmitScalarExpr(E->getArg(1)); 117 118 // Convert the type of the pointer to a pointer to the stored type. 119 Val = CGF.EmitToMemory(Val, E->getArg(0)->getType()); 120 Value *BC = CGF.Builder.CreateBitCast( 121 Address, llvm::PointerType::getUnqual(Val->getType()), "cast"); 122 LValue LV = CGF.MakeNaturalAlignAddrLValue(BC, E->getArg(0)->getType()); 123 LV.setNontemporal(true); 124 CGF.EmitStoreOfScalar(Val, LV, false); 125 return nullptr; 126 } 127 128 static Value *EmitNontemporalLoad(CodeGenFunction &CGF, const CallExpr *E) { 129 Value *Address = CGF.EmitScalarExpr(E->getArg(0)); 130 131 LValue LV = CGF.MakeNaturalAlignAddrLValue(Address, E->getType()); 132 LV.setNontemporal(true); 133 return CGF.EmitLoadOfScalar(LV, E->getExprLoc()); 134 } 135 136 static RValue EmitBinaryAtomic(CodeGenFunction &CGF, 137 llvm::AtomicRMWInst::BinOp Kind, 138 const CallExpr *E) { 139 return RValue::get(MakeBinaryAtomicValue(CGF, Kind, E)); 140 } 141 142 /// Utility to insert an atomic instruction based Instrinsic::ID and 143 /// the expression node, where the return value is the result of the 144 /// operation. 145 static RValue EmitBinaryAtomicPost(CodeGenFunction &CGF, 146 llvm::AtomicRMWInst::BinOp Kind, 147 const CallExpr *E, 148 Instruction::BinaryOps Op, 149 bool Invert = false) { 150 QualType T = E->getType(); 151 assert(E->getArg(0)->getType()->isPointerType()); 152 assert(CGF.getContext().hasSameUnqualifiedType(T, 153 E->getArg(0)->getType()->getPointeeType())); 154 assert(CGF.getContext().hasSameUnqualifiedType(T, E->getArg(1)->getType())); 155 156 llvm::Value *DestPtr = CGF.EmitScalarExpr(E->getArg(0)); 157 unsigned AddrSpace = DestPtr->getType()->getPointerAddressSpace(); 158 159 llvm::IntegerType *IntType = 160 llvm::IntegerType::get(CGF.getLLVMContext(), 161 CGF.getContext().getTypeSize(T)); 162 llvm::Type *IntPtrType = IntType->getPointerTo(AddrSpace); 163 164 llvm::Value *Args[2]; 165 Args[1] = CGF.EmitScalarExpr(E->getArg(1)); 166 llvm::Type *ValueType = Args[1]->getType(); 167 Args[1] = EmitToInt(CGF, Args[1], T, IntType); 168 Args[0] = CGF.Builder.CreateBitCast(DestPtr, IntPtrType); 169 170 llvm::Value *Result = 171 CGF.Builder.CreateAtomicRMW(Kind, Args[0], Args[1], 172 llvm::SequentiallyConsistent); 173 Result = CGF.Builder.CreateBinOp(Op, Result, Args[1]); 174 if (Invert) 175 Result = CGF.Builder.CreateBinOp(llvm::Instruction::Xor, Result, 176 llvm::ConstantInt::get(IntType, -1)); 177 Result = EmitFromInt(CGF, Result, T, ValueType); 178 return RValue::get(Result); 179 } 180 181 /// @brief Utility to insert an atomic cmpxchg instruction. 182 /// 183 /// @param CGF The current codegen function. 184 /// @param E Builtin call expression to convert to cmpxchg. 185 /// arg0 - address to operate on 186 /// arg1 - value to compare with 187 /// arg2 - new value 188 /// @param ReturnBool Specifies whether to return success flag of 189 /// cmpxchg result or the old value. 190 /// 191 /// @returns result of cmpxchg, according to ReturnBool 192 static Value *MakeAtomicCmpXchgValue(CodeGenFunction &CGF, const CallExpr *E, 193 bool ReturnBool) { 194 QualType T = ReturnBool ? E->getArg(1)->getType() : E->getType(); 195 llvm::Value *DestPtr = CGF.EmitScalarExpr(E->getArg(0)); 196 unsigned AddrSpace = DestPtr->getType()->getPointerAddressSpace(); 197 198 llvm::IntegerType *IntType = llvm::IntegerType::get( 199 CGF.getLLVMContext(), CGF.getContext().getTypeSize(T)); 200 llvm::Type *IntPtrType = IntType->getPointerTo(AddrSpace); 201 202 Value *Args[3]; 203 Args[0] = CGF.Builder.CreateBitCast(DestPtr, IntPtrType); 204 Args[1] = CGF.EmitScalarExpr(E->getArg(1)); 205 llvm::Type *ValueType = Args[1]->getType(); 206 Args[1] = EmitToInt(CGF, Args[1], T, IntType); 207 Args[2] = EmitToInt(CGF, CGF.EmitScalarExpr(E->getArg(2)), T, IntType); 208 209 Value *Pair = CGF.Builder.CreateAtomicCmpXchg(Args[0], Args[1], Args[2], 210 llvm::SequentiallyConsistent, 211 llvm::SequentiallyConsistent); 212 if (ReturnBool) 213 // Extract boolean success flag and zext it to int. 214 return CGF.Builder.CreateZExt(CGF.Builder.CreateExtractValue(Pair, 1), 215 CGF.ConvertType(E->getType())); 216 else 217 // Extract old value and emit it using the same type as compare value. 218 return EmitFromInt(CGF, CGF.Builder.CreateExtractValue(Pair, 0), T, 219 ValueType); 220 } 221 222 /// EmitFAbs - Emit a call to @llvm.fabs(). 223 static Value *EmitFAbs(CodeGenFunction &CGF, Value *V) { 224 Value *F = CGF.CGM.getIntrinsic(Intrinsic::fabs, V->getType()); 225 llvm::CallInst *Call = CGF.Builder.CreateCall(F, V); 226 Call->setDoesNotAccessMemory(); 227 return Call; 228 } 229 230 /// Emit the computation of the sign bit for a floating point value. Returns 231 /// the i1 sign bit value. 232 static Value *EmitSignBit(CodeGenFunction &CGF, Value *V) { 233 LLVMContext &C = CGF.CGM.getLLVMContext(); 234 235 llvm::Type *Ty = V->getType(); 236 int Width = Ty->getPrimitiveSizeInBits(); 237 llvm::Type *IntTy = llvm::IntegerType::get(C, Width); 238 V = CGF.Builder.CreateBitCast(V, IntTy); 239 if (Ty->isPPC_FP128Ty()) { 240 // We want the sign bit of the higher-order double. The bitcast we just 241 // did works as if the double-double was stored to memory and then 242 // read as an i128. The "store" will put the higher-order double in the 243 // lower address in both little- and big-Endian modes, but the "load" 244 // will treat those bits as a different part of the i128: the low bits in 245 // little-Endian, the high bits in big-Endian. Therefore, on big-Endian 246 // we need to shift the high bits down to the low before truncating. 247 Width >>= 1; 248 if (CGF.getTarget().isBigEndian()) { 249 Value *ShiftCst = llvm::ConstantInt::get(IntTy, Width); 250 V = CGF.Builder.CreateLShr(V, ShiftCst); 251 } 252 // We are truncating value in order to extract the higher-order 253 // double, which we will be using to extract the sign from. 254 IntTy = llvm::IntegerType::get(C, Width); 255 V = CGF.Builder.CreateTrunc(V, IntTy); 256 } 257 Value *Zero = llvm::Constant::getNullValue(IntTy); 258 return CGF.Builder.CreateICmpSLT(V, Zero); 259 } 260 261 static RValue emitLibraryCall(CodeGenFunction &CGF, const FunctionDecl *Fn, 262 const CallExpr *E, llvm::Value *calleeValue) { 263 return CGF.EmitCall(E->getCallee()->getType(), calleeValue, E, 264 ReturnValueSlot(), Fn); 265 } 266 267 /// \brief Emit a call to llvm.{sadd,uadd,ssub,usub,smul,umul}.with.overflow.* 268 /// depending on IntrinsicID. 269 /// 270 /// \arg CGF The current codegen function. 271 /// \arg IntrinsicID The ID for the Intrinsic we wish to generate. 272 /// \arg X The first argument to the llvm.*.with.overflow.*. 273 /// \arg Y The second argument to the llvm.*.with.overflow.*. 274 /// \arg Carry The carry returned by the llvm.*.with.overflow.*. 275 /// \returns The result (i.e. sum/product) returned by the intrinsic. 276 static llvm::Value *EmitOverflowIntrinsic(CodeGenFunction &CGF, 277 const llvm::Intrinsic::ID IntrinsicID, 278 llvm::Value *X, llvm::Value *Y, 279 llvm::Value *&Carry) { 280 // Make sure we have integers of the same width. 281 assert(X->getType() == Y->getType() && 282 "Arguments must be the same type. (Did you forget to make sure both " 283 "arguments have the same integer width?)"); 284 285 llvm::Value *Callee = CGF.CGM.getIntrinsic(IntrinsicID, X->getType()); 286 llvm::Value *Tmp = CGF.Builder.CreateCall(Callee, {X, Y}); 287 Carry = CGF.Builder.CreateExtractValue(Tmp, 1); 288 return CGF.Builder.CreateExtractValue(Tmp, 0); 289 } 290 291 namespace { 292 struct WidthAndSignedness { 293 unsigned Width; 294 bool Signed; 295 }; 296 } 297 298 static WidthAndSignedness 299 getIntegerWidthAndSignedness(const clang::ASTContext &context, 300 const clang::QualType Type) { 301 assert(Type->isIntegerType() && "Given type is not an integer."); 302 unsigned Width = Type->isBooleanType() ? 1 : context.getTypeInfo(Type).Width; 303 bool Signed = Type->isSignedIntegerType(); 304 return {Width, Signed}; 305 } 306 307 // Given one or more integer types, this function produces an integer type that 308 // encompasses them: any value in one of the given types could be expressed in 309 // the encompassing type. 310 static struct WidthAndSignedness 311 EncompassingIntegerType(ArrayRef<struct WidthAndSignedness> Types) { 312 assert(Types.size() > 0 && "Empty list of types."); 313 314 // If any of the given types is signed, we must return a signed type. 315 bool Signed = false; 316 for (const auto &Type : Types) { 317 Signed |= Type.Signed; 318 } 319 320 // The encompassing type must have a width greater than or equal to the width 321 // of the specified types. Aditionally, if the encompassing type is signed, 322 // its width must be strictly greater than the width of any unsigned types 323 // given. 324 unsigned Width = 0; 325 for (const auto &Type : Types) { 326 unsigned MinWidth = Type.Width + (Signed && !Type.Signed); 327 if (Width < MinWidth) { 328 Width = MinWidth; 329 } 330 } 331 332 return {Width, Signed}; 333 } 334 335 Value *CodeGenFunction::EmitVAStartEnd(Value *ArgValue, bool IsStart) { 336 llvm::Type *DestType = Int8PtrTy; 337 if (ArgValue->getType() != DestType) 338 ArgValue = 339 Builder.CreateBitCast(ArgValue, DestType, ArgValue->getName().data()); 340 341 Intrinsic::ID inst = IsStart ? Intrinsic::vastart : Intrinsic::vaend; 342 return Builder.CreateCall(CGM.getIntrinsic(inst), ArgValue); 343 } 344 345 /// Checks if using the result of __builtin_object_size(p, @p From) in place of 346 /// __builtin_object_size(p, @p To) is correct 347 static bool areBOSTypesCompatible(int From, int To) { 348 // Note: Our __builtin_object_size implementation currently treats Type=0 and 349 // Type=2 identically. Encoding this implementation detail here may make 350 // improving __builtin_object_size difficult in the future, so it's omitted. 351 return From == To || (From == 0 && To == 1) || (From == 3 && To == 2); 352 } 353 354 static llvm::Value * 355 getDefaultBuiltinObjectSizeResult(unsigned Type, llvm::IntegerType *ResType) { 356 return ConstantInt::get(ResType, (Type & 2) ? 0 : -1, /*isSigned=*/true); 357 } 358 359 llvm::Value * 360 CodeGenFunction::evaluateOrEmitBuiltinObjectSize(const Expr *E, unsigned Type, 361 llvm::IntegerType *ResType) { 362 uint64_t ObjectSize; 363 if (!E->tryEvaluateObjectSize(ObjectSize, getContext(), Type)) 364 return emitBuiltinObjectSize(E, Type, ResType); 365 return ConstantInt::get(ResType, ObjectSize, /*isSigned=*/true); 366 } 367 368 /// Returns a Value corresponding to the size of the given expression. 369 /// This Value may be either of the following: 370 /// - A llvm::Argument (if E is a param with the pass_object_size attribute on 371 /// it) 372 /// - A call to the @llvm.objectsize intrinsic 373 llvm::Value * 374 CodeGenFunction::emitBuiltinObjectSize(const Expr *E, unsigned Type, 375 llvm::IntegerType *ResType) { 376 // We need to reference an argument if the pointer is a parameter with the 377 // pass_object_size attribute. 378 if (auto *D = dyn_cast<DeclRefExpr>(E->IgnoreParenImpCasts())) { 379 auto *Param = dyn_cast<ParmVarDecl>(D->getDecl()); 380 auto *PS = D->getDecl()->getAttr<PassObjectSizeAttr>(); 381 if (Param != nullptr && PS != nullptr && 382 areBOSTypesCompatible(PS->getType(), Type)) { 383 auto Iter = SizeArguments.find(Param); 384 assert(Iter != SizeArguments.end()); 385 386 const ImplicitParamDecl *D = Iter->second; 387 auto DIter = LocalDeclMap.find(D); 388 assert(DIter != LocalDeclMap.end()); 389 390 return EmitLoadOfScalar(DIter->second, /*volatile=*/false, 391 getContext().getSizeType(), E->getLocStart()); 392 } 393 } 394 395 // LLVM can't handle Type=3 appropriately, and __builtin_object_size shouldn't 396 // evaluate E for side-effects. In either case, we shouldn't lower to 397 // @llvm.objectsize. 398 if (Type == 3 || E->HasSideEffects(getContext())) 399 return getDefaultBuiltinObjectSizeResult(Type, ResType); 400 401 // LLVM only supports 0 and 2, make sure that we pass along that 402 // as a boolean. 403 auto *CI = ConstantInt::get(Builder.getInt1Ty(), (Type & 2) >> 1); 404 // FIXME: Get right address space. 405 llvm::Type *Tys[] = {ResType, Builder.getInt8PtrTy(0)}; 406 Value *F = CGM.getIntrinsic(Intrinsic::objectsize, Tys); 407 return Builder.CreateCall(F, {EmitScalarExpr(E), CI}); 408 } 409 410 RValue CodeGenFunction::EmitBuiltinExpr(const FunctionDecl *FD, 411 unsigned BuiltinID, const CallExpr *E, 412 ReturnValueSlot ReturnValue) { 413 // See if we can constant fold this builtin. If so, don't emit it at all. 414 Expr::EvalResult Result; 415 if (E->EvaluateAsRValue(Result, CGM.getContext()) && 416 !Result.hasSideEffects()) { 417 if (Result.Val.isInt()) 418 return RValue::get(llvm::ConstantInt::get(getLLVMContext(), 419 Result.Val.getInt())); 420 if (Result.Val.isFloat()) 421 return RValue::get(llvm::ConstantFP::get(getLLVMContext(), 422 Result.Val.getFloat())); 423 } 424 425 switch (BuiltinID) { 426 default: break; // Handle intrinsics and libm functions below. 427 case Builtin::BI__builtin___CFStringMakeConstantString: 428 case Builtin::BI__builtin___NSStringMakeConstantString: 429 return RValue::get(CGM.EmitConstantExpr(E, E->getType(), nullptr)); 430 case Builtin::BI__builtin_stdarg_start: 431 case Builtin::BI__builtin_va_start: 432 case Builtin::BI__va_start: 433 case Builtin::BI__builtin_va_end: 434 return RValue::get( 435 EmitVAStartEnd(BuiltinID == Builtin::BI__va_start 436 ? EmitScalarExpr(E->getArg(0)) 437 : EmitVAListRef(E->getArg(0)).getPointer(), 438 BuiltinID != Builtin::BI__builtin_va_end)); 439 case Builtin::BI__builtin_va_copy: { 440 Value *DstPtr = EmitVAListRef(E->getArg(0)).getPointer(); 441 Value *SrcPtr = EmitVAListRef(E->getArg(1)).getPointer(); 442 443 llvm::Type *Type = Int8PtrTy; 444 445 DstPtr = Builder.CreateBitCast(DstPtr, Type); 446 SrcPtr = Builder.CreateBitCast(SrcPtr, Type); 447 return RValue::get(Builder.CreateCall(CGM.getIntrinsic(Intrinsic::vacopy), 448 {DstPtr, SrcPtr})); 449 } 450 case Builtin::BI__builtin_abs: 451 case Builtin::BI__builtin_labs: 452 case Builtin::BI__builtin_llabs: { 453 Value *ArgValue = EmitScalarExpr(E->getArg(0)); 454 455 Value *NegOp = Builder.CreateNeg(ArgValue, "neg"); 456 Value *CmpResult = 457 Builder.CreateICmpSGE(ArgValue, 458 llvm::Constant::getNullValue(ArgValue->getType()), 459 "abscond"); 460 Value *Result = 461 Builder.CreateSelect(CmpResult, ArgValue, NegOp, "abs"); 462 463 return RValue::get(Result); 464 } 465 case Builtin::BI__builtin_fabs: 466 case Builtin::BI__builtin_fabsf: 467 case Builtin::BI__builtin_fabsl: { 468 Value *Arg1 = EmitScalarExpr(E->getArg(0)); 469 Value *Result = EmitFAbs(*this, Arg1); 470 return RValue::get(Result); 471 } 472 case Builtin::BI__builtin_fmod: 473 case Builtin::BI__builtin_fmodf: 474 case Builtin::BI__builtin_fmodl: { 475 Value *Arg1 = EmitScalarExpr(E->getArg(0)); 476 Value *Arg2 = EmitScalarExpr(E->getArg(1)); 477 Value *Result = Builder.CreateFRem(Arg1, Arg2, "fmod"); 478 return RValue::get(Result); 479 } 480 481 case Builtin::BI__builtin_conj: 482 case Builtin::BI__builtin_conjf: 483 case Builtin::BI__builtin_conjl: { 484 ComplexPairTy ComplexVal = EmitComplexExpr(E->getArg(0)); 485 Value *Real = ComplexVal.first; 486 Value *Imag = ComplexVal.second; 487 Value *Zero = 488 Imag->getType()->isFPOrFPVectorTy() 489 ? llvm::ConstantFP::getZeroValueForNegation(Imag->getType()) 490 : llvm::Constant::getNullValue(Imag->getType()); 491 492 Imag = Builder.CreateFSub(Zero, Imag, "sub"); 493 return RValue::getComplex(std::make_pair(Real, Imag)); 494 } 495 case Builtin::BI__builtin_creal: 496 case Builtin::BI__builtin_crealf: 497 case Builtin::BI__builtin_creall: 498 case Builtin::BIcreal: 499 case Builtin::BIcrealf: 500 case Builtin::BIcreall: { 501 ComplexPairTy ComplexVal = EmitComplexExpr(E->getArg(0)); 502 return RValue::get(ComplexVal.first); 503 } 504 505 case Builtin::BI__builtin_cimag: 506 case Builtin::BI__builtin_cimagf: 507 case Builtin::BI__builtin_cimagl: 508 case Builtin::BIcimag: 509 case Builtin::BIcimagf: 510 case Builtin::BIcimagl: { 511 ComplexPairTy ComplexVal = EmitComplexExpr(E->getArg(0)); 512 return RValue::get(ComplexVal.second); 513 } 514 515 case Builtin::BI__builtin_ctzs: 516 case Builtin::BI__builtin_ctz: 517 case Builtin::BI__builtin_ctzl: 518 case Builtin::BI__builtin_ctzll: { 519 Value *ArgValue = EmitScalarExpr(E->getArg(0)); 520 521 llvm::Type *ArgType = ArgValue->getType(); 522 Value *F = CGM.getIntrinsic(Intrinsic::cttz, ArgType); 523 524 llvm::Type *ResultType = ConvertType(E->getType()); 525 Value *ZeroUndef = Builder.getInt1(getTarget().isCLZForZeroUndef()); 526 Value *Result = Builder.CreateCall(F, {ArgValue, ZeroUndef}); 527 if (Result->getType() != ResultType) 528 Result = Builder.CreateIntCast(Result, ResultType, /*isSigned*/true, 529 "cast"); 530 return RValue::get(Result); 531 } 532 case Builtin::BI__builtin_clzs: 533 case Builtin::BI__builtin_clz: 534 case Builtin::BI__builtin_clzl: 535 case Builtin::BI__builtin_clzll: { 536 Value *ArgValue = EmitScalarExpr(E->getArg(0)); 537 538 llvm::Type *ArgType = ArgValue->getType(); 539 Value *F = CGM.getIntrinsic(Intrinsic::ctlz, ArgType); 540 541 llvm::Type *ResultType = ConvertType(E->getType()); 542 Value *ZeroUndef = Builder.getInt1(getTarget().isCLZForZeroUndef()); 543 Value *Result = Builder.CreateCall(F, {ArgValue, ZeroUndef}); 544 if (Result->getType() != ResultType) 545 Result = Builder.CreateIntCast(Result, ResultType, /*isSigned*/true, 546 "cast"); 547 return RValue::get(Result); 548 } 549 case Builtin::BI__builtin_ffs: 550 case Builtin::BI__builtin_ffsl: 551 case Builtin::BI__builtin_ffsll: { 552 // ffs(x) -> x ? cttz(x) + 1 : 0 553 Value *ArgValue = EmitScalarExpr(E->getArg(0)); 554 555 llvm::Type *ArgType = ArgValue->getType(); 556 Value *F = CGM.getIntrinsic(Intrinsic::cttz, ArgType); 557 558 llvm::Type *ResultType = ConvertType(E->getType()); 559 Value *Tmp = 560 Builder.CreateAdd(Builder.CreateCall(F, {ArgValue, Builder.getTrue()}), 561 llvm::ConstantInt::get(ArgType, 1)); 562 Value *Zero = llvm::Constant::getNullValue(ArgType); 563 Value *IsZero = Builder.CreateICmpEQ(ArgValue, Zero, "iszero"); 564 Value *Result = Builder.CreateSelect(IsZero, Zero, Tmp, "ffs"); 565 if (Result->getType() != ResultType) 566 Result = Builder.CreateIntCast(Result, ResultType, /*isSigned*/true, 567 "cast"); 568 return RValue::get(Result); 569 } 570 case Builtin::BI__builtin_parity: 571 case Builtin::BI__builtin_parityl: 572 case Builtin::BI__builtin_parityll: { 573 // parity(x) -> ctpop(x) & 1 574 Value *ArgValue = EmitScalarExpr(E->getArg(0)); 575 576 llvm::Type *ArgType = ArgValue->getType(); 577 Value *F = CGM.getIntrinsic(Intrinsic::ctpop, ArgType); 578 579 llvm::Type *ResultType = ConvertType(E->getType()); 580 Value *Tmp = Builder.CreateCall(F, ArgValue); 581 Value *Result = Builder.CreateAnd(Tmp, llvm::ConstantInt::get(ArgType, 1)); 582 if (Result->getType() != ResultType) 583 Result = Builder.CreateIntCast(Result, ResultType, /*isSigned*/true, 584 "cast"); 585 return RValue::get(Result); 586 } 587 case Builtin::BI__builtin_popcount: 588 case Builtin::BI__builtin_popcountl: 589 case Builtin::BI__builtin_popcountll: { 590 Value *ArgValue = EmitScalarExpr(E->getArg(0)); 591 592 llvm::Type *ArgType = ArgValue->getType(); 593 Value *F = CGM.getIntrinsic(Intrinsic::ctpop, ArgType); 594 595 llvm::Type *ResultType = ConvertType(E->getType()); 596 Value *Result = Builder.CreateCall(F, ArgValue); 597 if (Result->getType() != ResultType) 598 Result = Builder.CreateIntCast(Result, ResultType, /*isSigned*/true, 599 "cast"); 600 return RValue::get(Result); 601 } 602 case Builtin::BI__builtin_unpredictable: { 603 // Always return the argument of __builtin_unpredictable. LLVM does not 604 // handle this builtin. Metadata for this builtin should be added directly 605 // to instructions such as branches or switches that use it. 606 return RValue::get(EmitScalarExpr(E->getArg(0))); 607 } 608 case Builtin::BI__builtin_expect: { 609 Value *ArgValue = EmitScalarExpr(E->getArg(0)); 610 llvm::Type *ArgType = ArgValue->getType(); 611 612 Value *ExpectedValue = EmitScalarExpr(E->getArg(1)); 613 // Don't generate llvm.expect on -O0 as the backend won't use it for 614 // anything. 615 // Note, we still IRGen ExpectedValue because it could have side-effects. 616 if (CGM.getCodeGenOpts().OptimizationLevel == 0) 617 return RValue::get(ArgValue); 618 619 Value *FnExpect = CGM.getIntrinsic(Intrinsic::expect, ArgType); 620 Value *Result = 621 Builder.CreateCall(FnExpect, {ArgValue, ExpectedValue}, "expval"); 622 return RValue::get(Result); 623 } 624 case Builtin::BI__builtin_assume_aligned: { 625 Value *PtrValue = EmitScalarExpr(E->getArg(0)); 626 Value *OffsetValue = 627 (E->getNumArgs() > 2) ? EmitScalarExpr(E->getArg(2)) : nullptr; 628 629 Value *AlignmentValue = EmitScalarExpr(E->getArg(1)); 630 ConstantInt *AlignmentCI = cast<ConstantInt>(AlignmentValue); 631 unsigned Alignment = (unsigned) AlignmentCI->getZExtValue(); 632 633 EmitAlignmentAssumption(PtrValue, Alignment, OffsetValue); 634 return RValue::get(PtrValue); 635 } 636 case Builtin::BI__assume: 637 case Builtin::BI__builtin_assume: { 638 if (E->getArg(0)->HasSideEffects(getContext())) 639 return RValue::get(nullptr); 640 641 Value *ArgValue = EmitScalarExpr(E->getArg(0)); 642 Value *FnAssume = CGM.getIntrinsic(Intrinsic::assume); 643 return RValue::get(Builder.CreateCall(FnAssume, ArgValue)); 644 } 645 case Builtin::BI__builtin_bswap16: 646 case Builtin::BI__builtin_bswap32: 647 case Builtin::BI__builtin_bswap64: { 648 Value *ArgValue = EmitScalarExpr(E->getArg(0)); 649 llvm::Type *ArgType = ArgValue->getType(); 650 Value *F = CGM.getIntrinsic(Intrinsic::bswap, ArgType); 651 return RValue::get(Builder.CreateCall(F, ArgValue)); 652 } 653 case Builtin::BI__builtin_object_size: { 654 unsigned Type = 655 E->getArg(1)->EvaluateKnownConstInt(getContext()).getZExtValue(); 656 auto *ResType = cast<llvm::IntegerType>(ConvertType(E->getType())); 657 658 // We pass this builtin onto the optimizer so that it can figure out the 659 // object size in more complex cases. 660 return RValue::get(emitBuiltinObjectSize(E->getArg(0), Type, ResType)); 661 } 662 case Builtin::BI__builtin_prefetch: { 663 Value *Locality, *RW, *Address = EmitScalarExpr(E->getArg(0)); 664 // FIXME: Technically these constants should of type 'int', yes? 665 RW = (E->getNumArgs() > 1) ? EmitScalarExpr(E->getArg(1)) : 666 llvm::ConstantInt::get(Int32Ty, 0); 667 Locality = (E->getNumArgs() > 2) ? EmitScalarExpr(E->getArg(2)) : 668 llvm::ConstantInt::get(Int32Ty, 3); 669 Value *Data = llvm::ConstantInt::get(Int32Ty, 1); 670 Value *F = CGM.getIntrinsic(Intrinsic::prefetch); 671 return RValue::get(Builder.CreateCall(F, {Address, RW, Locality, Data})); 672 } 673 case Builtin::BI__builtin_readcyclecounter: { 674 Value *F = CGM.getIntrinsic(Intrinsic::readcyclecounter); 675 return RValue::get(Builder.CreateCall(F)); 676 } 677 case Builtin::BI__builtin___clear_cache: { 678 Value *Begin = EmitScalarExpr(E->getArg(0)); 679 Value *End = EmitScalarExpr(E->getArg(1)); 680 Value *F = CGM.getIntrinsic(Intrinsic::clear_cache); 681 return RValue::get(Builder.CreateCall(F, {Begin, End})); 682 } 683 case Builtin::BI__builtin_trap: 684 return RValue::get(EmitTrapCall(Intrinsic::trap)); 685 case Builtin::BI__debugbreak: 686 return RValue::get(EmitTrapCall(Intrinsic::debugtrap)); 687 case Builtin::BI__builtin_unreachable: { 688 if (SanOpts.has(SanitizerKind::Unreachable)) { 689 SanitizerScope SanScope(this); 690 EmitCheck(std::make_pair(static_cast<llvm::Value *>(Builder.getFalse()), 691 SanitizerKind::Unreachable), 692 "builtin_unreachable", EmitCheckSourceLocation(E->getExprLoc()), 693 None); 694 } else 695 Builder.CreateUnreachable(); 696 697 // We do need to preserve an insertion point. 698 EmitBlock(createBasicBlock("unreachable.cont")); 699 700 return RValue::get(nullptr); 701 } 702 703 case Builtin::BI__builtin_powi: 704 case Builtin::BI__builtin_powif: 705 case Builtin::BI__builtin_powil: { 706 Value *Base = EmitScalarExpr(E->getArg(0)); 707 Value *Exponent = EmitScalarExpr(E->getArg(1)); 708 llvm::Type *ArgType = Base->getType(); 709 Value *F = CGM.getIntrinsic(Intrinsic::powi, ArgType); 710 return RValue::get(Builder.CreateCall(F, {Base, Exponent})); 711 } 712 713 case Builtin::BI__builtin_isgreater: 714 case Builtin::BI__builtin_isgreaterequal: 715 case Builtin::BI__builtin_isless: 716 case Builtin::BI__builtin_islessequal: 717 case Builtin::BI__builtin_islessgreater: 718 case Builtin::BI__builtin_isunordered: { 719 // Ordered comparisons: we know the arguments to these are matching scalar 720 // floating point values. 721 Value *LHS = EmitScalarExpr(E->getArg(0)); 722 Value *RHS = EmitScalarExpr(E->getArg(1)); 723 724 switch (BuiltinID) { 725 default: llvm_unreachable("Unknown ordered comparison"); 726 case Builtin::BI__builtin_isgreater: 727 LHS = Builder.CreateFCmpOGT(LHS, RHS, "cmp"); 728 break; 729 case Builtin::BI__builtin_isgreaterequal: 730 LHS = Builder.CreateFCmpOGE(LHS, RHS, "cmp"); 731 break; 732 case Builtin::BI__builtin_isless: 733 LHS = Builder.CreateFCmpOLT(LHS, RHS, "cmp"); 734 break; 735 case Builtin::BI__builtin_islessequal: 736 LHS = Builder.CreateFCmpOLE(LHS, RHS, "cmp"); 737 break; 738 case Builtin::BI__builtin_islessgreater: 739 LHS = Builder.CreateFCmpONE(LHS, RHS, "cmp"); 740 break; 741 case Builtin::BI__builtin_isunordered: 742 LHS = Builder.CreateFCmpUNO(LHS, RHS, "cmp"); 743 break; 744 } 745 // ZExt bool to int type. 746 return RValue::get(Builder.CreateZExt(LHS, ConvertType(E->getType()))); 747 } 748 case Builtin::BI__builtin_isnan: { 749 Value *V = EmitScalarExpr(E->getArg(0)); 750 V = Builder.CreateFCmpUNO(V, V, "cmp"); 751 return RValue::get(Builder.CreateZExt(V, ConvertType(E->getType()))); 752 } 753 754 case Builtin::BI__builtin_isinf: { 755 // isinf(x) --> fabs(x) == infinity 756 Value *V = EmitScalarExpr(E->getArg(0)); 757 V = EmitFAbs(*this, V); 758 759 V = Builder.CreateFCmpOEQ(V, ConstantFP::getInfinity(V->getType()),"isinf"); 760 return RValue::get(Builder.CreateZExt(V, ConvertType(E->getType()))); 761 } 762 763 case Builtin::BI__builtin_isinf_sign: { 764 // isinf_sign(x) -> fabs(x) == infinity ? (signbit(x) ? -1 : 1) : 0 765 Value *Arg = EmitScalarExpr(E->getArg(0)); 766 Value *AbsArg = EmitFAbs(*this, Arg); 767 Value *IsInf = Builder.CreateFCmpOEQ( 768 AbsArg, ConstantFP::getInfinity(Arg->getType()), "isinf"); 769 Value *IsNeg = EmitSignBit(*this, Arg); 770 771 llvm::Type *IntTy = ConvertType(E->getType()); 772 Value *Zero = Constant::getNullValue(IntTy); 773 Value *One = ConstantInt::get(IntTy, 1); 774 Value *NegativeOne = ConstantInt::get(IntTy, -1); 775 Value *SignResult = Builder.CreateSelect(IsNeg, NegativeOne, One); 776 Value *Result = Builder.CreateSelect(IsInf, SignResult, Zero); 777 return RValue::get(Result); 778 } 779 780 case Builtin::BI__builtin_isnormal: { 781 // isnormal(x) --> x == x && fabsf(x) < infinity && fabsf(x) >= float_min 782 Value *V = EmitScalarExpr(E->getArg(0)); 783 Value *Eq = Builder.CreateFCmpOEQ(V, V, "iseq"); 784 785 Value *Abs = EmitFAbs(*this, V); 786 Value *IsLessThanInf = 787 Builder.CreateFCmpULT(Abs, ConstantFP::getInfinity(V->getType()),"isinf"); 788 APFloat Smallest = APFloat::getSmallestNormalized( 789 getContext().getFloatTypeSemantics(E->getArg(0)->getType())); 790 Value *IsNormal = 791 Builder.CreateFCmpUGE(Abs, ConstantFP::get(V->getContext(), Smallest), 792 "isnormal"); 793 V = Builder.CreateAnd(Eq, IsLessThanInf, "and"); 794 V = Builder.CreateAnd(V, IsNormal, "and"); 795 return RValue::get(Builder.CreateZExt(V, ConvertType(E->getType()))); 796 } 797 798 case Builtin::BI__builtin_isfinite: { 799 // isfinite(x) --> x == x && fabs(x) != infinity; 800 Value *V = EmitScalarExpr(E->getArg(0)); 801 Value *Eq = Builder.CreateFCmpOEQ(V, V, "iseq"); 802 803 Value *Abs = EmitFAbs(*this, V); 804 Value *IsNotInf = 805 Builder.CreateFCmpUNE(Abs, ConstantFP::getInfinity(V->getType()),"isinf"); 806 807 V = Builder.CreateAnd(Eq, IsNotInf, "and"); 808 return RValue::get(Builder.CreateZExt(V, ConvertType(E->getType()))); 809 } 810 811 case Builtin::BI__builtin_fpclassify: { 812 Value *V = EmitScalarExpr(E->getArg(5)); 813 llvm::Type *Ty = ConvertType(E->getArg(5)->getType()); 814 815 // Create Result 816 BasicBlock *Begin = Builder.GetInsertBlock(); 817 BasicBlock *End = createBasicBlock("fpclassify_end", this->CurFn); 818 Builder.SetInsertPoint(End); 819 PHINode *Result = 820 Builder.CreatePHI(ConvertType(E->getArg(0)->getType()), 4, 821 "fpclassify_result"); 822 823 // if (V==0) return FP_ZERO 824 Builder.SetInsertPoint(Begin); 825 Value *IsZero = Builder.CreateFCmpOEQ(V, Constant::getNullValue(Ty), 826 "iszero"); 827 Value *ZeroLiteral = EmitScalarExpr(E->getArg(4)); 828 BasicBlock *NotZero = createBasicBlock("fpclassify_not_zero", this->CurFn); 829 Builder.CreateCondBr(IsZero, End, NotZero); 830 Result->addIncoming(ZeroLiteral, Begin); 831 832 // if (V != V) return FP_NAN 833 Builder.SetInsertPoint(NotZero); 834 Value *IsNan = Builder.CreateFCmpUNO(V, V, "cmp"); 835 Value *NanLiteral = EmitScalarExpr(E->getArg(0)); 836 BasicBlock *NotNan = createBasicBlock("fpclassify_not_nan", this->CurFn); 837 Builder.CreateCondBr(IsNan, End, NotNan); 838 Result->addIncoming(NanLiteral, NotZero); 839 840 // if (fabs(V) == infinity) return FP_INFINITY 841 Builder.SetInsertPoint(NotNan); 842 Value *VAbs = EmitFAbs(*this, V); 843 Value *IsInf = 844 Builder.CreateFCmpOEQ(VAbs, ConstantFP::getInfinity(V->getType()), 845 "isinf"); 846 Value *InfLiteral = EmitScalarExpr(E->getArg(1)); 847 BasicBlock *NotInf = createBasicBlock("fpclassify_not_inf", this->CurFn); 848 Builder.CreateCondBr(IsInf, End, NotInf); 849 Result->addIncoming(InfLiteral, NotNan); 850 851 // if (fabs(V) >= MIN_NORMAL) return FP_NORMAL else FP_SUBNORMAL 852 Builder.SetInsertPoint(NotInf); 853 APFloat Smallest = APFloat::getSmallestNormalized( 854 getContext().getFloatTypeSemantics(E->getArg(5)->getType())); 855 Value *IsNormal = 856 Builder.CreateFCmpUGE(VAbs, ConstantFP::get(V->getContext(), Smallest), 857 "isnormal"); 858 Value *NormalResult = 859 Builder.CreateSelect(IsNormal, EmitScalarExpr(E->getArg(2)), 860 EmitScalarExpr(E->getArg(3))); 861 Builder.CreateBr(End); 862 Result->addIncoming(NormalResult, NotInf); 863 864 // return Result 865 Builder.SetInsertPoint(End); 866 return RValue::get(Result); 867 } 868 869 case Builtin::BIalloca: 870 case Builtin::BI_alloca: 871 case Builtin::BI__builtin_alloca: { 872 Value *Size = EmitScalarExpr(E->getArg(0)); 873 return RValue::get(Builder.CreateAlloca(Builder.getInt8Ty(), Size)); 874 } 875 case Builtin::BIbzero: 876 case Builtin::BI__builtin_bzero: { 877 Address Dest = EmitPointerWithAlignment(E->getArg(0)); 878 Value *SizeVal = EmitScalarExpr(E->getArg(1)); 879 EmitNonNullArgCheck(RValue::get(Dest.getPointer()), E->getArg(0)->getType(), 880 E->getArg(0)->getExprLoc(), FD, 0); 881 Builder.CreateMemSet(Dest, Builder.getInt8(0), SizeVal, false); 882 return RValue::get(Dest.getPointer()); 883 } 884 case Builtin::BImemcpy: 885 case Builtin::BI__builtin_memcpy: { 886 Address Dest = EmitPointerWithAlignment(E->getArg(0)); 887 Address Src = EmitPointerWithAlignment(E->getArg(1)); 888 Value *SizeVal = EmitScalarExpr(E->getArg(2)); 889 EmitNonNullArgCheck(RValue::get(Dest.getPointer()), E->getArg(0)->getType(), 890 E->getArg(0)->getExprLoc(), FD, 0); 891 EmitNonNullArgCheck(RValue::get(Src.getPointer()), E->getArg(1)->getType(), 892 E->getArg(1)->getExprLoc(), FD, 1); 893 Builder.CreateMemCpy(Dest, Src, SizeVal, false); 894 return RValue::get(Dest.getPointer()); 895 } 896 897 case Builtin::BI__builtin___memcpy_chk: { 898 // fold __builtin_memcpy_chk(x, y, cst1, cst2) to memcpy iff cst1<=cst2. 899 llvm::APSInt Size, DstSize; 900 if (!E->getArg(2)->EvaluateAsInt(Size, CGM.getContext()) || 901 !E->getArg(3)->EvaluateAsInt(DstSize, CGM.getContext())) 902 break; 903 if (Size.ugt(DstSize)) 904 break; 905 Address Dest = EmitPointerWithAlignment(E->getArg(0)); 906 Address Src = EmitPointerWithAlignment(E->getArg(1)); 907 Value *SizeVal = llvm::ConstantInt::get(Builder.getContext(), Size); 908 Builder.CreateMemCpy(Dest, Src, SizeVal, false); 909 return RValue::get(Dest.getPointer()); 910 } 911 912 case Builtin::BI__builtin_objc_memmove_collectable: { 913 Address DestAddr = EmitPointerWithAlignment(E->getArg(0)); 914 Address SrcAddr = EmitPointerWithAlignment(E->getArg(1)); 915 Value *SizeVal = EmitScalarExpr(E->getArg(2)); 916 CGM.getObjCRuntime().EmitGCMemmoveCollectable(*this, 917 DestAddr, SrcAddr, SizeVal); 918 return RValue::get(DestAddr.getPointer()); 919 } 920 921 case Builtin::BI__builtin___memmove_chk: { 922 // fold __builtin_memmove_chk(x, y, cst1, cst2) to memmove iff cst1<=cst2. 923 llvm::APSInt Size, DstSize; 924 if (!E->getArg(2)->EvaluateAsInt(Size, CGM.getContext()) || 925 !E->getArg(3)->EvaluateAsInt(DstSize, CGM.getContext())) 926 break; 927 if (Size.ugt(DstSize)) 928 break; 929 Address Dest = EmitPointerWithAlignment(E->getArg(0)); 930 Address Src = EmitPointerWithAlignment(E->getArg(1)); 931 Value *SizeVal = llvm::ConstantInt::get(Builder.getContext(), Size); 932 Builder.CreateMemMove(Dest, Src, SizeVal, false); 933 return RValue::get(Dest.getPointer()); 934 } 935 936 case Builtin::BImemmove: 937 case Builtin::BI__builtin_memmove: { 938 Address Dest = EmitPointerWithAlignment(E->getArg(0)); 939 Address Src = EmitPointerWithAlignment(E->getArg(1)); 940 Value *SizeVal = EmitScalarExpr(E->getArg(2)); 941 EmitNonNullArgCheck(RValue::get(Dest.getPointer()), E->getArg(0)->getType(), 942 E->getArg(0)->getExprLoc(), FD, 0); 943 EmitNonNullArgCheck(RValue::get(Src.getPointer()), E->getArg(1)->getType(), 944 E->getArg(1)->getExprLoc(), FD, 1); 945 Builder.CreateMemMove(Dest, Src, SizeVal, false); 946 return RValue::get(Dest.getPointer()); 947 } 948 case Builtin::BImemset: 949 case Builtin::BI__builtin_memset: { 950 Address Dest = EmitPointerWithAlignment(E->getArg(0)); 951 Value *ByteVal = Builder.CreateTrunc(EmitScalarExpr(E->getArg(1)), 952 Builder.getInt8Ty()); 953 Value *SizeVal = EmitScalarExpr(E->getArg(2)); 954 EmitNonNullArgCheck(RValue::get(Dest.getPointer()), E->getArg(0)->getType(), 955 E->getArg(0)->getExprLoc(), FD, 0); 956 Builder.CreateMemSet(Dest, ByteVal, SizeVal, false); 957 return RValue::get(Dest.getPointer()); 958 } 959 case Builtin::BI__builtin___memset_chk: { 960 // fold __builtin_memset_chk(x, y, cst1, cst2) to memset iff cst1<=cst2. 961 llvm::APSInt Size, DstSize; 962 if (!E->getArg(2)->EvaluateAsInt(Size, CGM.getContext()) || 963 !E->getArg(3)->EvaluateAsInt(DstSize, CGM.getContext())) 964 break; 965 if (Size.ugt(DstSize)) 966 break; 967 Address Dest = EmitPointerWithAlignment(E->getArg(0)); 968 Value *ByteVal = Builder.CreateTrunc(EmitScalarExpr(E->getArg(1)), 969 Builder.getInt8Ty()); 970 Value *SizeVal = llvm::ConstantInt::get(Builder.getContext(), Size); 971 Builder.CreateMemSet(Dest, ByteVal, SizeVal, false); 972 return RValue::get(Dest.getPointer()); 973 } 974 case Builtin::BI__builtin_dwarf_cfa: { 975 // The offset in bytes from the first argument to the CFA. 976 // 977 // Why on earth is this in the frontend? Is there any reason at 978 // all that the backend can't reasonably determine this while 979 // lowering llvm.eh.dwarf.cfa()? 980 // 981 // TODO: If there's a satisfactory reason, add a target hook for 982 // this instead of hard-coding 0, which is correct for most targets. 983 int32_t Offset = 0; 984 985 Value *F = CGM.getIntrinsic(Intrinsic::eh_dwarf_cfa); 986 return RValue::get(Builder.CreateCall(F, 987 llvm::ConstantInt::get(Int32Ty, Offset))); 988 } 989 case Builtin::BI__builtin_return_address: { 990 Value *Depth = 991 CGM.EmitConstantExpr(E->getArg(0), getContext().UnsignedIntTy, this); 992 Value *F = CGM.getIntrinsic(Intrinsic::returnaddress); 993 return RValue::get(Builder.CreateCall(F, Depth)); 994 } 995 case Builtin::BI__builtin_frame_address: { 996 Value *Depth = 997 CGM.EmitConstantExpr(E->getArg(0), getContext().UnsignedIntTy, this); 998 Value *F = CGM.getIntrinsic(Intrinsic::frameaddress); 999 return RValue::get(Builder.CreateCall(F, Depth)); 1000 } 1001 case Builtin::BI__builtin_extract_return_addr: { 1002 Value *Address = EmitScalarExpr(E->getArg(0)); 1003 Value *Result = getTargetHooks().decodeReturnAddress(*this, Address); 1004 return RValue::get(Result); 1005 } 1006 case Builtin::BI__builtin_frob_return_addr: { 1007 Value *Address = EmitScalarExpr(E->getArg(0)); 1008 Value *Result = getTargetHooks().encodeReturnAddress(*this, Address); 1009 return RValue::get(Result); 1010 } 1011 case Builtin::BI__builtin_dwarf_sp_column: { 1012 llvm::IntegerType *Ty 1013 = cast<llvm::IntegerType>(ConvertType(E->getType())); 1014 int Column = getTargetHooks().getDwarfEHStackPointer(CGM); 1015 if (Column == -1) { 1016 CGM.ErrorUnsupported(E, "__builtin_dwarf_sp_column"); 1017 return RValue::get(llvm::UndefValue::get(Ty)); 1018 } 1019 return RValue::get(llvm::ConstantInt::get(Ty, Column, true)); 1020 } 1021 case Builtin::BI__builtin_init_dwarf_reg_size_table: { 1022 Value *Address = EmitScalarExpr(E->getArg(0)); 1023 if (getTargetHooks().initDwarfEHRegSizeTable(*this, Address)) 1024 CGM.ErrorUnsupported(E, "__builtin_init_dwarf_reg_size_table"); 1025 return RValue::get(llvm::UndefValue::get(ConvertType(E->getType()))); 1026 } 1027 case Builtin::BI__builtin_eh_return: { 1028 Value *Int = EmitScalarExpr(E->getArg(0)); 1029 Value *Ptr = EmitScalarExpr(E->getArg(1)); 1030 1031 llvm::IntegerType *IntTy = cast<llvm::IntegerType>(Int->getType()); 1032 assert((IntTy->getBitWidth() == 32 || IntTy->getBitWidth() == 64) && 1033 "LLVM's __builtin_eh_return only supports 32- and 64-bit variants"); 1034 Value *F = CGM.getIntrinsic(IntTy->getBitWidth() == 32 1035 ? Intrinsic::eh_return_i32 1036 : Intrinsic::eh_return_i64); 1037 Builder.CreateCall(F, {Int, Ptr}); 1038 Builder.CreateUnreachable(); 1039 1040 // We do need to preserve an insertion point. 1041 EmitBlock(createBasicBlock("builtin_eh_return.cont")); 1042 1043 return RValue::get(nullptr); 1044 } 1045 case Builtin::BI__builtin_unwind_init: { 1046 Value *F = CGM.getIntrinsic(Intrinsic::eh_unwind_init); 1047 return RValue::get(Builder.CreateCall(F)); 1048 } 1049 case Builtin::BI__builtin_extend_pointer: { 1050 // Extends a pointer to the size of an _Unwind_Word, which is 1051 // uint64_t on all platforms. Generally this gets poked into a 1052 // register and eventually used as an address, so if the 1053 // addressing registers are wider than pointers and the platform 1054 // doesn't implicitly ignore high-order bits when doing 1055 // addressing, we need to make sure we zext / sext based on 1056 // the platform's expectations. 1057 // 1058 // See: http://gcc.gnu.org/ml/gcc-bugs/2002-02/msg00237.html 1059 1060 // Cast the pointer to intptr_t. 1061 Value *Ptr = EmitScalarExpr(E->getArg(0)); 1062 Value *Result = Builder.CreatePtrToInt(Ptr, IntPtrTy, "extend.cast"); 1063 1064 // If that's 64 bits, we're done. 1065 if (IntPtrTy->getBitWidth() == 64) 1066 return RValue::get(Result); 1067 1068 // Otherwise, ask the codegen data what to do. 1069 if (getTargetHooks().extendPointerWithSExt()) 1070 return RValue::get(Builder.CreateSExt(Result, Int64Ty, "extend.sext")); 1071 else 1072 return RValue::get(Builder.CreateZExt(Result, Int64Ty, "extend.zext")); 1073 } 1074 case Builtin::BI__builtin_setjmp: { 1075 // Buffer is a void**. 