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_bitreverse8: 685 case Builtin::BI__builtin_bitreverse16: 686 case Builtin::BI__builtin_bitreverse32: 687 case Builtin::BI__builtin_bitreverse64: { 688 return RValue::get(emitUnaryBuiltin(*this, E, Intrinsic::bitreverse)); 689 } 690 case Builtin::BI__builtin_object_size: { 691 unsigned Type = 692 E->getArg(1)->EvaluateKnownConstInt(getContext()).getZExtValue(); 693 auto *ResType = cast<llvm::IntegerType>(ConvertType(E->getType())); 694 695 // We pass this builtin onto the optimizer so that it can figure out the 696 // object size in more complex cases. 697 return RValue::get(emitBuiltinObjectSize(E->getArg(0), Type, ResType)); 698 } 699 case Builtin::BI__builtin_prefetch: { 700 Value *Locality, *RW, *Address = EmitScalarExpr(E->getArg(0)); 701 // FIXME: Technically these constants should of type 'int', yes? 702 RW = (E->getNumArgs() > 1) ? EmitScalarExpr(E->getArg(1)) : 703 llvm::ConstantInt::get(Int32Ty, 0); 704 Locality = (E->getNumArgs() > 2) ? EmitScalarExpr(E->getArg(2)) : 705 llvm::ConstantInt::get(Int32Ty, 3); 706 Value *Data = llvm::ConstantInt::get(Int32Ty, 1); 707 Value *F = CGM.getIntrinsic(Intrinsic::prefetch); 708 return RValue::get(Builder.CreateCall(F, {Address, RW, Locality, Data})); 709 } 710 case Builtin::BI__builtin_readcyclecounter: { 711 Value *F = CGM.getIntrinsic(Intrinsic::readcyclecounter); 712 return RValue::get(Builder.CreateCall(F)); 713 } 714 case Builtin::BI__builtin___clear_cache: { 715 Value *Begin = EmitScalarExpr(E->getArg(0)); 716 Value *End = EmitScalarExpr(E->getArg(1)); 717 Value *F = CGM.getIntrinsic(Intrinsic::clear_cache); 718 return RValue::get(Builder.CreateCall(F, {Begin, End})); 719 } 720 case Builtin::BI__builtin_trap: 721 return RValue::get(EmitTrapCall(Intrinsic::trap)); 722 case Builtin::BI__debugbreak: 723 return RValue::get(EmitTrapCall(Intrinsic::debugtrap)); 724 case Builtin::BI__builtin_unreachable: { 725 if (SanOpts.has(SanitizerKind::Unreachable)) { 726 SanitizerScope SanScope(this); 727 EmitCheck(std::make_pair(static_cast<llvm::Value *>(Builder.getFalse()), 728 SanitizerKind::Unreachable), 729 "builtin_unreachable", EmitCheckSourceLocation(E->getExprLoc()), 730 None); 731 } else 732 Builder.CreateUnreachable(); 733 734 // We do need to preserve an insertion point. 735 EmitBlock(createBasicBlock("unreachable.cont")); 736 737 return RValue::get(nullptr); 738 } 739 740 case Builtin::BI__builtin_powi: 741 case Builtin::BI__builtin_powif: 742 case Builtin::BI__builtin_powil: { 743 Value *Base = EmitScalarExpr(E->getArg(0)); 744 Value *Exponent = EmitScalarExpr(E->getArg(1)); 745 llvm::Type *ArgType = Base->getType(); 746 Value *F = CGM.getIntrinsic(Intrinsic::powi, ArgType); 747 return RValue::get(Builder.CreateCall(F, {Base, Exponent})); 748 } 749 750 case Builtin::BI__builtin_isgreater: 751 case Builtin::BI__builtin_isgreaterequal: 752 case Builtin::BI__builtin_isless: 753 case Builtin::BI__builtin_islessequal: 754 case Builtin::BI__builtin_islessgreater: 755 case Builtin::BI__builtin_isunordered: { 756 // Ordered comparisons: we know the arguments to these are matching scalar 757 // floating point values. 758 Value *LHS = EmitScalarExpr(E->getArg(0)); 759 Value *RHS = EmitScalarExpr(E->getArg(1)); 760 761 switch (BuiltinID) { 762 default: llvm_unreachable("Unknown ordered comparison"); 763 case Builtin::BI__builtin_isgreater: 764 LHS = Builder.CreateFCmpOGT(LHS, RHS, "cmp"); 765 break; 766 case Builtin::BI__builtin_isgreaterequal: 767 LHS = Builder.CreateFCmpOGE(LHS, RHS, "cmp"); 768 break; 769 case Builtin::BI__builtin_isless: 770 LHS = Builder.CreateFCmpOLT(LHS, RHS, "cmp"); 771 break; 772 case Builtin::BI__builtin_islessequal: 773 LHS = Builder.CreateFCmpOLE(LHS, RHS, "cmp"); 774 break; 775 case Builtin::BI__builtin_islessgreater: 776 LHS = Builder.CreateFCmpONE(LHS, RHS, "cmp"); 777 break; 778 case Builtin::BI__builtin_isunordered: 779 LHS = Builder.CreateFCmpUNO(LHS, RHS, "cmp"); 780 break; 781 } 782 // ZExt bool to int type. 783 return RValue::get(Builder.CreateZExt(LHS, ConvertType(E->getType()))); 784 } 785 case Builtin::BI__builtin_isnan: { 786 Value *V = EmitScalarExpr(E->getArg(0)); 787 V = Builder.CreateFCmpUNO(V, V, "cmp"); 788 return RValue::get(Builder.CreateZExt(V, ConvertType(E->getType()))); 789 } 790 791 case Builtin::BI__builtin_isinf: { 792 // isinf(x) --> fabs(x) == infinity 793 Value *V = EmitScalarExpr(E->getArg(0)); 794 V = EmitFAbs(*this, V); 795 796 V = Builder.CreateFCmpOEQ(V, ConstantFP::getInfinity(V->getType()),"isinf"); 797 return RValue::get(Builder.CreateZExt(V, ConvertType(E->getType()))); 798 } 799 800 case Builtin::BI__builtin_isinf_sign: { 801 // isinf_sign(x) -> fabs(x) == infinity ? (signbit(x) ? -1 : 1) : 0 802 Value *Arg = EmitScalarExpr(E->getArg(0)); 803 Value *AbsArg = EmitFAbs(*this, Arg); 804 Value *IsInf = Builder.CreateFCmpOEQ( 805 AbsArg, ConstantFP::getInfinity(Arg->getType()), "isinf"); 806 Value *IsNeg = EmitSignBit(*this, Arg); 807 808 llvm::Type *IntTy = ConvertType(E->getType()); 809 Value *Zero = Constant::getNullValue(IntTy); 810 Value *One = ConstantInt::get(IntTy, 1); 811 Value *NegativeOne = ConstantInt::get(IntTy, -1); 812 Value *SignResult = Builder.CreateSelect(IsNeg, NegativeOne, One); 813 Value *Result = Builder.CreateSelect(IsInf, SignResult, Zero); 814 return RValue::get(Result); 815 } 816 817 case Builtin::BI__builtin_isnormal: { 818 // isnormal(x) --> x == x && fabsf(x) < infinity && fabsf(x) >= float_min 819 Value *V = EmitScalarExpr(E->getArg(0)); 820 Value *Eq = Builder.CreateFCmpOEQ(V, V, "iseq"); 821 822 Value *Abs = EmitFAbs(*this, V); 823 Value *IsLessThanInf = 824 Builder.CreateFCmpULT(Abs, ConstantFP::getInfinity(V->getType()),"isinf"); 825 APFloat Smallest = APFloat::getSmallestNormalized( 826 getContext().getFloatTypeSemantics(E->getArg(0)->getType())); 827 Value *IsNormal = 828 Builder.CreateFCmpUGE(Abs, ConstantFP::get(V->getContext(), Smallest), 829 "isnormal"); 830 V = Builder.CreateAnd(Eq, IsLessThanInf, "and"); 831 V = Builder.CreateAnd(V, IsNormal, "and"); 832 return RValue::get(Builder.CreateZExt(V, ConvertType(E->getType()))); 833 } 834 835 case Builtin::BI__builtin_isfinite: { 836 // isfinite(x) --> x == x && fabs(x) != infinity; 837 Value *V = EmitScalarExpr(E->getArg(0)); 838 Value *Eq = Builder.CreateFCmpOEQ(V, V, "iseq"); 839 840 Value *Abs = EmitFAbs(*this, V); 841 Value *IsNotInf = 842 Builder.CreateFCmpUNE(Abs, ConstantFP::getInfinity(V->getType()),"isinf"); 843 844 V = Builder.CreateAnd(Eq, IsNotInf, "and"); 845 return RValue::get(Builder.CreateZExt(V, ConvertType(E->getType()))); 846 } 847 848 case Builtin::BI__builtin_fpclassify: { 849 Value *V = EmitScalarExpr(E->getArg(5)); 850 llvm::Type *Ty = ConvertType(E->getArg(5)->getType()); 851 852 // Create Result 853 BasicBlock *Begin = Builder.GetInsertBlock(); 854 BasicBlock *End = createBasicBlock("fpclassify_end", this->CurFn); 855 Builder.SetInsertPoint(End); 856 PHINode *Result = 857 Builder.CreatePHI(ConvertType(E->getArg(0)->getType()), 4, 858 "fpclassify_result"); 859 860 // if (V==0) return FP_ZERO 861 Builder.SetInsertPoint(Begin); 862 Value *IsZero = Builder.CreateFCmpOEQ(V, Constant::getNullValue(Ty), 863 "iszero"); 864 Value *ZeroLiteral = EmitScalarExpr(E->getArg(4)); 865 BasicBlock *NotZero = createBasicBlock("fpclassify_not_zero", this->CurFn); 866 Builder.CreateCondBr(IsZero, End, NotZero); 867 Result->addIncoming(ZeroLiteral, Begin); 868 869 // if (V != V) return FP_NAN 870 Builder.SetInsertPoint(NotZero); 871 Value *IsNan = Builder.CreateFCmpUNO(V, V, "cmp"); 872 Value *NanLiteral = EmitScalarExpr(E->getArg(0)); 873 BasicBlock *NotNan = createBasicBlock("fpclassify_not_nan", this->CurFn); 874 Builder.CreateCondBr(IsNan, End, NotNan); 875 Result->addIncoming(NanLiteral, NotZero); 876 877 // if (fabs(V) == infinity) return FP_INFINITY 878 Builder.SetInsertPoint(NotNan); 879 Value *VAbs = EmitFAbs(*this, V); 880 Value *IsInf = 881 Builder.CreateFCmpOEQ(VAbs, ConstantFP::getInfinity(V->getType()), 882 "isinf"); 883 Value *InfLiteral = EmitScalarExpr(E->getArg(1)); 884 BasicBlock *NotInf = createBasicBlock("fpclassify_not_inf", this->CurFn); 885 Builder.CreateCondBr(IsInf, End, NotInf); 886 Result->addIncoming(InfLiteral, NotNan); 887 888 // if (fabs(V) >= MIN_NORMAL) return FP_NORMAL else FP_SUBNORMAL 889 Builder.SetInsertPoint(NotInf); 890 APFloat Smallest = APFloat::getSmallestNormalized( 891 getContext().getFloatTypeSemantics(E->getArg(5)->getType())); 892 Value *IsNormal = 893 Builder.CreateFCmpUGE(VAbs, ConstantFP::get(V->getContext(), Smallest), 894 "isnormal"); 895 Value *NormalResult = 896 Builder.CreateSelect(IsNormal, EmitScalarExpr(E->getArg(2)), 897 EmitScalarExpr(E->getArg(3))); 898 Builder.CreateBr(End); 899 Result->addIncoming(NormalResult, NotInf); 900 901 // return Result 902 Builder.SetInsertPoint(End); 903 return RValue::get(Result); 904 } 905 906 case Builtin::BIalloca: 907 case Builtin::BI_alloca: 908 case Builtin::BI__builtin_alloca: { 909 Value *Size = EmitScalarExpr(E->getArg(0)); 910 return RValue::get(Builder.CreateAlloca(Builder.getInt8Ty(), Size)); 911 } 912 case Builtin::BIbzero: 913 case Builtin::BI__builtin_bzero: { 914 Address Dest = EmitPointerWithAlignment(E->getArg(0)); 915 Value *SizeVal = EmitScalarExpr(E->getArg(1)); 916 EmitNonNullArgCheck(RValue::get(Dest.getPointer()), E->getArg(0)->getType(), 917 E->getArg(0)->getExprLoc(), FD, 0); 918 Builder.CreateMemSet(Dest, Builder.getInt8(0), SizeVal, false); 919 return RValue::get(Dest.getPointer()); 920 } 921 case Builtin::BImemcpy: 922 case Builtin::BI__builtin_memcpy: { 923 Address Dest = EmitPointerWithAlignment(E->getArg(0)); 924 Address Src = EmitPointerWithAlignment(E->getArg(1)); 925 Value *SizeVal = EmitScalarExpr(E->getArg(2)); 926 EmitNonNullArgCheck(RValue::get(Dest.getPointer()), E->getArg(0)->getType(), 927 E->getArg(0)->getExprLoc(), FD, 0); 928 EmitNonNullArgCheck(RValue::get(Src.getPointer()), E->getArg(1)->getType(), 929 E->getArg(1)->getExprLoc(), FD, 1); 930 Builder.CreateMemCpy(Dest, Src, SizeVal, false); 931 return RValue::get(Dest.getPointer()); 932 } 933 934 case Builtin::BI__builtin___memcpy_chk: { 935 // fold __builtin_memcpy_chk(x, y, cst1, cst2) to memcpy iff cst1<=cst2. 936 llvm::APSInt Size, DstSize; 937 if (!E->getArg(2)->EvaluateAsInt(Size, CGM.getContext()) || 938 !E->getArg(3)->EvaluateAsInt(DstSize, CGM.getContext())) 939 break; 940 if (Size.ugt(DstSize)) 941 break; 942 Address Dest = EmitPointerWithAlignment(E->getArg(0)); 943 Address Src = EmitPointerWithAlignment(E->getArg(1)); 944 Value *SizeVal = llvm::ConstantInt::get(Builder.getContext(), Size); 945 Builder.CreateMemCpy(Dest, Src, SizeVal, false); 946 return RValue::get(Dest.getPointer()); 947 } 948 949 case Builtin::BI__builtin_objc_memmove_collectable: { 950 Address DestAddr = EmitPointerWithAlignment(E->getArg(0)); 951 Address SrcAddr = EmitPointerWithAlignment(E->getArg(1)); 952 Value *SizeVal = EmitScalarExpr(E->getArg(2)); 953 CGM.getObjCRuntime().EmitGCMemmoveCollectable(*this, 954 DestAddr, SrcAddr, SizeVal); 955 return RValue::get(DestAddr.getPointer()); 956 } 957 958 case Builtin::BI__builtin___memmove_chk: { 959 // fold __builtin_memmove_chk(x, y, cst1, cst2) to memmove iff cst1<=cst2. 960 llvm::APSInt Size, DstSize; 961 if (!E->getArg(2)->EvaluateAsInt(Size, CGM.getContext()) || 962 !E->getArg(3)->EvaluateAsInt(DstSize, CGM.getContext())) 963 break; 964 if (Size.ugt(DstSize)) 965 break; 966 Address Dest = EmitPointerWithAlignment(E->getArg(0)); 967 Address Src = EmitPointerWithAlignment(E->getArg(1)); 968 Value *SizeVal = llvm::ConstantInt::get(Builder.getContext(), Size); 969 Builder.CreateMemMove(Dest, Src, SizeVal, false); 970 return RValue::get(Dest.getPointer()); 971 } 972 973 case Builtin::BImemmove: 974 case Builtin::BI__builtin_memmove: { 975 Address Dest = EmitPointerWithAlignment(E->getArg(0)); 976 Address Src = EmitPointerWithAlignment(E->getArg(1)); 977 Value *SizeVal = EmitScalarExpr(E->getArg(2)); 978 EmitNonNullArgCheck(RValue::get(Dest.getPointer()), E->getArg(0)->getType(), 979 E->getArg(0)->getExprLoc(), FD, 0); 980 EmitNonNullArgCheck(RValue::get(Src.getPointer()), E->getArg(1)->getType(), 981 E->getArg(1)->getExprLoc(), FD, 1); 982 Builder.CreateMemMove(Dest, Src, SizeVal, false); 983 return RValue::get(Dest.getPointer()); 984 } 985 case Builtin::BImemset: 986 case Builtin::BI__builtin_memset: { 987 Address Dest = EmitPointerWithAlignment(E->getArg(0)); 988 Value *ByteVal = Builder.CreateTrunc(EmitScalarExpr(E->getArg(1)), 989 Builder.getInt8Ty()); 990 Value *SizeVal = EmitScalarExpr(E->getArg(2)); 991 EmitNonNullArgCheck(RValue::get(Dest.getPointer()), E->getArg(0)->getType(), 992 E->getArg(0)->getExprLoc(), FD, 0); 993 Builder.CreateMemSet(Dest, ByteVal, SizeVal, false); 994 return RValue::get(Dest.getPointer()); 995 } 996 case Builtin::BI__builtin___memset_chk: { 997 // fold __builtin_memset_chk(x, y, cst1, cst2) to memset iff cst1<=cst2. 998 llvm::APSInt Size, DstSize; 999 if (!E->getArg(2)->EvaluateAsInt(Size, CGM.getContext()) || 1000 !E->getArg(3)->EvaluateAsInt(DstSize, CGM.getContext())) 1001 break; 1002 if (Size.ugt(DstSize)) 1003 break; 1004 Address Dest = EmitPointerWithAlignment(E->getArg(0)); 1005 Value *ByteVal = Builder.CreateTrunc(EmitScalarExpr(E->getArg(1)), 1006 Builder.getInt8Ty()); 1007 Value *SizeVal = llvm::ConstantInt::get(Builder.getContext(), Size); 1008 Builder.CreateMemSet(Dest, ByteVal, SizeVal, false); 1009 return RValue::get(Dest.getPointer()); 1010 } 1011 case Builtin::BI__builtin_dwarf_cfa: { 1012 // The offset in bytes from the first argument to the CFA. 1013 // 1014 // Why on earth is this in the frontend? Is there any reason at 1015 // all that the backend can't reasonably determine this while 1016 // lowering llvm.eh.dwarf.cfa()? 1017 // 1018 // TODO: If there's a satisfactory reason, add a target hook for 1019 // this instead of hard-coding 0, which is correct for most targets. 1020 int32_t Offset = 0; 1021 1022 Value *F = CGM.getIntrinsic(Intrinsic::eh_dwarf_cfa); 1023 return RValue::get(Builder.CreateCall(F, 1024 llvm::ConstantInt::get(Int32Ty, Offset))); 1025 } 1026 case Builtin::BI__builtin_return_address: { 1027 Value *Depth = 1028 CGM.EmitConstantExpr(E->getArg(0), getContext().UnsignedIntTy, this); 1029 Value *F = CGM.getIntrinsic(Intrinsic::returnaddress); 1030 return RValue::get(Builder.CreateCall(F, Depth)); 1031 } 1032 case Builtin::BI__builtin_frame_address: { 1033 Value *Depth = 1034 CGM.EmitConstantExpr(E->getArg(0), getContext().UnsignedIntTy, this); 1035 Value *F = CGM.getIntrinsic(Intrinsic::frameaddress); 1036 return RValue::get(Builder.CreateCall(F, Depth)); 1037 } 1038 case Builtin::BI__builtin_extract_return_addr: { 1039 Value *Address = EmitScalarExpr(E->getArg(0)); 1040 Value *Result = getTargetHooks().decodeReturnAddress(*this, Address); 1041 return RValue::get(Result); 1042 } 1043 case Builtin::BI__builtin_frob_return_addr: { 1044 Value *Address = EmitScalarExpr(E->getArg(0)); 1045 Value *Result = getTargetHooks().encodeReturnAddress(*this, Address); 1046 return RValue::get(Result); 1047 } 1048 case Builtin::BI__builtin_dwarf_sp_column: { 1049 llvm::IntegerType *Ty 1050 = cast<llvm::IntegerType>(ConvertType(E->getType())); 1051 int Column = getTargetHooks().getDwarfEHStackPointer(CGM); 1052 if (Column == -1) { 1053 CGM.ErrorUnsupported(E, "__builtin_dwarf_sp_column"); 1054 return RValue::get(llvm::UndefValue::get(Ty)); 1055 } 1056 return RValue::get(llvm::ConstantInt::get(Ty, Column, true)); 1057 } 1058 case Builtin::BI__builtin_init_dwarf_reg_size_table: { 1059 Value *Address = EmitScalarExpr(E->getArg(0)); 1060 if (getTargetHooks().initDwarfEHRegSizeTable(*this, Address)) 1061 CGM.ErrorUnsupported(E, "__builtin_init_dwarf_reg_size_table"); 1062 return RValue::get(llvm::UndefValue::get(ConvertType(E->getType()))); 1063 } 1064 case Builtin::BI__builtin_eh_return: { 1065 Value *Int = EmitScalarExpr(E->getArg(0)); 1066 Value *Ptr = EmitScalarExpr(E->getArg(1)); 1067 1068 llvm::IntegerType *IntTy = cast<llvm::IntegerType>(Int->getType()); 1069 assert((IntTy->getBitWidth() == 32 || IntTy->getBitWidth() == 64) && 1070 "LLVM's __builtin_eh_return only supports 32- and 64-bit variants"); 1071 Value *F = CGM.getIntrinsic(IntTy->getBitWidth() == 32 1072 ? Intrinsic::eh_return_i32 1073 : Intrinsic::eh_return_i64); 1074 Builder.CreateCall(F, {Int, Ptr}); 1075 Builder.CreateUnreachable(); 1076 1077 // We do need to preserve an insertion point. 1078 EmitBlock(createBasicBlock("builtin_eh_return.cont")); 1079 1080 return RValue::get(nullptr); 1081 } 1082 case Builtin::BI__builtin_unwind_init: { 1083 Value *F = CGM.getIntrinsic(Intrinsic::eh_unwind_init); 1084 return RValue::get(Builder.CreateCall(F)); 1085 } 1086 case Builtin::BI__builtin_extend_pointer: { 1087 // Extends a pointer to the size of an _Unwind_Word, which is 1088 // uint64_t on all platforms. Generally this gets poked into a 1089 // register and eventually used as an address, so if the 1090 // addressing registers are wider than pointers and the platform 1091 // doesn't implicitly ignore high-order bits when doing 1092 // addressing, we need to make sure we zext / sext based on 1093 // the platform's expectations. 1094 // 1095 // See: http://gcc.gnu.org/ml/gcc-bugs/2002-02/msg00237.html 1096 1097 // Cast the pointer to intptr_t. 1098 Value *Ptr = EmitScalarExpr(E->getArg(0)); 1099 Value *Result = Builder.CreatePtrToInt(Ptr, IntPtrTy, "extend.cast"); 1100 1101 // If that's 64 bits, we're done. 1102 if (IntPtrTy->getBitWidth() == 64) 1103 return RValue::get(Result); 1104 1105 // Otherwise, ask the codegen data what to do. 1106 if (getTargetHooks().extendPointerWithSExt()) 1107 return RValue::get(Builder.CreateSExt(Result, Int64Ty, "extend.sext")); 1108 else 1109 return RValue::get(Builder.CreateZExt(Result, Int64Ty, "extend.zext")); 1110 } 1111 case Builtin::BI__builtin_setjmp: { 1112 // Buffer is a void**. 1113 Address Buf = EmitPointerWithAlignment(E->getArg(0)); 1114 1115 // Store the frame pointer to the setjmp buffer. 1116 Value *FrameAddr = 1117 Builder.CreateCall(CGM.getIntrinsic(Intrinsic::frameaddress), 1118 ConstantInt::get(Int32Ty, 0)); 1119 Builder.CreateStore(FrameAddr, Buf); 1120 1121 // Store the stack pointer to the setjmp buffer. 1122 Value *StackAddr = 1123 Builder.CreateCall(CGM.getIntrinsic(Intrinsic::stacksave)); 1124 Address StackSaveSlot = 1125 Builder.CreateConstInBoundsGEP(Buf, 2, getPointerSize()); 1126 Builder.CreateStore(StackAddr, StackSaveSlot); 1127 1128 // Call LLVM's EH setjmp, which is lightweight. 1129 Value *F = CGM.getIntrinsic(Intrinsic::eh_sjlj_setjmp); 1130 Buf = Builder.CreateBitCast(Buf, Int8PtrTy); 1131 return RValue::get(Builder.CreateCall(F, Buf.getPointer())); 1132 } 1133 case Builtin::BI__builtin_longjmp: { 1134 Value *Buf = EmitScalarExpr(E->getArg(0)); 1135 Buf = Builder.CreateBitCast(Buf, Int8PtrTy); 1136 1137 // Call LLVM's EH longjmp, which is lightweight. 1138 Builder.CreateCall(CGM.getIntrinsic(Intrinsic::eh_sjlj_longjmp), Buf); 1139 1140 // longjmp doesn't return; mark this as unreachable. 1141 Builder.CreateUnreachable(); 1142 1143 // We do need to preserve an insertion point. 1144 EmitBlock(createBasicBlock("longjmp.cont")); 1145 1146 return RValue::get(nullptr); 1147 } 1148 case Builtin::BI__sync_fetch_and_add: 1149 case Builtin::BI__sync_fetch_and_sub: 1150 case Builtin::BI__sync_fetch_and_or: 1151 case Builtin::BI__sync_fetch_and_and: 1152 case Builtin::BI__sync_fetch_and_xor: 1153 case Builtin::BI__sync_fetch_and_nand: 1154 case Builtin::BI__sync_add_and_fetch: 1155 case Builtin::BI__sync_sub_and_fetch: 1156 case Builtin::BI__sync_and_and_fetch: 1157 case Builtin::BI__sync_or_and_fetch: 1158 case Builtin::BI__sync_xor_and_fetch: 1159 case Builtin::BI__sync_nand_and_fetch: 1160 case Builtin::BI__sync_val_compare_and_swap: 1161 case Builtin::BI__sync_bool_compare_and_swap: 1162 case Builtin::BI__sync_lock_test_and_set: 1163 case Builtin::BI__sync_lock_release: 1164 case Builtin::BI__sync_swap: 1165 llvm_unreachable("Shouldn't make it through sema"); 1166 case Builtin::BI__sync_fetch_and_add_1: 1167 case Builtin::BI__sync_fetch_and_add_2: 1168 case Builtin::BI__sync_fetch_and_add_4: 1169 case Builtin::BI__sync_fetch_and_add_8: 1170 case Builtin::BI__sync_fetch_and_add_16: 1171 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Add, E); 1172 case Builtin::BI__sync_fetch_and_sub_1: 1173 case Builtin::BI__sync_fetch_and_sub_2: 1174 case Builtin::BI__sync_fetch_and_sub_4: 1175 case Builtin::BI__sync_fetch_and_sub_8: 1176 case Builtin::BI__sync_fetch_and_sub_16: 1177 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Sub, E); 1178 case Builtin::BI__sync_fetch_and_or_1: 1179 case Builtin::BI__sync_fetch_and_or_2: 1180 case Builtin::BI__sync_fetch_and_or_4: 1181 case Builtin::BI__sync_fetch_and_or_8: 1182 case Builtin::BI__sync_fetch_and_or_16: 1183 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Or, E); 1184 case Builtin::BI__sync_fetch_and_and_1: 1185 case Builtin::BI__sync_fetch_and_and_2: 1186 case Builtin::BI__sync_fetch_and_and_4: 1187 case Builtin::BI__sync_fetch_and_and_8: 1188 case Builtin::BI__sync_fetch_and_and_16: 1189 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::And, E); 1190 case Builtin::BI__sync_fetch_and_xor_1: 1191 case Builtin::BI__sync_fetch_and_xor_2: 1192 case Builtin::BI__sync_fetch_and_xor_4: 1193 case Builtin::BI__sync_fetch_and_xor_8: 1194 case Builtin::BI__sync_fetch_and_xor_16: 1195 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Xor, E); 1196 case Builtin::BI__sync_fetch_and_nand_1: 1197 case Builtin::BI__sync_fetch_and_nand_2: 1198 case Builtin::BI__sync_fetch_and_nand_4: 1199 case Builtin::BI__sync_fetch_and_nand_8: 1200 case Builtin::BI__sync_fetch_and_nand_16: 1201 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Nand, E); 1202 1203 // Clang extensions: not overloaded yet. 1204 case Builtin::BI__sync_fetch_and_min: 1205 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Min, E); 1206 case Builtin::BI__sync_fetch_and_max: 1207 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Max, E); 1208 case Builtin::BI__sync_fetch_and_umin: 1209 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::UMin, E); 1210 case Builtin::BI__sync_fetch_and_umax: 1211 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::UMax, E); 1212 1213 case Builtin::BI__sync_add_and_fetch_1: 1214 case Builtin::BI__sync_add_and_fetch_2: 1215 case Builtin::BI__sync_add_and_fetch_4: 1216 case Builtin::BI__sync_add_and_fetch_8: 1217 case Builtin::BI__sync_add_and_fetch_16: 1218 return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::Add, E, 1219 llvm::Instruction::Add); 1220 case Builtin::BI__sync_sub_and_fetch_1: 1221 case Builtin::BI__sync_sub_and_fetch_2: 1222 case Builtin::BI__sync_sub_and_fetch_4: 1223 case Builtin::BI__sync_sub_and_fetch_8: 1224 case Builtin::BI__sync_sub_and_fetch_16: 1225 return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::Sub, E, 1226 llvm::Instruction::Sub); 1227 case Builtin::BI__sync_and_and_fetch_1: 1228 case Builtin::BI__sync_and_and_fetch_2: 1229 case Builtin::BI__sync_and_and_fetch_4: 1230 case Builtin::BI__sync_and_and_fetch_8: 1231 case Builtin::BI__sync_and_and_fetch_16: 1232 return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::And, E, 1233 llvm::Instruction::And); 1234 case Builtin::BI__sync_or_and_fetch_1: 1235 case Builtin::BI__sync_or_and_fetch_2: 1236 case Builtin::BI__sync_or_and_fetch_4: 1237 case Builtin::BI__sync_or_and_fetch_8: 1238 case Builtin::BI__sync_or_and_fetch_16: 1239 return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::Or, E, 1240 llvm::Instruction::Or); 1241 case Builtin::BI__sync_xor_and_fetch_1: 1242 case Builtin::BI__sync_xor_and_fetch_2: 1243 case Builtin::BI__sync_xor_and_fetch_4: 1244 case Builtin::BI__sync_xor_and_fetch_8: 1245 case Builtin::BI__sync_xor_and_fetch_16: 1246 return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::Xor, E, 1247 llvm::Instruction::Xor); 1248 case Builtin::BI__sync_nand_and_fetch_1: 1249 case Builtin::BI__sync_nand_and_fetch_2: 1250 case Builtin::BI__sync_nand_and_fetch_4: 1251 case Builtin::BI__sync_nand_and_fetch_8: 1252 case Builtin::BI__sync_nand_and_fetch_16: 1253 return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::Nand, E, 1254 llvm::Instruction::And, true); 1255 1256 case Builtin::BI__sync_val_compare_and_swap_1: 1257 case Builtin::BI__sync_val_compare_and_swap_2: 1258 case Builtin::BI__sync_val_compare_and_swap_4: 1259 case Builtin::BI__sync_val_compare_and_swap_8: 1260 case Builtin::BI__sync_val_compare_and_swap_16: 1261 return RValue::get(MakeAtomicCmpXchgValue(*this, E, false)); 1262 1263 case Builtin::BI__sync_bool_compare_and_swap_1: 1264 case Builtin::BI__sync_bool_compare_and_swap_2: 1265 case Builtin::BI__sync_bool_compare_and_swap_4: 1266 case Builtin::BI__sync_bool_compare_and_swap_8: 1267 case Builtin::BI__sync_bool_compare_and_swap_16: 1268 return RValue::get(MakeAtomicCmpXchgValue(*this, E, true)); 1269 1270 case Builtin::BI__sync_swap_1: 1271 case Builtin::BI__sync_swap_2: 1272 case Builtin::BI__sync_swap_4: 1273 case Builtin::BI__sync_swap_8: 1274 case Builtin::BI__sync_swap_16: 1275 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Xchg, E); 1276 1277 case Builtin::BI__sync_lock_test_and_set_1: 1278 case Builtin::BI__sync_lock_test_and_set_2: 1279 case Builtin::BI__sync_lock_test_and_set_4: 1280 case Builtin::BI__sync_lock_test_and_set_8: 1281 case Builtin::BI__sync_lock_test_and_set_16: 1282 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Xchg, E); 1283 1284 case Builtin::BI__sync_lock_release_1: 1285 case Builtin::BI__sync_lock_release_2: 1286 case Builtin::BI__sync_lock_release_4: 1287 case Builtin::BI__sync_lock_release_8: 1288 case Builtin::BI__sync_lock_release_16: { 1289 Value *Ptr = EmitScalarExpr(E->getArg(0)); 1290 QualType ElTy = E->getArg(0)->getType()->getPointeeType(); 1291 CharUnits StoreSize = getContext().getTypeSizeInChars(ElTy); 1292 llvm::Type *ITy = llvm::IntegerType::get(getLLVMContext(), 1293 StoreSize.getQuantity() * 8); 1294 Ptr = Builder.CreateBitCast(Ptr, ITy->getPointerTo()); 1295 llvm::StoreInst *Store = 1296 Builder.CreateAlignedStore(llvm::Constant::getNullValue(ITy), Ptr, 1297 StoreSize); 1298 Store->setAtomic(llvm::Release); 1299 return RValue::get(nullptr); 1300 } 1301 1302 case Builtin::BI__sync_synchronize: { 1303 // We assume this is supposed to correspond to a C++0x-style 1304 // sequentially-consistent fence (i.e. this is only usable for 1305 // synchonization, not device I/O or anything like that). This intrinsic 1306 // is really badly designed in the sense that in theory, there isn't 1307 // any way to safely use it... but in practice, it mostly works 1308 // to use it with non-atomic loads and stores to get acquire/release 1309 // semantics. 1310 Builder.CreateFence(llvm::SequentiallyConsistent); 1311 return RValue::get(nullptr); 1312 } 1313 1314 case Builtin::BI__builtin_nontemporal_load: 1315 return RValue::get(EmitNontemporalLoad(*this, E)); 1316 case Builtin::BI__builtin_nontemporal_store: 1317 return RValue::get(EmitNontemporalStore(*this, E)); 1318 case Builtin::BI__c11_atomic_is_lock_free: 1319 case Builtin::BI__atomic_is_lock_free: { 1320 // Call "bool __atomic_is_lock_free(size_t size, void *ptr)". For the 1321 // __c11 builtin, ptr is 0 (indicating a properly-aligned object), since 1322 // _Atomic(T) is always properly-aligned. 1323 const char *LibCallName = "__atomic_is_lock_free"; 1324 CallArgList Args; 1325 Args.add(RValue::get(EmitScalarExpr(E->getArg(0))), 1326 getContext().getSizeType()); 1327 if (BuiltinID == Builtin::BI__atomic_is_lock_free) 1328 Args.add(RValue::get(EmitScalarExpr(E->getArg(1))), 1329 getContext().VoidPtrTy); 1330 else 1331 Args.add(RValue::get(llvm::Constant::getNullValue(VoidPtrTy)), 1332 getContext().VoidPtrTy); 1333 const CGFunctionInfo &FuncInfo = 1334 CGM.getTypes().arrangeBuiltinFunctionCall(E->getType(), Args); 1335 llvm::FunctionType *FTy = CGM.getTypes().GetFunctionType(FuncInfo); 1336 llvm::Constant *Func = CGM.CreateRuntimeFunction(FTy, LibCallName); 1337 return EmitCall(FuncInfo, Func, ReturnValueSlot(), Args); 1338 } 1339 1340 case Builtin::BI__atomic_test_and_set: { 1341 // Look at the argument type to determine whether this is a volatile 1342 // operation. The parameter type is always volatile. 1343 QualType PtrTy = E->getArg(0)->IgnoreImpCasts()->getType(); 1344 bool Volatile = 1345 PtrTy->castAs<PointerType>()->getPointeeType().isVolatileQualified(); 1346 1347 Value *Ptr = EmitScalarExpr(E->getArg(0)); 1348 unsigned AddrSpace = Ptr->getType()->getPointerAddressSpace(); 1349 Ptr = Builder.CreateBitCast(Ptr, Int8Ty->getPointerTo(AddrSpace)); 1350 Value *NewVal = Builder.getInt8(1); 1351 Value *Order = EmitScalarExpr(E->getArg(1)); 1352 if (isa<llvm::ConstantInt>(Order)) { 1353 int ord = cast<llvm::ConstantInt>(Order)->getZExtValue(); 1354 AtomicRMWInst *Result = nullptr; 1355 switch (ord) { 1356 case 0: // memory_order_relaxed 1357 default: // invalid order 1358 Result = Builder.CreateAtomicRMW(llvm::AtomicRMWInst::Xchg, 1359 Ptr, NewVal, 1360 llvm::Monotonic); 1361 break; 1362 case 1: // memory_order_consume 1363 case 2: // memory_order_acquire 1364 Result = Builder.CreateAtomicRMW(llvm::AtomicRMWInst::Xchg, 1365 Ptr, NewVal, 1366 llvm::Acquire); 1367 break; 1368 case 3: // memory_order_release 1369 Result = Builder.CreateAtomicRMW(llvm::AtomicRMWInst::Xchg, 1370 Ptr, NewVal, 1371 llvm::Release); 1372 break; 1373 case 4: // memory_order_acq_rel 1374 Result = Builder.CreateAtomicRMW(llvm::AtomicRMWInst::Xchg, 1375 Ptr, NewVal, 1376 llvm::AcquireRelease); 1377 break; 1378 case 5: // memory_order_seq_cst 1379 Result = Builder.CreateAtomicRMW(llvm::AtomicRMWInst::Xchg, 1380 Ptr, NewVal, 1381 llvm::SequentiallyConsistent); 1382 break; 1383 } 1384 Result->setVolatile(Volatile); 1385 return RValue::get(Builder.CreateIsNotNull(Result, "tobool")); 1386 } 1387 1388 llvm::BasicBlock *ContBB = createBasicBlock("atomic.continue", CurFn); 1389 1390 llvm::BasicBlock *BBs[5] = { 1391 createBasicBlock("monotonic", CurFn), 1392 createBasicBlock("acquire", CurFn), 1393 createBasicBlock("release", CurFn), 1394 createBasicBlock("acqrel", CurFn), 1395 createBasicBlock("seqcst", CurFn) 1396 }; 1397 llvm::AtomicOrdering Orders[5] = { 1398 llvm::Monotonic, llvm::Acquire, llvm::Release, 1399 llvm::AcquireRelease, llvm::SequentiallyConsistent 1400 }; 1401 1402 Order = Builder.CreateIntCast(Order, Builder.getInt32Ty(), false); 1403 llvm::SwitchInst *SI = Builder.CreateSwitch(Order, BBs[0]); 1404 1405 Builder.SetInsertPoint(ContBB); 1406 PHINode *Result = Builder.CreatePHI(Int8Ty, 5, "was_set"); 1407 1408 for (unsigned i = 0; i < 5; ++i) { 1409 Builder.SetInsertPoint(BBs[i]); 1410 AtomicRMWInst *RMW = Builder.CreateAtomicRMW(llvm::AtomicRMWInst::Xchg, 1411 Ptr, NewVal, Orders[i]); 1412 RMW->setVolatile(Volatile); 1413 Result->addIncoming(RMW, BBs[i]); 1414 Builder.CreateBr(ContBB); 1415 } 1416 1417 SI->addCase(Builder.getInt32(0), BBs[0]); 1418 SI->addCase(Builder.getInt32(1), BBs[1]); 1419 SI->addCase(Builder.getInt32(2), BBs[1]); 1420 SI->addCase(Builder.getInt32(3), BBs[2]); 1421 SI->addCase(Builder.getInt32(4), BBs[3]); 1422 SI->addCase(Builder.getInt32(5), BBs[4]); 1423 1424 Builder.SetInsertPoint(ContBB); 1425 return RValue::get(Builder.CreateIsNotNull(Result, "tobool")); 1426 } 1427 1428 case Builtin::BI__atomic_clear: { 1429 QualType PtrTy = E->getArg(0)->IgnoreImpCasts()->getType(); 1430 bool Volatile = 1431 PtrTy->castAs<PointerType>()->getPointeeType().isVolatileQualified(); 1432 1433 Address Ptr = EmitPointerWithAlignment(E->getArg(0)); 1434 unsigned AddrSpace = Ptr.getPointer()->getType()->getPointerAddressSpace(); 1435 Ptr = Builder.CreateBitCast(Ptr, Int8Ty->getPointerTo(AddrSpace)); 1436 Value *NewVal = Builder.getInt8(0); 1437 Value *Order = EmitScalarExpr(E->getArg(1)); 1438 if (isa<llvm::ConstantInt>(Order)) { 1439 int ord = cast<llvm::ConstantInt>(Order)->getZExtValue(); 1440 StoreInst *Store = Builder.CreateStore(NewVal, Ptr, Volatile); 1441 switch (ord) { 1442 case 0: // memory_order_relaxed 1443 default: // invalid order 1444 Store->setOrdering(llvm::Monotonic); 1445 break; 1446 case 3: // memory_order_release 1447 Store->setOrdering(llvm::Release); 1448 break; 1449 case 5: // memory_order_seq_cst 1450 Store->setOrdering(llvm::SequentiallyConsistent); 1451 break; 1452 } 1453 return RValue::get(nullptr); 1454 } 1455 1456 llvm::BasicBlock *ContBB = createBasicBlock("atomic.continue", CurFn); 1457 1458 llvm::BasicBlock *BBs[3] = { 1459 createBasicBlock("monotonic", CurFn), 1460 createBasicBlock("release", CurFn), 1461 createBasicBlock("seqcst", CurFn) 1462 }; 1463 llvm::AtomicOrdering Orders[3] = { 1464 llvm::Monotonic, llvm::Release, llvm::SequentiallyConsistent 1465 }; 1466 1467 Order = Builder.CreateIntCast(Order, Builder.getInt32Ty(), false); 1468 llvm::SwitchInst *SI = Builder.CreateSwitch(Order, BBs[0]); 1469 1470 for (unsigned i = 0; i < 3; ++i) { 1471 Builder.SetInsertPoint(BBs[i]); 1472 StoreInst *Store = Builder.CreateStore(NewVal, Ptr, Volatile); 1473 Store->setOrdering(Orders[i]); 1474 Builder.CreateBr(ContBB); 1475 } 1476 1477 SI->addCase(Builder.getInt32(0), BBs[0]); 1478 SI->addCase(Builder.getInt32(3), BBs[1]); 1479 SI->addCase(Builder.getInt32(5), BBs[2]); 1480 1481 Builder.SetInsertPoint(ContBB); 1482 return RValue::get(nullptr); 1483 } 1484 1485 case Builtin::BI__atomic_thread_fence: 1486 case Builtin::BI__atomic_signal_fence: 1487 case Builtin::BI__c11_atomic_thread_fence: 1488 case Builtin::BI__c11_atomic_signal_fence: { 1489 llvm::SynchronizationScope Scope; 1490 if (BuiltinID == Builtin::BI__atomic_signal_fence || 1491 BuiltinID == Builtin::BI__c11_atomic_signal_fence) 1492 Scope = llvm::SingleThread; 1493 else 1494 Scope = llvm::CrossThread; 1495 Value *Order = EmitScalarExpr(E->getArg(0)); 1496 if (isa<llvm::ConstantInt>(Order)) { 1497 int ord = cast<llvm::ConstantInt>(Order)->getZExtValue(); 1498 switch (ord) { 1499 case 0: // memory_order_relaxed 1500 default: // invalid order 1501 break; 1502 case 1: // memory_order_consume 1503 case 2: // memory_order_acquire 1504 Builder.CreateFence(llvm::Acquire, Scope); 1505 break; 1506 case 3: // memory_order_release 1507 Builder.CreateFence(llvm::Release, Scope); 1508 break; 1509 case 4: // memory_order_acq_rel 1510 Builder.CreateFence(llvm::AcquireRelease, Scope); 1511 break; 1512 case 5: // memory_order_seq_cst 1513 Builder.CreateFence(llvm::SequentiallyConsistent, Scope); 1514 break; 1515 } 1516 return RValue::get(nullptr); 1517 } 1518 1519 llvm::BasicBlock *AcquireBB, *ReleaseBB, *AcqRelBB, *SeqCstBB; 1520 AcquireBB = createBasicBlock("acquire", CurFn); 1521 ReleaseBB = createBasicBlock("release", CurFn); 1522 AcqRelBB = createBasicBlock("acqrel", CurFn); 1523 SeqCstBB = createBasicBlock("seqcst", CurFn); 1524 llvm::BasicBlock *ContBB = createBasicBlock("atomic.continue", CurFn); 1525 1526 Order = Builder.CreateIntCast(Order, Builder.getInt32Ty(), false); 1527 llvm::SwitchInst *SI = Builder.CreateSwitch(Order, ContBB); 1528 1529 Builder.SetInsertPoint(AcquireBB); 1530 Builder.CreateFence(llvm::Acquire, Scope); 1531 Builder.CreateBr(ContBB); 1532 SI->addCase(Builder.getInt32(1), AcquireBB); 1533 SI->addCase(Builder.getInt32(2), AcquireBB); 1534 1535 Builder.SetInsertPoint(ReleaseBB); 1536 Builder.CreateFence(llvm::Release, Scope); 1537 Builder.CreateBr(ContBB); 1538 SI->addCase(Builder.getInt32(3), ReleaseBB); 1539 1540 Builder.SetInsertPoint(AcqRelBB); 1541 Builder.CreateFence(llvm::AcquireRelease, Scope); 1542 Builder.CreateBr(ContBB); 1543 SI->addCase(Builder.getInt32(4), AcqRelBB); 1544 1545 Builder.SetInsertPoint(SeqCstBB); 1546 Builder.CreateFence(llvm::SequentiallyConsistent, Scope); 1547 Builder.CreateBr(ContBB); 1548 SI->addCase(Builder.getInt32(5), SeqCstBB); 1549 1550 Builder.SetInsertPoint(ContBB); 1551 return RValue::get(nullptr); 1552 } 1553 1554 // Library functions with special handling. 1555 case Builtin::BIsqrt: 1556 case Builtin::BIsqrtf: 1557 case Builtin::BIsqrtl: { 1558 // Transform a call to sqrt* into a @llvm.sqrt.* intrinsic call, but only 1559 // in finite- or unsafe-math mode (the intrinsic has different semantics 1560 // for handling negative numbers compared to the library function, so 1561 // -fmath-errno=0 is not enough). 1562 if (!FD->hasAttr<ConstAttr>()) 1563 break; 1564 if (!