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 = CGF.Builder.CreateAtomicRMW( 109 Kind, Args[0], Args[1], llvm::AtomicOrdering::SequentiallyConsistent); 110 return EmitFromInt(CGF, Result, T, ValueType); 111 } 112 113 static Value *EmitNontemporalStore(CodeGenFunction &CGF, const CallExpr *E) { 114 Value *Val = CGF.EmitScalarExpr(E->getArg(0)); 115 Value *Address = CGF.EmitScalarExpr(E->getArg(1)); 116 117 // Convert the type of the pointer to a pointer to the stored type. 118 Val = CGF.EmitToMemory(Val, E->getArg(0)->getType()); 119 Value *BC = CGF.Builder.CreateBitCast( 120 Address, llvm::PointerType::getUnqual(Val->getType()), "cast"); 121 LValue LV = CGF.MakeNaturalAlignAddrLValue(BC, E->getArg(0)->getType()); 122 LV.setNontemporal(true); 123 CGF.EmitStoreOfScalar(Val, LV, false); 124 return nullptr; 125 } 126 127 static Value *EmitNontemporalLoad(CodeGenFunction &CGF, const CallExpr *E) { 128 Value *Address = CGF.EmitScalarExpr(E->getArg(0)); 129 130 LValue LV = CGF.MakeNaturalAlignAddrLValue(Address, E->getType()); 131 LV.setNontemporal(true); 132 return CGF.EmitLoadOfScalar(LV, E->getExprLoc()); 133 } 134 135 static RValue EmitBinaryAtomic(CodeGenFunction &CGF, 136 llvm::AtomicRMWInst::BinOp Kind, 137 const CallExpr *E) { 138 return RValue::get(MakeBinaryAtomicValue(CGF, Kind, E)); 139 } 140 141 /// Utility to insert an atomic instruction based Instrinsic::ID and 142 /// the expression node, where the return value is the result of the 143 /// operation. 144 static RValue EmitBinaryAtomicPost(CodeGenFunction &CGF, 145 llvm::AtomicRMWInst::BinOp Kind, 146 const CallExpr *E, 147 Instruction::BinaryOps Op, 148 bool Invert = false) { 149 QualType T = E->getType(); 150 assert(E->getArg(0)->getType()->isPointerType()); 151 assert(CGF.getContext().hasSameUnqualifiedType(T, 152 E->getArg(0)->getType()->getPointeeType())); 153 assert(CGF.getContext().hasSameUnqualifiedType(T, E->getArg(1)->getType())); 154 155 llvm::Value *DestPtr = CGF.EmitScalarExpr(E->getArg(0)); 156 unsigned AddrSpace = DestPtr->getType()->getPointerAddressSpace(); 157 158 llvm::IntegerType *IntType = 159 llvm::IntegerType::get(CGF.getLLVMContext(), 160 CGF.getContext().getTypeSize(T)); 161 llvm::Type *IntPtrType = IntType->getPointerTo(AddrSpace); 162 163 llvm::Value *Args[2]; 164 Args[1] = CGF.EmitScalarExpr(E->getArg(1)); 165 llvm::Type *ValueType = Args[1]->getType(); 166 Args[1] = EmitToInt(CGF, Args[1], T, IntType); 167 Args[0] = CGF.Builder.CreateBitCast(DestPtr, IntPtrType); 168 169 llvm::Value *Result = CGF.Builder.CreateAtomicRMW( 170 Kind, Args[0], Args[1], llvm::AtomicOrdering::SequentiallyConsistent); 171 Result = CGF.Builder.CreateBinOp(Op, Result, Args[1]); 172 if (Invert) 173 Result = CGF.Builder.CreateBinOp(llvm::Instruction::Xor, Result, 174 llvm::ConstantInt::get(IntType, -1)); 175 Result = EmitFromInt(CGF, Result, T, ValueType); 176 return RValue::get(Result); 177 } 178 179 /// @brief Utility to insert an atomic cmpxchg instruction. 180 /// 181 /// @param CGF The current codegen function. 182 /// @param E Builtin call expression to convert to cmpxchg. 183 /// arg0 - address to operate on 184 /// arg1 - value to compare with 185 /// arg2 - new value 186 /// @param ReturnBool Specifies whether to return success flag of 187 /// cmpxchg result or the old value. 188 /// 189 /// @returns result of cmpxchg, according to ReturnBool 190 static Value *MakeAtomicCmpXchgValue(CodeGenFunction &CGF, const CallExpr *E, 191 bool ReturnBool) { 192 QualType T = ReturnBool ? E->getArg(1)->getType() : E->getType(); 193 llvm::Value *DestPtr = CGF.EmitScalarExpr(E->getArg(0)); 194 unsigned AddrSpace = DestPtr->getType()->getPointerAddressSpace(); 195 196 llvm::IntegerType *IntType = llvm::IntegerType::get( 197 CGF.getLLVMContext(), CGF.getContext().getTypeSize(T)); 198 llvm::Type *IntPtrType = IntType->getPointerTo(AddrSpace); 199 200 Value *Args[3]; 201 Args[0] = CGF.Builder.CreateBitCast(DestPtr, IntPtrType); 202 Args[1] = CGF.EmitScalarExpr(E->getArg(1)); 203 llvm::Type *ValueType = Args[1]->getType(); 204 Args[1] = EmitToInt(CGF, Args[1], T, IntType); 205 Args[2] = EmitToInt(CGF, CGF.EmitScalarExpr(E->getArg(2)), T, IntType); 206 207 Value *Pair = CGF.Builder.CreateAtomicCmpXchg( 208 Args[0], Args[1], Args[2], llvm::AtomicOrdering::SequentiallyConsistent, 209 llvm::AtomicOrdering::SequentiallyConsistent); 210 if (ReturnBool) 211 // Extract boolean success flag and zext it to int. 212 return CGF.Builder.CreateZExt(CGF.Builder.CreateExtractValue(Pair, 1), 213 CGF.ConvertType(E->getType())); 214 else 215 // Extract old value and emit it using the same type as compare value. 216 return EmitFromInt(CGF, CGF.Builder.CreateExtractValue(Pair, 0), T, 217 ValueType); 218 } 219 220 /// EmitFAbs - Emit a call to @llvm.fabs(). 221 static Value *EmitFAbs(CodeGenFunction &CGF, Value *V) { 222 Value *F = CGF.CGM.getIntrinsic(Intrinsic::fabs, V->getType()); 223 llvm::CallInst *Call = CGF.Builder.CreateCall(F, V); 224 Call->setDoesNotAccessMemory(); 225 return Call; 226 } 227 228 /// Emit the computation of the sign bit for a floating point value. Returns 229 /// the i1 sign bit value. 230 static Value *EmitSignBit(CodeGenFunction &CGF, Value *V) { 231 LLVMContext &C = CGF.CGM.getLLVMContext(); 232 233 llvm::Type *Ty = V->getType(); 234 int Width = Ty->getPrimitiveSizeInBits(); 235 llvm::Type *IntTy = llvm::IntegerType::get(C, Width); 236 V = CGF.Builder.CreateBitCast(V, IntTy); 237 if (Ty->isPPC_FP128Ty()) { 238 // We want the sign bit of the higher-order double. The bitcast we just 239 // did works as if the double-double was stored to memory and then 240 // read as an i128. The "store" will put the higher-order double in the 241 // lower address in both little- and big-Endian modes, but the "load" 242 // will treat those bits as a different part of the i128: the low bits in 243 // little-Endian, the high bits in big-Endian. Therefore, on big-Endian 244 // we need to shift the high bits down to the low before truncating. 245 Width >>= 1; 246 if (CGF.getTarget().isBigEndian()) { 247 Value *ShiftCst = llvm::ConstantInt::get(IntTy, Width); 248 V = CGF.Builder.CreateLShr(V, ShiftCst); 249 } 250 // We are truncating value in order to extract the higher-order 251 // double, which we will be using to extract the sign from. 252 IntTy = llvm::IntegerType::get(C, Width); 253 V = CGF.Builder.CreateTrunc(V, IntTy); 254 } 255 Value *Zero = llvm::Constant::getNullValue(IntTy); 256 return CGF.Builder.CreateICmpSLT(V, Zero); 257 } 258 259 static RValue emitLibraryCall(CodeGenFunction &CGF, const FunctionDecl *Fn, 260 const CallExpr *E, llvm::Value *calleeValue) { 261 return CGF.EmitCall(E->getCallee()->getType(), calleeValue, E, 262 ReturnValueSlot(), Fn); 263 } 264 265 /// \brief Emit a call to llvm.{sadd,uadd,ssub,usub,smul,umul}.with.overflow.* 266 /// depending on IntrinsicID. 267 /// 268 /// \arg CGF The current codegen function. 269 /// \arg IntrinsicID The ID for the Intrinsic we wish to generate. 270 /// \arg X The first argument to the llvm.*.with.overflow.*. 271 /// \arg Y The second argument to the llvm.*.with.overflow.*. 272 /// \arg Carry The carry returned by the llvm.*.with.overflow.*. 273 /// \returns The result (i.e. sum/product) returned by the intrinsic. 274 static llvm::Value *EmitOverflowIntrinsic(CodeGenFunction &CGF, 275 const llvm::Intrinsic::ID IntrinsicID, 276 llvm::Value *X, llvm::Value *Y, 277 llvm::Value *&Carry) { 278 // Make sure we have integers of the same width. 279 assert(X->getType() == Y->getType() && 280 "Arguments must be the same type. (Did you forget to make sure both " 281 "arguments have the same integer width?)"); 282 283 llvm::Value *Callee = CGF.CGM.getIntrinsic(IntrinsicID, X->getType()); 284 llvm::Value *Tmp = CGF.Builder.CreateCall(Callee, {X, Y}); 285 Carry = CGF.Builder.CreateExtractValue(Tmp, 1); 286 return CGF.Builder.CreateExtractValue(Tmp, 0); 287 } 288 289 // Emit a simple mangled intrinsic that has 1 argument and a return type 290 // matching the argument type. 291 static Value *emitUnaryBuiltin(CodeGenFunction &CGF, 292 const CallExpr *E, 293 unsigned IntrinsicID) { 294 llvm::Value *Src0 = CGF.EmitScalarExpr(E->getArg(0)); 295 296 Value *F = CGF.CGM.getIntrinsic(IntrinsicID, Src0->getType()); 297 return CGF.Builder.CreateCall(F, Src0); 298 } 299 300 // Emit an intrinsic that has 3 float or double operands. 301 static Value *emitTernaryFPBuiltin(CodeGenFunction &CGF, 302 const CallExpr *E, 303 unsigned IntrinsicID) { 304 llvm::Value *Src0 = CGF.EmitScalarExpr(E->getArg(0)); 305 llvm::Value *Src1 = CGF.EmitScalarExpr(E->getArg(1)); 306 llvm::Value *Src2 = CGF.EmitScalarExpr(E->getArg(2)); 307 308 Value *F = CGF.CGM.getIntrinsic(IntrinsicID, Src0->getType()); 309 return CGF.Builder.CreateCall(F, {Src0, Src1, Src2}); 310 } 311 312 // Emit an intrinsic that has 1 float or double operand, and 1 integer. 313 static Value *emitFPIntBuiltin(CodeGenFunction &CGF, 314 const CallExpr *E, 315 unsigned IntrinsicID) { 316 llvm::Value *Src0 = CGF.EmitScalarExpr(E->getArg(0)); 317 llvm::Value *Src1 = CGF.EmitScalarExpr(E->getArg(1)); 318 319 Value *F = CGF.CGM.getIntrinsic(IntrinsicID, Src0->getType()); 320 return CGF.Builder.CreateCall(F, {Src0, Src1}); 321 } 322 323 namespace { 324 struct WidthAndSignedness { 325 unsigned Width; 326 bool Signed; 327 }; 328 } 329 330 static WidthAndSignedness 331 getIntegerWidthAndSignedness(const clang::ASTContext &context, 332 const clang::QualType Type) { 333 assert(Type->isIntegerType() && "Given type is not an integer."); 334 unsigned Width = Type->isBooleanType() ? 1 : context.getTypeInfo(Type).Width; 335 bool Signed = Type->isSignedIntegerType(); 336 return {Width, Signed}; 337 } 338 339 // Given one or more integer types, this function produces an integer type that 340 // encompasses them: any value in one of the given types could be expressed in 341 // the encompassing type. 342 static struct WidthAndSignedness 343 EncompassingIntegerType(ArrayRef<struct WidthAndSignedness> Types) { 344 assert(Types.size() > 0 && "Empty list of types."); 345 346 // If any of the given types is signed, we must return a signed type. 347 bool Signed = false; 348 for (const auto &Type : Types) { 349 Signed |= Type.Signed; 350 } 351 352 // The encompassing type must have a width greater than or equal to the width 353 // of the specified types. Aditionally, if the encompassing type is signed, 354 // its width must be strictly greater than the width of any unsigned types 355 // given. 356 unsigned Width = 0; 357 for (const auto &Type : Types) { 358 unsigned MinWidth = Type.Width + (Signed && !Type.Signed); 359 if (Width < MinWidth) { 360 Width = MinWidth; 361 } 362 } 363 364 return {Width, Signed}; 365 } 366 367 Value *CodeGenFunction::EmitVAStartEnd(Value *ArgValue, bool IsStart) { 368 llvm::Type *DestType = Int8PtrTy; 369 if (ArgValue->getType() != DestType) 370 ArgValue = 371 Builder.CreateBitCast(ArgValue, DestType, ArgValue->getName().data()); 372 373 Intrinsic::ID inst = IsStart ? Intrinsic::vastart : Intrinsic::vaend; 374 return Builder.CreateCall(CGM.getIntrinsic(inst), ArgValue); 375 } 376 377 /// Checks if using the result of __builtin_object_size(p, @p From) in place of 378 /// __builtin_object_size(p, @p To) is correct 379 static bool areBOSTypesCompatible(int From, int To) { 380 // Note: Our __builtin_object_size implementation currently treats Type=0 and 381 // Type=2 identically. Encoding this implementation detail here may make 382 // improving __builtin_object_size difficult in the future, so it's omitted. 383 return From == To || (From == 0 && To == 1) || (From == 3 && To == 2); 384 } 385 386 static llvm::Value * 387 getDefaultBuiltinObjectSizeResult(unsigned Type, llvm::IntegerType *ResType) { 388 return ConstantInt::get(ResType, (Type & 2) ? 0 : -1, /*isSigned=*/true); 389 } 390 391 llvm::Value * 392 CodeGenFunction::evaluateOrEmitBuiltinObjectSize(const Expr *E, unsigned Type, 393 llvm::IntegerType *ResType) { 394 uint64_t ObjectSize; 395 if (!E->tryEvaluateObjectSize(ObjectSize, getContext(), Type)) 396 return emitBuiltinObjectSize(E, Type, ResType); 397 return ConstantInt::get(ResType, ObjectSize, /*isSigned=*/true); 398 } 399 400 /// Returns a Value corresponding to the size of the given expression. 401 /// This Value may be either of the following: 402 /// - A llvm::Argument (if E is a param with the pass_object_size attribute on 403 /// it) 404 /// - A call to the @llvm.objectsize intrinsic 405 llvm::Value * 406 CodeGenFunction::emitBuiltinObjectSize(const Expr *E, unsigned Type, 407 llvm::IntegerType *ResType) { 408 // We need to reference an argument if the pointer is a parameter with the 409 // pass_object_size attribute. 410 if (auto *D = dyn_cast<DeclRefExpr>(E->IgnoreParenImpCasts())) { 411 auto *Param = dyn_cast<ParmVarDecl>(D->getDecl()); 412 auto *PS = D->getDecl()->getAttr<PassObjectSizeAttr>(); 413 if (Param != nullptr && PS != nullptr && 414 areBOSTypesCompatible(PS->getType(), Type)) { 415 auto Iter = SizeArguments.find(Param); 416 assert(Iter != SizeArguments.end()); 417 418 const ImplicitParamDecl *D = Iter->second; 419 auto DIter = LocalDeclMap.find(D); 420 assert(DIter != LocalDeclMap.end()); 421 422 return EmitLoadOfScalar(DIter->second, /*volatile=*/false, 423 getContext().getSizeType(), E->getLocStart()); 424 } 425 } 426 427 // LLVM can't handle Type=3 appropriately, and __builtin_object_size shouldn't 428 // evaluate E for side-effects. In either case, we shouldn't lower to 429 // @llvm.objectsize. 430 if (Type == 3 || E->HasSideEffects(getContext())) 431 return getDefaultBuiltinObjectSizeResult(Type, ResType); 432 433 // LLVM only supports 0 and 2, make sure that we pass along that 434 // as a boolean. 435 auto *CI = ConstantInt::get(Builder.getInt1Ty(), (Type & 2) >> 1); 436 // FIXME: Get right address space. 437 llvm::Type *Tys[] = {ResType, Builder.getInt8PtrTy(0)}; 438 Value *F = CGM.getIntrinsic(Intrinsic::objectsize, Tys); 439 return Builder.CreateCall(F, {EmitScalarExpr(E), CI}); 440 } 441 442 RValue CodeGenFunction::EmitBuiltinExpr(const FunctionDecl *FD, 443 unsigned BuiltinID, const CallExpr *E, 444 ReturnValueSlot ReturnValue) { 445 // See if we can constant fold this builtin. If so, don't emit it at all. 446 Expr::EvalResult Result; 447 if (E->EvaluateAsRValue(Result, CGM.getContext()) && 448 !Result.hasSideEffects()) { 449 if (Result.Val.isInt()) 450 return RValue::get(llvm::ConstantInt::get(getLLVMContext(), 451 Result.Val.getInt())); 452 if (Result.Val.isFloat()) 453 return RValue::get(llvm::ConstantFP::get(getLLVMContext(), 454 Result.Val.getFloat())); 455 } 456 457 switch (BuiltinID) { 458 default: break; // Handle intrinsics and libm functions below. 