1076 Address Buf = EmitPointerWithAlignment(E->getArg(0)); 1077 1078 // Store the frame pointer to the setjmp buffer. 1079 Value *FrameAddr = 1080 Builder.CreateCall(CGM.getIntrinsic(Intrinsic::frameaddress), 1081 ConstantInt::get(Int32Ty, 0)); 1082 Builder.CreateStore(FrameAddr, Buf); 1083 1084 // Store the stack pointer to the setjmp buffer. 1085 Value *StackAddr = 1086 Builder.CreateCall(CGM.getIntrinsic(Intrinsic::stacksave)); 1087 Address StackSaveSlot = 1088 Builder.CreateConstInBoundsGEP(Buf, 2, getPointerSize()); 1089 Builder.CreateStore(StackAddr, StackSaveSlot); 1090 1091 // Call LLVM's EH setjmp, which is lightweight. 1092 Value *F = CGM.getIntrinsic(Intrinsic::eh_sjlj_setjmp); 1093 Buf = Builder.CreateBitCast(Buf, Int8PtrTy); 1094 return RValue::get(Builder.CreateCall(F, Buf.getPointer())); 1095 } 1096 case Builtin::BI__builtin_longjmp: { 1097 Value *Buf = EmitScalarExpr(E->getArg(0)); 1098 Buf = Builder.CreateBitCast(Buf, Int8PtrTy); 1099 1100 // Call LLVM's EH longjmp, which is lightweight. 1101 Builder.CreateCall(CGM.getIntrinsic(Intrinsic::eh_sjlj_longjmp), Buf); 1102 1103 // longjmp doesn't return; mark this as unreachable. 1104 Builder.CreateUnreachable(); 1105 1106 // We do need to preserve an insertion point. 1107 EmitBlock(createBasicBlock("longjmp.cont")); 1108 1109 return RValue::get(nullptr); 1110 } 1111 case Builtin::BI__sync_fetch_and_add: 1112 case Builtin::BI__sync_fetch_and_sub: 1113 case Builtin::BI__sync_fetch_and_or: 1114 case Builtin::BI__sync_fetch_and_and: 1115 case Builtin::BI__sync_fetch_and_xor: 1116 case Builtin::BI__sync_fetch_and_nand: 1117 case Builtin::BI__sync_add_and_fetch: 1118 case Builtin::BI__sync_sub_and_fetch: 1119 case Builtin::BI__sync_and_and_fetch: 1120 case Builtin::BI__sync_or_and_fetch: 1121 case Builtin::BI__sync_xor_and_fetch: 1122 case Builtin::BI__sync_nand_and_fetch: 1123 case Builtin::BI__sync_val_compare_and_swap: 1124 case Builtin::BI__sync_bool_compare_and_swap: 1125 case Builtin::BI__sync_lock_test_and_set: 1126 case Builtin::BI__sync_lock_release: 1127 case Builtin::BI__sync_swap: 1128 llvm_unreachable("Shouldn't make it through sema"); 1129 case Builtin::BI__sync_fetch_and_add_1: 1130 case Builtin::BI__sync_fetch_and_add_2: 1131 case Builtin::BI__sync_fetch_and_add_4: 1132 case Builtin::BI__sync_fetch_and_add_8: 1133 case Builtin::BI__sync_fetch_and_add_16: 1134 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Add, E); 1135 case Builtin::BI__sync_fetch_and_sub_1: 1136 case Builtin::BI__sync_fetch_and_sub_2: 1137 case Builtin::BI__sync_fetch_and_sub_4: 1138 case Builtin::BI__sync_fetch_and_sub_8: 1139 case Builtin::BI__sync_fetch_and_sub_16: 1140 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Sub, E); 1141 case Builtin::BI__sync_fetch_and_or_1: 1142 case Builtin::BI__sync_fetch_and_or_2: 1143 case Builtin::BI__sync_fetch_and_or_4: 1144 case Builtin::BI__sync_fetch_and_or_8: 1145 case Builtin::BI__sync_fetch_and_or_16: 1146 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Or, E); 1147 case Builtin::BI__sync_fetch_and_and_1: 1148 case Builtin::BI__sync_fetch_and_and_2: 1149 case Builtin::BI__sync_fetch_and_and_4: 1150 case Builtin::BI__sync_fetch_and_and_8: 1151 case Builtin::BI__sync_fetch_and_and_16: 1152 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::And, E); 1153 case Builtin::BI__sync_fetch_and_xor_1: 1154 case Builtin::BI__sync_fetch_and_xor_2: 1155 case Builtin::BI__sync_fetch_and_xor_4: 1156 case Builtin::BI__sync_fetch_and_xor_8: 1157 case Builtin::BI__sync_fetch_and_xor_16: 1158 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Xor, E); 1159 case Builtin::BI__sync_fetch_and_nand_1: 1160 case Builtin::BI__sync_fetch_and_nand_2: 1161 case Builtin::BI__sync_fetch_and_nand_4: 1162 case Builtin::BI__sync_fetch_and_nand_8: 1163 case Builtin::BI__sync_fetch_and_nand_16: 1164 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Nand, E); 1165 1166 // Clang extensions: not overloaded yet. 1167 case Builtin::BI__sync_fetch_and_min: 1168 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Min, E); 1169 case Builtin::BI__sync_fetch_and_max: 1170 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Max, E); 1171 case Builtin::BI__sync_fetch_and_umin: 1172 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::UMin, E); 1173 case Builtin::BI__sync_fetch_and_umax: 1174 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::UMax, E); 1175 1176 case Builtin::BI__sync_add_and_fetch_1: 1177 case Builtin::BI__sync_add_and_fetch_2: 1178 case Builtin::BI__sync_add_and_fetch_4: 1179 case Builtin::BI__sync_add_and_fetch_8: 1180 case Builtin::BI__sync_add_and_fetch_16: 1181 return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::Add, E, 1182 llvm::Instruction::Add); 1183 case Builtin::BI__sync_sub_and_fetch_1: 1184 case Builtin::BI__sync_sub_and_fetch_2: 1185 case Builtin::BI__sync_sub_and_fetch_4: 1186 case Builtin::BI__sync_sub_and_fetch_8: 1187 case Builtin::BI__sync_sub_and_fetch_16: 1188 return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::Sub, E, 1189 llvm::Instruction::Sub); 1190 case Builtin::BI__sync_and_and_fetch_1: 1191 case Builtin::BI__sync_and_and_fetch_2: 1192 case Builtin::BI__sync_and_and_fetch_4: 1193 case Builtin::BI__sync_and_and_fetch_8: 1194 case Builtin::BI__sync_and_and_fetch_16: 1195 return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::And, E, 1196 llvm::Instruction::And); 1197 case Builtin::BI__sync_or_and_fetch_1: 1198 case Builtin::BI__sync_or_and_fetch_2: 1199 case Builtin::BI__sync_or_and_fetch_4: 1200 case Builtin::BI__sync_or_and_fetch_8: 1201 case Builtin::BI__sync_or_and_fetch_16: 1202 return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::Or, E, 1203 llvm::Instruction::Or); 1204 case Builtin::BI__sync_xor_and_fetch_1: 1205 case Builtin::BI__sync_xor_and_fetch_2: 1206 case Builtin::BI__sync_xor_and_fetch_4: 1207 case Builtin::BI__sync_xor_and_fetch_8: 1208 case Builtin::BI__sync_xor_and_fetch_16: 1209 return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::Xor, E, 1210 llvm::Instruction::Xor); 1211 case Builtin::BI__sync_nand_and_fetch_1: 1212 case Builtin::BI__sync_nand_and_fetch_2: 1213 case Builtin::BI__sync_nand_and_fetch_4: 1214 case Builtin::BI__sync_nand_and_fetch_8: 1215 case Builtin::BI__sync_nand_and_fetch_16: 1216 return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::Nand, E, 1217 llvm::Instruction::And, true); 1218 1219 case Builtin::BI__sync_val_compare_and_swap_1: 1220 case Builtin::BI__sync_val_compare_and_swap_2: 1221 case Builtin::BI__sync_val_compare_and_swap_4: 1222 case Builtin::BI__sync_val_compare_and_swap_8: 1223 case Builtin::BI__sync_val_compare_and_swap_16: 1224 return RValue::get(MakeAtomicCmpXchgValue(*this, E, false)); 1225 1226 case Builtin::BI__sync_bool_compare_and_swap_1: 1227 case Builtin::BI__sync_bool_compare_and_swap_2: 1228 case Builtin::BI__sync_bool_compare_and_swap_4: 1229 case Builtin::BI__sync_bool_compare_and_swap_8: 1230 case Builtin::BI__sync_bool_compare_and_swap_16: 1231 return RValue::get(MakeAtomicCmpXchgValue(*this, E, true)); 1232 1233 case Builtin::BI__sync_swap_1: 1234 case Builtin::BI__sync_swap_2: 1235 case Builtin::BI__sync_swap_4: 1236 case Builtin::BI__sync_swap_8: 1237 case Builtin::BI__sync_swap_16: 1238 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Xchg, E); 1239 1240 case Builtin::BI__sync_lock_test_and_set_1: 1241 case Builtin::BI__sync_lock_test_and_set_2: 1242 case Builtin::BI__sync_lock_test_and_set_4: 1243 case Builtin::BI__sync_lock_test_and_set_8: 1244 case Builtin::BI__sync_lock_test_and_set_16: 1245 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Xchg, E); 1246 1247 case Builtin::BI__sync_lock_release_1: 1248 case Builtin::BI__sync_lock_release_2: 1249 case Builtin::BI__sync_lock_release_4: 1250 case Builtin::BI__sync_lock_release_8: 1251 case Builtin::BI__sync_lock_release_16: { 1252 Value *Ptr = EmitScalarExpr(E->getArg(0)); 1253 QualType ElTy = E->getArg(0)->getType()->getPointeeType(); 1254 CharUnits StoreSize = getContext().getTypeSizeInChars(ElTy); 1255 llvm::Type *ITy = llvm::IntegerType::get(getLLVMContext(), 1256 StoreSize.getQuantity() * 8); 1257 Ptr = Builder.CreateBitCast(Ptr, ITy->getPointerTo()); 1258 llvm::StoreInst *Store = 1259 Builder.CreateAlignedStore(llvm::Constant::getNullValue(ITy), Ptr, 1260 StoreSize); 1261 Store->setAtomic(llvm::Release); 1262 return RValue::get(nullptr); 1263 } 1264 1265 case Builtin::BI__sync_synchronize: { 1266 // We assume this is supposed to correspond to a C++0x-style 1267 // sequentially-consistent fence (i.e. this is only usable for 1268 // synchonization, not device I/O or anything like that). This intrinsic 1269 // is really badly designed in the sense that in theory, there isn't 1270 // any way to safely use it... but in practice, it mostly works 1271 // to use it with non-atomic loads and stores to get acquire/release 1272 // semantics. 1273 Builder.CreateFence(llvm::SequentiallyConsistent); 1274 return RValue::get(nullptr); 1275 } 1276 1277 case Builtin::BI__builtin_nontemporal_load: 1278 return RValue::get(EmitNontemporalLoad(*this, E)); 1279 case Builtin::BI__builtin_nontemporal_store: 1280 return RValue::get(EmitNontemporalStore(*this, E)); 1281 case Builtin::BI__c11_atomic_is_lock_free: 1282 case Builtin::BI__atomic_is_lock_free: { 1283 // Call "bool __atomic_is_lock_free(size_t size, void *ptr)". For the 1284 // __c11 builtin, ptr is 0 (indicating a properly-aligned object), since 1285 // _Atomic(T) is always properly-aligned. 1286 const char *LibCallName = "__atomic_is_lock_free"; 1287 CallArgList Args; 1288 Args.add(RValue::get(EmitScalarExpr(E->getArg(0))), 1289 getContext().getSizeType()); 1290 if (BuiltinID == Builtin::BI__atomic_is_lock_free) 1291 Args.add(RValue::get(EmitScalarExpr(E->getArg(1))), 1292 getContext().VoidPtrTy); 1293 else 1294 Args.add(RValue::get(llvm::Constant::getNullValue(VoidPtrTy)), 1295 getContext().VoidPtrTy); 1296 const CGFunctionInfo &FuncInfo = 1297 CGM.getTypes().arrangeFreeFunctionCall(E->getType(), Args, 1298 FunctionType::ExtInfo(), 1299 RequiredArgs::All); 1300 llvm::FunctionType *FTy = CGM.getTypes().GetFunctionType(FuncInfo); 1301 llvm::Constant *Func = CGM.CreateRuntimeFunction(FTy, LibCallName); 1302 return EmitCall(FuncInfo, Func, ReturnValueSlot(), Args); 1303 } 1304 1305 case Builtin::BI__atomic_test_and_set: { 1306 // Look at the argument type to determine whether this is a volatile 1307 // operation. The parameter type is always volatile. 1308 QualType PtrTy = E->getArg(0)->IgnoreImpCasts()->getType(); 1309 bool Volatile = 1310 PtrTy->castAs<PointerType>()->getPointeeType().isVolatileQualified(); 1311 1312 Value *Ptr = EmitScalarExpr(E->getArg(0)); 1313 unsigned AddrSpace = Ptr->getType()->getPointerAddressSpace(); 1314 Ptr = Builder.CreateBitCast(Ptr, Int8Ty->getPointerTo(AddrSpace)); 1315 Value *NewVal = Builder.getInt8(1); 1316 Value *Order = EmitScalarExpr(E->getArg(1)); 1317 if (isa<llvm::ConstantInt>(Order)) { 1318 int ord = cast<llvm::ConstantInt>(Order)->getZExtValue(); 1319 AtomicRMWInst *Result = nullptr; 1320 switch (ord) { 1321 case 0: // memory_order_relaxed 1322 default: // invalid order 1323 Result = Builder.CreateAtomicRMW(llvm::AtomicRMWInst::Xchg, 1324 Ptr, NewVal, 1325 llvm::Monotonic); 1326 break; 1327 case 1: // memory_order_consume 1328 case 2: // memory_order_acquire 1329 Result = Builder.CreateAtomicRMW(llvm::AtomicRMWInst::Xchg, 1330 Ptr, NewVal, 1331 llvm::Acquire); 1332 break; 1333 case 3: // memory_order_release 1334 Result = Builder.CreateAtomicRMW(llvm::AtomicRMWInst::Xchg, 1335 Ptr, NewVal, 1336 llvm::Release); 1337 break; 1338 case 4: // memory_order_acq_rel 1339 Result = Builder.CreateAtomicRMW(llvm::AtomicRMWInst::Xchg, 1340 Ptr, NewVal, 1341 llvm::AcquireRelease); 1342 break; 1343 case 5: // memory_order_seq_cst 1344 Result = Builder.CreateAtomicRMW(llvm::AtomicRMWInst::Xchg, 1345 Ptr, NewVal, 1346 llvm::SequentiallyConsistent); 1347 break; 1348 } 1349 Result->setVolatile(Volatile); 1350 return RValue::get(Builder.CreateIsNotNull(Result, "tobool")); 1351 } 1352 1353 llvm::BasicBlock *ContBB = createBasicBlock("atomic.continue", CurFn); 1354 1355 llvm::BasicBlock *BBs[5] = { 1356 createBasicBlock("monotonic", CurFn), 1357 createBasicBlock("acquire", CurFn), 1358 createBasicBlock("release", CurFn), 1359 createBasicBlock("acqrel", CurFn), 1360 createBasicBlock("seqcst", CurFn) 1361 }; 1362 llvm::AtomicOrdering Orders[5] = { 1363 llvm::Monotonic, llvm::Acquire, llvm::Release, 1364 llvm::AcquireRelease, llvm::SequentiallyConsistent 1365 }; 1366 1367 Order = Builder.CreateIntCast(Order, Builder.getInt32Ty(), false); 1368 llvm::SwitchInst *SI = Builder.CreateSwitch(Order, BBs[0]); 1369 1370 Builder.SetInsertPoint(ContBB); 1371 PHINode *Result = Builder.CreatePHI(Int8Ty, 5, "was_set"); 1372 1373 for (unsigned i = 0; i < 5; ++i) { 1374 Builder.SetInsertPoint(BBs[i]); 1375 AtomicRMWInst *RMW = Builder.CreateAtomicRMW(llvm::AtomicRMWInst::Xchg, 1376 Ptr, NewVal, Orders[i]); 1377 RMW->setVolatile(Volatile); 1378 Result->addIncoming(RMW, BBs[i]); 1379 Builder.CreateBr(ContBB); 1380 } 1381 1382 SI->addCase(Builder.getInt32(0), BBs[0]); 1383 SI->addCase(Builder.getInt32(1), BBs[1]); 1384 SI->addCase(Builder.getInt32(2), BBs[1]); 1385 SI->addCase(Builder.getInt32(3), BBs[2]); 1386 SI->addCase(Builder.getInt32(4), BBs[3]); 1387 SI->addCase(Builder.getInt32(5), BBs[4]); 1388 1389 Builder.SetInsertPoint(ContBB); 1390 return RValue::get(Builder.CreateIsNotNull(Result, "tobool")); 1391 } 1392 1393 case Builtin::BI__atomic_clear: { 1394 QualType PtrTy = E->getArg(0)->IgnoreImpCasts()->getType(); 1395 bool Volatile = 1396 PtrTy->castAs<PointerType>()->getPointeeType().isVolatileQualified(); 1397 1398 Address Ptr = EmitPointerWithAlignment(E->getArg(0)); 1399 unsigned AddrSpace = Ptr.getPointer()->getType()->getPointerAddressSpace(); 1400 Ptr = Builder.CreateBitCast(Ptr, Int8Ty->getPointerTo(AddrSpace)); 1401 Value *NewVal = Builder.getInt8(0); 1402 Value *Order = EmitScalarExpr(E->getArg(1)); 1403 if (isa<llvm::ConstantInt>(Order)) { 1404 int ord = cast<llvm::ConstantInt>(Order)->getZExtValue(); 1405 StoreInst *Store = Builder.CreateStore(NewVal, Ptr, Volatile); 1406 switch (ord) { 1407 case 0: // memory_order_relaxed 1408 default: // invalid order 1409 Store->setOrdering(llvm::Monotonic); 1410 break; 1411 case 3: // memory_order_release 1412 Store->setOrdering(llvm::Release); 1413 break; 1414 case 5: // memory_order_seq_cst 1415 Store->setOrdering(llvm::SequentiallyConsistent); 1416 break; 1417 } 1418 return RValue::get(nullptr); 1419 } 1420 1421 llvm::BasicBlock *ContBB = createBasicBlock("atomic.continue", CurFn); 1422 1423 llvm::BasicBlock *BBs[3] = { 1424 createBasicBlock("monotonic", CurFn), 1425 createBasicBlock("release", CurFn), 1426 createBasicBlock("seqcst", CurFn) 1427 }; 1428 llvm::AtomicOrdering Orders[3] = { 1429 llvm::Monotonic, llvm::Release, llvm::SequentiallyConsistent 1430 }; 1431 1432 Order = Builder.CreateIntCast(Order, Builder.getInt32Ty(), false); 1433 llvm::SwitchInst *SI = Builder.CreateSwitch(Order, BBs[0]); 1434 1435 for (unsigned i = 0; i < 3; ++i) { 1436 Builder.SetInsertPoint(BBs[i]); 1437 StoreInst *Store = Builder.CreateStore(NewVal, Ptr, Volatile); 1438 Store->setOrdering(Orders[i]); 1439 Builder.CreateBr(ContBB); 1440 } 1441 1442 SI->addCase(Builder.getInt32(0), BBs[0]); 1443 SI->addCase(Builder.getInt32(3), BBs[1]); 1444 SI->addCase(Builder.getInt32(5), BBs[2]); 1445 1446 Builder.SetInsertPoint(ContBB); 1447 return RValue::get(nullptr); 1448 } 1449 1450 case Builtin::BI__atomic_thread_fence: 1451 case Builtin::BI__atomic_signal_fence: 1452 case Builtin::BI__c11_atomic_thread_fence: 1453 case Builtin::BI__c11_atomic_signal_fence: { 1454 llvm::SynchronizationScope Scope; 1455 if (BuiltinID == Builtin::BI__atomic_signal_fence || 1456 BuiltinID == Builtin::BI__c11_atomic_signal_fence) 1457 Scope = llvm::SingleThread; 1458 else 1459 Scope = llvm::CrossThread; 1460 Value *Order = EmitScalarExpr(E->getArg(0)); 1461 if (isa<llvm::ConstantInt>(Order)) { 1462 int ord = cast<llvm::ConstantInt>(Order)->getZExtValue(); 1463 switch (ord) { 1464 case 0: // memory_order_relaxed 1465 default: // invalid order 1466 break; 1467 case 1: // memory_order_consume 1468 case 2: // memory_order_acquire 1469 Builder.CreateFence(llvm::Acquire, Scope); 1470 break; 1471 case 3: // memory_order_release 1472 Builder.CreateFence(llvm::Release, Scope); 1473 break; 1474 case 4: // memory_order_acq_rel 1475 Builder.CreateFence(llvm::AcquireRelease, Scope); 1476 break; 1477 case 5: // memory_order_seq_cst 1478 Builder.CreateFence(llvm::SequentiallyConsistent, Scope); 1479 break; 1480 } 1481 return RValue::get(nullptr); 1482 } 1483 1484 llvm::BasicBlock *AcquireBB, *ReleaseBB, *AcqRelBB, *SeqCstBB; 1485 AcquireBB = createBasicBlock("acquire", CurFn); 1486 ReleaseBB = createBasicBlock("release", CurFn); 1487 AcqRelBB = createBasicBlock("acqrel", CurFn); 1488 SeqCstBB = createBasicBlock("seqcst", CurFn); 1489 llvm::BasicBlock *ContBB = createBasicBlock("atomic.continue", CurFn); 1490 1491 Order = Builder.CreateIntCast(Order, Builder.getInt32Ty(), false); 1492 llvm::SwitchInst *SI = Builder.CreateSwitch(Order, ContBB); 1493 1494 Builder.SetInsertPoint(AcquireBB); 1495 Builder.CreateFence(llvm::Acquire, Scope); 1496 Builder.CreateBr(ContBB); 1497 SI->addCase(Builder.getInt32(1), AcquireBB); 1498 SI->addCase(Builder.getInt32(2), AcquireBB); 1499 1500 Builder.SetInsertPoint(ReleaseBB); 1501 Builder.CreateFence(llvm::Release, Scope); 1502 Builder.CreateBr(ContBB); 1503 SI->addCase(Builder.getInt32(3), ReleaseBB); 1504 1505 Builder.SetInsertPoint(AcqRelBB); 1506 Builder.CreateFence(llvm::AcquireRelease, Scope); 1507 Builder.CreateBr(ContBB); 1508 SI->addCase(Builder.getInt32(4), AcqRelBB); 1509 1510 Builder.SetInsertPoint(SeqCstBB); 1511 Builder.CreateFence(llvm::SequentiallyConsistent, Scope); 1512 Builder.CreateBr(ContBB); 1513 SI->addCase(Builder.getInt32(5), SeqCstBB); 1514 1515 Builder.SetInsertPoint(ContBB); 1516 return RValue::get(nullptr); 1517 } 1518 1519 // Library functions with special handling. 1520 case Builtin::BIsqrt: 1521 case Builtin::BIsqrtf: 1522 case Builtin::BIsqrtl: { 1523 // Transform a call to sqrt* into a @llvm.sqrt.* intrinsic call, but only 1524 // in finite- or unsafe-math mode (the intrinsic has different semantics 1525 // for handling negative numbers compared to the library function, so 1526 // -fmath-errno=0 is not enough). 1527 if (!FD->hasAttr<ConstAttr>()) 1528 break; 1529 if (!(CGM.getCodeGenOpts().UnsafeFPMath || 1530 CGM.getCodeGenOpts().NoNaNsFPMath)) 1531 break; 1532 Value *Arg0 = EmitScalarExpr(E->getArg(0)); 1533 llvm::Type *ArgType = Arg0->getType(); 1534 Value *F = CGM.getIntrinsic(Intrinsic::sqrt, ArgType); 1535 return RValue::get(Builder.CreateCall(F, Arg0)); 1536 } 1537 1538 case Builtin::BI__builtin_pow: 1539 case Builtin::BI__builtin_powf: 1540 case Builtin::BI__builtin_powl: 1541 case Builtin::BIpow: 1542 case Builtin::BIpowf: 1543 case Builtin::BIpowl: { 1544 // Transform a call to pow* into a @llvm.pow.* intrinsic call. 1545 if (!FD->hasAttr<ConstAttr>()) 1546 break; 1547 Value *Base = EmitScalarExpr(E->getArg(0)); 1548 Value *Exponent = EmitScalarExpr(E->getArg(1)); 1549 llvm::Type *ArgType = Base->getType(); 1550 Value *F = CGM.getIntrinsic(Intrinsic::pow, ArgType); 1551 return RValue::get(Builder.CreateCall(F, {Base, Exponent})); 1552 } 1553 1554 case Builtin::BIfma: 1555 case Builtin::BIfmaf: 1556 case Builtin::BIfmal: 1557 case Builtin::BI__builtin_fma: 1558 case Builtin::BI__builtin_fmaf: 1559 case Builtin::BI__builtin_fmal: { 1560 // Rewrite fma to intrinsic. 1561 Value *FirstArg = EmitScalarExpr(E->getArg(0)); 1562 llvm::Type *ArgType = FirstArg->getType(); 1563 Value *F = CGM.getIntrinsic(Intrinsic::fma, ArgType); 1564 return RValue::get( 1565 Builder.CreateCall(F, {FirstArg, EmitScalarExpr(E->getArg(1)), 1566 EmitScalarExpr(E->getArg(2))})); 1567 } 1568 1569 case Builtin::BI__builtin_signbit: 1570 case Builtin::BI__builtin_signbitf: 1571 case Builtin::BI__builtin_signbitl: { 1572 return RValue::get( 1573 Builder.CreateZExt(EmitSignBit(*this, EmitScalarExpr(E->getArg(0))), 1574 ConvertType(E->getType()))); 1575 } 1576 case Builtin::BI__builtin_annotation: { 1577 llvm::Value *AnnVal = EmitScalarExpr(E->getArg(0)); 1578 llvm::Value *F = CGM.getIntrinsic(llvm::Intrinsic::annotation, 1579 AnnVal->getType()); 1580 1581 // Get the annotation string, go through casts. Sema requires this to be a 1582 // non-wide string literal, potentially casted, so the cast<> is safe. 1583 const Expr *AnnotationStrExpr = E->getArg(1)->IgnoreParenCasts(); 1584 StringRef Str = cast<StringLiteral>(AnnotationStrExpr)->getString(); 1585 return RValue::get(EmitAnnotationCall(F, AnnVal, Str, E->getExprLoc())); 1586 } 1587 case Builtin::BI__builtin_addcb: 1588 case Builtin::BI__builtin_addcs: 1589 case Builtin::BI__builtin_addc: 1590 case Builtin::BI__builtin_addcl: 1591 case Builtin::BI__builtin_addcll: 1592 case Builtin::BI__builtin_subcb: 1593 case Builtin::BI__builtin_subcs: 1594 case Builtin::BI__builtin_subc: 1595 case Builtin::BI__builtin_subcl: 1596 case Builtin::BI__builtin_subcll: { 1597 1598 // We translate all of these builtins from expressions of the form: 1599 // int x = ..., y = ..., carryin = ..., carryout, result; 1600 // result = __builtin_addc(x, y, carryin, &carryout); 1601 // 1602 // to LLVM IR of the form: 1603 // 1604 // %tmp1 = call {i32, i1} @llvm.uadd.with.overflow.i32(i32 %x, i32 %y) 1605 // %tmpsum1 = extractvalue {i32, i1} %tmp1, 0 1606 // %carry1 = extractvalue {i32, i1} %tmp1, 1 1607 // %tmp2 = call {i32, i1} @llvm.uadd.with.overflow.i32(i32 %tmpsum1, 1608 // i32 %carryin) 1609 // %result = extractvalue {i32, i1} %tmp2, 0 1610 // %carry2 = extractvalue {i32, i1} %tmp2, 1 1611 // %tmp3 = or i1 %carry1, %carry2 1612 // %tmp4 = zext i1 %tmp3 to i32 1613 // store i32 %tmp4, i32* %carryout 1614 1615 // Scalarize our inputs. 1616 llvm::Value *X = EmitScalarExpr(E->getArg(0)); 1617 llvm::Value *Y = EmitScalarExpr(E->getArg(1)); 1618 llvm::Value *Carryin = EmitScalarExpr(E->getArg(2)); 1619 Address CarryOutPtr = EmitPointerWithAlignment(E->getArg(3)); 1620 1621 // Decide if we are lowering to a uadd.with.overflow or usub.with.overflow. 1622 llvm::Intrinsic::ID IntrinsicId; 1623 switch (BuiltinID) { 1624 default: llvm_unreachable("Unknown multiprecision builtin id."); 1625 case Builtin::BI__builtin_addcb: 1626 case Builtin::BI__builtin_addcs: 1627 case Builtin::BI__builtin_addc: 1628 case Builtin::BI__builtin_addcl: 1629 case Builtin::BI__builtin_addcll: 1630 IntrinsicId = llvm::Intrinsic::uadd_with_overflow; 1631 break; 1632 case Builtin::BI__builtin_subcb: 1633 case Builtin::BI__builtin_subcs: 1634 case Builtin::BI__builtin_subc: 1635 case Builtin::BI__builtin_subcl: 1636 case Builtin::BI__builtin_subcll: 1637 IntrinsicId = llvm::Intrinsic::usub_with_overflow; 1638 break; 1639 } 1640 1641 // Construct our resulting LLVM IR expression. 1642 llvm::Value *Carry1; 1643 llvm::Value *Sum1 = EmitOverflowIntrinsic(*this, IntrinsicId, 1644 X, Y, Carry1); 1645 llvm::Value *Carry2; 1646 llvm::Value *Sum2 = EmitOverflowIntrinsic(*this, IntrinsicId, 1647 Sum1, Carryin, Carry2); 1648 llvm::Value *CarryOut = Builder.CreateZExt(Builder.CreateOr(Carry1, Carry2), 1649 X->getType()); 1650 Builder.CreateStore(CarryOut, CarryOutPtr); 1651 return RValue::get(Sum2); 1652 } 1653 1654 case Builtin::BI__builtin_add_overflow: 1655 case Builtin::BI__builtin_sub_overflow: 1656 case Builtin::BI__builtin_mul_overflow: { 1657 const clang::Expr *LeftArg = E->getArg(0); 1658 const clang::Expr *RightArg = E->getArg(1); 1659 const clang::Expr *ResultArg = E->getArg(2); 1660 1661 clang::QualType ResultQTy = 1662 ResultArg->getType()->castAs<PointerType>()->getPointeeType(); 1663 1664 WidthAndSignedness LeftInfo = 1665 getIntegerWidthAndSignedness(CGM.getContext(), LeftArg->getType()); 1666 WidthAndSignedness RightInfo = 1667 getIntegerWidthAndSignedness(CGM.getContext(), RightArg->getType()); 1668 WidthAndSignedness ResultInfo = 1669 getIntegerWidthAndSignedness(CGM.getContext(), ResultQTy); 1670 WidthAndSignedness EncompassingInfo = 1671 EncompassingIntegerType({LeftInfo, RightInfo, ResultInfo}); 1672 1673 llvm::Type *EncompassingLLVMTy = 1674 llvm::IntegerType::get(CGM.getLLVMContext(), EncompassingInfo.Width); 1675 1676 llvm::Type *ResultLLVMTy = CGM.getTypes().ConvertType(ResultQTy); 1677 1678 llvm::Intrinsic::ID IntrinsicId; 1679 switch (BuiltinID) { 1680 default: 1681 llvm_unreachable("Unknown overflow builtin id."); 1682 case Builtin::BI__builtin_add_overflow: 1683 IntrinsicId = EncompassingInfo.Signed 1684 ? llvm::Intrinsic::sadd_with_overflow 1685 : llvm::Intrinsic::uadd_with_overflow; 1686 break; 1687 case Builtin::BI__builtin_sub_overflow: 1688 IntrinsicId = EncompassingInfo.Signed 1689 ? llvm::Intrinsic::ssub_with_overflow 1690 : llvm::Intrinsic::usub_with_overflow; 1691 break; 1692 case Builtin::BI__builtin_mul_overflow: 1693 IntrinsicId = EncompassingInfo.Signed 1694 ? llvm::Intrinsic::smul_with_overflow 1695 : llvm::Intrinsic::umul_with_overflow; 1696 break; 1697 } 1698 1699 llvm::Value *Left = EmitScalarExpr(LeftArg); 1700 llvm::Value *Right = EmitScalarExpr(RightArg); 1701 Address ResultPtr = EmitPointerWithAlignment(ResultArg); 1702 1703 // Extend each operand to the encompassing type. 1704 Left = Builder.CreateIntCast(Left, EncompassingLLVMTy, LeftInfo.Signed); 1705 Right = Builder.CreateIntCast(Right, EncompassingLLVMTy, RightInfo.Signed); 1706 1707 // Perform the operation on the extended values. 1708 llvm::Value *Overflow, *Result; 1709 Result = EmitOverflowIntrinsic(*this, IntrinsicId, Left, Right, Overflow); 1710 1711 if (EncompassingInfo.Width > ResultInfo.Width) { 1712 // The encompassing type is wider than the result type, so we need to 1713 // truncate it. 1714 llvm::Value *ResultTrunc = Builder.CreateTrunc(Result, ResultLLVMTy); 1715 1716 // To see if the truncation caused an overflow, we will extend 1717 // the result and then compare it to the original result. 1718 llvm::Value *ResultTruncExt = Builder.CreateIntCast( 1719 ResultTrunc, EncompassingLLVMTy, ResultInfo.Signed); 1720 llvm::Value *TruncationOverflow = 1721 Builder.CreateICmpNE(Result, ResultTruncExt); 1722 1723 Overflow = Builder.CreateOr(Overflow, TruncationOverflow); 1724 Result = ResultTrunc; 1725 } 1726 1727 // Finally, store the result using the pointer. 1728 bool isVolatile = 1729 ResultArg->getType()->getPointeeType().isVolatileQualified(); 1730 Builder.CreateStore(EmitToMemory(Result, ResultQTy), ResultPtr, isVolatile); 1731 1732 return RValue::get(Overflow); 1733 } 1734 1735 case Builtin::BI__builtin_uadd_overflow: 1736 case Builtin::BI__builtin_uaddl_overflow: 1737 case Builtin::BI__builtin_uaddll_overflow: 1738 case Builtin::BI__builtin_usub_overflow: 1739 case Builtin::BI__builtin_usubl_overflow: 1740 case Builtin::BI__builtin_usubll_overflow: 1741 case Builtin::BI__builtin_umul_overflow: 1742 case Builtin::BI__builtin_umull_overflow: 1743 case Builtin::BI__builtin_umulll_overflow: 1744 case Builtin::BI__builtin_sadd_overflow: 1745 case Builtin::BI__builtin_saddl_overflow: 1746 case Builtin::BI__builtin_saddll_overflow: 1747 case Builtin::BI__builtin_ssub_overflow: 1748 case Builtin::BI__builtin_ssubl_overflow: 1749 case Builtin::BI__builtin_ssubll_overflow: 1750 case Builtin::BI__builtin_smul_overflow: 1751 case Builtin::BI__builtin_smull_overflow: 1752 case Builtin::BI__builtin_smulll_overflow: { 1753 1754 // We translate all of these builtins directly to the relevant llvm IR node. 1755 1756 // Scalarize our inputs. 1757 llvm::Value *X = EmitScalarExpr(E->getArg(0)); 1758 llvm::Value *Y = EmitScalarExpr(E->getArg(1)); 1759 Address SumOutPtr = EmitPointerWithAlignment(E->getArg(2)); 1760 1761 // Decide which of the overflow intrinsics we are lowering to: 1762 llvm::Intrinsic::ID IntrinsicId; 1763 switch (BuiltinID) { 1764 default: llvm_unreachable("Unknown overflow builtin id."); 1765 case Builtin::BI__builtin_uadd_overflow: 1766 case Builtin::BI__builtin_uaddl_overflow: 1767 case Builtin::BI__builtin_uaddll_overflow: 1768 IntrinsicId = llvm::Intrinsic::uadd_with_overflow; 1769 break; 1770 case Builtin::BI__builtin_usub_overflow: 1771 case Builtin::BI__builtin_usubl_overflow: 1772 case Builtin::BI__builtin_usubll_overflow: 1773 IntrinsicId = llvm::Intrinsic::usub_with_overflow; 1774 break; 1775 case Builtin::BI__builtin_umul_overflow: 1776 case Builtin::BI__builtin_umull_overflow: 1777 case Builtin::BI__builtin_umulll_overflow: 1778 IntrinsicId = llvm::Intrinsic::umul_with_overflow; 1779 break; 1780 case Builtin::BI__builtin_sadd_overflow: 1781 case Builtin::BI__builtin_saddl_overflow: 1782 case Builtin::BI__builtin_saddll_overflow: 1783 IntrinsicId = llvm::Intrinsic::sadd_with_overflow; 1784 break; 1785 case Builtin::BI__builtin_ssub_overflow: 1786 case Builtin::BI__builtin_ssubl_overflow: 1787 case Builtin::BI__builtin_ssubll_overflow: 1788 IntrinsicId = llvm::Intrinsic::ssub_with_overflow; 1789 break; 1790 case Builtin::BI__builtin_smul_overflow: 1791 case Builtin::BI__builtin_smull_overflow: 1792 case Builtin::BI__builtin_smulll_overflow: 1793 IntrinsicId = llvm::Intrinsic::smul_with_overflow; 1794 break; 1795 } 1796 1797 1798 llvm::Value *Carry; 1799 llvm::Value *Sum = EmitOverflowIntrinsic(*this, IntrinsicId, X, Y, Carry); 1800 Builder.CreateStore(Sum, SumOutPtr); 1801 1802 return RValue::get(Carry); 1803 } 1804 case Builtin::BI__builtin_addressof: 1805 return RValue::get(EmitLValue(E->getArg(0)).getPointer()); 1806 case Builtin::BI__builtin_operator_new: 1807 return EmitBuiltinNewDeleteCall(FD->getType()->castAs<FunctionProtoType>(), 1808 E->getArg(0), false); 1809 case Builtin::BI__builtin_operator_delete: 1810 return EmitBuiltinNewDeleteCall(FD->getType()->castAs<FunctionProtoType>(), 1811 E->getArg(0), true); 1812 case Builtin::BI__noop: 1813 // __noop always evaluates to an integer literal zero. 1814 return RValue::get(ConstantInt::get(IntTy, 0)); 1815 case Builtin::BI__builtin_call_with_static_chain: { 1816 const CallExpr *Call = cast<CallExpr>(E->getArg(0)); 1817 const Expr *Chain = E->getArg(1); 1818 return EmitCall(Call->getCallee()->getType(), 1819 EmitScalarExpr(Call->getCallee()), Call, ReturnValue, 1820 Call->getCalleeDecl(), EmitScalarExpr(Chain)); 1821 } 1822 case Builtin::BI_InterlockedExchange: 1823 case Builtin::BI_InterlockedExchangePointer: 1824 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Xchg, E); 1825 case Builtin::BI_InterlockedCompareExchangePointer: { 1826 llvm::Type *RTy; 1827 llvm::IntegerType *IntType = 1828 IntegerType::get(getLLVMContext(), 1829 getContext().getTypeSize(E->getType())); 1830 llvm::Type *IntPtrType = IntType->getPointerTo(); 1831 1832 llvm::Value *Destination = 1833 Builder.CreateBitCast(EmitScalarExpr(E->getArg(0)), IntPtrType); 1834 1835 llvm::Value *Exchange = EmitScalarExpr(E->getArg(1)); 1836 RTy = Exchange->getType(); 1837 Exchange = Builder.CreatePtrToInt(Exchange, IntType); 1838 1839 llvm::Value *Comparand = 1840 Builder.CreatePtrToInt(EmitScalarExpr(E->getArg(2)), IntType); 1841 1842 auto Result = Builder.CreateAtomicCmpXchg(Destination, Comparand, Exchange, 1843 SequentiallyConsistent, 1844 SequentiallyConsistent); 1845 Result->setVolatile(true); 1846 1847 return RValue::get(Builder.CreateIntToPtr(Builder.CreateExtractValue(Result, 1848 0), 1849 RTy)); 1850 } 1851 case Builtin::BI_InterlockedCompareExchange: { 1852 AtomicCmpXchgInst *CXI = Builder.CreateAtomicCmpXchg( 1853 EmitScalarExpr(E->getArg(0)), 1854 EmitScalarExpr(E->getArg(2)), 1855 EmitScalarExpr(E->getArg(1)), 1856 SequentiallyConsistent, 1857 SequentiallyConsistent); 1858 CXI->setVolatile(true); 1859 return RValue::get(Builder.CreateExtractValue(CXI, 0)); 1860 } 1861 case Builtin::BI_InterlockedIncrement: { 1862 AtomicRMWInst *RMWI = Builder.CreateAtomicRMW( 1863 AtomicRMWInst::Add, 1864 EmitScalarExpr(E->getArg(0)), 1865 ConstantInt::get(Int32Ty, 1), 1866 llvm::SequentiallyConsistent); 1867 RMWI->setVolatile(true); 1868 return RValue::get(Builder.CreateAdd(RMWI, ConstantInt::get(Int32Ty, 1))); 1869 } 1870 case Builtin::BI_InterlockedDecrement: { 1871 AtomicRMWInst *RMWI = Builder.CreateAtomicRMW( 1872 AtomicRMWInst::Sub, 1873 EmitScalarExpr(E->getArg(0)), 1874 ConstantInt::get(Int32Ty, 1), 1875 llvm::SequentiallyConsistent); 1876 RMWI->setVolatile(true); 1877 return RValue::get(Builder.CreateSub(RMWI, ConstantInt::get(Int32Ty, 1))); 1878 } 1879 case Builtin::BI_InterlockedExchangeAdd: { 1880 AtomicRMWInst *RMWI = Builder.CreateAtomicRMW( 1881 AtomicRMWInst::Add, 1882 EmitScalarExpr(E->getArg(0)), 1883 EmitScalarExpr(E->getArg(1)), 1884 llvm::SequentiallyConsistent); 1885 RMWI->setVolatile(true); 1886 return RValue::get(RMWI); 1887 } 1888 case Builtin::BI__readfsdword: { 1889 Value *IntToPtr = 1890 Builder.CreateIntToPtr(EmitScalarExpr(E->getArg(0)), 1891 llvm::PointerType::get(CGM.Int32Ty, 257)); 1892 LoadInst *Load = 1893 Builder.CreateAlignedLoad(IntToPtr, /*Align=*/4, /*isVolatile=*/true); 1894 return RValue::get(Load); 1895 } 1896 1897 case Builtin::BI__exception_code: 1898 case Builtin::BI_exception_code: 1899 return RValue::get(EmitSEHExceptionCode()); 1900 case Builtin::BI__exception_info: 1901 case Builtin::BI_exception_info: 1902 return RValue::get(EmitSEHExceptionInfo()); 1903 case Builtin::BI__abnormal_termination: 1904 case Builtin::BI_abnormal_termination: 1905 return RValue::get(EmitSEHAbnormalTermination()); 1906 case Builtin::BI_setjmpex: { 1907 if (getTarget().getTriple().isOSMSVCRT()) { 1908 llvm::Type *ArgTypes[] = {Int8PtrTy, Int8PtrTy}; 1909 llvm::AttributeSet ReturnsTwiceAttr = 1910 AttributeSet::get(getLLVMContext(), llvm::AttributeSet::FunctionIndex, 1911 llvm::Attribute::ReturnsTwice); 1912 llvm::Constant *SetJmpEx = CGM.CreateRuntimeFunction( 1913 llvm::FunctionType::get(IntTy, ArgTypes, /*isVarArg=*/false), 1914 "_setjmpex", ReturnsTwiceAttr); 1915 llvm::Value *Buf = Builder.CreateBitOrPointerCast( 1916 EmitScalarExpr(E->getArg(0)), Int8PtrTy); 1917 llvm::Value *FrameAddr = 1918 Builder.CreateCall(CGM.getIntrinsic(Intrinsic::frameaddress), 1919 ConstantInt::get(Int32Ty, 0)); 1920 llvm::Value *Args[] = {Buf, FrameAddr}; 1921 llvm::CallSite CS = EmitRuntimeCallOrInvoke(SetJmpEx, Args); 1922 CS.setAttributes(ReturnsTwiceAttr); 1923 return RValue::get(CS.getInstruction()); 1924 } 1925 break; 1926 } 1927 case Builtin::BI_setjmp: { 1928 if (getTarget().getTriple().isOSMSVCRT()) { 1929 llvm::AttributeSet ReturnsTwiceAttr = 1930 AttributeSet::get(getLLVMContext(), llvm::AttributeSet::FunctionIndex, 1931 llvm::Attribute::ReturnsTwice); 1932 llvm::Value *Buf = Builder.CreateBitOrPointerCast( 1933 EmitScalarExpr(E->getArg(0)), Int8PtrTy); 1934 llvm::CallSite CS; 1935 if (getTarget().getTriple().getArch() == llvm::Triple::x86) { 1936 llvm::Type *ArgTypes[] = {Int8PtrTy, IntTy}; 1937 llvm::Constant *SetJmp3 = CGM.CreateRuntimeFunction( 1938 llvm::FunctionType::get(IntTy, ArgTypes, /*isVarArg=*/true), 1939 "_setjmp3", ReturnsTwiceAttr); 1940 llvm::Value *Count = ConstantInt::get(IntTy, 0); 1941 llvm::Value *Args[] = {Buf, Count}; 1942 CS = EmitRuntimeCallOrInvoke(SetJmp3, Args); 1943 } else { 1944 llvm::Type *ArgTypes[] = {Int8PtrTy, Int8PtrTy}; 1945 llvm::Constant *SetJmp = CGM.CreateRuntimeFunction( 1946 llvm::FunctionType::get(IntTy, ArgTypes, /*isVarArg=*/false), 1947 "_setjmp", ReturnsTwiceAttr); 1948 llvm::Value *FrameAddr = 1949 Builder.CreateCall(CGM.getIntrinsic(Intrinsic::frameaddress), 1950 ConstantInt::get(Int32Ty, 0)); 1951 llvm::Value *Args[] = {Buf, FrameAddr}; 1952 CS = EmitRuntimeCallOrInvoke(SetJmp, Args); 1953 } 1954 CS.setAttributes(ReturnsTwiceAttr); 1955 return RValue::get(CS.getInstruction()); 1956 } 1957 break; 1958 } 1959 1960 case Builtin::BI__GetExceptionInfo: { 1961 if (llvm::GlobalVariable *GV = 1962 CGM.getCXXABI().getThrowInfo(FD->getParamDecl(0)->getType())) 1963 return RValue::get(llvm::ConstantExpr::getBitCast(GV, CGM.Int8PtrTy)); 1964 break; 1965 } 1966 1967 // OpenCL v2.0 s6.13.16.2, Built-in pipe read and write functions 1968 case Builtin::BIread_pipe: 1969 case Builtin::BIwrite_pipe: { 1970 Value *Arg0 = EmitScalarExpr(E->getArg(0)), 1971 *Arg1 = EmitScalarExpr(E->getArg(1)); 1972 1973 // Type of the generic packet parameter. 1974 unsigned GenericAS = 1975 getContext().getTargetAddressSpace(LangAS::opencl_generic); 1976 llvm::Type *I8PTy = llvm::PointerType::get( 1977 llvm::Type::getInt8Ty(getLLVMContext()), GenericAS); 1978 1979 // Testing which overloaded version we should generate the call for. 1980 if (2U == E->getNumArgs()) { 1981 const char *Name = (BuiltinID == Builtin::BIread_pipe) ? "__read_pipe_2" 1982 : "__write_pipe_2"; 1983 // Creating a generic function type to be able to call with any builtin or 1984 // user defined type. 1985 llvm::Type *ArgTys[] = {Arg0->getType(), I8PTy}; 1986 llvm::FunctionType *FTy = llvm::FunctionType::get( 1987 Int32Ty, llvm::ArrayRef<llvm::Type *>(ArgTys), false); 1988 Value *BCast = Builder.CreatePointerCast(Arg1, I8PTy); 1989 return RValue::get(Builder.CreateCall( 1990 CGM.CreateRuntimeFunction(FTy, Name), {Arg0, BCast})); 1991 } else { 1992 assert(4 == E->getNumArgs() && 1993 "Illegal number of parameters to pipe function"); 1994 const char *Name = (BuiltinID == Builtin::BIread_pipe) ? "__read_pipe_4" 1995 : "__write_pipe_4"; 1996 1997 llvm::Type *ArgTys[] = {Arg0->getType(), Arg1->getType(), Int32Ty, I8PTy}; 1998 Value *Arg2 = EmitScalarExpr(E->getArg(2)), 1999 *Arg3 = EmitScalarExpr(E->getArg(3)); 2000 llvm::FunctionType *FTy = llvm::FunctionType::get( 2001 Int32Ty, llvm::ArrayRef<llvm::Type *>(ArgTys), false); 2002 Value *BCast = Builder.CreatePointerCast(Arg3, I8PTy); 2003 // We know the third argument is an integer type, but we may need to cast 2004 // it to i32. 2005 if (Arg2->getType() != Int32Ty) 2006 Arg2 = Builder.CreateZExtOrTrunc(Arg2, Int32Ty); 2007 return RValue::get(Builder.CreateCall( 2008 CGM.CreateRuntimeFunction(FTy, Name), {Arg0, Arg1, Arg2, BCast})); 2009 } 2010 } 2011 // OpenCL v2.0 s6.13.16 ,s9.17.3.5 - Built-in pipe reserve read and write 2012 // functions 2013 case Builtin::BIreserve_read_pipe: 2014 case Builtin::BIreserve_write_pipe: 2015 case Builtin::BIwork_group_reserve_read_pipe: 2016 case Builtin::BIwork_group_reserve_write_pipe: 2017 case Builtin::BIsub_group_reserve_read_pipe: 2018 case Builtin::BIsub_group_reserve_write_pipe: { 2019 // Composing the mangled name for the function. 