(CGM.getCodeGenOpts().UnsafeFPMath || 1565 CGM.getCodeGenOpts().NoNaNsFPMath)) 1566 break; 1567 Value *Arg0 = EmitScalarExpr(E->getArg(0)); 1568 llvm::Type *ArgType = Arg0->getType(); 1569 Value *F = CGM.getIntrinsic(Intrinsic::sqrt, ArgType); 1570 return RValue::get(Builder.CreateCall(F, Arg0)); 1571 } 1572 1573 case Builtin::BI__builtin_pow: 1574 case Builtin::BI__builtin_powf: 1575 case Builtin::BI__builtin_powl: 1576 case Builtin::BIpow: 1577 case Builtin::BIpowf: 1578 case Builtin::BIpowl: { 1579 // Transform a call to pow* into a @llvm.pow.* intrinsic call. 1580 if (!FD->hasAttr<ConstAttr>()) 1581 break; 1582 Value *Base = EmitScalarExpr(E->getArg(0)); 1583 Value *Exponent = EmitScalarExpr(E->getArg(1)); 1584 llvm::Type *ArgType = Base->getType(); 1585 Value *F = CGM.getIntrinsic(Intrinsic::pow, ArgType); 1586 return RValue::get(Builder.CreateCall(F, {Base, Exponent})); 1587 } 1588 1589 case Builtin::BIfma: 1590 case Builtin::BIfmaf: 1591 case Builtin::BIfmal: 1592 case Builtin::BI__builtin_fma: 1593 case Builtin::BI__builtin_fmaf: 1594 case Builtin::BI__builtin_fmal: { 1595 // Rewrite fma to intrinsic. 1596 Value *FirstArg = EmitScalarExpr(E->getArg(0)); 1597 llvm::Type *ArgType = FirstArg->getType(); 1598 Value *F = CGM.getIntrinsic(Intrinsic::fma, ArgType); 1599 return RValue::get( 1600 Builder.CreateCall(F, {FirstArg, EmitScalarExpr(E->getArg(1)), 1601 EmitScalarExpr(E->getArg(2))})); 1602 } 1603 1604 case Builtin::BI__builtin_signbit: 1605 case Builtin::BI__builtin_signbitf: 1606 case Builtin::BI__builtin_signbitl: { 1607 return RValue::get( 1608 Builder.CreateZExt(EmitSignBit(*this, EmitScalarExpr(E->getArg(0))), 1609 ConvertType(E->getType()))); 1610 } 1611 case Builtin::BI__builtin_annotation: { 1612 llvm::Value *AnnVal = EmitScalarExpr(E->getArg(0)); 1613 llvm::Value *F = CGM.getIntrinsic(llvm::Intrinsic::annotation, 1614 AnnVal->getType()); 1615 1616 // Get the annotation string, go through casts. Sema requires this to be a 1617 // non-wide string literal, potentially casted, so the cast<> is safe. 1618 const Expr *AnnotationStrExpr = E->getArg(1)->IgnoreParenCasts(); 1619 StringRef Str = cast<StringLiteral>(AnnotationStrExpr)->getString(); 1620 return RValue::get(EmitAnnotationCall(F, AnnVal, Str, E->getExprLoc())); 1621 } 1622 case Builtin::BI__builtin_addcb: 1623 case Builtin::BI__builtin_addcs: 1624 case Builtin::BI__builtin_addc: 1625 case Builtin::BI__builtin_addcl: 1626 case Builtin::BI__builtin_addcll: 1627 case Builtin::BI__builtin_subcb: 1628 case Builtin::BI__builtin_subcs: 1629 case Builtin::BI__builtin_subc: 1630 case Builtin::BI__builtin_subcl: 1631 case Builtin::BI__builtin_subcll: { 1632 1633 // We translate all of these builtins from expressions of the form: 1634 // int x = ..., y = ..., carryin = ..., carryout, result; 1635 // result = __builtin_addc(x, y, carryin, &carryout); 1636 // 1637 // to LLVM IR of the form: 1638 // 1639 // %tmp1 = call {i32, i1} @llvm.uadd.with.overflow.i32(i32 %x, i32 %y) 1640 // %tmpsum1 = extractvalue {i32, i1} %tmp1, 0 1641 // %carry1 = extractvalue {i32, i1} %tmp1, 1 1642 // %tmp2 = call {i32, i1} @llvm.uadd.with.overflow.i32(i32 %tmpsum1, 1643 // i32 %carryin) 1644 // %result = extractvalue {i32, i1} %tmp2, 0 1645 // %carry2 = extractvalue {i32, i1} %tmp2, 1 1646 // %tmp3 = or i1 %carry1, %carry2 1647 // %tmp4 = zext i1 %tmp3 to i32 1648 // store i32 %tmp4, i32* %carryout 1649 1650 // Scalarize our inputs. 1651 llvm::Value *X = EmitScalarExpr(E->getArg(0)); 1652 llvm::Value *Y = EmitScalarExpr(E->getArg(1)); 1653 llvm::Value *Carryin = EmitScalarExpr(E->getArg(2)); 1654 Address CarryOutPtr = EmitPointerWithAlignment(E->getArg(3)); 1655 1656 // Decide if we are lowering to a uadd.with.overflow or usub.with.overflow. 1657 llvm::Intrinsic::ID IntrinsicId; 1658 switch (BuiltinID) { 1659 default: llvm_unreachable("Unknown multiprecision builtin id."); 1660 case Builtin::BI__builtin_addcb: 1661 case Builtin::BI__builtin_addcs: 1662 case Builtin::BI__builtin_addc: 1663 case Builtin::BI__builtin_addcl: 1664 case Builtin::BI__builtin_addcll: 1665 IntrinsicId = llvm::Intrinsic::uadd_with_overflow; 1666 break; 1667 case Builtin::BI__builtin_subcb: 1668 case Builtin::BI__builtin_subcs: 1669 case Builtin::BI__builtin_subc: 1670 case Builtin::BI__builtin_subcl: 1671 case Builtin::BI__builtin_subcll: 1672 IntrinsicId = llvm::Intrinsic::usub_with_overflow; 1673 break; 1674 } 1675 1676 // Construct our resulting LLVM IR expression. 1677 llvm::Value *Carry1; 1678 llvm::Value *Sum1 = EmitOverflowIntrinsic(*this, IntrinsicId, 1679 X, Y, Carry1); 1680 llvm::Value *Carry2; 1681 llvm::Value *Sum2 = EmitOverflowIntrinsic(*this, IntrinsicId, 1682 Sum1, Carryin, Carry2); 1683 llvm::Value *CarryOut = Builder.CreateZExt(Builder.CreateOr(Carry1, Carry2), 1684 X->getType()); 1685 Builder.CreateStore(CarryOut, CarryOutPtr); 1686 return RValue::get(Sum2); 1687 } 1688 1689 case Builtin::BI__builtin_add_overflow: 1690 case Builtin::BI__builtin_sub_overflow: 1691 case Builtin::BI__builtin_mul_overflow: { 1692 const clang::Expr *LeftArg = E->getArg(0); 1693 const clang::Expr *RightArg = E->getArg(1); 1694 const clang::Expr *ResultArg = E->getArg(2); 1695 1696 clang::QualType ResultQTy = 1697 ResultArg->getType()->castAs<PointerType>()->getPointeeType(); 1698 1699 WidthAndSignedness LeftInfo = 1700 getIntegerWidthAndSignedness(CGM.getContext(), LeftArg->getType()); 1701 WidthAndSignedness RightInfo = 1702 getIntegerWidthAndSignedness(CGM.getContext(), RightArg->getType()); 1703 WidthAndSignedness ResultInfo = 1704 getIntegerWidthAndSignedness(CGM.getContext(), ResultQTy); 1705 WidthAndSignedness EncompassingInfo = 1706 EncompassingIntegerType({LeftInfo, RightInfo, ResultInfo}); 1707 1708 llvm::Type *EncompassingLLVMTy = 1709 llvm::IntegerType::get(CGM.getLLVMContext(), EncompassingInfo.Width); 1710 1711 llvm::Type *ResultLLVMTy = CGM.getTypes().ConvertType(ResultQTy); 1712 1713 llvm::Intrinsic::ID IntrinsicId; 1714 switch (BuiltinID) { 1715 default: 1716 llvm_unreachable("Unknown overflow builtin id."); 1717 case Builtin::BI__builtin_add_overflow: 1718 IntrinsicId = EncompassingInfo.Signed 1719 ? llvm::Intrinsic::sadd_with_overflow 1720 : llvm::Intrinsic::uadd_with_overflow; 1721 break; 1722 case Builtin::BI__builtin_sub_overflow: 1723 IntrinsicId = EncompassingInfo.Signed 1724 ? llvm::Intrinsic::ssub_with_overflow 1725 : llvm::Intrinsic::usub_with_overflow; 1726 break; 1727 case Builtin::BI__builtin_mul_overflow: 1728 IntrinsicId = EncompassingInfo.Signed 1729 ? llvm::Intrinsic::smul_with_overflow 1730 : llvm::Intrinsic::umul_with_overflow; 1731 break; 1732 } 1733 1734 llvm::Value *Left = EmitScalarExpr(LeftArg); 1735 llvm::Value *Right = EmitScalarExpr(RightArg); 1736 Address ResultPtr = EmitPointerWithAlignment(ResultArg); 1737 1738 // Extend each operand to the encompassing type. 1739 Left = Builder.CreateIntCast(Left, EncompassingLLVMTy, LeftInfo.Signed); 1740 Right = Builder.CreateIntCast(Right, EncompassingLLVMTy, RightInfo.Signed); 1741 1742 // Perform the operation on the extended values. 1743 llvm::Value *Overflow, *Result; 1744 Result = EmitOverflowIntrinsic(*this, IntrinsicId, Left, Right, Overflow); 1745 1746 if (EncompassingInfo.Width > ResultInfo.Width) { 1747 // The encompassing type is wider than the result type, so we need to 1748 // truncate it. 1749 llvm::Value *ResultTrunc = Builder.CreateTrunc(Result, ResultLLVMTy); 1750 1751 // To see if the truncation caused an overflow, we will extend 1752 // the result and then compare it to the original result. 1753 llvm::Value *ResultTruncExt = Builder.CreateIntCast( 1754 ResultTrunc, EncompassingLLVMTy, ResultInfo.Signed); 1755 llvm::Value *TruncationOverflow = 1756 Builder.CreateICmpNE(Result, ResultTruncExt); 1757 1758 Overflow = Builder.CreateOr(Overflow, TruncationOverflow); 1759 Result = ResultTrunc; 1760 } 1761 1762 // Finally, store the result using the pointer. 1763 bool isVolatile = 1764 ResultArg->getType()->getPointeeType().isVolatileQualified(); 1765 Builder.CreateStore(EmitToMemory(Result, ResultQTy), ResultPtr, isVolatile); 1766 1767 return RValue::get(Overflow); 1768 } 1769 1770 case Builtin::BI__builtin_uadd_overflow: 1771 case Builtin::BI__builtin_uaddl_overflow: 1772 case Builtin::BI__builtin_uaddll_overflow: 1773 case Builtin::BI__builtin_usub_overflow: 1774 case Builtin::BI__builtin_usubl_overflow: 1775 case Builtin::BI__builtin_usubll_overflow: 1776 case Builtin::BI__builtin_umul_overflow: 1777 case Builtin::BI__builtin_umull_overflow: 1778 case Builtin::BI__builtin_umulll_overflow: 1779 case Builtin::BI__builtin_sadd_overflow: 1780 case Builtin::BI__builtin_saddl_overflow: 1781 case Builtin::BI__builtin_saddll_overflow: 1782 case Builtin::BI__builtin_ssub_overflow: 1783 case Builtin::BI__builtin_ssubl_overflow: 1784 case Builtin::BI__builtin_ssubll_overflow: 1785 case Builtin::BI__builtin_smul_overflow: 1786 case Builtin::BI__builtin_smull_overflow: 1787 case Builtin::BI__builtin_smulll_overflow: { 1788 1789 // We translate all of these builtins directly to the relevant llvm IR node. 1790 1791 // Scalarize our inputs. 1792 llvm::Value *X = EmitScalarExpr(E->getArg(0)); 1793 llvm::Value *Y = EmitScalarExpr(E->getArg(1)); 1794 Address SumOutPtr = EmitPointerWithAlignment(E->getArg(2)); 1795 1796 // Decide which of the overflow intrinsics we are lowering to: 1797 llvm::Intrinsic::ID IntrinsicId; 1798 switch (BuiltinID) { 1799 default: llvm_unreachable("Unknown overflow builtin id."); 1800 case Builtin::BI__builtin_uadd_overflow: 1801 case Builtin::BI__builtin_uaddl_overflow: 1802 case Builtin::BI__builtin_uaddll_overflow: 1803 IntrinsicId = llvm::Intrinsic::uadd_with_overflow; 1804 break; 1805 case Builtin::BI__builtin_usub_overflow: 1806 case Builtin::BI__builtin_usubl_overflow: 1807 case Builtin::BI__builtin_usubll_overflow: 1808 IntrinsicId = llvm::Intrinsic::usub_with_overflow; 1809 break; 1810 case Builtin::BI__builtin_umul_overflow: 1811 case Builtin::BI__builtin_umull_overflow: 1812 case Builtin::BI__builtin_umulll_overflow: 1813 IntrinsicId = llvm::Intrinsic::umul_with_overflow; 1814 break; 1815 case Builtin::BI__builtin_sadd_overflow: 1816 case Builtin::BI__builtin_saddl_overflow: 1817 case Builtin::BI__builtin_saddll_overflow: 1818 IntrinsicId = llvm::Intrinsic::sadd_with_overflow; 1819 break; 1820 case Builtin::BI__builtin_ssub_overflow: 1821 case Builtin::BI__builtin_ssubl_overflow: 1822 case Builtin::BI__builtin_ssubll_overflow: 1823 IntrinsicId = llvm::Intrinsic::ssub_with_overflow; 1824 break; 1825 case Builtin::BI__builtin_smul_overflow: 1826 case Builtin::BI__builtin_smull_overflow: 1827 case Builtin::BI__builtin_smulll_overflow: 1828 IntrinsicId = llvm::Intrinsic::smul_with_overflow; 1829 break; 1830 } 1831 1832 1833 llvm::Value *Carry; 1834 llvm::Value *Sum = EmitOverflowIntrinsic(*this, IntrinsicId, X, Y, Carry); 1835 Builder.CreateStore(Sum, SumOutPtr); 1836 1837 return RValue::get(Carry); 1838 } 1839 case Builtin::BI__builtin_addressof: 1840 return RValue::get(EmitLValue(E->getArg(0)).getPointer()); 1841 case Builtin::BI__builtin_operator_new: 1842 return EmitBuiltinNewDeleteCall(FD->getType()->castAs<FunctionProtoType>(), 1843 E->getArg(0), false); 1844 case Builtin::BI__builtin_operator_delete: 1845 return EmitBuiltinNewDeleteCall(FD->getType()->castAs<FunctionProtoType>(), 1846 E->getArg(0), true); 1847 case Builtin::BI__noop: 1848 // __noop always evaluates to an integer literal zero. 1849 return RValue::get(ConstantInt::get(IntTy, 0)); 1850 case Builtin::BI__builtin_call_with_static_chain: { 1851 const CallExpr *Call = cast<CallExpr>(E->getArg(0)); 1852 const Expr *Chain = E->getArg(1); 1853 return EmitCall(Call->getCallee()->getType(), 1854 EmitScalarExpr(Call->getCallee()), Call, ReturnValue, 1855 Call->getCalleeDecl(), EmitScalarExpr(Chain)); 1856 } 1857 case Builtin::BI_InterlockedExchange: 1858 case Builtin::BI_InterlockedExchangePointer: 1859 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Xchg, E); 1860 case Builtin::BI_InterlockedCompareExchangePointer: { 1861 llvm::Type *RTy; 1862 llvm::IntegerType *IntType = 1863 IntegerType::get(getLLVMContext(), 1864 getContext().getTypeSize(E->getType())); 1865 llvm::Type *IntPtrType = IntType->getPointerTo(); 1866 1867 llvm::Value *Destination = 1868 Builder.CreateBitCast(EmitScalarExpr(E->getArg(0)), IntPtrType); 1869 1870 llvm::Value *Exchange = EmitScalarExpr(E->getArg(1)); 1871 RTy = Exchange->getType(); 1872 Exchange = Builder.CreatePtrToInt(Exchange, IntType); 1873 1874 llvm::Value *Comparand = 1875 Builder.CreatePtrToInt(EmitScalarExpr(E->getArg(2)), IntType); 1876 1877 auto Result = Builder.CreateAtomicCmpXchg(Destination, Comparand, Exchange, 1878 SequentiallyConsistent, 1879 SequentiallyConsistent); 1880 Result->setVolatile(true); 1881 1882 return RValue::get(Builder.CreateIntToPtr(Builder.CreateExtractValue(Result, 1883 0), 1884 RTy)); 1885 } 1886 case Builtin::BI_InterlockedCompareExchange: { 1887 AtomicCmpXchgInst *CXI = Builder.CreateAtomicCmpXchg( 1888 EmitScalarExpr(E->getArg(0)), 1889 EmitScalarExpr(E->getArg(2)), 1890 EmitScalarExpr(E->getArg(1)), 1891 SequentiallyConsistent, 1892 SequentiallyConsistent); 1893 CXI->setVolatile(true); 1894 return RValue::get(Builder.CreateExtractValue(CXI, 0)); 1895 } 1896 case Builtin::BI_InterlockedIncrement: { 1897 AtomicRMWInst *RMWI = Builder.CreateAtomicRMW( 1898 AtomicRMWInst::Add, 1899 EmitScalarExpr(E->getArg(0)), 1900 ConstantInt::get(Int32Ty, 1), 1901 llvm::SequentiallyConsistent); 1902 RMWI->setVolatile(true); 1903 return RValue::get(Builder.CreateAdd(RMWI, ConstantInt::get(Int32Ty, 1))); 1904 } 1905 case Builtin::BI_InterlockedDecrement: { 1906 AtomicRMWInst *RMWI = Builder.CreateAtomicRMW( 1907 AtomicRMWInst::Sub, 1908 EmitScalarExpr(E->getArg(0)), 1909 ConstantInt::get(Int32Ty, 1), 1910 llvm::SequentiallyConsistent); 1911 RMWI->setVolatile(true); 1912 return RValue::get(Builder.CreateSub(RMWI, ConstantInt::get(Int32Ty, 1))); 1913 } 1914 case Builtin::BI_InterlockedExchangeAdd: { 1915 AtomicRMWInst *RMWI = Builder.CreateAtomicRMW( 1916 AtomicRMWInst::Add, 1917 EmitScalarExpr(E->getArg(0)), 1918 EmitScalarExpr(E->getArg(1)), 1919 llvm::SequentiallyConsistent); 1920 RMWI->setVolatile(true); 1921 return RValue::get(RMWI); 1922 } 1923 case Builtin::BI__readfsdword: { 1924 Value *IntToPtr = 1925 Builder.CreateIntToPtr(EmitScalarExpr(E->getArg(0)), 1926 llvm::PointerType::get(CGM.Int32Ty, 257)); 1927 LoadInst *Load = 1928 Builder.CreateAlignedLoad(IntToPtr, /*Align=*/4, /*isVolatile=*/true); 1929 return RValue::get(Load); 1930 } 1931 1932 case Builtin::BI__exception_code: 1933 case Builtin::BI_exception_code: 1934 return RValue::get(EmitSEHExceptionCode()); 1935 case Builtin::BI__exception_info: 1936 case Builtin::BI_exception_info: 1937 return RValue::get(EmitSEHExceptionInfo()); 1938 case Builtin::BI__abnormal_termination: 1939 case Builtin::BI_abnormal_termination: 1940 return RValue::get(EmitSEHAbnormalTermination()); 1941 case Builtin::BI_setjmpex: { 1942 if (getTarget().getTriple().isOSMSVCRT()) { 1943 llvm::Type *ArgTypes[] = {Int8PtrTy, Int8PtrTy}; 1944 llvm::AttributeSet ReturnsTwiceAttr = 1945 AttributeSet::get(getLLVMContext(), llvm::AttributeSet::FunctionIndex, 1946 llvm::Attribute::ReturnsTwice); 1947 llvm::Constant *SetJmpEx = CGM.CreateRuntimeFunction( 1948 llvm::FunctionType::get(IntTy, ArgTypes, /*isVarArg=*/false), 1949 "_setjmpex", ReturnsTwiceAttr); 1950 llvm::Value *Buf = Builder.CreateBitOrPointerCast( 1951 EmitScalarExpr(E->getArg(0)), Int8PtrTy); 1952 llvm::Value *FrameAddr = 1953 Builder.CreateCall(CGM.getIntrinsic(Intrinsic::frameaddress), 1954 ConstantInt::get(Int32Ty, 0)); 1955 llvm::Value *Args[] = {Buf, FrameAddr}; 1956 llvm::CallSite CS = EmitRuntimeCallOrInvoke(SetJmpEx, Args); 1957 CS.setAttributes(ReturnsTwiceAttr); 1958 return RValue::get(CS.getInstruction()); 1959 } 1960 break; 1961 } 1962 case Builtin::BI_setjmp: { 1963 if (getTarget().getTriple().isOSMSVCRT()) { 1964 llvm::AttributeSet ReturnsTwiceAttr = 1965 AttributeSet::get(getLLVMContext(), llvm::AttributeSet::FunctionIndex, 1966 llvm::Attribute::ReturnsTwice); 1967 llvm::Value *Buf = Builder.CreateBitOrPointerCast( 1968 EmitScalarExpr(E->getArg(0)), Int8PtrTy); 1969 llvm::CallSite CS; 1970 if (getTarget().getTriple().getArch() == llvm::Triple::x86) { 1971 llvm::Type *ArgTypes[] = {Int8PtrTy, IntTy}; 1972 llvm::Constant *SetJmp3 = CGM.CreateRuntimeFunction( 1973 llvm::FunctionType::get(IntTy, ArgTypes, /*isVarArg=*/true), 1974 "_setjmp3", ReturnsTwiceAttr); 1975 llvm::Value *Count = ConstantInt::get(IntTy, 0); 1976 llvm::Value *Args[] = {Buf, Count}; 1977 CS = EmitRuntimeCallOrInvoke(SetJmp3, Args); 1978 } else { 1979 llvm::Type *ArgTypes[] = {Int8PtrTy, Int8PtrTy}; 1980 llvm::Constant *SetJmp = CGM.CreateRuntimeFunction( 1981 llvm::FunctionType::get(IntTy, ArgTypes, /*isVarArg=*/false), 1982 "_setjmp", ReturnsTwiceAttr); 1983 llvm::Value *FrameAddr = 1984 Builder.CreateCall(CGM.getIntrinsic(Intrinsic::frameaddress), 1985 ConstantInt::get(Int32Ty, 0)); 1986 llvm::Value *Args[] = {Buf, FrameAddr}; 1987 CS = EmitRuntimeCallOrInvoke(SetJmp, Args); 1988 } 1989 CS.setAttributes(ReturnsTwiceAttr); 1990 return RValue::get(CS.getInstruction()); 1991 } 1992 break; 1993 } 1994 1995 case Builtin::BI__GetExceptionInfo: { 1996 if (llvm::GlobalVariable *GV = 1997 CGM.getCXXABI().getThrowInfo(FD->getParamDecl(0)->getType())) 1998 return RValue::get(llvm::ConstantExpr::getBitCast(GV, CGM.Int8PtrTy)); 1999 break; 2000 } 2001 2002 // OpenCL v2.0 s6.13.16.2, Built-in pipe read and write functions 2003 case Builtin::BIread_pipe: 2004 case Builtin::BIwrite_pipe: { 2005 Value *Arg0 = EmitScalarExpr(E->getArg(0)), 2006 *Arg1 = EmitScalarExpr(E->getArg(1)); 2007 2008 // Type of the generic packet parameter. 2009 unsigned GenericAS = 2010 getContext().getTargetAddressSpace(LangAS::opencl_generic); 2011 llvm::Type *I8PTy = llvm::PointerType::get( 2012 llvm::Type::getInt8Ty(getLLVMContext()), GenericAS); 2013 2014 // Testing which overloaded version we should generate the call for. 2015 if (2U == E->getNumArgs()) { 2016 const char *Name = (BuiltinID == Builtin::BIread_pipe) ? "__read_pipe_2" 2017 : "__write_pipe_2"; 2018 // Creating a generic function type to be able to call with any builtin or 2019 // user defined type. 2020 llvm::Type *ArgTys[] = {Arg0->getType(), I8PTy}; 2021 llvm::FunctionType *FTy = llvm::FunctionType::get( 2022 Int32Ty, llvm::ArrayRef<llvm::Type *>(ArgTys), false); 2023 Value *BCast = Builder.CreatePointerCast(Arg1, I8PTy); 2024 return RValue::get(Builder.CreateCall( 2025 CGM.CreateRuntimeFunction(FTy, Name), {Arg0, BCast})); 2026 } else { 2027 assert(4 == E->getNumArgs() && 2028 "Illegal number of parameters to pipe function"); 2029 const char *Name = (BuiltinID == Builtin::BIread_pipe) ? "__read_pipe_4" 2030 : "__write_pipe_4"; 2031 2032 llvm::Type *ArgTys[] = {Arg0->getType(), Arg1->getType(), Int32Ty, I8PTy}; 2033 Value *Arg2 = EmitScalarExpr(E->getArg(2)), 2034 *Arg3 = EmitScalarExpr(E->getArg(3)); 2035 llvm::FunctionType *FTy = llvm::FunctionType::get( 2036 Int32Ty, llvm::ArrayRef<llvm::Type *>(ArgTys), false); 2037 Value *BCast = Builder.CreatePointerCast(Arg3, I8PTy); 2038 // We know the third argument is an integer type, but we may need to cast 2039 // it to i32. 2040 if (Arg2->getType() != Int32Ty) 2041 Arg2 = Builder.CreateZExtOrTrunc(Arg2, Int32Ty); 2042 return RValue::get(Builder.CreateCall( 2043 CGM.CreateRuntimeFunction(FTy, Name), {Arg0, Arg1, Arg2, BCast})); 2044 } 2045 } 2046 // OpenCL v2.0 s6.13.16 ,s9.17.3.5 - Built-in pipe reserve read and write 2047 // functions 2048 case Builtin::BIreserve_read_pipe: 2049 case Builtin::BIreserve_write_pipe: 2050 case Builtin::BIwork_group_reserve_read_pipe: 2051 case Builtin::BIwork_group_reserve_write_pipe: 2052 case Builtin::BIsub_group_reserve_read_pipe: 2053 case Builtin::BIsub_group_reserve_write_pipe: { 2054 // Composing the mangled name for the function. 2055 const char *Name; 2056 if (BuiltinID == Builtin::BIreserve_read_pipe) 2057 Name = "__reserve_read_pipe"; 2058 else if (BuiltinID == Builtin::BIreserve_write_pipe) 2059 Name = "__reserve_write_pipe"; 2060 else if (BuiltinID == Builtin::BIwork_group_reserve_read_pipe) 2061 Name = "__work_group_reserve_read_pipe"; 2062 else if (BuiltinID == Builtin::BIwork_group_reserve_write_pipe) 2063 Name = "__work_group_reserve_write_pipe"; 2064 else if (BuiltinID == Builtin::BIsub_group_reserve_read_pipe) 2065 Name = "__sub_group_reserve_read_pipe"; 2066 else 2067 Name = "__sub_group_reserve_write_pipe"; 2068 2069 Value *Arg0 = EmitScalarExpr(E->getArg(0)), 2070 *Arg1 = EmitScalarExpr(E->getArg(1)); 2071 llvm::Type *ReservedIDTy = ConvertType(getContext().OCLReserveIDTy); 2072 2073 // Building the generic function prototype. 2074 llvm::Type *ArgTys[] = {Arg0->getType(), Int32Ty}; 2075 llvm::FunctionType *FTy = llvm::FunctionType::get( 2076 ReservedIDTy, llvm::ArrayRef<llvm::Type *>(ArgTys), false); 2077 // We know the second argument is an integer type, but we may need to cast 2078 // it to i32. 2079 if (Arg1->getType() != Int32Ty) 2080 Arg1 = Builder.CreateZExtOrTrunc(Arg1, Int32Ty); 2081 return RValue::get( 2082 Builder.CreateCall(CGM.CreateRuntimeFunction(FTy, Name), {Arg0, Arg1})); 2083 } 2084 // OpenCL v2.0 s6.13.16 ,s9.17.3.5 - Built-in pipe commit read and write 2085 // functions 2086 case Builtin::BIcommit_read_pipe: 2087 case Builtin::BIcommit_write_pipe: 2088 case Builtin::BIwork_group_commit_read_pipe: 2089 case Builtin::BIwork_group_commit_write_pipe: 2090 case Builtin::BIsub_group_commit_read_pipe: 2091 case Builtin::BIsub_group_commit_write_pipe: { 2092 const char *Name; 2093 if (BuiltinID == Builtin::BIcommit_read_pipe) 2094 Name = "__commit_read_pipe"; 2095 else if (BuiltinID == Builtin::BIcommit_write_pipe) 2096 Name = "__commit_write_pipe"; 2097 else if (BuiltinID == Builtin::BIwork_group_commit_read_pipe) 2098 Name = "__work_group_commit_read_pipe"; 2099 else if (BuiltinID == Builtin::BIwork_group_commit_write_pipe) 2100 Name = "__work_group_commit_write_pipe"; 2101 else if (BuiltinID == Builtin::BIsub_group_commit_read_pipe) 2102 Name = "__sub_group_commit_read_pipe"; 2103 else 2104 Name = "__sub_group_commit_write_pipe"; 2105 2106 Value *Arg0 = EmitScalarExpr(E->getArg(0)), 2107 *Arg1 = EmitScalarExpr(E->getArg(1)); 2108 2109 // Building the generic function prototype. 2110 llvm::Type *ArgTys[] = {Arg0->getType(), Arg1->getType()}; 2111 llvm::FunctionType *FTy = 2112 llvm::FunctionType::get(llvm::Type::getVoidTy(getLLVMContext()), 2113 llvm::ArrayRef<llvm::Type *>(ArgTys), false); 2114 2115 return RValue::get( 2116 Builder.CreateCall(CGM.CreateRuntimeFunction(FTy, Name), {Arg0, Arg1})); 2117 } 2118 // OpenCL v2.0 s6.13.16.4 Built-in pipe query functions 2119 case Builtin::BIget_pipe_num_packets: 2120 case Builtin::BIget_pipe_max_packets: { 2121 const char *Name; 2122 if (BuiltinID == Builtin::BIget_pipe_num_packets) 2123 Name = "__get_pipe_num_packets"; 2124 else 2125 Name = "__get_pipe_max_packets"; 2126 2127 // Building the generic function prototype. 2128 Value *Arg0 = EmitScalarExpr(E->getArg(0)); 2129 llvm::Type *ArgTys[] = {Arg0->getType()}; 2130 llvm::FunctionType *FTy = llvm::FunctionType::get( 2131 Int32Ty, llvm::ArrayRef<llvm::Type *>(ArgTys), false); 2132 2133 return RValue::get( 2134 Builder.CreateCall(CGM.CreateRuntimeFunction(FTy, Name), {Arg0})); 2135 } 2136 2137 case Builtin::BIprintf: 2138 if (getLangOpts().CUDA && getLangOpts().CUDAIsDevice) 2139 return EmitCUDADevicePrintfCallExpr(E, ReturnValue); 2140 break; 2141 case Builtin::BI__builtin_canonicalize: 2142 case Builtin::BI__builtin_canonicalizef: 2143 case Builtin::BI__builtin_canonicalizel: 2144 return RValue::get(emitUnaryBuiltin(*this, E, Intrinsic::canonicalize)); 2145 } 2146 2147 // If this is an alias for a lib function (e.g. __builtin_sin), emit 2148 // the call using the normal call path, but using the unmangled 2149 // version of the function name. 2150 if (getContext().BuiltinInfo.isLibFunction(BuiltinID)) 2151 return emitLibraryCall(*this, FD, E, 2152 CGM.getBuiltinLibFunction(FD, BuiltinID)); 2153 2154 // If this is a predefined lib function (e.g. malloc), emit the call 2155 // using exactly the normal call path. 2156 if (getContext().BuiltinInfo.isPredefinedLibFunction(BuiltinID)) 2157 return emitLibraryCall(*this, FD, E, EmitScalarExpr(E->getCallee())); 2158 2159 // Check that a call to a target specific builtin has the correct target 2160 // features. 2161 // This is down here to avoid non-target specific builtins, however, if 2162 // generic builtins start to require generic target features then we 2163 // can move this up to the beginning of the function. 2164 checkTargetFeatures(E, FD); 2165 2166 // See if we have a target specific intrinsic. 2167 const char *Name = getContext().BuiltinInfo.getName(BuiltinID); 2168 Intrinsic::ID IntrinsicID = Intrinsic::not_intrinsic; 2169 if (const char *Prefix = 2170 llvm::Triple::getArchTypePrefix(getTarget().getTriple().getArch())) { 2171 IntrinsicID = Intrinsic::getIntrinsicForGCCBuiltin(Prefix, Name); 2172 // NOTE we dont need to perform a compatibility flag check here since the 2173 // intrinsics are declared in Builtins*.def via LANGBUILTIN which filter the 2174 // MS builtins via ALL_MS_LANGUAGES and are filtered earlier. 2175 if (IntrinsicID == Intrinsic::not_intrinsic) 2176 IntrinsicID = Intrinsic::getIntrinsicForMSBuiltin(Prefix, Name); 2177 } 2178 2179 if (IntrinsicID != Intrinsic::not_intrinsic) { 2180 SmallVector<Value*, 16> Args; 2181 2182 // Find out if any arguments are required to be integer constant 2183 // expressions. 2184 unsigned ICEArguments = 0; 2185 ASTContext::GetBuiltinTypeError Error; 2186 getContext().GetBuiltinType(BuiltinID, Error, &ICEArguments); 2187 assert(Error == ASTContext::GE_None && "Should not codegen an error"); 2188 2189 Function *F = CGM.getIntrinsic(IntrinsicID); 2190 llvm::FunctionType *FTy = F->getFunctionType(); 2191 2192 for (unsigned i = 0, e = E->getNumArgs(); i != e; ++i) { 2193 Value *ArgValue; 2194 // If this is a normal argument, just emit it as a scalar. 2195 if ((ICEArguments & (1 << i)) == 0) { 2196 ArgValue = EmitScalarExpr(E->getArg(i)); 2197 } else { 2198 // If this is required to be a constant, constant fold it so that we 2199 // know that the generated intrinsic gets a ConstantInt. 2200 llvm::APSInt Result; 2201 bool IsConst = E->getArg(i)->isIntegerConstantExpr(Result,getContext()); 2202 assert(IsConst && "Constant arg isn't actually constant?"); 2203 (void)IsConst; 2204 ArgValue = llvm::ConstantInt::get(getLLVMContext(), Result); 2205 } 2206 2207 // If the intrinsic arg type is different from the builtin arg type 2208 // we need to do a bit cast. 2209 llvm::Type *PTy = FTy->getParamType(i); 2210 if (PTy != ArgValue->getType()) { 2211 assert(PTy->canLosslesslyBitCastTo(FTy->getParamType(i)) && 2212 "Must be able to losslessly bit cast to param"); 2213 ArgValue = Builder.CreateBitCast(ArgValue, PTy); 2214 } 2215 2216 Args.push_back(ArgValue); 2217 } 2218 2219 Value *V = Builder.CreateCall(F, Args); 2220 QualType BuiltinRetType = E->getType(); 2221 2222 llvm::Type *RetTy = VoidTy; 2223 if (!BuiltinRetType->isVoidType()) 2224 RetTy = ConvertType(BuiltinRetType); 2225 2226 if (RetTy != V->getType()) { 2227 assert(V->getType()->canLosslesslyBitCastTo(RetTy) && 2228 "Must be able to losslessly bit cast result type"); 2229 V = Builder.CreateBitCast(V, RetTy); 2230 } 2231 2232 return RValue::get(V); 2233 } 2234 2235 // See if we have a target specific builtin that needs to be lowered. 2236 if (Value *V = EmitTargetBuiltinExpr(BuiltinID, E)) 2237 return RValue::get(V); 2238 2239 ErrorUnsupported(E, "builtin function"); 2240 2241 // Unknown builtin, for now just dump it out and return undef. 2242 return GetUndefRValue(E->getType()); 2243 } 2244 2245 static Value *EmitTargetArchBuiltinExpr(CodeGenFunction *CGF, 2246 unsigned BuiltinID, const CallExpr *E, 2247 llvm::Triple::ArchType Arch) { 2248 switch (Arch) { 2249 case llvm::Triple::arm: 2250 case llvm::Triple::armeb: 2251 case llvm::Triple::thumb: 2252 case llvm::Triple::thumbeb: 2253 return CGF->EmitARMBuiltinExpr(BuiltinID, E); 2254 case llvm::Triple::aarch64: 2255 case llvm::Triple::aarch64_be: 2256 return CGF->EmitAArch64BuiltinExpr(BuiltinID, E); 2257 case llvm::Triple::x86: 2258 case llvm::Triple::x86_64: 2259 return CGF->EmitX86BuiltinExpr(BuiltinID, E); 2260 case llvm::Triple::ppc: 2261 case llvm::Triple::ppc64: 2262 case llvm::Triple::ppc64le: 2263 return CGF->EmitPPCBuiltinExpr(BuiltinID, E); 2264 case llvm::Triple::r600: 2265 case llvm::Triple::amdgcn: 2266 return CGF->EmitAMDGPUBuiltinExpr(BuiltinID, E); 2267 case llvm::Triple::systemz: 2268 return CGF->EmitSystemZBuiltinExpr(BuiltinID, E); 2269 case llvm::Triple::nvptx: 2270 case llvm::Triple::nvptx64: 2271 return CGF->EmitNVPTXBuiltinExpr(BuiltinID, E); 2272 case llvm::Triple::wasm32: 2273 case llvm::Triple::wasm64: 2274 return CGF->EmitWebAssemblyBuiltinExpr(BuiltinID, E); 2275 default: 2276 return nullptr; 2277 } 2278 } 2279 2280 Value *CodeGenFunction::EmitTargetBuiltinExpr(unsigned BuiltinID, 2281 const CallExpr *E) { 2282 if (getContext().BuiltinInfo.isAuxBuiltinID(BuiltinID)) { 2283 assert(getContext().getAuxTargetInfo() && "Missing aux target info"); 2284 return EmitTargetArchBuiltinExpr( 2285 this, getContext().BuiltinInfo.getAuxBuiltinID(BuiltinID), E, 2286 getContext().getAuxTargetInfo()->getTriple().getArch()); 2287 } 2288 2289 return EmitTargetArchBuiltinExpr(this, BuiltinID, E, 2290 getTarget().getTriple().getArch()); 2291 } 2292 2293 static llvm::VectorType *GetNeonType(CodeGenFunction *CGF, 2294 NeonTypeFlags TypeFlags, 2295 bool V1Ty=false) { 2296 int IsQuad = TypeFlags.isQuad(); 2297 switch (TypeFlags.getEltType()) { 2298 case NeonTypeFlags::Int8: 2299 case NeonTypeFlags::Poly8: 2300 return llvm::VectorType::get(CGF->Int8Ty, V1Ty ? 1 : (8 << IsQuad)); 2301 case NeonTypeFlags::Int16: 2302 case NeonTypeFlags::Poly16: 2303 case NeonTypeFlags::Float16: 2304 return llvm::VectorType::get(CGF->Int16Ty, V1Ty ? 1 : (4 << IsQuad)); 2305 case NeonTypeFlags::Int32: 2306 return llvm::VectorType::get(CGF->Int32Ty, V1Ty ? 1 : (2 << IsQuad)); 2307 case NeonTypeFlags::Int64: 2308 case NeonTypeFlags::Poly64: 2309 return llvm::VectorType::get(CGF->Int64Ty, V1Ty ? 1 : (1 << IsQuad)); 2310 case NeonTypeFlags::Poly128: 2311 // FIXME: i128 and f128 doesn't get fully support in Clang and llvm. 2312 // There is a lot of i128 and f128 API missing. 2313 // so we use v16i8 to represent poly128 and get pattern matched. 2314 return llvm::VectorType::get(CGF->Int8Ty, 16); 2315 case NeonTypeFlags::Float32: 2316 return llvm::VectorType::get(CGF->FloatTy, V1Ty ? 1 : (2 << IsQuad)); 2317 case NeonTypeFlags::Float64: 2318 return llvm::VectorType::get(CGF->DoubleTy, V1Ty ? 1 : (1 << IsQuad)); 2319 } 2320 llvm_unreachable("Unknown vector element type!"); 2321 } 2322 2323 static llvm::VectorType *GetFloatNeonType(CodeGenFunction *CGF, 2324 NeonTypeFlags IntTypeFlags) { 2325 int IsQuad = IntTypeFlags.isQuad(); 2326 switch (IntTypeFlags.getEltType()) { 2327 case NeonTypeFlags::Int32: 2328 return llvm::VectorType::get(CGF->FloatTy, (2 << IsQuad)); 2329 case NeonTypeFlags::Int64: 2330 return llvm::VectorType::get(CGF->DoubleTy, (1 << IsQuad)); 2331 default: 2332 llvm_unreachable("Type can't be converted to floating-point!"); 2333 } 2334 } 2335 2336 Value *CodeGenFunction::EmitNeonSplat(Value *V, Constant *C) { 2337 unsigned nElts = cast<llvm::VectorType>(V->getType())->getNumElements(); 2338 Value* SV = llvm::ConstantVector::getSplat(nElts, C); 2339 return Builder.CreateShuffleVector(V, V, SV, "lane"); 2340 } 2341 2342 Value *CodeGenFunction::EmitNeonCall(Function *F, SmallVectorImpl<Value*> &Ops, 2343 const char *name, 2344 unsigned shift, bool rightshift) { 2345 unsigned j = 0; 2346 for (Function::const_arg_iterator ai = F->arg_begin(), ae = F->arg_end(); 2347 ai != ae; ++ai, ++j) 2348 if (shift > 0 && shift == j) 2349 Ops[j] = EmitNeonShiftVector(Ops[j], ai->getType(), rightshift); 2350 else 2351 Ops[j] = Builder.CreateBitCast(Ops[j], ai->getType(), name); 2352 2353 return Builder.CreateCall(F, Ops, name); 2354 } 2355 2356 Value *CodeGenFunction::EmitNeonShiftVector(Value *V, llvm::Type *Ty, 2357 bool neg) { 2358 int SV = cast<ConstantInt>(V)->getSExtValue(); 2359 return ConstantInt::get(Ty, neg ? -SV : SV); 2360 } 2361 2362 // \brief Right-shift a vector by a constant. 2363 Value *CodeGenFunction::EmitNeonRShiftImm(Value *Vec, Value *Shift, 2364 llvm::Type *Ty, bool usgn, 2365 const char *name) { 2366 llvm::VectorType *VTy = cast<llvm::VectorType>(Ty); 2367 2368 int ShiftAmt = cast<ConstantInt>(Shift)->getSExtValue(); 2369 int EltSize = VTy->getScalarSizeInBits(); 2370 2371 Vec = Builder.CreateBitCast(Vec, Ty); 2372 2373 // lshr/ashr are undefined when the shift amount is equal to the vector 2374 // element size. 2375 if (ShiftAmt == EltSize) { 2376 if (usgn) { 2377 // Right-shifting an unsigned value by its size yields 0. 2378 return llvm::ConstantAggregateZero::get(VTy); 2379 } else { 2380 // Right-shifting a signed value by its size is equivalent 2381 // to a shift of size-1. 