459 case Builtin::BI__builtin___CFStringMakeConstantString: 460 case Builtin::BI__builtin___NSStringMakeConstantString: 461 return RValue::get(CGM.EmitConstantExpr(E, E->getType(), nullptr)); 462 case Builtin::BI__builtin_stdarg_start: 463 case Builtin::BI__builtin_va_start: 464 case Builtin::BI__va_start: 465 case Builtin::BI__builtin_va_end: 466 return RValue::get( 467 EmitVAStartEnd(BuiltinID == Builtin::BI__va_start 468 ? EmitScalarExpr(E->getArg(0)) 469 : EmitVAListRef(E->getArg(0)).getPointer(), 470 BuiltinID != Builtin::BI__builtin_va_end)); 471 case Builtin::BI__builtin_va_copy: { 472 Value *DstPtr = EmitVAListRef(E->getArg(0)).getPointer(); 473 Value *SrcPtr = EmitVAListRef(E->getArg(1)).getPointer(); 474 475 llvm::Type *Type = Int8PtrTy; 476 477 DstPtr = Builder.CreateBitCast(DstPtr, Type); 478 SrcPtr = Builder.CreateBitCast(SrcPtr, Type); 479 return RValue::get(Builder.CreateCall(CGM.getIntrinsic(Intrinsic::vacopy), 480 {DstPtr, SrcPtr})); 481 } 482 case Builtin::BI__builtin_abs: 483 case Builtin::BI__builtin_labs: 484 case Builtin::BI__builtin_llabs: { 485 Value *ArgValue = EmitScalarExpr(E->getArg(0)); 486 487 Value *NegOp = Builder.CreateNeg(ArgValue, "neg"); 488 Value *CmpResult = 489 Builder.CreateICmpSGE(ArgValue, 490 llvm::Constant::getNullValue(ArgValue->getType()), 491 "abscond"); 492 Value *Result = 493 Builder.CreateSelect(CmpResult, ArgValue, NegOp, "abs"); 494 495 return RValue::get(Result); 496 } 497 case Builtin::BI__builtin_fabs: 498 case Builtin::BI__builtin_fabsf: 499 case Builtin::BI__builtin_fabsl: { 500 Value *Arg1 = EmitScalarExpr(E->getArg(0)); 501 Value *Result = EmitFAbs(*this, Arg1); 502 return RValue::get(Result); 503 } 504 case Builtin::BI__builtin_fmod: 505 case Builtin::BI__builtin_fmodf: 506 case Builtin::BI__builtin_fmodl: { 507 Value *Arg1 = EmitScalarExpr(E->getArg(0)); 508 Value *Arg2 = EmitScalarExpr(E->getArg(1)); 509 Value *Result = Builder.CreateFRem(Arg1, Arg2, "fmod"); 510 return RValue::get(Result); 511 } 512 513 case Builtin::BI__builtin_conj: 514 case Builtin::BI__builtin_conjf: 515 case Builtin::BI__builtin_conjl: { 516 ComplexPairTy ComplexVal = EmitComplexExpr(E->getArg(0)); 517 Value *Real = ComplexVal.first; 518 Value *Imag = ComplexVal.second; 519 Value *Zero = 520 Imag->getType()->isFPOrFPVectorTy() 521 ? llvm::ConstantFP::getZeroValueForNegation(Imag->getType()) 522 : llvm::Constant::getNullValue(Imag->getType()); 523 524 Imag = Builder.CreateFSub(Zero, Imag, "sub"); 525 return RValue::getComplex(std::make_pair(Real, Imag)); 526 } 527 case Builtin::BI__builtin_creal: 528 case Builtin::BI__builtin_crealf: 529 case Builtin::BI__builtin_creall: 530 case Builtin::BIcreal: 531 case Builtin::BIcrealf: 532 case Builtin::BIcreall: { 533 ComplexPairTy ComplexVal = EmitComplexExpr(E->getArg(0)); 534 return RValue::get(ComplexVal.first); 535 } 536 537 case Builtin::BI__builtin_cimag: 538 case Builtin::BI__builtin_cimagf: 539 case Builtin::BI__builtin_cimagl: 540 case Builtin::BIcimag: 541 case Builtin::BIcimagf: 542 case Builtin::BIcimagl: { 543 ComplexPairTy ComplexVal = EmitComplexExpr(E->getArg(0)); 544 return RValue::get(ComplexVal.second); 545 } 546 547 case Builtin::BI__builtin_ctzs: 548 case Builtin::BI__builtin_ctz: 549 case Builtin::BI__builtin_ctzl: 550 case Builtin::BI__builtin_ctzll: { 551 Value *ArgValue = EmitScalarExpr(E->getArg(0)); 552 553 llvm::Type *ArgType = ArgValue->getType(); 554 Value *F = CGM.getIntrinsic(Intrinsic::cttz, ArgType); 555 556 llvm::Type *ResultType = ConvertType(E->getType()); 557 Value *ZeroUndef = Builder.getInt1(getTarget().isCLZForZeroUndef()); 558 Value *Result = Builder.CreateCall(F, {ArgValue, ZeroUndef}); 559 if (Result->getType() != ResultType) 560 Result = Builder.CreateIntCast(Result, ResultType, /*isSigned*/true, 561 "cast"); 562 return RValue::get(Result); 563 } 564 case Builtin::BI__builtin_clzs: 565 case Builtin::BI__builtin_clz: 566 case Builtin::BI__builtin_clzl: 567 case Builtin::BI__builtin_clzll: { 568 Value *ArgValue = EmitScalarExpr(E->getArg(0)); 569 570 llvm::Type *ArgType = ArgValue->getType(); 571 Value *F = CGM.getIntrinsic(Intrinsic::ctlz, ArgType); 572 573 llvm::Type *ResultType = ConvertType(E->getType()); 574 Value *ZeroUndef = Builder.getInt1(getTarget().isCLZForZeroUndef()); 575 Value *Result = Builder.CreateCall(F, {ArgValue, ZeroUndef}); 576 if (Result->getType() != ResultType) 577 Result = Builder.CreateIntCast(Result, ResultType, /*isSigned*/true, 578 "cast"); 579 return RValue::get(Result); 580 } 581 case Builtin::BI__builtin_ffs: 582 case Builtin::BI__builtin_ffsl: 583 case Builtin::BI__builtin_ffsll: { 584 // ffs(x) -> x ? cttz(x) + 1 : 0 585 Value *ArgValue = EmitScalarExpr(E->getArg(0)); 586 587 llvm::Type *ArgType = ArgValue->getType(); 588 Value *F = CGM.getIntrinsic(Intrinsic::cttz, ArgType); 589 590 llvm::Type *ResultType = ConvertType(E->getType()); 591 Value *Tmp = 592 Builder.CreateAdd(Builder.CreateCall(F, {ArgValue, Builder.getTrue()}), 593 llvm::ConstantInt::get(ArgType, 1)); 594 Value *Zero = llvm::Constant::getNullValue(ArgType); 595 Value *IsZero = Builder.CreateICmpEQ(ArgValue, Zero, "iszero"); 596 Value *Result = Builder.CreateSelect(IsZero, Zero, Tmp, "ffs"); 597 if (Result->getType() != ResultType) 598 Result = Builder.CreateIntCast(Result, ResultType, /*isSigned*/true, 599 "cast"); 600 return RValue::get(Result); 601 } 602 case Builtin::BI__builtin_parity: 603 case Builtin::BI__builtin_parityl: 604 case Builtin::BI__builtin_parityll: { 605 // parity(x) -> ctpop(x) & 1 606 Value *ArgValue = EmitScalarExpr(E->getArg(0)); 607 608 llvm::Type *ArgType = ArgValue->getType(); 609 Value *F = CGM.getIntrinsic(Intrinsic::ctpop, ArgType); 610 611 llvm::Type *ResultType = ConvertType(E->getType()); 612 Value *Tmp = Builder.CreateCall(F, ArgValue); 613 Value *Result = Builder.CreateAnd(Tmp, llvm::ConstantInt::get(ArgType, 1)); 614 if (Result->getType() != ResultType) 615 Result = Builder.CreateIntCast(Result, ResultType, /*isSigned*/true, 616 "cast"); 617 return RValue::get(Result); 618 } 619 case Builtin::BI__builtin_popcount: 620 case Builtin::BI__builtin_popcountl: 621 case Builtin::BI__builtin_popcountll: { 622 Value *ArgValue = EmitScalarExpr(E->getArg(0)); 623 624 llvm::Type *ArgType = ArgValue->getType(); 625 Value *F = CGM.getIntrinsic(Intrinsic::ctpop, ArgType); 626 627 llvm::Type *ResultType = ConvertType(E->getType()); 628 Value *Result = Builder.CreateCall(F, ArgValue); 629 if (Result->getType() != ResultType) 630 Result = Builder.CreateIntCast(Result, ResultType, /*isSigned*/true, 631 "cast"); 632 return RValue::get(Result); 633 } 634 case Builtin::BI__builtin_unpredictable: { 635 // Always return the argument of __builtin_unpredictable. LLVM does not 636 // handle this builtin. Metadata for this builtin should be added directly 637 // to instructions such as branches or switches that use it. 638 return RValue::get(EmitScalarExpr(E->getArg(0))); 639 } 640 case Builtin::BI__builtin_expect: { 641 Value *ArgValue = EmitScalarExpr(E->getArg(0)); 642 llvm::Type *ArgType = ArgValue->getType(); 643 644 Value *ExpectedValue = EmitScalarExpr(E->getArg(1)); 645 // Don't generate llvm.expect on -O0 as the backend won't use it for 646 // anything. 647 // Note, we still IRGen ExpectedValue because it could have side-effects. 648 if (CGM.getCodeGenOpts().OptimizationLevel == 0) 649 return RValue::get(ArgValue); 650 651 Value *FnExpect = CGM.getIntrinsic(Intrinsic::expect, ArgType); 652 Value *Result = 653 Builder.CreateCall(FnExpect, {ArgValue, ExpectedValue}, "expval"); 654 return RValue::get(Result); 655 } 656 case Builtin::BI__builtin_assume_aligned: { 657 Value *PtrValue = EmitScalarExpr(E->getArg(0)); 658 Value *OffsetValue = 659 (E->getNumArgs() > 2) ? EmitScalarExpr(E->getArg(2)) : nullptr; 660 661 Value *AlignmentValue = EmitScalarExpr(E->getArg(1)); 662 ConstantInt *AlignmentCI = cast<ConstantInt>(AlignmentValue); 663 unsigned Alignment = (unsigned) AlignmentCI->getZExtValue(); 664 665 EmitAlignmentAssumption(PtrValue, Alignment, OffsetValue); 666 return RValue::get(PtrValue); 667 } 668 case Builtin::BI__assume: 669 case Builtin::BI__builtin_assume: { 670 if (E->getArg(0)->HasSideEffects(getContext())) 671 return RValue::get(nullptr); 672 673 Value *ArgValue = EmitScalarExpr(E->getArg(0)); 674 Value *FnAssume = CGM.getIntrinsic(Intrinsic::assume); 675 return RValue::get(Builder.CreateCall(FnAssume, ArgValue)); 676 } 677 case Builtin::BI__builtin_bswap16: 678 case Builtin::BI__builtin_bswap32: 679 case Builtin::BI__builtin_bswap64: { 680 return RValue::get(emitUnaryBuiltin(*this, E, Intrinsic::bswap)); 681 } 682 case Builtin::BI__builtin_bitreverse8: 683 case Builtin::BI__builtin_bitreverse16: 684 case Builtin::BI__builtin_bitreverse32: 685 case Builtin::BI__builtin_bitreverse64: { 686 return RValue::get(emitUnaryBuiltin(*this, E, Intrinsic::bitreverse)); 687 } 688 case Builtin::BI__builtin_object_size: { 689 unsigned Type = 690 E->getArg(1)->EvaluateKnownConstInt(getContext()).getZExtValue(); 691 auto *ResType = cast<llvm::IntegerType>(ConvertType(E->getType())); 692 693 // We pass this builtin onto the optimizer so that it can figure out the 694 // object size in more complex cases. 695 return RValue::get(emitBuiltinObjectSize(E->getArg(0), Type, ResType)); 696 } 697 case Builtin::BI__builtin_prefetch: { 698 Value *Locality, *RW, *Address = EmitScalarExpr(E->getArg(0)); 699 // FIXME: Technically these constants should of type 'int', yes? 700 RW = (E->getNumArgs() > 1) ? EmitScalarExpr(E->getArg(1)) : 701 llvm::ConstantInt::get(Int32Ty, 0); 702 Locality = (E->getNumArgs() > 2) ? EmitScalarExpr(E->getArg(2)) : 703 llvm::ConstantInt::get(Int32Ty, 3); 704 Value *Data = llvm::ConstantInt::get(Int32Ty, 1); 705 Value *F = CGM.getIntrinsic(Intrinsic::prefetch); 706 return RValue::get(Builder.CreateCall(F, {Address, RW, Locality, Data})); 707 } 708 case Builtin::BI__builtin_readcyclecounter: { 709 Value *F = CGM.getIntrinsic(Intrinsic::readcyclecounter); 710 return RValue::get(Builder.CreateCall(F)); 711 } 712 case Builtin::BI__builtin___clear_cache: { 713 Value *Begin = EmitScalarExpr(E->getArg(0)); 714 Value *End = EmitScalarExpr(E->getArg(1)); 715 Value *F = CGM.getIntrinsic(Intrinsic::clear_cache); 716 return RValue::get(Builder.CreateCall(F, {Begin, End})); 717 } 718 case Builtin::BI__builtin_trap: 719 return RValue::get(EmitTrapCall(Intrinsic::trap)); 720 case Builtin::BI__debugbreak: 721 return RValue::get(EmitTrapCall(Intrinsic::debugtrap)); 722 case Builtin::BI__builtin_unreachable: { 723 if (SanOpts.has(SanitizerKind::Unreachable)) { 724 SanitizerScope SanScope(this); 725 EmitCheck(std::make_pair(static_cast<llvm::Value *>(Builder.getFalse()), 726 SanitizerKind::Unreachable), 727 "builtin_unreachable", EmitCheckSourceLocation(E->getExprLoc()), 728 None); 729 } else 730 Builder.CreateUnreachable(); 731 732 // We do need to preserve an insertion point. 733 EmitBlock(createBasicBlock("unreachable.cont")); 734 735 return RValue::get(nullptr); 736 } 737 738 case Builtin::BI__builtin_powi: 739 case Builtin::BI__builtin_powif: 740 case Builtin::BI__builtin_powil: { 741 Value *Base = EmitScalarExpr(E->getArg(0)); 742 Value *Exponent = EmitScalarExpr(E->getArg(1)); 743 llvm::Type *ArgType = Base->getType(); 744 Value *F = CGM.getIntrinsic(Intrinsic::powi, ArgType); 745 return RValue::get(Builder.CreateCall(F, {Base, Exponent})); 746 } 747 748 case Builtin::BI__builtin_isgreater: 749 case Builtin::BI__builtin_isgreaterequal: 750 case Builtin::BI__builtin_isless: 751 case Builtin::BI__builtin_islessequal: 752 case Builtin::BI__builtin_islessgreater: 753 case Builtin::BI__builtin_isunordered: { 754 // Ordered comparisons: we know the arguments to these are matching scalar 755 // floating point values. 756 Value *LHS = EmitScalarExpr(E->getArg(0)); 757 Value *RHS = EmitScalarExpr(E->getArg(1)); 758 759 switch (BuiltinID) { 760 default: llvm_unreachable("Unknown ordered comparison"); 761 case Builtin::BI__builtin_isgreater: 762 LHS = Builder.CreateFCmpOGT(LHS, RHS, "cmp"); 763 break; 764 case Builtin::BI__builtin_isgreaterequal: 765 LHS = Builder.CreateFCmpOGE(LHS, RHS, "cmp"); 766 break; 767 case Builtin::BI__builtin_isless: 768 LHS = Builder.CreateFCmpOLT(LHS, RHS, "cmp"); 769 break; 770 case Builtin::BI__builtin_islessequal: 771 LHS = Builder.CreateFCmpOLE(LHS, RHS, "cmp"); 772 break; 773 case Builtin::BI__builtin_islessgreater: 774 LHS = Builder.CreateFCmpONE(LHS, RHS, "cmp"); 775 break; 776 case Builtin::BI__builtin_isunordered: 777 LHS = Builder.CreateFCmpUNO(LHS, RHS, "cmp"); 778 break; 779 } 780 // ZExt bool to int type. 781 return RValue::get(Builder.CreateZExt(LHS, ConvertType(E->getType()))); 782 } 783 case Builtin::BI__builtin_isnan: { 784 Value *V = EmitScalarExpr(E->getArg(0)); 785 V = Builder.CreateFCmpUNO(V, V, "cmp"); 786 return RValue::get(Builder.CreateZExt(V, ConvertType(E->getType()))); 787 } 788 789 case Builtin::BI__builtin_isinf: 790 case Builtin::BI__builtin_isfinite: { 791 // isinf(x) --> fabs(x) == infinity 792 // isfinite(x) --> fabs(x) != infinity 793 // x != NaN via the ordered compare in either case. 794 Value *V = EmitScalarExpr(E->getArg(0)); 795 Value *Fabs = EmitFAbs(*this, V); 796 Constant *Infinity = ConstantFP::getInfinity(V->getType()); 797 CmpInst::Predicate Pred = (BuiltinID == Builtin::BI__builtin_isinf) 798 ? CmpInst::FCMP_OEQ 799 : CmpInst::FCMP_ONE; 800 Value *FCmp = Builder.CreateFCmp(Pred, Fabs, Infinity, "cmpinf"); 801 return RValue::get(Builder.CreateZExt(FCmp, ConvertType(E->getType()))); 802 } 803 804 case Builtin::BI__builtin_isinf_sign: { 805 // isinf_sign(x) -> fabs(x) == infinity ? (signbit(x) ? -1 : 1) : 0 806 Value *Arg = EmitScalarExpr(E->getArg(0)); 807 Value *AbsArg = EmitFAbs(*this, Arg); 808 Value *IsInf = Builder.CreateFCmpOEQ( 809 AbsArg, ConstantFP::getInfinity(Arg->getType()), "isinf"); 810 Value *IsNeg = EmitSignBit(*this, Arg); 811 812 llvm::Type *IntTy = ConvertType(E->getType()); 813 Value *Zero = Constant::getNullValue(IntTy); 814 Value *One = ConstantInt::get(IntTy, 1); 815 Value *NegativeOne = ConstantInt::get(IntTy, -1); 816 Value *SignResult = Builder.CreateSelect(IsNeg, NegativeOne, One); 817 Value *Result = Builder.CreateSelect(IsInf, SignResult, Zero); 818 return RValue::get(Result); 819 } 820 821 case Builtin::BI__builtin_isnormal: { 822 // isnormal(x) --> x == x && fabsf(x) < infinity && fabsf(x) >= float_min 823 Value *V = EmitScalarExpr(E->getArg(0)); 824 Value *Eq = Builder.CreateFCmpOEQ(V, V, "iseq"); 825 826 Value *Abs = EmitFAbs(*this, V); 827 Value *IsLessThanInf = 828 Builder.CreateFCmpULT(Abs, ConstantFP::getInfinity(V->getType()),"isinf"); 829 APFloat Smallest = APFloat::getSmallestNormalized( 830 getContext().getFloatTypeSemantics(E->getArg(0)->getType())); 831 Value *IsNormal = 832 Builder.CreateFCmpUGE(Abs, ConstantFP::get(V->getContext(), Smallest), 833 "isnormal"); 834 V = Builder.CreateAnd(Eq, IsLessThanInf, "and"); 835 V = Builder.CreateAnd(V, IsNormal, "and"); 836 return RValue::get(Builder.CreateZExt(V, ConvertType(E->getType()))); 837 } 838 839 case Builtin::BI__builtin_fpclassify: { 840 Value *V = EmitScalarExpr(E->getArg(5)); 841 llvm::Type *Ty = ConvertType(E->getArg(5)->getType()); 842 843 // Create Result 844 BasicBlock *Begin = Builder.GetInsertBlock(); 845 BasicBlock *End = createBasicBlock("fpclassify_end", this->CurFn); 846 Builder.SetInsertPoint(End); 847 PHINode *Result = 848 Builder.CreatePHI(ConvertType(E->getArg(0)->getType()), 4, 849 "fpclassify_result"); 850 851 // if (V==0) return FP_ZERO 852 Builder.SetInsertPoint(Begin); 853 Value *IsZero = Builder.CreateFCmpOEQ(V, Constant::getNullValue(Ty), 854 "iszero"); 855 Value *ZeroLiteral = EmitScalarExpr(E->getArg(4)); 856 BasicBlock *NotZero = createBasicBlock("fpclassify_not_zero", this->CurFn); 857 Builder.CreateCondBr(IsZero, End, NotZero); 858 Result->addIncoming(ZeroLiteral, Begin); 859 860 // if (V != V) return FP_NAN 861 Builder.SetInsertPoint(NotZero); 862 Value *IsNan = Builder.CreateFCmpUNO(V, V, "cmp"); 863 Value *NanLiteral = EmitScalarExpr(E->getArg(0)); 864 BasicBlock *NotNan = createBasicBlock("fpclassify_not_nan", this->CurFn); 865 Builder.CreateCondBr(IsNan, End, NotNan); 866 Result->addIncoming(NanLiteral, NotZero); 867 868 // if (fabs(V) == infinity) return FP_INFINITY 869 Builder.SetInsertPoint(NotNan); 870 Value *VAbs = EmitFAbs(*this, V); 871 Value *IsInf = 872 Builder.CreateFCmpOEQ(VAbs, ConstantFP::getInfinity(V->getType()), 873 "isinf"); 874 Value *InfLiteral = EmitScalarExpr(E->getArg(1)); 875 BasicBlock *NotInf = createBasicBlock("fpclassify_not_inf", this->CurFn); 876 Builder.CreateCondBr(IsInf, End, NotInf); 877 Result->addIncoming(InfLiteral, NotNan); 878 879 // if (fabs(V) >= MIN_NORMAL) return FP_NORMAL else FP_SUBNORMAL 880 Builder.SetInsertPoint(NotInf); 881 APFloat Smallest = APFloat::getSmallestNormalized( 882 getContext().getFloatTypeSemantics(E->getArg(5)->getType())); 883 Value *IsNormal = 884 Builder.CreateFCmpUGE(VAbs, ConstantFP::get(V->getContext(), Smallest), 885 "isnormal"); 886 Value *NormalResult = 887 Builder.CreateSelect(IsNormal, EmitScalarExpr(E->getArg(2)), 888 EmitScalarExpr(E->getArg(3))); 889 Builder.CreateBr(End); 890 Result->addIncoming(NormalResult, NotInf); 891 892 // return Result 893 Builder.SetInsertPoint(End); 894 return RValue::get(Result); 895 } 896 897 case Builtin::BIalloca: 898 case Builtin::BI_alloca: 899 case Builtin::BI__builtin_alloca: { 900 Value *Size = EmitScalarExpr(E->getArg(0)); 901 return RValue::get(Builder.CreateAlloca(Builder.getInt8Ty(), Size)); 902 } 903 case Builtin::BIbzero: 904 case Builtin::BI__builtin_bzero: { 905 Address Dest = EmitPointerWithAlignment(E->getArg(0)); 906 Value *SizeVal = EmitScalarExpr(E->getArg(1)); 907 EmitNonNullArgCheck(RValue::get(Dest.getPointer()), E->getArg(0)->getType(), 908 E->getArg(0)->getExprLoc(), FD, 0); 909 Builder.CreateMemSet(Dest, Builder.getInt8(0), SizeVal, false); 910 return RValue::get(Dest.getPointer()); 911 } 912 case Builtin::BImemcpy: 913 case Builtin::BI__builtin_memcpy: { 914 Address Dest = EmitPointerWithAlignment(E->getArg(0)); 915 Address Src = EmitPointerWithAlignment(E->getArg(1)); 916 Value *SizeVal = EmitScalarExpr(E->getArg(2)); 917 EmitNonNullArgCheck(RValue::get(Dest.getPointer()), E->getArg(0)->getType(), 918 E->getArg(0)->getExprLoc(), FD, 0); 919 EmitNonNullArgCheck(RValue::get(Src.getPointer()), E->getArg(1)->getType(), 920 E->getArg(1)->getExprLoc(), FD, 1); 921 Builder.CreateMemCpy(Dest, Src, SizeVal, false); 922 return RValue::get(Dest.getPointer()); 923 } 924 925 case Builtin::BI__builtin___memcpy_chk: { 926 // fold __builtin_memcpy_chk(x, y, cst1, cst2) to memcpy iff cst1<=cst2. 