2020 const char *Name; 2021 if (BuiltinID == Builtin::BIreserve_read_pipe) 2022 Name = "__reserve_read_pipe"; 2023 else if (BuiltinID == Builtin::BIreserve_write_pipe) 2024 Name = "__reserve_write_pipe"; 2025 else if (BuiltinID == Builtin::BIwork_group_reserve_read_pipe) 2026 Name = "__work_group_reserve_read_pipe"; 2027 else if (BuiltinID == Builtin::BIwork_group_reserve_write_pipe) 2028 Name = "__work_group_reserve_write_pipe"; 2029 else if (BuiltinID == Builtin::BIsub_group_reserve_read_pipe) 2030 Name = "__sub_group_reserve_read_pipe"; 2031 else 2032 Name = "__sub_group_reserve_write_pipe"; 2033 2034 Value *Arg0 = EmitScalarExpr(E->getArg(0)), 2035 *Arg1 = EmitScalarExpr(E->getArg(1)); 2036 llvm::Type *ReservedIDTy = ConvertType(getContext().OCLReserveIDTy); 2037 2038 // Building the generic function prototype. 2039 llvm::Type *ArgTys[] = {Arg0->getType(), Int32Ty}; 2040 llvm::FunctionType *FTy = llvm::FunctionType::get( 2041 ReservedIDTy, llvm::ArrayRef<llvm::Type *>(ArgTys), false); 2042 // We know the second argument is an integer type, but we may need to cast 2043 // it to i32. 2044 if (Arg1->getType() != Int32Ty) 2045 Arg1 = Builder.CreateZExtOrTrunc(Arg1, Int32Ty); 2046 return RValue::get( 2047 Builder.CreateCall(CGM.CreateRuntimeFunction(FTy, Name), {Arg0, Arg1})); 2048 } 2049 // OpenCL v2.0 s6.13.16 ,s9.17.3.5 - Built-in pipe commit read and write 2050 // functions 2051 case Builtin::BIcommit_read_pipe: 2052 case Builtin::BIcommit_write_pipe: 2053 case Builtin::BIwork_group_commit_read_pipe: 2054 case Builtin::BIwork_group_commit_write_pipe: 2055 case Builtin::BIsub_group_commit_read_pipe: 2056 case Builtin::BIsub_group_commit_write_pipe: { 2057 const char *Name; 2058 if (BuiltinID == Builtin::BIcommit_read_pipe) 2059 Name = "__commit_read_pipe"; 2060 else if (BuiltinID == Builtin::BIcommit_write_pipe) 2061 Name = "__commit_write_pipe"; 2062 else if (BuiltinID == Builtin::BIwork_group_commit_read_pipe) 2063 Name = "__work_group_commit_read_pipe"; 2064 else if (BuiltinID == Builtin::BIwork_group_commit_write_pipe) 2065 Name = "__work_group_commit_write_pipe"; 2066 else if (BuiltinID == Builtin::BIsub_group_commit_read_pipe) 2067 Name = "__sub_group_commit_read_pipe"; 2068 else 2069 Name = "__sub_group_commit_write_pipe"; 2070 2071 Value *Arg0 = EmitScalarExpr(E->getArg(0)), 2072 *Arg1 = EmitScalarExpr(E->getArg(1)); 2073 2074 // Building the generic function prototype. 2075 llvm::Type *ArgTys[] = {Arg0->getType(), Arg1->getType()}; 2076 llvm::FunctionType *FTy = 2077 llvm::FunctionType::get(llvm::Type::getVoidTy(getLLVMContext()), 2078 llvm::ArrayRef<llvm::Type *>(ArgTys), false); 2079 2080 return RValue::get( 2081 Builder.CreateCall(CGM.CreateRuntimeFunction(FTy, Name), {Arg0, Arg1})); 2082 } 2083 // OpenCL v2.0 s6.13.16.4 Built-in pipe query functions 2084 case Builtin::BIget_pipe_num_packets: 2085 case Builtin::BIget_pipe_max_packets: { 2086 const char *Name; 2087 if (BuiltinID == Builtin::BIget_pipe_num_packets) 2088 Name = "__get_pipe_num_packets"; 2089 else 2090 Name = "__get_pipe_max_packets"; 2091 2092 // Building the generic function prototype. 2093 Value *Arg0 = EmitScalarExpr(E->getArg(0)); 2094 llvm::Type *ArgTys[] = {Arg0->getType()}; 2095 llvm::FunctionType *FTy = llvm::FunctionType::get( 2096 Int32Ty, llvm::ArrayRef<llvm::Type *>(ArgTys), false); 2097 2098 return RValue::get( 2099 Builder.CreateCall(CGM.CreateRuntimeFunction(FTy, Name), {Arg0})); 2100 } 2101 2102 case Builtin::BIprintf: 2103 if (getLangOpts().CUDA && getLangOpts().CUDAIsDevice) 2104 return EmitCUDADevicePrintfCallExpr(E, ReturnValue); 2105 } 2106 2107 // If this is an alias for a lib function (e.g. __builtin_sin), emit 2108 // the call using the normal call path, but using the unmangled 2109 // version of the function name. 2110 if (getContext().BuiltinInfo.isLibFunction(BuiltinID)) 2111 return emitLibraryCall(*this, FD, E, 2112 CGM.getBuiltinLibFunction(FD, BuiltinID)); 2113 2114 // If this is a predefined lib function (e.g. malloc), emit the call 2115 // using exactly the normal call path. 2116 if (getContext().BuiltinInfo.isPredefinedLibFunction(BuiltinID)) 2117 return emitLibraryCall(*this, FD, E, EmitScalarExpr(E->getCallee())); 2118 2119 // Check that a call to a target specific builtin has the correct target 2120 // features. 2121 // This is down here to avoid non-target specific builtins, however, if 2122 // generic builtins start to require generic target features then we 2123 // can move this up to the beginning of the function. 2124 checkTargetFeatures(E, FD); 2125 2126 // See if we have a target specific intrinsic. 2127 const char *Name = getContext().BuiltinInfo.getName(BuiltinID); 2128 Intrinsic::ID IntrinsicID = Intrinsic::not_intrinsic; 2129 if (const char *Prefix = 2130 llvm::Triple::getArchTypePrefix(getTarget().getTriple().getArch())) { 2131 IntrinsicID = Intrinsic::getIntrinsicForGCCBuiltin(Prefix, Name); 2132 // NOTE we dont need to perform a compatibility flag check here since the 2133 // intrinsics are declared in Builtins*.def via LANGBUILTIN which filter the 2134 // MS builtins via ALL_MS_LANGUAGES and are filtered earlier. 2135 if (IntrinsicID == Intrinsic::not_intrinsic) 2136 IntrinsicID = Intrinsic::getIntrinsicForMSBuiltin(Prefix, Name); 2137 } 2138 2139 if (IntrinsicID != Intrinsic::not_intrinsic) { 2140 SmallVector<Value*, 16> Args; 2141 2142 // Find out if any arguments are required to be integer constant 2143 // expressions. 2144 unsigned ICEArguments = 0; 2145 ASTContext::GetBuiltinTypeError Error; 2146 getContext().GetBuiltinType(BuiltinID, Error, &ICEArguments); 2147 assert(Error == ASTContext::GE_None && "Should not codegen an error"); 2148 2149 Function *F = CGM.getIntrinsic(IntrinsicID); 2150 llvm::FunctionType *FTy = F->getFunctionType(); 2151 2152 for (unsigned i = 0, e = E->getNumArgs(); i != e; ++i) { 2153 Value *ArgValue; 2154 // If this is a normal argument, just emit it as a scalar. 2155 if ((ICEArguments & (1 << i)) == 0) { 2156 ArgValue = EmitScalarExpr(E->getArg(i)); 2157 } else { 2158 // If this is required to be a constant, constant fold it so that we 2159 // know that the generated intrinsic gets a ConstantInt. 2160 llvm::APSInt Result; 2161 bool IsConst = E->getArg(i)->isIntegerConstantExpr(Result,getContext()); 2162 assert(IsConst && "Constant arg isn't actually constant?"); 2163 (void)IsConst; 2164 ArgValue = llvm::ConstantInt::get(getLLVMContext(), Result); 2165 } 2166 2167 // If the intrinsic arg type is different from the builtin arg type 2168 // we need to do a bit cast. 2169 llvm::Type *PTy = FTy->getParamType(i); 2170 if (PTy != ArgValue->getType()) { 2171 assert(PTy->canLosslesslyBitCastTo(FTy->getParamType(i)) && 2172 "Must be able to losslessly bit cast to param"); 2173 ArgValue = Builder.CreateBitCast(ArgValue, PTy); 2174 } 2175 2176 Args.push_back(ArgValue); 2177 } 2178 2179 Value *V = Builder.CreateCall(F, Args); 2180 QualType BuiltinRetType = E->getType(); 2181 2182 llvm::Type *RetTy = VoidTy; 2183 if (!BuiltinRetType->isVoidType()) 2184 RetTy = ConvertType(BuiltinRetType); 2185 2186 if (RetTy != V->getType()) { 2187 assert(V->getType()->canLosslesslyBitCastTo(RetTy) && 2188 "Must be able to losslessly bit cast result type"); 2189 V = Builder.CreateBitCast(V, RetTy); 2190 } 2191 2192 return RValue::get(V); 2193 } 2194 2195 // See if we have a target specific builtin that needs to be lowered. 2196 if (Value *V = EmitTargetBuiltinExpr(BuiltinID, E)) 2197 return RValue::get(V); 2198 2199 ErrorUnsupported(E, "builtin function"); 2200 2201 // Unknown builtin, for now just dump it out and return undef. 2202 return GetUndefRValue(E->getType()); 2203 } 2204 2205 static Value *EmitTargetArchBuiltinExpr(CodeGenFunction *CGF, 2206 unsigned BuiltinID, const CallExpr *E, 2207 llvm::Triple::ArchType Arch) { 2208 switch (Arch) { 2209 case llvm::Triple::arm: 2210 case llvm::Triple::armeb: 2211 case llvm::Triple::thumb: 2212 case llvm::Triple::thumbeb: 2213 return CGF->EmitARMBuiltinExpr(BuiltinID, E); 2214 case llvm::Triple::aarch64: 2215 case llvm::Triple::aarch64_be: 2216 return CGF->EmitAArch64BuiltinExpr(BuiltinID, E); 2217 case llvm::Triple::x86: 2218 case llvm::Triple::x86_64: 2219 return CGF->EmitX86BuiltinExpr(BuiltinID, E); 2220 case llvm::Triple::ppc: 2221 case llvm::Triple::ppc64: 2222 case llvm::Triple::ppc64le: 2223 return CGF->EmitPPCBuiltinExpr(BuiltinID, E); 2224 case llvm::Triple::r600: 2225 case llvm::Triple::amdgcn: 2226 return CGF->EmitAMDGPUBuiltinExpr(BuiltinID, E); 2227 case llvm::Triple::systemz: 2228 return CGF->EmitSystemZBuiltinExpr(BuiltinID, E); 2229 case llvm::Triple::nvptx: 2230 case llvm::Triple::nvptx64: 2231 return CGF->EmitNVPTXBuiltinExpr(BuiltinID, E); 2232 case llvm::Triple::wasm32: 2233 case llvm::Triple::wasm64: 2234 return CGF->EmitWebAssemblyBuiltinExpr(BuiltinID, E); 2235 default: 2236 return nullptr; 2237 } 2238 } 2239 2240 Value *CodeGenFunction::EmitTargetBuiltinExpr(unsigned BuiltinID, 2241 const CallExpr *E) { 2242 if (getContext().BuiltinInfo.isAuxBuiltinID(BuiltinID)) { 2243 assert(getContext().getAuxTargetInfo() && "Missing aux target info"); 2244 return EmitTargetArchBuiltinExpr( 2245 this, getContext().BuiltinInfo.getAuxBuiltinID(BuiltinID), E, 2246 getContext().getAuxTargetInfo()->getTriple().getArch()); 2247 } 2248 2249 return EmitTargetArchBuiltinExpr(this, BuiltinID, E, 2250 getTarget().getTriple().getArch()); 2251 } 2252 2253 static llvm::VectorType *GetNeonType(CodeGenFunction *CGF, 2254 NeonTypeFlags TypeFlags, 2255 bool V1Ty=false) { 2256 int IsQuad = TypeFlags.isQuad(); 2257 switch (TypeFlags.getEltType()) { 2258 case NeonTypeFlags::Int8: 2259 case NeonTypeFlags::Poly8: 2260 return llvm::VectorType::get(CGF->Int8Ty, V1Ty ? 1 : (8 << IsQuad)); 2261 case NeonTypeFlags::Int16: 2262 case NeonTypeFlags::Poly16: 2263 case NeonTypeFlags::Float16: 2264 return llvm::VectorType::get(CGF->Int16Ty, V1Ty ? 1 : (4 << IsQuad)); 2265 case NeonTypeFlags::Int32: 2266 return llvm::VectorType::get(CGF->Int32Ty, V1Ty ? 1 : (2 << IsQuad)); 2267 case NeonTypeFlags::Int64: 2268 case NeonTypeFlags::Poly64: 2269 return llvm::VectorType::get(CGF->Int64Ty, V1Ty ? 1 : (1 << IsQuad)); 2270 case NeonTypeFlags::Poly128: 2271 // FIXME: i128 and f128 doesn't get fully support in Clang and llvm. 2272 // There is a lot of i128 and f128 API missing. 2273 // so we use v16i8 to represent poly128 and get pattern matched. 2274 return llvm::VectorType::get(CGF->Int8Ty, 16); 2275 case NeonTypeFlags::Float32: 2276 return llvm::VectorType::get(CGF->FloatTy, V1Ty ? 1 : (2 << IsQuad)); 2277 case NeonTypeFlags::Float64: 2278 return llvm::VectorType::get(CGF->DoubleTy, V1Ty ? 1 : (1 << IsQuad)); 2279 } 2280 llvm_unreachable("Unknown vector element type!"); 2281 } 2282 2283 static llvm::VectorType *GetFloatNeonType(CodeGenFunction *CGF, 2284 NeonTypeFlags IntTypeFlags) { 2285 int IsQuad = IntTypeFlags.isQuad(); 2286 switch (IntTypeFlags.getEltType()) { 2287 case NeonTypeFlags::Int32: 2288 return llvm::VectorType::get(CGF->FloatTy, (2 << IsQuad)); 2289 case NeonTypeFlags::Int64: 2290 return llvm::VectorType::get(CGF->DoubleTy, (1 << IsQuad)); 2291 default: 2292 llvm_unreachable("Type can't be converted to floating-point!"); 2293 } 2294 } 2295 2296 Value *CodeGenFunction::EmitNeonSplat(Value *V, Constant *C) { 2297 unsigned nElts = cast<llvm::VectorType>(V->getType())->getNumElements(); 2298 Value* SV = llvm::ConstantVector::getSplat(nElts, C); 2299 return Builder.CreateShuffleVector(V, V, SV, "lane"); 2300 } 2301 2302 Value *CodeGenFunction::EmitNeonCall(Function *F, SmallVectorImpl<Value*> &Ops, 2303 const char *name, 2304 unsigned shift, bool rightshift) { 2305 unsigned j = 0; 2306 for (Function::const_arg_iterator ai = F->arg_begin(), ae = F->arg_end(); 2307 ai != ae; ++ai, ++j) 2308 if (shift > 0 && shift == j) 2309 Ops[j] = EmitNeonShiftVector(Ops[j], ai->getType(), rightshift); 2310 else 2311 Ops[j] = Builder.CreateBitCast(Ops[j], ai->getType(), name); 2312 2313 return Builder.CreateCall(F, Ops, name); 2314 } 2315 2316 Value *CodeGenFunction::EmitNeonShiftVector(Value *V, llvm::Type *Ty, 2317 bool neg) { 2318 int SV = cast<ConstantInt>(V)->getSExtValue(); 2319 return ConstantInt::get(Ty, neg ? -SV : SV); 2320 } 2321 2322 // \brief Right-shift a vector by a constant. 2323 Value *CodeGenFunction::EmitNeonRShiftImm(Value *Vec, Value *Shift, 2324 llvm::Type *Ty, bool usgn, 2325 const char *name) { 2326 llvm::VectorType *VTy = cast<llvm::VectorType>(Ty); 2327 2328 int ShiftAmt = cast<ConstantInt>(Shift)->getSExtValue(); 2329 int EltSize = VTy->getScalarSizeInBits(); 2330 2331 Vec = Builder.CreateBitCast(Vec, Ty); 2332 2333 // lshr/ashr are undefined when the shift amount is equal to the vector 2334 // element size. 2335 if (ShiftAmt == EltSize) { 2336 if (usgn) { 2337 // Right-shifting an unsigned value by its size yields 0. 2338 return llvm::ConstantAggregateZero::get(VTy); 2339 } else { 2340 // Right-shifting a signed value by its size is equivalent 2341 // to a shift of size-1. 2342 --ShiftAmt; 2343 Shift = ConstantInt::get(VTy->getElementType(), ShiftAmt); 2344 } 2345 } 2346 2347 Shift = EmitNeonShiftVector(Shift, Ty, false); 2348 if (usgn) 2349 return Builder.CreateLShr(Vec, Shift, name); 2350 else 2351 return Builder.CreateAShr(Vec, Shift, name); 2352 } 2353 2354 enum { 2355 AddRetType = (1 << 0), 2356 Add1ArgType = (1 << 1), 2357 Add2ArgTypes = (1 << 2), 2358 2359 VectorizeRetType = (1 << 3), 2360 VectorizeArgTypes = (1 << 4), 2361 2362 InventFloatType = (1 << 5), 2363 UnsignedAlts = (1 << 6), 2364 2365 Use64BitVectors = (1 << 7), 2366 Use128BitVectors = (1 << 8), 2367 2368 Vectorize1ArgType = Add1ArgType | VectorizeArgTypes, 2369 VectorRet = AddRetType | VectorizeRetType, 2370 VectorRetGetArgs01 = 2371 AddRetType | Add2ArgTypes | VectorizeRetType | VectorizeArgTypes, 2372 FpCmpzModifiers = 2373 AddRetType | VectorizeRetType | Add1ArgType | InventFloatType 2374 }; 2375 2376 namespace { 2377 struct NeonIntrinsicInfo { 2378 const char *NameHint; 2379 unsigned BuiltinID; 2380 unsigned LLVMIntrinsic; 2381 unsigned AltLLVMIntrinsic; 2382 unsigned TypeModifier; 2383 2384 bool operator<(unsigned RHSBuiltinID) const { 2385 return BuiltinID < RHSBuiltinID; 2386 } 2387 bool operator<(const NeonIntrinsicInfo &TE) const { 2388 return BuiltinID < TE.BuiltinID; 2389 } 2390 }; 2391 } // end anonymous namespace 2392 2393 #define NEONMAP0(NameBase) \ 2394 { #NameBase, NEON::BI__builtin_neon_ ## NameBase, 0, 0, 0 } 2395 2396 #define NEONMAP1(NameBase, LLVMIntrinsic, TypeModifier) \ 2397 { #NameBase, NEON:: BI__builtin_neon_ ## NameBase, \ 2398 Intrinsic::LLVMIntrinsic, 0, TypeModifier } 2399 2400 #define NEONMAP2(NameBase, LLVMIntrinsic, AltLLVMIntrinsic, TypeModifier) \ 2401 { #NameBase, NEON:: BI__builtin_neon_ ## NameBase, \ 2402 Intrinsic::LLVMIntrinsic, Intrinsic::AltLLVMIntrinsic, \ 2403 TypeModifier } 2404 2405 static const NeonIntrinsicInfo ARMSIMDIntrinsicMap [] = { 2406 NEONMAP2(vabd_v, arm_neon_vabdu, arm_neon_vabds, Add1ArgType | UnsignedAlts), 2407 NEONMAP2(vabdq_v, arm_neon_vabdu, arm_neon_vabds, Add1ArgType | UnsignedAlts), 2408 NEONMAP1(vabs_v, arm_neon_vabs, 0), 2409 NEONMAP1(vabsq_v, arm_neon_vabs, 0), 2410 NEONMAP0(vaddhn_v), 2411 NEONMAP1(vaesdq_v, arm_neon_aesd, 0), 2412 NEONMAP1(vaeseq_v, arm_neon_aese, 0), 2413 NEONMAP1(vaesimcq_v, arm_neon_aesimc, 0), 2414 NEONMAP1(vaesmcq_v, arm_neon_aesmc, 0), 2415 NEONMAP1(vbsl_v, arm_neon_vbsl, AddRetType), 2416 NEONMAP1(vbslq_v, arm_neon_vbsl, AddRetType), 2417 NEONMAP1(vcage_v, arm_neon_vacge, 0), 2418 NEONMAP1(vcageq_v, arm_neon_vacge, 0), 2419 NEONMAP1(vcagt_v, arm_neon_vacgt, 0), 2420 NEONMAP1(vcagtq_v, arm_neon_vacgt, 0), 2421 NEONMAP1(vcale_v, arm_neon_vacge, 0), 2422 NEONMAP1(vcaleq_v, arm_neon_vacge, 0), 2423 NEONMAP1(vcalt_v, arm_neon_vacgt, 0), 2424 NEONMAP1(vcaltq_v, arm_neon_vacgt, 0), 2425 NEONMAP1(vcls_v, arm_neon_vcls, Add1ArgType), 2426 NEONMAP1(vclsq_v, arm_neon_vcls, Add1ArgType), 2427 NEONMAP1(vclz_v, ctlz, Add1ArgType), 2428 NEONMAP1(vclzq_v, ctlz, Add1ArgType), 2429 NEONMAP1(vcnt_v, ctpop, Add1ArgType), 2430 NEONMAP1(vcntq_v, ctpop, Add1ArgType), 2431 NEONMAP1(vcvt_f16_f32, arm_neon_vcvtfp2hf, 0), 2432 NEONMAP1(vcvt_f32_f16, arm_neon_vcvthf2fp, 0), 2433 NEONMAP0(vcvt_f32_v), 2434 NEONMAP2(vcvt_n_f32_v, arm_neon_vcvtfxu2fp, arm_neon_vcvtfxs2fp, 0), 2435 NEONMAP1(vcvt_n_s32_v, arm_neon_vcvtfp2fxs, 0), 2436 NEONMAP1(vcvt_n_s64_v, arm_neon_vcvtfp2fxs, 0), 2437 NEONMAP1(vcvt_n_u32_v, arm_neon_vcvtfp2fxu, 0), 2438 NEONMAP1(vcvt_n_u64_v, arm_neon_vcvtfp2fxu, 0), 2439 NEONMAP0(vcvt_s32_v), 2440 NEONMAP0(vcvt_s64_v), 2441 NEONMAP0(vcvt_u32_v), 2442 NEONMAP0(vcvt_u64_v), 2443 NEONMAP1(vcvta_s32_v, arm_neon_vcvtas, 0), 2444 NEONMAP1(vcvta_s64_v, arm_neon_vcvtas, 0), 2445 NEONMAP1(vcvta_u32_v, arm_neon_vcvtau, 0), 2446 NEONMAP1(vcvta_u64_v, arm_neon_vcvtau, 0), 2447 NEONMAP1(vcvtaq_s32_v, arm_neon_vcvtas, 0), 2448 NEONMAP1(vcvtaq_s64_v, arm_neon_vcvtas, 0), 2449 NEONMAP1(vcvtaq_u32_v, arm_neon_vcvtau, 0), 2450 NEONMAP1(vcvtaq_u64_v, arm_neon_vcvtau, 0), 2451 NEONMAP1(vcvtm_s32_v, arm_neon_vcvtms, 0), 2452 NEONMAP1(vcvtm_s64_v, arm_neon_vcvtms, 0), 2453 NEONMAP1(vcvtm_u32_v, arm_neon_vcvtmu, 0), 2454 NEONMAP1(vcvtm_u64_v, arm_neon_vcvtmu, 0), 2455 NEONMAP1(vcvtmq_s32_v, arm_neon_vcvtms, 0), 2456 NEONMAP1(vcvtmq_s64_v, arm_neon_vcvtms, 0), 2457 NEONMAP1(vcvtmq_u32_v, arm_neon_vcvtmu, 0), 2458 NEONMAP1(vcvtmq_u64_v, arm_neon_vcvtmu, 0), 2459 NEONMAP1(vcvtn_s32_v, arm_neon_vcvtns, 0), 2460 NEONMAP1(vcvtn_s64_v, arm_neon_vcvtns, 0), 2461 NEONMAP1(vcvtn_u32_v, arm_neon_vcvtnu, 0), 2462 NEONMAP1(vcvtn_u64_v, arm_neon_vcvtnu, 0), 2463 NEONMAP1(vcvtnq_s32_v, arm_neon_vcvtns, 0), 2464 NEONMAP1(vcvtnq_s64_v, arm_neon_vcvtns, 0), 2465 NEONMAP1(vcvtnq_u32_v, arm_neon_vcvtnu, 0), 2466 NEONMAP1(vcvtnq_u64_v, arm_neon_vcvtnu, 0), 2467 NEONMAP1(vcvtp_s32_v, arm_neon_vcvtps, 0), 2468 NEONMAP1(vcvtp_s64_v, arm_neon_vcvtps, 0), 2469 NEONMAP1(vcvtp_u32_v, arm_neon_vcvtpu, 0), 2470 NEONMAP1(vcvtp_u64_v, arm_neon_vcvtpu, 0), 2471 NEONMAP1(vcvtpq_s32_v, arm_neon_vcvtps, 0), 2472 NEONMAP1(vcvtpq_s64_v, arm_neon_vcvtps, 0), 2473 NEONMAP1(vcvtpq_u32_v, arm_neon_vcvtpu, 0), 2474 NEONMAP1(vcvtpq_u64_v, arm_neon_vcvtpu, 0), 2475 NEONMAP0(vcvtq_f32_v), 2476 NEONMAP2(vcvtq_n_f32_v, arm_neon_vcvtfxu2fp, arm_neon_vcvtfxs2fp, 0), 2477 NEONMAP1(vcvtq_n_s32_v, arm_neon_vcvtfp2fxs, 0), 2478 NEONMAP1(vcvtq_n_s64_v, arm_neon_vcvtfp2fxs, 0), 2479 NEONMAP1(vcvtq_n_u32_v, arm_neon_vcvtfp2fxu, 0), 2480 NEONMAP1(vcvtq_n_u64_v, arm_neon_vcvtfp2fxu, 0), 2481 NEONMAP0(vcvtq_s32_v), 2482 NEONMAP0(vcvtq_s64_v), 2483 NEONMAP0(vcvtq_u32_v), 2484 NEONMAP0(vcvtq_u64_v), 2485 NEONMAP0(vext_v), 2486 NEONMAP0(vextq_v), 2487 NEONMAP0(vfma_v), 2488 NEONMAP0(vfmaq_v), 2489 NEONMAP2(vhadd_v, arm_neon_vhaddu, arm_neon_vhadds, Add1ArgType | UnsignedAlts), 2490 NEONMAP2(vhaddq_v, arm_neon_vhaddu, arm_neon_vhadds, Add1ArgType | UnsignedAlts), 2491 NEONMAP2(vhsub_v, arm_neon_vhsubu, arm_neon_vhsubs, Add1ArgType | UnsignedAlts), 2492 NEONMAP2(vhsubq_v, arm_neon_vhsubu, arm_neon_vhsubs, Add1ArgType | UnsignedAlts), 2493 NEONMAP0(vld1_dup_v), 2494 NEONMAP1(vld1_v, arm_neon_vld1, 0), 2495 NEONMAP0(vld1q_dup_v), 2496 NEONMAP1(vld1q_v, arm_neon_vld1, 0), 2497 NEONMAP1(vld2_lane_v, arm_neon_vld2lane, 0), 2498 NEONMAP1(vld2_v, arm_neon_vld2, 0), 2499 NEONMAP1(vld2q_lane_v, arm_neon_vld2lane, 0), 2500 NEONMAP1(vld2q_v, arm_neon_vld2, 0), 2501 NEONMAP1(vld3_lane_v, arm_neon_vld3lane, 0), 2502 NEONMAP1(vld3_v, arm_neon_vld3, 0), 2503 NEONMAP1(vld3q_lane_v, arm_neon_vld3lane, 0), 2504 NEONMAP1(vld3q_v, arm_neon_vld3, 0), 2505 NEONMAP1(vld4_lane_v, arm_neon_vld4lane, 0), 2506 NEONMAP1(vld4_v, arm_neon_vld4, 0), 2507 NEONMAP1(vld4q_lane_v, arm_neon_vld4lane, 0), 2508 NEONMAP1(vld4q_v, arm_neon_vld4, 0), 2509 NEONMAP2(vmax_v, arm_neon_vmaxu, arm_neon_vmaxs, Add1ArgType | UnsignedAlts), 2510 NEONMAP1(vmaxnm_v, arm_neon_vmaxnm, Add1ArgType), 2511 NEONMAP1(vmaxnmq_v, arm_neon_vmaxnm, Add1ArgType), 2512 NEONMAP2(vmaxq_v, arm_neon_vmaxu, arm_neon_vmaxs, Add1ArgType | UnsignedAlts), 2513 NEONMAP2(vmin_v, arm_neon_vminu, arm_neon_vmins, Add1ArgType | UnsignedAlts), 2514 NEONMAP1(vminnm_v, arm_neon_vminnm, Add1ArgType), 2515 NEONMAP1(vminnmq_v, arm_neon_vminnm, Add1ArgType), 2516 NEONMAP2(vminq_v, arm_neon_vminu, arm_neon_vmins, Add1ArgType | UnsignedAlts), 2517 NEONMAP0(vmovl_v), 2518 NEONMAP0(vmovn_v), 2519 NEONMAP1(vmul_v, arm_neon_vmulp, Add1ArgType), 2520 NEONMAP0(vmull_v), 2521 NEONMAP1(vmulq_v, arm_neon_vmulp, Add1ArgType), 2522 NEONMAP2(vpadal_v, arm_neon_vpadalu, arm_neon_vpadals, UnsignedAlts), 2523 NEONMAP2(vpadalq_v, arm_neon_vpadalu, arm_neon_vpadals, UnsignedAlts), 2524 NEONMAP1(vpadd_v, arm_neon_vpadd, Add1ArgType), 2525 NEONMAP2(vpaddl_v, arm_neon_vpaddlu, arm_neon_vpaddls, UnsignedAlts), 2526 NEONMAP2(vpaddlq_v, arm_neon_vpaddlu, arm_neon_vpaddls, UnsignedAlts), 2527 NEONMAP1(vpaddq_v, arm_neon_vpadd, Add1ArgType), 2528 NEONMAP2(vpmax_v, arm_neon_vpmaxu, arm_neon_vpmaxs, Add1ArgType | UnsignedAlts), 2529 NEONMAP2(vpmin_v, arm_neon_vpminu, arm_neon_vpmins, Add1ArgType | UnsignedAlts), 2530 NEONMAP1(vqabs_v, arm_neon_vqabs, Add1ArgType), 2531 NEONMAP1(vqabsq_v, arm_neon_vqabs, Add1ArgType), 2532 NEONMAP2(vqadd_v, arm_neon_vqaddu, arm_neon_vqadds, Add1ArgType | UnsignedAlts), 2533 NEONMAP2(vqaddq_v, arm_neon_vqaddu, arm_neon_vqadds, Add1ArgType | UnsignedAlts), 2534 NEONMAP2(vqdmlal_v, arm_neon_vqdmull, arm_neon_vqadds, 0), 2535 NEONMAP2(vqdmlsl_v, arm_neon_vqdmull, arm_neon_vqsubs, 0), 2536 NEONMAP1(vqdmulh_v, arm_neon_vqdmulh, Add1ArgType), 2537 NEONMAP1(vqdmulhq_v, arm_neon_vqdmulh, Add1ArgType), 2538 NEONMAP1(vqdmull_v, arm_neon_vqdmull, Add1ArgType), 2539 NEONMAP2(vqmovn_v, arm_neon_vqmovnu, arm_neon_vqmovns, Add1ArgType | UnsignedAlts), 2540 NEONMAP1(vqmovun_v, arm_neon_vqmovnsu, Add1ArgType), 2541 NEONMAP1(vqneg_v, arm_neon_vqneg, Add1ArgType), 2542 NEONMAP1(vqnegq_v, arm_neon_vqneg, Add1ArgType), 2543 NEONMAP1(vqrdmulh_v, arm_neon_vqrdmulh, Add1ArgType), 2544 NEONMAP1(vqrdmulhq_v, arm_neon_vqrdmulh, Add1ArgType), 2545 NEONMAP2(vqrshl_v, arm_neon_vqrshiftu, arm_neon_vqrshifts, Add1ArgType | UnsignedAlts), 2546 NEONMAP2(vqrshlq_v, arm_neon_vqrshiftu, arm_neon_vqrshifts, Add1ArgType | UnsignedAlts), 2547 NEONMAP2(vqshl_n_v, arm_neon_vqshiftu, arm_neon_vqshifts, UnsignedAlts), 2548 NEONMAP2(vqshl_v, arm_neon_vqshiftu, arm_neon_vqshifts, Add1ArgType | UnsignedAlts), 2549 NEONMAP2(vqshlq_n_v, arm_neon_vqshiftu, arm_neon_vqshifts, UnsignedAlts), 2550 NEONMAP2(vqshlq_v, arm_neon_vqshiftu, arm_neon_vqshifts, Add1ArgType | UnsignedAlts), 2551 NEONMAP1(vqshlu_n_v, arm_neon_vqshiftsu, 0), 2552 NEONMAP1(vqshluq_n_v, arm_neon_vqshiftsu, 0), 2553 NEONMAP2(vqsub_v, arm_neon_vqsubu, arm_neon_vqsubs, Add1ArgType | UnsignedAlts), 2554 NEONMAP2(vqsubq_v, arm_neon_vqsubu, arm_neon_vqsubs, Add1ArgType | UnsignedAlts), 2555 NEONMAP1(vraddhn_v, arm_neon_vraddhn, Add1ArgType), 2556 NEONMAP2(vrecpe_v, arm_neon_vrecpe, arm_neon_vrecpe, 0), 2557 NEONMAP2(vrecpeq_v, arm_neon_vrecpe, arm_neon_vrecpe, 0), 2558 NEONMAP1(vrecps_v, arm_neon_vrecps, Add1ArgType), 2559 NEONMAP1(vrecpsq_v, arm_neon_vrecps, Add1ArgType), 2560 NEONMAP2(vrhadd_v, arm_neon_vrhaddu, arm_neon_vrhadds, Add1ArgType | UnsignedAlts), 2561 NEONMAP2(vrhaddq_v, arm_neon_vrhaddu, arm_neon_vrhadds, Add1ArgType | UnsignedAlts), 2562 NEONMAP1(vrnd_v, arm_neon_vrintz, Add1ArgType), 2563 NEONMAP1(vrnda_v, arm_neon_vrinta, Add1ArgType), 2564 NEONMAP1(vrndaq_v, arm_neon_vrinta, Add1ArgType), 2565 NEONMAP1(vrndm_v, arm_neon_vrintm, Add1ArgType), 2566 NEONMAP1(vrndmq_v, arm_neon_vrintm, Add1ArgType), 2567 NEONMAP1(vrndn_v, arm_neon_vrintn, Add1ArgType), 2568 NEONMAP1(vrndnq_v, arm_neon_vrintn, Add1ArgType), 2569 NEONMAP1(vrndp_v, arm_neon_vrintp, Add1ArgType), 2570 NEONMAP1(vrndpq_v, arm_neon_vrintp, Add1ArgType), 2571 NEONMAP1(vrndq_v, arm_neon_vrintz, Add1ArgType), 2572 NEONMAP1(vrndx_v, arm_neon_vrintx, Add1ArgType), 2573 NEONMAP1(vrndxq_v, arm_neon_vrintx, Add1ArgType), 2574 NEONMAP2(vrshl_v, arm_neon_vrshiftu, arm_neon_vrshifts, Add1ArgType | UnsignedAlts), 2575 NEONMAP2(vrshlq_v, arm_neon_vrshiftu, arm_neon_vrshifts, Add1ArgType | UnsignedAlts), 2576 NEONMAP2(vrshr_n_v, arm_neon_vrshiftu, arm_neon_vrshifts, UnsignedAlts), 2577 NEONMAP2(vrshrq_n_v, arm_neon_vrshiftu, arm_neon_vrshifts, UnsignedAlts), 2578 NEONMAP2(vrsqrte_v, arm_neon_vrsqrte, arm_neon_vrsqrte, 0), 2579 NEONMAP2(vrsqrteq_v, arm_neon_vrsqrte, arm_neon_vrsqrte, 0), 2580 NEONMAP1(vrsqrts_v, arm_neon_vrsqrts, Add1ArgType), 2581 NEONMAP1(vrsqrtsq_v, arm_neon_vrsqrts, Add1ArgType), 2582 NEONMAP1(vrsubhn_v, arm_neon_vrsubhn, Add1ArgType), 2583 NEONMAP1(vsha1su0q_v, arm_neon_sha1su0, 0), 2584 NEONMAP1(vsha1su1q_v, arm_neon_sha1su1, 0), 2585 NEONMAP1(vsha256h2q_v, arm_neon_sha256h2, 0), 2586 NEONMAP1(vsha256hq_v, arm_neon_sha256h, 0), 2587 NEONMAP1(vsha256su0q_v, arm_neon_sha256su0, 0), 2588 NEONMAP1(vsha256su1q_v, arm_neon_sha256su1, 0), 2589 NEONMAP0(vshl_n_v), 2590 NEONMAP2(vshl_v, arm_neon_vshiftu, arm_neon_vshifts, Add1ArgType | UnsignedAlts), 2591 NEONMAP0(vshll_n_v), 2592 NEONMAP0(vshlq_n_v), 2593 NEONMAP2(vshlq_v, arm_neon_vshiftu, arm_neon_vshifts, Add1ArgType | UnsignedAlts), 2594 NEONMAP0(vshr_n_v), 2595 NEONMAP0(vshrn_n_v), 2596 NEONMAP0(vshrq_n_v), 2597 NEONMAP1(vst1_v, arm_neon_vst1, 0), 2598 NEONMAP1(vst1q_v, arm_neon_vst1, 0), 2599 NEONMAP1(vst2_lane_v, arm_neon_vst2lane, 0), 2600 NEONMAP1(vst2_v, arm_neon_vst2, 0), 2601 NEONMAP1(vst2q_lane_v, arm_neon_vst2lane, 0), 2602 NEONMAP1(vst2q_v, arm_neon_vst2, 0), 2603 NEONMAP1(vst3_lane_v, arm_neon_vst3lane, 0), 2604 NEONMAP1(vst3_v, arm_neon_vst3, 0), 2605 NEONMAP1(vst3q_lane_v, arm_neon_vst3lane, 0), 2606 NEONMAP1(vst3q_v, arm_neon_vst3, 0), 2607 NEONMAP1(vst4_lane_v, arm_neon_vst4lane, 0), 2608 NEONMAP1(vst4_v, arm_neon_vst4, 0), 2609 NEONMAP1(vst4q_lane_v, arm_neon_vst4lane, 0), 2610 NEONMAP1(vst4q_v, arm_neon_vst4, 0), 2611 NEONMAP0(vsubhn_v), 2612 NEONMAP0(vtrn_v), 2613 NEONMAP0(vtrnq_v), 2614 NEONMAP0(vtst_v), 2615 NEONMAP0(vtstq_v), 2616 NEONMAP0(vuzp_v), 2617 NEONMAP0(vuzpq_v), 2618 NEONMAP0(vzip_v), 2619 NEONMAP0(vzipq_v) 2620 }; 2621 2622 static const NeonIntrinsicInfo AArch64SIMDIntrinsicMap[] = { 2623 NEONMAP1(vabs_v, aarch64_neon_abs, 0), 2624 NEONMAP1(vabsq_v, aarch64_neon_abs, 0), 2625 NEONMAP0(vaddhn_v), 2626 NEONMAP1(vaesdq_v, aarch64_crypto_aesd, 0), 2627 NEONMAP1(vaeseq_v, aarch64_crypto_aese, 0), 2628 NEONMAP1(vaesimcq_v, aarch64_crypto_aesimc, 0), 2629 NEONMAP1(vaesmcq_v, aarch64_crypto_aesmc, 0), 2630 NEONMAP1(vcage_v, aarch64_neon_facge, 0), 2631 NEONMAP1(vcageq_v, aarch64_neon_facge, 0), 2632 NEONMAP1(vcagt_v, aarch64_neon_facgt, 0), 2633 NEONMAP1(vcagtq_v, aarch64_neon_facgt, 0), 2634 NEONMAP1(vcale_v, aarch64_neon_facge, 0), 2635 NEONMAP1(vcaleq_v, aarch64_neon_facge, 0), 2636 NEONMAP1(vcalt_v, aarch64_neon_facgt, 0), 2637 NEONMAP1(vcaltq_v, aarch64_neon_facgt, 0), 2638 NEONMAP1(vcls_v, aarch64_neon_cls, Add1ArgType), 2639 NEONMAP1(vclsq_v, aarch64_neon_cls, Add1ArgType), 2640 NEONMAP1(vclz_v, ctlz, Add1ArgType), 2641 NEONMAP1(vclzq_v, ctlz, Add1ArgType), 2642 NEONMAP1(vcnt_v, ctpop, Add1ArgType), 2643 NEONMAP1(vcntq_v, ctpop, Add1ArgType), 2644 NEONMAP1(vcvt_f16_f32, aarch64_neon_vcvtfp2hf, 0), 2645 NEONMAP1(vcvt_f32_f16, aarch64_neon_vcvthf2fp, 0), 2646 NEONMAP0(vcvt_f32_v), 2647 NEONMAP2(vcvt_n_f32_v, aarch64_neon_vcvtfxu2fp, aarch64_neon_vcvtfxs2fp, 0), 2648 NEONMAP2(vcvt_n_f64_v, aarch64_neon_vcvtfxu2fp, aarch64_neon_vcvtfxs2fp, 0), 2649 NEONMAP1(vcvt_n_s32_v, aarch64_neon_vcvtfp2fxs, 0), 2650 NEONMAP1(vcvt_n_s64_v, aarch64_neon_vcvtfp2fxs, 0), 2651 NEONMAP1(vcvt_n_u32_v, aarch64_neon_vcvtfp2fxu, 0), 2652 NEONMAP1(vcvt_n_u64_v, aarch64_neon_vcvtfp2fxu, 0), 2653 NEONMAP0(vcvtq_f32_v), 2654 NEONMAP2(vcvtq_n_f32_v, aarch64_neon_vcvtfxu2fp, aarch64_neon_vcvtfxs2fp, 0), 2655 NEONMAP2(vcvtq_n_f64_v, aarch64_neon_vcvtfxu2fp, aarch64_neon_vcvtfxs2fp, 0), 2656 NEONMAP1(vcvtq_n_s32_v, aarch64_neon_vcvtfp2fxs, 0), 2657 NEONMAP1(vcvtq_n_s64_v, aarch64_neon_vcvtfp2fxs, 0), 2658 NEONMAP1(vcvtq_n_u32_v, aarch64_neon_vcvtfp2fxu, 0), 2659 NEONMAP1(vcvtq_n_u64_v, aarch64_neon_vcvtfp2fxu, 0), 2660 NEONMAP1(vcvtx_f32_v, aarch64_neon_fcvtxn, AddRetType | Add1ArgType), 2661 NEONMAP0(vext_v), 2662 NEONMAP0(vextq_v), 2663 NEONMAP0(vfma_v), 2664 NEONMAP0(vfmaq_v), 2665 NEONMAP2(vhadd_v, aarch64_neon_uhadd, aarch64_neon_shadd, Add1ArgType | UnsignedAlts), 2666 NEONMAP2(vhaddq_v, aarch64_neon_uhadd, aarch64_neon_shadd, Add1ArgType | UnsignedAlts), 2667 NEONMAP2(vhsub_v, aarch64_neon_uhsub, aarch64_neon_shsub, Add1ArgType | UnsignedAlts), 2668 NEONMAP2(vhsubq_v, aarch64_neon_uhsub, aarch64_neon_shsub, Add1ArgType | UnsignedAlts), 2669 NEONMAP0(vmovl_v), 2670 NEONMAP0(vmovn_v), 2671 NEONMAP1(vmul_v, aarch64_neon_pmul, Add1ArgType), 2672 NEONMAP1(vmulq_v, aarch64_neon_pmul, Add1ArgType), 2673 NEONMAP1(vpadd_v, aarch64_neon_addp, Add1ArgType), 2674 NEONMAP2(vpaddl_v, aarch64_neon_uaddlp, aarch64_neon_saddlp, UnsignedAlts), 2675 NEONMAP2(vpaddlq_v, aarch64_neon_uaddlp, aarch64_neon_saddlp, UnsignedAlts), 2676 NEONMAP1(vpaddq_v, aarch64_neon_addp, Add1ArgType), 2677 NEONMAP1(vqabs_v, aarch64_neon_sqabs, Add1ArgType), 2678 NEONMAP1(vqabsq_v, aarch64_neon_sqabs, Add1ArgType), 2679 NEONMAP2(vqadd_v, aarch64_neon_uqadd, aarch64_neon_sqadd, Add1ArgType | UnsignedAlts), 2680 NEONMAP2(vqaddq_v, aarch64_neon_uqadd, aarch64_neon_sqadd, Add1ArgType | UnsignedAlts), 2681 NEONMAP2(vqdmlal_v, aarch64_neon_sqdmull, aarch64_neon_sqadd, 0), 2682 NEONMAP2(vqdmlsl_v, aarch64_neon_sqdmull, aarch64_neon_sqsub, 0), 2683 NEONMAP1(vqdmulh_v, aarch64_neon_sqdmulh, Add1ArgType), 2684 NEONMAP1(vqdmulhq_v, aarch64_neon_sqdmulh, Add1ArgType), 2685 NEONMAP1(vqdmull_v, aarch64_neon_sqdmull, Add1ArgType), 2686 NEONMAP2(vqmovn_v, aarch64_neon_uqxtn, aarch64_neon_sqxtn, Add1ArgType | UnsignedAlts), 2687 NEONMAP1(vqmovun_v, aarch64_neon_sqxtun, Add1ArgType), 2688 NEONMAP1(vqneg_v, aarch64_neon_sqneg, Add1ArgType), 2689 NEONMAP1(vqnegq_v, aarch64_neon_sqneg, Add1ArgType), 2690 NEONMAP1(vqrdmulh_v, aarch64_neon_sqrdmulh, Add1ArgType), 2691 NEONMAP1(vqrdmulhq_v, aarch64_neon_sqrdmulh, Add1ArgType), 2692 NEONMAP2(vqrshl_v, aarch64_neon_uqrshl, aarch64_neon_sqrshl, Add1ArgType | UnsignedAlts), 2693 NEONMAP2(vqrshlq_v, aarch64_neon_uqrshl, aarch64_neon_sqrshl, Add1ArgType | UnsignedAlts), 2694 NEONMAP2(vqshl_n_v, aarch64_neon_uqshl, aarch64_neon_sqshl, UnsignedAlts), 2695 NEONMAP2(vqshl_v, aarch64_neon_uqshl, aarch64_neon_sqshl, Add1ArgType | UnsignedAlts), 2696 NEONMAP2(vqshlq_n_v, aarch64_neon_uqshl, aarch64_neon_sqshl,UnsignedAlts), 2697 NEONMAP2(vqshlq_v, aarch64_neon_uqshl, aarch64_neon_sqshl, Add1ArgType | UnsignedAlts), 2698 NEONMAP1(vqshlu_n_v, aarch64_neon_sqshlu, 0), 2699 NEONMAP1(vqshluq_n_v, aarch64_neon_sqshlu, 0), 2700 NEONMAP2(vqsub_v, aarch64_neon_uqsub, aarch64_neon_sqsub, Add1ArgType | UnsignedAlts), 2701 NEONMAP2(vqsubq_v, aarch64_neon_uqsub, aarch64_neon_sqsub, Add1ArgType | UnsignedAlts), 2702 NEONMAP1(vraddhn_v, aarch64_neon_raddhn, Add1ArgType), 2703 NEONMAP2(vrecpe_v, aarch64_neon_frecpe, aarch64_neon_urecpe, 0), 2704 NEONMAP2(vrecpeq_v, aarch64_neon_frecpe, aarch64_neon_urecpe, 0), 2705 NEONMAP1(vrecps_v, aarch64_neon_frecps, Add1ArgType), 2706 NEONMAP1(vrecpsq_v, aarch64_neon_frecps, Add1ArgType), 2707 NEONMAP2(vrhadd_v, aarch64_neon_urhadd, aarch64_neon_srhadd, Add1ArgType | UnsignedAlts), 2708 NEONMAP2(vrhaddq_v, aarch64_neon_urhadd, aarch64_neon_srhadd, Add1ArgType | UnsignedAlts), 2709 NEONMAP2(vrshl_v, aarch64_neon_urshl, aarch64_neon_srshl, Add1ArgType | UnsignedAlts), 2710 NEONMAP2(vrshlq_v, aarch64_neon_urshl, aarch64_neon_srshl, Add1ArgType | UnsignedAlts), 2711 NEONMAP2(vrshr_n_v, aarch64_neon_urshl, aarch64_neon_srshl, UnsignedAlts), 2712 NEONMAP2(vrshrq_n_v, aarch64_neon_urshl, aarch64_neon_srshl, UnsignedAlts), 2713 NEONMAP2(vrsqrte_v, aarch64_neon_frsqrte, aarch64_neon_ursqrte, 0), 2714 NEONMAP2(vrsqrteq_v, aarch64_neon_frsqrte, aarch64_neon_ursqrte, 0), 2715 NEONMAP1(vrsqrts_v, aarch64_neon_frsqrts, Add1ArgType), 2716 NEONMAP1(vrsqrtsq_v, aarch64_neon_frsqrts, Add1ArgType), 2717 NEONMAP1(vrsubhn_v, aarch64_neon_rsubhn, Add1ArgType), 2718 NEONMAP1(vsha1su0q_v, aarch64_crypto_sha1su0, 0), 2719 NEONMAP1(vsha1su1q_v, aarch64_crypto_sha1su1, 0), 2720 NEONMAP1(vsha256h2q_v, aarch64_crypto_sha256h2, 0), 2721 NEONMAP1(vsha256hq_v, aarch64_crypto_sha256h, 0), 2722 NEONMAP1(vsha256su0q_v, aarch64_crypto_sha256su0, 0), 2723 NEONMAP1(vsha256su1q_v, aarch64_crypto_sha256su1, 0), 2724 NEONMAP0(vshl_n_v), 2725 NEONMAP2(vshl_v, aarch64_neon_ushl, aarch64_neon_sshl, Add1ArgType | UnsignedAlts), 2726 NEONMAP0(vshll_n_v), 2727 NEONMAP0(vshlq_n_v), 2728 NEONMAP2(vshlq_v, aarch64_neon_ushl, aarch64_neon_sshl, Add1ArgType | UnsignedAlts), 2729 NEONMAP0(vshr_n_v), 2730 NEONMAP0(vshrn_n_v), 2731 NEONMAP0(vshrq_n_v), 2732 NEONMAP0(vsubhn_v), 2733 NEONMAP0(vtst_v), 2734 NEONMAP0(vtstq_v), 2735 }; 2736 2737 static const NeonIntrinsicInfo AArch64SISDIntrinsicMap[] = { 2738 NEONMAP1(vabdd_f64, aarch64_sisd_fabd, Add1ArgType), 2739 NEONMAP1(vabds_f32, aarch64_sisd_fabd, Add1ArgType), 2740 NEONMAP1(vabsd_s64, aarch64_neon_abs, Add1ArgType), 2741 NEONMAP1(vaddlv_s32, aarch64_neon_saddlv, AddRetType | Add1ArgType), 2742 NEONMAP1(vaddlv_u32, aarch64_neon_uaddlv, AddRetType | Add1ArgType), 2743 NEONMAP1(vaddlvq_s32, aarch64_neon_saddlv, AddRetType | Add1ArgType), 2744 NEONMAP1(vaddlvq_u32, aarch64_neon_uaddlv, AddRetType | Add1ArgType), 2745 NEONMAP1(vaddv_f32, aarch64_neon_faddv, AddRetType | Add1ArgType), 2746 NEONMAP1(vaddv_s32, aarch64_neon_saddv, AddRetType | Add1ArgType), 2747 NEONMAP1(vaddv_u32, aarch64_neon_uaddv, AddRetType | Add1ArgType), 2748 NEONMAP1(vaddvq_f32, aarch64_neon_faddv, AddRetType | Add1ArgType), 2749 NEONMAP1(vaddvq_f64, aarch64_neon_faddv, AddRetType | Add1ArgType), 2750 NEONMAP1(vaddvq_s32, aarch64_neon_saddv, AddRetType | Add1ArgType), 2751 NEONMAP1(vaddvq_s64, aarch64_neon_saddv, AddRetType | Add1ArgType), 2752 NEONMAP1(vaddvq_u32, aarch64_neon_uaddv, AddRetType | Add1ArgType), 2753 NEONMAP1(vaddvq_u64, aarch64_neon_uaddv, AddRetType | Add1ArgType), 2754 NEONMAP1(vcaged_f64, aarch64_neon_facge, AddRetType | Add1ArgType), 2755 NEONMAP1(vcages_f32, aarch64_neon_facge, AddRetType | Add1ArgType), 2756 NEONMAP1(vcagtd_f64, aarch64_neon_facgt, AddRetType | Add1ArgType), 2757 NEONMAP1(vcagts_f32, aarch64_neon_facgt, AddRetType | Add1ArgType), 2758 NEONMAP1(vcaled_f64, aarch64_neon_facge, AddRetType | Add1ArgType), 2759 NEONMAP1(vcales_f32, aarch64_neon_facge, AddRetType | Add1ArgType), 2760 NEONMAP1(vcaltd_f64, aarch64_neon_facgt, AddRetType | Add1ArgType), 2761 NEONMAP1(vcalts_f32, aarch64_neon_facgt, AddRetType | Add1ArgType), 2762 NEONMAP1(vcvtad_s64_f64, aarch64_neon_fcvtas, AddRetType | Add1ArgType), 2763 NEONMAP1(vcvtad_u64_f64, aarch64_neon_fcvtau, AddRetType | Add1ArgType), 2764 NEONMAP1(vcvtas_s32_f32, aarch64_neon_fcvtas, AddRetType | Add1ArgType), 2765 NEONMAP1(vcvtas_u32_f32, aarch64_neon_fcvtau, AddRetType | Add1ArgType), 2766 NEONMAP1(vcvtd_n_f64_s64, aarch64_neon_vcvtfxs2fp, AddRetType | Add1ArgType), 2767 NEONMAP1(vcvtd_n_f64_u64, aarch64_neon_vcvtfxu2fp, AddRetType | Add1ArgType), 2768 NEONMAP1(vcvtd_n_s64_f64, aarch64_neon_vcvtfp2fxs, AddRetType | Add1ArgType), 2769 NEONMAP1(vcvtd_n_u64_f64, aarch64_neon_vcvtfp2fxu, AddRetType | Add1ArgType), 2770 NEONMAP1(vcvtmd_s64_f64, aarch64_neon_fcvtms, AddRetType | Add1ArgType), 2771 NEONMAP1(vcvtmd_u64_f64, aarch64_neon_fcvtmu, AddRetType | Add1ArgType), 2772 NEONMAP1(vcvtms_s32_f32, aarch64_neon_fcvtms, AddRetType | Add1ArgType), 2773 NEONMAP1(vcvtms_u32_f32, aarch64_neon_fcvtmu, AddRetType | Add1ArgType), 2774 NEONMAP1(vcvtnd_s64_f64, aarch64_neon_fcvtns, AddRetType | Add1ArgType), 2775 NEONMAP1(vcvtnd_u64_f64, aarch64_neon_fcvtnu, AddRetType | Add1ArgType), 2776 NEONMAP1(vcvtns_s32_f32, aarch64_neon_fcvtns, AddRetType | Add1ArgType), 2777 NEONMAP1(vcvtns_u32_f32, aarch64_neon_fcvtnu, AddRetType | Add1ArgType), 2778 NEONMAP1(vcvtpd_s64_f64, aarch64_neon_fcvtps, AddRetType | Add1ArgType), 2779 NEONMAP1(vcvtpd_u64_f64, aarch64_neon_fcvtpu, AddRetType | Add1ArgType), 2780 NEONMAP1(vcvtps_s32_f32, aarch64_neon_fcvtps, AddRetType | Add1ArgType), 2781 NEONMAP1(vcvtps_u32_f32, aarch64_neon_fcvtpu, AddRetType | Add1ArgType), 2782 NEONMAP1(vcvts_n_f32_s32, aarch64_neon_vcvtfxs2fp, AddRetType | Add1ArgType), 2783 NEONMAP1(vcvts_n_f32_u32, aarch64_neon_vcvtfxu2fp, AddRetType | Add1ArgType), 2784 NEONMAP1(vcvts_n_s32_f32, aarch64_neon_vcvtfp2fxs, AddRetType | Add1ArgType), 2785 NEONMAP1(vcvts_n_u32_f32, aarch64_neon_vcvtfp2fxu, AddRetType | Add1ArgType), 2786 NEONMAP1(vcvtxd_f32_f64, aarch64_sisd_fcvtxn, 0), 2787 NEONMAP1(vmaxnmv_f32, aarch64_neon_fmaxnmv, AddRetType | Add1ArgType), 2788 NEONMAP1(vmaxnmvq_f32, aarch64_neon_fmaxnmv, AddRetType | Add1ArgType), 2789 NEONMAP1(vmaxnmvq_f64, aarch64_neon_fmaxnmv, AddRetType | Add1ArgType), 2790 NEONMAP1(vmaxv_f32, aarch64_neon_fmaxv, AddRetType | Add1ArgType), 2791 NEONMAP1(vmaxv_s32, aarch64_neon_smaxv, AddRetType | Add1ArgType), 2792 NEONMAP1(vmaxv_u32, aarch64_neon_umaxv, AddRetType | Add1ArgType), 2793 NEONMAP1(vmaxvq_f32, aarch64_neon_fmaxv, AddRetType | Add1ArgType), 2794 NEONMAP1(vmaxvq_f64, aarch64_neon_fmaxv, AddRetType | Add1ArgType), 2795 NEONMAP1(vmaxvq_s32, aarch64_neon_smaxv, AddRetType | Add1ArgType), 2796 NEONMAP1(vmaxvq_u32, aarch64_neon_umaxv, AddRetType | Add1ArgType), 2797 NEONMAP1(vminnmv_f32, aarch64_neon_fminnmv, AddRetType | Add1ArgType), 2798 NEONMAP1(vminnmvq_f32, aarch64_neon_fminnmv, AddRetType | Add1ArgType), 2799 NEONMAP1(vminnmvq_f64, aarch64_neon_fminnmv, AddRetType | Add1ArgType), 2800 NEONMAP1(vminv_f32, aarch64_neon_fminv, AddRetType | Add1ArgType), 2801 NEONMAP1(vminv_s32, aarch64_neon_sminv, AddRetType | Add1ArgType), 2802 NEONMAP1(vminv_u32, aarch64_neon_uminv, AddRetType | Add1ArgType), 2803 NEONMAP1(vminvq_f32, aarch64_neon_fminv, AddRetType | Add1ArgType), 2804 NEONMAP1(vminvq_f64, aarch64_neon_fminv, AddRetType | Add1ArgType), 2805 NEONMAP1(vminvq_s32, aarch64_neon_sminv, AddRetType | Add1ArgType), 2806 NEONMAP1(vminvq_u32, aarch64_neon_uminv, AddRetType | Add1ArgType), 2807 NEONMAP1(vmull_p64, aarch64_neon_pmull64, 0), 2808 NEONMAP1(vmulxd_f64, aarch64_neon_fmulx, Add1ArgType), 2809 NEONMAP1(vmulxs_f32, aarch64_neon_fmulx, Add1ArgType), 2810 NEONMAP1(vpaddd_s64, aarch64_neon_uaddv, AddRetType | Add1ArgType), 2811 NEONMAP1(vpaddd_u64, aarch64_neon_uaddv, AddRetType | Add1ArgType), 2812 NEONMAP1(vpmaxnmqd_f64, aarch64_neon_fmaxnmv, AddRetType | Add1ArgType), 2813 NEONMAP1(vpmaxnms_f32, aarch64_neon_fmaxnmv, AddRetType | Add1ArgType), 2814 NEONMAP1(vpmaxqd_f64, aarch64_neon_fmaxv, AddRetType | Add1ArgType), 2815 NEONMAP1(vpmaxs_f32, aarch64_neon_fmaxv, AddRetType | Add1ArgType), 2816 NEONMAP1(vpminnmqd_f64, aarch64_neon_fminnmv, AddRetType | Add1ArgType), 2817 NEONMAP1(vpminnms_f32, aarch64_neon_fminnmv, AddRetType | Add1ArgType), 2818 NEONMAP1(vpminqd_f64, aarch64_neon_fminv, AddRetType | Add1ArgType), 2819 NEONMAP1(vpmins_f32, aarch64_neon_fminv, AddRetType | Add1ArgType), 2820 NEONMAP1(vqabsb_s8, aarch64_neon_sqabs, Vectorize1ArgType | Use64BitVectors), 2821 NEONMAP1(vqabsd_s64, aarch64_neon_sqabs, Add1ArgType), 2822 NEONMAP1(vqabsh_s16, aarch64_neon_sqabs, Vectorize1ArgType | Use64BitVectors), 2823 NEONMAP1(vqabss_s32, aarch64_neon_sqabs, Add1ArgType), 2824 NEONMAP1(vqaddb_s8, aarch64_neon_sqadd, Vectorize1ArgType | Use64BitVectors), 2825 NEONMAP1(vqaddb_u8, aarch64_neon_uqadd, Vectorize1ArgType | Use64BitVectors), 2826 NEONMAP1(vqaddd_s64, aarch64_neon_sqadd, Add1ArgType), 2827 NEONMAP1(vqaddd_u64, aarch64_neon_uqadd, Add1ArgType), 2828 NEONMAP1(vqaddh_s16, aarch64_neon_sqadd, Vectorize1ArgType | Use64BitVectors), 2829 NEONMAP1(vqaddh_u16, aarch64_neon_uqadd, Vectorize1ArgType | Use64BitVectors), 2830 NEONMAP1(vqadds_s32, aarch64_neon_sqadd, Add1ArgType), 2831 NEONMAP1(vqadds_u32, aarch64_neon_uqadd, Add1ArgType), 2832 NEONMAP1(vqdmulhh_s16, aarch64_neon_sqdmulh, Vectorize1ArgType | Use64BitVectors), 2833 NEONMAP1(vqdmulhs_s32, aarch64_neon_sqdmulh, Add1ArgType), 2834 NEONMAP1(vqdmullh_s16, aarch64_neon_sqdmull, VectorRet | Use128BitVectors), 2835 NEONMAP1(vqdmulls_s32, aarch64_neon_sqdmulls_scalar, 0), 2836 NEONMAP1(vqmovnd_s64, aarch64_neon_scalar_sqxtn, AddRetType | Add1ArgType), 2837 NEONMAP1(vqmovnd_u64, aarch64_neon_scalar_uqxtn, AddRetType | Add1ArgType), 2838 NEONMAP1(vqmovnh_s16, aarch64_neon_sqxtn, VectorRet | Use64BitVectors), 2839 NEONMAP1(vqmovnh_u16, aarch64_neon_uqxtn, VectorRet | Use64BitVectors), 2840 NEONMAP1(vqmovns_s32, aarch64_neon_sqxtn, VectorRet | Use64BitVectors), 2841 NEONMAP1(vqmovns_u32, aarch64_neon_uqxtn, VectorRet | Use64BitVectors), 2842 NEONMAP1(vqmovund_s64, aarch64_neon_scalar_sqxtun, AddRetType | Add1ArgType), 2843 NEONMAP1(vqmovunh_s16, aarch64_neon_sqxtun, VectorRet | Use64BitVectors), 2844 NEONMAP1(vqmovuns_s32, aarch64_neon_sqxtun, VectorRet | Use64BitVectors), 2845 NEONMAP1(vqnegb_s8, aarch64_neon_sqneg, Vectorize1ArgType | Use64BitVectors), 2846 NEONMAP1(vqnegd_s64, aarch64_neon_sqneg, Add1ArgType), 2847 NEONMAP1(vqnegh_s16, aarch64_neon_sqneg, Vectorize1ArgType | Use64BitVectors), 2848 NEONMAP1(vqnegs_s32, aarch64_neon_sqneg, Add1ArgType), 2849 NEONMAP1(vqrdmulhh_s16, aarch64_neon_sqrdmulh, Vectorize1ArgType | Use64BitVectors), 2850 NEONMAP1(vqrdmulhs_s32, aarch64_neon_sqrdmulh, Add1ArgType), 2851 NEONMAP1(vqrshlb_s8, aarch64_neon_sqrshl, Vectorize1ArgType | Use64BitVectors), 2852 NEONMAP1(vqrshlb_u8, aarch64_neon_uqrshl, Vectorize1ArgType | Use64BitVectors), 2853 NEONMAP1(vqrshld_s64, aarch64_neon_sqrshl, Add1ArgType), 2854 NEONMAP1(vqrshld_u64, aarch64_neon_uqrshl, Add1ArgType), 2855 NEONMAP1(vqrshlh_s16, aarch64_neon_sqrshl, Vectorize1ArgType | Use64BitVectors), 2856 NEONMAP1(vqrshlh_u16, aarch64_neon_uqrshl, Vectorize1ArgType | Use64BitVectors), 2857 NEONMAP1(vqrshls_s32, aarch64_neon_sqrshl, Add1ArgType), 2858 NEONMAP1(vqrshls_u32, aarch64_neon_uqrshl, Add1ArgType), 2859 NEONMAP1(vqrshrnd_n_s64, aarch64_neon_sqrshrn, AddRetType), 2860 NEONMAP1(vqrshrnd_n_u64, aarch64_neon_uqrshrn, AddRetType), 2861 NEONMAP1(vqrshrnh_n_s16, aarch64_neon_sqrshrn, VectorRet | Use64BitVectors), 2862 NEONMAP1(vqrshrnh_n_u16, aarch64_neon_uqrshrn, VectorRet | Use64BitVectors), 2863 NEONMAP1(vqrshrns_n_s32, aarch64_neon_sqrshrn, VectorRet | Use64BitVectors), 2864 NEONMAP1(vqrshrns_n_u32, aarch64_neon_uqrshrn, VectorRet | Use64BitVectors), 2865 NEONMAP1(vqrshrund_n_s64, aarch64_neon_sqrshrun, AddRetType), 2866 NEONMAP1(vqrshrunh_n_s16, aarch64_neon_sqrshrun, VectorRet | Use64BitVectors), 2867 NEONMAP1(vqrshruns_n_s32, aarch64_neon_sqrshrun, VectorRet | Use64BitVectors), 2868 NEONMAP1(vqshlb_n_s8, aarch64_neon_sqshl, Vectorize1ArgType | Use64BitVectors), 2869 NEONMAP1(vqshlb_n_u8, aarch64_neon_uqshl, Vectorize1ArgType | Use64BitVectors), 2870 NEONMAP1(vqshlb_s8, aarch64_neon_sqshl, Vectorize1ArgType | Use64BitVectors), 2871 NEONMAP1(vqshlb_u8, aarch64_neon_uqshl, Vectorize1ArgType | Use64BitVectors), 2872 NEONMAP1(vqshld_s64, aarch64_neon_sqshl, Add1ArgType), 2873 NEONMAP1(vqshld_u64, aarch64_neon_uqshl, Add1ArgType), 2874 NEONMAP1(vqshlh_n_s16, aarch64_neon_sqshl, Vectorize1ArgType | Use64BitVectors), 2875 NEONMAP1(vqshlh_n_u16, aarch64_neon_uqshl, Vectorize1ArgType | Use64BitVectors), 2876 NEONMAP1(vqshlh_s16, aarch64_neon_sqshl, Vectorize1ArgType | Use64BitVectors), 2877 NEONMAP1(vqshlh_u16, aarch64_neon_uqshl, Vectorize1ArgType | Use64BitVectors), 2878 NEONMAP1(vqshls_n_s32, aarch64_neon_sqshl, Add1ArgType), 2879 NEONMAP1(vqshls_n_u32, aarch64_neon_uqshl, Add1ArgType), 2880 NEONMAP1(vqshls_s32, aarch64_neon_sqshl, Add1ArgType), 2881 NEONMAP1(vqshls_u32, aarch64_neon_uqshl, Add1ArgType), 2882 NEONMAP1(vqshlub_n_s8, aarch64_neon_sqshlu, Vectorize1ArgType | Use64BitVectors), 2883 NEONMAP1(vqshluh_n_s16, aarch64_neon_sqshlu, Vectorize1ArgType | Use64BitVectors), 2884 NEONMAP1(vqshlus_n_s32, aarch64_neon_sqshlu, Add1ArgType), 2885 NEONMAP1(vqshrnd_n_s64, aarch64_neon_sqshrn, AddRetType), 2886 NEONMAP1(vqshrnd_n_u64, aarch64_neon_uqshrn, AddRetType), 2887 NEONMAP1(vqshrnh_n_s16, aarch64_neon_sqshrn, VectorRet | Use64BitVectors), 2888 NEONMAP1(vqshrnh_n_u16, aarch64_neon_uqshrn, VectorRet | Use64BitVectors), 2889 NEONMAP1(vqshrns_n_s32, aarch64_neon_sqshrn, VectorRet | Use64BitVectors), 2890 NEONMAP1(vqshrns_n_u32, aarch64_neon_uqshrn, VectorRet | Use64BitVectors), 2891 NEONMAP1(vqshrund_n_s64, aarch64_neon_sqshrun, AddRetType), 2892 NEONMAP1(vqshrunh_n_s16, aarch64_neon_sqshrun, VectorRet | Use64BitVectors), 2893 NEONMAP1(vqshruns_n_s32, aarch64_neon_sqshrun, VectorRet | Use64BitVectors), 2894 NEONMAP1(vqsubb_s8, aarch64_neon_sqsub, Vectorize1ArgType | Use64BitVectors), 2895 NEONMAP1(vqsubb_u8, aarch64_neon_uqsub, Vectorize1ArgType | Use64BitVectors), 2896 NEONMAP1(vqsubd_s64, aarch64_neon_sqsub, Add1ArgType), 2897 NEONMAP1(vqsubd_u64, aarch64_neon_uqsub, Add1ArgType), 2898 NEONMAP1(vqsubh_s16, aarch64_neon_sqsub, Vectorize1ArgType | Use64BitVectors), 2899 NEONMAP1(vqsubh_u16, aarch64_neon_uqsub, Vectorize1ArgType | Use64BitVectors), 2900 NEONMAP1(vqsubs_s32, aarch64_neon_sqsub, Add1ArgType), 2901 NEONMAP1(vqsubs_u32, aarch64_neon_uqsub, Add1ArgType), 2902 NEONMAP1(vrecped_f64, aarch64_neon_frecpe, Add1ArgType), 2903 NEONMAP1(vrecpes_f32, aarch64_neon_frecpe, Add1ArgType), 2904 NEONMAP1(vrecpxd_f64, aarch64_neon_frecpx, Add1ArgType), 2905 NEONMAP1(vrecpxs_f32, aarch64_neon_frecpx, Add1ArgType), 2906 NEONMAP1(vrshld_s64, aarch64_neon_srshl, Add1ArgType), 2907 NEONMAP1(vrshld_u64, aarch64_neon_urshl, Add1ArgType), 2908 NEONMAP1(vrsqrted_f64, aarch64_neon_frsqrte, Add1ArgType), 2909 NEONMAP1(vrsqrtes_f32, aarch64_neon_frsqrte, Add1ArgType), 2910 NEONMAP1(vrsqrtsd_f64, aarch64_neon_frsqrts, Add1ArgType), 2911 NEONMAP1(vrsqrtss_f32, aarch64_neon_frsqrts, Add1ArgType), 2912 NEONMAP1(vsha1cq_u32, aarch64_crypto_sha1c, 0), 2913 NEONMAP1(vsha1h_u32, aarch64_crypto_sha1h, 0), 2914 NEONMAP1(vsha1mq_u32, aarch64_crypto_sha1m, 0), 2915 NEONMAP1(vsha1pq_u32, aarch64_crypto_sha1p, 0), 2916 NEONMAP1(vshld_s64, aarch64_neon_sshl, Add1ArgType), 2917 NEONMAP1(vshld_u64, aarch64_neon_ushl, Add1ArgType), 2918 NEONMAP1(vslid_n_s64, aarch64_neon_vsli, Vectorize1ArgType), 2919 NEONMAP1(vslid_n_u64, aarch64_neon_vsli, Vectorize1ArgType), 2920 NEONMAP1(vsqaddb_u8, aarch64_neon_usqadd, Vectorize1ArgType | Use64BitVectors), 2921 NEONMAP1(vsqaddd_u64, aarch64_neon_usqadd, Add1ArgType), 2922 NEONMAP1(vsqaddh_u16, aarch64_neon_usqadd, Vectorize1ArgType | Use64BitVectors), 2923 NEONMAP1(vsqadds_u32, aarch64_neon_usqadd, Add1ArgType), 2924 NEONMAP1(vsrid_n_s64, aarch64_neon_vsri, Vectorize1ArgType), 2925 NEONMAP1(vsrid_n_u64, aarch64_neon_vsri, Vectorize1ArgType), 2926 NEONMAP1(vuqaddb_s8, aarch64_neon_suqadd, Vectorize1ArgType | Use64BitVectors), 2927 NEONMAP1(vuqaddd_s64, aarch64_neon_suqadd, Add1ArgType), 2928 NEONMAP1(vuqaddh_s16, aarch64_neon_suqadd, Vectorize1ArgType | Use64BitVectors), 2929 NEONMAP1(vuqadds_s32, aarch64_neon_suqadd, Add1ArgType), 2930 }; 2931 2932 #undef NEONMAP0 2933 #undef NEONMAP1 2934 #undef NEONMAP2 2935 2936 static bool NEONSIMDIntrinsicsProvenSorted = false; 2937 2938 static bool AArch64SIMDIntrinsicsProvenSorted = false; 2939 static bool AArch64SISDIntrinsicsProvenSorted = false; 2940 2941 2942 static const NeonIntrinsicInfo * 2943 findNeonIntrinsicInMap(ArrayRef<NeonIntrinsicInfo> IntrinsicMap, 2944 unsigned BuiltinID, bool &MapProvenSorted) { 2945 2946 #ifndef NDEBUG 2947 if (!MapProvenSorted) { 2948 assert(std::is_sorted(std::begin(IntrinsicMap), std::end(IntrinsicMap))); 2949 MapProvenSorted = true; 2950 } 2951 #endif 2952 2953 const NeonIntrinsicInfo *Builtin = 2954 std::lower_bound(IntrinsicMap.begin(), IntrinsicMap.end(), BuiltinID); 2955 2956 if (Builtin != IntrinsicMap.end() && Builtin->BuiltinID == BuiltinID) 2957 return Builtin; 2958 2959 return nullptr; 2960 } 2961 2962 Function *CodeGenFunction::LookupNeonLLVMIntrinsic(unsigned IntrinsicID, 2963 unsigned Modifier, 2964 llvm::Type *ArgType, 2965 const CallExpr *E) { 2966 int VectorSize = 0; 2967 if (Modifier & Use64BitVectors) 2968 VectorSize = 64; 2969 else if (Modifier & Use128BitVectors) 2970 VectorSize = 128; 2971 2972 // Return type. 2973 SmallVector<llvm::Type *, 3> Tys; 2974 if (Modifier & AddRetType) { 2975 llvm::Type *Ty = ConvertType(E->getCallReturnType(getContext())); 2976 if (Modifier & VectorizeRetType) 2977 Ty = llvm::VectorType::get( 2978 Ty, VectorSize ? VectorSize / Ty->getPrimitiveSizeInBits() : 1); 2979 2980 Tys.push_back(Ty); 2981 } 2982 2983 // Arguments. 2984 if (Modifier & VectorizeArgTypes) { 2985 int Elts = VectorSize ? VectorSize / ArgType->getPrimitiveSizeInBits() : 1; 2986 ArgType = llvm::VectorType::get(ArgType, Elts); 2987 } 2988 2989 if (Modifier & (Add1ArgType | Add2ArgTypes)) 2990 Tys.push_back(ArgType); 2991 2992 if (Modifier & Add2ArgTypes) 2993 Tys.push_back(ArgType); 2994 2995 if (Modifier & InventFloatType) 2996 Tys.push_back(FloatTy); 2997 2998 return CGM.getIntrinsic(IntrinsicID, Tys); 2999 } 3000 3001 static Value *EmitCommonNeonSISDBuiltinExpr(CodeGenFunction &CGF, 3002 const NeonIntrinsicInfo &SISDInfo, 3003 SmallVectorImpl<Value *> &Ops, 3004 const CallExpr *E) { 3005 unsigned BuiltinID = SISDInfo.BuiltinID; 3006 unsigned int Int = SISDInfo.LLVMIntrinsic; 3007 unsigned Modifier = SISDInfo.TypeModifier; 3008 const char *s = SISDInfo.NameHint; 3009 3010 switch (BuiltinID) { 3011 case NEON::BI__builtin_neon_vcled_s64: 3012 case NEON::BI__builtin_neon_vcled_u64: 3013 case NEON::BI__builtin_neon_vcles_f32: 3014 case NEON::BI__builtin_neon_vcled_f64: 3015 case NEON::BI__builtin_neon_vcltd_s64: 3016 case NEON::BI__builtin_neon_vcltd_u64: 3017 case NEON::BI__builtin_neon_vclts_f32: 3018 case NEON::BI__builtin_neon_vcltd_f64: 3019 case NEON::BI__builtin_neon_vcales_f32: 3020 case NEON::BI__builtin_neon_vcaled_f64: 3021 case NEON::BI__builtin_neon_vcalts_f32: 3022 case NEON::BI__builtin_neon_vcaltd_f64: 3023 // Only one direction of comparisons actually exist, cmle is actually a cmge 3024 // with swapped operands. The table gives us the right intrinsic but we 3025 // still need to do the swap. 3026 std::swap(Ops[0], Ops[1]); 3027 break; 3028 } 3029 3030 assert(Int && "Generic code assumes a valid intrinsic"); 3031 3032 // Determine the type(s) of this overloaded AArch64 intrinsic. 3033 const Expr *Arg = E->getArg(0); 3034 llvm::Type *ArgTy = CGF.ConvertType(Arg->getType()); 3035 Function *F = CGF.LookupNeonLLVMIntrinsic(Int, Modifier, ArgTy, E); 3036 3037 int j = 0; 3038 ConstantInt *C0 = ConstantInt::get(CGF.SizeTy, 0); 3039 for (Function::const_arg_iterator ai = F->arg_begin(), ae = F->arg_end(); 3040 ai != ae; ++ai, ++j) { 3041 llvm::Type *ArgTy = ai->getType(); 3042 if (Ops[j]->getType()->getPrimitiveSizeInBits() == 3043 ArgTy->getPrimitiveSizeInBits()) 3044 continue; 3045 3046 assert(ArgTy->isVectorTy() && !Ops[j]->getType()->isVectorTy()); 3047 // The constant argument to an _n_ intrinsic always has Int32Ty, so truncate 3048 // it before inserting. 3049 Ops[j] = 3050 CGF.Builder.CreateTruncOrBitCast(Ops[j], ArgTy->getVectorElementType()); 3051 Ops[j] = 3052 CGF.Builder.CreateInsertElement(UndefValue::get(ArgTy), Ops[j], C0); 3053 } 3054 3055 Value *Result = CGF.EmitNeonCall(F, Ops, s); 3056 llvm::Type *ResultType = CGF.ConvertType(E->getType()); 3057 if (ResultType->getPrimitiveSizeInBits() < 3058 Result->getType()->getPrimitiveSizeInBits()) 3059 return CGF.Builder.CreateExtractElement(Result, C0); 3060 3061 return CGF.Builder.CreateBitCast(Result, ResultType, s); 3062 } 3063 3064 Value *CodeGenFunction::EmitCommonNeonBuiltinExpr( 3065 unsigned BuiltinID, unsigned LLVMIntrinsic, unsigned AltLLVMIntrinsic, 3066 const char *NameHint, unsigned Modifier, const CallExpr *E, 3067 SmallVectorImpl<llvm::Value *> &Ops, Address PtrOp0, Address PtrOp1) { 3068 // Get the last argument, which specifies the vector type. 3069 llvm::APSInt NeonTypeConst; 3070 const Expr *Arg = E->getArg(E->getNumArgs() - 1); 3071 if (!Arg->isIntegerConstantExpr(NeonTypeConst, getContext())) 3072 return nullptr; 3073 3074 // Determine the type of this overloaded NEON intrinsic. 3075 NeonTypeFlags Type(NeonTypeConst.getZExtValue()); 3076 bool Usgn = Type.isUnsigned(); 3077 bool Quad = Type.isQuad(); 3078 3079 llvm::VectorType *VTy = GetNeonType(this, Type); 3080 llvm::Type *Ty = VTy; 3081 if (!Ty) 3082 return nullptr; 3083 3084 auto getAlignmentValue32 = [&](Address addr) -> Value* { 3085 return Builder.getInt32(addr.getAlignment().getQuantity()); 3086 }; 3087 3088 unsigned Int = LLVMIntrinsic; 3089 if ((Modifier & UnsignedAlts) && !Usgn) 3090 Int = AltLLVMIntrinsic; 3091 3092 switch (BuiltinID) { 3093 default: break; 3094 case NEON::BI__builtin_neon_vabs_v: 3095 case NEON::BI__builtin_neon_vabsq_v: 3096 if (VTy->getElementType()->isFloatingPointTy()) 3097 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::fabs, Ty), Ops, "vabs"); 3098 return EmitNeonCall(CGM.getIntrinsic(LLVMIntrinsic, Ty), Ops, "vabs"); 3099 case NEON::BI__builtin_neon_vaddhn_v: { 3100 llvm::VectorType *SrcTy = 3101 llvm::VectorType::getExtendedElementVectorType(VTy); 3102 3103 // %sum = add <4 x i32> %lhs, %rhs 3104 Ops[0] = Builder.CreateBitCast(Ops[0], SrcTy); 3105 Ops[1] = Builder.CreateBitCast(Ops[1], SrcTy); 3106 Ops[0] = Builder.CreateAdd(Ops[0], Ops[1], "vaddhn"); 3107 3108 // %high = lshr <4 x i32> %sum, <i32 16, i32 16, i32 16, i32 16> 3109 Constant *ShiftAmt = 3110 ConstantInt::get(SrcTy, SrcTy->getScalarSizeInBits() / 2); 3111 Ops[0] = Builder.CreateLShr(Ops[0], ShiftAmt, "vaddhn"); 3112 3113 // %res = trunc <4 x i32> %high to <4 x i16> 3114 return Builder.CreateTrunc(Ops[0], VTy, "vaddhn"); 3115 } 3116 case NEON::BI__builtin_neon_vcale_v: 3117 case NEON::BI__builtin_neon_vcaleq_v: 3118 case NEON::BI__builtin_neon_vcalt_v: 3119 case NEON::BI__builtin_neon_vcaltq_v: 3120 std::swap(Ops[0], Ops[1]); 3121 case NEON::BI__builtin_neon_vcage_v: 3122 case NEON::BI__builtin_neon_vcageq_v: 3123 case NEON::BI__builtin_neon_vcagt_v: 3124 case NEON::BI__builtin_neon_vcagtq_v: { 3125 llvm::Type *VecFlt = llvm::VectorType::get( 3126 VTy->getScalarSizeInBits() == 32 ? FloatTy : DoubleTy, 3127 VTy->getNumElements()); 3128 llvm::Type *Tys[] = { VTy, VecFlt }; 3129 Function *F = CGM.getIntrinsic(LLVMIntrinsic, Tys); 3130 return EmitNeonCall(F, Ops, NameHint); 3131 } 3132 case NEON::BI__builtin_neon_vclz_v: 3133 case NEON::BI__builtin_neon_vclzq_v: 3134 // We generate target-independent intrinsic, which needs a second argument 3135 // for whether or not clz of zero is undefined; on ARM it isn't. 3136 Ops.push_back(Builder.getInt1(getTarget().isCLZForZeroUndef())); 3137 break; 3138 case NEON::BI__builtin_neon_vcvt_f32_v: 3139 case NEON::BI__builtin_neon_vcvtq_f32_v: 3140 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 3141 Ty = GetNeonType(this, NeonTypeFlags(NeonTypeFlags::Float32, false, Quad)); 3142 return Usgn ? Builder.CreateUIToFP(Ops[0], Ty, "vcvt") 3143 : Builder.CreateSIToFP(Ops[0], Ty, "vcvt"); 3144 case NEON::BI__builtin_neon_vcvt_n_f32_v: 3145 case NEON::BI__builtin_neon_vcvt_n_f64_v: 3146 case NEON::BI__builtin_neon_vcvtq_n_f32_v: 3147 case NEON::BI__builtin_neon_vcvtq_n_f64_v: { 3148 llvm::Type *Tys[2] = { GetFloatNeonType(this, Type), Ty }; 3149 Int = Usgn ? LLVMIntrinsic : AltLLVMIntrinsic; 3150 Function *F = CGM.getIntrinsic(Int, Tys); 3151 return EmitNeonCall(F, Ops, "vcvt_n"); 3152 } 3153 case NEON::BI__builtin_neon_vcvt_n_s32_v: 3154 case NEON::BI__builtin_neon_vcvt_n_u32_v: 3155 case NEON::BI__builtin_neon_vcvt_n_s64_v: 3156 case NEON::BI__builtin_neon_vcvt_n_u64_v: 3157 case NEON::BI__builtin_neon_vcvtq_n_s32_v: 3158 case NEON::BI__builtin_neon_vcvtq_n_u32_v: 3159 case NEON::BI__builtin_neon_vcvtq_n_s64_v: 3160 case NEON::BI__builtin_neon_vcvtq_n_u64_v: { 3161 llvm::Type *Tys[2] = { Ty, GetFloatNeonType(this, Type) }; 3162 Function *F = CGM.getIntrinsic(LLVMIntrinsic, Tys); 3163 return EmitNeonCall(F, Ops, "vcvt_n"); 3164 } 3165 case NEON::BI__builtin_neon_vcvt_s32_v: 3166 case NEON::BI__builtin_neon_vcvt_u32_v: 3167 case NEON::BI__builtin_neon_vcvt_s64_v: 3168 case NEON::BI__builtin_neon_vcvt_u64_v: 3169 case NEON::BI__builtin_neon_vcvtq_s32_v: 3170 case NEON::BI__builtin_neon_vcvtq_u32_v: 3171 case NEON::BI__builtin_neon_vcvtq_s64_v: 3172 case NEON::BI__builtin_neon_vcvtq_u64_v: { 3173 Ops[0] = Builder.CreateBitCast(Ops[0], GetFloatNeonType(this, Type)); 3174 return Usgn ? Builder.CreateFPToUI(Ops[0], Ty, "vcvt") 3175 : Builder.CreateFPToSI(Ops[0], Ty, "vcvt"); 3176 } 3177 case NEON::BI__builtin_neon_vcvta_s32_v: 3178 case NEON::BI__builtin_neon_vcvta_s64_v: 3179 case NEON::BI__builtin_neon_vcvta_u32_v: 3180 case NEON::BI__builtin_neon_vcvta_u64_v: 3181 case NEON::BI__builtin_neon_vcvtaq_s32_v: 3182 case NEON::BI__builtin_neon_vcvtaq_s64_v: 3183 case NEON::BI__builtin_neon_vcvtaq_u32_v: 3184 case NEON::BI__builtin_neon_vcvtaq_u64_v: 3185 case NEON::BI__builtin_neon_vcvtn_s32_v: 3186 case NEON::BI__builtin_neon_vcvtn_s64_v: 3187 case NEON::BI__builtin_neon_vcvtn_u32_v: 3188 case NEON::BI__builtin_neon_vcvtn_u64_v: 3189 case NEON::BI__builtin_neon_vcvtnq_s32_v: 3190 case NEON::BI__builtin_neon_vcvtnq_s64_v: 3191 case NEON::BI__builtin_neon_vcvtnq_u32_v: 3192 case NEON::BI__builtin_neon_vcvtnq_u64_v: 3193 case NEON::BI__builtin_neon_vcvtp_s32_v: 3194 case NEON::BI__builtin_neon_vcvtp_s64_v: 3195 case NEON::BI__builtin_neon_vcvtp_u32_v: 3196 case NEON::BI__builtin_neon_vcvtp_u64_v: 3197 case NEON::BI__builtin_neon_vcvtpq_s32_v: 3198 case NEON::BI__builtin_neon_vcvtpq_s64_v: 3199 case NEON::BI__builtin_neon_vcvtpq_u32_v: 3200 case NEON::BI__builtin_neon_vcvtpq_u64_v: 3201 case NEON::BI__builtin_neon_vcvtm_s32_v: 3202 case NEON::BI__builtin_neon_vcvtm_s64_v: 3203 case NEON::BI__builtin_neon_vcvtm_u32_v: 3204 case NEON::BI__builtin_neon_vcvtm_u64_v: 3205 case NEON::BI__builtin_neon_vcvtmq_s32_v: 3206 case NEON::BI__builtin_neon_vcvtmq_s64_v: 3207 case NEON::BI__builtin_neon_vcvtmq_u32_v: 3208 case NEON::BI__builtin_neon_vcvtmq_u64_v: { 3209 llvm::Type *Tys[2] = { Ty, GetFloatNeonType(this, Type) }; 3210 return EmitNeonCall(CGM.getIntrinsic(LLVMIntrinsic, Tys), Ops, NameHint); 3211 } 3212 case NEON::BI__builtin_neon_vext_v: 3213 case NEON::BI__builtin_neon_vextq_v: { 3214 int CV = cast<ConstantInt>(Ops[2])->getSExtValue(); 3215 SmallVector<Constant*, 16> Indices; 3216 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) 3217 Indices.push_back(ConstantInt::get(Int32Ty, i+CV)); 3218 3219 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 3220 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 3221 Value *SV = llvm::ConstantVector::get(Indices); 3222 return Builder.CreateShuffleVector(Ops[0], Ops[1], SV, "vext"); 3223 } 3224 case NEON::BI__builtin_neon_vfma_v: 3225 case NEON::BI__builtin_neon_vfmaq_v: { 3226 Value *F = CGM.getIntrinsic(Intrinsic::fma, Ty); 3227 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 3228 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 3229 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 3230 3231 // NEON intrinsic puts accumulator first, unlike the LLVM fma. 3232 return Builder.CreateCall(F, {Ops[1], Ops[2], Ops[0]}); 3233 } 3234 case NEON::BI__builtin_neon_vld1_v: 3235 case NEON::BI__builtin_neon_vld1q_v: { 3236 llvm::Type *Tys[] = {Ty, Int8PtrTy}; 3237 Ops.push_back(getAlignmentValue32(PtrOp0)); 3238 return EmitNeonCall(CGM.getIntrinsic(LLVMIntrinsic, Tys), Ops, "vld1"); 3239 } 3240 case NEON::BI__builtin_neon_vld2_v: 3241 case NEON::BI__builtin_neon_vld2q_v: 3242 case NEON::BI__builtin_neon_vld3_v: 3243 case NEON::BI__builtin_neon_vld3q_v: 3244 case NEON::BI__builtin_neon_vld4_v: 3245 case NEON::BI__builtin_neon_vld4q_v: { 3246 llvm::Type *Tys[] = {Ty, Int8PtrTy}; 3247 Function *F = CGM.getIntrinsic(LLVMIntrinsic, Tys); 3248 Value *Align = getAlignmentValue32(PtrOp1); 3249 Ops[1] = Builder.CreateCall(F, {Ops[1], Align}, NameHint); 3250 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 3251 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 3252 return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]); 3253 } 3254 case NEON::BI__builtin_neon_vld1_dup_v: 3255 case NEON::BI__builtin_neon_vld1q_dup_v: { 3256 Value *V = UndefValue::get(Ty); 3257 Ty = llvm::PointerType::getUnqual(VTy->getElementType()); 3258 PtrOp0 = Builder.CreateBitCast(PtrOp0, Ty); 3259 LoadInst *Ld = Builder.CreateLoad(PtrOp0); 3260 llvm::Constant *CI = ConstantInt::get(SizeTy, 0); 3261 Ops[0] = Builder.CreateInsertElement(V, Ld, CI); 3262 return EmitNeonSplat(Ops[0], CI); 3263 } 3264 case NEON::BI__builtin_neon_vld2_lane_v: 3265 case NEON::BI__builtin_neon_vld2q_lane_v: 3266 case NEON::BI__builtin_neon_vld3_lane_v: 3267 case NEON::BI__builtin_neon_vld3q_lane_v: 3268 case NEON::BI__builtin_neon_vld4_lane_v: 3269 case NEON::BI__builtin_neon_vld4q_lane_v: { 3270 llvm::Type *Tys[] = {Ty, Int8PtrTy}; 3271 Function *F = CGM.getIntrinsic(LLVMIntrinsic, Tys); 3272 for (unsigned I = 2; I < Ops.size() - 1; ++I) 3273 Ops[I] = Builder.CreateBitCast(Ops[I], Ty); 3274 Ops.push_back(getAlignmentValue32(PtrOp1)); 3275 Ops[1] = Builder.CreateCall(F, makeArrayRef(Ops).slice(1), NameHint); 3276 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 3277 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 3278 return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]); 3279 } 3280 case NEON::BI__builtin_neon_vmovl_v: { 3281 llvm::Type *DTy =llvm::VectorType::getTruncatedElementVectorType(VTy); 3282 Ops[0] = Builder.CreateBitCast(Ops[0], DTy); 3283 if (Usgn) 3284 return Builder.CreateZExt(Ops[0], Ty, "vmovl"); 3285 return Builder.CreateSExt(Ops[0], Ty, "vmovl"); 3286 } 3287 case NEON::BI__builtin_neon_vmovn_v: { 3288 llvm::Type *QTy = llvm::VectorType::getExtendedElementVectorType(VTy); 3289 Ops[0] = Builder.CreateBitCast(Ops[0], QTy); 3290 return Builder.CreateTrunc(Ops[0], Ty, "vmovn"); 3291 } 3292 case NEON::BI__builtin_neon_vmull_v: 3293 // FIXME: the integer vmull operations could be emitted in terms of pure 3294 // LLVM IR (2 exts followed by a mul). Unfortunately LLVM has a habit of 3295 // hoisting the exts outside loops. Until global ISel comes along that can 3296 // see through such movement this leads to bad CodeGen. So we need an 3297 // intrinsic for now. 3298 Int = Usgn ? Intrinsic::arm_neon_vmullu : Intrinsic::arm_neon_vmulls; 3299 Int = Type.isPoly() ? (unsigned)Intrinsic::arm_neon_vmullp : Int; 3300 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmull"); 3301 case NEON::BI__builtin_neon_vpadal_v: 3302 case NEON::BI__builtin_neon_vpadalq_v: { 3303 // The source operand type has twice as many elements of half the size. 3304 unsigned EltBits = VTy->getElementType()->getPrimitiveSizeInBits(); 3305 llvm::Type *EltTy = 3306 llvm::IntegerType::get(getLLVMContext(), EltBits / 2); 3307 llvm::Type *NarrowTy = 3308 llvm::VectorType::get(EltTy, VTy->getNumElements() * 2); 3309 llvm::Type *Tys[2] = { Ty, NarrowTy }; 3310 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, NameHint); 3311 } 3312 case NEON::BI__builtin_neon_vpaddl_v: 3313 case NEON::BI__builtin_neon_vpaddlq_v: { 3314 // The source operand type has twice as many elements of half the size. 3315 unsigned EltBits = VTy->getElementType()->getPrimitiveSizeInBits(); 3316 llvm::Type *EltTy = llvm::IntegerType::get(getLLVMContext(), EltBits / 2); 3317 llvm::Type *NarrowTy = 3318 llvm::VectorType::get(EltTy, VTy->getNumElements() * 2); 3319 llvm::Type *Tys[2] = { Ty, NarrowTy }; 3320 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vpaddl"); 3321 } 3322 case NEON::BI__builtin_neon_vqdmlal_v: 3323 case NEON::BI__builtin_neon_vqdmlsl_v: { 3324 SmallVector<Value *, 2> MulOps(Ops.begin() + 1, Ops.end()); 3325 Ops[1] = 3326 EmitNeonCall(CGM.getIntrinsic(LLVMIntrinsic, Ty), MulOps, "vqdmlal"); 3327 Ops.resize(2); 3328 return EmitNeonCall(CGM.getIntrinsic(AltLLVMIntrinsic, Ty), Ops, NameHint); 3329 } 3330 case NEON::BI__builtin_neon_vqshl_n_v: 3331 case NEON::BI__builtin_neon_vqshlq_n_v: 3332 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqshl_n", 3333 1, false); 3334 case NEON::BI__builtin_neon_vqshlu_n_v: 3335 case NEON::BI__builtin_neon_vqshluq_n_v: 3336 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqshlu_n", 3337 1, false); 3338 case NEON::BI__builtin_neon_vrecpe_v: 3339 case NEON::BI__builtin_neon_vrecpeq_v: 3340 case NEON::BI__builtin_neon_vrsqrte_v: 3341 case NEON::BI__builtin_neon_vrsqrteq_v: 3342 Int = Ty->isFPOrFPVectorTy() ? LLVMIntrinsic : AltLLVMIntrinsic; 3343 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, NameHint); 3344 3345 case NEON::BI__builtin_neon_vrshr_n_v: 3346 case NEON::BI__builtin_neon_vrshrq_n_v: 3347 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrshr_n", 3348 1, true); 3349 case NEON::BI__builtin_neon_vshl_n_v: 3350 case NEON::BI__builtin_neon_vshlq_n_v: 3351 Ops[1] = EmitNeonShiftVector(Ops[1], Ty, false); 3352 return Builder.CreateShl(Builder.CreateBitCast(Ops[0],Ty), Ops[1], 3353 "vshl_n"); 3354 case NEON::BI__builtin_neon_vshll_n_v: { 3355 llvm::Type *SrcTy = llvm::VectorType::getTruncatedElementVectorType(VTy); 3356 Ops[0] = Builder.CreateBitCast(Ops[0], SrcTy); 3357 if (Usgn) 3358 Ops[0] = Builder.CreateZExt(Ops[0], VTy); 3359 else 3360 Ops[0] = Builder.CreateSExt(Ops[0], VTy); 3361 Ops[1] = EmitNeonShiftVector(Ops[1], VTy, false); 3362 return Builder.CreateShl(Ops[0], Ops[1], "vshll_n"); 3363 } 3364 case NEON::BI__builtin_neon_vshrn_n_v: { 3365 llvm::Type *SrcTy = llvm::VectorType::getExtendedElementVectorType(VTy); 3366 Ops[0] = Builder.CreateBitCast(Ops[0], SrcTy); 3367 Ops[1] = EmitNeonShiftVector(Ops[1], SrcTy, false); 3368 if (Usgn) 3369 Ops[0] = Builder.CreateLShr(Ops[0], Ops[1]); 3370 else 3371 Ops[0] = Builder.CreateAShr(Ops[0], Ops[1]); 3372 return Builder.CreateTrunc(Ops[0], Ty, "vshrn_n"); 3373 } 3374 case NEON::BI__builtin_neon_vshr_n_v: 3375 case NEON::BI__builtin_neon_vshrq_n_v: 3376 return EmitNeonRShiftImm(Ops[0], Ops[1], Ty, Usgn, "vshr_n"); 3377 case NEON::BI__builtin_neon_vst1_v: 3378 case NEON::BI__builtin_neon_vst1q_v: 3379 case NEON::BI__builtin_neon_vst2_v: 3380 case NEON::BI__builtin_neon_vst2q_v: 3381 case NEON::BI__builtin_neon_vst3_v: 3382 case NEON::BI__builtin_neon_vst3q_v: 3383 case NEON::BI__builtin_neon_vst4_v: 3384 case NEON::BI__builtin_neon_vst4q_v: 3385 case NEON::BI__builtin_neon_vst2_lane_v: 3386 case NEON::BI__builtin_neon_vst2q_lane_v: 3387 case NEON::BI__builtin_neon_vst3_lane_v: 3388 case NEON::BI__builtin_neon_vst3q_lane_v: 3389 case NEON::BI__builtin_neon_vst4_lane_v: 3390 case NEON::BI__builtin_neon_vst4q_lane_v: { 3391 llvm::Type *Tys[] = {Int8PtrTy, Ty}; 3392 Ops.push_back(getAlignmentValue32(PtrOp0)); 3393 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, ""); 3394 } 3395 case NEON::BI__builtin_neon_vsubhn_v: { 3396 llvm::VectorType *SrcTy = 3397 llvm::VectorType::getExtendedElementVectorType(VTy); 3398 3399 // %sum = add <4 x i32> %lhs, %rhs 3400 Ops[0] = Builder.CreateBitCast(Ops[0], SrcTy); 3401 Ops[1] = Builder.CreateBitCast(Ops[1], SrcTy); 3402 Ops[0] = Builder.CreateSub(Ops[0], Ops[1], "vsubhn"); 3403 3404 // %high = lshr <4 x i32> %sum, <i32 16, i32 16, i32 16, i32 16> 3405 Constant *ShiftAmt = 3406 ConstantInt::get(SrcTy, SrcTy->getScalarSizeInBits() / 2); 3407 Ops[0] = Builder.CreateLShr(Ops[0], ShiftAmt, "vsubhn"); 3408 3409 // %res = trunc <4 x i32> %high to <4 x i16> 3410 return Builder.CreateTrunc(Ops[0], VTy, "vsubhn"); 3411 } 3412 case NEON::BI__builtin_neon_vtrn_v: 3413 case NEON::BI__builtin_neon_vtrnq_v: { 3414 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty)); 3415 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 3416 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 3417 Value *SV = nullptr; 3418 3419 for (unsigned vi = 0; vi != 2; ++vi) { 3420 SmallVector<Constant*, 16> Indices; 3421 for (unsigned i = 0, e = VTy->getNumElements(); i != e; i += 2) { 3422 Indices.push_back(Builder.getInt32(i+vi)); 3423 Indices.push_back(Builder.getInt32(i+e+vi)); 3424 } 3425 Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ty, Ops[0], vi); 3426 SV = llvm::ConstantVector::get(Indices); 3427 SV = Builder.CreateShuffleVector(Ops[1], Ops[2], SV, "vtrn"); 3428 SV = Builder.CreateDefaultAlignedStore(SV, Addr); 3429 } 3430 return SV; 3431 } 3432 case NEON::BI__builtin_neon_vtst_v: 3433 case NEON::BI__builtin_neon_vtstq_v: { 3434 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 3435 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 3436 Ops[0] = Builder.CreateAnd(Ops[0], Ops[1]); 3437 Ops[0] = Builder.CreateICmp(ICmpInst::ICMP_NE, Ops[0], 3438 ConstantAggregateZero::get(Ty)); 3439 return Builder.CreateSExt(Ops[0], Ty, "vtst"); 3440 } 3441 case NEON::BI__builtin_neon_vuzp_v: 3442 case NEON::BI__builtin_neon_vuzpq_v: { 3443 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty)); 3444 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 3445 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 3446 Value *SV = nullptr; 3447 3448 for (unsigned vi = 0; vi != 2; ++vi) { 3449 SmallVector<Constant*, 16> Indices; 3450 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) 3451 Indices.push_back(ConstantInt::get(Int32Ty, 2*i+vi)); 3452 3453 Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ty, Ops[0], vi); 3454 SV = llvm::ConstantVector::get(Indices); 3455 SV = Builder.CreateShuffleVector(Ops[1], Ops[2], SV, "vuzp"); 3456 SV = Builder.CreateDefaultAlignedStore(SV, Addr); 3457 } 3458 return SV; 3459 } 3460 case NEON::BI__builtin_neon_vzip_v: 3461 case NEON::BI__builtin_neon_vzipq_v: { 3462 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty)); 3463 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 3464 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 3465 Value *SV = nullptr; 3466 3467 for (unsigned vi = 0; vi != 2; ++vi) { 3468 SmallVector<Constant*, 16> Indices; 3469 for (unsigned i = 0, e = VTy->getNumElements(); i != e; i += 2) { 3470 Indices.push_back(ConstantInt::get(Int32Ty, (i + vi*e) >> 1)); 3471 Indices.push_back(ConstantInt::get(Int32Ty, ((i + vi*e) >> 1)+e)); 3472 } 3473 Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ty, Ops[0], vi); 3474 SV = llvm::ConstantVector::get(Indices); 3475 SV = Builder.CreateShuffleVector(Ops[1], Ops[2], SV, "vzip"); 3476 SV = Builder.CreateDefaultAlignedStore(SV, Addr); 3477 } 3478 return SV; 3479 } 3480 } 3481 3482 assert(Int && "Expected valid intrinsic number"); 3483 3484 // Determine the type(s) of this overloaded AArch64 intrinsic. 3485 Function *F = LookupNeonLLVMIntrinsic(Int, Modifier, Ty, E); 3486 3487 Value *Result = EmitNeonCall(F, Ops, NameHint); 3488 llvm::Type *ResultType = ConvertType(E->getType()); 3489 // AArch64 intrinsic one-element vector type cast to 3490 // scalar type expected by the builtin 3491 return Builder.CreateBitCast(Result, ResultType, NameHint); 3492 } 3493 3494 Value *CodeGenFunction::EmitAArch64CompareBuiltinExpr( 3495 Value *Op, llvm::Type *Ty, const CmpInst::Predicate Fp, 3496 const CmpInst::Predicate Ip, const Twine &Name) { 3497 llvm::Type *OTy = Op->getType(); 3498 3499 // FIXME: this is utterly horrific. We should not be looking at previous 3500 // codegen context to find out what needs doing. Unfortunately TableGen 3501 // currently gives us exactly the same calls for vceqz_f32 and vceqz_s32 3502 // (etc). 3503 if (BitCastInst *BI = dyn_cast<BitCastInst>(Op)) 3504 OTy = BI->getOperand(0)->getType(); 3505 3506 Op = Builder.CreateBitCast(Op, OTy); 3507 if (OTy->getScalarType()->isFloatingPointTy()) { 3508 Op = Builder.CreateFCmp(Fp, Op, Constant::getNullValue(OTy)); 3509 } else { 3510 Op = Builder.CreateICmp(Ip, Op, Constant::getNullValue(OTy)); 3511 } 3512 return Builder.CreateSExt(Op, Ty, Name); 3513 } 3514 3515 static Value *packTBLDVectorList(CodeGenFunction &CGF, ArrayRef<Value *> Ops, 3516 Value *ExtOp, Value *IndexOp, 3517 llvm::Type *ResTy, unsigned IntID, 3518 const char *Name) { 3519 SmallVector<Value *, 2> TblOps; 3520 if (ExtOp) 3521 TblOps.push_back(ExtOp); 3522 3523 // Build a vector containing sequential number like (0, 1, 2, ..., 15) 3524 SmallVector<Constant*, 16> Indices; 3525 llvm::VectorType *TblTy = cast<llvm::VectorType>(Ops[0]->getType()); 3526 for (unsigned i = 0, e = TblTy->getNumElements(); i != e; ++i) { 3527 Indices.push_back(ConstantInt::get(CGF.Int32Ty, 2*i)); 3528 Indices.push_back(ConstantInt::get(CGF.Int32Ty, 2*i+1)); 3529 } 3530 Value *SV = llvm::ConstantVector::get(Indices); 3531 3532 int PairPos = 0, End = Ops.size() - 1; 3533 while (PairPos < End) { 3534 TblOps.push_back(CGF.Builder.CreateShuffleVector(Ops[PairPos], 3535 Ops[PairPos+1], SV, Name)); 3536 PairPos += 2; 3537 } 3538 3539 // If there's an odd number of 64-bit lookup table, fill the high 64-bit 3540 // of the 128-bit lookup table with zero. 3541 if (PairPos == End) { 3542 Value *ZeroTbl = ConstantAggregateZero::get(TblTy); 3543 TblOps.push_back(CGF.Builder.CreateShuffleVector(Ops[PairPos], 3544 ZeroTbl, SV, Name)); 3545 } 3546 3547 Function *TblF; 3548 TblOps.push_back(IndexOp); 3549 TblF = CGF.CGM.getIntrinsic(IntID, ResTy); 3550 3551 return CGF.EmitNeonCall(TblF, TblOps, Name); 3552 } 3553 3554 Value *CodeGenFunction::GetValueForARMHint(unsigned BuiltinID) { 3555 unsigned Value; 3556 switch (BuiltinID) { 3557 default: 3558 return nullptr; 3559 case ARM::BI__builtin_arm_nop: 3560 Value = 0; 3561 break; 3562 case ARM::BI__builtin_arm_yield: 3563 case ARM::BI__yield: 3564 Value = 1; 3565 break; 3566 case ARM::BI__builtin_arm_wfe: 3567 case ARM::BI__wfe: 3568 Value = 2; 3569 break; 3570 case ARM::BI__builtin_arm_wfi: 3571 case ARM::BI__wfi: 3572 Value = 3; 3573 break; 3574 case ARM::BI__builtin_arm_sev: 3575 case ARM::BI__sev: 3576 Value = 4; 3577 break; 3578 case ARM::BI__builtin_arm_sevl: 3579 case ARM::BI__sevl: 3580 Value = 5; 3581 break; 3582 } 3583 3584 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::arm_hint), 3585 llvm::ConstantInt::get(Int32Ty, Value)); 3586 } 3587 3588 // Generates the IR for the read/write special register builtin, 3589 // ValueType is the type of the value that is to be written or read, 3590 // RegisterType is the type of the register being written to or read from. 3591 static Value *EmitSpecialRegisterBuiltin(CodeGenFunction &CGF, 3592 const CallExpr *E, 3593 llvm::Type *RegisterType, 3594 llvm::Type *ValueType, bool IsRead) { 3595 // write and register intrinsics only support 32 and 64 bit operations. 3596 assert((RegisterType->isIntegerTy(32) || RegisterType->isIntegerTy(64)) 3597 && "Unsupported size for register."); 3598 3599 CodeGen::CGBuilderTy &Builder = CGF.Builder; 3600 CodeGen::CodeGenModule &CGM = CGF.CGM; 3601 LLVMContext &Context = CGM.getLLVMContext(); 3602 3603 const Expr *SysRegStrExpr = E->getArg(0)->IgnoreParenCasts(); 3604 StringRef SysReg = cast<StringLiteral>(SysRegStrExpr)->getString(); 3605 3606 llvm::Metadata *Ops[] = { llvm::MDString::get(Context, SysReg) }; 3607 llvm::MDNode *RegName = llvm::MDNode::get(Context, Ops); 3608 llvm::Value *Metadata = llvm::MetadataAsValue::get(Context, RegName); 3609 3610 llvm::Type *Types[] = { RegisterType }; 3611 3612 bool MixedTypes = RegisterType->isIntegerTy(64) && ValueType->isIntegerTy(32); 3613 assert(!(RegisterType->isIntegerTy(32) && ValueType->isIntegerTy(64)) 3614 && "Can't fit 64-bit value in 32-bit register"); 3615 3616 if (IsRead) { 3617 llvm::Value *F = CGM.getIntrinsic(llvm::Intrinsic::read_register, Types); 3618 llvm::Value *Call = Builder.CreateCall(F, Metadata); 3619 3620 if (MixedTypes) 3621 // Read into 64 bit register and then truncate result to 32 bit. 3622 return Builder.CreateTrunc(Call, ValueType); 3623 3624 if (ValueType->isPointerTy()) 3625 // Have i32/i64 result (Call) but want to return a VoidPtrTy (i8*). 3626 return Builder.CreateIntToPtr(Call, ValueType); 3627 3628 return Call; 3629 } 3630 3631 llvm::Value *F = CGM.getIntrinsic(llvm::Intrinsic::write_register, Types); 3632 llvm::Value *ArgValue = CGF.EmitScalarExpr(E->getArg(1)); 3633 if (MixedTypes) { 3634 // Extend 32 bit write value to 64 bit to pass to write. 3635 ArgValue = Builder.CreateZExt(ArgValue, RegisterType); 3636 return Builder.CreateCall(F, { Metadata, ArgValue }); 3637 } 3638 3639 if (ValueType->isPointerTy()) { 3640 // Have VoidPtrTy ArgValue but want to return an i32/i64. 3641 ArgValue = Builder.CreatePtrToInt(ArgValue, RegisterType); 3642 return Builder.CreateCall(F, { Metadata, ArgValue }); 3643 } 3644 3645 return Builder.CreateCall(F, { Metadata, ArgValue }); 3646 } 3647 3648 /// Return true if BuiltinID is an overloaded Neon intrinsic with an extra 3649 /// argument that specifies the vector type. 3650 static bool HasExtraNeonArgument(unsigned BuiltinID) { 3651 switch (BuiltinID) { 3652 default: break; 3653 case NEON::BI__builtin_neon_vget_lane_i8: 3654 case NEON::BI__builtin_neon_vget_lane_i16: 3655 case NEON::BI__builtin_neon_vget_lane_i32: 3656 case NEON::BI__builtin_neon_vget_lane_i64: 3657 case NEON::BI__builtin_neon_vget_lane_f32: 3658 case NEON::BI__builtin_neon_vgetq_lane_i8: 3659 case NEON::BI__builtin_neon_vgetq_lane_i16: 3660 case NEON::BI__builtin_neon_vgetq_lane_i32: 3661 case NEON::BI__builtin_neon_vgetq_lane_i64: 3662 case NEON::BI__builtin_neon_vgetq_lane_f32: 3663 case NEON::BI__builtin_neon_vset_lane_i8: 3664 case NEON::BI__builtin_neon_vset_lane_i16: 3665 case NEON::BI__builtin_neon_vset_lane_i32: 3666 case NEON::BI__builtin_neon_vset_lane_i64: 3667 case NEON::BI__builtin_neon_vset_lane_f32: 3668 case NEON::BI__builtin_neon_vsetq_lane_i8: 3669 case NEON::BI__builtin_neon_vsetq_lane_i16: 3670 case NEON::BI__builtin_neon_vsetq_lane_i32: 3671 case NEON::BI__builtin_neon_vsetq_lane_i64: 3672 case NEON::BI__builtin_neon_vsetq_lane_f32: 3673 case NEON::BI__builtin_neon_vsha1h_u32: 3674 case NEON::BI__builtin_neon_vsha1cq_u32: 3675 case NEON::BI__builtin_neon_vsha1pq_u32: 3676 case NEON::BI__builtin_neon_vsha1mq_u32: 3677 case ARM::BI_MoveToCoprocessor: 3678 case ARM::BI_MoveToCoprocessor2: 3679 return false; 3680 } 3681 return true; 3682 } 3683 3684 Value *CodeGenFunction::EmitARMBuiltinExpr(unsigned BuiltinID, 3685 const CallExpr *E) { 3686 if (auto Hint = GetValueForARMHint(BuiltinID)) 3687 return Hint; 3688 3689 if (BuiltinID == ARM::BI__emit) { 3690 bool IsThumb = getTarget().getTriple().getArch() == llvm::Triple::thumb; 3691 llvm::FunctionType *FTy = 3692 llvm::FunctionType::get(VoidTy, /*Variadic=*/false); 3693 3694 APSInt Value; 3695 if (!E->getArg(0)->EvaluateAsInt(Value, CGM.getContext())) 3696 llvm_unreachable("Sema will ensure that the parameter is constant"); 3697 3698 uint64_t ZExtValue = Value.zextOrTrunc(IsThumb ? 