2382 --ShiftAmt; 2383 Shift = ConstantInt::get(VTy->getElementType(), ShiftAmt); 2384 } 2385 } 2386 2387 Shift = EmitNeonShiftVector(Shift, Ty, false); 2388 if (usgn) 2389 return Builder.CreateLShr(Vec, Shift, name); 2390 else 2391 return Builder.CreateAShr(Vec, Shift, name); 2392 } 2393 2394 enum { 2395 AddRetType = (1 << 0), 2396 Add1ArgType = (1 << 1), 2397 Add2ArgTypes = (1 << 2), 2398 2399 VectorizeRetType = (1 << 3), 2400 VectorizeArgTypes = (1 << 4), 2401 2402 InventFloatType = (1 << 5), 2403 UnsignedAlts = (1 << 6), 2404 2405 Use64BitVectors = (1 << 7), 2406 Use128BitVectors = (1 << 8), 2407 2408 Vectorize1ArgType = Add1ArgType | VectorizeArgTypes, 2409 VectorRet = AddRetType | VectorizeRetType, 2410 VectorRetGetArgs01 = 2411 AddRetType | Add2ArgTypes | VectorizeRetType | VectorizeArgTypes, 2412 FpCmpzModifiers = 2413 AddRetType | VectorizeRetType | Add1ArgType | InventFloatType 2414 }; 2415 2416 namespace { 2417 struct NeonIntrinsicInfo { 2418 const char *NameHint; 2419 unsigned BuiltinID; 2420 unsigned LLVMIntrinsic; 2421 unsigned AltLLVMIntrinsic; 2422 unsigned TypeModifier; 2423 2424 bool operator<(unsigned RHSBuiltinID) const { 2425 return BuiltinID < RHSBuiltinID; 2426 } 2427 bool operator<(const NeonIntrinsicInfo &TE) const { 2428 return BuiltinID < TE.BuiltinID; 2429 } 2430 }; 2431 } // end anonymous namespace 2432 2433 #define NEONMAP0(NameBase) \ 2434 { #NameBase, NEON::BI__builtin_neon_ ## NameBase, 0, 0, 0 } 2435 2436 #define NEONMAP1(NameBase, LLVMIntrinsic, TypeModifier) \ 2437 { #NameBase, NEON:: BI__builtin_neon_ ## NameBase, \ 2438 Intrinsic::LLVMIntrinsic, 0, TypeModifier } 2439 2440 #define NEONMAP2(NameBase, LLVMIntrinsic, AltLLVMIntrinsic, TypeModifier) \ 2441 { #NameBase, NEON:: BI__builtin_neon_ ## NameBase, \ 2442 Intrinsic::LLVMIntrinsic, Intrinsic::AltLLVMIntrinsic, \ 2443 TypeModifier } 2444 2445 static const NeonIntrinsicInfo ARMSIMDIntrinsicMap [] = { 2446 NEONMAP2(vabd_v, arm_neon_vabdu, arm_neon_vabds, Add1ArgType | UnsignedAlts), 2447 NEONMAP2(vabdq_v, arm_neon_vabdu, arm_neon_vabds, Add1ArgType | UnsignedAlts), 2448 NEONMAP1(vabs_v, arm_neon_vabs, 0), 2449 NEONMAP1(vabsq_v, arm_neon_vabs, 0), 2450 NEONMAP0(vaddhn_v), 2451 NEONMAP1(vaesdq_v, arm_neon_aesd, 0), 2452 NEONMAP1(vaeseq_v, arm_neon_aese, 0), 2453 NEONMAP1(vaesimcq_v, arm_neon_aesimc, 0), 2454 NEONMAP1(vaesmcq_v, arm_neon_aesmc, 0), 2455 NEONMAP1(vbsl_v, arm_neon_vbsl, AddRetType), 2456 NEONMAP1(vbslq_v, arm_neon_vbsl, AddRetType), 2457 NEONMAP1(vcage_v, arm_neon_vacge, 0), 2458 NEONMAP1(vcageq_v, arm_neon_vacge, 0), 2459 NEONMAP1(vcagt_v, arm_neon_vacgt, 0), 2460 NEONMAP1(vcagtq_v, arm_neon_vacgt, 0), 2461 NEONMAP1(vcale_v, arm_neon_vacge, 0), 2462 NEONMAP1(vcaleq_v, arm_neon_vacge, 0), 2463 NEONMAP1(vcalt_v, arm_neon_vacgt, 0), 2464 NEONMAP1(vcaltq_v, arm_neon_vacgt, 0), 2465 NEONMAP1(vcls_v, arm_neon_vcls, Add1ArgType), 2466 NEONMAP1(vclsq_v, arm_neon_vcls, Add1ArgType), 2467 NEONMAP1(vclz_v, ctlz, Add1ArgType), 2468 NEONMAP1(vclzq_v, ctlz, Add1ArgType), 2469 NEONMAP1(vcnt_v, ctpop, Add1ArgType), 2470 NEONMAP1(vcntq_v, ctpop, Add1ArgType), 2471 NEONMAP1(vcvt_f16_f32, arm_neon_vcvtfp2hf, 0), 2472 NEONMAP1(vcvt_f32_f16, arm_neon_vcvthf2fp, 0), 2473 NEONMAP0(vcvt_f32_v), 2474 NEONMAP2(vcvt_n_f32_v, arm_neon_vcvtfxu2fp, arm_neon_vcvtfxs2fp, 0), 2475 NEONMAP1(vcvt_n_s32_v, arm_neon_vcvtfp2fxs, 0), 2476 NEONMAP1(vcvt_n_s64_v, arm_neon_vcvtfp2fxs, 0), 2477 NEONMAP1(vcvt_n_u32_v, arm_neon_vcvtfp2fxu, 0), 2478 NEONMAP1(vcvt_n_u64_v, arm_neon_vcvtfp2fxu, 0), 2479 NEONMAP0(vcvt_s32_v), 2480 NEONMAP0(vcvt_s64_v), 2481 NEONMAP0(vcvt_u32_v), 2482 NEONMAP0(vcvt_u64_v), 2483 NEONMAP1(vcvta_s32_v, arm_neon_vcvtas, 0), 2484 NEONMAP1(vcvta_s64_v, arm_neon_vcvtas, 0), 2485 NEONMAP1(vcvta_u32_v, arm_neon_vcvtau, 0), 2486 NEONMAP1(vcvta_u64_v, arm_neon_vcvtau, 0), 2487 NEONMAP1(vcvtaq_s32_v, arm_neon_vcvtas, 0), 2488 NEONMAP1(vcvtaq_s64_v, arm_neon_vcvtas, 0), 2489 NEONMAP1(vcvtaq_u32_v, arm_neon_vcvtau, 0), 2490 NEONMAP1(vcvtaq_u64_v, arm_neon_vcvtau, 0), 2491 NEONMAP1(vcvtm_s32_v, arm_neon_vcvtms, 0), 2492 NEONMAP1(vcvtm_s64_v, arm_neon_vcvtms, 0), 2493 NEONMAP1(vcvtm_u32_v, arm_neon_vcvtmu, 0), 2494 NEONMAP1(vcvtm_u64_v, arm_neon_vcvtmu, 0), 2495 NEONMAP1(vcvtmq_s32_v, arm_neon_vcvtms, 0), 2496 NEONMAP1(vcvtmq_s64_v, arm_neon_vcvtms, 0), 2497 NEONMAP1(vcvtmq_u32_v, arm_neon_vcvtmu, 0), 2498 NEONMAP1(vcvtmq_u64_v, arm_neon_vcvtmu, 0), 2499 NEONMAP1(vcvtn_s32_v, arm_neon_vcvtns, 0), 2500 NEONMAP1(vcvtn_s64_v, arm_neon_vcvtns, 0), 2501 NEONMAP1(vcvtn_u32_v, arm_neon_vcvtnu, 0), 2502 NEONMAP1(vcvtn_u64_v, arm_neon_vcvtnu, 0), 2503 NEONMAP1(vcvtnq_s32_v, arm_neon_vcvtns, 0), 2504 NEONMAP1(vcvtnq_s64_v, arm_neon_vcvtns, 0), 2505 NEONMAP1(vcvtnq_u32_v, arm_neon_vcvtnu, 0), 2506 NEONMAP1(vcvtnq_u64_v, arm_neon_vcvtnu, 0), 2507 NEONMAP1(vcvtp_s32_v, arm_neon_vcvtps, 0), 2508 NEONMAP1(vcvtp_s64_v, arm_neon_vcvtps, 0), 2509 NEONMAP1(vcvtp_u32_v, arm_neon_vcvtpu, 0), 2510 NEONMAP1(vcvtp_u64_v, arm_neon_vcvtpu, 0), 2511 NEONMAP1(vcvtpq_s32_v, arm_neon_vcvtps, 0), 2512 NEONMAP1(vcvtpq_s64_v, arm_neon_vcvtps, 0), 2513 NEONMAP1(vcvtpq_u32_v, arm_neon_vcvtpu, 0), 2514 NEONMAP1(vcvtpq_u64_v, arm_neon_vcvtpu, 0), 2515 NEONMAP0(vcvtq_f32_v), 2516 NEONMAP2(vcvtq_n_f32_v, arm_neon_vcvtfxu2fp, arm_neon_vcvtfxs2fp, 0), 2517 NEONMAP1(vcvtq_n_s32_v, arm_neon_vcvtfp2fxs, 0), 2518 NEONMAP1(vcvtq_n_s64_v, arm_neon_vcvtfp2fxs, 0), 2519 NEONMAP1(vcvtq_n_u32_v, arm_neon_vcvtfp2fxu, 0), 2520 NEONMAP1(vcvtq_n_u64_v, arm_neon_vcvtfp2fxu, 0), 2521 NEONMAP0(vcvtq_s32_v), 2522 NEONMAP0(vcvtq_s64_v), 2523 NEONMAP0(vcvtq_u32_v), 2524 NEONMAP0(vcvtq_u64_v), 2525 NEONMAP0(vext_v), 2526 NEONMAP0(vextq_v), 2527 NEONMAP0(vfma_v), 2528 NEONMAP0(vfmaq_v), 2529 NEONMAP2(vhadd_v, arm_neon_vhaddu, arm_neon_vhadds, Add1ArgType | UnsignedAlts), 2530 NEONMAP2(vhaddq_v, arm_neon_vhaddu, arm_neon_vhadds, Add1ArgType | UnsignedAlts), 2531 NEONMAP2(vhsub_v, arm_neon_vhsubu, arm_neon_vhsubs, Add1ArgType | UnsignedAlts), 2532 NEONMAP2(vhsubq_v, arm_neon_vhsubu, arm_neon_vhsubs, Add1ArgType | UnsignedAlts), 2533 NEONMAP0(vld1_dup_v), 2534 NEONMAP1(vld1_v, arm_neon_vld1, 0), 2535 NEONMAP0(vld1q_dup_v), 2536 NEONMAP1(vld1q_v, arm_neon_vld1, 0), 2537 NEONMAP1(vld2_lane_v, arm_neon_vld2lane, 0), 2538 NEONMAP1(vld2_v, arm_neon_vld2, 0), 2539 NEONMAP1(vld2q_lane_v, arm_neon_vld2lane, 0), 2540 NEONMAP1(vld2q_v, arm_neon_vld2, 0), 2541 NEONMAP1(vld3_lane_v, arm_neon_vld3lane, 0), 2542 NEONMAP1(vld3_v, arm_neon_vld3, 0), 2543 NEONMAP1(vld3q_lane_v, arm_neon_vld3lane, 0), 2544 NEONMAP1(vld3q_v, arm_neon_vld3, 0), 2545 NEONMAP1(vld4_lane_v, arm_neon_vld4lane, 0), 2546 NEONMAP1(vld4_v, arm_neon_vld4, 0), 2547 NEONMAP1(vld4q_lane_v, arm_neon_vld4lane, 0), 2548 NEONMAP1(vld4q_v, arm_neon_vld4, 0), 2549 NEONMAP2(vmax_v, arm_neon_vmaxu, arm_neon_vmaxs, Add1ArgType | UnsignedAlts), 2550 NEONMAP1(vmaxnm_v, arm_neon_vmaxnm, Add1ArgType), 2551 NEONMAP1(vmaxnmq_v, arm_neon_vmaxnm, Add1ArgType), 2552 NEONMAP2(vmaxq_v, arm_neon_vmaxu, arm_neon_vmaxs, Add1ArgType | UnsignedAlts), 2553 NEONMAP2(vmin_v, arm_neon_vminu, arm_neon_vmins, Add1ArgType | UnsignedAlts), 2554 NEONMAP1(vminnm_v, arm_neon_vminnm, Add1ArgType), 2555 NEONMAP1(vminnmq_v, arm_neon_vminnm, Add1ArgType), 2556 NEONMAP2(vminq_v, arm_neon_vminu, arm_neon_vmins, Add1ArgType | UnsignedAlts), 2557 NEONMAP0(vmovl_v), 2558 NEONMAP0(vmovn_v), 2559 NEONMAP1(vmul_v, arm_neon_vmulp, Add1ArgType), 2560 NEONMAP0(vmull_v), 2561 NEONMAP1(vmulq_v, arm_neon_vmulp, Add1ArgType), 2562 NEONMAP2(vpadal_v, arm_neon_vpadalu, arm_neon_vpadals, UnsignedAlts), 2563 NEONMAP2(vpadalq_v, arm_neon_vpadalu, arm_neon_vpadals, UnsignedAlts), 2564 NEONMAP1(vpadd_v, arm_neon_vpadd, Add1ArgType), 2565 NEONMAP2(vpaddl_v, arm_neon_vpaddlu, arm_neon_vpaddls, UnsignedAlts), 2566 NEONMAP2(vpaddlq_v, arm_neon_vpaddlu, arm_neon_vpaddls, UnsignedAlts), 2567 NEONMAP1(vpaddq_v, arm_neon_vpadd, Add1ArgType), 2568 NEONMAP2(vpmax_v, arm_neon_vpmaxu, arm_neon_vpmaxs, Add1ArgType | UnsignedAlts), 2569 NEONMAP2(vpmin_v, arm_neon_vpminu, arm_neon_vpmins, Add1ArgType | UnsignedAlts), 2570 NEONMAP1(vqabs_v, arm_neon_vqabs, Add1ArgType), 2571 NEONMAP1(vqabsq_v, arm_neon_vqabs, Add1ArgType), 2572 NEONMAP2(vqadd_v, arm_neon_vqaddu, arm_neon_vqadds, Add1ArgType | UnsignedAlts), 2573 NEONMAP2(vqaddq_v, arm_neon_vqaddu, arm_neon_vqadds, Add1ArgType | UnsignedAlts), 2574 NEONMAP2(vqdmlal_v, arm_neon_vqdmull, arm_neon_vqadds, 0), 2575 NEONMAP2(vqdmlsl_v, arm_neon_vqdmull, arm_neon_vqsubs, 0), 2576 NEONMAP1(vqdmulh_v, arm_neon_vqdmulh, Add1ArgType), 2577 NEONMAP1(vqdmulhq_v, arm_neon_vqdmulh, Add1ArgType), 2578 NEONMAP1(vqdmull_v, arm_neon_vqdmull, Add1ArgType), 2579 NEONMAP2(vqmovn_v, arm_neon_vqmovnu, arm_neon_vqmovns, Add1ArgType | UnsignedAlts), 2580 NEONMAP1(vqmovun_v, arm_neon_vqmovnsu, Add1ArgType), 2581 NEONMAP1(vqneg_v, arm_neon_vqneg, Add1ArgType), 2582 NEONMAP1(vqnegq_v, arm_neon_vqneg, Add1ArgType), 2583 NEONMAP1(vqrdmulh_v, arm_neon_vqrdmulh, Add1ArgType), 2584 NEONMAP1(vqrdmulhq_v, arm_neon_vqrdmulh, Add1ArgType), 2585 NEONMAP2(vqrshl_v, arm_neon_vqrshiftu, arm_neon_vqrshifts, Add1ArgType | UnsignedAlts), 2586 NEONMAP2(vqrshlq_v, arm_neon_vqrshiftu, arm_neon_vqrshifts, Add1ArgType | UnsignedAlts), 2587 NEONMAP2(vqshl_n_v, arm_neon_vqshiftu, arm_neon_vqshifts, UnsignedAlts), 2588 NEONMAP2(vqshl_v, arm_neon_vqshiftu, arm_neon_vqshifts, Add1ArgType | UnsignedAlts), 2589 NEONMAP2(vqshlq_n_v, arm_neon_vqshiftu, arm_neon_vqshifts, UnsignedAlts), 2590 NEONMAP2(vqshlq_v, arm_neon_vqshiftu, arm_neon_vqshifts, Add1ArgType | UnsignedAlts), 2591 NEONMAP1(vqshlu_n_v, arm_neon_vqshiftsu, 0), 2592 NEONMAP1(vqshluq_n_v, arm_neon_vqshiftsu, 0), 2593 NEONMAP2(vqsub_v, arm_neon_vqsubu, arm_neon_vqsubs, Add1ArgType | UnsignedAlts), 2594 NEONMAP2(vqsubq_v, arm_neon_vqsubu, arm_neon_vqsubs, Add1ArgType | UnsignedAlts), 2595 NEONMAP1(vraddhn_v, arm_neon_vraddhn, Add1ArgType), 2596 NEONMAP2(vrecpe_v, arm_neon_vrecpe, arm_neon_vrecpe, 0), 2597 NEONMAP2(vrecpeq_v, arm_neon_vrecpe, arm_neon_vrecpe, 0), 2598 NEONMAP1(vrecps_v, arm_neon_vrecps, Add1ArgType), 2599 NEONMAP1(vrecpsq_v, arm_neon_vrecps, Add1ArgType), 2600 NEONMAP2(vrhadd_v, arm_neon_vrhaddu, arm_neon_vrhadds, Add1ArgType | UnsignedAlts), 2601 NEONMAP2(vrhaddq_v, arm_neon_vrhaddu, arm_neon_vrhadds, Add1ArgType | UnsignedAlts), 2602 NEONMAP1(vrnd_v, arm_neon_vrintz, Add1ArgType), 2603 NEONMAP1(vrnda_v, arm_neon_vrinta, Add1ArgType), 2604 NEONMAP1(vrndaq_v, arm_neon_vrinta, Add1ArgType), 2605 NEONMAP1(vrndm_v, arm_neon_vrintm, Add1ArgType), 2606 NEONMAP1(vrndmq_v, arm_neon_vrintm, Add1ArgType), 2607 NEONMAP1(vrndn_v, arm_neon_vrintn, Add1ArgType), 2608 NEONMAP1(vrndnq_v, arm_neon_vrintn, Add1ArgType), 2609 NEONMAP1(vrndp_v, arm_neon_vrintp, Add1ArgType), 2610 NEONMAP1(vrndpq_v, arm_neon_vrintp, Add1ArgType), 2611 NEONMAP1(vrndq_v, arm_neon_vrintz, Add1ArgType), 2612 NEONMAP1(vrndx_v, arm_neon_vrintx, Add1ArgType), 2613 NEONMAP1(vrndxq_v, arm_neon_vrintx, Add1ArgType), 2614 NEONMAP2(vrshl_v, arm_neon_vrshiftu, arm_neon_vrshifts, Add1ArgType | UnsignedAlts), 2615 NEONMAP2(vrshlq_v, arm_neon_vrshiftu, arm_neon_vrshifts, Add1ArgType | UnsignedAlts), 2616 NEONMAP2(vrshr_n_v, arm_neon_vrshiftu, arm_neon_vrshifts, UnsignedAlts), 2617 NEONMAP2(vrshrq_n_v, arm_neon_vrshiftu, arm_neon_vrshifts, UnsignedAlts), 2618 NEONMAP2(vrsqrte_v, arm_neon_vrsqrte, arm_neon_vrsqrte, 0), 2619 NEONMAP2(vrsqrteq_v, arm_neon_vrsqrte, arm_neon_vrsqrte, 0), 2620 NEONMAP1(vrsqrts_v, arm_neon_vrsqrts, Add1ArgType), 2621 NEONMAP1(vrsqrtsq_v, arm_neon_vrsqrts, Add1ArgType), 2622 NEONMAP1(vrsubhn_v, arm_neon_vrsubhn, Add1ArgType), 2623 NEONMAP1(vsha1su0q_v, arm_neon_sha1su0, 0), 2624 NEONMAP1(vsha1su1q_v, arm_neon_sha1su1, 0), 2625 NEONMAP1(vsha256h2q_v, arm_neon_sha256h2, 0), 2626 NEONMAP1(vsha256hq_v, arm_neon_sha256h, 0), 2627 NEONMAP1(vsha256su0q_v, arm_neon_sha256su0, 0), 2628 NEONMAP1(vsha256su1q_v, arm_neon_sha256su1, 0), 2629 NEONMAP0(vshl_n_v), 2630 NEONMAP2(vshl_v, arm_neon_vshiftu, arm_neon_vshifts, Add1ArgType | UnsignedAlts), 2631 NEONMAP0(vshll_n_v), 2632 NEONMAP0(vshlq_n_v), 2633 NEONMAP2(vshlq_v, arm_neon_vshiftu, arm_neon_vshifts, Add1ArgType | UnsignedAlts), 2634 NEONMAP0(vshr_n_v), 2635 NEONMAP0(vshrn_n_v), 2636 NEONMAP0(vshrq_n_v), 2637 NEONMAP1(vst1_v, arm_neon_vst1, 0), 2638 NEONMAP1(vst1q_v, arm_neon_vst1, 0), 2639 NEONMAP1(vst2_lane_v, arm_neon_vst2lane, 0), 2640 NEONMAP1(vst2_v, arm_neon_vst2, 0), 2641 NEONMAP1(vst2q_lane_v, arm_neon_vst2lane, 0), 2642 NEONMAP1(vst2q_v, arm_neon_vst2, 0), 2643 NEONMAP1(vst3_lane_v, arm_neon_vst3lane, 0), 2644 NEONMAP1(vst3_v, arm_neon_vst3, 0), 2645 NEONMAP1(vst3q_lane_v, arm_neon_vst3lane, 0), 2646 NEONMAP1(vst3q_v, arm_neon_vst3, 0), 2647 NEONMAP1(vst4_lane_v, arm_neon_vst4lane, 0), 2648 NEONMAP1(vst4_v, arm_neon_vst4, 0), 2649 NEONMAP1(vst4q_lane_v, arm_neon_vst4lane, 0), 2650 NEONMAP1(vst4q_v, arm_neon_vst4, 0), 2651 NEONMAP0(vsubhn_v), 2652 NEONMAP0(vtrn_v), 2653 NEONMAP0(vtrnq_v), 2654 NEONMAP0(vtst_v), 2655 NEONMAP0(vtstq_v), 2656 NEONMAP0(vuzp_v), 2657 NEONMAP0(vuzpq_v), 2658 NEONMAP0(vzip_v), 2659 NEONMAP0(vzipq_v) 2660 }; 2661 2662 static const NeonIntrinsicInfo AArch64SIMDIntrinsicMap[] = { 2663 NEONMAP1(vabs_v, aarch64_neon_abs, 0), 2664 NEONMAP1(vabsq_v, aarch64_neon_abs, 0), 2665 NEONMAP0(vaddhn_v), 2666 NEONMAP1(vaesdq_v, aarch64_crypto_aesd, 0), 2667 NEONMAP1(vaeseq_v, aarch64_crypto_aese, 0), 2668 NEONMAP1(vaesimcq_v, aarch64_crypto_aesimc, 0), 2669 NEONMAP1(vaesmcq_v, aarch64_crypto_aesmc, 0), 2670 NEONMAP1(vcage_v, aarch64_neon_facge, 0), 2671 NEONMAP1(vcageq_v, aarch64_neon_facge, 0), 2672 NEONMAP1(vcagt_v, aarch64_neon_facgt, 0), 2673 NEONMAP1(vcagtq_v, aarch64_neon_facgt, 0), 2674 NEONMAP1(vcale_v, aarch64_neon_facge, 0), 2675 NEONMAP1(vcaleq_v, aarch64_neon_facge, 0), 2676 NEONMAP1(vcalt_v, aarch64_neon_facgt, 0), 2677 NEONMAP1(vcaltq_v, aarch64_neon_facgt, 0), 2678 NEONMAP1(vcls_v, aarch64_neon_cls, Add1ArgType), 2679 NEONMAP1(vclsq_v, aarch64_neon_cls, Add1ArgType), 2680 NEONMAP1(vclz_v, ctlz, Add1ArgType), 2681 NEONMAP1(vclzq_v, ctlz, Add1ArgType), 2682 NEONMAP1(vcnt_v, ctpop, Add1ArgType), 2683 NEONMAP1(vcntq_v, ctpop, Add1ArgType), 2684 NEONMAP1(vcvt_f16_f32, aarch64_neon_vcvtfp2hf, 0), 2685 NEONMAP1(vcvt_f32_f16, aarch64_neon_vcvthf2fp, 0), 2686 NEONMAP0(vcvt_f32_v), 2687 NEONMAP2(vcvt_n_f32_v, aarch64_neon_vcvtfxu2fp, aarch64_neon_vcvtfxs2fp, 0), 2688 NEONMAP2(vcvt_n_f64_v, aarch64_neon_vcvtfxu2fp, aarch64_neon_vcvtfxs2fp, 0), 2689 NEONMAP1(vcvt_n_s32_v, aarch64_neon_vcvtfp2fxs, 0), 2690 NEONMAP1(vcvt_n_s64_v, aarch64_neon_vcvtfp2fxs, 0), 2691 NEONMAP1(vcvt_n_u32_v, aarch64_neon_vcvtfp2fxu, 0), 2692 NEONMAP1(vcvt_n_u64_v, aarch64_neon_vcvtfp2fxu, 0), 2693 NEONMAP0(vcvtq_f32_v), 2694 NEONMAP2(vcvtq_n_f32_v, aarch64_neon_vcvtfxu2fp, aarch64_neon_vcvtfxs2fp, 0), 2695 NEONMAP2(vcvtq_n_f64_v, aarch64_neon_vcvtfxu2fp, aarch64_neon_vcvtfxs2fp, 0), 2696 NEONMAP1(vcvtq_n_s32_v, aarch64_neon_vcvtfp2fxs, 0), 2697 NEONMAP1(vcvtq_n_s64_v, aarch64_neon_vcvtfp2fxs, 0), 2698 NEONMAP1(vcvtq_n_u32_v, aarch64_neon_vcvtfp2fxu, 0), 2699 NEONMAP1(vcvtq_n_u64_v, aarch64_neon_vcvtfp2fxu, 0), 2700 NEONMAP1(vcvtx_f32_v, aarch64_neon_fcvtxn, AddRetType | Add1ArgType), 2701 NEONMAP0(vext_v), 2702 NEONMAP0(vextq_v), 2703 NEONMAP0(vfma_v), 2704 NEONMAP0(vfmaq_v), 2705 NEONMAP2(vhadd_v, aarch64_neon_uhadd, aarch64_neon_shadd, Add1ArgType | UnsignedAlts), 2706 NEONMAP2(vhaddq_v, aarch64_neon_uhadd, aarch64_neon_shadd, Add1ArgType | UnsignedAlts), 2707 NEONMAP2(vhsub_v, aarch64_neon_uhsub, aarch64_neon_shsub, Add1ArgType | UnsignedAlts), 2708 NEONMAP2(vhsubq_v, aarch64_neon_uhsub, aarch64_neon_shsub, Add1ArgType | UnsignedAlts), 2709 NEONMAP0(vmovl_v), 2710 NEONMAP0(vmovn_v), 2711 NEONMAP1(vmul_v, aarch64_neon_pmul, Add1ArgType), 2712 NEONMAP1(vmulq_v, aarch64_neon_pmul, Add1ArgType), 2713 NEONMAP1(vpadd_v, aarch64_neon_addp, Add1ArgType), 2714 NEONMAP2(vpaddl_v, aarch64_neon_uaddlp, aarch64_neon_saddlp, UnsignedAlts), 2715 NEONMAP2(vpaddlq_v, aarch64_neon_uaddlp, aarch64_neon_saddlp, UnsignedAlts), 2716 NEONMAP1(vpaddq_v, aarch64_neon_addp, Add1ArgType), 2717 NEONMAP1(vqabs_v, aarch64_neon_sqabs, Add1ArgType), 2718 NEONMAP1(vqabsq_v, aarch64_neon_sqabs, Add1ArgType), 2719 NEONMAP2(vqadd_v, aarch64_neon_uqadd, aarch64_neon_sqadd, Add1ArgType | UnsignedAlts), 2720 NEONMAP2(vqaddq_v, aarch64_neon_uqadd, aarch64_neon_sqadd, Add1ArgType | UnsignedAlts), 2721 NEONMAP2(vqdmlal_v, aarch64_neon_sqdmull, aarch64_neon_sqadd, 0), 2722 NEONMAP2(vqdmlsl_v, aarch64_neon_sqdmull, aarch64_neon_sqsub, 0), 2723 NEONMAP1(vqdmulh_v, aarch64_neon_sqdmulh, Add1ArgType), 2724 NEONMAP1(vqdmulhq_v, aarch64_neon_sqdmulh, Add1ArgType), 2725 NEONMAP1(vqdmull_v, aarch64_neon_sqdmull, Add1ArgType), 2726 NEONMAP2(vqmovn_v, aarch64_neon_uqxtn, aarch64_neon_sqxtn, Add1ArgType | UnsignedAlts), 2727 NEONMAP1(vqmovun_v, aarch64_neon_sqxtun, Add1ArgType), 2728 NEONMAP1(vqneg_v, aarch64_neon_sqneg, Add1ArgType), 2729 NEONMAP1(vqnegq_v, aarch64_neon_sqneg, Add1ArgType), 2730 NEONMAP1(vqrdmulh_v, aarch64_neon_sqrdmulh, Add1ArgType), 2731 NEONMAP1(vqrdmulhq_v, aarch64_neon_sqrdmulh, Add1ArgType), 2732 NEONMAP2(vqrshl_v, aarch64_neon_uqrshl, aarch64_neon_sqrshl, Add1ArgType | UnsignedAlts), 2733 NEONMAP2(vqrshlq_v, aarch64_neon_uqrshl, aarch64_neon_sqrshl, Add1ArgType | UnsignedAlts), 2734 NEONMAP2(vqshl_n_v, aarch64_neon_uqshl, aarch64_neon_sqshl, UnsignedAlts), 2735 NEONMAP2(vqshl_v, aarch64_neon_uqshl, aarch64_neon_sqshl, Add1ArgType | UnsignedAlts), 2736 NEONMAP2(vqshlq_n_v, aarch64_neon_uqshl, aarch64_neon_sqshl,UnsignedAlts), 2737 NEONMAP2(vqshlq_v, aarch64_neon_uqshl, aarch64_neon_sqshl, Add1ArgType | UnsignedAlts), 2738 NEONMAP1(vqshlu_n_v, aarch64_neon_sqshlu, 0), 2739 NEONMAP1(vqshluq_n_v, aarch64_neon_sqshlu, 0), 2740 NEONMAP2(vqsub_v, aarch64_neon_uqsub, aarch64_neon_sqsub, Add1ArgType | UnsignedAlts), 2741 NEONMAP2(vqsubq_v, aarch64_neon_uqsub, aarch64_neon_sqsub, Add1ArgType | UnsignedAlts), 2742 NEONMAP1(vraddhn_v, aarch64_neon_raddhn, Add1ArgType), 2743 NEONMAP2(vrecpe_v, aarch64_neon_frecpe, aarch64_neon_urecpe, 0), 2744 NEONMAP2(vrecpeq_v, aarch64_neon_frecpe, aarch64_neon_urecpe, 0), 2745 NEONMAP1(vrecps_v, aarch64_neon_frecps, Add1ArgType), 2746 NEONMAP1(vrecpsq_v, aarch64_neon_frecps, Add1ArgType), 2747 NEONMAP2(vrhadd_v, aarch64_neon_urhadd, aarch64_neon_srhadd, Add1ArgType | UnsignedAlts), 2748 NEONMAP2(vrhaddq_v, aarch64_neon_urhadd, aarch64_neon_srhadd, Add1ArgType | UnsignedAlts), 2749 NEONMAP2(vrshl_v, aarch64_neon_urshl, aarch64_neon_srshl, Add1ArgType | UnsignedAlts), 2750 NEONMAP2(vrshlq_v, aarch64_neon_urshl, aarch64_neon_srshl, Add1ArgType | UnsignedAlts), 2751 NEONMAP2(vrshr_n_v, aarch64_neon_urshl, aarch64_neon_srshl, UnsignedAlts), 2752 NEONMAP2(vrshrq_n_v, aarch64_neon_urshl, aarch64_neon_srshl, UnsignedAlts), 2753 NEONMAP2(vrsqrte_v, aarch64_neon_frsqrte, aarch64_neon_ursqrte, 0), 2754 NEONMAP2(vrsqrteq_v, aarch64_neon_frsqrte, aarch64_neon_ursqrte, 0), 2755 NEONMAP1(vrsqrts_v, aarch64_neon_frsqrts, Add1ArgType), 2756 NEONMAP1(vrsqrtsq_v, aarch64_neon_frsqrts, Add1ArgType), 2757 NEONMAP1(vrsubhn_v, aarch64_neon_rsubhn, Add1ArgType), 2758 NEONMAP1(vsha1su0q_v, aarch64_crypto_sha1su0, 0), 2759 NEONMAP1(vsha1su1q_v, aarch64_crypto_sha1su1, 0), 2760 NEONMAP1(vsha256h2q_v, aarch64_crypto_sha256h2, 0), 2761 NEONMAP1(vsha256hq_v, aarch64_crypto_sha256h, 0), 2762 NEONMAP1(vsha256su0q_v, aarch64_crypto_sha256su0, 0), 2763 NEONMAP1(vsha256su1q_v, aarch64_crypto_sha256su1, 0), 2764 NEONMAP0(vshl_n_v), 2765 NEONMAP2(vshl_v, aarch64_neon_ushl, aarch64_neon_sshl, Add1ArgType | UnsignedAlts), 2766 NEONMAP0(vshll_n_v), 2767 NEONMAP0(vshlq_n_v), 2768 NEONMAP2(vshlq_v, aarch64_neon_ushl, aarch64_neon_sshl, Add1ArgType | UnsignedAlts), 2769 NEONMAP0(vshr_n_v), 2770 NEONMAP0(vshrn_n_v), 2771 NEONMAP0(vshrq_n_v), 2772 NEONMAP0(vsubhn_v), 2773 NEONMAP0(vtst_v), 2774 NEONMAP0(vtstq_v), 2775 }; 2776 2777 static const NeonIntrinsicInfo AArch64SISDIntrinsicMap[] = { 2778 NEONMAP1(vabdd_f64, aarch64_sisd_fabd, Add1ArgType), 2779 NEONMAP1(vabds_f32, aarch64_sisd_fabd, Add1ArgType), 2780 NEONMAP1(vabsd_s64, aarch64_neon_abs, Add1ArgType), 2781 NEONMAP1(vaddlv_s32, aarch64_neon_saddlv, AddRetType | Add1ArgType), 2782 NEONMAP1(vaddlv_u32, aarch64_neon_uaddlv, AddRetType | Add1ArgType), 2783 NEONMAP1(vaddlvq_s32, aarch64_neon_saddlv, AddRetType | Add1ArgType), 2784 NEONMAP1(vaddlvq_u32, aarch64_neon_uaddlv, AddRetType | Add1ArgType), 2785 NEONMAP1(vaddv_f32, aarch64_neon_faddv, AddRetType | Add1ArgType), 2786 NEONMAP1(vaddv_s32, aarch64_neon_saddv, AddRetType | Add1ArgType), 2787 NEONMAP1(vaddv_u32, aarch64_neon_uaddv, AddRetType | Add1ArgType), 2788 NEONMAP1(vaddvq_f32, aarch64_neon_faddv, AddRetType | Add1ArgType), 2789 NEONMAP1(vaddvq_f64, aarch64_neon_faddv, AddRetType | Add1ArgType), 2790 NEONMAP1(vaddvq_s32, aarch64_neon_saddv, AddRetType | Add1ArgType), 2791 NEONMAP1(vaddvq_s64, aarch64_neon_saddv, AddRetType | Add1ArgType), 2792 NEONMAP1(vaddvq_u32, aarch64_neon_uaddv, AddRetType | Add1ArgType), 2793 NEONMAP1(vaddvq_u64, aarch64_neon_uaddv, AddRetType | Add1ArgType), 2794 NEONMAP1(vcaged_f64, aarch64_neon_facge, AddRetType | Add1ArgType), 2795 NEONMAP1(vcages_f32, aarch64_neon_facge, AddRetType | Add1ArgType), 2796 NEONMAP1(vcagtd_f64, aarch64_neon_facgt, AddRetType | Add1ArgType), 2797 NEONMAP1(vcagts_f32, aarch64_neon_facgt, AddRetType | Add1ArgType), 2798 NEONMAP1(vcaled_f64, aarch64_neon_facge, AddRetType | Add1ArgType), 2799 NEONMAP1(vcales_f32, aarch64_neon_facge, AddRetType | Add1ArgType), 2800 NEONMAP1(vcaltd_f64, aarch64_neon_facgt, AddRetType | Add1ArgType), 2801 NEONMAP1(vcalts_f32, aarch64_neon_facgt, AddRetType | Add1ArgType), 2802 NEONMAP1(vcvtad_s64_f64, aarch64_neon_fcvtas, AddRetType | Add1ArgType), 2803 NEONMAP1(vcvtad_u64_f64, aarch64_neon_fcvtau, AddRetType | Add1ArgType), 2804 NEONMAP1(vcvtas_s32_f32, aarch64_neon_fcvtas, AddRetType | Add1ArgType), 2805 NEONMAP1(vcvtas_u32_f32, aarch64_neon_fcvtau, AddRetType | Add1ArgType), 2806 NEONMAP1(vcvtd_n_f64_s64, aarch64_neon_vcvtfxs2fp, AddRetType | Add1ArgType), 2807 NEONMAP1(vcvtd_n_f64_u64, aarch64_neon_vcvtfxu2fp, AddRetType | Add1ArgType), 2808 NEONMAP1(vcvtd_n_s64_f64, aarch64_neon_vcvtfp2fxs, AddRetType | Add1ArgType), 2809 NEONMAP1(vcvtd_n_u64_f64, aarch64_neon_vcvtfp2fxu, AddRetType | Add1ArgType), 2810 NEONMAP1(vcvtmd_s64_f64, aarch64_neon_fcvtms, AddRetType | Add1ArgType), 2811 NEONMAP1(vcvtmd_u64_f64, aarch64_neon_fcvtmu, AddRetType | Add1ArgType), 2812 NEONMAP1(vcvtms_s32_f32, aarch64_neon_fcvtms, AddRetType | Add1ArgType), 2813 NEONMAP1(vcvtms_u32_f32, aarch64_neon_fcvtmu, AddRetType | Add1ArgType), 2814 NEONMAP1(vcvtnd_s64_f64, aarch64_neon_fcvtns, AddRetType | Add1ArgType), 2815 NEONMAP1(vcvtnd_u64_f64, aarch64_neon_fcvtnu, AddRetType | Add1ArgType), 2816 NEONMAP1(vcvtns_s32_f32, aarch64_neon_fcvtns, AddRetType | Add1ArgType), 2817 NEONMAP1(vcvtns_u32_f32, aarch64_neon_fcvtnu, AddRetType | Add1ArgType), 2818 NEONMAP1(vcvtpd_s64_f64, aarch64_neon_fcvtps, AddRetType | Add1ArgType), 2819 NEONMAP1(vcvtpd_u64_f64, aarch64_neon_fcvtpu, AddRetType | Add1ArgType), 2820 NEONMAP1(vcvtps_s32_f32, aarch64_neon_fcvtps, AddRetType | Add1ArgType), 2821 NEONMAP1(vcvtps_u32_f32, aarch64_neon_fcvtpu, AddRetType | Add1ArgType), 2822 NEONMAP1(vcvts_n_f32_s32, aarch64_neon_vcvtfxs2fp, AddRetType | Add1ArgType), 2823 NEONMAP1(vcvts_n_f32_u32, aarch64_neon_vcvtfxu2fp, AddRetType | Add1ArgType), 2824 NEONMAP1(vcvts_n_s32_f32, aarch64_neon_vcvtfp2fxs, AddRetType | Add1ArgType), 2825 NEONMAP1(vcvts_n_u32_f32, aarch64_neon_vcvtfp2fxu, AddRetType | Add1ArgType), 2826 NEONMAP1(vcvtxd_f32_f64, aarch64_sisd_fcvtxn, 0), 2827 NEONMAP1(vmaxnmv_f32, aarch64_neon_fmaxnmv, AddRetType | Add1ArgType), 2828 NEONMAP1(vmaxnmvq_f32, aarch64_neon_fmaxnmv, AddRetType | Add1ArgType), 2829 NEONMAP1(vmaxnmvq_f64, aarch64_neon_fmaxnmv, AddRetType | Add1ArgType), 2830 NEONMAP1(vmaxv_f32, aarch64_neon_fmaxv, AddRetType | Add1ArgType), 2831 NEONMAP1(vmaxv_s32, aarch64_neon_smaxv, AddRetType | Add1ArgType), 2832 NEONMAP1(vmaxv_u32, aarch64_neon_umaxv, AddRetType | Add1ArgType), 2833 NEONMAP1(vmaxvq_f32, aarch64_neon_fmaxv, AddRetType | Add1ArgType), 2834 NEONMAP1(vmaxvq_f64, aarch64_neon_fmaxv, AddRetType | Add1ArgType), 2835 NEONMAP1(vmaxvq_s32, aarch64_neon_smaxv, AddRetType | Add1ArgType), 2836 NEONMAP1(vmaxvq_u32, aarch64_neon_umaxv, AddRetType | Add1ArgType), 2837 NEONMAP1(vminnmv_f32, aarch64_neon_fminnmv, AddRetType | Add1ArgType), 2838 NEONMAP1(vminnmvq_f32, aarch64_neon_fminnmv, AddRetType | Add1ArgType), 2839 NEONMAP1(vminnmvq_f64, aarch64_neon_fminnmv, AddRetType | Add1ArgType), 2840 NEONMAP1(vminv_f32, aarch64_neon_fminv, AddRetType | Add1ArgType), 2841 NEONMAP1(vminv_s32, aarch64_neon_sminv, AddRetType | Add1ArgType), 2842 NEONMAP1(vminv_u32, aarch64_neon_uminv, AddRetType | Add1ArgType), 2843 NEONMAP1(vminvq_f32, aarch64_neon_fminv, AddRetType | Add1ArgType), 2844 NEONMAP1(vminvq_f64, aarch64_neon_fminv, AddRetType | Add1ArgType), 2845 NEONMAP1(vminvq_s32, aarch64_neon_sminv, AddRetType | Add1ArgType), 2846 NEONMAP1(vminvq_u32, aarch64_neon_uminv, AddRetType | Add1ArgType), 2847 NEONMAP1(vmull_p64, aarch64_neon_pmull64, 0), 2848 NEONMAP1(vmulxd_f64, aarch64_neon_fmulx, Add1ArgType), 2849 NEONMAP1(vmulxs_f32, aarch64_neon_fmulx, Add1ArgType), 2850 NEONMAP1(vpaddd_s64, aarch64_neon_uaddv, AddRetType | Add1ArgType), 2851 NEONMAP1(vpaddd_u64, aarch64_neon_uaddv, AddRetType | Add1ArgType), 2852 NEONMAP1(vpmaxnmqd_f64, aarch64_neon_fmaxnmv, AddRetType | Add1ArgType), 2853 NEONMAP1(vpmaxnms_f32, aarch64_neon_fmaxnmv, AddRetType | Add1ArgType), 2854 NEONMAP1(vpmaxqd_f64, aarch64_neon_fmaxv, AddRetType | Add1ArgType), 2855 NEONMAP1(vpmaxs_f32, aarch64_neon_fmaxv, AddRetType | Add1ArgType), 2856 NEONMAP1(vpminnmqd_f64, aarch64_neon_fminnmv, AddRetType | Add1ArgType), 2857 NEONMAP1(vpminnms_f32, aarch64_neon_fminnmv, AddRetType | Add1ArgType), 2858 NEONMAP1(vpminqd_f64, aarch64_neon_fminv, AddRetType | Add1ArgType), 2859 NEONMAP1(vpmins_f32, aarch64_neon_fminv, AddRetType | Add1ArgType), 2860 NEONMAP1(vqabsb_s8, aarch64_neon_sqabs, Vectorize1ArgType | Use64BitVectors), 2861 NEONMAP1(vqabsd_s64, aarch64_neon_sqabs, Add1ArgType), 2862 NEONMAP1(vqabsh_s16, aarch64_neon_sqabs, Vectorize1ArgType | Use64BitVectors), 2863 NEONMAP1(vqabss_s32, aarch64_neon_sqabs, Add1ArgType), 2864 NEONMAP1(vqaddb_s8, aarch64_neon_sqadd, Vectorize1ArgType | Use64BitVectors), 2865 NEONMAP1(vqaddb_u8, aarch64_neon_uqadd, Vectorize1ArgType | Use64BitVectors), 2866 NEONMAP1(vqaddd_s64, aarch64_neon_sqadd, Add1ArgType), 2867 NEONMAP1(vqaddd_u64, aarch64_neon_uqadd, Add1ArgType), 2868 NEONMAP1(vqaddh_s16, aarch64_neon_sqadd, Vectorize1ArgType | Use64BitVectors), 2869 NEONMAP1(vqaddh_u16, aarch64_neon_uqadd, Vectorize1ArgType | Use64BitVectors), 2870 NEONMAP1(vqadds_s32, aarch64_neon_sqadd, Add1ArgType), 2871 NEONMAP1(vqadds_u32, aarch64_neon_uqadd, Add1ArgType), 2872 NEONMAP1(vqdmulhh_s16, aarch64_neon_sqdmulh, Vectorize1ArgType | Use64BitVectors), 2873 NEONMAP1(vqdmulhs_s32, aarch64_neon_sqdmulh, Add1ArgType), 2874 NEONMAP1(vqdmullh_s16, aarch64_neon_sqdmull, VectorRet | Use128BitVectors), 2875 NEONMAP1(vqdmulls_s32, aarch64_neon_sqdmulls_scalar, 0), 2876 NEONMAP1(vqmovnd_s64, aarch64_neon_scalar_sqxtn, AddRetType | Add1ArgType), 2877 NEONMAP1(vqmovnd_u64, aarch64_neon_scalar_uqxtn, AddRetType | Add1ArgType), 2878 NEONMAP1(vqmovnh_s16, aarch64_neon_sqxtn, VectorRet | Use64BitVectors), 2879 NEONMAP1(vqmovnh_u16, aarch64_neon_uqxtn, VectorRet | Use64BitVectors), 2880 NEONMAP1(vqmovns_s32, aarch64_neon_sqxtn, VectorRet | Use64BitVectors), 2881 NEONMAP1(vqmovns_u32, aarch64_neon_uqxtn, VectorRet | Use64BitVectors), 2882 NEONMAP1(vqmovund_s64, aarch64_neon_scalar_sqxtun, AddRetType | Add1ArgType), 2883 NEONMAP1(vqmovunh_s16, aarch64_neon_sqxtun, VectorRet | Use64BitVectors), 2884 NEONMAP1(vqmovuns_s32, aarch64_neon_sqxtun, VectorRet | Use64BitVectors), 2885 NEONMAP1(vqnegb_s8, aarch64_neon_sqneg, Vectorize1ArgType | Use64BitVectors), 2886 NEONMAP1(vqnegd_s64, aarch64_neon_sqneg, Add1ArgType), 2887 NEONMAP1(vqnegh_s16, aarch64_neon_sqneg, Vectorize1ArgType | Use64BitVectors), 2888 NEONMAP1(vqnegs_s32, aarch64_neon_sqneg, Add1ArgType), 2889 NEONMAP1(vqrdmulhh_s16, aarch64_neon_sqrdmulh, Vectorize1ArgType | Use64BitVectors), 2890 NEONMAP1(vqrdmulhs_s32, aarch64_neon_sqrdmulh, Add1ArgType), 2891 NEONMAP1(vqrshlb_s8, aarch64_neon_sqrshl, Vectorize1ArgType | Use64BitVectors), 2892 NEONMAP1(vqrshlb_u8, aarch64_neon_uqrshl, Vectorize1ArgType | Use64BitVectors), 2893 NEONMAP1(vqrshld_s64, aarch64_neon_sqrshl, Add1ArgType), 2894 NEONMAP1(vqrshld_u64, aarch64_neon_uqrshl, Add1ArgType), 2895 NEONMAP1(vqrshlh_s16, aarch64_neon_sqrshl, Vectorize1ArgType | Use64BitVectors), 2896 NEONMAP1(vqrshlh_u16, aarch64_neon_uqrshl, Vectorize1ArgType | Use64BitVectors), 2897 NEONMAP1(vqrshls_s32, aarch64_neon_sqrshl, Add1ArgType), 2898 NEONMAP1(vqrshls_u32, aarch64_neon_uqrshl, Add1ArgType), 2899 NEONMAP1(vqrshrnd_n_s64, aarch64_neon_sqrshrn, AddRetType), 2900 NEONMAP1(vqrshrnd_n_u64, aarch64_neon_uqrshrn, AddRetType), 2901 NEONMAP1(vqrshrnh_n_s16, aarch64_neon_sqrshrn, VectorRet | Use64BitVectors), 2902 NEONMAP1(vqrshrnh_n_u16, aarch64_neon_uqrshrn, VectorRet | Use64BitVectors), 2903 NEONMAP1(vqrshrns_n_s32, aarch64_neon_sqrshrn, VectorRet | Use64BitVectors), 2904 NEONMAP1(vqrshrns_n_u32, aarch64_neon_uqrshrn, VectorRet | Use64BitVectors), 2905 NEONMAP1(vqrshrund_n_s64, aarch64_neon_sqrshrun, AddRetType), 2906 NEONMAP1(vqrshrunh_n_s16, aarch64_neon_sqrshrun, VectorRet | Use64BitVectors), 2907 NEONMAP1(vqrshruns_n_s32, aarch64_neon_sqrshrun, VectorRet | Use64BitVectors), 2908 NEONMAP1(vqshlb_n_s8, aarch64_neon_sqshl, Vectorize1ArgType | Use64BitVectors), 2909 NEONMAP1(vqshlb_n_u8, aarch64_neon_uqshl, Vectorize1ArgType | Use64BitVectors), 2910 NEONMAP1(vqshlb_s8, aarch64_neon_sqshl, Vectorize1ArgType | Use64BitVectors), 2911 NEONMAP1(vqshlb_u8, aarch64_neon_uqshl, Vectorize1ArgType | Use64BitVectors), 2912 NEONMAP1(vqshld_s64, aarch64_neon_sqshl, Add1ArgType), 2913 NEONMAP1(vqshld_u64, aarch64_neon_uqshl, Add1ArgType), 2914 NEONMAP1(vqshlh_n_s16, aarch64_neon_sqshl, Vectorize1ArgType | Use64BitVectors), 2915 NEONMAP1(vqshlh_n_u16, aarch64_neon_uqshl, Vectorize1ArgType | Use64BitVectors), 2916 NEONMAP1(vqshlh_s16, aarch64_neon_sqshl, Vectorize1ArgType | Use64BitVectors), 2917 NEONMAP1(vqshlh_u16, aarch64_neon_uqshl, Vectorize1ArgType | Use64BitVectors), 2918 NEONMAP1(vqshls_n_s32, aarch64_neon_sqshl, Add1ArgType), 2919 NEONMAP1(vqshls_n_u32, aarch64_neon_uqshl, Add1ArgType), 2920 NEONMAP1(vqshls_s32, aarch64_neon_sqshl, Add1ArgType), 2921 NEONMAP1(vqshls_u32, aarch64_neon_uqshl, Add1ArgType), 2922 NEONMAP1(vqshlub_n_s8, aarch64_neon_sqshlu, Vectorize1ArgType | Use64BitVectors), 2923 NEONMAP1(vqshluh_n_s16, aarch64_neon_sqshlu, Vectorize1ArgType | Use64BitVectors), 2924 NEONMAP1(vqshlus_n_s32, aarch64_neon_sqshlu, Add1ArgType), 2925 NEONMAP1(vqshrnd_n_s64, aarch64_neon_sqshrn, AddRetType), 2926 NEONMAP1(vqshrnd_n_u64, aarch64_neon_uqshrn, AddRetType), 2927 NEONMAP1(vqshrnh_n_s16, aarch64_neon_sqshrn, VectorRet | Use64BitVectors), 2928 NEONMAP1(vqshrnh_n_u16, aarch64_neon_uqshrn, VectorRet | Use64BitVectors), 2929 NEONMAP1(vqshrns_n_s32, aarch64_neon_sqshrn, VectorRet | Use64BitVectors), 2930 NEONMAP1(vqshrns_n_u32, aarch64_neon_uqshrn, VectorRet | Use64BitVectors), 2931 NEONMAP1(vqshrund_n_s64, aarch64_neon_sqshrun, AddRetType), 2932 NEONMAP1(vqshrunh_n_s16, aarch64_neon_sqshrun, VectorRet | Use64BitVectors), 2933 NEONMAP1(vqshruns_n_s32, aarch64_neon_sqshrun, VectorRet | Use64BitVectors), 2934 NEONMAP1(vqsubb_s8, aarch64_neon_sqsub, Vectorize1ArgType | Use64BitVectors), 2935 NEONMAP1(vqsubb_u8, aarch64_neon_uqsub, Vectorize1ArgType | Use64BitVectors), 2936 NEONMAP1(vqsubd_s64, aarch64_neon_sqsub, Add1ArgType), 2937 NEONMAP1(vqsubd_u64, aarch64_neon_uqsub, Add1ArgType), 2938 NEONMAP1(vqsubh_s16, aarch64_neon_sqsub, Vectorize1ArgType | Use64BitVectors), 2939 NEONMAP1(vqsubh_u16, aarch64_neon_uqsub, Vectorize1ArgType | Use64BitVectors), 2940 NEONMAP1(vqsubs_s32, aarch64_neon_sqsub, Add1ArgType), 2941 NEONMAP1(vqsubs_u32, aarch64_neon_uqsub, Add1ArgType), 2942 NEONMAP1(vrecped_f64, aarch64_neon_frecpe, Add1ArgType), 2943 NEONMAP1(vrecpes_f32, aarch64_neon_frecpe, Add1ArgType), 2944 NEONMAP1(vrecpxd_f64, aarch64_neon_frecpx, Add1ArgType), 2945 NEONMAP1(vrecpxs_f32, aarch64_neon_frecpx, Add1ArgType), 2946 NEONMAP1(vrshld_s64, aarch64_neon_srshl, Add1ArgType), 2947 NEONMAP1(vrshld_u64, aarch64_neon_urshl, Add1ArgType), 2948 NEONMAP1(vrsqrted_f64, aarch64_neon_frsqrte, Add1ArgType), 2949 NEONMAP1(vrsqrtes_f32, aarch64_neon_frsqrte, Add1ArgType), 2950 NEONMAP1(vrsqrtsd_f64, aarch64_neon_frsqrts, Add1ArgType), 2951 NEONMAP1(vrsqrtss_f32, aarch64_neon_frsqrts, Add1ArgType), 2952 NEONMAP1(vsha1cq_u32, aarch64_crypto_sha1c, 0), 2953 NEONMAP1(vsha1h_u32, aarch64_crypto_sha1h, 0), 2954 NEONMAP1(vsha1mq_u32, aarch64_crypto_sha1m, 0), 2955 NEONMAP1(vsha1pq_u32, aarch64_crypto_sha1p, 0), 2956 NEONMAP1(vshld_s64, aarch64_neon_sshl, Add1ArgType), 2957 NEONMAP1(vshld_u64, aarch64_neon_ushl, Add1ArgType), 2958 NEONMAP1(vslid_n_s64, aarch64_neon_vsli, Vectorize1ArgType), 2959 NEONMAP1(vslid_n_u64, aarch64_neon_vsli, Vectorize1ArgType), 2960 NEONMAP1(vsqaddb_u8, aarch64_neon_usqadd, Vectorize1ArgType | Use64BitVectors), 2961 NEONMAP1(vsqaddd_u64, aarch64_neon_usqadd, Add1ArgType), 2962 NEONMAP1(vsqaddh_u16, aarch64_neon_usqadd, Vectorize1ArgType | Use64BitVectors), 2963 NEONMAP1(vsqadds_u32, aarch64_neon_usqadd, Add1ArgType), 2964 NEONMAP1(vsrid_n_s64, aarch64_neon_vsri, Vectorize1ArgType), 2965 NEONMAP1(vsrid_n_u64, aarch64_neon_vsri, Vectorize1ArgType), 2966 NEONMAP1(vuqaddb_s8, aarch64_neon_suqadd, Vectorize1ArgType | Use64BitVectors), 2967 NEONMAP1(vuqaddd_s64, aarch64_neon_suqadd, Add1ArgType), 2968 NEONMAP1(vuqaddh_s16, aarch64_neon_suqadd, Vectorize1ArgType | Use64BitVectors), 2969 NEONMAP1(vuqadds_s32, aarch64_neon_suqadd, Add1ArgType), 2970 }; 2971 2972 #undef NEONMAP0 2973 #undef NEONMAP1 2974 #undef NEONMAP2 2975 2976 static bool NEONSIMDIntrinsicsProvenSorted = false; 2977 2978 static bool AArch64SIMDIntrinsicsProvenSorted = false; 2979 static bool AArch64SISDIntrinsicsProvenSorted = false; 2980 2981 2982 static const NeonIntrinsicInfo * 2983 findNeonIntrinsicInMap(ArrayRef<NeonIntrinsicInfo> IntrinsicMap, 2984 unsigned BuiltinID, bool &MapProvenSorted) { 2985 2986 #ifndef NDEBUG 2987 if (!MapProvenSorted) { 2988 assert(std::is_sorted(std::begin(IntrinsicMap), std::end(IntrinsicMap))); 2989 MapProvenSorted = true; 2990 } 2991 #endif 2992 2993 const NeonIntrinsicInfo *Builtin = 2994 std::lower_bound(IntrinsicMap.begin(), IntrinsicMap.end(), BuiltinID); 2995 2996 if (Builtin != IntrinsicMap.end() && Builtin->BuiltinID == BuiltinID) 2997 return Builtin; 2998 2999 return nullptr; 3000 } 3001 3002 Function *CodeGenFunction::LookupNeonLLVMIntrinsic(unsigned IntrinsicID, 3003 unsigned Modifier, 3004 llvm::Type *ArgType, 3005 const CallExpr *E) { 3006 int VectorSize = 0; 3007 if (Modifier & Use64BitVectors) 3008 VectorSize = 64; 3009 else if (Modifier & Use128BitVectors) 3010 VectorSize = 128; 3011 3012 // Return type. 3013 SmallVector<llvm::Type *, 3> Tys; 3014 if (Modifier & AddRetType) { 3015 llvm::Type *Ty = ConvertType(E->getCallReturnType(getContext())); 3016 if (Modifier & VectorizeRetType) 3017 Ty = llvm::VectorType::get( 3018 Ty, VectorSize ? VectorSize / Ty->getPrimitiveSizeInBits() : 1); 3019 3020 Tys.push_back(Ty); 3021 } 3022 3023 // Arguments. 3024 if (Modifier & VectorizeArgTypes) { 3025 int Elts = VectorSize ? VectorSize / ArgType->getPrimitiveSizeInBits() : 1; 3026 ArgType = llvm::VectorType::get(ArgType, Elts); 3027 } 3028 3029 if (Modifier & (Add1ArgType | Add2ArgTypes)) 3030 Tys.push_back(ArgType); 3031 3032 if (Modifier & Add2ArgTypes) 3033 Tys.push_back(ArgType); 3034 3035 if (Modifier & InventFloatType) 3036 Tys.push_back(FloatTy); 3037 3038 return CGM.getIntrinsic(IntrinsicID, Tys); 3039 } 3040 3041 static Value *EmitCommonNeonSISDBuiltinExpr(CodeGenFunction &CGF, 3042 const NeonIntrinsicInfo &SISDInfo, 3043 SmallVectorImpl<Value *> &Ops, 3044 const CallExpr *E) { 3045 unsigned BuiltinID = SISDInfo.BuiltinID; 3046 unsigned int Int = SISDInfo.LLVMIntrinsic; 3047 unsigned Modifier = SISDInfo.TypeModifier; 3048 const char *s = SISDInfo.NameHint; 3049 3050 switch (BuiltinID) { 3051 case NEON::BI__builtin_neon_vcled_s64: 3052 case NEON::BI__builtin_neon_vcled_u64: 3053 case NEON::BI__builtin_neon_vcles_f32: 3054 case NEON::BI__builtin_neon_vcled_f64: 3055 case NEON::BI__builtin_neon_vcltd_s64: 3056 case NEON::BI__builtin_neon_vcltd_u64: 3057 case NEON::BI__builtin_neon_vclts_f32: 3058 case NEON::BI__builtin_neon_vcltd_f64: 3059 case NEON::BI__builtin_neon_vcales_f32: 3060 case NEON::BI__builtin_neon_vcaled_f64: 3061 case NEON::BI__builtin_neon_vcalts_f32: 3062 case NEON::BI__builtin_neon_vcaltd_f64: 3063 // Only one direction of comparisons actually exist, cmle is actually a cmge 3064 // with swapped operands. The table gives us the right intrinsic but we 3065 // still need to do the swap. 3066 std::swap(Ops[0], Ops[1]); 3067 break; 3068 } 3069 3070 assert(Int && "Generic code assumes a valid intrinsic"); 3071 3072 // Determine the type(s) of this overloaded AArch64 intrinsic. 3073 const Expr *Arg = E->getArg(0); 3074 llvm::Type *ArgTy = CGF.ConvertType(Arg->getType()); 3075 Function *F = CGF.LookupNeonLLVMIntrinsic(Int, Modifier, ArgTy, E); 3076 3077 int j = 0; 3078 ConstantInt *C0 = ConstantInt::get(CGF.SizeTy, 0); 3079 for (Function::const_arg_iterator ai = F->arg_begin(), ae = F->arg_end(); 3080 ai != ae; ++ai, ++j) { 3081 llvm::Type *ArgTy = ai->getType(); 3082 if (Ops[j]->getType()->getPrimitiveSizeInBits() == 3083 ArgTy->getPrimitiveSizeInBits()) 3084 continue; 3085 3086 assert(ArgTy->isVectorTy() && !Ops[j]->getType()->isVectorTy()); 3087 // The constant argument to an _n_ intrinsic always has Int32Ty, so truncate 3088 // it before inserting. 3089 Ops[j] = 3090 CGF.Builder.CreateTruncOrBitCast(Ops[j], ArgTy->getVectorElementType()); 3091 Ops[j] = 3092 CGF.Builder.CreateInsertElement(UndefValue::get(ArgTy), Ops[j], C0); 3093 } 3094 3095 Value *Result = CGF.EmitNeonCall(F, Ops, s); 3096 llvm::Type *ResultType = CGF.ConvertType(E->getType()); 3097 if (ResultType->getPrimitiveSizeInBits() < 3098 Result->getType()->getPrimitiveSizeInBits()) 3099 return CGF.Builder.CreateExtractElement(Result, C0); 3100 3101 return CGF.Builder.CreateBitCast(Result, ResultType, s); 3102 } 3103 3104 Value *CodeGenFunction::EmitCommonNeonBuiltinExpr( 3105 unsigned BuiltinID, unsigned LLVMIntrinsic, unsigned AltLLVMIntrinsic, 3106 const char *NameHint, unsigned Modifier, const CallExpr *E, 3107 SmallVectorImpl<llvm::Value *> &Ops, Address PtrOp0, Address PtrOp1) { 3108 // Get the last argument, which specifies the vector type. 3109 llvm::APSInt NeonTypeConst; 3110 const Expr *Arg = E->getArg(E->getNumArgs() - 1); 3111 if (!Arg->isIntegerConstantExpr(NeonTypeConst, getContext())) 3112 return nullptr; 3113 3114 // Determine the type of this overloaded NEON intrinsic. 