927 llvm::APSInt Size, DstSize; 928 if (!E->getArg(2)->EvaluateAsInt(Size, CGM.getContext()) || 929 !E->getArg(3)->EvaluateAsInt(DstSize, CGM.getContext())) 930 break; 931 if (Size.ugt(DstSize)) 932 break; 933 Address Dest = EmitPointerWithAlignment(E->getArg(0)); 934 Address Src = EmitPointerWithAlignment(E->getArg(1)); 935 Value *SizeVal = llvm::ConstantInt::get(Builder.getContext(), Size); 936 Builder.CreateMemCpy(Dest, Src, SizeVal, false); 937 return RValue::get(Dest.getPointer()); 938 } 939 940 case Builtin::BI__builtin_objc_memmove_collectable: { 941 Address DestAddr = EmitPointerWithAlignment(E->getArg(0)); 942 Address SrcAddr = EmitPointerWithAlignment(E->getArg(1)); 943 Value *SizeVal = EmitScalarExpr(E->getArg(2)); 944 CGM.getObjCRuntime().EmitGCMemmoveCollectable(*this, 945 DestAddr, SrcAddr, SizeVal); 946 return RValue::get(DestAddr.getPointer()); 947 } 948 949 case Builtin::BI__builtin___memmove_chk: { 950 // fold __builtin_memmove_chk(x, y, cst1, cst2) to memmove iff cst1<=cst2. 951 llvm::APSInt Size, DstSize; 952 if (!E->getArg(2)->EvaluateAsInt(Size, CGM.getContext()) || 953 !E->getArg(3)->EvaluateAsInt(DstSize, CGM.getContext())) 954 break; 955 if (Size.ugt(DstSize)) 956 break; 957 Address Dest = EmitPointerWithAlignment(E->getArg(0)); 958 Address Src = EmitPointerWithAlignment(E->getArg(1)); 959 Value *SizeVal = llvm::ConstantInt::get(Builder.getContext(), Size); 960 Builder.CreateMemMove(Dest, Src, SizeVal, false); 961 return RValue::get(Dest.getPointer()); 962 } 963 964 case Builtin::BImemmove: 965 case Builtin::BI__builtin_memmove: { 966 Address Dest = EmitPointerWithAlignment(E->getArg(0)); 967 Address Src = EmitPointerWithAlignment(E->getArg(1)); 968 Value *SizeVal = EmitScalarExpr(E->getArg(2)); 969 EmitNonNullArgCheck(RValue::get(Dest.getPointer()), E->getArg(0)->getType(), 970 E->getArg(0)->getExprLoc(), FD, 0); 971 EmitNonNullArgCheck(RValue::get(Src.getPointer()), E->getArg(1)->getType(), 972 E->getArg(1)->getExprLoc(), FD, 1); 973 Builder.CreateMemMove(Dest, Src, SizeVal, false); 974 return RValue::get(Dest.getPointer()); 975 } 976 case Builtin::BImemset: 977 case Builtin::BI__builtin_memset: { 978 Address Dest = EmitPointerWithAlignment(E->getArg(0)); 979 Value *ByteVal = Builder.CreateTrunc(EmitScalarExpr(E->getArg(1)), 980 Builder.getInt8Ty()); 981 Value *SizeVal = EmitScalarExpr(E->getArg(2)); 982 EmitNonNullArgCheck(RValue::get(Dest.getPointer()), E->getArg(0)->getType(), 983 E->getArg(0)->getExprLoc(), FD, 0); 984 Builder.CreateMemSet(Dest, ByteVal, SizeVal, false); 985 return RValue::get(Dest.getPointer()); 986 } 987 case Builtin::BI__builtin___memset_chk: { 988 // fold __builtin_memset_chk(x, y, cst1, cst2) to memset iff cst1<=cst2. 989 llvm::APSInt Size, DstSize; 990 if (!E->getArg(2)->EvaluateAsInt(Size, CGM.getContext()) || 991 !E->getArg(3)->EvaluateAsInt(DstSize, CGM.getContext())) 992 break; 993 if (Size.ugt(DstSize)) 994 break; 995 Address Dest = EmitPointerWithAlignment(E->getArg(0)); 996 Value *ByteVal = Builder.CreateTrunc(EmitScalarExpr(E->getArg(1)), 997 Builder.getInt8Ty()); 998 Value *SizeVal = llvm::ConstantInt::get(Builder.getContext(), Size); 999 Builder.CreateMemSet(Dest, ByteVal, SizeVal, false); 1000 return RValue::get(Dest.getPointer()); 1001 } 1002 case Builtin::BI__builtin_dwarf_cfa: { 1003 // The offset in bytes from the first argument to the CFA. 1004 // 1005 // Why on earth is this in the frontend? Is there any reason at 1006 // all that the backend can't reasonably determine this while 1007 // lowering llvm.eh.dwarf.cfa()? 1008 // 1009 // TODO: If there's a satisfactory reason, add a target hook for 1010 // this instead of hard-coding 0, which is correct for most targets. 1011 int32_t Offset = 0; 1012 1013 Value *F = CGM.getIntrinsic(Intrinsic::eh_dwarf_cfa); 1014 return RValue::get(Builder.CreateCall(F, 1015 llvm::ConstantInt::get(Int32Ty, Offset))); 1016 } 1017 case Builtin::BI__builtin_return_address: { 1018 Value *Depth = 1019 CGM.EmitConstantExpr(E->getArg(0), getContext().UnsignedIntTy, this); 1020 Value *F = CGM.getIntrinsic(Intrinsic::returnaddress); 1021 return RValue::get(Builder.CreateCall(F, Depth)); 1022 } 1023 case Builtin::BI__builtin_frame_address: { 1024 Value *Depth = 1025 CGM.EmitConstantExpr(E->getArg(0), getContext().UnsignedIntTy, this); 1026 Value *F = CGM.getIntrinsic(Intrinsic::frameaddress); 1027 return RValue::get(Builder.CreateCall(F, Depth)); 1028 } 1029 case Builtin::BI__builtin_extract_return_addr: { 1030 Value *Address = EmitScalarExpr(E->getArg(0)); 1031 Value *Result = getTargetHooks().decodeReturnAddress(*this, Address); 1032 return RValue::get(Result); 1033 } 1034 case Builtin::BI__builtin_frob_return_addr: { 1035 Value *Address = EmitScalarExpr(E->getArg(0)); 1036 Value *Result = getTargetHooks().encodeReturnAddress(*this, Address); 1037 return RValue::get(Result); 1038 } 1039 case Builtin::BI__builtin_dwarf_sp_column: { 1040 llvm::IntegerType *Ty 1041 = cast<llvm::IntegerType>(ConvertType(E->getType())); 1042 int Column = getTargetHooks().getDwarfEHStackPointer(CGM); 1043 if (Column == -1) { 1044 CGM.ErrorUnsupported(E, "__builtin_dwarf_sp_column"); 1045 return RValue::get(llvm::UndefValue::get(Ty)); 1046 } 1047 return RValue::get(llvm::ConstantInt::get(Ty, Column, true)); 1048 } 1049 case Builtin::BI__builtin_init_dwarf_reg_size_table: { 1050 Value *Address = EmitScalarExpr(E->getArg(0)); 1051 if (getTargetHooks().initDwarfEHRegSizeTable(*this, Address)) 1052 CGM.ErrorUnsupported(E, "__builtin_init_dwarf_reg_size_table"); 1053 return RValue::get(llvm::UndefValue::get(ConvertType(E->getType()))); 1054 } 1055 case Builtin::BI__builtin_eh_return: { 1056 Value *Int = EmitScalarExpr(E->getArg(0)); 1057 Value *Ptr = EmitScalarExpr(E->getArg(1)); 1058 1059 llvm::IntegerType *IntTy = cast<llvm::IntegerType>(Int->getType()); 1060 assert((IntTy->getBitWidth() == 32 || IntTy->getBitWidth() == 64) && 1061 "LLVM's __builtin_eh_return only supports 32- and 64-bit variants"); 1062 Value *F = CGM.getIntrinsic(IntTy->getBitWidth() == 32 1063 ? Intrinsic::eh_return_i32 1064 : Intrinsic::eh_return_i64); 1065 Builder.CreateCall(F, {Int, Ptr}); 1066 Builder.CreateUnreachable(); 1067 1068 // We do need to preserve an insertion point. 1069 EmitBlock(createBasicBlock("builtin_eh_return.cont")); 1070 1071 return RValue::get(nullptr); 1072 } 1073 case Builtin::BI__builtin_unwind_init: { 1074 Value *F = CGM.getIntrinsic(Intrinsic::eh_unwind_init); 1075 return RValue::get(Builder.CreateCall(F)); 1076 } 1077 case Builtin::BI__builtin_extend_pointer: { 1078 // Extends a pointer to the size of an _Unwind_Word, which is 1079 // uint64_t on all platforms. Generally this gets poked into a 1080 // register and eventually used as an address, so if the 1081 // addressing registers are wider than pointers and the platform 1082 // doesn't implicitly ignore high-order bits when doing 1083 // addressing, we need to make sure we zext / sext based on 1084 // the platform's expectations. 1085 // 1086 // See: http://gcc.gnu.org/ml/gcc-bugs/2002-02/msg00237.html 1087 1088 // Cast the pointer to intptr_t. 1089 Value *Ptr = EmitScalarExpr(E->getArg(0)); 1090 Value *Result = Builder.CreatePtrToInt(Ptr, IntPtrTy, "extend.cast"); 1091 1092 // If that's 64 bits, we're done. 1093 if (IntPtrTy->getBitWidth() == 64) 1094 return RValue::get(Result); 1095 1096 // Otherwise, ask the codegen data what to do. 1097 if (getTargetHooks().extendPointerWithSExt()) 1098 return RValue::get(Builder.CreateSExt(Result, Int64Ty, "extend.sext")); 1099 else 1100 return RValue::get(Builder.CreateZExt(Result, Int64Ty, "extend.zext")); 1101 } 1102 case Builtin::BI__builtin_setjmp: { 1103 // Buffer is a void**. 1104 Address Buf = EmitPointerWithAlignment(E->getArg(0)); 1105 1106 // Store the frame pointer to the setjmp buffer. 1107 Value *FrameAddr = 1108 Builder.CreateCall(CGM.getIntrinsic(Intrinsic::frameaddress), 1109 ConstantInt::get(Int32Ty, 0)); 1110 Builder.CreateStore(FrameAddr, Buf); 1111 1112 // Store the stack pointer to the setjmp buffer. 1113 Value *StackAddr = 1114 Builder.CreateCall(CGM.getIntrinsic(Intrinsic::stacksave)); 1115 Address StackSaveSlot = 1116 Builder.CreateConstInBoundsGEP(Buf, 2, getPointerSize()); 1117 Builder.CreateStore(StackAddr, StackSaveSlot); 1118 1119 // Call LLVM's EH setjmp, which is lightweight. 1120 Value *F = CGM.getIntrinsic(Intrinsic::eh_sjlj_setjmp); 1121 Buf = Builder.CreateBitCast(Buf, Int8PtrTy); 1122 return RValue::get(Builder.CreateCall(F, Buf.getPointer())); 1123 } 1124 case Builtin::BI__builtin_longjmp: { 1125 Value *Buf = EmitScalarExpr(E->getArg(0)); 1126 Buf = Builder.CreateBitCast(Buf, Int8PtrTy); 1127 1128 // Call LLVM's EH longjmp, which is lightweight. 1129 Builder.CreateCall(CGM.getIntrinsic(Intrinsic::eh_sjlj_longjmp), Buf); 1130 1131 // longjmp doesn't return; mark this as unreachable. 1132 Builder.CreateUnreachable(); 1133 1134 // We do need to preserve an insertion point. 1135 EmitBlock(createBasicBlock("longjmp.cont")); 1136 1137 return RValue::get(nullptr); 1138 } 1139 case Builtin::BI__sync_fetch_and_add: 1140 case Builtin::BI__sync_fetch_and_sub: 1141 case Builtin::BI__sync_fetch_and_or: 1142 case Builtin::BI__sync_fetch_and_and: 1143 case Builtin::BI__sync_fetch_and_xor: 1144 case Builtin::BI__sync_fetch_and_nand: 1145 case Builtin::BI__sync_add_and_fetch: 1146 case Builtin::BI__sync_sub_and_fetch: 1147 case Builtin::BI__sync_and_and_fetch: 1148 case Builtin::BI__sync_or_and_fetch: 1149 case Builtin::BI__sync_xor_and_fetch: 1150 case Builtin::BI__sync_nand_and_fetch: 1151 case Builtin::BI__sync_val_compare_and_swap: 1152 case Builtin::BI__sync_bool_compare_and_swap: 1153 case Builtin::BI__sync_lock_test_and_set: 1154 case Builtin::BI__sync_lock_release: 1155 case Builtin::BI__sync_swap: 1156 llvm_unreachable("Shouldn't make it through sema"); 1157 case Builtin::BI__sync_fetch_and_add_1: 1158 case Builtin::BI__sync_fetch_and_add_2: 1159 case Builtin::BI__sync_fetch_and_add_4: 1160 case Builtin::BI__sync_fetch_and_add_8: 1161 case Builtin::BI__sync_fetch_and_add_16: 1162 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Add, E); 1163 case Builtin::BI__sync_fetch_and_sub_1: 1164 case Builtin::BI__sync_fetch_and_sub_2: 1165 case Builtin::BI__sync_fetch_and_sub_4: 1166 case Builtin::BI__sync_fetch_and_sub_8: 1167 case Builtin::BI__sync_fetch_and_sub_16: 1168 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Sub, E); 1169 case Builtin::BI__sync_fetch_and_or_1: 1170 case Builtin::BI__sync_fetch_and_or_2: 1171 case Builtin::BI__sync_fetch_and_or_4: 1172 case Builtin::BI__sync_fetch_and_or_8: 1173 case Builtin::BI__sync_fetch_and_or_16: 1174 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Or, E); 1175 case Builtin::BI__sync_fetch_and_and_1: 1176 case Builtin::BI__sync_fetch_and_and_2: 1177 case Builtin::BI__sync_fetch_and_and_4: 1178 case Builtin::BI__sync_fetch_and_and_8: 1179 case Builtin::BI__sync_fetch_and_and_16: 1180 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::And, E); 1181 case Builtin::BI__sync_fetch_and_xor_1: 1182 case Builtin::BI__sync_fetch_and_xor_2: 1183 case Builtin::BI__sync_fetch_and_xor_4: 1184 case Builtin::BI__sync_fetch_and_xor_8: 1185 case Builtin::BI__sync_fetch_and_xor_16: 1186 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Xor, E); 1187 case Builtin::BI__sync_fetch_and_nand_1: 1188 case Builtin::BI__sync_fetch_and_nand_2: 1189 case Builtin::BI__sync_fetch_and_nand_4: 1190 case Builtin::BI__sync_fetch_and_nand_8: 1191 case Builtin::BI__sync_fetch_and_nand_16: 1192 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Nand, E); 1193 1194 // Clang extensions: not overloaded yet. 1195 case Builtin::BI__sync_fetch_and_min: 1196 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Min, E); 1197 case Builtin::BI__sync_fetch_and_max: 1198 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Max, E); 1199 case Builtin::BI__sync_fetch_and_umin: 1200 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::UMin, E); 1201 case Builtin::BI__sync_fetch_and_umax: 1202 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::UMax, E); 1203 1204 case Builtin::BI__sync_add_and_fetch_1: 1205 case Builtin::BI__sync_add_and_fetch_2: 1206 case Builtin::BI__sync_add_and_fetch_4: 1207 case Builtin::BI__sync_add_and_fetch_8: 1208 case Builtin::BI__sync_add_and_fetch_16: 1209 return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::Add, E, 1210 llvm::Instruction::Add); 1211 case Builtin::BI__sync_sub_and_fetch_1: 1212 case Builtin::BI__sync_sub_and_fetch_2: 1213 case Builtin::BI__sync_sub_and_fetch_4: 1214 case Builtin::BI__sync_sub_and_fetch_8: 1215 case Builtin::BI__sync_sub_and_fetch_16: 1216 return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::Sub, E, 1217 llvm::Instruction::Sub); 1218 case Builtin::BI__sync_and_and_fetch_1: 1219 case Builtin::BI__sync_and_and_fetch_2: 1220 case Builtin::BI__sync_and_and_fetch_4: 1221 case Builtin::BI__sync_and_and_fetch_8: 1222 case Builtin::BI__sync_and_and_fetch_16: 1223 return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::And, E, 1224 llvm::Instruction::And); 1225 case Builtin::BI__sync_or_and_fetch_1: 1226 case Builtin::BI__sync_or_and_fetch_2: 1227 case Builtin::BI__sync_or_and_fetch_4: 1228 case Builtin::BI__sync_or_and_fetch_8: 1229 case Builtin::BI__sync_or_and_fetch_16: 1230 return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::Or, E, 1231 llvm::Instruction::Or); 1232 case Builtin::BI__sync_xor_and_fetch_1: 1233 case Builtin::BI__sync_xor_and_fetch_2: 1234 case Builtin::BI__sync_xor_and_fetch_4: 1235 case Builtin::BI__sync_xor_and_fetch_8: 1236 case Builtin::BI__sync_xor_and_fetch_16: 1237 return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::Xor, E, 1238 llvm::Instruction::Xor); 1239 case Builtin::BI__sync_nand_and_fetch_1: 1240 case Builtin::BI__sync_nand_and_fetch_2: 1241 case Builtin::BI__sync_nand_and_fetch_4: 1242 case Builtin::BI__sync_nand_and_fetch_8: 1243 case Builtin::BI__sync_nand_and_fetch_16: 1244 return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::Nand, E, 1245 llvm::Instruction::And, true); 1246 1247 case