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 // Emit a simple mangled intrinsic that has 1 argument and a return type 292 // matching the argument type. 293 static Value *emitUnaryBuiltin(CodeGenFunction &CGF, 294 const CallExpr *E, 295 unsigned IntrinsicID) { 296 llvm::Value *Src0 = CGF.EmitScalarExpr(E->getArg(0)); 297 298 Value *F = CGF.CGM.getIntrinsic(IntrinsicID, Src0->getType()); 299 return CGF.Builder.CreateCall(F, Src0); 300 } 301 302 // Emit an intrinsic that has 3 float or double operands. 303 static Value *emitTernaryFPBuiltin(CodeGenFunction &CGF, 304 const CallExpr *E, 305 unsigned IntrinsicID) { 306 llvm::Value *Src0 = CGF.EmitScalarExpr(E->getArg(0)); 307 llvm::Value *Src1 = CGF.EmitScalarExpr(E->getArg(1)); 308 llvm::Value *Src2 = CGF.EmitScalarExpr(E->getArg(2)); 309 310 Value *F = CGF.CGM.getIntrinsic(IntrinsicID, Src0->getType()); 311 return CGF.Builder.CreateCall(F, {Src0, Src1, Src2}); 312 } 313 314 // Emit an intrinsic that has 1 float or double operand, and 1 integer. 315 static Value *emitFPIntBuiltin(CodeGenFunction &CGF, 316 const CallExpr *E, 317 unsigned IntrinsicID) { 318 llvm::Value *Src0 = CGF.EmitScalarExpr(E->getArg(0)); 319 llvm::Value *Src1 = CGF.EmitScalarExpr(E->getArg(1)); 320 321 Value *F = CGF.CGM.getIntrinsic(IntrinsicID, Src0->getType()); 322 return CGF.Builder.CreateCall(F, {Src0, Src1}); 323 } 324 325 namespace { 326 struct WidthAndSignedness { 327 unsigned Width; 328 bool Signed; 329 }; 330 } 331 332 static WidthAndSignedness 333 getIntegerWidthAndSignedness(const clang::ASTContext &context, 334 const clang::QualType Type) { 335 assert(Type->isIntegerType() && "Given type is not an integer."); 336 unsigned Width = Type->isBooleanType() ? 1 : context.getTypeInfo(Type).Width; 337 bool Signed = Type->isSignedIntegerType(); 338 return {Width, Signed}; 339 } 340 341 // Given one or more integer types, this function produces an integer type that 342 // encompasses them: any value in one of the given types could be expressed in 343 // the encompassing type. 344 static struct WidthAndSignedness 345 EncompassingIntegerType(ArrayRef<struct WidthAndSignedness> Types) { 346 assert(Types.size() > 0 && "Empty list of types."); 347 348 // If any of the given types is signed, we must return a signed type. 349 bool Signed = false; 350 for (const auto &Type : Types) { 351 Signed |= Type.Signed; 352 } 353 354 // The encompassing type must have a width greater than or equal to the width 355 // of the specified types. Aditionally, if the encompassing type is signed, 356 // its width must be strictly greater than the width of any unsigned types 357 // given. 358 unsigned Width = 0; 359 for (const auto &Type : Types) { 360 unsigned MinWidth = Type.Width + (Signed && !Type.Signed); 361 if (Width < MinWidth) { 362 Width = MinWidth; 363 } 364 } 365 366 return {Width, Signed}; 367 } 368 369 Value *CodeGenFunction::EmitVAStartEnd(Value *ArgValue, bool IsStart) { 370 llvm::Type *DestType = Int8PtrTy; 371 if (ArgValue->getType() != DestType) 372 ArgValue = 373 Builder.CreateBitCast(ArgValue, DestType, ArgValue->getName().data()); 374 375 Intrinsic::ID inst = IsStart ? Intrinsic::vastart : Intrinsic::vaend; 376 return Builder.CreateCall(CGM.getIntrinsic(inst), ArgValue); 377 } 378 379 /// Checks if using the result of __builtin_object_size(p, @p From) in place of 380 /// __builtin_object_size(p, @p To) is correct 381 static bool areBOSTypesCompatible(int From, int To) { 382 // Note: Our __builtin_object_size implementation currently treats Type=0 and 383 // Type=2 identically. Encoding this implementation detail here may make 384 // improving __builtin_object_size difficult in the future, so it's omitted. 385 return From == To || (From == 0 && To == 1) || (From == 3 && To == 2); 386 } 387 388 static llvm::Value * 389 getDefaultBuiltinObjectSizeResult(unsigned Type, llvm::IntegerType *ResType) { 390 return ConstantInt::get(ResType, (Type & 2) ? 0 : -1, /*isSigned=*/true); 391 } 392 393 llvm::Value * 394 CodeGenFunction::evaluateOrEmitBuiltinObjectSize(const Expr *E, unsigned Type, 395 llvm::IntegerType *ResType) { 396 uint64_t ObjectSize; 397 if (!E->tryEvaluateObjectSize(ObjectSize, getContext(), Type)) 398 return emitBuiltinObjectSize(E, Type, ResType); 399 return ConstantInt::get(ResType, ObjectSize, /*isSigned=*/true); 400 } 401 402 /// Returns a Value corresponding to the size of the given expression. 403 /// This Value may be either of the following: 404 /// - A llvm::Argument (if E is a param with the pass_object_size attribute on 405 /// it) 406 /// - A call to the @llvm.objectsize intrinsic 407 llvm::Value * 408 CodeGenFunction::emitBuiltinObjectSize(const Expr *E, unsigned Type, 409 llvm::IntegerType *ResType) { 410 // We need to reference an argument if the pointer is a parameter with the 411 // pass_object_size attribute. 412 if (auto *D = dyn_cast<DeclRefExpr>(E->IgnoreParenImpCasts())) { 413 auto *Param = dyn_cast<ParmVarDecl>(D->getDecl()); 414 auto *PS = D->getDecl()->getAttr<PassObjectSizeAttr>(); 415 if (Param != nullptr && PS != nullptr && 416 areBOSTypesCompatible(PS->getType(), Type)) { 417 auto Iter = SizeArguments.find(Param); 418 assert(Iter != SizeArguments.end()); 419 420 const ImplicitParamDecl *D = Iter->second; 421 auto DIter = LocalDeclMap.find(D); 422 assert(DIter != LocalDeclMap.end()); 423 424 return EmitLoadOfScalar(DIter->second, /*volatile=*/false, 425 getContext().getSizeType(), E->getLocStart()); 426 } 427 } 428 429 // LLVM can't handle Type=3 appropriately, and __builtin_object_size shouldn't 430 // evaluate E for side-effects. In either case, we shouldn't lower to 431 // @llvm.objectsize. 432 if (Type == 3 || E->HasSideEffects(getContext())) 433 return getDefaultBuiltinObjectSizeResult(Type, ResType); 434 435 // LLVM only supports 0 and 2, make sure that we pass along that 436 // as a boolean. 437 auto *CI = ConstantInt::get(Builder.getInt1Ty(), (Type & 2) >> 1); 438 // FIXME: Get right address space. 439 llvm::Type *Tys[] = {ResType, Builder.getInt8PtrTy(0)}; 440 Value *F = CGM.getIntrinsic(Intrinsic::objectsize, Tys); 441 return Builder.CreateCall(F, {EmitScalarExpr(E), CI}); 442 } 443 444 RValue CodeGenFunction::EmitBuiltinExpr(const FunctionDecl *FD, 445 unsigned BuiltinID, const CallExpr *E, 446 ReturnValueSlot ReturnValue) { 447 // See if we can constant fold this builtin. If so, don't emit it at all. 448 Expr::EvalResult Result; 449 if (E->EvaluateAsRValue(Result, CGM.getContext()) && 450 !Result.hasSideEffects()) { 451 if (Result.Val.isInt()) 452 return RValue::get(llvm::ConstantInt::get(getLLVMContext(), 453 Result.Val.getInt())); 454 if (Result.Val.isFloat()) 455 return RValue::get(llvm::ConstantFP::get(getLLVMContext(), 456 Result.Val.getFloat())); 457 } 458 459 switch (BuiltinID) { 460 default: break; // Handle intrinsics and libm functions below. 461 case Builtin::BI__builtin___CFStringMakeConstantString: 462 case Builtin::BI__builtin___NSStringMakeConstantString: 463 return RValue::get(CGM.EmitConstantExpr(E, E->getType(), nullptr)); 464 case Builtin::BI__builtin_stdarg_start: 465 case Builtin::BI__builtin_va_start: 466 case Builtin::BI__va_start: 467 case Builtin::BI__builtin_va_end: 468 return RValue::get( 469 EmitVAStartEnd(BuiltinID == Builtin::BI__va_start 470 ? EmitScalarExpr(E->getArg(0)) 471 : EmitVAListRef(E->getArg(0)).getPointer(), 472 BuiltinID != Builtin::BI__builtin_va_end)); 473 case Builtin::BI__builtin_va_copy: { 474 Value *DstPtr = EmitVAListRef(E->getArg(0)).getPointer(); 475 Value *SrcPtr = EmitVAListRef(E->getArg(1)).getPointer(); 476 477 llvm::Type *Type = Int8PtrTy; 478 479 DstPtr = Builder.CreateBitCast(DstPtr, Type); 480 SrcPtr = Builder.CreateBitCast(SrcPtr, Type); 481 return RValue::get(Builder.CreateCall(CGM.getIntrinsic(Intrinsic::vacopy), 482 {DstPtr, SrcPtr})); 483 } 484 case Builtin::BI__builtin_abs: 485 case Builtin::BI__builtin_labs: 486 case Builtin::BI__builtin_llabs: { 487 Value *ArgValue = EmitScalarExpr(E->getArg(0)); 488 489 Value *NegOp = Builder.CreateNeg(ArgValue, "neg"); 490 Value *CmpResult = 491 Builder.CreateICmpSGE(ArgValue, 492 llvm::Constant::getNullValue(ArgValue->getType()), 493 "abscond"); 494 Value *Result = 495 Builder.CreateSelect(CmpResult, ArgValue, NegOp, "abs"); 496 497 return RValue::get(Result); 498 } 499 case Builtin::BI__builtin_fabs: 500 case Builtin::BI__builtin_fabsf: 501 case Builtin::BI__builtin_fabsl: { 502 Value *Arg1 = EmitScalarExpr(E->getArg(0)); 503 Value *Result = EmitFAbs(*this, Arg1); 504 return RValue::get(Result); 505 } 506 case Builtin::BI__builtin_fmod: 507 case Builtin::BI__builtin_fmodf: 508 case Builtin::BI__builtin_fmodl: { 509 Value *Arg1 = EmitScalarExpr(E->getArg(0)); 510 Value *Arg2 = EmitScalarExpr(E->getArg(1)); 511 Value *Result = Builder.CreateFRem(Arg1, Arg2, "fmod"); 512 return RValue::get(Result); 513 } 514 515 case Builtin::BI__builtin_conj: 516 case Builtin::BI__builtin_conjf: 517 case Builtin::BI__builtin_conjl: { 518 ComplexPairTy ComplexVal = EmitComplexExpr(E->getArg(0)); 519 Value *Real = ComplexVal.first; 520 Value *Imag = ComplexVal.second; 521 Value *Zero = 522 Imag->getType()->isFPOrFPVectorTy() 523 ? llvm::ConstantFP::getZeroValueForNegation(Imag->getType()) 524 : llvm::Constant::getNullValue(Imag->getType()); 525 526 Imag = Builder.CreateFSub(Zero, Imag, "sub"); 527 return RValue::getComplex(std::make_pair(Real, Imag)); 528 } 529 case Builtin::BI__builtin_creal: 530 case Builtin::BI__builtin_crealf: 531 case Builtin::BI__builtin_creall: 532 case Builtin::BIcreal: 533 case Builtin::BIcrealf: 534 case Builtin::BIcreall: { 535 ComplexPairTy ComplexVal = EmitComplexExpr(E->getArg(0)); 536 return RValue::get(ComplexVal.first); 537 } 538 539 case Builtin::BI__builtin_cimag: 540 case Builtin::BI__builtin_cimagf: 541 case Builtin::BI__builtin_cimagl: 542 case Builtin::BIcimag: 543 case Builtin::BIcimagf: 544 case Builtin::BIcimagl: { 545 ComplexPairTy ComplexVal = EmitComplexExpr(E->getArg(0)); 546 return RValue::get(ComplexVal.second); 547 } 548 549 case Builtin::BI__builtin_ctzs: 550 case Builtin::BI__builtin_ctz: 551 case Builtin::BI__builtin_ctzl: 552 case Builtin::BI__builtin_ctzll: { 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 *ZeroUndef = Builder.getInt1(getTarget().isCLZForZeroUndef()); 560 Value *Result = Builder.CreateCall(F, {ArgValue, ZeroUndef}); 561 if (Result->getType() != ResultType) 562 Result = Builder.CreateIntCast(Result, ResultType, /*isSigned*/true, 563 "cast"); 564 return RValue::get(Result); 565 } 566 case Builtin::BI__builtin_clzs: 567 case Builtin::BI__builtin_clz: 568 case Builtin::BI__builtin_clzl: 569 case Builtin::BI__builtin_clzll: { 570 Value *ArgValue = EmitScalarExpr(E->getArg(0)); 571 572 llvm::Type *ArgType = ArgValue->getType(); 573 Value *F = CGM.getIntrinsic(Intrinsic::ctlz, ArgType); 574 575 llvm::Type *ResultType = ConvertType(E->getType()); 576 Value *ZeroUndef = Builder.getInt1(getTarget().isCLZForZeroUndef()); 577 Value *Result = Builder.CreateCall(F, {ArgValue, ZeroUndef}); 578 if (Result->getType() != ResultType) 579 Result = Builder.CreateIntCast(Result, ResultType, /*isSigned*/true, 580 "cast"); 581 return RValue::get(Result); 582 } 583 case Builtin::BI__builtin_ffs: 584 case Builtin::BI__builtin_ffsl: 585 case Builtin::BI__builtin_ffsll: { 586 // ffs(x) -> x ? cttz(x) + 1 : 0 587 Value *ArgValue = EmitScalarExpr(E->getArg(0)); 588 589 llvm::Type *ArgType = ArgValue->getType(); 590 Value *F = CGM.getIntrinsic(Intrinsic::cttz, ArgType); 591 592 llvm::Type *ResultType = ConvertType(E->getType()); 593 Value *Tmp = 594 Builder.CreateAdd(Builder.CreateCall(F, {ArgValue, Builder.getTrue()}), 595 llvm::ConstantInt::get(ArgType, 1)); 596 Value *Zero = llvm::Constant::getNullValue(ArgType); 597 Value *IsZero = Builder.CreateICmpEQ(ArgValue, Zero, "iszero"); 598 Value *Result = Builder.CreateSelect(IsZero, Zero, Tmp, "ffs"); 599 if (Result->getType() != ResultType) 600 Result = Builder.CreateIntCast(Result, ResultType, /*isSigned*/true, 601 "cast"); 602 return RValue::get(Result); 603 } 604 case Builtin::BI__builtin_parity: 605 case Builtin::BI__builtin_parityl: 606 case Builtin::BI__builtin_parityll: { 607 // parity(x) -> ctpop(x) & 1 608 Value *ArgValue = EmitScalarExpr(E->getArg(0)); 609 610 llvm::Type *ArgType = ArgValue->getType(); 611 Value *F = CGM.getIntrinsic(Intrinsic::ctpop, ArgType); 612 613 llvm::Type *ResultType = ConvertType(E->getType()); 614 Value *Tmp = Builder.CreateCall(F, ArgValue); 615 Value *Result = Builder.CreateAnd(Tmp, llvm::ConstantInt::get(ArgType, 1)); 616 if (Result->getType() != ResultType) 617 Result = Builder.CreateIntCast(Result, ResultType, /*isSigned*/true, 618 "cast"); 619 return RValue::get(Result); 620 } 621 case Builtin::BI__builtin_popcount: 622 case Builtin::BI__builtin_popcountl: 623 case Builtin::BI__builtin_popcountll: { 624 Value *ArgValue = EmitScalarExpr(E->getArg(0)); 625 626 llvm::Type *ArgType = ArgValue->getType(); 627 Value *F = CGM.getIntrinsic(Intrinsic::ctpop, ArgType); 628 629 llvm::Type *ResultType = ConvertType(E->getType()); 630 Value *Result = Builder.CreateCall(F, ArgValue); 631 if (Result->getType() != ResultType) 632 Result = Builder.CreateIntCast(Result, ResultType, /*isSigned*/true, 633 "cast"); 634 return RValue::get(Result); 635 } 636 case Builtin::BI__builtin_unpredictable: { 637 // Always return the argument of __builtin_unpredictable. LLVM does not 638 // handle this builtin. Metadata for this builtin should be added directly 639 // to instructions such as branches or switches that use it. 640 return RValue::get(EmitScalarExpr(E->getArg(0))); 641 } 642 case Builtin::BI__builtin_expect: { 643 Value *ArgValue = EmitScalarExpr(E->getArg(0)); 644 llvm::Type *ArgType = ArgValue->getType(); 645 646 Value *ExpectedValue = EmitScalarExpr(E->getArg(1)); 647 // Don't generate llvm.expect on -O0 as the backend won't use it for 648 // anything. 649 // Note, we still IRGen ExpectedValue because it could have side-effects. 650 if (CGM.getCodeGenOpts().OptimizationLevel == 0) 651 return RValue::get(ArgValue); 652 653 Value *FnExpect = CGM.getIntrinsic(Intrinsic::expect, ArgType); 654 Value *Result = 655 Builder.CreateCall(FnExpect, {ArgValue, ExpectedValue}, "expval"); 656 return RValue::get(Result); 657 } 658 case Builtin::BI__builtin_assume_aligned: { 659 Value *PtrValue = EmitScalarExpr(E->getArg(0)); 660 Value *OffsetValue = 661 (E->getNumArgs() > 2) ? EmitScalarExpr(E->getArg(2)) : nullptr; 662 663 Value *AlignmentValue = EmitScalarExpr(E->getArg(1)); 664 ConstantInt *AlignmentCI = cast<ConstantInt>(AlignmentValue); 665 unsigned Alignment = (unsigned) AlignmentCI->getZExtValue(); 666 667 EmitAlignmentAssumption(PtrValue, Alignment, OffsetValue); 668 return RValue::get(PtrValue); 669 } 670 case Builtin::BI__assume: 671 case Builtin::BI__builtin_assume: { 672 if (E->getArg(0)->HasSideEffects(getContext())) 673 return RValue::get(nullptr); 674 675 Value *ArgValue = EmitScalarExpr(E->getArg(0)); 676 Value *FnAssume = CGM.getIntrinsic(Intrinsic::assume); 677 return RValue::get(Builder.CreateCall(FnAssume, ArgValue)); 678 } 679 case Builtin::BI__builtin_bswap16: 680 case Builtin::BI__builtin_bswap32: 681 case Builtin::BI__builtin_bswap64: { 682 return RValue::get(emitUnaryBuiltin(*this, E, Intrinsic::bswap)); 683 } 684 case Builtin::BI__builtin_bitreverse16: 685 case Builtin::BI__builtin_bitreverse32: 686 case Builtin::BI__builtin_bitreverse64: { 687 return RValue::get(emitUnaryBuiltin(*this, E, Intrinsic::bitreverse)); 688 } 689 case Builtin::BI__builtin_object_size: { 690 unsigned Type = 691 E->getArg(1)->EvaluateKnownConstInt(getContext()).getZExtValue(); 692 auto *ResType = cast<llvm::IntegerType>(ConvertType(E->getType())); 693 694 // We pass this builtin onto the optimizer so that it can figure out the 695 // object size in more complex cases. 696 return RValue::get(emitBuiltinObjectSize(E->getArg(0), Type, ResType)); 697 } 698 case Builtin::BI__builtin_prefetch: { 699 Value *Locality, *RW, *Address = EmitScalarExpr(E->getArg(0)); 700 // FIXME: Technically these constants should of type 'int', yes? 701 RW = (E->getNumArgs() > 1) ? EmitScalarExpr(E->getArg(1)) : 702 llvm::ConstantInt::get(Int32Ty, 0); 703 Locality = (E->getNumArgs() > 2) ? EmitScalarExpr(E->getArg(2)) : 704 llvm::ConstantInt::get(Int32Ty, 3); 705 Value *Data = llvm::ConstantInt::get(Int32Ty, 1); 706 Value *F = CGM.getIntrinsic(Intrinsic::prefetch); 707 return RValue::get(Builder.CreateCall(F, {Address, RW, Locality, Data})); 708 } 709 case Builtin::BI__builtin_readcyclecounter: { 710 Value *F = CGM.getIntrinsic(Intrinsic::readcyclecounter); 711 return RValue::get(Builder.CreateCall(F)); 712 } 713 case Builtin::BI__builtin___clear_cache: { 714 Value *Begin = EmitScalarExpr(E->getArg(0)); 715 Value *End = EmitScalarExpr(E->getArg(1)); 716 Value *F = CGM.getIntrinsic(Intrinsic::clear_cache); 717 return RValue::get(Builder.CreateCall(F, {Begin, End})); 718 } 719 case Builtin::BI__builtin_trap: 720 return RValue::get(EmitTrapCall(Intrinsic::trap)); 721 case Builtin::BI__debugbreak: 722 return RValue::get(EmitTrapCall(Intrinsic::debugtrap)); 723 case Builtin::BI__builtin_unreachable: { 724 if (SanOpts.has(SanitizerKind::Unreachable)) { 725 SanitizerScope SanScope(this); 726 EmitCheck(std::make_pair(static_cast<llvm::Value *>(Builder.getFalse()), 727 SanitizerKind::Unreachable), 728 "builtin_unreachable", EmitCheckSourceLocation(E->getExprLoc()), 729 None); 730 } else 731 Builder.CreateUnreachable(); 732 733 // We do need to preserve an insertion point. 734 EmitBlock(createBasicBlock("unreachable.cont")); 735 736 return RValue::get(nullptr); 737 } 738 739 case Builtin::BI__builtin_powi: 740 case Builtin::BI__builtin_powif: 741 case Builtin::BI__builtin_powil: { 742 Value *Base = EmitScalarExpr(E->getArg(0)); 743 Value *Exponent = EmitScalarExpr(E->getArg(1)); 744 llvm::Type *ArgType = Base->getType(); 745 Value *F = CGM.getIntrinsic(Intrinsic::powi, ArgType); 746 return RValue::get(Builder.CreateCall(F, {Base, Exponent})); 747 } 748 749 case Builtin::BI__builtin_isgreater: 750 case Builtin::BI__builtin_isgreaterequal: 751 case Builtin::BI__builtin_isless: 752 case Builtin::BI__builtin_islessequal: 753 case Builtin::BI__builtin_islessgreater: 754 case Builtin::BI__builtin_isunordered: { 755 // Ordered comparisons: we know the arguments to these are matching scalar 756 // floating point values. 757 Value *LHS = EmitScalarExpr(E->getArg(0)); 758 Value *RHS = EmitScalarExpr(E->getArg(1)); 759 760 switch (BuiltinID) { 761 default: llvm_unreachable("Unknown ordered comparison"); 762 case Builtin::BI__builtin_isgreater: 763 LHS = Builder.CreateFCmpOGT(LHS, RHS, "cmp"); 764 break; 765 case Builtin::BI__builtin_isgreaterequal: 766 LHS = Builder.CreateFCmpOGE(LHS, RHS, "cmp"); 767 break; 768 case Builtin::BI__builtin_isless: 769 LHS = Builder.CreateFCmpOLT(LHS, RHS, "cmp"); 770 break; 771 case Builtin::BI__builtin_islessequal: 772 LHS = Builder.CreateFCmpOLE(LHS, RHS, "cmp"); 773 break; 774 case Builtin::BI__builtin_islessgreater: 775 LHS = Builder.CreateFCmpONE(LHS, RHS, "cmp"); 776 break; 777 case Builtin::BI__builtin_isunordered: 778 LHS = Builder.CreateFCmpUNO(LHS, RHS, "cmp"); 779 break; 780 } 781 // ZExt bool to int type. 782 return RValue::get(Builder.CreateZExt(LHS, ConvertType(E->getType()))); 783 } 784 case Builtin::BI__builtin_isnan: { 785 Value *V = EmitScalarExpr(E->getArg(0)); 786 V = Builder.CreateFCmpUNO(V, V, "cmp"); 787 return RValue::get(Builder.CreateZExt(V, ConvertType(E->getType()))); 788 } 789 790 case Builtin::BI__builtin_isinf: { 791 // isinf(x) --> fabs(x) == infinity 792 Value *V = EmitScalarExpr(E->getArg(0)); 793 V = EmitFAbs(*this, V); 794 795 V = Builder.CreateFCmpOEQ(V, ConstantFP::getInfinity(V->getType()),"isinf"); 796 return RValue::get(Builder.CreateZExt(V, ConvertType(E->getType()))); 797 } 798 799 case Builtin::BI__builtin_isinf_sign: { 800 // isinf_sign(x) -> fabs(x) == infinity ? (signbit(x) ? -1 : 1) : 0 801 Value *Arg = EmitScalarExpr(E->getArg(0)); 802 Value *AbsArg = EmitFAbs(*this, Arg); 803 Value *IsInf = Builder.CreateFCmpOEQ( 804 AbsArg, ConstantFP::getInfinity(Arg->getType()), "isinf"); 805 Value *IsNeg = EmitSignBit(*this, Arg); 806 807 llvm::Type *IntTy = ConvertType(E->getType()); 808 Value *Zero = Constant::getNullValue(IntTy); 809 Value *One = ConstantInt::get(IntTy, 1); 810 Value *NegativeOne = ConstantInt::get(IntTy, -1); 811 Value *SignResult = Builder.CreateSelect(IsNeg, NegativeOne, One); 812 Value *Result = Builder.CreateSelect(IsInf, SignResult, Zero); 813 return RValue::get(Result); 814 } 815 816 case Builtin::BI__builtin_isnormal: { 817 // isnormal(x) --> x == x && fabsf(x) < infinity && fabsf(x) >= float_min 818 Value *V = EmitScalarExpr(E->getArg(0)); 819 Value *Eq = Builder.CreateFCmpOEQ(V, V, "iseq"); 820 821 Value *Abs = EmitFAbs(*this, V); 822 Value *IsLessThanInf = 823 Builder.CreateFCmpULT(Abs, ConstantFP::getInfinity(V->getType()),"isinf"); 824 APFloat Smallest = APFloat::getSmallestNormalized( 825 getContext().getFloatTypeSemantics(E->getArg(0)->getType())); 826 Value *IsNormal = 827 Builder.CreateFCmpUGE(Abs, ConstantFP::get(V->getContext(), Smallest), 828 "isnormal"); 829 V = Builder.CreateAnd(Eq, IsLessThanInf, "and"); 830 V = Builder.CreateAnd(V, IsNormal, "and"); 831 return RValue::get(Builder.CreateZExt(V, ConvertType(E->getType()))); 832 } 833 834 case Builtin::BI__builtin_isfinite: { 835 // isfinite(x) --> x == x && fabs(x) != infinity; 836 Value *V = EmitScalarExpr(E->getArg(0)); 837 Value *Eq = Builder.CreateFCmpOEQ(V, V, "iseq"); 838 839 Value *Abs = EmitFAbs(*this, V); 840 Value *IsNotInf = 841 Builder.CreateFCmpUNE(Abs, ConstantFP::getInfinity(V->getType()),"isinf"); 842 843 V = Builder.CreateAnd(Eq, IsNotInf, "and"); 844 return RValue::get(Builder.CreateZExt(V, ConvertType(E->getType()))); 845 } 846 847 case Builtin::BI__builtin_fpclassify: { 848 Value *V = EmitScalarExpr(E->getArg(5)); 849 llvm::Type *Ty = ConvertType(E->getArg(5)->getType()); 850 851 // Create Result 852 BasicBlock *Begin = Builder.GetInsertBlock(); 853 BasicBlock *End = createBasicBlock("fpclassify_end", this->CurFn); 854 Builder.SetInsertPoint(End); 855 PHINode *Result = 856 Builder.CreatePHI(ConvertType(E->getArg(0)->getType()), 4, 857 "fpclassify_result"); 858 859 // if (V==0) return FP_ZERO 860 Builder.SetInsertPoint(Begin); 861 Value *IsZero = Builder.CreateFCmpOEQ(V, Constant::getNullValue(Ty), 862 "iszero"); 863 Value *ZeroLiteral = EmitScalarExpr(E->getArg(4)); 864 BasicBlock *NotZero = createBasicBlock("fpclassify_not_zero", this->CurFn); 865 Builder.CreateCondBr(IsZero, End, NotZero); 866 Result->addIncoming(ZeroLiteral, Begin); 867 868 // if (V != V) return FP_NAN 869 Builder.SetInsertPoint(NotZero); 870 Value *IsNan = Builder.CreateFCmpUNO(V, V, "cmp"); 871 Value *NanLiteral = EmitScalarExpr(E->getArg(0)); 872 BasicBlock *NotNan = createBasicBlock("fpclassify_not_nan", this->CurFn); 873 Builder.CreateCondBr(IsNan, End, NotNan); 874 Result->addIncoming(NanLiteral, NotZero); 875 876 // if (fabs(V) == infinity) return FP_INFINITY 877 Builder.SetInsertPoint(NotNan); 878 Value *VAbs = EmitFAbs(*this, V); 879 Value *IsInf = 880 Builder.CreateFCmpOEQ(VAbs, ConstantFP::getInfinity(V->getType()), 881 "isinf"); 882 Value *InfLiteral = EmitScalarExpr(E->getArg(1)); 883 BasicBlock *NotInf = createBasicBlock("fpclassify_not_inf", this->CurFn); 884 Builder.CreateCondBr(IsInf, End, NotInf); 885 Result->addIncoming(InfLiteral, NotNan); 886 887 // if (fabs(V) >= MIN_NORMAL) return FP_NORMAL else FP_SUBNORMAL 888 Builder.SetInsertPoint(NotInf); 889 APFloat Smallest = APFloat::getSmallestNormalized( 890 getContext().getFloatTypeSemantics(E->getArg(5)->getType())); 891 Value *IsNormal = 892 Builder.CreateFCmpUGE(VAbs, ConstantFP::get(V->getContext(), Smallest), 893 "isnormal"); 894 Value *NormalResult = 895 Builder.CreateSelect(IsNormal, EmitScalarExpr(E->getArg(2)), 896 EmitScalarExpr(E->getArg(3))); 897 Builder.CreateBr(End); 898 Result->addIncoming(NormalResult, NotInf); 899 900 // return Result 901 Builder.SetInsertPoint(End); 902 return RValue::get(Result); 903 } 904 905 case Builtin::BIalloca: 906 case Builtin::BI_alloca: 907 case Builtin::BI__builtin_alloca: { 908 Value *Size = EmitScalarExpr(E->getArg(0)); 909 return RValue::get(Builder.CreateAlloca(Builder.getInt8Ty(), Size)); 910 } 911 case Builtin::BIbzero: 912 case Builtin::BI__builtin_bzero: { 913 Address Dest = EmitPointerWithAlignment(E->getArg(0)); 914 Value *SizeVal = EmitScalarExpr(E->getArg(1)); 915 EmitNonNullArgCheck(RValue::get(Dest.getPointer()), E->getArg(0)->getType(), 916 E->getArg(0)->getExprLoc(), FD, 0); 917 Builder.CreateMemSet(Dest, Builder.getInt8(0), SizeVal, false); 918 return RValue::get(Dest.getPointer()); 919 } 920 case Builtin::BImemcpy: 921 case Builtin::BI__builtin_memcpy: { 922 Address Dest = EmitPointerWithAlignment(E->getArg(0)); 923 Address Src = EmitPointerWithAlignment(E->getArg(1)); 924 Value *SizeVal = EmitScalarExpr(E->getArg(2)); 925 EmitNonNullArgCheck(RValue::get(Dest.getPointer()), E->getArg(0)->getType(), 926 E->getArg(0)->getExprLoc(), FD, 0); 927 EmitNonNullArgCheck(RValue::get(Src.getPointer()), E->getArg(1)->getType(), 928 E->getArg(1)->getExprLoc(), FD, 1); 929 Builder.CreateMemCpy(Dest, Src, SizeVal, false); 930 return RValue::get(Dest.getPointer()); 931 } 932 933 case Builtin::BI__builtin___memcpy_chk: { 934 // fold __builtin_memcpy_chk(x, y, cst1, cst2) to memcpy iff cst1<=cst2. 935 llvm::APSInt Size, DstSize; 936 if (!E->getArg(2)->EvaluateAsInt(Size, CGM.getContext()) || 937 !E->getArg(3)->EvaluateAsInt(DstSize, CGM.getContext())) 938 break; 939 if (Size.ugt(DstSize)) 940 break; 941 Address Dest = EmitPointerWithAlignment(E->getArg(0)); 942 Address Src = EmitPointerWithAlignment(E->getArg(1)); 943 Value *SizeVal = llvm::ConstantInt::get(Builder.getContext(), Size); 944 Builder.CreateMemCpy(Dest, Src, SizeVal, false); 945 return RValue::get(Dest.getPointer()); 946 } 947 948 case Builtin::BI__builtin_objc_memmove_collectable: { 949 Address DestAddr = EmitPointerWithAlignment(E->getArg(0)); 950 Address SrcAddr = EmitPointerWithAlignment(E->getArg(1)); 951 Value *SizeVal = EmitScalarExpr(E->getArg(2)); 952 CGM.getObjCRuntime().EmitGCMemmoveCollectable(*this, 953 DestAddr, SrcAddr, SizeVal); 954 return RValue::get(DestAddr.getPointer()); 955 } 956 957 case Builtin::BI__builtin___memmove_chk: { 958 // fold __builtin_memmove_chk(x, y, cst1, cst2) to memmove iff cst1<=cst2. 959 llvm::APSInt Size, DstSize; 960 if (!E->getArg(2)->EvaluateAsInt(Size, CGM.getContext()) || 961 !E->getArg(3)->EvaluateAsInt(DstSize, CGM.getContext())) 962 break; 963 if (Size.ugt(DstSize)) 964 break; 965 Address Dest = EmitPointerWithAlignment(E->getArg(0)); 966 Address Src = EmitPointerWithAlignment(E->getArg(1)); 967 Value *SizeVal = llvm::ConstantInt::get(Builder.getContext(), Size); 968 Builder.CreateMemMove(Dest, Src, SizeVal, false); 969 return RValue::get(Dest.getPointer()); 970 } 971 972 case Builtin::BImemmove: 973 case Builtin::BI__builtin_memmove: { 974 Address Dest = EmitPointerWithAlignment(E->getArg(0)); 975 Address Src = EmitPointerWithAlignment(E->getArg(1)); 976 Value *SizeVal = EmitScalarExpr(E->getArg(2)); 977 EmitNonNullArgCheck(RValue::get(Dest.getPointer()), E->getArg(0)->getType(), 978 E->getArg(0)->getExprLoc(), FD, 0); 979 EmitNonNullArgCheck(RValue::get(Src.getPointer()), E->getArg(1)->getType(), 980 E->getArg(1)->getExprLoc(), FD, 1); 981 Builder.CreateMemMove(Dest, Src, SizeVal, false); 982 return RValue::get(Dest.getPointer()); 983 } 984 case Builtin::BImemset: 985 case Builtin::BI__builtin_memset: { 986 Address Dest = EmitPointerWithAlignment(E->getArg(0)); 987 Value *ByteVal = Builder.CreateTrunc(EmitScalarExpr(E->getArg(1)), 988 Builder.getInt8Ty()); 989 Value *SizeVal = EmitScalarExpr(E->getArg(2)); 990 EmitNonNullArgCheck(RValue::get(Dest.getPointer()), E->getArg(0)->getType(), 991 E->getArg(0)->getExprLoc(), FD, 0); 992 Builder.CreateMemSet(Dest, ByteVal, SizeVal, false); 993 return RValue::get(Dest.getPointer()); 994 } 995 case Builtin::BI__builtin___memset_chk: { 996 // fold __builtin_memset_chk(x, y, cst1, cst2) to memset iff cst1<=cst2. 997 llvm::APSInt Size, DstSize; 998 if (!E->getArg(2)->EvaluateAsInt(Size, CGM.getContext()) || 999 !E->getArg(3)->EvaluateAsInt(DstSize, CGM.getContext())) 1000 break; 1001 if (Size.ugt(DstSize)) 1002 break; 1003 Address Dest = EmitPointerWithAlignment(E->getArg(0)); 1004 Value *ByteVal = Builder.CreateTrunc(EmitScalarExpr(E->getArg(1)), 1005 Builder.getInt8Ty()); 1006 Value *SizeVal = llvm::ConstantInt::get(Builder.getContext(), Size); 1007 Builder.CreateMemSet(Dest, ByteVal, SizeVal, false); 1008 return RValue::get(Dest.getPointer()); 1009 } 1010 case Builtin::BI__builtin_dwarf_cfa: { 1011 // The offset in bytes from the first argument to the CFA. 1012 // 1013 // Why on earth is this in the frontend? Is there any reason at 1014 // all that the backend can't reasonably determine this while 1015 // lowering llvm.eh.dwarf.cfa()? 1016 // 1017 // TODO: If there's a satisfactory reason, add a target hook for 1018 // this instead of hard-coding 0, which is correct for most targets. 1019 int32_t Offset = 0; 1020 1021 Value *F = CGM.getIntrinsic(Intrinsic::eh_dwarf_cfa); 1022 return RValue::get(Builder.CreateCall(F, 1023 llvm::ConstantInt::get(Int32Ty, Offset))); 1024 } 1025 case Builtin::BI__builtin_return_address: { 1026 Value *Depth = 1027 CGM.EmitConstantExpr(E->getArg(0), getContext().UnsignedIntTy, this); 1028 Value *F = CGM.getIntrinsic(Intrinsic::returnaddress); 1029 return RValue::get(Builder.CreateCall(F, Depth)); 1030 } 1031 case Builtin::BI__builtin_frame_address: { 1032 Value *Depth = 1033 CGM.EmitConstantExpr(E->getArg(0), getContext().UnsignedIntTy, this); 1034 Value *F = CGM.getIntrinsic(Intrinsic::frameaddress); 1035 return RValue::get(Builder.CreateCall(F, Depth)); 1036 } 1037 case Builtin::BI__builtin_extract_return_addr: { 1038 Value *Address = EmitScalarExpr(E->getArg(0)); 1039 Value *Result = getTargetHooks().decodeReturnAddress(*this, Address); 1040 return RValue::get(Result); 1041 } 1042 case Builtin::BI__builtin_frob_return_addr: { 1043 Value *Address = EmitScalarExpr(E->getArg(0)); 1044 Value *Result = getTargetHooks().encodeReturnAddress(*this, Address); 1045 return RValue::get(Result); 1046 } 1047 case Builtin::BI__builtin_dwarf_sp_column: { 1048 llvm::IntegerType *Ty 1049 = cast<llvm::IntegerType>(ConvertType(E->getType())); 1050 int Column = getTargetHooks().getDwarfEHStackPointer(CGM); 1051 if (Column == -1) { 1052 CGM.ErrorUnsupported(E, "__builtin_dwarf_sp_column"); 1053 return RValue::get(llvm::UndefValue::get(Ty)); 1054 } 1055 return RValue::get(llvm::ConstantInt::get(Ty, Column, true)); 1056 } 1057 case Builtin::BI__builtin_init_dwarf_reg_size_table: { 1058 Value *Address = EmitScalarExpr(E->getArg(0)); 1059 if (getTargetHooks().initDwarfEHRegSizeTable(*this, Address)) 1060 CGM.ErrorUnsupported(E, "__builtin_init_dwarf_reg_size_table"); 1061 return RValue::get(llvm::UndefValue::get(ConvertType(E->getType()))); 1062 } 1063 case Builtin::BI__builtin_eh_return: { 1064 Value *Int = EmitScalarExpr(E->getArg(0)); 1065 Value *Ptr = EmitScalarExpr(E->getArg(1)); 1066 1067 llvm::IntegerType *IntTy = cast<llvm::IntegerType>(Int->getType()); 1068 assert((IntTy->getBitWidth() == 32 || IntTy->getBitWidth() == 64) && 1069 "LLVM's __builtin_eh_return only supports 32- and 64-bit variants"); 1070 Value *F = CGM.getIntrinsic(IntTy->getBitWidth() == 32 1071 ? Intrinsic::eh_return_i32 1072 : Intrinsic::eh_return_i64); 1073 Builder.CreateCall(F, {Int, Ptr}); 1074 Builder.CreateUnreachable(); 1075 1076 // We do need to preserve an insertion point. 1077 EmitBlock(createBasicBlock("builtin_eh_return.cont")); 1078 1079 return RValue::get(nullptr); 1080 } 1081 case Builtin::BI__builtin_unwind_init: { 1082 Value *F = CGM.getIntrinsic(Intrinsic::eh_unwind_init); 1083 return RValue::get(Builder.CreateCall(F)); 1084 } 1085 case Builtin::BI__builtin_extend_pointer: { 1086 // Extends a pointer to the size of an _Unwind_Word, which is 1087 // uint64_t on all platforms. Generally this gets poked into a 1088 // register and eventually used as an address, so if the 1089 // addressing registers are wider than pointers and the platform 1090 // doesn't implicitly ignore high-order bits when doing 1091 // addressing, we need to make sure we zext / sext based on 1092 // the platform's expectations. 1093 // 1094 // See: http://gcc.gnu.org/ml/gcc-bugs/2002-02/msg00237.html 1095 1096 // Cast the pointer to intptr_t. 1097 Value *Ptr = EmitScalarExpr(E->getArg(0)); 1098 Value *Result = Builder.CreatePtrToInt(Ptr, IntPtrTy, "extend.cast"); 1099 1100 // If that's 64 bits, we're done. 1101 if (IntPtrTy->getBitWidth() == 64) 1102 return RValue::get(Result); 1103 1104 // Otherwise, ask the codegen data what to do. 1105 if (getTargetHooks().extendPointerWithSExt()) 1106 return RValue::get(Builder.CreateSExt(Result, Int64Ty, "extend.sext")); 1107 else 1108 return RValue::get(Builder.CreateZExt(Result, Int64Ty, "extend.zext")); 1109 } 1110 case Builtin::BI__builtin_setjmp: { 1111 // Buffer is a void**. 1112 Address Buf = EmitPointerWithAlignment(E->getArg(0)); 1113 1114 // Store the frame pointer to the setjmp buffer. 1115 Value *FrameAddr = 1116 Builder.CreateCall(CGM.getIntrinsic(Intrinsic::frameaddress), 1117 ConstantInt::get(Int32Ty, 0)); 1118 Builder.CreateStore(FrameAddr, Buf); 1119 1120 // Store the stack pointer to the setjmp buffer. 1121 Value *StackAddr = 1122 Builder.CreateCall(CGM.getIntrinsic(Intrinsic::stacksave)); 1123 Address StackSaveSlot = 1124 Builder.CreateConstInBoundsGEP(Buf, 2, getPointerSize()); 1125 Builder.CreateStore(StackAddr, StackSaveSlot); 1126 1127 // Call LLVM's EH setjmp, which is lightweight. 1128 Value *F = CGM.getIntrinsic(Intrinsic::eh_sjlj_setjmp); 1129 Buf = Builder.CreateBitCast(Buf, Int8PtrTy); 1130 return RValue::get(Builder.CreateCall(F, Buf.getPointer())); 1131 } 1132 case Builtin::BI__builtin_longjmp: { 1133 Value *Buf = EmitScalarExpr(E->getArg(0)); 1134 Buf = Builder.CreateBitCast(Buf, Int8PtrTy); 1135 1136 // Call LLVM's EH longjmp, which is lightweight. 1137 Builder.CreateCall(CGM.getIntrinsic(Intrinsic::eh_sjlj_longjmp), Buf); 1138 1139 // longjmp doesn't return; mark this as unreachable. 1140 Builder.CreateUnreachable(); 1141 1142 // We do need to preserve an insertion point. 