3115 NeonTypeFlags Type(NeonTypeConst.getZExtValue()); 3116 bool Usgn = Type.isUnsigned(); 3117 bool Quad = Type.isQuad(); 3118 3119 llvm::VectorType *VTy = GetNeonType(this, Type); 3120 llvm::Type *Ty = VTy; 3121 if (!Ty) 3122 return nullptr; 3123 3124 auto getAlignmentValue32 = [&](Address addr) -> Value* { 3125 return Builder.getInt32(addr.getAlignment().getQuantity()); 3126 }; 3127 3128 unsigned Int = LLVMIntrinsic; 3129 if ((Modifier & UnsignedAlts) && !Usgn) 3130 Int = AltLLVMIntrinsic; 3131 3132 switch (BuiltinID) { 3133 default: break; 3134 case NEON::BI__builtin_neon_vabs_v: 3135 case NEON::BI__builtin_neon_vabsq_v: 3136 if (VTy->getElementType()->isFloatingPointTy()) 3137 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::fabs, Ty), Ops, "vabs"); 3138 return EmitNeonCall(CGM.getIntrinsic(LLVMIntrinsic, Ty), Ops, "vabs"); 3139 case NEON::BI__builtin_neon_vaddhn_v: { 3140 llvm::VectorType *SrcTy = 3141 llvm::VectorType::getExtendedElementVectorType(VTy); 3142 3143 // %sum = add <4 x i32> %lhs, %rhs 3144 Ops[0] = Builder.CreateBitCast(Ops[0], SrcTy); 3145 Ops[1] = Builder.CreateBitCast(Ops[1], SrcTy); 3146 Ops[0] = Builder.CreateAdd(Ops[0], Ops[1], "vaddhn"); 3147 3148 // %high = lshr <4 x i32> %sum, <i32 16, i32 16, i32 16, i32 16> 3149 Constant *ShiftAmt = 3150 ConstantInt::get(SrcTy, SrcTy->getScalarSizeInBits() / 2); 3151 Ops[0] = Builder.CreateLShr(Ops[0], ShiftAmt, "vaddhn"); 3152 3153 // %res = trunc <4 x i32> %high to <4 x i16> 3154 return Builder.CreateTrunc(Ops[0], VTy, "vaddhn"); 3155 } 3156 case NEON::BI__builtin_neon_vcale_v: 3157 case NEON::BI__builtin_neon_vcaleq_v: 3158 case NEON::BI__builtin_neon_vcalt_v: 3159 case NEON::BI__builtin_neon_vcaltq_v: 3160 std::swap(Ops[0], Ops[1]); 3161 case NEON::BI__builtin_neon_vcage_v: 3162 case NEON::BI__builtin_neon_vcageq_v: 3163 case NEON::BI__builtin_neon_vcagt_v: 3164 case NEON::BI__builtin_neon_vcagtq_v: { 3165 llvm::Type *VecFlt = llvm::VectorType::get( 3166 VTy->getScalarSizeInBits() == 32 ? FloatTy : DoubleTy, 3167 VTy->getNumElements()); 3168 llvm::Type *Tys[] = { VTy, VecFlt }; 3169 Function *F = CGM.getIntrinsic(LLVMIntrinsic, Tys); 3170 return EmitNeonCall(F, Ops, NameHint); 3171 } 3172 case NEON::BI__builtin_neon_vclz_v: 3173 case NEON::BI__builtin_neon_vclzq_v: 3174 // We generate target-independent intrinsic, which needs a second argument 3175 // for whether or not clz of zero is undefined; on ARM it isn't. 3176 Ops.push_back(Builder.getInt1(getTarget().isCLZForZeroUndef())); 3177 break; 3178 case NEON::BI__builtin_neon_vcvt_f32_v: 3179 case NEON::BI__builtin_neon_vcvtq_f32_v: 3180 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 3181 Ty = GetNeonType(this, NeonTypeFlags(NeonTypeFlags::Float32, false, Quad)); 3182 return Usgn ? Builder.CreateUIToFP(Ops[0], Ty, "vcvt") 3183 : Builder.CreateSIToFP(Ops[0], Ty, "vcvt"); 3184 case NEON::BI__builtin_neon_vcvt_n_f32_v: 3185 case NEON::BI__builtin_neon_vcvt_n_f64_v: 3186 case NEON::BI__builtin_neon_vcvtq_n_f32_v: 3187 case NEON::BI__builtin_neon_vcvtq_n_f64_v: { 3188 llvm::Type *Tys[2] = { GetFloatNeonType(this, Type), Ty }; 3189 Int = Usgn ? LLVMIntrinsic : AltLLVMIntrinsic; 3190 Function *F = CGM.getIntrinsic(Int, Tys); 3191 return EmitNeonCall(F, Ops, "vcvt_n"); 3192 } 3193 case NEON::BI__builtin_neon_vcvt_n_s32_v: 3194 case NEON::BI__builtin_neon_vcvt_n_u32_v: 3195 case NEON::BI__builtin_neon_vcvt_n_s64_v: 3196 case NEON::BI__builtin_neon_vcvt_n_u64_v: 3197 case NEON::BI__builtin_neon_vcvtq_n_s32_v: 3198 case NEON::BI__builtin_neon_vcvtq_n_u32_v: 3199 case NEON::BI__builtin_neon_vcvtq_n_s64_v: 3200 case NEON::BI__builtin_neon_vcvtq_n_u64_v: { 3201 llvm::Type *Tys[2] = { Ty, GetFloatNeonType(this, Type) }; 3202 Function *F = CGM.getIntrinsic(LLVMIntrinsic, Tys); 3203 return EmitNeonCall(F, Ops, "vcvt_n"); 3204 } 3205 case NEON::BI__builtin_neon_vcvt_s32_v: 3206 case NEON::BI__builtin_neon_vcvt_u32_v: 3207 case NEON::BI__builtin_neon_vcvt_s64_v: 3208 case NEON::BI__builtin_neon_vcvt_u64_v: 3209 case NEON::BI__builtin_neon_vcvtq_s32_v: 3210 case NEON::BI__builtin_neon_vcvtq_u32_v: 3211 case NEON::BI__builtin_neon_vcvtq_s64_v: 3212 case NEON::BI__builtin_neon_vcvtq_u64_v: { 3213 Ops[0] = Builder.CreateBitCast(Ops[0], GetFloatNeonType(this, Type)); 3214 return Usgn ? Builder.CreateFPToUI(Ops[0], Ty, "vcvt") 3215 : Builder.CreateFPToSI(Ops[0], Ty, "vcvt"); 3216 } 3217 case NEON::BI__builtin_neon_vcvta_s32_v: 3218 case NEON::BI__builtin_neon_vcvta_s64_v: 3219 case NEON::BI__builtin_neon_vcvta_u32_v: 3220 case NEON::BI__builtin_neon_vcvta_u64_v: 3221 case NEON::BI__builtin_neon_vcvtaq_s32_v: 3222 case NEON::BI__builtin_neon_vcvtaq_s64_v: 3223 case NEON::BI__builtin_neon_vcvtaq_u32_v: 3224 case NEON::BI__builtin_neon_vcvtaq_u64_v: 3225 case NEON::BI__builtin_neon_vcvtn_s32_v: 3226 case NEON::BI__builtin_neon_vcvtn_s64_v: 3227 case NEON::BI__builtin_neon_vcvtn_u32_v: 3228 case NEON::BI__builtin_neon_vcvtn_u64_v: 3229 case NEON::BI__builtin_neon_vcvtnq_s32_v: 3230 case NEON::BI__builtin_neon_vcvtnq_s64_v: 3231 case NEON::BI__builtin_neon_vcvtnq_u32_v: 3232 case NEON::BI__builtin_neon_vcvtnq_u64_v: 3233 case NEON::BI__builtin_neon_vcvtp_s32_v: 3234 case NEON::BI__builtin_neon_vcvtp_s64_v: 3235 case NEON::BI__builtin_neon_vcvtp_u32_v: 3236 case NEON::BI__builtin_neon_vcvtp_u64_v: 3237 case NEON::BI__builtin_neon_vcvtpq_s32_v: 3238 case NEON::BI__builtin_neon_vcvtpq_s64_v: 3239 case NEON::BI__builtin_neon_vcvtpq_u32_v: 3240 case NEON::BI__builtin_neon_vcvtpq_u64_v: 3241 case NEON::BI__builtin_neon_vcvtm_s32_v: 3242 case NEON::BI__builtin_neon_vcvtm_s64_v: 3243 case NEON::BI__builtin_neon_vcvtm_u32_v: 3244 case NEON::BI__builtin_neon_vcvtm_u64_v: 3245 case NEON::BI__builtin_neon_vcvtmq_s32_v: 3246 case NEON::BI__builtin_neon_vcvtmq_s64_v: 3247 case NEON::BI__builtin_neon_vcvtmq_u32_v: 3248 case NEON::BI__builtin_neon_vcvtmq_u64_v: { 3249 llvm::Type *Tys[2] = { Ty, GetFloatNeonType(this, Type) }; 3250 return EmitNeonCall(CGM.getIntrinsic(LLVMIntrinsic, Tys), Ops, NameHint); 3251 } 3252 case NEON::BI__builtin_neon_vext_v: 3253 case NEON::BI__builtin_neon_vextq_v: { 3254 int CV = cast<ConstantInt>(Ops[2])->getSExtValue(); 3255 SmallVector<Constant*, 16> Indices; 3256 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) 3257 Indices.push_back(ConstantInt::get(Int32Ty, i+CV)); 3258 3259 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 3260 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 3261 Value *SV = llvm::ConstantVector::get(Indices); 3262 return Builder.CreateShuffleVector(Ops[0], Ops[1], SV, "vext"); 3263 } 3264 case NEON::BI__builtin_neon_vfma_v: 3265 case NEON::BI__builtin_neon_vfmaq_v: { 3266 Value *F = CGM.getIntrinsic(Intrinsic::fma, Ty); 3267 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 3268 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 3269 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 3270 3271 // NEON intrinsic puts accumulator first, unlike the LLVM fma. 3272 return Builder.CreateCall(F, {Ops[1], Ops[2], Ops[0]}); 3273 } 3274 case NEON::BI__builtin_neon_vld1_v: 3275 case NEON::BI__builtin_neon_vld1q_v: { 3276 llvm::Type *Tys[] = {Ty, Int8PtrTy}; 3277 Ops.push_back(getAlignmentValue32(PtrOp0)); 3278 return EmitNeonCall(CGM.getIntrinsic(LLVMIntrinsic, Tys), Ops, "vld1"); 3279 } 3280 case NEON::BI__builtin_neon_vld2_v: 3281 case NEON::BI__builtin_neon_vld2q_v: 3282 case NEON::BI__builtin_neon_vld3_v: 3283 case NEON::BI__builtin_neon_vld3q_v: 3284 case NEON::BI__builtin_neon_vld4_v: 3285 case NEON::BI__builtin_neon_vld4q_v: { 3286 llvm::Type *Tys[] = {Ty, Int8PtrTy}; 3287 Function *F = CGM.getIntrinsic(LLVMIntrinsic, Tys); 3288 Value *Align = getAlignmentValue32(PtrOp1); 3289 Ops[1] = Builder.CreateCall(F, {Ops[1], Align}, NameHint); 3290 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 3291 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 3292 return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]); 3293 } 3294 case NEON::BI__builtin_neon_vld1_dup_v: 3295 case NEON::BI__builtin_neon_vld1q_dup_v: { 3296 Value *V = UndefValue::get(Ty); 3297 Ty = llvm::PointerType::getUnqual(VTy->getElementType()); 3298 PtrOp0 = Builder.CreateBitCast(PtrOp0, Ty); 3299 LoadInst *Ld = Builder.CreateLoad(PtrOp0); 3300 llvm::Constant *CI = ConstantInt::get(SizeTy, 0); 3301 Ops[0] = Builder.CreateInsertElement(V, Ld, CI); 3302 return EmitNeonSplat(Ops[0], CI); 3303 } 3304 case NEON::BI__builtin_neon_vld2_lane_v: 3305 case NEON::BI__builtin_neon_vld2q_lane_v: 3306 case NEON::BI__builtin_neon_vld3_lane_v: 3307 case NEON::BI__builtin_neon_vld3q_lane_v: 3308 case NEON::BI__builtin_neon_vld4_lane_v: 3309 case NEON::BI__builtin_neon_vld4q_lane_v: { 3310 llvm::Type *Tys[] = {Ty, Int8PtrTy}; 3311 Function *F = CGM.getIntrinsic(LLVMIntrinsic, Tys); 3312 for (unsigned I = 2; I < Ops.size() - 1; ++I) 3313 Ops[I] = Builder.CreateBitCast(Ops[I], Ty); 3314 Ops.push_back(getAlignmentValue32(PtrOp1)); 3315 Ops[1] = Builder.CreateCall(F, makeArrayRef(Ops).slice(1), NameHint); 3316 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 3317 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 3318 return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]); 3319 } 3320 case NEON::BI__builtin_neon_vmovl_v: { 3321 llvm::Type *DTy =llvm::VectorType::getTruncatedElementVectorType(VTy); 3322 Ops[0] = Builder.CreateBitCast(Ops[0], DTy); 3323 if (Usgn) 3324 return Builder.CreateZExt(Ops[0], Ty, "vmovl"); 3325 return Builder.CreateSExt(Ops[0], Ty, "vmovl"); 3326 } 3327 case NEON::BI__builtin_neon_vmovn_v: { 3328 llvm::Type *QTy = llvm::VectorType::getExtendedElementVectorType(VTy); 3329 Ops[0] = Builder.CreateBitCast(Ops[0], QTy); 3330 return Builder.CreateTrunc(Ops[0], Ty, "vmovn"); 3331 } 3332 case NEON::BI__builtin_neon_vmull_v: 3333 // FIXME: the integer vmull operations could be emitted in terms of pure 3334 // LLVM IR (2 exts followed by a mul). Unfortunately LLVM has a habit of 3335 // hoisting the exts outside loops. Until global ISel comes along that can 3336 // see through such movement this leads to bad CodeGen. So we need an 3337 // intrinsic for now. 3338 Int = Usgn ? Intrinsic::arm_neon_vmullu : Intrinsic::arm_neon_vmulls; 3339 Int = Type.isPoly() ? (unsigned)Intrinsic::arm_neon_vmullp : Int; 3340 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmull"); 3341 case NEON::BI__builtin_neon_vpadal_v: 3342 case NEON::BI__builtin_neon_vpadalq_v: { 3343 // The source operand type has twice as many elements of half the size. 3344 unsigned EltBits = VTy->getElementType()->getPrimitiveSizeInBits(); 3345 llvm::Type *EltTy = 3346 llvm::IntegerType::get(getLLVMContext(), EltBits / 2); 3347 llvm::Type *NarrowTy = 3348 llvm::VectorType::get(EltTy, VTy->getNumElements() * 2); 3349 llvm::Type *Tys[2] = { Ty, NarrowTy }; 3350 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, NameHint); 3351 } 3352 case NEON::BI__builtin_neon_vpaddl_v: 3353 case NEON::BI__builtin_neon_vpaddlq_v: { 3354 // The source operand type has twice as many elements of half the size. 3355 unsigned EltBits = VTy->getElementType()->getPrimitiveSizeInBits(); 3356 llvm::Type *EltTy = llvm::IntegerType::get(getLLVMContext(), EltBits / 2); 3357 llvm::Type *NarrowTy = 3358 llvm::VectorType::get(EltTy, VTy->getNumElements() * 2); 3359 llvm::Type *Tys[2] = { Ty, NarrowTy }; 3360 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vpaddl"); 3361 } 3362 case NEON::BI__builtin_neon_vqdmlal_v: 3363 case NEON::BI__builtin_neon_vqdmlsl_v: { 3364 SmallVector<Value *, 2> MulOps(Ops.begin() + 1, Ops.end()); 3365 Ops[1] = 3366 EmitNeonCall(CGM.getIntrinsic(LLVMIntrinsic, Ty), MulOps, "vqdmlal"); 3367 Ops.resize(2); 3368 return EmitNeonCall(CGM.getIntrinsic(AltLLVMIntrinsic, Ty), Ops, NameHint); 3369 } 3370 case NEON::BI__builtin_neon_vqshl_n_v: 3371 case NEON::BI__builtin_neon_vqshlq_n_v: 3372 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqshl_n", 3373 1, false); 3374 case NEON::BI__builtin_neon_vqshlu_n_v: 3375 case NEON::BI__builtin_neon_vqshluq_n_v: 3376 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqshlu_n", 3377 1, false); 3378 case NEON::BI__builtin_neon_vrecpe_v: 3379 case NEON::BI__builtin_neon_vrecpeq_v: 3380 case NEON::BI__builtin_neon_vrsqrte_v: 3381 case NEON::BI__builtin_neon_vrsqrteq_v: 3382 Int = Ty->isFPOrFPVectorTy() ? LLVMIntrinsic : AltLLVMIntrinsic; 3383 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, NameHint); 3384 3385 case NEON::BI__builtin_neon_vrshr_n_v: 3386 case NEON::BI__builtin_neon_vrshrq_n_v: 3387 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrshr_n", 3388 1, true); 3389 case NEON::BI__builtin_neon_vshl_n_v: 3390 case NEON::BI__builtin_neon_vshlq_n_v: 3391 Ops[1] = EmitNeonShiftVector(Ops[1], Ty, false); 3392 return Builder.CreateShl(Builder.CreateBitCast(Ops[0],Ty), Ops[1], 3393 "vshl_n"); 3394 case NEON::BI__builtin_neon_vshll_n_v: { 3395 llvm::Type *SrcTy = llvm::VectorType::getTruncatedElementVectorType(VTy); 3396 Ops[0] = Builder.CreateBitCast(Ops[0], SrcTy); 3397 if (Usgn) 3398 Ops[0] = Builder.CreateZExt(Ops[0], VTy); 3399 else 3400 Ops[0] = Builder.CreateSExt(Ops[0], VTy); 3401 Ops[1] = EmitNeonShiftVector(Ops[1], VTy, false); 3402 return Builder.CreateShl(Ops[0], Ops[1], "vshll_n"); 3403 } 3404 case NEON::BI__builtin_neon_vshrn_n_v: { 3405 llvm::Type *SrcTy = llvm::VectorType::getExtendedElementVectorType(VTy); 3406 Ops[0] = Builder.CreateBitCast(Ops[0], SrcTy); 3407 Ops[1] = EmitNeonShiftVector(Ops[1], SrcTy, false); 3408 if (Usgn) 3409 Ops[0] = Builder.CreateLShr(Ops[0], Ops[1]); 3410 else 3411 Ops[0] = Builder.CreateAShr(Ops[0], Ops[1]); 3412 return Builder.CreateTrunc(Ops[0], Ty, "vshrn_n"); 3413 } 3414 case NEON::BI__builtin_neon_vshr_n_v: 3415 case NEON::BI__builtin_neon_vshrq_n_v: 3416 return EmitNeonRShiftImm(Ops[0], Ops[1], Ty, Usgn, "vshr_n"); 3417 case NEON::BI__builtin_neon_vst1_v: 3418 case NEON::BI__builtin_neon_vst1q_v: 3419 case NEON::BI__builtin_neon_vst2_v: 3420 case NEON::BI__builtin_neon_vst2q_v: 3421 case NEON::BI__builtin_neon_vst3_v: 3422 case NEON::BI__builtin_neon_vst3q_v: 3423 case NEON::BI__builtin_neon_vst4_v: 3424 case NEON::BI__builtin_neon_vst4q_v: 3425 case NEON::BI__builtin_neon_vst2_lane_v: 3426 case NEON::BI__builtin_neon_vst2q_lane_v: 3427 case NEON::BI__builtin_neon_vst3_lane_v: 3428 case NEON::BI__builtin_neon_vst3q_lane_v: 3429 case NEON::BI__builtin_neon_vst4_lane_v: 3430 case NEON::BI__builtin_neon_vst4q_lane_v: { 3431 llvm::Type *Tys[] = {Int8PtrTy, Ty}; 3432 Ops.push_back(getAlignmentValue32(PtrOp0)); 3433 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, ""); 3434 } 3435 case NEON::BI__builtin_neon_vsubhn_v: { 3436 llvm::VectorType *SrcTy = 3437 llvm::VectorType::getExtendedElementVectorType(VTy); 3438 3439 // %sum = add <4 x i32> %lhs, %rhs 3440 Ops[0] = Builder.CreateBitCast(Ops[0], SrcTy); 3441 Ops[1] = Builder.CreateBitCast(Ops[1], SrcTy); 3442 Ops[0] = Builder.CreateSub(Ops[0], Ops[1], "vsubhn"); 3443 3444 // %high = lshr <4 x i32> %sum, <i32 16, i32 16, i32 16, i32 16> 3445 Constant *ShiftAmt = 3446 ConstantInt::get(SrcTy, SrcTy->getScalarSizeInBits() / 2); 3447 Ops[0] = Builder.CreateLShr(Ops[0], ShiftAmt, "vsubhn"); 3448 3449 // %res = trunc <4 x i32> %high to <4 x i16> 3450 return Builder.CreateTrunc(Ops[0], VTy, "vsubhn"); 3451 } 3452 case NEON::BI__builtin_neon_vtrn_v: 3453 case NEON::BI__builtin_neon_vtrnq_v: { 3454 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty)); 3455 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 3456 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 3457 Value *SV = nullptr; 3458 3459 for (unsigned vi = 0; vi != 2; ++vi) { 3460 SmallVector<Constant*, 16> Indices; 3461 for (unsigned i = 0, e = VTy->getNumElements(); i != e; i += 2) { 3462 Indices.push_back(Builder.getInt32(i+vi)); 3463 Indices.push_back(Builder.getInt32(i+e+vi)); 3464 } 3465 Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ty, Ops[0], vi); 3466 SV = llvm::ConstantVector::get(Indices); 3467 SV = Builder.CreateShuffleVector(Ops[1], Ops[2], SV, "vtrn"); 3468 SV = Builder.CreateDefaultAlignedStore(SV, Addr); 3469 } 3470 return SV; 3471 } 3472 case NEON::BI__builtin_neon_vtst_v: 3473 case NEON::BI__builtin_neon_vtstq_v: { 3474 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 3475 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 3476 Ops[0] = Builder.CreateAnd(Ops[0], Ops[1]); 3477 Ops[0] = Builder.CreateICmp(ICmpInst::ICMP_NE, Ops[0], 3478 ConstantAggregateZero::get(Ty)); 3479 return Builder.CreateSExt(Ops[0], Ty, "vtst"); 3480 } 3481 case NEON::BI__builtin_neon_vuzp_v: 3482 case NEON::BI__builtin_neon_vuzpq_v: { 3483 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty)); 3484 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 3485 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 3486 Value *SV = nullptr; 3487 3488 for (unsigned vi = 0; vi != 2; ++vi) { 3489 SmallVector<Constant*, 16> Indices; 3490 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) 3491 Indices.push_back(ConstantInt::get(Int32Ty, 2*i+vi)); 3492 3493 Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ty, Ops[0], vi); 3494 SV = llvm::ConstantVector::get(Indices); 3495 SV = Builder.CreateShuffleVector(Ops[1], Ops[2], SV, "vuzp"); 3496 SV = Builder.CreateDefaultAlignedStore(SV, Addr); 3497 } 3498 return SV; 3499 } 3500 case NEON::BI__builtin_neon_vzip_v: 3501 case NEON::BI__builtin_neon_vzipq_v: { 3502 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty)); 3503 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 3504 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 3505 Value *SV = nullptr; 3506 3507 for (unsigned vi = 0; vi != 2; ++vi) { 3508 SmallVector<Constant*, 16> Indices; 3509 for (unsigned i = 0, e = VTy->getNumElements(); i != e; i += 2) { 3510 Indices.push_back(ConstantInt::get(Int32Ty, (i + vi*e) >> 1)); 3511 Indices.push_back(ConstantInt::get(Int32Ty, ((i + vi*e) >> 1)+e)); 3512 } 3513 Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ty, Ops[0], vi); 3514 SV = llvm::ConstantVector::get(Indices); 3515 SV = Builder.CreateShuffleVector(Ops[1], Ops[2], SV, "vzip"); 3516 SV = Builder.CreateDefaultAlignedStore(SV, Addr); 3517 } 3518 return SV; 3519 } 3520 } 3521 3522 assert(Int && "Expected valid intrinsic number"); 3523 3524 // Determine the type(s) of this overloaded AArch64 intrinsic. 3525 Function *F = LookupNeonLLVMIntrinsic(Int, Modifier, Ty, E); 3526 3527 Value *Result = EmitNeonCall(F, Ops, NameHint); 3528 llvm::Type *ResultType = ConvertType(E->getType()); 3529 // AArch64 intrinsic one-element vector type cast to 3530 // scalar type expected by the builtin 3531 return Builder.CreateBitCast(Result, ResultType, NameHint); 3532 } 3533 3534 Value *CodeGenFunction::EmitAArch64CompareBuiltinExpr( 3535 Value *Op, llvm::Type *Ty, const CmpInst::Predicate Fp, 3536 const CmpInst::Predicate Ip, const Twine &Name) { 3537 llvm::Type *OTy = Op->getType(); 3538 3539 // FIXME: this is utterly horrific. We should not be looking at previous 3540 // codegen context to find out what needs doing. Unfortunately TableGen 3541 // currently gives us exactly the same calls for vceqz_f32 and vceqz_s32 3542 // (etc). 3543 if (BitCastInst *BI = dyn_cast<BitCastInst>(Op)) 3544 OTy = BI->getOperand(0)->getType(); 3545 3546 Op = Builder.CreateBitCast(Op, OTy); 3547 if (OTy->getScalarType()->isFloatingPointTy()) { 3548 Op = Builder.CreateFCmp(Fp, Op, Constant::getNullValue(OTy)); 3549 } else { 3550 Op = Builder.CreateICmp(Ip, Op, Constant::getNullValue(OTy)); 3551 } 3552 return Builder.CreateSExt(Op, Ty, Name); 3553 } 3554 3555 static Value *packTBLDVectorList(CodeGenFunction &CGF, ArrayRef<Value *> Ops, 3556 Value *ExtOp, Value *IndexOp, 3557 llvm::Type *ResTy, unsigned IntID, 3558 const char *Name) { 3559 SmallVector<Value *, 2> TblOps; 3560 if (ExtOp) 3561 TblOps.push_back(ExtOp); 3562 3563 // Build a vector containing sequential number like (0, 1, 2, ..., 15) 3564 SmallVector<Constant*, 16> Indices; 3565 llvm::VectorType *TblTy = cast<llvm::VectorType>(Ops[0]->getType()); 3566 for (unsigned i = 0, e = TblTy->getNumElements(); i != e; ++i) { 3567 Indices.push_back(ConstantInt::get(CGF.Int32Ty, 2*i)); 3568 Indices.push_back(ConstantInt::get(CGF.Int32Ty, 2*i+1)); 3569 } 3570 Value *SV = llvm::ConstantVector::get(Indices); 3571 3572 int PairPos = 0, End = Ops.size() - 1; 3573 while (PairPos < End) { 3574 TblOps.push_back(CGF.Builder.CreateShuffleVector(Ops[PairPos], 3575 Ops[PairPos+1], SV, Name)); 3576 PairPos += 2; 3577 } 3578 3579 // If there's an odd number of 64-bit lookup table, fill the high 64-bit 3580 // of the 128-bit lookup table with zero. 3581 if (PairPos == End) { 3582 Value *ZeroTbl = ConstantAggregateZero::get(TblTy); 3583 TblOps.push_back(CGF.Builder.CreateShuffleVector(Ops[PairPos], 3584 ZeroTbl, SV, Name)); 3585 } 3586 3587 Function *TblF; 3588 TblOps.push_back(IndexOp); 3589 TblF = CGF.CGM.getIntrinsic(IntID, ResTy); 3590 3591 return CGF.EmitNeonCall(TblF, TblOps, Name); 3592 } 3593 3594 Value *CodeGenFunction::GetValueForARMHint(unsigned BuiltinID) { 3595 unsigned Value; 3596 switch (BuiltinID) { 3597 default: 3598 return nullptr; 3599 case ARM::BI__builtin_arm_nop: 3600 Value = 0; 3601 break; 3602 case ARM::BI__builtin_arm_yield: 3603 case ARM::BI__yield: 3604 Value = 1; 3605 break; 3606 case ARM::BI__builtin_arm_wfe: 3607 case ARM::BI__wfe: 3608 Value = 2; 3609 break; 3610 case ARM::BI__builtin_arm_wfi: 3611 case ARM::BI__wfi: 3612 Value = 3; 3613 break; 3614 case ARM::BI__builtin_arm_sev: 3615 case ARM::BI__sev: 3616 Value = 4; 3617 break; 3618 case ARM::BI__builtin_arm_sevl: 3619 case ARM::BI__sevl: 3620 Value = 5; 3621 break; 3622 } 3623 3624 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::arm_hint), 3625 llvm::ConstantInt::get(Int32Ty, Value)); 3626 } 3627 3628 // Generates the IR for the read/write special register builtin, 3629 // ValueType is the type of the value that is to be written or read, 3630 // RegisterType is the type of the register being written to or read from. 3631 static Value *EmitSpecialRegisterBuiltin(CodeGenFunction &CGF, 3632 const CallExpr *E, 3633 llvm::Type *RegisterType, 3634 llvm::Type *ValueType, bool IsRead) { 3635 // write and register intrinsics only support 32 and 64 bit operations. 3636 assert((RegisterType->isIntegerTy(32) || RegisterType->isIntegerTy(64)) 3637 && "Unsupported size for register."); 3638 3639 CodeGen::CGBuilderTy &Builder = CGF.Builder; 3640 CodeGen::CodeGenModule &CGM = CGF.CGM; 3641 LLVMContext &Context = CGM.getLLVMContext(); 3642 3643 const Expr *SysRegStrExpr = E->getArg(0)->IgnoreParenCasts(); 3644 StringRef SysReg = cast<StringLiteral>(SysRegStrExpr)->getString(); 3645 3646 llvm::Metadata *Ops[] = { llvm::MDString::get(Context, SysReg) }; 3647 llvm::MDNode *RegName = llvm::MDNode::get(Context, Ops); 3648 llvm::Value *Metadata = llvm::MetadataAsValue::get(Context, RegName); 3649 3650 llvm::Type *Types[] = { RegisterType }; 3651 3652 bool MixedTypes = RegisterType->isIntegerTy(64) && ValueType->isIntegerTy(32); 3653 assert(!(RegisterType->isIntegerTy(32) && ValueType->isIntegerTy(64)) 3654 && "Can't fit 64-bit value in 32-bit register"); 3655 3656 if (IsRead) { 3657 llvm::Value *F = CGM.getIntrinsic(llvm::Intrinsic::read_register, Types); 3658 llvm::Value *Call = Builder.CreateCall(F, Metadata); 3659 3660 if (MixedTypes) 3661 // Read into 64 bit register and then truncate result to 32 bit. 3662 return Builder.CreateTrunc(Call, ValueType); 3663 3664 if (ValueType->isPointerTy()) 3665 // Have i32/i64 result (Call) but want to return a VoidPtrTy (i8*). 3666 return Builder.CreateIntToPtr(Call, ValueType); 3667 3668 return Call; 3669 } 3670 3671 llvm::Value *F = CGM.getIntrinsic(llvm::Intrinsic::write_register, Types); 3672 llvm::Value *ArgValue = CGF.EmitScalarExpr(E->getArg(1)); 3673 if (MixedTypes) { 3674 // Extend 32 bit write value to 64 bit to pass to write. 3675 ArgValue = Builder.CreateZExt(ArgValue, RegisterType); 3676 return Builder.CreateCall(F, { Metadata, ArgValue }); 3677 } 3678 3679 if (ValueType->isPointerTy()) { 3680 // Have VoidPtrTy ArgValue but want to return an i32/i64. 3681 ArgValue = Builder.CreatePtrToInt(ArgValue, RegisterType); 3682 return Builder.CreateCall(F, { Metadata, ArgValue }); 3683 } 3684 3685 return Builder.CreateCall(F, { Metadata, ArgValue }); 3686 } 3687 3688 /// Return true if BuiltinID is an overloaded Neon intrinsic with an extra 3689 /// argument that specifies the vector type. 3690 static bool HasExtraNeonArgument(unsigned BuiltinID) { 3691 switch (BuiltinID) { 3692 default: break; 3693 case NEON::BI__builtin_neon_vget_lane_i8: 3694 case NEON::BI__builtin_neon_vget_lane_i16: 3695 case NEON::BI__builtin_neon_vget_lane_i32: 3696 case NEON::BI__builtin_neon_vget_lane_i64: 3697 case NEON::BI__builtin_neon_vget_lane_f32: 3698 case NEON::BI__builtin_neon_vgetq_lane_i8: 3699 case NEON::BI__builtin_neon_vgetq_lane_i16: 3700 case NEON::BI__builtin_neon_vgetq_lane_i32: 3701 case NEON::BI__builtin_neon_vgetq_lane_i64: 3702 case NEON::BI__builtin_neon_vgetq_lane_f32: 3703 case NEON::BI__builtin_neon_vset_lane_i8: 3704 case NEON::BI__builtin_neon_vset_lane_i16: 3705 case NEON::BI__builtin_neon_vset_lane_i32: 3706 case NEON::BI__builtin_neon_vset_lane_i64: 3707 case NEON::BI__builtin_neon_vset_lane_f32: 3708 case NEON::BI__builtin_neon_vsetq_lane_i8: 3709 case NEON::BI__builtin_neon_vsetq_lane_i16: 3710 case NEON::BI__builtin_neon_vsetq_lane_i32: 3711 case NEON::BI__builtin_neon_vsetq_lane_i64: 3712 case NEON::BI__builtin_neon_vsetq_lane_f32: 3713 case NEON::BI__builtin_neon_vsha1h_u32: 3714 case NEON::BI__builtin_neon_vsha1cq_u32: 3715 case NEON::BI__builtin_neon_vsha1pq_u32: 3716 case NEON::BI__builtin_neon_vsha1mq_u32: 3717 case ARM::BI_MoveToCoprocessor: 3718 case ARM::BI_MoveToCoprocessor2: 3719 return false; 3720 } 3721 return true; 3722 } 3723 3724 Value *CodeGenFunction::EmitARMBuiltinExpr(unsigned BuiltinID, 3725 const CallExpr *E) { 3726 if (auto Hint = GetValueForARMHint(BuiltinID)) 3727 return Hint; 3728 3729 if (BuiltinID == ARM::BI__emit) { 3730 bool IsThumb = getTarget().getTriple().getArch() == llvm::Triple::thumb; 3731 llvm::FunctionType *FTy = 3732 llvm::FunctionType::get(VoidTy, /*Variadic=*/false); 3733 3734 APSInt Value; 3735 if (!E->getArg(0)->EvaluateAsInt(Value, CGM.getContext())) 3736 llvm_unreachable("Sema will ensure that the parameter is constant"); 3737 3738 uint64_t ZExtValue = Value.zextOrTrunc(IsThumb ? 16 : 32).getZExtValue(); 3739 3740 llvm::InlineAsm *Emit = 3741 IsThumb ? InlineAsm::get(FTy, ".inst.n 0x" + utohexstr(ZExtValue), "", 3742 /*SideEffects=*/true) 3743 : InlineAsm::get(FTy, ".inst 0x" + utohexstr(ZExtValue), "", 3744 /*SideEffects=*/true); 3745 3746 return Builder.CreateCall(Emit); 3747 } 3748 3749 if (BuiltinID == ARM::BI__builtin_arm_dbg) { 3750 Value *Option = EmitScalarExpr(E->getArg(0)); 3751 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::arm_dbg), Option); 3752 } 3753 3754 if (BuiltinID == ARM::BI__builtin_arm_prefetch) { 3755 Value *Address = EmitScalarExpr(E->getArg(0)); 3756 Value *RW = EmitScalarExpr(E->getArg(1)); 3757 Value *IsData = EmitScalarExpr(E->getArg(2)); 3758 3759 // Locality is not supported on ARM target 3760 Value *Locality = llvm::ConstantInt::get(Int32Ty, 3); 3761 3762 Value *F = CGM.getIntrinsic(Intrinsic::prefetch); 3763 return Builder.CreateCall(F, {Address, RW, Locality, IsData}); 3764 } 3765 3766 if (BuiltinID == ARM::BI__builtin_arm_rbit) { 3767 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::arm_rbit), 3768 EmitScalarExpr(E->getArg(0)), 3769 "rbit"); 3770 } 3771 3772 if (BuiltinID == ARM::BI__clear_cache) { 3773 assert(E->getNumArgs() == 2 && "__clear_cache takes 2 arguments"); 3774 const FunctionDecl *FD = E->getDirectCallee(); 3775 Value *Ops[2]; 3776 for (unsigned i = 0; i < 2; i++) 3777 Ops[i] = EmitScalarExpr(E->getArg(i)); 3778 llvm::Type *Ty = CGM.getTypes().ConvertType(FD->getType()); 3779 llvm::FunctionType *FTy = cast<llvm::FunctionType>(Ty); 3780 StringRef Name = FD->getName(); 3781 return EmitNounwindRuntimeCall(CGM.CreateRuntimeFunction(FTy, Name), Ops); 3782 } 3783 3784 if (BuiltinID == ARM::BI__builtin_arm_ldrexd || 3785 ((BuiltinID == ARM::BI__builtin_arm_ldrex || 3786 BuiltinID == ARM::BI__builtin_arm_ldaex) && 3787 getContext().getTypeSize(E->getType()) == 64) || 3788 BuiltinID == ARM::BI__ldrexd) { 3789 Function *F; 3790 3791 switch (BuiltinID) { 3792 default: llvm_unreachable("unexpected builtin"); 3793 case ARM::BI__builtin_arm_ldaex: 3794 F = CGM.getIntrinsic(Intrinsic::arm_ldaexd); 3795 break; 3796 case ARM::BI__builtin_arm_ldrexd: 3797 case ARM::BI__builtin_arm_ldrex: 3798 case ARM::BI__ldrexd: 3799 F = CGM.getIntrinsic(Intrinsic::arm_ldrexd); 3800 break; 3801 } 3802 3803 Value *LdPtr = EmitScalarExpr(E->getArg(0)); 3804 Value *Val = Builder.CreateCall(F, Builder.CreateBitCast(LdPtr, Int8PtrTy), 3805 "ldrexd"); 3806 3807 Value *Val0 = Builder.CreateExtractValue(Val, 1); 3808 Value *Val1 = Builder.CreateExtractValue(Val, 0); 3809 Val0 = Builder.CreateZExt(Val0, Int64Ty); 3810 Val1 = Builder.CreateZExt(Val1, Int64Ty); 3811 3812 Value *ShiftCst = llvm::ConstantInt::get(Int64Ty, 32); 3813 Val = Builder.CreateShl(Val0, ShiftCst, "shl", true /* nuw */); 3814 Val = Builder.CreateOr(Val, Val1); 3815 return Builder.CreateBitCast(Val, ConvertType(E->getType())); 3816 } 3817 3818 if (BuiltinID == ARM::BI__builtin_arm_ldrex || 3819 BuiltinID == ARM::BI__builtin_arm_ldaex) { 3820 Value *LoadAddr = EmitScalarExpr(E->getArg(0)); 3821 3822 QualType Ty = E->getType(); 3823 llvm::Type *RealResTy = ConvertType(Ty); 3824 llvm::Type *IntResTy = llvm::IntegerType::get(getLLVMContext(), 3825 getContext().getTypeSize(Ty)); 3826 LoadAddr = Builder.CreateBitCast(LoadAddr, IntResTy->getPointerTo()); 3827 3828 Function *F = CGM.getIntrinsic(BuiltinID == ARM::BI__builtin_arm_ldaex 3829 ? Intrinsic::arm_ldaex 3830 : Intrinsic::arm_ldrex, 3831 LoadAddr->getType()); 3832 Value *Val = Builder.CreateCall(F, LoadAddr, "ldrex"); 3833 3834 if (RealResTy->isPointerTy()) 3835 return Builder.CreateIntToPtr(Val, RealResTy); 3836 else { 3837 Val = Builder.CreateTruncOrBitCast(Val, IntResTy); 3838 return Builder.CreateBitCast(Val, RealResTy); 3839 } 3840 } 3841 3842 if (BuiltinID == ARM::BI__builtin_arm_strexd || 3843 ((BuiltinID == ARM::BI__builtin_arm_stlex || 3844 BuiltinID == ARM::BI__builtin_arm_strex) && 3845 getContext().getTypeSize(E->getArg(0)->getType()) == 64)) { 3846 Function *F = CGM.getIntrinsic(BuiltinID == ARM::BI__builtin_arm_stlex 3847 ? Intrinsic::arm_stlexd 3848 : Intrinsic::arm_strexd); 3849 llvm::Type *STy = llvm::StructType::get(Int32Ty, Int32Ty, nullptr); 3850 3851 Address Tmp = CreateMemTemp(E->getArg(0)->getType()); 3852 Value *Val = EmitScalarExpr(E->getArg(0)); 3853 Builder.CreateStore(Val, Tmp); 3854 3855 Address LdPtr = Builder.CreateBitCast(Tmp,llvm::PointerType::getUnqual(STy)); 3856 Val = Builder.CreateLoad(LdPtr); 3857 3858 Value *Arg0 = Builder.CreateExtractValue(Val, 0); 3859 Value *Arg1 = Builder.CreateExtractValue(Val, 1); 3860 Value *StPtr = Builder.CreateBitCast(EmitScalarExpr(E->getArg(1)), Int8PtrTy); 3861 return Builder.CreateCall(F, {Arg0, Arg1, StPtr}, "strexd"); 3862 } 3863 3864 if (BuiltinID == ARM::BI__builtin_arm_strex || 3865 BuiltinID == ARM::BI__builtin_arm_stlex) { 3866 Value *StoreVal = EmitScalarExpr(E->getArg(0)); 3867 Value *StoreAddr = EmitScalarExpr(E->getArg(1)); 3868 3869 QualType Ty = E->getArg(0)->getType(); 3870 llvm::Type *StoreTy = llvm::IntegerType::get(getLLVMContext(), 3871 getContext().getTypeSize(Ty)); 3872 StoreAddr = Builder.CreateBitCast(StoreAddr, StoreTy->getPointerTo()); 3873 3874 if (StoreVal->getType()->isPointerTy()) 3875 StoreVal = Builder.CreatePtrToInt(StoreVal, Int32Ty); 3876 else { 3877 StoreVal = Builder.CreateBitCast(StoreVal, StoreTy); 3878 StoreVal = Builder.CreateZExtOrBitCast(StoreVal, Int32Ty); 3879 } 3880 3881 Function *F = CGM.getIntrinsic(BuiltinID == ARM::BI__builtin_arm_stlex 3882 ? Intrinsic::arm_stlex 3883 : Intrinsic::arm_strex, 3884 StoreAddr->getType()); 3885 return Builder.CreateCall(F, {StoreVal, StoreAddr}, "strex"); 3886 } 3887 3888 if (BuiltinID == ARM::BI__builtin_arm_clrex) { 3889 Function *F = CGM.getIntrinsic(Intrinsic::arm_clrex); 3890 return Builder.CreateCall(F); 3891 } 3892 3893 // CRC32 3894 Intrinsic::ID CRCIntrinsicID = Intrinsic::not_intrinsic; 3895 switch (BuiltinID) { 3896 case ARM::BI__builtin_arm_crc32b: 3897 CRCIntrinsicID = Intrinsic::arm_crc32b; break; 3898 case ARM::BI__builtin_arm_crc32cb: 3899 CRCIntrinsicID = Intrinsic::arm_crc32cb; break; 3900 case ARM::BI__builtin_arm_crc32h: 3901 CRCIntrinsicID = Intrinsic::arm_crc32h; break; 3902 case ARM::BI__builtin_arm_crc32ch: 3903 CRCIntrinsicID = Intrinsic::arm_crc32ch; break; 3904 case ARM::BI__builtin_arm_crc32w: 3905 case ARM::BI__builtin_arm_crc32d: 3906 CRCIntrinsicID = Intrinsic::arm_crc32w; break; 3907 case ARM::BI__builtin_arm_crc32cw: 3908 case ARM::BI__builtin_arm_crc32cd: 3909 CRCIntrinsicID = Intrinsic::arm_crc32cw; break; 3910 } 3911 3912 if (CRCIntrinsicID != Intrinsic::not_intrinsic) { 3913 Value *Arg0 = EmitScalarExpr(E->getArg(0)); 3914 Value *Arg1 = EmitScalarExpr(E->getArg(1)); 3915 3916 // crc32{c,}d intrinsics are implemnted as two calls to crc32{c,}w 3917 // intrinsics, hence we need different codegen for these cases. 3918 if (BuiltinID == ARM::BI__builtin_arm_crc32d || 3919 BuiltinID == ARM::BI__builtin_arm_crc32cd) { 3920 Value *C1 = llvm::ConstantInt::get(Int64Ty, 32); 3921 Value *Arg1a = Builder.CreateTruncOrBitCast(Arg1, Int32Ty); 3922 Value *Arg1b = Builder.CreateLShr(Arg1, C1); 3923 Arg1b = Builder.CreateTruncOrBitCast(Arg1b, Int32Ty); 3924 3925 Function *F = CGM.getIntrinsic(CRCIntrinsicID); 3926 Value *Res = Builder.CreateCall(F, {Arg0, Arg1a}); 3927 return Builder.CreateCall(F, {Res, Arg1b}); 3928 } else { 3929 Arg1 = Builder.CreateZExtOrBitCast(Arg1, Int32Ty); 3930 3931 Function *F = CGM.getIntrinsic(CRCIntrinsicID); 3932 return Builder.CreateCall(F, {Arg0, Arg1}); 3933 } 3934 } 3935 3936 if (BuiltinID == ARM::BI__builtin_arm_rsr || 3937 BuiltinID == ARM::BI__builtin_arm_rsr64 || 3938 BuiltinID == ARM::BI__builtin_arm_rsrp || 3939 BuiltinID == ARM::BI__builtin_arm_wsr || 3940 BuiltinID == ARM::BI__builtin_arm_wsr64 || 3941 BuiltinID == ARM::BI__builtin_arm_wsrp) { 3942 3943 bool IsRead = BuiltinID == ARM::BI__builtin_arm_rsr || 3944 BuiltinID == ARM::BI__builtin_arm_rsr64 || 3945 BuiltinID == ARM::BI__builtin_arm_rsrp; 3946 3947 bool IsPointerBuiltin = BuiltinID == ARM::BI__builtin_arm_rsrp || 3948 BuiltinID == ARM::BI__builtin_arm_wsrp; 3949 3950 bool Is64Bit = BuiltinID == ARM::BI__builtin_arm_rsr64 || 3951 BuiltinID == ARM::BI__builtin_arm_wsr64; 3952 3953 llvm::Type *ValueType; 3954 llvm::Type *RegisterType; 3955 if (IsPointerBuiltin) { 3956 ValueType = VoidPtrTy; 3957 RegisterType = Int32Ty; 3958 } else if (Is64Bit) { 3959 ValueType = RegisterType = Int64Ty; 3960 } else { 3961 ValueType = RegisterType = Int32Ty; 3962 } 3963 3964 return EmitSpecialRegisterBuiltin(*this, E, RegisterType, ValueType, IsRead); 3965 } 3966 3967 // Find out if any arguments are required to be integer constant 3968 // expressions. 3969 unsigned ICEArguments = 0; 3970 ASTContext::GetBuiltinTypeError Error; 3971 getContext().GetBuiltinType(BuiltinID, Error, &ICEArguments); 3972 assert(Error == ASTContext::GE_None && "Should not codegen an error"); 3973 3974 auto getAlignmentValue32 = [&](Address addr) -> Value* { 3975 return Builder.getInt32(addr.getAlignment().getQuantity()); 3976 }; 3977 3978 Address PtrOp0 = Address::invalid(); 3979 Address PtrOp1 = Address::invalid(); 3980 SmallVector<Value*, 4> Ops; 3981 bool HasExtraArg = HasExtraNeonArgument(BuiltinID); 3982 unsigned NumArgs = E->getNumArgs() - (HasExtraArg ? 1 : 0); 3983 for (unsigned i = 0, e = NumArgs; i != e; i++) { 3984 if (i == 0) { 3985 switch (BuiltinID) { 3986 case NEON::BI__builtin_neon_vld1_v: 3987 case NEON::BI__builtin_neon_vld1q_v: 3988 case NEON::BI__builtin_neon_vld1q_lane_v: 3989 case NEON::BI__builtin_neon_vld1_lane_v: 3990 case NEON::BI__builtin_neon_vld1_dup_v: 3991 case NEON::BI__builtin_neon_vld1q_dup_v: 3992 case NEON::BI__builtin_neon_vst1_v: 3993 case NEON::BI__builtin_neon_vst1q_v: 3994 case NEON::BI__builtin_neon_vst1q_lane_v: 3995 case NEON::BI__builtin_neon_vst1_lane_v: 3996 case NEON::BI__builtin_neon_vst2_v: 3997 case NEON::BI__builtin_neon_vst2q_v: 3998 case NEON::BI__builtin_neon_vst2_lane_v: 3999 case NEON::BI__builtin_neon_vst2q_lane_v: 4000 case NEON::BI__builtin_neon_vst3_v: 4001 case NEON::BI__builtin_neon_vst3q_v: 4002 case NEON::BI__builtin_neon_vst3_lane_v: 4003 case NEON::BI__builtin_neon_vst3q_lane_v: 4004 case NEON::BI__builtin_neon_vst4_v: 4005 case NEON::BI__builtin_neon_vst4q_v: 4006 case NEON::BI__builtin_neon_vst4_lane_v: 4007 case NEON::BI__builtin_neon_vst4q_lane_v: 4008 // Get the alignment for the argument in addition to the value; 4009 // we'll use it later. 