Builtin::BI__sync_val_compare_and_swap_1: 1248 case Builtin::BI__sync_val_compare_and_swap_2: 1249 case Builtin::BI__sync_val_compare_and_swap_4: 1250 case Builtin::BI__sync_val_compare_and_swap_8: 1251 case Builtin::BI__sync_val_compare_and_swap_16: 1252 return RValue::get(MakeAtomicCmpXchgValue(*this, E, false)); 1253 1254 case Builtin::BI__sync_bool_compare_and_swap_1: 1255 case Builtin::BI__sync_bool_compare_and_swap_2: 1256 case Builtin::BI__sync_bool_compare_and_swap_4: 1257 case Builtin::BI__sync_bool_compare_and_swap_8: 1258 case Builtin::BI__sync_bool_compare_and_swap_16: 1259 return RValue::get(MakeAtomicCmpXchgValue(*this, E, true)); 1260 1261 case Builtin::BI__sync_swap_1: 1262 case Builtin::BI__sync_swap_2: 1263 case Builtin::BI__sync_swap_4: 1264 case Builtin::BI__sync_swap_8: 1265 case Builtin::BI__sync_swap_16: 1266 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Xchg, E); 1267 1268 case Builtin::BI__sync_lock_test_and_set_1: 1269 case Builtin::BI__sync_lock_test_and_set_2: 1270 case Builtin::BI__sync_lock_test_and_set_4: 1271 case Builtin::BI__sync_lock_test_and_set_8: 1272 case Builtin::BI__sync_lock_test_and_set_16: 1273 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Xchg, E); 1274 1275 case Builtin::BI__sync_lock_release_1: 1276 case Builtin::BI__sync_lock_release_2: 1277 case Builtin::BI__sync_lock_release_4: 1278 case Builtin::BI__sync_lock_release_8: 1279 case Builtin::BI__sync_lock_release_16: { 1280 Value *Ptr = EmitScalarExpr(E->getArg(0)); 1281 QualType ElTy = E->getArg(0)->getType()->getPointeeType(); 1282 CharUnits StoreSize = getContext().getTypeSizeInChars(ElTy); 1283 llvm::Type *ITy = llvm::IntegerType::get(getLLVMContext(), 1284 StoreSize.getQuantity() * 8); 1285 Ptr = Builder.CreateBitCast(Ptr, ITy->getPointerTo()); 1286 llvm::StoreInst *Store = 1287 Builder.CreateAlignedStore(llvm::Constant::getNullValue(ITy), Ptr, 1288 StoreSize); 1289 Store->setAtomic(llvm::AtomicOrdering::Release); 1290 return RValue::get(nullptr); 1291 } 1292 1293 case Builtin::BI__sync_synchronize: { 1294 // We assume this is supposed to correspond to a C++0x-style 1295 // sequentially-consistent fence (i.e. this is only usable for 1296 // synchonization, not device I/O or anything like that). This intrinsic 1297 // is really badly designed in the sense that in theory, there isn't 1298 // any way to safely use it... but in practice, it mostly works 1299 // to use it with non-atomic loads and stores to get acquire/release 1300 // semantics. 1301 Builder.CreateFence(llvm::AtomicOrdering::SequentiallyConsistent); 1302 return RValue::get(nullptr); 1303 } 1304 1305 case Builtin::BI__builtin_nontemporal_load: 1306 return RValue::get(EmitNontemporalLoad(*this, E)); 1307 case Builtin::BI__builtin_nontemporal_store: 1308 return RValue::get(EmitNontemporalStore(*this, E)); 1309 case Builtin::BI__c11_atomic_is_lock_free: 1310 case Builtin::BI__atomic_is_lock_free: { 1311 // Call "bool __atomic_is_lock_free(size_t size, void *ptr)". For the 1312 // __c11 builtin, ptr is 0 (indicating a properly-aligned object), since 1313 // _Atomic(T) is always properly-aligned. 1314 const char *LibCallName = "__atomic_is_lock_free"; 1315 CallArgList Args; 1316 Args.add(RValue::get(EmitScalarExpr(E->getArg(0))), 1317 getContext().getSizeType()); 1318 if (BuiltinID == Builtin::BI__atomic_is_lock_free) 1319 Args.add(RValue::get(EmitScalarExpr(E->getArg(1))), 1320 getContext().VoidPtrTy); 1321 else 1322 Args.add(RValue::get(llvm::Constant::getNullValue(VoidPtrTy)), 1323 getContext().VoidPtrTy); 1324 const CGFunctionInfo &FuncInfo = 1325 CGM.getTypes().arrangeBuiltinFunctionCall(E->getType(), Args); 1326 llvm::FunctionType *FTy = CGM.getTypes().GetFunctionType(FuncInfo); 1327 llvm::Constant *Func = CGM.CreateRuntimeFunction(FTy, LibCallName); 1328 return EmitCall(FuncInfo, Func, ReturnValueSlot(), Args); 1329 } 1330 1331 case Builtin::BI__atomic_test_and_set: { 1332 // Look at the argument type to determine whether this is a volatile 1333 // operation. The parameter type is always volatile. 1334 QualType PtrTy = E->getArg(0)->IgnoreImpCasts()->getType(); 1335 bool Volatile = 1336 PtrTy->castAs<PointerType>()->getPointeeType().isVolatileQualified(); 1337 1338 Value *Ptr = EmitScalarExpr(E->getArg(0)); 1339 unsigned AddrSpace = Ptr->getType()->getPointerAddressSpace(); 1340 Ptr = Builder.CreateBitCast(Ptr, Int8Ty->getPointerTo(AddrSpace)); 1341 Value *NewVal = Builder.getInt8(1); 1342 Value *Order = EmitScalarExpr(E->getArg(1)); 1343 if (isa<llvm::ConstantInt>(Order)) { 1344 int ord = cast<llvm::ConstantInt>(Order)->getZExtValue(); 1345 AtomicRMWInst *Result = nullptr; 1346 switch (ord) { 1347 case 0: // memory_order_relaxed 1348 default: // invalid order 1349 Result = Builder.CreateAtomicRMW(llvm::AtomicRMWInst::Xchg, Ptr, NewVal, 1350 llvm::AtomicOrdering::Monotonic); 1351 break; 1352 case 1: // memory_order_consume 1353 case 2: // memory_order_acquire 1354 Result = Builder.CreateAtomicRMW(llvm::AtomicRMWInst::Xchg, Ptr, NewVal, 1355 llvm::AtomicOrdering::Acquire); 1356 break; 1357 case 3: // memory_order_release 1358 Result = Builder.CreateAtomicRMW(llvm::AtomicRMWInst::Xchg, Ptr, NewVal, 1359 llvm::AtomicOrdering::Release); 1360 break; 1361 case 4: // memory_order_acq_rel 1362 1363 Result = Builder.CreateAtomicRMW(llvm::AtomicRMWInst::Xchg, Ptr, NewVal, 1364 llvm::AtomicOrdering::AcquireRelease); 1365 break; 1366 case 5: // memory_order_seq_cst 1367 Result = Builder.CreateAtomicRMW( 1368 llvm::AtomicRMWInst::Xchg, Ptr, NewVal, 1369 llvm::AtomicOrdering::SequentiallyConsistent); 1370 break; 1371 } 1372 Result->setVolatile(Volatile); 1373 return RValue::get(Builder.CreateIsNotNull(Result, "tobool")); 1374 } 1375 1376 llvm::BasicBlock *ContBB = createBasicBlock("atomic.continue", CurFn); 1377 1378 llvm::BasicBlock *BBs[5] = { 1379 createBasicBlock("monotonic", CurFn), 1380 createBasicBlock("acquire", CurFn), 1381 createBasicBlock("release", CurFn), 1382 createBasicBlock("acqrel", CurFn), 1383 createBasicBlock("seqcst", CurFn) 1384 }; 1385 llvm::AtomicOrdering Orders[5] = { 1386 llvm::AtomicOrdering::Monotonic, llvm::AtomicOrdering::Acquire, 1387 llvm::AtomicOrdering::Release, llvm::AtomicOrdering::AcquireRelease, 1388 llvm::AtomicOrdering::SequentiallyConsistent}; 1389 1390 Order = Builder.CreateIntCast(Order, Builder.getInt32Ty(), false); 1391 llvm::SwitchInst *SI = Builder.CreateSwitch(Order, BBs[0]); 1392 1393 Builder.SetInsertPoint(ContBB); 1394 PHINode *Result = Builder.CreatePHI(Int8Ty, 5, "was_set"); 1395 1396 for (unsigned i = 0; i < 5; ++i) { 1397 Builder.SetInsertPoint(BBs[i]); 1398 AtomicRMWInst *RMW = Builder.CreateAtomicRMW(llvm::AtomicRMWInst::Xchg, 1399 Ptr, NewVal, Orders[i]); 1400 RMW->setVolatile(Volatile); 1401 Result->addIncoming(RMW, BBs[i]); 1402 Builder.CreateBr(ContBB); 1403 } 1404 1405 SI->addCase(Builder.getInt32(0), BBs[0]); 1406 SI->addCase(Builder.getInt32(1), BBs[1]); 1407 SI->addCase(Builder.getInt32(2), BBs[1]); 1408 SI->addCase(Builder.getInt32(3), BBs[2]); 1409 SI->addCase(Builder.getInt32(4), BBs[3]); 1410 SI->addCase(Builder.getInt32(5), BBs[4]); 1411 1412 Builder.SetInsertPoint(ContBB); 1413 return RValue::get(Builder.CreateIsNotNull(Result, "tobool")); 1414 } 1415 1416 case Builtin::BI__atomic_clear: { 1417 QualType PtrTy = E->getArg(0)->IgnoreImpCasts()->getType(); 1418 bool Volatile = 1419 PtrTy->castAs<PointerType>()->getPointeeType().isVolatileQualified(); 1420 1421 Address Ptr = EmitPointerWithAlignment(E->getArg(0)); 1422 unsigned AddrSpace = Ptr.getPointer()->getType()->getPointerAddressSpace(); 1423 Ptr = Builder.CreateBitCast(Ptr, Int8Ty->getPointerTo(AddrSpace)); 1424 Value *NewVal = Builder.getInt8(0); 1425 Value *Order = EmitScalarExpr(E->getArg(1)); 1426 if (isa<llvm::ConstantInt>(Order)) { 1427 int ord = cast<llvm::ConstantInt>(Order)->getZExtValue(); 1428 StoreInst *Store = Builder.CreateStore(NewVal, Ptr, Volatile); 1429 switch (ord) { 1430 case 0: // memory_order_relaxed 1431 default: // invalid order 1432 Store->setOrdering(llvm::AtomicOrdering::Monotonic); 1433 break; 1434 case 3: // memory_order_release 1435 Store->setOrdering(llvm::AtomicOrdering::Release); 1436 break; 1437 case 5: // memory_order_seq_cst 1438 Store->setOrdering(llvm::AtomicOrdering::SequentiallyConsistent); 1439 break; 1440 } 1441 return RValue::get(nullptr); 1442 } 1443 1444 llvm::BasicBlock *ContBB = createBasicBlock("atomic.continue", CurFn); 1445 1446 llvm::BasicBlock *BBs[3] = { 1447 createBasicBlock("monotonic", CurFn), 1448 createBasicBlock("release", CurFn), 1449 createBasicBlock("seqcst", CurFn) 1450 }; 1451 llvm::AtomicOrdering Orders[3] = { 1452 llvm::AtomicOrdering::Monotonic, llvm::AtomicOrdering::Release, 1453 llvm::AtomicOrdering::SequentiallyConsistent}; 1454 1455 Order = Builder.CreateIntCast(Order, Builder.getInt32Ty(), false); 1456 llvm::SwitchInst *SI = Builder.CreateSwitch(Order, BBs[0]); 1457 1458 for (unsigned i = 0; i < 3; ++i) { 1459 Builder.SetInsertPoint(BBs[i]); 1460 StoreInst *Store = Builder.CreateStore(NewVal, Ptr, Volatile); 1461 Store->setOrdering(Orders[i]); 1462 Builder.CreateBr(ContBB); 1463 } 1464 1465 SI->addCase(Builder.getInt32(0), BBs[0]); 1466 SI->addCase(Builder.getInt32(3), BBs[1]); 1467 SI->addCase(Builder.getInt32(5), BBs[2]); 1468 1469 Builder.SetInsertPoint(ContBB); 1470 return RValue::get(nullptr); 1471 } 1472 1473 case Builtin::BI__atomic_thread_fence: 1474 case Builtin::BI__atomic_signal_fence: 1475 case Builtin::BI__c11_atomic_thread_fence: 1476 case Builtin::BI__c11_atomic_signal_fence: { 1477 llvm::SynchronizationScope Scope; 1478 if (BuiltinID == Builtin::BI__atomic_signal_fence || 1479 BuiltinID == Builtin::BI__c11_atomic_signal_fence) 1480 Scope = llvm::SingleThread; 1481 else 1482 Scope = llvm::CrossThread; 1483 Value *Order = EmitScalarExpr(E->getArg(0)); 1484 if (isa<llvm::ConstantInt>(Order)) { 1485 int ord = cast<llvm::ConstantInt>(Order)->getZExtValue(); 1486 switch (ord) { 1487 case 0: // memory_order_relaxed 1488 default: // invalid order 1489 break; 1490 case 1: // memory_order_consume 1491 case 2: // memory_order_acquire 1492 Builder.CreateFence(llvm::AtomicOrdering::Acquire, Scope); 1493 break; 1494 case 3: // memory_order_release 1495 Builder.CreateFence(llvm::AtomicOrdering::Release, Scope); 1496 break; 1497 case 4: // memory_order_acq_rel 1498 Builder.CreateFence(llvm::AtomicOrdering::AcquireRelease, Scope); 1499 break; 1500 case 5: // memory_order_seq_cst 1501 Builder.CreateFence(llvm::AtomicOrdering::SequentiallyConsistent, 1502 Scope); 1503 break; 1504 } 1505 return RValue::get(nullptr); 1506 } 1507 1508 llvm::BasicBlock *AcquireBB, *ReleaseBB, *AcqRelBB, *SeqCstBB; 1509 AcquireBB = createBasicBlock("acquire", CurFn); 1510 ReleaseBB = createBasicBlock("release", CurFn); 1511 AcqRelBB = createBasicBlock("acqrel", CurFn); 1512 SeqCstBB = createBasicBlock("seqcst", CurFn); 1513 llvm::BasicBlock *ContBB = createBasicBlock("atomic.continue", CurFn); 1514 1515 Order = Builder.CreateIntCast(Order, Builder.getInt32Ty(), false); 1516 llvm::SwitchInst *SI = Builder.CreateSwitch(Order, ContBB); 1517 1518 Builder.SetInsertPoint(AcquireBB); 1519 Builder.CreateFence(llvm::AtomicOrdering::Acquire, Scope); 1520 Builder.CreateBr(ContBB); 1521 SI->addCase(Builder.getInt32(1), AcquireBB); 1522 SI->addCase(Builder.getInt32(2), AcquireBB); 1523 1524 Builder.SetInsertPoint(ReleaseBB); 1525 Builder.CreateFence(llvm::AtomicOrdering::Release, Scope); 1526 Builder.CreateBr(ContBB); 1527 SI->addCase(Builder.getInt32(3), ReleaseBB); 1528 1529 Builder.SetInsertPoint(AcqRelBB); 1530 Builder.CreateFence(llvm::AtomicOrdering::AcquireRelease, Scope); 1531 Builder.CreateBr(ContBB); 1532 SI->addCase(Builder.getInt32(4), AcqRelBB); 1533 1534 Builder.SetInsertPoint(SeqCstBB); 1535 Builder.CreateFence(llvm::AtomicOrdering::SequentiallyConsistent, Scope); 1536 Builder.CreateBr(ContBB); 1537 SI->addCase(Builder.getInt32(5), SeqCstBB); 1538 1539 Builder.SetInsertPoint(ContBB); 1540 return RValue::get(nullptr); 1541 } 1542 1543 // Library functions with special handling. 1544 case Builtin::BIsqrt: 1545 case Builtin::BIsqrtf: 1546 case Builtin::BIsqrtl: { 1547 // Transform a call to sqrt* into a @llvm.sqrt.* intrinsic call, but only 1548 // in finite- or unsafe-math mode (the intrinsic has different semantics 1549 // for handling negative numbers compared to the library function, so 1550 // -fmath-errno=0 is not enough). 1551 if (!FD->hasAttr<ConstAttr>()) 1552 break; 1553 if (!(CGM.getCodeGenOpts().UnsafeFPMath || 1554 CGM.getCodeGenOpts().NoNaNsFPMath)) 1555 break; 1556 Value *Arg0 = EmitScalarExpr(E->getArg(0)); 1557 llvm::Type *ArgType = Arg0->getType(); 1558 Value *F = CGM.getIntrinsic(Intrinsic::sqrt, ArgType); 1559 return RValue::get(Builder.CreateCall(F, Arg0)); 1560 } 1561 1562 case Builtin::BI__builtin_pow: 1563 case Builtin::BI__builtin_powf: 1564 case Builtin::BI__builtin_powl: 1565 case Builtin::BIpow: 1566 case Builtin::BIpowf: 1567 case Builtin::BIpowl: { 1568 // Transform a call to pow* into a @llvm.pow.* intrinsic call. 1569 if (!FD->hasAttr<ConstAttr>()) 1570 break; 1571 Value *Base = EmitScalarExpr(E->getArg(0)); 1572 Value *Exponent = EmitScalarExpr(E->getArg(1)); 1573 llvm::Type *ArgType = Base->getType(); 1574 Value *F = CGM.getIntrinsic(Intrinsic::pow, ArgType); 1575 return RValue::get(Builder.CreateCall(F, {Base, Exponent})); 1576 } 1577 1578 case Builtin::BIfma: 1579 case Builtin::BIfmaf: 1580 case Builtin::BIfmal: 1581 case Builtin::BI__builtin_fma: 1582 case Builtin::BI__builtin_fmaf: 1583 case Builtin::BI__builtin_fmal: { 1584 // Rewrite fma to intrinsic. 1585 Value *FirstArg = EmitScalarExpr(E->getArg(0)); 1586 llvm::Type *ArgType = FirstArg->getType(); 1587 Value *F = CGM.getIntrinsic(Intrinsic::fma, ArgType); 1588 return RValue::get( 1589 Builder.CreateCall(F, {FirstArg, EmitScalarExpr(E->getArg(1)), 1590 EmitScalarExpr(E->getArg(2))})); 1591 } 1592 1593 case Builtin::BI__builtin_signbit: 1594 case Builtin::BI__builtin_signbitf: 1595 case Builtin::BI__builtin_signbitl: { 1596 return RValue::get( 1597 Builder.CreateZExt(EmitSignBit(*this, EmitScalarExpr(E->getArg(0))), 1598 ConvertType(E->getType()))); 1599 } 1600 case Builtin::BI__builtin_annotation: { 1601 llvm::Value *AnnVal = EmitScalarExpr(E->getArg(0)); 1602 llvm::Value *F = CGM.getIntrinsic(llvm::Intrinsic::annotation, 1603 AnnVal->getType()); 1604 1605 // Get the annotation string, go through casts. Sema requires this to be a 1606 // non-wide string literal, potentially casted, so the cast<> is safe. 1607 const Expr *AnnotationStrExpr = E->getArg(1)->IgnoreParenCasts(); 1608 StringRef Str = cast<StringLiteral>(AnnotationStrExpr)->getString(); 1609 return RValue::get(EmitAnnotationCall(F, AnnVal, Str, E->getExprLoc())); 1610 } 1611 case Builtin::BI__builtin_addcb: 1612 case Builtin::BI__builtin_addcs: 1613 case Builtin::BI__builtin_addc: 1614 case Builtin::BI__builtin_addcl: 1615 case Builtin::BI__builtin_addcll: 1616 case Builtin::BI__builtin_subcb: 1617 case Builtin::BI__builtin_subcs: 1618 case Builtin::BI__builtin_subc: 1619 case Builtin::BI__builtin_subcl: 1620 case Builtin::BI__builtin_subcll: { 1621 1622 // We translate all of these builtins from expressions of the form: 1623 // int x = ..., y = ..., carryin = ..., carryout, result; 1624 // result = __builtin_addc(x, y, carryin, &carryout); 1625 // 1626 // to LLVM IR of the form: 1627 // 1628 // %tmp1 = call {i32, i1} @llvm.uadd.with.overflow.i32(i32 %x, i32 %y) 1629 // %tmpsum1 = extractvalue {i32, i1} %tmp1, 0 1630 // %carry1 = extractvalue {i32, i1} %tmp1, 1 1631 // %tmp2 = call {i32, i1} @llvm.uadd.with.overflow.i32(i32 %tmpsum1, 1632 // i32 %carryin) 1633 // %result = extractvalue {i32, i1} %tmp2, 0 1634 // %carry2 = extractvalue {i32, i1} %tmp2, 1 1635 // %tmp3 = or i1 %carry1, %carry2 1636 // %tmp4 = zext i1 %tmp3 to i32 1637 // store i32 %tmp4, i32* %carryout 1638 1639 // Scalarize our inputs. 