16 : 32).getZExtValue(); 3699 3700 llvm::InlineAsm *Emit = 3701 IsThumb ? InlineAsm::get(FTy, ".inst.n 0x" + utohexstr(ZExtValue), "", 3702 /*SideEffects=*/true) 3703 : InlineAsm::get(FTy, ".inst 0x" + utohexstr(ZExtValue), "", 3704 /*SideEffects=*/true); 3705 3706 return Builder.CreateCall(Emit); 3707 } 3708 3709 if (BuiltinID == ARM::BI__builtin_arm_dbg) { 3710 Value *Option = EmitScalarExpr(E->getArg(0)); 3711 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::arm_dbg), Option); 3712 } 3713 3714 if (BuiltinID == ARM::BI__builtin_arm_prefetch) { 3715 Value *Address = EmitScalarExpr(E->getArg(0)); 3716 Value *RW = EmitScalarExpr(E->getArg(1)); 3717 Value *IsData = EmitScalarExpr(E->getArg(2)); 3718 3719 // Locality is not supported on ARM target 3720 Value *Locality = llvm::ConstantInt::get(Int32Ty, 3); 3721 3722 Value *F = CGM.getIntrinsic(Intrinsic::prefetch); 3723 return Builder.CreateCall(F, {Address, RW, Locality, IsData}); 3724 } 3725 3726 if (BuiltinID == ARM::BI__builtin_arm_rbit) { 3727 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::arm_rbit), 3728 EmitScalarExpr(E->getArg(0)), 3729 "rbit"); 3730 } 3731 3732 if (BuiltinID == ARM::BI__clear_cache) { 3733 assert(E->getNumArgs() == 2 && "__clear_cache takes 2 arguments"); 3734 const FunctionDecl *FD = E->getDirectCallee(); 3735 Value *Ops[2]; 3736 for (unsigned i = 0; i < 2; i++) 3737 Ops[i] = EmitScalarExpr(E->getArg(i)); 3738 llvm::Type *Ty = CGM.getTypes().ConvertType(FD->getType()); 3739 llvm::FunctionType *FTy = cast<llvm::FunctionType>(Ty); 3740 StringRef Name = FD->getName(); 3741 return EmitNounwindRuntimeCall(CGM.CreateRuntimeFunction(FTy, Name), Ops); 3742 } 3743 3744 if (BuiltinID == ARM::BI__builtin_arm_ldrexd || 3745 ((BuiltinID == ARM::BI__builtin_arm_ldrex || 3746 BuiltinID == ARM::BI__builtin_arm_ldaex) && 3747 getContext().getTypeSize(E->getType()) == 64) || 3748 BuiltinID == ARM::BI__ldrexd) { 3749 Function *F; 3750 3751 switch (BuiltinID) { 3752 default: llvm_unreachable("unexpected builtin"); 3753 case ARM::BI__builtin_arm_ldaex: 3754 F = CGM.getIntrinsic(Intrinsic::arm_ldaexd); 3755 break; 3756 case ARM::BI__builtin_arm_ldrexd: 3757 case ARM::BI__builtin_arm_ldrex: 3758 case ARM::BI__ldrexd: 3759 F = CGM.getIntrinsic(Intrinsic::arm_ldrexd); 3760 break; 3761 } 3762 3763 Value *LdPtr = EmitScalarExpr(E->getArg(0)); 3764 Value *Val = Builder.CreateCall(F, Builder.CreateBitCast(LdPtr, Int8PtrTy), 3765 "ldrexd"); 3766 3767 Value *Val0 = Builder.CreateExtractValue(Val, 1); 3768 Value *Val1 = Builder.CreateExtractValue(Val, 0); 3769 Val0 = Builder.CreateZExt(Val0, Int64Ty); 3770 Val1 = Builder.CreateZExt(Val1, Int64Ty); 3771 3772 Value *ShiftCst = llvm::ConstantInt::get(Int64Ty, 32); 3773 Val = Builder.CreateShl(Val0, ShiftCst, "shl", true /* nuw */); 3774 Val = Builder.CreateOr(Val, Val1); 3775 return Builder.CreateBitCast(Val, ConvertType(E->getType())); 3776 } 3777 3778 if (BuiltinID == ARM::BI__builtin_arm_ldrex || 3779 BuiltinID == ARM::BI__builtin_arm_ldaex) { 3780 Value *LoadAddr = EmitScalarExpr(E->getArg(0)); 3781 3782 QualType Ty = E->getType(); 3783 llvm::Type *RealResTy = ConvertType(Ty); 3784 llvm::Type *IntResTy = llvm::IntegerType::get(getLLVMContext(), 3785 getContext().getTypeSize(Ty)); 3786 LoadAddr = Builder.CreateBitCast(LoadAddr, IntResTy->getPointerTo()); 3787 3788 Function *F = CGM.getIntrinsic(BuiltinID == ARM::BI__builtin_arm_ldaex 3789 ? Intrinsic::arm_ldaex 3790 : Intrinsic::arm_ldrex, 3791 LoadAddr->getType()); 3792 Value *Val = Builder.CreateCall(F, LoadAddr, "ldrex"); 3793 3794 if (RealResTy->isPointerTy()) 3795 return Builder.CreateIntToPtr(Val, RealResTy); 3796 else { 3797 Val = Builder.CreateTruncOrBitCast(Val, IntResTy); 3798 return Builder.CreateBitCast(Val, RealResTy); 3799 } 3800 } 3801 3802 if (BuiltinID == ARM::BI__builtin_arm_strexd || 3803 ((BuiltinID == ARM::BI__builtin_arm_stlex || 3804 BuiltinID == ARM::BI__builtin_arm_strex) && 3805 getContext().getTypeSize(E->getArg(0)->getType()) == 64)) { 3806 Function *F = CGM.getIntrinsic(BuiltinID == ARM::BI__builtin_arm_stlex 3807 ? Intrinsic::arm_stlexd 3808 : Intrinsic::arm_strexd); 3809 llvm::Type *STy = llvm::StructType::get(Int32Ty, Int32Ty, nullptr); 3810 3811 Address Tmp = CreateMemTemp(E->getArg(0)->getType()); 3812 Value *Val = EmitScalarExpr(E->getArg(0)); 3813 Builder.CreateStore(Val, Tmp); 3814 3815 Address LdPtr = Builder.CreateBitCast(Tmp,llvm::PointerType::getUnqual(STy)); 3816 Val = Builder.CreateLoad(LdPtr); 3817 3818 Value *Arg0 = Builder.CreateExtractValue(Val, 0); 3819 Value *Arg1 = Builder.CreateExtractValue(Val, 1); 3820 Value *StPtr = Builder.CreateBitCast(EmitScalarExpr(E->getArg(1)), Int8PtrTy); 3821 return Builder.CreateCall(F, {Arg0, Arg1, StPtr}, "strexd"); 3822 } 3823 3824 if (BuiltinID == ARM::BI__builtin_arm_strex || 3825 BuiltinID == ARM::BI__builtin_arm_stlex) { 3826 Value *StoreVal = EmitScalarExpr(E->getArg(0)); 3827 Value *StoreAddr = EmitScalarExpr(E->getArg(1)); 3828 3829 QualType Ty = E->getArg(0)->getType(); 3830 llvm::Type *StoreTy = llvm::IntegerType::get(getLLVMContext(), 3831 getContext().getTypeSize(Ty)); 3832 StoreAddr = Builder.CreateBitCast(StoreAddr, StoreTy->getPointerTo()); 3833 3834 if (StoreVal->getType()->isPointerTy()) 3835 StoreVal = Builder.CreatePtrToInt(StoreVal, Int32Ty); 3836 else { 3837 StoreVal = Builder.CreateBitCast(StoreVal, StoreTy); 3838 StoreVal = Builder.CreateZExtOrBitCast(StoreVal, Int32Ty); 3839 } 3840 3841 Function *F = CGM.getIntrinsic(BuiltinID == ARM::BI__builtin_arm_stlex 3842 ? Intrinsic::arm_stlex 3843 : Intrinsic::arm_strex, 3844 StoreAddr->getType()); 3845 return Builder.CreateCall(F, {StoreVal, StoreAddr}, "strex"); 3846 } 3847 3848 if (BuiltinID == ARM::BI__builtin_arm_clrex) { 3849 Function *F = CGM.getIntrinsic(Intrinsic::arm_clrex); 3850 return Builder.CreateCall(F); 3851 } 3852 3853 // CRC32 3854 Intrinsic::ID CRCIntrinsicID = Intrinsic::not_intrinsic; 3855 switch (BuiltinID) { 3856 case ARM::BI__builtin_arm_crc32b: 3857 CRCIntrinsicID = Intrinsic::arm_crc32b; break; 3858 case ARM::BI__builtin_arm_crc32cb: 3859 CRCIntrinsicID = Intrinsic::arm_crc32cb; break; 3860 case ARM::BI__builtin_arm_crc32h: 3861 CRCIntrinsicID = Intrinsic::arm_crc32h; break; 3862 case ARM::BI__builtin_arm_crc32ch: 3863 CRCIntrinsicID = Intrinsic::arm_crc32ch; break; 3864 case ARM::BI__builtin_arm_crc32w: 3865 case ARM::BI__builtin_arm_crc32d: 3866 CRCIntrinsicID = Intrinsic::arm_crc32w; break; 3867 case ARM::BI__builtin_arm_crc32cw: 3868 case ARM::BI__builtin_arm_crc32cd: 3869 CRCIntrinsicID = Intrinsic::arm_crc32cw; break; 3870 } 3871 3872 if (CRCIntrinsicID != Intrinsic::not_intrinsic) { 3873 Value *Arg0 = EmitScalarExpr(E->getArg(0)); 3874 Value *Arg1 = EmitScalarExpr(E->getArg(1)); 3875 3876 // crc32{c,}d intrinsics are implemnted as two calls to crc32{c,}w 3877 // intrinsics, hence we need different codegen for these cases. 3878 if (BuiltinID == ARM::BI__builtin_arm_crc32d || 3879 BuiltinID == ARM::BI__builtin_arm_crc32cd) { 3880 Value *C1 = llvm::ConstantInt::get(Int64Ty, 32); 3881 Value *Arg1a = Builder.CreateTruncOrBitCast(Arg1, Int32Ty); 3882 Value *Arg1b = Builder.CreateLShr(Arg1, C1); 3883 Arg1b = Builder.CreateTruncOrBitCast(Arg1b, Int32Ty); 3884 3885 Function *F = CGM.getIntrinsic(CRCIntrinsicID); 3886 Value *Res = Builder.CreateCall(F, {Arg0, Arg1a}); 3887 return Builder.CreateCall(F, {Res, Arg1b}); 3888 } else { 3889 Arg1 = Builder.CreateZExtOrBitCast(Arg1, Int32Ty); 3890 3891 Function *F = CGM.getIntrinsic(CRCIntrinsicID); 3892 return Builder.CreateCall(F, {Arg0, Arg1}); 3893 } 3894 } 3895 3896 if (BuiltinID == ARM::BI__builtin_arm_rsr || 3897 BuiltinID == ARM::BI__builtin_arm_rsr64 || 3898 BuiltinID == ARM::BI__builtin_arm_rsrp || 3899 BuiltinID == ARM::BI__builtin_arm_wsr || 3900 BuiltinID == ARM::BI__builtin_arm_wsr64 || 3901 BuiltinID == ARM::BI__builtin_arm_wsrp) { 3902 3903 bool IsRead = BuiltinID == ARM::BI__builtin_arm_rsr || 3904 BuiltinID == ARM::BI__builtin_arm_rsr64 || 3905 BuiltinID == ARM::BI__builtin_arm_rsrp; 3906 3907 bool IsPointerBuiltin = BuiltinID == ARM::BI__builtin_arm_rsrp || 3908 BuiltinID == ARM::BI__builtin_arm_wsrp; 3909 3910 bool Is64Bit = BuiltinID == ARM::BI__builtin_arm_rsr64 || 3911 BuiltinID == ARM::BI__builtin_arm_wsr64; 3912 3913 llvm::Type *ValueType; 3914 llvm::Type *RegisterType; 3915 if (IsPointerBuiltin) { 3916 ValueType = VoidPtrTy; 3917 RegisterType = Int32Ty; 3918 } else if (Is64Bit) { 3919 ValueType = RegisterType = Int64Ty; 3920 } else { 3921 ValueType = RegisterType = Int32Ty; 3922 } 3923 3924 return EmitSpecialRegisterBuiltin(*this, E, RegisterType, ValueType, IsRead); 3925 } 3926 3927 // Find out if any arguments are required to be integer constant 3928 // expressions. 3929 unsigned ICEArguments = 0; 3930 ASTContext::GetBuiltinTypeError Error; 3931 getContext().GetBuiltinType(BuiltinID, Error, &ICEArguments); 3932 assert(Error == ASTContext::GE_None && "Should not codegen an error"); 3933 3934 auto getAlignmentValue32 = [&](Address addr) -> Value* { 3935 return Builder.getInt32(addr.getAlignment().getQuantity()); 3936 }; 3937 3938 Address PtrOp0 = Address::invalid(); 3939 Address PtrOp1 = Address::invalid(); 3940 SmallVector<Value*, 4> Ops; 3941 bool HasExtraArg = HasExtraNeonArgument(BuiltinID); 3942 unsigned NumArgs = E->getNumArgs() - (HasExtraArg ? 1 : 0); 3943 for (unsigned i = 0, e = NumArgs; i != e; i++) { 3944 if (i == 0) { 3945 switch (BuiltinID) { 3946 case NEON::BI__builtin_neon_vld1_v: 3947 case NEON::BI__builtin_neon_vld1q_v: 3948 case NEON::BI__builtin_neon_vld1q_lane_v: 3949 case NEON::BI__builtin_neon_vld1_lane_v: 3950 case NEON::BI__builtin_neon_vld1_dup_v: 3951 case NEON::BI__builtin_neon_vld1q_dup_v: 3952 case NEON::BI__builtin_neon_vst1_v: 3953 case NEON::BI__builtin_neon_vst1q_v: 3954 case NEON::BI__builtin_neon_vst1q_lane_v: 3955 case NEON::BI__builtin_neon_vst1_lane_v: 3956 case NEON::BI__builtin_neon_vst2_v: 3957 case NEON::BI__builtin_neon_vst2q_v: 3958 case NEON::BI__builtin_neon_vst2_lane_v: 3959 case NEON::BI__builtin_neon_vst2q_lane_v: 3960 case NEON::BI__builtin_neon_vst3_v: 3961 case NEON::BI__builtin_neon_vst3q_v: 3962 case NEON::BI__builtin_neon_vst3_lane_v: 3963 case NEON::BI__builtin_neon_vst3q_lane_v: 3964 case NEON::BI__builtin_neon_vst4_v: 3965 case NEON::BI__builtin_neon_vst4q_v: 3966 case NEON::BI__builtin_neon_vst4_lane_v: 3967 case NEON::BI__builtin_neon_vst4q_lane_v: 3968 // Get the alignment for the argument in addition to the value; 3969 // we'll use it later. 3970 PtrOp0 = EmitPointerWithAlignment(E->getArg(0)); 3971 Ops.push_back(PtrOp0.getPointer()); 3972 continue; 3973 } 3974 } 3975 if (i == 1) { 3976 switch (BuiltinID) { 3977 case NEON::BI__builtin_neon_vld2_v: 3978 case NEON::BI__builtin_neon_vld2q_v: 3979 case NEON::BI__builtin_neon_vld3_v: 3980 case NEON::BI__builtin_neon_vld3q_v: 3981 case NEON::BI__builtin_neon_vld4_v: 3982 case NEON::BI__builtin_neon_vld4q_v: 3983 case NEON::BI__builtin_neon_vld2_lane_v: 3984 case NEON::BI__builtin_neon_vld2q_lane_v: 3985 case NEON::BI__builtin_neon_vld3_lane_v: 3986 case NEON::BI__builtin_neon_vld3q_lane_v: 3987 case NEON::BI__builtin_neon_vld4_lane_v: 3988 case NEON::BI__builtin_neon_vld4q_lane_v: 3989 case NEON::BI__builtin_neon_vld2_dup_v: 3990 case NEON::BI__builtin_neon_vld3_dup_v: 3991 case NEON::BI__builtin_neon_vld4_dup_v: 3992 // Get the alignment for the argument in addition to the value; 3993 // we'll use it later. 3994 PtrOp1 = EmitPointerWithAlignment(E->getArg(1)); 3995 Ops.push_back(PtrOp1.getPointer()); 3996 continue; 3997 } 3998 } 3999 4000 if ((ICEArguments & (1 << i)) == 0) { 4001 Ops.push_back(EmitScalarExpr(E->getArg(i))); 4002 } else { 4003 // If this is required to be a constant, constant fold it so that we know 4004 // that the generated intrinsic gets a ConstantInt. 4005 llvm::APSInt Result; 4006 bool IsConst = E->getArg(i)->isIntegerConstantExpr(Result, getContext()); 4007 assert(IsConst && "Constant arg isn't actually constant?"); (void)IsConst; 4008 Ops.push_back(llvm::ConstantInt::get(getLLVMContext(), Result)); 4009 } 4010 } 4011 4012 switch (BuiltinID) { 4013 default: break; 4014 4015 case NEON::BI__builtin_neon_vget_lane_i8: 4016 case NEON::BI__builtin_neon_vget_lane_i16: 4017 case NEON::BI__builtin_neon_vget_lane_i32: 4018 case NEON::BI__builtin_neon_vget_lane_i64: 4019 case NEON::BI__builtin_neon_vget_lane_f32: 4020 case NEON::BI__builtin_neon_vgetq_lane_i8: 4021 case NEON::BI__builtin_neon_vgetq_lane_i16: 4022 case NEON::BI__builtin_neon_vgetq_lane_i32: 4023 case NEON::BI__builtin_neon_vgetq_lane_i64: 4024 case NEON::BI__builtin_neon_vgetq_lane_f32: 4025 return Builder.CreateExtractElement(Ops[0], Ops[1], "vget_lane"); 4026 4027 case NEON::BI__builtin_neon_vset_lane_i8: 4028 case NEON::BI__builtin_neon_vset_lane_i16: 4029 case NEON::BI__builtin_neon_vset_lane_i32: 4030 case NEON::BI__builtin_neon_vset_lane_i64: 4031 case NEON::BI__builtin_neon_vset_lane_f32: 4032 case NEON::BI__builtin_neon_vsetq_lane_i8: 4033 case NEON::BI__builtin_neon_vsetq_lane_i16: 4034 case NEON::BI__builtin_neon_vsetq_lane_i32: 4035 case NEON::BI__builtin_neon_vsetq_lane_i64: 4036 case NEON::BI__builtin_neon_vsetq_lane_f32: 4037 return Builder.CreateInsertElement(Ops[1], Ops[0], Ops[2], "vset_lane"); 4038 4039 case NEON::BI__builtin_neon_vsha1h_u32: 4040 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_sha1h), Ops, 4041 "vsha1h"); 4042 case NEON::BI__builtin_neon_vsha1cq_u32: 4043 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_sha1c), Ops, 4044 "vsha1h"); 4045 case NEON::BI__builtin_neon_vsha1pq_u32: 4046 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_sha1p), Ops, 4047 "vsha1h"); 4048 case NEON::BI__builtin_neon_vsha1mq_u32: 4049 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_sha1m), Ops, 4050 "vsha1h"); 4051 4052 // The ARM _MoveToCoprocessor builtins put the input register value as 4053 // the first argument, but the LLVM intrinsic expects it as the third one. 4054 case ARM::BI_MoveToCoprocessor: 4055 case ARM::BI_MoveToCoprocessor2: { 4056 Function *F = CGM.getIntrinsic(BuiltinID == ARM::BI_MoveToCoprocessor ? 4057 Intrinsic::arm_mcr : Intrinsic::arm_mcr2); 4058 return Builder.CreateCall(F, {Ops[1], Ops[2], Ops[0], 4059 Ops[3], Ops[4], Ops[5]}); 4060 } 4061 } 4062 4063 // Get the last argument, which specifies the vector type. 4064 assert(HasExtraArg); 4065 llvm::APSInt Result; 4066 const Expr *Arg = E->getArg(E->getNumArgs()-1); 4067 if (!Arg->isIntegerConstantExpr(Result, getContext())) 4068 return nullptr; 4069 4070 if (BuiltinID == ARM::BI__builtin_arm_vcvtr_f || 4071 BuiltinID == ARM::BI__builtin_arm_vcvtr_d) { 4072 // Determine the overloaded type of this builtin. 4073 llvm::Type *Ty; 4074 if (BuiltinID == ARM::BI__builtin_arm_vcvtr_f) 4075 Ty = FloatTy; 4076 else 4077 Ty = DoubleTy; 4078 4079 // Determine whether this is an unsigned conversion or not. 4080 bool usgn = Result.getZExtValue() == 1; 4081 unsigned Int = usgn ? Intrinsic::arm_vcvtru : Intrinsic::arm_vcvtr; 4082 4083 // Call the appropriate intrinsic. 4084 Function *F = CGM.getIntrinsic(Int, Ty); 4085 return Builder.CreateCall(F, Ops, "vcvtr"); 4086 } 4087 4088 // Determine the type of this overloaded NEON intrinsic. 4089 NeonTypeFlags Type(Result.getZExtValue()); 4090 bool usgn = Type.isUnsigned(); 4091 bool rightShift = false; 4092 4093 llvm::VectorType *VTy = GetNeonType(this, Type); 4094 llvm::Type *Ty = VTy; 4095 if (!Ty) 4096 return nullptr; 4097 4098 // Many NEON builtins have identical semantics and uses in ARM and 4099 // AArch64. Emit these in a single function. 4100 auto IntrinsicMap = makeArrayRef(ARMSIMDIntrinsicMap); 4101 const NeonIntrinsicInfo *Builtin = findNeonIntrinsicInMap( 4102 IntrinsicMap, BuiltinID, NEONSIMDIntrinsicsProvenSorted); 4103 if (Builtin) 4104 return EmitCommonNeonBuiltinExpr( 4105 Builtin->BuiltinID, Builtin->LLVMIntrinsic, Builtin->AltLLVMIntrinsic, 4106 Builtin->NameHint, Builtin->TypeModifier, E, Ops, PtrOp0, PtrOp1); 4107 4108 unsigned Int; 4109 switch (BuiltinID) { 4110 default: return nullptr; 4111 case NEON::BI__builtin_neon_vld1q_lane_v: 4112 // Handle 64-bit integer elements as a special case. Use shuffles of 4113 // one-element vectors to avoid poor code for i64 in the backend. 4114 if (VTy->getElementType()->isIntegerTy(64)) { 4115 // Extract the other lane. 4116 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 4117 uint32_t Lane = cast<ConstantInt>(Ops[2])->getZExtValue(); 4118 Value *SV = llvm::ConstantVector::get(ConstantInt::get(Int32Ty, 1-Lane)); 4119 Ops[1] = Builder.CreateShuffleVector(Ops[1], Ops[1], SV); 4120 // Load the value as a one-element vector. 4121 Ty = llvm::VectorType::get(VTy->getElementType(), 1); 4122 llvm::Type *Tys[] = {Ty, Int8PtrTy}; 4123 Function *F = CGM.getIntrinsic(Intrinsic::arm_neon_vld1, Tys); 4124 Value *Align = getAlignmentValue32(PtrOp0); 4125 Value *Ld = Builder.CreateCall(F, {Ops[0], Align}); 4126 // Combine them. 4127 uint32_t Indices[] = {1 - Lane, Lane}; 4128 SV = llvm::ConstantDataVector::get(getLLVMContext(), Indices); 4129 return Builder.CreateShuffleVector(Ops[1], Ld, SV, "vld1q_lane"); 4130 } 4131 // fall through 4132 case NEON::BI__builtin_neon_vld1_lane_v: { 4133 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 4134 PtrOp0 = Builder.CreateElementBitCast(PtrOp0, VTy->getElementType()); 4135 Value *Ld = Builder.CreateLoad(PtrOp0); 4136 return Builder.CreateInsertElement(Ops[1], Ld, Ops[2], "vld1_lane"); 4137 } 4138 case NEON::BI__builtin_neon_vld2_dup_v: 4139 case NEON::BI__builtin_neon_vld3_dup_v: 4140 case NEON::BI__builtin_neon_vld4_dup_v: { 4141 // Handle 64-bit elements as a special-case. There is no "dup" needed. 4142 if (VTy->getElementType()->getPrimitiveSizeInBits() == 64) { 4143 switch (BuiltinID) { 4144 case NEON::BI__builtin_neon_vld2_dup_v: 4145 Int = Intrinsic::arm_neon_vld2; 4146 break; 4147 case NEON::BI__builtin_neon_vld3_dup_v: 4148 Int = Intrinsic::arm_neon_vld3; 4149 break; 4150 case NEON::BI__builtin_neon_vld4_dup_v: 4151 Int = Intrinsic::arm_neon_vld4; 4152 break; 4153 default: llvm_unreachable("unknown vld_dup intrinsic?"); 4154 } 4155 llvm::Type *Tys[] = {Ty, Int8PtrTy}; 4156 Function *F = CGM.getIntrinsic(Int, Tys); 4157 llvm::Value *Align = getAlignmentValue32(PtrOp1); 4158 Ops[1] = Builder.CreateCall(F, {Ops[1], Align}, "vld_dup"); 4159 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 4160 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 4161 return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]); 4162 } 4163 switch (BuiltinID) { 4164 case NEON::BI__builtin_neon_vld2_dup_v: 4165 Int = Intrinsic::arm_neon_vld2lane; 4166 break; 4167 case NEON::BI__builtin_neon_vld3_dup_v: 4168 Int = Intrinsic::arm_neon_vld3lane; 4169 break; 4170 case NEON::BI__builtin_neon_vld4_dup_v: 4171 Int = Intrinsic::arm_neon_vld4lane; 4172 break; 4173 default: llvm_unreachable("unknown vld_dup intrinsic?"); 4174 } 4175 llvm::Type *Tys[] = {Ty, Int8PtrTy}; 4176 Function *F = CGM.getIntrinsic(Int, Tys); 4177 llvm::StructType *STy = cast<llvm::StructType>(F->getReturnType()); 4178 4179 SmallVector<Value*, 6> Args; 4180 Args.push_back(Ops[1]); 4181 Args.append(STy->getNumElements(), UndefValue::get(Ty)); 4182 4183 llvm::Constant *CI = ConstantInt::get(Int32Ty, 0); 4184 Args.push_back(CI); 4185 Args.push_back(getAlignmentValue32(PtrOp1)); 4186 4187 Ops[1] = Builder.CreateCall(F, Args, "vld_dup"); 4188 // splat lane 0 to all elts in each vector of the result. 4189 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 4190 Value *Val = Builder.CreateExtractValue(Ops[1], i); 4191 Value *Elt = Builder.CreateBitCast(Val, Ty); 4192 Elt = EmitNeonSplat(Elt, CI); 4193 Elt = Builder.CreateBitCast(Elt, Val->getType()); 4194 Ops[1] = Builder.CreateInsertValue(Ops[1], Elt, i); 4195 } 4196 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 4197 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 4198 return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]); 4199 } 4200 case NEON::BI__builtin_neon_vqrshrn_n_v: 4201 Int = 4202 usgn ? Intrinsic::arm_neon_vqrshiftnu : Intrinsic::arm_neon_vqrshiftns; 4203 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqrshrn_n", 4204 1, true); 4205 case NEON::BI__builtin_neon_vqrshrun_n_v: 4206 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vqrshiftnsu, Ty), 4207 Ops, "vqrshrun_n", 1, true); 4208 case NEON::BI__builtin_neon_vqshrn_n_v: 4209 Int = usgn ? Intrinsic::arm_neon_vqshiftnu : Intrinsic::arm_neon_vqshiftns; 4210 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqshrn_n", 4211 1, true); 4212 case NEON::BI__builtin_neon_vqshrun_n_v: 4213 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vqshiftnsu, Ty), 4214 Ops, "vqshrun_n", 1, true); 4215 case NEON::BI__builtin_neon_vrecpe_v: 4216 case NEON::BI__builtin_neon_vrecpeq_v: 4217 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vrecpe, Ty), 4218 Ops, "vrecpe"); 4219 case NEON::BI__builtin_neon_vrshrn_n_v: 4220 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vrshiftn, Ty), 4221 Ops, "vrshrn_n", 1, true); 4222 case NEON::BI__builtin_neon_vrsra_n_v: 4223 case NEON::BI__builtin_neon_vrsraq_n_v: 4224 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 4225 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 4226 Ops[2] = EmitNeonShiftVector(Ops[2], Ty, true); 4227 Int = usgn ? Intrinsic::arm_neon_vrshiftu : Intrinsic::arm_neon_vrshifts; 4228 Ops[1] = Builder.CreateCall(CGM.getIntrinsic(Int, Ty), {Ops[1], Ops[2]}); 4229 return Builder.CreateAdd(Ops[0], Ops[1], "vrsra_n"); 4230 case NEON::BI__builtin_neon_vsri_n_v: 4231 case NEON::BI__builtin_neon_vsriq_n_v: 4232 rightShift = true; 4233 case NEON::BI__builtin_neon_vsli_n_v: 4234 case NEON::BI__builtin_neon_vsliq_n_v: 4235 Ops[2] = EmitNeonShiftVector(Ops[2], Ty, rightShift); 4236 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vshiftins, Ty), 4237 Ops, "vsli_n"); 4238 case NEON::BI__builtin_neon_vsra_n_v: 4239 case NEON::BI__builtin_neon_vsraq_n_v: 4240 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 4241 Ops[1] = EmitNeonRShiftImm(Ops[1], Ops[2], Ty, usgn, "vsra_n"); 4242 return Builder.CreateAdd(Ops[0], Ops[1]); 4243 case NEON::BI__builtin_neon_vst1q_lane_v: 4244 // Handle 64-bit integer elements as a special case. Use a shuffle to get 4245 // a one-element vector and avoid poor code for i64 in the backend. 4246 if (VTy->getElementType()->isIntegerTy(64)) { 4247 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 4248 Value *SV = llvm::ConstantVector::get(cast<llvm::Constant>(Ops[2])); 4249 Ops[1] = Builder.CreateShuffleVector(Ops[1], Ops[1], SV); 4250 Ops[2] = getAlignmentValue32(PtrOp0); 4251 llvm::Type *Tys[] = {Int8PtrTy, Ops[1]->getType()}; 4252 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::arm_neon_vst1, 4253 Tys), Ops); 4254 } 4255 // fall through 4256 case NEON::BI__builtin_neon_vst1_lane_v: { 4257 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 4258 Ops[1] = Builder.CreateExtractElement(Ops[1], Ops[2]); 4259 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 4260 auto St = Builder.CreateStore(Ops[1], Builder.CreateBitCast(PtrOp0, Ty)); 4261 return St; 4262 } 4263 case NEON::BI__builtin_neon_vtbl1_v: 4264 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbl1), 4265 Ops, "vtbl1"); 4266 case NEON::BI__builtin_neon_vtbl2_v: 4267 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbl2), 4268 Ops, "vtbl2"); 4269 case NEON::BI__builtin_neon_vtbl3_v: 4270 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbl3), 4271 Ops, "vtbl3"); 4272 case NEON::BI__builtin_neon_vtbl4_v: 4273 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbl4), 4274 Ops, "vtbl4"); 4275 case NEON::BI__builtin_neon_vtbx1_v: 4276 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbx1), 4277 Ops, "vtbx1"); 4278 case NEON::BI__builtin_neon_vtbx2_v: 4279 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbx2), 4280 Ops, "vtbx2"); 4281 case NEON::BI__builtin_neon_vtbx3_v: 4282 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbx3), 4283 Ops, "vtbx3"); 4284 case NEON::BI__builtin_neon_vtbx4_v: 4285 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbx4), 4286 Ops, "vtbx4"); 4287 } 4288 } 4289 4290 static Value *EmitAArch64TblBuiltinExpr(CodeGenFunction &CGF, unsigned BuiltinID, 4291 const CallExpr *E, 4292 SmallVectorImpl<Value *> &Ops) { 4293 unsigned int Int = 0; 4294 const char *s = nullptr; 4295 4296 switch (BuiltinID) { 4297 default: 4298 return nullptr; 4299 case NEON::BI__builtin_neon_vtbl1_v: 4300 case NEON::BI__builtin_neon_vqtbl1_v: 4301 case NEON::BI__builtin_neon_vqtbl1q_v: 4302 case NEON::BI__builtin_neon_vtbl2_v: 4303 case NEON::BI__builtin_neon_vqtbl2_v: 4304 case NEON::BI__builtin_neon_vqtbl2q_v: 4305 case NEON::BI__builtin_neon_vtbl3_v: 4306 case NEON::BI__builtin_neon_vqtbl3_v: 4307 case NEON::BI__builtin_neon_vqtbl3q_v: 4308 case NEON::BI__builtin_neon_vtbl4_v: 4309 case NEON::BI__builtin_neon_vqtbl4_v: 4310 case NEON::BI__builtin_neon_vqtbl4q_v: 4311 break; 4312 case NEON::BI__builtin_neon_vtbx1_v: 4313 case NEON::BI__builtin_neon_vqtbx1_v: 4314 case NEON::BI__builtin_neon_vqtbx1q_v: 4315 case NEON::BI__builtin_neon_vtbx2_v: 4316 case NEON::BI__builtin_neon_vqtbx2_v: 4317 case NEON::BI__builtin_neon_vqtbx2q_v: 4318 case NEON::BI__builtin_neon_vtbx3_v: 4319 case NEON::BI__builtin_neon_vqtbx3_v: 4320 case NEON::BI__builtin_neon_vqtbx3q_v: 4321 case NEON::BI__builtin_neon_vtbx4_v: 4322 case NEON::BI__builtin_neon_vqtbx4_v: 4323 case NEON::BI__builtin_neon_vqtbx4q_v: 4324 break; 4325 } 4326 4327 assert(E->getNumArgs() >= 3); 4328 4329 // Get the last argument, which specifies the vector type. 4330 llvm::APSInt Result; 4331 const Expr *Arg = E->getArg(E->getNumArgs() - 1); 4332 if (!Arg->isIntegerConstantExpr(Result, CGF.getContext())) 4333 return nullptr; 4334 4335 // Determine the type of this overloaded NEON intrinsic. 4336 NeonTypeFlags Type(Result.getZExtValue()); 4337 llvm::VectorType *Ty = GetNeonType(&CGF, Type); 4338 if (!Ty) 4339 return nullptr; 4340 4341 CodeGen::CGBuilderTy &Builder = CGF.Builder; 4342 4343 // AArch64 scalar builtins are not overloaded, they do not have an extra 4344 // argument that specifies the vector type, need to handle each case. 4345 switch (BuiltinID) { 4346 case NEON::BI__builtin_neon_vtbl1_v: { 4347 return packTBLDVectorList(CGF, makeArrayRef(Ops).slice(0, 1), nullptr, 4348 Ops[1], Ty, Intrinsic::aarch64_neon_tbl1, 4349 "vtbl1"); 4350 } 4351 case NEON::BI__builtin_neon_vtbl2_v: { 4352 return packTBLDVectorList(CGF, makeArrayRef(Ops).slice(0, 2), nullptr, 4353 Ops[2], Ty, Intrinsic::aarch64_neon_tbl1, 4354 "vtbl1"); 4355 } 4356 case NEON::BI__builtin_neon_vtbl3_v: { 4357 return packTBLDVectorList(CGF, makeArrayRef(Ops).slice(0, 3), nullptr, 4358 Ops[3], Ty, Intrinsic::aarch64_neon_tbl2, 4359 "vtbl2"); 4360 } 4361 case NEON::BI__builtin_neon_vtbl4_v: { 4362 return packTBLDVectorList(CGF, makeArrayRef(Ops).slice(0, 4), nullptr, 4363 Ops[4], Ty, Intrinsic::aarch64_neon_tbl2, 4364 "vtbl2"); 4365 } 4366 case NEON::BI__builtin_neon_vtbx1_v: { 4367 Value *TblRes = 4368 packTBLDVectorList(CGF, makeArrayRef(Ops).slice(1, 1), nullptr, Ops[2], 4369 Ty, Intrinsic::aarch64_neon_tbl1, "vtbl1"); 4370 4371 llvm::Constant *EightV = ConstantInt::get(Ty, 8); 4372 Value *CmpRes = Builder.CreateICmp(ICmpInst::ICMP_UGE, Ops[2], EightV); 4373 CmpRes = Builder.CreateSExt(CmpRes, Ty); 4374 4375 Value *EltsFromInput = Builder.CreateAnd(CmpRes, Ops[0]); 4376 Value *EltsFromTbl = Builder.CreateAnd(Builder.CreateNot(CmpRes), TblRes); 4377 return Builder.CreateOr(EltsFromInput, EltsFromTbl, "vtbx"); 4378 } 4379 case NEON::BI__builtin_neon_vtbx2_v: { 4380 return packTBLDVectorList(CGF, makeArrayRef(Ops).slice(1, 2), Ops[0], 4381 Ops[3], Ty, Intrinsic::aarch64_neon_tbx1, 4382 "vtbx1"); 4383 } 4384 case NEON::BI__builtin_neon_vtbx3_v: { 4385 Value *TblRes = 4386 packTBLDVectorList(CGF, makeArrayRef(Ops).slice(1, 3), nullptr, Ops[4], 4387 Ty, Intrinsic::aarch64_neon_tbl2, "vtbl2"); 4388 4389 llvm::Constant *TwentyFourV = ConstantInt::get(Ty, 24); 4390 Value *CmpRes = Builder.CreateICmp(ICmpInst::ICMP_UGE, Ops[4], 4391 TwentyFourV); 4392 CmpRes = Builder.CreateSExt(CmpRes, Ty); 4393 4394 Value *EltsFromInput = Builder.CreateAnd(CmpRes, Ops[0]); 4395 Value *EltsFromTbl = Builder.CreateAnd(Builder.CreateNot(CmpRes), TblRes); 4396 return Builder.CreateOr(EltsFromInput, EltsFromTbl, "vtbx"); 4397 } 4398 case NEON::BI__builtin_neon_vtbx4_v: { 4399 return packTBLDVectorList(CGF, makeArrayRef(Ops).slice(1, 4), Ops[0], 4400 Ops[5], Ty, Intrinsic::aarch64_neon_tbx2, 4401 "vtbx2"); 4402 } 4403 case NEON::BI__builtin_neon_vqtbl1_v: 4404 case NEON::BI__builtin_neon_vqtbl1q_v: 4405 Int = Intrinsic::aarch64_neon_tbl1; s = "vtbl1"; break; 4406 case NEON::BI__builtin_neon_vqtbl2_v: 4407 case NEON::BI__builtin_neon_vqtbl2q_v: { 4408 Int = Intrinsic::aarch64_neon_tbl2; s = "vtbl2"; break; 4409 case NEON::BI__builtin_neon_vqtbl3_v: 4410 case NEON::BI__builtin_neon_vqtbl3q_v: 4411 Int = Intrinsic::aarch64_neon_tbl3; s = "vtbl3"; break; 4412 case NEON::BI__builtin_neon_vqtbl4_v: 4413 case NEON::BI__builtin_neon_vqtbl4q_v: 4414 Int = Intrinsic::aarch64_neon_tbl4; s = "vtbl4"; break; 4415 case NEON::BI__builtin_neon_vqtbx1_v: 4416 case NEON::BI__builtin_neon_vqtbx1q_v: 4417 Int = Intrinsic::aarch64_neon_tbx1; s = "vtbx1"; break; 4418 case NEON::BI__builtin_neon_vqtbx2_v: 4419 case NEON::BI__builtin_neon_vqtbx2q_v: 4420 Int = Intrinsic::aarch64_neon_tbx2; s = "vtbx2"; break; 4421 case NEON::BI__builtin_neon_vqtbx3_v: 4422 case NEON::BI__builtin_neon_vqtbx3q_v: 4423 Int = Intrinsic::aarch64_neon_tbx3; s = "vtbx3"; break; 4424 case NEON::BI__builtin_neon_vqtbx4_v: 4425 case NEON::BI__builtin_neon_vqtbx4q_v: 4426 Int = Intrinsic::aarch64_neon_tbx4; s = "vtbx4"; break; 4427 } 4428 } 4429 4430 if (!Int) 4431 return nullptr; 4432 4433 Function *F = CGF.CGM.getIntrinsic(Int, Ty); 4434 return CGF.EmitNeonCall(F, Ops, s); 4435 } 4436 4437 Value *CodeGenFunction::vectorWrapScalar16(Value *Op) { 4438 llvm::Type *VTy = llvm::VectorType::get(Int16Ty, 4); 4439 Op = Builder.CreateBitCast(Op, Int16Ty); 4440 Value *V = UndefValue::get(VTy); 4441 llvm::Constant *CI = ConstantInt::get(SizeTy, 0); 4442 Op = Builder.CreateInsertElement(V, Op, CI); 4443 return Op; 4444 } 4445 4446 Value *CodeGenFunction::EmitAArch64BuiltinExpr(unsigned BuiltinID, 4447 const CallExpr *E) { 4448 unsigned HintID = static_cast<unsigned>(-1); 4449 switch (BuiltinID) { 4450 default: break; 4451 case AArch64::BI__builtin_arm_nop: 4452 HintID = 0; 4453 break; 4454 case AArch64::BI__builtin_arm_yield: 4455 HintID = 1; 4456 break; 4457 case AArch64::BI__builtin_arm_wfe: 4458 HintID = 2; 4459 break; 4460 case AArch64::BI__builtin_arm_wfi: 4461 HintID = 3; 4462 break; 4463 case AArch64::BI__builtin_arm_sev: 4464 HintID = 4; 4465 break; 4466 case AArch64::BI__builtin_arm_sevl: 4467 HintID = 5; 4468 break; 4469 } 4470 4471 if (HintID != static_cast<unsigned>(-1)) { 4472 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_hint); 4473 return Builder.CreateCall(F, llvm::ConstantInt::get(Int32Ty, HintID)); 4474 } 4475 4476 if (BuiltinID == AArch64::BI__builtin_arm_prefetch) { 4477 Value *Address = EmitScalarExpr(E->getArg(0)); 4478 Value *RW = EmitScalarExpr(E->getArg(1)); 4479 Value *CacheLevel = EmitScalarExpr(E->getArg(2)); 4480 Value *RetentionPolicy = EmitScalarExpr(E->getArg(3)); 4481 Value *IsData = EmitScalarExpr(E->getArg(4)); 4482 4483 Value *Locality = nullptr; 4484 if (cast<llvm::ConstantInt>(RetentionPolicy)->isZero()) { 4485 // Temporal fetch, needs to convert cache level to locality. 4486 Locality = llvm::ConstantInt::get(Int32Ty, 4487 -cast<llvm::ConstantInt>(CacheLevel)->getValue() + 3); 4488 } else { 4489 // Streaming fetch. 4490 Locality = llvm::ConstantInt::get(Int32Ty, 0); 4491 } 4492 4493 // FIXME: We need AArch64 specific LLVM intrinsic if we want to specify 4494 // PLDL3STRM or PLDL2STRM. 4495 Value *F = CGM.getIntrinsic(Intrinsic::prefetch); 4496 return Builder.CreateCall(F, {Address, RW, Locality, IsData}); 4497 } 4498 4499 if (BuiltinID == AArch64::BI__builtin_arm_rbit) { 4500 assert((getContext().getTypeSize(E->getType()) == 32) && 4501 "rbit of unusual size!"); 4502 llvm::Value *Arg = EmitScalarExpr(E->getArg(0)); 4503 return Builder.CreateCall( 4504 CGM.getIntrinsic(Intrinsic::aarch64_rbit, Arg->getType()), Arg, "rbit"); 4505 } 4506 if (BuiltinID == AArch64::BI__builtin_arm_rbit64) { 4507 assert((getContext().getTypeSize(E->getType()) == 64) && 4508 "rbit of unusual size!"); 4509 llvm::Value *Arg = EmitScalarExpr(E->getArg(0)); 4510 return Builder.CreateCall( 4511 CGM.getIntrinsic(Intrinsic::aarch64_rbit, Arg->getType()), Arg, "rbit"); 4512 } 4513 4514 if (BuiltinID == AArch64::BI__clear_cache) { 4515 assert(E->getNumArgs() == 2 && "__clear_cache takes 2 arguments"); 4516 const FunctionDecl *FD = E->getDirectCallee(); 4517 Value *Ops[2]; 4518 for (unsigned i = 0; i < 2; i++) 4519 Ops[i] = EmitScalarExpr(E->getArg(i)); 4520 llvm::Type *Ty = CGM.getTypes().ConvertType(FD->getType()); 4521 llvm::FunctionType *FTy = cast<llvm::FunctionType>(Ty); 4522 StringRef Name = FD->getName(); 4523 return EmitNounwindRuntimeCall(CGM.CreateRuntimeFunction(FTy, Name), Ops); 4524 } 4525 4526 if ((BuiltinID == AArch64::BI__builtin_arm_ldrex || 4527 BuiltinID == AArch64::BI__builtin_arm_ldaex) && 4528 getContext().getTypeSize(E->getType()) == 128) { 4529 Function *F = CGM.getIntrinsic(BuiltinID == AArch64::BI__builtin_arm_ldaex 4530 ? Intrinsic::aarch64_ldaxp 4531 : Intrinsic::aarch64_ldxp); 4532 4533 Value *LdPtr = EmitScalarExpr(E->getArg(0)); 4534 Value *Val = Builder.CreateCall(F, Builder.CreateBitCast(LdPtr, Int8PtrTy), 4535 "ldxp"); 4536 4537 Value *Val0 = Builder.CreateExtractValue(Val, 1); 4538 Value *Val1 = Builder.CreateExtractValue(Val, 0); 4539 llvm::Type *Int128Ty = llvm::IntegerType::get(getLLVMContext(), 128); 4540 Val0 = Builder.CreateZExt(Val0, Int128Ty); 4541 Val1 = Builder.CreateZExt(Val1, Int128Ty); 4542 4543 Value *ShiftCst = llvm::ConstantInt::get(Int128Ty, 64); 4544 Val = Builder.CreateShl(Val0, ShiftCst, "shl", true /* nuw */); 4545 Val = Builder.CreateOr(Val, Val1); 4546 return Builder.CreateBitCast(Val, ConvertType(E->getType())); 4547 } else if (BuiltinID == AArch64::BI__builtin_arm_ldrex || 4548 BuiltinID == AArch64::BI__builtin_arm_ldaex) { 4549 Value *LoadAddr = EmitScalarExpr(E->getArg(0)); 4550 4551 QualType Ty = E->getType(); 4552 llvm::Type *RealResTy = ConvertType(Ty); 4553 llvm::Type *IntResTy = llvm::IntegerType::get(getLLVMContext(), 4554 getContext().getTypeSize(Ty)); 4555 LoadAddr = Builder.CreateBitCast(LoadAddr, IntResTy->getPointerTo()); 4556 4557 Function *F = CGM.getIntrinsic(BuiltinID == AArch64::BI__builtin_arm_ldaex 4558 ? Intrinsic::aarch64_ldaxr 4559 : Intrinsic::aarch64_ldxr, 4560 LoadAddr->getType()); 4561 Value *Val = Builder.CreateCall(F, LoadAddr, "ldxr"); 4562 4563 if (RealResTy->isPointerTy()) 4564 return Builder.CreateIntToPtr(Val, RealResTy); 4565 4566 Val = Builder.CreateTruncOrBitCast(Val, IntResTy); 4567 return Builder.CreateBitCast(Val, RealResTy); 4568 } 4569 4570 if ((BuiltinID == AArch64::BI__builtin_arm_strex || 4571 BuiltinID == AArch64::BI__builtin_arm_stlex) && 4572 getContext().getTypeSize(E->getArg(0)->getType()) == 128) { 4573 Function *F = CGM.getIntrinsic(BuiltinID == AArch64::BI__builtin_arm_stlex 4574 ? Intrinsic::aarch64_stlxp 4575 : Intrinsic::aarch64_stxp); 4576 llvm::Type *STy = llvm::StructType::get(Int64Ty, Int64Ty, nullptr); 4577 4578 Address Tmp = CreateMemTemp(E->getArg(0)->getType()); 4579 EmitAnyExprToMem(E->getArg(0), Tmp, Qualifiers(), /*init*/ true); 4580 4581 Tmp = Builder.CreateBitCast(Tmp, llvm::PointerType::getUnqual(STy)); 4582 llvm::Value *Val = Builder.CreateLoad(Tmp); 4583 4584 Value *Arg0 = Builder.CreateExtractValue(Val, 0); 4585 Value *Arg1 = Builder.CreateExtractValue(Val, 1); 4586 Value *StPtr = Builder.CreateBitCast(EmitScalarExpr(E->getArg(1)), 4587 Int8PtrTy); 4588 return Builder.CreateCall(F, {Arg0, Arg1, StPtr}, "stxp"); 4589 } 4590 4591 if (BuiltinID == AArch64::BI__builtin_arm_strex || 4592 BuiltinID == AArch64::BI__builtin_arm_stlex) { 4593 Value *StoreVal = EmitScalarExpr(E->getArg(0)); 4594 Value *StoreAddr = EmitScalarExpr(E->getArg(1)); 4595 4596 QualType Ty = E->getArg(0)->getType(); 4597 llvm::Type *StoreTy = llvm::IntegerType::get(getLLVMContext(), 4598 getContext().getTypeSize(Ty)); 4599 StoreAddr = Builder.CreateBitCast(StoreAddr, StoreTy->getPointerTo()); 4600 4601 if (StoreVal->getType()->isPointerTy()) 4602 StoreVal = Builder.CreatePtrToInt(StoreVal, Int64Ty); 4603 else { 4604 StoreVal = Builder.CreateBitCast(StoreVal, StoreTy); 4605 StoreVal = Builder.CreateZExtOrBitCast(StoreVal, Int64Ty); 4606 } 4607 4608 Function *F = CGM.getIntrinsic(BuiltinID == AArch64::BI__builtin_arm_stlex 4609 ? Intrinsic::aarch64_stlxr 4610 : Intrinsic::aarch64_stxr, 4611 StoreAddr->getType()); 4612 return Builder.CreateCall(F, {StoreVal, StoreAddr}, "stxr"); 4613 } 4614 4615 if (BuiltinID == AArch64::BI__builtin_arm_clrex) { 4616 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_clrex); 4617 return Builder.CreateCall(F); 4618 } 4619 4620 if (BuiltinID == AArch64::BI__builtin_thread_pointer) { 4621 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_thread_pointer); 4622 return Builder.CreateCall(F); 4623 } 4624 4625 // CRC32 4626 Intrinsic::ID CRCIntrinsicID = Intrinsic::not_intrinsic; 4627 switch (BuiltinID) { 4628 case AArch64::BI__builtin_arm_crc32b: 4629 CRCIntrinsicID = Intrinsic::aarch64_crc32b; break; 4630 case AArch64::BI__builtin_arm_crc32cb: 4631 CRCIntrinsicID = Intrinsic::aarch64_crc32cb; break; 4632 case AArch64::BI__builtin_arm_crc32h: 4633 CRCIntrinsicID = Intrinsic::aarch64_crc32h; break; 4634 case AArch64::BI__builtin_arm_crc32ch: 4635 CRCIntrinsicID = Intrinsic::aarch64_crc32ch; break; 4636 case AArch64::BI__builtin_arm_crc32w: 4637 CRCIntrinsicID = Intrinsic::aarch64_crc32w; break; 4638 case AArch64::BI__builtin_arm_crc32cw: 4639 CRCIntrinsicID = Intrinsic::aarch64_crc32cw; break; 4640 case AArch64::BI__builtin_arm_crc32d: 4641 CRCIntrinsicID = Intrinsic::aarch64_crc32x; break; 4642 case AArch64::BI__builtin_arm_crc32cd: 4643 CRCIntrinsicID = Intrinsic::aarch64_crc32cx; break; 4644 } 4645 4646 if (CRCIntrinsicID != Intrinsic::not_intrinsic) { 4647 Value *Arg0 = EmitScalarExpr(E->getArg(0)); 4648 Value *Arg1 = EmitScalarExpr(E->getArg(1)); 4649 Function *F = CGM.getIntrinsic(CRCIntrinsicID); 4650 4651 llvm::Type *DataTy = F->getFunctionType()->getParamType(1); 4652 Arg1 = Builder.CreateZExtOrBitCast(Arg1, DataTy); 4653 4654 return Builder.CreateCall(F, {Arg0, Arg1}); 4655 } 4656 4657 if (BuiltinID == AArch64::BI__builtin_arm_rsr || 4658 BuiltinID == AArch64::BI__builtin_arm_rsr64 || 4659 BuiltinID == AArch64::BI__builtin_arm_rsrp || 4660 BuiltinID == AArch64::BI__builtin_arm_wsr || 4661 BuiltinID == AArch64::BI__builtin_arm_wsr64 || 4662 BuiltinID == AArch64::BI__builtin_arm_wsrp) { 4663 4664 bool IsRead = BuiltinID == AArch64::BI__builtin_arm_rsr || 4665 BuiltinID == AArch64::BI__builtin_arm_rsr64 || 4666 BuiltinID == AArch64::BI__builtin_arm_rsrp; 4667 4668 bool IsPointerBuiltin = BuiltinID == AArch64::BI__builtin_arm_rsrp || 4669 BuiltinID == AArch64::BI__builtin_arm_wsrp; 4670 4671 bool Is64Bit = BuiltinID != AArch64::BI__builtin_arm_rsr && 4672 BuiltinID != AArch64::BI__builtin_arm_wsr; 4673 4674 llvm::Type *ValueType; 4675 llvm::Type *RegisterType = Int64Ty; 4676 if (IsPointerBuiltin) { 4677 ValueType = VoidPtrTy; 4678 } else if (Is64Bit) { 4679 ValueType = Int64Ty; 4680 } else { 4681 ValueType = Int32Ty; 4682 } 4683 4684 return EmitSpecialRegisterBuiltin(*this, E, RegisterType, ValueType, IsRead); 4685 } 4686 4687 // Find out if any arguments are required to be integer constant 4688 // expressions. 