1143 EmitBlock(createBasicBlock("longjmp.cont")); 1144 1145 return RValue::get(nullptr); 1146 } 1147 case Builtin::BI__sync_fetch_and_add: 1148 case Builtin::BI__sync_fetch_and_sub: 1149 case Builtin::BI__sync_fetch_and_or: 1150 case Builtin::BI__sync_fetch_and_and: 1151 case Builtin::BI__sync_fetch_and_xor: 1152 case Builtin::BI__sync_fetch_and_nand: 1153 case Builtin::BI__sync_add_and_fetch: 1154 case Builtin::BI__sync_sub_and_fetch: 1155 case Builtin::BI__sync_and_and_fetch: 1156 case Builtin::BI__sync_or_and_fetch: 1157 case Builtin::BI__sync_xor_and_fetch: 1158 case Builtin::BI__sync_nand_and_fetch: 1159 case Builtin::BI__sync_val_compare_and_swap: 1160 case Builtin::BI__sync_bool_compare_and_swap: 1161 case Builtin::BI__sync_lock_test_and_set: 1162 case Builtin::BI__sync_lock_release: 1163 case Builtin::BI__sync_swap: 1164 llvm_unreachable("Shouldn't make it through sema"); 1165 case Builtin::BI__sync_fetch_and_add_1: 1166 case Builtin::BI__sync_fetch_and_add_2: 1167 case Builtin::BI__sync_fetch_and_add_4: 1168 case Builtin::BI__sync_fetch_and_add_8: 1169 case Builtin::BI__sync_fetch_and_add_16: 1170 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Add, E); 1171 case Builtin::BI__sync_fetch_and_sub_1: 1172 case Builtin::BI__sync_fetch_and_sub_2: 1173 case Builtin::BI__sync_fetch_and_sub_4: 1174 case Builtin::BI__sync_fetch_and_sub_8: 1175 case Builtin::BI__sync_fetch_and_sub_16: 1176 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Sub, E); 1177 case Builtin::BI__sync_fetch_and_or_1: 1178 case Builtin::BI__sync_fetch_and_or_2: 1179 case Builtin::BI__sync_fetch_and_or_4: 1180 case Builtin::BI__sync_fetch_and_or_8: 1181 case Builtin::BI__sync_fetch_and_or_16: 1182 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Or, E); 1183 case Builtin::BI__sync_fetch_and_and_1: 1184 case Builtin::BI__sync_fetch_and_and_2: 1185 case Builtin::BI__sync_fetch_and_and_4: 1186 case Builtin::BI__sync_fetch_and_and_8: 1187 case Builtin::BI__sync_fetch_and_and_16: 1188 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::And, E); 1189 case Builtin::BI__sync_fetch_and_xor_1: 1190 case Builtin::BI__sync_fetch_and_xor_2: 1191 case Builtin::BI__sync_fetch_and_xor_4: 1192 case Builtin::BI__sync_fetch_and_xor_8: 1193 case Builtin::BI__sync_fetch_and_xor_16: 1194 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Xor, E); 1195 case Builtin::BI__sync_fetch_and_nand_1: 1196 case Builtin::BI__sync_fetch_and_nand_2: 1197 case Builtin::BI__sync_fetch_and_nand_4: 1198 case Builtin::BI__sync_fetch_and_nand_8: 1199 case Builtin::BI__sync_fetch_and_nand_16: 1200 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Nand, E); 1201 1202 // Clang extensions: not overloaded yet. 1203 case Builtin::BI__sync_fetch_and_min: 1204 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Min, E); 1205 case Builtin::BI__sync_fetch_and_max: 1206 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Max, E); 1207 case Builtin::BI__sync_fetch_and_umin: 1208 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::UMin, E); 1209 case Builtin::BI__sync_fetch_and_umax: 1210 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::UMax, E); 1211 1212 case Builtin::BI__sync_add_and_fetch_1: 1213 case Builtin::BI__sync_add_and_fetch_2: 1214 case Builtin::BI__sync_add_and_fetch_4: 1215 case Builtin::BI__sync_add_and_fetch_8: 1216 case Builtin::BI__sync_add_and_fetch_16: 1217 return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::Add, E, 1218 llvm::Instruction::Add); 1219 case Builtin::BI__sync_sub_and_fetch_1: 1220 case Builtin::BI__sync_sub_and_fetch_2: 1221 case Builtin::BI__sync_sub_and_fetch_4: 1222 case Builtin::BI__sync_sub_and_fetch_8: 1223 case Builtin::BI__sync_sub_and_fetch_16: 1224 return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::Sub, E, 1225 llvm::Instruction::Sub); 1226 case Builtin::BI__sync_and_and_fetch_1: 1227 case Builtin::BI__sync_and_and_fetch_2: 1228 case Builtin::BI__sync_and_and_fetch_4: 1229 case Builtin::BI__sync_and_and_fetch_8: 1230 case Builtin::BI__sync_and_and_fetch_16: 1231 return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::And, E, 1232 llvm::Instruction::And); 1233 case Builtin::BI__sync_or_and_fetch_1: 1234 case Builtin::BI__sync_or_and_fetch_2: 1235 case Builtin::BI__sync_or_and_fetch_4: 1236 case Builtin::BI__sync_or_and_fetch_8: 1237 case Builtin::BI__sync_or_and_fetch_16: 1238 return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::Or, E, 1239 llvm::Instruction::Or); 1240 case Builtin::BI__sync_xor_and_fetch_1: 1241 case Builtin::BI__sync_xor_and_fetch_2: 1242 case Builtin::BI__sync_xor_and_fetch_4: 1243 case Builtin::BI__sync_xor_and_fetch_8: 1244 case Builtin::BI__sync_xor_and_fetch_16: 1245 return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::Xor, E, 1246 llvm::Instruction::Xor); 1247 case Builtin::BI__sync_nand_and_fetch_1: 1248 case Builtin::BI__sync_nand_and_fetch_2: 1249 case Builtin::BI__sync_nand_and_fetch_4: 1250 case Builtin::BI__sync_nand_and_fetch_8: 1251 case Builtin::BI__sync_nand_and_fetch_16: 1252 return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::Nand, E, 1253 llvm::Instruction::And, true); 1254 1255 case Builtin::BI__sync_val_compare_and_swap_1: 1256 case Builtin::BI__sync_val_compare_and_swap_2: 1257 case Builtin::BI__sync_val_compare_and_swap_4: 1258 case Builtin::BI__sync_val_compare_and_swap_8: 1259 case Builtin::BI__sync_val_compare_and_swap_16: 1260 return RValue::get(MakeAtomicCmpXchgValue(*this, E, false)); 1261 1262 case Builtin::BI__sync_bool_compare_and_swap_1: 1263 case Builtin::BI__sync_bool_compare_and_swap_2: 1264 case Builtin::BI__sync_bool_compare_and_swap_4: 1265 case Builtin::BI__sync_bool_compare_and_swap_8: 1266 case Builtin::BI__sync_bool_compare_and_swap_16: 1267 return RValue::get(MakeAtomicCmpXchgValue(*this, E, true)); 1268 1269 case Builtin::BI__sync_swap_1: 1270 case Builtin::BI__sync_swap_2: 1271 case Builtin::BI__sync_swap_4: 1272 case Builtin::BI__sync_swap_8: 1273 case Builtin::BI__sync_swap_16: 1274 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Xchg, E); 1275 1276 case Builtin::BI__sync_lock_test_and_set_1: 1277 case Builtin::BI__sync_lock_test_and_set_2: 1278 case Builtin::BI__sync_lock_test_and_set_4: 1279 case Builtin::BI__sync_lock_test_and_set_8: 1280 case Builtin::BI__sync_lock_test_and_set_16: 1281 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Xchg, E); 1282 1283 case Builtin::BI__sync_lock_release_1: 1284 case Builtin::BI__sync_lock_release_2: 1285 case Builtin::BI__sync_lock_release_4: 1286 case Builtin::BI__sync_lock_release_8: 1287 case Builtin::BI__sync_lock_release_16: { 1288 Value *Ptr = EmitScalarExpr(E->getArg(0)); 1289 QualType ElTy = E->getArg(0)->getType()->getPointeeType(); 1290 CharUnits StoreSize = getContext().getTypeSizeInChars(ElTy); 1291 llvm::Type *ITy = llvm::IntegerType::get(getLLVMContext(), 1292 StoreSize.getQuantity() * 8); 1293 Ptr = Builder.CreateBitCast(Ptr, ITy->getPointerTo()); 1294 llvm::StoreInst *Store = 1295 Builder.CreateAlignedStore(llvm::Constant::getNullValue(ITy), Ptr, 1296 StoreSize); 1297 Store->setAtomic(llvm::Release); 1298 return RValue::get(nullptr); 1299 } 1300 1301 case Builtin::BI__sync_synchronize: { 1302 // We assume this is supposed to correspond to a C++0x-style 1303 // sequentially-consistent fence (i.e. this is only usable for 1304 // synchonization, not device I/O or anything like that). This intrinsic 1305 // is really badly designed in the sense that in theory, there isn't 1306 // any way to safely use it... but in practice, it mostly works 1307 // to use it with non-atomic loads and stores to get acquire/release 1308 // semantics. 1309 Builder.CreateFence(llvm::SequentiallyConsistent); 1310 return RValue::get(nullptr); 1311 } 1312 1313 case Builtin::BI__builtin_nontemporal_load: 1314 return RValue::get(EmitNontemporalLoad(*this, E)); 1315 case Builtin::BI__builtin_nontemporal_store: 1316 return RValue::get(EmitNontemporalStore(*this, E)); 1317 case Builtin::BI__c11_atomic_is_lock_free: 1318 case Builtin::BI__atomic_is_lock_free: { 1319 // Call "bool __atomic_is_lock_free(size_t size, void *ptr)". For the 1320 // __c11 builtin, ptr is 0 (indicating a properly-aligned object), since 1321 // _Atomic(T) is always properly-aligned. 1322 const char *LibCallName = "__atomic_is_lock_free"; 1323 CallArgList Args; 1324 Args.add(RValue::get(EmitScalarExpr(E->getArg(0))), 1325 getContext().getSizeType()); 1326 if (BuiltinID == Builtin::BI__atomic_is_lock_free) 1327 Args.add(RValue::get(EmitScalarExpr(E->getArg(1))), 1328 getContext().VoidPtrTy); 1329 else 1330 Args.add(RValue::get(llvm::Constant::getNullValue(VoidPtrTy)), 1331 getContext().VoidPtrTy); 1332 const CGFunctionInfo &FuncInfo = 1333 CGM.getTypes().arrangeFreeFunctionCall(E->getType(), Args, 1334 FunctionType::ExtInfo(), 1335 RequiredArgs::All); 1336 llvm::FunctionType *FTy = CGM.getTypes().GetFunctionType(FuncInfo); 1337 llvm::Constant *Func = CGM.CreateRuntimeFunction(FTy, LibCallName); 1338 return EmitCall(FuncInfo, Func, ReturnValueSlot(), Args); 1339 } 1340 1341 case Builtin::BI__atomic_test_and_set: { 1342 // Look at the argument type to determine whether this is a volatile 1343 // operation. The parameter type is always volatile. 1344 QualType PtrTy = E->getArg(0)->IgnoreImpCasts()->getType(); 1345 bool Volatile = 1346 PtrTy->castAs<PointerType>()->getPointeeType().isVolatileQualified(); 1347 1348 Value *Ptr = EmitScalarExpr(E->getArg(0)); 1349 unsigned AddrSpace = Ptr->getType()->getPointerAddressSpace(); 1350 Ptr = Builder.CreateBitCast(Ptr, Int8Ty->getPointerTo(AddrSpace)); 1351 Value *NewVal = Builder.getInt8(1); 1352 Value *Order = EmitScalarExpr(E->getArg(1)); 1353 if (isa<llvm::ConstantInt>(Order)) { 1354 int ord = cast<llvm::ConstantInt>(Order)->getZExtValue(); 1355 AtomicRMWInst *Result = nullptr; 1356 switch (ord) { 1357 case 0: // memory_order_relaxed 1358 default: // invalid order 1359 Result = Builder.CreateAtomicRMW(llvm::AtomicRMWInst::Xchg, 1360 Ptr, NewVal, 1361 llvm::Monotonic); 1362 break; 1363 case 1: // memory_order_consume 1364 case 2: // memory_order_acquire 1365 Result = Builder.CreateAtomicRMW(llvm::AtomicRMWInst::Xchg, 1366 Ptr, NewVal, 1367 llvm::Acquire); 1368 break; 1369 case 3: // memory_order_release 1370 Result = Builder.CreateAtomicRMW(llvm::AtomicRMWInst::Xchg, 1371 Ptr, NewVal, 1372 llvm::Release); 1373 break; 1374 case 4: // memory_order_acq_rel 1375 Result = Builder.CreateAtomicRMW(llvm::AtomicRMWInst::Xchg, 1376 Ptr, NewVal, 1377 llvm::AcquireRelease); 1378 break; 1379 case 5: // memory_order_seq_cst 1380 Result = Builder.CreateAtomicRMW(llvm::AtomicRMWInst::Xchg, 1381 Ptr, NewVal, 1382 llvm::SequentiallyConsistent); 1383 break; 1384 } 1385 Result->setVolatile(Volatile); 1386 return RValue::get(Builder.CreateIsNotNull(Result, "tobool")); 1387 } 1388 1389 llvm::BasicBlock *ContBB = createBasicBlock("atomic.continue", CurFn); 1390 1391 llvm::BasicBlock *BBs[5] = { 1392 createBasicBlock("monotonic", CurFn), 1393 createBasicBlock("acquire", CurFn), 1394 createBasicBlock("release", CurFn), 1395 createBasicBlock("acqrel", CurFn), 1396 createBasicBlock("seqcst", CurFn) 1397 }; 1398 llvm::AtomicOrdering Orders[5] = { 1399 llvm::Monotonic, llvm::Acquire, llvm::Release, 1400 llvm::AcquireRelease, llvm::SequentiallyConsistent 1401 }; 1402 1403 Order = Builder.CreateIntCast(Order, Builder.getInt32Ty(), false); 1404 llvm::SwitchInst *SI = Builder.CreateSwitch(Order, BBs[0]); 1405 1406 Builder.SetInsertPoint(ContBB); 1407 PHINode *Result = Builder.CreatePHI(Int8Ty, 5, "was_set"); 1408 1409 for (unsigned i = 0; i < 5; ++i) { 1410 Builder.SetInsertPoint(BBs[i]); 1411 AtomicRMWInst *RMW = Builder.CreateAtomicRMW(llvm::AtomicRMWInst::Xchg, 1412 Ptr, NewVal, Orders[i]); 1413 RMW->setVolatile(Volatile); 1414 Result->addIncoming(RMW, BBs[i]); 1415 Builder.CreateBr(ContBB); 1416 } 1417 1418 SI->addCase(Builder.getInt32(0), BBs[0]); 1419 SI->addCase(Builder.getInt32(1), BBs[1]); 1420 SI->addCase(Builder.getInt32(2), BBs[1]); 1421 SI->addCase(Builder.getInt32(3), BBs[2]); 1422 SI->addCase(Builder.getInt32(4), BBs[3]); 1423 SI->addCase(Builder.getInt32(5), BBs[4]); 1424 1425 Builder.SetInsertPoint(ContBB); 1426 return RValue::get(Builder.CreateIsNotNull(Result, "tobool")); 1427 } 1428 1429 case Builtin::BI__atomic_clear: { 1430 QualType PtrTy = E->getArg(0)->IgnoreImpCasts()->getType(); 1431 bool Volatile = 1432 PtrTy->castAs<PointerType>()->getPointeeType().isVolatileQualified(); 1433 1434 Address Ptr = EmitPointerWithAlignment(E->getArg(0)); 1435 unsigned AddrSpace = Ptr.getPointer()->getType()->getPointerAddressSpace(); 1436 Ptr = Builder.CreateBitCast(Ptr, Int8Ty->getPointerTo(AddrSpace)); 1437 Value *NewVal = Builder.getInt8(0); 1438 Value *Order = EmitScalarExpr(E->getArg(1)); 1439 if (isa<llvm::ConstantInt>(Order)) { 1440 int ord = cast<llvm::ConstantInt>(Order)->getZExtValue(); 1441 StoreInst *Store = Builder.CreateStore(NewVal, Ptr, Volatile); 1442 switch (ord) { 1443 case 0: // memory_order_relaxed 1444 default: // invalid order 1445 Store->setOrdering(llvm::Monotonic); 1446 break; 1447 case 3: // memory_order_release 1448 Store->setOrdering(llvm::Release); 1449 break; 1450 case 5: // memory_order_seq_cst 1451 Store->setOrdering(llvm::SequentiallyConsistent); 1452 break; 1453 } 1454 return RValue::get(nullptr); 1455 } 1456 1457 llvm::BasicBlock *ContBB = createBasicBlock("atomic.continue", CurFn); 1458 1459 llvm::BasicBlock *BBs[3] = { 1460 createBasicBlock("monotonic", CurFn), 1461 createBasicBlock("release", CurFn), 1462 createBasicBlock("seqcst", CurFn) 1463 }; 1464 llvm::AtomicOrdering Orders[3] = { 1465 llvm::Monotonic, llvm::Release, llvm::SequentiallyConsistent 1466 }; 1467 1468 Order = Builder.CreateIntCast(Order, Builder.getInt32Ty(), false); 1469 llvm::SwitchInst *SI = Builder.CreateSwitch(Order, BBs[0]); 1470 1471 for (unsigned i = 0; i < 3; ++i) { 1472 Builder.SetInsertPoint(BBs[i]); 1473 StoreInst *Store = Builder.CreateStore(NewVal, Ptr, Volatile); 1474 Store->setOrdering(Orders[i]); 1475 Builder.CreateBr(ContBB); 1476 } 1477 1478 SI->addCase(Builder.getInt32(0), BBs[0]); 1479 SI->addCase(Builder.getInt32(3), BBs[1]); 1480 SI->addCase(Builder.getInt32(5), BBs[2]); 1481 1482 Builder.SetInsertPoint(ContBB); 1483 return RValue::get(nullptr); 1484 } 1485 1486 case Builtin::BI__atomic_thread_fence: 1487 case Builtin::BI__atomic_signal_fence: 1488 case Builtin::BI__c11_atomic_thread_fence: 1489 case Builtin::BI__c11_atomic_signal_fence: { 1490 llvm::SynchronizationScope Scope; 1491 if (BuiltinID == Builtin::BI__atomic_signal_fence || 1492 BuiltinID == Builtin::BI__c11_atomic_signal_fence) 1493 Scope = llvm::SingleThread; 1494 else 1495 Scope = llvm::CrossThread; 1496 Value *Order = EmitScalarExpr(E->getArg(0)); 1497 if (isa<llvm::ConstantInt>(Order)) { 1498 int ord = cast<llvm::ConstantInt>(Order)->getZExtValue(); 1499 switch (ord) { 1500 case 0: // memory_order_relaxed 1501 default: // invalid order 1502 break; 1503 case 1: // memory_order_consume 1504 case 2: // memory_order_acquire 1505 Builder.CreateFence(llvm::Acquire, Scope); 1506 break; 1507 case 3: // memory_order_release 1508 Builder.CreateFence(llvm::Release, Scope); 1509 break; 1510 case 4: // memory_order_acq_rel 1511 Builder.CreateFence(llvm::AcquireRelease, Scope); 1512 break; 1513 case 5: // memory_order_seq_cst 1514 Builder.CreateFence(llvm::SequentiallyConsistent, Scope); 1515 break; 1516 } 1517 return RValue::get(nullptr); 1518 } 1519 1520 llvm::BasicBlock *AcquireBB, *ReleaseBB, *AcqRelBB, *SeqCstBB; 1521 AcquireBB = createBasicBlock("acquire", CurFn); 1522 ReleaseBB = createBasicBlock("release", CurFn); 1523 AcqRelBB = createBasicBlock("acqrel", CurFn); 1524 SeqCstBB = createBasicBlock("seqcst", CurFn); 1525 llvm::BasicBlock *ContBB = createBasicBlock("atomic.continue", CurFn); 1526 1527 Order = Builder.CreateIntCast(Order, Builder.getInt32Ty(), false); 1528 llvm::SwitchInst *SI = Builder.CreateSwitch(Order, ContBB); 1529 1530 Builder.SetInsertPoint(AcquireBB); 1531 Builder.CreateFence(llvm::Acquire, Scope); 1532 Builder.CreateBr(ContBB); 1533 SI->addCase(Builder.getInt32(1), AcquireBB); 1534 SI->addCase(Builder.getInt32(2), AcquireBB); 1535 1536 Builder.SetInsertPoint(ReleaseBB); 1537 Builder.CreateFence(llvm::Release, Scope); 1538 Builder.CreateBr(ContBB); 1539 SI->addCase(Builder.getInt32(3), ReleaseBB); 1540 1541 Builder.SetInsertPoint(AcqRelBB); 1542 Builder.CreateFence(llvm::AcquireRelease, Scope); 1543 Builder.CreateBr(ContBB); 1544 SI->addCase(Builder.getInt32(4), AcqRelBB); 1545 1546 Builder.SetInsertPoint(SeqCstBB); 1547 Builder.CreateFence(llvm::SequentiallyConsistent, Scope); 1548 Builder.CreateBr(ContBB); 1549 SI->addCase(Builder.getInt32(5), SeqCstBB); 1550 1551 Builder.SetInsertPoint(ContBB); 1552 return RValue::get(nullptr); 1553 } 1554 1555 // Library functions with special handling. 1556 case Builtin::BIsqrt: 1557 case Builtin::BIsqrtf: 1558 case Builtin::BIsqrtl: { 1559 // Transform a call to sqrt* into a @llvm.sqrt.* intrinsic call, but only 1560 // in finite- or unsafe-math mode (the intrinsic has different semantics 1561 // for handling negative numbers compared to the library function, so 1562 // -fmath-errno=0 is not enough). 1563 if (!FD->hasAttr<ConstAttr>()) 1564 break; 1565 if (!(CGM.getCodeGenOpts().UnsafeFPMath || 1566 CGM.getCodeGenOpts().NoNaNsFPMath)) 1567 break; 1568 Value *Arg0 = EmitScalarExpr(E->getArg(0)); 1569 llvm::Type *ArgType = Arg0->getType(); 1570 Value *F = CGM.getIntrinsic(Intrinsic::sqrt, ArgType); 1571 return RValue::get(Builder.CreateCall(F, Arg0)); 1572 } 1573 1574 case Builtin::BI__builtin_pow: 1575 case Builtin::BI__builtin_powf: 1576 case Builtin::BI__builtin_powl: 1577 case Builtin::BIpow: 1578 case Builtin::BIpowf: 1579 case Builtin::BIpowl: { 1580 // Transform a call to pow* into a @llvm.pow.* intrinsic call. 1581 if (!FD->hasAttr<ConstAttr>()) 1582 break; 1583 Value *Base = EmitScalarExpr(E->getArg(0)); 1584 Value *Exponent = EmitScalarExpr(E->getArg(1)); 1585 llvm::Type *ArgType = Base->getType(); 1586 Value *F = CGM.getIntrinsic(Intrinsic::pow, ArgType); 1587 return RValue::get(Builder.CreateCall(F, {Base, Exponent})); 1588 } 1589 1590 case Builtin::BIfma: 1591 case Builtin::BIfmaf: 1592 case Builtin::BIfmal: 1593 case Builtin::BI__builtin_fma: 1594 case Builtin::BI__builtin_fmaf: 1595 case Builtin::BI__builtin_fmal: { 1596 // Rewrite fma to intrinsic. 1597 Value *FirstArg = EmitScalarExpr(E->getArg(0)); 1598 llvm::Type *ArgType = FirstArg->getType(); 1599 Value *F = CGM.getIntrinsic(Intrinsic::fma, ArgType); 1600 return RValue::get( 1601 Builder.CreateCall(F, {FirstArg, EmitScalarExpr(E->getArg(1)), 1602 EmitScalarExpr(E->getArg(2))})); 1603 } 1604 1605 case Builtin::BI__builtin_signbit: 1606 case Builtin::BI__builtin_signbitf: 1607 case Builtin::BI__builtin_signbitl: { 1608 return RValue::get( 1609 Builder.CreateZExt(EmitSignBit(*this, EmitScalarExpr(E->getArg(0))), 1610 ConvertType(E->getType()))); 1611 } 1612 case Builtin::BI__builtin_annotation: { 1613 llvm::Value *AnnVal = EmitScalarExpr(E->getArg(0)); 1614 llvm::Value *F = CGM.getIntrinsic(llvm::Intrinsic::annotation, 1615 AnnVal->getType()); 1616 1617 // Get the annotation string, go through casts. Sema requires this to be a 1618 // non-wide string literal, potentially casted, so the cast<> is safe. 1619 const Expr *AnnotationStrExpr = E->getArg(1)->IgnoreParenCasts(); 1620 StringRef Str = cast<StringLiteral>(AnnotationStrExpr)->getString(); 1621 return RValue::get(EmitAnnotationCall(F, AnnVal, Str, E->getExprLoc())); 1622 } 1623 case Builtin::BI__builtin_addcb: 1624 case Builtin::BI__builtin_addcs: 1625 case Builtin::BI__builtin_addc: 1626 case Builtin::BI__builtin_addcl: 1627 case Builtin::BI__builtin_addcll: 1628 case Builtin::BI__builtin_subcb: 1629 case Builtin::BI__builtin_subcs: 1630 case Builtin::BI__builtin_subc: 1631 case Builtin::BI__builtin_subcl: 1632 case Builtin::BI__builtin_subcll: { 1633 1634 // We translate all of these builtins from expressions of the form: 1635 // int x = ..., y = ..., carryin = ..., carryout, result; 1636 // result = __builtin_addc(x, y, carryin, &carryout); 1637 // 1638 // to LLVM IR of the form: 1639 // 1640 // %tmp1 = call {i32, i1} @llvm.uadd.with.overflow.i32(i32 %x, i32 %y) 1641 // %tmpsum1 = extractvalue {i32, i1} %tmp1, 0 1642 // %carry1 = extractvalue {i32, i1} %tmp1, 1 1643 // %tmp2 = call {i32, i1} @llvm.uadd.with.overflow.i32(i32 %tmpsum1, 1644 // i32 %carryin) 1645 // %result = extractvalue {i32, i1} %tmp2, 0 1646 // %carry2 = extractvalue {i32, i1} %tmp2, 1 1647 // %tmp3 = or i1 %carry1, %carry2 1648 // %tmp4 = zext i1 %tmp3 to i32 1649 // store i32 %tmp4, i32* %carryout 1650 1651 // Scalarize our inputs. 1652 llvm::Value *X = EmitScalarExpr(E->getArg(0)); 1653 llvm::Value *Y = EmitScalarExpr(E->getArg(1)); 1654 llvm::Value *Carryin = EmitScalarExpr(E->getArg(2)); 1655 Address CarryOutPtr = EmitPointerWithAlignment(E->getArg(3)); 1656 1657 // Decide if we are lowering to a uadd.with.overflow or usub.with.overflow. 1658 llvm::Intrinsic::ID IntrinsicId; 1659 switch (BuiltinID) { 1660 default: llvm_unreachable("Unknown multiprecision builtin id."); 1661 case Builtin::BI__builtin_addcb: 1662 case Builtin::BI__builtin_addcs: 1663 case Builtin::BI__builtin_addc: 1664 case Builtin::BI__builtin_addcl: 1665 case Builtin::BI__builtin_addcll: 1666 IntrinsicId = llvm::Intrinsic::uadd_with_overflow; 1667 break; 1668 case Builtin::BI__builtin_subcb: 1669 case Builtin::BI__builtin_subcs: 1670 case Builtin::BI__builtin_subc: 1671 case Builtin::BI__builtin_subcl: 1672 case Builtin::BI__builtin_subcll: 1673 IntrinsicId = llvm::Intrinsic::usub_with_overflow; 1674 break; 1675 } 1676 1677 // Construct our resulting LLVM IR expression. 1678 llvm::Value *Carry1; 1679 llvm::Value *Sum1 = EmitOverflowIntrinsic(*this, IntrinsicId, 1680 X, Y, Carry1); 1681 llvm::Value *Carry2; 1682 llvm::Value *Sum2 = EmitOverflowIntrinsic(*this, IntrinsicId, 1683 Sum1, Carryin, Carry2); 1684 llvm::Value *CarryOut = Builder.CreateZExt(Builder.CreateOr(Carry1, Carry2), 1685 X->getType()); 1686 Builder.CreateStore(CarryOut, CarryOutPtr); 1687 return RValue::get(Sum2); 1688 } 1689 1690 case Builtin::BI__builtin_add_overflow: 1691 case Builtin::BI__builtin_sub_overflow: 1692 case Builtin::BI__builtin_mul_overflow: { 1693 const clang::Expr *LeftArg = E->getArg(0); 1694 const clang::Expr *RightArg = E->getArg(1); 1695 const clang::Expr *ResultArg = E->getArg(2); 1696 1697 clang::QualType ResultQTy = 1698 ResultArg->getType()->castAs<PointerType>()->getPointeeType(); 1699 1700 WidthAndSignedness LeftInfo = 1701 getIntegerWidthAndSignedness(CGM.getContext(), LeftArg->getType()); 1702 WidthAndSignedness RightInfo = 1703 getIntegerWidthAndSignedness(CGM.getContext(), RightArg->getType()); 1704 WidthAndSignedness ResultInfo = 1705 getIntegerWidthAndSignedness(CGM.getContext(), ResultQTy); 1706 WidthAndSignedness EncompassingInfo = 1707 EncompassingIntegerType({LeftInfo, RightInfo, ResultInfo}); 1708 1709 llvm::Type *EncompassingLLVMTy = 1710 llvm::IntegerType::get(CGM.getLLVMContext(), EncompassingInfo.Width); 1711 1712 llvm::Type *ResultLLVMTy = CGM.getTypes().ConvertType(ResultQTy); 1713 1714 llvm::Intrinsic::ID IntrinsicId; 1715 switch (BuiltinID) { 1716 default: 1717 llvm_unreachable("Unknown overflow builtin id."); 1718 case Builtin::BI__builtin_add_overflow: 1719 IntrinsicId = EncompassingInfo.Signed 1720 ? llvm::Intrinsic::sadd_with_overflow 1721 : llvm::Intrinsic::uadd_with_overflow; 1722 break; 1723 case Builtin::BI__builtin_sub_overflow: 1724 IntrinsicId = EncompassingInfo.Signed 1725 ? llvm::Intrinsic::ssub_with_overflow 1726 : llvm::Intrinsic::usub_with_overflow; 1727 break; 1728 case Builtin::BI__builtin_mul_overflow: 1729 IntrinsicId = EncompassingInfo.Signed 1730 ? llvm::Intrinsic::smul_with_overflow 1731 : llvm::Intrinsic::umul_with_overflow; 1732 break; 1733 } 1734 1735 llvm::Value *Left = EmitScalarExpr(LeftArg); 1736 llvm::Value *Right = EmitScalarExpr(RightArg); 1737 Address ResultPtr = EmitPointerWithAlignment(ResultArg); 1738 1739 // Extend each operand to the encompassing type. 1740 Left = Builder.CreateIntCast(Left, EncompassingLLVMTy, LeftInfo.Signed); 1741 Right = Builder.CreateIntCast(Right, EncompassingLLVMTy, RightInfo.Signed); 1742 1743 // Perform the operation on the extended values. 1744 llvm::Value *Overflow, *Result; 1745 Result = EmitOverflowIntrinsic(*this, IntrinsicId, Left, Right, Overflow); 1746 1747 if (EncompassingInfo.Width > ResultInfo.Width) { 1748 // The encompassing type is wider than the result type, so we need to 1749 // truncate it. 1750 llvm::Value *ResultTrunc = Builder.CreateTrunc(Result, ResultLLVMTy); 1751 1752 // To see if the truncation caused an overflow, we will extend 1753 // the result and then compare it to the original result. 1754 llvm::Value *ResultTruncExt = Builder.CreateIntCast( 1755 ResultTrunc, EncompassingLLVMTy, ResultInfo.Signed); 1756 llvm::Value *TruncationOverflow = 1757 Builder.CreateICmpNE(Result, ResultTruncExt); 1758 1759 Overflow = Builder.CreateOr(Overflow, TruncationOverflow); 1760 Result = ResultTrunc; 1761 } 1762 1763 // Finally, store the result using the pointer. 1764 bool isVolatile = 1765 ResultArg->getType()->getPointeeType().isVolatileQualified(); 1766 Builder.CreateStore(EmitToMemory(Result, ResultQTy), ResultPtr, isVolatile); 1767 1768 return RValue::get(Overflow); 1769 } 1770 1771 case Builtin::BI__builtin_uadd_overflow: 1772 case Builtin::BI__builtin_uaddl_overflow: 1773 case Builtin::BI__builtin_uaddll_overflow: 1774 case Builtin::BI__builtin_usub_overflow: 1775 case Builtin::BI__builtin_usubl_overflow: 1776 case Builtin::BI__builtin_usubll_overflow: 1777 case Builtin::BI__builtin_umul_overflow: 1778 case Builtin::BI__builtin_umull_overflow: 1779 case Builtin::BI__builtin_umulll_overflow: 1780 case Builtin::BI__builtin_sadd_overflow: 1781 case Builtin::BI__builtin_saddl_overflow: 1782 case Builtin::BI__builtin_saddll_overflow: 1783 case Builtin::BI__builtin_ssub_overflow: 1784 case Builtin::BI__builtin_ssubl_overflow: 1785 case Builtin::BI__builtin_ssubll_overflow: 1786 case Builtin::BI__builtin_smul_overflow: 1787 case Builtin::BI__builtin_smull_overflow: 1788 case Builtin::BI__builtin_smulll_overflow: { 1789 1790 // We translate all of these builtins directly to the relevant llvm IR node. 1791 1792 // Scalarize our inputs. 1793 llvm::Value *X = EmitScalarExpr(E->getArg(0)); 1794 llvm::Value *Y = EmitScalarExpr(E->getArg(1)); 1795 Address SumOutPtr = EmitPointerWithAlignment(E->getArg(2)); 1796 1797 // Decide which of the overflow intrinsics we are lowering to: 1798 llvm::Intrinsic::ID IntrinsicId; 1799 switch (BuiltinID) { 1800 default: llvm_unreachable("Unknown overflow builtin id."); 1801 case Builtin::BI__builtin_uadd_overflow: 1802 case Builtin::BI__builtin_uaddl_overflow: 1803 case Builtin::BI__builtin_uaddll_overflow: 1804 IntrinsicId = llvm::Intrinsic::uadd_with_overflow; 1805 break; 1806 case Builtin::BI__builtin_usub_overflow: 1807 case Builtin::BI__builtin_usubl_overflow: 1808 case Builtin::BI__builtin_usubll_overflow: 1809 IntrinsicId = llvm::Intrinsic::usub_with_overflow; 1810 break; 1811 case Builtin::BI__builtin_umul_overflow: 1812 case Builtin::BI__builtin_umull_overflow: 1813 case Builtin::BI__builtin_umulll_overflow: 1814 IntrinsicId = llvm::Intrinsic::umul_with_overflow; 1815 break; 1816 case Builtin::BI__builtin_sadd_overflow: 1817 case Builtin::BI__builtin_saddl_overflow: 1818 case Builtin::BI__builtin_saddll_overflow: 1819 IntrinsicId = llvm::Intrinsic::sadd_with_overflow; 1820 break; 1821 case Builtin::BI__builtin_ssub_overflow: 1822 case Builtin::BI__builtin_ssubl_overflow: 1823 case Builtin::BI__builtin_ssubll_overflow: 1824 IntrinsicId = llvm::Intrinsic::ssub_with_overflow; 1825 break; 1826 case Builtin::BI__builtin_smul_overflow: 1827 case Builtin::BI__builtin_smull_overflow: 1828 case Builtin::BI__builtin_smulll_overflow: 1829 IntrinsicId = llvm::Intrinsic::smul_with_overflow; 1830 break; 1831 } 1832 1833 1834 llvm::Value *Carry; 1835 llvm::Value *Sum = EmitOverflowIntrinsic(*this, IntrinsicId, X, Y, Carry); 1836 Builder.CreateStore(Sum, SumOutPtr); 1837 1838 return RValue::get(Carry); 1839 } 1840 case Builtin::BI__builtin_addressof: 1841 return RValue::get(EmitLValue(E->getArg(0)).getPointer()); 1842 case Builtin::BI__builtin_operator_new: 1843 return EmitBuiltinNewDeleteCall(FD->getType()->castAs<FunctionProtoType>(), 1844 E->getArg(0), false); 1845 case Builtin::BI__builtin_operator_delete: 1846 return EmitBuiltinNewDeleteCall(FD->getType()->castAs<FunctionProtoType>(), 1847 E->getArg(0), true); 1848 case Builtin::BI__noop: 1849 // __noop always evaluates to an integer literal zero. 1850 return RValue::get(ConstantInt::get(IntTy, 0)); 1851 case Builtin::BI__builtin_call_with_static_chain: { 1852 const CallExpr *Call = cast<CallExpr>(E->getArg(0)); 1853 const Expr *Chain = E->getArg(1); 1854 return EmitCall(Call->getCallee()->getType(), 1855 EmitScalarExpr(Call->getCallee()), Call, ReturnValue, 1856 Call->getCalleeDecl(), EmitScalarExpr(Chain)); 1857 } 1858 case Builtin::BI_InterlockedExchange: 1859 case Builtin::BI_InterlockedExchangePointer: 1860 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Xchg, E); 1861 case Builtin::BI_InterlockedCompareExchangePointer: { 1862 llvm::Type *RTy; 1863 llvm::IntegerType *IntType = 1864 IntegerType::get(getLLVMContext(), 1865 getContext().getTypeSize(E->getType())); 1866 llvm::Type *IntPtrType = IntType->getPointerTo(); 1867 1868 llvm::Value *Destination = 1869 Builder.CreateBitCast(EmitScalarExpr(E->getArg(0)), IntPtrType); 1870 1871 llvm::Value *Exchange = EmitScalarExpr(E->getArg(1)); 1872 RTy = Exchange->getType(); 1873 Exchange = Builder.CreatePtrToInt(Exchange, IntType); 1874 1875 llvm::Value *Comparand = 1876 Builder.CreatePtrToInt(EmitScalarExpr(E->getArg(2)), IntType); 1877 1878 auto Result = Builder.CreateAtomicCmpXchg(Destination, Comparand, Exchange, 1879 SequentiallyConsistent, 1880 SequentiallyConsistent); 1881 Result->setVolatile(true); 1882 1883 return RValue::get(Builder.CreateIntToPtr(Builder.CreateExtractValue(Result, 1884 0), 1885 RTy)); 1886 } 1887 case Builtin::BI_InterlockedCompareExchange: { 1888 AtomicCmpXchgInst *CXI = Builder.CreateAtomicCmpXchg( 1889 EmitScalarExpr(E->getArg(0)), 1890 EmitScalarExpr(E->getArg(2)), 1891 EmitScalarExpr(E->getArg(1)), 1892 SequentiallyConsistent, 1893 SequentiallyConsistent); 1894 CXI->setVolatile(true); 1895 return RValue::get(Builder.CreateExtractValue(CXI, 0)); 1896 } 1897 case Builtin::BI_InterlockedIncrement: { 1898 AtomicRMWInst *RMWI = Builder.CreateAtomicRMW( 1899 AtomicRMWInst::Add, 1900 EmitScalarExpr(E->getArg(0)), 1901 ConstantInt::get(Int32Ty, 1), 1902 llvm::SequentiallyConsistent); 1903 RMWI->setVolatile(true); 1904 return RValue::get(Builder.CreateAdd(RMWI, ConstantInt::get(Int32Ty, 1))); 1905 } 1906 case Builtin::BI_InterlockedDecrement: { 1907 AtomicRMWInst *RMWI = Builder.CreateAtomicRMW( 1908 AtomicRMWInst::Sub, 1909 EmitScalarExpr(E->getArg(0)), 1910 ConstantInt::get(Int32Ty, 1), 1911 llvm::SequentiallyConsistent); 1912 RMWI->setVolatile(true); 1913 return RValue::get(Builder.CreateSub(RMWI, ConstantInt::get(Int32Ty, 1))); 1914 } 1915 case Builtin::BI_InterlockedExchangeAdd: { 1916 AtomicRMWInst *RMWI = Builder.CreateAtomicRMW( 1917 AtomicRMWInst::Add, 1918 EmitScalarExpr(E->getArg(0)), 1919 EmitScalarExpr(E->getArg(1)), 1920 llvm::SequentiallyConsistent); 1921 RMWI->setVolatile(true); 1922 return RValue::get(RMWI); 1923 } 1924 case Builtin::BI__readfsdword: { 1925 Value *IntToPtr = 1926 Builder.CreateIntToPtr(EmitScalarExpr(E->getArg(0)), 1927 llvm::PointerType::get(CGM.Int32Ty, 257)); 1928 LoadInst *Load = 1929 Builder.CreateAlignedLoad(IntToPtr, /*Align=*/4, /*isVolatile=*/true); 1930 return RValue::get(Load); 1931 } 1932 1933 case Builtin::BI__exception_code: 1934 case Builtin::BI_exception_code: 1935 return RValue::get(EmitSEHExceptionCode()); 1936 case Builtin::BI__exception_info: 1937 case Builtin::BI_exception_info: 1938 return RValue::get(EmitSEHExceptionInfo()); 1939 case Builtin::BI__abnormal_termination: 1940 case Builtin::BI_abnormal_termination: 1941 return RValue::get(EmitSEHAbnormalTermination()); 1942 case Builtin::BI_setjmpex: { 1943 if (getTarget().getTriple().isOSMSVCRT()) { 1944 llvm::Type *ArgTypes[] = {Int8PtrTy, Int8PtrTy}; 1945 llvm::AttributeSet ReturnsTwiceAttr = 1946 AttributeSet::get(getLLVMContext(), llvm::AttributeSet::FunctionIndex, 1947 llvm::Attribute::ReturnsTwice); 1948 llvm::Constant *SetJmpEx = CGM.CreateRuntimeFunction( 1949 llvm::FunctionType::get(IntTy, ArgTypes, /*isVarArg=*/false), 1950 "_setjmpex", ReturnsTwiceAttr); 1951 llvm::Value *Buf = Builder.CreateBitOrPointerCast( 1952 EmitScalarExpr(E->getArg(0)), Int8PtrTy); 1953 llvm::Value *FrameAddr = 1954 Builder.CreateCall(CGM.getIntrinsic(Intrinsic::frameaddress), 1955 ConstantInt::get(Int32Ty, 0)); 1956 llvm::Value *Args[] = {Buf, FrameAddr}; 1957 llvm::CallSite CS = EmitRuntimeCallOrInvoke(SetJmpEx, Args); 1958 CS.setAttributes(ReturnsTwiceAttr); 1959 return RValue::get(CS.getInstruction()); 1960 } 1961 break; 1962 } 1963 case Builtin::BI_setjmp: { 1964 if (getTarget().getTriple().isOSMSVCRT()) { 1965 llvm::AttributeSet ReturnsTwiceAttr = 1966 AttributeSet::get(getLLVMContext(), llvm::AttributeSet::FunctionIndex, 1967 llvm::Attribute::ReturnsTwice); 1968 llvm::Value *Buf = Builder.CreateBitOrPointerCast( 1969 EmitScalarExpr(E->getArg(0)), Int8PtrTy); 1970 llvm::CallSite CS; 1971 if (getTarget().getTriple().getArch() == llvm::Triple::x86) { 1972 llvm::Type *ArgTypes[] = {Int8PtrTy, IntTy}; 1973 llvm::Constant *SetJmp3 = CGM.CreateRuntimeFunction( 1974 llvm::FunctionType::get(IntTy, ArgTypes, /*isVarArg=*/true), 1975 "_setjmp3", ReturnsTwiceAttr); 1976 llvm::Value *Count = ConstantInt::get(IntTy, 0); 1977 llvm::Value *Args[] = {Buf, Count}; 1978 CS = EmitRuntimeCallOrInvoke(SetJmp3, Args); 1979 } else { 1980 llvm::Type *ArgTypes[] = {Int8PtrTy, Int8PtrTy}; 1981 llvm::Constant *SetJmp = CGM.CreateRuntimeFunction( 1982 llvm::FunctionType::get(IntTy, ArgTypes, /*isVarArg=*/false), 1983 "_setjmp", ReturnsTwiceAttr); 1984 llvm::Value *FrameAddr = 1985 Builder.CreateCall(CGM.getIntrinsic(Intrinsic::frameaddress), 1986 ConstantInt::get(Int32Ty, 0)); 1987 llvm::Value *Args[] = {Buf, FrameAddr}; 1988 CS = EmitRuntimeCallOrInvoke(SetJmp, Args); 1989 } 1990 CS.setAttributes(ReturnsTwiceAttr); 1991 return RValue::get(CS.getInstruction()); 1992 } 1993 break; 1994 } 1995 1996 case Builtin::BI__GetExceptionInfo: { 1997 if (llvm::GlobalVariable *GV = 1998 CGM.getCXXABI().getThrowInfo(FD->getParamDecl(0)->getType())) 1999 return RValue::get(llvm::ConstantExpr::getBitCast(GV, CGM.Int8PtrTy)); 2000 break; 2001 } 2002 2003 // OpenCL v2.0 s6.13.16.2, Built-in pipe read and write functions 2004 case Builtin::BIread_pipe: 2005 case Builtin::BIwrite_pipe: { 2006 Value *Arg0 = EmitScalarExpr(E->getArg(0)), 2007 *Arg1 = EmitScalarExpr(E->getArg(1)); 2008 2009 // Type of the generic packet parameter. 2010 unsigned GenericAS = 2011 getContext().getTargetAddressSpace(LangAS::opencl_generic); 2012 llvm::Type *I8PTy = llvm::PointerType::get( 2013 llvm::Type::getInt8Ty(getLLVMContext()), GenericAS); 2014 2015 // Testing which overloaded version we should generate the call for. 2016 if (2U == E->getNumArgs()) { 2017 const char *Name = (BuiltinID == Builtin::BIread_pipe) ? "__read_pipe_2" 2018 : "__write_pipe_2"; 2019 // Creating a generic function type to be able to call with any builtin or 2020 // user defined type. 2021 llvm::Type *ArgTys[] = {Arg0->getType(), I8PTy}; 2022 llvm::FunctionType *FTy = llvm::FunctionType::get( 2023 Int32Ty, llvm::ArrayRef<llvm::Type *>(ArgTys), false); 2024 Value *BCast = Builder.CreatePointerCast(Arg1, I8PTy); 2025 return RValue::get(Builder.CreateCall( 2026 CGM.CreateRuntimeFunction(FTy, Name), {Arg0, BCast})); 2027 } else { 2028 assert(4 == E->getNumArgs() && 2029 "Illegal number of parameters to pipe function"); 2030 const char *Name = (BuiltinID == Builtin::BIread_pipe) ? "__read_pipe_4" 2031 : "__write_pipe_4"; 2032 2033 llvm::Type *ArgTys[] = {Arg0->getType(), Arg1->getType(), Int32Ty, I8PTy}; 2034 Value *Arg2 = EmitScalarExpr(E->getArg(2)), 2035 *Arg3 = EmitScalarExpr(E->getArg(3)); 2036 llvm::FunctionType *FTy = llvm::FunctionType::get( 2037 Int32Ty, llvm::ArrayRef<llvm::Type *>(ArgTys), false); 2038 Value *BCast = Builder.CreatePointerCast(Arg3, I8PTy); 2039 // We know the third argument is an integer type, but we may need to cast 2040 // it to i32. 2041 if (Arg2->getType() != Int32Ty) 2042 Arg2 = Builder.CreateZExtOrTrunc(Arg2, Int32Ty); 2043 return RValue::get(Builder.CreateCall( 2044 CGM.CreateRuntimeFunction(FTy, Name), {Arg0, Arg1, Arg2, BCast})); 2045 } 2046 } 2047 // OpenCL v2.0 s6.13.16 ,s9.17.3.5 - Built-in pipe reserve read and write 2048 // functions 2049 case Builtin::BIreserve_read_pipe: 2050 case Builtin::BIreserve_write_pipe: 2051 case Builtin::BIwork_group_reserve_read_pipe: 2052 case Builtin::BIwork_group_reserve_write_pipe: 2053 case Builtin::BIsub_group_reserve_read_pipe: 2054 case Builtin::BIsub_group_reserve_write_pipe: { 2055 // Composing the mangled name for the function. 2056 const char *Name; 2057 if (BuiltinID == Builtin::BIreserve_read_pipe) 2058 Name = "__reserve_read_pipe"; 2059 else if (BuiltinID == Builtin::BIreserve_write_pipe) 2060 Name = "__reserve_write_pipe"; 2061 else if (BuiltinID == Builtin::BIwork_group_reserve_read_pipe) 2062 Name = "__work_group_reserve_read_pipe"; 2063 else if (BuiltinID == Builtin::BIwork_group_reserve_write_pipe) 2064 Name = "__work_group_reserve_write_pipe"; 2065 else if (BuiltinID == Builtin::BIsub_group_reserve_read_pipe) 2066 Name = "__sub_group_reserve_read_pipe"; 2067 else 2068 Name = "__sub_group_reserve_write_pipe"; 2069 2070 Value *Arg0 = EmitScalarExpr(E->getArg(0)), 2071 *Arg1 = EmitScalarExpr(E->getArg(1)); 2072 llvm::Type *ReservedIDTy = ConvertType(getContext().OCLReserveIDTy); 2073 2074 // Building the generic function prototype. 2075 llvm::Type *ArgTys[] = {Arg0->getType(), Int32Ty}; 2076 llvm::FunctionType *FTy = llvm::FunctionType::get( 2077 ReservedIDTy, llvm::ArrayRef<llvm::Type *>(ArgTys), false); 2078 // We know the second argument is an integer type, but we may need to cast 2079 // it to i32. 2080 if (Arg1->getType() != Int32Ty) 2081 Arg1 = Builder.CreateZExtOrTrunc(Arg1, Int32Ty); 2082 return RValue::get( 2083 Builder.CreateCall(CGM.CreateRuntimeFunction(FTy, Name), {Arg0, Arg1})); 2084 } 2085 // OpenCL v2.0 s6.13.16 ,s9.17.3.5 - Built-in pipe commit read and write 2086 // functions 2087 case Builtin::BIcommit_read_pipe: 2088 case Builtin::BIcommit_write_pipe: 2089 case Builtin::BIwork_group_commit_read_pipe: 2090 case Builtin::BIwork_group_commit_write_pipe: 2091 case Builtin::BIsub_group_commit_read_pipe: 2092 case Builtin::BIsub_group_commit_write_pipe: { 2093 const char *Name; 2094 if (BuiltinID == Builtin::BIcommit_read_pipe) 2095 Name = "__commit_read_pipe"; 2096 else if (BuiltinID == Builtin::BIcommit_write_pipe) 2097 Name = "__commit_write_pipe"; 2098 else if (BuiltinID == Builtin::BIwork_group_commit_read_pipe) 2099 Name = "__work_group_commit_read_pipe"; 2100 else if (BuiltinID == Builtin::BIwork_group_commit_write_pipe) 2101 Name = "__work_group_commit_write_pipe"; 2102 else if (BuiltinID == Builtin::BIsub_group_commit_read_pipe) 2103 Name = "__sub_group_commit_read_pipe"; 2104 else 2105 Name = "__sub_group_commit_write_pipe"; 2106 2107 Value *Arg0 = EmitScalarExpr(E->getArg(0)), 2108 *Arg1 = EmitScalarExpr(E->getArg(1)); 2109 2110 // Building the generic function prototype. 2111 llvm::Type *ArgTys[] = {Arg0->getType(), Arg1->getType()}; 2112 llvm::FunctionType *FTy = 2113 llvm::FunctionType::get(llvm::Type::getVoidTy(getLLVMContext()), 2114 llvm::ArrayRef<llvm::Type *>(ArgTys), false); 2115 2116 return RValue::get( 2117 Builder.CreateCall(CGM.CreateRuntimeFunction(FTy, Name), {Arg0, Arg1})); 2118 } 2119 // OpenCL v2.0 s6.13.16.4 Built-in pipe query functions 2120 case Builtin::BIget_pipe_num_packets: 2121 case Builtin::BIget_pipe_max_packets: { 2122 const char *Name; 2123 if (BuiltinID == Builtin::BIget_pipe_num_packets) 2124 Name = "__get_pipe_num_packets"; 2125 else 2126 Name = "__get_pipe_max_packets"; 2127 2128 // Building the generic function prototype. 2129 Value *Arg0 = EmitScalarExpr(E->getArg(0)); 2130 llvm::Type *ArgTys[] = {Arg0->getType()}; 2131 llvm::FunctionType *FTy = llvm::FunctionType::get( 2132 Int32Ty, llvm::ArrayRef<llvm::Type *>(ArgTys), false); 2133 2134 return RValue::get( 2135 Builder.CreateCall(CGM.CreateRuntimeFunction(FTy, Name), {Arg0})); 2136 } 2137 2138 case Builtin::BIprintf: 2139 if (getLangOpts().CUDA && getLangOpts().CUDAIsDevice) 2140 return EmitCUDADevicePrintfCallExpr(E, ReturnValue); 2141 break; 2142 case Builtin::BI__builtin_canonicalize: 2143 case Builtin::BI__builtin_canonicalizef: 2144 case Builtin::BI__builtin_canonicalizel: 2145 return RValue::get(emitUnaryBuiltin(*this, E, Intrinsic::canonicalize)); 2146 } 2147 2148 // If this is an alias for a lib function (e.g. __builtin_sin), emit 2149 // the call using the normal call path, but using the unmangled 2150 // version of the function name. 2151 if (getContext().BuiltinInfo.isLibFunction(BuiltinID)) 2152 return emitLibraryCall(*this, FD, E, 2153 CGM.getBuiltinLibFunction(FD, BuiltinID)); 2154 2155 // If this is a predefined lib function (e.g. malloc), emit the call 2156 // using exactly the normal call path. 2157 if (getContext().BuiltinInfo.isPredefinedLibFunction(BuiltinID)) 2158 return emitLibraryCall(*this, FD, E, EmitScalarExpr(E->getCallee())); 2159 2160 // Check that a call to a target specific builtin has the correct target 2161 // features. 2162 // This is down here to avoid non-target specific builtins, however, if 2163 // generic builtins start to require generic target features then we 2164 // can move this up to the beginning of the function. 