1640 llvm::Value *X = EmitScalarExpr(E->getArg(0)); 1641 llvm::Value *Y = EmitScalarExpr(E->getArg(1)); 1642 llvm::Value *Carryin = EmitScalarExpr(E->getArg(2)); 1643 Address CarryOutPtr = EmitPointerWithAlignment(E->getArg(3)); 1644 1645 // Decide if we are lowering to a uadd.with.overflow or usub.with.overflow. 1646 llvm::Intrinsic::ID IntrinsicId; 1647 switch (BuiltinID) { 1648 default: llvm_unreachable("Unknown multiprecision builtin id."); 1649 case Builtin::BI__builtin_addcb: 1650 case Builtin::BI__builtin_addcs: 1651 case Builtin::BI__builtin_addc: 1652 case Builtin::BI__builtin_addcl: 1653 case Builtin::BI__builtin_addcll: 1654 IntrinsicId = llvm::Intrinsic::uadd_with_overflow; 1655 break; 1656 case Builtin::BI__builtin_subcb: 1657 case Builtin::BI__builtin_subcs: 1658 case Builtin::BI__builtin_subc: 1659 case Builtin::BI__builtin_subcl: 1660 case Builtin::BI__builtin_subcll: 1661 IntrinsicId = llvm::Intrinsic::usub_with_overflow; 1662 break; 1663 } 1664 1665 // Construct our resulting LLVM IR expression. 1666 llvm::Value *Carry1; 1667 llvm::Value *Sum1 = EmitOverflowIntrinsic(*this, IntrinsicId, 1668 X, Y, Carry1); 1669 llvm::Value *Carry2; 1670 llvm::Value *Sum2 = EmitOverflowIntrinsic(*this, IntrinsicId, 1671 Sum1, Carryin, Carry2); 1672 llvm::Value *CarryOut = Builder.CreateZExt(Builder.CreateOr(Carry1, Carry2), 1673 X->getType()); 1674 Builder.CreateStore(CarryOut, CarryOutPtr); 1675 return RValue::get(Sum2); 1676 } 1677 1678 case Builtin::BI__builtin_add_overflow: 1679 case Builtin::BI__builtin_sub_overflow: 1680 case Builtin::BI__builtin_mul_overflow: { 1681 const clang::Expr *LeftArg = E->getArg(0); 1682 const clang::Expr *RightArg = E->getArg(1); 1683 const clang::Expr *ResultArg = E->getArg(2); 1684 1685 clang::QualType ResultQTy = 1686 ResultArg->getType()->castAs<PointerType>()->getPointeeType(); 1687 1688 WidthAndSignedness LeftInfo = 1689 getIntegerWidthAndSignedness(CGM.getContext(), LeftArg->getType()); 1690 WidthAndSignedness RightInfo = 1691 getIntegerWidthAndSignedness(CGM.getContext(), RightArg->getType()); 1692 WidthAndSignedness ResultInfo = 1693 getIntegerWidthAndSignedness(CGM.getContext(), ResultQTy); 1694 WidthAndSignedness EncompassingInfo = 1695 EncompassingIntegerType({LeftInfo, RightInfo, ResultInfo}); 1696 1697 llvm::Type *EncompassingLLVMTy = 1698 llvm::IntegerType::get(CGM.getLLVMContext(), EncompassingInfo.Width); 1699 1700 llvm::Type *ResultLLVMTy = CGM.getTypes().ConvertType(ResultQTy); 1701 1702 llvm::Intrinsic::ID IntrinsicId; 1703 switch (BuiltinID) { 1704 default: 1705 llvm_unreachable("Unknown overflow builtin id."); 1706 case Builtin::BI__builtin_add_overflow: 1707 IntrinsicId = EncompassingInfo.Signed 1708 ? llvm::Intrinsic::sadd_with_overflow 1709 : llvm::Intrinsic::uadd_with_overflow; 1710 break; 1711 case Builtin::BI__builtin_sub_overflow: 1712 IntrinsicId = EncompassingInfo.Signed 1713 ? llvm::Intrinsic::ssub_with_overflow 1714 : llvm::Intrinsic::usub_with_overflow; 1715 break; 1716 case Builtin::BI__builtin_mul_overflow: 1717 IntrinsicId = EncompassingInfo.Signed 1718 ? llvm::Intrinsic::smul_with_overflow 1719 : llvm::Intrinsic::umul_with_overflow; 1720 break; 1721 } 1722 1723 llvm::Value *Left = EmitScalarExpr(LeftArg); 1724 llvm::Value *Right = EmitScalarExpr(RightArg); 1725 Address ResultPtr = EmitPointerWithAlignment(ResultArg); 1726 1727 // Extend each operand to the encompassing type. 1728 Left = Builder.CreateIntCast(Left, EncompassingLLVMTy, LeftInfo.Signed); 1729 Right = Builder.CreateIntCast(Right, EncompassingLLVMTy, RightInfo.Signed); 1730 1731 // Perform the operation on the extended values. 1732 llvm::Value *Overflow, *Result; 1733 Result = EmitOverflowIntrinsic(*this, IntrinsicId, Left, Right, Overflow); 1734 1735 if (EncompassingInfo.Width > ResultInfo.Width) { 1736 // The encompassing type is wider than the result type, so we need to 1737 // truncate it. 1738 llvm::Value *ResultTrunc = Builder.CreateTrunc(Result, ResultLLVMTy); 1739 1740 // To see if the truncation caused an overflow, we will extend 1741 // the result and then compare it to the original result. 1742 llvm::Value *ResultTruncExt = Builder.CreateIntCast( 1743 ResultTrunc, EncompassingLLVMTy, ResultInfo.Signed); 1744 llvm::Value *TruncationOverflow = 1745 Builder.CreateICmpNE(Result, ResultTruncExt); 1746 1747 Overflow = Builder.CreateOr(Overflow, TruncationOverflow); 1748 Result = ResultTrunc; 1749 } 1750 1751 // Finally, store the result using the pointer. 1752 bool isVolatile = 1753 ResultArg->getType()->getPointeeType().isVolatileQualified(); 1754 Builder.CreateStore(EmitToMemory(Result, ResultQTy), ResultPtr, isVolatile); 1755 1756 return RValue::get(Overflow); 1757 } 1758 1759 case Builtin::BI__builtin_uadd_overflow: 1760 case Builtin::BI__builtin_uaddl_overflow: 1761 case Builtin::BI__builtin_uaddll_overflow: 1762 case Builtin::BI__builtin_usub_overflow: 1763 case Builtin::BI__builtin_usubl_overflow: 1764 case Builtin::BI__builtin_usubll_overflow: 1765 case Builtin::BI__builtin_umul_overflow: 1766 case Builtin::BI__builtin_umull_overflow: 1767 case Builtin::BI__builtin_umulll_overflow: 1768 case Builtin::BI__builtin_sadd_overflow: 1769 case Builtin::BI__builtin_saddl_overflow: 1770 case Builtin::BI__builtin_saddll_overflow: 1771 case Builtin::BI__builtin_ssub_overflow: 1772 case Builtin::BI__builtin_ssubl_overflow: 1773 case Builtin::BI__builtin_ssubll_overflow: 1774 case Builtin::BI__builtin_smul_overflow: 1775 case Builtin::BI__builtin_smull_overflow: 1776 case Builtin::BI__builtin_smulll_overflow: { 1777 1778 // We translate all of these builtins directly to the relevant llvm IR node. 1779 1780 // Scalarize our inputs. 1781 llvm::Value *X = EmitScalarExpr(E->getArg(0)); 1782 llvm::Value *Y = EmitScalarExpr(E->getArg(1)); 1783 Address SumOutPtr = EmitPointerWithAlignment(E->getArg(2)); 1784 1785 // Decide which of the overflow intrinsics we are lowering to: 1786 llvm::Intrinsic::ID IntrinsicId; 1787 switch (BuiltinID) { 1788 default: llvm_unreachable("Unknown overflow builtin id."); 1789 case Builtin::BI__builtin_uadd_overflow: 1790 case Builtin::BI__builtin_uaddl_overflow: 1791 case Builtin::BI__builtin_uaddll_overflow: 1792 IntrinsicId = llvm::Intrinsic::uadd_with_overflow; 1793 break; 1794 case Builtin::BI__builtin_usub_overflow: 1795 case Builtin::BI__builtin_usubl_overflow: 1796 case Builtin::BI__builtin_usubll_overflow: 1797 IntrinsicId = llvm::Intrinsic::usub_with_overflow; 1798 break; 1799 case Builtin::BI__builtin_umul_overflow: 1800 case Builtin::BI__builtin_umull_overflow: 1801 case Builtin::BI__builtin_umulll_overflow: 1802 IntrinsicId = llvm::Intrinsic::umul_with_overflow; 1803 break; 1804 case Builtin::BI__builtin_sadd_overflow: 1805 case Builtin::BI__builtin_saddl_overflow: 1806 case Builtin::BI__builtin_saddll_overflow: 1807 IntrinsicId = llvm::Intrinsic::sadd_with_overflow; 1808 break; 1809 case Builtin::BI__builtin_ssub_overflow: 1810 case Builtin::BI__builtin_ssubl_overflow: 1811 case Builtin::BI__builtin_ssubll_overflow: 1812 IntrinsicId = llvm::Intrinsic::ssub_with_overflow; 1813 break; 1814 case Builtin::BI__builtin_smul_overflow: 1815 case Builtin::BI__builtin_smull_overflow: 1816 case Builtin::BI__builtin_smulll_overflow: 1817 IntrinsicId = llvm::Intrinsic::smul_with_overflow; 1818 break; 1819 } 1820 1821 1822 llvm::Value *Carry; 1823 llvm::Value *Sum = EmitOverflowIntrinsic(*this, IntrinsicId, X, Y, Carry); 1824 Builder.CreateStore(Sum, SumOutPtr); 1825 1826 return RValue::get(Carry); 1827 } 1828 case Builtin::BI__builtin_addressof: 1829 return RValue::get(EmitLValue(E->getArg(0)).getPointer()); 1830 case Builtin::BI__builtin_operator_new: 1831 return EmitBuiltinNewDeleteCall(FD->getType()->castAs<FunctionProtoType>(), 1832 E->getArg(0), false); 1833 case Builtin::BI__builtin_operator_delete: 1834 return EmitBuiltinNewDeleteCall(FD->getType()->castAs<FunctionProtoType>(), 1835 E->getArg(0), true); 1836 case Builtin::BI__noop: 1837 // __noop always evaluates to an integer literal zero. 1838 return RValue::get(ConstantInt::get(IntTy, 0)); 1839 case Builtin::BI__builtin_call_with_static_chain: { 1840 const CallExpr *Call = cast<CallExpr>(E->getArg(0)); 1841 const Expr *Chain = E->getArg(1); 1842 return EmitCall(Call->getCallee()->getType(), 1843 EmitScalarExpr(Call->getCallee()), Call, ReturnValue, 1844 Call->getCalleeDecl(), EmitScalarExpr(Chain)); 1845 } 1846 case Builtin::BI_InterlockedExchange: 1847 case Builtin::BI_InterlockedExchangePointer: 1848 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Xchg, E); 1849 case Builtin::BI_InterlockedCompareExchangePointer: { 1850 llvm::Type *RTy; 1851 llvm::IntegerType *IntType = 1852 IntegerType::get(getLLVMContext(), 1853 getContext().getTypeSize(E->getType())); 1854 llvm::Type *IntPtrType = IntType->getPointerTo(); 1855 1856 llvm::Value *Destination = 1857 Builder.CreateBitCast(EmitScalarExpr(E->getArg(0)), IntPtrType); 1858 1859 llvm::Value *Exchange = EmitScalarExpr(E->getArg(1)); 1860 RTy = Exchange->getType(); 1861 Exchange = Builder.CreatePtrToInt(Exchange, IntType); 1862 1863 llvm::Value *Comparand = 1864 Builder.CreatePtrToInt(EmitScalarExpr(E->getArg(2)), IntType); 1865 1866 auto Result = 1867 Builder.CreateAtomicCmpXchg(Destination, Comparand, Exchange, 1868 AtomicOrdering::SequentiallyConsistent, 1869 AtomicOrdering::SequentiallyConsistent); 1870 Result->setVolatile(true); 1871 1872 return RValue::get(Builder.CreateIntToPtr(Builder.CreateExtractValue(Result, 1873 0), 1874 RTy)); 1875 } 1876 case Builtin::BI_InterlockedCompareExchange: { 1877 AtomicCmpXchgInst *CXI = Builder.CreateAtomicCmpXchg( 1878 EmitScalarExpr(E->getArg(0)), 1879 EmitScalarExpr(E->getArg(2)), 1880 EmitScalarExpr(E->getArg(1)), 1881 AtomicOrdering::SequentiallyConsistent, 1882 AtomicOrdering::SequentiallyConsistent); 1883 CXI->setVolatile(true); 1884 return RValue::get(Builder.CreateExtractValue(CXI, 0)); 1885 } 1886 case Builtin::BI_InterlockedIncrement: { 1887 AtomicRMWInst *RMWI = Builder.CreateAtomicRMW( 1888 AtomicRMWInst::Add, 1889 EmitScalarExpr(E->getArg(0)), 1890 ConstantInt::get(Int32Ty, 1), 1891 llvm::AtomicOrdering::SequentiallyConsistent); 1892 RMWI->setVolatile(true); 1893 return RValue::get(Builder.CreateAdd(RMWI, ConstantInt::get(Int32Ty, 1))); 1894 } 1895 case Builtin::BI_InterlockedDecrement: { 1896 AtomicRMWInst *RMWI = Builder.CreateAtomicRMW( 1897 AtomicRMWInst::Sub, 1898 EmitScalarExpr(E->getArg(0)), 1899 ConstantInt::get(Int32Ty, 1), 1900 llvm::AtomicOrdering::SequentiallyConsistent); 1901 RMWI->setVolatile(true); 1902 return RValue::get(Builder.CreateSub(RMWI, ConstantInt::get(Int32Ty, 1))); 1903 } 1904 case Builtin::BI_InterlockedExchangeAdd: { 1905 AtomicRMWInst *RMWI = Builder.CreateAtomicRMW( 1906 AtomicRMWInst::Add, 1907 EmitScalarExpr(E->getArg(0)), 1908 EmitScalarExpr(E->getArg(1)), 1909 llvm::AtomicOrdering::SequentiallyConsistent); 1910 RMWI->setVolatile(true); 1911 return RValue::get(RMWI); 1912 } 1913 case Builtin::BI__readfsdword: { 1914 Value *IntToPtr = 1915 Builder.CreateIntToPtr(EmitScalarExpr(E->getArg(0)), 1916 llvm::PointerType::get(CGM.Int32Ty, 257)); 1917 LoadInst *Load = 1918 Builder.CreateAlignedLoad(IntToPtr, /*Align=*/4, /*isVolatile=*/true); 1919 return RValue::get(Load); 1920 } 1921 1922 case Builtin::BI__exception_code: 1923 case Builtin::BI_exception_code: 1924 return RValue::get(EmitSEHExceptionCode()); 1925 case Builtin::BI__exception_info: 1926 case Builtin::BI_exception_info: 1927 return RValue::get(EmitSEHExceptionInfo()); 1928 case Builtin::BI__abnormal_termination: 1929 case Builtin::BI_abnormal_termination: 1930 return RValue::get(EmitSEHAbnormalTermination()); 1931 case Builtin::BI_setjmpex: { 1932 if (getTarget().getTriple().isOSMSVCRT()) { 1933 llvm::Type *ArgTypes[] = {Int8PtrTy, Int8PtrTy}; 1934 llvm::AttributeSet ReturnsTwiceAttr = 1935 AttributeSet::get(getLLVMContext(), llvm::AttributeSet::FunctionIndex, 1936 llvm::Attribute::ReturnsTwice); 1937 llvm::Constant *SetJmpEx = CGM.CreateRuntimeFunction( 1938 llvm::FunctionType::get(IntTy, ArgTypes, /*isVarArg=*/false), 1939 "_setjmpex", ReturnsTwiceAttr); 1940 llvm::Value *Buf = Builder.CreateBitOrPointerCast( 1941 EmitScalarExpr(E->getArg(0)), Int8PtrTy); 1942 llvm::Value *FrameAddr = 1943 Builder.CreateCall(CGM.getIntrinsic(Intrinsic::frameaddress), 1944 ConstantInt::get(Int32Ty, 0)); 1945 llvm::Value *Args[] = {Buf, FrameAddr}; 1946 llvm::CallSite CS = EmitRuntimeCallOrInvoke(SetJmpEx, Args); 1947 CS.setAttributes(ReturnsTwiceAttr); 1948 return RValue::get(CS.getInstruction()); 1949 } 1950 break; 1951 } 1952 case Builtin::BI_setjmp: { 1953 if (getTarget().getTriple().isOSMSVCRT()) { 1954 llvm::AttributeSet ReturnsTwiceAttr = 1955 AttributeSet::get(getLLVMContext(), llvm::AttributeSet::FunctionIndex, 1956 llvm::Attribute::ReturnsTwice); 1957 llvm::Value *Buf = Builder.CreateBitOrPointerCast( 1958 EmitScalarExpr(E->getArg(0)), Int8PtrTy); 1959 llvm::CallSite CS; 1960 if (getTarget().getTriple().getArch() == llvm::Triple::x86) { 1961 llvm::Type *ArgTypes[] = {Int8PtrTy, IntTy}; 1962 llvm::Constant *SetJmp3 = CGM.CreateRuntimeFunction( 1963 llvm::FunctionType::get(IntTy, ArgTypes, /*isVarArg=*/true), 1964 "_setjmp3", ReturnsTwiceAttr); 1965 llvm::Value *Count = ConstantInt::get(IntTy, 0); 1966 llvm::Value *Args[] = {Buf, Count}; 1967 CS = EmitRuntimeCallOrInvoke(SetJmp3, Args); 1968 } else { 1969 llvm::Type *ArgTypes[] = {Int8PtrTy, Int8PtrTy}; 1970 llvm::Constant *SetJmp = CGM.CreateRuntimeFunction( 1971 llvm::FunctionType::get(IntTy, ArgTypes, /*isVarArg=*/false), 1972 "_setjmp", ReturnsTwiceAttr); 1973 llvm::Value *FrameAddr = 1974 Builder.CreateCall(CGM.getIntrinsic(Intrinsic::frameaddress), 1975 ConstantInt::get(Int32Ty, 0)); 1976 llvm::Value *Args[] = {Buf, FrameAddr}; 1977 CS = EmitRuntimeCallOrInvoke(SetJmp, Args); 1978 } 1979 CS.setAttributes(ReturnsTwiceAttr); 1980 return RValue::get(CS.getInstruction()); 1981 } 1982 break; 1983 } 1984 1985 case Builtin::BI__GetExceptionInfo: { 1986 if (llvm::GlobalVariable *GV = 1987 CGM.getCXXABI().getThrowInfo(FD->getParamDecl(0)->getType())) 1988 return RValue::get(llvm::ConstantExpr::getBitCast(GV, CGM.Int8PtrTy)); 1989 break; 1990 } 1991 1992 // OpenCL v2.0 s6.13.16.2, Built-in pipe read and write functions 1993 case Builtin::BIread_pipe: 1994 case Builtin::BIwrite_pipe: { 1995 Value *Arg0 = EmitScalarExpr(E->getArg(0)), 1996 *Arg1 = EmitScalarExpr(E->getArg(1)); 1997 1998 // Type of the generic packet parameter. 1999 unsigned GenericAS = 2000 getContext().getTargetAddressSpace(LangAS::opencl_generic); 2001 llvm::Type *I8PTy = llvm::PointerType::get( 2002 llvm::Type::getInt8Ty(getLLVMContext()), GenericAS); 2003 2004 // Testing which overloaded version we should generate the call for. 2005 if (2U == E->getNumArgs()) { 2006 const char *Name = (BuiltinID == Builtin::BIread_pipe) ? "__read_pipe_2" 2007 : "__write_pipe_2"; 2008 // Creating a generic function type to be able to call with any builtin or 2009 // user defined type. 2010 llvm::Type *ArgTys[] = {Arg0->getType(), I8PTy}; 2011 llvm::FunctionType *FTy = llvm::FunctionType::get( 2012 Int32Ty, llvm::ArrayRef<llvm::Type *>(ArgTys), false); 2013 Value *BCast = Builder.CreatePointerCast(Arg1, I8PTy); 2014 return RValue::get(Builder.CreateCall( 2015 CGM.CreateRuntimeFunction(FTy, Name), {Arg0, BCast})); 2016 } else { 2017 assert(4 == E->getNumArgs() && 2018 "Illegal number of parameters to pipe function"); 2019 const char *Name = (BuiltinID == Builtin::BIread_pipe) ? "__read_pipe_4" 2020 : "__write_pipe_4"; 2021 2022 llvm::Type *ArgTys[] = {Arg0->getType(), Arg1->getType(), Int32Ty, I8PTy}; 2023 Value *Arg2 = EmitScalarExpr(E->getArg(2)), 2024 *Arg3 = EmitScalarExpr(E->getArg(3)); 2025 llvm::FunctionType *FTy = llvm::FunctionType::get( 2026 Int32Ty, llvm::ArrayRef<llvm::Type *>(ArgTys), false); 2027 Value *BCast = Builder.CreatePointerCast(Arg3, I8PTy); 2028 // We know the third argument is an integer type, but we may need to cast 2029 // it to i32. 