4689 unsigned ICEArguments = 0; 4690 ASTContext::GetBuiltinTypeError Error; 4691 getContext().GetBuiltinType(BuiltinID, Error, &ICEArguments); 4692 assert(Error == ASTContext::GE_None && "Should not codegen an error"); 4693 4694 llvm::SmallVector<Value*, 4> Ops; 4695 for (unsigned i = 0, e = E->getNumArgs() - 1; i != e; i++) { 4696 if ((ICEArguments & (1 << i)) == 0) { 4697 Ops.push_back(EmitScalarExpr(E->getArg(i))); 4698 } else { 4699 // If this is required to be a constant, constant fold it so that we know 4700 // that the generated intrinsic gets a ConstantInt. 4701 llvm::APSInt Result; 4702 bool IsConst = E->getArg(i)->isIntegerConstantExpr(Result, getContext()); 4703 assert(IsConst && "Constant arg isn't actually constant?"); 4704 (void)IsConst; 4705 Ops.push_back(llvm::ConstantInt::get(getLLVMContext(), Result)); 4706 } 4707 } 4708 4709 auto SISDMap = makeArrayRef(AArch64SISDIntrinsicMap); 4710 const NeonIntrinsicInfo *Builtin = findNeonIntrinsicInMap( 4711 SISDMap, BuiltinID, AArch64SISDIntrinsicsProvenSorted); 4712 4713 if (Builtin) { 4714 Ops.push_back(EmitScalarExpr(E->getArg(E->getNumArgs() - 1))); 4715 Value *Result = EmitCommonNeonSISDBuiltinExpr(*this, *Builtin, Ops, E); 4716 assert(Result && "SISD intrinsic should have been handled"); 4717 return Result; 4718 } 4719 4720 llvm::APSInt Result; 4721 const Expr *Arg = E->getArg(E->getNumArgs()-1); 4722 NeonTypeFlags Type(0); 4723 if (Arg->isIntegerConstantExpr(Result, getContext())) 4724 // Determine the type of this overloaded NEON intrinsic. 4725 Type = NeonTypeFlags(Result.getZExtValue()); 4726 4727 bool usgn = Type.isUnsigned(); 4728 bool quad = Type.isQuad(); 4729 4730 // Handle non-overloaded intrinsics first. 4731 switch (BuiltinID) { 4732 default: break; 4733 case NEON::BI__builtin_neon_vldrq_p128: { 4734 llvm::Type *Int128PTy = llvm::Type::getIntNPtrTy(getLLVMContext(), 128); 4735 Value *Ptr = Builder.CreateBitCast(EmitScalarExpr(E->getArg(0)), Int128PTy); 4736 return Builder.CreateDefaultAlignedLoad(Ptr); 4737 } 4738 case NEON::BI__builtin_neon_vstrq_p128: { 4739 llvm::Type *Int128PTy = llvm::Type::getIntNPtrTy(getLLVMContext(), 128); 4740 Value *Ptr = Builder.CreateBitCast(Ops[0], Int128PTy); 4741 return Builder.CreateDefaultAlignedStore(EmitScalarExpr(E->getArg(1)), Ptr); 4742 } 4743 case NEON::BI__builtin_neon_vcvts_u32_f32: 4744 case NEON::BI__builtin_neon_vcvtd_u64_f64: 4745 usgn = true; 4746 // FALL THROUGH 4747 case NEON::BI__builtin_neon_vcvts_s32_f32: 4748 case NEON::BI__builtin_neon_vcvtd_s64_f64: { 4749 Ops.push_back(EmitScalarExpr(E->getArg(0))); 4750 bool Is64 = Ops[0]->getType()->getPrimitiveSizeInBits() == 64; 4751 llvm::Type *InTy = Is64 ? Int64Ty : Int32Ty; 4752 llvm::Type *FTy = Is64 ? DoubleTy : FloatTy; 4753 Ops[0] = Builder.CreateBitCast(Ops[0], FTy); 4754 if (usgn) 4755 return Builder.CreateFPToUI(Ops[0], InTy); 4756 return Builder.CreateFPToSI(Ops[0], InTy); 4757 } 4758 case NEON::BI__builtin_neon_vcvts_f32_u32: 4759 case NEON::BI__builtin_neon_vcvtd_f64_u64: 4760 usgn = true; 4761 // FALL THROUGH 4762 case NEON::BI__builtin_neon_vcvts_f32_s32: 4763 case NEON::BI__builtin_neon_vcvtd_f64_s64: { 4764 Ops.push_back(EmitScalarExpr(E->getArg(0))); 4765 bool Is64 = Ops[0]->getType()->getPrimitiveSizeInBits() == 64; 4766 llvm::Type *InTy = Is64 ? Int64Ty : Int32Ty; 4767 llvm::Type *FTy = Is64 ? DoubleTy : FloatTy; 4768 Ops[0] = Builder.CreateBitCast(Ops[0], InTy); 4769 if (usgn) 4770 return Builder.CreateUIToFP(Ops[0], FTy); 4771 return Builder.CreateSIToFP(Ops[0], FTy); 4772 } 4773 case NEON::BI__builtin_neon_vpaddd_s64: { 4774 llvm::Type *Ty = llvm::VectorType::get(Int64Ty, 2); 4775 Value *Vec = EmitScalarExpr(E->getArg(0)); 4776 // The vector is v2f64, so make sure it's bitcast to that. 4777 Vec = Builder.CreateBitCast(Vec, Ty, "v2i64"); 4778 llvm::Value *Idx0 = llvm::ConstantInt::get(SizeTy, 0); 4779 llvm::Value *Idx1 = llvm::ConstantInt::get(SizeTy, 1); 4780 Value *Op0 = Builder.CreateExtractElement(Vec, Idx0, "lane0"); 4781 Value *Op1 = Builder.CreateExtractElement(Vec, Idx1, "lane1"); 4782 // Pairwise addition of a v2f64 into a scalar f64. 4783 return Builder.CreateAdd(Op0, Op1, "vpaddd"); 4784 } 4785 case NEON::BI__builtin_neon_vpaddd_f64: { 4786 llvm::Type *Ty = 4787 llvm::VectorType::get(DoubleTy, 2); 4788 Value *Vec = EmitScalarExpr(E->getArg(0)); 4789 // The vector is v2f64, so make sure it's bitcast to that. 4790 Vec = Builder.CreateBitCast(Vec, Ty, "v2f64"); 4791 llvm::Value *Idx0 = llvm::ConstantInt::get(SizeTy, 0); 4792 llvm::Value *Idx1 = llvm::ConstantInt::get(SizeTy, 1); 4793 Value *Op0 = Builder.CreateExtractElement(Vec, Idx0, "lane0"); 4794 Value *Op1 = Builder.CreateExtractElement(Vec, Idx1, "lane1"); 4795 // Pairwise addition of a v2f64 into a scalar f64. 4796 return Builder.CreateFAdd(Op0, Op1, "vpaddd"); 4797 } 4798 case NEON::BI__builtin_neon_vpadds_f32: { 4799 llvm::Type *Ty = 4800 llvm::VectorType::get(FloatTy, 2); 4801 Value *Vec = EmitScalarExpr(E->getArg(0)); 4802 // The vector is v2f32, so make sure it's bitcast to that. 4803 Vec = Builder.CreateBitCast(Vec, Ty, "v2f32"); 4804 llvm::Value *Idx0 = llvm::ConstantInt::get(SizeTy, 0); 4805 llvm::Value *Idx1 = llvm::ConstantInt::get(SizeTy, 1); 4806 Value *Op0 = Builder.CreateExtractElement(Vec, Idx0, "lane0"); 4807 Value *Op1 = Builder.CreateExtractElement(Vec, Idx1, "lane1"); 4808 // Pairwise addition of a v2f32 into a scalar f32. 4809 return Builder.CreateFAdd(Op0, Op1, "vpaddd"); 4810 } 4811 case NEON::BI__builtin_neon_vceqzd_s64: 4812 case NEON::BI__builtin_neon_vceqzd_f64: 4813 case NEON::BI__builtin_neon_vceqzs_f32: 4814 Ops.push_back(EmitScalarExpr(E->getArg(0))); 4815 return EmitAArch64CompareBuiltinExpr( 4816 Ops[0], ConvertType(E->getCallReturnType(getContext())), 4817 ICmpInst::FCMP_OEQ, ICmpInst::ICMP_EQ, "vceqz"); 4818 case NEON::BI__builtin_neon_vcgezd_s64: 4819 case NEON::BI__builtin_neon_vcgezd_f64: 4820 case NEON::BI__builtin_neon_vcgezs_f32: 4821 Ops.push_back(EmitScalarExpr(E->getArg(0))); 4822 return EmitAArch64CompareBuiltinExpr( 4823 Ops[0], ConvertType(E->getCallReturnType(getContext())), 4824 ICmpInst::FCMP_OGE, ICmpInst::ICMP_SGE, "vcgez"); 4825 case NEON::BI__builtin_neon_vclezd_s64: 4826 case NEON::BI__builtin_neon_vclezd_f64: 4827 case NEON::BI__builtin_neon_vclezs_f32: 4828 Ops.push_back(EmitScalarExpr(E->getArg(0))); 4829 return EmitAArch64CompareBuiltinExpr( 4830 Ops[0], ConvertType(E->getCallReturnType(getContext())), 4831 ICmpInst::FCMP_OLE, ICmpInst::ICMP_SLE, "vclez"); 4832 case NEON::BI__builtin_neon_vcgtzd_s64: 4833 case NEON::BI__builtin_neon_vcgtzd_f64: 4834 case NEON::BI__builtin_neon_vcgtzs_f32: 4835 Ops.push_back(EmitScalarExpr(E->getArg(0))); 4836 return EmitAArch64CompareBuiltinExpr( 4837 Ops[0], ConvertType(E->getCallReturnType(getContext())), 4838 ICmpInst::FCMP_OGT, ICmpInst::ICMP_SGT, "vcgtz"); 4839 case NEON::BI__builtin_neon_vcltzd_s64: 4840 case NEON::BI__builtin_neon_vcltzd_f64: 4841 case NEON::BI__builtin_neon_vcltzs_f32: 4842 Ops.push_back(EmitScalarExpr(E->getArg(0))); 4843 return EmitAArch64CompareBuiltinExpr( 4844 Ops[0], ConvertType(E->getCallReturnType(getContext())), 4845 ICmpInst::FCMP_OLT, ICmpInst::ICMP_SLT, "vcltz"); 4846 4847 case NEON::BI__builtin_neon_vceqzd_u64: { 4848 Ops.push_back(EmitScalarExpr(E->getArg(0))); 4849 Ops[0] = Builder.CreateBitCast(Ops[0], Int64Ty); 4850 Ops[0] = 4851 Builder.CreateICmpEQ(Ops[0], llvm::Constant::getNullValue(Int64Ty)); 4852 return Builder.CreateSExt(Ops[0], Int64Ty, "vceqzd"); 4853 } 4854 case NEON::BI__builtin_neon_vceqd_f64: 4855 case NEON::BI__builtin_neon_vcled_f64: 4856 case NEON::BI__builtin_neon_vcltd_f64: 4857 case NEON::BI__builtin_neon_vcged_f64: 4858 case NEON::BI__builtin_neon_vcgtd_f64: { 4859 llvm::CmpInst::Predicate P; 4860 switch (BuiltinID) { 4861 default: llvm_unreachable("missing builtin ID in switch!"); 4862 case NEON::BI__builtin_neon_vceqd_f64: P = llvm::FCmpInst::FCMP_OEQ; break; 4863 case NEON::BI__builtin_neon_vcled_f64: P = llvm::FCmpInst::FCMP_OLE; break; 4864 case NEON::BI__builtin_neon_vcltd_f64: P = llvm::FCmpInst::FCMP_OLT; break; 4865 case NEON::BI__builtin_neon_vcged_f64: P = llvm::FCmpInst::FCMP_OGE; break; 4866 case NEON::BI__builtin_neon_vcgtd_f64: P = llvm::FCmpInst::FCMP_OGT; break; 4867 } 4868 Ops.push_back(EmitScalarExpr(E->getArg(1))); 4869 Ops[0] = Builder.CreateBitCast(Ops[0], DoubleTy); 4870 Ops[1] = Builder.CreateBitCast(Ops[1], DoubleTy); 4871 Ops[0] = Builder.CreateFCmp(P, Ops[0], Ops[1]); 4872 return Builder.CreateSExt(Ops[0], Int64Ty, "vcmpd"); 4873 } 4874 case NEON::BI__builtin_neon_vceqs_f32: 4875 case NEON::BI__builtin_neon_vcles_f32: 4876 case NEON::BI__builtin_neon_vclts_f32: 4877 case NEON::BI__builtin_neon_vcges_f32: 4878 case NEON::BI__builtin_neon_vcgts_f32: { 4879 llvm::CmpInst::Predicate P; 4880 switch (BuiltinID) { 4881 default: llvm_unreachable("missing builtin ID in switch!"); 4882 case NEON::BI__builtin_neon_vceqs_f32: P = llvm::FCmpInst::FCMP_OEQ; break; 4883 case NEON::BI__builtin_neon_vcles_f32: P = llvm::FCmpInst::FCMP_OLE; break; 4884 case NEON::BI__builtin_neon_vclts_f32: P = llvm::FCmpInst::FCMP_OLT; break; 4885 case NEON::BI__builtin_neon_vcges_f32: P = llvm::FCmpInst::FCMP_OGE; break; 4886 case NEON::BI__builtin_neon_vcgts_f32: P = llvm::FCmpInst::FCMP_OGT; break; 4887 } 4888 Ops.push_back(EmitScalarExpr(E->getArg(1))); 4889 Ops[0] = Builder.CreateBitCast(Ops[0], FloatTy); 4890 Ops[1] = Builder.CreateBitCast(Ops[1], FloatTy); 4891 Ops[0] = Builder.CreateFCmp(P, Ops[0], Ops[1]); 4892 return Builder.CreateSExt(Ops[0], Int32Ty, "vcmpd"); 4893 } 4894 case NEON::BI__builtin_neon_vceqd_s64: 4895 case NEON::BI__builtin_neon_vceqd_u64: 4896 case NEON::BI__builtin_neon_vcgtd_s64: 4897 case NEON::BI__builtin_neon_vcgtd_u64: 4898 case NEON::BI__builtin_neon_vcltd_s64: 4899 case NEON::BI__builtin_neon_vcltd_u64: 4900 case NEON::BI__builtin_neon_vcged_u64: 4901 case NEON::BI__builtin_neon_vcged_s64: 4902 case NEON::BI__builtin_neon_vcled_u64: 4903 case NEON::BI__builtin_neon_vcled_s64: { 4904 llvm::CmpInst::Predicate P; 4905 switch (BuiltinID) { 4906 default: llvm_unreachable("missing builtin ID in switch!"); 4907 case NEON::BI__builtin_neon_vceqd_s64: 4908 case NEON::BI__builtin_neon_vceqd_u64:P = llvm::ICmpInst::ICMP_EQ;break; 4909 case NEON::BI__builtin_neon_vcgtd_s64:P = llvm::ICmpInst::ICMP_SGT;break; 4910 case NEON::BI__builtin_neon_vcgtd_u64:P = llvm::ICmpInst::ICMP_UGT;break; 4911 case NEON::BI__builtin_neon_vcltd_s64:P = llvm::ICmpInst::ICMP_SLT;break; 4912 case NEON::BI__builtin_neon_vcltd_u64:P = llvm::ICmpInst::ICMP_ULT;break; 4913 case NEON::BI__builtin_neon_vcged_u64:P = llvm::ICmpInst::ICMP_UGE;break; 4914 case NEON::BI__builtin_neon_vcged_s64:P = llvm::ICmpInst::ICMP_SGE;break; 4915 case NEON::BI__builtin_neon_vcled_u64:P = llvm::ICmpInst::ICMP_ULE;break; 4916 case NEON::BI__builtin_neon_vcled_s64:P = llvm::ICmpInst::ICMP_SLE;break; 4917 } 4918 Ops.push_back(EmitScalarExpr(E->getArg(1))); 4919 Ops[0] = Builder.CreateBitCast(Ops[0], Int64Ty); 4920 Ops[1] = Builder.CreateBitCast(Ops[1], Int64Ty); 4921 Ops[0] = Builder.CreateICmp(P, Ops[0], Ops[1]); 4922 return Builder.CreateSExt(Ops[0], Int64Ty, "vceqd"); 4923 } 4924 case NEON::BI__builtin_neon_vtstd_s64: 4925 case NEON::BI__builtin_neon_vtstd_u64: { 4926 Ops.push_back(EmitScalarExpr(E->getArg(1))); 4927 Ops[0] = Builder.CreateBitCast(Ops[0], Int64Ty); 4928 Ops[1] = Builder.CreateBitCast(Ops[1], Int64Ty); 4929 Ops[0] = Builder.CreateAnd(Ops[0], Ops[1]); 4930 Ops[0] = Builder.CreateICmp(ICmpInst::ICMP_NE, Ops[0], 4931 llvm::Constant::getNullValue(Int64Ty)); 4932 return Builder.CreateSExt(Ops[0], Int64Ty, "vtstd"); 4933 } 4934 case NEON::BI__builtin_neon_vset_lane_i8: 4935 case NEON::BI__builtin_neon_vset_lane_i16: 4936 case NEON::BI__builtin_neon_vset_lane_i32: 4937 case NEON::BI__builtin_neon_vset_lane_i64: 4938 case NEON::BI__builtin_neon_vset_lane_f32: 4939 case NEON::BI__builtin_neon_vsetq_lane_i8: 4940 case NEON::BI__builtin_neon_vsetq_lane_i16: 4941 case NEON::BI__builtin_neon_vsetq_lane_i32: 4942 case NEON::BI__builtin_neon_vsetq_lane_i64: 4943 case NEON::BI__builtin_neon_vsetq_lane_f32: 4944 Ops.push_back(EmitScalarExpr(E->getArg(2))); 4945 return Builder.CreateInsertElement(Ops[1], Ops[0], Ops[2], "vset_lane"); 4946 case NEON::BI__builtin_neon_vset_lane_f64: 4947 // The vector type needs a cast for the v1f64 variant. 4948 Ops[1] = Builder.CreateBitCast(Ops[1], 4949 llvm::VectorType::get(DoubleTy, 1)); 4950 Ops.push_back(EmitScalarExpr(E->getArg(2))); 4951 return Builder.CreateInsertElement(Ops[1], Ops[0], Ops[2], "vset_lane"); 4952 case NEON::BI__builtin_neon_vsetq_lane_f64: 4953 // The vector type needs a cast for the v2f64 variant. 4954 Ops[1] = Builder.CreateBitCast(Ops[1], 4955 llvm::VectorType::get(DoubleTy, 2)); 4956 Ops.push_back(EmitScalarExpr(E->getArg(2))); 4957 return Builder.CreateInsertElement(Ops[1], Ops[0], Ops[2], "vset_lane"); 4958 4959 case NEON::BI__builtin_neon_vget_lane_i8: 4960 case NEON::BI__builtin_neon_vdupb_lane_i8: 4961 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::VectorType::get(Int8Ty, 8)); 4962 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4963 "vget_lane"); 4964 case NEON::BI__builtin_neon_vgetq_lane_i8: 4965 case NEON::BI__builtin_neon_vdupb_laneq_i8: 4966 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::VectorType::get(Int8Ty, 16)); 4967 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4968 "vgetq_lane"); 4969 case NEON::BI__builtin_neon_vget_lane_i16: 4970 case NEON::BI__builtin_neon_vduph_lane_i16: 4971 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::VectorType::get(Int16Ty, 4)); 4972 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4973 "vget_lane"); 4974 case NEON::BI__builtin_neon_vgetq_lane_i16: 4975 case NEON::BI__builtin_neon_vduph_laneq_i16: 4976 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::VectorType::get(Int16Ty, 8)); 4977 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4978 "vgetq_lane"); 4979 case NEON::BI__builtin_neon_vget_lane_i32: 4980 case NEON::BI__builtin_neon_vdups_lane_i32: 4981 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::VectorType::get(Int32Ty, 2)); 4982 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4983 "vget_lane"); 4984 case NEON::BI__builtin_neon_vdups_lane_f32: 4985 Ops[0] = Builder.CreateBitCast(Ops[0], 4986 llvm::VectorType::get(FloatTy, 2)); 4987 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4988 "vdups_lane"); 4989 case NEON::BI__builtin_neon_vgetq_lane_i32: 4990 case NEON::BI__builtin_neon_vdups_laneq_i32: 4991 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::VectorType::get(Int32Ty, 4)); 4992 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4993 "vgetq_lane"); 4994 case NEON::BI__builtin_neon_vget_lane_i64: 4995 case NEON::BI__builtin_neon_vdupd_lane_i64: 4996 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::VectorType::get(Int64Ty, 1)); 4997 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4998 "vget_lane"); 4999 case NEON::BI__builtin_neon_vdupd_lane_f64: 5000 Ops[0] = Builder.CreateBitCast(Ops[0], 5001 llvm::VectorType::get(DoubleTy, 1)); 5002 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 5003 "vdupd_lane"); 5004 case NEON::BI__builtin_neon_vgetq_lane_i64: 5005 case NEON::BI__builtin_neon_vdupd_laneq_i64: 5006 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::VectorType::get(Int64Ty, 2)); 5007 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 5008 "vgetq_lane"); 5009 case NEON::BI__builtin_neon_vget_lane_f32: 5010 Ops[0] = Builder.CreateBitCast(Ops[0], 5011 llvm::VectorType::get(FloatTy, 2)); 5012 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 5013 "vget_lane"); 5014 case NEON::BI__builtin_neon_vget_lane_f64: 5015 Ops[0] = Builder.CreateBitCast(Ops[0], 5016 llvm::VectorType::get(DoubleTy, 1)); 5017 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 5018 "vget_lane"); 5019 case NEON::BI__builtin_neon_vgetq_lane_f32: 5020 case NEON::BI__builtin_neon_vdups_laneq_f32: 5021 Ops[0] = Builder.CreateBitCast(Ops[0], 5022 llvm::VectorType::get(FloatTy, 4)); 5023 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 5024 "vgetq_lane"); 5025 case NEON::BI__builtin_neon_vgetq_lane_f64: 5026 case NEON::BI__builtin_neon_vdupd_laneq_f64: 5027 Ops[0] = Builder.CreateBitCast(Ops[0], 5028 llvm::VectorType::get(DoubleTy, 2)); 5029 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 5030 "vgetq_lane"); 5031 case NEON::BI__builtin_neon_vaddd_s64: 5032 case NEON::BI__builtin_neon_vaddd_u64: 5033 return Builder.CreateAdd(Ops[0], EmitScalarExpr(E->getArg(1)), "vaddd"); 5034 case NEON::BI__builtin_neon_vsubd_s64: 5035 case NEON::BI__builtin_neon_vsubd_u64: 5036 return Builder.CreateSub(Ops[0], EmitScalarExpr(E->getArg(1)), "vsubd"); 5037 case NEON::BI__builtin_neon_vqdmlalh_s16: 5038 case NEON::BI__builtin_neon_vqdmlslh_s16: { 5039 SmallVector<Value *, 2> ProductOps; 5040 ProductOps.push_back(vectorWrapScalar16(Ops[1])); 5041 ProductOps.push_back(vectorWrapScalar16(EmitScalarExpr(E->getArg(2)))); 5042 llvm::Type *VTy = llvm::VectorType::get(Int32Ty, 4); 5043 Ops[1] = EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_sqdmull, VTy), 5044 ProductOps, "vqdmlXl"); 5045 Constant *CI = ConstantInt::get(SizeTy, 0); 5046 Ops[1] = Builder.CreateExtractElement(Ops[1], CI, "lane0"); 5047 5048 unsigned AccumInt = BuiltinID == NEON::BI__builtin_neon_vqdmlalh_s16 5049 ? Intrinsic::aarch64_neon_sqadd 5050 : Intrinsic::aarch64_neon_sqsub; 5051 return EmitNeonCall(CGM.getIntrinsic(AccumInt, Int32Ty), Ops, "vqdmlXl"); 5052 } 5053 case NEON::BI__builtin_neon_vqshlud_n_s64: { 5054 Ops.push_back(EmitScalarExpr(E->getArg(1))); 5055 Ops[1] = Builder.CreateZExt(Ops[1], Int64Ty); 5056 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_sqshlu, Int64Ty), 5057 Ops, "vqshlu_n"); 5058 } 5059 case NEON::BI__builtin_neon_vqshld_n_u64: 5060 case NEON::BI__builtin_neon_vqshld_n_s64: { 5061 unsigned Int = BuiltinID == NEON::BI__builtin_neon_vqshld_n_u64 5062 ? Intrinsic::aarch64_neon_uqshl 5063 : Intrinsic::aarch64_neon_sqshl; 5064 Ops.push_back(EmitScalarExpr(E->getArg(1))); 5065 Ops[1] = Builder.CreateZExt(Ops[1], Int64Ty); 5066 return EmitNeonCall(CGM.getIntrinsic(Int, Int64Ty), Ops, "vqshl_n"); 5067 } 5068 case NEON::BI__builtin_neon_vrshrd_n_u64: 5069 case NEON::BI__builtin_neon_vrshrd_n_s64: { 5070 unsigned Int = BuiltinID == NEON::BI__builtin_neon_vrshrd_n_u64 5071 ? Intrinsic::aarch64_neon_urshl 5072 : Intrinsic::aarch64_neon_srshl; 5073 Ops.push_back(EmitScalarExpr(E->getArg(1))); 5074 int SV = cast<ConstantInt>(Ops[1])->getSExtValue(); 5075 Ops[1] = ConstantInt::get(Int64Ty, -SV); 5076 return EmitNeonCall(CGM.getIntrinsic(Int, Int64Ty), Ops, "vrshr_n"); 5077 } 5078 case NEON::BI__builtin_neon_vrsrad_n_u64: 5079 case NEON::BI__builtin_neon_vrsrad_n_s64: { 5080 unsigned Int = BuiltinID == NEON::BI__builtin_neon_vrsrad_n_u64 5081 ? Intrinsic::aarch64_neon_urshl 5082 : Intrinsic::aarch64_neon_srshl; 5083 Ops[1] = Builder.CreateBitCast(Ops[1], Int64Ty); 5084 Ops.push_back(Builder.CreateNeg(EmitScalarExpr(E->getArg(2)))); 5085 Ops[1] = Builder.CreateCall(CGM.getIntrinsic(Int, Int64Ty), 5086 {Ops[1], Builder.CreateSExt(Ops[2], Int64Ty)}); 5087 return Builder.CreateAdd(Ops[0], Builder.CreateBitCast(Ops[1], Int64Ty)); 5088 } 5089 case NEON::BI__builtin_neon_vshld_n_s64: 5090 case NEON::BI__builtin_neon_vshld_n_u64: { 5091 llvm::ConstantInt *Amt = cast<ConstantInt>(EmitScalarExpr(E->getArg(1))); 5092 return Builder.CreateShl( 5093 Ops[0], ConstantInt::get(Int64Ty, Amt->getZExtValue()), "shld_n"); 5094 } 5095 case NEON::BI__builtin_neon_vshrd_n_s64: { 5096 llvm::ConstantInt *Amt = cast<ConstantInt>(EmitScalarExpr(E->getArg(1))); 5097 return Builder.CreateAShr( 5098 Ops[0], ConstantInt::get(Int64Ty, std::min(static_cast<uint64_t>(63), 5099 Amt->getZExtValue())), 5100 "shrd_n"); 5101 } 5102 case NEON::BI__builtin_neon_vshrd_n_u64: { 5103 llvm::ConstantInt *Amt = cast<ConstantInt>(EmitScalarExpr(E->getArg(1))); 5104 uint64_t ShiftAmt = Amt->getZExtValue(); 5105 // Right-shifting an unsigned value by its size yields 0. 5106 if (ShiftAmt == 64) 5107 return ConstantInt::get(Int64Ty, 0); 5108 return Builder.CreateLShr(Ops[0], ConstantInt::get(Int64Ty, ShiftAmt), 5109 "shrd_n"); 5110 } 5111 case NEON::BI__builtin_neon_vsrad_n_s64: { 5112 llvm::ConstantInt *Amt = cast<ConstantInt>(EmitScalarExpr(E->getArg(2))); 5113 Ops[1] = Builder.CreateAShr( 5114 Ops[1], ConstantInt::get(Int64Ty, std::min(static_cast<uint64_t>(63), 5115 Amt->getZExtValue())), 5116 "shrd_n"); 5117 return Builder.CreateAdd(Ops[0], Ops[1]); 5118 } 5119 case NEON::BI__builtin_neon_vsrad_n_u64: { 5120 llvm::ConstantInt *Amt = cast<ConstantInt>(EmitScalarExpr(E->getArg(2))); 5121 uint64_t ShiftAmt = Amt->getZExtValue(); 5122 // Right-shifting an unsigned value by its size yields 0. 5123 // As Op + 0 = Op, return Ops[0] directly. 5124 if (ShiftAmt == 64) 5125 return Ops[0]; 5126 Ops[1] = Builder.CreateLShr(Ops[1], ConstantInt::get(Int64Ty, ShiftAmt), 5127 "shrd_n"); 5128 return Builder.CreateAdd(Ops[0], Ops[1]); 5129 } 5130 case NEON::BI__builtin_neon_vqdmlalh_lane_s16: 5131 case NEON::BI__builtin_neon_vqdmlalh_laneq_s16: 5132 case NEON::BI__builtin_neon_vqdmlslh_lane_s16: 5133 case NEON::BI__builtin_neon_vqdmlslh_laneq_s16: { 5134 Ops[2] = Builder.CreateExtractElement(Ops[2], EmitScalarExpr(E->getArg(3)), 5135 "lane"); 5136 SmallVector<Value *, 2> ProductOps; 5137 ProductOps.push_back(vectorWrapScalar16(Ops[1])); 5138 ProductOps.push_back(vectorWrapScalar16(Ops[2])); 5139 llvm::Type *VTy = llvm::VectorType::get(Int32Ty, 4); 5140 Ops[1] = EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_sqdmull, VTy), 5141 ProductOps, "vqdmlXl"); 5142 Constant *CI = ConstantInt::get(SizeTy, 0); 5143 Ops[1] = Builder.CreateExtractElement(Ops[1], CI, "lane0"); 5144 Ops.pop_back(); 5145 5146 unsigned AccInt = (BuiltinID == NEON::BI__builtin_neon_vqdmlalh_lane_s16 || 5147 BuiltinID == NEON::BI__builtin_neon_vqdmlalh_laneq_s16) 5148 ? Intrinsic::aarch64_neon_sqadd 5149 : Intrinsic::aarch64_neon_sqsub; 5150 return EmitNeonCall(CGM.getIntrinsic(AccInt, Int32Ty), Ops, "vqdmlXl"); 5151 } 5152 case NEON::BI__builtin_neon_vqdmlals_s32: 5153 case NEON::BI__builtin_neon_vqdmlsls_s32: { 5154 SmallVector<Value *, 2> ProductOps; 5155 ProductOps.push_back(Ops[1]); 5156 ProductOps.push_back(EmitScalarExpr(E->getArg(2))); 5157 Ops[1] = 5158 EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_sqdmulls_scalar), 5159 ProductOps, "vqdmlXl"); 5160 5161 unsigned AccumInt = BuiltinID == NEON::BI__builtin_neon_vqdmlals_s32 5162 ? Intrinsic::aarch64_neon_sqadd 5163 : Intrinsic::aarch64_neon_sqsub; 5164 return EmitNeonCall(CGM.getIntrinsic(AccumInt, Int64Ty), Ops, "vqdmlXl"); 5165 } 5166 case NEON::BI__builtin_neon_vqdmlals_lane_s32: 5167 case NEON::BI__builtin_neon_vqdmlals_laneq_s32: 5168 case NEON::BI__builtin_neon_vqdmlsls_lane_s32: 5169 case NEON::BI__builtin_neon_vqdmlsls_laneq_s32: { 5170 Ops[2] = Builder.CreateExtractElement(Ops[2], EmitScalarExpr(E->getArg(3)), 5171 "lane"); 5172 SmallVector<Value *, 2> ProductOps; 5173 ProductOps.push_back(Ops[1]); 5174 ProductOps.push_back(Ops[2]); 5175 Ops[1] = 5176 EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_sqdmulls_scalar), 5177 ProductOps, "vqdmlXl"); 5178 Ops.pop_back(); 5179 5180 unsigned AccInt = (BuiltinID == NEON::BI__builtin_neon_vqdmlals_lane_s32 || 5181 BuiltinID == NEON::BI__builtin_neon_vqdmlals_laneq_s32) 5182 ? Intrinsic::aarch64_neon_sqadd 5183 : Intrinsic::aarch64_neon_sqsub; 5184 return EmitNeonCall(CGM.getIntrinsic(AccInt, Int64Ty), Ops, "vqdmlXl"); 5185 } 5186 } 5187 5188 llvm::VectorType *VTy = GetNeonType(this, Type); 5189 llvm::Type *Ty = VTy; 5190 if (!Ty) 5191 return nullptr; 5192 5193 // Not all intrinsics handled by the common case work for AArch64 yet, so only 5194 // defer to common code if it's been added to our special map. 5195 Builtin = findNeonIntrinsicInMap(AArch64SIMDIntrinsicMap, BuiltinID, 5196 AArch64SIMDIntrinsicsProvenSorted); 5197 5198 if (Builtin) 5199 return EmitCommonNeonBuiltinExpr( 5200 Builtin->BuiltinID, Builtin->LLVMIntrinsic, Builtin->AltLLVMIntrinsic, 5201 Builtin->NameHint, Builtin->TypeModifier, E, Ops, 5202 /*never use addresses*/ Address::invalid(), Address::invalid()); 5203 5204 if (Value *V = EmitAArch64TblBuiltinExpr(*this, BuiltinID, E, Ops)) 5205 return V; 5206 5207 unsigned Int; 5208 switch (BuiltinID) { 5209 default: return nullptr; 5210 case NEON::BI__builtin_neon_vbsl_v: 5211 case NEON::BI__builtin_neon_vbslq_v: { 5212 llvm::Type *BitTy = llvm::VectorType::getInteger(VTy); 5213 Ops[0] = Builder.CreateBitCast(Ops[0], BitTy, "vbsl"); 5214 Ops[1] = Builder.CreateBitCast(Ops[1], BitTy, "vbsl"); 5215 Ops[2] = Builder.CreateBitCast(Ops[2], BitTy, "vbsl"); 5216 5217 Ops[1] = Builder.CreateAnd(Ops[0], Ops[1], "vbsl"); 5218 Ops[2] = Builder.CreateAnd(Builder.CreateNot(Ops[0]), Ops[2], "vbsl"); 5219 Ops[0] = Builder.CreateOr(Ops[1], Ops[2], "vbsl"); 5220 return Builder.CreateBitCast(Ops[0], Ty); 5221 } 5222 case NEON::BI__builtin_neon_vfma_lane_v: 5223 case NEON::BI__builtin_neon_vfmaq_lane_v: { // Only used for FP types 5224 // The ARM builtins (and instructions) have the addend as the first 5225 // operand, but the 'fma' intrinsics have it last. Swap it around here. 5226 Value *Addend = Ops[0]; 5227 Value *Multiplicand = Ops[1]; 5228 Value *LaneSource = Ops[2]; 5229 Ops[0] = Multiplicand; 5230 Ops[1] = LaneSource; 5231 Ops[2] = Addend; 5232 5233 // Now adjust things to handle the lane access. 5234 llvm::Type *SourceTy = BuiltinID == NEON::BI__builtin_neon_vfmaq_lane_v ? 5235 llvm::VectorType::get(VTy->getElementType(), VTy->getNumElements() / 2) : 5236 VTy; 5237 llvm::Constant *cst = cast<Constant>(Ops[3]); 5238 Value *SV = llvm::ConstantVector::getSplat(VTy->getNumElements(), cst); 5239 Ops[1] = Builder.CreateBitCast(Ops[1], SourceTy); 5240 Ops[1] = Builder.CreateShuffleVector(Ops[1], Ops[1], SV, "lane"); 5241 5242 Ops.pop_back(); 5243 Int = Intrinsic::fma; 5244 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "fmla"); 5245 } 5246 case NEON::BI__builtin_neon_vfma_laneq_v: { 5247 llvm::VectorType *VTy = cast<llvm::VectorType>(Ty); 5248 // v1f64 fma should be mapped to Neon scalar f64 fma 5249 if (VTy && VTy->getElementType() == DoubleTy) { 5250 Ops[0] = Builder.CreateBitCast(Ops[0], DoubleTy); 5251 Ops[1] = Builder.CreateBitCast(Ops[1], DoubleTy); 5252 llvm::Type *VTy = GetNeonType(this, 5253 NeonTypeFlags(NeonTypeFlags::Float64, false, true)); 5254 Ops[2] = Builder.CreateBitCast(Ops[2], VTy); 5255 Ops[2] = Builder.CreateExtractElement(Ops[2], Ops[3], "extract"); 5256 Value *F = CGM.getIntrinsic(Intrinsic::fma, DoubleTy); 5257 Value *Result = Builder.CreateCall(F, {Ops[1], Ops[2], Ops[0]}); 5258 return Builder.CreateBitCast(Result, Ty); 5259 } 5260 Value *F = CGM.getIntrinsic(Intrinsic::fma, Ty); 5261 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 5262 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 5263 5264 llvm::Type *STy = llvm::VectorType::get(VTy->getElementType(), 5265 VTy->getNumElements() * 2); 5266 Ops[2] = Builder.CreateBitCast(Ops[2], STy); 5267 Value* SV = llvm::ConstantVector::getSplat(VTy->getNumElements(), 5268 cast<ConstantInt>(Ops[3])); 5269 Ops[2] = Builder.CreateShuffleVector(Ops[2], Ops[2], SV, "lane"); 5270 5271 return Builder.CreateCall(F, {Ops[2], Ops[1], Ops[0]}); 5272 } 5273 case NEON::BI__builtin_neon_vfmaq_laneq_v: { 5274 Value *F = CGM.getIntrinsic(Intrinsic::fma, Ty); 5275 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 5276 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 5277 5278 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 5279 Ops[2] = EmitNeonSplat(Ops[2], cast<ConstantInt>(Ops[3])); 5280 return Builder.CreateCall(F, {Ops[2], Ops[1], Ops[0]}); 5281 } 5282 case NEON::BI__builtin_neon_vfmas_lane_f32: 5283 case NEON::BI__builtin_neon_vfmas_laneq_f32: 5284 case NEON::BI__builtin_neon_vfmad_lane_f64: 5285 case NEON::BI__builtin_neon_vfmad_laneq_f64: { 5286 Ops.push_back(EmitScalarExpr(E->getArg(3))); 5287 llvm::Type *Ty = ConvertType(E->getCallReturnType(getContext())); 5288 Value *F = CGM.getIntrinsic(Intrinsic::fma, Ty); 5289 Ops[2] = Builder.CreateExtractElement(Ops[2], Ops[3], "extract"); 5290 return Builder.CreateCall(F, {Ops[1], Ops[2], Ops[0]}); 5291 } 5292 case NEON::BI__builtin_neon_vfms_v: 5293 case NEON::BI__builtin_neon_vfmsq_v: { // Only used for FP types 5294 // FIXME: probably remove when we no longer support aarch64_simd.h 5295 // (arm_neon.h delegates to vfma). 5296 5297 // The ARM builtins (and instructions) have the addend as the first 5298 // operand, but the 'fma' intrinsics have it last. Swap it around here. 5299 Value *Subtrahend = Ops[0]; 5300 Value *Multiplicand = Ops[2]; 5301 Ops[0] = Multiplicand; 5302 Ops[2] = Subtrahend; 5303 Ops[1] = Builder.CreateBitCast(Ops[1], VTy); 5304 Ops[1] = Builder.CreateFNeg(Ops[1]); 5305 Int = Intrinsic::fma; 5306 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "fmls"); 5307 } 5308 case NEON::BI__builtin_neon_vmull_v: 5309 // FIXME: improve sharing scheme to cope with 3 alternative LLVM intrinsics. 5310 Int = usgn ? Intrinsic::aarch64_neon_umull : Intrinsic::aarch64_neon_smull; 5311 if (Type.isPoly()) Int = Intrinsic::aarch64_neon_pmull; 5312 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmull"); 5313 case NEON::BI__builtin_neon_vmax_v: 5314 case NEON::BI__builtin_neon_vmaxq_v: 5315 // FIXME: improve sharing scheme to cope with 3 alternative LLVM intrinsics. 5316 Int = usgn ? Intrinsic::aarch64_neon_umax : Intrinsic::aarch64_neon_smax; 5317 if (Ty->isFPOrFPVectorTy()) Int = Intrinsic::aarch64_neon_fmax; 5318 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmax"); 5319 case NEON::BI__builtin_neon_vmin_v: 5320 case NEON::BI__builtin_neon_vminq_v: 5321 // FIXME: improve sharing scheme to cope with 3 alternative LLVM intrinsics. 5322 Int = usgn ? Intrinsic::aarch64_neon_umin : Intrinsic::aarch64_neon_smin; 5323 if (Ty->isFPOrFPVectorTy()) Int = Intrinsic::aarch64_neon_fmin; 5324 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmin"); 5325 case NEON::BI__builtin_neon_vabd_v: 5326 case NEON::BI__builtin_neon_vabdq_v: 5327 // FIXME: improve sharing scheme to cope with 3 alternative LLVM intrinsics. 5328 Int = usgn ? Intrinsic::aarch64_neon_uabd : Intrinsic::aarch64_neon_sabd; 5329 if (Ty->isFPOrFPVectorTy()) Int = Intrinsic::aarch64_neon_fabd; 5330 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vabd"); 5331 case NEON::BI__builtin_neon_vpadal_v: 5332 case NEON::BI__builtin_neon_vpadalq_v: { 5333 unsigned ArgElts = VTy->getNumElements(); 5334 llvm::IntegerType *EltTy = cast<IntegerType>(VTy->getElementType()); 5335 unsigned BitWidth = EltTy->getBitWidth(); 5336 llvm::Type *ArgTy = llvm::VectorType::get( 5337 llvm::IntegerType::get(getLLVMContext(), BitWidth/2), 2*ArgElts); 5338 llvm::Type* Tys[2] = { VTy, ArgTy }; 5339 Int = usgn ? Intrinsic::aarch64_neon_uaddlp : Intrinsic::aarch64_neon_saddlp; 5340 SmallVector<llvm::Value*, 1> TmpOps; 5341 TmpOps.push_back(Ops[1]); 5342 Function *F = CGM.getIntrinsic(Int, Tys); 5343 llvm::Value *tmp = EmitNeonCall(F, TmpOps, "vpadal"); 5344 llvm::Value *addend = Builder.CreateBitCast(Ops[0], tmp->getType()); 5345 return Builder.CreateAdd(tmp, addend); 5346 } 5347 case NEON::BI__builtin_neon_vpmin_v: 5348 case NEON::BI__builtin_neon_vpminq_v: 5349 // FIXME: improve sharing scheme to cope with 3 alternative LLVM intrinsics. 5350 Int = usgn ? Intrinsic::aarch64_neon_uminp : Intrinsic::aarch64_neon_sminp; 5351 if (Ty->isFPOrFPVectorTy()) Int = Intrinsic::aarch64_neon_fminp; 5352 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vpmin"); 5353 case NEON::BI__builtin_neon_vpmax_v: 5354 case NEON::BI__builtin_neon_vpmaxq_v: 5355 // FIXME: improve sharing scheme to cope with 3 alternative LLVM intrinsics. 5356 Int = usgn ? Intrinsic::aarch64_neon_umaxp : Intrinsic::aarch64_neon_smaxp; 5357 if (Ty->isFPOrFPVectorTy()) Int = Intrinsic::aarch64_neon_fmaxp; 5358 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vpmax"); 5359 case NEON::BI__builtin_neon_vminnm_v: 5360 case NEON::BI__builtin_neon_vminnmq_v: 5361 Int = Intrinsic::aarch64_neon_fminnm; 5362 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vminnm"); 5363 case NEON::BI__builtin_neon_vmaxnm_v: 5364 case NEON::BI__builtin_neon_vmaxnmq_v: 5365 Int = Intrinsic::aarch64_neon_fmaxnm; 5366 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmaxnm"); 5367 case NEON::BI__builtin_neon_vrecpss_f32: { 5368 Ops.push_back(EmitScalarExpr(E->getArg(1))); 5369 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_frecps, FloatTy), 5370 Ops, "vrecps"); 5371 } 5372 case NEON::BI__builtin_neon_vrecpsd_f64: { 5373 Ops.push_back(EmitScalarExpr(E->getArg(1))); 5374 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_frecps, DoubleTy), 5375 Ops, "vrecps"); 5376 } 5377 case NEON::BI__builtin_neon_vqshrun_n_v: 5378 Int = Intrinsic::aarch64_neon_sqshrun; 5379 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqshrun_n"); 5380 case NEON::BI__builtin_neon_vqrshrun_n_v: 5381 Int = Intrinsic::aarch64_neon_sqrshrun; 5382 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqrshrun_n"); 5383 case NEON::BI__builtin_neon_vqshrn_n_v: 5384 Int = usgn ? Intrinsic::aarch64_neon_uqshrn : Intrinsic::aarch64_neon_sqshrn; 5385 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqshrn_n"); 5386 case NEON::BI__builtin_neon_vrshrn_n_v: 5387 Int = Intrinsic::aarch64_neon_rshrn; 5388 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrshrn_n"); 5389 case NEON::BI__builtin_neon_vqrshrn_n_v: 5390 Int = usgn ? Intrinsic::aarch64_neon_uqrshrn : Intrinsic::aarch64_neon_sqrshrn; 5391 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqrshrn_n"); 5392 case NEON::BI__builtin_neon_vrnda_v: 5393 case NEON::BI__builtin_neon_vrndaq_v: { 5394 Int = Intrinsic::round; 5395 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrnda"); 5396 } 5397 case NEON::BI__builtin_neon_vrndi_v: 5398 case NEON::BI__builtin_neon_vrndiq_v: { 5399 Int = Intrinsic::nearbyint; 5400 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrndi"); 5401 } 5402 case NEON::BI__builtin_neon_vrndm_v: 5403 case NEON::BI__builtin_neon_vrndmq_v: { 5404 Int = Intrinsic::floor; 5405 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrndm"); 5406 } 5407 case NEON::BI__builtin_neon_vrndn_v: 5408 case NEON::BI__builtin_neon_vrndnq_v: { 5409 Int = Intrinsic::aarch64_neon_frintn; 5410 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrndn"); 5411 } 5412 case NEON::BI__builtin_neon_vrndp_v: 5413 case NEON::BI__builtin_neon_vrndpq_v: { 5414 Int = Intrinsic::ceil; 5415 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrndp"); 5416 } 5417 case NEON::BI__builtin_neon_vrndx_v: 5418 case NEON::BI__builtin_neon_vrndxq_v: { 5419 Int = Intrinsic::rint; 5420 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrndx"); 5421 } 5422 case NEON::BI__builtin_neon_vrnd_v: 5423 case NEON::BI__builtin_neon_vrndq_v: { 5424 Int = Intrinsic::trunc; 5425 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrndz"); 5426 } 5427 case NEON::BI__builtin_neon_vceqz_v: 5428 case NEON::BI__builtin_neon_vceqzq_v: 5429 return EmitAArch64CompareBuiltinExpr(Ops[0], Ty, ICmpInst::FCMP_OEQ, 5430 ICmpInst::ICMP_EQ, "vceqz"); 5431 case NEON::BI__builtin_neon_vcgez_v: 5432 case NEON::BI__builtin_neon_vcgezq_v: 5433 return EmitAArch64CompareBuiltinExpr(Ops[0], Ty, ICmpInst::FCMP_OGE, 5434 ICmpInst::ICMP_SGE, "vcgez"); 5435 case NEON::BI__builtin_neon_vclez_v: 5436 case NEON::BI__builtin_neon_vclezq_v: 5437 return EmitAArch64CompareBuiltinExpr(Ops[0], Ty, ICmpInst::FCMP_OLE, 5438 ICmpInst::ICMP_SLE, "vclez"); 5439 case NEON::BI__builtin_neon_vcgtz_v: 5440 case NEON::BI__builtin_neon_vcgtzq_v: 5441 return EmitAArch64CompareBuiltinExpr(Ops[0], Ty, ICmpInst::FCMP_OGT, 5442 ICmpInst::ICMP_SGT, "vcgtz"); 5443 case NEON::BI__builtin_neon_vcltz_v: 5444 case NEON::BI__builtin_neon_vcltzq_v: 5445 return EmitAArch64CompareBuiltinExpr(Ops[0], Ty, ICmpInst::FCMP_OLT, 5446 ICmpInst::ICMP_SLT, "vcltz"); 5447 case NEON::BI__builtin_neon_vcvt_f64_v: 5448 case NEON::BI__builtin_neon_vcvtq_f64_v: 5449 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 5450 Ty = GetNeonType(this, NeonTypeFlags(NeonTypeFlags::Float64, false, quad)); 5451 return usgn ? Builder.CreateUIToFP(Ops[0], Ty, "vcvt") 5452 : Builder.CreateSIToFP(Ops[0], Ty, "vcvt"); 5453 case NEON::BI__builtin_neon_vcvt_f64_f32: { 5454 assert(Type.getEltType() == NeonTypeFlags::Float64 && quad && 5455 "unexpected vcvt_f64_f32 builtin"); 5456 NeonTypeFlags SrcFlag = NeonTypeFlags(NeonTypeFlags::Float32, false, false); 5457 Ops[0] = Builder.CreateBitCast(Ops[0], GetNeonType(this, SrcFlag)); 5458 5459 return Builder.CreateFPExt(Ops[0], Ty, "vcvt"); 5460 } 5461 case NEON::BI__builtin_neon_vcvt_f32_f64: { 5462 assert(Type.getEltType() == NeonTypeFlags::Float32 && 5463 "unexpected vcvt_f32_f64 builtin"); 5464 NeonTypeFlags SrcFlag = NeonTypeFlags(NeonTypeFlags::Float64, false, true); 5465 Ops[0] = Builder.CreateBitCast(Ops[0], GetNeonType(this, SrcFlag)); 5466 5467 return Builder.CreateFPTrunc(Ops[0], Ty, "vcvt"); 5468 } 5469 case NEON::BI__builtin_neon_vcvt_s32_v: 5470 case NEON::BI__builtin_neon_vcvt_u32_v: 5471 case NEON::BI__builtin_neon_vcvt_s64_v: 5472 case NEON::BI__builtin_neon_vcvt_u64_v: 5473 case NEON::BI__builtin_neon_vcvtq_s32_v: 5474 case NEON::BI__builtin_neon_vcvtq_u32_v: 5475 case NEON::BI__builtin_neon_vcvtq_s64_v: 5476 case NEON::BI__builtin_neon_vcvtq_u64_v: { 5477 Ops[0] = Builder.CreateBitCast(Ops[0], GetFloatNeonType(this, Type)); 5478 if (usgn) 5479 return Builder.CreateFPToUI(Ops[0], Ty); 5480 return Builder.CreateFPToSI(Ops[0], Ty); 5481 } 5482 case NEON::BI__builtin_neon_vcvta_s32_v: 5483 case NEON::BI__builtin_neon_vcvtaq_s32_v: 5484 case NEON::BI__builtin_neon_vcvta_u32_v: 5485 case NEON::BI__builtin_neon_vcvtaq_u32_v: 5486 case NEON::BI__builtin_neon_vcvta_s64_v: 5487 case NEON::BI__builtin_neon_vcvtaq_s64_v: 5488 case NEON::BI__builtin_neon_vcvta_u64_v: 5489 case NEON::BI__builtin_neon_vcvtaq_u64_v: { 5490 Int = usgn ? Intrinsic::aarch64_neon_fcvtau : Intrinsic::aarch64_neon_fcvtas; 5491 llvm::Type *Tys[2] = { Ty, GetFloatNeonType(this, Type) }; 5492 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vcvta"); 5493 } 5494 case NEON::BI__builtin_neon_vcvtm_s32_v: 5495 case NEON::BI__builtin_neon_vcvtmq_s32_v: 5496 case NEON::BI__builtin_neon_vcvtm_u32_v: 5497 case NEON::BI__builtin_neon_vcvtmq_u32_v: 5498 case NEON::BI__builtin_neon_vcvtm_s64_v: 5499 case NEON::BI__builtin_neon_vcvtmq_s64_v: 