2165 checkTargetFeatures(E, FD); 2166 2167 // See if we have a target specific intrinsic. 2168 const char *Name = getContext().BuiltinInfo.getName(BuiltinID); 2169 Intrinsic::ID IntrinsicID = Intrinsic::not_intrinsic; 2170 if (const char *Prefix = 2171 llvm::Triple::getArchTypePrefix(getTarget().getTriple().getArch())) { 2172 IntrinsicID = Intrinsic::getIntrinsicForGCCBuiltin(Prefix, Name); 2173 // NOTE we dont need to perform a compatibility flag check here since the 2174 // intrinsics are declared in Builtins*.def via LANGBUILTIN which filter the 2175 // MS builtins via ALL_MS_LANGUAGES and are filtered earlier. 2176 if (IntrinsicID == Intrinsic::not_intrinsic) 2177 IntrinsicID = Intrinsic::getIntrinsicForMSBuiltin(Prefix, Name); 2178 } 2179 2180 if (IntrinsicID != Intrinsic::not_intrinsic) { 2181 SmallVector<Value*, 16> Args; 2182 2183 // Find out if any arguments are required to be integer constant 2184 // expressions. 2185 unsigned ICEArguments = 0; 2186 ASTContext::GetBuiltinTypeError Error; 2187 getContext().GetBuiltinType(BuiltinID, Error, &ICEArguments); 2188 assert(Error == ASTContext::GE_None && "Should not codegen an error"); 2189 2190 Function *F = CGM.getIntrinsic(IntrinsicID); 2191 llvm::FunctionType *FTy = F->getFunctionType(); 2192 2193 for (unsigned i = 0, e = E->getNumArgs(); i != e; ++i) { 2194 Value *ArgValue; 2195 // If this is a normal argument, just emit it as a scalar. 2196 if ((ICEArguments & (1 << i)) == 0) { 2197 ArgValue = EmitScalarExpr(E->getArg(i)); 2198 } else { 2199 // If this is required to be a constant, constant fold it so that we 2200 // know that the generated intrinsic gets a ConstantInt. 2201 llvm::APSInt Result; 2202 bool IsConst = E->getArg(i)->isIntegerConstantExpr(Result,getContext()); 2203 assert(IsConst && "Constant arg isn't actually constant?"); 2204 (void)IsConst; 2205 ArgValue = llvm::ConstantInt::get(getLLVMContext(), Result); 2206 } 2207 2208 // If the intrinsic arg type is different from the builtin arg type 2209 // we need to do a bit cast. 2210 llvm::Type *PTy = FTy->getParamType(i); 2211 if (PTy != ArgValue->getType()) { 2212 assert(PTy->canLosslesslyBitCastTo(FTy->getParamType(i)) && 2213 "Must be able to losslessly bit cast to param"); 2214 ArgValue = Builder.CreateBitCast(ArgValue, PTy); 2215 } 2216 2217 Args.push_back(ArgValue); 2218 } 2219 2220 Value *V = Builder.CreateCall(F, Args); 2221 QualType BuiltinRetType = E->getType(); 2222 2223 llvm::Type *RetTy = VoidTy; 2224 if (!BuiltinRetType->isVoidType()) 2225 RetTy = ConvertType(BuiltinRetType); 2226 2227 if (RetTy != V->getType()) { 2228 assert(V->getType()->canLosslesslyBitCastTo(RetTy) && 2229 "Must be able to losslessly bit cast result type"); 2230 V = Builder.CreateBitCast(V, RetTy); 2231 } 2232 2233 return RValue::get(V); 2234 } 2235 2236 // See if we have a target specific builtin that needs to be lowered. 2237 if (Value *V = EmitTargetBuiltinExpr(BuiltinID, E)) 2238 return RValue::get(V); 2239 2240 ErrorUnsupported(E, "builtin function"); 2241 2242 // Unknown builtin, for now just dump it out and return undef. 2243 return GetUndefRValue(E->getType()); 2244 } 2245 2246 static Value *EmitTargetArchBuiltinExpr(CodeGenFunction *CGF, 2247 unsigned BuiltinID, const CallExpr *E, 2248 llvm::Triple::ArchType Arch) { 2249 switch (Arch) { 2250 case llvm::Triple::arm: 2251 case llvm::Triple::armeb: 2252 case llvm::Triple::thumb: 2253 case llvm::Triple::thumbeb: 2254 return CGF->EmitARMBuiltinExpr(BuiltinID, E); 2255 case llvm::Triple::aarch64: 2256 case llvm::Triple::aarch64_be: 2257 return CGF->EmitAArch64BuiltinExpr(BuiltinID, E); 2258 case llvm::Triple::x86: 2259 case llvm::Triple::x86_64: 2260 return CGF->EmitX86BuiltinExpr(BuiltinID, E); 2261 case llvm::Triple::ppc: 2262 case llvm::Triple::ppc64: 2263 case llvm::Triple::ppc64le: 2264 return CGF->EmitPPCBuiltinExpr(BuiltinID, E); 2265 case llvm::Triple::r600: 2266 case llvm::Triple::amdgcn: 2267 return CGF->EmitAMDGPUBuiltinExpr(BuiltinID, E); 2268 case llvm::Triple::systemz: 2269 return CGF->EmitSystemZBuiltinExpr(BuiltinID, E); 2270 case llvm::Triple::nvptx: 2271 case llvm::Triple::nvptx64: 2272 return CGF->EmitNVPTXBuiltinExpr(BuiltinID, E); 2273 case llvm::Triple::wasm32: 2274 case llvm::Triple::wasm64: 2275 return CGF->EmitWebAssemblyBuiltinExpr(BuiltinID, E); 2276 default: 2277 return nullptr; 2278 } 2279 } 2280 2281 Value *CodeGenFunction::EmitTargetBuiltinExpr(unsigned BuiltinID, 2282 const CallExpr *E) { 2283 if (getContext().BuiltinInfo.isAuxBuiltinID(BuiltinID)) { 2284 assert(getContext().getAuxTargetInfo() && "Missing aux target info"); 2285 return EmitTargetArchBuiltinExpr( 2286 this, getContext().BuiltinInfo.getAuxBuiltinID(BuiltinID), E, 2287 getContext().getAuxTargetInfo()->getTriple().getArch()); 2288 } 2289 2290 return EmitTargetArchBuiltinExpr(this, BuiltinID, E, 2291 getTarget().getTriple().getArch()); 2292 } 2293 2294 static llvm::VectorType *GetNeonType(CodeGenFunction *CGF, 2295 NeonTypeFlags TypeFlags, 2296 bool V1Ty=false) { 2297 int IsQuad = TypeFlags.isQuad(); 2298 switch (TypeFlags.getEltType()) { 2299 case NeonTypeFlags::Int8: 2300 case NeonTypeFlags::Poly8: 2301 return llvm::VectorType::get(CGF->Int8Ty, V1Ty ? 1 : (8 << IsQuad)); 2302 case NeonTypeFlags::Int16: 2303 case NeonTypeFlags::Poly16: 2304 case NeonTypeFlags::Float16: 2305 return llvm::VectorType::get(CGF->Int16Ty, V1Ty ? 1 : (4 << IsQuad)); 2306 case NeonTypeFlags::Int32: 2307 return llvm::VectorType::get(CGF->Int32Ty, V1Ty ? 1 : (2 << IsQuad)); 2308 case NeonTypeFlags::Int64: 2309 case NeonTypeFlags::Poly64: 2310 return llvm::VectorType::get(CGF->Int64Ty, V1Ty ? 1 : (1 << IsQuad)); 2311 case NeonTypeFlags::Poly128: 2312 // FIXME: i128 and f128 doesn't get fully support in Clang and llvm. 2313 // There is a lot of i128 and f128 API missing. 2314 // so we use v16i8 to represent poly128 and get pattern matched. 2315 return llvm::VectorType::get(CGF->Int8Ty, 16); 2316 case NeonTypeFlags::Float32: 2317 return llvm::VectorType::get(CGF->FloatTy, V1Ty ? 1 : (2 << IsQuad)); 2318 case NeonTypeFlags::Float64: 2319 return llvm::VectorType::get(CGF->DoubleTy, V1Ty ? 1 : (1 << IsQuad)); 2320 } 2321 llvm_unreachable("Unknown vector element type!"); 2322 } 2323 2324 static llvm::VectorType *GetFloatNeonType(CodeGenFunction *CGF, 2325 NeonTypeFlags IntTypeFlags) { 2326 int IsQuad = IntTypeFlags.isQuad(); 2327 switch (IntTypeFlags.getEltType()) { 2328 case NeonTypeFlags::Int32: 2329 return llvm::VectorType::get(CGF->FloatTy, (2 << IsQuad)); 2330 case NeonTypeFlags::Int64: 2331 return llvm::VectorType::get(CGF->DoubleTy, (1 << IsQuad)); 2332 default: 2333 llvm_unreachable("Type can't be converted to floating-point!"); 2334 } 2335 } 2336 2337 Value *CodeGenFunction::EmitNeonSplat(Value *V, Constant *C) { 2338 unsigned nElts = cast<llvm::VectorType>(V->getType())->getNumElements(); 2339 Value* SV = llvm::ConstantVector::getSplat(nElts, C); 2340 return Builder.CreateShuffleVector(V, V, SV, "lane"); 2341 } 2342 2343 Value *CodeGenFunction::EmitNeonCall(Function *F, SmallVectorImpl<Value*> &Ops, 2344 const char *name, 2345 unsigned shift, bool rightshift) { 2346 unsigned j = 0; 2347 for (Function::const_arg_iterator ai = F->arg_begin(), ae = F->arg_end(); 2348 ai != ae; ++ai, ++j) 2349 if (shift > 0 && shift == j) 2350 Ops[j] = EmitNeonShiftVector(Ops[j], ai->getType(), rightshift); 2351 else 2352 Ops[j] = Builder.CreateBitCast(Ops[j], ai->getType(), name); 2353 2354 return Builder.CreateCall(F, Ops, name); 2355 } 2356 2357 Value *CodeGenFunction::EmitNeonShiftVector(Value *V, llvm::Type *Ty, 2358 bool neg) { 2359 int SV = cast<ConstantInt>(V)->getSExtValue(); 2360 return ConstantInt::get(Ty, neg ? -SV : SV); 2361 } 2362 2363 // \brief Right-shift a vector by a constant. 2364 Value *CodeGenFunction::EmitNeonRShiftImm(Value *Vec, Value *Shift, 2365 llvm::Type *Ty, bool usgn, 2366 const char *name) { 2367 llvm::VectorType *VTy = cast<llvm::VectorType>(Ty); 2368 2369 int ShiftAmt = cast<ConstantInt>(Shift)->getSExtValue(); 2370 int EltSize = VTy->getScalarSizeInBits(); 2371 2372 Vec = Builder.CreateBitCast(Vec, Ty); 2373 2374 // lshr/ashr are undefined when the shift amount is equal to the vector 2375 // element size. 2376 if (ShiftAmt == EltSize) { 2377 if (usgn) { 2378 // Right-shifting an unsigned value by its size yields 0. 2379 return llvm::ConstantAggregateZero::get(VTy); 2380 } else { 2381 // Right-shifting a signed value by its size is equivalent 2382 // to a shift of size-1. 2383 --ShiftAmt; 2384 Shift = ConstantInt::get(VTy->getElementType(), ShiftAmt); 2385 } 2386 } 2387 2388 Shift = EmitNeonShiftVector(Shift, Ty, false); 2389 if (usgn) 2390 return Builder.CreateLShr(Vec, Shift, name); 2391 else 2392 return Builder.CreateAShr(Vec, Shift, name); 2393 } 2394 2395 enum { 2396 AddRetType = (1 << 0), 2397 Add1ArgType = (1 << 1), 2398 Add2ArgTypes = (1 << 2), 2399 2400 VectorizeRetType = (1 << 3), 2401 VectorizeArgTypes = (1 << 4), 2402 2403 InventFloatType = (1 << 5), 2404 UnsignedAlts = (1 << 6), 2405 2406 Use64BitVectors = (1 << 7), 2407 Use128BitVectors = (1 << 8), 2408 2409 Vectorize1ArgType = Add1ArgType | VectorizeArgTypes, 2410 VectorRet = AddRetType | VectorizeRetType, 2411 VectorRetGetArgs01 = 2412 AddRetType | Add2ArgTypes | VectorizeRetType | VectorizeArgTypes, 2413 FpCmpzModifiers = 2414 AddRetType | VectorizeRetType | Add1ArgType | InventFloatType 2415 }; 2416 2417 namespace { 2418 struct NeonIntrinsicInfo { 2419 const char *NameHint; 2420 unsigned BuiltinID; 2421 unsigned LLVMIntrinsic; 2422 unsigned AltLLVMIntrinsic; 2423 unsigned TypeModifier; 2424 2425 bool operator<(unsigned RHSBuiltinID) const { 2426 return BuiltinID < RHSBuiltinID; 2427 } 2428 bool operator<(const NeonIntrinsicInfo &TE) const { 2429 return BuiltinID < TE.BuiltinID; 2430 } 2431 }; 2432 } // end anonymous namespace 2433 2434 #define NEONMAP0(NameBase) \ 2435 { #NameBase, NEON::BI__builtin_neon_ ## NameBase, 0, 0, 0 } 2436 2437 #define NEONMAP1(NameBase, LLVMIntrinsic, TypeModifier) \ 2438 { #NameBase, NEON:: BI__builtin_neon_ ## NameBase, \ 2439 Intrinsic::LLVMIntrinsic, 0, TypeModifier } 2440 2441 #define NEONMAP2(NameBase, LLVMIntrinsic, AltLLVMIntrinsic, TypeModifier) \ 2442 { #NameBase, NEON:: BI__builtin_neon_ ## NameBase, \ 2443 Intrinsic::LLVMIntrinsic, Intrinsic::AltLLVMIntrinsic, \ 2444 TypeModifier } 2445 2446 static const NeonIntrinsicInfo ARMSIMDIntrinsicMap [] = { 2447 NEONMAP2(vabd_v, arm_neon_vabdu, arm_neon_vabds, Add1ArgType | UnsignedAlts), 2448 NEONMAP2(vabdq_v, arm_neon_vabdu, arm_neon_vabds, Add1ArgType | UnsignedAlts), 2449 NEONMAP1(vabs_v, arm_neon_vabs, 0), 2450 NEONMAP1(vabsq_v, arm_neon_vabs, 0), 2451 NEONMAP0(vaddhn_v), 2452 NEONMAP1(vaesdq_v, arm_neon_aesd, 0), 2453 NEONMAP1(vaeseq_v, arm_neon_aese, 0), 2454 NEONMAP1(vaesimcq_v, arm_neon_aesimc, 0), 2455 NEONMAP1(vaesmcq_v, arm_neon_aesmc, 0), 2456 NEONMAP1(vbsl_v, arm_neon_vbsl, AddRetType), 2457 NEONMAP1(vbslq_v, arm_neon_vbsl, AddRetType), 2458 NEONMAP1(vcage_v, arm_neon_vacge, 0), 2459 NEONMAP1(vcageq_v, arm_neon_vacge, 0), 2460 NEONMAP1(vcagt_v, arm_neon_vacgt, 0), 2461 NEONMAP1(vcagtq_v, arm_neon_vacgt, 0), 2462 NEONMAP1(vcale_v, arm_neon_vacge, 0), 2463 NEONMAP1(vcaleq_v, arm_neon_vacge, 0), 2464 NEONMAP1(vcalt_v, arm_neon_vacgt, 0), 2465 NEONMAP1(vcaltq_v, arm_neon_vacgt, 0), 2466 NEONMAP1(vcls_v, arm_neon_vcls, Add1ArgType), 2467 NEONMAP1(vclsq_v, arm_neon_vcls, Add1ArgType), 2468 NEONMAP1(vclz_v, ctlz, Add1ArgType), 2469 NEONMAP1(vclzq_v, ctlz, Add1ArgType), 2470 NEONMAP1(vcnt_v, ctpop, Add1ArgType), 2471 NEONMAP1(vcntq_v, ctpop, Add1ArgType), 2472 NEONMAP1(vcvt_f16_f32, arm_neon_vcvtfp2hf, 0), 2473 NEONMAP1(vcvt_f32_f16, arm_neon_vcvthf2fp, 0), 2474 NEONMAP0(vcvt_f32_v), 2475 NEONMAP2(vcvt_n_f32_v, arm_neon_vcvtfxu2fp, arm_neon_vcvtfxs2fp, 0), 2476 NEONMAP1(vcvt_n_s32_v, arm_neon_vcvtfp2fxs, 0), 2477 NEONMAP1(vcvt_n_s64_v, arm_neon_vcvtfp2fxs, 0), 2478 NEONMAP1(vcvt_n_u32_v, arm_neon_vcvtfp2fxu, 0), 2479 NEONMAP1(vcvt_n_u64_v, arm_neon_vcvtfp2fxu, 0), 2480 NEONMAP0(vcvt_s32_v), 2481 NEONMAP0(vcvt_s64_v), 2482 NEONMAP0(vcvt_u32_v), 2483 NEONMAP0(vcvt_u64_v), 2484 NEONMAP1(vcvta_s32_v, arm_neon_vcvtas, 0), 2485 NEONMAP1(vcvta_s64_v, arm_neon_vcvtas, 0), 2486 NEONMAP1(vcvta_u32_v, arm_neon_vcvtau, 0), 2487 NEONMAP1(vcvta_u64_v, arm_neon_vcvtau, 0), 2488 NEONMAP1(vcvtaq_s32_v, arm_neon_vcvtas, 0), 2489 NEONMAP1(vcvtaq_s64_v, arm_neon_vcvtas, 0), 2490 NEONMAP1(vcvtaq_u32_v, arm_neon_vcvtau, 0), 2491 NEONMAP1(vcvtaq_u64_v, arm_neon_vcvtau, 0), 2492 NEONMAP1(vcvtm_s32_v, arm_neon_vcvtms, 0), 2493 NEONMAP1(vcvtm_s64_v, arm_neon_vcvtms, 0), 2494 NEONMAP1(vcvtm_u32_v, arm_neon_vcvtmu, 0), 2495 NEONMAP1(vcvtm_u64_v, arm_neon_vcvtmu, 0), 2496 NEONMAP1(vcvtmq_s32_v, arm_neon_vcvtms, 0), 2497 NEONMAP1(vcvtmq_s64_v, arm_neon_vcvtms, 0), 2498 NEONMAP1(vcvtmq_u32_v, arm_neon_vcvtmu, 0), 2499 NEONMAP1(vcvtmq_u64_v, arm_neon_vcvtmu, 0), 2500 NEONMAP1(vcvtn_s32_v, arm_neon_vcvtns, 0), 2501 NEONMAP1(vcvtn_s64_v, arm_neon_vcvtns, 0), 2502 NEONMAP1(vcvtn_u32_v, arm_neon_vcvtnu, 0), 2503 NEONMAP1(vcvtn_u64_v, arm_neon_vcvtnu, 0), 2504 NEONMAP1(vcvtnq_s32_v, arm_neon_vcvtns, 0), 2505 NEONMAP1(vcvtnq_s64_v, arm_neon_vcvtns, 0), 2506 NEONMAP1(vcvtnq_u32_v, arm_neon_vcvtnu, 0), 2507 NEONMAP1(vcvtnq_u64_v, arm_neon_vcvtnu, 0), 2508 NEONMAP1(vcvtp_s32_v, arm_neon_vcvtps, 0), 2509 NEONMAP1(vcvtp_s64_v, arm_neon_vcvtps, 0), 2510 NEONMAP1(vcvtp_u32_v, arm_neon_vcvtpu, 0), 2511 NEONMAP1(vcvtp_u64_v, arm_neon_vcvtpu, 0), 2512 NEONMAP1(vcvtpq_s32_v, arm_neon_vcvtps, 0), 2513 NEONMAP1(vcvtpq_s64_v, arm_neon_vcvtps, 0), 2514 NEONMAP1(vcvtpq_u32_v, arm_neon_vcvtpu, 0), 2515 NEONMAP1(vcvtpq_u64_v, arm_neon_vcvtpu, 0), 2516 NEONMAP0(vcvtq_f32_v), 2517 NEONMAP2(vcvtq_n_f32_v, arm_neon_vcvtfxu2fp, arm_neon_vcvtfxs2fp, 0), 2518 NEONMAP1(vcvtq_n_s32_v, arm_neon_vcvtfp2fxs, 0), 2519 NEONMAP1(vcvtq_n_s64_v, arm_neon_vcvtfp2fxs, 0), 2520 NEONMAP1(vcvtq_n_u32_v, arm_neon_vcvtfp2fxu, 0), 2521 NEONMAP1(vcvtq_n_u64_v, arm_neon_vcvtfp2fxu, 0), 2522 NEONMAP0(vcvtq_s32_v), 2523 NEONMAP0(vcvtq_s64_v), 2524 NEONMAP0(vcvtq_u32_v), 2525 NEONMAP0(vcvtq_u64_v), 2526 NEONMAP0(vext_v), 2527 NEONMAP0(vextq_v), 2528 NEONMAP0(vfma_v), 2529 NEONMAP0(vfmaq_v), 2530 NEONMAP2(vhadd_v, arm_neon_vhaddu, arm_neon_vhadds, Add1ArgType | UnsignedAlts), 2531 NEONMAP2(vhaddq_v, arm_neon_vhaddu, arm_neon_vhadds, Add1ArgType | UnsignedAlts), 2532 NEONMAP2(vhsub_v, arm_neon_vhsubu, arm_neon_vhsubs, Add1ArgType | UnsignedAlts), 2533 NEONMAP2(vhsubq_v, arm_neon_vhsubu, arm_neon_vhsubs, Add1ArgType | UnsignedAlts), 2534 NEONMAP0(vld1_dup_v), 2535 NEONMAP1(vld1_v, arm_neon_vld1, 0), 2536 NEONMAP0(vld1q_dup_v), 2537 NEONMAP1(vld1q_v, arm_neon_vld1, 0), 2538 NEONMAP1(vld2_lane_v, arm_neon_vld2lane, 0), 2539 NEONMAP1(vld2_v, arm_neon_vld2, 0), 2540 NEONMAP1(vld2q_lane_v, arm_neon_vld2lane, 0), 2541 NEONMAP1(vld2q_v, arm_neon_vld2, 0), 2542 NEONMAP1(vld3_lane_v, arm_neon_vld3lane, 0), 2543 NEONMAP1(vld3_v, arm_neon_vld3, 0), 2544 NEONMAP1(vld3q_lane_v, arm_neon_vld3lane, 0), 2545 NEONMAP1(vld3q_v, arm_neon_vld3, 0), 2546 NEONMAP1(vld4_lane_v, arm_neon_vld4lane, 0), 2547 NEONMAP1(vld4_v, arm_neon_vld4, 0), 2548 NEONMAP1(vld4q_lane_v, arm_neon_vld4lane, 0), 2549 NEONMAP1(vld4q_v, arm_neon_vld4, 0), 2550 NEONMAP2(vmax_v, arm_neon_vmaxu, arm_neon_vmaxs, Add1ArgType | UnsignedAlts), 2551 NEONMAP1(vmaxnm_v, arm_neon_vmaxnm, Add1ArgType), 2552 NEONMAP1(vmaxnmq_v, arm_neon_vmaxnm, Add1ArgType), 2553 NEONMAP2(vmaxq_v, arm_neon_vmaxu, arm_neon_vmaxs, Add1ArgType | UnsignedAlts), 2554 NEONMAP2(vmin_v, arm_neon_vminu, arm_neon_vmins, Add1ArgType | UnsignedAlts), 2555 NEONMAP1(vminnm_v, arm_neon_vminnm, Add1ArgType), 2556 NEONMAP1(vminnmq_v, arm_neon_vminnm, Add1ArgType), 2557 NEONMAP2(vminq_v, arm_neon_vminu, arm_neon_vmins, Add1ArgType | UnsignedAlts), 2558 NEONMAP0(vmovl_v), 2559 NEONMAP0(vmovn_v), 2560 NEONMAP1(vmul_v, arm_neon_vmulp, Add1ArgType), 2561 NEONMAP0(vmull_v), 2562 NEONMAP1(vmulq_v, arm_neon_vmulp, Add1ArgType), 2563 NEONMAP2(vpadal_v, arm_neon_vpadalu, arm_neon_vpadals, UnsignedAlts), 2564 NEONMAP2(vpadalq_v, arm_neon_vpadalu, arm_neon_vpadals, UnsignedAlts), 2565 NEONMAP1(vpadd_v, arm_neon_vpadd, Add1ArgType), 2566 NEONMAP2(vpaddl_v, arm_neon_vpaddlu, arm_neon_vpaddls, UnsignedAlts), 2567 NEONMAP2(vpaddlq_v, arm_neon_vpaddlu, arm_neon_vpaddls, UnsignedAlts), 2568 NEONMAP1(vpaddq_v, arm_neon_vpadd, Add1ArgType), 2569 NEONMAP2(vpmax_v, arm_neon_vpmaxu, arm_neon_vpmaxs, Add1ArgType | UnsignedAlts), 2570 NEONMAP2(vpmin_v, arm_neon_vpminu, arm_neon_vpmins, Add1ArgType | UnsignedAlts), 2571 NEONMAP1(vqabs_v, arm_neon_vqabs, Add1ArgType), 2572 NEONMAP1(vqabsq_v, arm_neon_vqabs, Add1ArgType), 2573 NEONMAP2(vqadd_v, arm_neon_vqaddu, arm_neon_vqadds, Add1ArgType | UnsignedAlts), 2574 NEONMAP2(vqaddq_v, arm_neon_vqaddu, arm_neon_vqadds, Add1ArgType | UnsignedAlts), 2575 NEONMAP2(vqdmlal_v, arm_neon_vqdmull, arm_neon_vqadds, 0), 2576 NEONMAP2(vqdmlsl_v, arm_neon_vqdmull, arm_neon_vqsubs, 0), 2577 NEONMAP1(vqdmulh_v, arm_neon_vqdmulh, Add1ArgType), 2578 NEONMAP1(vqdmulhq_v, arm_neon_vqdmulh, Add1ArgType), 2579 NEONMAP1(vqdmull_v, arm_neon_vqdmull, Add1ArgType), 2580 NEONMAP2(vqmovn_v, arm_neon_vqmovnu, arm_neon_vqmovns, Add1ArgType | UnsignedAlts), 2581 NEONMAP1(vqmovun_v, arm_neon_vqmovnsu, Add1ArgType), 2582 NEONMAP1(vqneg_v, arm_neon_vqneg, Add1ArgType), 2583 NEONMAP1(vqnegq_v, arm_neon_vqneg, Add1ArgType), 2584 NEONMAP1(vqrdmulh_v, arm_neon_vqrdmulh, Add1ArgType), 2585 NEONMAP1(vqrdmulhq_v, arm_neon_vqrdmulh, Add1ArgType), 2586 NEONMAP2(vqrshl_v, arm_neon_vqrshiftu, arm_neon_vqrshifts, Add1ArgType | UnsignedAlts), 2587 NEONMAP2(vqrshlq_v, arm_neon_vqrshiftu, arm_neon_vqrshifts, Add1ArgType | UnsignedAlts), 2588 NEONMAP2(vqshl_n_v, arm_neon_vqshiftu, arm_neon_vqshifts, UnsignedAlts), 2589 NEONMAP2(vqshl_v, arm_neon_vqshiftu, arm_neon_vqshifts, Add1ArgType | UnsignedAlts), 2590 NEONMAP2(vqshlq_n_v, arm_neon_vqshiftu, arm_neon_vqshifts, UnsignedAlts), 2591 NEONMAP2(vqshlq_v, arm_neon_vqshiftu, arm_neon_vqshifts, Add1ArgType | UnsignedAlts), 2592 NEONMAP1(vqshlu_n_v, arm_neon_vqshiftsu, 0), 2593 NEONMAP1(vqshluq_n_v, arm_neon_vqshiftsu, 0), 2594 NEONMAP2(vqsub_v, arm_neon_vqsubu, arm_neon_vqsubs, Add1ArgType | UnsignedAlts), 2595 NEONMAP2(vqsubq_v, arm_neon_vqsubu, arm_neon_vqsubs, Add1ArgType | UnsignedAlts), 2596 NEONMAP1(vraddhn_v, arm_neon_vraddhn, Add1ArgType), 2597 NEONMAP2(vrecpe_v, arm_neon_vrecpe, arm_neon_vrecpe, 0), 2598 NEONMAP2(vrecpeq_v, arm_neon_vrecpe, arm_neon_vrecpe, 0), 2599 NEONMAP1(vrecps_v, arm_neon_vrecps, Add1ArgType), 2600 NEONMAP1(vrecpsq_v, arm_neon_vrecps, Add1ArgType), 2601 NEONMAP2(vrhadd_v, arm_neon_vrhaddu, arm_neon_vrhadds, Add1ArgType | UnsignedAlts), 2602 NEONMAP2(vrhaddq_v, arm_neon_vrhaddu, arm_neon_vrhadds, Add1ArgType | UnsignedAlts), 2603 NEONMAP1(vrnd_v, arm_neon_vrintz, Add1ArgType), 2604 NEONMAP1(vrnda_v, arm_neon_vrinta, Add1ArgType), 2605 NEONMAP1(vrndaq_v, arm_neon_vrinta, Add1ArgType), 2606 NEONMAP1(vrndm_v, arm_neon_vrintm, Add1ArgType), 2607 NEONMAP1(vrndmq_v, arm_neon_vrintm, Add1ArgType), 2608 NEONMAP1(vrndn_v, arm_neon_vrintn, Add1ArgType), 2609 NEONMAP1(vrndnq_v, arm_neon_vrintn, Add1ArgType), 2610 NEONMAP1(vrndp_v, arm_neon_vrintp, Add1ArgType), 2611 NEONMAP1(vrndpq_v, arm_neon_vrintp, Add1ArgType), 2612 NEONMAP1(vrndq_v, arm_neon_vrintz, Add1ArgType), 2613 NEONMAP1(vrndx_v, arm_neon_vrintx, Add1ArgType), 2614 NEONMAP1(vrndxq_v, arm_neon_vrintx, Add1ArgType), 2615 NEONMAP2(vrshl_v, arm_neon_vrshiftu, arm_neon_vrshifts, Add1ArgType | UnsignedAlts), 2616 NEONMAP2(vrshlq_v, arm_neon_vrshiftu, arm_neon_vrshifts, Add1ArgType | UnsignedAlts), 2617 NEONMAP2(vrshr_n_v, arm_neon_vrshiftu, arm_neon_vrshifts, UnsignedAlts), 2618 NEONMAP2(vrshrq_n_v, arm_neon_vrshiftu, arm_neon_vrshifts, UnsignedAlts), 2619 NEONMAP2(vrsqrte_v, arm_neon_vrsqrte, arm_neon_vrsqrte, 0), 2620 NEONMAP2(vrsqrteq_v, arm_neon_vrsqrte, arm_neon_vrsqrte, 0), 2621 NEONMAP1(vrsqrts_v, arm_neon_vrsqrts, Add1ArgType), 2622 NEONMAP1(vrsqrtsq_v, arm_neon_vrsqrts, Add1ArgType), 2623 NEONMAP1(vrsubhn_v, arm_neon_vrsubhn, Add1ArgType), 2624 NEONMAP1(vsha1su0q_v, arm_neon_sha1su0, 0), 2625 NEONMAP1(vsha1su1q_v, arm_neon_sha1su1, 0), 2626 NEONMAP1(vsha256h2q_v, arm_neon_sha256h2, 0), 2627 NEONMAP1(vsha256hq_v, arm_neon_sha256h, 0), 2628 NEONMAP1(vsha256su0q_v, arm_neon_sha256su0, 0), 2629 NEONMAP1(vsha256su1q_v, arm_neon_sha256su1, 0), 2630 NEONMAP0(vshl_n_v), 2631 NEONMAP2(vshl_v, arm_neon_vshiftu, arm_neon_vshifts, Add1ArgType | UnsignedAlts), 2632 NEONMAP0(vshll_n_v), 2633 NEONMAP0(vshlq_n_v), 2634 NEONMAP2(vshlq_v, arm_neon_vshiftu, arm_neon_vshifts, Add1ArgType | UnsignedAlts), 2635 NEONMAP0(vshr_n_v), 2636 NEONMAP0(vshrn_n_v), 2637 NEONMAP0(vshrq_n_v), 2638 NEONMAP1(vst1_v, arm_neon_vst1, 0), 2639 NEONMAP1(vst1q_v, arm_neon_vst1, 0), 2640 NEONMAP1(vst2_lane_v, arm_neon_vst2lane, 0), 2641 NEONMAP1(vst2_v, arm_neon_vst2, 0), 2642 NEONMAP1(vst2q_lane_v, arm_neon_vst2lane, 0), 2643 NEONMAP1(vst2q_v, arm_neon_vst2, 0), 2644 NEONMAP1(vst3_lane_v, arm_neon_vst3lane, 0), 2645 NEONMAP1(vst3_v, arm_neon_vst3, 0), 2646 NEONMAP1(vst3q_lane_v, arm_neon_vst3lane, 0), 2647 NEONMAP1(vst3q_v, arm_neon_vst3, 0), 2648 NEONMAP1(vst4_lane_v, arm_neon_vst4lane, 0), 2649 NEONMAP1(vst4_v, arm_neon_vst4, 0), 2650 NEONMAP1(vst4q_lane_v, arm_neon_vst4lane, 0), 2651 NEONMAP1(vst4q_v, arm_neon_vst4, 0), 2652 NEONMAP0(vsubhn_v), 2653 NEONMAP0(vtrn_v), 2654 NEONMAP0(vtrnq_v), 2655 NEONMAP0(vtst_v), 2656 NEONMAP0(vtstq_v), 2657 NEONMAP0(vuzp_v), 2658 NEONMAP0(vuzpq_v), 2659 NEONMAP0(vzip_v), 2660 NEONMAP0(vzipq_v) 2661 }; 2662 2663 static const NeonIntrinsicInfo AArch64SIMDIntrinsicMap[] = { 2664 NEONMAP1(vabs_v, aarch64_neon_abs, 0), 2665 NEONMAP1(vabsq_v, aarch64_neon_abs, 0), 2666 NEONMAP0(vaddhn_v), 2667 NEONMAP1(vaesdq_v, aarch64_crypto_aesd, 0), 2668 NEONMAP1(vaeseq_v, aarch64_crypto_aese, 0), 2669 NEONMAP1(vaesimcq_v, aarch64_crypto_aesimc, 0), 2670 NEONMAP1(vaesmcq_v, aarch64_crypto_aesmc, 0), 2671 NEONMAP1(vcage_v, aarch64_neon_facge, 0), 2672 NEONMAP1(vcageq_v, aarch64_neon_facge, 0), 2673 NEONMAP1(vcagt_v, aarch64_neon_facgt, 0), 2674 NEONMAP1(vcagtq_v, aarch64_neon_facgt, 0), 2675 NEONMAP1(vcale_v, aarch64_neon_facge, 0), 2676 NEONMAP1(vcaleq_v, aarch64_neon_facge, 0), 2677 NEONMAP1(vcalt_v, aarch64_neon_facgt, 0), 2678 NEONMAP1(vcaltq_v, aarch64_neon_facgt, 0), 2679 NEONMAP1(vcls_v, aarch64_neon_cls, Add1ArgType), 2680 NEONMAP1(vclsq_v, aarch64_neon_cls, Add1ArgType), 2681 NEONMAP1(vclz_v, ctlz, Add1ArgType), 2682 NEONMAP1(vclzq_v, ctlz, Add1ArgType), 2683 NEONMAP1(vcnt_v, ctpop, Add1ArgType), 2684 NEONMAP1(vcntq_v, ctpop, Add1ArgType), 2685 NEONMAP1(vcvt_f16_f32, aarch64_neon_vcvtfp2hf, 0), 2686 NEONMAP1(vcvt_f32_f16, aarch64_neon_vcvthf2fp, 0), 2687 NEONMAP0(vcvt_f32_v), 2688 NEONMAP2(vcvt_n_f32_v, aarch64_neon_vcvtfxu2fp, aarch64_neon_vcvtfxs2fp, 0), 2689 NEONMAP2(vcvt_n_f64_v, aarch64_neon_vcvtfxu2fp, aarch64_neon_vcvtfxs2fp, 0), 2690 NEONMAP1(vcvt_n_s32_v, aarch64_neon_vcvtfp2fxs, 0), 2691 NEONMAP1(vcvt_n_s64_v, aarch64_neon_vcvtfp2fxs, 0), 2692 NEONMAP1(vcvt_n_u32_v, aarch64_neon_vcvtfp2fxu, 0), 2693 NEONMAP1(vcvt_n_u64_v, aarch64_neon_vcvtfp2fxu, 0), 2694 NEONMAP0(vcvtq_f32_v), 2695 NEONMAP2(vcvtq_n_f32_v, aarch64_neon_vcvtfxu2fp, aarch64_neon_vcvtfxs2fp, 0), 2696 NEONMAP2(vcvtq_n_f64_v, aarch64_neon_vcvtfxu2fp, aarch64_neon_vcvtfxs2fp, 0), 2697 NEONMAP1(vcvtq_n_s32_v, aarch64_neon_vcvtfp2fxs, 0), 2698 NEONMAP1(vcvtq_n_s64_v, aarch64_neon_vcvtfp2fxs, 0), 2699 NEONMAP1(vcvtq_n_u32_v, aarch64_neon_vcvtfp2fxu, 0), 2700 NEONMAP1(vcvtq_n_u64_v, aarch64_neon_vcvtfp2fxu, 0), 2701 NEONMAP1(vcvtx_f32_v, aarch64_neon_fcvtxn, AddRetType | Add1ArgType), 2702 NEONMAP0(vext_v), 2703 NEONMAP0(vextq_v), 2704 NEONMAP0(vfma_v), 2705 NEONMAP0(vfmaq_v), 2706 NEONMAP2(vhadd_v, aarch64_neon_uhadd, aarch64_neon_shadd, Add1ArgType | UnsignedAlts), 2707 NEONMAP2(vhaddq_v, aarch64_neon_uhadd, aarch64_neon_shadd, Add1ArgType | UnsignedAlts), 2708 NEONMAP2(vhsub_v, aarch64_neon_uhsub, aarch64_neon_shsub, Add1ArgType | UnsignedAlts), 2709 NEONMAP2(vhsubq_v, aarch64_neon_uhsub, aarch64_neon_shsub, Add1ArgType | UnsignedAlts), 2710 NEONMAP0(vmovl_v), 2711 NEONMAP0(vmovn_v), 2712 NEONMAP1(vmul_v, aarch64_neon_pmul, Add1ArgType), 2713 NEONMAP1(vmulq_v, aarch64_neon_pmul, Add1ArgType), 2714 NEONMAP1(vpadd_v, aarch64_neon_addp, Add1ArgType), 2715 NEONMAP2(vpaddl_v, aarch64_neon_uaddlp, aarch64_neon_saddlp, UnsignedAlts), 2716 NEONMAP2(vpaddlq_v, aarch64_neon_uaddlp, aarch64_neon_saddlp, UnsignedAlts), 2717 NEONMAP1(vpaddq_v, aarch64_neon_addp, Add1ArgType), 2718 NEONMAP1(vqabs_v, aarch64_neon_sqabs, Add1ArgType), 2719 NEONMAP1(vqabsq_v, aarch64_neon_sqabs, Add1ArgType), 2720 NEONMAP2(vqadd_v, aarch64_neon_uqadd, aarch64_neon_sqadd, Add1ArgType | UnsignedAlts), 2721 NEONMAP2(vqaddq_v, aarch64_neon_uqadd, aarch64_neon_sqadd, Add1ArgType | UnsignedAlts), 2722 NEONMAP2(vqdmlal_v, aarch64_neon_sqdmull, aarch64_neon_sqadd, 0), 2723 NEONMAP2(vqdmlsl_v, aarch64_neon_sqdmull, aarch64_neon_sqsub, 0), 2724 NEONMAP1(vqdmulh_v, aarch64_neon_sqdmulh, Add1ArgType), 2725 NEONMAP1(vqdmulhq_v, aarch64_neon_sqdmulh, Add1ArgType), 2726 NEONMAP1(vqdmull_v, aarch64_neon_sqdmull, Add1ArgType), 2727 NEONMAP2(vqmovn_v, aarch64_neon_uqxtn, aarch64_neon_sqxtn, Add1ArgType | UnsignedAlts), 2728 NEONMAP1(vqmovun_v, aarch64_neon_sqxtun, Add1ArgType), 2729 NEONMAP1(vqneg_v, aarch64_neon_sqneg, Add1ArgType), 2730 NEONMAP1(vqnegq_v, aarch64_neon_sqneg, Add1ArgType), 2731 NEONMAP1(vqrdmulh_v, aarch64_neon_sqrdmulh, Add1ArgType), 2732 NEONMAP1(vqrdmulhq_v, aarch64_neon_sqrdmulh, Add1ArgType), 2733 NEONMAP2(vqrshl_v, aarch64_neon_uqrshl, aarch64_neon_sqrshl, Add1ArgType | UnsignedAlts), 2734 NEONMAP2(vqrshlq_v, aarch64_neon_uqrshl, aarch64_neon_sqrshl, Add1ArgType | UnsignedAlts), 2735 NEONMAP2(vqshl_n_v, aarch64_neon_uqshl, aarch64_neon_sqshl, UnsignedAlts), 2736 NEONMAP2(vqshl_v, aarch64_neon_uqshl, aarch64_neon_sqshl, Add1ArgType | UnsignedAlts), 2737 NEONMAP2(vqshlq_n_v, aarch64_neon_uqshl, aarch64_neon_sqshl,UnsignedAlts), 2738 NEONMAP2(vqshlq_v, aarch64_neon_uqshl, aarch64_neon_sqshl, Add1ArgType | UnsignedAlts), 2739 NEONMAP1(vqshlu_n_v, aarch64_neon_sqshlu, 0), 2740 NEONMAP1(vqshluq_n_v, aarch64_neon_sqshlu, 0), 2741 NEONMAP2(vqsub_v, aarch64_neon_uqsub, aarch64_neon_sqsub, Add1ArgType | UnsignedAlts), 2742 NEONMAP2(vqsubq_v, aarch64_neon_uqsub, aarch64_neon_sqsub, Add1ArgType | UnsignedAlts), 2743 NEONMAP1(vraddhn_v, aarch64_neon_raddhn, Add1ArgType), 2744 NEONMAP2(vrecpe_v, aarch64_neon_frecpe, aarch64_neon_urecpe, 0), 2745 NEONMAP2(vrecpeq_v, aarch64_neon_frecpe, aarch64_neon_urecpe, 0), 2746 NEONMAP1(vrecps_v, aarch64_neon_frecps, Add1ArgType), 2747 NEONMAP1(vrecpsq_v, aarch64_neon_frecps, Add1ArgType), 2748 NEONMAP2(vrhadd_v, aarch64_neon_urhadd, aarch64_neon_srhadd, Add1ArgType | UnsignedAlts), 2749 NEONMAP2(vrhaddq_v, aarch64_neon_urhadd, aarch64_neon_srhadd, Add1ArgType | UnsignedAlts), 2750 NEONMAP2(vrshl_v, aarch64_neon_urshl, aarch64_neon_srshl, Add1ArgType | UnsignedAlts), 2751 NEONMAP2(vrshlq_v, aarch64_neon_urshl, aarch64_neon_srshl, Add1ArgType | UnsignedAlts), 2752 NEONMAP2(vrshr_n_v, aarch64_neon_urshl, aarch64_neon_srshl, UnsignedAlts), 2753 NEONMAP2(vrshrq_n_v, aarch64_neon_urshl, aarch64_neon_srshl, UnsignedAlts), 2754 NEONMAP2(vrsqrte_v, aarch64_neon_frsqrte, aarch64_neon_ursqrte, 0), 2755 NEONMAP2(vrsqrteq_v, aarch64_neon_frsqrte, aarch64_neon_ursqrte, 0), 2756 NEONMAP1(vrsqrts_v, aarch64_neon_frsqrts, Add1ArgType), 2757 NEONMAP1(vrsqrtsq_v, aarch64_neon_frsqrts, Add1ArgType), 2758 NEONMAP1(vrsubhn_v, aarch64_neon_rsubhn, Add1ArgType), 2759 NEONMAP1(vsha1su0q_v, aarch64_crypto_sha1su0, 0), 2760 NEONMAP1(vsha1su1q_v, aarch64_crypto_sha1su1, 0), 2761 NEONMAP1(vsha256h2q_v, aarch64_crypto_sha256h2, 0), 2762 NEONMAP1(vsha256hq_v, aarch64_crypto_sha256h, 0), 2763 NEONMAP1(vsha256su0q_v, aarch64_crypto_sha256su0, 0), 2764 NEONMAP1(vsha256su1q_v, aarch64_crypto_sha256su1, 0), 2765 NEONMAP0(vshl_n_v), 2766 NEONMAP2(vshl_v, aarch64_neon_ushl, aarch64_neon_sshl, Add1ArgType | UnsignedAlts), 2767 NEONMAP0(vshll_n_v), 2768 NEONMAP0(vshlq_n_v), 2769 NEONMAP2(vshlq_v, aarch64_neon_ushl, aarch64_neon_sshl, Add1ArgType | UnsignedAlts), 2770 NEONMAP0(vshr_n_v), 2771 NEONMAP0(vshrn_n_v), 2772 NEONMAP0(vshrq_n_v), 2773 NEONMAP0(vsubhn_v), 2774 NEONMAP0(vtst_v), 2775 NEONMAP0(vtstq_v), 2776 }; 2777 2778 static const NeonIntrinsicInfo AArch64SISDIntrinsicMap[] = { 2779 NEONMAP1(vabdd_f64, aarch64_sisd_fabd, Add1ArgType), 2780 NEONMAP1(vabds_f32, aarch64_sisd_fabd, Add1ArgType), 2781 NEONMAP1(vabsd_s64, aarch64_neon_abs, Add1ArgType), 2782 NEONMAP1(vaddlv_s32, aarch64_neon_saddlv, AddRetType | Add1ArgType), 2783 NEONMAP1(vaddlv_u32, aarch64_neon_uaddlv, AddRetType | Add1ArgType), 2784 NEONMAP1(vaddlvq_s32, aarch64_neon_saddlv, AddRetType | Add1ArgType), 2785 NEONMAP1(vaddlvq_u32, aarch64_neon_uaddlv, AddRetType | Add1ArgType), 2786 NEONMAP1(vaddv_f32, aarch64_neon_faddv, AddRetType | Add1ArgType), 2787 NEONMAP1(vaddv_s32, aarch64_neon_saddv, AddRetType | Add1ArgType), 2788 NEONMAP1(vaddv_u32, aarch64_neon_uaddv, AddRetType | Add1ArgType), 2789 NEONMAP1(vaddvq_f32, aarch64_neon_faddv, AddRetType | Add1ArgType), 2790 NEONMAP1(vaddvq_f64, aarch64_neon_faddv, AddRetType | Add1ArgType), 2791 NEONMAP1(vaddvq_s32, aarch64_neon_saddv, AddRetType | Add1ArgType), 2792 NEONMAP1(vaddvq_s64, aarch64_neon_saddv, AddRetType | Add1ArgType), 2793 NEONMAP1(vaddvq_u32, aarch64_neon_uaddv, AddRetType | Add1ArgType), 2794 NEONMAP1(vaddvq_u64, aarch64_neon_uaddv, AddRetType | Add1ArgType), 2795 NEONMAP1(vcaged_f64, aarch64_neon_facge, AddRetType | Add1ArgType), 2796 NEONMAP1(vcages_f32, aarch64_neon_facge, AddRetType | Add1ArgType), 2797 NEONMAP1(vcagtd_f64, aarch64_neon_facgt, AddRetType | Add1ArgType), 2798 NEONMAP1(vcagts_f32, aarch64_neon_facgt, AddRetType | Add1ArgType), 2799 NEONMAP1(vcaled_f64, aarch64_neon_facge, AddRetType | Add1ArgType), 2800 NEONMAP1(vcales_f32, aarch64_neon_facge, AddRetType | Add1ArgType), 2801 NEONMAP1(vcaltd_f64, aarch64_neon_facgt, AddRetType | Add1ArgType), 2802 NEONMAP1(vcalts_f32, aarch64_neon_facgt, AddRetType | Add1ArgType), 2803 NEONMAP1(vcvtad_s64_f64, aarch64_neon_fcvtas, AddRetType | Add1ArgType), 2804 NEONMAP1(vcvtad_u64_f64, aarch64_neon_fcvtau, AddRetType | Add1ArgType), 2805 NEONMAP1(vcvtas_s32_f32, aarch64_neon_fcvtas, AddRetType | Add1ArgType), 2806 NEONMAP1(vcvtas_u32_f32, aarch64_neon_fcvtau, AddRetType | Add1ArgType), 2807 NEONMAP1(vcvtd_n_f64_s64, aarch64_neon_vcvtfxs2fp, AddRetType | Add1ArgType), 2808 NEONMAP1(vcvtd_n_f64_u64, aarch64_neon_vcvtfxu2fp, AddRetType | Add1ArgType), 2809 NEONMAP1(vcvtd_n_s64_f64, aarch64_neon_vcvtfp2fxs, AddRetType | Add1ArgType), 2810 NEONMAP1(vcvtd_n_u64_f64, aarch64_neon_vcvtfp2fxu, AddRetType | Add1ArgType), 2811 NEONMAP1(vcvtmd_s64_f64, aarch64_neon_fcvtms, AddRetType | Add1ArgType), 2812 NEONMAP1(vcvtmd_u64_f64, aarch64_neon_fcvtmu, AddRetType | Add1ArgType), 2813 NEONMAP1(vcvtms_s32_f32, aarch64_neon_fcvtms, AddRetType | Add1ArgType), 2814 NEONMAP1(vcvtms_u32_f32, aarch64_neon_fcvtmu, AddRetType | Add1ArgType), 2815 NEONMAP1(vcvtnd_s64_f64, aarch64_neon_fcvtns, AddRetType | Add1ArgType), 2816 NEONMAP1(vcvtnd_u64_f64, aarch64_neon_fcvtnu, AddRetType | Add1ArgType), 2817 NEONMAP1(vcvtns_s32_f32, aarch64_neon_fcvtns, AddRetType | Add1ArgType), 2818 NEONMAP1(vcvtns_u32_f32, aarch64_neon_fcvtnu, AddRetType | Add1ArgType), 2819 NEONMAP1(vcvtpd_s64_f64, aarch64_neon_fcvtps, AddRetType | Add1ArgType), 2820 NEONMAP1(vcvtpd_u64_f64, aarch64_neon_fcvtpu, AddRetType | Add1ArgType), 2821 NEONMAP1(vcvtps_s32_f32, aarch64_neon_fcvtps, AddRetType | Add1ArgType), 2822 NEONMAP1(vcvtps_u32_f32, aarch64_neon_fcvtpu, AddRetType | Add1ArgType), 2823 NEONMAP1(vcvts_n_f32_s32, aarch64_neon_vcvtfxs2fp, AddRetType | Add1ArgType), 2824 NEONMAP1(vcvts_n_f32_u32, aarch64_neon_vcvtfxu2fp, AddRetType | Add1ArgType), 2825 NEONMAP1(vcvts_n_s32_f32, aarch64_neon_vcvtfp2fxs, AddRetType | Add1ArgType), 2826 NEONMAP1(vcvts_n_u32_f32, aarch64_neon_vcvtfp2fxu, AddRetType | Add1ArgType), 2827 NEONMAP1(vcvtxd_f32_f64, aarch64_sisd_fcvtxn, 0), 2828 NEONMAP1(vmaxnmv_f32, aarch64_neon_fmaxnmv, AddRetType | Add1ArgType), 2829 NEONMAP1(vmaxnmvq_f32, aarch64_neon_fmaxnmv, AddRetType | Add1ArgType), 2830 NEONMAP1(vmaxnmvq_f64, aarch64_neon_fmaxnmv, AddRetType | Add1ArgType), 2831 NEONMAP1(vmaxv_f32, aarch64_neon_fmaxv, AddRetType | Add1ArgType), 2832 NEONMAP1(vmaxv_s32, aarch64_neon_smaxv, AddRetType | Add1ArgType), 2833 NEONMAP1(vmaxv_u32, aarch64_neon_umaxv, AddRetType | Add1ArgType), 2834 NEONMAP1(vmaxvq_f32, aarch64_neon_fmaxv, AddRetType | Add1ArgType), 2835 NEONMAP1(vmaxvq_f64, aarch64_neon_fmaxv, AddRetType | Add1ArgType), 2836 NEONMAP1(vmaxvq_s32, aarch64_neon_smaxv, AddRetType | Add1ArgType), 2837 NEONMAP1(vmaxvq_u32, aarch64_neon_umaxv, AddRetType | Add1ArgType), 2838 NEONMAP1(vminnmv_f32, aarch64_neon_fminnmv, AddRetType | Add1ArgType), 2839 NEONMAP1(vminnmvq_f32, aarch64_neon_fminnmv, AddRetType | Add1ArgType), 2840 NEONMAP1(vminnmvq_f64, aarch64_neon_fminnmv, AddRetType | Add1ArgType), 2841 NEONMAP1(vminv_f32, aarch64_neon_fminv, AddRetType | Add1ArgType), 2842 NEONMAP1(vminv_s32, aarch64_neon_sminv, AddRetType | Add1ArgType), 2843 NEONMAP1(vminv_u32, aarch64_neon_uminv, AddRetType | Add1ArgType), 2844 NEONMAP1(vminvq_f32, aarch64_neon_fminv, AddRetType | Add1ArgType), 2845 NEONMAP1(vminvq_f64, aarch64_neon_fminv, AddRetType | Add1ArgType), 2846 NEONMAP1(vminvq_s32, aarch64_neon_sminv, AddRetType | Add1ArgType), 2847 NEONMAP1(vminvq_u32, aarch64_neon_uminv, AddRetType | Add1ArgType), 2848 NEONMAP1(vmull_p64, aarch64_neon_pmull64, 0), 2849 NEONMAP1(vmulxd_f64, aarch64_neon_fmulx, Add1ArgType), 2850 NEONMAP1(vmulxs_f32, aarch64_neon_fmulx, Add1ArgType), 2851 NEONMAP1(vpaddd_s64, aarch64_neon_uaddv, AddRetType | Add1ArgType), 2852 NEONMAP1(vpaddd_u64, aarch64_neon_uaddv, AddRetType | Add1ArgType), 2853 NEONMAP1(vpmaxnmqd_f64, aarch64_neon_fmaxnmv, AddRetType | Add1ArgType), 2854 NEONMAP1(vpmaxnms_f32, aarch64_neon_fmaxnmv, AddRetType | Add1ArgType), 2855 NEONMAP1(vpmaxqd_f64, aarch64_neon_fmaxv, AddRetType | Add1ArgType), 2856 NEONMAP1(vpmaxs_f32, aarch64_neon_fmaxv, AddRetType | Add1ArgType), 2857 NEONMAP1(vpminnmqd_f64, aarch64_neon_fminnmv, AddRetType | Add1ArgType), 2858 NEONMAP1(vpminnms_f32, aarch64_neon_fminnmv, AddRetType | Add1ArgType), 2859 NEONMAP1(vpminqd_f64, aarch64_neon_fminv, AddRetType | Add1ArgType), 2860 NEONMAP1(vpmins_f32, aarch64_neon_fminv, AddRetType | Add1ArgType), 2861 NEONMAP1(vqabsb_s8, aarch64_neon_sqabs, Vectorize1ArgType | Use64BitVectors), 2862 NEONMAP1(vqabsd_s64, aarch64_neon_sqabs, Add1ArgType), 2863 NEONMAP1(vqabsh_s16, aarch64_neon_sqabs, Vectorize1ArgType | Use64BitVectors), 2864 NEONMAP1(vqabss_s32, aarch64_neon_sqabs, Add1ArgType), 2865 NEONMAP1(vqaddb_s8, aarch64_neon_sqadd, Vectorize1ArgType | Use64BitVectors), 2866 NEONMAP1(vqaddb_u8, aarch64_neon_uqadd, Vectorize1ArgType | Use64BitVectors), 2867 NEONMAP1(vqaddd_s64, aarch64_neon_sqadd, Add1ArgType), 2868 NEONMAP1(vqaddd_u64, aarch64_neon_uqadd, Add1ArgType), 2869 NEONMAP1(vqaddh_s16, aarch64_neon_sqadd, Vectorize1ArgType | Use64BitVectors), 2870 NEONMAP1(vqaddh_u16, aarch64_neon_uqadd, Vectorize1ArgType | Use64BitVectors), 2871 NEONMAP1(vqadds_s32, aarch64_neon_sqadd, Add1ArgType), 2872 NEONMAP1(vqadds_u32, aarch64_neon_uqadd, Add1ArgType), 2873 NEONMAP1(vqdmulhh_s16, aarch64_neon_sqdmulh, Vectorize1ArgType | Use64BitVectors), 2874 NEONMAP1(vqdmulhs_s32, aarch64_neon_sqdmulh, Add1ArgType), 2875 NEONMAP1(vqdmullh_s16, aarch64_neon_sqdmull, VectorRet | Use128BitVectors), 2876 NEONMAP1(vqdmulls_s32, aarch64_neon_sqdmulls_scalar, 0), 2877 NEONMAP1(vqmovnd_s64, aarch64_neon_scalar_sqxtn, AddRetType | Add1ArgType), 2878 NEONMAP1(vqmovnd_u64, aarch64_neon_scalar_uqxtn, AddRetType | Add1ArgType), 2879 NEONMAP1(vqmovnh_s16, aarch64_neon_sqxtn, VectorRet | Use64BitVectors), 2880 NEONMAP1(vqmovnh_u16, aarch64_neon_uqxtn, VectorRet | Use64BitVectors), 2881 NEONMAP1(vqmovns_s32, aarch64_neon_sqxtn, VectorRet | Use64BitVectors), 2882 NEONMAP1(vqmovns_u32, aarch64_neon_uqxtn, VectorRet | Use64BitVectors), 2883 NEONMAP1(vqmovund_s64, aarch64_neon_scalar_sqxtun, AddRetType | Add1ArgType), 2884 NEONMAP1(vqmovunh_s16, aarch64_neon_sqxtun, VectorRet | Use64BitVectors), 2885 NEONMAP1(vqmovuns_s32, aarch64_neon_sqxtun, VectorRet | Use64BitVectors), 2886 NEONMAP1(vqnegb_s8, aarch64_neon_sqneg, Vectorize1ArgType | Use64BitVectors), 2887 NEONMAP1(vqnegd_s64, aarch64_neon_sqneg, Add1ArgType), 2888 NEONMAP1(vqnegh_s16, aarch64_neon_sqneg, Vectorize1ArgType | Use64BitVectors), 2889 NEONMAP1(vqnegs_s32, aarch64_neon_sqneg, Add1ArgType), 2890 NEONMAP1(vqrdmulhh_s16, aarch64_neon_sqrdmulh, Vectorize1ArgType | Use64BitVectors), 2891 NEONMAP1(vqrdmulhs_s32, aarch64_neon_sqrdmulh, Add1ArgType), 2892 NEONMAP1(vqrshlb_s8, aarch64_neon_sqrshl, Vectorize1ArgType | Use64BitVectors), 2893 NEONMAP1(vqrshlb_u8, aarch64_neon_uqrshl, Vectorize1ArgType | Use64BitVectors), 2894 NEONMAP1(vqrshld_s64, aarch64_neon_sqrshl, Add1ArgType), 2895 NEONMAP1(vqrshld_u64, aarch64_neon_uqrshl, Add1ArgType), 2896 NEONMAP1(vqrshlh_s16, aarch64_neon_sqrshl, Vectorize1ArgType | Use64BitVectors), 2897 NEONMAP1(vqrshlh_u16, aarch64_neon_uqrshl, Vectorize1ArgType | Use64BitVectors), 2898 NEONMAP1(vqrshls_s32, aarch64_neon_sqrshl, Add1ArgType), 2899 NEONMAP1(vqrshls_u32, aarch64_neon_uqrshl, Add1ArgType), 2900 NEONMAP1(vqrshrnd_n_s64, aarch64_neon_sqrshrn, AddRetType), 2901 NEONMAP1(vqrshrnd_n_u64, aarch64_neon_uqrshrn, AddRetType), 2902 NEONMAP1(vqrshrnh_n_s16, aarch64_neon_sqrshrn, VectorRet | Use64BitVectors), 2903 NEONMAP1(vqrshrnh_n_u16, aarch64_neon_uqrshrn, VectorRet | Use64BitVectors), 2904 NEONMAP1(vqrshrns_n_s32, aarch64_neon_sqrshrn, VectorRet | Use64BitVectors), 2905 NEONMAP1(vqrshrns_n_u32, aarch64_neon_uqrshrn, VectorRet | Use64BitVectors), 2906 NEONMAP1(vqrshrund_n_s64, aarch64_neon_sqrshrun, AddRetType), 2907 NEONMAP1(vqrshrunh_n_s16, aarch64_neon_sqrshrun, VectorRet | Use64BitVectors), 2908 NEONMAP1(vqrshruns_n_s32, aarch64_neon_sqrshrun, VectorRet | Use64BitVectors), 2909 NEONMAP1(vqshlb_n_s8, aarch64_neon_sqshl, Vectorize1ArgType | Use64BitVectors), 2910 