2030 if (Arg2->getType() != Int32Ty) 2031 Arg2 = Builder.CreateZExtOrTrunc(Arg2, Int32Ty); 2032 return RValue::get(Builder.CreateCall( 2033 CGM.CreateRuntimeFunction(FTy, Name), {Arg0, Arg1, Arg2, BCast})); 2034 } 2035 } 2036 // OpenCL v2.0 s6.13.16 ,s9.17.3.5 - Built-in pipe reserve read and write 2037 // functions 2038 case Builtin::BIreserve_read_pipe: 2039 case Builtin::BIreserve_write_pipe: 2040 case Builtin::BIwork_group_reserve_read_pipe: 2041 case Builtin::BIwork_group_reserve_write_pipe: 2042 case Builtin::BIsub_group_reserve_read_pipe: 2043 case Builtin::BIsub_group_reserve_write_pipe: { 2044 // Composing the mangled name for the function. 2045 const char *Name; 2046 if (BuiltinID == Builtin::BIreserve_read_pipe) 2047 Name = "__reserve_read_pipe"; 2048 else if (BuiltinID == Builtin::BIreserve_write_pipe) 2049 Name = "__reserve_write_pipe"; 2050 else if (BuiltinID == Builtin::BIwork_group_reserve_read_pipe) 2051 Name = "__work_group_reserve_read_pipe"; 2052 else if (BuiltinID == Builtin::BIwork_group_reserve_write_pipe) 2053 Name = "__work_group_reserve_write_pipe"; 2054 else if (BuiltinID == Builtin::BIsub_group_reserve_read_pipe) 2055 Name = "__sub_group_reserve_read_pipe"; 2056 else 2057 Name = "__sub_group_reserve_write_pipe"; 2058 2059 Value *Arg0 = EmitScalarExpr(E->getArg(0)), 2060 *Arg1 = EmitScalarExpr(E->getArg(1)); 2061 llvm::Type *ReservedIDTy = ConvertType(getContext().OCLReserveIDTy); 2062 2063 // Building the generic function prototype. 2064 llvm::Type *ArgTys[] = {Arg0->getType(), Int32Ty}; 2065 llvm::FunctionType *FTy = llvm::FunctionType::get( 2066 ReservedIDTy, llvm::ArrayRef<llvm::Type *>(ArgTys), false); 2067 // We know the second argument is an integer type, but we may need to cast 2068 // it to i32. 2069 if (Arg1->getType() != Int32Ty) 2070 Arg1 = Builder.CreateZExtOrTrunc(Arg1, Int32Ty); 2071 return RValue::get( 2072 Builder.CreateCall(CGM.CreateRuntimeFunction(FTy, Name), {Arg0, Arg1})); 2073 } 2074 // OpenCL v2.0 s6.13.16 ,s9.17.3.5 - Built-in pipe commit read and write 2075 // functions 2076 case Builtin::BIcommit_read_pipe: 2077 case Builtin::BIcommit_write_pipe: 2078 case Builtin::BIwork_group_commit_read_pipe: 2079 case Builtin::BIwork_group_commit_write_pipe: 2080 case Builtin::BIsub_group_commit_read_pipe: 2081 case Builtin::BIsub_group_commit_write_pipe: { 2082 const char *Name; 2083 if (BuiltinID == Builtin::BIcommit_read_pipe) 2084 Name = "__commit_read_pipe"; 2085 else if (BuiltinID == Builtin::BIcommit_write_pipe) 2086 Name = "__commit_write_pipe"; 2087 else if (BuiltinID == Builtin::BIwork_group_commit_read_pipe) 2088 Name = "__work_group_commit_read_pipe"; 2089 else if (BuiltinID == Builtin::BIwork_group_commit_write_pipe) 2090 Name = "__work_group_commit_write_pipe"; 2091 else if (BuiltinID == Builtin::BIsub_group_commit_read_pipe) 2092 Name = "__sub_group_commit_read_pipe"; 2093 else 2094 Name = "__sub_group_commit_write_pipe"; 2095 2096 Value *Arg0 = EmitScalarExpr(E->getArg(0)), 2097 *Arg1 = EmitScalarExpr(E->getArg(1)); 2098 2099 // Building the generic function prototype. 2100 llvm::Type *ArgTys[] = {Arg0->getType(), Arg1->getType()}; 2101 llvm::FunctionType *FTy = 2102 llvm::FunctionType::get(llvm::Type::getVoidTy(getLLVMContext()), 2103 llvm::ArrayRef<llvm::Type *>(ArgTys), false); 2104 2105 return RValue::get( 2106 Builder.CreateCall(CGM.CreateRuntimeFunction(FTy, Name), {Arg0, Arg1})); 2107 } 2108 // OpenCL v2.0 s6.13.16.4 Built-in pipe query functions 2109 case Builtin::BIget_pipe_num_packets: 2110 case Builtin::BIget_pipe_max_packets: { 2111 const char *Name; 2112 if (BuiltinID == Builtin::BIget_pipe_num_packets) 2113 Name = "__get_pipe_num_packets"; 2114 else 2115 Name = "__get_pipe_max_packets"; 2116 2117 // Building the generic function prototype. 2118 Value *Arg0 = EmitScalarExpr(E->getArg(0)); 2119 llvm::Type *ArgTys[] = {Arg0->getType()}; 2120 llvm::FunctionType *FTy = llvm::FunctionType::get( 2121 Int32Ty, llvm::ArrayRef<llvm::Type *>(ArgTys), false); 2122 2123 return RValue::get( 2124 Builder.CreateCall(CGM.CreateRuntimeFunction(FTy, Name), {Arg0})); 2125 } 2126 2127 case Builtin::BIprintf: 2128 if (getLangOpts().CUDA && getLangOpts().CUDAIsDevice) 2129 return EmitCUDADevicePrintfCallExpr(E, ReturnValue); 2130 break; 2131 case Builtin::BI__builtin_canonicalize: 2132 case Builtin::BI__builtin_canonicalizef: 2133 case Builtin::BI__builtin_canonicalizel: 2134 return RValue::get(emitUnaryBuiltin(*this, E, Intrinsic::canonicalize)); 2135 } 2136 2137 // If this is an alias for a lib function (e.g. __builtin_sin), emit 2138 // the call using the normal call path, but using the unmangled 2139 // version of the function name. 2140 if (getContext().BuiltinInfo.isLibFunction(BuiltinID)) 2141 return emitLibraryCall(*this, FD, E, 2142 CGM.getBuiltinLibFunction(FD, BuiltinID)); 2143 2144 // If this is a predefined lib function (e.g. malloc), emit the call 2145 // using exactly the normal call path. 2146 if (getContext().BuiltinInfo.isPredefinedLibFunction(BuiltinID)) 2147 return emitLibraryCall(*this, FD, E, EmitScalarExpr(E->getCallee())); 2148 2149 // Check that a call to a target specific builtin has the correct target 2150 // features. 2151 // This is down here to avoid non-target specific builtins, however, if 2152 // generic builtins start to require generic target features then we 2153 // can move this up to the beginning of the function. 2154 checkTargetFeatures(E, FD); 2155 2156 // See if we have a target specific intrinsic. 2157 const char *Name = getContext().BuiltinInfo.getName(BuiltinID); 2158 Intrinsic::ID IntrinsicID = Intrinsic::not_intrinsic; 2159 if (const char *Prefix = 2160 llvm::Triple::getArchTypePrefix(getTarget().getTriple().getArch())) { 2161 IntrinsicID = Intrinsic::getIntrinsicForGCCBuiltin(Prefix, Name); 2162 // NOTE we dont need to perform a compatibility flag check here since the 2163 // intrinsics are declared in Builtins*.def via LANGBUILTIN which filter the 2164 // MS builtins via ALL_MS_LANGUAGES and are filtered earlier. 2165 if (IntrinsicID == Intrinsic::not_intrinsic) 2166 IntrinsicID = Intrinsic::getIntrinsicForMSBuiltin(Prefix, Name); 2167 } 2168 2169 if (IntrinsicID != Intrinsic::not_intrinsic) { 2170 SmallVector<Value*, 16> Args; 2171 2172 // Find out if any arguments are required to be integer constant 2173 // expressions. 2174 unsigned ICEArguments = 0; 2175 ASTContext::GetBuiltinTypeError Error; 2176 getContext().GetBuiltinType(BuiltinID, Error, &ICEArguments); 2177 assert(Error == ASTContext::GE_None && "Should not codegen an error"); 2178 2179 Function *F = CGM.getIntrinsic(IntrinsicID); 2180 llvm::FunctionType *FTy = F->getFunctionType(); 2181 2182 for (unsigned i = 0, e = E->getNumArgs(); i != e; ++i) { 2183 Value *ArgValue; 2184 // If this is a normal argument, just emit it as a scalar. 2185 if ((ICEArguments & (1 << i)) == 0) { 2186 ArgValue = EmitScalarExpr(E->getArg(i)); 2187 } else { 2188 // If this is required to be a constant, constant fold it so that we 2189 // know that the generated intrinsic gets a ConstantInt. 2190 llvm::APSInt Result; 2191 bool IsConst = E->getArg(i)->isIntegerConstantExpr(Result,getContext()); 2192 assert(IsConst && "Constant arg isn't actually constant?"); 2193 (void)IsConst; 2194 ArgValue = llvm::ConstantInt::get(getLLVMContext(), Result); 2195 } 2196 2197 // If the intrinsic arg type is different from the builtin arg type 2198 // we need to do a bit cast. 2199 llvm::Type *PTy = FTy->getParamType(i); 2200 if (PTy != ArgValue->getType()) { 2201 assert(PTy->canLosslesslyBitCastTo(FTy->getParamType(i)) && 2202 "Must be able to losslessly bit cast to param"); 2203 ArgValue = Builder.CreateBitCast(ArgValue, PTy); 2204 } 2205 2206 Args.push_back(ArgValue); 2207 } 2208 2209 Value *V = Builder.CreateCall(F, Args); 2210 QualType BuiltinRetType = E->getType(); 2211 2212 llvm::Type *RetTy = VoidTy; 2213 if (!BuiltinRetType->isVoidType()) 2214 RetTy = ConvertType(BuiltinRetType); 2215 2216 if (RetTy != V->getType()) { 2217 assert(V->getType()->canLosslesslyBitCastTo(RetTy) && 2218 "Must be able to losslessly bit cast result type"); 2219 V = Builder.CreateBitCast(V, RetTy); 2220 } 2221 2222 return RValue::get(V); 2223 } 2224 2225 // See if we have a target specific builtin that needs to be lowered. 2226 if (Value *V = EmitTargetBuiltinExpr(BuiltinID, E)) 2227 return RValue::get(V); 2228 2229 ErrorUnsupported(E, "builtin function"); 2230 2231 // Unknown builtin, for now just dump it out and return undef. 2232 return GetUndefRValue(E->getType()); 2233 } 2234 2235 static Value *EmitTargetArchBuiltinExpr(CodeGenFunction *CGF, 2236 unsigned BuiltinID, const CallExpr *E, 2237 llvm::Triple::ArchType Arch) { 2238 switch (Arch) { 2239 case llvm::Triple::arm: 2240 case llvm::Triple::armeb: 2241 case llvm::Triple::thumb: 2242 case llvm::Triple::thumbeb: 2243 return CGF->EmitARMBuiltinExpr(BuiltinID, E); 2244 case llvm::Triple::aarch64: 2245 case llvm::Triple::aarch64_be: 2246 return CGF->EmitAArch64BuiltinExpr(BuiltinID, E); 2247 case llvm::Triple::x86: 2248 case llvm::Triple::x86_64: 2249 return CGF->EmitX86BuiltinExpr(BuiltinID, E); 2250 case llvm::Triple::ppc: 2251 case llvm::Triple::ppc64: 2252 case llvm::Triple::ppc64le: 2253 return CGF->EmitPPCBuiltinExpr(BuiltinID, E); 2254 case llvm::Triple::r600: 2255 case llvm::Triple::amdgcn: 2256 return CGF->EmitAMDGPUBuiltinExpr(BuiltinID, E); 2257 case llvm::Triple::systemz: 2258 return CGF->EmitSystemZBuiltinExpr(BuiltinID, E); 2259 case llvm::Triple::nvptx: 2260 case llvm::Triple::nvptx64: 2261 return CGF->EmitNVPTXBuiltinExpr(BuiltinID, E); 2262 case llvm::Triple::wasm32: 2263 case llvm::Triple::wasm64: 2264 return CGF->EmitWebAssemblyBuiltinExpr(BuiltinID, E); 2265 default: 2266 return nullptr; 2267 } 2268 } 2269 2270 Value *CodeGenFunction::EmitTargetBuiltinExpr(unsigned BuiltinID, 2271 const CallExpr *E) { 2272 if (getContext().BuiltinInfo.isAuxBuiltinID(BuiltinID)) { 2273 assert(getContext().getAuxTargetInfo() && "Missing aux target info"); 2274 return EmitTargetArchBuiltinExpr( 2275 this, getContext().BuiltinInfo.getAuxBuiltinID(BuiltinID), E, 2276 getContext().getAuxTargetInfo()->getTriple().getArch()); 2277 } 2278 2279 return EmitTargetArchBuiltinExpr(this, BuiltinID, E, 2280 getTarget().getTriple().getArch()); 2281 } 2282 2283 static llvm::VectorType *GetNeonType(CodeGenFunction *CGF, 2284 NeonTypeFlags TypeFlags, 2285 bool V1Ty=false) { 2286 int IsQuad = TypeFlags.isQuad(); 2287 switch (TypeFlags.getEltType()) { 2288 case NeonTypeFlags::Int8: 2289 case NeonTypeFlags::Poly8: 2290 return llvm::VectorType::get(CGF->Int8Ty, V1Ty ? 1 : (8 << IsQuad)); 2291 case NeonTypeFlags::Int16: 2292 case NeonTypeFlags::Poly16: 2293 case NeonTypeFlags::Float16: 2294 return llvm::VectorType::get(CGF->Int16Ty, V1Ty ? 1 : (4 << IsQuad)); 2295 case NeonTypeFlags::Int32: 2296 return llvm::VectorType::get(CGF->Int32Ty, V1Ty ? 1 : (2 << IsQuad)); 2297 case NeonTypeFlags::Int64: 2298 case NeonTypeFlags::Poly64: 2299 return llvm::VectorType::get(CGF->Int64Ty, V1Ty ? 1 : (1 << IsQuad)); 2300 case NeonTypeFlags::Poly128: 2301 // FIXME: i128 and f128 doesn't get fully support in Clang and llvm. 2302 // There is a lot of i128 and f128 API missing. 2303 // so we use v16i8 to represent poly128 and get pattern matched. 2304 return llvm::VectorType::get(CGF->Int8Ty, 16); 2305 case NeonTypeFlags::Float32: 2306 return llvm::VectorType::get(CGF->FloatTy, V1Ty ? 1 : (2 << IsQuad)); 2307 case NeonTypeFlags::Float64: 2308 return llvm::VectorType::get(CGF->DoubleTy, V1Ty ? 1 : (1 << IsQuad)); 2309 } 2310 llvm_unreachable("Unknown vector element type!"); 2311 } 2312 2313 static llvm::VectorType *GetFloatNeonType(CodeGenFunction *CGF, 2314 NeonTypeFlags IntTypeFlags) { 2315 int IsQuad = IntTypeFlags.isQuad(); 2316 switch (IntTypeFlags.getEltType()) { 2317 case NeonTypeFlags::Int32: 2318 return llvm::VectorType::get(CGF->FloatTy, (2 << IsQuad)); 2319 case NeonTypeFlags::Int64: 2320 return llvm::VectorType::get(CGF->DoubleTy, (1 << IsQuad)); 2321 default: 2322 llvm_unreachable("Type can't be converted to floating-point!"); 2323 } 2324 } 2325 2326 Value *CodeGenFunction::EmitNeonSplat(Value *V, Constant *C) { 2327 unsigned nElts = cast<llvm::VectorType>(V->getType())->getNumElements(); 2328 Value* SV = llvm::ConstantVector::getSplat(nElts, C); 2329 return Builder.CreateShuffleVector(V, V, SV, "lane"); 2330 } 2331 2332 Value *CodeGenFunction::EmitNeonCall(Function *F, SmallVectorImpl<Value*> &Ops, 2333 const char *name, 2334 unsigned shift, bool rightshift) { 2335 unsigned j = 0; 2336 for (Function::const_arg_iterator ai = F->arg_begin(), ae = F->arg_end(); 2337 ai != ae; ++ai, ++j) 2338 if (shift > 0 && shift == j) 2339 Ops[j] = EmitNeonShiftVector(Ops[j], ai->getType(), rightshift); 2340 else 2341 Ops[j] = Builder.CreateBitCast(Ops[j], ai->getType(), name); 2342 2343 return Builder.CreateCall(F, Ops, name); 2344 } 2345 2346 Value *CodeGenFunction::EmitNeonShiftVector(Value *V, llvm::Type *Ty, 2347 bool neg) { 2348 int SV = cast<ConstantInt>(V)->getSExtValue(); 2349 return ConstantInt::get(Ty, neg ? -SV : SV); 2350 } 2351 2352 // \brief Right-shift a vector by a constant. 2353 Value *CodeGenFunction::EmitNeonRShiftImm(Value *Vec, Value *Shift, 2354 llvm::Type *Ty, bool usgn, 2355 const char *name) { 2356 llvm::VectorType *VTy = cast<llvm::VectorType>(Ty); 2357 2358 int ShiftAmt = cast<ConstantInt>(Shift)->getSExtValue(); 2359 int EltSize = VTy->getScalarSizeInBits(); 2360 2361 Vec = Builder.CreateBitCast(Vec, Ty); 2362 2363 // lshr/ashr are undefined when the shift amount is equal to the vector 2364 // element size. 2365 if (ShiftAmt == EltSize) { 2366 if (usgn) { 2367 // Right-shifting an unsigned value by its size yields 0. 2368 return llvm::ConstantAggregateZero::get(VTy); 2369 } else { 2370 // Right-shifting a signed value by its size is equivalent 2371 // to a shift of size-1. 