4010 PtrOp0 = EmitPointerWithAlignment(E->getArg(0)); 4011 Ops.push_back(PtrOp0.getPointer()); 4012 continue; 4013 } 4014 } 4015 if (i == 1) { 4016 switch (BuiltinID) { 4017 case NEON::BI__builtin_neon_vld2_v: 4018 case NEON::BI__builtin_neon_vld2q_v: 4019 case NEON::BI__builtin_neon_vld3_v: 4020 case NEON::BI__builtin_neon_vld3q_v: 4021 case NEON::BI__builtin_neon_vld4_v: 4022 case NEON::BI__builtin_neon_vld4q_v: 4023 case NEON::BI__builtin_neon_vld2_lane_v: 4024 case NEON::BI__builtin_neon_vld2q_lane_v: 4025 case NEON::BI__builtin_neon_vld3_lane_v: 4026 case NEON::BI__builtin_neon_vld3q_lane_v: 4027 case NEON::BI__builtin_neon_vld4_lane_v: 4028 case NEON::BI__builtin_neon_vld4q_lane_v: 4029 case NEON::BI__builtin_neon_vld2_dup_v: 4030 case NEON::BI__builtin_neon_vld3_dup_v: 4031 case NEON::BI__builtin_neon_vld4_dup_v: 4032 // Get the alignment for the argument in addition to the value; 4033 // we'll use it later. 4034 PtrOp1 = EmitPointerWithAlignment(E->getArg(1)); 4035 Ops.push_back(PtrOp1.getPointer()); 4036 continue; 4037 } 4038 } 4039 4040 if ((ICEArguments & (1 << i)) == 0) { 4041 Ops.push_back(EmitScalarExpr(E->getArg(i))); 4042 } else { 4043 // If this is required to be a constant, constant fold it so that we know 4044 // that the generated intrinsic gets a ConstantInt. 4045 llvm::APSInt Result; 4046 bool IsConst = E->getArg(i)->isIntegerConstantExpr(Result, getContext()); 4047 assert(IsConst && "Constant arg isn't actually constant?"); (void)IsConst; 4048 Ops.push_back(llvm::ConstantInt::get(getLLVMContext(), Result)); 4049 } 4050 } 4051 4052 switch (BuiltinID) { 4053 default: break; 4054 4055 case NEON::BI__builtin_neon_vget_lane_i8: 4056 case NEON::BI__builtin_neon_vget_lane_i16: 4057 case NEON::BI__builtin_neon_vget_lane_i32: 4058 case NEON::BI__builtin_neon_vget_lane_i64: 4059 case NEON::BI__builtin_neon_vget_lane_f32: 4060 case NEON::BI__builtin_neon_vgetq_lane_i8: 4061 case NEON::BI__builtin_neon_vgetq_lane_i16: 4062 case NEON::BI__builtin_neon_vgetq_lane_i32: 4063 case NEON::BI__builtin_neon_vgetq_lane_i64: 4064 case NEON::BI__builtin_neon_vgetq_lane_f32: 4065 return Builder.CreateExtractElement(Ops[0], Ops[1], "vget_lane"); 4066 4067 case NEON::BI__builtin_neon_vset_lane_i8: 4068 case NEON::BI__builtin_neon_vset_lane_i16: 4069 case NEON::BI__builtin_neon_vset_lane_i32: 4070 case NEON::BI__builtin_neon_vset_lane_i64: 4071 case NEON::BI__builtin_neon_vset_lane_f32: 4072 case NEON::BI__builtin_neon_vsetq_lane_i8: 4073 case NEON::BI__builtin_neon_vsetq_lane_i16: 4074 case NEON::BI__builtin_neon_vsetq_lane_i32: 4075 case NEON::BI__builtin_neon_vsetq_lane_i64: 4076 case NEON::BI__builtin_neon_vsetq_lane_f32: 4077 return Builder.CreateInsertElement(Ops[1], Ops[0], Ops[2], "vset_lane"); 4078 4079 case NEON::BI__builtin_neon_vsha1h_u32: 4080 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_sha1h), Ops, 4081 "vsha1h"); 4082 case NEON::BI__builtin_neon_vsha1cq_u32: 4083 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_sha1c), Ops, 4084 "vsha1h"); 4085 case NEON::BI__builtin_neon_vsha1pq_u32: 4086 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_sha1p), Ops, 4087 "vsha1h"); 4088 case NEON::BI__builtin_neon_vsha1mq_u32: 4089 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_sha1m), Ops, 4090 "vsha1h"); 4091 4092 // The ARM _MoveToCoprocessor builtins put the input register value as 4093 // the first argument, but the LLVM intrinsic expects it as the third one. 4094 case ARM::BI_MoveToCoprocessor: 4095 case ARM::BI_MoveToCoprocessor2: { 4096 Function *F = CGM.getIntrinsic(BuiltinID == ARM::BI_MoveToCoprocessor ? 4097 Intrinsic::arm_mcr : Intrinsic::arm_mcr2); 4098 return Builder.CreateCall(F, {Ops[1], Ops[2], Ops[0], 4099 Ops[3], Ops[4], Ops[5]}); 4100 } 4101 } 4102 4103 // Get the last argument, which specifies the vector type. 4104 assert(HasExtraArg); 4105 llvm::APSInt Result; 4106 const Expr *Arg = E->getArg(E->getNumArgs()-1); 4107 if (!Arg->isIntegerConstantExpr(Result, getContext())) 4108 return nullptr; 4109 4110 if (BuiltinID == ARM::BI__builtin_arm_vcvtr_f || 4111 BuiltinID == ARM::BI__builtin_arm_vcvtr_d) { 4112 // Determine the overloaded type of this builtin. 4113 llvm::Type *Ty; 4114 if (BuiltinID == ARM::BI__builtin_arm_vcvtr_f) 4115 Ty = FloatTy; 4116 else 4117 Ty = DoubleTy; 4118 4119 // Determine whether this is an unsigned conversion or not. 4120 bool usgn = Result.getZExtValue() == 1; 4121 unsigned Int = usgn ? Intrinsic::arm_vcvtru : Intrinsic::arm_vcvtr; 4122 4123 // Call the appropriate intrinsic. 4124 Function *F = CGM.getIntrinsic(Int, Ty); 4125 return Builder.CreateCall(F, Ops, "vcvtr"); 4126 } 4127 4128 // Determine the type of this overloaded NEON intrinsic. 4129 NeonTypeFlags Type(Result.getZExtValue()); 4130 bool usgn = Type.isUnsigned(); 4131 bool rightShift = false; 4132 4133 llvm::VectorType *VTy = GetNeonType(this, Type); 4134 llvm::Type *Ty = VTy; 4135 if (!Ty) 4136 return nullptr; 4137 4138 // Many NEON builtins have identical semantics and uses in ARM and 4139 // AArch64. Emit these in a single function. 4140 auto IntrinsicMap = makeArrayRef(ARMSIMDIntrinsicMap); 4141 const NeonIntrinsicInfo *Builtin = findNeonIntrinsicInMap( 4142 IntrinsicMap, BuiltinID, NEONSIMDIntrinsicsProvenSorted); 4143 if (Builtin) 4144 return EmitCommonNeonBuiltinExpr( 4145 Builtin->BuiltinID, Builtin->LLVMIntrinsic, Builtin->AltLLVMIntrinsic, 4146 Builtin->NameHint, Builtin->TypeModifier, E, Ops, PtrOp0, PtrOp1); 4147 4148 unsigned Int; 4149 switch (BuiltinID) { 4150 default: return nullptr; 4151 case NEON::BI__builtin_neon_vld1q_lane_v: 4152 // Handle 64-bit integer elements as a special case. Use shuffles of 4153 // one-element vectors to avoid poor code for i64 in the backend. 4154 if (VTy->getElementType()->isIntegerTy(64)) { 4155 // Extract the other lane. 4156 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 4157 uint32_t Lane = cast<ConstantInt>(Ops[2])->getZExtValue(); 4158 Value *SV = llvm::ConstantVector::get(ConstantInt::get(Int32Ty, 1-Lane)); 4159 Ops[1] = Builder.CreateShuffleVector(Ops[1], Ops[1], SV); 4160 // Load the value as a one-element vector. 4161 Ty = llvm::VectorType::get(VTy->getElementType(), 1); 4162 llvm::Type *Tys[] = {Ty, Int8PtrTy}; 4163 Function *F = CGM.getIntrinsic(Intrinsic::arm_neon_vld1, Tys); 4164 Value *Align = getAlignmentValue32(PtrOp0); 4165 Value *Ld = Builder.CreateCall(F, {Ops[0], Align}); 4166 // Combine them. 4167 uint32_t Indices[] = {1 - Lane, Lane}; 4168 SV = llvm::ConstantDataVector::get(getLLVMContext(), Indices); 4169 return Builder.CreateShuffleVector(Ops[1], Ld, SV, "vld1q_lane"); 4170 } 4171 // fall through 4172 case NEON::BI__builtin_neon_vld1_lane_v: { 4173 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 4174 PtrOp0 = Builder.CreateElementBitCast(PtrOp0, VTy->getElementType()); 4175 Value *Ld = Builder.CreateLoad(PtrOp0); 4176 return Builder.CreateInsertElement(Ops[1], Ld, Ops[2], "vld1_lane"); 4177 } 4178 case NEON::BI__builtin_neon_vld2_dup_v: 4179 case NEON::BI__builtin_neon_vld3_dup_v: 4180 case NEON::BI__builtin_neon_vld4_dup_v: { 4181 // Handle 64-bit elements as a special-case. There is no "dup" needed. 4182 if (VTy->getElementType()->getPrimitiveSizeInBits() == 64) { 4183 switch (BuiltinID) { 4184 case NEON::BI__builtin_neon_vld2_dup_v: 4185 Int = Intrinsic::arm_neon_vld2; 4186 break; 4187 case NEON::BI__builtin_neon_vld3_dup_v: 4188 Int = Intrinsic::arm_neon_vld3; 4189 break; 4190 case NEON::BI__builtin_neon_vld4_dup_v: 4191 Int = Intrinsic::arm_neon_vld4; 4192 break; 4193 default: llvm_unreachable("unknown vld_dup intrinsic?"); 4194 } 4195 llvm::Type *Tys[] = {Ty, Int8PtrTy}; 4196 Function *F = CGM.getIntrinsic(Int, Tys); 4197 llvm::Value *Align = getAlignmentValue32(PtrOp1); 4198 Ops[1] = Builder.CreateCall(F, {Ops[1], Align}, "vld_dup"); 4199 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 4200 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 4201 return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]); 4202 } 4203 switch (BuiltinID) { 4204 case NEON::BI__builtin_neon_vld2_dup_v: 4205 Int = Intrinsic::arm_neon_vld2lane; 4206 break; 4207 case NEON::BI__builtin_neon_vld3_dup_v: 4208 Int = Intrinsic::arm_neon_vld3lane; 4209 break; 4210 case NEON::BI__builtin_neon_vld4_dup_v: 4211 Int = Intrinsic::arm_neon_vld4lane; 4212 break; 4213 default: llvm_unreachable("unknown vld_dup intrinsic?"); 4214 } 4215 llvm::Type *Tys[] = {Ty, Int8PtrTy}; 4216 Function *F = CGM.getIntrinsic(Int, Tys); 4217 llvm::StructType *STy = cast<llvm::StructType>(F->getReturnType()); 4218 4219 SmallVector<Value*, 6> Args; 4220 Args.push_back(Ops[1]); 4221 Args.append(STy->getNumElements(), UndefValue::get(Ty)); 4222 4223 llvm::Constant *CI = ConstantInt::get(Int32Ty, 0); 4224 Args.push_back(CI); 4225 Args.push_back(getAlignmentValue32(PtrOp1)); 4226 4227 Ops[1] = Builder.CreateCall(F, Args, "vld_dup"); 4228 // splat lane 0 to all elts in each vector of the result. 4229 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 4230 Value *Val = Builder.CreateExtractValue(Ops[1], i); 4231 Value *Elt = Builder.CreateBitCast(Val, Ty); 4232 Elt = EmitNeonSplat(Elt, CI); 4233 Elt = Builder.CreateBitCast(Elt, Val->getType()); 4234 Ops[1] = Builder.CreateInsertValue(Ops[1], Elt, i); 4235 } 4236 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 4237 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 4238 return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]); 4239 } 4240 case NEON::BI__builtin_neon_vqrshrn_n_v: 4241 Int = 4242 usgn ? Intrinsic::arm_neon_vqrshiftnu : Intrinsic::arm_neon_vqrshiftns; 4243 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqrshrn_n", 4244 1, true); 4245 case NEON::BI__builtin_neon_vqrshrun_n_v: 4246 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vqrshiftnsu, Ty), 4247 Ops, "vqrshrun_n", 1, true); 4248 case NEON::BI__builtin_neon_vqshrn_n_v: 4249 Int = usgn ? Intrinsic::arm_neon_vqshiftnu : Intrinsic::arm_neon_vqshiftns; 4250 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqshrn_n", 4251 1, true); 4252 case NEON::BI__builtin_neon_vqshrun_n_v: 4253 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vqshiftnsu, Ty), 4254 Ops, "vqshrun_n", 1, true); 4255 case NEON::BI__builtin_neon_vrecpe_v: 4256 case NEON::BI__builtin_neon_vrecpeq_v: 4257 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vrecpe, Ty), 4258 Ops, "vrecpe"); 4259 case NEON::BI__builtin_neon_vrshrn_n_v: 4260 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vrshiftn, Ty), 4261 Ops, "vrshrn_n", 1, true); 4262 case NEON::BI__builtin_neon_vrsra_n_v: 4263 case NEON::BI__builtin_neon_vrsraq_n_v: 4264 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 4265 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 4266 Ops[2] = EmitNeonShiftVector(Ops[2], Ty, true); 4267 Int = usgn ? Intrinsic::arm_neon_vrshiftu : Intrinsic::arm_neon_vrshifts; 4268 Ops[1] = Builder.CreateCall(CGM.getIntrinsic(Int, Ty), {Ops[1], Ops[2]}); 4269 return Builder.CreateAdd(Ops[0], Ops[1], "vrsra_n"); 4270 case NEON::BI__builtin_neon_vsri_n_v: 4271 case NEON::BI__builtin_neon_vsriq_n_v: 4272 rightShift = true; 4273 case NEON::BI__builtin_neon_vsli_n_v: 4274 case NEON::BI__builtin_neon_vsliq_n_v: 4275 Ops[2] = EmitNeonShiftVector(Ops[2], Ty, rightShift); 4276 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vshiftins, Ty), 4277 Ops, "vsli_n"); 4278 case NEON::BI__builtin_neon_vsra_n_v: 4279 case NEON::BI__builtin_neon_vsraq_n_v: 4280 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 4281 Ops[1] = EmitNeonRShiftImm(Ops[1], Ops[2], Ty, usgn, "vsra_n"); 4282 return Builder.CreateAdd(Ops[0], Ops[1]); 4283 case NEON::BI__builtin_neon_vst1q_lane_v: 4284 // Handle 64-bit integer elements as a special case. Use a shuffle to get 4285 // a one-element vector and avoid poor code for i64 in the backend. 4286 if (VTy->getElementType()->isIntegerTy(64)) { 4287 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 4288 Value *SV = llvm::ConstantVector::get(cast<llvm::Constant>(Ops[2])); 4289 Ops[1] = Builder.CreateShuffleVector(Ops[1], Ops[1], SV); 4290 Ops[2] = getAlignmentValue32(PtrOp0); 4291 llvm::Type *Tys[] = {Int8PtrTy, Ops[1]->getType()}; 4292 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::arm_neon_vst1, 4293 Tys), Ops); 4294 } 4295 // fall through 4296 case NEON::BI__builtin_neon_vst1_lane_v: { 4297 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 4298 Ops[1] = Builder.CreateExtractElement(Ops[1], Ops[2]); 4299 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 4300 auto St = Builder.CreateStore(Ops[1], Builder.CreateBitCast(PtrOp0, Ty)); 4301 return St; 4302 } 4303 case NEON::BI__builtin_neon_vtbl1_v: 4304 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbl1), 4305 Ops, "vtbl1"); 4306 case NEON::BI__builtin_neon_vtbl2_v: 4307 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbl2), 4308 Ops, "vtbl2"); 4309 case NEON::BI__builtin_neon_vtbl3_v: 4310 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbl3), 4311 Ops, "vtbl3"); 4312 case NEON::BI__builtin_neon_vtbl4_v: 4313 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbl4), 4314 Ops, "vtbl4"); 4315 case NEON::BI__builtin_neon_vtbx1_v: 4316 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbx1), 4317 Ops, "vtbx1"); 4318 case NEON::BI__builtin_neon_vtbx2_v: 4319 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbx2), 4320 Ops, "vtbx2"); 4321 case NEON::BI__builtin_neon_vtbx3_v: 4322 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbx3), 4323 Ops, "vtbx3"); 4324 case NEON::BI__builtin_neon_vtbx4_v: 4325 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbx4), 4326 Ops, "vtbx4"); 4327 } 4328 } 4329 4330 static Value *EmitAArch64TblBuiltinExpr(CodeGenFunction &CGF, unsigned BuiltinID, 4331 const CallExpr *E, 4332 SmallVectorImpl<Value *> &Ops) { 4333 unsigned int Int = 0; 4334 const char *s = nullptr; 4335 4336 switch (BuiltinID) { 4337 default: 4338 return nullptr; 4339 case NEON::BI__builtin_neon_vtbl1_v: 4340 case NEON::BI__builtin_neon_vqtbl1_v: 4341 case NEON::BI__builtin_neon_vqtbl1q_v: 4342 case NEON::BI__builtin_neon_vtbl2_v: 4343 case NEON::BI__builtin_neon_vqtbl2_v: 4344 case NEON::BI__builtin_neon_vqtbl2q_v: 4345 case NEON::BI__builtin_neon_vtbl3_v: 4346 case NEON::BI__builtin_neon_vqtbl3_v: 4347 case NEON::BI__builtin_neon_vqtbl3q_v: 4348 case NEON::BI__builtin_neon_vtbl4_v: 4349 case NEON::BI__builtin_neon_vqtbl4_v: 4350 case NEON::BI__builtin_neon_vqtbl4q_v: 4351 break; 4352 case NEON::BI__builtin_neon_vtbx1_v: 4353 case NEON::BI__builtin_neon_vqtbx1_v: 4354 case NEON::BI__builtin_neon_vqtbx1q_v: 4355 case NEON::BI__builtin_neon_vtbx2_v: 4356 case NEON::BI__builtin_neon_vqtbx2_v: 4357 case NEON::BI__builtin_neon_vqtbx2q_v: 4358 case NEON::BI__builtin_neon_vtbx3_v: 4359 case NEON::BI__builtin_neon_vqtbx3_v: 4360 case NEON::BI__builtin_neon_vqtbx3q_v: 4361 case NEON::BI__builtin_neon_vtbx4_v: 4362 case NEON::BI__builtin_neon_vqtbx4_v: 4363 case NEON::BI__builtin_neon_vqtbx4q_v: 4364 break; 4365 } 4366 4367 assert(E->getNumArgs() >= 3); 4368 4369 // Get the last argument, which specifies the vector type. 4370 llvm::APSInt Result; 4371 const Expr *Arg = E->getArg(E->getNumArgs() - 1); 4372 if (!Arg->isIntegerConstantExpr(Result, CGF.getContext())) 4373 return nullptr; 4374 4375 // Determine the type of this overloaded NEON intrinsic. 4376 NeonTypeFlags Type(Result.getZExtValue()); 4377 llvm::VectorType *Ty = GetNeonType(&CGF, Type); 4378 if (!Ty) 4379 return nullptr; 4380 4381 CodeGen::CGBuilderTy &Builder = CGF.Builder; 4382 4383 // AArch64 scalar builtins are not overloaded, they do not have an extra 4384 // argument that specifies the vector type, need to handle each case. 4385 switch (BuiltinID) { 4386 case NEON::BI__builtin_neon_vtbl1_v: { 4387 return packTBLDVectorList(CGF, makeArrayRef(Ops).slice(0, 1), nullptr, 4388 Ops[1], Ty, Intrinsic::aarch64_neon_tbl1, 4389 "vtbl1"); 4390 } 4391 case NEON::BI__builtin_neon_vtbl2_v: { 4392 return packTBLDVectorList(CGF, makeArrayRef(Ops).slice(0, 2), nullptr, 4393 Ops[2], Ty, Intrinsic::aarch64_neon_tbl1, 4394 "vtbl1"); 4395 } 4396 case NEON::BI__builtin_neon_vtbl3_v: { 4397 return packTBLDVectorList(CGF, makeArrayRef(Ops).slice(0, 3), nullptr, 4398 Ops[3], Ty, Intrinsic::aarch64_neon_tbl2, 4399 "vtbl2"); 4400 } 4401 case NEON::BI__builtin_neon_vtbl4_v: { 4402 return packTBLDVectorList(CGF, makeArrayRef(Ops).slice(0, 4), nullptr, 4403 Ops[4], Ty, Intrinsic::aarch64_neon_tbl2, 4404 "vtbl2"); 4405 } 4406 case NEON::BI__builtin_neon_vtbx1_v: { 4407 Value *TblRes = 4408 packTBLDVectorList(CGF, makeArrayRef(Ops).slice(1, 1), nullptr, Ops[2], 4409 Ty, Intrinsic::aarch64_neon_tbl1, "vtbl1"); 4410 4411 llvm::Constant *EightV = ConstantInt::get(Ty, 8); 4412 Value *CmpRes = Builder.CreateICmp(ICmpInst::ICMP_UGE, Ops[2], EightV); 4413 CmpRes = Builder.CreateSExt(CmpRes, Ty); 4414 4415 Value *EltsFromInput = Builder.CreateAnd(CmpRes, Ops[0]); 4416 Value *EltsFromTbl = Builder.CreateAnd(Builder.CreateNot(CmpRes), TblRes); 4417 return Builder.CreateOr(EltsFromInput, EltsFromTbl, "vtbx"); 4418 } 4419 case NEON::BI__builtin_neon_vtbx2_v: { 4420 return packTBLDVectorList(CGF, makeArrayRef(Ops).slice(1, 2), Ops[0], 4421 Ops[3], Ty, Intrinsic::aarch64_neon_tbx1, 4422 "vtbx1"); 4423 } 4424 case NEON::BI__builtin_neon_vtbx3_v: { 4425 Value *TblRes = 4426 packTBLDVectorList(CGF, makeArrayRef(Ops).slice(1, 3), nullptr, Ops[4], 4427 Ty, Intrinsic::aarch64_neon_tbl2, "vtbl2"); 4428 4429 llvm::Constant *TwentyFourV = ConstantInt::get(Ty, 24); 4430 Value *CmpRes = Builder.CreateICmp(ICmpInst::ICMP_UGE, Ops[4], 4431 TwentyFourV); 4432 CmpRes = Builder.CreateSExt(CmpRes, Ty); 4433 4434 Value *EltsFromInput = Builder.CreateAnd(CmpRes, Ops[0]); 4435 Value *EltsFromTbl = Builder.CreateAnd(Builder.CreateNot(CmpRes), TblRes); 4436 return Builder.CreateOr(EltsFromInput, EltsFromTbl, "vtbx"); 4437 } 4438 case NEON::BI__builtin_neon_vtbx4_v: { 4439 return packTBLDVectorList(CGF, makeArrayRef(Ops).slice(1, 4), Ops[0], 4440 Ops[5], Ty, Intrinsic::aarch64_neon_tbx2, 4441 "vtbx2"); 4442 } 4443 case NEON::BI__builtin_neon_vqtbl1_v: 4444 case NEON::BI__builtin_neon_vqtbl1q_v: 4445 Int = Intrinsic::aarch64_neon_tbl1; s = "vtbl1"; break; 4446 case NEON::BI__builtin_neon_vqtbl2_v: 4447 case NEON::BI__builtin_neon_vqtbl2q_v: { 4448 Int = Intrinsic::aarch64_neon_tbl2; s = "vtbl2"; break; 4449 case NEON::BI__builtin_neon_vqtbl3_v: 4450 case NEON::BI__builtin_neon_vqtbl3q_v: 4451 Int = Intrinsic::aarch64_neon_tbl3; s = "vtbl3"; break; 4452 case NEON::BI__builtin_neon_vqtbl4_v: 4453 case NEON::BI__builtin_neon_vqtbl4q_v: 4454 Int = Intrinsic::aarch64_neon_tbl4; s = "vtbl4"; break; 4455 case NEON::BI__builtin_neon_vqtbx1_v: 4456 case NEON::BI__builtin_neon_vqtbx1q_v: 4457 Int = Intrinsic::aarch64_neon_tbx1; s = "vtbx1"; break; 4458 case NEON::BI__builtin_neon_vqtbx2_v: 4459 case NEON::BI__builtin_neon_vqtbx2q_v: 4460 Int = Intrinsic::aarch64_neon_tbx2; s = "vtbx2"; break; 4461 case NEON::BI__builtin_neon_vqtbx3_v: 4462 case NEON::BI__builtin_neon_vqtbx3q_v: 4463 Int = Intrinsic::aarch64_neon_tbx3; s = "vtbx3"; break; 4464 case NEON::BI__builtin_neon_vqtbx4_v: 4465 case NEON::BI__builtin_neon_vqtbx4q_v: 4466 Int = Intrinsic::aarch64_neon_tbx4; s = "vtbx4"; break; 4467 } 4468 } 4469 4470 if (!Int) 4471 return nullptr; 4472 4473 Function *F = CGF.CGM.getIntrinsic(Int, Ty); 4474 return CGF.EmitNeonCall(F, Ops, s); 4475 } 4476 4477 Value *CodeGenFunction::vectorWrapScalar16(Value *Op) { 4478 llvm::Type *VTy = llvm::VectorType::get(Int16Ty, 4); 4479 Op = Builder.CreateBitCast(Op, Int16Ty); 4480 Value *V = UndefValue::get(VTy); 4481 llvm::Constant *CI = ConstantInt::get(SizeTy, 0); 4482 Op = Builder.CreateInsertElement(V, Op, CI); 4483 return Op; 4484 } 4485 4486 Value *CodeGenFunction::EmitAArch64BuiltinExpr(unsigned BuiltinID, 4487 const CallExpr *E) { 4488 unsigned HintID = static_cast<unsigned>(-1); 4489 switch (BuiltinID) { 4490 default: break; 4491 case AArch64::BI__builtin_arm_nop: 4492 HintID = 0; 4493 break; 4494 case AArch64::BI__builtin_arm_yield: 4495 HintID = 1; 4496 break; 4497 case AArch64::BI__builtin_arm_wfe: 4498 HintID = 2; 4499 break; 4500 case AArch64::BI__builtin_arm_wfi: 4501 HintID = 3; 4502 break; 4503 case AArch64::BI__builtin_arm_sev: 4504 HintID = 4; 4505 break; 4506 case AArch64::BI__builtin_arm_sevl: 4507 HintID = 5; 4508 break; 4509 } 4510 4511 if (HintID != static_cast<unsigned>(-1)) { 4512 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_hint); 4513 return Builder.CreateCall(F, llvm::ConstantInt::get(Int32Ty, HintID)); 4514 } 4515 4516 if (BuiltinID == AArch64::BI__builtin_arm_prefetch) { 4517 Value *Address = EmitScalarExpr(E->getArg(0)); 4518 Value *RW = EmitScalarExpr(E->getArg(1)); 4519 Value *CacheLevel = EmitScalarExpr(E->getArg(2)); 4520 Value *RetentionPolicy = EmitScalarExpr(E->getArg(3)); 4521 Value *IsData = EmitScalarExpr(E->getArg(4)); 4522 4523 Value *Locality = nullptr; 4524 if (cast<llvm::ConstantInt>(RetentionPolicy)->isZero()) { 4525 // Temporal fetch, needs to convert cache level to locality. 4526 Locality = llvm::ConstantInt::get(Int32Ty, 4527 -cast<llvm::ConstantInt>(CacheLevel)->getValue() + 3); 4528 } else { 4529 // Streaming fetch. 4530 Locality = llvm::ConstantInt::get(Int32Ty, 0); 4531 } 4532 4533 // FIXME: We need AArch64 specific LLVM intrinsic if we want to specify 4534 // PLDL3STRM or PLDL2STRM. 4535 Value *F = CGM.getIntrinsic(Intrinsic::prefetch); 4536 return Builder.CreateCall(F, {Address, RW, Locality, IsData}); 4537 } 4538 4539 if (BuiltinID == AArch64::BI__builtin_arm_rbit) { 4540 assert((getContext().getTypeSize(E->getType()) == 32) && 4541 "rbit of unusual size!"); 4542 llvm::Value *Arg = EmitScalarExpr(E->getArg(0)); 4543 return Builder.CreateCall( 4544 CGM.getIntrinsic(Intrinsic::aarch64_rbit, Arg->getType()), Arg, "rbit"); 4545 } 4546 if (BuiltinID == AArch64::BI__builtin_arm_rbit64) { 4547 assert((getContext().getTypeSize(E->getType()) == 64) && 4548 "rbit of unusual size!"); 4549 llvm::Value *Arg = EmitScalarExpr(E->getArg(0)); 4550 return Builder.CreateCall( 4551 CGM.getIntrinsic(Intrinsic::aarch64_rbit, Arg->getType()), Arg, "rbit"); 4552 } 4553 4554 if (BuiltinID == AArch64::BI__clear_cache) { 4555 assert(E->getNumArgs() == 2 && "__clear_cache takes 2 arguments"); 4556 const FunctionDecl *FD = E->getDirectCallee(); 4557 Value *Ops[2]; 4558 for (unsigned i = 0; i < 2; i++) 4559 Ops[i] = EmitScalarExpr(E->getArg(i)); 4560 llvm::Type *Ty = CGM.getTypes().ConvertType(FD->getType()); 4561 llvm::FunctionType *FTy = cast<llvm::FunctionType>(Ty); 4562 StringRef Name = FD->getName(); 4563 return EmitNounwindRuntimeCall(CGM.CreateRuntimeFunction(FTy, Name), Ops); 4564 } 4565 4566 if ((BuiltinID == AArch64::BI__builtin_arm_ldrex || 4567 BuiltinID == AArch64::BI__builtin_arm_ldaex) && 4568 getContext().getTypeSize(E->getType()) == 128) { 4569 Function *F = CGM.getIntrinsic(BuiltinID == AArch64::BI__builtin_arm_ldaex 4570 ? Intrinsic::aarch64_ldaxp 4571 : Intrinsic::aarch64_ldxp); 4572 4573 Value *LdPtr = EmitScalarExpr(E->getArg(0)); 4574 Value *Val = Builder.CreateCall(F, Builder.CreateBitCast(LdPtr, Int8PtrTy), 4575 "ldxp"); 4576 4577 Value *Val0 = Builder.CreateExtractValue(Val, 1); 4578 Value *Val1 = Builder.CreateExtractValue(Val, 0); 4579 llvm::Type *Int128Ty = llvm::IntegerType::get(getLLVMContext(), 128); 4580 Val0 = Builder.CreateZExt(Val0, Int128Ty); 4581 Val1 = Builder.CreateZExt(Val1, Int128Ty); 4582 4583 Value *ShiftCst = llvm::ConstantInt::get(Int128Ty, 64); 4584 Val = Builder.CreateShl(Val0, ShiftCst, "shl", true /* nuw */); 4585 Val = Builder.CreateOr(Val, Val1); 4586 return Builder.CreateBitCast(Val, ConvertType(E->getType())); 4587 } else if (BuiltinID == AArch64::BI__builtin_arm_ldrex || 4588 BuiltinID == AArch64::BI__builtin_arm_ldaex) { 4589 Value *LoadAddr = EmitScalarExpr(E->getArg(0)); 4590 4591 QualType Ty = E->getType(); 4592 llvm::Type *RealResTy = ConvertType(Ty); 4593 llvm::Type *IntResTy = llvm::IntegerType::get(getLLVMContext(), 4594 getContext().getTypeSize(Ty)); 4595 LoadAddr = Builder.CreateBitCast(LoadAddr, IntResTy->getPointerTo()); 4596 4597 Function *F = CGM.getIntrinsic(BuiltinID == AArch64::BI__builtin_arm_ldaex 4598 ? Intrinsic::aarch64_ldaxr 4599 : Intrinsic::aarch64_ldxr, 4600 LoadAddr->getType()); 4601 Value *Val = Builder.CreateCall(F, LoadAddr, "ldxr"); 4602 4603 if (RealResTy->isPointerTy()) 4604 return Builder.CreateIntToPtr(Val, RealResTy); 4605 4606 Val = Builder.CreateTruncOrBitCast(Val, IntResTy); 4607 return Builder.CreateBitCast(Val, RealResTy); 4608 } 4609 4610 if ((BuiltinID == AArch64::BI__builtin_arm_strex || 4611 BuiltinID == AArch64::BI__builtin_arm_stlex) && 4612 getContext().getTypeSize(E->getArg(0)->getType()) == 128) { 4613 Function *F = CGM.getIntrinsic(BuiltinID == AArch64::BI__builtin_arm_stlex 4614 ? Intrinsic::aarch64_stlxp 4615 : Intrinsic::aarch64_stxp); 4616 llvm::Type *STy = llvm::StructType::get(Int64Ty, Int64Ty, nullptr); 4617 4618 Address Tmp = CreateMemTemp(E->getArg(0)->getType()); 4619 EmitAnyExprToMem(E->getArg(0), Tmp, Qualifiers(), /*init*/ true); 4620 4621 Tmp = Builder.CreateBitCast(Tmp, llvm::PointerType::getUnqual(STy)); 4622 llvm::Value *Val = Builder.CreateLoad(Tmp); 4623 4624 Value *Arg0 = Builder.CreateExtractValue(Val, 0); 4625 Value *Arg1 = Builder.CreateExtractValue(Val, 1); 4626 Value *StPtr = Builder.CreateBitCast(EmitScalarExpr(E->getArg(1)), 4627 Int8PtrTy); 4628 return Builder.CreateCall(F, {Arg0, Arg1, StPtr}, "stxp"); 4629 } 4630 4631 if (BuiltinID == AArch64::BI__builtin_arm_strex || 4632 BuiltinID == AArch64::BI__builtin_arm_stlex) { 4633 Value *StoreVal = EmitScalarExpr(E->getArg(0)); 4634 Value *StoreAddr = EmitScalarExpr(E->getArg(1)); 4635 4636 QualType Ty = E->getArg(0)->getType(); 4637 llvm::Type *StoreTy = llvm::IntegerType::get(getLLVMContext(), 4638 getContext().getTypeSize(Ty)); 4639 StoreAddr = Builder.CreateBitCast(StoreAddr, StoreTy->getPointerTo()); 4640 4641 if (StoreVal->getType()->isPointerTy()) 4642 StoreVal = Builder.CreatePtrToInt(StoreVal, Int64Ty); 4643 else { 4644 StoreVal = Builder.CreateBitCast(StoreVal, StoreTy); 4645 StoreVal = Builder.CreateZExtOrBitCast(StoreVal, Int64Ty); 4646 } 4647 4648 Function *F = CGM.getIntrinsic(BuiltinID == AArch64::BI__builtin_arm_stlex 4649 ? Intrinsic::aarch64_stlxr 4650 : Intrinsic::aarch64_stxr, 4651 StoreAddr->getType()); 4652 return Builder.CreateCall(F, {StoreVal, StoreAddr}, "stxr"); 4653 } 4654 4655 if (BuiltinID == AArch64::BI__builtin_arm_clrex) { 4656 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_clrex); 4657 return Builder.CreateCall(F); 4658 } 4659 4660 if (BuiltinID == AArch64::BI__builtin_thread_pointer) { 4661 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_thread_pointer); 4662 return Builder.CreateCall(F); 4663 } 4664 4665 // CRC32 4666 Intrinsic::ID CRCIntrinsicID = Intrinsic::not_intrinsic; 4667 switch (BuiltinID) { 4668 case AArch64::BI__builtin_arm_crc32b: 4669 CRCIntrinsicID = Intrinsic::aarch64_crc32b; break; 4670 case AArch64::BI__builtin_arm_crc32cb: 4671 CRCIntrinsicID = Intrinsic::aarch64_crc32cb; break; 4672 case AArch64::BI__builtin_arm_crc32h: 4673 CRCIntrinsicID = Intrinsic::aarch64_crc32h; break; 4674 case AArch64::BI__builtin_arm_crc32ch: 4675 CRCIntrinsicID = Intrinsic::aarch64_crc32ch; break; 4676 case AArch64::BI__builtin_arm_crc32w: 4677 CRCIntrinsicID = Intrinsic::aarch64_crc32w; break; 4678 case AArch64::BI__builtin_arm_crc32cw: 4679 CRCIntrinsicID = Intrinsic::aarch64_crc32cw; break; 4680 case AArch64::BI__builtin_arm_crc32d: 4681 CRCIntrinsicID = Intrinsic::aarch64_crc32x; break; 4682 case AArch64::BI__builtin_arm_crc32cd: 4683 CRCIntrinsicID = Intrinsic::aarch64_crc32cx; break; 4684 } 4685 4686 if (CRCIntrinsicID != Intrinsic::not_intrinsic) { 4687 Value *Arg0 = EmitScalarExpr(E->getArg(0)); 4688 Value *Arg1 = EmitScalarExpr(E->getArg(1)); 4689 Function *F = CGM.getIntrinsic(CRCIntrinsicID); 4690 4691 llvm::Type *DataTy = F->getFunctionType()->getParamType(1); 4692 Arg1 = Builder.CreateZExtOrBitCast(Arg1, DataTy); 4693 4694 return Builder.CreateCall(F, {Arg0, Arg1}); 4695 } 4696 4697 if (BuiltinID == AArch64::BI__builtin_arm_rsr || 4698 BuiltinID == AArch64::BI__builtin_arm_rsr64 || 4699 BuiltinID == AArch64::BI__builtin_arm_rsrp || 4700 BuiltinID == AArch64::BI__builtin_arm_wsr || 4701 BuiltinID == AArch64::BI__builtin_arm_wsr64 || 4702 BuiltinID == AArch64::BI__builtin_arm_wsrp) { 4703 4704 bool IsRead = BuiltinID == AArch64::BI__builtin_arm_rsr || 4705 BuiltinID == AArch64::BI__builtin_arm_rsr64 || 4706 BuiltinID == AArch64::BI__builtin_arm_rsrp; 4707 4708 bool IsPointerBuiltin = BuiltinID == AArch64::BI__builtin_arm_rsrp || 4709 BuiltinID == AArch64::BI__builtin_arm_wsrp; 4710 4711 bool Is64Bit = BuiltinID != AArch64::BI__builtin_arm_rsr && 4712 BuiltinID != AArch64::BI__builtin_arm_wsr; 4713 4714 llvm::Type *ValueType; 4715 llvm::Type *RegisterType = Int64Ty; 4716 if (IsPointerBuiltin) { 4717 ValueType = VoidPtrTy; 4718 } else if (Is64Bit) { 4719 ValueType = Int64Ty; 4720 } else { 4721 ValueType = Int32Ty; 4722 } 4723 4724 return EmitSpecialRegisterBuiltin(*this, E, RegisterType, ValueType, IsRead); 4725 } 4726 4727 // Find out if any arguments are required to be integer constant 4728 // expressions. 4729 unsigned ICEArguments = 0; 4730 ASTContext::GetBuiltinTypeError Error; 4731 getContext().GetBuiltinType(BuiltinID, Error, &ICEArguments); 4732 assert(Error == ASTContext::GE_None && "Should not codegen an error"); 4733 4734 llvm::SmallVector<Value*, 4> Ops; 4735 for (unsigned i = 0, e = E->getNumArgs() - 1; i != e; i++) { 4736 if ((ICEArguments & (1 << i)) == 0) { 4737 Ops.push_back(EmitScalarExpr(E->getArg(i))); 4738 } else { 4739 // If this is required to be a constant, constant fold it so that we know 4740 // that the generated intrinsic gets a ConstantInt. 4741 llvm::APSInt Result; 4742 bool IsConst = E->getArg(i)->isIntegerConstantExpr(Result, getContext()); 4743 assert(IsConst && "Constant arg isn't actually constant?"); 4744 (void)IsConst; 4745 Ops.push_back(llvm::ConstantInt::get(getLLVMContext(), Result)); 4746 } 4747 } 4748 4749 auto SISDMap = makeArrayRef(AArch64SISDIntrinsicMap); 4750 const NeonIntrinsicInfo *Builtin = findNeonIntrinsicInMap( 4751 SISDMap, BuiltinID, AArch64SISDIntrinsicsProvenSorted); 4752 4753 if (Builtin) { 4754 Ops.push_back(EmitScalarExpr(E->getArg(E->getNumArgs() - 1))); 4755 Value *Result = EmitCommonNeonSISDBuiltinExpr(*this, *Builtin, Ops, E); 4756 assert(Result && "SISD intrinsic should have been handled"); 4757 return Result; 4758 } 4759 4760 llvm::APSInt Result; 4761 const Expr *Arg = E->getArg(E->getNumArgs()-1); 4762 NeonTypeFlags Type(0); 4763 if (Arg->isIntegerConstantExpr(Result, getContext())) 4764 // Determine the type of this overloaded NEON intrinsic. 4765 Type = NeonTypeFlags(Result.getZExtValue()); 4766 4767 bool usgn = Type.isUnsigned(); 4768 bool quad = Type.isQuad(); 4769 4770 // Handle non-overloaded intrinsics first. 4771 switch (BuiltinID) { 4772 default: break; 4773 case NEON::BI__builtin_neon_vldrq_p128: { 4774 llvm::Type *Int128PTy = llvm::Type::getIntNPtrTy(getLLVMContext(), 128); 4775 Value *Ptr = Builder.CreateBitCast(EmitScalarExpr(E->getArg(0)), Int128PTy); 4776 return Builder.CreateDefaultAlignedLoad(Ptr); 4777 } 4778 case NEON::BI__builtin_neon_vstrq_p128: { 4779 llvm::Type *Int128PTy = llvm::Type::getIntNPtrTy(getLLVMContext(), 128); 4780 Value *Ptr = Builder.CreateBitCast(Ops[0], Int128PTy); 4781 return Builder.CreateDefaultAlignedStore(EmitScalarExpr(E->getArg(1)), Ptr); 4782 } 4783 case NEON::BI__builtin_neon_vcvts_u32_f32: 4784 case NEON::BI__builtin_neon_vcvtd_u64_f64: 4785 usgn = true; 4786 // FALL THROUGH 4787 case NEON::BI__builtin_neon_vcvts_s32_f32: 4788 case NEON::BI__builtin_neon_vcvtd_s64_f64: { 4789 Ops.push_back(EmitScalarExpr(E->getArg(0))); 4790 bool Is64 = Ops[0]->getType()->getPrimitiveSizeInBits() == 64; 4791 llvm::Type *InTy = Is64 ? Int64Ty : Int32Ty; 4792 llvm::Type *FTy = Is64 ? DoubleTy : FloatTy; 4793 Ops[0] = Builder.CreateBitCast(Ops[0], FTy); 4794 if (usgn) 4795 return Builder.CreateFPToUI(Ops[0], InTy); 4796 return Builder.CreateFPToSI(Ops[0], InTy); 4797 } 4798 case NEON::BI__builtin_neon_vcvts_f32_u32: 4799 case NEON::BI__builtin_neon_vcvtd_f64_u64: 4800 usgn = true; 4801 // FALL THROUGH 4802 case NEON::BI__builtin_neon_vcvts_f32_s32: 4803 case NEON::BI__builtin_neon_vcvtd_f64_s64: { 4804 Ops.push_back(EmitScalarExpr(E->getArg(0))); 4805 bool Is64 = Ops[0]->getType()->getPrimitiveSizeInBits() == 64; 4806 llvm::Type *InTy = Is64 ? Int64Ty : Int32Ty; 4807 llvm::Type *FTy = Is64 ? DoubleTy : FloatTy; 4808 Ops[0] = Builder.CreateBitCast(Ops[0], InTy); 4809 if (usgn) 4810 return Builder.CreateUIToFP(Ops[0], FTy); 4811 return Builder.CreateSIToFP(Ops[0], FTy); 4812 } 4813 case NEON::BI__builtin_neon_vpaddd_s64: { 4814 llvm::Type *Ty = llvm::VectorType::get(Int64Ty, 2); 4815 Value *Vec = EmitScalarExpr(E->getArg(0)); 4816 // The vector is v2f64, so make sure it's bitcast to that. 4817 Vec = Builder.CreateBitCast(Vec, Ty, "v2i64"); 4818 llvm::Value *Idx0 = llvm::ConstantInt::get(SizeTy, 0); 4819 llvm::Value *Idx1 = llvm::ConstantInt::get(SizeTy, 1); 4820 Value *Op0 = Builder.CreateExtractElement(Vec, Idx0, "lane0"); 4821 Value *Op1 = Builder.CreateExtractElement(Vec, Idx1, "lane1"); 4822 // Pairwise addition of a v2f64 into a scalar f64. 4823 return Builder.CreateAdd(Op0, Op1, "vpaddd"); 4824 } 4825 case NEON::BI__builtin_neon_vpaddd_f64: { 4826 llvm::Type *Ty = 4827 llvm::VectorType::get(DoubleTy, 2); 4828 Value *Vec = EmitScalarExpr(E->getArg(0)); 4829 // The vector is v2f64, so make sure it's bitcast to that. 4830 Vec = Builder.CreateBitCast(Vec, Ty, "v2f64"); 4831 llvm::Value *Idx0 = llvm::ConstantInt::get(SizeTy, 0); 4832 llvm::Value *Idx1 = llvm::ConstantInt::get(SizeTy, 1); 4833 Value *Op0 = Builder.CreateExtractElement(Vec, Idx0, "lane0"); 4834 Value *Op1 = Builder.CreateExtractElement(Vec, Idx1, "lane1"); 4835 // Pairwise addition of a v2f64 into a scalar f64. 4836 return Builder.CreateFAdd(Op0, Op1, "vpaddd"); 4837 } 4838 case NEON::BI__builtin_neon_vpadds_f32: { 4839 llvm::Type *Ty = 4840 llvm::VectorType::get(FloatTy, 2); 4841 Value *Vec = EmitScalarExpr(E->getArg(0)); 4842 // The vector is v2f32, so make sure it's bitcast to that. 4843 Vec = Builder.CreateBitCast(Vec, Ty, "v2f32"); 4844 llvm::Value *Idx0 = llvm::ConstantInt::get(SizeTy, 0); 4845 llvm::Value *Idx1 = llvm::ConstantInt::get(SizeTy, 1); 4846 Value *Op0 = Builder.CreateExtractElement(Vec, Idx0, "lane0"); 4847 Value *Op1 = Builder.CreateExtractElement(Vec, Idx1, "lane1"); 4848 // Pairwise addition of a v2f32 into a scalar f32. 