NEONMAP1(vqshlb_n_u8, aarch64_neon_uqshl, Vectorize1ArgType | Use64BitVectors), 2911 NEONMAP1(vqshlb_s8, aarch64_neon_sqshl, Vectorize1ArgType | Use64BitVectors), 2912 NEONMAP1(vqshlb_u8, aarch64_neon_uqshl, Vectorize1ArgType | Use64BitVectors), 2913 NEONMAP1(vqshld_s64, aarch64_neon_sqshl, Add1ArgType), 2914 NEONMAP1(vqshld_u64, aarch64_neon_uqshl, Add1ArgType), 2915 NEONMAP1(vqshlh_n_s16, aarch64_neon_sqshl, Vectorize1ArgType | Use64BitVectors), 2916 NEONMAP1(vqshlh_n_u16, aarch64_neon_uqshl, Vectorize1ArgType | Use64BitVectors), 2917 NEONMAP1(vqshlh_s16, aarch64_neon_sqshl, Vectorize1ArgType | Use64BitVectors), 2918 NEONMAP1(vqshlh_u16, aarch64_neon_uqshl, Vectorize1ArgType | Use64BitVectors), 2919 NEONMAP1(vqshls_n_s32, aarch64_neon_sqshl, Add1ArgType), 2920 NEONMAP1(vqshls_n_u32, aarch64_neon_uqshl, Add1ArgType), 2921 NEONMAP1(vqshls_s32, aarch64_neon_sqshl, Add1ArgType), 2922 NEONMAP1(vqshls_u32, aarch64_neon_uqshl, Add1ArgType), 2923 NEONMAP1(vqshlub_n_s8, aarch64_neon_sqshlu, Vectorize1ArgType | Use64BitVectors), 2924 NEONMAP1(vqshluh_n_s16, aarch64_neon_sqshlu, Vectorize1ArgType | Use64BitVectors), 2925 NEONMAP1(vqshlus_n_s32, aarch64_neon_sqshlu, Add1ArgType), 2926 NEONMAP1(vqshrnd_n_s64, aarch64_neon_sqshrn, AddRetType), 2927 NEONMAP1(vqshrnd_n_u64, aarch64_neon_uqshrn, AddRetType), 2928 NEONMAP1(vqshrnh_n_s16, aarch64_neon_sqshrn, VectorRet | Use64BitVectors), 2929 NEONMAP1(vqshrnh_n_u16, aarch64_neon_uqshrn, VectorRet | Use64BitVectors), 2930 NEONMAP1(vqshrns_n_s32, aarch64_neon_sqshrn, VectorRet | Use64BitVectors), 2931 NEONMAP1(vqshrns_n_u32, aarch64_neon_uqshrn, VectorRet | Use64BitVectors), 2932 NEONMAP1(vqshrund_n_s64, aarch64_neon_sqshrun, AddRetType), 2933 NEONMAP1(vqshrunh_n_s16, aarch64_neon_sqshrun, VectorRet | Use64BitVectors), 2934 NEONMAP1(vqshruns_n_s32, aarch64_neon_sqshrun, VectorRet | Use64BitVectors), 2935 NEONMAP1(vqsubb_s8, aarch64_neon_sqsub, Vectorize1ArgType | Use64BitVectors), 2936 NEONMAP1(vqsubb_u8, aarch64_neon_uqsub, Vectorize1ArgType | Use64BitVectors), 2937 NEONMAP1(vqsubd_s64, aarch64_neon_sqsub, Add1ArgType), 2938 NEONMAP1(vqsubd_u64, aarch64_neon_uqsub, Add1ArgType), 2939 NEONMAP1(vqsubh_s16, aarch64_neon_sqsub, Vectorize1ArgType | Use64BitVectors), 2940 NEONMAP1(vqsubh_u16, aarch64_neon_uqsub, Vectorize1ArgType | Use64BitVectors), 2941 NEONMAP1(vqsubs_s32, aarch64_neon_sqsub, Add1ArgType), 2942 NEONMAP1(vqsubs_u32, aarch64_neon_uqsub, Add1ArgType), 2943 NEONMAP1(vrecped_f64, aarch64_neon_frecpe, Add1ArgType), 2944 NEONMAP1(vrecpes_f32, aarch64_neon_frecpe, Add1ArgType), 2945 NEONMAP1(vrecpxd_f64, aarch64_neon_frecpx, Add1ArgType), 2946 NEONMAP1(vrecpxs_f32, aarch64_neon_frecpx, Add1ArgType), 2947 NEONMAP1(vrshld_s64, aarch64_neon_srshl, Add1ArgType), 2948 NEONMAP1(vrshld_u64, aarch64_neon_urshl, Add1ArgType), 2949 NEONMAP1(vrsqrted_f64, aarch64_neon_frsqrte, Add1ArgType), 2950 NEONMAP1(vrsqrtes_f32, aarch64_neon_frsqrte, Add1ArgType), 2951 NEONMAP1(vrsqrtsd_f64, aarch64_neon_frsqrts, Add1ArgType), 2952 NEONMAP1(vrsqrtss_f32, aarch64_neon_frsqrts, Add1ArgType), 2953 NEONMAP1(vsha1cq_u32, aarch64_crypto_sha1c, 0), 2954 NEONMAP1(vsha1h_u32, aarch64_crypto_sha1h, 0), 2955 NEONMAP1(vsha1mq_u32, aarch64_crypto_sha1m, 0), 2956 NEONMAP1(vsha1pq_u32, aarch64_crypto_sha1p, 0), 2957 NEONMAP1(vshld_s64, aarch64_neon_sshl, Add1ArgType), 2958 NEONMAP1(vshld_u64, aarch64_neon_ushl, Add1ArgType), 2959 NEONMAP1(vslid_n_s64, aarch64_neon_vsli, Vectorize1ArgType), 2960 NEONMAP1(vslid_n_u64, aarch64_neon_vsli, Vectorize1ArgType), 2961 NEONMAP1(vsqaddb_u8, aarch64_neon_usqadd, Vectorize1ArgType | Use64BitVectors), 2962 NEONMAP1(vsqaddd_u64, aarch64_neon_usqadd, Add1ArgType), 2963 NEONMAP1(vsqaddh_u16, aarch64_neon_usqadd, Vectorize1ArgType | Use64BitVectors), 2964 NEONMAP1(vsqadds_u32, aarch64_neon_usqadd, Add1ArgType), 2965 NEONMAP1(vsrid_n_s64, aarch64_neon_vsri, Vectorize1ArgType), 2966 NEONMAP1(vsrid_n_u64, aarch64_neon_vsri, Vectorize1ArgType), 2967 NEONMAP1(vuqaddb_s8, aarch64_neon_suqadd, Vectorize1ArgType | Use64BitVectors), 2968 NEONMAP1(vuqaddd_s64, aarch64_neon_suqadd, Add1ArgType), 2969 NEONMAP1(vuqaddh_s16, aarch64_neon_suqadd, Vectorize1ArgType | Use64BitVectors), 2970 NEONMAP1(vuqadds_s32, aarch64_neon_suqadd, Add1ArgType), 2971 }; 2972 2973 #undef NEONMAP0 2974 #undef NEONMAP1 2975 #undef NEONMAP2 2976 2977 static bool NEONSIMDIntrinsicsProvenSorted = false; 2978 2979 static bool AArch64SIMDIntrinsicsProvenSorted = false; 2980 static bool AArch64SISDIntrinsicsProvenSorted = false; 2981 2982 2983 static const NeonIntrinsicInfo * 2984 findNeonIntrinsicInMap(ArrayRef<NeonIntrinsicInfo> IntrinsicMap, 2985 unsigned BuiltinID, bool &MapProvenSorted) { 2986 2987 #ifndef NDEBUG 2988 if (!MapProvenSorted) { 2989 assert(std::is_sorted(std::begin(IntrinsicMap), std::end(IntrinsicMap))); 2990 MapProvenSorted = true; 2991 } 2992 #endif 2993 2994 const NeonIntrinsicInfo *Builtin = 2995 std::lower_bound(IntrinsicMap.begin(), IntrinsicMap.end(), BuiltinID); 2996 2997 if (Builtin != IntrinsicMap.end() && Builtin->BuiltinID == BuiltinID) 2998 return Builtin; 2999 3000 return nullptr; 3001 } 3002 3003 Function *CodeGenFunction::LookupNeonLLVMIntrinsic(unsigned IntrinsicID, 3004 unsigned Modifier, 3005 llvm::Type *ArgType, 3006 const CallExpr *E) { 3007 int VectorSize = 0; 3008 if (Modifier & Use64BitVectors) 3009 VectorSize = 64; 3010 else if (Modifier & Use128BitVectors) 3011 VectorSize = 128; 3012 3013 // Return type. 3014 SmallVector<llvm::Type *, 3> Tys; 3015 if (Modifier & AddRetType) { 3016 llvm::Type *Ty = ConvertType(E->getCallReturnType(getContext())); 3017 if (Modifier & VectorizeRetType) 3018 Ty = llvm::VectorType::get( 3019 Ty, VectorSize ? VectorSize / Ty->getPrimitiveSizeInBits() : 1); 3020 3021 Tys.push_back(Ty); 3022 } 3023 3024 // Arguments. 3025 if (Modifier & VectorizeArgTypes) { 3026 int Elts = VectorSize ? VectorSize / ArgType->getPrimitiveSizeInBits() : 1; 3027 ArgType = llvm::VectorType::get(ArgType, Elts); 3028 } 3029 3030 if (Modifier & (Add1ArgType | Add2ArgTypes)) 3031 Tys.push_back(ArgType); 3032 3033 if (Modifier & Add2ArgTypes) 3034 Tys.push_back(ArgType); 3035 3036 if (Modifier & InventFloatType) 3037 Tys.push_back(FloatTy); 3038 3039 return CGM.getIntrinsic(IntrinsicID, Tys); 3040 } 3041 3042 static Value *EmitCommonNeonSISDBuiltinExpr(CodeGenFunction &CGF, 3043 const NeonIntrinsicInfo &SISDInfo, 3044 SmallVectorImpl<Value *> &Ops, 3045 const CallExpr *E) { 3046 unsigned BuiltinID = SISDInfo.BuiltinID; 3047 unsigned int Int = SISDInfo.LLVMIntrinsic; 3048 unsigned Modifier = SISDInfo.TypeModifier; 3049 const char *s = SISDInfo.NameHint; 3050 3051 switch (BuiltinID) { 3052 case NEON::BI__builtin_neon_vcled_s64: 3053 case NEON::BI__builtin_neon_vcled_u64: 3054 case NEON::BI__builtin_neon_vcles_f32: 3055 case NEON::BI__builtin_neon_vcled_f64: 3056 case NEON::BI__builtin_neon_vcltd_s64: 3057 case NEON::BI__builtin_neon_vcltd_u64: 3058 case NEON::BI__builtin_neon_vclts_f32: 3059 case NEON::BI__builtin_neon_vcltd_f64: 3060 case NEON::BI__builtin_neon_vcales_f32: 3061 case NEON::BI__builtin_neon_vcaled_f64: 3062 case NEON::BI__builtin_neon_vcalts_f32: 3063 case NEON::BI__builtin_neon_vcaltd_f64: 3064 // Only one direction of comparisons actually exist, cmle is actually a cmge 3065 // with swapped operands. The table gives us the right intrinsic but we 3066 // still need to do the swap. 3067 std::swap(Ops[0], Ops[1]); 3068 break; 3069 } 3070 3071 assert(Int && "Generic code assumes a valid intrinsic"); 3072 3073 // Determine the type(s) of this overloaded AArch64 intrinsic. 3074 const Expr *Arg = E->getArg(0); 3075 llvm::Type *ArgTy = CGF.ConvertType(Arg->getType()); 3076 Function *F = CGF.LookupNeonLLVMIntrinsic(Int, Modifier, ArgTy, E); 3077 3078 int j = 0; 3079 ConstantInt *C0 = ConstantInt::get(CGF.SizeTy, 0); 3080 for (Function::const_arg_iterator ai = F->arg_begin(), ae = F->arg_end(); 3081 ai != ae; ++ai, ++j) { 3082 llvm::Type *ArgTy = ai->getType(); 3083 if (Ops[j]->getType()->getPrimitiveSizeInBits() == 3084 ArgTy->getPrimitiveSizeInBits()) 3085 continue; 3086 3087 assert(ArgTy->isVectorTy() && !Ops[j]->getType()->isVectorTy()); 3088 // The constant argument to an _n_ intrinsic always has Int32Ty, so truncate 3089 // it before inserting. 3090 Ops[j] = 3091 CGF.Builder.CreateTruncOrBitCast(Ops[j], ArgTy->getVectorElementType()); 3092 Ops[j] = 3093 CGF.Builder.CreateInsertElement(UndefValue::get(ArgTy), Ops[j], C0); 3094 } 3095 3096 Value *Result = CGF.EmitNeonCall(F, Ops, s); 3097 llvm::Type *ResultType = CGF.ConvertType(E->getType()); 3098 if (ResultType->getPrimitiveSizeInBits() < 3099 Result->getType()->getPrimitiveSizeInBits()) 3100 return CGF.Builder.CreateExtractElement(Result, C0); 3101 3102 return CGF.Builder.CreateBitCast(Result, ResultType, s); 3103 } 3104 3105 Value *CodeGenFunction::EmitCommonNeonBuiltinExpr( 3106 unsigned BuiltinID, unsigned LLVMIntrinsic, unsigned AltLLVMIntrinsic, 3107 const char *NameHint, unsigned Modifier, const CallExpr *E, 3108 SmallVectorImpl<llvm::Value *> &Ops, Address PtrOp0, Address PtrOp1) { 3109 // Get the last argument, which specifies the vector type. 3110 llvm::APSInt NeonTypeConst; 3111 const Expr *Arg = E->getArg(E->getNumArgs() - 1); 3112 if (!Arg->isIntegerConstantExpr(NeonTypeConst, getContext())) 3113 return nullptr; 3114 3115 // Determine the type of this overloaded NEON intrinsic. 3116 NeonTypeFlags Type(NeonTypeConst.getZExtValue()); 3117 bool Usgn = Type.isUnsigned(); 3118 bool Quad = Type.isQuad(); 3119 3120 llvm::VectorType *VTy = GetNeonType(this, Type); 3121 llvm::Type *Ty = VTy; 3122 if (!Ty) 3123 return nullptr; 3124 3125 auto getAlignmentValue32 = [&](Address addr) -> Value* { 3126 return Builder.getInt32(addr.getAlignment().getQuantity()); 3127 }; 3128 3129 unsigned Int = LLVMIntrinsic; 3130 if ((Modifier & UnsignedAlts) && !Usgn) 3131 Int = AltLLVMIntrinsic; 3132 3133 switch (BuiltinID) { 3134 default: break; 3135 case NEON::BI__builtin_neon_vabs_v: 3136 case NEON::BI__builtin_neon_vabsq_v: 3137 if (VTy->getElementType()->isFloatingPointTy()) 3138 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::fabs, Ty), Ops, "vabs"); 3139 return EmitNeonCall(CGM.getIntrinsic(LLVMIntrinsic, Ty), Ops, "vabs"); 3140 case NEON::BI__builtin_neon_vaddhn_v: { 3141 llvm::VectorType *SrcTy = 3142 llvm::VectorType::getExtendedElementVectorType(VTy); 3143 3144 // %sum = add <4 x i32> %lhs, %rhs 3145 Ops[0] = Builder.CreateBitCast(Ops[0], SrcTy); 3146 Ops[1] = Builder.CreateBitCast(Ops[1], SrcTy); 3147 Ops[0] = Builder.CreateAdd(Ops[0], Ops[1], "vaddhn"); 3148 3149 // %high = lshr <4 x i32> %sum, <i32 16, i32 16, i32 16, i32 16> 3150 Constant *ShiftAmt = 3151 ConstantInt::get(SrcTy, SrcTy->getScalarSizeInBits() / 2); 3152 Ops[0] = Builder.CreateLShr(Ops[0], ShiftAmt, "vaddhn"); 3153 3154 // %res = trunc <4 x i32> %high to <4 x i16> 3155 return Builder.CreateTrunc(Ops[0], VTy, "vaddhn"); 3156 } 3157 case NEON::BI__builtin_neon_vcale_v: 3158 case NEON::BI__builtin_neon_vcaleq_v: 3159 case NEON::BI__builtin_neon_vcalt_v: 3160 case NEON::BI__builtin_neon_vcaltq_v: 3161 std::swap(Ops[0], Ops[1]); 3162 case NEON::BI__builtin_neon_vcage_v: 3163 case NEON::BI__builtin_neon_vcageq_v: 3164 case NEON::BI__builtin_neon_vcagt_v: 3165 case NEON::BI__builtin_neon_vcagtq_v: { 3166 llvm::Type *VecFlt = llvm::VectorType::get( 3167 VTy->getScalarSizeInBits() == 32 ? FloatTy : DoubleTy, 3168 VTy->getNumElements()); 3169 llvm::Type *Tys[] = { VTy, VecFlt }; 3170 Function *F = CGM.getIntrinsic(LLVMIntrinsic, Tys); 3171 return EmitNeonCall(F, Ops, NameHint); 3172 } 3173 case NEON::BI__builtin_neon_vclz_v: 3174 case NEON::BI__builtin_neon_vclzq_v: 3175 // We generate target-independent intrinsic, which needs a second argument 3176 // for whether or not clz of zero is undefined; on ARM it isn't. 3177 Ops.push_back(Builder.getInt1(getTarget().isCLZForZeroUndef())); 3178 break; 3179 case NEON::BI__builtin_neon_vcvt_f32_v: 3180 case NEON::BI__builtin_neon_vcvtq_f32_v: 3181 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 3182 Ty = GetNeonType(this, NeonTypeFlags(NeonTypeFlags::Float32, false, Quad)); 3183 return Usgn ? Builder.CreateUIToFP(Ops[0], Ty, "vcvt") 3184 : Builder.CreateSIToFP(Ops[0], Ty, "vcvt"); 3185 case NEON::BI__builtin_neon_vcvt_n_f32_v: 3186 case NEON::BI__builtin_neon_vcvt_n_f64_v: 3187 case NEON::BI__builtin_neon_vcvtq_n_f32_v: 3188 case NEON::BI__builtin_neon_vcvtq_n_f64_v: { 3189 llvm::Type *Tys[2] = { GetFloatNeonType(this, Type), Ty }; 3190 Int = Usgn ? LLVMIntrinsic : AltLLVMIntrinsic; 3191 Function *F = CGM.getIntrinsic(Int, Tys); 3192 return EmitNeonCall(F, Ops, "vcvt_n"); 3193 } 3194 case NEON::BI__builtin_neon_vcvt_n_s32_v: 3195 case NEON::BI__builtin_neon_vcvt_n_u32_v: 3196 case NEON::BI__builtin_neon_vcvt_n_s64_v: 3197 case NEON::BI__builtin_neon_vcvt_n_u64_v: 3198 case NEON::BI__builtin_neon_vcvtq_n_s32_v: 3199 case NEON::BI__builtin_neon_vcvtq_n_u32_v: 3200 case NEON::BI__builtin_neon_vcvtq_n_s64_v: 3201 case NEON::BI__builtin_neon_vcvtq_n_u64_v: { 3202 llvm::Type *Tys[2] = { Ty, GetFloatNeonType(this, Type) }; 3203 Function *F = CGM.getIntrinsic(LLVMIntrinsic, Tys); 3204 return EmitNeonCall(F, Ops, "vcvt_n"); 3205 } 3206 case NEON::BI__builtin_neon_vcvt_s32_v: 3207 case NEON::BI__builtin_neon_vcvt_u32_v: 3208 case NEON::BI__builtin_neon_vcvt_s64_v: 3209 case NEON::BI__builtin_neon_vcvt_u64_v: 3210 case NEON::BI__builtin_neon_vcvtq_s32_v: 3211 case NEON::BI__builtin_neon_vcvtq_u32_v: 3212 case NEON::BI__builtin_neon_vcvtq_s64_v: 3213 case NEON::BI__builtin_neon_vcvtq_u64_v: { 3214 Ops[0] = Builder.CreateBitCast(Ops[0], GetFloatNeonType(this, Type)); 3215 return Usgn ? Builder.CreateFPToUI(Ops[0], Ty, "vcvt") 3216 : Builder.CreateFPToSI(Ops[0], Ty, "vcvt"); 3217 } 3218 case NEON::BI__builtin_neon_vcvta_s32_v: 3219 case NEON::BI__builtin_neon_vcvta_s64_v: 3220 case NEON::BI__builtin_neon_vcvta_u32_v: 3221 case NEON::BI__builtin_neon_vcvta_u64_v: 3222 case NEON::BI__builtin_neon_vcvtaq_s32_v: 3223 case NEON::BI__builtin_neon_vcvtaq_s64_v: 3224 case NEON::BI__builtin_neon_vcvtaq_u32_v: 3225 case NEON::BI__builtin_neon_vcvtaq_u64_v: 3226 case NEON::BI__builtin_neon_vcvtn_s32_v: 3227 case NEON::BI__builtin_neon_vcvtn_s64_v: 3228 case NEON::BI__builtin_neon_vcvtn_u32_v: 3229 case NEON::BI__builtin_neon_vcvtn_u64_v: 3230 case NEON::BI__builtin_neon_vcvtnq_s32_v: 3231 case NEON::BI__builtin_neon_vcvtnq_s64_v: 3232 case NEON::BI__builtin_neon_vcvtnq_u32_v: 3233 case NEON::BI__builtin_neon_vcvtnq_u64_v: 3234 case NEON::BI__builtin_neon_vcvtp_s32_v: 3235 case NEON::BI__builtin_neon_vcvtp_s64_v: 3236 case NEON::BI__builtin_neon_vcvtp_u32_v: 3237 case NEON::BI__builtin_neon_vcvtp_u64_v: 3238 case NEON::BI__builtin_neon_vcvtpq_s32_v: 3239 case NEON::BI__builtin_neon_vcvtpq_s64_v: 3240 case NEON::BI__builtin_neon_vcvtpq_u32_v: 3241 case NEON::BI__builtin_neon_vcvtpq_u64_v: 3242 case NEON::BI__builtin_neon_vcvtm_s32_v: 3243 case NEON::BI__builtin_neon_vcvtm_s64_v: 3244 case NEON::BI__builtin_neon_vcvtm_u32_v: 3245 case NEON::BI__builtin_neon_vcvtm_u64_v: 3246 case NEON::BI__builtin_neon_vcvtmq_s32_v: 3247 case NEON::BI__builtin_neon_vcvtmq_s64_v: 3248 case NEON::BI__builtin_neon_vcvtmq_u32_v: 3249 case NEON::BI__builtin_neon_vcvtmq_u64_v: { 3250 llvm::Type *Tys[2] = { Ty, GetFloatNeonType(this, Type) }; 3251 return EmitNeonCall(CGM.getIntrinsic(LLVMIntrinsic, Tys), Ops, NameHint); 3252 } 3253 case NEON::BI__builtin_neon_vext_v: 3254 case NEON::BI__builtin_neon_vextq_v: { 3255 int CV = cast<ConstantInt>(Ops[2])->getSExtValue(); 3256 SmallVector<Constant*, 16> Indices; 3257 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) 3258 Indices.push_back(ConstantInt::get(Int32Ty, i+CV)); 3259 3260 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 3261 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 3262 Value *SV = llvm::ConstantVector::get(Indices); 3263 return Builder.CreateShuffleVector(Ops[0], Ops[1], SV, "vext"); 3264 } 3265 case NEON::BI__builtin_neon_vfma_v: 3266 case NEON::BI__builtin_neon_vfmaq_v: { 3267 Value *F = CGM.getIntrinsic(Intrinsic::fma, Ty); 3268 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 3269 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 3270 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 3271 3272 // NEON intrinsic puts accumulator first, unlike the LLVM fma. 3273 return Builder.CreateCall(F, {Ops[1], Ops[2], Ops[0]}); 3274 } 3275 case NEON::BI__builtin_neon_vld1_v: 3276 case NEON::BI__builtin_neon_vld1q_v: { 3277 llvm::Type *Tys[] = {Ty, Int8PtrTy}; 3278 Ops.push_back(getAlignmentValue32(PtrOp0)); 3279 return EmitNeonCall(CGM.getIntrinsic(LLVMIntrinsic, Tys), Ops, "vld1"); 3280 } 3281 case NEON::BI__builtin_neon_vld2_v: 3282 case NEON::BI__builtin_neon_vld2q_v: 3283 case NEON::BI__builtin_neon_vld3_v: 3284 case NEON::BI__builtin_neon_vld3q_v: 3285 case NEON::BI__builtin_neon_vld4_v: 3286 case NEON::BI__builtin_neon_vld4q_v: { 3287 llvm::Type *Tys[] = {Ty, Int8PtrTy}; 3288 Function *F = CGM.getIntrinsic(LLVMIntrinsic, Tys); 3289 Value *Align = getAlignmentValue32(PtrOp1); 3290 Ops[1] = Builder.CreateCall(F, {Ops[1], Align}, NameHint); 3291 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 3292 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 3293 return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]); 3294 } 3295 case NEON::BI__builtin_neon_vld1_dup_v: 3296 case NEON::BI__builtin_neon_vld1q_dup_v: { 3297 Value *V = UndefValue::get(Ty); 3298 Ty = llvm::PointerType::getUnqual(VTy->getElementType()); 3299 PtrOp0 = Builder.CreateBitCast(PtrOp0, Ty); 3300 LoadInst *Ld = Builder.CreateLoad(PtrOp0); 3301 llvm::Constant *CI = ConstantInt::get(SizeTy, 0); 3302 Ops[0] = Builder.CreateInsertElement(V, Ld, CI); 3303 return EmitNeonSplat(Ops[0], CI); 3304 } 3305 case NEON::BI__builtin_neon_vld2_lane_v: 3306 case NEON::BI__builtin_neon_vld2q_lane_v: 3307 case NEON::BI__builtin_neon_vld3_lane_v: 3308 case NEON::BI__builtin_neon_vld3q_lane_v: 3309 case NEON::BI__builtin_neon_vld4_lane_v: 3310 case NEON::BI__builtin_neon_vld4q_lane_v: { 3311 llvm::Type *Tys[] = {Ty, Int8PtrTy}; 3312 Function *F = CGM.getIntrinsic(LLVMIntrinsic, Tys); 3313 for (unsigned I = 2; I < Ops.size() - 1; ++I) 3314 Ops[I] = Builder.CreateBitCast(Ops[I], Ty); 3315 Ops.push_back(getAlignmentValue32(PtrOp1)); 3316 Ops[1] = Builder.CreateCall(F, makeArrayRef(Ops).slice(1), NameHint); 3317 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 3318 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 3319 return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]); 3320 } 3321 case NEON::BI__builtin_neon_vmovl_v: { 3322 llvm::Type *DTy =llvm::VectorType::getTruncatedElementVectorType(VTy); 3323 Ops[0] = Builder.CreateBitCast(Ops[0], DTy); 3324 if (Usgn) 3325 return Builder.CreateZExt(Ops[0], Ty, "vmovl"); 3326 return Builder.CreateSExt(Ops[0], Ty, "vmovl"); 3327 } 3328 case NEON::BI__builtin_neon_vmovn_v: { 3329 llvm::Type *QTy = llvm::VectorType::getExtendedElementVectorType(VTy); 3330 Ops[0] = Builder.CreateBitCast(Ops[0], QTy); 3331 return Builder.CreateTrunc(Ops[0], Ty, "vmovn"); 3332 } 3333 case NEON::BI__builtin_neon_vmull_v: 3334 // FIXME: the integer vmull operations could be emitted in terms of pure 3335 // LLVM IR (2 exts followed by a mul). Unfortunately LLVM has a habit of 3336 // hoisting the exts outside loops. Until global ISel comes along that can 3337 // see through such movement this leads to bad CodeGen. So we need an 3338 // intrinsic for now. 3339 Int = Usgn ? Intrinsic::arm_neon_vmullu : Intrinsic::arm_neon_vmulls; 3340 Int = Type.isPoly() ? (unsigned)Intrinsic::arm_neon_vmullp : Int; 3341 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmull"); 3342 case NEON::BI__builtin_neon_vpadal_v: 3343 case NEON::BI__builtin_neon_vpadalq_v: { 3344 // The source operand type has twice as many elements of half the size. 3345 unsigned EltBits = VTy->getElementType()->getPrimitiveSizeInBits(); 3346 llvm::Type *EltTy = 3347 llvm::IntegerType::get(getLLVMContext(), EltBits / 2); 3348 llvm::Type *NarrowTy = 3349 llvm::VectorType::get(EltTy, VTy->getNumElements() * 2); 3350 llvm::Type *Tys[2] = { Ty, NarrowTy }; 3351 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, NameHint); 3352 } 3353 case NEON::BI__builtin_neon_vpaddl_v: 3354 case NEON::BI__builtin_neon_vpaddlq_v: { 3355 // The source operand type has twice as many elements of half the size. 3356 unsigned EltBits = VTy->getElementType()->getPrimitiveSizeInBits(); 3357 llvm::Type *EltTy = llvm::IntegerType::get(getLLVMContext(), EltBits / 2); 3358 llvm::Type *NarrowTy = 3359 llvm::VectorType::get(EltTy, VTy->getNumElements() * 2); 3360 llvm::Type *Tys[2] = { Ty, NarrowTy }; 3361 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vpaddl"); 3362 } 3363 case NEON::BI__builtin_neon_vqdmlal_v: 3364 case NEON::BI__builtin_neon_vqdmlsl_v: { 3365 SmallVector<Value *, 2> MulOps(Ops.begin() + 1, Ops.end()); 3366 Ops[1] = 3367 EmitNeonCall(CGM.getIntrinsic(LLVMIntrinsic, Ty), MulOps, "vqdmlal"); 3368 Ops.resize(2); 3369 return EmitNeonCall(CGM.getIntrinsic(AltLLVMIntrinsic, Ty), Ops, NameHint); 3370 } 3371 case NEON::BI__builtin_neon_vqshl_n_v: 3372 case NEON::BI__builtin_neon_vqshlq_n_v: 3373 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqshl_n", 3374 1, false); 3375 case NEON::BI__builtin_neon_vqshlu_n_v: 3376 case NEON::BI__builtin_neon_vqshluq_n_v: 3377 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqshlu_n", 3378 1, false); 3379 case NEON::BI__builtin_neon_vrecpe_v: 3380 case NEON::BI__builtin_neon_vrecpeq_v: 3381 case NEON::BI__builtin_neon_vrsqrte_v: 3382 case NEON::BI__builtin_neon_vrsqrteq_v: 3383 Int = Ty->isFPOrFPVectorTy() ? LLVMIntrinsic : AltLLVMIntrinsic; 3384 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, NameHint); 3385 3386 case NEON::BI__builtin_neon_vrshr_n_v: 3387 case NEON::BI__builtin_neon_vrshrq_n_v: 3388 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrshr_n", 3389 1, true); 3390 case NEON::BI__builtin_neon_vshl_n_v: 3391 case NEON::BI__builtin_neon_vshlq_n_v: 3392 Ops[1] = EmitNeonShiftVector(Ops[1], Ty, false); 3393 return Builder.CreateShl(Builder.CreateBitCast(Ops[0],Ty), Ops[1], 3394 "vshl_n"); 3395 case NEON::BI__builtin_neon_vshll_n_v: { 3396 llvm::Type *SrcTy = llvm::VectorType::getTruncatedElementVectorType(VTy); 3397 Ops[0] = Builder.CreateBitCast(Ops[0], SrcTy); 3398 if (Usgn) 3399 Ops[0] = Builder.CreateZExt(Ops[0], VTy); 3400 else 3401 Ops[0] = Builder.CreateSExt(Ops[0], VTy); 3402 Ops[1] = EmitNeonShiftVector(Ops[1], VTy, false); 3403 return Builder.CreateShl(Ops[0], Ops[1], "vshll_n"); 3404 } 3405 case NEON::BI__builtin_neon_vshrn_n_v: { 3406 llvm::Type *SrcTy = llvm::VectorType::getExtendedElementVectorType(VTy); 3407 Ops[0] = Builder.CreateBitCast(Ops[0], SrcTy); 3408 Ops[1] = EmitNeonShiftVector(Ops[1], SrcTy, false); 3409 if (Usgn) 3410 Ops[0] = Builder.CreateLShr(Ops[0], Ops[1]); 3411 else 3412 Ops[0] = Builder.CreateAShr(Ops[0], Ops[1]); 3413 return Builder.CreateTrunc(Ops[0], Ty, "vshrn_n"); 3414 } 3415 case NEON::BI__builtin_neon_vshr_n_v: 3416 case NEON::BI__builtin_neon_vshrq_n_v: 3417 return EmitNeonRShiftImm(Ops[0], Ops[1], Ty, Usgn, "vshr_n"); 3418 case NEON::BI__builtin_neon_vst1_v: 3419 case NEON::BI__builtin_neon_vst1q_v: 3420 case NEON::BI__builtin_neon_vst2_v: 3421 case NEON::BI__builtin_neon_vst2q_v: 3422 case NEON::BI__builtin_neon_vst3_v: 3423 case NEON::BI__builtin_neon_vst3q_v: 3424 case NEON::BI__builtin_neon_vst4_v: 3425 case NEON::BI__builtin_neon_vst4q_v: 3426 case NEON::BI__builtin_neon_vst2_lane_v: 3427 case NEON::BI__builtin_neon_vst2q_lane_v: 3428 case NEON::BI__builtin_neon_vst3_lane_v: 3429 case NEON::BI__builtin_neon_vst3q_lane_v: 3430 case NEON::BI__builtin_neon_vst4_lane_v: 3431 case NEON::BI__builtin_neon_vst4q_lane_v: { 3432 llvm::Type *Tys[] = {Int8PtrTy, Ty}; 3433 Ops.push_back(getAlignmentValue32(PtrOp0)); 3434 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, ""); 3435 } 3436 case NEON::BI__builtin_neon_vsubhn_v: { 3437 llvm::VectorType *SrcTy = 3438 llvm::VectorType::getExtendedElementVectorType(VTy); 3439 3440 // %sum = add <4 x i32> %lhs, %rhs 3441 Ops[0] = Builder.CreateBitCast(Ops[0], SrcTy); 3442 Ops[1] = Builder.CreateBitCast(Ops[1], SrcTy); 3443 Ops[0] = Builder.CreateSub(Ops[0], Ops[1], "vsubhn"); 3444 3445 // %high = lshr <4 x i32> %sum, <i32 16, i32 16, i32 16, i32 16> 3446 Constant *ShiftAmt = 3447 ConstantInt::get(SrcTy, SrcTy->getScalarSizeInBits() / 2); 3448 Ops[0] = Builder.CreateLShr(Ops[0], ShiftAmt, "vsubhn"); 3449 3450 // %res = trunc <4 x i32> %high to <4 x i16> 3451 return Builder.CreateTrunc(Ops[0], VTy, "vsubhn"); 3452 } 3453 case NEON::BI__builtin_neon_vtrn_v: 3454 case NEON::BI__builtin_neon_vtrnq_v: { 3455 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty)); 3456 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 3457 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 3458 Value *SV = nullptr; 3459 3460 for (unsigned vi = 0; vi != 2; ++vi) { 3461 SmallVector<Constant*, 16> Indices; 3462 for (unsigned i = 0, e = VTy->getNumElements(); i != e; i += 2) { 3463 Indices.push_back(Builder.getInt32(i+vi)); 3464 Indices.push_back(Builder.getInt32(i+e+vi)); 3465 } 3466 Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ty, Ops[0], vi); 3467 SV = llvm::ConstantVector::get(Indices); 3468 SV = Builder.CreateShuffleVector(Ops[1], Ops[2], SV, "vtrn"); 3469 SV = Builder.CreateDefaultAlignedStore(SV, Addr); 3470 } 3471 return SV; 3472 } 3473 case NEON::BI__builtin_neon_vtst_v: 3474 case NEON::BI__builtin_neon_vtstq_v: { 3475 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 3476 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 3477 Ops[0] = Builder.CreateAnd(Ops[0], Ops[1]); 3478 Ops[0] = Builder.CreateICmp(ICmpInst::ICMP_NE, Ops[0], 3479 ConstantAggregateZero::get(Ty)); 3480 return Builder.CreateSExt(Ops[0], Ty, "vtst"); 3481 } 3482 case NEON::BI__builtin_neon_vuzp_v: 3483 case NEON::BI__builtin_neon_vuzpq_v: { 3484 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty)); 3485 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 3486 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 3487 Value *SV = nullptr; 3488 3489 for (unsigned vi = 0; vi != 2; ++vi) { 3490 SmallVector<Constant*, 16> Indices; 3491 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) 3492 Indices.push_back(ConstantInt::get(Int32Ty, 2*i+vi)); 3493 3494 Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ty, Ops[0], vi); 3495 SV = llvm::ConstantVector::get(Indices); 3496 SV = Builder.CreateShuffleVector(Ops[1], Ops[2], SV, "vuzp"); 3497 SV = Builder.CreateDefaultAlignedStore(SV, Addr); 3498 } 3499 return SV; 3500 } 3501 case NEON::BI__builtin_neon_vzip_v: 3502 case NEON::BI__builtin_neon_vzipq_v: { 3503 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty)); 3504 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 3505 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 3506 Value *SV = nullptr; 3507 3508 for (unsigned vi = 0; vi != 2; ++vi) { 3509 SmallVector<Constant*, 16> Indices; 3510 for (unsigned i = 0, e = VTy->getNumElements(); i != e; i += 2) { 3511 Indices.push_back(ConstantInt::get(Int32Ty, (i + vi*e) >> 1)); 3512 Indices.push_back(ConstantInt::get(Int32Ty, ((i + vi*e) >> 1)+e)); 3513 } 3514 Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ty, Ops[0], vi); 3515 SV = llvm::ConstantVector::get(Indices); 3516 SV = Builder.CreateShuffleVector(Ops[1], Ops[2], SV, "vzip"); 3517 SV = Builder.CreateDefaultAlignedStore(SV, Addr); 3518 } 3519 return SV; 3520 } 3521 } 3522 3523 assert(Int && "Expected valid intrinsic number"); 3524 3525 // Determine the type(s) of this overloaded AArch64 intrinsic. 3526 Function *F = LookupNeonLLVMIntrinsic(Int, Modifier, Ty, E); 3527 3528 Value *Result = EmitNeonCall(F, Ops, NameHint); 3529 llvm::Type *ResultType = ConvertType(E->getType()); 3530 // AArch64 intrinsic one-element vector type cast to 3531 // scalar type expected by the builtin 3532 return Builder.CreateBitCast(Result, ResultType, NameHint); 3533 } 3534 3535 Value *CodeGenFunction::EmitAArch64CompareBuiltinExpr( 3536 Value *Op, llvm::Type *Ty, const CmpInst::Predicate Fp, 3537 const CmpInst::Predicate Ip, const Twine &Name) { 3538 llvm::Type *OTy = Op->getType(); 3539 3540 // FIXME: this is utterly horrific. We should not be looking at previous 3541 // codegen context to find out what needs doing. Unfortunately TableGen 3542 // currently gives us exactly the same calls for vceqz_f32 and vceqz_s32 3543 // (etc). 3544 if (BitCastInst *BI = dyn_cast<BitCastInst>(Op)) 3545 OTy = BI->getOperand(0)->getType(); 3546 3547 Op = Builder.CreateBitCast(Op, OTy); 3548 if (OTy->getScalarType()->isFloatingPointTy()) { 3549 Op = Builder.CreateFCmp(Fp, Op, Constant::getNullValue(OTy)); 3550 } else { 3551 Op = Builder.CreateICmp(Ip, Op, Constant::getNullValue(OTy)); 3552 } 3553 return Builder.CreateSExt(Op, Ty, Name); 3554 } 3555 3556 static Value *packTBLDVectorList(CodeGenFunction &CGF, ArrayRef<Value *> Ops, 3557 Value *ExtOp, Value *IndexOp, 3558 llvm::Type *ResTy, unsigned IntID, 3559 const char *Name) { 3560 SmallVector<Value *, 2> TblOps; 3561 if (ExtOp) 3562 TblOps.push_back(ExtOp); 3563 3564 // Build a vector containing sequential number like (0, 1, 2, ..., 15) 3565 SmallVector<Constant*, 16> Indices; 3566 llvm::VectorType *TblTy = cast<llvm::VectorType>(Ops[0]->getType()); 3567 for (unsigned i = 0, e = TblTy->getNumElements(); i != e; ++i) { 3568 Indices.push_back(ConstantInt::get(CGF.Int32Ty, 2*i)); 3569 Indices.push_back(ConstantInt::get(CGF.Int32Ty, 2*i+1)); 3570 } 3571 Value *SV = llvm::ConstantVector::get(Indices); 3572 3573 int PairPos = 0, End = Ops.size() - 1; 3574 while (PairPos < End) { 3575 TblOps.push_back(CGF.Builder.CreateShuffleVector(Ops[PairPos], 3576 Ops[PairPos+1], SV, Name)); 3577 PairPos += 2; 3578 } 3579 3580 // If there's an odd number of 64-bit lookup table, fill the high 64-bit 3581 // of the 128-bit lookup table with zero. 3582 if (PairPos == End) { 3583 Value *ZeroTbl = ConstantAggregateZero::get(TblTy); 3584 TblOps.push_back(CGF.Builder.CreateShuffleVector(Ops[PairPos], 3585 ZeroTbl, SV, Name)); 3586 } 3587 3588 Function *TblF; 3589 TblOps.push_back(IndexOp); 3590 TblF = CGF.CGM.getIntrinsic(IntID, ResTy); 3591 3592 return CGF.EmitNeonCall(TblF, TblOps, Name); 3593 } 3594 3595 Value *CodeGenFunction::GetValueForARMHint(unsigned BuiltinID) { 3596 unsigned Value; 3597 switch (BuiltinID) { 3598 default: 3599 return nullptr; 3600 case ARM::BI__builtin_arm_nop: 3601 Value = 0; 3602 break; 3603 case ARM::BI__builtin_arm_yield: 3604 case ARM::BI__yield: 3605 Value = 1; 3606 break; 3607 case ARM::BI__builtin_arm_wfe: 3608 case ARM::BI__wfe: 3609 Value = 2; 3610 break; 3611 case ARM::BI__builtin_arm_wfi: 3612 case ARM::BI__wfi: 3613 Value = 3; 3614 break; 3615 case ARM::BI__builtin_arm_sev: 3616 case ARM::BI__sev: 3617 Value = 4; 3618 break; 3619 case ARM::BI__builtin_arm_sevl: 3620 case ARM::BI__sevl: 3621 Value = 5; 3622 break; 3623 } 3624 3625 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::arm_hint), 3626 llvm::ConstantInt::get(Int32Ty, Value)); 3627 } 3628 3629 // Generates the IR for the read/write special register builtin, 3630 // ValueType is the type of the value that is to be written or read, 3631 // RegisterType is the type of the register being written to or read from. 3632 static Value *EmitSpecialRegisterBuiltin(CodeGenFunction &CGF, 3633 const CallExpr *E, 3634 llvm::Type *RegisterType, 3635 llvm::Type *ValueType, bool IsRead) { 3636 // write and register intrinsics only support 32 and 64 bit operations. 3637 assert((RegisterType->isIntegerTy(32) || RegisterType->isIntegerTy(64)) 3638 && "Unsupported size for register."); 3639 3640 CodeGen::CGBuilderTy &Builder = CGF.Builder; 3641 CodeGen::CodeGenModule &CGM = CGF.CGM; 3642 LLVMContext &Context = CGM.getLLVMContext(); 3643 3644 const Expr *SysRegStrExpr = E->getArg(0)->IgnoreParenCasts(); 3645 StringRef SysReg = cast<StringLiteral>(SysRegStrExpr)->getString(); 3646 3647 llvm::Metadata *Ops[] = { llvm::MDString::get(Context, SysReg) }; 3648 llvm::MDNode *RegName = llvm::MDNode::get(Context, Ops); 3649 llvm::Value *Metadata = llvm::MetadataAsValue::get(Context, RegName); 3650 3651 llvm::Type *Types[] = { RegisterType }; 3652 3653 bool MixedTypes = RegisterType->isIntegerTy(64) && ValueType->isIntegerTy(32); 3654 assert(!(RegisterType->isIntegerTy(32) && ValueType->isIntegerTy(64)) 3655 && "Can't fit 64-bit value in 32-bit register"); 3656 3657 if (IsRead) { 3658 llvm::Value *F = CGM.getIntrinsic(llvm::Intrinsic::read_register, Types); 3659 llvm::Value *Call = Builder.CreateCall(F, Metadata); 3660 3661 if (MixedTypes) 3662 // Read into 64 bit register and then truncate result to 32 bit. 3663 return Builder.CreateTrunc(Call, ValueType); 3664 3665 if (ValueType->isPointerTy()) 3666 // Have i32/i64 result (Call) but want to return a VoidPtrTy (i8*). 3667 return Builder.CreateIntToPtr(Call, ValueType); 3668 3669 return Call; 3670 } 3671 3672 llvm::Value *F = CGM.getIntrinsic(llvm::Intrinsic::write_register, Types); 3673 llvm::Value *ArgValue = CGF.EmitScalarExpr(E->getArg(1)); 3674 if (MixedTypes) { 3675 // Extend 32 bit write value to 64 bit to pass to write. 3676 ArgValue = Builder.CreateZExt(ArgValue, RegisterType); 3677 return Builder.CreateCall(F, { Metadata, ArgValue }); 3678 } 3679 3680 if (ValueType->isPointerTy()) { 3681 // Have VoidPtrTy ArgValue but want to return an i32/i64. 3682 ArgValue = Builder.CreatePtrToInt(ArgValue, RegisterType); 3683 return Builder.CreateCall(F, { Metadata, ArgValue }); 3684 } 3685 3686 return Builder.CreateCall(F, { Metadata, ArgValue }); 3687 } 3688 3689 /// Return true if BuiltinID is an overloaded Neon intrinsic with an extra 3690 /// argument that specifies the vector type. 3691 static bool HasExtraNeonArgument(unsigned BuiltinID) { 3692 switch (BuiltinID) { 3693 default: break; 3694 case NEON::BI__builtin_neon_vget_lane_i8: 3695 case NEON::BI__builtin_neon_vget_lane_i16: 3696 case NEON::BI__builtin_neon_vget_lane_i32: 3697 case NEON::BI__builtin_neon_vget_lane_i64: 3698 case NEON::BI__builtin_neon_vget_lane_f32: 3699 case NEON::BI__builtin_neon_vgetq_lane_i8: 3700 case NEON::BI__builtin_neon_vgetq_lane_i16: 3701 case NEON::BI__builtin_neon_vgetq_lane_i32: 3702 case NEON::BI__builtin_neon_vgetq_lane_i64: 3703 case NEON::BI__builtin_neon_vgetq_lane_f32: 3704 case NEON::BI__builtin_neon_vset_lane_i8: 3705 case NEON::BI__builtin_neon_vset_lane_i16: 3706 case NEON::BI__builtin_neon_vset_lane_i32: 3707 case NEON::BI__builtin_neon_vset_lane_i64: 3708 case NEON::BI__builtin_neon_vset_lane_f32: 3709 case NEON::BI__builtin_neon_vsetq_lane_i8: 3710 case NEON::BI__builtin_neon_vsetq_lane_i16: 3711 case NEON::BI__builtin_neon_vsetq_lane_i32: 3712 case NEON::BI__builtin_neon_vsetq_lane_i64: 3713 case NEON::BI__builtin_neon_vsetq_lane_f32: 3714 case NEON::BI__builtin_neon_vsha1h_u32: 3715 case NEON::BI__builtin_neon_vsha1cq_u32: 3716 case NEON::BI__builtin_neon_vsha1pq_u32: 3717 case NEON::BI__builtin_neon_vsha1mq_u32: 3718 case ARM::BI_MoveToCoprocessor: 3719 case ARM::BI_MoveToCoprocessor2: 3720 return false; 3721 } 3722 return true; 3723 } 3724 3725 Value *CodeGenFunction::EmitARMBuiltinExpr(unsigned BuiltinID, 3726 const CallExpr *E) { 3727 if (auto Hint = GetValueForARMHint(BuiltinID)) 3728 return Hint; 3729 3730 if (BuiltinID == ARM::BI__emit) { 3731 bool IsThumb = getTarget().getTriple().getArch() == llvm::Triple::thumb; 3732 llvm::FunctionType *FTy = 3733 llvm::FunctionType::get(VoidTy, /*Variadic=*/false); 3734 3735 APSInt Value; 3736 if (!E->getArg(0)->EvaluateAsInt(Value, CGM.getContext())) 3737 llvm_unreachable("Sema will ensure that the parameter is constant"); 3738 3739 uint64_t ZExtValue = Value.zextOrTrunc(IsThumb ? 