2372 --ShiftAmt; 2373 Shift = ConstantInt::get(VTy->getElementType(), ShiftAmt); 2374 } 2375 } 2376 2377 Shift = EmitNeonShiftVector(Shift, Ty, false); 2378 if (usgn) 2379 return Builder.CreateLShr(Vec, Shift, name); 2380 else 2381 return Builder.CreateAShr(Vec, Shift, name); 2382 } 2383 2384 enum { 2385 AddRetType = (1 << 0), 2386 Add1ArgType = (1 << 1), 2387 Add2ArgTypes = (1 << 2), 2388 2389 VectorizeRetType = (1 << 3), 2390 VectorizeArgTypes = (1 << 4), 2391 2392 InventFloatType = (1 << 5), 2393 UnsignedAlts = (1 << 6), 2394 2395 Use64BitVectors = (1 << 7), 2396 Use128BitVectors = (1 << 8), 2397 2398 Vectorize1ArgType = Add1ArgType | VectorizeArgTypes, 2399 VectorRet = AddRetType | VectorizeRetType, 2400 VectorRetGetArgs01 = 2401 AddRetType | Add2ArgTypes | VectorizeRetType | VectorizeArgTypes, 2402 FpCmpzModifiers = 2403 AddRetType | VectorizeRetType | Add1ArgType | InventFloatType 2404 }; 2405 2406 namespace { 2407 struct NeonIntrinsicInfo { 2408 const char *NameHint; 2409 unsigned BuiltinID; 2410 unsigned LLVMIntrinsic; 2411 unsigned AltLLVMIntrinsic; 2412 unsigned TypeModifier; 2413 2414 bool operator<(unsigned RHSBuiltinID) const { 2415 return BuiltinID < RHSBuiltinID; 2416 } 2417 bool operator<(const NeonIntrinsicInfo &TE) const { 2418 return BuiltinID < TE.BuiltinID; 2419 } 2420 }; 2421 } // end anonymous namespace 2422 2423 #define NEONMAP0(NameBase) \ 2424 { #NameBase, NEON::BI__builtin_neon_ ## NameBase, 0, 0, 0 } 2425 2426 #define NEONMAP1(NameBase, LLVMIntrinsic, TypeModifier) \ 2427 { #NameBase, NEON:: BI__builtin_neon_ ## NameBase, \ 2428 Intrinsic::LLVMIntrinsic, 0, TypeModifier } 2429 2430 #define NEONMAP2(NameBase, LLVMIntrinsic, AltLLVMIntrinsic, TypeModifier) \ 2431 { #NameBase, NEON:: BI__builtin_neon_ ## NameBase, \ 2432 Intrinsic::LLVMIntrinsic, Intrinsic::AltLLVMIntrinsic, \ 2433 TypeModifier } 2434 2435 static const NeonIntrinsicInfo ARMSIMDIntrinsicMap [] = { 2436 NEONMAP2(vabd_v, arm_neon_vabdu, arm_neon_vabds, Add1ArgType | UnsignedAlts), 2437 NEONMAP2(vabdq_v, arm_neon_vabdu, arm_neon_vabds, Add1ArgType | UnsignedAlts), 2438 NEONMAP1(vabs_v, arm_neon_vabs, 0), 2439 NEONMAP1(vabsq_v, arm_neon_vabs, 0), 2440 NEONMAP0(vaddhn_v), 2441 NEONMAP1(vaesdq_v, arm_neon_aesd, 0), 2442 NEONMAP1(vaeseq_v, arm_neon_aese, 0), 2443 NEONMAP1(vaesimcq_v, arm_neon_aesimc, 0), 2444 NEONMAP1(vaesmcq_v, arm_neon_aesmc, 0), 2445 NEONMAP1(vbsl_v, arm_neon_vbsl, AddRetType), 2446 NEONMAP1(vbslq_v, arm_neon_vbsl, AddRetType), 2447 NEONMAP1(vcage_v, arm_neon_vacge, 0), 2448 NEONMAP1(vcageq_v, arm_neon_vacge, 0), 2449 NEONMAP1(vcagt_v, arm_neon_vacgt, 0), 2450 NEONMAP1(vcagtq_v, arm_neon_vacgt, 0), 2451 NEONMAP1(vcale_v, arm_neon_vacge, 0), 2452 NEONMAP1(vcaleq_v, arm_neon_vacge, 0), 2453 NEONMAP1(vcalt_v, arm_neon_vacgt, 0), 2454 NEONMAP1(vcaltq_v, arm_neon_vacgt, 0), 2455 NEONMAP1(vcls_v, arm_neon_vcls, Add1ArgType), 2456 NEONMAP1(vclsq_v, arm_neon_vcls, Add1ArgType), 2457 NEONMAP1(vclz_v, ctlz, Add1ArgType), 2458 NEONMAP1(vclzq_v, ctlz, Add1ArgType), 2459 NEONMAP1(vcnt_v, ctpop, Add1ArgType), 2460 NEONMAP1(vcntq_v, ctpop, Add1ArgType), 2461 NEONMAP1(vcvt_f16_f32, arm_neon_vcvtfp2hf, 0), 2462 NEONMAP1(vcvt_f32_f16, arm_neon_vcvthf2fp, 0), 2463 NEONMAP0(vcvt_f32_v), 2464 NEONMAP2(vcvt_n_f32_v, arm_neon_vcvtfxu2fp, arm_neon_vcvtfxs2fp, 0), 2465 NEONMAP1(vcvt_n_s32_v, arm_neon_vcvtfp2fxs, 0), 2466 NEONMAP1(vcvt_n_s64_v, arm_neon_vcvtfp2fxs, 0), 2467 NEONMAP1(vcvt_n_u32_v, arm_neon_vcvtfp2fxu, 0), 2468 NEONMAP1(vcvt_n_u64_v, arm_neon_vcvtfp2fxu, 0), 2469 NEONMAP0(vcvt_s32_v), 2470 NEONMAP0(vcvt_s64_v), 2471 NEONMAP0(vcvt_u32_v), 2472 NEONMAP0(vcvt_u64_v), 2473 NEONMAP1(vcvta_s32_v, arm_neon_vcvtas, 0), 2474 NEONMAP1(vcvta_s64_v, arm_neon_vcvtas, 0), 2475 NEONMAP1(vcvta_u32_v, arm_neon_vcvtau, 0), 2476 NEONMAP1(vcvta_u64_v, arm_neon_vcvtau, 0), 2477 NEONMAP1(vcvtaq_s32_v, arm_neon_vcvtas, 0), 2478 NEONMAP1(vcvtaq_s64_v, arm_neon_vcvtas, 0), 2479 NEONMAP1(vcvtaq_u32_v, arm_neon_vcvtau, 0), 2480 NEONMAP1(vcvtaq_u64_v, arm_neon_vcvtau, 0), 2481 NEONMAP1(vcvtm_s32_v, arm_neon_vcvtms, 0), 2482 NEONMAP1(vcvtm_s64_v, arm_neon_vcvtms, 0), 2483 NEONMAP1(vcvtm_u32_v, arm_neon_vcvtmu, 0), 2484 NEONMAP1(vcvtm_u64_v, arm_neon_vcvtmu, 0), 2485 NEONMAP1(vcvtmq_s32_v, arm_neon_vcvtms, 0), 2486 NEONMAP1(vcvtmq_s64_v, arm_neon_vcvtms, 0), 2487 NEONMAP1(vcvtmq_u32_v, arm_neon_vcvtmu, 0), 2488 NEONMAP1(vcvtmq_u64_v, arm_neon_vcvtmu, 0), 2489 NEONMAP1(vcvtn_s32_v, arm_neon_vcvtns, 0), 2490 NEONMAP1(vcvtn_s64_v, arm_neon_vcvtns, 0), 2491 NEONMAP1(vcvtn_u32_v, arm_neon_vcvtnu, 0), 2492 NEONMAP1(vcvtn_u64_v, arm_neon_vcvtnu, 0), 2493 NEONMAP1(vcvtnq_s32_v, arm_neon_vcvtns, 0), 2494 NEONMAP1(vcvtnq_s64_v, arm_neon_vcvtns, 0), 2495 NEONMAP1(vcvtnq_u32_v, arm_neon_vcvtnu, 0), 2496 NEONMAP1(vcvtnq_u64_v, arm_neon_vcvtnu, 0), 2497 NEONMAP1(vcvtp_s32_v, arm_neon_vcvtps, 0), 2498 NEONMAP1(vcvtp_s64_v, arm_neon_vcvtps, 0), 2499 NEONMAP1(vcvtp_u32_v, arm_neon_vcvtpu, 0), 2500 NEONMAP1(vcvtp_u64_v, arm_neon_vcvtpu, 0), 2501 NEONMAP1(vcvtpq_s32_v, arm_neon_vcvtps, 0), 2502 NEONMAP1(vcvtpq_s64_v, arm_neon_vcvtps, 0), 2503 NEONMAP1(vcvtpq_u32_v, arm_neon_vcvtpu, 0), 2504 NEONMAP1(vcvtpq_u64_v, arm_neon_vcvtpu, 0), 2505 NEONMAP0(vcvtq_f32_v), 2506 NEONMAP2(vcvtq_n_f32_v, arm_neon_vcvtfxu2fp, arm_neon_vcvtfxs2fp, 0), 2507 NEONMAP1(vcvtq_n_s32_v, arm_neon_vcvtfp2fxs, 0), 2508 NEONMAP1(vcvtq_n_s64_v, arm_neon_vcvtfp2fxs, 0), 2509 NEONMAP1(vcvtq_n_u32_v, arm_neon_vcvtfp2fxu, 0), 2510 NEONMAP1(vcvtq_n_u64_v, arm_neon_vcvtfp2fxu, 0), 2511 NEONMAP0(vcvtq_s32_v), 2512 NEONMAP0(vcvtq_s64_v), 2513 NEONMAP0(vcvtq_u32_v), 2514 NEONMAP0(vcvtq_u64_v), 2515 NEONMAP0(vext_v), 2516 NEONMAP0(vextq_v), 2517 NEONMAP0(vfma_v), 2518 NEONMAP0(vfmaq_v), 2519 NEONMAP2(vhadd_v, arm_neon_vhaddu, arm_neon_vhadds, Add1ArgType | UnsignedAlts), 2520 NEONMAP2(vhaddq_v, arm_neon_vhaddu, arm_neon_vhadds, Add1ArgType | UnsignedAlts), 2521 NEONMAP2(vhsub_v, arm_neon_vhsubu, arm_neon_vhsubs, Add1ArgType | UnsignedAlts), 2522 NEONMAP2(vhsubq_v, arm_neon_vhsubu, arm_neon_vhsubs, Add1ArgType | UnsignedAlts), 2523 NEONMAP0(vld1_dup_v), 2524 NEONMAP1(vld1_v, arm_neon_vld1, 0), 2525 NEONMAP0(vld1q_dup_v), 2526 NEONMAP1(vld1q_v, arm_neon_vld1, 0), 2527 NEONMAP1(vld2_lane_v, arm_neon_vld2lane, 0), 2528 NEONMAP1(vld2_v, arm_neon_vld2, 0), 2529 NEONMAP1(vld2q_lane_v, arm_neon_vld2lane, 0), 2530 NEONMAP1(vld2q_v, arm_neon_vld2, 0), 2531 NEONMAP1(vld3_lane_v, arm_neon_vld3lane, 0), 2532 NEONMAP1(vld3_v, arm_neon_vld3, 0), 2533 NEONMAP1(vld3q_lane_v, arm_neon_vld3lane, 0), 2534 NEONMAP1(vld3q_v, arm_neon_vld3, 0), 2535 NEONMAP1(vld4_lane_v, arm_neon_vld4lane, 0), 2536 NEONMAP1(vld4_v, arm_neon_vld4, 0), 2537 NEONMAP1(vld4q_lane_v, arm_neon_vld4lane, 0), 2538 NEONMAP1(vld4q_v, arm_neon_vld4, 0), 2539 NEONMAP2(vmax_v, arm_neon_vmaxu, arm_neon_vmaxs, Add1ArgType | UnsignedAlts), 2540 NEONMAP1(vmaxnm_v, arm_neon_vmaxnm, Add1ArgType), 2541 NEONMAP1(vmaxnmq_v, arm_neon_vmaxnm, Add1ArgType), 2542 NEONMAP2(vmaxq_v, arm_neon_vmaxu, arm_neon_vmaxs, Add1ArgType | UnsignedAlts), 2543 NEONMAP2(vmin_v, arm_neon_vminu, arm_neon_vmins, Add1ArgType | UnsignedAlts), 2544 NEONMAP1(vminnm_v, arm_neon_vminnm, Add1ArgType), 2545 NEONMAP1(vminnmq_v, arm_neon_vminnm, Add1ArgType), 2546 NEONMAP2(vminq_v, arm_neon_vminu, arm_neon_vmins, Add1ArgType | UnsignedAlts), 2547 NEONMAP0(vmovl_v), 2548 NEONMAP0(vmovn_v), 2549 NEONMAP1(vmul_v, arm_neon_vmulp, Add1ArgType), 2550 NEONMAP0(vmull_v), 2551 NEONMAP1(vmulq_v, arm_neon_vmulp, Add1ArgType), 2552 NEONMAP2(vpadal_v, arm_neon_vpadalu, arm_neon_vpadals, UnsignedAlts), 2553 NEONMAP2(vpadalq_v, arm_neon_vpadalu, arm_neon_vpadals, UnsignedAlts), 2554 NEONMAP1(vpadd_v, arm_neon_vpadd, Add1ArgType), 2555 NEONMAP2(vpaddl_v, arm_neon_vpaddlu, arm_neon_vpaddls, UnsignedAlts), 2556 NEONMAP2(vpaddlq_v, arm_neon_vpaddlu, arm_neon_vpaddls, UnsignedAlts), 2557 NEONMAP1(vpaddq_v, arm_neon_vpadd, Add1ArgType), 2558 NEONMAP2(vpmax_v, arm_neon_vpmaxu, arm_neon_vpmaxs, Add1ArgType | UnsignedAlts), 2559 NEONMAP2(vpmin_v, arm_neon_vpminu, arm_neon_vpmins, Add1ArgType | UnsignedAlts), 2560 NEONMAP1(vqabs_v, arm_neon_vqabs, Add1ArgType), 2561 NEONMAP1(vqabsq_v, arm_neon_vqabs, Add1ArgType), 2562 NEONMAP2(vqadd_v, arm_neon_vqaddu, arm_neon_vqadds, Add1ArgType | UnsignedAlts), 2563 NEONMAP2(vqaddq_v, arm_neon_vqaddu, arm_neon_vqadds, Add1ArgType | UnsignedAlts), 2564 NEONMAP2(vqdmlal_v, arm_neon_vqdmull, arm_neon_vqadds, 0), 2565 NEONMAP2(vqdmlsl_v, arm_neon_vqdmull, arm_neon_vqsubs, 0), 2566 NEONMAP1(vqdmulh_v, arm_neon_vqdmulh, Add1ArgType), 2567 NEONMAP1(vqdmulhq_v, arm_neon_vqdmulh, Add1ArgType), 2568 NEONMAP1(vqdmull_v, arm_neon_vqdmull, Add1ArgType), 2569 NEONMAP2(vqmovn_v, arm_neon_vqmovnu, arm_neon_vqmovns, Add1ArgType | UnsignedAlts), 2570 NEONMAP1(vqmovun_v, arm_neon_vqmovnsu, Add1ArgType), 2571 NEONMAP1(vqneg_v, arm_neon_vqneg, Add1ArgType), 2572 NEONMAP1(vqnegq_v, arm_neon_vqneg, Add1ArgType), 2573 NEONMAP1(vqrdmulh_v, arm_neon_vqrdmulh, Add1ArgType), 2574 NEONMAP1(vqrdmulhq_v, arm_neon_vqrdmulh, Add1ArgType), 2575 NEONMAP2(vqrshl_v, arm_neon_vqrshiftu, arm_neon_vqrshifts, Add1ArgType | UnsignedAlts), 2576 NEONMAP2(vqrshlq_v, arm_neon_vqrshiftu, arm_neon_vqrshifts, Add1ArgType | UnsignedAlts), 2577 NEONMAP2(vqshl_n_v, arm_neon_vqshiftu, arm_neon_vqshifts, UnsignedAlts), 2578 NEONMAP2(vqshl_v, arm_neon_vqshiftu, arm_neon_vqshifts, Add1ArgType | UnsignedAlts), 2579 NEONMAP2(vqshlq_n_v, arm_neon_vqshiftu, arm_neon_vqshifts, UnsignedAlts), 2580 NEONMAP2(vqshlq_v, arm_neon_vqshiftu, arm_neon_vqshifts, Add1ArgType | UnsignedAlts), 2581 NEONMAP1(vqshlu_n_v, arm_neon_vqshiftsu, 0), 2582 NEONMAP1(vqshluq_n_v, arm_neon_vqshiftsu, 0), 2583 NEONMAP2(vqsub_v, arm_neon_vqsubu, arm_neon_vqsubs, Add1ArgType | UnsignedAlts), 2584 NEONMAP2(vqsubq_v, arm_neon_vqsubu, arm_neon_vqsubs, Add1ArgType | UnsignedAlts), 2585 NEONMAP1(vraddhn_v, arm_neon_vraddhn, Add1ArgType), 2586 NEONMAP2(vrecpe_v, arm_neon_vrecpe, arm_neon_vrecpe, 0), 2587 NEONMAP2(vrecpeq_v, arm_neon_vrecpe, arm_neon_vrecpe, 0), 2588 NEONMAP1(vrecps_v, arm_neon_vrecps, Add1ArgType), 2589 NEONMAP1(vrecpsq_v, arm_neon_vrecps, Add1ArgType), 2590 NEONMAP2(vrhadd_v, arm_neon_vrhaddu, arm_neon_vrhadds, Add1ArgType | UnsignedAlts), 2591 NEONMAP2(vrhaddq_v, arm_neon_vrhaddu, arm_neon_vrhadds, Add1ArgType | UnsignedAlts), 2592 NEONMAP1(vrnd_v, arm_neon_vrintz, Add1ArgType), 2593 NEONMAP1(vrnda_v, arm_neon_vrinta, Add1ArgType), 2594 NEONMAP1(vrndaq_v, arm_neon_vrinta, Add1ArgType), 2595 NEONMAP1(vrndm_v, arm_neon_vrintm, Add1ArgType), 2596 NEONMAP1(vrndmq_v, arm_neon_vrintm, Add1ArgType), 2597 NEONMAP1(vrndn_v, arm_neon_vrintn, Add1ArgType), 2598 NEONMAP1(vrndnq_v, arm_neon_vrintn, Add1ArgType), 2599 NEONMAP1(vrndp_v, arm_neon_vrintp, Add1ArgType), 2600 NEONMAP1(vrndpq_v, arm_neon_vrintp, Add1ArgType), 2601 NEONMAP1(vrndq_v, arm_neon_vrintz, Add1ArgType), 2602 NEONMAP1(vrndx_v, arm_neon_vrintx, Add1ArgType), 2603 NEONMAP1(vrndxq_v, arm_neon_vrintx, Add1ArgType), 2604 NEONMAP2(vrshl_v, arm_neon_vrshiftu, arm_neon_vrshifts, Add1ArgType | UnsignedAlts), 2605 NEONMAP2(vrshlq_v, arm_neon_vrshiftu, arm_neon_vrshifts, Add1ArgType | UnsignedAlts), 2606 NEONMAP2(vrshr_n_v, arm_neon_vrshiftu, arm_neon_vrshifts, UnsignedAlts), 2607 NEONMAP2(vrshrq_n_v, arm_neon_vrshiftu, arm_neon_vrshifts, UnsignedAlts), 2608 NEONMAP2(vrsqrte_v, arm_neon_vrsqrte, arm_neon_vrsqrte, 0), 2609 NEONMAP2(vrsqrteq_v, arm_neon_vrsqrte, arm_neon_vrsqrte, 0), 2610 NEONMAP1(vrsqrts_v, arm_neon_vrsqrts, Add1ArgType), 2611 NEONMAP1(vrsqrtsq_v, arm_neon_vrsqrts, Add1ArgType), 2612 NEONMAP1(vrsubhn_v, arm_neon_vrsubhn, Add1ArgType), 2613 NEONMAP1(vsha1su0q_v, arm_neon_sha1su0, 0), 2614 NEONMAP1(vsha1su1q_v, arm_neon_sha1su1, 0), 2615 NEONMAP1(vsha256h2q_v, arm_neon_sha256h2, 0), 2616 NEONMAP1(vsha256hq_v, arm_neon_sha256h, 0), 2617 NEONMAP1(vsha256su0q_v, arm_neon_sha256su0, 0), 2618 NEONMAP1(vsha256su1q_v, arm_neon_sha256su1, 0), 2619 NEONMAP0(vshl_n_v), 2620 NEONMAP2(vshl_v, arm_neon_vshiftu, arm_neon_vshifts, Add1ArgType | UnsignedAlts), 2621 NEONMAP0(vshll_n_v), 2622 NEONMAP0(vshlq_n_v), 2623 NEONMAP2(vshlq_v, arm_neon_vshiftu, arm_neon_vshifts, Add1ArgType | UnsignedAlts), 2624 NEONMAP0(vshr_n_v), 2625 NEONMAP0(vshrn_n_v), 2626 NEONMAP0(vshrq_n_v), 2627 NEONMAP1(vst1_v, arm_neon_vst1, 0), 2628 NEONMAP1(vst1q_v, arm_neon_vst1, 0), 2629 NEONMAP1(vst2_lane_v, arm_neon_vst2lane, 0), 2630 NEONMAP1(vst2_v, arm_neon_vst2, 0), 2631 NEONMAP1(vst2q_lane_v, arm_neon_vst2lane, 0), 2632 NEONMAP1(vst2q_v, arm_neon_vst2, 0), 2633 NEONMAP1(vst3_lane_v, arm_neon_vst3lane, 0), 2634 NEONMAP1(vst3_v, arm_neon_vst3, 0), 2635 NEONMAP1(vst3q_lane_v, arm_neon_vst3lane, 0), 2636 NEONMAP1(vst3q_v, arm_neon_vst3, 0), 2637 NEONMAP1(vst4_lane_v, arm_neon_vst4lane, 0), 2638 NEONMAP1(vst4_v, arm_neon_vst4, 0), 2639 NEONMAP1(vst4q_lane_v, arm_neon_vst4lane, 0), 2640 NEONMAP1(vst4q_v, arm_neon_vst4, 0), 2641 NEONMAP0(vsubhn_v), 2642 NEONMAP0(vtrn_v), 2643 NEONMAP0(vtrnq_v), 2644 NEONMAP0(vtst_v), 2645 NEONMAP0(vtstq_v), 2646 NEONMAP0(vuzp_v), 2647 NEONMAP0(vuzpq_v), 2648 NEONMAP0(vzip_v), 2649 NEONMAP0(vzipq_v) 2650 }; 2651 2652 static const NeonIntrinsicInfo AArch64SIMDIntrinsicMap[] = { 2653 NEONMAP1(vabs_v, aarch64_neon_abs, 0), 2654 NEONMAP1(vabsq_v, aarch64_neon_abs, 0), 2655 NEONMAP0(vaddhn_v), 2656 NEONMAP1(vaesdq_v, aarch64_crypto_aesd, 0), 2657 NEONMAP1(vaeseq_v, aarch64_crypto_aese, 0), 2658 NEONMAP1(vaesimcq_v, aarch64_crypto_aesimc, 0), 2659 NEONMAP1(vaesmcq_v, aarch64_crypto_aesmc, 0), 2660 NEONMAP1(vcage_v, aarch64_neon_facge, 0), 2661 NEONMAP1(vcageq_v, aarch64_neon_facge, 0), 2662 NEONMAP1(vcagt_v, aarch64_neon_facgt, 0), 2663 NEONMAP1(vcagtq_v, aarch64_neon_facgt, 0), 2664 NEONMAP1(vcale_v, aarch64_neon_facge, 0), 2665 NEONMAP1(vcaleq_v, aarch64_neon_facge, 0), 2666 NEONMAP1(vcalt_v, aarch64_neon_facgt, 0), 2667 NEONMAP1(vcaltq_v, aarch64_neon_facgt, 0), 2668 NEONMAP1(vcls_v, aarch64_neon_cls, Add1ArgType), 2669 NEONMAP1(vclsq_v, aarch64_neon_cls, Add1ArgType), 2670 NEONMAP1(vclz_v, ctlz, Add1ArgType), 2671 NEONMAP1(vclzq_v, ctlz, Add1ArgType), 2672 NEONMAP1(vcnt_v, ctpop, Add1ArgType), 2673 NEONMAP1(vcntq_v, ctpop, Add1ArgType), 2674 NEONMAP1(vcvt_f16_f32, aarch64_neon_vcvtfp2hf, 0), 2675 NEONMAP1(vcvt_f32_f16, aarch64_neon_vcvthf2fp, 0), 2676 NEONMAP0(vcvt_f32_v), 2677 NEONMAP2(vcvt_n_f32_v, aarch64_neon_vcvtfxu2fp, aarch64_neon_vcvtfxs2fp, 0), 2678 NEONMAP2(vcvt_n_f64_v, aarch64_neon_vcvtfxu2fp, aarch64_neon_vcvtfxs2fp, 0), 2679 NEONMAP1(vcvt_n_s32_v, aarch64_neon_vcvtfp2fxs, 0), 2680 NEONMAP1(vcvt_n_s64_v, aarch64_neon_vcvtfp2fxs, 0), 2681 NEONMAP1(vcvt_n_u32_v, aarch64_neon_vcvtfp2fxu, 0), 2682 NEONMAP1(vcvt_n_u64_v, aarch64_neon_vcvtfp2fxu, 0), 2683 NEONMAP0(vcvtq_f32_v), 2684 NEONMAP2(vcvtq_n_f32_v, aarch64_neon_vcvtfxu2fp, aarch64_neon_vcvtfxs2fp, 0), 2685 NEONMAP2(vcvtq_n_f64_v, aarch64_neon_vcvtfxu2fp, aarch64_neon_vcvtfxs2fp, 0), 2686 NEONMAP1(vcvtq_n_s32_v, aarch64_neon_vcvtfp2fxs, 0), 2687 NEONMAP1(vcvtq_n_s64_v, aarch64_neon_vcvtfp2fxs, 0), 2688 NEONMAP1(vcvtq_n_u32_v, aarch64_neon_vcvtfp2fxu, 0), 2689 NEONMAP1(vcvtq_n_u64_v, aarch64_neon_vcvtfp2fxu, 0), 2690 NEONMAP1(vcvtx_f32_v, aarch64_neon_fcvtxn, AddRetType | Add1ArgType), 2691 NEONMAP0(vext_v), 2692 NEONMAP0(vextq_v), 2693 NEONMAP0(vfma_v), 2694 NEONMAP0(vfmaq_v), 2695 NEONMAP2(vhadd_v, aarch64_neon_uhadd, aarch64_neon_shadd, Add1ArgType | UnsignedAlts), 2696 NEONMAP2(vhaddq_v, aarch64_neon_uhadd, aarch64_neon_shadd, Add1ArgType | UnsignedAlts), 2697 NEONMAP2(vhsub_v, aarch64_neon_uhsub, aarch64_neon_shsub, Add1ArgType | UnsignedAlts), 2698 NEONMAP2(vhsubq_v, aarch64_neon_uhsub, aarch64_neon_shsub, Add1ArgType | UnsignedAlts), 2699 NEONMAP0(vmovl_v), 2700 NEONMAP0(vmovn_v), 2701 NEONMAP1(vmul_v, aarch64_neon_pmul, Add1ArgType), 2702 NEONMAP1(vmulq_v, aarch64_neon_pmul, Add1ArgType), 2703 NEONMAP1(vpadd_v, aarch64_neon_addp, Add1ArgType), 2704 NEONMAP2(vpaddl_v, aarch64_neon_uaddlp, aarch64_neon_saddlp, UnsignedAlts), 2705 NEONMAP2(vpaddlq_v, aarch64_neon_uaddlp, aarch64_neon_saddlp, UnsignedAlts), 2706 NEONMAP1(vpaddq_v, aarch64_neon_addp, Add1ArgType), 2707 NEONMAP1(vqabs_v, aarch64_neon_sqabs, Add1ArgType), 2708 NEONMAP1(vqabsq_v, aarch64_neon_sqabs, Add1ArgType), 2709 NEONMAP2(vqadd_v, aarch64_neon_uqadd, aarch64_neon_sqadd, Add1ArgType | UnsignedAlts), 2710 NEONMAP2(vqaddq_v, aarch64_neon_uqadd, aarch64_neon_sqadd, Add1ArgType | UnsignedAlts), 2711 NEONMAP2(vqdmlal_v, aarch64_neon_sqdmull, aarch64_neon_sqadd, 0), 2712 NEONMAP2(vqdmlsl_v, aarch64_neon_sqdmull, aarch64_neon_sqsub, 0), 2713 NEONMAP1(vqdmulh_v, aarch64_neon_sqdmulh, Add1ArgType), 2714 NEONMAP1(vqdmulhq_v, aarch64_neon_sqdmulh, Add1ArgType), 2715 NEONMAP1(vqdmull_v, aarch64_neon_sqdmull, Add1ArgType), 2716 NEONMAP2(vqmovn_v, aarch64_neon_uqxtn, aarch64_neon_sqxtn, Add1ArgType | UnsignedAlts), 2717 NEONMAP1(vqmovun_v, aarch64_neon_sqxtun, Add1ArgType), 2718 NEONMAP1(vqneg_v, aarch64_neon_sqneg, Add1ArgType), 2719 NEONMAP1(vqnegq_v, aarch64_neon_sqneg, Add1ArgType), 2720 NEONMAP1(vqrdmulh_v, aarch64_neon_sqrdmulh, Add1ArgType), 2721 NEONMAP1(vqrdmulhq_v, aarch64_neon_sqrdmulh, Add1ArgType), 2722 NEONMAP2(vqrshl_v, aarch64_neon_uqrshl, aarch64_neon_sqrshl, Add1ArgType | UnsignedAlts), 2723 NEONMAP2(vqrshlq_v, aarch64_neon_uqrshl, aarch64_neon_sqrshl, Add1ArgType | UnsignedAlts), 2724 NEONMAP2(vqshl_n_v, aarch64_neon_uqshl, aarch64_neon_sqshl, UnsignedAlts), 2725 NEONMAP2(vqshl_v, aarch64_neon_uqshl, aarch64_neon_sqshl, Add1ArgType | UnsignedAlts), 2726 NEONMAP2(vqshlq_n_v, aarch64_neon_uqshl, aarch64_neon_sqshl,UnsignedAlts), 2727 NEONMAP2(vqshlq_v, aarch64_neon_uqshl, aarch64_neon_sqshl, Add1ArgType | UnsignedAlts), 2728 NEONMAP1(vqshlu_n_v, aarch64_neon_sqshlu, 