5500 case NEON::BI__builtin_neon_vcvtm_u64_v: 5501 case NEON::BI__builtin_neon_vcvtmq_u64_v: { 5502 Int = usgn ? Intrinsic::aarch64_neon_fcvtmu : Intrinsic::aarch64_neon_fcvtms; 5503 llvm::Type *Tys[2] = { Ty, GetFloatNeonType(this, Type) }; 5504 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vcvtm"); 5505 } 5506 case NEON::BI__builtin_neon_vcvtn_s32_v: 5507 case NEON::BI__builtin_neon_vcvtnq_s32_v: 5508 case NEON::BI__builtin_neon_vcvtn_u32_v: 5509 case NEON::BI__builtin_neon_vcvtnq_u32_v: 5510 case NEON::BI__builtin_neon_vcvtn_s64_v: 5511 case NEON::BI__builtin_neon_vcvtnq_s64_v: 5512 case NEON::BI__builtin_neon_vcvtn_u64_v: 5513 case NEON::BI__builtin_neon_vcvtnq_u64_v: { 5514 Int = usgn ? Intrinsic::aarch64_neon_fcvtnu : Intrinsic::aarch64_neon_fcvtns; 5515 llvm::Type *Tys[2] = { Ty, GetFloatNeonType(this, Type) }; 5516 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vcvtn"); 5517 } 5518 case NEON::BI__builtin_neon_vcvtp_s32_v: 5519 case NEON::BI__builtin_neon_vcvtpq_s32_v: 5520 case NEON::BI__builtin_neon_vcvtp_u32_v: 5521 case NEON::BI__builtin_neon_vcvtpq_u32_v: 5522 case NEON::BI__builtin_neon_vcvtp_s64_v: 5523 case NEON::BI__builtin_neon_vcvtpq_s64_v: 5524 case NEON::BI__builtin_neon_vcvtp_u64_v: 5525 case NEON::BI__builtin_neon_vcvtpq_u64_v: { 5526 Int = usgn ? Intrinsic::aarch64_neon_fcvtpu : Intrinsic::aarch64_neon_fcvtps; 5527 llvm::Type *Tys[2] = { Ty, GetFloatNeonType(this, Type) }; 5528 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vcvtp"); 5529 } 5530 case NEON::BI__builtin_neon_vmulx_v: 5531 case NEON::BI__builtin_neon_vmulxq_v: { 5532 Int = Intrinsic::aarch64_neon_fmulx; 5533 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmulx"); 5534 } 5535 case NEON::BI__builtin_neon_vmul_lane_v: 5536 case NEON::BI__builtin_neon_vmul_laneq_v: { 5537 // v1f64 vmul_lane should be mapped to Neon scalar mul lane 5538 bool Quad = false; 5539 if (BuiltinID == NEON::BI__builtin_neon_vmul_laneq_v) 5540 Quad = true; 5541 Ops[0] = Builder.CreateBitCast(Ops[0], DoubleTy); 5542 llvm::Type *VTy = GetNeonType(this, 5543 NeonTypeFlags(NeonTypeFlags::Float64, false, Quad)); 5544 Ops[1] = Builder.CreateBitCast(Ops[1], VTy); 5545 Ops[1] = Builder.CreateExtractElement(Ops[1], Ops[2], "extract"); 5546 Value *Result = Builder.CreateFMul(Ops[0], Ops[1]); 5547 return Builder.CreateBitCast(Result, Ty); 5548 } 5549 case NEON::BI__builtin_neon_vnegd_s64: 5550 return Builder.CreateNeg(EmitScalarExpr(E->getArg(0)), "vnegd"); 5551 case NEON::BI__builtin_neon_vpmaxnm_v: 5552 case NEON::BI__builtin_neon_vpmaxnmq_v: { 5553 Int = Intrinsic::aarch64_neon_fmaxnmp; 5554 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vpmaxnm"); 5555 } 5556 case NEON::BI__builtin_neon_vpminnm_v: 5557 case NEON::BI__builtin_neon_vpminnmq_v: { 5558 Int = Intrinsic::aarch64_neon_fminnmp; 5559 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vpminnm"); 5560 } 5561 case NEON::BI__builtin_neon_vsqrt_v: 5562 case NEON::BI__builtin_neon_vsqrtq_v: { 5563 Int = Intrinsic::sqrt; 5564 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 5565 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vsqrt"); 5566 } 5567 case NEON::BI__builtin_neon_vrbit_v: 5568 case NEON::BI__builtin_neon_vrbitq_v: { 5569 Int = Intrinsic::aarch64_neon_rbit; 5570 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrbit"); 5571 } 5572 case NEON::BI__builtin_neon_vaddv_u8: 5573 // FIXME: These are handled by the AArch64 scalar code. 5574 usgn = true; 5575 // FALLTHROUGH 5576 case NEON::BI__builtin_neon_vaddv_s8: { 5577 Int = usgn ? Intrinsic::aarch64_neon_uaddv : Intrinsic::aarch64_neon_saddv; 5578 Ty = Int32Ty; 5579 VTy = llvm::VectorType::get(Int8Ty, 8); 5580 llvm::Type *Tys[2] = { Ty, VTy }; 5581 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5582 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddv"); 5583 return Builder.CreateTrunc(Ops[0], Int8Ty); 5584 } 5585 case NEON::BI__builtin_neon_vaddv_u16: 5586 usgn = true; 5587 // FALLTHROUGH 5588 case NEON::BI__builtin_neon_vaddv_s16: { 5589 Int = usgn ? Intrinsic::aarch64_neon_uaddv : Intrinsic::aarch64_neon_saddv; 5590 Ty = Int32Ty; 5591 VTy = llvm::VectorType::get(Int16Ty, 4); 5592 llvm::Type *Tys[2] = { Ty, VTy }; 5593 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5594 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddv"); 5595 return Builder.CreateTrunc(Ops[0], Int16Ty); 5596 } 5597 case NEON::BI__builtin_neon_vaddvq_u8: 5598 usgn = true; 5599 // FALLTHROUGH 5600 case NEON::BI__builtin_neon_vaddvq_s8: { 5601 Int = usgn ? Intrinsic::aarch64_neon_uaddv : Intrinsic::aarch64_neon_saddv; 5602 Ty = Int32Ty; 5603 VTy = llvm::VectorType::get(Int8Ty, 16); 5604 llvm::Type *Tys[2] = { Ty, VTy }; 5605 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5606 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddv"); 5607 return Builder.CreateTrunc(Ops[0], Int8Ty); 5608 } 5609 case NEON::BI__builtin_neon_vaddvq_u16: 5610 usgn = true; 5611 // FALLTHROUGH 5612 case NEON::BI__builtin_neon_vaddvq_s16: { 5613 Int = usgn ? Intrinsic::aarch64_neon_uaddv : Intrinsic::aarch64_neon_saddv; 5614 Ty = Int32Ty; 5615 VTy = llvm::VectorType::get(Int16Ty, 8); 5616 llvm::Type *Tys[2] = { Ty, VTy }; 5617 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5618 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddv"); 5619 return Builder.CreateTrunc(Ops[0], Int16Ty); 5620 } 5621 case NEON::BI__builtin_neon_vmaxv_u8: { 5622 Int = Intrinsic::aarch64_neon_umaxv; 5623 Ty = Int32Ty; 5624 VTy = llvm::VectorType::get(Int8Ty, 8); 5625 llvm::Type *Tys[2] = { Ty, VTy }; 5626 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5627 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv"); 5628 return Builder.CreateTrunc(Ops[0], Int8Ty); 5629 } 5630 case NEON::BI__builtin_neon_vmaxv_u16: { 5631 Int = Intrinsic::aarch64_neon_umaxv; 5632 Ty = Int32Ty; 5633 VTy = llvm::VectorType::get(Int16Ty, 4); 5634 llvm::Type *Tys[2] = { Ty, VTy }; 5635 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5636 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv"); 5637 return Builder.CreateTrunc(Ops[0], Int16Ty); 5638 } 5639 case NEON::BI__builtin_neon_vmaxvq_u8: { 5640 Int = Intrinsic::aarch64_neon_umaxv; 5641 Ty = Int32Ty; 5642 VTy = llvm::VectorType::get(Int8Ty, 16); 5643 llvm::Type *Tys[2] = { Ty, VTy }; 5644 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5645 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv"); 5646 return Builder.CreateTrunc(Ops[0], Int8Ty); 5647 } 5648 case NEON::BI__builtin_neon_vmaxvq_u16: { 5649 Int = Intrinsic::aarch64_neon_umaxv; 5650 Ty = Int32Ty; 5651 VTy = llvm::VectorType::get(Int16Ty, 8); 5652 llvm::Type *Tys[2] = { Ty, VTy }; 5653 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5654 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv"); 5655 return Builder.CreateTrunc(Ops[0], Int16Ty); 5656 } 5657 case NEON::BI__builtin_neon_vmaxv_s8: { 5658 Int = Intrinsic::aarch64_neon_smaxv; 5659 Ty = Int32Ty; 5660 VTy = llvm::VectorType::get(Int8Ty, 8); 5661 llvm::Type *Tys[2] = { Ty, VTy }; 5662 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5663 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv"); 5664 return Builder.CreateTrunc(Ops[0], Int8Ty); 5665 } 5666 case NEON::BI__builtin_neon_vmaxv_s16: { 5667 Int = Intrinsic::aarch64_neon_smaxv; 5668 Ty = Int32Ty; 5669 VTy = llvm::VectorType::get(Int16Ty, 4); 5670 llvm::Type *Tys[2] = { Ty, VTy }; 5671 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5672 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv"); 5673 return Builder.CreateTrunc(Ops[0], Int16Ty); 5674 } 5675 case NEON::BI__builtin_neon_vmaxvq_s8: { 5676 Int = Intrinsic::aarch64_neon_smaxv; 5677 Ty = Int32Ty; 5678 VTy = llvm::VectorType::get(Int8Ty, 16); 5679 llvm::Type *Tys[2] = { Ty, VTy }; 5680 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5681 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv"); 5682 return Builder.CreateTrunc(Ops[0], Int8Ty); 5683 } 5684 case NEON::BI__builtin_neon_vmaxvq_s16: { 5685 Int = Intrinsic::aarch64_neon_smaxv; 5686 Ty = Int32Ty; 5687 VTy = llvm::VectorType::get(Int16Ty, 8); 5688 llvm::Type *Tys[2] = { Ty, VTy }; 5689 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5690 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv"); 5691 return Builder.CreateTrunc(Ops[0], Int16Ty); 5692 } 5693 case NEON::BI__builtin_neon_vminv_u8: { 5694 Int = Intrinsic::aarch64_neon_uminv; 5695 Ty = Int32Ty; 5696 VTy = llvm::VectorType::get(Int8Ty, 8); 5697 llvm::Type *Tys[2] = { Ty, VTy }; 5698 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5699 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv"); 5700 return Builder.CreateTrunc(Ops[0], Int8Ty); 5701 } 5702 case NEON::BI__builtin_neon_vminv_u16: { 5703 Int = Intrinsic::aarch64_neon_uminv; 5704 Ty = Int32Ty; 5705 VTy = llvm::VectorType::get(Int16Ty, 4); 5706 llvm::Type *Tys[2] = { Ty, VTy }; 5707 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5708 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv"); 5709 return Builder.CreateTrunc(Ops[0], Int16Ty); 5710 } 5711 case NEON::BI__builtin_neon_vminvq_u8: { 5712 Int = Intrinsic::aarch64_neon_uminv; 5713 Ty = Int32Ty; 5714 VTy = llvm::VectorType::get(Int8Ty, 16); 5715 llvm::Type *Tys[2] = { Ty, VTy }; 5716 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5717 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv"); 5718 return Builder.CreateTrunc(Ops[0], Int8Ty); 5719 } 5720 case NEON::BI__builtin_neon_vminvq_u16: { 5721 Int = Intrinsic::aarch64_neon_uminv; 5722 Ty = Int32Ty; 5723 VTy = llvm::VectorType::get(Int16Ty, 8); 5724 llvm::Type *Tys[2] = { Ty, VTy }; 5725 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5726 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv"); 5727 return Builder.CreateTrunc(Ops[0], Int16Ty); 5728 } 5729 case NEON::BI__builtin_neon_vminv_s8: { 5730 Int = Intrinsic::aarch64_neon_sminv; 5731 Ty = Int32Ty; 5732 VTy = llvm::VectorType::get(Int8Ty, 8); 5733 llvm::Type *Tys[2] = { Ty, VTy }; 5734 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5735 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv"); 5736 return Builder.CreateTrunc(Ops[0], Int8Ty); 5737 } 5738 case NEON::BI__builtin_neon_vminv_s16: { 5739 Int = Intrinsic::aarch64_neon_sminv; 5740 Ty = Int32Ty; 5741 VTy = llvm::VectorType::get(Int16Ty, 4); 5742 llvm::Type *Tys[2] = { Ty, VTy }; 5743 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5744 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv"); 5745 return Builder.CreateTrunc(Ops[0], Int16Ty); 5746 } 5747 case NEON::BI__builtin_neon_vminvq_s8: { 5748 Int = Intrinsic::aarch64_neon_sminv; 5749 Ty = Int32Ty; 5750 VTy = llvm::VectorType::get(Int8Ty, 16); 5751 llvm::Type *Tys[2] = { Ty, VTy }; 5752 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5753 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv"); 5754 return Builder.CreateTrunc(Ops[0], Int8Ty); 5755 } 5756 case NEON::BI__builtin_neon_vminvq_s16: { 5757 Int = Intrinsic::aarch64_neon_sminv; 5758 Ty = Int32Ty; 5759 VTy = llvm::VectorType::get(Int16Ty, 8); 5760 llvm::Type *Tys[2] = { Ty, VTy }; 5761 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5762 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv"); 5763 return Builder.CreateTrunc(Ops[0], Int16Ty); 5764 } 5765 case NEON::BI__builtin_neon_vmul_n_f64: { 5766 Ops[0] = Builder.CreateBitCast(Ops[0], DoubleTy); 5767 Value *RHS = Builder.CreateBitCast(EmitScalarExpr(E->getArg(1)), DoubleTy); 5768 return Builder.CreateFMul(Ops[0], RHS); 5769 } 5770 case NEON::BI__builtin_neon_vaddlv_u8: { 5771 Int = Intrinsic::aarch64_neon_uaddlv; 5772 Ty = Int32Ty; 5773 VTy = llvm::VectorType::get(Int8Ty, 8); 5774 llvm::Type *Tys[2] = { Ty, VTy }; 5775 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5776 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv"); 5777 return Builder.CreateTrunc(Ops[0], Int16Ty); 5778 } 5779 case NEON::BI__builtin_neon_vaddlv_u16: { 5780 Int = Intrinsic::aarch64_neon_uaddlv; 5781 Ty = Int32Ty; 5782 VTy = llvm::VectorType::get(Int16Ty, 4); 5783 llvm::Type *Tys[2] = { Ty, VTy }; 5784 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5785 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv"); 5786 } 5787 case NEON::BI__builtin_neon_vaddlvq_u8: { 5788 Int = Intrinsic::aarch64_neon_uaddlv; 5789 Ty = Int32Ty; 5790 VTy = llvm::VectorType::get(Int8Ty, 16); 5791 llvm::Type *Tys[2] = { Ty, VTy }; 5792 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5793 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv"); 5794 return Builder.CreateTrunc(Ops[0], Int16Ty); 5795 } 5796 case NEON::BI__builtin_neon_vaddlvq_u16: { 5797 Int = Intrinsic::aarch64_neon_uaddlv; 5798 Ty = Int32Ty; 5799 VTy = llvm::VectorType::get(Int16Ty, 8); 5800 llvm::Type *Tys[2] = { Ty, VTy }; 5801 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5802 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv"); 5803 } 5804 case NEON::BI__builtin_neon_vaddlv_s8: { 5805 Int = Intrinsic::aarch64_neon_saddlv; 5806 Ty = Int32Ty; 5807 VTy = llvm::VectorType::get(Int8Ty, 8); 5808 llvm::Type *Tys[2] = { Ty, VTy }; 5809 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5810 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv"); 5811 return Builder.CreateTrunc(Ops[0], Int16Ty); 5812 } 5813 case NEON::BI__builtin_neon_vaddlv_s16: { 5814 Int = Intrinsic::aarch64_neon_saddlv; 5815 Ty = Int32Ty; 5816 VTy = llvm::VectorType::get(Int16Ty, 4); 5817 llvm::Type *Tys[2] = { Ty, VTy }; 5818 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5819 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv"); 5820 } 5821 case NEON::BI__builtin_neon_vaddlvq_s8: { 5822 Int = Intrinsic::aarch64_neon_saddlv; 5823 Ty = Int32Ty; 5824 VTy = llvm::VectorType::get(Int8Ty, 16); 5825 llvm::Type *Tys[2] = { Ty, VTy }; 5826 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5827 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv"); 5828 return Builder.CreateTrunc(Ops[0], Int16Ty); 5829 } 5830 case NEON::BI__builtin_neon_vaddlvq_s16: { 5831 Int = Intrinsic::aarch64_neon_saddlv; 5832 Ty = Int32Ty; 5833 VTy = llvm::VectorType::get(Int16Ty, 8); 5834 llvm::Type *Tys[2] = { Ty, VTy }; 5835 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5836 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv"); 5837 } 5838 case NEON::BI__builtin_neon_vsri_n_v: 5839 case NEON::BI__builtin_neon_vsriq_n_v: { 5840 Int = Intrinsic::aarch64_neon_vsri; 5841 llvm::Function *Intrin = CGM.getIntrinsic(Int, Ty); 5842 return EmitNeonCall(Intrin, Ops, "vsri_n"); 5843 } 5844 case NEON::BI__builtin_neon_vsli_n_v: 5845 case NEON::BI__builtin_neon_vsliq_n_v: { 5846 Int = Intrinsic::aarch64_neon_vsli; 5847 llvm::Function *Intrin = CGM.getIntrinsic(Int, Ty); 5848 return EmitNeonCall(Intrin, Ops, "vsli_n"); 5849 } 5850 case NEON::BI__builtin_neon_vsra_n_v: 5851 case NEON::BI__builtin_neon_vsraq_n_v: 5852 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 5853 Ops[1] = EmitNeonRShiftImm(Ops[1], Ops[2], Ty, usgn, "vsra_n"); 5854 return Builder.CreateAdd(Ops[0], Ops[1]); 5855 case NEON::BI__builtin_neon_vrsra_n_v: 5856 case NEON::BI__builtin_neon_vrsraq_n_v: { 5857 Int = usgn ? Intrinsic::aarch64_neon_urshl : Intrinsic::aarch64_neon_srshl; 5858 SmallVector<llvm::Value*,2> TmpOps; 5859 TmpOps.push_back(Ops[1]); 5860 TmpOps.push_back(Ops[2]); 5861 Function* F = CGM.getIntrinsic(Int, Ty); 5862 llvm::Value *tmp = EmitNeonCall(F, TmpOps, "vrshr_n", 1, true); 5863 Ops[0] = Builder.CreateBitCast(Ops[0], VTy); 5864 return Builder.CreateAdd(Ops[0], tmp); 5865 } 5866 // FIXME: Sharing loads & stores with 32-bit is complicated by the absence 5867 // of an Align parameter here. 5868 case NEON::BI__builtin_neon_vld1_x2_v: 5869 case NEON::BI__builtin_neon_vld1q_x2_v: 5870 case NEON::BI__builtin_neon_vld1_x3_v: 5871 case NEON::BI__builtin_neon_vld1q_x3_v: 5872 case NEON::BI__builtin_neon_vld1_x4_v: 5873 case NEON::BI__builtin_neon_vld1q_x4_v: { 5874 llvm::Type *PTy = llvm::PointerType::getUnqual(VTy->getVectorElementType()); 5875 Ops[1] = Builder.CreateBitCast(Ops[1], PTy); 5876 llvm::Type *Tys[2] = { VTy, PTy }; 5877 unsigned Int; 5878 switch (BuiltinID) { 5879 case NEON::BI__builtin_neon_vld1_x2_v: 5880 case NEON::BI__builtin_neon_vld1q_x2_v: 5881 Int = Intrinsic::aarch64_neon_ld1x2; 5882 break; 5883 case NEON::BI__builtin_neon_vld1_x3_v: 5884 case NEON::BI__builtin_neon_vld1q_x3_v: 5885 Int = Intrinsic::aarch64_neon_ld1x3; 5886 break; 5887 case NEON::BI__builtin_neon_vld1_x4_v: 5888 case NEON::BI__builtin_neon_vld1q_x4_v: 5889 Int = Intrinsic::aarch64_neon_ld1x4; 5890 break; 5891 } 5892 Function *F = CGM.getIntrinsic(Int, Tys); 5893 Ops[1] = Builder.CreateCall(F, Ops[1], "vld1xN"); 5894 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 5895 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 5896 return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]); 5897 } 5898 case NEON::BI__builtin_neon_vst1_x2_v: 5899 case NEON::BI__builtin_neon_vst1q_x2_v: 5900 case NEON::BI__builtin_neon_vst1_x3_v: 5901 case NEON::BI__builtin_neon_vst1q_x3_v: 5902 case NEON::BI__builtin_neon_vst1_x4_v: 5903 case NEON::BI__builtin_neon_vst1q_x4_v: { 5904 llvm::Type *PTy = llvm::PointerType::getUnqual(VTy->getVectorElementType()); 5905 llvm::Type *Tys[2] = { VTy, PTy }; 5906 unsigned Int; 5907 switch (BuiltinID) { 5908 case NEON::BI__builtin_neon_vst1_x2_v: 5909 case NEON::BI__builtin_neon_vst1q_x2_v: 5910 Int = Intrinsic::aarch64_neon_st1x2; 5911 break; 5912 case NEON::BI__builtin_neon_vst1_x3_v: 5913 case NEON::BI__builtin_neon_vst1q_x3_v: 5914 Int = Intrinsic::aarch64_neon_st1x3; 5915 break; 5916 case NEON::BI__builtin_neon_vst1_x4_v: 5917 case NEON::BI__builtin_neon_vst1q_x4_v: 5918 Int = Intrinsic::aarch64_neon_st1x4; 5919 break; 5920 } 5921 std::rotate(Ops.begin(), Ops.begin() + 1, Ops.end()); 5922 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, ""); 5923 } 5924 case NEON::BI__builtin_neon_vld1_v: 5925 case NEON::BI__builtin_neon_vld1q_v: 5926 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(VTy)); 5927 return Builder.CreateDefaultAlignedLoad(Ops[0]); 5928 case NEON::BI__builtin_neon_vst1_v: 5929 case NEON::BI__builtin_neon_vst1q_v: 5930 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(VTy)); 5931 Ops[1] = Builder.CreateBitCast(Ops[1], VTy); 5932 return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]); 5933 case NEON::BI__builtin_neon_vld1_lane_v: 5934 case NEON::BI__builtin_neon_vld1q_lane_v: 5935 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 5936 Ty = llvm::PointerType::getUnqual(VTy->getElementType()); 5937 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 5938 Ops[0] = Builder.CreateDefaultAlignedLoad(Ops[0]); 5939 return Builder.CreateInsertElement(Ops[1], Ops[0], Ops[2], "vld1_lane"); 5940 case NEON::BI__builtin_neon_vld1_dup_v: 5941 case NEON::BI__builtin_neon_vld1q_dup_v: { 5942 Value *V = UndefValue::get(Ty); 5943 Ty = llvm::PointerType::getUnqual(VTy->getElementType()); 5944 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 5945 Ops[0] = Builder.CreateDefaultAlignedLoad(Ops[0]); 5946 llvm::Constant *CI = ConstantInt::get(Int32Ty, 0); 5947 Ops[0] = Builder.CreateInsertElement(V, Ops[0], CI); 5948 return EmitNeonSplat(Ops[0], CI); 5949 } 5950 case NEON::BI__builtin_neon_vst1_lane_v: 5951 case NEON::BI__builtin_neon_vst1q_lane_v: 5952 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 5953 Ops[1] = Builder.CreateExtractElement(Ops[1], Ops[2]); 5954 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 5955 return Builder.CreateDefaultAlignedStore(Ops[1], 5956 Builder.CreateBitCast(Ops[0], Ty)); 5957 case NEON::BI__builtin_neon_vld2_v: 5958 case NEON::BI__builtin_neon_vld2q_v: { 5959 llvm::Type *PTy = llvm::PointerType::getUnqual(VTy); 5960 Ops[1] = Builder.CreateBitCast(Ops[1], PTy); 5961 llvm::Type *Tys[2] = { VTy, PTy }; 5962 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld2, Tys); 5963 Ops[1] = Builder.CreateCall(F, Ops[1], "vld2"); 5964 Ops[0] = Builder.CreateBitCast(Ops[0], 5965 llvm::PointerType::getUnqual(Ops[1]->getType())); 5966 return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]); 5967 } 5968 case NEON::BI__builtin_neon_vld3_v: 5969 case NEON::BI__builtin_neon_vld3q_v: { 5970 llvm::Type *PTy = llvm::PointerType::getUnqual(VTy); 5971 Ops[1] = Builder.CreateBitCast(Ops[1], PTy); 5972 llvm::Type *Tys[2] = { VTy, PTy }; 5973 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld3, Tys); 5974 Ops[1] = Builder.CreateCall(F, Ops[1], "vld3"); 5975 Ops[0] = Builder.CreateBitCast(Ops[0], 5976 llvm::PointerType::getUnqual(Ops[1]->getType())); 5977 return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]); 5978 } 5979 case NEON::BI__builtin_neon_vld4_v: 5980 case NEON::BI__builtin_neon_vld4q_v: { 5981 llvm::Type *PTy = llvm::PointerType::getUnqual(VTy); 5982 Ops[1] = Builder.CreateBitCast(Ops[1], PTy); 5983 llvm::Type *Tys[2] = { VTy, PTy }; 5984 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld4, Tys); 5985 Ops[1] = Builder.CreateCall(F, Ops[1], "vld4"); 5986 Ops[0] = Builder.CreateBitCast(Ops[0], 5987 llvm::PointerType::getUnqual(Ops[1]->getType())); 5988 return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]); 5989 } 5990 case NEON::BI__builtin_neon_vld2_dup_v: 5991 case NEON::BI__builtin_neon_vld2q_dup_v: { 5992 llvm::Type *PTy = 5993 llvm::PointerType::getUnqual(VTy->getElementType()); 5994 Ops[1] = Builder.CreateBitCast(Ops[1], PTy); 5995 llvm::Type *Tys[2] = { VTy, PTy }; 5996 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld2r, Tys); 5997 Ops[1] = Builder.CreateCall(F, Ops[1], "vld2"); 5998 Ops[0] = Builder.CreateBitCast(Ops[0], 5999 llvm::PointerType::getUnqual(Ops[1]->getType())); 6000 return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]); 6001 } 6002 case NEON::BI__builtin_neon_vld3_dup_v: 6003 case NEON::BI__builtin_neon_vld3q_dup_v: { 6004 llvm::Type *PTy = 6005 llvm::PointerType::getUnqual(VTy->getElementType()); 6006 Ops[1] = Builder.CreateBitCast(Ops[1], PTy); 6007 llvm::Type *Tys[2] = { VTy, PTy }; 6008 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld3r, Tys); 6009 Ops[1] = Builder.CreateCall(F, Ops[1], "vld3"); 6010 Ops[0] = Builder.CreateBitCast(Ops[0], 6011 llvm::PointerType::getUnqual(Ops[1]->getType())); 6012 return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]); 6013 } 6014 case NEON::BI__builtin_neon_vld4_dup_v: 6015 case NEON::BI__builtin_neon_vld4q_dup_v: { 6016 llvm::Type *PTy = 6017 llvm::PointerType::getUnqual(VTy->getElementType()); 6018 Ops[1] = Builder.CreateBitCast(Ops[1], PTy); 6019 llvm::Type *Tys[2] = { VTy, PTy }; 6020 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld4r, Tys); 6021 Ops[1] = Builder.CreateCall(F, Ops[1], "vld4"); 6022 Ops[0] = Builder.CreateBitCast(Ops[0], 6023 llvm::PointerType::getUnqual(Ops[1]->getType())); 6024 return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]); 6025 } 6026 case NEON::BI__builtin_neon_vld2_lane_v: 6027 case NEON::BI__builtin_neon_vld2q_lane_v: { 6028 llvm::Type *Tys[2] = { VTy, Ops[1]->getType() }; 6029 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld2lane, Tys); 6030 Ops.push_back(Ops[1]); 6031 Ops.erase(Ops.begin()+1); 6032 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 6033 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 6034 Ops[3] = Builder.CreateZExt(Ops[3], Int64Ty); 6035 Ops[1] = Builder.CreateCall(F, makeArrayRef(Ops).slice(1), "vld2_lane"); 6036 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 6037 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 6038 return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]); 6039 } 6040 case NEON::BI__builtin_neon_vld3_lane_v: 6041 case NEON::BI__builtin_neon_vld3q_lane_v: { 6042 llvm::Type *Tys[2] = { VTy, Ops[1]->getType() }; 6043 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld3lane, Tys); 6044 Ops.push_back(Ops[1]); 6045 Ops.erase(Ops.begin()+1); 6046 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 6047 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 6048 Ops[3] = Builder.CreateBitCast(Ops[3], Ty); 6049 Ops[4] = Builder.CreateZExt(Ops[4], Int64Ty); 6050 Ops[1] = Builder.CreateCall(F, makeArrayRef(Ops).slice(1), "vld3_lane"); 6051 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 6052 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 6053 return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]); 6054 } 6055 case NEON::BI__builtin_neon_vld4_lane_v: 6056 case NEON::BI__builtin_neon_vld4q_lane_v: { 6057 llvm::Type *Tys[2] = { VTy, Ops[1]->getType() }; 6058 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld4lane, Tys); 6059 Ops.push_back(Ops[1]); 6060 Ops.erase(Ops.begin()+1); 6061 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 6062 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 6063 Ops[3] = Builder.CreateBitCast(Ops[3], Ty); 6064 Ops[4] = Builder.CreateBitCast(Ops[4], Ty); 6065 Ops[5] = Builder.CreateZExt(Ops[5], Int64Ty); 6066 Ops[1] = Builder.CreateCall(F, makeArrayRef(Ops).slice(1), "vld4_lane"); 6067 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 6068 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 6069 return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]); 6070 } 6071 case NEON::BI__builtin_neon_vst2_v: 6072 case NEON::BI__builtin_neon_vst2q_v: { 6073 Ops.push_back(Ops[0]); 6074 Ops.erase(Ops.begin()); 6075 llvm::Type *Tys[2] = { VTy, Ops[2]->getType() }; 6076 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_st2, Tys), 6077 Ops, ""); 6078 } 6079 case NEON::BI__builtin_neon_vst2_lane_v: 6080 case NEON::BI__builtin_neon_vst2q_lane_v: { 6081 Ops.push_back(Ops[0]); 6082 Ops.erase(Ops.begin()); 6083 Ops[2] = Builder.CreateZExt(Ops[2], Int64Ty); 6084 llvm::Type *Tys[2] = { VTy, Ops[3]->getType() }; 6085 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_st2lane, Tys), 6086 Ops, ""); 6087 } 6088 case NEON::BI__builtin_neon_vst3_v: 6089 case NEON::BI__builtin_neon_vst3q_v: { 6090 Ops.push_back(Ops[0]); 6091 Ops.erase(Ops.begin()); 6092 llvm::Type *Tys[2] = { VTy, Ops[3]->getType() }; 6093 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_st3, Tys), 6094 Ops, ""); 6095 } 6096 case NEON::BI__builtin_neon_vst3_lane_v: 6097 case NEON::BI__builtin_neon_vst3q_lane_v: { 6098 Ops.push_back(Ops[0]); 6099 Ops.erase(Ops.begin()); 6100 Ops[3] = Builder.CreateZExt(Ops[3], Int64Ty); 6101 llvm::Type *Tys[2] = { VTy, Ops[4]->getType() }; 6102 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_st3lane, Tys), 6103 Ops, ""); 6104 } 6105 case NEON::BI__builtin_neon_vst4_v: 6106 case NEON::BI__builtin_neon_vst4q_v: { 6107 Ops.push_back(Ops[0]); 6108 Ops.erase(Ops.begin()); 6109 llvm::Type *Tys[2] = { VTy, Ops[4]->getType() }; 6110 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_st4, Tys), 6111 Ops, ""); 6112 } 6113 case NEON::BI__builtin_neon_vst4_lane_v: 6114 case NEON::BI__builtin_neon_vst4q_lane_v: { 6115 Ops.push_back(Ops[0]); 6116 Ops.erase(Ops.begin()); 6117 Ops[4] = Builder.CreateZExt(Ops[4], Int64Ty); 6118 llvm::Type *Tys[2] = { VTy, Ops[5]->getType() }; 6119 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_st4lane, Tys), 6120 Ops, ""); 6121 } 6122 case NEON::BI__builtin_neon_vtrn_v: 6123 case NEON::BI__builtin_neon_vtrnq_v: { 6124 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty)); 6125 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 6126 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 6127 Value *SV = nullptr; 6128 6129 for (unsigned vi = 0; vi != 2; ++vi) { 6130 SmallVector<Constant*, 16> Indices; 6131 for (unsigned i = 0, e = VTy->getNumElements(); i != e; i += 2) { 6132 Indices.push_back(ConstantInt::get(Int32Ty, i+vi)); 6133 Indices.push_back(ConstantInt::get(Int32Ty, i+e+vi)); 6134 } 6135 Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ty, Ops[0], vi); 6136 SV = llvm::ConstantVector::get(Indices); 6137 SV = Builder.CreateShuffleVector(Ops[1], Ops[2], SV, "vtrn"); 6138 SV = Builder.CreateDefaultAlignedStore(SV, Addr); 6139 } 6140 return SV; 6141 } 6142 case NEON::BI__builtin_neon_vuzp_v: 6143 case NEON::BI__builtin_neon_vuzpq_v: { 6144 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty)); 6145 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 6146 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 6147 Value *SV = nullptr; 6148 6149 for (unsigned vi = 0; vi != 2; ++vi) { 6150 SmallVector<Constant*, 16> Indices; 6151 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) 6152 Indices.push_back(ConstantInt::get(Int32Ty, 2*i+vi)); 6153 6154 Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ty, Ops[0], vi); 6155 SV = llvm::ConstantVector::get(Indices); 6156 SV = Builder.CreateShuffleVector(Ops[1], Ops[2], SV, "vuzp"); 6157 SV = Builder.CreateDefaultAlignedStore(SV, Addr); 6158 } 6159 return SV; 6160 } 6161 case NEON::BI__builtin_neon_vzip_v: 6162 case NEON::BI__builtin_neon_vzipq_v: { 6163 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty)); 6164 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 6165 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 6166 Value *SV = nullptr; 6167 6168 for (unsigned vi = 0; vi != 2; ++vi) { 6169 SmallVector<Constant*, 16> Indices; 6170 for (unsigned i = 0, e = VTy->getNumElements(); i != e; i += 2) { 6171 Indices.push_back(ConstantInt::get(Int32Ty, (i + vi*e) >> 1)); 6172 Indices.push_back(ConstantInt::get(Int32Ty, ((i + vi*e) >> 1)+e)); 6173 } 6174 Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ty, Ops[0], vi); 6175 SV = llvm::ConstantVector::get(Indices); 6176 SV = Builder.CreateShuffleVector(Ops[1], Ops[2], SV, "vzip"); 6177 SV = Builder.CreateDefaultAlignedStore(SV, Addr); 6178 } 6179 return SV; 6180 } 6181 case NEON::BI__builtin_neon_vqtbl1q_v: { 6182 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbl1, Ty), 6183 Ops, "vtbl1"); 6184 } 6185 case NEON::BI__builtin_neon_vqtbl2q_v: { 6186 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbl2, Ty), 6187 Ops, "vtbl2"); 6188 } 6189 case NEON::BI__builtin_neon_vqtbl3q_v: { 6190 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbl3, Ty), 6191 Ops, "vtbl3"); 6192 } 6193 case NEON::BI__builtin_neon_vqtbl4q_v: { 6194 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbl4, Ty), 6195 Ops, "vtbl4"); 6196 } 6197 case NEON::BI__builtin_neon_vqtbx1q_v: { 6198 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbx1, Ty), 6199 Ops, "vtbx1"); 6200 } 6201 case NEON::BI__builtin_neon_vqtbx2q_v: { 6202 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbx2, Ty), 6203 Ops, "vtbx2"); 6204 } 6205 case NEON::BI__builtin_neon_vqtbx3q_v: { 6206 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbx3, Ty), 6207 Ops, "vtbx3"); 6208 } 6209 case NEON::BI__builtin_neon_vqtbx4q_v: { 6210 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbx4, Ty), 6211 Ops, "vtbx4"); 6212 } 6213 case NEON::BI__builtin_neon_vsqadd_v: 6214 case NEON::BI__builtin_neon_vsqaddq_v: { 6215 Int = Intrinsic::aarch64_neon_usqadd; 6216 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vsqadd"); 6217 } 6218 case NEON::BI__builtin_neon_vuqadd_v: 6219 case NEON::BI__builtin_neon_vuqaddq_v: { 6220 Int = Intrinsic::aarch64_neon_suqadd; 6221 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vuqadd"); 6222 } 6223 } 6224 } 6225 6226 llvm::Value *CodeGenFunction:: 6227 BuildVector(ArrayRef<llvm::Value*> Ops) { 6228 assert((Ops.size() & (Ops.size() - 1)) == 0 && 6229 "Not a power-of-two sized vector!"); 6230 bool AllConstants = true; 6231 for (unsigned i = 0, e = Ops.size(); i != e && AllConstants; ++i) 6232 AllConstants &= isa<Constant>(Ops[i]); 6233 6234 // If this is a constant vector, create a ConstantVector. 6235 if (AllConstants) { 6236 SmallVector<llvm::Constant*, 16> CstOps; 6237 for (unsigned i = 0, e = Ops.size(); i != e; ++i) 6238 CstOps.push_back(cast<Constant>(Ops[i])); 6239 return llvm::ConstantVector::get(CstOps); 6240 } 6241 6242 // Otherwise, insertelement the values to build the vector. 6243 Value *Result = 6244 llvm::UndefValue::get(llvm::VectorType::get(Ops[0]->getType(), Ops.size())); 6245 6246 for (unsigned i = 0, e = Ops.size(); i != e; ++i) 6247 Result = Builder.CreateInsertElement(Result, Ops[i], Builder.getInt32(i)); 6248 6249 return Result; 6250 } 6251 6252 Value *CodeGenFunction::EmitX86BuiltinExpr(unsigned BuiltinID, 6253 const CallExpr *E) { 6254 if (BuiltinID == X86::BI__builtin_ms_va_start || 6255 BuiltinID == X86::BI__builtin_ms_va_end) 6256 return EmitVAStartEnd(EmitMSVAListRef(E->getArg(0)).getPointer(), 6257 BuiltinID == X86::BI__builtin_ms_va_start); 6258 if (BuiltinID == X86::BI__builtin_ms_va_copy) { 6259 // Lower this manually. We can't reliably determine whether or not any 6260 // given va_copy() is for a Win64 va_list from the calling convention 6261 // alone, because it's legal to do this from a System V ABI function. 6262 // With opaque pointer types, we won't have enough information in LLVM 6263 // IR to determine this from the argument types, either. Best to do it 6264 // now, while we have enough information. 6265 Address DestAddr = EmitMSVAListRef(E->getArg(0)); 6266 Address SrcAddr = EmitMSVAListRef(E->getArg(1)); 6267 6268 llvm::Type *BPP = Int8PtrPtrTy; 6269 6270 DestAddr = Address(Builder.CreateBitCast(DestAddr.getPointer(), BPP, "cp"), 6271 DestAddr.getAlignment()); 6272 SrcAddr = Address(Builder.CreateBitCast(SrcAddr.getPointer(), BPP, "ap"), 6273 SrcAddr.getAlignment()); 6274 6275 Value *ArgPtr = Builder.CreateLoad(SrcAddr, "ap.val"); 6276 return Builder.CreateStore(ArgPtr, DestAddr); 6277 } 6278 6279 SmallVector<Value*, 4> Ops; 6280 6281 // Find out if any arguments are required to be integer constant expressions. 6282 unsigned ICEArguments = 0; 6283 ASTContext::GetBuiltinTypeError Error; 6284 getContext().GetBuiltinType(BuiltinID, Error, &ICEArguments); 6285 assert(Error == ASTContext::GE_None && "Should not codegen an error"); 6286 6287 for (unsigned i = 0, e = E->getNumArgs(); i != e; i++) { 6288 // If this is a normal argument, just emit it as a scalar. 6289 if ((ICEArguments & (1 << i)) == 0) { 6290 Ops.push_back(EmitScalarExpr(E->getArg(i))); 6291 continue; 6292 } 6293 6294 // If this is required to be a constant, constant fold it so that we know 6295 // that the generated intrinsic gets a ConstantInt. 6296 llvm::APSInt Result; 6297 bool IsConst = E->getArg(i)->isIntegerConstantExpr(Result, getContext()); 6298 assert(IsConst && "Constant arg isn't actually constant?"); (void)IsConst; 6299 Ops.push_back(llvm::ConstantInt::get(getLLVMContext(), Result)); 6300 } 6301 6302 switch (BuiltinID) { 6303 default: return nullptr; 6304 case X86::BI__builtin_cpu_supports: { 6305 const Expr *FeatureExpr = E->getArg(0)->IgnoreParenCasts(); 6306 StringRef FeatureStr = cast<StringLiteral>(FeatureExpr)->getString(); 6307 6308 // TODO: When/if this becomes more than x86 specific then use a TargetInfo 6309 // based mapping. 6310 // Processor features and mapping to processor feature value. 6311 enum X86Features { 6312 CMOV = 0, 6313 MMX, 6314 POPCNT, 6315 SSE, 6316 SSE2, 6317 SSE3, 6318 SSSE3, 6319 SSE4_1, 6320 SSE4_2, 6321 AVX, 6322 AVX2, 6323 SSE4_A, 6324 FMA4, 6325 XOP, 6326 FMA, 6327 AVX512F, 6328 BMI, 6329 BMI2, 6330 MAX 6331 }; 6332 6333 X86Features Feature = StringSwitch<X86Features>(FeatureStr) 6334 .Case("cmov", X86Features::CMOV) 6335 .Case("mmx", X86Features::MMX) 6336 .Case("popcnt", X86Features::POPCNT) 6337 .Case("sse", X86Features::SSE) 6338 .Case("sse2", X86Features::SSE2) 6339 .Case("sse3", X86Features::SSE3) 6340 .Case("sse4.1", X86Features::SSE4_1) 6341 .Case("sse4.2", X86Features::SSE4_2) 6342 .Case("avx", X86Features::AVX) 6343 .Case("avx2", X86Features::AVX2) 6344 .Case("sse4a", X86Features::SSE4_A) 6345 .Case("fma4", X86Features::FMA4) 6346 .Case("xop", X86Features::XOP) 6347 .Case("fma", X86Features::FMA) 6348 .Case("avx512f", X86Features::AVX512F) 6349 .Case("bmi", X86Features::BMI) 6350 .Case("bmi2", X86Features::BMI2) 6351 .Default(X86Features::MAX); 6352 assert(Feature != X86Features::MAX && "Invalid feature!"); 6353 6354 // Matching the struct layout from the compiler-rt/libgcc structure that is 6355 // filled in: 6356 // unsigned int __cpu_vendor; 6357 // unsigned int __cpu_type; 6358 // unsigned int __cpu_subtype; 6359 // unsigned int __cpu_features[1]; 6360 llvm::Type *STy = llvm::StructType::get( 6361 Int32Ty, Int32Ty, Int32Ty, llvm::ArrayType::get(Int32Ty, 1), nullptr); 6362 6363 // Grab the global __cpu_model. 6364 llvm::Constant *CpuModel = CGM.CreateRuntimeVariable(STy, "__cpu_model"); 6365 6366 // Grab the first (0th) element from the field __cpu_features off of the 6367 // global in the struct STy. 6368 Value *Idxs[] = { 6369 ConstantInt::get(Int32Ty, 0), 6370 ConstantInt::get(Int32Ty, 3), 6371 ConstantInt::get(Int32Ty, 0) 6372 }; 6373 Value *CpuFeatures = Builder.CreateGEP(STy, CpuModel, Idxs); 6374 Value *Features = Builder.CreateAlignedLoad(CpuFeatures, 6375 CharUnits::fromQuantity(4)); 6376 6377 // Check the value of the bit corresponding to the feature requested. 6378 Value *Bitset = Builder.CreateAnd( 6379 Features, llvm::ConstantInt::get(Int32Ty, 1 << Feature)); 6380 return Builder.CreateICmpNE(Bitset, llvm::ConstantInt::get(Int32Ty, 0)); 6381 } 6382 case X86::BI_mm_prefetch: { 6383 Value *Address = Ops[0]; 6384 Value *RW = ConstantInt::get(Int32Ty, 0); 6385 Value *Locality = Ops[1]; 6386 Value *Data = ConstantInt::get(Int32Ty, 1); 6387 Value *F = CGM.getIntrinsic(Intrinsic::prefetch); 6388 return Builder.CreateCall(F, {Address, RW, Locality, Data}); 6389 } 6390 case X86::BI__builtin_ia32_undef128: 6391 case X86::BI__builtin_ia32_undef256: 6392 case X86::BI__builtin_ia32_undef512: 6393 return UndefValue::get(ConvertType(E->getType())); 6394 case X86::BI__builtin_ia32_vec_init_v8qi: 6395 case X86::BI__builtin_ia32_vec_init_v4hi: 6396 case X86::BI__builtin_ia32_vec_init_v2si: 6397 return Builder.CreateBitCast(BuildVector(Ops), 6398 llvm::Type::getX86_MMXTy(getLLVMContext())); 6399 case X86::BI__builtin_ia32_vec_ext_v2si: 6400 return Builder.CreateExtractElement(Ops[0], 6401 llvm::ConstantInt::get(Ops[1]->getType(), 0)); 6402 case X86::BI__builtin_ia32_ldmxcsr: { 6403 Address Tmp = CreateMemTemp(E->getArg(0)->getType()); 6404 Builder.CreateStore(Ops[0], Tmp); 6405 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::x86_sse_ldmxcsr), 6406 Builder.CreateBitCast(Tmp.getPointer(), Int8PtrTy)); 6407 } 6408 case X86::BI__builtin_ia32_stmxcsr: { 6409 Address Tmp = CreateMemTemp(E->getType()); 6410 Builder.CreateCall(CGM.getIntrinsic(Intrinsic::x86_sse_stmxcsr), 6411 Builder.CreateBitCast(Tmp.getPointer(), Int8PtrTy)); 6412 return Builder.CreateLoad(Tmp, "stmxcsr"); 6413 } 6414 case X86::BI__builtin_ia32_xsave: 6415 case X86::BI__builtin_ia32_xsave64: 6416 case X86::BI__builtin_ia32_xrstor: 6417 case X86::BI__builtin_ia32_xrstor64: 6418 case X86::BI__builtin_ia32_xsaveopt: 6419 case X86::BI__builtin_ia32_xsaveopt64: 6420 case X86::BI__builtin_ia32_xrstors: 6421 case X86::BI__builtin_ia32_xrstors64: 6422 case X86::BI__builtin_ia32_xsavec: 6423 case X86::BI__builtin_ia32_xsavec64: 6424 case X86::BI__builtin_ia32_xsaves: 6425 case X86::BI__builtin_ia32_xsaves64: { 6426 Intrinsic::ID ID; 6427 #define INTRINSIC_X86_XSAVE_ID(NAME) \ 6428 case X86::BI__builtin_ia32_##NAME: \ 6429 ID = Intrinsic::x86_##NAME; \ 6430 break 6431 switch (BuiltinID) { 6432 default: llvm_unreachable("Unsupported intrinsic!"); 6433 INTRINSIC_X86_XSAVE_ID(xsave); 6434 INTRINSIC_X86_XSAVE_ID(xsave64); 6435 INTRINSIC_X86_XSAVE_ID(xrstor); 6436 INTRINSIC_X86_XSAVE_ID(xrstor64); 6437 INTRINSIC_X86_XSAVE_ID(xsaveopt); 6438 INTRINSIC_X86_XSAVE_ID(xsaveopt64); 6439 INTRINSIC_X86_XSAVE_ID(xrstors); 6440 INTRINSIC_X86_XSAVE_ID(xrstors64); 6441 INTRINSIC_X86_XSAVE_ID(xsavec); 6442 INTRINSIC_X86_XSAVE_ID(xsavec64); 6443 INTRINSIC_X86_XSAVE_ID(xsaves); 6444 INTRINSIC_X86_XSAVE_ID(xsaves64); 6445 } 6446 #undef INTRINSIC_X86_XSAVE_ID 6447 Value *Mhi = Builder.CreateTrunc( 6448 Builder.CreateLShr(Ops[1], ConstantInt::get(Int64Ty, 32)), Int32Ty); 6449 Value *Mlo = Builder.CreateTrunc(Ops[1], Int32Ty); 6450 Ops[1] = Mhi; 6451 Ops.push_back(Mlo); 6452 return Builder.CreateCall(CGM.getIntrinsic(ID), Ops); 6453 } 6454 case X86::BI__builtin_ia32_storehps: 6455 case X86::BI__builtin_ia32_storelps: { 6456 llvm::Type *PtrTy = llvm::PointerType::getUnqual(Int64Ty); 6457 llvm::Type *VecTy = llvm::VectorType::get(Int64Ty, 2); 6458 6459 // cast val v2i64 6460 Ops[1] = Builder.CreateBitCast(Ops[1], VecTy, "cast"); 6461 6462 // extract (0, 1) 6463 unsigned Index = BuiltinID == X86::BI__builtin_ia32_storelps ? 