16 : 32).getZExtValue(); 3740 3741 llvm::InlineAsm *Emit = 3742 IsThumb ? InlineAsm::get(FTy, ".inst.n 0x" + utohexstr(ZExtValue), "", 3743 /*SideEffects=*/true) 3744 : InlineAsm::get(FTy, ".inst 0x" + utohexstr(ZExtValue), "", 3745 /*SideEffects=*/true); 3746 3747 return Builder.CreateCall(Emit); 3748 } 3749 3750 if (BuiltinID == ARM::BI__builtin_arm_dbg) { 3751 Value *Option = EmitScalarExpr(E->getArg(0)); 3752 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::arm_dbg), Option); 3753 } 3754 3755 if (BuiltinID == ARM::BI__builtin_arm_prefetch) { 3756 Value *Address = EmitScalarExpr(E->getArg(0)); 3757 Value *RW = EmitScalarExpr(E->getArg(1)); 3758 Value *IsData = EmitScalarExpr(E->getArg(2)); 3759 3760 // Locality is not supported on ARM target 3761 Value *Locality = llvm::ConstantInt::get(Int32Ty, 3); 3762 3763 Value *F = CGM.getIntrinsic(Intrinsic::prefetch); 3764 return Builder.CreateCall(F, {Address, RW, Locality, IsData}); 3765 } 3766 3767 if (BuiltinID == ARM::BI__builtin_arm_rbit) { 3768 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::arm_rbit), 3769 EmitScalarExpr(E->getArg(0)), 3770 "rbit"); 3771 } 3772 3773 if (BuiltinID == ARM::BI__clear_cache) { 3774 assert(E->getNumArgs() == 2 && "__clear_cache takes 2 arguments"); 3775 const FunctionDecl *FD = E->getDirectCallee(); 3776 Value *Ops[2]; 3777 for (unsigned i = 0; i < 2; i++) 3778 Ops[i] = EmitScalarExpr(E->getArg(i)); 3779 llvm::Type *Ty = CGM.getTypes().ConvertType(FD->getType()); 3780 llvm::FunctionType *FTy = cast<llvm::FunctionType>(Ty); 3781 StringRef Name = FD->getName(); 3782 return EmitNounwindRuntimeCall(CGM.CreateRuntimeFunction(FTy, Name), Ops); 3783 } 3784 3785 if (BuiltinID == ARM::BI__builtin_arm_ldrexd || 3786 ((BuiltinID == ARM::BI__builtin_arm_ldrex || 3787 BuiltinID == ARM::BI__builtin_arm_ldaex) && 3788 getContext().getTypeSize(E->getType()) == 64) || 3789 BuiltinID == ARM::BI__ldrexd) { 3790 Function *F; 3791 3792 switch (BuiltinID) { 3793 default: llvm_unreachable("unexpected builtin"); 3794 case ARM::BI__builtin_arm_ldaex: 3795 F = CGM.getIntrinsic(Intrinsic::arm_ldaexd); 3796 break; 3797 case ARM::BI__builtin_arm_ldrexd: 3798 case ARM::BI__builtin_arm_ldrex: 3799 case ARM::BI__ldrexd: 3800 F = CGM.getIntrinsic(Intrinsic::arm_ldrexd); 3801 break; 3802 } 3803 3804 Value *LdPtr = EmitScalarExpr(E->getArg(0)); 3805 Value *Val = Builder.CreateCall(F, Builder.CreateBitCast(LdPtr, Int8PtrTy), 3806 "ldrexd"); 3807 3808 Value *Val0 = Builder.CreateExtractValue(Val, 1); 3809 Value *Val1 = Builder.CreateExtractValue(Val, 0); 3810 Val0 = Builder.CreateZExt(Val0, Int64Ty); 3811 Val1 = Builder.CreateZExt(Val1, Int64Ty); 3812 3813 Value *ShiftCst = llvm::ConstantInt::get(Int64Ty, 32); 3814 Val = Builder.CreateShl(Val0, ShiftCst, "shl", true /* nuw */); 3815 Val = Builder.CreateOr(Val, Val1); 3816 return Builder.CreateBitCast(Val, ConvertType(E->getType())); 3817 } 3818 3819 if (BuiltinID == ARM::BI__builtin_arm_ldrex || 3820 BuiltinID == ARM::BI__builtin_arm_ldaex) { 3821 Value *LoadAddr = EmitScalarExpr(E->getArg(0)); 3822 3823 QualType Ty = E->getType(); 3824 llvm::Type *RealResTy = ConvertType(Ty); 3825 llvm::Type *IntResTy = llvm::IntegerType::get(getLLVMContext(), 3826 getContext().getTypeSize(Ty)); 3827 LoadAddr = Builder.CreateBitCast(LoadAddr, IntResTy->getPointerTo()); 3828 3829 Function *F = CGM.getIntrinsic(BuiltinID == ARM::BI__builtin_arm_ldaex 3830 ? Intrinsic::arm_ldaex 3831 : Intrinsic::arm_ldrex, 3832 LoadAddr->getType()); 3833 Value *Val = Builder.CreateCall(F, LoadAddr, "ldrex"); 3834 3835 if (RealResTy->isPointerTy()) 3836 return Builder.CreateIntToPtr(Val, RealResTy); 3837 else { 3838 Val = Builder.CreateTruncOrBitCast(Val, IntResTy); 3839 return Builder.CreateBitCast(Val, RealResTy); 3840 } 3841 } 3842 3843 if (BuiltinID == ARM::BI__builtin_arm_strexd || 3844 ((BuiltinID == ARM::BI__builtin_arm_stlex || 3845 BuiltinID == ARM::BI__builtin_arm_strex) && 3846 getContext().getTypeSize(E->getArg(0)->getType()) == 64)) { 3847 Function *F = CGM.getIntrinsic(BuiltinID == ARM::BI__builtin_arm_stlex 3848 ? Intrinsic::arm_stlexd 3849 : Intrinsic::arm_strexd); 3850 llvm::Type *STy = llvm::StructType::get(Int32Ty, Int32Ty, nullptr); 3851 3852 Address Tmp = CreateMemTemp(E->getArg(0)->getType()); 3853 Value *Val = EmitScalarExpr(E->getArg(0)); 3854 Builder.CreateStore(Val, Tmp); 3855 3856 Address LdPtr = Builder.CreateBitCast(Tmp,llvm::PointerType::getUnqual(STy)); 3857 Val = Builder.CreateLoad(LdPtr); 3858 3859 Value *Arg0 = Builder.CreateExtractValue(Val, 0); 3860 Value *Arg1 = Builder.CreateExtractValue(Val, 1); 3861 Value *StPtr = Builder.CreateBitCast(EmitScalarExpr(E->getArg(1)), Int8PtrTy); 3862 return Builder.CreateCall(F, {Arg0, Arg1, StPtr}, "strexd"); 3863 } 3864 3865 if (BuiltinID == ARM::BI__builtin_arm_strex || 3866 BuiltinID == ARM::BI__builtin_arm_stlex) { 3867 Value *StoreVal = EmitScalarExpr(E->getArg(0)); 3868 Value *StoreAddr = EmitScalarExpr(E->getArg(1)); 3869 3870 QualType Ty = E->getArg(0)->getType(); 3871 llvm::Type *StoreTy = llvm::IntegerType::get(getLLVMContext(), 3872 getContext().getTypeSize(Ty)); 3873 StoreAddr = Builder.CreateBitCast(StoreAddr, StoreTy->getPointerTo()); 3874 3875 if (StoreVal->getType()->isPointerTy()) 3876 StoreVal = Builder.CreatePtrToInt(StoreVal, Int32Ty); 3877 else { 3878 StoreVal = Builder.CreateBitCast(StoreVal, StoreTy); 3879 StoreVal = Builder.CreateZExtOrBitCast(StoreVal, Int32Ty); 3880 } 3881 3882 Function *F = CGM.getIntrinsic(BuiltinID == ARM::BI__builtin_arm_stlex 3883 ? Intrinsic::arm_stlex 3884 : Intrinsic::arm_strex, 3885 StoreAddr->getType()); 3886 return Builder.CreateCall(F, {StoreVal, StoreAddr}, "strex"); 3887 } 3888 3889 if (BuiltinID == ARM::BI__builtin_arm_clrex) { 3890 Function *F = CGM.getIntrinsic(Intrinsic::arm_clrex); 3891 return Builder.CreateCall(F); 3892 } 3893 3894 // CRC32 3895 Intrinsic::ID CRCIntrinsicID = Intrinsic::not_intrinsic; 3896 switch (BuiltinID) { 3897 case ARM::BI__builtin_arm_crc32b: 3898 CRCIntrinsicID = Intrinsic::arm_crc32b; break; 3899 case ARM::BI__builtin_arm_crc32cb: 3900 CRCIntrinsicID = Intrinsic::arm_crc32cb; break; 3901 case ARM::BI__builtin_arm_crc32h: 3902 CRCIntrinsicID = Intrinsic::arm_crc32h; break; 3903 case ARM::BI__builtin_arm_crc32ch: 3904 CRCIntrinsicID = Intrinsic::arm_crc32ch; break; 3905 case ARM::BI__builtin_arm_crc32w: 3906 case ARM::BI__builtin_arm_crc32d: 3907 CRCIntrinsicID = Intrinsic::arm_crc32w; break; 3908 case ARM::BI__builtin_arm_crc32cw: 3909 case ARM::BI__builtin_arm_crc32cd: 3910 CRCIntrinsicID = Intrinsic::arm_crc32cw; break; 3911 } 3912 3913 if (CRCIntrinsicID != Intrinsic::not_intrinsic) { 3914 Value *Arg0 = EmitScalarExpr(E->getArg(0)); 3915 Value *Arg1 = EmitScalarExpr(E->getArg(1)); 3916 3917 // crc32{c,}d intrinsics are implemnted as two calls to crc32{c,}w 3918 // intrinsics, hence we need different codegen for these cases. 3919 if (BuiltinID == ARM::BI__builtin_arm_crc32d || 3920 BuiltinID == ARM::BI__builtin_arm_crc32cd) { 3921 Value *C1 = llvm::ConstantInt::get(Int64Ty, 32); 3922 Value *Arg1a = Builder.CreateTruncOrBitCast(Arg1, Int32Ty); 3923 Value *Arg1b = Builder.CreateLShr(Arg1, C1); 3924 Arg1b = Builder.CreateTruncOrBitCast(Arg1b, Int32Ty); 3925 3926 Function *F = CGM.getIntrinsic(CRCIntrinsicID); 3927 Value *Res = Builder.CreateCall(F, {Arg0, Arg1a}); 3928 return Builder.CreateCall(F, {Res, Arg1b}); 3929 } else { 3930 Arg1 = Builder.CreateZExtOrBitCast(Arg1, Int32Ty); 3931 3932 Function *F = CGM.getIntrinsic(CRCIntrinsicID); 3933 return Builder.CreateCall(F, {Arg0, Arg1}); 3934 } 3935 } 3936 3937 if (BuiltinID == ARM::BI__builtin_arm_rsr || 3938 BuiltinID == ARM::BI__builtin_arm_rsr64 || 3939 BuiltinID == ARM::BI__builtin_arm_rsrp || 3940 BuiltinID == ARM::BI__builtin_arm_wsr || 3941 BuiltinID == ARM::BI__builtin_arm_wsr64 || 3942 BuiltinID == ARM::BI__builtin_arm_wsrp) { 3943 3944 bool IsRead = BuiltinID == ARM::BI__builtin_arm_rsr || 3945 BuiltinID == ARM::BI__builtin_arm_rsr64 || 3946 BuiltinID == ARM::BI__builtin_arm_rsrp; 3947 3948 bool IsPointerBuiltin = BuiltinID == ARM::BI__builtin_arm_rsrp || 3949 BuiltinID == ARM::BI__builtin_arm_wsrp; 3950 3951 bool Is64Bit = BuiltinID == ARM::BI__builtin_arm_rsr64 || 3952 BuiltinID == ARM::BI__builtin_arm_wsr64; 3953 3954 llvm::Type *ValueType; 3955 llvm::Type *RegisterType; 3956 if (IsPointerBuiltin) { 3957 ValueType = VoidPtrTy; 3958 RegisterType = Int32Ty; 3959 } else if (Is64Bit) { 3960 ValueType = RegisterType = Int64Ty; 3961 } else { 3962 ValueType = RegisterType = Int32Ty; 3963 } 3964 3965 return EmitSpecialRegisterBuiltin(*this, E, RegisterType, ValueType, IsRead); 3966 } 3967 3968 // Find out if any arguments are required to be integer constant 3969 // expressions. 3970 unsigned ICEArguments = 0; 3971 ASTContext::GetBuiltinTypeError Error; 3972 getContext().GetBuiltinType(BuiltinID, Error, &ICEArguments); 3973 assert(Error == ASTContext::GE_None && "Should not codegen an error"); 3974 3975 auto getAlignmentValue32 = [&](Address addr) -> Value* { 3976 return Builder.getInt32(addr.getAlignment().getQuantity()); 3977 }; 3978 3979 Address PtrOp0 = Address::invalid(); 3980 Address PtrOp1 = Address::invalid(); 3981 SmallVector<Value*, 4> Ops; 3982 bool HasExtraArg = HasExtraNeonArgument(BuiltinID); 3983 unsigned NumArgs = E->getNumArgs() - (HasExtraArg ? 1 : 0); 3984 for (unsigned i = 0, e = NumArgs; i != e; i++) { 3985 if (i == 0) { 3986 switch (BuiltinID) { 3987 case NEON::BI__builtin_neon_vld1_v: 3988 case NEON::BI__builtin_neon_vld1q_v: 3989 case NEON::BI__builtin_neon_vld1q_lane_v: 3990 case NEON::BI__builtin_neon_vld1_lane_v: 3991 case NEON::BI__builtin_neon_vld1_dup_v: 3992 case NEON::BI__builtin_neon_vld1q_dup_v: 3993 case NEON::BI__builtin_neon_vst1_v: 3994 case NEON::BI__builtin_neon_vst1q_v: 3995 case NEON::BI__builtin_neon_vst1q_lane_v: 3996 case NEON::BI__builtin_neon_vst1_lane_v: 3997 case NEON::BI__builtin_neon_vst2_v: 3998 case NEON::BI__builtin_neon_vst2q_v: 3999 case NEON::BI__builtin_neon_vst2_lane_v: 4000 case NEON::BI__builtin_neon_vst2q_lane_v: 4001 case NEON::BI__builtin_neon_vst3_v: 4002 case NEON::BI__builtin_neon_vst3q_v: 4003 case NEON::BI__builtin_neon_vst3_lane_v: 4004 case NEON::BI__builtin_neon_vst3q_lane_v: 4005 case NEON::BI__builtin_neon_vst4_v: 4006 case NEON::BI__builtin_neon_vst4q_v: 4007 case NEON::BI__builtin_neon_vst4_lane_v: 4008 case NEON::BI__builtin_neon_vst4q_lane_v: 4009 // Get the alignment for the argument in addition to the value; 4010 // we'll use it later. 4011 PtrOp0 = EmitPointerWithAlignment(E->getArg(0)); 4012 Ops.push_back(PtrOp0.getPointer()); 4013 continue; 4014 } 4015 } 4016 if (i == 1) { 4017 switch (BuiltinID) { 4018 case NEON::BI__builtin_neon_vld2_v: 4019 case NEON::BI__builtin_neon_vld2q_v: 4020 case NEON::BI__builtin_neon_vld3_v: 4021 case NEON::BI__builtin_neon_vld3q_v: 4022 case NEON::BI__builtin_neon_vld4_v: 4023 case NEON::BI__builtin_neon_vld4q_v: 4024 case NEON::BI__builtin_neon_vld2_lane_v: 4025 case NEON::BI__builtin_neon_vld2q_lane_v: 4026 case NEON::BI__builtin_neon_vld3_lane_v: 4027 case NEON::BI__builtin_neon_vld3q_lane_v: 4028 case NEON::BI__builtin_neon_vld4_lane_v: 4029 case NEON::BI__builtin_neon_vld4q_lane_v: 4030 case NEON::BI__builtin_neon_vld2_dup_v: 4031 case NEON::BI__builtin_neon_vld3_dup_v: 4032 case NEON::BI__builtin_neon_vld4_dup_v: 4033 // Get the alignment for the argument in addition to the value; 4034 // we'll use it later. 4035 PtrOp1 = EmitPointerWithAlignment(E->getArg(1)); 4036 Ops.push_back(PtrOp1.getPointer()); 4037 continue; 4038 } 4039 } 4040 4041 if ((ICEArguments & (1 << i)) == 0) { 4042 Ops.push_back(EmitScalarExpr(E->getArg(i))); 4043 } else { 4044 // If this is required to be a constant, constant fold it so that we know 4045 // that the generated intrinsic gets a ConstantInt. 4046 llvm::APSInt Result; 4047 bool IsConst = E->getArg(i)->isIntegerConstantExpr(Result, getContext()); 4048 assert(IsConst && "Constant arg isn't actually constant?"); (void)IsConst; 4049 Ops.push_back(llvm::ConstantInt::get(getLLVMContext(), Result)); 4050 } 4051 } 4052 4053 switch (BuiltinID) { 4054 default: break; 4055 4056 case NEON::BI__builtin_neon_vget_lane_i8: 4057 case NEON::BI__builtin_neon_vget_lane_i16: 4058 case NEON::BI__builtin_neon_vget_lane_i32: 4059 case NEON::BI__builtin_neon_vget_lane_i64: 4060 case NEON::BI__builtin_neon_vget_lane_f32: 4061 case NEON::BI__builtin_neon_vgetq_lane_i8: 4062 case NEON::BI__builtin_neon_vgetq_lane_i16: 4063 case NEON::BI__builtin_neon_vgetq_lane_i32: 4064 case NEON::BI__builtin_neon_vgetq_lane_i64: 4065 case NEON::BI__builtin_neon_vgetq_lane_f32: 4066 return Builder.CreateExtractElement(Ops[0], Ops[1], "vget_lane"); 4067 4068 case NEON::BI__builtin_neon_vset_lane_i8: 4069 case NEON::BI__builtin_neon_vset_lane_i16: 4070 case NEON::BI__builtin_neon_vset_lane_i32: 4071 case NEON::BI__builtin_neon_vset_lane_i64: 4072 case NEON::BI__builtin_neon_vset_lane_f32: 4073 case NEON::BI__builtin_neon_vsetq_lane_i8: 4074 case NEON::BI__builtin_neon_vsetq_lane_i16: 4075 case NEON::BI__builtin_neon_vsetq_lane_i32: 4076 case NEON::BI__builtin_neon_vsetq_lane_i64: 4077 case NEON::BI__builtin_neon_vsetq_lane_f32: 4078 return Builder.CreateInsertElement(Ops[1], Ops[0], Ops[2], "vset_lane"); 4079 4080 case NEON::BI__builtin_neon_vsha1h_u32: 4081 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_sha1h), Ops, 4082 "vsha1h"); 4083 case NEON::BI__builtin_neon_vsha1cq_u32: 4084 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_sha1c), Ops, 4085 "vsha1h"); 4086 case NEON::BI__builtin_neon_vsha1pq_u32: 4087 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_sha1p), Ops, 4088 "vsha1h"); 4089 case NEON::BI__builtin_neon_vsha1mq_u32: 4090 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_sha1m), Ops, 4091 "vsha1h"); 4092 4093 // The ARM _MoveToCoprocessor builtins put the input register value as 4094 // the first argument, but the LLVM intrinsic expects it as the third one. 4095 case ARM::BI_MoveToCoprocessor: 4096 case ARM::BI_MoveToCoprocessor2: { 4097 Function *F = CGM.getIntrinsic(BuiltinID == ARM::BI_MoveToCoprocessor ? 4098 Intrinsic::arm_mcr : Intrinsic::arm_mcr2); 4099 return Builder.CreateCall(F, {Ops[1], Ops[2], Ops[0], 4100 Ops[3], Ops[4], Ops[5]}); 4101 } 4102 } 4103 4104 // Get the last argument, which specifies the vector type. 4105 assert(HasExtraArg); 4106 llvm::APSInt Result; 4107 const Expr *Arg = E->getArg(E->getNumArgs()-1); 4108 if (!Arg->isIntegerConstantExpr(Result, getContext())) 4109 return nullptr; 4110 4111 if (BuiltinID == ARM::BI__builtin_arm_vcvtr_f || 4112 BuiltinID == ARM::BI__builtin_arm_vcvtr_d) { 4113 // Determine the overloaded type of this builtin. 4114 llvm::Type *Ty; 4115 if (BuiltinID == ARM::BI__builtin_arm_vcvtr_f) 4116 Ty = FloatTy; 4117 else 4118 Ty = DoubleTy; 4119 4120 // Determine whether this is an unsigned conversion or not. 4121 bool usgn = Result.getZExtValue() == 1; 4122 unsigned Int = usgn ? Intrinsic::arm_vcvtru : Intrinsic::arm_vcvtr; 4123 4124 // Call the appropriate intrinsic. 4125 Function *F = CGM.getIntrinsic(Int, Ty); 4126 return Builder.CreateCall(F, Ops, "vcvtr"); 4127 } 4128 4129 // Determine the type of this overloaded NEON intrinsic. 4130 NeonTypeFlags Type(Result.getZExtValue()); 4131 bool usgn = Type.isUnsigned(); 4132 bool rightShift = false; 4133 4134 llvm::VectorType *VTy = GetNeonType(this, Type); 4135 llvm::Type *Ty = VTy; 4136 if (!Ty) 4137 return nullptr; 4138 4139 // Many NEON builtins have identical semantics and uses in ARM and 4140 // AArch64. Emit these in a single function. 4141 auto IntrinsicMap = makeArrayRef(ARMSIMDIntrinsicMap); 4142 const NeonIntrinsicInfo *Builtin = findNeonIntrinsicInMap( 4143 IntrinsicMap, BuiltinID, NEONSIMDIntrinsicsProvenSorted); 4144 if (Builtin) 4145 return EmitCommonNeonBuiltinExpr( 4146 Builtin->BuiltinID, Builtin->LLVMIntrinsic, Builtin->AltLLVMIntrinsic, 4147 Builtin->NameHint, Builtin->TypeModifier, E, Ops, PtrOp0, PtrOp1); 4148 4149 unsigned Int; 4150 switch (BuiltinID) { 4151 default: return nullptr; 4152 case NEON::BI__builtin_neon_vld1q_lane_v: 4153 // Handle 64-bit integer elements as a special case. Use shuffles of 4154 // one-element vectors to avoid poor code for i64 in the backend. 4155 if (VTy->getElementType()->isIntegerTy(64)) { 4156 // Extract the other lane. 4157 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 4158 uint32_t Lane = cast<ConstantInt>(Ops[2])->getZExtValue(); 4159 Value *SV = llvm::ConstantVector::get(ConstantInt::get(Int32Ty, 1-Lane)); 4160 Ops[1] = Builder.CreateShuffleVector(Ops[1], Ops[1], SV); 4161 // Load the value as a one-element vector. 4162 Ty = llvm::VectorType::get(VTy->getElementType(), 1); 4163 llvm::Type *Tys[] = {Ty, Int8PtrTy}; 4164 Function *F = CGM.getIntrinsic(Intrinsic::arm_neon_vld1, Tys); 4165 Value *Align = getAlignmentValue32(PtrOp0); 4166 Value *Ld = Builder.CreateCall(F, {Ops[0], Align}); 4167 // Combine them. 4168 uint32_t Indices[] = {1 - Lane, Lane}; 4169 SV = llvm::ConstantDataVector::get(getLLVMContext(), Indices); 4170 return Builder.CreateShuffleVector(Ops[1], Ld, SV, "vld1q_lane"); 4171 } 4172 // fall through 4173 case NEON::BI__builtin_neon_vld1_lane_v: { 4174 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 4175 PtrOp0 = Builder.CreateElementBitCast(PtrOp0, VTy->getElementType()); 4176 Value *Ld = Builder.CreateLoad(PtrOp0); 4177 return Builder.CreateInsertElement(Ops[1], Ld, Ops[2], "vld1_lane"); 4178 } 4179 case NEON::BI__builtin_neon_vld2_dup_v: 4180 case NEON::BI__builtin_neon_vld3_dup_v: 4181 case NEON::BI__builtin_neon_vld4_dup_v: { 4182 // Handle 64-bit elements as a special-case. There is no "dup" needed. 4183 if (VTy->getElementType()->getPrimitiveSizeInBits() == 64) { 4184 switch (BuiltinID) { 4185 case NEON::BI__builtin_neon_vld2_dup_v: 4186 Int = Intrinsic::arm_neon_vld2; 4187 break; 4188 case NEON::BI__builtin_neon_vld3_dup_v: 4189 Int = Intrinsic::arm_neon_vld3; 4190 break; 4191 case NEON::BI__builtin_neon_vld4_dup_v: 4192 Int = Intrinsic::arm_neon_vld4; 4193 break; 4194 default: llvm_unreachable("unknown vld_dup intrinsic?"); 4195 } 4196 llvm::Type *Tys[] = {Ty, Int8PtrTy}; 4197 Function *F = CGM.getIntrinsic(Int, Tys); 4198 llvm::Value *Align = getAlignmentValue32(PtrOp1); 4199 Ops[1] = Builder.CreateCall(F, {Ops[1], Align}, "vld_dup"); 4200 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 4201 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 4202 return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]); 4203 } 4204 switch (BuiltinID) { 4205 case NEON::BI__builtin_neon_vld2_dup_v: 4206 Int = Intrinsic::arm_neon_vld2lane; 4207 break; 4208 case NEON::BI__builtin_neon_vld3_dup_v: 4209 Int = Intrinsic::arm_neon_vld3lane; 4210 break; 4211 case NEON::BI__builtin_neon_vld4_dup_v: 4212 Int = Intrinsic::arm_neon_vld4lane; 4213 break; 4214 default: llvm_unreachable("unknown vld_dup intrinsic?"); 4215 } 4216 llvm::Type *Tys[] = {Ty, Int8PtrTy}; 4217 Function *F = CGM.getIntrinsic(Int, Tys); 4218 llvm::StructType *STy = cast<llvm::StructType>(F->getReturnType()); 4219 4220 SmallVector<Value*, 6> Args; 4221 Args.push_back(Ops[1]); 4222 Args.append(STy->getNumElements(), UndefValue::get(Ty)); 4223 4224 llvm::Constant *CI = ConstantInt::get(Int32Ty, 0); 4225 Args.push_back(CI); 4226 Args.push_back(getAlignmentValue32(PtrOp1)); 4227 4228 Ops[1] = Builder.CreateCall(F, Args, "vld_dup"); 4229 // splat lane 0 to all elts in each vector of the result. 4230 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 4231 Value *Val = Builder.CreateExtractValue(Ops[1], i); 4232 Value *Elt = Builder.CreateBitCast(Val, Ty); 4233 Elt = EmitNeonSplat(Elt, CI); 4234 Elt = Builder.CreateBitCast(Elt, Val->getType()); 4235 Ops[1] = Builder.CreateInsertValue(Ops[1], Elt, i); 4236 } 4237 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 4238 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 4239 return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]); 4240 } 4241 case NEON::BI__builtin_neon_vqrshrn_n_v: 4242 Int = 4243 usgn ? Intrinsic::arm_neon_vqrshiftnu : Intrinsic::arm_neon_vqrshiftns; 4244 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqrshrn_n", 4245 1, true); 4246 case NEON::BI__builtin_neon_vqrshrun_n_v: 4247 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vqrshiftnsu, Ty), 4248 Ops, "vqrshrun_n", 1, true); 4249 case NEON::BI__builtin_neon_vqshrn_n_v: 4250 Int = usgn ? Intrinsic::arm_neon_vqshiftnu : Intrinsic::arm_neon_vqshiftns; 4251 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqshrn_n", 4252 1, true); 4253 case NEON::BI__builtin_neon_vqshrun_n_v: 4254 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vqshiftnsu, Ty), 4255 Ops, "vqshrun_n", 1, true); 4256 case NEON::BI__builtin_neon_vrecpe_v: 4257 case NEON::BI__builtin_neon_vrecpeq_v: 4258 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vrecpe, Ty), 4259 Ops, "vrecpe"); 4260 case NEON::BI__builtin_neon_vrshrn_n_v: 4261 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vrshiftn, Ty), 4262 Ops, "vrshrn_n", 1, true); 4263 case NEON::BI__builtin_neon_vrsra_n_v: 4264 case NEON::BI__builtin_neon_vrsraq_n_v: 4265 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 4266 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 4267 Ops[2] = EmitNeonShiftVector(Ops[2], Ty, true); 4268 Int = usgn ? Intrinsic::arm_neon_vrshiftu : Intrinsic::arm_neon_vrshifts; 4269 Ops[1] = Builder.CreateCall(CGM.getIntrinsic(Int, Ty), {Ops[1], Ops[2]}); 4270 return Builder.CreateAdd(Ops[0], Ops[1], "vrsra_n"); 4271 case NEON::BI__builtin_neon_vsri_n_v: 4272 case NEON::BI__builtin_neon_vsriq_n_v: 4273 rightShift = true; 4274 case NEON::BI__builtin_neon_vsli_n_v: 4275 case NEON::BI__builtin_neon_vsliq_n_v: 4276 Ops[2] = EmitNeonShiftVector(Ops[2], Ty, rightShift); 4277 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vshiftins, Ty), 4278 Ops, "vsli_n"); 4279 case NEON::BI__builtin_neon_vsra_n_v: 4280 case NEON::BI__builtin_neon_vsraq_n_v: 4281 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 4282 Ops[1] = EmitNeonRShiftImm(Ops[1], Ops[2], Ty, usgn, "vsra_n"); 4283 return Builder.CreateAdd(Ops[0], Ops[1]); 4284 case NEON::BI__builtin_neon_vst1q_lane_v: 4285 // Handle 64-bit integer elements as a special case. Use a shuffle to get 4286 // a one-element vector and avoid poor code for i64 in the backend. 4287 if (VTy->getElementType()->isIntegerTy(64)) { 4288 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 4289 Value *SV = llvm::ConstantVector::get(cast<llvm::Constant>(Ops[2])); 4290 Ops[1] = Builder.CreateShuffleVector(Ops[1], Ops[1], SV); 4291 Ops[2] = getAlignmentValue32(PtrOp0); 4292 llvm::Type *Tys[] = {Int8PtrTy, Ops[1]->getType()}; 4293 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::arm_neon_vst1, 4294 Tys), Ops); 4295 } 4296 // fall through 4297 case NEON::BI__builtin_neon_vst1_lane_v: { 4298 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 4299 Ops[1] = Builder.CreateExtractElement(Ops[1], Ops[2]); 4300 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 4301 auto St = Builder.CreateStore(Ops[1], Builder.CreateBitCast(PtrOp0, Ty)); 4302 return St; 4303 } 4304 case NEON::BI__builtin_neon_vtbl1_v: 4305 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbl1), 4306 Ops, "vtbl1"); 4307 case NEON::BI__builtin_neon_vtbl2_v: 4308 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbl2), 4309 Ops, "vtbl2"); 4310 case NEON::BI__builtin_neon_vtbl3_v: 4311 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbl3), 4312 Ops, "vtbl3"); 4313 case NEON::BI__builtin_neon_vtbl4_v: 4314 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbl4), 4315 Ops, "vtbl4"); 4316 case NEON::BI__builtin_neon_vtbx1_v: 4317 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbx1), 4318 Ops, "vtbx1"); 4319 case NEON::BI__builtin_neon_vtbx2_v: 4320 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbx2), 4321 Ops, "vtbx2"); 4322 case NEON::BI__builtin_neon_vtbx3_v: 4323 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbx3), 4324 Ops, "vtbx3"); 4325 case NEON::BI__builtin_neon_vtbx4_v: 4326 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbx4), 4327 Ops, "vtbx4"); 4328 } 4329 } 4330 4331 static Value *EmitAArch64TblBuiltinExpr(CodeGenFunction &CGF, unsigned BuiltinID, 4332 const CallExpr *E, 4333 SmallVectorImpl<Value *> &Ops) { 4334 unsigned int Int = 0; 4335 const char *s = nullptr; 4336 4337 switch (BuiltinID) { 4338 default: 4339 return nullptr; 4340 case NEON::BI__builtin_neon_vtbl1_v: 4341 case NEON::BI__builtin_neon_vqtbl1_v: 4342 case NEON::BI__builtin_neon_vqtbl1q_v: 4343 case NEON::BI__builtin_neon_vtbl2_v: 4344 case NEON::BI__builtin_neon_vqtbl2_v: 4345 case NEON::BI__builtin_neon_vqtbl2q_v: 4346 case NEON::BI__builtin_neon_vtbl3_v: 4347 case NEON::BI__builtin_neon_vqtbl3_v: 4348 case NEON::BI__builtin_neon_vqtbl3q_v: 4349 case NEON::BI__builtin_neon_vtbl4_v: 4350 case NEON::BI__builtin_neon_vqtbl4_v: 4351 case NEON::BI__builtin_neon_vqtbl4q_v: 4352 break; 4353 case NEON::BI__builtin_neon_vtbx1_v: 4354 case NEON::BI__builtin_neon_vqtbx1_v: 4355 case NEON::BI__builtin_neon_vqtbx1q_v: 4356 case NEON::BI__builtin_neon_vtbx2_v: 4357 case NEON::BI__builtin_neon_vqtbx2_v: 4358 case NEON::BI__builtin_neon_vqtbx2q_v: 4359 case NEON::BI__builtin_neon_vtbx3_v: 4360 case NEON::BI__builtin_neon_vqtbx3_v: 4361 case NEON::BI__builtin_neon_vqtbx3q_v: 4362 case NEON::BI__builtin_neon_vtbx4_v: 4363 case NEON::BI__builtin_neon_vqtbx4_v: 4364 case NEON::BI__builtin_neon_vqtbx4q_v: 4365 break; 4366 } 4367 4368 assert(E->getNumArgs() >= 3); 4369 4370 // Get the last argument, which specifies the vector type. 4371 llvm::APSInt Result; 4372 const Expr *Arg = E->getArg(E->getNumArgs() - 1); 4373 if (!Arg->isIntegerConstantExpr(Result, CGF.getContext())) 4374 return nullptr; 4375 4376 // Determine the type of this overloaded NEON intrinsic. 4377 NeonTypeFlags Type(Result.getZExtValue()); 4378 llvm::VectorType *Ty = GetNeonType(&CGF, Type); 4379 if (!Ty) 4380 return nullptr; 4381 4382 CodeGen::CGBuilderTy &Builder = CGF.Builder; 4383 4384 // AArch64 scalar builtins are not overloaded, they do not have an extra 4385 // argument that specifies the vector type, need to handle each case. 4386 switch (BuiltinID) { 4387 case NEON::BI__builtin_neon_vtbl1_v: { 4388 return packTBLDVectorList(CGF, makeArrayRef(Ops).slice(0, 1), nullptr, 4389 Ops[1], Ty, Intrinsic::aarch64_neon_tbl1, 4390 "vtbl1"); 4391 } 4392 case NEON::BI__builtin_neon_vtbl2_v: { 4393 return packTBLDVectorList(CGF, makeArrayRef(Ops).slice(0, 2), nullptr, 4394 Ops[2], Ty, Intrinsic::aarch64_neon_tbl1, 4395 "vtbl1"); 4396 } 4397 case NEON::BI__builtin_neon_vtbl3_v: { 4398 return packTBLDVectorList(CGF, makeArrayRef(Ops).slice(0, 3), nullptr, 4399 Ops[3], Ty, Intrinsic::aarch64_neon_tbl2, 4400 "vtbl2"); 4401 } 4402 case NEON::BI__builtin_neon_vtbl4_v: { 4403 return packTBLDVectorList(CGF, makeArrayRef(Ops).slice(0, 4), nullptr, 4404 Ops[4], Ty, Intrinsic::aarch64_neon_tbl2, 4405 "vtbl2"); 4406 } 4407 case NEON::BI__builtin_neon_vtbx1_v: { 4408 Value *TblRes = 4409 packTBLDVectorList(CGF, makeArrayRef(Ops).slice(1, 1), nullptr, Ops[2], 4410 Ty, Intrinsic::aarch64_neon_tbl1, "vtbl1"); 4411 4412 llvm::Constant *EightV = ConstantInt::get(Ty, 8); 4413 Value *CmpRes = Builder.CreateICmp(ICmpInst::ICMP_UGE, Ops[2], EightV); 4414 CmpRes = Builder.CreateSExt(CmpRes, Ty); 4415 4416 Value *EltsFromInput = Builder.CreateAnd(CmpRes, Ops[0]); 4417 Value *EltsFromTbl = Builder.CreateAnd(Builder.CreateNot(CmpRes), TblRes); 4418 return Builder.CreateOr(EltsFromInput, EltsFromTbl, "vtbx"); 4419 } 4420 case NEON::BI__builtin_neon_vtbx2_v: { 4421 return packTBLDVectorList(CGF, makeArrayRef(Ops).slice(1, 2), Ops[0], 4422 Ops[3], Ty, Intrinsic::aarch64_neon_tbx1, 4423 "vtbx1"); 4424 } 4425 case NEON::BI__builtin_neon_vtbx3_v: { 4426 Value *TblRes = 4427 packTBLDVectorList(CGF, makeArrayRef(Ops).slice(1, 3), nullptr, Ops[4], 4428 Ty, Intrinsic::aarch64_neon_tbl2, "vtbl2"); 4429 4430 llvm::Constant *TwentyFourV = ConstantInt::get(Ty, 24); 4431 Value *CmpRes = Builder.CreateICmp(ICmpInst::ICMP_UGE, Ops[4], 4432 TwentyFourV); 4433 CmpRes = Builder.CreateSExt(CmpRes, Ty); 4434 4435 Value *EltsFromInput = Builder.CreateAnd(CmpRes, Ops[0]); 4436 Value *EltsFromTbl = Builder.CreateAnd(Builder.CreateNot(CmpRes), TblRes); 4437 return Builder.CreateOr(EltsFromInput, EltsFromTbl, "vtbx"); 4438 } 4439 case NEON::BI__builtin_neon_vtbx4_v: { 4440 return packTBLDVectorList(CGF, makeArrayRef(Ops).slice(1, 4), Ops[0], 4441 Ops[5], Ty, Intrinsic::aarch64_neon_tbx2, 4442 "vtbx2"); 4443 } 4444 case NEON::BI__builtin_neon_vqtbl1_v: 4445 case NEON::BI__builtin_neon_vqtbl1q_v: 4446 Int = Intrinsic::aarch64_neon_tbl1; s = "vtbl1"; break; 4447 case NEON::BI__builtin_neon_vqtbl2_v: 4448 case NEON::BI__builtin_neon_vqtbl2q_v: { 4449 Int = Intrinsic::aarch64_neon_tbl2; s = "vtbl2"; break; 4450 case NEON::BI__builtin_neon_vqtbl3_v: 4451 case NEON::BI__builtin_neon_vqtbl3q_v: 4452 Int = Intrinsic::aarch64_neon_tbl3; s = "vtbl3"; break; 4453 case NEON::BI__builtin_neon_vqtbl4_v: 4454 case NEON::BI__builtin_neon_vqtbl4q_v: 4455 Int = Intrinsic::aarch64_neon_tbl4; s = "vtbl4"; break; 4456 case NEON::BI__builtin_neon_vqtbx1_v: 4457 case NEON::BI__builtin_neon_vqtbx1q_v: 4458 Int = Intrinsic::aarch64_neon_tbx1; s = "vtbx1"; break; 4459 case NEON::BI__builtin_neon_vqtbx2_v: 4460 case NEON::BI__builtin_neon_vqtbx2q_v: 4461 Int = Intrinsic::aarch64_neon_tbx2; s = "vtbx2"; break; 4462 case NEON::BI__builtin_neon_vqtbx3_v: 4463 case NEON::BI__builtin_neon_vqtbx3q_v: 4464 Int = Intrinsic::aarch64_neon_tbx3; s = "vtbx3"; break; 4465 case NEON::BI__builtin_neon_vqtbx4_v: 4466 case NEON::BI__builtin_neon_vqtbx4q_v: 4467 Int = Intrinsic::aarch64_neon_tbx4; s = "vtbx4"; break; 4468 } 4469 } 4470 4471 if (!Int) 4472 return nullptr; 4473 4474 Function *F = CGF.CGM.getIntrinsic(Int, Ty); 4475 return CGF.EmitNeonCall(F, Ops, s); 4476 } 4477 4478 Value *CodeGenFunction::vectorWrapScalar16(Value *Op) { 4479 llvm::Type *VTy = llvm::VectorType::get(Int16Ty, 4); 4480 Op = Builder.CreateBitCast(Op, Int16Ty); 4481 Value *V = UndefValue::get(VTy); 4482 llvm::Constant *CI = ConstantInt::get(SizeTy, 0); 4483 Op = Builder.CreateInsertElement(V, Op, CI); 4484 return Op; 4485 } 4486 4487 Value *CodeGenFunction::EmitAArch64BuiltinExpr(unsigned BuiltinID, 4488 const CallExpr *E) { 4489 unsigned HintID = static_cast<unsigned>(-1); 4490 switch (BuiltinID) { 4491 default: break; 4492 case AArch64::BI__builtin_arm_nop: 4493 HintID = 0; 4494 break; 4495 case AArch64::BI__builtin_arm_yield: 4496 HintID = 1; 4497 break; 4498 case AArch64::BI__builtin_arm_wfe: 4499 HintID = 2; 4500 break; 4501 case AArch64::BI__builtin_arm_wfi: 4502 HintID = 3; 4503 break; 4504 case AArch64::BI__builtin_arm_sev: 4505 HintID = 4; 4506 break; 4507 case AArch64::BI__builtin_arm_sevl: 4508 HintID = 5; 4509 break; 4510 } 4511 4512 if (HintID != static_cast<unsigned>(-1)) { 4513 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_hint); 4514 return Builder.CreateCall(F, llvm::ConstantInt::get(Int32Ty, HintID)); 4515 } 4516 4517 if (BuiltinID == AArch64::BI__builtin_arm_prefetch) { 4518 Value *Address = EmitScalarExpr(E->getArg(0)); 4519 Value *RW = EmitScalarExpr(E->getArg(1)); 4520 Value *CacheLevel = EmitScalarExpr(E->getArg(2)); 4521 Value *RetentionPolicy = EmitScalarExpr(E->getArg(3)); 4522 Value *IsData = EmitScalarExpr(E->getArg(4)); 4523 4524 Value *Locality = nullptr; 4525 if (cast<llvm::ConstantInt>(RetentionPolicy)->isZero()) { 4526 // Temporal fetch, needs to convert cache level to locality. 4527 Locality = llvm::ConstantInt::get(Int32Ty, 4528 -cast<llvm::ConstantInt>(CacheLevel)->getValue() + 3); 4529 } else { 4530 // Streaming fetch. 4531 Locality = llvm::ConstantInt::get(Int32Ty, 0); 4532 } 4533 4534 // FIXME: We need AArch64 specific LLVM intrinsic if we want to specify 4535 // PLDL3STRM or PLDL2STRM. 4536 Value *F = CGM.getIntrinsic(Intrinsic::prefetch); 4537 return Builder.CreateCall(F, {Address, RW, Locality, IsData}); 4538 } 4539 4540 if (BuiltinID == AArch64::BI__builtin_arm_rbit) { 4541 assert((getContext().getTypeSize(E->getType()) == 32) && 4542 "rbit of unusual size!"); 4543 llvm::Value *Arg = EmitScalarExpr(E->getArg(0)); 4544 return Builder.CreateCall( 4545 CGM.getIntrinsic(Intrinsic::aarch64_rbit, Arg->getType()), Arg, "rbit"); 4546 } 4547 if (BuiltinID == AArch64::BI__builtin_arm_rbit64) { 4548 assert((getContext().getTypeSize(E->getType()) == 64) && 4549 "rbit of unusual size!"); 4550 llvm::Value *Arg = EmitScalarExpr(E->getArg(0)); 4551 return Builder.CreateCall( 4552 CGM.getIntrinsic(Intrinsic::aarch64_rbit, Arg->getType()), Arg, "rbit"); 4553 } 4554 4555 if (BuiltinID == AArch64::BI__clear_cache) { 4556 assert(E->getNumArgs() == 2 && "__clear_cache takes 2 arguments"); 4557 const FunctionDecl *FD = E->getDirectCallee(); 4558 Value *Ops[2]; 4559 for (unsigned i = 0; i < 2; i++) 4560 Ops[i] = EmitScalarExpr(E->getArg(i)); 4561 llvm::Type *Ty = CGM.getTypes().ConvertType(FD->getType()); 4562 llvm::FunctionType *FTy = cast<llvm::FunctionType>(Ty); 4563 StringRef Name = FD->getName(); 4564 return EmitNounwindRuntimeCall(CGM.CreateRuntimeFunction(FTy, Name), Ops); 4565 } 4566 4567 if ((BuiltinID == AArch64::BI__builtin_arm_ldrex || 4568 BuiltinID == AArch64::BI__builtin_arm_ldaex) && 4569 getContext().getTypeSize(E->getType()) == 128) { 4570 Function *F = CGM.getIntrinsic(BuiltinID == AArch64::BI__builtin_arm_ldaex 4571 ? Intrinsic::aarch64_ldaxp 4572 : Intrinsic::aarch64_ldxp); 4573 4574 Value *LdPtr = EmitScalarExpr(E->getArg(0)); 4575 Value *Val = Builder.CreateCall(F, Builder.CreateBitCast(LdPtr, Int8PtrTy), 4576 "ldxp"); 4577 4578 Value *Val0 = Builder.CreateExtractValue(Val, 1); 4579 Value *Val1 = Builder.CreateExtractValue(Val, 0); 4580 llvm::Type *Int128Ty = llvm::IntegerType::get(getLLVMContext(), 128); 4581 Val0 = Builder.CreateZExt(Val0, Int128Ty); 4582 Val1 = Builder.CreateZExt(Val1, Int128Ty); 4583 4584 Value *ShiftCst = llvm::ConstantInt::get(Int128Ty, 64); 4585 Val = Builder.CreateShl(Val0, ShiftCst, "shl", true /* nuw */); 4586 Val = Builder.CreateOr(Val, Val1); 4587 return Builder.CreateBitCast(Val, ConvertType(E->getType())); 4588 } else if (BuiltinID == AArch64::BI__builtin_arm_ldrex || 4589 BuiltinID == AArch64::BI__builtin_arm_ldaex) { 4590 Value *LoadAddr = EmitScalarExpr(E->getArg(0)); 4591 4592 QualType Ty = E->getType(); 4593 llvm::Type *RealResTy = ConvertType(Ty); 4594 llvm::Type *IntResTy = llvm::IntegerType::get(getLLVMContext(), 4595 getContext().getTypeSize(Ty)); 4596 LoadAddr = Builder.CreateBitCast(LoadAddr, IntResTy->getPointerTo()); 4597 4598 Function *F = CGM.getIntrinsic(BuiltinID == AArch64::BI__builtin_arm_ldaex 4599 ? Intrinsic::aarch64_ldaxr 4600 : Intrinsic::aarch64_ldxr, 4601 LoadAddr->getType()); 4602 Value *Val = Builder.CreateCall(F, LoadAddr, "ldxr"); 4603 4604 if (RealResTy->isPointerTy()) 4605 return Builder.CreateIntToPtr(Val, RealResTy); 4606 4607 Val = Builder.CreateTruncOrBitCast(Val, IntResTy); 4608 return Builder.CreateBitCast(Val, RealResTy); 4609 } 4610 4611 if ((BuiltinID == AArch64::BI__builtin_arm_strex || 4612 BuiltinID == AArch64::BI__builtin_arm_stlex) && 4613 getContext().getTypeSize(E->getArg(0)->getType()) == 128) { 4614 Function *F = CGM.getIntrinsic(BuiltinID == AArch64::BI__builtin_arm_stlex 4615 ? Intrinsic::aarch64_stlxp 4616 : Intrinsic::aarch64_stxp); 4617 llvm::Type *STy = llvm::StructType::get(Int64Ty, Int64Ty, nullptr); 4618 4619 Address Tmp = CreateMemTemp(E->getArg(0)->getType()); 4620 EmitAnyExprToMem(E->getArg(0), Tmp, Qualifiers(), /*init*/ true); 4621 4622 Tmp = Builder.CreateBitCast(Tmp, llvm::PointerType::getUnqual(STy)); 4623 llvm::Value *Val = Builder.CreateLoad(Tmp); 4624 4625 Value *Arg0 = Builder.CreateExtractValue(Val, 0); 4626 Value *Arg1 = Builder.CreateExtractValue(Val, 1); 4627 Value *StPtr = Builder.CreateBitCast(EmitScalarExpr(E->getArg(1)), 4628 Int8PtrTy); 4629 return Builder.CreateCall(F, {Arg0, Arg1, StPtr}, "stxp"); 4630 } 4631 4632 if (BuiltinID == AArch64::BI__builtin_arm_strex || 4633 BuiltinID == AArch64::BI__builtin_arm_stlex) { 4634 Value *StoreVal = EmitScalarExpr(E->getArg(0)); 4635 Value *StoreAddr = EmitScalarExpr(E->getArg(1)); 4636 4637 QualType Ty = E->getArg(0)->getType(); 4638 llvm::Type *StoreTy = llvm::IntegerType::get(getLLVMContext(), 4639 getContext().getTypeSize(Ty)); 4640 StoreAddr = Builder.CreateBitCast(StoreAddr, StoreTy->getPointerTo()); 4641 4642 if (StoreVal->getType()->isPointerTy()) 4643 StoreVal = Builder.CreatePtrToInt(StoreVal, Int64Ty); 4644 else { 4645 StoreVal = Builder.CreateBitCast(StoreVal, StoreTy); 4646 StoreVal = Builder.CreateZExtOrBitCast(StoreVal, Int64Ty); 4647 } 4648 4649 Function *F = CGM.getIntrinsic(BuiltinID == AArch64::BI__builtin_arm_stlex 4650 ? Intrinsic::aarch64_stlxr 4651 : Intrinsic::aarch64_stxr, 4652 StoreAddr->getType()); 4653 return Builder.CreateCall(F, {StoreVal, StoreAddr}, "stxr"); 4654 } 4655 4656 if (BuiltinID == AArch64::BI__builtin_arm_clrex) { 4657 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_clrex); 4658 return Builder.CreateCall(F); 4659 } 4660 4661 if (BuiltinID == AArch64::BI__builtin_thread_pointer) { 4662 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_thread_pointer); 4663 return Builder.CreateCall(F); 4664 } 4665 4666 // CRC32 4667 Intrinsic::ID CRCIntrinsicID = Intrinsic::not_intrinsic; 4668 switch (BuiltinID) { 4669 case AArch64::BI__builtin_arm_crc32b: 4670 CRCIntrinsicID = Intrinsic::aarch64_crc32b; break; 4671 case AArch64::BI__builtin_arm_crc32cb: 4672 CRCIntrinsicID = Intrinsic::aarch64_crc32cb; break; 4673 case AArch64::BI__builtin_arm_crc32h: 4674 CRCIntrinsicID = Intrinsic::aarch64_crc32h; break; 4675 case AArch64::BI__builtin_arm_crc32ch: 4676 CRCIntrinsicID = Intrinsic::aarch64_crc32ch; break; 4677 case AArch64::BI__builtin_arm_crc32w: 4678 CRCIntrinsicID = Intrinsic::aarch64_crc32w; break; 4679 case AArch64::BI__builtin_arm_crc32cw: 4680 CRCIntrinsicID = Intrinsic::aarch64_crc32cw; break; 4681 case AArch64::BI__builtin_arm_crc32d: 4682 CRCIntrinsicID = Intrinsic::aarch64_crc32x; break; 4683 case AArch64::BI__builtin_arm_crc32cd: 4684 CRCIntrinsicID = Intrinsic::aarch64_crc32cx; break; 4685 } 4686 4687 if (CRCIntrinsicID != Intrinsic::not_intrinsic) { 4688 Value *Arg0 = EmitScalarExpr(E->getArg(0)); 4689 Value *Arg1 = EmitScalarExpr(E->getArg(1)); 4690 Function *F = CGM.getIntrinsic(CRCIntrinsicID); 4691 4692 llvm::Type *DataTy = F->getFunctionType()->getParamType(1); 4693 Arg1 = Builder.CreateZExtOrBitCast(Arg1, DataTy); 4694 4695 return Builder.CreateCall(F, {Arg0, Arg1}); 4696 } 4697 4698 if (BuiltinID == AArch64::BI__builtin_arm_rsr || 4699 BuiltinID == AArch64::BI__builtin_arm_rsr64 || 4700 BuiltinID == AArch64::BI__builtin_arm_rsrp || 4701 BuiltinID == AArch64::BI__builtin_arm_wsr || 4702 BuiltinID == AArch64::BI__builtin_arm_wsr64 || 4703 BuiltinID == AArch64::BI__builtin_arm_wsrp) { 4704 4705 bool IsRead = BuiltinID == AArch64::BI__builtin_arm_rsr || 4706 BuiltinID == AArch64::BI__builtin_arm_rsr64 || 4707 BuiltinID == AArch64::BI__builtin_arm_rsrp; 4708 4709 bool IsPointerBuiltin = BuiltinID == AArch64::BI__builtin_arm_rsrp || 4710 BuiltinID == AArch64::BI__builtin_arm_wsrp; 4711 4712 bool Is64Bit = BuiltinID != AArch64::BI__builtin_arm_rsr && 4713 BuiltinID != AArch64::BI__builtin_arm_wsr; 4714 4715 llvm::Type *ValueType; 4716 llvm::Type *RegisterType = Int64Ty; 4717 if (IsPointerBuiltin) { 4718 ValueType = VoidPtrTy; 4719 } else if (Is64Bit) { 4720 ValueType = Int64Ty; 4721 } else { 4722 ValueType = Int32Ty; 4723 } 4724 4725 return EmitSpecialRegisterBuiltin(*this, E, RegisterType, ValueType, IsRead); 4726 } 4727 4728 // Find out if any arguments are required to be integer constant 4729 // expressions. 4730 unsigned ICEArguments = 0; 4731 ASTContext::GetBuiltinTypeError Error; 4732 getContext().GetBuiltinType(BuiltinID, Error, &ICEArguments); 4733 assert(Error == ASTContext::GE_None && "Should not codegen an error"); 4734 4735 llvm::SmallVector<Value*, 4> Ops; 4736 for (unsigned i = 0, e = E->getNumArgs() - 1; i != e; i++) { 4737 if ((ICEArguments & (1 << i)) == 0) { 4738 Ops.push_back(EmitScalarExpr(E->getArg(i))); 4739 } else { 4740 // If this is required to be a constant, constant fold it so that we know 4741 // that the generated intrinsic gets a ConstantInt. 