4849 return Builder.CreateFAdd(Op0, Op1, "vpaddd"); 4850 } 4851 case NEON::BI__builtin_neon_vceqzd_s64: 4852 case NEON::BI__builtin_neon_vceqzd_f64: 4853 case NEON::BI__builtin_neon_vceqzs_f32: 4854 Ops.push_back(EmitScalarExpr(E->getArg(0))); 4855 return EmitAArch64CompareBuiltinExpr( 4856 Ops[0], ConvertType(E->getCallReturnType(getContext())), 4857 ICmpInst::FCMP_OEQ, ICmpInst::ICMP_EQ, "vceqz"); 4858 case NEON::BI__builtin_neon_vcgezd_s64: 4859 case NEON::BI__builtin_neon_vcgezd_f64: 4860 case NEON::BI__builtin_neon_vcgezs_f32: 4861 Ops.push_back(EmitScalarExpr(E->getArg(0))); 4862 return EmitAArch64CompareBuiltinExpr( 4863 Ops[0], ConvertType(E->getCallReturnType(getContext())), 4864 ICmpInst::FCMP_OGE, ICmpInst::ICMP_SGE, "vcgez"); 4865 case NEON::BI__builtin_neon_vclezd_s64: 4866 case NEON::BI__builtin_neon_vclezd_f64: 4867 case NEON::BI__builtin_neon_vclezs_f32: 4868 Ops.push_back(EmitScalarExpr(E->getArg(0))); 4869 return EmitAArch64CompareBuiltinExpr( 4870 Ops[0], ConvertType(E->getCallReturnType(getContext())), 4871 ICmpInst::FCMP_OLE, ICmpInst::ICMP_SLE, "vclez"); 4872 case NEON::BI__builtin_neon_vcgtzd_s64: 4873 case NEON::BI__builtin_neon_vcgtzd_f64: 4874 case NEON::BI__builtin_neon_vcgtzs_f32: 4875 Ops.push_back(EmitScalarExpr(E->getArg(0))); 4876 return EmitAArch64CompareBuiltinExpr( 4877 Ops[0], ConvertType(E->getCallReturnType(getContext())), 4878 ICmpInst::FCMP_OGT, ICmpInst::ICMP_SGT, "vcgtz"); 4879 case NEON::BI__builtin_neon_vcltzd_s64: 4880 case NEON::BI__builtin_neon_vcltzd_f64: 4881 case NEON::BI__builtin_neon_vcltzs_f32: 4882 Ops.push_back(EmitScalarExpr(E->getArg(0))); 4883 return EmitAArch64CompareBuiltinExpr( 4884 Ops[0], ConvertType(E->getCallReturnType(getContext())), 4885 ICmpInst::FCMP_OLT, ICmpInst::ICMP_SLT, "vcltz"); 4886 4887 case NEON::BI__builtin_neon_vceqzd_u64: { 4888 Ops.push_back(EmitScalarExpr(E->getArg(0))); 4889 Ops[0] = Builder.CreateBitCast(Ops[0], Int64Ty); 4890 Ops[0] = 4891 Builder.CreateICmpEQ(Ops[0], llvm::Constant::getNullValue(Int64Ty)); 4892 return Builder.CreateSExt(Ops[0], Int64Ty, "vceqzd"); 4893 } 4894 case NEON::BI__builtin_neon_vceqd_f64: 4895 case NEON::BI__builtin_neon_vcled_f64: 4896 case NEON::BI__builtin_neon_vcltd_f64: 4897 case NEON::BI__builtin_neon_vcged_f64: 4898 case NEON::BI__builtin_neon_vcgtd_f64: { 4899 llvm::CmpInst::Predicate P; 4900 switch (BuiltinID) { 4901 default: llvm_unreachable("missing builtin ID in switch!"); 4902 case NEON::BI__builtin_neon_vceqd_f64: P = llvm::FCmpInst::FCMP_OEQ; break; 4903 case NEON::BI__builtin_neon_vcled_f64: P = llvm::FCmpInst::FCMP_OLE; break; 4904 case NEON::BI__builtin_neon_vcltd_f64: P = llvm::FCmpInst::FCMP_OLT; break; 4905 case NEON::BI__builtin_neon_vcged_f64: P = llvm::FCmpInst::FCMP_OGE; break; 4906 case NEON::BI__builtin_neon_vcgtd_f64: P = llvm::FCmpInst::FCMP_OGT; break; 4907 } 4908 Ops.push_back(EmitScalarExpr(E->getArg(1))); 4909 Ops[0] = Builder.CreateBitCast(Ops[0], DoubleTy); 4910 Ops[1] = Builder.CreateBitCast(Ops[1], DoubleTy); 4911 Ops[0] = Builder.CreateFCmp(P, Ops[0], Ops[1]); 4912 return Builder.CreateSExt(Ops[0], Int64Ty, "vcmpd"); 4913 } 4914 case NEON::BI__builtin_neon_vceqs_f32: 4915 case NEON::BI__builtin_neon_vcles_f32: 4916 case NEON::BI__builtin_neon_vclts_f32: 4917 case NEON::BI__builtin_neon_vcges_f32: 4918 case NEON::BI__builtin_neon_vcgts_f32: { 4919 llvm::CmpInst::Predicate P; 4920 switch (BuiltinID) { 4921 default: llvm_unreachable("missing builtin ID in switch!"); 4922 case NEON::BI__builtin_neon_vceqs_f32: P = llvm::FCmpInst::FCMP_OEQ; break; 4923 case NEON::BI__builtin_neon_vcles_f32: P = llvm::FCmpInst::FCMP_OLE; break; 4924 case NEON::BI__builtin_neon_vclts_f32: P = llvm::FCmpInst::FCMP_OLT; break; 4925 case NEON::BI__builtin_neon_vcges_f32: P = llvm::FCmpInst::FCMP_OGE; break; 4926 case NEON::BI__builtin_neon_vcgts_f32: P = llvm::FCmpInst::FCMP_OGT; break; 4927 } 4928 Ops.push_back(EmitScalarExpr(E->getArg(1))); 4929 Ops[0] = Builder.CreateBitCast(Ops[0], FloatTy); 4930 Ops[1] = Builder.CreateBitCast(Ops[1], FloatTy); 4931 Ops[0] = Builder.CreateFCmp(P, Ops[0], Ops[1]); 4932 return Builder.CreateSExt(Ops[0], Int32Ty, "vcmpd"); 4933 } 4934 case NEON::BI__builtin_neon_vceqd_s64: 4935 case NEON::BI__builtin_neon_vceqd_u64: 4936 case NEON::BI__builtin_neon_vcgtd_s64: 4937 case NEON::BI__builtin_neon_vcgtd_u64: 4938 case NEON::BI__builtin_neon_vcltd_s64: 4939 case NEON::BI__builtin_neon_vcltd_u64: 4940 case NEON::BI__builtin_neon_vcged_u64: 4941 case NEON::BI__builtin_neon_vcged_s64: 4942 case NEON::BI__builtin_neon_vcled_u64: 4943 case NEON::BI__builtin_neon_vcled_s64: { 4944 llvm::CmpInst::Predicate P; 4945 switch (BuiltinID) { 4946 default: llvm_unreachable("missing builtin ID in switch!"); 4947 case NEON::BI__builtin_neon_vceqd_s64: 4948 case NEON::BI__builtin_neon_vceqd_u64:P = llvm::ICmpInst::ICMP_EQ;break; 4949 case NEON::BI__builtin_neon_vcgtd_s64:P = llvm::ICmpInst::ICMP_SGT;break; 4950 case NEON::BI__builtin_neon_vcgtd_u64:P = llvm::ICmpInst::ICMP_UGT;break; 4951 case NEON::BI__builtin_neon_vcltd_s64:P = llvm::ICmpInst::ICMP_SLT;break; 4952 case NEON::BI__builtin_neon_vcltd_u64:P = llvm::ICmpInst::ICMP_ULT;break; 4953 case NEON::BI__builtin_neon_vcged_u64:P = llvm::ICmpInst::ICMP_UGE;break; 4954 case NEON::BI__builtin_neon_vcged_s64:P = llvm::ICmpInst::ICMP_SGE;break; 4955 case NEON::BI__builtin_neon_vcled_u64:P = llvm::ICmpInst::ICMP_ULE;break; 4956 case NEON::BI__builtin_neon_vcled_s64:P = llvm::ICmpInst::ICMP_SLE;break; 4957 } 4958 Ops.push_back(EmitScalarExpr(E->getArg(1))); 4959 Ops[0] = Builder.CreateBitCast(Ops[0], Int64Ty); 4960 Ops[1] = Builder.CreateBitCast(Ops[1], Int64Ty); 4961 Ops[0] = Builder.CreateICmp(P, Ops[0], Ops[1]); 4962 return Builder.CreateSExt(Ops[0], Int64Ty, "vceqd"); 4963 } 4964 case NEON::BI__builtin_neon_vtstd_s64: 4965 case NEON::BI__builtin_neon_vtstd_u64: { 4966 Ops.push_back(EmitScalarExpr(E->getArg(1))); 4967 Ops[0] = Builder.CreateBitCast(Ops[0], Int64Ty); 4968 Ops[1] = Builder.CreateBitCast(Ops[1], Int64Ty); 4969 Ops[0] = Builder.CreateAnd(Ops[0], Ops[1]); 4970 Ops[0] = Builder.CreateICmp(ICmpInst::ICMP_NE, Ops[0], 4971 llvm::Constant::getNullValue(Int64Ty)); 4972 return Builder.CreateSExt(Ops[0], Int64Ty, "vtstd"); 4973 } 4974 case NEON::BI__builtin_neon_vset_lane_i8: 4975 case NEON::BI__builtin_neon_vset_lane_i16: 4976 case NEON::BI__builtin_neon_vset_lane_i32: 4977 case NEON::BI__builtin_neon_vset_lane_i64: 4978 case NEON::BI__builtin_neon_vset_lane_f32: 4979 case NEON::BI__builtin_neon_vsetq_lane_i8: 4980 case NEON::BI__builtin_neon_vsetq_lane_i16: 4981 case NEON::BI__builtin_neon_vsetq_lane_i32: 4982 case NEON::BI__builtin_neon_vsetq_lane_i64: 4983 case NEON::BI__builtin_neon_vsetq_lane_f32: 4984 Ops.push_back(EmitScalarExpr(E->getArg(2))); 4985 return Builder.CreateInsertElement(Ops[1], Ops[0], Ops[2], "vset_lane"); 4986 case NEON::BI__builtin_neon_vset_lane_f64: 4987 // The vector type needs a cast for the v1f64 variant. 4988 Ops[1] = Builder.CreateBitCast(Ops[1], 4989 llvm::VectorType::get(DoubleTy, 1)); 4990 Ops.push_back(EmitScalarExpr(E->getArg(2))); 4991 return Builder.CreateInsertElement(Ops[1], Ops[0], Ops[2], "vset_lane"); 4992 case NEON::BI__builtin_neon_vsetq_lane_f64: 4993 // The vector type needs a cast for the v2f64 variant. 4994 Ops[1] = Builder.CreateBitCast(Ops[1], 4995 llvm::VectorType::get(DoubleTy, 2)); 4996 Ops.push_back(EmitScalarExpr(E->getArg(2))); 4997 return Builder.CreateInsertElement(Ops[1], Ops[0], Ops[2], "vset_lane"); 4998 4999 case NEON::BI__builtin_neon_vget_lane_i8: 5000 case NEON::BI__builtin_neon_vdupb_lane_i8: 5001 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::VectorType::get(Int8Ty, 8)); 5002 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 5003 "vget_lane"); 5004 case NEON::BI__builtin_neon_vgetq_lane_i8: 5005 case NEON::BI__builtin_neon_vdupb_laneq_i8: 5006 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::VectorType::get(Int8Ty, 16)); 5007 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 5008 "vgetq_lane"); 5009 case NEON::BI__builtin_neon_vget_lane_i16: 5010 case NEON::BI__builtin_neon_vduph_lane_i16: 5011 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::VectorType::get(Int16Ty, 4)); 5012 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 5013 "vget_lane"); 5014 case NEON::BI__builtin_neon_vgetq_lane_i16: 5015 case NEON::BI__builtin_neon_vduph_laneq_i16: 5016 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::VectorType::get(Int16Ty, 8)); 5017 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 5018 "vgetq_lane"); 5019 case NEON::BI__builtin_neon_vget_lane_i32: 5020 case NEON::BI__builtin_neon_vdups_lane_i32: 5021 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::VectorType::get(Int32Ty, 2)); 5022 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 5023 "vget_lane"); 5024 case NEON::BI__builtin_neon_vdups_lane_f32: 5025 Ops[0] = Builder.CreateBitCast(Ops[0], 5026 llvm::VectorType::get(FloatTy, 2)); 5027 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 5028 "vdups_lane"); 5029 case NEON::BI__builtin_neon_vgetq_lane_i32: 5030 case NEON::BI__builtin_neon_vdups_laneq_i32: 5031 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::VectorType::get(Int32Ty, 4)); 5032 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 5033 "vgetq_lane"); 5034 case NEON::BI__builtin_neon_vget_lane_i64: 5035 case NEON::BI__builtin_neon_vdupd_lane_i64: 5036 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::VectorType::get(Int64Ty, 1)); 5037 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 5038 "vget_lane"); 5039 case NEON::BI__builtin_neon_vdupd_lane_f64: 5040 Ops[0] = Builder.CreateBitCast(Ops[0], 5041 llvm::VectorType::get(DoubleTy, 1)); 5042 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 5043 "vdupd_lane"); 5044 case NEON::BI__builtin_neon_vgetq_lane_i64: 5045 case NEON::BI__builtin_neon_vdupd_laneq_i64: 5046 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::VectorType::get(Int64Ty, 2)); 5047 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 5048 "vgetq_lane"); 5049 case NEON::BI__builtin_neon_vget_lane_f32: 5050 Ops[0] = Builder.CreateBitCast(Ops[0], 5051 llvm::VectorType::get(FloatTy, 2)); 5052 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 5053 "vget_lane"); 5054 case NEON::BI__builtin_neon_vget_lane_f64: 5055 Ops[0] = Builder.CreateBitCast(Ops[0], 5056 llvm::VectorType::get(DoubleTy, 1)); 5057 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 5058 "vget_lane"); 5059 case NEON::BI__builtin_neon_vgetq_lane_f32: 5060 case NEON::BI__builtin_neon_vdups_laneq_f32: 5061 Ops[0] = Builder.CreateBitCast(Ops[0], 5062 llvm::VectorType::get(FloatTy, 4)); 5063 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 5064 "vgetq_lane"); 5065 case NEON::BI__builtin_neon_vgetq_lane_f64: 5066 case NEON::BI__builtin_neon_vdupd_laneq_f64: 5067 Ops[0] = Builder.CreateBitCast(Ops[0], 5068 llvm::VectorType::get(DoubleTy, 2)); 5069 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 5070 "vgetq_lane"); 5071 case NEON::BI__builtin_neon_vaddd_s64: 5072 case NEON::BI__builtin_neon_vaddd_u64: 5073 return Builder.CreateAdd(Ops[0], EmitScalarExpr(E->getArg(1)), "vaddd"); 5074 case NEON::BI__builtin_neon_vsubd_s64: 5075 case NEON::BI__builtin_neon_vsubd_u64: 5076 return Builder.CreateSub(Ops[0], EmitScalarExpr(E->getArg(1)), "vsubd"); 5077 case NEON::BI__builtin_neon_vqdmlalh_s16: 5078 case NEON::BI__builtin_neon_vqdmlslh_s16: { 5079 SmallVector<Value *, 2> ProductOps; 5080 ProductOps.push_back(vectorWrapScalar16(Ops[1])); 5081 ProductOps.push_back(vectorWrapScalar16(EmitScalarExpr(E->getArg(2)))); 5082 llvm::Type *VTy = llvm::VectorType::get(Int32Ty, 4); 5083 Ops[1] = EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_sqdmull, VTy), 5084 ProductOps, "vqdmlXl"); 5085 Constant *CI = ConstantInt::get(SizeTy, 0); 5086 Ops[1] = Builder.CreateExtractElement(Ops[1], CI, "lane0"); 5087 5088 unsigned AccumInt = BuiltinID == NEON::BI__builtin_neon_vqdmlalh_s16 5089 ? Intrinsic::aarch64_neon_sqadd 5090 : Intrinsic::aarch64_neon_sqsub; 5091 return EmitNeonCall(CGM.getIntrinsic(AccumInt, Int32Ty), Ops, "vqdmlXl"); 5092 } 5093 case NEON::BI__builtin_neon_vqshlud_n_s64: { 5094 Ops.push_back(EmitScalarExpr(E->getArg(1))); 5095 Ops[1] = Builder.CreateZExt(Ops[1], Int64Ty); 5096 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_sqshlu, Int64Ty), 5097 Ops, "vqshlu_n"); 5098 } 5099 case NEON::BI__builtin_neon_vqshld_n_u64: 5100 case NEON::BI__builtin_neon_vqshld_n_s64: { 5101 unsigned Int = BuiltinID == NEON::BI__builtin_neon_vqshld_n_u64 5102 ? Intrinsic::aarch64_neon_uqshl 5103 : Intrinsic::aarch64_neon_sqshl; 5104 Ops.push_back(EmitScalarExpr(E->getArg(1))); 5105 Ops[1] = Builder.CreateZExt(Ops[1], Int64Ty); 5106 return EmitNeonCall(CGM.getIntrinsic(Int, Int64Ty), Ops, "vqshl_n"); 5107 } 5108 case NEON::BI__builtin_neon_vrshrd_n_u64: 5109 case NEON::BI__builtin_neon_vrshrd_n_s64: { 5110 unsigned Int = BuiltinID == NEON::BI__builtin_neon_vrshrd_n_u64 5111 ? Intrinsic::aarch64_neon_urshl 5112 : Intrinsic::aarch64_neon_srshl; 5113 Ops.push_back(EmitScalarExpr(E->getArg(1))); 5114 int SV = cast<ConstantInt>(Ops[1])->getSExtValue(); 5115 Ops[1] = ConstantInt::get(Int64Ty, -SV); 5116 return EmitNeonCall(CGM.getIntrinsic(Int, Int64Ty), Ops, "vrshr_n"); 5117 } 5118 case NEON::BI__builtin_neon_vrsrad_n_u64: 5119 case NEON::BI__builtin_neon_vrsrad_n_s64: { 5120 unsigned Int = BuiltinID == NEON::BI__builtin_neon_vrsrad_n_u64 5121 ? Intrinsic::aarch64_neon_urshl 5122 : Intrinsic::aarch64_neon_srshl; 5123 Ops[1] = Builder.CreateBitCast(Ops[1], Int64Ty); 5124 Ops.push_back(Builder.CreateNeg(EmitScalarExpr(E->getArg(2)))); 5125 Ops[1] = Builder.CreateCall(CGM.getIntrinsic(Int, Int64Ty), 5126 {Ops[1], Builder.CreateSExt(Ops[2], Int64Ty)}); 5127 return Builder.CreateAdd(Ops[0], Builder.CreateBitCast(Ops[1], Int64Ty)); 5128 } 5129 case NEON::BI__builtin_neon_vshld_n_s64: 5130 case NEON::BI__builtin_neon_vshld_n_u64: { 5131 llvm::ConstantInt *Amt = cast<ConstantInt>(EmitScalarExpr(E->getArg(1))); 5132 return Builder.CreateShl( 5133 Ops[0], ConstantInt::get(Int64Ty, Amt->getZExtValue()), "shld_n"); 5134 } 5135 case NEON::BI__builtin_neon_vshrd_n_s64: { 5136 llvm::ConstantInt *Amt = cast<ConstantInt>(EmitScalarExpr(E->getArg(1))); 5137 return Builder.CreateAShr( 5138 Ops[0], ConstantInt::get(Int64Ty, std::min(static_cast<uint64_t>(63), 5139 Amt->getZExtValue())), 5140 "shrd_n"); 5141 } 5142 case NEON::BI__builtin_neon_vshrd_n_u64: { 5143 llvm::ConstantInt *Amt = cast<ConstantInt>(EmitScalarExpr(E->getArg(1))); 5144 uint64_t ShiftAmt = Amt->getZExtValue(); 5145 // Right-shifting an unsigned value by its size yields 0. 5146 if (ShiftAmt == 64) 5147 return ConstantInt::get(Int64Ty, 0); 5148 return Builder.CreateLShr(Ops[0], ConstantInt::get(Int64Ty, ShiftAmt), 5149 "shrd_n"); 5150 } 5151 case NEON::BI__builtin_neon_vsrad_n_s64: { 5152 llvm::ConstantInt *Amt = cast<ConstantInt>(EmitScalarExpr(E->getArg(2))); 5153 Ops[1] = Builder.CreateAShr( 5154 Ops[1], ConstantInt::get(Int64Ty, std::min(static_cast<uint64_t>(63), 5155 Amt->getZExtValue())), 5156 "shrd_n"); 5157 return Builder.CreateAdd(Ops[0], Ops[1]); 5158 } 5159 case NEON::BI__builtin_neon_vsrad_n_u64: { 5160 llvm::ConstantInt *Amt = cast<ConstantInt>(EmitScalarExpr(E->getArg(2))); 5161 uint64_t ShiftAmt = Amt->getZExtValue(); 5162 // Right-shifting an unsigned value by its size yields 0. 5163 // As Op + 0 = Op, return Ops[0] directly. 5164 if (ShiftAmt == 64) 5165 return Ops[0]; 5166 Ops[1] = Builder.CreateLShr(Ops[1], ConstantInt::get(Int64Ty, ShiftAmt), 5167 "shrd_n"); 5168 return Builder.CreateAdd(Ops[0], Ops[1]); 5169 } 5170 case NEON::BI__builtin_neon_vqdmlalh_lane_s16: 5171 case NEON::BI__builtin_neon_vqdmlalh_laneq_s16: 5172 case NEON::BI__builtin_neon_vqdmlslh_lane_s16: 5173 case NEON::BI__builtin_neon_vqdmlslh_laneq_s16: { 5174 Ops[2] = Builder.CreateExtractElement(Ops[2], EmitScalarExpr(E->getArg(3)), 5175 "lane"); 5176 SmallVector<Value *, 2> ProductOps; 5177 ProductOps.push_back(vectorWrapScalar16(Ops[1])); 5178 ProductOps.push_back(vectorWrapScalar16(Ops[2])); 5179 llvm::Type *VTy = llvm::VectorType::get(Int32Ty, 4); 5180 Ops[1] = EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_sqdmull, VTy), 5181 ProductOps, "vqdmlXl"); 5182 Constant *CI = ConstantInt::get(SizeTy, 0); 5183 Ops[1] = Builder.CreateExtractElement(Ops[1], CI, "lane0"); 5184 Ops.pop_back(); 5185 5186 unsigned AccInt = (BuiltinID == NEON::BI__builtin_neon_vqdmlalh_lane_s16 || 5187 BuiltinID == NEON::BI__builtin_neon_vqdmlalh_laneq_s16) 5188 ? Intrinsic::aarch64_neon_sqadd 5189 : Intrinsic::aarch64_neon_sqsub; 5190 return EmitNeonCall(CGM.getIntrinsic(AccInt, Int32Ty), Ops, "vqdmlXl"); 5191 } 5192 case NEON::BI__builtin_neon_vqdmlals_s32: 5193 case NEON::BI__builtin_neon_vqdmlsls_s32: { 5194 SmallVector<Value *, 2> ProductOps; 5195 ProductOps.push_back(Ops[1]); 5196 ProductOps.push_back(EmitScalarExpr(E->getArg(2))); 5197 Ops[1] = 5198 EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_sqdmulls_scalar), 5199 ProductOps, "vqdmlXl"); 5200 5201 unsigned AccumInt = BuiltinID == NEON::BI__builtin_neon_vqdmlals_s32 5202 ? Intrinsic::aarch64_neon_sqadd 5203 : Intrinsic::aarch64_neon_sqsub; 5204 return EmitNeonCall(CGM.getIntrinsic(AccumInt, Int64Ty), Ops, "vqdmlXl"); 5205 } 5206 case NEON::BI__builtin_neon_vqdmlals_lane_s32: 5207 case NEON::BI__builtin_neon_vqdmlals_laneq_s32: 5208 case NEON::BI__builtin_neon_vqdmlsls_lane_s32: 5209 case NEON::BI__builtin_neon_vqdmlsls_laneq_s32: { 5210 Ops[2] = Builder.CreateExtractElement(Ops[2], EmitScalarExpr(E->getArg(3)), 5211 "lane"); 5212 SmallVector<Value *, 2> ProductOps; 5213 ProductOps.push_back(Ops[1]); 5214 ProductOps.push_back(Ops[2]); 5215 Ops[1] = 5216 EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_sqdmulls_scalar), 5217 ProductOps, "vqdmlXl"); 5218 Ops.pop_back(); 5219 5220 unsigned AccInt = (BuiltinID == NEON::BI__builtin_neon_vqdmlals_lane_s32 || 5221 BuiltinID == NEON::BI__builtin_neon_vqdmlals_laneq_s32) 5222 ? Intrinsic::aarch64_neon_sqadd 5223 : Intrinsic::aarch64_neon_sqsub; 5224 return EmitNeonCall(CGM.getIntrinsic(AccInt, Int64Ty), Ops, "vqdmlXl"); 5225 } 5226 } 5227 5228 llvm::VectorType *VTy = GetNeonType(this, Type); 5229 llvm::Type *Ty = VTy; 5230 if (!Ty) 5231 return nullptr; 5232 5233 // Not all intrinsics handled by the common case work for AArch64 yet, so only 5234 // defer to common code if it's been added to our special map. 5235 Builtin = findNeonIntrinsicInMap(AArch64SIMDIntrinsicMap, BuiltinID, 5236 AArch64SIMDIntrinsicsProvenSorted); 5237 5238 if (Builtin) 5239 return EmitCommonNeonBuiltinExpr( 5240 Builtin->BuiltinID, Builtin->LLVMIntrinsic, Builtin->AltLLVMIntrinsic, 5241 Builtin->NameHint, Builtin->TypeModifier, E, Ops, 5242 /*never use addresses*/ Address::invalid(), Address::invalid()); 5243 5244 if (Value *V = EmitAArch64TblBuiltinExpr(*this, BuiltinID, E, Ops)) 5245 return V; 5246 5247 unsigned Int; 5248 switch (BuiltinID) { 5249 default: return nullptr; 5250 case NEON::BI__builtin_neon_vbsl_v: 5251 case NEON::BI__builtin_neon_vbslq_v: { 5252 llvm::Type *BitTy = llvm::VectorType::getInteger(VTy); 5253 Ops[0] = Builder.CreateBitCast(Ops[0], BitTy, "vbsl"); 5254 Ops[1] = Builder.CreateBitCast(Ops[1], BitTy, "vbsl"); 5255 Ops[2] = Builder.CreateBitCast(Ops[2], BitTy, "vbsl"); 5256 5257 Ops[1] = Builder.CreateAnd(Ops[0], Ops[1], "vbsl"); 5258 Ops[2] = Builder.CreateAnd(Builder.CreateNot(Ops[0]), Ops[2], "vbsl"); 5259 Ops[0] = Builder.CreateOr(Ops[1], Ops[2], "vbsl"); 5260 return Builder.CreateBitCast(Ops[0], Ty); 5261 } 5262 case NEON::BI__builtin_neon_vfma_lane_v: 5263 case NEON::BI__builtin_neon_vfmaq_lane_v: { // Only used for FP types 5264 // The ARM builtins (and instructions) have the addend as the first 5265 // operand, but the 'fma' intrinsics have it last. Swap it around here. 5266 Value *Addend = Ops[0]; 5267 Value *Multiplicand = Ops[1]; 5268 Value *LaneSource = Ops[2]; 5269 Ops[0] = Multiplicand; 5270 Ops[1] = LaneSource; 5271 Ops[2] = Addend; 5272 5273 // Now adjust things to handle the lane access. 5274 llvm::Type *SourceTy = BuiltinID == NEON::BI__builtin_neon_vfmaq_lane_v ? 5275 llvm::VectorType::get(VTy->getElementType(), VTy->getNumElements() / 2) : 5276 VTy; 5277 llvm::Constant *cst = cast<Constant>(Ops[3]); 5278 Value *SV = llvm::ConstantVector::getSplat(VTy->getNumElements(), cst); 5279 Ops[1] = Builder.CreateBitCast(Ops[1], SourceTy); 5280 Ops[1] = Builder.CreateShuffleVector(Ops[1], Ops[1], SV, "lane"); 5281 5282 Ops.pop_back(); 5283 Int = Intrinsic::fma; 5284 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "fmla"); 5285 } 5286 case NEON::BI__builtin_neon_vfma_laneq_v: { 5287 llvm::VectorType *VTy = cast<llvm::VectorType>(Ty); 5288 // v1f64 fma should be mapped to Neon scalar f64 fma 5289 if (VTy && VTy->getElementType() == DoubleTy) { 5290 Ops[0] = Builder.CreateBitCast(Ops[0], DoubleTy); 5291 Ops[1] = Builder.CreateBitCast(Ops[1], DoubleTy); 5292 llvm::Type *VTy = GetNeonType(this, 5293 NeonTypeFlags(NeonTypeFlags::Float64, false, true)); 5294 Ops[2] = Builder.CreateBitCast(Ops[2], VTy); 5295 Ops[2] = Builder.CreateExtractElement(Ops[2], Ops[3], "extract"); 5296 Value *F = CGM.getIntrinsic(Intrinsic::fma, DoubleTy); 5297 Value *Result = Builder.CreateCall(F, {Ops[1], Ops[2], Ops[0]}); 5298 return Builder.CreateBitCast(Result, Ty); 5299 } 5300 Value *F = CGM.getIntrinsic(Intrinsic::fma, Ty); 5301 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 5302 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 5303 5304 llvm::Type *STy = llvm::VectorType::get(VTy->getElementType(), 5305 VTy->getNumElements() * 2); 5306 Ops[2] = Builder.CreateBitCast(Ops[2], STy); 5307 Value* SV = llvm::ConstantVector::getSplat(VTy->getNumElements(), 5308 cast<ConstantInt>(Ops[3])); 5309 Ops[2] = Builder.CreateShuffleVector(Ops[2], Ops[2], SV, "lane"); 5310 5311 return Builder.CreateCall(F, {Ops[2], Ops[1], Ops[0]}); 5312 } 5313 case NEON::BI__builtin_neon_vfmaq_laneq_v: { 5314 Value *F = CGM.getIntrinsic(Intrinsic::fma, Ty); 5315 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 5316 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 5317 5318 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 5319 Ops[2] = EmitNeonSplat(Ops[2], cast<ConstantInt>(Ops[3])); 5320 return Builder.CreateCall(F, {Ops[2], Ops[1], Ops[0]}); 5321 } 5322 case NEON::BI__builtin_neon_vfmas_lane_f32: 5323 case NEON::BI__builtin_neon_vfmas_laneq_f32: 5324 case NEON::BI__builtin_neon_vfmad_lane_f64: 5325 case NEON::BI__builtin_neon_vfmad_laneq_f64: { 5326 Ops.push_back(EmitScalarExpr(E->getArg(3))); 5327 llvm::Type *Ty = ConvertType(E->getCallReturnType(getContext())); 5328 Value *F = CGM.getIntrinsic(Intrinsic::fma, Ty); 5329 Ops[2] = Builder.CreateExtractElement(Ops[2], Ops[3], "extract"); 5330 return Builder.CreateCall(F, {Ops[1], Ops[2], Ops[0]}); 5331 } 5332 case NEON::BI__builtin_neon_vfms_v: 5333 case NEON::BI__builtin_neon_vfmsq_v: { // Only used for FP types 5334 // FIXME: probably remove when we no longer support aarch64_simd.h 5335 // (arm_neon.h delegates to vfma). 5336 5337 // The ARM builtins (and instructions) have the addend as the first 5338 // operand, but the 'fma' intrinsics have it last. Swap it around here. 5339 Value *Subtrahend = Ops[0]; 5340 Value *Multiplicand = Ops[2]; 5341 Ops[0] = Multiplicand; 5342 Ops[2] = Subtrahend; 5343 Ops[1] = Builder.CreateBitCast(Ops[1], VTy); 5344 Ops[1] = Builder.CreateFNeg(Ops[1]); 5345 Int = Intrinsic::fma; 5346 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "fmls"); 5347 } 5348 case NEON::BI__builtin_neon_vmull_v: 5349 // FIXME: improve sharing scheme to cope with 3 alternative LLVM intrinsics. 5350 Int = usgn ? Intrinsic::aarch64_neon_umull : Intrinsic::aarch64_neon_smull; 5351 if (Type.isPoly()) Int = Intrinsic::aarch64_neon_pmull; 5352 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmull"); 5353 case NEON::BI__builtin_neon_vmax_v: 5354 case NEON::BI__builtin_neon_vmaxq_v: 5355 // FIXME: improve sharing scheme to cope with 3 alternative LLVM intrinsics. 5356 Int = usgn ? Intrinsic::aarch64_neon_umax : Intrinsic::aarch64_neon_smax; 5357 if (Ty->isFPOrFPVectorTy()) Int = Intrinsic::aarch64_neon_fmax; 5358 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmax"); 5359 case NEON::BI__builtin_neon_vmin_v: 5360 case NEON::BI__builtin_neon_vminq_v: 5361 // FIXME: improve sharing scheme to cope with 3 alternative LLVM intrinsics. 5362 Int = usgn ? Intrinsic::aarch64_neon_umin : Intrinsic::aarch64_neon_smin; 5363 if (Ty->isFPOrFPVectorTy()) Int = Intrinsic::aarch64_neon_fmin; 5364 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmin"); 5365 case NEON::BI__builtin_neon_vabd_v: 5366 case NEON::BI__builtin_neon_vabdq_v: 5367 // FIXME: improve sharing scheme to cope with 3 alternative LLVM intrinsics. 5368 Int = usgn ? Intrinsic::aarch64_neon_uabd : Intrinsic::aarch64_neon_sabd; 5369 if (Ty->isFPOrFPVectorTy()) Int = Intrinsic::aarch64_neon_fabd; 5370 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vabd"); 5371 case NEON::BI__builtin_neon_vpadal_v: 5372 case NEON::BI__builtin_neon_vpadalq_v: { 5373 unsigned ArgElts = VTy->getNumElements(); 5374 llvm::IntegerType *EltTy = cast<IntegerType>(VTy->getElementType()); 5375 unsigned BitWidth = EltTy->getBitWidth(); 5376 llvm::Type *ArgTy = llvm::VectorType::get( 5377 llvm::IntegerType::get(getLLVMContext(), BitWidth/2), 2*ArgElts); 5378 llvm::Type* Tys[2] = { VTy, ArgTy }; 5379 Int = usgn ? Intrinsic::aarch64_neon_uaddlp : Intrinsic::aarch64_neon_saddlp; 5380 SmallVector<llvm::Value*, 1> TmpOps; 5381 TmpOps.push_back(Ops[1]); 5382 Function *F = CGM.getIntrinsic(Int, Tys); 5383 llvm::Value *tmp = EmitNeonCall(F, TmpOps, "vpadal"); 5384 llvm::Value *addend = Builder.CreateBitCast(Ops[0], tmp->getType()); 5385 return Builder.CreateAdd(tmp, addend); 5386 } 5387 case NEON::BI__builtin_neon_vpmin_v: 5388 case NEON::BI__builtin_neon_vpminq_v: 5389 // FIXME: improve sharing scheme to cope with 3 alternative LLVM intrinsics. 5390 Int = usgn ? Intrinsic::aarch64_neon_uminp : Intrinsic::aarch64_neon_sminp; 5391 if (Ty->isFPOrFPVectorTy()) Int = Intrinsic::aarch64_neon_fminp; 5392 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vpmin"); 5393 case NEON::BI__builtin_neon_vpmax_v: 5394 case NEON::BI__builtin_neon_vpmaxq_v: 5395 // FIXME: improve sharing scheme to cope with 3 alternative LLVM intrinsics. 5396 Int = usgn ? Intrinsic::aarch64_neon_umaxp : Intrinsic::aarch64_neon_smaxp; 5397 if (Ty->isFPOrFPVectorTy()) Int = Intrinsic::aarch64_neon_fmaxp; 5398 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vpmax"); 5399 case NEON::BI__builtin_neon_vminnm_v: 5400 case NEON::BI__builtin_neon_vminnmq_v: 5401 Int = Intrinsic::aarch64_neon_fminnm; 5402 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vminnm"); 5403 case NEON::BI__builtin_neon_vmaxnm_v: 5404 case NEON::BI__builtin_neon_vmaxnmq_v: 5405 Int = Intrinsic::aarch64_neon_fmaxnm; 5406 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmaxnm"); 5407 case NEON::BI__builtin_neon_vrecpss_f32: { 5408 Ops.push_back(EmitScalarExpr(E->getArg(1))); 5409 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_frecps, FloatTy), 5410 Ops, "vrecps"); 5411 } 5412 case NEON::BI__builtin_neon_vrecpsd_f64: { 5413 Ops.push_back(EmitScalarExpr(E->getArg(1))); 5414 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_frecps, DoubleTy), 5415 Ops, "vrecps"); 5416 } 5417 case NEON::BI__builtin_neon_vqshrun_n_v: 5418 Int = Intrinsic::aarch64_neon_sqshrun; 5419 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqshrun_n"); 5420 case NEON::BI__builtin_neon_vqrshrun_n_v: 5421 Int = Intrinsic::aarch64_neon_sqrshrun; 5422 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqrshrun_n"); 5423 case NEON::BI__builtin_neon_vqshrn_n_v: 5424 Int = usgn ? Intrinsic::aarch64_neon_uqshrn : Intrinsic::aarch64_neon_sqshrn; 5425 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqshrn_n"); 5426 case NEON::BI__builtin_neon_vrshrn_n_v: 5427 Int = Intrinsic::aarch64_neon_rshrn; 5428 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrshrn_n"); 5429 case NEON::BI__builtin_neon_vqrshrn_n_v: 5430 Int = usgn ? Intrinsic::aarch64_neon_uqrshrn : Intrinsic::aarch64_neon_sqrshrn; 5431 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqrshrn_n"); 5432 case NEON::BI__builtin_neon_vrnda_v: 5433 case NEON::BI__builtin_neon_vrndaq_v: { 5434 Int = Intrinsic::round; 5435 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrnda"); 5436 } 5437 case NEON::BI__builtin_neon_vrndi_v: 5438 case NEON::BI__builtin_neon_vrndiq_v: { 5439 Int = Intrinsic::nearbyint; 5440 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrndi"); 5441 } 5442 case NEON::BI__builtin_neon_vrndm_v: 5443 case NEON::BI__builtin_neon_vrndmq_v: { 5444 Int = Intrinsic::floor; 5445 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrndm"); 5446 } 5447 case NEON::BI__builtin_neon_vrndn_v: 5448 case NEON::BI__builtin_neon_vrndnq_v: { 5449 Int = Intrinsic::aarch64_neon_frintn; 5450 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrndn"); 5451 } 5452 case NEON::BI__builtin_neon_vrndp_v: 5453 case NEON::BI__builtin_neon_vrndpq_v: { 5454 Int = Intrinsic::ceil; 5455 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrndp"); 5456 } 5457 case NEON::BI__builtin_neon_vrndx_v: 5458 case NEON::BI__builtin_neon_vrndxq_v: { 5459 Int = Intrinsic::rint; 5460 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrndx"); 5461 } 5462 case NEON::BI__builtin_neon_vrnd_v: 5463 case NEON::BI__builtin_neon_vrndq_v: { 5464 Int = Intrinsic::trunc; 5465 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrndz"); 5466 } 5467 case NEON::BI__builtin_neon_vceqz_v: 5468 case NEON::BI__builtin_neon_vceqzq_v: 5469 return EmitAArch64CompareBuiltinExpr(Ops[0], Ty, ICmpInst::FCMP_OEQ, 5470 ICmpInst::ICMP_EQ, "vceqz"); 5471 case NEON::BI__builtin_neon_vcgez_v: 5472 case NEON::BI__builtin_neon_vcgezq_v: 5473 return EmitAArch64CompareBuiltinExpr(Ops[0], Ty, ICmpInst::FCMP_OGE, 5474 ICmpInst::ICMP_SGE, "vcgez"); 5475 case NEON::BI__builtin_neon_vclez_v: 5476 case NEON::BI__builtin_neon_vclezq_v: 5477 return EmitAArch64CompareBuiltinExpr(Ops[0], Ty, ICmpInst::FCMP_OLE, 5478 ICmpInst::ICMP_SLE, "vclez"); 5479 case NEON::BI__builtin_neon_vcgtz_v: 5480 case NEON::BI__builtin_neon_vcgtzq_v: 5481 return EmitAArch64CompareBuiltinExpr(Ops[0], Ty, ICmpInst::FCMP_OGT, 5482 ICmpInst::ICMP_SGT, "vcgtz"); 5483 case NEON::BI__builtin_neon_vcltz_v: 5484 case NEON::BI__builtin_neon_vcltzq_v: 5485 return EmitAArch64CompareBuiltinExpr(Ops[0], Ty, ICmpInst::FCMP_OLT, 5486 ICmpInst::ICMP_SLT, "vcltz"); 5487 case NEON::BI__builtin_neon_vcvt_f64_v: 5488 case NEON::BI__builtin_neon_vcvtq_f64_v: 5489 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 5490 Ty = GetNeonType(this, NeonTypeFlags(NeonTypeFlags::Float64, false, quad)); 5491 return usgn ? Builder.CreateUIToFP(Ops[0], Ty, "vcvt") 5492 : Builder.CreateSIToFP(Ops[0], Ty, "vcvt"); 5493 case NEON::BI__builtin_neon_vcvt_f64_f32: { 5494 assert(Type.getEltType() == NeonTypeFlags::Float64 && quad && 5495 "unexpected vcvt_f64_f32 builtin"); 5496 NeonTypeFlags SrcFlag = NeonTypeFlags(NeonTypeFlags::Float32, false, false); 5497 Ops[0] = Builder.CreateBitCast(Ops[0], GetNeonType(this, SrcFlag)); 5498 5499 return Builder.CreateFPExt(Ops[0], Ty, "vcvt"); 5500 } 5501 case NEON::BI__builtin_neon_vcvt_f32_f64: { 5502 assert(Type.getEltType() == NeonTypeFlags::Float32 && 5503 "unexpected vcvt_f32_f64 builtin"); 5504 NeonTypeFlags SrcFlag = NeonTypeFlags(NeonTypeFlags::Float64, false, true); 5505 Ops[0] = Builder.CreateBitCast(Ops[0], GetNeonType(this, SrcFlag)); 5506 5507 return Builder.CreateFPTrunc(Ops[0], Ty, "vcvt"); 5508 } 5509 case NEON::BI__builtin_neon_vcvt_s32_v: 5510 case NEON::BI__builtin_neon_vcvt_u32_v: 5511 case NEON::BI__builtin_neon_vcvt_s64_v: 5512 case NEON::BI__builtin_neon_vcvt_u64_v: 5513 case NEON::BI__builtin_neon_vcvtq_s32_v: 5514 case NEON::BI__builtin_neon_vcvtq_u32_v: 5515 case NEON::BI__builtin_neon_vcvtq_s64_v: 5516 case NEON::BI__builtin_neon_vcvtq_u64_v: { 5517 Ops[0] = Builder.CreateBitCast(Ops[0], GetFloatNeonType(this, Type)); 5518 if (usgn) 5519 return Builder.CreateFPToUI(Ops[0], Ty); 5520 return Builder.CreateFPToSI(Ops[0], Ty); 5521 } 5522 case NEON::BI__builtin_neon_vcvta_s32_v: 5523 case NEON::BI__builtin_neon_vcvtaq_s32_v: 5524 case NEON::BI__builtin_neon_vcvta_u32_v: 5525 case NEON::BI__builtin_neon_vcvtaq_u32_v: 5526 case NEON::BI__builtin_neon_vcvta_s64_v: 5527 case NEON::BI__builtin_neon_vcvtaq_s64_v: 5528 case NEON::BI__builtin_neon_vcvta_u64_v: 5529 case NEON::BI__builtin_neon_vcvtaq_u64_v: { 5530 Int = usgn ? Intrinsic::aarch64_neon_fcvtau : Intrinsic::aarch64_neon_fcvtas; 5531 llvm::Type *Tys[2] = { Ty, GetFloatNeonType(this, Type) }; 5532 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vcvta"); 5533 } 5534 case NEON::BI__builtin_neon_vcvtm_s32_v: 5535 case NEON::BI__builtin_neon_vcvtmq_s32_v: 5536 case NEON::BI__builtin_neon_vcvtm_u32_v: 5537 case NEON::BI__builtin_neon_vcvtmq_u32_v: 5538 case NEON::BI__builtin_neon_vcvtm_s64_v: 5539 case NEON::BI__builtin_neon_vcvtmq_s64_v: 5540 case NEON::BI__builtin_neon_vcvtm_u64_v: 5541 case NEON::BI__builtin_neon_vcvtmq_u64_v: { 5542 Int = usgn ? Intrinsic::aarch64_neon_fcvtmu : Intrinsic::aarch64_neon_fcvtms; 5543 llvm::Type *Tys[2] = { Ty, GetFloatNeonType(this, Type) }; 5544 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vcvtm"); 5545 } 5546 case NEON::BI__builtin_neon_vcvtn_s32_v: 5547 case NEON::BI__builtin_neon_vcvtnq_s32_v: 5548 case NEON::BI__builtin_neon_vcvtn_u32_v: 5549 case NEON::BI__builtin_neon_vcvtnq_u32_v: 5550 case NEON::BI__builtin_neon_vcvtn_s64_v: 5551 case NEON::BI__builtin_neon_vcvtnq_s64_v: 5552 case NEON::BI__builtin_neon_vcvtn_u64_v: 5553 case NEON::BI__builtin_neon_vcvtnq_u64_v: { 5554 Int = usgn ? Intrinsic::aarch64_neon_fcvtnu : Intrinsic::aarch64_neon_fcvtns; 5555 llvm::Type *Tys[2] = { Ty, GetFloatNeonType(this, Type) }; 5556 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vcvtn"); 5557 } 5558 case NEON::BI__builtin_neon_vcvtp_s32_v: 5559 case NEON::BI__builtin_neon_vcvtpq_s32_v: 5560 case NEON::BI__builtin_neon_vcvtp_u32_v: 5561 case NEON::BI__builtin_neon_vcvtpq_u32_v: 5562 case NEON::BI__builtin_neon_vcvtp_s64_v: 5563 case NEON::BI__builtin_neon_vcvtpq_s64_v: 5564 case NEON::BI__builtin_neon_vcvtp_u64_v: 5565 case NEON::BI__builtin_neon_vcvtpq_u64_v: { 5566 Int = usgn ? Intrinsic::aarch64_neon_fcvtpu : Intrinsic::aarch64_neon_fcvtps; 5567 llvm::Type *Tys[2] = { Ty, GetFloatNeonType(this, Type) }; 5568 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vcvtp"); 5569 } 5570 case NEON::BI__builtin_neon_vmulx_v: 5571 case NEON::BI__builtin_neon_vmulxq_v: { 5572 Int = Intrinsic::aarch64_neon_fmulx; 5573 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmulx"); 5574 } 5575 case NEON::BI__builtin_neon_vmul_lane_v: 5576 case NEON::BI__builtin_neon_vmul_laneq_v: { 5577 // v1f64 vmul_lane should be mapped to Neon scalar mul lane 5578 bool Quad = false; 5579 if (BuiltinID == NEON::BI__builtin_neon_vmul_laneq_v) 5580 Quad = true; 5581 Ops[0] = Builder.CreateBitCast(Ops[0], DoubleTy); 5582 llvm::Type *VTy = GetNeonType(this, 5583 NeonTypeFlags(NeonTypeFlags::Float64, false, Quad)); 5584 Ops[1] = Builder.CreateBitCast(Ops[1], VTy); 5585 Ops[1] = Builder.CreateExtractElement(Ops[1], Ops[2], "extract"); 5586 Value *Result = Builder.CreateFMul(Ops[0], Ops[1]); 5587 return Builder.CreateBitCast(Result, Ty); 5588 } 5589 case NEON::BI__builtin_neon_vnegd_s64: 5590 return Builder.CreateNeg(EmitScalarExpr(E->getArg(0)), "vnegd"); 5591 case NEON::BI__builtin_neon_vpmaxnm_v: 5592 case NEON::BI__builtin_neon_vpmaxnmq_v: { 5593 Int = Intrinsic::aarch64_neon_fmaxnmp; 5594 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vpmaxnm"); 5595 } 5596 case NEON::BI__builtin_neon_vpminnm_v: 5597 case NEON::BI__builtin_neon_vpminnmq_v: { 5598 Int = Intrinsic::aarch64_neon_fminnmp; 5599 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vpminnm"); 5600 } 5601 case NEON::BI__builtin_neon_vsqrt_v: 5602 case NEON::BI__builtin_neon_vsqrtq_v: { 5603 Int = Intrinsic::sqrt; 5604 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 5605 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vsqrt"); 5606 } 5607 case NEON::BI__builtin_neon_vrbit_v: 5608 case NEON::BI__builtin_neon_vrbitq_v: { 5609 Int = Intrinsic::aarch64_neon_rbit; 5610 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrbit"); 5611 } 5612 case NEON::BI__builtin_neon_vaddv_u8: 5613 // FIXME: These are handled by the AArch64 scalar code. 5614 usgn = true; 5615 // FALLTHROUGH 5616 case NEON::BI__builtin_neon_vaddv_s8: { 5617 Int = usgn ? Intrinsic::aarch64_neon_uaddv : Intrinsic::aarch64_neon_saddv; 5618 Ty = Int32Ty; 5619 VTy = llvm::VectorType::get(Int8Ty, 8); 5620 llvm::Type *Tys[2] = { Ty, VTy }; 5621 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5622 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddv"); 5623 return Builder.CreateTrunc(Ops[0], Int8Ty); 5624 } 5625 case NEON::BI__builtin_neon_vaddv_u16: 5626 usgn = true; 5627 // FALLTHROUGH 5628 case NEON::BI__builtin_neon_vaddv_s16: { 5629 Int = usgn ? Intrinsic::aarch64_neon_uaddv : Intrinsic::aarch64_neon_saddv; 5630 Ty = Int32Ty; 5631 VTy = llvm::VectorType::get(Int16Ty, 4); 5632 llvm::Type *Tys[2] = { Ty, VTy }; 5633 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5634 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddv"); 5635 return Builder.CreateTrunc(Ops[0], Int16Ty); 5636 } 5637 case NEON::BI__builtin_neon_vaddvq_u8: 5638 usgn = true; 5639 // FALLTHROUGH 5640 case NEON::BI__builtin_neon_vaddvq_s8: { 5641 Int = usgn ? Intrinsic::aarch64_neon_uaddv : Intrinsic::aarch64_neon_saddv; 5642 Ty = Int32Ty; 5643 VTy = llvm::VectorType::get(Int8Ty, 16); 5644 llvm::Type *Tys[2] = { Ty, VTy }; 5645 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5646 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddv"); 5647 return Builder.CreateTrunc(Ops[0], Int8Ty); 5648 } 5649 case NEON::BI__builtin_neon_vaddvq_u16: 5650 usgn = true; 5651 // FALLTHROUGH 5652 case NEON::BI__builtin_neon_vaddvq_s16: { 5653 Int = usgn ? Intrinsic::aarch64_neon_uaddv : Intrinsic::aarch64_neon_saddv; 5654 Ty = Int32Ty; 5655 VTy = llvm::VectorType::get(Int16Ty, 8); 5656 llvm::Type *Tys[2] = { Ty, VTy }; 5657 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5658 