0 : 1; 6464 llvm::Value *Idx = llvm::ConstantInt::get(SizeTy, Index); 6465 Ops[1] = Builder.CreateExtractElement(Ops[1], Idx, "extract"); 6466 6467 // cast pointer to i64 & store 6468 Ops[0] = Builder.CreateBitCast(Ops[0], PtrTy); 6469 return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]); 6470 } 6471 case X86::BI__builtin_ia32_palignr128: 6472 case X86::BI__builtin_ia32_palignr256: { 6473 unsigned ShiftVal = cast<llvm::ConstantInt>(Ops[2])->getZExtValue(); 6474 6475 unsigned NumElts = 6476 cast<llvm::VectorType>(Ops[0]->getType())->getNumElements(); 6477 assert(NumElts % 16 == 0); 6478 unsigned NumLanes = NumElts / 16; 6479 unsigned NumLaneElts = NumElts / NumLanes; 6480 6481 // If palignr is shifting the pair of vectors more than the size of two 6482 // lanes, emit zero. 6483 if (ShiftVal >= (2 * NumLaneElts)) 6484 return llvm::Constant::getNullValue(ConvertType(E->getType())); 6485 6486 // If palignr is shifting the pair of input vectors more than one lane, 6487 // but less than two lanes, convert to shifting in zeroes. 6488 if (ShiftVal > NumLaneElts) { 6489 ShiftVal -= NumLaneElts; 6490 Ops[1] = Ops[0]; 6491 Ops[0] = llvm::Constant::getNullValue(Ops[0]->getType()); 6492 } 6493 6494 uint32_t Indices[32]; 6495 // 256-bit palignr operates on 128-bit lanes so we need to handle that 6496 for (unsigned l = 0; l != NumElts; l += NumLaneElts) { 6497 for (unsigned i = 0; i != NumLaneElts; ++i) { 6498 unsigned Idx = ShiftVal + i; 6499 if (Idx >= NumLaneElts) 6500 Idx += NumElts - NumLaneElts; // End of lane, switch operand. 6501 Indices[l + i] = Idx + l; 6502 } 6503 } 6504 6505 Value *SV = llvm::ConstantDataVector::get(getLLVMContext(), 6506 makeArrayRef(Indices, NumElts)); 6507 return Builder.CreateShuffleVector(Ops[1], Ops[0], SV, "palignr"); 6508 } 6509 case X86::BI__builtin_ia32_pslldqi256: { 6510 // Shift value is in bits so divide by 8. 6511 unsigned shiftVal = cast<llvm::ConstantInt>(Ops[1])->getZExtValue() >> 3; 6512 6513 // If pslldq is shifting the vector more than 15 bytes, emit zero. 6514 if (shiftVal >= 16) 6515 return llvm::Constant::getNullValue(ConvertType(E->getType())); 6516 6517 uint32_t Indices[32]; 6518 // 256-bit pslldq operates on 128-bit lanes so we need to handle that 6519 for (unsigned l = 0; l != 32; l += 16) { 6520 for (unsigned i = 0; i != 16; ++i) { 6521 unsigned Idx = 32 + i - shiftVal; 6522 if (Idx < 32) Idx -= 16; // end of lane, switch operand. 6523 Indices[l + i] = Idx + l; 6524 } 6525 } 6526 6527 llvm::Type *VecTy = llvm::VectorType::get(Int8Ty, 32); 6528 Ops[0] = Builder.CreateBitCast(Ops[0], VecTy, "cast"); 6529 Value *Zero = llvm::Constant::getNullValue(VecTy); 6530 6531 Value *SV = llvm::ConstantDataVector::get(getLLVMContext(), Indices); 6532 SV = Builder.CreateShuffleVector(Zero, Ops[0], SV, "pslldq"); 6533 llvm::Type *ResultType = ConvertType(E->getType()); 6534 return Builder.CreateBitCast(SV, ResultType, "cast"); 6535 } 6536 case X86::BI__builtin_ia32_psrldqi256: { 6537 // Shift value is in bits so divide by 8. 6538 unsigned shiftVal = cast<llvm::ConstantInt>(Ops[1])->getZExtValue() >> 3; 6539 6540 // If psrldq is shifting the vector more than 15 bytes, emit zero. 6541 if (shiftVal >= 16) 6542 return llvm::Constant::getNullValue(ConvertType(E->getType())); 6543 6544 uint32_t Indices[32]; 6545 // 256-bit psrldq operates on 128-bit lanes so we need to handle that 6546 for (unsigned l = 0; l != 32; l += 16) { 6547 for (unsigned i = 0; i != 16; ++i) { 6548 unsigned Idx = i + shiftVal; 6549 if (Idx >= 16) Idx += 16; // end of lane, switch operand. 6550 Indices[l + i] = Idx + l; 6551 } 6552 } 6553 6554 llvm::Type *VecTy = llvm::VectorType::get(Int8Ty, 32); 6555 Ops[0] = Builder.CreateBitCast(Ops[0], VecTy, "cast"); 6556 Value *Zero = llvm::Constant::getNullValue(VecTy); 6557 6558 Value *SV = llvm::ConstantDataVector::get(getLLVMContext(), Indices); 6559 SV = Builder.CreateShuffleVector(Ops[0], Zero, SV, "psrldq"); 6560 llvm::Type *ResultType = ConvertType(E->getType()); 6561 return Builder.CreateBitCast(SV, ResultType, "cast"); 6562 } 6563 case X86::BI__builtin_ia32_movntps: 6564 case X86::BI__builtin_ia32_movntps256: 6565 case X86::BI__builtin_ia32_movntpd: 6566 case X86::BI__builtin_ia32_movntpd256: 6567 case X86::BI__builtin_ia32_movntdq: 6568 case X86::BI__builtin_ia32_movntdq256: 6569 case X86::BI__builtin_ia32_movnti: 6570 case X86::BI__builtin_ia32_movnti64: { 6571 llvm::MDNode *Node = llvm::MDNode::get( 6572 getLLVMContext(), llvm::ConstantAsMetadata::get(Builder.getInt32(1))); 6573 6574 // Convert the type of the pointer to a pointer to the stored type. 6575 Value *BC = Builder.CreateBitCast(Ops[0], 6576 llvm::PointerType::getUnqual(Ops[1]->getType()), 6577 "cast"); 6578 StoreInst *SI = Builder.CreateDefaultAlignedStore(Ops[1], BC); 6579 SI->setMetadata(CGM.getModule().getMDKindID("nontemporal"), Node); 6580 6581 // If the operand is an integer, we can't assume alignment. Otherwise, 6582 // assume natural alignment. 6583 QualType ArgTy = E->getArg(1)->getType(); 6584 unsigned Align; 6585 if (ArgTy->isIntegerType()) 6586 Align = 1; 6587 else 6588 Align = getContext().getTypeSizeInChars(ArgTy).getQuantity(); 6589 SI->setAlignment(Align); 6590 return SI; 6591 } 6592 // 3DNow! 6593 case X86::BI__builtin_ia32_pswapdsf: 6594 case X86::BI__builtin_ia32_pswapdsi: { 6595 llvm::Type *MMXTy = llvm::Type::getX86_MMXTy(getLLVMContext()); 6596 Ops[0] = Builder.CreateBitCast(Ops[0], MMXTy, "cast"); 6597 llvm::Function *F = CGM.getIntrinsic(Intrinsic::x86_3dnowa_pswapd); 6598 return Builder.CreateCall(F, Ops, "pswapd"); 6599 } 6600 case X86::BI__builtin_ia32_rdrand16_step: 6601 case X86::BI__builtin_ia32_rdrand32_step: 6602 case X86::BI__builtin_ia32_rdrand64_step: 6603 case X86::BI__builtin_ia32_rdseed16_step: 6604 case X86::BI__builtin_ia32_rdseed32_step: 6605 case X86::BI__builtin_ia32_rdseed64_step: { 6606 Intrinsic::ID ID; 6607 switch (BuiltinID) { 6608 default: llvm_unreachable("Unsupported intrinsic!"); 6609 case X86::BI__builtin_ia32_rdrand16_step: 6610 ID = Intrinsic::x86_rdrand_16; 6611 break; 6612 case X86::BI__builtin_ia32_rdrand32_step: 6613 ID = Intrinsic::x86_rdrand_32; 6614 break; 6615 case X86::BI__builtin_ia32_rdrand64_step: 6616 ID = Intrinsic::x86_rdrand_64; 6617 break; 6618 case X86::BI__builtin_ia32_rdseed16_step: 6619 ID = Intrinsic::x86_rdseed_16; 6620 break; 6621 case X86::BI__builtin_ia32_rdseed32_step: 6622 ID = Intrinsic::x86_rdseed_32; 6623 break; 6624 case X86::BI__builtin_ia32_rdseed64_step: 6625 ID = Intrinsic::x86_rdseed_64; 6626 break; 6627 } 6628 6629 Value *Call = Builder.CreateCall(CGM.getIntrinsic(ID)); 6630 Builder.CreateDefaultAlignedStore(Builder.CreateExtractValue(Call, 0), 6631 Ops[0]); 6632 return Builder.CreateExtractValue(Call, 1); 6633 } 6634 // SSE comparison intrisics 6635 case X86::BI__builtin_ia32_cmpeqps: 6636 case X86::BI__builtin_ia32_cmpltps: 6637 case X86::BI__builtin_ia32_cmpleps: 6638 case X86::BI__builtin_ia32_cmpunordps: 6639 case X86::BI__builtin_ia32_cmpneqps: 6640 case X86::BI__builtin_ia32_cmpnltps: 6641 case X86::BI__builtin_ia32_cmpnleps: 6642 case X86::BI__builtin_ia32_cmpordps: 6643 case X86::BI__builtin_ia32_cmpeqss: 6644 case X86::BI__builtin_ia32_cmpltss: 6645 case X86::BI__builtin_ia32_cmpless: 6646 case X86::BI__builtin_ia32_cmpunordss: 6647 case X86::BI__builtin_ia32_cmpneqss: 6648 case X86::BI__builtin_ia32_cmpnltss: 6649 case X86::BI__builtin_ia32_cmpnless: 6650 case X86::BI__builtin_ia32_cmpordss: 6651 case X86::BI__builtin_ia32_cmpeqpd: 6652 case X86::BI__builtin_ia32_cmpltpd: 6653 case X86::BI__builtin_ia32_cmplepd: 6654 case X86::BI__builtin_ia32_cmpunordpd: 6655 case X86::BI__builtin_ia32_cmpneqpd: 6656 case X86::BI__builtin_ia32_cmpnltpd: 6657 case X86::BI__builtin_ia32_cmpnlepd: 6658 case X86::BI__builtin_ia32_cmpordpd: 6659 case X86::BI__builtin_ia32_cmpeqsd: 6660 case X86::BI__builtin_ia32_cmpltsd: 6661 case X86::BI__builtin_ia32_cmplesd: 6662 case X86::BI__builtin_ia32_cmpunordsd: 6663 case X86::BI__builtin_ia32_cmpneqsd: 6664 case X86::BI__builtin_ia32_cmpnltsd: 6665 case X86::BI__builtin_ia32_cmpnlesd: 6666 case X86::BI__builtin_ia32_cmpordsd: 6667 // These exist so that the builtin that takes an immediate can be bounds 6668 // checked by clang to avoid passing bad immediates to the backend. Since 6669 // AVX has a larger immediate than SSE we would need separate builtins to 6670 // do the different bounds checking. Rather than create a clang specific 6671 // SSE only builtin, this implements eight separate builtins to match gcc 6672 // implementation. 6673 6674 // Choose the immediate. 6675 unsigned Imm; 6676 switch (BuiltinID) { 6677 default: llvm_unreachable("Unsupported intrinsic!"); 6678 case X86::BI__builtin_ia32_cmpeqps: 6679 case X86::BI__builtin_ia32_cmpeqss: 6680 case X86::BI__builtin_ia32_cmpeqpd: 6681 case X86::BI__builtin_ia32_cmpeqsd: 6682 Imm = 0; 6683 break; 6684 case X86::BI__builtin_ia32_cmpltps: 6685 case X86::BI__builtin_ia32_cmpltss: 6686 case X86::BI__builtin_ia32_cmpltpd: 6687 case X86::BI__builtin_ia32_cmpltsd: 6688 Imm = 1; 6689 break; 6690 case X86::BI__builtin_ia32_cmpleps: 6691 case X86::BI__builtin_ia32_cmpless: 6692 case X86::BI__builtin_ia32_cmplepd: 6693 case X86::BI__builtin_ia32_cmplesd: 6694 Imm = 2; 6695 break; 6696 case X86::BI__builtin_ia32_cmpunordps: 6697 case X86::BI__builtin_ia32_cmpunordss: 6698 case X86::BI__builtin_ia32_cmpunordpd: 6699 case X86::BI__builtin_ia32_cmpunordsd: 6700 Imm = 3; 6701 break; 6702 case X86::BI__builtin_ia32_cmpneqps: 6703 case X86::BI__builtin_ia32_cmpneqss: 6704 case X86::BI__builtin_ia32_cmpneqpd: 6705 case X86::BI__builtin_ia32_cmpneqsd: 6706 Imm = 4; 6707 break; 6708 case X86::BI__builtin_ia32_cmpnltps: 6709 case X86::BI__builtin_ia32_cmpnltss: 6710 case X86::BI__builtin_ia32_cmpnltpd: 6711 case X86::BI__builtin_ia32_cmpnltsd: 6712 Imm = 5; 6713 break; 6714 case X86::BI__builtin_ia32_cmpnleps: 6715 case X86::BI__builtin_ia32_cmpnless: 6716 case X86::BI__builtin_ia32_cmpnlepd: 6717 case X86::BI__builtin_ia32_cmpnlesd: 6718 Imm = 6; 6719 break; 6720 case X86::BI__builtin_ia32_cmpordps: 6721 case X86::BI__builtin_ia32_cmpordss: 6722 case X86::BI__builtin_ia32_cmpordpd: 6723 case X86::BI__builtin_ia32_cmpordsd: 6724 Imm = 7; 6725 break; 6726 } 6727 6728 // Choose the intrinsic ID. 6729 const char *name; 6730 Intrinsic::ID ID; 6731 switch (BuiltinID) { 6732 default: llvm_unreachable("Unsupported intrinsic!"); 6733 case X86::BI__builtin_ia32_cmpeqps: 6734 case X86::BI__builtin_ia32_cmpltps: 6735 case X86::BI__builtin_ia32_cmpleps: 6736 case X86::BI__builtin_ia32_cmpunordps: 6737 case X86::BI__builtin_ia32_cmpneqps: 6738 case X86::BI__builtin_ia32_cmpnltps: 6739 case X86::BI__builtin_ia32_cmpnleps: 6740 case X86::BI__builtin_ia32_cmpordps: 6741 name = "cmpps"; 6742 ID = Intrinsic::x86_sse_cmp_ps; 6743 break; 6744 case X86::BI__builtin_ia32_cmpeqss: 6745 case X86::BI__builtin_ia32_cmpltss: 6746 case X86::BI__builtin_ia32_cmpless: 6747 case X86::BI__builtin_ia32_cmpunordss: 6748 case X86::BI__builtin_ia32_cmpneqss: 6749 case X86::BI__builtin_ia32_cmpnltss: 6750 case X86::BI__builtin_ia32_cmpnless: 6751 case X86::BI__builtin_ia32_cmpordss: 6752 name = "cmpss"; 6753 ID = Intrinsic::x86_sse_cmp_ss; 6754 break; 6755 case X86::BI__builtin_ia32_cmpeqpd: 6756 case X86::BI__builtin_ia32_cmpltpd: 6757 case X86::BI__builtin_ia32_cmplepd: 6758 case X86::BI__builtin_ia32_cmpunordpd: 6759 case X86::BI__builtin_ia32_cmpneqpd: 6760 case X86::BI__builtin_ia32_cmpnltpd: 6761 case X86::BI__builtin_ia32_cmpnlepd: 6762 case X86::BI__builtin_ia32_cmpordpd: 6763 name = "cmppd"; 6764 ID = Intrinsic::x86_sse2_cmp_pd; 6765 break; 6766 case X86::BI__builtin_ia32_cmpeqsd: 6767 case X86::BI__builtin_ia32_cmpltsd: 6768 case X86::BI__builtin_ia32_cmplesd: 6769 case X86::BI__builtin_ia32_cmpunordsd: 6770 case X86::BI__builtin_ia32_cmpneqsd: 6771 case X86::BI__builtin_ia32_cmpnltsd: 6772 case X86::BI__builtin_ia32_cmpnlesd: 6773 case X86::BI__builtin_ia32_cmpordsd: 6774 name = "cmpsd"; 6775 ID = Intrinsic::x86_sse2_cmp_sd; 6776 break; 6777 } 6778 6779 Ops.push_back(llvm::ConstantInt::get(Int8Ty, Imm)); 6780 llvm::Function *F = CGM.getIntrinsic(ID); 6781 return Builder.CreateCall(F, Ops, name); 6782 } 6783 } 6784 6785 6786 Value *CodeGenFunction::EmitPPCBuiltinExpr(unsigned BuiltinID, 6787 const CallExpr *E) { 6788 SmallVector<Value*, 4> Ops; 6789 6790 for (unsigned i = 0, e = E->getNumArgs(); i != e; i++) 6791 Ops.push_back(EmitScalarExpr(E->getArg(i))); 6792 6793 Intrinsic::ID ID = Intrinsic::not_intrinsic; 6794 6795 switch (BuiltinID) { 6796 default: return nullptr; 6797 6798 // __builtin_ppc_get_timebase is GCC 4.8+'s PowerPC-specific name for what we 6799 // call __builtin_readcyclecounter. 6800 case PPC::BI__builtin_ppc_get_timebase: 6801 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::readcyclecounter)); 6802 6803 // vec_ld, vec_lvsl, vec_lvsr 6804 case PPC::BI__builtin_altivec_lvx: 6805 case PPC::BI__builtin_altivec_lvxl: 6806 case PPC::BI__builtin_altivec_lvebx: 6807 case PPC::BI__builtin_altivec_lvehx: 6808 case PPC::BI__builtin_altivec_lvewx: 6809 case PPC::BI__builtin_altivec_lvsl: 6810 case PPC::BI__builtin_altivec_lvsr: 6811 case PPC::BI__builtin_vsx_lxvd2x: 6812 case PPC::BI__builtin_vsx_lxvw4x: 6813 { 6814 Ops[1] = Builder.CreateBitCast(Ops[1], Int8PtrTy); 6815 6816 Ops[0] = Builder.CreateGEP(Ops[1], Ops[0]); 6817 Ops.pop_back(); 6818 6819 switch (BuiltinID) { 6820 default: llvm_unreachable("Unsupported ld/lvsl/lvsr intrinsic!"); 6821 case PPC::BI__builtin_altivec_lvx: 6822 ID = Intrinsic::ppc_altivec_lvx; 6823 break; 6824 case PPC::BI__builtin_altivec_lvxl: 6825 ID = Intrinsic::ppc_altivec_lvxl; 6826 break; 6827 case PPC::BI__builtin_altivec_lvebx: 6828 ID = Intrinsic::ppc_altivec_lvebx; 6829 break; 6830 case PPC::BI__builtin_altivec_lvehx: 6831 ID = Intrinsic::ppc_altivec_lvehx; 6832 break; 6833 case PPC::BI__builtin_altivec_lvewx: 6834 ID = Intrinsic::ppc_altivec_lvewx; 6835 break; 6836 case PPC::BI__builtin_altivec_lvsl: 6837 ID = Intrinsic::ppc_altivec_lvsl; 6838 break; 6839 case PPC::BI__builtin_altivec_lvsr: 6840 ID = Intrinsic::ppc_altivec_lvsr; 6841 break; 6842 case PPC::BI__builtin_vsx_lxvd2x: 6843 ID = Intrinsic::ppc_vsx_lxvd2x; 6844 break; 6845 case PPC::BI__builtin_vsx_lxvw4x: 6846 ID = Intrinsic::ppc_vsx_lxvw4x; 6847 break; 6848 } 6849 llvm::Function *F = CGM.getIntrinsic(ID); 6850 return Builder.CreateCall(F, Ops, ""); 6851 } 6852 6853 // vec_st 6854 case PPC::BI__builtin_altivec_stvx: 6855 case PPC::BI__builtin_altivec_stvxl: 6856 case PPC::BI__builtin_altivec_stvebx: 6857 case PPC::BI__builtin_altivec_stvehx: 6858 case PPC::BI__builtin_altivec_stvewx: 6859 case PPC::BI__builtin_vsx_stxvd2x: 6860 case PPC::BI__builtin_vsx_stxvw4x: 6861 { 6862 Ops[2] = Builder.CreateBitCast(Ops[2], Int8PtrTy); 6863 Ops[1] = Builder.CreateGEP(Ops[2], Ops[1]); 6864 Ops.pop_back(); 6865 6866 switch (BuiltinID) { 6867 default: llvm_unreachable("Unsupported st intrinsic!"); 6868 case PPC::BI__builtin_altivec_stvx: 6869 ID = Intrinsic::ppc_altivec_stvx; 6870 break; 6871 case PPC::BI__builtin_altivec_stvxl: 6872 ID = Intrinsic::ppc_altivec_stvxl; 6873 break; 6874 case PPC::BI__builtin_altivec_stvebx: 6875 ID = Intrinsic::ppc_altivec_stvebx; 6876 break; 6877 case PPC::BI__builtin_altivec_stvehx: 6878 ID = Intrinsic::ppc_altivec_stvehx; 6879 break; 6880 case PPC::BI__builtin_altivec_stvewx: 6881 ID = Intrinsic::ppc_altivec_stvewx; 6882 break; 6883 case PPC::BI__builtin_vsx_stxvd2x: 6884 ID = Intrinsic::ppc_vsx_stxvd2x; 6885 break; 6886 case PPC::BI__builtin_vsx_stxvw4x: 6887 ID = Intrinsic::ppc_vsx_stxvw4x; 6888 break; 6889 } 6890 llvm::Function *F = CGM.getIntrinsic(ID); 6891 return Builder.CreateCall(F, Ops, ""); 6892 } 6893 // Square root 6894 case PPC::BI__builtin_vsx_xvsqrtsp: 6895 case PPC::BI__builtin_vsx_xvsqrtdp: { 6896 llvm::Type *ResultType = ConvertType(E->getType()); 6897 Value *X = EmitScalarExpr(E->getArg(0)); 6898 ID = Intrinsic::sqrt; 6899 llvm::Function *F = CGM.getIntrinsic(ID, ResultType); 6900 return Builder.CreateCall(F, X); 6901 } 6902 // Count leading zeros 6903 case PPC::BI__builtin_altivec_vclzb: 6904 case PPC::BI__builtin_altivec_vclzh: 6905 case PPC::BI__builtin_altivec_vclzw: 6906 case PPC::BI__builtin_altivec_vclzd: { 6907 llvm::Type *ResultType = ConvertType(E->getType()); 6908 Value *X = EmitScalarExpr(E->getArg(0)); 6909 Value *Undef = ConstantInt::get(Builder.getInt1Ty(), false); 6910 Function *F = CGM.getIntrinsic(Intrinsic::ctlz, ResultType); 6911 return Builder.CreateCall(F, {X, Undef}); 6912 } 6913 // Copy sign 6914 case PPC::BI__builtin_vsx_xvcpsgnsp: 6915 case PPC::BI__builtin_vsx_xvcpsgndp: { 6916 llvm::Type *ResultType = ConvertType(E->getType()); 6917 Value *X = EmitScalarExpr(E->getArg(0)); 6918 Value *Y = EmitScalarExpr(E->getArg(1)); 6919 ID = Intrinsic::copysign; 6920 llvm::Function *F = CGM.getIntrinsic(ID, ResultType); 6921 return Builder.CreateCall(F, {X, Y}); 6922 } 6923 // Rounding/truncation 6924 case PPC::BI__builtin_vsx_xvrspip: 6925 case PPC::BI__builtin_vsx_xvrdpip: 6926 case PPC::BI__builtin_vsx_xvrdpim: 6927 case PPC::BI__builtin_vsx_xvrspim: 6928 case PPC::BI__builtin_vsx_xvrdpi: 6929 case PPC::BI__builtin_vsx_xvrspi: 6930 case PPC::BI__builtin_vsx_xvrdpic: 6931 case PPC::BI__builtin_vsx_xvrspic: 6932 case PPC::BI__builtin_vsx_xvrdpiz: 6933 case PPC::BI__builtin_vsx_xvrspiz: { 6934 llvm::Type *ResultType = ConvertType(E->getType()); 6935 Value *X = EmitScalarExpr(E->getArg(0)); 6936 if (BuiltinID == PPC::BI__builtin_vsx_xvrdpim || 6937 BuiltinID == PPC::BI__builtin_vsx_xvrspim) 6938 ID = Intrinsic::floor; 6939 else if (BuiltinID == PPC::BI__builtin_vsx_xvrdpi || 6940 BuiltinID == PPC::BI__builtin_vsx_xvrspi) 6941 ID = Intrinsic::round; 6942 else if (BuiltinID == PPC::BI__builtin_vsx_xvrdpic || 6943 BuiltinID == PPC::BI__builtin_vsx_xvrspic) 6944 ID = Intrinsic::nearbyint; 6945 else if (BuiltinID == PPC::BI__builtin_vsx_xvrdpip || 6946 BuiltinID == PPC::BI__builtin_vsx_xvrspip) 6947 ID = Intrinsic::ceil; 6948 else if (BuiltinID == PPC::BI__builtin_vsx_xvrdpiz || 6949 BuiltinID == PPC::BI__builtin_vsx_xvrspiz) 6950 ID = Intrinsic::trunc; 6951 llvm::Function *F = CGM.getIntrinsic(ID, ResultType); 6952 return Builder.CreateCall(F, X); 6953 } 6954 // FMA variations 6955 case PPC::BI__builtin_vsx_xvmaddadp: 6956 case PPC::BI__builtin_vsx_xvmaddasp: 6957 case PPC::BI__builtin_vsx_xvnmaddadp: 6958 case PPC::BI__builtin_vsx_xvnmaddasp: 6959 case PPC::BI__builtin_vsx_xvmsubadp: 6960 case PPC::BI__builtin_vsx_xvmsubasp: 6961 case PPC::BI__builtin_vsx_xvnmsubadp: 6962 case PPC::BI__builtin_vsx_xvnmsubasp: { 6963 llvm::Type *ResultType = ConvertType(E->getType()); 6964 Value *X = EmitScalarExpr(E->getArg(0)); 6965 Value *Y = EmitScalarExpr(E->getArg(1)); 6966 Value *Z = EmitScalarExpr(E->getArg(2)); 6967 Value *Zero = llvm::ConstantFP::getZeroValueForNegation(ResultType); 6968 llvm::Function *F = CGM.getIntrinsic(Intrinsic::fma, ResultType); 6969 switch (BuiltinID) { 6970 case PPC::BI__builtin_vsx_xvmaddadp: 6971 case PPC::BI__builtin_vsx_xvmaddasp: 6972 return Builder.CreateCall(F, {X, Y, Z}); 6973 case PPC::BI__builtin_vsx_xvnmaddadp: 6974 case PPC::BI__builtin_vsx_xvnmaddasp: 6975 return Builder.CreateFSub(Zero, 6976 Builder.CreateCall(F, {X, Y, Z}), "sub"); 6977 case PPC::BI__builtin_vsx_xvmsubadp: 6978 case PPC::BI__builtin_vsx_xvmsubasp: 6979 return Builder.CreateCall(F, 6980 {X, Y, Builder.CreateFSub(Zero, Z, "sub")}); 6981 case PPC::BI__builtin_vsx_xvnmsubadp: 6982 case PPC::BI__builtin_vsx_xvnmsubasp: 6983 Value *FsubRes = 6984 Builder.CreateCall(F, {X, Y, Builder.CreateFSub(Zero, Z, "sub")}); 6985 return Builder.CreateFSub(Zero, FsubRes, "sub"); 6986 } 6987 llvm_unreachable("Unknown FMA operation"); 6988 return nullptr; // Suppress no-return warning 6989 } 6990 } 6991 } 6992 6993 // Emit an intrinsic that has 1 float or double. 6994 static Value *emitUnaryFPBuiltin(CodeGenFunction &CGF, 6995 const CallExpr *E, 6996 unsigned IntrinsicID) { 6997 llvm::Value *Src0 = CGF.EmitScalarExpr(E->getArg(0)); 6998 6999 Value *F = CGF.CGM.getIntrinsic(IntrinsicID, Src0->getType()); 7000 return CGF.Builder.CreateCall(F, Src0); 7001 } 7002 7003 // Emit an intrinsic that has 3 float or double operands. 7004 static Value *emitTernaryFPBuiltin(CodeGenFunction &CGF, 7005 const CallExpr *E, 7006 unsigned IntrinsicID) { 7007 llvm::Value *Src0 = CGF.EmitScalarExpr(E->getArg(0)); 7008 llvm::Value *Src1 = CGF.EmitScalarExpr(E->getArg(1)); 7009 llvm::Value *Src2 = CGF.EmitScalarExpr(E->getArg(2)); 7010 7011 Value *F = CGF.CGM.getIntrinsic(IntrinsicID, Src0->getType()); 7012 return CGF.Builder.CreateCall(F, {Src0, Src1, Src2}); 7013 } 7014 7015 // Emit an intrinsic that has 1 float or double operand, and 1 integer. 7016 static Value *emitFPIntBuiltin(CodeGenFunction &CGF, 7017 const CallExpr *E, 7018 unsigned IntrinsicID) { 7019 llvm::Value *Src0 = CGF.EmitScalarExpr(E->getArg(0)); 7020 llvm::Value *Src1 = CGF.EmitScalarExpr(E->getArg(1)); 7021 7022 Value *F = CGF.CGM.getIntrinsic(IntrinsicID, Src0->getType()); 7023 return CGF.Builder.CreateCall(F, {Src0, Src1}); 7024 } 7025 7026 Value *CodeGenFunction::EmitAMDGPUBuiltinExpr(unsigned BuiltinID, 7027 const CallExpr *E) { 7028 switch (BuiltinID) { 7029 case AMDGPU::BI__builtin_amdgcn_div_scale: 7030 case AMDGPU::BI__builtin_amdgcn_div_scalef: { 7031 // Translate from the intrinsics's struct return to the builtin's out 7032 // argument. 7033 7034 Address FlagOutPtr = EmitPointerWithAlignment(E->getArg(3)); 7035 7036 llvm::Value *X = EmitScalarExpr(E->getArg(0)); 7037 llvm::Value *Y = EmitScalarExpr(E->getArg(1)); 7038 llvm::Value *Z = EmitScalarExpr(E->getArg(2)); 7039 7040 llvm::Value *Callee = CGM.getIntrinsic(Intrinsic::amdgcn_div_scale, 7041 X->getType()); 7042 7043 llvm::Value *Tmp = Builder.CreateCall(Callee, {X, Y, Z}); 7044 7045 llvm::Value *Result = Builder.CreateExtractValue(Tmp, 0); 7046 llvm::Value *Flag = Builder.CreateExtractValue(Tmp, 1); 7047 7048 llvm::Type *RealFlagType 7049 = FlagOutPtr.getPointer()->getType()->getPointerElementType(); 7050 7051 llvm::Value *FlagExt = Builder.CreateZExt(Flag, RealFlagType); 7052 Builder.CreateStore(FlagExt, FlagOutPtr); 7053 return Result; 7054 } 7055 case AMDGPU::BI__builtin_amdgcn_div_fmas: 7056 case AMDGPU::BI__builtin_amdgcn_div_fmasf: { 7057 llvm::Value *Src0 = EmitScalarExpr(E->getArg(0)); 7058 llvm::Value *Src1 = EmitScalarExpr(E->getArg(1)); 7059 llvm::Value *Src2 = EmitScalarExpr(E->getArg(2)); 7060 llvm::Value *Src3 = EmitScalarExpr(E->getArg(3)); 7061 7062 llvm::Value *F = CGM.getIntrinsic(Intrinsic::amdgcn_div_fmas, 7063 Src0->getType()); 7064 llvm::Value *Src3ToBool = Builder.CreateIsNotNull(Src3); 7065 return Builder.CreateCall(F, {Src0, Src1, Src2, Src3ToBool}); 7066 } 7067 case AMDGPU::BI__builtin_amdgcn_div_fixup: 7068 case AMDGPU::BI__builtin_amdgcn_div_fixupf: 7069 return emitTernaryFPBuiltin(*this, E, Intrinsic::amdgcn_div_fixup); 7070 case AMDGPU::BI__builtin_amdgcn_trig_preop: 7071 case AMDGPU::BI__builtin_amdgcn_trig_preopf: 7072 return emitFPIntBuiltin(*this, E, Intrinsic::amdgcn_trig_preop); 7073 case AMDGPU::BI__builtin_amdgcn_rcp: 7074 case AMDGPU::BI__builtin_amdgcn_rcpf: 7075 return emitUnaryFPBuiltin(*this, E, Intrinsic::amdgcn_rcp); 7076 case AMDGPU::BI__builtin_amdgcn_rsq: 7077 case AMDGPU::BI__builtin_amdgcn_rsqf: 7078 return emitUnaryFPBuiltin(*this, E, Intrinsic::amdgcn_rsq); 7079 case AMDGPU::BI__builtin_amdgcn_rsq_clamped: 7080 case AMDGPU::BI__builtin_amdgcn_rsq_clampedf: 7081 return emitUnaryFPBuiltin(*this, E, Intrinsic::amdgcn_rsq_clamped); 7082 case AMDGPU::BI__builtin_amdgcn_ldexp: 7083 case AMDGPU::BI__builtin_amdgcn_ldexpf: 7084 return emitFPIntBuiltin(*this, E, Intrinsic::amdgcn_ldexp); 7085 case AMDGPU::BI__builtin_amdgcn_class: 7086 case AMDGPU::BI__builtin_amdgcn_classf: 7087 return emitFPIntBuiltin(*this, E, Intrinsic::amdgcn_class); 7088 7089 // Legacy amdgpu prefix 7090 case AMDGPU::BI__builtin_amdgpu_rsq: 7091 case AMDGPU::BI__builtin_amdgpu_rsqf: { 7092 if (getTarget().getTriple().getArch() == Triple::amdgcn) 7093 return emitUnaryFPBuiltin(*this, E, Intrinsic::amdgcn_rsq); 7094 return emitUnaryFPBuiltin(*this, E, Intrinsic::r600_rsq); 7095 } 7096 case AMDGPU::BI__builtin_amdgpu_ldexp: 7097 case AMDGPU::BI__builtin_amdgpu_ldexpf: { 7098 if (getTarget().getTriple().getArch() == Triple::amdgcn) 7099 return emitFPIntBuiltin(*this, E, Intrinsic::amdgcn_ldexp); 7100 return emitFPIntBuiltin(*this, E, Intrinsic::AMDGPU_ldexp); 7101 } 7102 default: 7103 return nullptr; 7104 } 7105 } 7106 7107 /// Handle a SystemZ function in which the final argument is a pointer 7108 /// to an int that receives the post-instruction CC value. At the LLVM level 7109 /// this is represented as a function that returns a {result, cc} pair. 7110 static Value *EmitSystemZIntrinsicWithCC(CodeGenFunction &CGF, 7111 unsigned IntrinsicID, 7112 const CallExpr *E) { 7113 unsigned NumArgs = E->getNumArgs() - 1; 7114 SmallVector<Value *, 8> Args(NumArgs); 7115 for (unsigned I = 0; I < NumArgs; ++I) 7116 Args[I] = CGF.EmitScalarExpr(E->getArg(I)); 7117 Address CCPtr = CGF.EmitPointerWithAlignment(E->getArg(NumArgs)); 7118 Value *F = CGF.CGM.getIntrinsic(IntrinsicID); 7119 Value *Call = CGF.Builder.CreateCall(F, Args); 7120 Value *CC = CGF.Builder.CreateExtractValue(Call, 1); 7121 CGF.Builder.CreateStore(CC, CCPtr); 7122 return CGF.Builder.CreateExtractValue(Call, 0); 7123 } 7124 7125 Value *CodeGenFunction::EmitSystemZBuiltinExpr(unsigned BuiltinID, 7126 const CallExpr *E) { 7127 switch (BuiltinID) { 7128 case SystemZ::BI__builtin_tbegin: { 7129 Value *TDB = EmitScalarExpr(E->getArg(0)); 7130 Value *Control = llvm::ConstantInt::get(Int32Ty, 0xff0c); 7131 Value *F = CGM.getIntrinsic(Intrinsic::s390_tbegin); 7132 return Builder.CreateCall(F, {TDB, Control}); 7133 } 7134 case SystemZ::BI__builtin_tbegin_nofloat: { 7135 Value *TDB = EmitScalarExpr(E->getArg(0)); 7136 Value *Control = llvm::ConstantInt::get(Int32Ty, 0xff0c); 7137 Value *F = CGM.getIntrinsic(Intrinsic::s390_tbegin_nofloat); 7138 return Builder.CreateCall(F, {TDB, Control}); 7139 } 7140 case SystemZ::BI__builtin_tbeginc: { 7141 Value *TDB = llvm::ConstantPointerNull::get(Int8PtrTy); 7142 Value *Control = llvm::ConstantInt::get(Int32Ty, 0xff08); 7143 Value *F = CGM.getIntrinsic(Intrinsic::s390_tbeginc); 7144 return Builder.CreateCall(F, {TDB, Control}); 7145 } 7146 case SystemZ::BI__builtin_tabort: { 7147 Value *Data = EmitScalarExpr(E->getArg(0)); 7148 Value *F = CGM.getIntrinsic(Intrinsic::s390_tabort); 7149 return Builder.CreateCall(F, Builder.CreateSExt(Data, Int64Ty, "tabort")); 7150 } 7151 case SystemZ::BI__builtin_non_tx_store: { 7152 Value *Address = EmitScalarExpr(E->getArg(0)); 7153 Value *Data = EmitScalarExpr(E->getArg(1)); 7154 Value *F = CGM.getIntrinsic(Intrinsic::s390_ntstg); 7155 return Builder.CreateCall(F, {Data, Address}); 7156 } 7157 7158 // Vector builtins. Note that most vector builtins are mapped automatically 7159 // to target-specific LLVM intrinsics. The ones handled specially here can 7160 // be represented via standard LLVM IR, which is preferable to enable common 7161 // LLVM optimizations. 7162 7163 case SystemZ::BI__builtin_s390_vpopctb: 7164 case SystemZ::BI__builtin_s390_vpopcth: 7165 case SystemZ::BI__builtin_s390_vpopctf: 7166 case SystemZ::BI__builtin_s390_vpopctg: { 7167 llvm::Type *ResultType = ConvertType(E->getType()); 7168 Value *X = EmitScalarExpr(E->getArg(0)); 7169 Function *F = CGM.getIntrinsic(Intrinsic::ctpop, ResultType); 7170 return Builder.CreateCall(F, X); 7171 } 7172 7173 case SystemZ::BI__builtin_s390_vclzb: 7174 case SystemZ::BI__builtin_s390_vclzh: 7175 case SystemZ::BI__builtin_s390_vclzf: 7176 case SystemZ::BI__builtin_s390_vclzg: { 7177 llvm::Type *ResultType = ConvertType(E->getType()); 7178 Value *X = EmitScalarExpr(E->getArg(0)); 7179 Value *Undef = ConstantInt::get(Builder.getInt1Ty(), false); 7180 Function *F = CGM.getIntrinsic(Intrinsic::ctlz, ResultType); 7181 return Builder.CreateCall(F, {X, Undef}); 7182 } 7183 7184 case SystemZ::BI__builtin_s390_vctzb: 7185 case SystemZ::BI__builtin_s390_vctzh: 7186 case SystemZ::BI__builtin_s390_vctzf: 7187 case SystemZ::BI__builtin_s390_vctzg: { 7188 llvm::Type *ResultType = ConvertType(E->getType()); 7189 Value *X = EmitScalarExpr(E->getArg(0)); 7190 Value *Undef = ConstantInt::get(Builder.getInt1Ty(), false); 7191 Function *F = CGM.getIntrinsic(Intrinsic::cttz, ResultType); 7192 return Builder.CreateCall(F, {X, Undef}); 7193 } 7194 7195 case SystemZ::BI__builtin_s390_vfsqdb: { 7196 llvm::Type *ResultType = ConvertType(E->getType()); 7197 Value *X = EmitScalarExpr(E->getArg(0)); 7198 Function *F = CGM.getIntrinsic(Intrinsic::sqrt, ResultType); 7199 return Builder.CreateCall(F, X); 7200 } 7201 case SystemZ::BI__builtin_s390_vfmadb: { 7202 llvm::Type *ResultType = ConvertType(E->getType()); 7203 Value *X = EmitScalarExpr(E->getArg(0)); 7204 Value *Y = EmitScalarExpr(E->getArg(1)); 7205 Value *Z = EmitScalarExpr(E->getArg(2)); 7206 Function *F = CGM.getIntrinsic(Intrinsic::fma, ResultType); 7207 return Builder.CreateCall(F, {X, Y, Z}); 7208 } 7209 case SystemZ::BI__builtin_s390_vfmsdb: { 7210 llvm::Type *ResultType = ConvertType(E->getType()); 7211 Value *X = EmitScalarExpr(E->getArg(0)); 7212 Value *Y = EmitScalarExpr(E->getArg(1)); 7213 Value *Z = EmitScalarExpr(E->getArg(2)); 7214 Value *Zero = llvm::ConstantFP::getZeroValueForNegation(ResultType); 7215 Function *F = CGM.getIntrinsic(Intrinsic::fma, ResultType); 7216 return Builder.CreateCall(F, {X, Y, Builder.CreateFSub(Zero, Z, "sub")}); 7217 } 7218 case SystemZ::BI__builtin_s390_vflpdb: { 7219 llvm::Type *ResultType = ConvertType(E->getType()); 7220 Value *X = EmitScalarExpr(E->getArg(0)); 7221 Function *F = CGM.getIntrinsic(Intrinsic::fabs, ResultType); 7222 return Builder.CreateCall(F, X); 7223 } 7224 case SystemZ::BI__builtin_s390_vflndb: { 7225 llvm::Type *ResultType = ConvertType(E->getType()); 7226 Value *X = EmitScalarExpr(E->getArg(0)); 7227 Value *Zero = llvm::ConstantFP::getZeroValueForNegation(ResultType); 7228 Function *F = CGM.getIntrinsic(Intrinsic::fabs, ResultType); 7229 return Builder.CreateFSub(Zero, Builder.CreateCall(F, X), "sub"); 7230 } 7231 case SystemZ::BI__builtin_s390_vfidb: { 7232 llvm::Type *ResultType = ConvertType(E->getType()); 7233 Value *X = EmitScalarExpr(E->getArg(0)); 7234 // Constant-fold the M4 and M5 mask arguments. 7235 llvm::APSInt M4, M5; 7236 bool IsConstM4 = E->getArg(1)->isIntegerConstantExpr(M4, getContext()); 7237 bool IsConstM5 = E->getArg(2)->isIntegerConstantExpr(M5, getContext()); 7238 assert(IsConstM4 && IsConstM5 && "Constant arg isn't actually constant?"); 7239 (void)IsConstM4; (void)IsConstM5; 7240 // Check whether this instance of vfidb can be represented via a LLVM 7241 // standard intrinsic. We only support some combinations of M4 and M5. 7242 Intrinsic::ID ID = Intrinsic::not_intrinsic; 7243 switch (M4.getZExtValue()) { 7244 default: break; 7245 case 0: // IEEE-inexact exception allowed 7246 switch (M5.getZExtValue()) { 7247 default: break; 7248 case 0: ID = Intrinsic::rint; break; 7249 } 7250 break; 7251 case 4: // IEEE-inexact exception suppressed 7252 switch (M5.getZExtValue()) { 7253 default: break; 7254 case 0: ID = Intrinsic::nearbyint; break; 7255 case 1: ID = Intrinsic::round; break; 7256 case 5: ID = Intrinsic::trunc; break; 7257 case 6: ID = Intrinsic::ceil; break; 7258 case 7: ID = Intrinsic::floor; break; 7259 } 7260 break; 7261 } 7262 if (ID != Intrinsic::not_intrinsic) { 7263 Function *F = CGM.getIntrinsic(ID, ResultType); 7264 return Builder.CreateCall(F, X); 7265 } 7266 Function *F = CGM.getIntrinsic(Intrinsic::s390_vfidb); 7267 Value *M4Value = llvm::ConstantInt::get(getLLVMContext(), M4); 7268 Value *M5Value = llvm::ConstantInt::get(getLLVMContext(), M5); 7269 return Builder.CreateCall(F, {X, M4Value, M5Value}); 7270 } 7271 7272 // Vector intrisincs that output the post-instruction CC value. 7273 7274 #define INTRINSIC_WITH_CC(NAME) \ 7275 case SystemZ::BI__builtin_##NAME: \ 7276 return EmitSystemZIntrinsicWithCC(*this, Intrinsic::NAME, E) 7277 7278 INTRINSIC_WITH_CC(s390_vpkshs); 7279 INTRINSIC_WITH_CC(s390_vpksfs); 7280 INTRINSIC_WITH_CC(s390_vpksgs); 7281 7282 INTRINSIC_WITH_CC(s390_vpklshs); 7283 INTRINSIC_WITH_CC(s390_vpklsfs); 7284 INTRINSIC_WITH_CC(s390_vpklsgs); 7285 7286 INTRINSIC_WITH_CC(s390_vceqbs); 7287 INTRINSIC_WITH_CC(s390_vceqhs); 7288 INTRINSIC_WITH_CC(s390_vceqfs); 7289 INTRINSIC_WITH_CC(s390_vceqgs); 7290 7291 INTRINSIC_WITH_CC(s390_vchbs); 7292 INTRINSIC_WITH_CC(s390_vchhs); 7293 INTRINSIC_WITH_CC(s390_vchfs); 7294 INTRINSIC_WITH_CC(s390_vchgs); 7295 7296 INTRINSIC_WITH_CC(s390_vchlbs); 7297 INTRINSIC_WITH_CC(s390_vchlhs); 7298 INTRINSIC_WITH_CC(s390_vchlfs); 7299 INTRINSIC_WITH_CC(s390_vchlgs); 7300 7301 INTRINSIC_WITH_CC(s390_vfaebs); 7302 INTRINSIC_WITH_CC(s390_vfaehs); 7303 INTRINSIC_WITH_CC(s390_vfaefs); 7304 7305 INTRINSIC_WITH_CC(s390_vfaezbs); 7306 INTRINSIC_WITH_CC(s390_vfaezhs); 7307 INTRINSIC_WITH_CC(s390_vfaezfs); 7308 7309 INTRINSIC_WITH_CC(s390_vfeebs); 7310 INTRINSIC_WITH_CC(s390_vfeehs); 7311 INTRINSIC_WITH_CC(s390_vfeefs); 7312 7313 INTRINSIC_WITH_CC(s390_vfeezbs); 7314 INTRINSIC_WITH_CC(s390_vfeezhs); 7315 INTRINSIC_WITH_CC(s390_vfeezfs); 7316 7317 INTRINSIC_WITH_CC(s390_vfenebs); 7318 INTRINSIC_WITH_CC(s390_vfenehs); 7319 INTRINSIC_WITH_CC(s390_vfenefs); 7320 7321 INTRINSIC_WITH_CC(s390_vfenezbs); 7322 INTRINSIC_WITH_CC(s390_vfenezhs); 7323 INTRINSIC_WITH_CC(s390_vfenezfs); 7324 7325 INTRINSIC_WITH_CC(s390_vistrbs); 7326 INTRINSIC_WITH_CC(s390_vistrhs); 7327 INTRINSIC_WITH_CC(s390_vistrfs); 7328 7329 INTRINSIC_WITH_CC(s390_vstrcbs); 7330 INTRINSIC_WITH_CC(s390_vstrchs); 7331 INTRINSIC_WITH_CC(s390_vstrcfs); 7332 7333 INTRINSIC_WITH_CC(s390_vstrczbs); 7334 INTRINSIC_WITH_CC(s390_vstrczhs); 7335 INTRINSIC_WITH_CC(s390_vstrczfs); 7336 7337 INTRINSIC_WITH_CC(s390_vfcedbs); 7338 INTRINSIC_WITH_CC(s390_vfchdbs); 7339 INTRINSIC_WITH_CC(s390_vfchedbs); 7340 7341 INTRINSIC_WITH_CC(s390_vftcidb); 7342 7343 #undef INTRINSIC_WITH_CC 7344 7345 default: 7346 return nullptr; 7347 } 7348 } 7349 7350 Value *CodeGenFunction::EmitNVPTXBuiltinExpr(unsigned BuiltinID, 7351 const CallExpr *E) { 7352 switch (BuiltinID) { 7353 case NVPTX::BI__nvvm_atom_add_gen_i: 7354 case NVPTX::BI__nvvm_atom_add_gen_l: 7355 case NVPTX::BI__nvvm_atom_add_gen_ll: 7356 return MakeBinaryAtomicValue(*this, llvm::AtomicRMWInst::Add, E); 7357 7358 case NVPTX::BI__nvvm_atom_sub_gen_i: 7359 case NVPTX::BI__nvvm_atom_sub_gen_l: 7360 case NVPTX::BI__nvvm_atom_sub_gen_ll: 7361 return MakeBinaryAtomicValue(*this, llvm::AtomicRMWInst::Sub, E); 7362 7363 case NVPTX::BI__nvvm_atom_and_gen_i: 7364 case NVPTX::BI__nvvm_atom_and_gen_l: 7365 case NVPTX::BI__nvvm_atom_and_gen_ll: 7366 return MakeBinaryAtomicValue(*this, llvm::AtomicRMWInst::And, E); 7367 7368 case NVPTX::BI__nvvm_atom_or_gen_i: 7369 case NVPTX::BI__nvvm_atom_or_gen_l: 7370 case NVPTX::BI__nvvm_atom_or_gen_ll: 7371 return MakeBinaryAtomicValue(*this, llvm::AtomicRMWInst::Or, E); 7372 7373 case NVPTX::BI__nvvm_atom_xor_gen_i: 7374 case NVPTX::BI__nvvm_atom_xor_gen_l: 7375 case NVPTX::BI__nvvm_atom_xor_gen_ll: 7376 return MakeBinaryAtomicValue(*this, llvm::AtomicRMWInst::Xor, E); 7377 7378 case NVPTX::BI__nvvm_atom_xchg_gen_i: 7379 case NVPTX::BI__nvvm_atom_xchg_gen_l: 7380 case NVPTX::BI__nvvm_atom_xchg_gen_ll: 7381 return MakeBinaryAtomicValue(*this, llvm::AtomicRMWInst::Xchg, E); 7382 7383 case NVPTX::BI__nvvm_atom_max_gen_i: 7384 case NVPTX::BI__nvvm_atom_max_gen_l: 7385 case NVPTX::BI__nvvm_atom_max_gen_ll: 7386 return MakeBinaryAtomicValue(*this, llvm::AtomicRMWInst::Max, E); 7387 7388 case NVPTX::BI__nvvm_atom_max_gen_ui: 7389 case NVPTX::BI__nvvm_atom_max_gen_ul: 7390 case NVPTX::BI__nvvm_atom_max_gen_ull: 7391 return MakeBinaryAtomicValue(*this, llvm::AtomicRMWInst::UMax, E); 7392 7393 case NVPTX::BI__nvvm_atom_min_gen_i: 7394 case NVPTX::BI__nvvm_atom_min_gen_l: 7395 case NVPTX::BI__nvvm_atom_min_gen_ll: 7396 return MakeBinaryAtomicValue(*this, llvm::AtomicRMWInst::Min, E); 7397 7398 case NVPTX::BI__nvvm_atom_min_gen_ui: 7399 case NVPTX::BI__nvvm_atom_min_gen_ul: 7400 case NVPTX::BI__nvvm_atom_min_gen_ull: 7401 return MakeBinaryAtomicValue(*this, llvm::AtomicRMWInst::UMin, E); 7402 7403 case NVPTX::BI__nvvm_atom_cas_gen_i: 7404 case NVPTX::BI__nvvm_atom_cas_gen_l: 7405 case NVPTX::BI__nvvm_atom_cas_gen_ll: 7406 // __nvvm_atom_cas_gen_* should return the old value rather than the 7407 // success flag. 7408 return MakeAtomicCmpXchgValue(*this, E, /*ReturnBool=*/false); 7409 7410 case NVPTX::BI__nvvm_atom_add_gen_f: { 7411 Value *Ptr = EmitScalarExpr(E->getArg(0)); 7412 Value *Val = EmitScalarExpr(E->getArg(1)); 7413 // atomicrmw only deals with integer arguments so we need to use 7414 // LLVM's nvvm_atomic_load_add_f32 intrinsic for that. 7415 Value *FnALAF32 = 7416 CGM.getIntrinsic(Intrinsic::nvvm_atomic_load_add_f32, Ptr->getType()); 7417 return Builder.CreateCall(FnALAF32, {Ptr, Val}); 7418 } 7419 7420 default: 7421 return nullptr; 7422 } 7423 } 7424 7425 Value *CodeGenFunction::EmitWebAssemblyBuiltinExpr(unsigned BuiltinID, 7426 const CallExpr *E) { 7427 switch (BuiltinID) { 7428 case WebAssembly::BI__builtin_wasm_memory_size: { 7429 llvm::Type *ResultType = ConvertType(E->getType()); 7430 Value *Callee = CGM.getIntrinsic(Intrinsic::wasm_memory_size, ResultType); 7431 return Builder.CreateCall(Callee); 7432 } 7433 case WebAssembly::BI__builtin_wasm_grow_memory: { 7434 Value *X = EmitScalarExpr(E->getArg(0)); 7435 Value *Callee = CGM.getIntrinsic(Intrinsic::wasm_grow_memory, X->getType()); 7436 return Builder.CreateCall(Callee, X); 7437 } 7438 7439 default: 7440 return nullptr; 7441 } 7442 } 7443