4742 llvm::APSInt Result; 4743 bool IsConst = E->getArg(i)->isIntegerConstantExpr(Result, getContext()); 4744 assert(IsConst && "Constant arg isn't actually constant?"); 4745 (void)IsConst; 4746 Ops.push_back(llvm::ConstantInt::get(getLLVMContext(), Result)); 4747 } 4748 } 4749 4750 auto SISDMap = makeArrayRef(AArch64SISDIntrinsicMap); 4751 const NeonIntrinsicInfo *Builtin = findNeonIntrinsicInMap( 4752 SISDMap, BuiltinID, AArch64SISDIntrinsicsProvenSorted); 4753 4754 if (Builtin) { 4755 Ops.push_back(EmitScalarExpr(E->getArg(E->getNumArgs() - 1))); 4756 Value *Result = EmitCommonNeonSISDBuiltinExpr(*this, *Builtin, Ops, E); 4757 assert(Result && "SISD intrinsic should have been handled"); 4758 return Result; 4759 } 4760 4761 llvm::APSInt Result; 4762 const Expr *Arg = E->getArg(E->getNumArgs()-1); 4763 NeonTypeFlags Type(0); 4764 if (Arg->isIntegerConstantExpr(Result, getContext())) 4765 // Determine the type of this overloaded NEON intrinsic. 4766 Type = NeonTypeFlags(Result.getZExtValue()); 4767 4768 bool usgn = Type.isUnsigned(); 4769 bool quad = Type.isQuad(); 4770 4771 // Handle non-overloaded intrinsics first. 4772 switch (BuiltinID) { 4773 default: break; 4774 case NEON::BI__builtin_neon_vldrq_p128: { 4775 llvm::Type *Int128PTy = llvm::Type::getIntNPtrTy(getLLVMContext(), 128); 4776 Value *Ptr = Builder.CreateBitCast(EmitScalarExpr(E->getArg(0)), Int128PTy); 4777 return Builder.CreateDefaultAlignedLoad(Ptr); 4778 } 4779 case NEON::BI__builtin_neon_vstrq_p128: { 4780 llvm::Type *Int128PTy = llvm::Type::getIntNPtrTy(getLLVMContext(), 128); 4781 Value *Ptr = Builder.CreateBitCast(Ops[0], Int128PTy); 4782 return Builder.CreateDefaultAlignedStore(EmitScalarExpr(E->getArg(1)), Ptr); 4783 } 4784 case NEON::BI__builtin_neon_vcvts_u32_f32: 4785 case NEON::BI__builtin_neon_vcvtd_u64_f64: 4786 usgn = true; 4787 // FALL THROUGH 4788 case NEON::BI__builtin_neon_vcvts_s32_f32: 4789 case NEON::BI__builtin_neon_vcvtd_s64_f64: { 4790 Ops.push_back(EmitScalarExpr(E->getArg(0))); 4791 bool Is64 = Ops[0]->getType()->getPrimitiveSizeInBits() == 64; 4792 llvm::Type *InTy = Is64 ? Int64Ty : Int32Ty; 4793 llvm::Type *FTy = Is64 ? DoubleTy : FloatTy; 4794 Ops[0] = Builder.CreateBitCast(Ops[0], FTy); 4795 if (usgn) 4796 return Builder.CreateFPToUI(Ops[0], InTy); 4797 return Builder.CreateFPToSI(Ops[0], InTy); 4798 } 4799 case NEON::BI__builtin_neon_vcvts_f32_u32: 4800 case NEON::BI__builtin_neon_vcvtd_f64_u64: 4801 usgn = true; 4802 // FALL THROUGH 4803 case NEON::BI__builtin_neon_vcvts_f32_s32: 4804 case NEON::BI__builtin_neon_vcvtd_f64_s64: { 4805 Ops.push_back(EmitScalarExpr(E->getArg(0))); 4806 bool Is64 = Ops[0]->getType()->getPrimitiveSizeInBits() == 64; 4807 llvm::Type *InTy = Is64 ? Int64Ty : Int32Ty; 4808 llvm::Type *FTy = Is64 ? DoubleTy : FloatTy; 4809 Ops[0] = Builder.CreateBitCast(Ops[0], InTy); 4810 if (usgn) 4811 return Builder.CreateUIToFP(Ops[0], FTy); 4812 return Builder.CreateSIToFP(Ops[0], FTy); 4813 } 4814 case NEON::BI__builtin_neon_vpaddd_s64: { 4815 llvm::Type *Ty = llvm::VectorType::get(Int64Ty, 2); 4816 Value *Vec = EmitScalarExpr(E->getArg(0)); 4817 // The vector is v2f64, so make sure it's bitcast to that. 4818 Vec = Builder.CreateBitCast(Vec, Ty, "v2i64"); 4819 llvm::Value *Idx0 = llvm::ConstantInt::get(SizeTy, 0); 4820 llvm::Value *Idx1 = llvm::ConstantInt::get(SizeTy, 1); 4821 Value *Op0 = Builder.CreateExtractElement(Vec, Idx0, "lane0"); 4822 Value *Op1 = Builder.CreateExtractElement(Vec, Idx1, "lane1"); 4823 // Pairwise addition of a v2f64 into a scalar f64. 4824 return Builder.CreateAdd(Op0, Op1, "vpaddd"); 4825 } 4826 case NEON::BI__builtin_neon_vpaddd_f64: { 4827 llvm::Type *Ty = 4828 llvm::VectorType::get(DoubleTy, 2); 4829 Value *Vec = EmitScalarExpr(E->getArg(0)); 4830 // The vector is v2f64, so make sure it's bitcast to that. 4831 Vec = Builder.CreateBitCast(Vec, Ty, "v2f64"); 4832 llvm::Value *Idx0 = llvm::ConstantInt::get(SizeTy, 0); 4833 llvm::Value *Idx1 = llvm::ConstantInt::get(SizeTy, 1); 4834 Value *Op0 = Builder.CreateExtractElement(Vec, Idx0, "lane0"); 4835 Value *Op1 = Builder.CreateExtractElement(Vec, Idx1, "lane1"); 4836 // Pairwise addition of a v2f64 into a scalar f64. 4837 return Builder.CreateFAdd(Op0, Op1, "vpaddd"); 4838 } 4839 case NEON::BI__builtin_neon_vpadds_f32: { 4840 llvm::Type *Ty = 4841 llvm::VectorType::get(FloatTy, 2); 4842 Value *Vec = EmitScalarExpr(E->getArg(0)); 4843 // The vector is v2f32, so make sure it's bitcast to that. 4844 Vec = Builder.CreateBitCast(Vec, Ty, "v2f32"); 4845 llvm::Value *Idx0 = llvm::ConstantInt::get(SizeTy, 0); 4846 llvm::Value *Idx1 = llvm::ConstantInt::get(SizeTy, 1); 4847 Value *Op0 = Builder.CreateExtractElement(Vec, Idx0, "lane0"); 4848 Value *Op1 = Builder.CreateExtractElement(Vec, Idx1, "lane1"); 4849 // Pairwise addition of a v2f32 into a scalar f32. 4850 return Builder.CreateFAdd(Op0, Op1, "vpaddd"); 4851 } 4852 case NEON::BI__builtin_neon_vceqzd_s64: 4853 case NEON::BI__builtin_neon_vceqzd_f64: 4854 case NEON::BI__builtin_neon_vceqzs_f32: 4855 Ops.push_back(EmitScalarExpr(E->getArg(0))); 4856 return EmitAArch64CompareBuiltinExpr( 4857 Ops[0], ConvertType(E->getCallReturnType(getContext())), 4858 ICmpInst::FCMP_OEQ, ICmpInst::ICMP_EQ, "vceqz"); 4859 case NEON::BI__builtin_neon_vcgezd_s64: 4860 case NEON::BI__builtin_neon_vcgezd_f64: 4861 case NEON::BI__builtin_neon_vcgezs_f32: 4862 Ops.push_back(EmitScalarExpr(E->getArg(0))); 4863 return EmitAArch64CompareBuiltinExpr( 4864 Ops[0], ConvertType(E->getCallReturnType(getContext())), 4865 ICmpInst::FCMP_OGE, ICmpInst::ICMP_SGE, "vcgez"); 4866 case NEON::BI__builtin_neon_vclezd_s64: 4867 case NEON::BI__builtin_neon_vclezd_f64: 4868 case NEON::BI__builtin_neon_vclezs_f32: 4869 Ops.push_back(EmitScalarExpr(E->getArg(0))); 4870 return EmitAArch64CompareBuiltinExpr( 4871 Ops[0], ConvertType(E->getCallReturnType(getContext())), 4872 ICmpInst::FCMP_OLE, ICmpInst::ICMP_SLE, "vclez"); 4873 case NEON::BI__builtin_neon_vcgtzd_s64: 4874 case NEON::BI__builtin_neon_vcgtzd_f64: 4875 case NEON::BI__builtin_neon_vcgtzs_f32: 4876 Ops.push_back(EmitScalarExpr(E->getArg(0))); 4877 return EmitAArch64CompareBuiltinExpr( 4878 Ops[0], ConvertType(E->getCallReturnType(getContext())), 4879 ICmpInst::FCMP_OGT, ICmpInst::ICMP_SGT, "vcgtz"); 4880 case NEON::BI__builtin_neon_vcltzd_s64: 4881 case NEON::BI__builtin_neon_vcltzd_f64: 4882 case NEON::BI__builtin_neon_vcltzs_f32: 4883 Ops.push_back(EmitScalarExpr(E->getArg(0))); 4884 return EmitAArch64CompareBuiltinExpr( 4885 Ops[0], ConvertType(E->getCallReturnType(getContext())), 4886 ICmpInst::FCMP_OLT, ICmpInst::ICMP_SLT, "vcltz"); 4887 4888 case NEON::BI__builtin_neon_vceqzd_u64: { 4889 Ops.push_back(EmitScalarExpr(E->getArg(0))); 4890 Ops[0] = Builder.CreateBitCast(Ops[0], Int64Ty); 4891 Ops[0] = 4892 Builder.CreateICmpEQ(Ops[0], llvm::Constant::getNullValue(Int64Ty)); 4893 return Builder.CreateSExt(Ops[0], Int64Ty, "vceqzd"); 4894 } 4895 case NEON::BI__builtin_neon_vceqd_f64: 4896 case NEON::BI__builtin_neon_vcled_f64: 4897 case NEON::BI__builtin_neon_vcltd_f64: 4898 case NEON::BI__builtin_neon_vcged_f64: 4899 case NEON::BI__builtin_neon_vcgtd_f64: { 4900 llvm::CmpInst::Predicate P; 4901 switch (BuiltinID) { 4902 default: llvm_unreachable("missing builtin ID in switch!"); 4903 case NEON::BI__builtin_neon_vceqd_f64: P = llvm::FCmpInst::FCMP_OEQ; break; 4904 case NEON::BI__builtin_neon_vcled_f64: P = llvm::FCmpInst::FCMP_OLE; break; 4905 case NEON::BI__builtin_neon_vcltd_f64: P = llvm::FCmpInst::FCMP_OLT; break; 4906 case NEON::BI__builtin_neon_vcged_f64: P = llvm::FCmpInst::FCMP_OGE; break; 4907 case NEON::BI__builtin_neon_vcgtd_f64: P = llvm::FCmpInst::FCMP_OGT; break; 4908 } 4909 Ops.push_back(EmitScalarExpr(E->getArg(1))); 4910 Ops[0] = Builder.CreateBitCast(Ops[0], DoubleTy); 4911 Ops[1] = Builder.CreateBitCast(Ops[1], DoubleTy); 4912 Ops[0] = Builder.CreateFCmp(P, Ops[0], Ops[1]); 4913 return Builder.CreateSExt(Ops[0], Int64Ty, "vcmpd"); 4914 } 4915 case NEON::BI__builtin_neon_vceqs_f32: 4916 case NEON::BI__builtin_neon_vcles_f32: 4917 case NEON::BI__builtin_neon_vclts_f32: 4918 case NEON::BI__builtin_neon_vcges_f32: 4919 case NEON::BI__builtin_neon_vcgts_f32: { 4920 llvm::CmpInst::Predicate P; 4921 switch (BuiltinID) { 4922 default: llvm_unreachable("missing builtin ID in switch!"); 4923 case NEON::BI__builtin_neon_vceqs_f32: P = llvm::FCmpInst::FCMP_OEQ; break; 4924 case NEON::BI__builtin_neon_vcles_f32: P = llvm::FCmpInst::FCMP_OLE; break; 4925 case NEON::BI__builtin_neon_vclts_f32: P = llvm::FCmpInst::FCMP_OLT; break; 4926 case NEON::BI__builtin_neon_vcges_f32: P = llvm::FCmpInst::FCMP_OGE; break; 4927 case NEON::BI__builtin_neon_vcgts_f32: P = llvm::FCmpInst::FCMP_OGT; break; 4928 } 4929 Ops.push_back(EmitScalarExpr(E->getArg(1))); 4930 Ops[0] = Builder.CreateBitCast(Ops[0], FloatTy); 4931 Ops[1] = Builder.CreateBitCast(Ops[1], FloatTy); 4932 Ops[0] = Builder.CreateFCmp(P, Ops[0], Ops[1]); 4933 return Builder.CreateSExt(Ops[0], Int32Ty, "vcmpd"); 4934 } 4935 case NEON::BI__builtin_neon_vceqd_s64: 4936 case NEON::BI__builtin_neon_vceqd_u64: 4937 case NEON::BI__builtin_neon_vcgtd_s64: 4938 case NEON::BI__builtin_neon_vcgtd_u64: 4939 case NEON::BI__builtin_neon_vcltd_s64: 4940 case NEON::BI__builtin_neon_vcltd_u64: 4941 case NEON::BI__builtin_neon_vcged_u64: 4942 case NEON::BI__builtin_neon_vcged_s64: 4943 case NEON::BI__builtin_neon_vcled_u64: 4944 case NEON::BI__builtin_neon_vcled_s64: { 4945 llvm::CmpInst::Predicate P; 4946 switch (BuiltinID) { 4947 default: llvm_unreachable("missing builtin ID in switch!"); 4948 case NEON::BI__builtin_neon_vceqd_s64: 4949 case NEON::BI__builtin_neon_vceqd_u64:P = llvm::ICmpInst::ICMP_EQ;break; 4950 case NEON::BI__builtin_neon_vcgtd_s64:P = llvm::ICmpInst::ICMP_SGT;break; 4951 case NEON::BI__builtin_neon_vcgtd_u64:P = llvm::ICmpInst::ICMP_UGT;break; 4952 case NEON::BI__builtin_neon_vcltd_s64:P = llvm::ICmpInst::ICMP_SLT;break; 4953 case NEON::BI__builtin_neon_vcltd_u64:P = llvm::ICmpInst::ICMP_ULT;break; 4954 case NEON::BI__builtin_neon_vcged_u64:P = llvm::ICmpInst::ICMP_UGE;break; 4955 case NEON::BI__builtin_neon_vcged_s64:P = llvm::ICmpInst::ICMP_SGE;break; 4956 case NEON::BI__builtin_neon_vcled_u64:P = llvm::ICmpInst::ICMP_ULE;break; 4957 case NEON::BI__builtin_neon_vcled_s64:P = llvm::ICmpInst::ICMP_SLE;break; 4958 } 4959 Ops.push_back(EmitScalarExpr(E->getArg(1))); 4960 Ops[0] = Builder.CreateBitCast(Ops[0], Int64Ty); 4961 Ops[1] = Builder.CreateBitCast(Ops[1], Int64Ty); 4962 Ops[0] = Builder.CreateICmp(P, Ops[0], Ops[1]); 4963 return Builder.CreateSExt(Ops[0], Int64Ty, "vceqd"); 4964 } 4965 case NEON::BI__builtin_neon_vtstd_s64: 4966 case NEON::BI__builtin_neon_vtstd_u64: { 4967 Ops.push_back(EmitScalarExpr(E->getArg(1))); 4968 Ops[0] = Builder.CreateBitCast(Ops[0], Int64Ty); 4969 Ops[1] = Builder.CreateBitCast(Ops[1], Int64Ty); 4970 Ops[0] = Builder.CreateAnd(Ops[0], Ops[1]); 4971 Ops[0] = Builder.CreateICmp(ICmpInst::ICMP_NE, Ops[0], 4972 llvm::Constant::getNullValue(Int64Ty)); 4973 return Builder.CreateSExt(Ops[0], Int64Ty, "vtstd"); 4974 } 4975 case NEON::BI__builtin_neon_vset_lane_i8: 4976 case NEON::BI__builtin_neon_vset_lane_i16: 4977 case NEON::BI__builtin_neon_vset_lane_i32: 4978 case NEON::BI__builtin_neon_vset_lane_i64: 4979 case NEON::BI__builtin_neon_vset_lane_f32: 4980 case NEON::BI__builtin_neon_vsetq_lane_i8: 4981 case NEON::BI__builtin_neon_vsetq_lane_i16: 4982 case NEON::BI__builtin_neon_vsetq_lane_i32: 4983 case NEON::BI__builtin_neon_vsetq_lane_i64: 4984 case NEON::BI__builtin_neon_vsetq_lane_f32: 4985 Ops.push_back(EmitScalarExpr(E->getArg(2))); 4986 return Builder.CreateInsertElement(Ops[1], Ops[0], Ops[2], "vset_lane"); 4987 case NEON::BI__builtin_neon_vset_lane_f64: 4988 // The vector type needs a cast for the v1f64 variant. 4989 Ops[1] = Builder.CreateBitCast(Ops[1], 4990 llvm::VectorType::get(DoubleTy, 1)); 4991 Ops.push_back(EmitScalarExpr(E->getArg(2))); 4992 return Builder.CreateInsertElement(Ops[1], Ops[0], Ops[2], "vset_lane"); 4993 case NEON::BI__builtin_neon_vsetq_lane_f64: 4994 // The vector type needs a cast for the v2f64 variant. 4995 Ops[1] = Builder.CreateBitCast(Ops[1], 4996 llvm::VectorType::get(DoubleTy, 2)); 4997 Ops.push_back(EmitScalarExpr(E->getArg(2))); 4998 return Builder.CreateInsertElement(Ops[1], Ops[0], Ops[2], "vset_lane"); 4999 5000 case NEON::BI__builtin_neon_vget_lane_i8: 5001 case NEON::BI__builtin_neon_vdupb_lane_i8: 5002 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::VectorType::get(Int8Ty, 8)); 5003 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 5004 "vget_lane"); 5005 case NEON::BI__builtin_neon_vgetq_lane_i8: 5006 case NEON::BI__builtin_neon_vdupb_laneq_i8: 5007 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::VectorType::get(Int8Ty, 16)); 5008 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 5009 "vgetq_lane"); 5010 case NEON::BI__builtin_neon_vget_lane_i16: 5011 case NEON::BI__builtin_neon_vduph_lane_i16: 5012 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::VectorType::get(Int16Ty, 4)); 5013 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 5014 "vget_lane"); 5015 case NEON::BI__builtin_neon_vgetq_lane_i16: 5016 case NEON::BI__builtin_neon_vduph_laneq_i16: 5017 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::VectorType::get(Int16Ty, 8)); 5018 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 5019 "vgetq_lane"); 5020 case NEON::BI__builtin_neon_vget_lane_i32: 5021 case NEON::BI__builtin_neon_vdups_lane_i32: 5022 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::VectorType::get(Int32Ty, 2)); 5023 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 5024 "vget_lane"); 5025 case NEON::BI__builtin_neon_vdups_lane_f32: 5026 Ops[0] = Builder.CreateBitCast(Ops[0], 5027 llvm::VectorType::get(FloatTy, 2)); 5028 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 5029 "vdups_lane"); 5030 case NEON::BI__builtin_neon_vgetq_lane_i32: 5031 case NEON::BI__builtin_neon_vdups_laneq_i32: 5032 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::VectorType::get(Int32Ty, 4)); 5033 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 5034 "vgetq_lane"); 5035 case NEON::BI__builtin_neon_vget_lane_i64: 5036 case NEON::BI__builtin_neon_vdupd_lane_i64: 5037 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::VectorType::get(Int64Ty, 1)); 5038 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 5039 "vget_lane"); 5040 case NEON::BI__builtin_neon_vdupd_lane_f64: 5041 Ops[0] = Builder.CreateBitCast(Ops[0], 5042 llvm::VectorType::get(DoubleTy, 1)); 5043 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 5044 "vdupd_lane"); 5045 case NEON::BI__builtin_neon_vgetq_lane_i64: 5046 case NEON::BI__builtin_neon_vdupd_laneq_i64: 5047 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::VectorType::get(Int64Ty, 2)); 5048 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 5049 "vgetq_lane"); 5050 case NEON::BI__builtin_neon_vget_lane_f32: 5051 Ops[0] = Builder.CreateBitCast(Ops[0], 5052 llvm::VectorType::get(FloatTy, 2)); 5053 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 5054 "vget_lane"); 5055 case NEON::BI__builtin_neon_vget_lane_f64: 5056 Ops[0] = Builder.CreateBitCast(Ops[0], 5057 llvm::VectorType::get(DoubleTy, 1)); 5058 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 5059 "vget_lane"); 5060 case NEON::BI__builtin_neon_vgetq_lane_f32: 5061 case NEON::BI__builtin_neon_vdups_laneq_f32: 5062 Ops[0] = Builder.CreateBitCast(Ops[0], 5063 llvm::VectorType::get(FloatTy, 4)); 5064 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 5065 "vgetq_lane"); 5066 case NEON::BI__builtin_neon_vgetq_lane_f64: 5067 case NEON::BI__builtin_neon_vdupd_laneq_f64: 5068 Ops[0] = Builder.CreateBitCast(Ops[0], 5069 llvm::VectorType::get(DoubleTy, 2)); 5070 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 5071 "vgetq_lane"); 5072 case NEON::BI__builtin_neon_vaddd_s64: 5073 case NEON::BI__builtin_neon_vaddd_u64: 5074 return Builder.CreateAdd(Ops[0], EmitScalarExpr(E->getArg(1)), "vaddd"); 5075 case NEON::BI__builtin_neon_vsubd_s64: 5076 case NEON::BI__builtin_neon_vsubd_u64: 5077 return Builder.CreateSub(Ops[0], EmitScalarExpr(E->getArg(1)), "vsubd"); 5078 case NEON::BI__builtin_neon_vqdmlalh_s16: 5079 case NEON::BI__builtin_neon_vqdmlslh_s16: { 5080 SmallVector<Value *, 2> ProductOps; 5081 ProductOps.push_back(vectorWrapScalar16(Ops[1])); 5082 ProductOps.push_back(vectorWrapScalar16(EmitScalarExpr(E->getArg(2)))); 5083 llvm::Type *VTy = llvm::VectorType::get(Int32Ty, 4); 5084 Ops[1] = EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_sqdmull, VTy), 5085 ProductOps, "vqdmlXl"); 5086 Constant *CI = ConstantInt::get(SizeTy, 0); 5087 Ops[1] = Builder.CreateExtractElement(Ops[1], CI, "lane0"); 5088 5089 unsigned AccumInt = BuiltinID == NEON::BI__builtin_neon_vqdmlalh_s16 5090 ? Intrinsic::aarch64_neon_sqadd 5091 : Intrinsic::aarch64_neon_sqsub; 5092 return EmitNeonCall(CGM.getIntrinsic(AccumInt, Int32Ty), Ops, "vqdmlXl"); 5093 } 5094 case NEON::BI__builtin_neon_vqshlud_n_s64: { 5095 Ops.push_back(EmitScalarExpr(E->getArg(1))); 5096 Ops[1] = Builder.CreateZExt(Ops[1], Int64Ty); 5097 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_sqshlu, Int64Ty), 5098 Ops, "vqshlu_n"); 5099 } 5100 case NEON::BI__builtin_neon_vqshld_n_u64: 5101 case NEON::BI__builtin_neon_vqshld_n_s64: { 5102 unsigned Int = BuiltinID == NEON::BI__builtin_neon_vqshld_n_u64 5103 ? Intrinsic::aarch64_neon_uqshl 5104 : Intrinsic::aarch64_neon_sqshl; 5105 Ops.push_back(EmitScalarExpr(E->getArg(1))); 5106 Ops[1] = Builder.CreateZExt(Ops[1], Int64Ty); 5107 return EmitNeonCall(CGM.getIntrinsic(Int, Int64Ty), Ops, "vqshl_n"); 5108 } 5109 case NEON::BI__builtin_neon_vrshrd_n_u64: 5110 case NEON::BI__builtin_neon_vrshrd_n_s64: { 5111 unsigned Int = BuiltinID == NEON::BI__builtin_neon_vrshrd_n_u64 5112 ? Intrinsic::aarch64_neon_urshl 5113 : Intrinsic::aarch64_neon_srshl; 5114 Ops.push_back(EmitScalarExpr(E->getArg(1))); 5115 int SV = cast<ConstantInt>(Ops[1])->getSExtValue(); 5116 Ops[1] = ConstantInt::get(Int64Ty, -SV); 5117 return EmitNeonCall(CGM.getIntrinsic(Int, Int64Ty), Ops, "vrshr_n"); 5118 } 5119 case NEON::BI__builtin_neon_vrsrad_n_u64: 5120 case NEON::BI__builtin_neon_vrsrad_n_s64: { 5121 unsigned Int = BuiltinID == NEON::BI__builtin_neon_vrsrad_n_u64 5122 ? Intrinsic::aarch64_neon_urshl 5123 : Intrinsic::aarch64_neon_srshl; 5124 Ops[1] = Builder.CreateBitCast(Ops[1], Int64Ty); 5125 Ops.push_back(Builder.CreateNeg(EmitScalarExpr(E->getArg(2)))); 5126 Ops[1] = Builder.CreateCall(CGM.getIntrinsic(Int, Int64Ty), 5127 {Ops[1], Builder.CreateSExt(Ops[2], Int64Ty)}); 5128 return Builder.CreateAdd(Ops[0], Builder.CreateBitCast(Ops[1], Int64Ty)); 5129 } 5130 case NEON::BI__builtin_neon_vshld_n_s64: 5131 case NEON::BI__builtin_neon_vshld_n_u64: { 5132 llvm::ConstantInt *Amt = cast<ConstantInt>(EmitScalarExpr(E->getArg(1))); 5133 return Builder.CreateShl( 5134 Ops[0], ConstantInt::get(Int64Ty, Amt->getZExtValue()), "shld_n"); 5135 } 5136 case NEON::BI__builtin_neon_vshrd_n_s64: { 5137 llvm::ConstantInt *Amt = cast<ConstantInt>(EmitScalarExpr(E->getArg(1))); 5138 return Builder.CreateAShr( 5139 Ops[0], ConstantInt::get(Int64Ty, std::min(static_cast<uint64_t>(63), 5140 Amt->getZExtValue())), 5141 "shrd_n"); 5142 } 5143 case NEON::BI__builtin_neon_vshrd_n_u64: { 5144 llvm::ConstantInt *Amt = cast<ConstantInt>(EmitScalarExpr(E->getArg(1))); 5145 uint64_t ShiftAmt = Amt->getZExtValue(); 5146 // Right-shifting an unsigned value by its size yields 0. 5147 if (ShiftAmt == 64) 5148 return ConstantInt::get(Int64Ty, 0); 5149 return Builder.CreateLShr(Ops[0], ConstantInt::get(Int64Ty, ShiftAmt), 5150 "shrd_n"); 5151 } 5152 case NEON::BI__builtin_neon_vsrad_n_s64: { 5153 llvm::ConstantInt *Amt = cast<ConstantInt>(EmitScalarExpr(E->getArg(2))); 5154 Ops[1] = Builder.CreateAShr( 5155 Ops[1], ConstantInt::get(Int64Ty, std::min(static_cast<uint64_t>(63), 5156 Amt->getZExtValue())), 5157 "shrd_n"); 5158 return Builder.CreateAdd(Ops[0], Ops[1]); 5159 } 5160 case NEON::BI__builtin_neon_vsrad_n_u64: { 5161 llvm::ConstantInt *Amt = cast<ConstantInt>(EmitScalarExpr(E->getArg(2))); 5162 uint64_t ShiftAmt = Amt->getZExtValue(); 5163 // Right-shifting an unsigned value by its size yields 0. 5164 // As Op + 0 = Op, return Ops[0] directly. 5165 if (ShiftAmt == 64) 5166 return Ops[0]; 5167 Ops[1] = Builder.CreateLShr(Ops[1], ConstantInt::get(Int64Ty, ShiftAmt), 5168 "shrd_n"); 5169 return Builder.CreateAdd(Ops[0], Ops[1]); 5170 } 5171 case NEON::BI__builtin_neon_vqdmlalh_lane_s16: 5172 case NEON::BI__builtin_neon_vqdmlalh_laneq_s16: 5173 case NEON::BI__builtin_neon_vqdmlslh_lane_s16: 5174 case NEON::BI__builtin_neon_vqdmlslh_laneq_s16: { 5175 Ops[2] = Builder.CreateExtractElement(Ops[2], EmitScalarExpr(E->getArg(3)), 5176 "lane"); 5177 SmallVector<Value *, 2> ProductOps; 5178 ProductOps.push_back(vectorWrapScalar16(Ops[1])); 5179 ProductOps.push_back(vectorWrapScalar16(Ops[2])); 5180 llvm::Type *VTy = llvm::VectorType::get(Int32Ty, 4); 5181 Ops[1] = EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_sqdmull, VTy), 5182 ProductOps, "vqdmlXl"); 5183 Constant *CI = ConstantInt::get(SizeTy, 0); 5184 Ops[1] = Builder.CreateExtractElement(Ops[1], CI, "lane0"); 5185 Ops.pop_back(); 5186 5187 unsigned AccInt = (BuiltinID == NEON::BI__builtin_neon_vqdmlalh_lane_s16 || 5188 BuiltinID == NEON::BI__builtin_neon_vqdmlalh_laneq_s16) 5189 ? Intrinsic::aarch64_neon_sqadd 5190 : Intrinsic::aarch64_neon_sqsub; 5191 return EmitNeonCall(CGM.getIntrinsic(AccInt, Int32Ty), Ops, "vqdmlXl"); 5192 } 5193 case NEON::BI__builtin_neon_vqdmlals_s32: 5194 case NEON::BI__builtin_neon_vqdmlsls_s32: { 5195 SmallVector<Value *, 2> ProductOps; 5196 ProductOps.push_back(Ops[1]); 5197 ProductOps.push_back(EmitScalarExpr(E->getArg(2))); 5198 Ops[1] = 5199 EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_sqdmulls_scalar), 5200 ProductOps, "vqdmlXl"); 5201 5202 unsigned AccumInt = BuiltinID == NEON::BI__builtin_neon_vqdmlals_s32 5203 ? Intrinsic::aarch64_neon_sqadd 5204 : Intrinsic::aarch64_neon_sqsub; 5205 return EmitNeonCall(CGM.getIntrinsic(AccumInt, Int64Ty), Ops, "vqdmlXl"); 5206 } 5207 case NEON::BI__builtin_neon_vqdmlals_lane_s32: 5208 case NEON::BI__builtin_neon_vqdmlals_laneq_s32: 5209 case NEON::BI__builtin_neon_vqdmlsls_lane_s32: 5210 case NEON::BI__builtin_neon_vqdmlsls_laneq_s32: { 5211 Ops[2] = Builder.CreateExtractElement(Ops[2], EmitScalarExpr(E->getArg(3)), 5212 "lane"); 5213 SmallVector<Value *, 2> ProductOps; 5214 ProductOps.push_back(Ops[1]); 5215 ProductOps.push_back(Ops[2]); 5216 Ops[1] = 5217 EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_sqdmulls_scalar), 5218 ProductOps, "vqdmlXl"); 5219 Ops.pop_back(); 5220 5221 unsigned AccInt = (BuiltinID == NEON::BI__builtin_neon_vqdmlals_lane_s32 || 5222 BuiltinID == NEON::BI__builtin_neon_vqdmlals_laneq_s32) 5223 ? Intrinsic::aarch64_neon_sqadd 5224 : Intrinsic::aarch64_neon_sqsub; 5225 return EmitNeonCall(CGM.getIntrinsic(AccInt, Int64Ty), Ops, "vqdmlXl"); 5226 } 5227 } 5228 5229 llvm::VectorType *VTy = GetNeonType(this, Type); 5230 llvm::Type *Ty = VTy; 5231 if (!Ty) 5232 return nullptr; 5233 5234 // Not all intrinsics handled by the common case work for AArch64 yet, so only 5235 // defer to common code if it's been added to our special map. 5236 Builtin = findNeonIntrinsicInMap(AArch64SIMDIntrinsicMap, BuiltinID, 5237 AArch64SIMDIntrinsicsProvenSorted); 5238 5239 if (Builtin) 5240 return EmitCommonNeonBuiltinExpr( 5241 Builtin->BuiltinID, Builtin->LLVMIntrinsic, Builtin->AltLLVMIntrinsic, 5242 Builtin->NameHint, Builtin->TypeModifier, E, Ops, 5243 /*never use addresses*/ Address::invalid(), Address::invalid()); 5244 5245 if (Value *V = EmitAArch64TblBuiltinExpr(*this, BuiltinID, E, Ops)) 5246 return V; 5247 5248 unsigned Int; 5249 switch (BuiltinID) { 5250 default: return nullptr; 5251 case NEON::BI__builtin_neon_vbsl_v: 5252 case NEON::BI__builtin_neon_vbslq_v: { 5253 llvm::Type *BitTy = llvm::VectorType::getInteger(VTy); 5254 Ops[0] = Builder.CreateBitCast(Ops[0], BitTy, "vbsl"); 5255 Ops[1] = Builder.CreateBitCast(Ops[1], BitTy, "vbsl"); 5256 Ops[2] = Builder.CreateBitCast(Ops[2], BitTy, "vbsl"); 5257 5258 Ops[1] = Builder.CreateAnd(Ops[0], Ops[1], "vbsl"); 5259 Ops[2] = Builder.CreateAnd(Builder.CreateNot(Ops[0]), Ops[2], "vbsl"); 5260 Ops[0] = Builder.CreateOr(Ops[1], Ops[2], "vbsl"); 5261 return Builder.CreateBitCast(Ops[0], Ty); 5262 } 5263 case NEON::BI__builtin_neon_vfma_lane_v: 5264 case NEON::BI__builtin_neon_vfmaq_lane_v: { // Only used for FP types 5265 // The ARM builtins (and instructions) have the addend as the first 5266 // operand, but the 'fma' intrinsics have it last. Swap it around here. 5267 Value *Addend = Ops[0]; 5268 Value *Multiplicand = Ops[1]; 5269 Value *LaneSource = Ops[2]; 5270 Ops[0] = Multiplicand; 5271 Ops[1] = LaneSource; 5272 Ops[2] = Addend; 5273 5274 // Now adjust things to handle the lane access. 5275 llvm::Type *SourceTy = BuiltinID == NEON::BI__builtin_neon_vfmaq_lane_v ? 5276 llvm::VectorType::get(VTy->getElementType(), VTy->getNumElements() / 2) : 5277 VTy; 5278 llvm::Constant *cst = cast<Constant>(Ops[3]); 5279 Value *SV = llvm::ConstantVector::getSplat(VTy->getNumElements(), cst); 5280 Ops[1] = Builder.CreateBitCast(Ops[1], SourceTy); 5281 Ops[1] = Builder.CreateShuffleVector(Ops[1], Ops[1], SV, "lane"); 5282 5283 Ops.pop_back(); 5284 Int = Intrinsic::fma; 5285 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "fmla"); 5286 } 5287 case NEON::BI__builtin_neon_vfma_laneq_v: { 5288 llvm::VectorType *VTy = cast<llvm::VectorType>(Ty); 5289 // v1f64 fma should be mapped to Neon scalar f64 fma 5290 if (VTy && VTy->getElementType() == DoubleTy) { 5291 Ops[0] = Builder.CreateBitCast(Ops[0], DoubleTy); 5292 Ops[1] = Builder.CreateBitCast(Ops[1], DoubleTy); 5293 llvm::Type *VTy = GetNeonType(this, 5294 NeonTypeFlags(NeonTypeFlags::Float64, false, true)); 5295 Ops[2] = Builder.CreateBitCast(Ops[2], VTy); 5296 Ops[2] = Builder.CreateExtractElement(Ops[2], Ops[3], "extract"); 5297 Value *F = CGM.getIntrinsic(Intrinsic::fma, DoubleTy); 5298 Value *Result = Builder.CreateCall(F, {Ops[1], Ops[2], Ops[0]}); 5299 return Builder.CreateBitCast(Result, Ty); 5300 } 5301 Value *F = CGM.getIntrinsic(Intrinsic::fma, Ty); 5302 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 5303 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 5304 5305 llvm::Type *STy = llvm::VectorType::get(VTy->getElementType(), 5306 VTy->getNumElements() * 2); 5307 Ops[2] = Builder.CreateBitCast(Ops[2], STy); 5308 Value* SV = llvm::ConstantVector::getSplat(VTy->getNumElements(), 5309 cast<ConstantInt>(Ops[3])); 5310 Ops[2] = Builder.CreateShuffleVector(Ops[2], Ops[2], SV, "lane"); 5311 5312 return Builder.CreateCall(F, {Ops[2], Ops[1], Ops[0]}); 5313 } 5314 case NEON::BI__builtin_neon_vfmaq_laneq_v: { 5315 Value *F = CGM.getIntrinsic(Intrinsic::fma, Ty); 5316 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 5317 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 5318 5319 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 5320 Ops[2] = EmitNeonSplat(Ops[2], cast<ConstantInt>(Ops[3])); 5321 return Builder.CreateCall(F, {Ops[2], Ops[1], Ops[0]}); 5322 } 5323 case NEON::BI__builtin_neon_vfmas_lane_f32: 5324 case NEON::BI__builtin_neon_vfmas_laneq_f32: 5325 case NEON::BI__builtin_neon_vfmad_lane_f64: 5326 case NEON::BI__builtin_neon_vfmad_laneq_f64: { 5327 Ops.push_back(EmitScalarExpr(E->getArg(3))); 5328 llvm::Type *Ty = ConvertType(E->getCallReturnType(getContext())); 5329 Value *F = CGM.getIntrinsic(Intrinsic::fma, Ty); 5330 Ops[2] = Builder.CreateExtractElement(Ops[2], Ops[3], "extract"); 5331 return Builder.CreateCall(F, {Ops[1], Ops[2], Ops[0]}); 5332 } 5333 case NEON::BI__builtin_neon_vfms_v: 5334 case NEON::BI__builtin_neon_vfmsq_v: { // Only used for FP types 5335 // FIXME: probably remove when we no longer support aarch64_simd.h 5336 // (arm_neon.h delegates to vfma). 5337 5338 // The ARM builtins (and instructions) have the addend as the first 5339 // operand, but the 'fma' intrinsics have it last. Swap it around here. 5340 Value *Subtrahend = Ops[0]; 5341 Value *Multiplicand = Ops[2]; 5342 Ops[0] = Multiplicand; 5343 Ops[2] = Subtrahend; 5344 Ops[1] = Builder.CreateBitCast(Ops[1], VTy); 5345 Ops[1] = Builder.CreateFNeg(Ops[1]); 5346 Int = Intrinsic::fma; 5347 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "fmls"); 5348 } 5349 case NEON::BI__builtin_neon_vmull_v: 5350 // FIXME: improve sharing scheme to cope with 3 alternative LLVM intrinsics. 5351 Int = usgn ? Intrinsic::aarch64_neon_umull : Intrinsic::aarch64_neon_smull; 5352 if (Type.isPoly()) Int = Intrinsic::aarch64_neon_pmull; 5353 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmull"); 5354 case NEON::BI__builtin_neon_vmax_v: 5355 case NEON::BI__builtin_neon_vmaxq_v: 5356 // FIXME: improve sharing scheme to cope with 3 alternative LLVM intrinsics. 5357 Int = usgn ? Intrinsic::aarch64_neon_umax : Intrinsic::aarch64_neon_smax; 5358 if (Ty->isFPOrFPVectorTy()) Int = Intrinsic::aarch64_neon_fmax; 5359 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmax"); 5360 case NEON::BI__builtin_neon_vmin_v: 5361 case NEON::BI__builtin_neon_vminq_v: 5362 // FIXME: improve sharing scheme to cope with 3 alternative LLVM intrinsics. 5363 Int = usgn ? Intrinsic::aarch64_neon_umin : Intrinsic::aarch64_neon_smin; 5364 if (Ty->isFPOrFPVectorTy()) Int = Intrinsic::aarch64_neon_fmin; 5365 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmin"); 5366 case NEON::BI__builtin_neon_vabd_v: 5367 case NEON::BI__builtin_neon_vabdq_v: 5368 // FIXME: improve sharing scheme to cope with 3 alternative LLVM intrinsics. 5369 Int = usgn ? Intrinsic::aarch64_neon_uabd : Intrinsic::aarch64_neon_sabd; 5370 if (Ty->isFPOrFPVectorTy()) Int = Intrinsic::aarch64_neon_fabd; 5371 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vabd"); 5372 case NEON::BI__builtin_neon_vpadal_v: 5373 case NEON::BI__builtin_neon_vpadalq_v: { 5374 unsigned ArgElts = VTy->getNumElements(); 5375 llvm::IntegerType *EltTy = cast<IntegerType>(VTy->getElementType()); 5376 unsigned BitWidth = EltTy->getBitWidth(); 5377 llvm::Type *ArgTy = llvm::VectorType::get( 5378 llvm::IntegerType::get(getLLVMContext(), BitWidth/2), 2*ArgElts); 5379 llvm::Type* Tys[2] = { VTy, ArgTy }; 5380 Int = usgn ? Intrinsic::aarch64_neon_uaddlp : Intrinsic::aarch64_neon_saddlp; 5381 SmallVector<llvm::Value*, 1> TmpOps; 5382 TmpOps.push_back(Ops[1]); 5383 Function *F = CGM.getIntrinsic(Int, Tys); 5384 llvm::Value *tmp = EmitNeonCall(F, TmpOps, "vpadal"); 5385 llvm::Value *addend = Builder.CreateBitCast(Ops[0], tmp->getType()); 5386 return Builder.CreateAdd(tmp, addend); 5387 } 5388 case NEON::BI__builtin_neon_vpmin_v: 5389 case NEON::BI__builtin_neon_vpminq_v: 5390 // FIXME: improve sharing scheme to cope with 3 alternative LLVM intrinsics. 5391 Int = usgn ? Intrinsic::aarch64_neon_uminp : Intrinsic::aarch64_neon_sminp; 5392 if (Ty->isFPOrFPVectorTy()) Int = Intrinsic::aarch64_neon_fminp; 5393 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vpmin"); 5394 case NEON::BI__builtin_neon_vpmax_v: 5395 case NEON::BI__builtin_neon_vpmaxq_v: 5396 // FIXME: improve sharing scheme to cope with 3 alternative LLVM intrinsics. 5397 Int = usgn ? Intrinsic::aarch64_neon_umaxp : Intrinsic::aarch64_neon_smaxp; 5398 if (Ty->isFPOrFPVectorTy()) Int = Intrinsic::aarch64_neon_fmaxp; 5399 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vpmax"); 5400 case NEON::BI__builtin_neon_vminnm_v: 5401 case NEON::BI__builtin_neon_vminnmq_v: 5402 Int = Intrinsic::aarch64_neon_fminnm; 5403 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vminnm"); 5404 case NEON::BI__builtin_neon_vmaxnm_v: 5405 case NEON::BI__builtin_neon_vmaxnmq_v: 5406 Int = Intrinsic::aarch64_neon_fmaxnm; 5407 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmaxnm"); 5408 case NEON::BI__builtin_neon_vrecpss_f32: { 5409 Ops.push_back(EmitScalarExpr(E->getArg(1))); 5410 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_frecps, FloatTy), 5411 Ops, "vrecps"); 5412 } 5413 case NEON::BI__builtin_neon_vrecpsd_f64: { 5414 Ops.push_back(EmitScalarExpr(E->getArg(1))); 5415 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_frecps, DoubleTy), 5416 Ops, "vrecps"); 5417 } 5418 case NEON::BI__builtin_neon_vqshrun_n_v: 5419 Int = Intrinsic::aarch64_neon_sqshrun; 5420 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqshrun_n"); 5421 case NEON::BI__builtin_neon_vqrshrun_n_v: 5422 Int = Intrinsic::aarch64_neon_sqrshrun; 5423 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqrshrun_n"); 5424 case NEON::BI__builtin_neon_vqshrn_n_v: 5425 Int = usgn ? Intrinsic::aarch64_neon_uqshrn : Intrinsic::aarch64_neon_sqshrn; 5426 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqshrn_n"); 5427 case NEON::BI__builtin_neon_vrshrn_n_v: 5428 Int = Intrinsic::aarch64_neon_rshrn; 5429 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrshrn_n"); 5430 case NEON::BI__builtin_neon_vqrshrn_n_v: 5431 Int = usgn ? Intrinsic::aarch64_neon_uqrshrn : Intrinsic::aarch64_neon_sqrshrn; 5432 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqrshrn_n"); 5433 case NEON::BI__builtin_neon_vrnda_v: 5434 case NEON::BI__builtin_neon_vrndaq_v: { 5435 Int = Intrinsic::round; 5436 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrnda"); 5437 } 5438 case NEON::BI__builtin_neon_vrndi_v: 5439 case NEON::BI__builtin_neon_vrndiq_v: { 5440 Int = Intrinsic::nearbyint; 5441 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrndi"); 5442 } 5443 case NEON::BI__builtin_neon_vrndm_v: 5444 case NEON::BI__builtin_neon_vrndmq_v: { 5445 Int = Intrinsic::floor; 5446 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrndm"); 5447 } 5448 case NEON::BI__builtin_neon_vrndn_v: 5449 case NEON::BI__builtin_neon_vrndnq_v: { 5450 Int = Intrinsic::aarch64_neon_frintn; 5451 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrndn"); 5452 } 5453 case NEON::BI__builtin_neon_vrndp_v: 5454 case NEON::BI__builtin_neon_vrndpq_v: { 5455 Int = Intrinsic::ceil; 5456 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrndp"); 5457 } 5458 case NEON::BI__builtin_neon_vrndx_v: 5459 case NEON::BI__builtin_neon_vrndxq_v: { 5460 Int = Intrinsic::rint; 5461 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrndx"); 5462 } 5463 case NEON::BI__builtin_neon_vrnd_v: 5464 case NEON::BI__builtin_neon_vrndq_v: { 5465 Int = Intrinsic::trunc; 5466 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrndz"); 5467 } 5468 case NEON::BI__builtin_neon_vceqz_v: 5469 case NEON::BI__builtin_neon_vceqzq_v: 5470 return EmitAArch64CompareBuiltinExpr(Ops[0], Ty, ICmpInst::FCMP_OEQ, 5471 ICmpInst::ICMP_EQ, "vceqz"); 5472 case NEON::BI__builtin_neon_vcgez_v: 5473 case NEON::BI__builtin_neon_vcgezq_v: 5474 return EmitAArch64CompareBuiltinExpr(Ops[0], Ty, ICmpInst::FCMP_OGE, 5475 ICmpInst::ICMP_SGE, "vcgez"); 5476 case NEON::BI__builtin_neon_vclez_v: 5477 case NEON::BI__builtin_neon_vclezq_v: 5478 return EmitAArch64CompareBuiltinExpr(Ops[0], Ty, ICmpInst::FCMP_OLE, 5479 ICmpInst::ICMP_SLE, "vclez"); 5480 case NEON::BI__builtin_neon_vcgtz_v: 5481 case NEON::BI__builtin_neon_vcgtzq_v: 5482 return EmitAArch64CompareBuiltinExpr(Ops[0], Ty, ICmpInst::FCMP_OGT, 5483 ICmpInst::ICMP_SGT, "vcgtz"); 5484 case NEON::BI__builtin_neon_vcltz_v: 5485 case NEON::BI__builtin_neon_vcltzq_v: 5486 return EmitAArch64CompareBuiltinExpr(Ops[0], Ty, ICmpInst::FCMP_OLT, 5487 ICmpInst::ICMP_SLT, "vcltz"); 5488 case NEON::BI__builtin_neon_vcvt_f64_v: 5489 case NEON::BI__builtin_neon_vcvtq_f64_v: 5490 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 5491 Ty = GetNeonType(this, NeonTypeFlags(NeonTypeFlags::Float64, false, quad)); 5492 return usgn ? Builder.CreateUIToFP(Ops[0], Ty, "vcvt") 5493 : Builder.CreateSIToFP(Ops[0], Ty, "vcvt"); 5494 case NEON::BI__builtin_neon_vcvt_f64_f32: { 5495 assert(Type.getEltType() == NeonTypeFlags::Float64 && quad && 5496 "unexpected vcvt_f64_f32 builtin"); 5497 NeonTypeFlags SrcFlag = NeonTypeFlags(NeonTypeFlags::Float32, false, false); 5498 Ops[0] = Builder.CreateBitCast(Ops[0], GetNeonType(this, SrcFlag)); 5499 5500 return Builder.CreateFPExt(Ops[0], Ty, "vcvt"); 5501 } 5502 case NEON::BI__builtin_neon_vcvt_f32_f64: { 5503 assert(Type.getEltType() == NeonTypeFlags::Float32 && 5504 "unexpected vcvt_f32_f64 builtin"); 5505 NeonTypeFlags SrcFlag = NeonTypeFlags(NeonTypeFlags::Float64, false, true); 5506 Ops[0] = Builder.CreateBitCast(Ops[0], GetNeonType(this, SrcFlag)); 5507 5508 return Builder.CreateFPTrunc(Ops[0], Ty, "vcvt"); 5509 } 5510 case NEON::BI__builtin_neon_vcvt_s32_v: 5511 case NEON::BI__builtin_neon_vcvt_u32_v: 5512 case NEON::BI__builtin_neon_vcvt_s64_v: 5513 case NEON::BI__builtin_neon_vcvt_u64_v: 5514 case NEON::BI__builtin_neon_vcvtq_s32_v: 5515 case NEON::BI__builtin_neon_vcvtq_u32_v: 5516 case NEON::BI__builtin_neon_vcvtq_s64_v: 5517 case NEON::BI__builtin_neon_vcvtq_u64_v: { 5518 Ops[0] = Builder.CreateBitCast(Ops[0], GetFloatNeonType(this, Type)); 5519 if (usgn) 5520 return Builder.CreateFPToUI(Ops[0], Ty); 5521 return Builder.CreateFPToSI(Ops[0], Ty); 5522 } 5523 case NEON::BI__builtin_neon_vcvta_s32_v: 5524 case NEON::BI__builtin_neon_vcvtaq_s32_v: 5525 case NEON::BI__builtin_neon_vcvta_u32_v: 5526 case NEON::BI__builtin_neon_vcvtaq_u32_v: 5527 case NEON::BI__builtin_neon_vcvta_s64_v: 5528 case NEON::BI__builtin_neon_vcvtaq_s64_v: 5529 case NEON::BI__builtin_neon_vcvta_u64_v: 5530 case NEON::BI__builtin_neon_vcvtaq_u64_v: { 5531 Int = usgn ? Intrinsic::aarch64_neon_fcvtau : Intrinsic::aarch64_neon_fcvtas; 5532 llvm::Type *Tys[2] = { Ty, GetFloatNeonType(this, Type) }; 5533 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vcvta"); 5534 } 5535 case NEON::BI__builtin_neon_vcvtm_s32_v: 5536 case NEON::BI__builtin_neon_vcvtmq_s32_v: 5537 case NEON::BI__builtin_neon_vcvtm_u32_v: 5538 case NEON::BI__builtin_neon_vcvtmq_u32_v: 5539 case NEON::BI__builtin_neon_vcvtm_s64_v: 5540 case NEON::BI__builtin_neon_vcvtmq_s64_v: 5541 case NEON::BI__builtin_neon_vcvtm_u64_v: 5542 case NEON::BI__builtin_neon_vcvtmq_u64_v: { 5543 Int = usgn ? Intrinsic::aarch64_neon_fcvtmu : Intrinsic::aarch64_neon_fcvtms; 5544 llvm::Type *Tys[2] = { Ty, GetFloatNeonType(this, Type) }; 5545 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vcvtm"); 5546 } 5547 case NEON::BI__builtin_neon_vcvtn_s32_v: 5548 case NEON::BI__builtin_neon_vcvtnq_s32_v: 5549 case NEON::BI__builtin_neon_vcvtn_u32_v: 5550 case NEON::BI__builtin_neon_vcvtnq_u32_v: 5551 case NEON::BI__builtin_neon_vcvtn_s64_v: 