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddv"); 5659 return Builder.CreateTrunc(Ops[0], Int16Ty); 5660 } 5661 case NEON::BI__builtin_neon_vmaxv_u8: { 5662 Int = Intrinsic::aarch64_neon_umaxv; 5663 Ty = Int32Ty; 5664 VTy = llvm::VectorType::get(Int8Ty, 8); 5665 llvm::Type *Tys[2] = { Ty, VTy }; 5666 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5667 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv"); 5668 return Builder.CreateTrunc(Ops[0], Int8Ty); 5669 } 5670 case NEON::BI__builtin_neon_vmaxv_u16: { 5671 Int = Intrinsic::aarch64_neon_umaxv; 5672 Ty = Int32Ty; 5673 VTy = llvm::VectorType::get(Int16Ty, 4); 5674 llvm::Type *Tys[2] = { Ty, VTy }; 5675 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5676 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv"); 5677 return Builder.CreateTrunc(Ops[0], Int16Ty); 5678 } 5679 case NEON::BI__builtin_neon_vmaxvq_u8: { 5680 Int = Intrinsic::aarch64_neon_umaxv; 5681 Ty = Int32Ty; 5682 VTy = llvm::VectorType::get(Int8Ty, 16); 5683 llvm::Type *Tys[2] = { Ty, VTy }; 5684 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5685 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv"); 5686 return Builder.CreateTrunc(Ops[0], Int8Ty); 5687 } 5688 case NEON::BI__builtin_neon_vmaxvq_u16: { 5689 Int = Intrinsic::aarch64_neon_umaxv; 5690 Ty = Int32Ty; 5691 VTy = llvm::VectorType::get(Int16Ty, 8); 5692 llvm::Type *Tys[2] = { Ty, VTy }; 5693 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5694 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv"); 5695 return Builder.CreateTrunc(Ops[0], Int16Ty); 5696 } 5697 case NEON::BI__builtin_neon_vmaxv_s8: { 5698 Int = Intrinsic::aarch64_neon_smaxv; 5699 Ty = Int32Ty; 5700 VTy = llvm::VectorType::get(Int8Ty, 8); 5701 llvm::Type *Tys[2] = { Ty, VTy }; 5702 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5703 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv"); 5704 return Builder.CreateTrunc(Ops[0], Int8Ty); 5705 } 5706 case NEON::BI__builtin_neon_vmaxv_s16: { 5707 Int = Intrinsic::aarch64_neon_smaxv; 5708 Ty = Int32Ty; 5709 VTy = llvm::VectorType::get(Int16Ty, 4); 5710 llvm::Type *Tys[2] = { Ty, VTy }; 5711 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5712 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv"); 5713 return Builder.CreateTrunc(Ops[0], Int16Ty); 5714 } 5715 case NEON::BI__builtin_neon_vmaxvq_s8: { 5716 Int = Intrinsic::aarch64_neon_smaxv; 5717 Ty = Int32Ty; 5718 VTy = llvm::VectorType::get(Int8Ty, 16); 5719 llvm::Type *Tys[2] = { Ty, VTy }; 5720 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5721 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv"); 5722 return Builder.CreateTrunc(Ops[0], Int8Ty); 5723 } 5724 case NEON::BI__builtin_neon_vmaxvq_s16: { 5725 Int = Intrinsic::aarch64_neon_smaxv; 5726 Ty = Int32Ty; 5727 VTy = llvm::VectorType::get(Int16Ty, 8); 5728 llvm::Type *Tys[2] = { Ty, VTy }; 5729 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5730 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv"); 5731 return Builder.CreateTrunc(Ops[0], Int16Ty); 5732 } 5733 case NEON::BI__builtin_neon_vminv_u8: { 5734 Int = Intrinsic::aarch64_neon_uminv; 5735 Ty = Int32Ty; 5736 VTy = llvm::VectorType::get(Int8Ty, 8); 5737 llvm::Type *Tys[2] = { Ty, VTy }; 5738 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5739 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv"); 5740 return Builder.CreateTrunc(Ops[0], Int8Ty); 5741 } 5742 case NEON::BI__builtin_neon_vminv_u16: { 5743 Int = Intrinsic::aarch64_neon_uminv; 5744 Ty = Int32Ty; 5745 VTy = llvm::VectorType::get(Int16Ty, 4); 5746 llvm::Type *Tys[2] = { Ty, VTy }; 5747 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5748 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv"); 5749 return Builder.CreateTrunc(Ops[0], Int16Ty); 5750 } 5751 case NEON::BI__builtin_neon_vminvq_u8: { 5752 Int = Intrinsic::aarch64_neon_uminv; 5753 Ty = Int32Ty; 5754 VTy = llvm::VectorType::get(Int8Ty, 16); 5755 llvm::Type *Tys[2] = { Ty, VTy }; 5756 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5757 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv"); 5758 return Builder.CreateTrunc(Ops[0], Int8Ty); 5759 } 5760 case NEON::BI__builtin_neon_vminvq_u16: { 5761 Int = Intrinsic::aarch64_neon_uminv; 5762 Ty = Int32Ty; 5763 VTy = llvm::VectorType::get(Int16Ty, 8); 5764 llvm::Type *Tys[2] = { Ty, VTy }; 5765 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5766 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv"); 5767 return Builder.CreateTrunc(Ops[0], Int16Ty); 5768 } 5769 case NEON::BI__builtin_neon_vminv_s8: { 5770 Int = Intrinsic::aarch64_neon_sminv; 5771 Ty = Int32Ty; 5772 VTy = llvm::VectorType::get(Int8Ty, 8); 5773 llvm::Type *Tys[2] = { Ty, VTy }; 5774 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5775 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv"); 5776 return Builder.CreateTrunc(Ops[0], Int8Ty); 5777 } 5778 case NEON::BI__builtin_neon_vminv_s16: { 5779 Int = Intrinsic::aarch64_neon_sminv; 5780 Ty = Int32Ty; 5781 VTy = llvm::VectorType::get(Int16Ty, 4); 5782 llvm::Type *Tys[2] = { Ty, VTy }; 5783 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5784 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv"); 5785 return Builder.CreateTrunc(Ops[0], Int16Ty); 5786 } 5787 case NEON::BI__builtin_neon_vminvq_s8: { 5788 Int = Intrinsic::aarch64_neon_sminv; 5789 Ty = Int32Ty; 5790 VTy = llvm::VectorType::get(Int8Ty, 16); 5791 llvm::Type *Tys[2] = { Ty, VTy }; 5792 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5793 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv"); 5794 return Builder.CreateTrunc(Ops[0], Int8Ty); 5795 } 5796 case NEON::BI__builtin_neon_vminvq_s16: { 5797 Int = Intrinsic::aarch64_neon_sminv; 5798 Ty = Int32Ty; 5799 VTy = llvm::VectorType::get(Int16Ty, 8); 5800 llvm::Type *Tys[2] = { Ty, VTy }; 5801 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5802 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv"); 5803 return Builder.CreateTrunc(Ops[0], Int16Ty); 5804 } 5805 case NEON::BI__builtin_neon_vmul_n_f64: { 5806 Ops[0] = Builder.CreateBitCast(Ops[0], DoubleTy); 5807 Value *RHS = Builder.CreateBitCast(EmitScalarExpr(E->getArg(1)), DoubleTy); 5808 return Builder.CreateFMul(Ops[0], RHS); 5809 } 5810 case NEON::BI__builtin_neon_vaddlv_u8: { 5811 Int = Intrinsic::aarch64_neon_uaddlv; 5812 Ty = Int32Ty; 5813 VTy = llvm::VectorType::get(Int8Ty, 8); 5814 llvm::Type *Tys[2] = { Ty, VTy }; 5815 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5816 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv"); 5817 return Builder.CreateTrunc(Ops[0], Int16Ty); 5818 } 5819 case NEON::BI__builtin_neon_vaddlv_u16: { 5820 Int = Intrinsic::aarch64_neon_uaddlv; 5821 Ty = Int32Ty; 5822 VTy = llvm::VectorType::get(Int16Ty, 4); 5823 llvm::Type *Tys[2] = { Ty, VTy }; 5824 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5825 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv"); 5826 } 5827 case NEON::BI__builtin_neon_vaddlvq_u8: { 5828 Int = Intrinsic::aarch64_neon_uaddlv; 5829 Ty = Int32Ty; 5830 VTy = llvm::VectorType::get(Int8Ty, 16); 5831 llvm::Type *Tys[2] = { Ty, VTy }; 5832 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5833 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv"); 5834 return Builder.CreateTrunc(Ops[0], Int16Ty); 5835 } 5836 case NEON::BI__builtin_neon_vaddlvq_u16: { 5837 Int = Intrinsic::aarch64_neon_uaddlv; 5838 Ty = Int32Ty; 5839 VTy = llvm::VectorType::get(Int16Ty, 8); 5840 llvm::Type *Tys[2] = { Ty, VTy }; 5841 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5842 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv"); 5843 } 5844 case NEON::BI__builtin_neon_vaddlv_s8: { 5845 Int = Intrinsic::aarch64_neon_saddlv; 5846 Ty = Int32Ty; 5847 VTy = llvm::VectorType::get(Int8Ty, 8); 5848 llvm::Type *Tys[2] = { Ty, VTy }; 5849 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5850 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv"); 5851 return Builder.CreateTrunc(Ops[0], Int16Ty); 5852 } 5853 case NEON::BI__builtin_neon_vaddlv_s16: { 5854 Int = Intrinsic::aarch64_neon_saddlv; 5855 Ty = Int32Ty; 5856 VTy = llvm::VectorType::get(Int16Ty, 4); 5857 llvm::Type *Tys[2] = { Ty, VTy }; 5858 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5859 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv"); 5860 } 5861 case NEON::BI__builtin_neon_vaddlvq_s8: { 5862 Int = Intrinsic::aarch64_neon_saddlv; 5863 Ty = Int32Ty; 5864 VTy = llvm::VectorType::get(Int8Ty, 16); 5865 llvm::Type *Tys[2] = { Ty, VTy }; 5866 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5867 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv"); 5868 return Builder.CreateTrunc(Ops[0], Int16Ty); 5869 } 5870 case NEON::BI__builtin_neon_vaddlvq_s16: { 5871 Int = Intrinsic::aarch64_neon_saddlv; 5872 Ty = Int32Ty; 5873 VTy = llvm::VectorType::get(Int16Ty, 8); 5874 llvm::Type *Tys[2] = { Ty, VTy }; 5875 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5876 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv"); 5877 } 5878 case NEON::BI__builtin_neon_vsri_n_v: 5879 case NEON::BI__builtin_neon_vsriq_n_v: { 5880 Int = Intrinsic::aarch64_neon_vsri; 5881 llvm::Function *Intrin = CGM.getIntrinsic(Int, Ty); 5882 return EmitNeonCall(Intrin, Ops, "vsri_n"); 5883 } 5884 case NEON::BI__builtin_neon_vsli_n_v: 5885 case NEON::BI__builtin_neon_vsliq_n_v: { 5886 Int = Intrinsic::aarch64_neon_vsli; 5887 llvm::Function *Intrin = CGM.getIntrinsic(Int, Ty); 5888 return EmitNeonCall(Intrin, Ops, "vsli_n"); 5889 } 5890 case NEON::BI__builtin_neon_vsra_n_v: 5891 case NEON::BI__builtin_neon_vsraq_n_v: 5892 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 5893 Ops[1] = EmitNeonRShiftImm(Ops[1], Ops[2], Ty, usgn, "vsra_n"); 5894 return Builder.CreateAdd(Ops[0], Ops[1]); 5895 case NEON::BI__builtin_neon_vrsra_n_v: 5896 case NEON::BI__builtin_neon_vrsraq_n_v: { 5897 Int = usgn ? Intrinsic::aarch64_neon_urshl : Intrinsic::aarch64_neon_srshl; 5898 SmallVector<llvm::Value*,2> TmpOps; 5899 TmpOps.push_back(Ops[1]); 5900 TmpOps.push_back(Ops[2]); 5901 Function* F = CGM.getIntrinsic(Int, Ty); 5902 llvm::Value *tmp = EmitNeonCall(F, TmpOps, "vrshr_n", 1, true); 5903 Ops[0] = Builder.CreateBitCast(Ops[0], VTy); 5904 return Builder.CreateAdd(Ops[0], tmp); 5905 } 5906 // FIXME: Sharing loads & stores with 32-bit is complicated by the absence 5907 // of an Align parameter here. 5908 case NEON::BI__builtin_neon_vld1_x2_v: 5909 case NEON::BI__builtin_neon_vld1q_x2_v: 5910 case NEON::BI__builtin_neon_vld1_x3_v: 5911 case NEON::BI__builtin_neon_vld1q_x3_v: 5912 case NEON::BI__builtin_neon_vld1_x4_v: 5913 case NEON::BI__builtin_neon_vld1q_x4_v: { 5914 llvm::Type *PTy = llvm::PointerType::getUnqual(VTy->getVectorElementType()); 5915 Ops[1] = Builder.CreateBitCast(Ops[1], PTy); 5916 llvm::Type *Tys[2] = { VTy, PTy }; 5917 unsigned Int; 5918 switch (BuiltinID) { 5919 case NEON::BI__builtin_neon_vld1_x2_v: 5920 case NEON::BI__builtin_neon_vld1q_x2_v: 5921 Int = Intrinsic::aarch64_neon_ld1x2; 5922 break; 5923 case NEON::BI__builtin_neon_vld1_x3_v: 5924 case NEON::BI__builtin_neon_vld1q_x3_v: 5925 Int = Intrinsic::aarch64_neon_ld1x3; 5926 break; 5927 case NEON::BI__builtin_neon_vld1_x4_v: 5928 case NEON::BI__builtin_neon_vld1q_x4_v: 5929 Int = Intrinsic::aarch64_neon_ld1x4; 5930 break; 5931 } 5932 Function *F = CGM.getIntrinsic(Int, Tys); 5933 Ops[1] = Builder.CreateCall(F, Ops[1], "vld1xN"); 5934 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 5935 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 5936 return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]); 5937 } 5938 case NEON::BI__builtin_neon_vst1_x2_v: 5939 case NEON::BI__builtin_neon_vst1q_x2_v: 5940 case NEON::BI__builtin_neon_vst1_x3_v: 5941 case NEON::BI__builtin_neon_vst1q_x3_v: 5942 case NEON::BI__builtin_neon_vst1_x4_v: 5943 case NEON::BI__builtin_neon_vst1q_x4_v: { 5944 llvm::Type *PTy = llvm::PointerType::getUnqual(VTy->getVectorElementType()); 5945 llvm::Type *Tys[2] = { VTy, PTy }; 5946 unsigned Int; 5947 switch (BuiltinID) { 5948 case NEON::BI__builtin_neon_vst1_x2_v: 5949 case NEON::BI__builtin_neon_vst1q_x2_v: 5950 Int = Intrinsic::aarch64_neon_st1x2; 5951 break; 5952 case NEON::BI__builtin_neon_vst1_x3_v: 5953 case NEON::BI__builtin_neon_vst1q_x3_v: 5954 Int = Intrinsic::aarch64_neon_st1x3; 5955 break; 5956 case NEON::BI__builtin_neon_vst1_x4_v: 5957 case NEON::BI__builtin_neon_vst1q_x4_v: 5958 Int = Intrinsic::aarch64_neon_st1x4; 5959 break; 5960 } 5961 std::rotate(Ops.begin(), Ops.begin() + 1, Ops.end()); 5962 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, ""); 5963 } 5964 case NEON::BI__builtin_neon_vld1_v: 5965 case NEON::BI__builtin_neon_vld1q_v: 5966 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(VTy)); 5967 return Builder.CreateDefaultAlignedLoad(Ops[0]); 5968 case NEON::BI__builtin_neon_vst1_v: 5969 case NEON::BI__builtin_neon_vst1q_v: 5970 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(VTy)); 5971 Ops[1] = Builder.CreateBitCast(Ops[1], VTy); 5972 return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]); 5973 case NEON::BI__builtin_neon_vld1_lane_v: 5974 case NEON::BI__builtin_neon_vld1q_lane_v: 5975 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 5976 Ty = llvm::PointerType::getUnqual(VTy->getElementType()); 5977 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 5978 Ops[0] = Builder.CreateDefaultAlignedLoad(Ops[0]); 5979 return Builder.CreateInsertElement(Ops[1], Ops[0], Ops[2], "vld1_lane"); 5980 case NEON::BI__builtin_neon_vld1_dup_v: 5981 case NEON::BI__builtin_neon_vld1q_dup_v: { 5982 Value *V = UndefValue::get(Ty); 5983 Ty = llvm::PointerType::getUnqual(VTy->getElementType()); 5984 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 5985 Ops[0] = Builder.CreateDefaultAlignedLoad(Ops[0]); 5986 llvm::Constant *CI = ConstantInt::get(Int32Ty, 0); 5987 Ops[0] = Builder.CreateInsertElement(V, Ops[0], CI); 5988 return EmitNeonSplat(Ops[0], CI); 5989 } 5990 case NEON::BI__builtin_neon_vst1_lane_v: 5991 case NEON::BI__builtin_neon_vst1q_lane_v: 5992 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 5993 Ops[1] = Builder.CreateExtractElement(Ops[1], Ops[2]); 5994 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 5995 return Builder.CreateDefaultAlignedStore(Ops[1], 5996 Builder.CreateBitCast(Ops[0], Ty)); 5997 case NEON::BI__builtin_neon_vld2_v: 5998 case NEON::BI__builtin_neon_vld2q_v: { 5999 llvm::Type *PTy = llvm::PointerType::getUnqual(VTy); 6000 Ops[1] = Builder.CreateBitCast(Ops[1], PTy); 6001 llvm::Type *Tys[2] = { VTy, PTy }; 6002 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld2, Tys); 6003 Ops[1] = Builder.CreateCall(F, Ops[1], "vld2"); 6004 Ops[0] = Builder.CreateBitCast(Ops[0], 6005 llvm::PointerType::getUnqual(Ops[1]->getType())); 6006 return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]); 6007 } 6008 case NEON::BI__builtin_neon_vld3_v: 6009 case NEON::BI__builtin_neon_vld3q_v: { 6010 llvm::Type *PTy = llvm::PointerType::getUnqual(VTy); 6011 Ops[1] = Builder.CreateBitCast(Ops[1], PTy); 6012 llvm::Type *Tys[2] = { VTy, PTy }; 6013 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld3, Tys); 6014 Ops[1] = Builder.CreateCall(F, Ops[1], "vld3"); 6015 Ops[0] = Builder.CreateBitCast(Ops[0], 6016 llvm::PointerType::getUnqual(Ops[1]->getType())); 6017 return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]); 6018 } 6019 case NEON::BI__builtin_neon_vld4_v: 6020 case NEON::BI__builtin_neon_vld4q_v: { 6021 llvm::Type *PTy = llvm::PointerType::getUnqual(VTy); 6022 Ops[1] = Builder.CreateBitCast(Ops[1], PTy); 6023 llvm::Type *Tys[2] = { VTy, PTy }; 6024 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld4, Tys); 6025 Ops[1] = Builder.CreateCall(F, Ops[1], "vld4"); 6026 Ops[0] = Builder.CreateBitCast(Ops[0], 6027 llvm::PointerType::getUnqual(Ops[1]->getType())); 6028 return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]); 6029 } 6030 case NEON::BI__builtin_neon_vld2_dup_v: 6031 case NEON::BI__builtin_neon_vld2q_dup_v: { 6032 llvm::Type *PTy = 6033 llvm::PointerType::getUnqual(VTy->getElementType()); 6034 Ops[1] = Builder.CreateBitCast(Ops[1], PTy); 6035 llvm::Type *Tys[2] = { VTy, PTy }; 6036 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld2r, Tys); 6037 Ops[1] = Builder.CreateCall(F, Ops[1], "vld2"); 6038 Ops[0] = Builder.CreateBitCast(Ops[0], 6039 llvm::PointerType::getUnqual(Ops[1]->getType())); 6040 return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]); 6041 } 6042 case NEON::BI__builtin_neon_vld3_dup_v: 6043 case NEON::BI__builtin_neon_vld3q_dup_v: { 6044 llvm::Type *PTy = 6045 llvm::PointerType::getUnqual(VTy->getElementType()); 6046 Ops[1] = Builder.CreateBitCast(Ops[1], PTy); 6047 llvm::Type *Tys[2] = { VTy, PTy }; 6048 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld3r, Tys); 6049 Ops[1] = Builder.CreateCall(F, Ops[1], "vld3"); 6050 Ops[0] = Builder.CreateBitCast(Ops[0], 6051 llvm::PointerType::getUnqual(Ops[1]->getType())); 6052 return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]); 6053 } 6054 case NEON::BI__builtin_neon_vld4_dup_v: 6055 case NEON::BI__builtin_neon_vld4q_dup_v: { 6056 llvm::Type *PTy = 6057 llvm::PointerType::getUnqual(VTy->getElementType()); 6058 Ops[1] = Builder.CreateBitCast(Ops[1], PTy); 6059 llvm::Type *Tys[2] = { VTy, PTy }; 6060 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld4r, Tys); 6061 Ops[1] = Builder.CreateCall(F, Ops[1], "vld4"); 6062 Ops[0] = Builder.CreateBitCast(Ops[0], 6063 llvm::PointerType::getUnqual(Ops[1]->getType())); 6064 return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]); 6065 } 6066 case NEON::BI__builtin_neon_vld2_lane_v: 6067 case NEON::BI__builtin_neon_vld2q_lane_v: { 6068 llvm::Type *Tys[2] = { VTy, Ops[1]->getType() }; 6069 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld2lane, Tys); 6070 Ops.push_back(Ops[1]); 6071 Ops.erase(Ops.begin()+1); 6072 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 6073 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 6074 Ops[3] = Builder.CreateZExt(Ops[3], Int64Ty); 6075 Ops[1] = Builder.CreateCall(F, makeArrayRef(Ops).slice(1), "vld2_lane"); 6076 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 6077 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 6078 return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]); 6079 } 6080 case NEON::BI__builtin_neon_vld3_lane_v: 6081 case NEON::BI__builtin_neon_vld3q_lane_v: { 6082 llvm::Type *Tys[2] = { VTy, Ops[1]->getType() }; 6083 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld3lane, Tys); 6084 Ops.push_back(Ops[1]); 6085 Ops.erase(Ops.begin()+1); 6086 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 6087 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 6088 Ops[3] = Builder.CreateBitCast(Ops[3], Ty); 6089 Ops[4] = Builder.CreateZExt(Ops[4], Int64Ty); 6090 Ops[1] = Builder.CreateCall(F, makeArrayRef(Ops).slice(1), "vld3_lane"); 6091 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 6092 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 6093 return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]); 6094 } 6095 case NEON::BI__builtin_neon_vld4_lane_v: 6096 case NEON::BI__builtin_neon_vld4q_lane_v: { 6097 llvm::Type *Tys[2] = { VTy, Ops[1]->getType() }; 6098 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld4lane, Tys); 6099 Ops.push_back(Ops[1]); 6100 Ops.erase(Ops.begin()+1); 6101 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 6102 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 6103 Ops[3] = Builder.CreateBitCast(Ops[3], Ty); 6104 Ops[4] = Builder.CreateBitCast(Ops[4], Ty); 6105 Ops[5] = Builder.CreateZExt(Ops[5], Int64Ty); 6106 Ops[1] = Builder.CreateCall(F, makeArrayRef(Ops).slice(1), "vld4_lane"); 6107 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 6108 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 6109 return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]); 6110 } 6111 case NEON::BI__builtin_neon_vst2_v: 6112 case NEON::BI__builtin_neon_vst2q_v: { 6113 Ops.push_back(Ops[0]); 6114 Ops.erase(Ops.begin()); 6115 llvm::Type *Tys[2] = { VTy, Ops[2]->getType() }; 6116 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_st2, Tys), 6117 Ops, ""); 6118 } 6119 case NEON::BI__builtin_neon_vst2_lane_v: 6120 case NEON::BI__builtin_neon_vst2q_lane_v: { 6121 Ops.push_back(Ops[0]); 6122 Ops.erase(Ops.begin()); 6123 Ops[2] = Builder.CreateZExt(Ops[2], Int64Ty); 6124 llvm::Type *Tys[2] = { VTy, Ops[3]->getType() }; 6125 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_st2lane, Tys), 6126 Ops, ""); 6127 } 6128 case NEON::BI__builtin_neon_vst3_v: 6129 case NEON::BI__builtin_neon_vst3q_v: { 6130 Ops.push_back(Ops[0]); 6131 Ops.erase(Ops.begin()); 6132 llvm::Type *Tys[2] = { VTy, Ops[3]->getType() }; 6133 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_st3, Tys), 6134 Ops, ""); 6135 } 6136 case NEON::BI__builtin_neon_vst3_lane_v: 6137 case NEON::BI__builtin_neon_vst3q_lane_v: { 6138 Ops.push_back(Ops[0]); 6139 Ops.erase(Ops.begin()); 6140 Ops[3] = Builder.CreateZExt(Ops[3], Int64Ty); 6141 llvm::Type *Tys[2] = { VTy, Ops[4]->getType() }; 6142 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_st3lane, Tys), 6143 Ops, ""); 6144 } 6145 case NEON::BI__builtin_neon_vst4_v: 6146 case NEON::BI__builtin_neon_vst4q_v: { 6147 Ops.push_back(Ops[0]); 6148 Ops.erase(Ops.begin()); 6149 llvm::Type *Tys[2] = { VTy, Ops[4]->getType() }; 6150 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_st4, Tys), 6151 Ops, ""); 6152 } 6153 case NEON::BI__builtin_neon_vst4_lane_v: 6154 case NEON::BI__builtin_neon_vst4q_lane_v: { 6155 Ops.push_back(Ops[0]); 6156 Ops.erase(Ops.begin()); 6157 Ops[4] = Builder.CreateZExt(Ops[4], Int64Ty); 6158 llvm::Type *Tys[2] = { VTy, Ops[5]->getType() }; 6159 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_st4lane, Tys), 6160 Ops, ""); 6161 } 6162 case NEON::BI__builtin_neon_vtrn_v: 6163 case NEON::BI__builtin_neon_vtrnq_v: { 6164 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty)); 6165 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 6166 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 6167 Value *SV = nullptr; 6168 6169 for (unsigned vi = 0; vi != 2; ++vi) { 6170 SmallVector<Constant*, 16> Indices; 6171 for (unsigned i = 0, e = VTy->getNumElements(); i != e; i += 2) { 6172 Indices.push_back(ConstantInt::get(Int32Ty, i+vi)); 6173 Indices.push_back(ConstantInt::get(Int32Ty, i+e+vi)); 6174 } 6175 Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ty, Ops[0], vi); 6176 SV = llvm::ConstantVector::get(Indices); 6177 SV = Builder.CreateShuffleVector(Ops[1], Ops[2], SV, "vtrn"); 6178 SV = Builder.CreateDefaultAlignedStore(SV, Addr); 6179 } 6180 return SV; 6181 } 6182 case NEON::BI__builtin_neon_vuzp_v: 6183 case NEON::BI__builtin_neon_vuzpq_v: { 6184 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty)); 6185 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 6186 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 6187 Value *SV = nullptr; 6188 6189 for (unsigned vi = 0; vi != 2; ++vi) { 6190 SmallVector<Constant*, 16> Indices; 6191 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) 6192 Indices.push_back(ConstantInt::get(Int32Ty, 2*i+vi)); 6193 6194 Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ty, Ops[0], vi); 6195 SV = llvm::ConstantVector::get(Indices); 6196 SV = Builder.CreateShuffleVector(Ops[1], Ops[2], SV, "vuzp"); 6197 SV = Builder.CreateDefaultAlignedStore(SV, Addr); 6198 } 6199 return SV; 6200 } 6201 case NEON::BI__builtin_neon_vzip_v: 6202 case NEON::BI__builtin_neon_vzipq_v: { 6203 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty)); 6204 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 6205 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 6206 Value *SV = nullptr; 6207 6208 for (unsigned vi = 0; vi != 2; ++vi) { 6209 SmallVector<Constant*, 16> Indices; 6210 for (unsigned i = 0, e = VTy->getNumElements(); i != e; i += 2) { 6211 Indices.push_back(ConstantInt::get(Int32Ty, (i + vi*e) >> 1)); 6212 Indices.push_back(ConstantInt::get(Int32Ty, ((i + vi*e) >> 1)+e)); 6213 } 6214 Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ty, Ops[0], vi); 6215 SV = llvm::ConstantVector::get(Indices); 6216 SV = Builder.CreateShuffleVector(Ops[1], Ops[2], SV, "vzip"); 6217 SV = Builder.CreateDefaultAlignedStore(SV, Addr); 6218 } 6219 return SV; 6220 } 6221 case NEON::BI__builtin_neon_vqtbl1q_v: { 6222 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbl1, Ty), 6223 Ops, "vtbl1"); 6224 } 6225 case NEON::BI__builtin_neon_vqtbl2q_v: { 6226 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbl2, Ty), 6227 Ops, "vtbl2"); 6228 } 6229 case NEON::BI__builtin_neon_vqtbl3q_v: { 6230 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbl3, Ty), 6231 Ops, "vtbl3"); 6232 } 6233 case NEON::BI__builtin_neon_vqtbl4q_v: { 6234 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbl4, Ty), 6235 Ops, "vtbl4"); 6236 } 6237 case NEON::BI__builtin_neon_vqtbx1q_v: { 6238 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbx1, Ty), 6239 Ops, "vtbx1"); 6240 } 6241 case NEON::BI__builtin_neon_vqtbx2q_v: { 6242 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbx2, Ty), 6243 Ops, "vtbx2"); 6244 } 6245 case NEON::BI__builtin_neon_vqtbx3q_v: { 6246 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbx3, Ty), 6247 Ops, "vtbx3"); 6248 } 6249 case NEON::BI__builtin_neon_vqtbx4q_v: { 6250 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbx4, Ty), 6251 Ops, "vtbx4"); 6252 } 6253 case NEON::BI__builtin_neon_vsqadd_v: 6254 case NEON::BI__builtin_neon_vsqaddq_v: { 6255 Int = Intrinsic::aarch64_neon_usqadd; 6256 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vsqadd"); 6257 } 6258 case NEON::BI__builtin_neon_vuqadd_v: 6259 case NEON::BI__builtin_neon_vuqaddq_v: { 6260 Int = Intrinsic::aarch64_neon_suqadd; 6261 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vuqadd"); 6262 } 6263 } 6264 } 6265 6266 llvm::Value *CodeGenFunction:: 6267 BuildVector(ArrayRef<llvm::Value*> Ops) { 6268 assert((Ops.size() & (Ops.size() - 1)) == 0 && 6269 "Not a power-of-two sized vector!"); 6270 bool AllConstants = true; 6271 for (unsigned i = 0, e = Ops.size(); i != e && AllConstants; ++i) 6272 AllConstants &= isa<Constant>(Ops[i]); 6273 6274 // If this is a constant vector, create a ConstantVector. 6275 if (AllConstants) { 6276 SmallVector<llvm::Constant*, 16> CstOps; 6277 for (unsigned i = 0, e = Ops.size(); i != e; ++i) 6278 CstOps.push_back(cast<Constant>(Ops[i])); 6279 return llvm::ConstantVector::get(CstOps); 6280 } 6281 6282 // Otherwise, insertelement the values to build the vector. 6283 Value *Result = 6284 llvm::UndefValue::get(llvm::VectorType::get(Ops[0]->getType(), Ops.size())); 6285 6286 for (unsigned i = 0, e = Ops.size(); i != e; ++i) 6287 Result = Builder.CreateInsertElement(Result, Ops[i], Builder.getInt32(i)); 6288 6289 return Result; 6290 } 6291 6292 Value *CodeGenFunction::EmitX86BuiltinExpr(unsigned BuiltinID, 6293 const CallExpr *E) { 6294 if (BuiltinID == X86::BI__builtin_ms_va_start || 6295 BuiltinID == X86::BI__builtin_ms_va_end) 6296 return EmitVAStartEnd(EmitMSVAListRef(E->getArg(0)).getPointer(), 6297 BuiltinID == X86::BI__builtin_ms_va_start); 6298 if (BuiltinID == X86::BI__builtin_ms_va_copy) { 6299 // Lower this manually. We can't reliably determine whether or not any 6300 // given va_copy() is for a Win64 va_list from the calling convention 6301 // alone, because it's legal to do this from a System V ABI function. 6302 // With opaque pointer types, we won't have enough information in LLVM 6303 // IR to determine this from the argument types, either. Best to do it 6304 // now, while we have enough information. 6305 Address DestAddr = EmitMSVAListRef(E->getArg(0)); 6306 Address SrcAddr = EmitMSVAListRef(E->getArg(1)); 6307 6308 llvm::Type *BPP = Int8PtrPtrTy; 6309 6310 DestAddr = Address(Builder.CreateBitCast(DestAddr.getPointer(), BPP, "cp"), 6311 DestAddr.getAlignment()); 6312 SrcAddr = Address(Builder.CreateBitCast(SrcAddr.getPointer(), BPP, "ap"), 6313 SrcAddr.getAlignment()); 6314 6315 Value *ArgPtr = Builder.CreateLoad(SrcAddr, "ap.val"); 6316 return Builder.CreateStore(ArgPtr, DestAddr); 6317 } 6318 6319 SmallVector<Value*, 4> Ops; 6320 6321 // Find out if any arguments are required to be integer constant expressions. 6322 unsigned ICEArguments = 0; 6323 ASTContext::GetBuiltinTypeError Error; 6324 getContext().GetBuiltinType(BuiltinID, Error, &ICEArguments); 6325 assert(Error == ASTContext::GE_None && "Should not codegen an error"); 6326 6327 for (unsigned i = 0, e = E->getNumArgs(); i != e; i++) { 6328 // If this is a normal argument, just emit it as a scalar. 6329 if ((ICEArguments & (1 << i)) == 0) { 6330 Ops.push_back(EmitScalarExpr(E->getArg(i))); 6331 continue; 6332 } 6333 6334 // If this is required to be a constant, constant fold it so that we know 6335 // that the generated intrinsic gets a ConstantInt. 6336 llvm::APSInt Result; 6337 bool IsConst = E->getArg(i)->isIntegerConstantExpr(Result, getContext()); 6338 assert(IsConst && "Constant arg isn't actually constant?"); (void)IsConst; 6339 Ops.push_back(llvm::ConstantInt::get(getLLVMContext(), Result)); 6340 } 6341 6342 switch (BuiltinID) { 6343 default: return nullptr; 6344 case X86::BI__builtin_cpu_supports: { 6345 const Expr *FeatureExpr = E->getArg(0)->IgnoreParenCasts(); 6346 StringRef FeatureStr = cast<StringLiteral>(FeatureExpr)->getString(); 6347 6348 // TODO: When/if this becomes more than x86 specific then use a TargetInfo 6349 // based mapping. 6350 // Processor features and mapping to processor feature value. 6351 enum X86Features { 6352 CMOV = 0, 6353 MMX, 6354 POPCNT, 6355 SSE, 6356 SSE2, 6357 SSE3, 6358 SSSE3, 6359 SSE4_1, 6360 SSE4_2, 6361 AVX, 6362 AVX2, 6363 SSE4_A, 6364 FMA4, 6365 XOP, 6366 FMA, 6367 AVX512F, 6368 BMI, 6369 BMI2, 6370 MAX 6371 }; 6372 6373 X86Features Feature = StringSwitch<X86Features>(FeatureStr) 6374 .Case("cmov", X86Features::CMOV) 6375 .Case("mmx", X86Features::MMX) 6376 .Case("popcnt", X86Features::POPCNT) 6377 .Case("sse", X86Features::SSE) 6378 .Case("sse2", X86Features::SSE2) 6379 .Case("sse3", X86Features::SSE3) 6380 .Case("sse4.1", X86Features::SSE4_1) 6381 .Case("sse4.2", X86Features::SSE4_2) 6382 .Case("avx", X86Features::AVX) 6383 .Case("avx2", X86Features::AVX2) 6384 .Case("sse4a", X86Features::SSE4_A) 6385 .Case("fma4", X86Features::FMA4) 6386 .Case("xop", X86Features::XOP) 6387 .Case("fma", X86Features::FMA) 6388 .Case("avx512f", X86Features::AVX512F) 6389 .Case("bmi", X86Features::BMI) 6390 .Case("bmi2", X86Features::BMI2) 6391 .Default(X86Features::MAX); 6392 assert(Feature != X86Features::MAX && "Invalid feature!"); 6393 6394 // Matching the struct layout from the compiler-rt/libgcc structure that is 6395 // filled in: 6396 // unsigned int __cpu_vendor; 6397 // unsigned int __cpu_type; 6398 // unsigned int __cpu_subtype; 6399 // unsigned int __cpu_features[1]; 6400 llvm::Type *STy = llvm::StructType::get( 6401 Int32Ty, Int32Ty, Int32Ty, llvm::ArrayType::get(Int32Ty, 1), nullptr); 6402 6403 // Grab the global __cpu_model. 6404 llvm::Constant *CpuModel = CGM.CreateRuntimeVariable(STy, "__cpu_model"); 6405 6406 // Grab the first (0th) element from the field __cpu_features off of the 6407 // global in the struct STy. 6408 Value *Idxs[] = { 6409 ConstantInt::get(Int32Ty, 0), 6410 ConstantInt::get(Int32Ty, 3), 6411 ConstantInt::get(Int32Ty, 0) 6412 }; 6413 Value *CpuFeatures = Builder.CreateGEP(STy, CpuModel, Idxs); 6414 Value *Features = Builder.CreateAlignedLoad(CpuFeatures, 6415 CharUnits::fromQuantity(4)); 6416 6417 // Check the value of the bit corresponding to the feature requested. 6418 Value *Bitset = Builder.CreateAnd( 6419 Features, llvm::ConstantInt::get(Int32Ty, 1 << Feature)); 6420 return Builder.CreateICmpNE(Bitset, llvm::ConstantInt::get(Int32Ty, 0)); 6421 } 6422 case X86::BI_mm_prefetch: { 6423 Value *Address = Ops[0]; 6424 Value *RW = ConstantInt::get(Int32Ty, 0); 6425 Value *Locality = Ops[1]; 6426 Value *Data = ConstantInt::get(Int32Ty, 1); 6427 Value *F = CGM.getIntrinsic(Intrinsic::prefetch); 6428 return Builder.CreateCall(F, {Address, RW, Locality, Data}); 6429 } 6430 case X86::BI__builtin_ia32_undef128: 6431 case X86::BI__builtin_ia32_undef256: 6432 case X86::BI__builtin_ia32_undef512: 6433 return UndefValue::get(ConvertType(E->getType())); 6434 case X86::BI__builtin_ia32_vec_init_v8qi: 6435 case X86::BI__builtin_ia32_vec_init_v4hi: 6436 case X86::BI__builtin_ia32_vec_init_v2si: 6437 return Builder.CreateBitCast(BuildVector(Ops), 6438 llvm::Type::getX86_MMXTy(getLLVMContext())); 6439 case X86::BI__builtin_ia32_vec_ext_v2si: 6440 return Builder.CreateExtractElement(Ops[0], 6441 llvm::ConstantInt::get(Ops[1]->getType(), 0)); 6442 case X86::BI__builtin_ia32_ldmxcsr: { 6443 Address Tmp = CreateMemTemp(E->getArg(0)->getType()); 6444 Builder.CreateStore(Ops[0], Tmp); 6445 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::x86_sse_ldmxcsr), 6446 Builder.CreateBitCast(Tmp.getPointer(), Int8PtrTy)); 6447 } 6448 case X86::BI__builtin_ia32_stmxcsr: { 6449 Address Tmp = CreateMemTemp(E->getType()); 6450 Builder.CreateCall(CGM.getIntrinsic(Intrinsic::x86_sse_stmxcsr), 6451 Builder.CreateBitCast(Tmp.getPointer(), Int8PtrTy)); 6452 return Builder.CreateLoad(Tmp, "stmxcsr"); 6453 } 6454 case X86::BI__builtin_ia32_xsave: 6455 case X86::BI__builtin_ia32_xsave64: 6456 case X86::BI__builtin_ia32_xrstor: 6457 case X86::BI__builtin_ia32_xrstor64: 6458 case X86::BI__builtin_ia32_xsaveopt: 6459 case X86::BI__builtin_ia32_xsaveopt64: 6460 case X86::BI__builtin_ia32_xrstors: 6461 case X86::BI__builtin_ia32_xrstors64: 6462 case X86::BI__builtin_ia32_xsavec: 6463 case X86::BI__builtin_ia32_xsavec64: 6464 case X86::BI__builtin_ia32_xsaves: 6465 case X86::BI__builtin_ia32_xsaves64: { 6466 Intrinsic::ID ID; 6467 #define INTRINSIC_X86_XSAVE_ID(NAME) \ 6468 case X86::BI__builtin_ia32_##NAME: \ 6469 ID = Intrinsic::x86_##NAME; \ 6470 break 6471 switch (BuiltinID) { 6472 default: llvm_unreachable("Unsupported intrinsic!"); 6473 INTRINSIC_X86_XSAVE_ID(xsave); 6474 INTRINSIC_X86_XSAVE_ID(xsave64); 6475 INTRINSIC_X86_XSAVE_ID(xrstor); 6476 INTRINSIC_X86_XSAVE_ID(xrstor64); 6477 INTRINSIC_X86_XSAVE_ID(xsaveopt); 6478 INTRINSIC_X86_XSAVE_ID(xsaveopt64); 6479 INTRINSIC_X86_XSAVE_ID(xrstors); 6480 INTRINSIC_X86_XSAVE_ID(xrstors64); 6481 INTRINSIC_X86_XSAVE_ID(xsavec); 6482 INTRINSIC_X86_XSAVE_ID(xsavec64); 6483 INTRINSIC_X86_XSAVE_ID(xsaves); 6484 INTRINSIC_X86_XSAVE_ID(xsaves64); 6485 } 6486 #undef INTRINSIC_X86_XSAVE_ID 6487 Value *Mhi = Builder.CreateTrunc( 6488 Builder.CreateLShr(Ops[1], ConstantInt::get(Int64Ty, 32)), Int32Ty); 6489 Value *Mlo = Builder.CreateTrunc(Ops[1], Int32Ty); 6490 Ops[1] = Mhi; 6491 Ops.push_back(Mlo); 6492 return Builder.CreateCall(CGM.getIntrinsic(ID), Ops); 6493 } 6494 case X86::BI__builtin_ia32_storehps: 6495 case X86::BI__builtin_ia32_storelps: { 6496 llvm::Type *PtrTy = llvm::PointerType::getUnqual(Int64Ty); 6497 llvm::Type *VecTy = llvm::VectorType::get(Int64Ty, 2); 6498 6499 // cast val v2i64 6500 Ops[1] = Builder.CreateBitCast(Ops[1], VecTy, "cast"); 6501 6502 // extract (0, 1) 6503 unsigned Index = BuiltinID == X86::BI__builtin_ia32_storelps ? 0 : 1; 6504 llvm::Value *Idx = llvm::ConstantInt::get(SizeTy, Index); 6505 Ops[1] = Builder.CreateExtractElement(Ops[1], Idx, "extract"); 6506 6507 // cast pointer to i64 & store 6508 Ops[0] = Builder.CreateBitCast(Ops[0], PtrTy); 6509 return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]); 6510 } 6511 case X86::BI__builtin_ia32_palignr128: 6512 case X86::BI__builtin_ia32_palignr256: { 6513 unsigned ShiftVal = cast<llvm::ConstantInt>(Ops[2])->getZExtValue(); 6514 6515 unsigned NumElts = 6516 cast<llvm::VectorType>(Ops[0]->getType())->getNumElements(); 6517 assert(NumElts % 16 == 0); 6518 unsigned NumLanes = NumElts / 16; 6519 unsigned NumLaneElts = NumElts / NumLanes; 6520 6521 // If palignr is shifting the pair of vectors more than the size of two 6522 // lanes, emit zero. 6523 if (ShiftVal >= (2 * NumLaneElts)) 6524 return llvm::Constant::getNullValue(ConvertType(E->getType())); 6525 6526 // If palignr is shifting the pair of input vectors more than one lane, 6527 // but less than two lanes, convert to shifting in zeroes. 