5552 case NEON::BI__builtin_neon_vcvtnq_s64_v: 5553 case NEON::BI__builtin_neon_vcvtn_u64_v: 5554 case NEON::BI__builtin_neon_vcvtnq_u64_v: { 5555 Int = usgn ? Intrinsic::aarch64_neon_fcvtnu : Intrinsic::aarch64_neon_fcvtns; 5556 llvm::Type *Tys[2] = { Ty, GetFloatNeonType(this, Type) }; 5557 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vcvtn"); 5558 } 5559 case NEON::BI__builtin_neon_vcvtp_s32_v: 5560 case NEON::BI__builtin_neon_vcvtpq_s32_v: 5561 case NEON::BI__builtin_neon_vcvtp_u32_v: 5562 case NEON::BI__builtin_neon_vcvtpq_u32_v: 5563 case NEON::BI__builtin_neon_vcvtp_s64_v: 5564 case NEON::BI__builtin_neon_vcvtpq_s64_v: 5565 case NEON::BI__builtin_neon_vcvtp_u64_v: 5566 case NEON::BI__builtin_neon_vcvtpq_u64_v: { 5567 Int = usgn ? Intrinsic::aarch64_neon_fcvtpu : Intrinsic::aarch64_neon_fcvtps; 5568 llvm::Type *Tys[2] = { Ty, GetFloatNeonType(this, Type) }; 5569 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vcvtp"); 5570 } 5571 case NEON::BI__builtin_neon_vmulx_v: 5572 case NEON::BI__builtin_neon_vmulxq_v: { 5573 Int = Intrinsic::aarch64_neon_fmulx; 5574 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmulx"); 5575 } 5576 case NEON::BI__builtin_neon_vmul_lane_v: 5577 case NEON::BI__builtin_neon_vmul_laneq_v: { 5578 // v1f64 vmul_lane should be mapped to Neon scalar mul lane 5579 bool Quad = false; 5580 if (BuiltinID == NEON::BI__builtin_neon_vmul_laneq_v) 5581 Quad = true; 5582 Ops[0] = Builder.CreateBitCast(Ops[0], DoubleTy); 5583 llvm::Type *VTy = GetNeonType(this, 5584 NeonTypeFlags(NeonTypeFlags::Float64, false, Quad)); 5585 Ops[1] = Builder.CreateBitCast(Ops[1], VTy); 5586 Ops[1] = Builder.CreateExtractElement(Ops[1], Ops[2], "extract"); 5587 Value *Result = Builder.CreateFMul(Ops[0], Ops[1]); 5588 return Builder.CreateBitCast(Result, Ty); 5589 } 5590 case NEON::BI__builtin_neon_vnegd_s64: 5591 return Builder.CreateNeg(EmitScalarExpr(E->getArg(0)), "vnegd"); 5592 case NEON::BI__builtin_neon_vpmaxnm_v: 5593 case NEON::BI__builtin_neon_vpmaxnmq_v: { 5594 Int = Intrinsic::aarch64_neon_fmaxnmp; 5595 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vpmaxnm"); 5596 } 5597 case NEON::BI__builtin_neon_vpminnm_v: 5598 case NEON::BI__builtin_neon_vpminnmq_v: { 5599 Int = Intrinsic::aarch64_neon_fminnmp; 5600 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vpminnm"); 5601 } 5602 case NEON::BI__builtin_neon_vsqrt_v: 5603 case NEON::BI__builtin_neon_vsqrtq_v: { 5604 Int = Intrinsic::sqrt; 5605 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 5606 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vsqrt"); 5607 } 5608 case NEON::BI__builtin_neon_vrbit_v: 5609 case NEON::BI__builtin_neon_vrbitq_v: { 5610 Int = Intrinsic::aarch64_neon_rbit; 5611 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrbit"); 5612 } 5613 case NEON::BI__builtin_neon_vaddv_u8: 5614 // FIXME: These are handled by the AArch64 scalar code. 5615 usgn = true; 5616 // FALLTHROUGH 5617 case NEON::BI__builtin_neon_vaddv_s8: { 5618 Int = usgn ? Intrinsic::aarch64_neon_uaddv : Intrinsic::aarch64_neon_saddv; 5619 Ty = Int32Ty; 5620 VTy = llvm::VectorType::get(Int8Ty, 8); 5621 llvm::Type *Tys[2] = { Ty, VTy }; 5622 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5623 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddv"); 5624 return Builder.CreateTrunc(Ops[0], Int8Ty); 5625 } 5626 case NEON::BI__builtin_neon_vaddv_u16: 5627 usgn = true; 5628 // FALLTHROUGH 5629 case NEON::BI__builtin_neon_vaddv_s16: { 5630 Int = usgn ? Intrinsic::aarch64_neon_uaddv : Intrinsic::aarch64_neon_saddv; 5631 Ty = Int32Ty; 5632 VTy = llvm::VectorType::get(Int16Ty, 4); 5633 llvm::Type *Tys[2] = { Ty, VTy }; 5634 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5635 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddv"); 5636 return Builder.CreateTrunc(Ops[0], Int16Ty); 5637 } 5638 case NEON::BI__builtin_neon_vaddvq_u8: 5639 usgn = true; 5640 // FALLTHROUGH 5641 case NEON::BI__builtin_neon_vaddvq_s8: { 5642 Int = usgn ? Intrinsic::aarch64_neon_uaddv : Intrinsic::aarch64_neon_saddv; 5643 Ty = Int32Ty; 5644 VTy = llvm::VectorType::get(Int8Ty, 16); 5645 llvm::Type *Tys[2] = { Ty, VTy }; 5646 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5647 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddv"); 5648 return Builder.CreateTrunc(Ops[0], Int8Ty); 5649 } 5650 case NEON::BI__builtin_neon_vaddvq_u16: 5651 usgn = true; 5652 // FALLTHROUGH 5653 case NEON::BI__builtin_neon_vaddvq_s16: { 5654 Int = usgn ? Intrinsic::aarch64_neon_uaddv : Intrinsic::aarch64_neon_saddv; 5655 Ty = Int32Ty; 5656 VTy = llvm::VectorType::get(Int16Ty, 8); 5657 llvm::Type *Tys[2] = { Ty, VTy }; 5658 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5659 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddv"); 5660 return Builder.CreateTrunc(Ops[0], Int16Ty); 5661 } 5662 case NEON::BI__builtin_neon_vmaxv_u8: { 5663 Int = Intrinsic::aarch64_neon_umaxv; 5664 Ty = Int32Ty; 5665 VTy = llvm::VectorType::get(Int8Ty, 8); 5666 llvm::Type *Tys[2] = { Ty, VTy }; 5667 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5668 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv"); 5669 return Builder.CreateTrunc(Ops[0], Int8Ty); 5670 } 5671 case NEON::BI__builtin_neon_vmaxv_u16: { 5672 Int = Intrinsic::aarch64_neon_umaxv; 5673 Ty = Int32Ty; 5674 VTy = llvm::VectorType::get(Int16Ty, 4); 5675 llvm::Type *Tys[2] = { Ty, VTy }; 5676 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5677 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv"); 5678 return Builder.CreateTrunc(Ops[0], Int16Ty); 5679 } 5680 case NEON::BI__builtin_neon_vmaxvq_u8: { 5681 Int = Intrinsic::aarch64_neon_umaxv; 5682 Ty = Int32Ty; 5683 VTy = llvm::VectorType::get(Int8Ty, 16); 5684 llvm::Type *Tys[2] = { Ty, VTy }; 5685 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5686 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv"); 5687 return Builder.CreateTrunc(Ops[0], Int8Ty); 5688 } 5689 case NEON::BI__builtin_neon_vmaxvq_u16: { 5690 Int = Intrinsic::aarch64_neon_umaxv; 5691 Ty = Int32Ty; 5692 VTy = llvm::VectorType::get(Int16Ty, 8); 5693 llvm::Type *Tys[2] = { Ty, VTy }; 5694 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5695 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv"); 5696 return Builder.CreateTrunc(Ops[0], Int16Ty); 5697 } 5698 case NEON::BI__builtin_neon_vmaxv_s8: { 5699 Int = Intrinsic::aarch64_neon_smaxv; 5700 Ty = Int32Ty; 5701 VTy = llvm::VectorType::get(Int8Ty, 8); 5702 llvm::Type *Tys[2] = { Ty, VTy }; 5703 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5704 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv"); 5705 return Builder.CreateTrunc(Ops[0], Int8Ty); 5706 } 5707 case NEON::BI__builtin_neon_vmaxv_s16: { 5708 Int = Intrinsic::aarch64_neon_smaxv; 5709 Ty = Int32Ty; 5710 VTy = llvm::VectorType::get(Int16Ty, 4); 5711 llvm::Type *Tys[2] = { Ty, VTy }; 5712 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5713 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv"); 5714 return Builder.CreateTrunc(Ops[0], Int16Ty); 5715 } 5716 case NEON::BI__builtin_neon_vmaxvq_s8: { 5717 Int = Intrinsic::aarch64_neon_smaxv; 5718 Ty = Int32Ty; 5719 VTy = llvm::VectorType::get(Int8Ty, 16); 5720 llvm::Type *Tys[2] = { Ty, VTy }; 5721 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5722 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv"); 5723 return Builder.CreateTrunc(Ops[0], Int8Ty); 5724 } 5725 case NEON::BI__builtin_neon_vmaxvq_s16: { 5726 Int = Intrinsic::aarch64_neon_smaxv; 5727 Ty = Int32Ty; 5728 VTy = llvm::VectorType::get(Int16Ty, 8); 5729 llvm::Type *Tys[2] = { Ty, VTy }; 5730 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5731 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv"); 5732 return Builder.CreateTrunc(Ops[0], Int16Ty); 5733 } 5734 case NEON::BI__builtin_neon_vminv_u8: { 5735 Int = Intrinsic::aarch64_neon_uminv; 5736 Ty = Int32Ty; 5737 VTy = llvm::VectorType::get(Int8Ty, 8); 5738 llvm::Type *Tys[2] = { Ty, VTy }; 5739 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5740 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv"); 5741 return Builder.CreateTrunc(Ops[0], Int8Ty); 5742 } 5743 case NEON::BI__builtin_neon_vminv_u16: { 5744 Int = Intrinsic::aarch64_neon_uminv; 5745 Ty = Int32Ty; 5746 VTy = llvm::VectorType::get(Int16Ty, 4); 5747 llvm::Type *Tys[2] = { Ty, VTy }; 5748 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5749 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv"); 5750 return Builder.CreateTrunc(Ops[0], Int16Ty); 5751 } 5752 case NEON::BI__builtin_neon_vminvq_u8: { 5753 Int = Intrinsic::aarch64_neon_uminv; 5754 Ty = Int32Ty; 5755 VTy = llvm::VectorType::get(Int8Ty, 16); 5756 llvm::Type *Tys[2] = { Ty, VTy }; 5757 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5758 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv"); 5759 return Builder.CreateTrunc(Ops[0], Int8Ty); 5760 } 5761 case NEON::BI__builtin_neon_vminvq_u16: { 5762 Int = Intrinsic::aarch64_neon_uminv; 5763 Ty = Int32Ty; 5764 VTy = llvm::VectorType::get(Int16Ty, 8); 5765 llvm::Type *Tys[2] = { Ty, VTy }; 5766 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5767 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv"); 5768 return Builder.CreateTrunc(Ops[0], Int16Ty); 5769 } 5770 case NEON::BI__builtin_neon_vminv_s8: { 5771 Int = Intrinsic::aarch64_neon_sminv; 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, "vminv"); 5777 return Builder.CreateTrunc(Ops[0], Int8Ty); 5778 } 5779 case NEON::BI__builtin_neon_vminv_s16: { 5780 Int = Intrinsic::aarch64_neon_sminv; 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 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv"); 5786 return Builder.CreateTrunc(Ops[0], Int16Ty); 5787 } 5788 case NEON::BI__builtin_neon_vminvq_s8: { 5789 Int = Intrinsic::aarch64_neon_sminv; 5790 Ty = Int32Ty; 5791 VTy = llvm::VectorType::get(Int8Ty, 16); 5792 llvm::Type *Tys[2] = { Ty, VTy }; 5793 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5794 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv"); 5795 return Builder.CreateTrunc(Ops[0], Int8Ty); 5796 } 5797 case NEON::BI__builtin_neon_vminvq_s16: { 5798 Int = Intrinsic::aarch64_neon_sminv; 5799 Ty = Int32Ty; 5800 VTy = llvm::VectorType::get(Int16Ty, 8); 5801 llvm::Type *Tys[2] = { Ty, VTy }; 5802 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5803 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv"); 5804 return Builder.CreateTrunc(Ops[0], Int16Ty); 5805 } 5806 case NEON::BI__builtin_neon_vmul_n_f64: { 5807 Ops[0] = Builder.CreateBitCast(Ops[0], DoubleTy); 5808 Value *RHS = Builder.CreateBitCast(EmitScalarExpr(E->getArg(1)), DoubleTy); 5809 return Builder.CreateFMul(Ops[0], RHS); 5810 } 5811 case NEON::BI__builtin_neon_vaddlv_u8: { 5812 Int = Intrinsic::aarch64_neon_uaddlv; 5813 Ty = Int32Ty; 5814 VTy = llvm::VectorType::get(Int8Ty, 8); 5815 llvm::Type *Tys[2] = { Ty, VTy }; 5816 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5817 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv"); 5818 return Builder.CreateTrunc(Ops[0], Int16Ty); 5819 } 5820 case NEON::BI__builtin_neon_vaddlv_u16: { 5821 Int = Intrinsic::aarch64_neon_uaddlv; 5822 Ty = Int32Ty; 5823 VTy = llvm::VectorType::get(Int16Ty, 4); 5824 llvm::Type *Tys[2] = { Ty, VTy }; 5825 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5826 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv"); 5827 } 5828 case NEON::BI__builtin_neon_vaddlvq_u8: { 5829 Int = Intrinsic::aarch64_neon_uaddlv; 5830 Ty = Int32Ty; 5831 VTy = llvm::VectorType::get(Int8Ty, 16); 5832 llvm::Type *Tys[2] = { Ty, VTy }; 5833 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5834 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv"); 5835 return Builder.CreateTrunc(Ops[0], Int16Ty); 5836 } 5837 case NEON::BI__builtin_neon_vaddlvq_u16: { 5838 Int = Intrinsic::aarch64_neon_uaddlv; 5839 Ty = Int32Ty; 5840 VTy = llvm::VectorType::get(Int16Ty, 8); 5841 llvm::Type *Tys[2] = { Ty, VTy }; 5842 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5843 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv"); 5844 } 5845 case NEON::BI__builtin_neon_vaddlv_s8: { 5846 Int = Intrinsic::aarch64_neon_saddlv; 5847 Ty = Int32Ty; 5848 VTy = llvm::VectorType::get(Int8Ty, 8); 5849 llvm::Type *Tys[2] = { Ty, VTy }; 5850 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5851 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv"); 5852 return Builder.CreateTrunc(Ops[0], Int16Ty); 5853 } 5854 case NEON::BI__builtin_neon_vaddlv_s16: { 5855 Int = Intrinsic::aarch64_neon_saddlv; 5856 Ty = Int32Ty; 5857 VTy = llvm::VectorType::get(Int16Ty, 4); 5858 llvm::Type *Tys[2] = { Ty, VTy }; 5859 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5860 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv"); 5861 } 5862 case NEON::BI__builtin_neon_vaddlvq_s8: { 5863 Int = Intrinsic::aarch64_neon_saddlv; 5864 Ty = Int32Ty; 5865 VTy = llvm::VectorType::get(Int8Ty, 16); 5866 llvm::Type *Tys[2] = { Ty, VTy }; 5867 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5868 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv"); 5869 return Builder.CreateTrunc(Ops[0], Int16Ty); 5870 } 5871 case NEON::BI__builtin_neon_vaddlvq_s16: { 5872 Int = Intrinsic::aarch64_neon_saddlv; 5873 Ty = Int32Ty; 5874 VTy = llvm::VectorType::get(Int16Ty, 8); 5875 llvm::Type *Tys[2] = { Ty, VTy }; 5876 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5877 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv"); 5878 } 5879 case NEON::BI__builtin_neon_vsri_n_v: 5880 case NEON::BI__builtin_neon_vsriq_n_v: { 5881 Int = Intrinsic::aarch64_neon_vsri; 5882 llvm::Function *Intrin = CGM.getIntrinsic(Int, Ty); 5883 return EmitNeonCall(Intrin, Ops, "vsri_n"); 5884 } 5885 case NEON::BI__builtin_neon_vsli_n_v: 5886 case NEON::BI__builtin_neon_vsliq_n_v: { 5887 Int = Intrinsic::aarch64_neon_vsli; 5888 llvm::Function *Intrin = CGM.getIntrinsic(Int, Ty); 5889 return EmitNeonCall(Intrin, Ops, "vsli_n"); 5890 } 5891 case NEON::BI__builtin_neon_vsra_n_v: 5892 case NEON::BI__builtin_neon_vsraq_n_v: 5893 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 5894 Ops[1] = EmitNeonRShiftImm(Ops[1], Ops[2], Ty, usgn, "vsra_n"); 5895 return Builder.CreateAdd(Ops[0], Ops[1]); 5896 case NEON::BI__builtin_neon_vrsra_n_v: 5897 case NEON::BI__builtin_neon_vrsraq_n_v: { 5898 Int = usgn ? Intrinsic::aarch64_neon_urshl : Intrinsic::aarch64_neon_srshl; 5899 SmallVector<llvm::Value*,2> TmpOps; 5900 TmpOps.push_back(Ops[1]); 5901 TmpOps.push_back(Ops[2]); 5902 Function* F = CGM.getIntrinsic(Int, Ty); 5903 llvm::Value *tmp = EmitNeonCall(F, TmpOps, "vrshr_n", 1, true); 5904 Ops[0] = Builder.CreateBitCast(Ops[0], VTy); 5905 return Builder.CreateAdd(Ops[0], tmp); 5906 } 5907 // FIXME: Sharing loads & stores with 32-bit is complicated by the absence 5908 // of an Align parameter here. 5909 case NEON::BI__builtin_neon_vld1_x2_v: 5910 case NEON::BI__builtin_neon_vld1q_x2_v: 5911 case NEON::BI__builtin_neon_vld1_x3_v: 5912 case NEON::BI__builtin_neon_vld1q_x3_v: 5913 case NEON::BI__builtin_neon_vld1_x4_v: 5914 case NEON::BI__builtin_neon_vld1q_x4_v: { 5915 llvm::Type *PTy = llvm::PointerType::getUnqual(VTy->getVectorElementType()); 5916 Ops[1] = Builder.CreateBitCast(Ops[1], PTy); 5917 llvm::Type *Tys[2] = { VTy, PTy }; 5918 unsigned Int; 5919 switch (BuiltinID) { 5920 case NEON::BI__builtin_neon_vld1_x2_v: 5921 case NEON::BI__builtin_neon_vld1q_x2_v: 5922 Int = Intrinsic::aarch64_neon_ld1x2; 5923 break; 5924 case NEON::BI__builtin_neon_vld1_x3_v: 5925 case NEON::BI__builtin_neon_vld1q_x3_v: 5926 Int = Intrinsic::aarch64_neon_ld1x3; 5927 break; 5928 case NEON::BI__builtin_neon_vld1_x4_v: 5929 case NEON::BI__builtin_neon_vld1q_x4_v: 5930 Int = Intrinsic::aarch64_neon_ld1x4; 5931 break; 5932 } 5933 Function *F = CGM.getIntrinsic(Int, Tys); 5934 Ops[1] = Builder.CreateCall(F, Ops[1], "vld1xN"); 5935 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 5936 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 5937 return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]); 5938 } 5939 case NEON::BI__builtin_neon_vst1_x2_v: 5940 case NEON::BI__builtin_neon_vst1q_x2_v: 5941 case NEON::BI__builtin_neon_vst1_x3_v: 5942 case NEON::BI__builtin_neon_vst1q_x3_v: 5943 case NEON::BI__builtin_neon_vst1_x4_v: 5944 case NEON::BI__builtin_neon_vst1q_x4_v: { 5945 llvm::Type *PTy = llvm::PointerType::getUnqual(VTy->getVectorElementType()); 5946 llvm::Type *Tys[2] = { VTy, PTy }; 5947 unsigned Int; 5948 switch (BuiltinID) { 5949 case NEON::BI__builtin_neon_vst1_x2_v: 5950 case NEON::BI__builtin_neon_vst1q_x2_v: 5951 Int = Intrinsic::aarch64_neon_st1x2; 5952 break; 5953 case NEON::BI__builtin_neon_vst1_x3_v: 5954 case NEON::BI__builtin_neon_vst1q_x3_v: 5955 Int = Intrinsic::aarch64_neon_st1x3; 5956 break; 5957 case NEON::BI__builtin_neon_vst1_x4_v: 5958 case NEON::BI__builtin_neon_vst1q_x4_v: 5959 Int = Intrinsic::aarch64_neon_st1x4; 5960 break; 5961 } 5962 std::rotate(Ops.begin(), Ops.begin() + 1, Ops.end()); 5963 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, ""); 5964 } 5965 case NEON::BI__builtin_neon_vld1_v: 5966 case NEON::BI__builtin_neon_vld1q_v: 5967 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(VTy)); 5968 return Builder.CreateDefaultAlignedLoad(Ops[0]); 5969 case NEON::BI__builtin_neon_vst1_v: 5970 case NEON::BI__builtin_neon_vst1q_v: 5971 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(VTy)); 5972 Ops[1] = Builder.CreateBitCast(Ops[1], VTy); 5973 return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]); 5974 case NEON::BI__builtin_neon_vld1_lane_v: 5975 case NEON::BI__builtin_neon_vld1q_lane_v: 5976 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 5977 Ty = llvm::PointerType::getUnqual(VTy->getElementType()); 5978 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 5979 Ops[0] = Builder.CreateDefaultAlignedLoad(Ops[0]); 5980 return Builder.CreateInsertElement(Ops[1], Ops[0], Ops[2], "vld1_lane"); 5981 case NEON::BI__builtin_neon_vld1_dup_v: 5982 case NEON::BI__builtin_neon_vld1q_dup_v: { 5983 Value *V = UndefValue::get(Ty); 5984 Ty = llvm::PointerType::getUnqual(VTy->getElementType()); 5985 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 5986 Ops[0] = Builder.CreateDefaultAlignedLoad(Ops[0]); 5987 llvm::Constant *CI = ConstantInt::get(Int32Ty, 0); 5988 Ops[0] = Builder.CreateInsertElement(V, Ops[0], CI); 5989 return EmitNeonSplat(Ops[0], CI); 5990 } 5991 case NEON::BI__builtin_neon_vst1_lane_v: 5992 case NEON::BI__builtin_neon_vst1q_lane_v: 5993 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 5994 Ops[1] = Builder.CreateExtractElement(Ops[1], Ops[2]); 5995 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 5996 return Builder.CreateDefaultAlignedStore(Ops[1], 5997 Builder.CreateBitCast(Ops[0], Ty)); 5998 case NEON::BI__builtin_neon_vld2_v: 5999 case NEON::BI__builtin_neon_vld2q_v: { 6000 llvm::Type *PTy = llvm::PointerType::getUnqual(VTy); 6001 Ops[1] = Builder.CreateBitCast(Ops[1], PTy); 6002 llvm::Type *Tys[2] = { VTy, PTy }; 6003 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld2, Tys); 6004 Ops[1] = Builder.CreateCall(F, Ops[1], "vld2"); 6005 Ops[0] = Builder.CreateBitCast(Ops[0], 6006 llvm::PointerType::getUnqual(Ops[1]->getType())); 6007 return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]); 6008 } 6009 case NEON::BI__builtin_neon_vld3_v: 6010 case NEON::BI__builtin_neon_vld3q_v: { 6011 llvm::Type *PTy = llvm::PointerType::getUnqual(VTy); 6012 Ops[1] = Builder.CreateBitCast(Ops[1], PTy); 6013 llvm::Type *Tys[2] = { VTy, PTy }; 6014 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld3, Tys); 6015 Ops[1] = Builder.CreateCall(F, Ops[1], "vld3"); 6016 Ops[0] = Builder.CreateBitCast(Ops[0], 6017 llvm::PointerType::getUnqual(Ops[1]->getType())); 6018 return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]); 6019 } 6020 case NEON::BI__builtin_neon_vld4_v: 6021 case NEON::BI__builtin_neon_vld4q_v: { 6022 llvm::Type *PTy = llvm::PointerType::getUnqual(VTy); 6023 Ops[1] = Builder.CreateBitCast(Ops[1], PTy); 6024 llvm::Type *Tys[2] = { VTy, PTy }; 6025 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld4, Tys); 6026 Ops[1] = Builder.CreateCall(F, Ops[1], "vld4"); 6027 Ops[0] = Builder.CreateBitCast(Ops[0], 6028 llvm::PointerType::getUnqual(Ops[1]->getType())); 6029 return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]); 6030 } 6031 case NEON::BI__builtin_neon_vld2_dup_v: 6032 case NEON::BI__builtin_neon_vld2q_dup_v: { 6033 llvm::Type *PTy = 6034 llvm::PointerType::getUnqual(VTy->getElementType()); 6035 Ops[1] = Builder.CreateBitCast(Ops[1], PTy); 6036 llvm::Type *Tys[2] = { VTy, PTy }; 6037 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld2r, Tys); 6038 Ops[1] = Builder.CreateCall(F, Ops[1], "vld2"); 6039 Ops[0] = Builder.CreateBitCast(Ops[0], 6040 llvm::PointerType::getUnqual(Ops[1]->getType())); 6041 return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]); 6042 } 6043 case NEON::BI__builtin_neon_vld3_dup_v: 6044 case NEON::BI__builtin_neon_vld3q_dup_v: { 6045 llvm::Type *PTy = 6046 llvm::PointerType::getUnqual(VTy->getElementType()); 6047 Ops[1] = Builder.CreateBitCast(Ops[1], PTy); 6048 llvm::Type *Tys[2] = { VTy, PTy }; 6049 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld3r, Tys); 6050 Ops[1] = Builder.CreateCall(F, Ops[1], "vld3"); 6051 Ops[0] = Builder.CreateBitCast(Ops[0], 6052 llvm::PointerType::getUnqual(Ops[1]->getType())); 6053 return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]); 6054 } 6055 case NEON::BI__builtin_neon_vld4_dup_v: 6056 case NEON::BI__builtin_neon_vld4q_dup_v: { 6057 llvm::Type *PTy = 6058 llvm::PointerType::getUnqual(VTy->getElementType()); 6059 Ops[1] = Builder.CreateBitCast(Ops[1], PTy); 6060 llvm::Type *Tys[2] = { VTy, PTy }; 6061 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld4r, Tys); 6062 Ops[1] = Builder.CreateCall(F, Ops[1], "vld4"); 6063 Ops[0] = Builder.CreateBitCast(Ops[0], 6064 llvm::PointerType::getUnqual(Ops[1]->getType())); 6065 return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]); 6066 } 6067 case NEON::BI__builtin_neon_vld2_lane_v: 6068 case NEON::BI__builtin_neon_vld2q_lane_v: { 6069 llvm::Type *Tys[2] = { VTy, Ops[1]->getType() }; 6070 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld2lane, Tys); 6071 Ops.push_back(Ops[1]); 6072 Ops.erase(Ops.begin()+1); 6073 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 6074 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 6075 Ops[3] = Builder.CreateZExt(Ops[3], Int64Ty); 6076 Ops[1] = Builder.CreateCall(F, makeArrayRef(Ops).slice(1), "vld2_lane"); 6077 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 6078 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 6079 return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]); 6080 } 6081 case NEON::BI__builtin_neon_vld3_lane_v: 6082 case NEON::BI__builtin_neon_vld3q_lane_v: { 6083 llvm::Type *Tys[2] = { VTy, Ops[1]->getType() }; 6084 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld3lane, Tys); 6085 Ops.push_back(Ops[1]); 6086 Ops.erase(Ops.begin()+1); 6087 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 6088 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 6089 Ops[3] = Builder.CreateBitCast(Ops[3], Ty); 6090 Ops[4] = Builder.CreateZExt(Ops[4], Int64Ty); 6091 Ops[1] = Builder.CreateCall(F, makeArrayRef(Ops).slice(1), "vld3_lane"); 6092 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 6093 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 6094 return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]); 6095 } 6096 case NEON::BI__builtin_neon_vld4_lane_v: 6097 case NEON::BI__builtin_neon_vld4q_lane_v: { 6098 llvm::Type *Tys[2] = { VTy, Ops[1]->getType() }; 6099 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld4lane, Tys); 6100 Ops.push_back(Ops[1]); 6101 Ops.erase(Ops.begin()+1); 6102 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 6103 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 6104 Ops[3] = Builder.CreateBitCast(Ops[3], Ty); 6105 Ops[4] = Builder.CreateBitCast(Ops[4], Ty); 6106 Ops[5] = Builder.CreateZExt(Ops[5], Int64Ty); 6107 Ops[1] = Builder.CreateCall(F, makeArrayRef(Ops).slice(1), "vld4_lane"); 6108 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 6109 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 6110 return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]); 6111 } 6112 case NEON::BI__builtin_neon_vst2_v: 6113 case NEON::BI__builtin_neon_vst2q_v: { 6114 Ops.push_back(Ops[0]); 6115 Ops.erase(Ops.begin()); 6116 llvm::Type *Tys[2] = { VTy, Ops[2]->getType() }; 6117 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_st2, Tys), 6118 Ops, ""); 6119 } 6120 case NEON::BI__builtin_neon_vst2_lane_v: 6121 case NEON::BI__builtin_neon_vst2q_lane_v: { 6122 Ops.push_back(Ops[0]); 6123 Ops.erase(Ops.begin()); 6124 Ops[2] = Builder.CreateZExt(Ops[2], Int64Ty); 6125 llvm::Type *Tys[2] = { VTy, Ops[3]->getType() }; 6126 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_st2lane, Tys), 6127 Ops, ""); 6128 } 6129 case NEON::BI__builtin_neon_vst3_v: 6130 case NEON::BI__builtin_neon_vst3q_v: { 6131 Ops.push_back(Ops[0]); 6132 Ops.erase(Ops.begin()); 6133 llvm::Type *Tys[2] = { VTy, Ops[3]->getType() }; 6134 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_st3, Tys), 6135 Ops, ""); 6136 } 6137 case NEON::BI__builtin_neon_vst3_lane_v: 6138 case NEON::BI__builtin_neon_vst3q_lane_v: { 6139 Ops.push_back(Ops[0]); 6140 Ops.erase(Ops.begin()); 6141 Ops[3] = Builder.CreateZExt(Ops[3], Int64Ty); 6142 llvm::Type *Tys[2] = { VTy, Ops[4]->getType() }; 6143 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_st3lane, Tys), 6144 Ops, ""); 6145 } 6146 case NEON::BI__builtin_neon_vst4_v: 6147 case NEON::BI__builtin_neon_vst4q_v: { 6148 Ops.push_back(Ops[0]); 6149 Ops.erase(Ops.begin()); 6150 llvm::Type *Tys[2] = { VTy, Ops[4]->getType() }; 6151 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_st4, Tys), 6152 Ops, ""); 6153 } 6154 case NEON::BI__builtin_neon_vst4_lane_v: 6155 case NEON::BI__builtin_neon_vst4q_lane_v: { 6156 Ops.push_back(Ops[0]); 6157 Ops.erase(Ops.begin()); 6158 Ops[4] = Builder.CreateZExt(Ops[4], Int64Ty); 6159 llvm::Type *Tys[2] = { VTy, Ops[5]->getType() }; 6160 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_st4lane, Tys), 6161 Ops, ""); 6162 } 6163 case NEON::BI__builtin_neon_vtrn_v: 6164 case NEON::BI__builtin_neon_vtrnq_v: { 6165 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty)); 6166 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 6167 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 6168 Value *SV = nullptr; 6169 6170 for (unsigned vi = 0; vi != 2; ++vi) { 6171 SmallVector<Constant*, 16> Indices; 6172 for (unsigned i = 0, e = VTy->getNumElements(); i != e; i += 2) { 6173 Indices.push_back(ConstantInt::get(Int32Ty, i+vi)); 6174 Indices.push_back(ConstantInt::get(Int32Ty, i+e+vi)); 6175 } 6176 Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ty, Ops[0], vi); 6177 SV = llvm::ConstantVector::get(Indices); 6178 SV = Builder.CreateShuffleVector(Ops[1], Ops[2], SV, "vtrn"); 6179 SV = Builder.CreateDefaultAlignedStore(SV, Addr); 6180 } 6181 return SV; 6182 } 6183 case NEON::BI__builtin_neon_vuzp_v: 6184 case NEON::BI__builtin_neon_vuzpq_v: { 6185 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty)); 6186 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 6187 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 6188 Value *SV = nullptr; 6189 6190 for (unsigned vi = 0; vi != 2; ++vi) { 6191 SmallVector<Constant*, 16> Indices; 6192 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) 6193 Indices.push_back(ConstantInt::get(Int32Ty, 2*i+vi)); 6194 6195 Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ty, Ops[0], vi); 6196 SV = llvm::ConstantVector::get(Indices); 6197 SV = Builder.CreateShuffleVector(Ops[1], Ops[2], SV, "vuzp"); 6198 SV = Builder.CreateDefaultAlignedStore(SV, Addr); 6199 } 6200 return SV; 6201 } 6202 case NEON::BI__builtin_neon_vzip_v: 6203 case NEON::BI__builtin_neon_vzipq_v: { 6204 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty)); 6205 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 6206 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 6207 Value *SV = nullptr; 6208 6209 for (unsigned vi = 0; vi != 2; ++vi) { 6210 SmallVector<Constant*, 16> Indices; 6211 for (unsigned i = 0, e = VTy->getNumElements(); i != e; i += 2) { 6212 Indices.push_back(ConstantInt::get(Int32Ty, (i + vi*e) >> 1)); 6213 Indices.push_back(ConstantInt::get(Int32Ty, ((i + vi*e) >> 1)+e)); 6214 } 6215 Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ty, Ops[0], vi); 6216 SV = llvm::ConstantVector::get(Indices); 6217 SV = Builder.CreateShuffleVector(Ops[1], Ops[2], SV, "vzip"); 6218 SV = Builder.CreateDefaultAlignedStore(SV, Addr); 6219 } 6220 return SV; 6221 } 6222 case NEON::BI__builtin_neon_vqtbl1q_v: { 6223 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbl1, Ty), 6224 Ops, "vtbl1"); 6225 } 6226 case NEON::BI__builtin_neon_vqtbl2q_v: { 6227 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbl2, Ty), 6228 Ops, "vtbl2"); 6229 } 6230 case NEON::BI__builtin_neon_vqtbl3q_v: { 6231 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbl3, Ty), 6232 Ops, "vtbl3"); 6233 } 6234 case NEON::BI__builtin_neon_vqtbl4q_v: { 6235 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbl4, Ty), 6236 Ops, "vtbl4"); 6237 } 6238 case NEON::BI__builtin_neon_vqtbx1q_v: { 6239 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbx1, Ty), 6240 Ops, "vtbx1"); 6241 } 6242 case NEON::BI__builtin_neon_vqtbx2q_v: { 6243 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbx2, Ty), 6244 Ops, "vtbx2"); 6245 } 6246 case NEON::BI__builtin_neon_vqtbx3q_v: { 6247 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbx3, Ty), 6248 Ops, "vtbx3"); 6249 } 6250 case NEON::BI__builtin_neon_vqtbx4q_v: { 6251 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbx4, Ty), 6252 Ops, "vtbx4"); 6253 } 6254 case NEON::BI__builtin_neon_vsqadd_v: 6255 case NEON::BI__builtin_neon_vsqaddq_v: { 6256 Int = Intrinsic::aarch64_neon_usqadd; 6257 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vsqadd"); 6258 } 6259 case NEON::BI__builtin_neon_vuqadd_v: 6260 case NEON::BI__builtin_neon_vuqaddq_v: { 6261 Int = Intrinsic::aarch64_neon_suqadd; 6262 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vuqadd"); 6263 } 6264 } 6265 } 6266 6267 llvm::Value *CodeGenFunction:: 6268 BuildVector(ArrayRef<llvm::Value*> Ops) { 6269 assert((Ops.size() & (Ops.size() - 1)) == 0 && 6270 "Not a power-of-two sized vector!"); 6271 bool AllConstants = true; 6272 for (unsigned i = 0, e = Ops.size(); i != e && AllConstants; ++i) 6273 AllConstants &= isa<Constant>(Ops[i]); 6274 6275 // If this is a constant vector, create a ConstantVector. 6276 if (AllConstants) { 6277 SmallVector<llvm::Constant*, 16> CstOps; 6278 for (unsigned i = 0, e = Ops.size(); i != e; ++i) 6279 CstOps.push_back(cast<Constant>(Ops[i])); 6280 return llvm::ConstantVector::get(CstOps); 6281 } 6282 6283 // Otherwise, insertelement the values to build the vector. 6284 Value *Result = 6285 llvm::UndefValue::get(llvm::VectorType::get(Ops[0]->getType(), Ops.size())); 6286 6287 for (unsigned i = 0, e = Ops.size(); i != e; ++i) 6288 Result = Builder.CreateInsertElement(Result, Ops[i], Builder.getInt32(i)); 6289 6290 return Result; 6291 } 6292 6293 Value *CodeGenFunction::EmitX86BuiltinExpr(unsigned BuiltinID, 6294 const CallExpr *E) { 6295 if (BuiltinID == X86::BI__builtin_ms_va_start || 6296 BuiltinID == X86::BI__builtin_ms_va_end) 6297 return EmitVAStartEnd(EmitMSVAListRef(E->getArg(0)).getPointer(), 6298 BuiltinID == X86::BI__builtin_ms_va_start); 6299 if (BuiltinID == X86::BI__builtin_ms_va_copy) { 6300 // Lower this manually. We can't reliably determine whether or not any 6301 // given va_copy() is for a Win64 va_list from the calling convention 6302 // alone, because it's legal to do this from a System V ABI function. 6303 // With opaque pointer types, we won't have enough information in LLVM 6304 // IR to determine this from the argument types, either. Best to do it 6305 // now, while we have enough information. 6306 Address DestAddr = EmitMSVAListRef(E->getArg(0)); 6307 Address SrcAddr = EmitMSVAListRef(E->getArg(1)); 6308 6309 llvm::Type *BPP = Int8PtrPtrTy; 6310 6311 DestAddr = Address(Builder.CreateBitCast(DestAddr.getPointer(), BPP, "cp"), 6312 DestAddr.getAlignment()); 6313 SrcAddr = Address(Builder.CreateBitCast(SrcAddr.getPointer(), BPP, "ap"), 6314 SrcAddr.getAlignment()); 6315 6316 Value *ArgPtr = Builder.CreateLoad(SrcAddr, "ap.val"); 6317 return Builder.CreateStore(ArgPtr, DestAddr); 6318 } 6319 6320 SmallVector<Value*, 4> Ops; 6321 6322 // Find out if any arguments are required to be integer constant expressions. 6323 unsigned ICEArguments = 0; 6324 ASTContext::GetBuiltinTypeError Error; 6325 getContext().GetBuiltinType(BuiltinID, Error, &ICEArguments); 6326 assert(Error == ASTContext::GE_None && "Should not codegen an error"); 6327 6328 for (unsigned i = 0, e = E->getNumArgs(); i != e; i++) { 6329 // If this is a normal argument, just emit it as a scalar. 6330 if ((ICEArguments & (1 << i)) == 0) { 6331 Ops.push_back(EmitScalarExpr(E->getArg(i))); 6332 continue; 6333 } 6334 6335 // If this is required to be a constant, constant fold it so that we know 6336 // that the generated intrinsic gets a ConstantInt. 6337 llvm::APSInt Result; 6338 bool IsConst = E->getArg(i)->isIntegerConstantExpr(Result, getContext()); 6339 assert(IsConst && "Constant arg isn't actually constant?"); (void)IsConst; 6340 Ops.push_back(llvm::ConstantInt::get(getLLVMContext(), Result)); 6341 } 6342 6343 switch (BuiltinID) { 6344 default: return nullptr; 6345 case X86::BI__builtin_cpu_supports: { 6346 const Expr *FeatureExpr = E->getArg(0)->IgnoreParenCasts(); 6347 StringRef FeatureStr = cast<StringLiteral>(FeatureExpr)->getString(); 6348 6349 // TODO: When/if this becomes more than x86 specific then use a TargetInfo 6350 // based mapping. 6351 // Processor features and mapping to processor feature value. 6352 enum X86Features { 6353 CMOV = 0, 6354 MMX, 6355 POPCNT, 6356 SSE, 6357 SSE2, 6358 SSE3, 6359 SSSE3, 6360 SSE4_1, 6361 SSE4_2, 6362 AVX, 6363 AVX2, 6364 SSE4_A, 6365 FMA4, 6366 XOP, 6367 FMA, 6368 AVX512F, 6369 BMI, 6370 BMI2, 6371 MAX 6372 }; 6373 6374 X86Features Feature = StringSwitch<X86Features>(FeatureStr) 6375 .Case("cmov", X86Features::CMOV) 6376 .Case("mmx", X86Features::MMX) 6377 .Case("popcnt", X86Features::POPCNT) 6378 .Case("sse", X86Features::SSE) 6379 .Case("sse2", X86Features::SSE2) 6380 .Case("sse3", X86Features::SSE3) 6381 .Case("sse4.1", X86Features::SSE4_1) 6382 .Case("sse4.2", X86Features::SSE4_2) 6383 .Case("avx", X86Features::AVX) 6384 .Case("avx2", X86Features::AVX2) 6385 .Case("sse4a", X86Features::SSE4_A) 6386 .Case("fma4", X86Features::FMA4) 6387 .Case("xop", X86Features::XOP) 6388 .Case("fma", X86Features::FMA) 6389 .Case("avx512f", X86Features::AVX512F) 6390 .Case("bmi", X86Features::BMI) 6391 .Case("bmi2", X86Features::BMI2) 6392 .Default(X86Features::MAX); 6393 assert(Feature != X86Features::MAX && "Invalid feature!"); 6394 6395 // Matching the struct layout from the compiler-rt/libgcc structure that is 6396 // filled in: 6397 // unsigned int __cpu_vendor; 6398 // unsigned int __cpu_type; 6399 // unsigned int __cpu_subtype; 6400 // unsigned int __cpu_features[1]; 6401 llvm::Type *STy = llvm::StructType::get( 6402 Int32Ty, Int32Ty, Int32Ty, llvm::ArrayType::get(Int32Ty, 1), nullptr); 6403 6404 // Grab the global __cpu_model. 6405 llvm::Constant *CpuModel = CGM.CreateRuntimeVariable(STy, "__cpu_model"); 6406 6407 // Grab the first (0th) element from the field __cpu_features off of the 6408 // global in the struct STy. 6409 Value *Idxs[] = { 6410 ConstantInt::get(Int32Ty, 0), 6411 ConstantInt::get(Int32Ty, 3), 6412 ConstantInt::get(Int32Ty, 0) 6413 }; 6414 Value *CpuFeatures = Builder.CreateGEP(STy, CpuModel, Idxs); 6415 Value *Features = Builder.CreateAlignedLoad(CpuFeatures, 6416 CharUnits::fromQuantity(4)); 6417 6418 // Check the value of the bit corresponding to the feature requested. 6419 Value *Bitset = Builder.CreateAnd( 6420 Features, llvm::ConstantInt::get(Int32Ty, 1 << Feature)); 6421 return Builder.CreateICmpNE(Bitset, llvm::ConstantInt::get(Int32Ty, 0)); 6422 } 6423 case X86::BI_mm_prefetch: { 6424 Value *Address = Ops[0]; 6425 Value *RW = ConstantInt::get(Int32Ty, 0); 6426 Value *Locality = Ops[1]; 6427 Value *Data = ConstantInt::get(Int32Ty, 1); 6428 Value *F = CGM.getIntrinsic(Intrinsic::prefetch); 6429 return Builder.CreateCall(F, {Address, RW, Locality, Data}); 6430 } 6431 case X86::BI__builtin_ia32_undef128: 6432 case X86::BI__builtin_ia32_undef256: 6433 case X86::BI__builtin_ia32_undef512: 6434 return UndefValue::get(ConvertType(E->getType())); 6435 case X86::BI__builtin_ia32_vec_init_v8qi: 6436 case X86::BI__builtin_ia32_vec_init_v4hi: 6437 case X86::BI__builtin_ia32_vec_init_v2si: 6438 return Builder.CreateBitCast(BuildVector(Ops), 6439 llvm::Type::getX86_MMXTy(getLLVMContext())); 6440 case X86::BI__builtin_ia32_vec_ext_v2si: 6441 return Builder.CreateExtractElement(Ops[0], 6442 llvm::ConstantInt::get(Ops[1]->getType(), 0)); 6443 case X86::BI__builtin_ia32_ldmxcsr: { 6444 Address Tmp = CreateMemTemp(E->getArg(0)->getType()); 6445 Builder.CreateStore(Ops[0], Tmp); 6446 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::x86_sse_ldmxcsr), 6447 Builder.CreateBitCast(Tmp.getPointer(), Int8PtrTy)); 6448 } 6449 case X86::BI__builtin_ia32_stmxcsr: { 6450 Address Tmp = CreateMemTemp(E->getType()); 6451 Builder.CreateCall(CGM.getIntrinsic(Intrinsic::x86_sse_stmxcsr), 6452 Builder.CreateBitCast(Tmp.getPointer(), Int8PtrTy)); 6453 return Builder.CreateLoad(Tmp, "stmxcsr"); 6454 } 6455 case X86::BI__builtin_ia32_xsave: 6456 case X86::BI__builtin_ia32_xsave64: 6457 case X86::BI__builtin_ia32_xrstor: 6458 case X86::BI__builtin_ia32_xrstor64: 6459 case X86::BI__builtin_ia32_xsaveopt: 6460 case X86::BI__builtin_ia32_xsaveopt64: 6461 case X86::BI__builtin_ia32_xrstors: 6462 case X86::BI__builtin_ia32_xrstors64: 6463 case X86::BI__builtin_ia32_xsavec: 6464 case X86::BI__builtin_ia32_xsavec64: 6465 case X86::BI__builtin_ia32_xsaves: 6466 case X86::BI__builtin_ia32_xsaves64: { 6467 Intrinsic::ID ID; 6468 #define INTRINSIC_X86_XSAVE_ID(NAME) \ 6469 case X86::BI__builtin_ia32_##NAME: \ 6470 ID = Intrinsic::x86_##NAME; \ 6471 break 6472 switch (BuiltinID) { 6473 default: llvm_unreachable("Unsupported intrinsic!"); 6474 INTRINSIC_X86_XSAVE_ID(xsave); 6475 INTRINSIC_X86_XSAVE_ID(xsave64); 6476 INTRINSIC_X86_XSAVE_ID(xrstor); 6477 INTRINSIC_X86_XSAVE_ID(xrstor64); 6478 INTRINSIC_X86_XSAVE_ID(xsaveopt); 6479 INTRINSIC_X86_XSAVE_ID(xsaveopt64); 6480 INTRINSIC_X86_XSAVE_ID(xrstors); 6481 INTRINSIC_X86_XSAVE_ID(xrstors64); 6482 INTRINSIC_X86_XSAVE_ID(xsavec); 6483 INTRINSIC_X86_XSAVE_ID(xsavec64); 6484 INTRINSIC_X86_XSAVE_ID(xsaves); 6485 INTRINSIC_X86_XSAVE_ID(xsaves64); 6486 } 6487 #undef INTRINSIC_X86_XSAVE_ID 6488 Value *Mhi = Builder.CreateTrunc( 6489 Builder.CreateLShr(Ops[1], ConstantInt::get(Int64Ty, 32)), Int32Ty); 6490 Value *Mlo = Builder.CreateTrunc(Ops[1], Int32Ty); 6491 Ops[1] = Mhi; 6492 Ops.push_back(Mlo); 6493 return Builder.CreateCall(CGM.getIntrinsic(ID), Ops); 6494 } 6495 case X86::BI__builtin_ia32_storehps: 6496 case X86::BI__builtin_ia32_storelps: { 6497 llvm::Type *PtrTy = llvm::PointerType::getUnqual(Int64Ty); 6498 llvm::Type *VecTy = llvm::VectorType::get(Int64Ty, 2); 6499 6500 // cast val v2i64 6501 Ops[1] = Builder.CreateBitCast(Ops[1], VecTy, "cast"); 6502 6503 // extract (0, 1) 6504 unsigned Index = BuiltinID == X86::BI__builtin_ia32_storelps ? 0 : 1; 6505 llvm::Value *Idx = llvm::ConstantInt::get(SizeTy, Index); 6506 Ops[1] = Builder.CreateExtractElement(Ops[1], Idx, "extract"); 6507 6508 // cast pointer to i64 & store 6509 Ops[0] = Builder.CreateBitCast(Ops[0], PtrTy); 6510 return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]); 6511 } 6512 case X86::BI__builtin_ia32_palignr128: 6513 case X86::BI__builtin_ia32_palignr256: { 6514 unsigned ShiftVal = cast<llvm::ConstantInt>(Ops[2])->getZExtValue(); 6515 6516 unsigned NumElts = 6517 cast<llvm::VectorType>(Ops[0]->getType())->getNumElements(); 6518 assert(NumElts % 16 == 0); 6519 unsigned NumLanes = NumElts / 16; 6520 unsigned NumLaneElts = NumElts / NumLanes; 6521 6522 // If palignr is shifting the pair of vectors more than the size of two 6523 // lanes, emit zero. 6524 if (ShiftVal >= (2 * NumLaneElts)) 6525 return llvm::Constant::getNullValue(ConvertType(E->getType())); 6526 6527 // If palignr is shifting the pair of input vectors more than one lane, 6528 // but less than two lanes, convert to shifting in zeroes. 6529 if (ShiftVal > NumLaneElts) { 6530 ShiftVal -= NumLaneElts; 6531 Ops[1] = Ops[0]; 6532 Ops[0] = llvm::Constant::getNullValue(Ops[0]->getType()); 6533 } 6534 6535 uint32_t Indices[32]; 6536 // 256-bit palignr operates on 128-bit lanes so we need to handle that 6537 for (unsigned l = 0; l != NumElts; l += NumLaneElts) { 6538 for (unsigned i = 0; i != NumLaneElts; ++i) { 6539 unsigned Idx = ShiftVal + i; 6540 if (Idx >= NumLaneElts) 6541 Idx += NumElts - NumLaneElts; // End of lane, switch operand. 6542 Indices[l + i] = Idx + l; 6543 } 6544 } 6545 6546 Value *SV = llvm::ConstantDataVector::get(getLLVMContext(), 6547 makeArrayRef(Indices, NumElts)); 6548 return Builder.CreateShuffleVector(Ops[1], Ops[0], SV, "palignr"); 6549 } 6550 case X86::BI__builtin_ia32_pslldqi256: { 6551 // Shift value is in bits so divide by 8. 6552 unsigned shiftVal = cast<llvm::ConstantInt>(Ops[1])->getZExtValue() >> 3; 6553 6554 // If pslldq is shifting the vector more than 15 bytes, emit zero. 6555 if (shiftVal >= 16) 6556 return llvm::Constant::getNullValue(ConvertType(E->getType())); 6557 6558 uint32_t Indices[32]; 6559 // 256-bit pslldq operates on 128-bit lanes so we need to handle that 6560 for (unsigned l = 0; l != 32; l += 16) { 6561 for (unsigned i = 0; i != 16; ++i) { 6562 unsigned Idx = 32 + i - shiftVal; 6563 if (Idx < 32) Idx -= 16; // end of lane, switch operand. 6564 Indices[l + i] = Idx + l; 6565 } 6566 } 6567 6568 llvm::Type *VecTy = llvm::VectorType::get(Int8Ty, 32); 6569 Ops[0] = Builder.CreateBitCast(Ops[0], VecTy, "cast"); 6570 Value *Zero = llvm::Constant::getNullValue(VecTy); 6571 6572 Value *SV = llvm::ConstantDataVector::get(getLLVMContext(), Indices); 6573 SV = Builder.CreateShuffleVector(Zero, Ops[0], SV, "pslldq"); 6574 llvm::Type *ResultType = ConvertType(E->getType()); 6575 return Builder.CreateBitCast(SV, ResultType, "cast"); 6576 } 6577 case X86::BI__builtin_ia32_psrldqi256: { 6578 // Shift value is in bits so divide by 8. 6579 unsigned shiftVal = cast<llvm::ConstantInt>(Ops[1])->getZExtValue() >> 3; 6580 6581 // If psrldq is shifting the vector more than 15 bytes, emit zero. 6582 if (shiftVal >= 16) 6583 return llvm::Constant::getNullValue(ConvertType(E->getType())); 6584 6585 uint32_t Indices[32]; 6586 // 256-bit psrldq operates on 128-bit lanes so we need to handle that 6587 for (unsigned l = 0; l != 32; l += 16) { 6588 for (unsigned i = 0; i != 16; ++i) { 6589 unsigned Idx = i + shiftVal; 6590 if (Idx >= 16) Idx += 16; // end of lane, switch operand. 6591 Indices[l + i] = Idx + l; 6592 } 6593 } 6594 6595 llvm::Type *VecTy = llvm::VectorType::get(Int8Ty, 32); 6596 Ops[0] = Builder.CreateBitCast(Ops[0], VecTy, "cast"); 6597 Value *Zero = llvm::Constant::getNullValue(VecTy); 6598 6599 Value *SV = llvm::ConstantDataVector::get(getLLVMContext(), Indices); 6600 SV = Builder.CreateShuffleVector(Ops[0], Zero, SV, "psrldq"); 6601 llvm::Type *ResultType = ConvertType(E->getType()); 6602 return Builder.CreateBitCast(SV, ResultType, "cast"); 6603 } 6604 case X86::BI__builtin_ia32_movntps: 6605 case X86::BI__builtin_ia32_movntps256: 6606 case X86::BI__builtin_ia32_movntpd: 6607 case X86::BI__builtin_ia32_movntpd256: 6608 case X86::BI__builtin_ia32_movntdq: 6609 case X86::BI__builtin_ia32_movntdq256: 6610 case X86::BI__builtin_ia32_movnti: 6611 case X86::BI__builtin_ia32_movnti64: { 6612 llvm::MDNode *Node = llvm::MDNode::get( 6613 getLLVMContext(), llvm::ConstantAsMetadata::get(Builder.getInt32(1))); 6614 6615 // Convert the type of the pointer to a pointer to the stored type. 6616 Value *BC = Builder.CreateBitCast(Ops[0], 6617 llvm::PointerType::getUnqual(Ops[1]->getType()), 6618 "cast"); 6619 StoreInst *SI = Builder.CreateDefaultAlignedStore(Ops[1], BC); 6620 SI->setMetadata(CGM.getModule().getMDKindID("nontemporal"), Node); 6621 6622 // If the operand is an integer, we can't assume alignment. Otherwise, 6623 // assume natural alignment. 6624 QualType ArgTy = E->getArg(1)->getType(); 6625 unsigned Align; 6626 if (ArgTy->isIntegerType()) 6627 Align = 1; 6628 else 6629 Align = getContext().getTypeSizeInChars(ArgTy).getQuantity(); 6630 SI->setAlignment(Align); 6631 return SI; 6632 } 6633 // 3DNow! 6634 case X86::BI__builtin_ia32_pswapdsf: 6635 case X86::BI__builtin_ia32_pswapdsi: { 6636 llvm::Type *MMXTy = llvm::Type::getX86_MMXTy(getLLVMContext()); 6637 Ops[0] = Builder.CreateBitCast(Ops[0], MMXTy, "cast"); 6638 llvm::Function *F = CGM.getIntrinsic(Intrinsic::x86_3dnowa_pswapd); 6639 return Builder.CreateCall(F, Ops, "pswapd"); 6640 } 6641 case X86::BI__builtin_ia32_rdrand16_step: 6642 case X86::BI__builtin_ia32_rdrand32_step: 6643 case X86::BI__builtin_ia32_rdrand64_step: 6644 case X86::BI__builtin_ia32_rdseed16_step: 6645 case X86::BI__builtin_ia32_rdseed32_step: 6646 case X86::BI__builtin_ia32_rdseed64_step: { 6647 Intrinsic::ID ID; 6648 switch (BuiltinID) { 6649 default: llvm_unreachable("Unsupported intrinsic!"); 6650 case X86::BI__builtin_ia32_rdrand16_step: 6651 ID = Intrinsic::x86_rdrand_16; 6652 break; 6653 case X86::BI__builtin_ia32_rdrand32_step: 6654 ID = Intrinsic::x86_rdrand_32; 6655 break; 6656 case X86::BI__builtin_ia32_rdrand64_step: 6657 ID = Intrinsic::x86_rdrand_64; 6658 break; 6659 case X86::BI__builtin_ia32_rdseed16_step: 6660 ID = Intrinsic::x86_rdseed_16; 6661 break; 6662 case X86::BI__builtin_ia32_rdseed32_step: 6663 ID = Intrinsic::x86_rdseed_32; 6664 break; 6665 case X86::BI__builtin_ia32_rdseed64_step: 6666 ID = Intrinsic::x86_rdseed_64; 6667 break; 6668 } 6669 6670 Value *Call = Builder.CreateCall(CGM.getIntrinsic(ID)); 6671 Builder.CreateDefaultAlignedStore(Builder.CreateExtractValue(Call, 0), 6672 Ops[0]); 6673 return Builder.CreateExtractValue(Call, 1); 6674 } 6675 // SSE comparison intrisics 6676 case X86::BI__builtin_ia32_cmpeqps: 6677 case X86::BI__builtin_ia32_cmpltps: 6678 case X86::BI__builtin_ia32_cmpleps: 6679 case X86::BI__builtin_ia32_cmpunordps: 6680 case X86::BI__builtin_ia32_cmpneqps: 6681 case X86::BI__builtin_ia32_cmpnltps: 6682 case X86::BI__builtin_ia32_cmpnleps: 6683 case X86::BI__builtin_ia32_cmpordps: 6684 case X86::BI__builtin_ia32_cmpeqss: 6685 case X86::BI__builtin_ia32_cmpltss: 6686 case X86::BI__builtin_ia32_cmpless: 6687 case X86::BI__builtin_ia32_cmpunordss: 6688 case X86::BI__builtin_ia32_cmpneqss: 6689 case X86::BI__builtin_ia32_cmpnltss: 6690 case X86::BI__builtin_ia32_cmpnless: 6691 case X86::BI__builtin_ia32_cmpordss: 6692 case X86::BI__builtin_ia32_cmpeqpd: 6693 case X86::BI__builtin_ia32_cmpltpd: 6694 case X86::BI__builtin_ia32_cmplepd: 6695 case X86::BI__builtin_ia32_cmpunordpd: 6696 case X86::BI__builtin_ia32_cmpneqpd: 6697 case X86::BI__builtin_ia32_cmpnltpd: 6698 case X86::BI__builtin_ia32_cmpnlepd: 6699 case X86::BI__builtin_ia32_cmpordpd: 6700 case X86::BI__builtin_ia32_cmpeqsd: 6701 case X86::BI__builtin_ia32_cmpltsd: 6702 case X86::BI__builtin_ia32_cmplesd: 6703 case X86::BI__builtin_ia32_cmpunordsd: 6704 case X86::BI__builtin_ia32_cmpneqsd: 6705 case X86::BI__builtin_ia32_cmpnltsd: 6706 case X86::BI__builtin_ia32_cmpnlesd: 6707 case X86::BI__builtin_ia32_cmpordsd: 6708 // These exist so that the builtin that takes an immediate can be bounds 6709 // checked by clang to avoid passing bad immediates to the backend. Since 6710 // AVX has a larger immediate than SSE we would need separate builtins to 6711 // do the different bounds checking. Rather than create a clang specific 6712 // SSE only builtin, this implements eight separate builtins to match gcc 6713 // implementation. 6714 6715 // Choose the immediate. 6716 unsigned Imm; 6717 switch (BuiltinID) { 6718 default: llvm_unreachable("Unsupported intrinsic!"); 6719 case X86::BI__builtin_ia32_cmpeqps: 6720 case X86::BI__builtin_ia32_cmpeqss: 6721 case X86::BI__builtin_ia32_cmpeqpd: 6722 case X86::BI__builtin_ia32_cmpeqsd: 6723 Imm = 0; 6724 break; 6725 case X86::BI__builtin_ia32_cmpltps: 6726 case X86::BI__builtin_ia32_cmpltss: 6727 case X86::BI__builtin_ia32_cmpltpd: 6728 case X86::BI__builtin_ia32_cmpltsd: 6729 Imm = 1; 6730 break; 6731 case X86::BI__builtin_ia32_cmpleps: 6732 case X86::BI__builtin_ia32_cmpless: 6733 case X86::BI__builtin_ia32_cmplepd: 6734 case X86::BI__builtin_ia32_cmplesd: 6735 Imm = 2; 6736 break; 6737 case X86::BI__builtin_ia32_cmpunordps: 6738 case X86::BI__builtin_ia32_cmpunordss: 6739 case X86::BI__builtin_ia32_cmpunordpd: 6740 case X86::BI__builtin_ia32_cmpunordsd: 6741 Imm = 3; 6742 break; 6743 case X86::BI__builtin_ia32_cmpneqps: 6744 case X86::BI__builtin_ia32_cmpneqss: 6745 case X86::BI__builtin_ia32_cmpneqpd: 6746 case X86::BI__builtin_ia32_cmpneqsd: 6747 Imm = 4; 6748 break; 6749 case X86::BI__builtin_ia32_cmpnltps: 6750 case X86::BI__builtin_ia32_cmpnltss: 6751 case X86::BI__builtin_ia32_cmpnltpd: 6752 case X86::BI__builtin_ia32_cmpnltsd: 6753 Imm = 5; 6754 break; 6755 case X86::BI__builtin_ia32_cmpnleps: 6756 case X86::BI__builtin_ia32_cmpnless: 6757 case X86::BI__builtin_ia32_cmpnlepd: 6758 case X86::BI__builtin_ia32_cmpnlesd: 6759 Imm = 6; 6760 break; 6761 case X86::BI__builtin_ia32_cmpordps: 6762 case X86::BI__builtin_ia32_cmpordss: 6763 case X86::BI__builtin_ia32_cmpordpd: 6764 case X86::BI__builtin_ia32_cmpordsd: 6765 Imm = 7; 6766 break; 6767 } 6768 6769 // Choose the intrinsic ID. 6770 const char *name; 6771 Intrinsic::ID ID; 6772 switch (BuiltinID) { 6773 default: llvm_unreachable("Unsupported intrinsic!"); 6774 case X86::BI__builtin_ia32_cmpeqps: 6775 case X86::BI__builtin_ia32_cmpltps: 6776 case X86::BI__builtin_ia32_cmpleps: 6777 case X86::BI__builtin_ia32_cmpunordps: 6778 case X86::BI__builtin_ia32_cmpneqps: 6779 case X86::BI__builtin_ia32_cmpnltps: 6780 case X86::BI__builtin_ia32_cmpnleps: 6781 case X86::BI__builtin_ia32_cmpordps: 6782 name = "cmpps"; 6783 ID = Intrinsic::x86_sse_cmp_ps; 6784 break; 6785 case X86::BI__builtin_ia32_cmpeqss: 6786 case X86::BI__builtin_ia32_cmpltss: 6787 case X86::BI__builtin_ia32_cmpless: 6788 case X86::BI__builtin_ia32_cmpunordss: 6789 case X86::BI__builtin_ia32_cmpneqss: 6790 case X86::BI__builtin_ia32_cmpnltss: 6791 case X86::BI__builtin_ia32_cmpnless: 6792 case X86::BI__builtin_ia32_cmpordss: 6793 name = "cmpss"; 6794 ID = Intrinsic::x86_sse_cmp_ss; 6795 break; 6796 case X86::BI__builtin_ia32_cmpeqpd: 6797 case X86::BI__builtin_ia32_cmpltpd: 6798 case X86::BI__builtin_ia32_cmplepd: 6799 case X86::BI__builtin_ia32_cmpunordpd: 6800 case X86::BI__builtin_ia32_cmpneqpd: 6801 case X86::BI__builtin_ia32_cmpnltpd: 6802 case X86::BI__builtin_ia32_cmpnlepd: 6803 case X86::BI__builtin_ia32_cmpordpd: 6804 name = "cmppd"; 6805 ID = Intrinsic::x86_sse2_cmp_pd; 6806 break; 6807 case X86::BI__builtin_ia32_cmpeqsd: 6808 case X86::BI__builtin_ia32_cmpltsd: 6809 case X86::BI__builtin_ia32_cmplesd: 6810 case X86::BI__builtin_ia32_cmpunordsd: 6811 case X86::BI__builtin_ia32_cmpneqsd: 6812 case X86::BI__builtin_ia32_cmpnltsd: 6813 case X86::BI__builtin_ia32_cmpnlesd: 6814 case X86::BI__builtin_ia32_cmpordsd: 6815 name = "cmpsd"; 6816 ID = Intrinsic::x86_sse2_cmp_sd; 6817 break; 6818 } 6819 6820 Ops.push_back(llvm::ConstantInt::get(Int8Ty, Imm)); 6821 llvm::Function *F = CGM.getIntrinsic(ID); 6822 return Builder.CreateCall(F, Ops, name); 6823 } 6824 } 6825 6826 6827 Value *CodeGenFunction::EmitPPCBuiltinExpr(unsigned BuiltinID, 6828 const CallExpr *E) { 6829 SmallVector<Value*, 4> Ops; 6830 6831 for (unsigned i = 0, e = E->getNumArgs(); i != e; i++) 6832 Ops.push_back(EmitScalarExpr(E->getArg(i))); 6833 6834 Intrinsic::ID ID = Intrinsic::not_intrinsic; 6835 6836 switch (BuiltinID) { 6837 default: return nullptr; 6838 6839 // __builtin_ppc_get_timebase is GCC 4.8+'s PowerPC-specific name for what we 6840 // call __builtin_readcyclecounter. 6841 case PPC::BI__builtin_ppc_get_timebase: 6842 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::readcyclecounter)); 6843 6844 // vec_ld, vec_lvsl, vec_lvsr 6845 case PPC::BI__builtin_altivec_lvx: 6846 case PPC::BI__builtin_altivec_lvxl: 6847 case PPC::BI__builtin_altivec_lvebx: 6848 case PPC::BI__builtin_altivec_lvehx: 6849 case PPC::BI__builtin_altivec_lvewx: 6850 case PPC::BI__builtin_altivec_lvsl: 6851 case PPC::BI__builtin_altivec_lvsr: 6852 case PPC::BI__builtin_vsx_lxvd2x: 6853 case PPC::BI__builtin_vsx_lxvw4x: 6854 { 6855 Ops[1] = Builder.CreateBitCast(Ops[1], Int8PtrTy); 6856 6857 Ops[0] = Builder.CreateGEP(Ops[1], Ops[0]); 6858 Ops.pop_back(); 6859 6860 switch (BuiltinID) { 6861 default: llvm_unreachable("Unsupported ld/lvsl/lvsr intrinsic!"); 6862 case PPC::BI__builtin_altivec_lvx: 6863 ID = Intrinsic::ppc_altivec_lvx; 6864 break; 6865 case PPC::BI__builtin_altivec_lvxl: 6866 ID = Intrinsic::ppc_altivec_lvxl; 6867 break; 6868 case PPC::BI__builtin_altivec_lvebx: 6869 ID = Intrinsic::ppc_altivec_lvebx; 6870 break; 6871 case PPC::BI__builtin_altivec_lvehx: 6872 ID = Intrinsic::ppc_altivec_lvehx; 6873 break; 6874 case PPC::BI__builtin_altivec_lvewx: 6875 ID = Intrinsic::ppc_altivec_lvewx; 6876 break; 6877 case PPC::BI__builtin_altivec_lvsl: 6878 ID = Intrinsic::ppc_altivec_lvsl; 6879 break; 6880 case PPC::BI__builtin_altivec_lvsr: 6881 ID = Intrinsic::ppc_altivec_lvsr; 6882 break; 6883 case PPC::BI__builtin_vsx_lxvd2x: 6884 ID = Intrinsic::ppc_vsx_lxvd2x; 6885 break; 6886 case PPC::BI__builtin_vsx_lxvw4x: 6887 ID = Intrinsic::ppc_vsx_lxvw4x; 6888 break; 6889 } 6890 llvm::Function *F = CGM.getIntrinsic(ID); 6891 return Builder.CreateCall(F, Ops, ""); 6892 } 6893 6894 // vec_st 6895 case PPC::BI__builtin_altivec_stvx: 6896 case PPC::BI__builtin_altivec_stvxl: 6897 case PPC::BI__builtin_altivec_stvebx: 6898 case PPC::BI__builtin_altivec_stvehx: 6899 case PPC::BI__builtin_altivec_stvewx: 6900 case PPC::BI__builtin_vsx_stxvd2x: 6901 case PPC::BI__builtin_vsx_stxvw4x: 6902 { 6903 Ops[2] = Builder.CreateBitCast(Ops[2], Int8PtrTy); 6904 Ops[1] = Builder.CreateGEP(Ops[2], Ops[1]); 6905 Ops.pop_back(); 6906 6907 switch (BuiltinID) { 6908 default: llvm_unreachable("Unsupported st intrinsic!"); 6909 case PPC::BI__builtin_altivec_stvx: 6910 ID = Intrinsic::ppc_altivec_stvx; 6911 break; 6912 case PPC::BI__builtin_altivec_stvxl: 6913 ID = Intrinsic::ppc_altivec_stvxl; 6914 break; 6915 case PPC::BI__builtin_altivec_stvebx: 6916 ID = Intrinsic::ppc_altivec_stvebx; 6917 break; 6918 case PPC::BI__builtin_altivec_stvehx: 6919 ID = Intrinsic::ppc_altivec_stvehx; 6920 break; 6921 case PPC::BI__builtin_altivec_stvewx: 6922 ID = Intrinsic::ppc_altivec_stvewx; 6923 break; 6924 case PPC::BI__builtin_vsx_stxvd2x: 6925 ID = Intrinsic::ppc_vsx_stxvd2x; 6926 break; 6927 case PPC::BI__builtin_vsx_stxvw4x: 6928 ID = Intrinsic::ppc_vsx_stxvw4x; 6929 break; 6930 } 6931 llvm::Function *F = CGM.getIntrinsic(ID); 6932 return Builder.CreateCall(F, Ops, ""); 6933 } 6934 // Square root 6935 case PPC::BI__builtin_vsx_xvsqrtsp: 6936 case PPC::BI__builtin_vsx_xvsqrtdp: { 6937 llvm::Type *ResultType = ConvertType(E->getType()); 6938 Value *X = EmitScalarExpr(E->getArg(0)); 6939 ID = Intrinsic::sqrt; 6940 llvm::Function *F = CGM.getIntrinsic(ID, ResultType); 6941 return Builder.CreateCall(F, X); 6942 } 6943 // Count leading zeros 6944 case PPC::BI__builtin_altivec_vclzb: 6945 case PPC::BI__builtin_altivec_vclzh: 6946 case PPC::BI__builtin_altivec_vclzw: 6947 case PPC::BI__builtin_altivec_vclzd: { 6948 llvm::Type *ResultType = ConvertType(E->getType()); 6949 Value *X = EmitScalarExpr(E->getArg(0)); 6950 Value *Undef = ConstantInt::get(Builder.getInt1Ty(), false); 6951 Function *F = CGM.getIntrinsic(Intrinsic::ctlz, ResultType); 6952 return Builder.CreateCall(F, {X, Undef}); 6953 } 6954 // Copy sign 6955 case PPC::BI__builtin_vsx_xvcpsgnsp: 6956 case PPC::BI__builtin_vsx_xvcpsgndp: { 6957 llvm::Type *ResultType = ConvertType(E->getType()); 6958 Value *X = EmitScalarExpr(E->getArg(0)); 6959 Value *Y = EmitScalarExpr(E->getArg(1)); 6960 ID = Intrinsic::copysign; 6961 llvm::Function *F = CGM.getIntrinsic(ID, ResultType); 6962 return Builder.CreateCall(F, {X, Y}); 6963 } 6964 // Rounding/truncation 6965 case PPC::BI__builtin_vsx_xvrspip: 6966 case PPC::BI__builtin_vsx_xvrdpip: 6967 case PPC::BI__builtin_vsx_xvrdpim: 6968 case PPC::BI__builtin_vsx_xvrspim: 6969 case PPC::BI__builtin_vsx_xvrdpi: 6970 case PPC::BI__builtin_vsx_xvrspi: 6971 case PPC::BI__builtin_vsx_xvrdpic: 6972 case PPC::BI__builtin_vsx_xvrspic: 6973 case PPC::BI__builtin_vsx_xvrdpiz: 6974 case PPC::BI__builtin_vsx_xvrspiz: { 6975 llvm::Type *ResultType = ConvertType(E->getType()); 6976 Value *X = EmitScalarExpr(E->getArg(0)); 6977 if (BuiltinID == PPC::BI__builtin_vsx_xvrdpim || 6978 BuiltinID == PPC::BI__builtin_vsx_xvrspim) 6979 ID = Intrinsic::floor; 6980 else if (BuiltinID == PPC::BI__builtin_vsx_xvrdpi || 6981 BuiltinID == PPC::BI__builtin_vsx_xvrspi) 6982 ID = Intrinsic::round; 6983 else if (BuiltinID == PPC::BI__builtin_vsx_xvrdpic || 6984 BuiltinID == PPC::BI__builtin_vsx_xvrspic) 6985 ID = Intrinsic::nearbyint; 6986 else if (BuiltinID == PPC::BI__builtin_vsx_xvrdpip || 6987 BuiltinID == PPC::BI__builtin_vsx_xvrspip) 6988 ID = Intrinsic::ceil; 6989 else if (BuiltinID == PPC::BI__builtin_vsx_xvrdpiz || 6990 BuiltinID == PPC::BI__builtin_vsx_xvrspiz) 6991 ID = Intrinsic::trunc; 6992 llvm::Function *F = CGM.getIntrinsic(ID, ResultType); 6993 return Builder.CreateCall(F, X); 6994 } 6995 // FMA variations 6996 case PPC::BI__builtin_vsx_xvmaddadp: 6997 case PPC::BI__builtin_vsx_xvmaddasp: 6998 case PPC::BI__builtin_vsx_xvnmaddadp: 6999 case PPC::BI__builtin_vsx_xvnmaddasp: 7000 case PPC::BI__builtin_vsx_xvmsubadp: 7001 case PPC::BI__builtin_vsx_xvmsubasp: 7002 case PPC::BI__builtin_vsx_xvnmsubadp: 7003 case PPC::BI__builtin_vsx_xvnmsubasp: { 7004 llvm::Type *ResultType = ConvertType(E->getType()); 7005 Value *X = EmitScalarExpr(E->getArg(0)); 7006 Value *Y = EmitScalarExpr(E->getArg(1)); 7007 Value *Z = EmitScalarExpr(E->getArg(2)); 7008 Value *Zero = llvm::ConstantFP::getZeroValueForNegation(ResultType); 7009 llvm::Function *F = CGM.getIntrinsic(Intrinsic::fma, ResultType); 7010 switch (BuiltinID) { 7011 case PPC::BI__builtin_vsx_xvmaddadp: 7012 case PPC::BI__builtin_vsx_xvmaddasp: 7013 return Builder.CreateCall(F, {X, Y, Z}); 7014 case PPC::BI__builtin_vsx_xvnmaddadp: 7015 case PPC::BI__builtin_vsx_xvnmaddasp: 7016 return Builder.CreateFSub(Zero, 7017 Builder.CreateCall(F, {X, Y, Z}), "sub"); 7018 case PPC::BI__builtin_vsx_xvmsubadp: 7019 case PPC::BI__builtin_vsx_xvmsubasp: 7020 return Builder.CreateCall(F, 7021 {X, Y, Builder.CreateFSub(Zero, Z, "sub")}); 7022 case PPC::BI__builtin_vsx_xvnmsubadp: 7023 case PPC::BI__builtin_vsx_xvnmsubasp: 7024 Value *FsubRes = 7025 Builder.CreateCall(F, {X, Y, Builder.CreateFSub(Zero, Z, "sub")}); 7026 return Builder.CreateFSub(Zero, FsubRes, "sub"); 7027 } 7028 llvm_unreachable("Unknown FMA operation"); 7029 return nullptr; // Suppress no-return warning 7030 } 7031 } 7032 } 7033 7034 Value *CodeGenFunction::EmitAMDGPUBuiltinExpr(unsigned BuiltinID, 7035 const CallExpr *E) { 7036 switch (BuiltinID) { 7037 case AMDGPU::BI__builtin_amdgcn_div_scale: 7038 case AMDGPU::BI__builtin_amdgcn_div_scalef: { 7039 // Translate from the intrinsics's struct return to the builtin's out 7040 // argument. 7041 7042 Address FlagOutPtr = EmitPointerWithAlignment(E->getArg(3)); 7043 7044 llvm::Value *X = EmitScalarExpr(E->getArg(0)); 7045 llvm::Value *Y = EmitScalarExpr(E->getArg(1)); 7046 llvm::Value *Z = EmitScalarExpr(E->getArg(2)); 7047 7048 llvm::Value *Callee = CGM.getIntrinsic(Intrinsic::amdgcn_div_scale, 7049 X->getType()); 7050 7051 llvm::Value *Tmp = Builder.CreateCall(Callee, {X, Y, Z}); 7052 7053 llvm::Value *Result = Builder.CreateExtractValue(Tmp, 0); 7054 llvm::Value *Flag = Builder.CreateExtractValue(Tmp, 1); 7055 7056 llvm::Type *RealFlagType 7057 = FlagOutPtr.getPointer()->getType()->getPointerElementType(); 7058 7059 llvm::Value *FlagExt = Builder.CreateZExt(Flag, RealFlagType); 7060 Builder.CreateStore(FlagExt, FlagOutPtr); 7061 return Result; 7062 } 7063 case AMDGPU::BI__builtin_amdgcn_div_fmas: 7064 case AMDGPU::BI__builtin_amdgcn_div_fmasf: { 7065 llvm::Value *Src0 = EmitScalarExpr(E->getArg(0)); 7066 llvm::Value *Src1 = EmitScalarExpr(E->getArg(1)); 7067 llvm::Value *Src2 = EmitScalarExpr(E->getArg(2)); 7068 llvm::Value *Src3 = EmitScalarExpr(E->getArg(3)); 7069 7070 llvm::Value *F = CGM.getIntrinsic(Intrinsic::amdgcn_div_fmas, 7071 Src0->getType()); 7072 llvm::Value *Src3ToBool = Builder.CreateIsNotNull(Src3); 7073 return Builder.CreateCall(F, {Src0, Src1, Src2, Src3ToBool}); 7074 } 7075 case AMDGPU::BI__builtin_amdgcn_div_fixup: 7076 case AMDGPU::BI__builtin_amdgcn_div_fixupf: 7077 return emitTernaryFPBuiltin(*this, E, Intrinsic::amdgcn_div_fixup); 7078 case AMDGPU::BI__builtin_amdgcn_trig_preop: 7079 case AMDGPU::BI__builtin_amdgcn_trig_preopf: 7080 return emitFPIntBuiltin(*this, E, Intrinsic::amdgcn_trig_preop); 7081 case AMDGPU::BI__builtin_amdgcn_rcp: 7082 case AMDGPU::BI__builtin_amdgcn_rcpf: 7083 return emitUnaryBuiltin(*this, E, Intrinsic::amdgcn_rcp); 7084 case AMDGPU::BI__builtin_amdgcn_rsq: 7085 case AMDGPU::BI__builtin_amdgcn_rsqf: 7086 return emitUnaryBuiltin(*this, E, Intrinsic::amdgcn_rsq); 7087 case AMDGPU::BI__builtin_amdgcn_rsq_clamp: 7088 case AMDGPU::BI__builtin_amdgcn_rsq_clampf: 7089 return emitUnaryBuiltin(*this, E, Intrinsic::amdgcn_rsq_clamp); 7090 case AMDGPU::BI__builtin_amdgcn_sinf: 7091 return emitUnaryBuiltin(*this, E, Intrinsic::amdgcn_sin); 7092 case AMDGPU::BI__builtin_amdgcn_cosf: 7093 return emitUnaryBuiltin(*this, E, Intrinsic::amdgcn_cos); 7094 case AMDGPU::BI__builtin_amdgcn_log_clampf: 7095 return emitUnaryBuiltin(*this, E, Intrinsic::amdgcn_log_clamp); 7096 case AMDGPU::BI__builtin_amdgcn_ldexp: 7097 case AMDGPU::BI__builtin_amdgcn_ldexpf: 7098 return emitFPIntBuiltin(*this, E, Intrinsic::amdgcn_ldexp); 7099 case AMDGPU::BI__builtin_amdgcn_class: 7100 case AMDGPU::BI__builtin_amdgcn_classf: 7101 return emitFPIntBuiltin(*this, E, Intrinsic::amdgcn_class); 7102 7103 // Legacy amdgpu prefix 7104 case AMDGPU::BI__builtin_amdgpu_rsq: 7105 case AMDGPU::BI__builtin_amdgpu_rsqf: { 7106 if (getTarget().getTriple().getArch() == Triple::amdgcn) 7107 return emitUnaryBuiltin(*this, E, Intrinsic::amdgcn_rsq); 7108 return emitUnaryBuiltin(*this, E, Intrinsic::r600_rsq); 7109 } 7110 case AMDGPU::BI__builtin_amdgpu_ldexp: 7111 case AMDGPU::BI__builtin_amdgpu_ldexpf: { 7112 if (getTarget().getTriple().getArch() == Triple::amdgcn) 7113 return emitFPIntBuiltin(*this, E, Intrinsic::amdgcn_ldexp); 7114 return emitFPIntBuiltin(*this, E, Intrinsic::AMDGPU_ldexp); 7115 } 7116 default: 7117 return nullptr; 7118 } 7119 } 7120 7121 /// Handle a SystemZ function in which the final argument is a pointer 7122 /// to an int that receives the post-instruction CC value. At the LLVM level 7123 /// this is represented as a function that returns a {result, cc} pair. 7124 static Value *EmitSystemZIntrinsicWithCC(CodeGenFunction &CGF, 7125 unsigned IntrinsicID, 7126 const CallExpr *E) { 7127 unsigned NumArgs = E->getNumArgs() - 1; 7128 SmallVector<Value *, 8> Args(NumArgs); 7129 for (unsigned I = 0; I < NumArgs; ++I) 7130 Args[I] = CGF.EmitScalarExpr(E->getArg(I)); 7131 Address CCPtr = CGF.EmitPointerWithAlignment(E->getArg(NumArgs)); 7132 Value *F = CGF.CGM.getIntrinsic(IntrinsicID); 7133 Value *Call = CGF.Builder.CreateCall(F, Args); 7134 Value *CC = CGF.Builder.CreateExtractValue(Call, 1); 7135 CGF.Builder.CreateStore(CC, CCPtr); 7136 return CGF.Builder.CreateExtractValue(Call, 0); 7137 } 7138 7139 Value *CodeGenFunction::EmitSystemZBuiltinExpr(unsigned BuiltinID, 7140 const CallExpr *E) { 7141 switch (BuiltinID) { 7142 case SystemZ::BI__builtin_tbegin: { 7143 Value *TDB = EmitScalarExpr(E->getArg(0)); 7144 Value *Control = llvm::ConstantInt::get(Int32Ty, 0xff0c); 7145 Value *F = CGM.getIntrinsic(Intrinsic::s390_tbegin); 7146 return Builder.CreateCall(F, {TDB, Control}); 7147 } 7148 case SystemZ::BI__builtin_tbegin_nofloat: { 7149 Value *TDB = EmitScalarExpr(E->getArg(0)); 7150 Value *Control = llvm::ConstantInt::get(Int32Ty, 0xff0c); 7151 Value *F = CGM.getIntrinsic(Intrinsic::s390_tbegin_nofloat); 7152 return Builder.CreateCall(F, {TDB, Control}); 7153 } 7154 case SystemZ::BI__builtin_tbeginc: { 7155 Value *TDB = llvm::ConstantPointerNull::get(Int8PtrTy); 7156 Value *Control = llvm::ConstantInt::get(Int32Ty, 0xff08); 7157 Value *F = CGM.getIntrinsic(Intrinsic::s390_tbeginc); 7158 return Builder.CreateCall(F, {TDB, Control}); 7159 } 7160 case SystemZ::BI__builtin_tabort: { 7161 Value *Data = EmitScalarExpr(E->getArg(0)); 7162 Value *F = CGM.getIntrinsic(Intrinsic::s390_tabort); 7163 return Builder.CreateCall(F, Builder.CreateSExt(Data, Int64Ty, "tabort")); 7164 } 7165 case SystemZ::BI__builtin_non_tx_store: { 7166 Value *Address = EmitScalarExpr(E->getArg(0)); 7167 Value *Data = EmitScalarExpr(E->getArg(1)); 7168 Value *F = CGM.getIntrinsic(Intrinsic::s390_ntstg); 7169 return Builder.CreateCall(F, {Data, Address}); 7170 } 7171 7172 // Vector builtins. Note that most vector builtins are mapped automatically 7173 // to target-specific LLVM intrinsics. The ones handled specially here can 7174 // be represented via standard LLVM IR, which is preferable to enable common 7175 // LLVM optimizations. 7176 7177 case SystemZ::BI__builtin_s390_vpopctb: 7178 case SystemZ::BI__builtin_s390_vpopcth: 7179 case SystemZ::BI__builtin_s390_vpopctf: 7180 case SystemZ::BI__builtin_s390_vpopctg: { 7181 llvm::Type *ResultType = ConvertType(E->getType()); 7182 Value *X = EmitScalarExpr(E->getArg(0)); 7183 Function *F = CGM.getIntrinsic(Intrinsic::ctpop, ResultType); 7184 return Builder.CreateCall(F, X); 7185 } 7186 7187 case SystemZ::BI__builtin_s390_vclzb: 7188 case SystemZ::BI__builtin_s390_vclzh: 7189 case SystemZ::BI__builtin_s390_vclzf: 7190 case SystemZ::BI__builtin_s390_vclzg: { 7191 llvm::Type *ResultType = ConvertType(E->getType()); 7192 Value *X = EmitScalarExpr(E->getArg(0)); 7193 Value *Undef = ConstantInt::get(Builder.getInt1Ty(), false); 7194 Function *F = CGM.getIntrinsic(Intrinsic::ctlz, ResultType); 7195 return Builder.CreateCall(F, {X, Undef}); 7196 } 7197 7198 case SystemZ::BI__builtin_s390_vctzb: 7199 case SystemZ::BI__builtin_s390_vctzh: 7200 case SystemZ::BI__builtin_s390_vctzf: 7201 case SystemZ::BI__builtin_s390_vctzg: { 7202 llvm::Type *ResultType = ConvertType(E->getType()); 7203 Value *X = EmitScalarExpr(E->getArg(0)); 7204 Value *Undef = ConstantInt::get(Builder.getInt1Ty(), false); 7205 Function *F = CGM.getIntrinsic(Intrinsic::cttz, ResultType); 7206 return Builder.CreateCall(F, {X, Undef}); 7207 } 7208 7209 case SystemZ::BI__builtin_s390_vfsqdb: { 7210 llvm::Type *ResultType = ConvertType(E->getType()); 7211 Value *X = EmitScalarExpr(E->getArg(0)); 7212 Function *F = CGM.getIntrinsic(Intrinsic::sqrt, ResultType); 7213 return Builder.CreateCall(F, X); 7214 } 7215 case SystemZ::BI__builtin_s390_vfmadb: { 7216 llvm::Type *ResultType = ConvertType(E->getType()); 7217 Value *X = EmitScalarExpr(E->getArg(0)); 7218 Value *Y = EmitScalarExpr(E->getArg(1)); 7219 Value *Z = EmitScalarExpr(E->getArg(2)); 7220 Function *F = CGM.getIntrinsic(Intrinsic::fma, ResultType); 7221 return Builder.CreateCall(F, {X, Y, Z}); 7222 } 7223 case SystemZ::BI__builtin_s390_vfmsdb: { 7224 llvm::Type *ResultType = ConvertType(E->getType()); 7225 Value *X = EmitScalarExpr(E->getArg(0)); 7226 Value *Y = EmitScalarExpr(E->getArg(1)); 7227 Value *Z = EmitScalarExpr(E->getArg(2)); 7228 Value *Zero = llvm::ConstantFP::getZeroValueForNegation(ResultType); 7229 Function *F = CGM.getIntrinsic(Intrinsic::fma, ResultType); 7230 return Builder.CreateCall(F, {X, Y, Builder.CreateFSub(Zero, Z, "sub")}); 7231 } 7232 case SystemZ::BI__builtin_s390_vflpdb: { 7233 llvm::Type *ResultType = ConvertType(E->getType()); 7234 Value *X = EmitScalarExpr(E->getArg(0)); 7235 Function *F = CGM.getIntrinsic(Intrinsic::fabs, ResultType); 7236 return Builder.CreateCall(F, X); 7237 } 7238 case SystemZ::BI__builtin_s390_vflndb: { 7239 llvm::Type *ResultType = ConvertType(E->getType()); 7240 Value *X = EmitScalarExpr(E->getArg(0)); 7241 Value *Zero = llvm::ConstantFP::getZeroValueForNegation(ResultType); 7242 Function *F = CGM.getIntrinsic(Intrinsic::fabs, ResultType); 7243 return Builder.CreateFSub(Zero, Builder.CreateCall(F, X), "sub"); 7244 } 7245 case SystemZ::BI__builtin_s390_vfidb: { 7246 llvm::Type *ResultType = ConvertType(E->getType()); 7247 Value *X = EmitScalarExpr(E->getArg(0)); 7248 // Constant-fold the M4 and M5 mask arguments. 7249 llvm::APSInt M4, M5; 7250 bool IsConstM4 = E->getArg(1)->isIntegerConstantExpr(M4, getContext()); 7251 bool IsConstM5 = E->getArg(2)->isIntegerConstantExpr(M5, getContext()); 7252 assert(IsConstM4 && IsConstM5 && "Constant arg isn't actually constant?"); 7253 (void)IsConstM4; (void)IsConstM5; 7254 // Check whether this instance of vfidb can be represented via a LLVM 7255 // standard intrinsic. We only support some combinations of M4 and M5. 7256 Intrinsic::ID ID = Intrinsic::not_intrinsic; 7257 switch (M4.getZExtValue()) { 7258 default: break; 7259 case 0: // IEEE-inexact exception allowed 7260 switch (M5.getZExtValue()) { 7261 default: break; 7262 case 0: ID = Intrinsic::rint; break; 7263 } 7264 break; 7265 case 4: // IEEE-inexact exception suppressed 7266 switch (M5.getZExtValue()) { 7267 default: break; 7268 case 0: ID = Intrinsic::nearbyint; break; 7269 case 1: ID = Intrinsic::round; break; 7270 case 5: ID = Intrinsic::trunc; break; 7271 case 6: ID = Intrinsic::ceil; break; 7272 case 7: ID = Intrinsic::floor; break; 7273 } 7274 break; 7275 } 7276 if (ID != Intrinsic::not_intrinsic) { 7277 Function *F = CGM.getIntrinsic(ID, ResultType); 7278 return Builder.CreateCall(F, X); 7279 } 7280 Function *F = CGM.getIntrinsic(Intrinsic::s390_vfidb); 7281 Value *M4Value = llvm::ConstantInt::get(getLLVMContext(), M4); 7282 Value *M5Value = llvm::ConstantInt::get(getLLVMContext(), M5); 7283 return Builder.CreateCall(F, {X, M4Value, M5Value}); 7284 } 7285 7286 // Vector intrisincs that output the post-instruction CC value. 7287 7288 #define INTRINSIC_WITH_CC(NAME) \ 7289 case SystemZ::BI__builtin_##NAME: \ 7290 return EmitSystemZIntrinsicWithCC(*this, Intrinsic::NAME, E) 7291 7292 INTRINSIC_WITH_CC(s390_vpkshs); 7293 INTRINSIC_WITH_CC(s390_vpksfs); 7294 INTRINSIC_WITH_CC(s390_vpksgs); 7295 7296 INTRINSIC_WITH_CC(s390_vpklshs); 7297 INTRINSIC_WITH_CC(s390_vpklsfs); 7298 INTRINSIC_WITH_CC(s390_vpklsgs); 7299 7300 INTRINSIC_WITH_CC(s390_vceqbs); 7301 INTRINSIC_WITH_CC(s390_vceqhs); 7302 INTRINSIC_WITH_CC(s390_vceqfs); 7303 INTRINSIC_WITH_CC(s390_vceqgs); 7304 7305 INTRINSIC_WITH_CC(s390_vchbs); 7306 INTRINSIC_WITH_CC(s390_vchhs); 7307 INTRINSIC_WITH_CC(s390_vchfs); 7308 INTRINSIC_WITH_CC(s390_vchgs); 7309 7310 INTRINSIC_WITH_CC(s390_vchlbs); 7311 INTRINSIC_WITH_CC(s390_vchlhs); 7312 INTRINSIC_WITH_CC(s390_vchlfs); 7313 INTRINSIC_WITH_CC(s390_vchlgs); 7314 7315 INTRINSIC_WITH_CC(s390_vfaebs); 7316 INTRINSIC_WITH_CC(s390_vfaehs); 7317 INTRINSIC_WITH_CC(s390_vfaefs); 7318 7319 INTRINSIC_WITH_CC(s390_vfaezbs); 7320 INTRINSIC_WITH_CC(s390_vfaezhs); 7321 INTRINSIC_WITH_CC(s390_vfaezfs); 7322 7323 INTRINSIC_WITH_CC(s390_vfeebs); 7324 INTRINSIC_WITH_CC(s390_vfeehs); 7325 INTRINSIC_WITH_CC(s390_vfeefs); 7326 7327 INTRINSIC_WITH_CC(s390_vfeezbs); 7328 INTRINSIC_WITH_CC(s390_vfeezhs); 7329 INTRINSIC_WITH_CC(s390_vfeezfs); 7330 7331 INTRINSIC_WITH_CC(s390_vfenebs); 7332 INTRINSIC_WITH_CC(s390_vfenehs); 7333 INTRINSIC_WITH_CC(s390_vfenefs); 7334 7335 INTRINSIC_WITH_CC(s390_vfenezbs); 7336 INTRINSIC_WITH_CC(s390_vfenezhs); 7337 INTRINSIC_WITH_CC(s390_vfenezfs); 7338 7339 INTRINSIC_WITH_CC(s390_vistrbs); 7340 INTRINSIC_WITH_CC(s390_vistrhs); 7341 INTRINSIC_WITH_CC(s390_vistrfs); 7342 7343 INTRINSIC_WITH_CC(s390_vstrcbs); 7344 INTRINSIC_WITH_CC(s390_vstrchs); 7345 INTRINSIC_WITH_CC(s390_vstrcfs); 7346 7347 INTRINSIC_WITH_CC(s390_vstrczbs); 7348 INTRINSIC_WITH_CC(s390_vstrczhs); 7349 INTRINSIC_WITH_CC(s390_vstrczfs); 7350 7351 INTRINSIC_WITH_CC(s390_vfcedbs); 7352 INTRINSIC_WITH_CC(s390_vfchdbs); 7353 INTRINSIC_WITH_CC(s390_vfchedbs); 7354 7355 INTRINSIC_WITH_CC(s390_vftcidb); 7356 7357 #undef INTRINSIC_WITH_CC 7358 7359 default: 7360 return nullptr; 7361 } 7362 } 7363 7364 Value *CodeGenFunction::EmitNVPTXBuiltinExpr(unsigned BuiltinID, 7365 const CallExpr *E) { 7366 switch (BuiltinID) { 7367 case NVPTX::BI__nvvm_atom_add_gen_i: 7368 case NVPTX::BI__nvvm_atom_add_gen_l: 7369 case NVPTX::BI__nvvm_atom_add_gen_ll: 7370 return MakeBinaryAtomicValue(*this, llvm::AtomicRMWInst::Add, E); 7371 7372 case NVPTX::BI__nvvm_atom_sub_gen_i: 7373 case NVPTX::BI__nvvm_atom_sub_gen_l: 7374 case NVPTX::BI__nvvm_atom_sub_gen_ll: 7375 return MakeBinaryAtomicValue(*this, llvm::AtomicRMWInst::Sub, E); 7376 7377 case NVPTX::BI__nvvm_atom_and_gen_i: 7378 case NVPTX::BI__nvvm_atom_and_gen_l: 7379 case NVPTX::BI__nvvm_atom_and_gen_ll: 7380 return MakeBinaryAtomicValue(*this, llvm::AtomicRMWInst::And, E); 7381 7382 case NVPTX::BI__nvvm_atom_or_gen_i: 7383 case NVPTX::BI__nvvm_atom_or_gen_l: 7384 case NVPTX::BI__nvvm_atom_or_gen_ll: 7385 return MakeBinaryAtomicValue(*this, llvm::AtomicRMWInst::Or, E); 7386 7387 case NVPTX::BI__nvvm_atom_xor_gen_i: 7388 case NVPTX::BI__nvvm_atom_xor_gen_l: 7389 case NVPTX::BI__nvvm_atom_xor_gen_ll: 7390 return MakeBinaryAtomicValue(*this, llvm::AtomicRMWInst::Xor, E); 7391 7392 case NVPTX::BI__nvvm_atom_xchg_gen_i: 7393 case NVPTX::BI__nvvm_atom_xchg_gen_l: 7394 case NVPTX::BI__nvvm_atom_xchg_gen_ll: 7395 return MakeBinaryAtomicValue(*this, llvm::AtomicRMWInst::Xchg, E); 7396 7397 case NVPTX::BI__nvvm_atom_max_gen_i: 7398 case NVPTX::BI__nvvm_atom_max_gen_l: 7399 case NVPTX::BI__nvvm_atom_max_gen_ll: 7400 return MakeBinaryAtomicValue(*this, llvm::AtomicRMWInst::Max, E); 7401 7402 case NVPTX::BI__nvvm_atom_max_gen_ui: 7403 case NVPTX::BI__nvvm_atom_max_gen_ul: 7404 case NVPTX::BI__nvvm_atom_max_gen_ull: 7405 return MakeBinaryAtomicValue(*this, llvm::AtomicRMWInst::UMax, E); 7406 7407 case NVPTX::BI__nvvm_atom_min_gen_i: 7408 case NVPTX::BI__nvvm_atom_min_gen_l: 7409 case NVPTX::BI__nvvm_atom_min_gen_ll: 7410 return MakeBinaryAtomicValue(*this, llvm::AtomicRMWInst::Min, E); 7411 7412 case NVPTX::BI__nvvm_atom_min_gen_ui: 7413 case NVPTX::BI__nvvm_atom_min_gen_ul: 7414 case NVPTX::BI__nvvm_atom_min_gen_ull: 7415 return MakeBinaryAtomicValue(*this, llvm::AtomicRMWInst::UMin, E); 7416 7417 case NVPTX::BI__nvvm_atom_cas_gen_i: 7418 case NVPTX::BI__nvvm_atom_cas_gen_l: 7419 case NVPTX::BI__nvvm_atom_cas_gen_ll: 7420 // __nvvm_atom_cas_gen_* should return the old value rather than the 7421 // success flag. 7422 return MakeAtomicCmpXchgValue(*this, E, /*ReturnBool=*/false); 7423 7424 case NVPTX::BI__nvvm_atom_add_gen_f: { 7425 Value *Ptr = EmitScalarExpr(E->getArg(0)); 7426 Value *Val = EmitScalarExpr(E->getArg(1)); 7427 // atomicrmw only deals with integer arguments so we need to use 7428 // LLVM's nvvm_atomic_load_add_f32 intrinsic for that. 7429 Value *FnALAF32 = 7430 CGM.getIntrinsic(Intrinsic::nvvm_atomic_load_add_f32, Ptr->getType()); 7431 return Builder.CreateCall(FnALAF32, {Ptr, Val}); 7432 } 7433 7434 default: 7435 return nullptr; 7436 } 7437 } 7438 7439 Value *CodeGenFunction::EmitWebAssemblyBuiltinExpr(unsigned BuiltinID, 7440 const CallExpr *E) { 7441 switch (BuiltinID) { 7442 case WebAssembly::BI__builtin_wasm_memory_size: { 7443 llvm::Type *ResultType = ConvertType(E->getType()); 7444 Value *Callee = CGM.getIntrinsic(Intrinsic::wasm_memory_size, ResultType); 7445 return Builder.CreateCall(Callee); 7446 } 7447 case WebAssembly::BI__builtin_wasm_grow_memory: { 7448 Value *X = EmitScalarExpr(E->getArg(0)); 7449 Value *Callee = CGM.getIntrinsic(Intrinsic::wasm_grow_memory, X->getType()); 7450 return Builder.CreateCall(Callee, X); 7451 } 7452 7453 default: 7454 return nullptr; 7455 } 7456 } 7457