6528 if (ShiftVal > NumLaneElts) { 6529 ShiftVal -= NumLaneElts; 6530 Ops[1] = Ops[0]; 6531 Ops[0] = llvm::Constant::getNullValue(Ops[0]->getType()); 6532 } 6533 6534 uint32_t Indices[32]; 6535 // 256-bit palignr operates on 128-bit lanes so we need to handle that 6536 for (unsigned l = 0; l != NumElts; l += NumLaneElts) { 6537 for (unsigned i = 0; i != NumLaneElts; ++i) { 6538 unsigned Idx = ShiftVal + i; 6539 if (Idx >= NumLaneElts) 6540 Idx += NumElts - NumLaneElts; // End of lane, switch operand. 6541 Indices[l + i] = Idx + l; 6542 } 6543 } 6544 6545 Value *SV = llvm::ConstantDataVector::get(getLLVMContext(), 6546 makeArrayRef(Indices, NumElts)); 6547 return Builder.CreateShuffleVector(Ops[1], Ops[0], SV, "palignr"); 6548 } 6549 case X86::BI__builtin_ia32_pslldqi256: { 6550 // Shift value is in bits so divide by 8. 6551 unsigned shiftVal = cast<llvm::ConstantInt>(Ops[1])->getZExtValue() >> 3; 6552 6553 // If pslldq is shifting the vector more than 15 bytes, emit zero. 6554 if (shiftVal >= 16) 6555 return llvm::Constant::getNullValue(ConvertType(E->getType())); 6556 6557 uint32_t Indices[32]; 6558 // 256-bit pslldq operates on 128-bit lanes so we need to handle that 6559 for (unsigned l = 0; l != 32; l += 16) { 6560 for (unsigned i = 0; i != 16; ++i) { 6561 unsigned Idx = 32 + i - shiftVal; 6562 if (Idx < 32) Idx -= 16; // end of lane, switch operand. 6563 Indices[l + i] = Idx + l; 6564 } 6565 } 6566 6567 llvm::Type *VecTy = llvm::VectorType::get(Int8Ty, 32); 6568 Ops[0] = Builder.CreateBitCast(Ops[0], VecTy, "cast"); 6569 Value *Zero = llvm::Constant::getNullValue(VecTy); 6570 6571 Value *SV = llvm::ConstantDataVector::get(getLLVMContext(), Indices); 6572 SV = Builder.CreateShuffleVector(Zero, Ops[0], SV, "pslldq"); 6573 llvm::Type *ResultType = ConvertType(E->getType()); 6574 return Builder.CreateBitCast(SV, ResultType, "cast"); 6575 } 6576 case X86::BI__builtin_ia32_psrldqi256: { 6577 // Shift value is in bits so divide by 8. 6578 unsigned shiftVal = cast<llvm::ConstantInt>(Ops[1])->getZExtValue() >> 3; 6579 6580 // If psrldq is shifting the vector more than 15 bytes, emit zero. 6581 if (shiftVal >= 16) 6582 return llvm::Constant::getNullValue(ConvertType(E->getType())); 6583 6584 uint32_t Indices[32]; 6585 // 256-bit psrldq operates on 128-bit lanes so we need to handle that 6586 for (unsigned l = 0; l != 32; l += 16) { 6587 for (unsigned i = 0; i != 16; ++i) { 6588 unsigned Idx = i + shiftVal; 6589 if (Idx >= 16) Idx += 16; // end of lane, switch operand. 6590 Indices[l + i] = Idx + l; 6591 } 6592 } 6593 6594 llvm::Type *VecTy = llvm::VectorType::get(Int8Ty, 32); 6595 Ops[0] = Builder.CreateBitCast(Ops[0], VecTy, "cast"); 6596 Value *Zero = llvm::Constant::getNullValue(VecTy); 6597 6598 Value *SV = llvm::ConstantDataVector::get(getLLVMContext(), Indices); 6599 SV = Builder.CreateShuffleVector(Ops[0], Zero, SV, "psrldq"); 6600 llvm::Type *ResultType = ConvertType(E->getType()); 6601 return Builder.CreateBitCast(SV, ResultType, "cast"); 6602 } 6603 case X86::BI__builtin_ia32_movntps: 6604 case X86::BI__builtin_ia32_movntps256: 6605 case X86::BI__builtin_ia32_movntpd: 6606 case X86::BI__builtin_ia32_movntpd256: 6607 case X86::BI__builtin_ia32_movntdq: 6608 case X86::BI__builtin_ia32_movntdq256: 6609 case X86::BI__builtin_ia32_movnti: 6610 case X86::BI__builtin_ia32_movnti64: { 6611 llvm::MDNode *Node = llvm::MDNode::get( 6612 getLLVMContext(), llvm::ConstantAsMetadata::get(Builder.getInt32(1))); 6613 6614 // Convert the type of the pointer to a pointer to the stored type. 6615 Value *BC = Builder.CreateBitCast(Ops[0], 6616 llvm::PointerType::getUnqual(Ops[1]->getType()), 6617 "cast"); 6618 StoreInst *SI = Builder.CreateDefaultAlignedStore(Ops[1], BC); 6619 SI->setMetadata(CGM.getModule().getMDKindID("nontemporal"), Node); 6620 6621 // If the operand is an integer, we can't assume alignment. Otherwise, 6622 // assume natural alignment. 6623 QualType ArgTy = E->getArg(1)->getType(); 6624 unsigned Align; 6625 if (ArgTy->isIntegerType()) 6626 Align = 1; 6627 else 6628 Align = getContext().getTypeSizeInChars(ArgTy).getQuantity(); 6629 SI->setAlignment(Align); 6630 return SI; 6631 } 6632 // 3DNow! 6633 case X86::BI__builtin_ia32_pswapdsf: 6634 case X86::BI__builtin_ia32_pswapdsi: { 6635 llvm::Type *MMXTy = llvm::Type::getX86_MMXTy(getLLVMContext()); 6636 Ops[0] = Builder.CreateBitCast(Ops[0], MMXTy, "cast"); 6637 llvm::Function *F = CGM.getIntrinsic(Intrinsic::x86_3dnowa_pswapd); 6638 return Builder.CreateCall(F, Ops, "pswapd"); 6639 } 6640 case X86::BI__builtin_ia32_rdrand16_step: 6641 case X86::BI__builtin_ia32_rdrand32_step: 6642 case X86::BI__builtin_ia32_rdrand64_step: 6643 case X86::BI__builtin_ia32_rdseed16_step: 6644 case X86::BI__builtin_ia32_rdseed32_step: 6645 case X86::BI__builtin_ia32_rdseed64_step: { 6646 Intrinsic::ID ID; 6647 switch (BuiltinID) { 6648 default: llvm_unreachable("Unsupported intrinsic!"); 6649 case X86::BI__builtin_ia32_rdrand16_step: 6650 ID = Intrinsic::x86_rdrand_16; 6651 break; 6652 case X86::BI__builtin_ia32_rdrand32_step: 6653 ID = Intrinsic::x86_rdrand_32; 6654 break; 6655 case X86::BI__builtin_ia32_rdrand64_step: 6656 ID = Intrinsic::x86_rdrand_64; 6657 break; 6658 case X86::BI__builtin_ia32_rdseed16_step: 6659 ID = Intrinsic::x86_rdseed_16; 6660 break; 6661 case X86::BI__builtin_ia32_rdseed32_step: 6662 ID = Intrinsic::x86_rdseed_32; 6663 break; 6664 case X86::BI__builtin_ia32_rdseed64_step: 6665 ID = Intrinsic::x86_rdseed_64; 6666 break; 6667 } 6668 6669 Value *Call = Builder.CreateCall(CGM.getIntrinsic(ID)); 6670 Builder.CreateDefaultAlignedStore(Builder.CreateExtractValue(Call, 0), 6671 Ops[0]); 6672 return Builder.CreateExtractValue(Call, 1); 6673 } 6674 // SSE comparison intrisics 6675 case X86::BI__builtin_ia32_cmpeqps: 6676 case X86::BI__builtin_ia32_cmpltps: 6677 case X86::BI__builtin_ia32_cmpleps: 6678 case X86::BI__builtin_ia32_cmpunordps: 6679 case X86::BI__builtin_ia32_cmpneqps: 6680 case X86::BI__builtin_ia32_cmpnltps: 6681 case X86::BI__builtin_ia32_cmpnleps: 6682 case X86::BI__builtin_ia32_cmpordps: 6683 case X86::BI__builtin_ia32_cmpeqss: 6684 case X86::BI__builtin_ia32_cmpltss: 6685 case X86::BI__builtin_ia32_cmpless: 6686 case X86::BI__builtin_ia32_cmpunordss: 6687 case X86::BI__builtin_ia32_cmpneqss: 6688 case X86::BI__builtin_ia32_cmpnltss: 6689 case X86::BI__builtin_ia32_cmpnless: 6690 case X86::BI__builtin_ia32_cmpordss: 6691 case X86::BI__builtin_ia32_cmpeqpd: 6692 case X86::BI__builtin_ia32_cmpltpd: 6693 case X86::BI__builtin_ia32_cmplepd: 6694 case X86::BI__builtin_ia32_cmpunordpd: 6695 case X86::BI__builtin_ia32_cmpneqpd: 6696 case X86::BI__builtin_ia32_cmpnltpd: 6697 case X86::BI__builtin_ia32_cmpnlepd: 6698 case X86::BI__builtin_ia32_cmpordpd: 6699 case X86::BI__builtin_ia32_cmpeqsd: 6700 case X86::BI__builtin_ia32_cmpltsd: 6701 case X86::BI__builtin_ia32_cmplesd: 6702 case X86::BI__builtin_ia32_cmpunordsd: 6703 case X86::BI__builtin_ia32_cmpneqsd: 6704 case X86::BI__builtin_ia32_cmpnltsd: 6705 case X86::BI__builtin_ia32_cmpnlesd: 6706 case X86::BI__builtin_ia32_cmpordsd: 6707 // These exist so that the builtin that takes an immediate can be bounds 6708 // checked by clang to avoid passing bad immediates to the backend. Since 6709 // AVX has a larger immediate than SSE we would need separate builtins to 6710 // do the different bounds checking. Rather than create a clang specific 6711 // SSE only builtin, this implements eight separate builtins to match gcc 6712 // implementation. 6713 6714 // Choose the immediate. 6715 unsigned Imm; 6716 switch (BuiltinID) { 6717 default: llvm_unreachable("Unsupported intrinsic!"); 6718 case X86::BI__builtin_ia32_cmpeqps: 6719 case X86::BI__builtin_ia32_cmpeqss: 6720 case X86::BI__builtin_ia32_cmpeqpd: 6721 case X86::BI__builtin_ia32_cmpeqsd: 6722 Imm = 0; 6723 break; 6724 case X86::BI__builtin_ia32_cmpltps: 6725 case X86::BI__builtin_ia32_cmpltss: 6726 case X86::BI__builtin_ia32_cmpltpd: 6727 case X86::BI__builtin_ia32_cmpltsd: 6728 Imm = 1; 6729 break; 6730 case X86::BI__builtin_ia32_cmpleps: 6731 case X86::BI__builtin_ia32_cmpless: 6732 case X86::BI__builtin_ia32_cmplepd: 6733 case X86::BI__builtin_ia32_cmplesd: 6734 Imm = 2; 6735 break; 6736 case X86::BI__builtin_ia32_cmpunordps: 6737 case X86::BI__builtin_ia32_cmpunordss: 6738 case X86::BI__builtin_ia32_cmpunordpd: 6739 case X86::BI__builtin_ia32_cmpunordsd: 6740 Imm = 3; 6741 break; 6742 case X86::BI__builtin_ia32_cmpneqps: 6743 case X86::BI__builtin_ia32_cmpneqss: 6744 case X86::BI__builtin_ia32_cmpneqpd: 6745 case X86::BI__builtin_ia32_cmpneqsd: 6746 Imm = 4; 6747 break; 6748 case X86::BI__builtin_ia32_cmpnltps: 6749 case X86::BI__builtin_ia32_cmpnltss: 6750 case X86::BI__builtin_ia32_cmpnltpd: 6751 case X86::BI__builtin_ia32_cmpnltsd: 6752 Imm = 5; 6753 break; 6754 case X86::BI__builtin_ia32_cmpnleps: 6755 case X86::BI__builtin_ia32_cmpnless: 6756 case X86::BI__builtin_ia32_cmpnlepd: 6757 case X86::BI__builtin_ia32_cmpnlesd: 6758 Imm = 6; 6759 break; 6760 case X86::BI__builtin_ia32_cmpordps: 6761 case X86::BI__builtin_ia32_cmpordss: 6762 case X86::BI__builtin_ia32_cmpordpd: 6763 case X86::BI__builtin_ia32_cmpordsd: 6764 Imm = 7; 6765 break; 6766 } 6767 6768 // Choose the intrinsic ID. 6769 const char *name; 6770 Intrinsic::ID ID; 6771 switch (BuiltinID) { 6772 default: llvm_unreachable("Unsupported intrinsic!"); 6773 case X86::BI__builtin_ia32_cmpeqps: 6774 case X86::BI__builtin_ia32_cmpltps: 6775 case X86::BI__builtin_ia32_cmpleps: 6776 case X86::BI__builtin_ia32_cmpunordps: 6777 case X86::BI__builtin_ia32_cmpneqps: 6778 case X86::BI__builtin_ia32_cmpnltps: 6779 case X86::BI__builtin_ia32_cmpnleps: 6780 case X86::BI__builtin_ia32_cmpordps: 6781 name = "cmpps"; 6782 ID = Intrinsic::x86_sse_cmp_ps; 6783 break; 6784 case X86::BI__builtin_ia32_cmpeqss: 6785 case X86::BI__builtin_ia32_cmpltss: 6786 case X86::BI__builtin_ia32_cmpless: 6787 case X86::BI__builtin_ia32_cmpunordss: 6788 case X86::BI__builtin_ia32_cmpneqss: 6789 case X86::BI__builtin_ia32_cmpnltss: 6790 case X86::BI__builtin_ia32_cmpnless: 6791 case X86::BI__builtin_ia32_cmpordss: 6792 name = "cmpss"; 6793 ID = Intrinsic::x86_sse_cmp_ss; 6794 break; 6795 case X86::BI__builtin_ia32_cmpeqpd: 6796 case X86::BI__builtin_ia32_cmpltpd: 6797 case X86::BI__builtin_ia32_cmplepd: 6798 case X86::BI__builtin_ia32_cmpunordpd: 6799 case X86::BI__builtin_ia32_cmpneqpd: 6800 case X86::BI__builtin_ia32_cmpnltpd: 6801 case X86::BI__builtin_ia32_cmpnlepd: 6802 case X86::BI__builtin_ia32_cmpordpd: 6803 name = "cmppd"; 6804 ID = Intrinsic::x86_sse2_cmp_pd; 6805 break; 6806 case X86::BI__builtin_ia32_cmpeqsd: 6807 case X86::BI__builtin_ia32_cmpltsd: 6808 case X86::BI__builtin_ia32_cmplesd: 6809 case X86::BI__builtin_ia32_cmpunordsd: 6810 case X86::BI__builtin_ia32_cmpneqsd: 6811 case X86::BI__builtin_ia32_cmpnltsd: 6812 case X86::BI__builtin_ia32_cmpnlesd: 6813 case X86::BI__builtin_ia32_cmpordsd: 6814 name = "cmpsd"; 6815 ID = Intrinsic::x86_sse2_cmp_sd; 6816 break; 6817 } 6818 6819 Ops.push_back(llvm::ConstantInt::get(Int8Ty, Imm)); 6820 llvm::Function *F = CGM.getIntrinsic(ID); 6821 return Builder.CreateCall(F, Ops, name); 6822 } 6823 } 6824 6825 6826 Value *CodeGenFunction::EmitPPCBuiltinExpr(unsigned BuiltinID, 6827 const CallExpr *E) { 6828 SmallVector<Value*, 4> Ops; 6829 6830 for (unsigned i = 0, e = E->getNumArgs(); i != e; i++) 6831 Ops.push_back(EmitScalarExpr(E->getArg(i))); 6832 6833 Intrinsic::ID ID = Intrinsic::not_intrinsic; 6834 6835 switch (BuiltinID) { 6836 default: return nullptr; 6837 6838 // __builtin_ppc_get_timebase is GCC 4.8+'s PowerPC-specific name for what we 6839 // call __builtin_readcyclecounter. 6840 case PPC::BI__builtin_ppc_get_timebase: 6841 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::readcyclecounter)); 6842 6843 // vec_ld, vec_lvsl, vec_lvsr 6844 case PPC::BI__builtin_altivec_lvx: 6845 case PPC::BI__builtin_altivec_lvxl: 6846 case PPC::BI__builtin_altivec_lvebx: 6847 case PPC::BI__builtin_altivec_lvehx: 6848 case PPC::BI__builtin_altivec_lvewx: 6849 case PPC::BI__builtin_altivec_lvsl: 6850 case PPC::BI__builtin_altivec_lvsr: 6851 case PPC::BI__builtin_vsx_lxvd2x: 6852 case PPC::BI__builtin_vsx_lxvw4x: 6853 { 6854 Ops[1] = Builder.CreateBitCast(Ops[1], Int8PtrTy); 6855 6856 Ops[0] = Builder.CreateGEP(Ops[1], Ops[0]); 6857 Ops.pop_back(); 6858 6859 switch (BuiltinID) { 6860 default: llvm_unreachable("Unsupported ld/lvsl/lvsr intrinsic!"); 6861 case PPC::BI__builtin_altivec_lvx: 6862 ID = Intrinsic::ppc_altivec_lvx; 6863 break; 6864 case PPC::BI__builtin_altivec_lvxl: 6865 ID = Intrinsic::ppc_altivec_lvxl; 6866 break; 6867 case PPC::BI__builtin_altivec_lvebx: 6868 ID = Intrinsic::ppc_altivec_lvebx; 6869 break; 6870 case PPC::BI__builtin_altivec_lvehx: 6871 ID = Intrinsic::ppc_altivec_lvehx; 6872 break; 6873 case PPC::BI__builtin_altivec_lvewx: 6874 ID = Intrinsic::ppc_altivec_lvewx; 6875 break; 6876 case PPC::BI__builtin_altivec_lvsl: 6877 ID = Intrinsic::ppc_altivec_lvsl; 6878 break; 6879 case PPC::BI__builtin_altivec_lvsr: 6880 ID = Intrinsic::ppc_altivec_lvsr; 6881 break; 6882 case PPC::BI__builtin_vsx_lxvd2x: 6883 ID = Intrinsic::ppc_vsx_lxvd2x; 6884 break; 6885 case PPC::BI__builtin_vsx_lxvw4x: 6886 ID = Intrinsic::ppc_vsx_lxvw4x; 6887 break; 6888 } 6889 llvm::Function *F = CGM.getIntrinsic(ID); 6890 return Builder.CreateCall(F, Ops, ""); 6891 } 6892 6893 // vec_st 6894 case PPC::BI__builtin_altivec_stvx: 6895 case PPC::BI__builtin_altivec_stvxl: 6896 case PPC::BI__builtin_altivec_stvebx: 6897 case PPC::BI__builtin_altivec_stvehx: 6898 case PPC::BI__builtin_altivec_stvewx: 6899 case PPC::BI__builtin_vsx_stxvd2x: 6900 case PPC::BI__builtin_vsx_stxvw4x: 6901 { 6902 Ops[2] = Builder.CreateBitCast(Ops[2], Int8PtrTy); 6903 Ops[1] = Builder.CreateGEP(Ops[2], Ops[1]); 6904 Ops.pop_back(); 6905 6906 switch (BuiltinID) { 6907 default: llvm_unreachable("Unsupported st intrinsic!"); 6908 case PPC::BI__builtin_altivec_stvx: 6909 ID = Intrinsic::ppc_altivec_stvx; 6910 break; 6911 case PPC::BI__builtin_altivec_stvxl: 6912 ID = Intrinsic::ppc_altivec_stvxl; 6913 break; 6914 case PPC::BI__builtin_altivec_stvebx: 6915 ID = Intrinsic::ppc_altivec_stvebx; 6916 break; 6917 case PPC::BI__builtin_altivec_stvehx: 6918 ID = Intrinsic::ppc_altivec_stvehx; 6919 break; 6920 case PPC::BI__builtin_altivec_stvewx: 6921 ID = Intrinsic::ppc_altivec_stvewx; 6922 break; 6923 case PPC::BI__builtin_vsx_stxvd2x: 6924 ID = Intrinsic::ppc_vsx_stxvd2x; 6925 break; 6926 case PPC::BI__builtin_vsx_stxvw4x: 6927 ID = Intrinsic::ppc_vsx_stxvw4x; 6928 break; 6929 } 6930 llvm::Function *F = CGM.getIntrinsic(ID); 6931 return Builder.CreateCall(F, Ops, ""); 6932 } 6933 // Square root 6934 case PPC::BI__builtin_vsx_xvsqrtsp: 6935 case PPC::BI__builtin_vsx_xvsqrtdp: { 6936 llvm::Type *ResultType = ConvertType(E->getType()); 6937 Value *X = EmitScalarExpr(E->getArg(0)); 6938 ID = Intrinsic::sqrt; 6939 llvm::Function *F = CGM.getIntrinsic(ID, ResultType); 6940 return Builder.CreateCall(F, X); 6941 } 6942 // Count leading zeros 6943 case PPC::BI__builtin_altivec_vclzb: 6944 case PPC::BI__builtin_altivec_vclzh: 6945 case PPC::BI__builtin_altivec_vclzw: 6946 case PPC::BI__builtin_altivec_vclzd: { 6947 llvm::Type *ResultType = ConvertType(E->getType()); 6948 Value *X = EmitScalarExpr(E->getArg(0)); 6949 Value *Undef = ConstantInt::get(Builder.getInt1Ty(), false); 6950 Function *F = CGM.getIntrinsic(Intrinsic::ctlz, ResultType); 6951 return Builder.CreateCall(F, {X, Undef}); 6952 } 6953 // Copy sign 6954 case PPC::BI__builtin_vsx_xvcpsgnsp: 6955 case PPC::BI__builtin_vsx_xvcpsgndp: { 6956 llvm::Type *ResultType = ConvertType(E->getType()); 6957 Value *X = EmitScalarExpr(E->getArg(0)); 6958 Value *Y = EmitScalarExpr(E->getArg(1)); 6959 ID = Intrinsic::copysign; 6960 llvm::Function *F = CGM.getIntrinsic(ID, ResultType); 6961 return Builder.CreateCall(F, {X, Y}); 6962 } 6963 // Rounding/truncation 6964 case PPC::BI__builtin_vsx_xvrspip: 6965 case PPC::BI__builtin_vsx_xvrdpip: 6966 case PPC::BI__builtin_vsx_xvrdpim: 6967 case PPC::BI__builtin_vsx_xvrspim: 6968 case PPC::BI__builtin_vsx_xvrdpi: 6969 case PPC::BI__builtin_vsx_xvrspi: 6970 case PPC::BI__builtin_vsx_xvrdpic: 6971 case PPC::BI__builtin_vsx_xvrspic: 6972 case PPC::BI__builtin_vsx_xvrdpiz: 6973 case PPC::BI__builtin_vsx_xvrspiz: { 6974 llvm::Type *ResultType = ConvertType(E->getType()); 6975 Value *X = EmitScalarExpr(E->getArg(0)); 6976 if (BuiltinID == PPC::BI__builtin_vsx_xvrdpim || 6977 BuiltinID == PPC::BI__builtin_vsx_xvrspim) 6978 ID = Intrinsic::floor; 6979 else if (BuiltinID == PPC::BI__builtin_vsx_xvrdpi || 6980 BuiltinID == PPC::BI__builtin_vsx_xvrspi) 6981 ID = Intrinsic::round; 6982 else if (BuiltinID == PPC::BI__builtin_vsx_xvrdpic || 6983 BuiltinID == PPC::BI__builtin_vsx_xvrspic) 6984 ID = Intrinsic::nearbyint; 6985 else if (BuiltinID == PPC::BI__builtin_vsx_xvrdpip || 6986 BuiltinID == PPC::BI__builtin_vsx_xvrspip) 6987 ID = Intrinsic::ceil; 6988 else if (BuiltinID == PPC::BI__builtin_vsx_xvrdpiz || 6989 BuiltinID == PPC::BI__builtin_vsx_xvrspiz) 6990 ID = Intrinsic::trunc; 6991 llvm::Function *F = CGM.getIntrinsic(ID, ResultType); 6992 return Builder.CreateCall(F, X); 6993 } 6994 6995 // Absolute value 6996 case PPC::BI__builtin_vsx_xvabsdp: 6997 case PPC::BI__builtin_vsx_xvabssp: { 6998 llvm::Type *ResultType = ConvertType(E->getType()); 6999 Value *X = EmitScalarExpr(E->getArg(0)); 7000 llvm::Function *F = CGM.getIntrinsic(Intrinsic::fabs, ResultType); 7001 return Builder.CreateCall(F, X); 7002 } 7003 7004 // FMA variations 7005 case PPC::BI__builtin_vsx_xvmaddadp: 7006 case PPC::BI__builtin_vsx_xvmaddasp: 7007 case PPC::BI__builtin_vsx_xvnmaddadp: 7008 case PPC::BI__builtin_vsx_xvnmaddasp: 7009 case PPC::BI__builtin_vsx_xvmsubadp: 7010 case PPC::BI__builtin_vsx_xvmsubasp: 7011 case PPC::BI__builtin_vsx_xvnmsubadp: 7012 case PPC::BI__builtin_vsx_xvnmsubasp: { 7013 llvm::Type *ResultType = ConvertType(E->getType()); 7014 Value *X = EmitScalarExpr(E->getArg(0)); 7015 Value *Y = EmitScalarExpr(E->getArg(1)); 7016 Value *Z = EmitScalarExpr(E->getArg(2)); 7017 Value *Zero = llvm::ConstantFP::getZeroValueForNegation(ResultType); 7018 llvm::Function *F = CGM.getIntrinsic(Intrinsic::fma, ResultType); 7019 switch (BuiltinID) { 7020 case PPC::BI__builtin_vsx_xvmaddadp: 7021 case PPC::BI__builtin_vsx_xvmaddasp: 7022 return Builder.CreateCall(F, {X, Y, Z}); 7023 case PPC::BI__builtin_vsx_xvnmaddadp: 7024 case PPC::BI__builtin_vsx_xvnmaddasp: 7025 return Builder.CreateFSub(Zero, 7026 Builder.CreateCall(F, {X, Y, Z}), "sub"); 7027 case PPC::BI__builtin_vsx_xvmsubadp: 7028 case PPC::BI__builtin_vsx_xvmsubasp: 7029 return Builder.CreateCall(F, 7030 {X, Y, Builder.CreateFSub(Zero, Z, "sub")}); 7031 case PPC::BI__builtin_vsx_xvnmsubadp: 7032 case PPC::BI__builtin_vsx_xvnmsubasp: 7033 Value *FsubRes = 7034 Builder.CreateCall(F, {X, Y, Builder.CreateFSub(Zero, Z, "sub")}); 7035 return Builder.CreateFSub(Zero, FsubRes, "sub"); 7036 } 7037 llvm_unreachable("Unknown FMA operation"); 7038 return nullptr; // Suppress no-return warning 7039 } 7040 } 7041 } 7042 7043 Value *CodeGenFunction::EmitAMDGPUBuiltinExpr(unsigned BuiltinID, 7044 const CallExpr *E) { 7045 switch (BuiltinID) { 7046 case AMDGPU::BI__builtin_amdgcn_div_scale: 7047 case AMDGPU::BI__builtin_amdgcn_div_scalef: { 7048 // Translate from the intrinsics's struct return to the builtin's out 7049 // argument. 7050 7051 Address FlagOutPtr = EmitPointerWithAlignment(E->getArg(3)); 7052 7053 llvm::Value *X = EmitScalarExpr(E->getArg(0)); 7054 llvm::Value *Y = EmitScalarExpr(E->getArg(1)); 7055 llvm::Value *Z = EmitScalarExpr(E->getArg(2)); 7056 7057 llvm::Value *Callee = CGM.getIntrinsic(Intrinsic::amdgcn_div_scale, 7058 X->getType()); 7059 7060 llvm::Value *Tmp = Builder.CreateCall(Callee, {X, Y, Z}); 7061 7062 llvm::Value *Result = Builder.CreateExtractValue(Tmp, 0); 7063 llvm::Value *Flag = Builder.CreateExtractValue(Tmp, 1); 7064 7065 llvm::Type *RealFlagType 7066 = FlagOutPtr.getPointer()->getType()->getPointerElementType(); 7067 7068 llvm::Value *FlagExt = Builder.CreateZExt(Flag, RealFlagType); 7069 Builder.CreateStore(FlagExt, FlagOutPtr); 7070 return Result; 7071 } 7072 case AMDGPU::BI__builtin_amdgcn_div_fmas: 7073 case AMDGPU::BI__builtin_amdgcn_div_fmasf: { 7074 llvm::Value *Src0 = EmitScalarExpr(E->getArg(0)); 7075 llvm::Value *Src1 = EmitScalarExpr(E->getArg(1)); 7076 llvm::Value *Src2 = EmitScalarExpr(E->getArg(2)); 7077 llvm::Value *Src3 = EmitScalarExpr(E->getArg(3)); 7078 7079 llvm::Value *F = CGM.getIntrinsic(Intrinsic::amdgcn_div_fmas, 7080 Src0->getType()); 7081 llvm::Value *Src3ToBool = Builder.CreateIsNotNull(Src3); 7082 return Builder.CreateCall(F, {Src0, Src1, Src2, Src3ToBool}); 7083 } 7084 case AMDGPU::BI__builtin_amdgcn_div_fixup: 7085 case AMDGPU::BI__builtin_amdgcn_div_fixupf: 7086 return emitTernaryFPBuiltin(*this, E, Intrinsic::amdgcn_div_fixup); 7087 case AMDGPU::BI__builtin_amdgcn_trig_preop: 7088 case AMDGPU::BI__builtin_amdgcn_trig_preopf: 7089 return emitFPIntBuiltin(*this, E, Intrinsic::amdgcn_trig_preop); 7090 case AMDGPU::BI__builtin_amdgcn_rcp: 7091 case AMDGPU::BI__builtin_amdgcn_rcpf: 7092 return emitUnaryBuiltin(*this, E, Intrinsic::amdgcn_rcp); 7093 case AMDGPU::BI__builtin_amdgcn_rsq: 7094 case AMDGPU::BI__builtin_amdgcn_rsqf: 7095 return emitUnaryBuiltin(*this, E, Intrinsic::amdgcn_rsq); 7096 case AMDGPU::BI__builtin_amdgcn_rsq_clamp: 7097 case AMDGPU::BI__builtin_amdgcn_rsq_clampf: 7098 return emitUnaryBuiltin(*this, E, Intrinsic::amdgcn_rsq_clamp); 7099 case AMDGPU::BI__builtin_amdgcn_sinf: 7100 return emitUnaryBuiltin(*this, E, Intrinsic::amdgcn_sin); 7101 case AMDGPU::BI__builtin_amdgcn_cosf: 7102 return emitUnaryBuiltin(*this, E, Intrinsic::amdgcn_cos); 7103 case AMDGPU::BI__builtin_amdgcn_log_clampf: 7104 return emitUnaryBuiltin(*this, E, Intrinsic::amdgcn_log_clamp); 7105 case AMDGPU::BI__builtin_amdgcn_ldexp: 7106 case AMDGPU::BI__builtin_amdgcn_ldexpf: 7107 return emitFPIntBuiltin(*this, E, Intrinsic::amdgcn_ldexp); 7108 case AMDGPU::BI__builtin_amdgcn_class: 7109 case AMDGPU::BI__builtin_amdgcn_classf: 7110 return emitFPIntBuiltin(*this, E, Intrinsic::amdgcn_class); 7111 7112 // Legacy amdgpu prefix 7113 case AMDGPU::BI__builtin_amdgpu_rsq: 7114 case AMDGPU::BI__builtin_amdgpu_rsqf: { 7115 if (getTarget().getTriple().getArch() == Triple::amdgcn) 7116 return emitUnaryBuiltin(*this, E, Intrinsic::amdgcn_rsq); 7117 return emitUnaryBuiltin(*this, E, Intrinsic::r600_rsq); 7118 } 7119 case AMDGPU::BI__builtin_amdgpu_ldexp: 7120 case AMDGPU::BI__builtin_amdgpu_ldexpf: { 7121 if (getTarget().getTriple().getArch() == Triple::amdgcn) 7122 return emitFPIntBuiltin(*this, E, Intrinsic::amdgcn_ldexp); 7123 return emitFPIntBuiltin(*this, E, Intrinsic::AMDGPU_ldexp); 7124 } 7125 default: 7126 return nullptr; 7127 } 7128 } 7129 7130 /// Handle a SystemZ function in which the final argument is a pointer 7131 /// to an int that receives the post-instruction CC value. At the LLVM level 7132 /// this is represented as a function that returns a {result, cc} pair. 7133 static Value *EmitSystemZIntrinsicWithCC(CodeGenFunction &CGF, 7134 unsigned IntrinsicID, 7135 const CallExpr *E) { 7136 unsigned NumArgs = E->getNumArgs() - 1; 7137 SmallVector<Value *, 8> Args(NumArgs); 7138 for (unsigned I = 0; I < NumArgs; ++I) 7139 Args[I] = CGF.EmitScalarExpr(E->getArg(I)); 7140 Address CCPtr = CGF.EmitPointerWithAlignment(E->getArg(NumArgs)); 7141 Value *F = CGF.CGM.getIntrinsic(IntrinsicID); 7142 Value *Call = CGF.Builder.CreateCall(F, Args); 7143 Value *CC = CGF.Builder.CreateExtractValue(Call, 1); 7144 CGF.Builder.CreateStore(CC, CCPtr); 7145 return CGF.Builder.CreateExtractValue(Call, 0); 7146 } 7147 7148 Value *CodeGenFunction::EmitSystemZBuiltinExpr(unsigned BuiltinID, 7149 const CallExpr *E) { 7150 switch (BuiltinID) { 7151 case SystemZ::BI__builtin_tbegin: { 7152 Value *TDB = EmitScalarExpr(E->getArg(0)); 7153 Value *Control = llvm::ConstantInt::get(Int32Ty, 0xff0c); 7154 Value *F = CGM.getIntrinsic(Intrinsic::s390_tbegin); 7155 return Builder.CreateCall(F, {TDB, Control}); 7156 } 7157 case SystemZ::BI__builtin_tbegin_nofloat: { 7158 Value *TDB = EmitScalarExpr(E->getArg(0)); 7159 Value *Control = llvm::ConstantInt::get(Int32Ty, 0xff0c); 7160 Value *F = CGM.getIntrinsic(Intrinsic::s390_tbegin_nofloat); 7161 return Builder.CreateCall(F, {TDB, Control}); 7162 } 7163 case SystemZ::BI__builtin_tbeginc: { 7164 Value *TDB = llvm::ConstantPointerNull::get(Int8PtrTy); 7165 Value *Control = llvm::ConstantInt::get(Int32Ty, 0xff08); 7166 Value *F = CGM.getIntrinsic(Intrinsic::s390_tbeginc); 7167 return Builder.CreateCall(F, {TDB, Control}); 7168 } 7169 case SystemZ::BI__builtin_tabort: { 7170 Value *Data = EmitScalarExpr(E->getArg(0)); 7171 Value *F = CGM.getIntrinsic(Intrinsic::s390_tabort); 7172 return Builder.CreateCall(F, Builder.CreateSExt(Data, Int64Ty, "tabort")); 7173 } 7174 case SystemZ::BI__builtin_non_tx_store: { 7175 Value *Address = EmitScalarExpr(E->getArg(0)); 7176 Value *Data = EmitScalarExpr(E->getArg(1)); 7177 Value *F = CGM.getIntrinsic(Intrinsic::s390_ntstg); 7178 return Builder.CreateCall(F, {Data, Address}); 7179 } 7180 7181 // Vector builtins. Note that most vector builtins are mapped automatically 7182 // to target-specific LLVM intrinsics. The ones handled specially here can 7183 // be represented via standard LLVM IR, which is preferable to enable common 7184 // LLVM optimizations. 7185 7186 case SystemZ::BI__builtin_s390_vpopctb: 7187 case SystemZ::BI__builtin_s390_vpopcth: 7188 case SystemZ::BI__builtin_s390_vpopctf: 7189 case SystemZ::BI__builtin_s390_vpopctg: { 7190 llvm::Type *ResultType = ConvertType(E->getType()); 7191 Value *X = EmitScalarExpr(E->getArg(0)); 7192 Function *F = CGM.getIntrinsic(Intrinsic::ctpop, ResultType); 7193 return Builder.CreateCall(F, X); 7194 } 7195 7196 case SystemZ::BI__builtin_s390_vclzb: 7197 case SystemZ::BI__builtin_s390_vclzh: 7198 case SystemZ::BI__builtin_s390_vclzf: 7199 case SystemZ::BI__builtin_s390_vclzg: { 7200 llvm::Type *ResultType = ConvertType(E->getType()); 7201 Value *X = EmitScalarExpr(E->getArg(0)); 7202 Value *Undef = ConstantInt::get(Builder.getInt1Ty(), false); 7203 Function *F = CGM.getIntrinsic(Intrinsic::ctlz, ResultType); 7204 return Builder.CreateCall(F, {X, Undef}); 7205 } 7206 7207 case SystemZ::BI__builtin_s390_vctzb: 7208 case SystemZ::BI__builtin_s390_vctzh: 7209 case SystemZ::BI__builtin_s390_vctzf: 7210 case SystemZ::BI__builtin_s390_vctzg: { 7211 llvm::Type *ResultType = ConvertType(E->getType()); 7212 Value *X = EmitScalarExpr(E->getArg(0)); 7213 Value *Undef = ConstantInt::get(Builder.getInt1Ty(), false); 7214 Function *F = CGM.getIntrinsic(Intrinsic::cttz, ResultType); 7215 return Builder.CreateCall(F, {X, Undef}); 7216 } 7217 7218 case SystemZ::BI__builtin_s390_vfsqdb: { 7219 llvm::Type *ResultType = ConvertType(E->getType()); 7220 Value *X = EmitScalarExpr(E->getArg(0)); 7221 Function *F = CGM.getIntrinsic(Intrinsic::sqrt, ResultType); 7222 return Builder.CreateCall(F, X); 7223 } 7224 case SystemZ::BI__builtin_s390_vfmadb: { 7225 llvm::Type *ResultType = ConvertType(E->getType()); 7226 Value *X = EmitScalarExpr(E->getArg(0)); 7227 Value *Y = EmitScalarExpr(E->getArg(1)); 7228 Value *Z = EmitScalarExpr(E->getArg(2)); 7229 Function *F = CGM.getIntrinsic(Intrinsic::fma, ResultType); 7230 return Builder.CreateCall(F, {X, Y, Z}); 7231 } 7232 case SystemZ::BI__builtin_s390_vfmsdb: { 7233 llvm::Type *ResultType = ConvertType(E->getType()); 7234 Value *X = EmitScalarExpr(E->getArg(0)); 7235 Value *Y = EmitScalarExpr(E->getArg(1)); 7236 Value *Z = EmitScalarExpr(E->getArg(2)); 7237 Value *Zero = llvm::ConstantFP::getZeroValueForNegation(ResultType); 7238 Function *F = CGM.getIntrinsic(Intrinsic::fma, ResultType); 7239 return Builder.CreateCall(F, {X, Y, Builder.CreateFSub(Zero, Z, "sub")}); 7240 } 7241 case SystemZ::BI__builtin_s390_vflpdb: { 7242 llvm::Type *ResultType = ConvertType(E->getType()); 7243 Value *X = EmitScalarExpr(E->getArg(0)); 7244 Function *F = CGM.getIntrinsic(Intrinsic::fabs, ResultType); 7245 return Builder.CreateCall(F, X); 7246 } 7247 case SystemZ::BI__builtin_s390_vflndb: { 7248 llvm::Type *ResultType = ConvertType(E->getType()); 7249 Value *X = EmitScalarExpr(E->getArg(0)); 7250 Value *Zero = llvm::ConstantFP::getZeroValueForNegation(ResultType); 7251 Function *F = CGM.getIntrinsic(Intrinsic::fabs, ResultType); 7252 return Builder.CreateFSub(Zero, Builder.CreateCall(F, X), "sub"); 7253 } 7254 case SystemZ::BI__builtin_s390_vfidb: { 7255 llvm::Type *ResultType = ConvertType(E->getType()); 7256 Value *X = EmitScalarExpr(E->getArg(0)); 7257 // Constant-fold the M4 and M5 mask arguments. 7258 llvm::APSInt M4, M5; 7259 bool IsConstM4 = E->getArg(1)->isIntegerConstantExpr(M4, getContext()); 7260 bool IsConstM5 = E->getArg(2)->isIntegerConstantExpr(M5, getContext()); 7261 assert(IsConstM4 && IsConstM5 && "Constant arg isn't actually constant?"); 7262 (void)IsConstM4; (void)IsConstM5; 7263 // Check whether this instance of vfidb can be represented via a LLVM 7264 // standard intrinsic. We only support some combinations of M4 and M5. 7265 Intrinsic::ID ID = Intrinsic::not_intrinsic; 7266 switch (M4.getZExtValue()) { 7267 default: break; 7268 case 0: // IEEE-inexact exception allowed 7269 switch (M5.getZExtValue()) { 7270 default: break; 7271 case 0: ID = Intrinsic::rint; break; 7272 } 7273 break; 7274 case 4: // IEEE-inexact exception suppressed 7275 switch (M5.getZExtValue()) { 7276 default: break; 7277 case 0: ID = Intrinsic::nearbyint; break; 7278 case 1: ID = Intrinsic::round; break; 7279 case 5: ID = Intrinsic::trunc; break; 7280 case 6: ID = Intrinsic::ceil; break; 7281 case 7: ID = Intrinsic::floor; break; 7282 } 7283 break; 7284 } 7285 if (ID != Intrinsic::not_intrinsic) { 7286 Function *F = CGM.getIntrinsic(ID, ResultType); 7287 return Builder.CreateCall(F, X); 7288 } 7289 Function *F = CGM.getIntrinsic(Intrinsic::s390_vfidb); 7290 Value *M4Value = llvm::ConstantInt::get(getLLVMContext(), M4); 7291 Value *M5Value = llvm::ConstantInt::get(getLLVMContext(), M5); 7292 return Builder.CreateCall(F, {X, M4Value, M5Value}); 7293 } 7294 7295 // Vector intrisincs that output the post-instruction CC value. 7296 7297 #define INTRINSIC_WITH_CC(NAME) \ 7298 case SystemZ::BI__builtin_##NAME: \ 7299 return EmitSystemZIntrinsicWithCC(*this, Intrinsic::NAME, E) 7300 7301 INTRINSIC_WITH_CC(s390_vpkshs); 7302 INTRINSIC_WITH_CC(s390_vpksfs); 7303 INTRINSIC_WITH_CC(s390_vpksgs); 7304 7305 INTRINSIC_WITH_CC(s390_vpklshs); 7306 INTRINSIC_WITH_CC(s390_vpklsfs); 7307 INTRINSIC_WITH_CC(s390_vpklsgs); 7308 7309 INTRINSIC_WITH_CC(s390_vceqbs); 7310 INTRINSIC_WITH_CC(s390_vceqhs); 7311 INTRINSIC_WITH_CC(s390_vceqfs); 7312 INTRINSIC_WITH_CC(s390_vceqgs); 7313 7314 INTRINSIC_WITH_CC(s390_vchbs); 7315 INTRINSIC_WITH_CC(s390_vchhs); 7316 INTRINSIC_WITH_CC(s390_vchfs); 7317 INTRINSIC_WITH_CC(s390_vchgs); 7318 7319 INTRINSIC_WITH_CC(s390_vchlbs); 7320 INTRINSIC_WITH_CC(s390_vchlhs); 7321 INTRINSIC_WITH_CC(s390_vchlfs); 7322 INTRINSIC_WITH_CC(s390_vchlgs); 7323 7324 INTRINSIC_WITH_CC(s390_vfaebs); 7325 INTRINSIC_WITH_CC(s390_vfaehs); 7326 INTRINSIC_WITH_CC(s390_vfaefs); 7327 7328 INTRINSIC_WITH_CC(s390_vfaezbs); 7329 INTRINSIC_WITH_CC(s390_vfaezhs); 7330 INTRINSIC_WITH_CC(s390_vfaezfs); 7331 7332 INTRINSIC_WITH_CC(s390_vfeebs); 7333 INTRINSIC_WITH_CC(s390_vfeehs); 7334 INTRINSIC_WITH_CC(s390_vfeefs); 7335 7336 INTRINSIC_WITH_CC(s390_vfeezbs); 7337 INTRINSIC_WITH_CC(s390_vfeezhs); 7338 INTRINSIC_WITH_CC(s390_vfeezfs); 7339 7340 INTRINSIC_WITH_CC(s390_vfenebs); 7341 INTRINSIC_WITH_CC(s390_vfenehs); 7342 INTRINSIC_WITH_CC(s390_vfenefs); 7343 7344 INTRINSIC_WITH_CC(s390_vfenezbs); 7345 INTRINSIC_WITH_CC(s390_vfenezhs); 7346 INTRINSIC_WITH_CC(s390_vfenezfs); 7347 7348 INTRINSIC_WITH_CC(s390_vistrbs); 7349 INTRINSIC_WITH_CC(s390_vistrhs); 7350 INTRINSIC_WITH_CC(s390_vistrfs); 7351 7352 INTRINSIC_WITH_CC(s390_vstrcbs); 7353 INTRINSIC_WITH_CC(s390_vstrchs); 7354 INTRINSIC_WITH_CC(s390_vstrcfs); 7355 7356 INTRINSIC_WITH_CC(s390_vstrczbs); 7357 INTRINSIC_WITH_CC(s390_vstrczhs); 7358 INTRINSIC_WITH_CC(s390_vstrczfs); 7359 7360 INTRINSIC_WITH_CC(s390_vfcedbs); 7361 INTRINSIC_WITH_CC(s390_vfchdbs); 7362 INTRINSIC_WITH_CC(s390_vfchedbs); 7363 7364 INTRINSIC_WITH_CC(s390_vftcidb); 7365 7366 #undef INTRINSIC_WITH_CC 7367 7368 default: 7369 return nullptr; 7370 } 7371 } 7372 7373 Value *CodeGenFunction::EmitNVPTXBuiltinExpr(unsigned BuiltinID, 7374 const CallExpr *E) { 7375 switch (BuiltinID) { 7376 case NVPTX::BI__nvvm_atom_add_gen_i: 7377 case NVPTX::BI__nvvm_atom_add_gen_l: 7378 case NVPTX::BI__nvvm_atom_add_gen_ll: 7379 return MakeBinaryAtomicValue(*this, llvm::AtomicRMWInst::Add, E); 7380 7381 case NVPTX::BI__nvvm_atom_sub_gen_i: 7382 case NVPTX::BI__nvvm_atom_sub_gen_l: 7383 case NVPTX::BI__nvvm_atom_sub_gen_ll: 7384 return MakeBinaryAtomicValue(*this, llvm::AtomicRMWInst::Sub, E); 7385 7386 case NVPTX::BI__nvvm_atom_and_gen_i: 7387 case NVPTX::BI__nvvm_atom_and_gen_l: 7388 case NVPTX::BI__nvvm_atom_and_gen_ll: 7389 return MakeBinaryAtomicValue(*this, llvm::AtomicRMWInst::And, E); 7390 7391 case NVPTX::BI__nvvm_atom_or_gen_i: 7392 case NVPTX::BI__nvvm_atom_or_gen_l: 7393 case NVPTX::BI__nvvm_atom_or_gen_ll: 7394 return MakeBinaryAtomicValue(*this, llvm::AtomicRMWInst::Or, E); 7395 7396 case NVPTX::BI__nvvm_atom_xor_gen_i: 7397 case NVPTX::BI__nvvm_atom_xor_gen_l: 7398 case NVPTX::BI__nvvm_atom_xor_gen_ll: 7399 return MakeBinaryAtomicValue(*this, llvm::AtomicRMWInst::Xor, E); 7400 7401 case NVPTX::BI__nvvm_atom_xchg_gen_i: 7402 case NVPTX::BI__nvvm_atom_xchg_gen_l: 7403 case NVPTX::BI__nvvm_atom_xchg_gen_ll: 7404 return MakeBinaryAtomicValue(*this, llvm::AtomicRMWInst::Xchg, E); 7405 7406 case NVPTX::BI__nvvm_atom_max_gen_i: 7407 case NVPTX::BI__nvvm_atom_max_gen_l: 7408 case NVPTX::BI__nvvm_atom_max_gen_ll: 7409 return MakeBinaryAtomicValue(*this, llvm::AtomicRMWInst::Max, E); 7410 7411 case NVPTX::BI__nvvm_atom_max_gen_ui: 7412 case NVPTX::BI__nvvm_atom_max_gen_ul: 7413 case NVPTX::BI__nvvm_atom_max_gen_ull: 7414 return MakeBinaryAtomicValue(*this, llvm::AtomicRMWInst::UMax, E); 7415 7416 case NVPTX::BI__nvvm_atom_min_gen_i: 7417 case NVPTX::BI__nvvm_atom_min_gen_l: 7418 case NVPTX::BI__nvvm_atom_min_gen_ll: 7419 return MakeBinaryAtomicValue(*this, llvm::AtomicRMWInst::Min, E); 7420 7421 case NVPTX::BI__nvvm_atom_min_gen_ui: 7422 case NVPTX::BI__nvvm_atom_min_gen_ul: 7423 case NVPTX::BI__nvvm_atom_min_gen_ull: 7424 return MakeBinaryAtomicValue(*this, llvm::AtomicRMWInst::UMin, E); 7425 7426 case NVPTX::BI__nvvm_atom_cas_gen_i: 7427 case NVPTX::BI__nvvm_atom_cas_gen_l: 7428 case NVPTX::BI__nvvm_atom_cas_gen_ll: 7429 // __nvvm_atom_cas_gen_* should return the old value rather than the 7430 // success flag. 7431 return MakeAtomicCmpXchgValue(*this, E, /*ReturnBool=*/false); 7432 7433 case NVPTX::BI__nvvm_atom_add_gen_f: { 7434 Value *Ptr = EmitScalarExpr(E->getArg(0)); 7435 Value *Val = EmitScalarExpr(E->getArg(1)); 7436 // atomicrmw only deals with integer arguments so we need to use 7437 // LLVM's nvvm_atomic_load_add_f32 intrinsic for that. 7438 Value *FnALAF32 = 7439 CGM.getIntrinsic(Intrinsic::nvvm_atomic_load_add_f32, Ptr->getType()); 7440 return Builder.CreateCall(FnALAF32, {Ptr, Val}); 7441 } 7442 7443 case NVPTX::BI__nvvm_atom_inc_gen_ui: { 7444 Value *Ptr = EmitScalarExpr(E->getArg(0)); 7445 Value *Val = EmitScalarExpr(E->getArg(1)); 7446 Value *FnALI32 = 7447 CGM.getIntrinsic(Intrinsic::nvvm_atomic_load_inc_32, Ptr->getType()); 7448 return Builder.CreateCall(FnALI32, {Ptr, Val}); 7449 } 7450 7451 case NVPTX::BI__nvvm_atom_dec_gen_ui: { 7452 Value *Ptr = EmitScalarExpr(E->getArg(0)); 7453 Value *Val = EmitScalarExpr(E->getArg(1)); 7454 Value *FnALD32 = 7455 CGM.getIntrinsic(Intrinsic::nvvm_atomic_load_dec_32, Ptr->getType()); 7456 return Builder.CreateCall(FnALD32, {Ptr, Val}); 7457 } 7458 7459 default: 7460 return nullptr; 7461 } 7462 } 7463 7464 Value *CodeGenFunction::EmitWebAssemblyBuiltinExpr(unsigned BuiltinID, 7465 const CallExpr *E) { 7466 switch (BuiltinID) { 7467 case WebAssembly::BI__builtin_wasm_memory_size: { 7468 llvm::Type *ResultType = ConvertType(E->getType()); 7469 Value *Callee = CGM.getIntrinsic(Intrinsic::wasm_memory_size, ResultType); 7470 return Builder.CreateCall(Callee); 7471 } 7472 case WebAssembly::BI__builtin_wasm_grow_memory: { 7473 Value *X = EmitScalarExpr(E->getArg(0)); 7474 Value *Callee = CGM.getIntrinsic(Intrinsic::wasm_grow_memory, X->getType()); 7475 return Builder.CreateCall(Callee, X); 7476 } 7477 7478 default: 7479 return nullptr; 7480 } 7481 } 7482