0), 2729 NEONMAP1(vqshluq_n_v, aarch64_neon_sqshlu, 0), 2730 NEONMAP2(vqsub_v, aarch64_neon_uqsub, aarch64_neon_sqsub, Add1ArgType | UnsignedAlts), 2731 NEONMAP2(vqsubq_v, aarch64_neon_uqsub, aarch64_neon_sqsub, Add1ArgType | UnsignedAlts), 2732 NEONMAP1(vraddhn_v, aarch64_neon_raddhn, Add1ArgType), 2733 NEONMAP2(vrecpe_v, aarch64_neon_frecpe, aarch64_neon_urecpe, 0), 2734 NEONMAP2(vrecpeq_v, aarch64_neon_frecpe, aarch64_neon_urecpe, 0), 2735 NEONMAP1(vrecps_v, aarch64_neon_frecps, Add1ArgType), 2736 NEONMAP1(vrecpsq_v, aarch64_neon_frecps, Add1ArgType), 2737 NEONMAP2(vrhadd_v, aarch64_neon_urhadd, aarch64_neon_srhadd, Add1ArgType | UnsignedAlts), 2738 NEONMAP2(vrhaddq_v, aarch64_neon_urhadd, aarch64_neon_srhadd, Add1ArgType | UnsignedAlts), 2739 NEONMAP2(vrshl_v, aarch64_neon_urshl, aarch64_neon_srshl, Add1ArgType | UnsignedAlts), 2740 NEONMAP2(vrshlq_v, aarch64_neon_urshl, aarch64_neon_srshl, Add1ArgType | UnsignedAlts), 2741 NEONMAP2(vrshr_n_v, aarch64_neon_urshl, aarch64_neon_srshl, UnsignedAlts), 2742 NEONMAP2(vrshrq_n_v, aarch64_neon_urshl, aarch64_neon_srshl, UnsignedAlts), 2743 NEONMAP2(vrsqrte_v, aarch64_neon_frsqrte, aarch64_neon_ursqrte, 0), 2744 NEONMAP2(vrsqrteq_v, aarch64_neon_frsqrte, aarch64_neon_ursqrte, 0), 2745 NEONMAP1(vrsqrts_v, aarch64_neon_frsqrts, Add1ArgType), 2746 NEONMAP1(vrsqrtsq_v, aarch64_neon_frsqrts, Add1ArgType), 2747 NEONMAP1(vrsubhn_v, aarch64_neon_rsubhn, Add1ArgType), 2748 NEONMAP1(vsha1su0q_v, aarch64_crypto_sha1su0, 0), 2749 NEONMAP1(vsha1su1q_v, aarch64_crypto_sha1su1, 0), 2750 NEONMAP1(vsha256h2q_v, aarch64_crypto_sha256h2, 0), 2751 NEONMAP1(vsha256hq_v, aarch64_crypto_sha256h, 0), 2752 NEONMAP1(vsha256su0q_v, aarch64_crypto_sha256su0, 0), 2753 NEONMAP1(vsha256su1q_v, aarch64_crypto_sha256su1, 0), 2754 NEONMAP0(vshl_n_v), 2755 NEONMAP2(vshl_v, aarch64_neon_ushl, aarch64_neon_sshl, Add1ArgType | UnsignedAlts), 2756 NEONMAP0(vshll_n_v), 2757 NEONMAP0(vshlq_n_v), 2758 NEONMAP2(vshlq_v, aarch64_neon_ushl, aarch64_neon_sshl, Add1ArgType | UnsignedAlts), 2759 NEONMAP0(vshr_n_v), 2760 NEONMAP0(vshrn_n_v), 2761 NEONMAP0(vshrq_n_v), 2762 NEONMAP0(vsubhn_v), 2763 NEONMAP0(vtst_v), 2764 NEONMAP0(vtstq_v), 2765 }; 2766 2767 static const NeonIntrinsicInfo AArch64SISDIntrinsicMap[] = { 2768 NEONMAP1(vabdd_f64, aarch64_sisd_fabd, Add1ArgType), 2769 NEONMAP1(vabds_f32, aarch64_sisd_fabd, Add1ArgType), 2770 NEONMAP1(vabsd_s64, aarch64_neon_abs, Add1ArgType), 2771 NEONMAP1(vaddlv_s32, aarch64_neon_saddlv, AddRetType | Add1ArgType), 2772 NEONMAP1(vaddlv_u32, aarch64_neon_uaddlv, AddRetType | Add1ArgType), 2773 NEONMAP1(vaddlvq_s32, aarch64_neon_saddlv, AddRetType | Add1ArgType), 2774 NEONMAP1(vaddlvq_u32, aarch64_neon_uaddlv, AddRetType | Add1ArgType), 2775 NEONMAP1(vaddv_f32, aarch64_neon_faddv, AddRetType | Add1ArgType), 2776 NEONMAP1(vaddv_s32, aarch64_neon_saddv, AddRetType | Add1ArgType), 2777 NEONMAP1(vaddv_u32, aarch64_neon_uaddv, AddRetType | Add1ArgType), 2778 NEONMAP1(vaddvq_f32, aarch64_neon_faddv, AddRetType | Add1ArgType), 2779 NEONMAP1(vaddvq_f64, aarch64_neon_faddv, AddRetType | Add1ArgType), 2780 NEONMAP1(vaddvq_s32, aarch64_neon_saddv, AddRetType | Add1ArgType), 2781 NEONMAP1(vaddvq_s64, aarch64_neon_saddv, AddRetType | Add1ArgType), 2782 NEONMAP1(vaddvq_u32, aarch64_neon_uaddv, AddRetType | Add1ArgType), 2783 NEONMAP1(vaddvq_u64, aarch64_neon_uaddv, AddRetType | Add1ArgType), 2784 NEONMAP1(vcaged_f64, aarch64_neon_facge, AddRetType | Add1ArgType), 2785 NEONMAP1(vcages_f32, aarch64_neon_facge, AddRetType | Add1ArgType), 2786 NEONMAP1(vcagtd_f64, aarch64_neon_facgt, AddRetType | Add1ArgType), 2787 NEONMAP1(vcagts_f32, aarch64_neon_facgt, AddRetType | Add1ArgType), 2788 NEONMAP1(vcaled_f64, aarch64_neon_facge, AddRetType | Add1ArgType), 2789 NEONMAP1(vcales_f32, aarch64_neon_facge, AddRetType | Add1ArgType), 2790 NEONMAP1(vcaltd_f64, aarch64_neon_facgt, AddRetType | Add1ArgType), 2791 NEONMAP1(vcalts_f32, aarch64_neon_facgt, AddRetType | Add1ArgType), 2792 NEONMAP1(vcvtad_s64_f64, aarch64_neon_fcvtas, AddRetType | Add1ArgType), 2793 NEONMAP1(vcvtad_u64_f64, aarch64_neon_fcvtau, AddRetType | Add1ArgType), 2794 NEONMAP1(vcvtas_s32_f32, aarch64_neon_fcvtas, AddRetType | Add1ArgType), 2795 NEONMAP1(vcvtas_u32_f32, aarch64_neon_fcvtau, AddRetType | Add1ArgType), 2796 NEONMAP1(vcvtd_n_f64_s64, aarch64_neon_vcvtfxs2fp, AddRetType | Add1ArgType), 2797 NEONMAP1(vcvtd_n_f64_u64, aarch64_neon_vcvtfxu2fp, AddRetType | Add1ArgType), 2798 NEONMAP1(vcvtd_n_s64_f64, aarch64_neon_vcvtfp2fxs, AddRetType | Add1ArgType), 2799 NEONMAP1(vcvtd_n_u64_f64, aarch64_neon_vcvtfp2fxu, AddRetType | Add1ArgType), 2800 NEONMAP1(vcvtmd_s64_f64, aarch64_neon_fcvtms, AddRetType | Add1ArgType), 2801 NEONMAP1(vcvtmd_u64_f64, aarch64_neon_fcvtmu, AddRetType | Add1ArgType), 2802 NEONMAP1(vcvtms_s32_f32, aarch64_neon_fcvtms, AddRetType | Add1ArgType), 2803 NEONMAP1(vcvtms_u32_f32, aarch64_neon_fcvtmu, AddRetType | Add1ArgType), 2804 NEONMAP1(vcvtnd_s64_f64, aarch64_neon_fcvtns, AddRetType | Add1ArgType), 2805 NEONMAP1(vcvtnd_u64_f64, aarch64_neon_fcvtnu, AddRetType | Add1ArgType), 2806 NEONMAP1(vcvtns_s32_f32, aarch64_neon_fcvtns, AddRetType | Add1ArgType), 2807 NEONMAP1(vcvtns_u32_f32, aarch64_neon_fcvtnu, AddRetType | Add1ArgType), 2808 NEONMAP1(vcvtpd_s64_f64, aarch64_neon_fcvtps, AddRetType | Add1ArgType), 2809 NEONMAP1(vcvtpd_u64_f64, aarch64_neon_fcvtpu, AddRetType | Add1ArgType), 2810 NEONMAP1(vcvtps_s32_f32, aarch64_neon_fcvtps, AddRetType | Add1ArgType), 2811 NEONMAP1(vcvtps_u32_f32, aarch64_neon_fcvtpu, AddRetType | Add1ArgType), 2812 NEONMAP1(vcvts_n_f32_s32, aarch64_neon_vcvtfxs2fp, AddRetType | Add1ArgType), 2813 NEONMAP1(vcvts_n_f32_u32, aarch64_neon_vcvtfxu2fp, AddRetType | Add1ArgType), 2814 NEONMAP1(vcvts_n_s32_f32, aarch64_neon_vcvtfp2fxs, AddRetType | Add1ArgType), 2815 NEONMAP1(vcvts_n_u32_f32, aarch64_neon_vcvtfp2fxu, AddRetType | Add1ArgType), 2816 NEONMAP1(vcvtxd_f32_f64, aarch64_sisd_fcvtxn, 0), 2817 NEONMAP1(vmaxnmv_f32, aarch64_neon_fmaxnmv, AddRetType | Add1ArgType), 2818 NEONMAP1(vmaxnmvq_f32, aarch64_neon_fmaxnmv, AddRetType | Add1ArgType), 2819 NEONMAP1(vmaxnmvq_f64, aarch64_neon_fmaxnmv, AddRetType | Add1ArgType), 2820 NEONMAP1(vmaxv_f32, aarch64_neon_fmaxv, AddRetType | Add1ArgType), 2821 NEONMAP1(vmaxv_s32, aarch64_neon_smaxv, AddRetType | Add1ArgType), 2822 NEONMAP1(vmaxv_u32, aarch64_neon_umaxv, AddRetType | Add1ArgType), 2823 NEONMAP1(vmaxvq_f32, aarch64_neon_fmaxv, AddRetType | Add1ArgType), 2824 NEONMAP1(vmaxvq_f64, aarch64_neon_fmaxv, AddRetType | Add1ArgType), 2825 NEONMAP1(vmaxvq_s32, aarch64_neon_smaxv, AddRetType | Add1ArgType), 2826 NEONMAP1(vmaxvq_u32, aarch64_neon_umaxv, AddRetType | Add1ArgType), 2827 NEONMAP1(vminnmv_f32, aarch64_neon_fminnmv, AddRetType | Add1ArgType), 2828 NEONMAP1(vminnmvq_f32, aarch64_neon_fminnmv, AddRetType | Add1ArgType), 2829 NEONMAP1(vminnmvq_f64, aarch64_neon_fminnmv, AddRetType | Add1ArgType), 2830 NEONMAP1(vminv_f32, aarch64_neon_fminv, AddRetType | Add1ArgType), 2831 NEONMAP1(vminv_s32, aarch64_neon_sminv, AddRetType | Add1ArgType), 2832 NEONMAP1(vminv_u32, aarch64_neon_uminv, AddRetType | Add1ArgType), 2833 NEONMAP1(vminvq_f32, aarch64_neon_fminv, AddRetType | Add1ArgType), 2834 NEONMAP1(vminvq_f64, aarch64_neon_fminv, AddRetType | Add1ArgType), 2835 NEONMAP1(vminvq_s32, aarch64_neon_sminv, AddRetType | Add1ArgType), 2836 NEONMAP1(vminvq_u32, aarch64_neon_uminv, AddRetType | Add1ArgType), 2837 NEONMAP1(vmull_p64, aarch64_neon_pmull64, 0), 2838 NEONMAP1(vmulxd_f64, aarch64_neon_fmulx, Add1ArgType), 2839 NEONMAP1(vmulxs_f32, aarch64_neon_fmulx, Add1ArgType), 2840 NEONMAP1(vpaddd_s64, aarch64_neon_uaddv, AddRetType | Add1ArgType), 2841 NEONMAP1(vpaddd_u64, aarch64_neon_uaddv, AddRetType | Add1ArgType), 2842 NEONMAP1(vpmaxnmqd_f64, aarch64_neon_fmaxnmv, AddRetType | Add1ArgType), 2843 NEONMAP1(vpmaxnms_f32, aarch64_neon_fmaxnmv, AddRetType | Add1ArgType), 2844 NEONMAP1(vpmaxqd_f64, aarch64_neon_fmaxv, AddRetType | Add1ArgType), 2845 NEONMAP1(vpmaxs_f32, aarch64_neon_fmaxv, AddRetType | Add1ArgType), 2846 NEONMAP1(vpminnmqd_f64, aarch64_neon_fminnmv, AddRetType | Add1ArgType), 2847 NEONMAP1(vpminnms_f32, aarch64_neon_fminnmv, AddRetType | Add1ArgType), 2848 NEONMAP1(vpminqd_f64, aarch64_neon_fminv, AddRetType | Add1ArgType), 2849 NEONMAP1(vpmins_f32, aarch64_neon_fminv, AddRetType | Add1ArgType), 2850 NEONMAP1(vqabsb_s8, aarch64_neon_sqabs, Vectorize1ArgType | Use64BitVectors), 2851 NEONMAP1(vqabsd_s64, aarch64_neon_sqabs, Add1ArgType), 2852 NEONMAP1(vqabsh_s16, aarch64_neon_sqabs, Vectorize1ArgType | Use64BitVectors), 2853 NEONMAP1(vqabss_s32, aarch64_neon_sqabs, Add1ArgType), 2854 NEONMAP1(vqaddb_s8, aarch64_neon_sqadd, Vectorize1ArgType | Use64BitVectors), 2855 NEONMAP1(vqaddb_u8, aarch64_neon_uqadd, Vectorize1ArgType | Use64BitVectors), 2856 NEONMAP1(vqaddd_s64, aarch64_neon_sqadd, Add1ArgType), 2857 NEONMAP1(vqaddd_u64, aarch64_neon_uqadd, Add1ArgType), 2858 NEONMAP1(vqaddh_s16, aarch64_neon_sqadd, Vectorize1ArgType | Use64BitVectors), 2859 NEONMAP1(vqaddh_u16, aarch64_neon_uqadd, Vectorize1ArgType | Use64BitVectors), 2860 NEONMAP1(vqadds_s32, aarch64_neon_sqadd, Add1ArgType), 2861 NEONMAP1(vqadds_u32, aarch64_neon_uqadd, Add1ArgType), 2862 NEONMAP1(vqdmulhh_s16, aarch64_neon_sqdmulh, Vectorize1ArgType | Use64BitVectors), 2863 NEONMAP1(vqdmulhs_s32, aarch64_neon_sqdmulh, Add1ArgType), 2864 NEONMAP1(vqdmullh_s16, aarch64_neon_sqdmull, VectorRet | Use128BitVectors), 2865 NEONMAP1(vqdmulls_s32, aarch64_neon_sqdmulls_scalar, 0), 2866 NEONMAP1(vqmovnd_s64, aarch64_neon_scalar_sqxtn, AddRetType | Add1ArgType), 2867 NEONMAP1(vqmovnd_u64, aarch64_neon_scalar_uqxtn, AddRetType | Add1ArgType), 2868 NEONMAP1(vqmovnh_s16, aarch64_neon_sqxtn, VectorRet | Use64BitVectors), 2869 NEONMAP1(vqmovnh_u16, aarch64_neon_uqxtn, VectorRet | Use64BitVectors), 2870 NEONMAP1(vqmovns_s32, aarch64_neon_sqxtn, VectorRet | Use64BitVectors), 2871 NEONMAP1(vqmovns_u32, aarch64_neon_uqxtn, VectorRet | Use64BitVectors), 2872 NEONMAP1(vqmovund_s64, aarch64_neon_scalar_sqxtun, AddRetType | Add1ArgType), 2873 NEONMAP1(vqmovunh_s16, aarch64_neon_sqxtun, VectorRet | Use64BitVectors), 2874 NEONMAP1(vqmovuns_s32, aarch64_neon_sqxtun, VectorRet | Use64BitVectors), 2875 NEONMAP1(vqnegb_s8, aarch64_neon_sqneg, Vectorize1ArgType | Use64BitVectors), 2876 NEONMAP1(vqnegd_s64, aarch64_neon_sqneg, Add1ArgType), 2877 NEONMAP1(vqnegh_s16, aarch64_neon_sqneg, Vectorize1ArgType | Use64BitVectors), 2878 NEONMAP1(vqnegs_s32, aarch64_neon_sqneg, Add1ArgType), 2879 NEONMAP1(vqrdmulhh_s16, aarch64_neon_sqrdmulh, Vectorize1ArgType | Use64BitVectors), 2880 NEONMAP1(vqrdmulhs_s32, aarch64_neon_sqrdmulh, Add1ArgType), 2881 NEONMAP1(vqrshlb_s8, aarch64_neon_sqrshl, Vectorize1ArgType | Use64BitVectors), 2882 NEONMAP1(vqrshlb_u8, aarch64_neon_uqrshl, Vectorize1ArgType | Use64BitVectors), 2883 NEONMAP1(vqrshld_s64, aarch64_neon_sqrshl, Add1ArgType), 2884 NEONMAP1(vqrshld_u64, aarch64_neon_uqrshl, Add1ArgType), 2885 NEONMAP1(vqrshlh_s16, aarch64_neon_sqrshl, Vectorize1ArgType | Use64BitVectors), 2886 NEONMAP1(vqrshlh_u16, aarch64_neon_uqrshl, Vectorize1ArgType | Use64BitVectors), 2887 NEONMAP1(vqrshls_s32, aarch64_neon_sqrshl, Add1ArgType), 2888 NEONMAP1(vqrshls_u32, aarch64_neon_uqrshl, Add1ArgType), 2889 NEONMAP1(vqrshrnd_n_s64, aarch64_neon_sqrshrn, AddRetType), 2890 NEONMAP1(vqrshrnd_n_u64, aarch64_neon_uqrshrn, AddRetType), 2891 NEONMAP1(vqrshrnh_n_s16, aarch64_neon_sqrshrn, VectorRet | Use64BitVectors), 2892 NEONMAP1(vqrshrnh_n_u16, aarch64_neon_uqrshrn, VectorRet | Use64BitVectors), 2893 NEONMAP1(vqrshrns_n_s32, aarch64_neon_sqrshrn, VectorRet | Use64BitVectors), 2894 NEONMAP1(vqrshrns_n_u32, aarch64_neon_uqrshrn, VectorRet | Use64BitVectors), 2895 NEONMAP1(vqrshrund_n_s64, aarch64_neon_sqrshrun, AddRetType), 2896 NEONMAP1(vqrshrunh_n_s16, aarch64_neon_sqrshrun, VectorRet | Use64BitVectors), 2897 NEONMAP1(vqrshruns_n_s32, aarch64_neon_sqrshrun, VectorRet | Use64BitVectors), 2898 NEONMAP1(vqshlb_n_s8, aarch64_neon_sqshl, Vectorize1ArgType | Use64BitVectors), 2899 NEONMAP1(vqshlb_n_u8, aarch64_neon_uqshl, Vectorize1ArgType | Use64BitVectors), 2900 NEONMAP1(vqshlb_s8, aarch64_neon_sqshl, Vectorize1ArgType | Use64BitVectors), 2901 NEONMAP1(vqshlb_u8, aarch64_neon_uqshl, Vectorize1ArgType | Use64BitVectors), 2902 NEONMAP1(vqshld_s64, aarch64_neon_sqshl, Add1ArgType), 2903 NEONMAP1(vqshld_u64, aarch64_neon_uqshl, Add1ArgType), 2904 NEONMAP1(vqshlh_n_s16, aarch64_neon_sqshl, Vectorize1ArgType | Use64BitVectors), 2905 NEONMAP1(vqshlh_n_u16, aarch64_neon_uqshl, Vectorize1ArgType | Use64BitVectors), 2906 NEONMAP1(vqshlh_s16, aarch64_neon_sqshl, Vectorize1ArgType | Use64BitVectors), 2907 NEONMAP1(vqshlh_u16, aarch64_neon_uqshl, Vectorize1ArgType | Use64BitVectors), 2908 NEONMAP1(vqshls_n_s32, aarch64_neon_sqshl, Add1ArgType), 2909 NEONMAP1(vqshls_n_u32, aarch64_neon_uqshl, Add1ArgType), 2910 NEONMAP1(vqshls_s32, aarch64_neon_sqshl, Add1ArgType), 2911 NEONMAP1(vqshls_u32, aarch64_neon_uqshl, Add1ArgType), 2912 NEONMAP1(vqshlub_n_s8, aarch64_neon_sqshlu, Vectorize1ArgType | Use64BitVectors), 2913 NEONMAP1(vqshluh_n_s16, aarch64_neon_sqshlu, Vectorize1ArgType | Use64BitVectors), 2914 NEONMAP1(vqshlus_n_s32, aarch64_neon_sqshlu, Add1ArgType), 2915 NEONMAP1(vqshrnd_n_s64, aarch64_neon_sqshrn, AddRetType), 2916 NEONMAP1(vqshrnd_n_u64, aarch64_neon_uqshrn, AddRetType), 2917 NEONMAP1(vqshrnh_n_s16, aarch64_neon_sqshrn, VectorRet | Use64BitVectors), 2918 NEONMAP1(vqshrnh_n_u16, aarch64_neon_uqshrn, VectorRet | Use64BitVectors), 2919 NEONMAP1(vqshrns_n_s32, aarch64_neon_sqshrn, VectorRet | Use64BitVectors), 2920 NEONMAP1(vqshrns_n_u32, aarch64_neon_uqshrn, VectorRet | Use64BitVectors), 2921 NEONMAP1(vqshrund_n_s64, aarch64_neon_sqshrun, AddRetType), 2922 NEONMAP1(vqshrunh_n_s16, aarch64_neon_sqshrun, VectorRet | Use64BitVectors), 2923 NEONMAP1(vqshruns_n_s32, aarch64_neon_sqshrun, VectorRet | Use64BitVectors), 2924 NEONMAP1(vqsubb_s8, aarch64_neon_sqsub, Vectorize1ArgType | Use64BitVectors), 2925 NEONMAP1(vqsubb_u8, aarch64_neon_uqsub, Vectorize1ArgType | Use64BitVectors), 2926 NEONMAP1(vqsubd_s64, aarch64_neon_sqsub, Add1ArgType), 2927 NEONMAP1(vqsubd_u64, aarch64_neon_uqsub, Add1ArgType), 2928 NEONMAP1(vqsubh_s16, aarch64_neon_sqsub, Vectorize1ArgType | Use64BitVectors), 2929 NEONMAP1(vqsubh_u16, aarch64_neon_uqsub, Vectorize1ArgType | Use64BitVectors), 2930 NEONMAP1(vqsubs_s32, aarch64_neon_sqsub, Add1ArgType), 2931 NEONMAP1(vqsubs_u32, aarch64_neon_uqsub, Add1ArgType), 2932 NEONMAP1(vrecped_f64, aarch64_neon_frecpe, Add1ArgType), 2933 NEONMAP1(vrecpes_f32, aarch64_neon_frecpe, Add1ArgType), 2934 NEONMAP1(vrecpxd_f64, aarch64_neon_frecpx, Add1ArgType), 2935 NEONMAP1(vrecpxs_f32, aarch64_neon_frecpx, Add1ArgType), 2936 NEONMAP1(vrshld_s64, aarch64_neon_srshl, Add1ArgType), 2937 NEONMAP1(vrshld_u64, aarch64_neon_urshl, Add1ArgType), 2938 NEONMAP1(vrsqrted_f64, aarch64_neon_frsqrte, Add1ArgType), 2939 NEONMAP1(vrsqrtes_f32, aarch64_neon_frsqrte, Add1ArgType), 2940 NEONMAP1(vrsqrtsd_f64, aarch64_neon_frsqrts, Add1ArgType), 2941 NEONMAP1(vrsqrtss_f32, aarch64_neon_frsqrts, Add1ArgType), 2942 NEONMAP1(vsha1cq_u32, aarch64_crypto_sha1c, 0), 2943 NEONMAP1(vsha1h_u32, aarch64_crypto_sha1h, 0), 2944 NEONMAP1(vsha1mq_u32, aarch64_crypto_sha1m, 0), 2945 NEONMAP1(vsha1pq_u32, aarch64_crypto_sha1p, 0), 2946 NEONMAP1(vshld_s64, aarch64_neon_sshl, Add1ArgType), 2947 NEONMAP1(vshld_u64, aarch64_neon_ushl, Add1ArgType), 2948 NEONMAP1(vslid_n_s64, aarch64_neon_vsli, Vectorize1ArgType), 2949 NEONMAP1(vslid_n_u64, aarch64_neon_vsli, Vectorize1ArgType), 2950 NEONMAP1(vsqaddb_u8, aarch64_neon_usqadd, Vectorize1ArgType | Use64BitVectors), 2951 NEONMAP1(vsqaddd_u64, aarch64_neon_usqadd, Add1ArgType), 2952 NEONMAP1(vsqaddh_u16, aarch64_neon_usqadd, Vectorize1ArgType | Use64BitVectors), 2953 NEONMAP1(vsqadds_u32, aarch64_neon_usqadd, Add1ArgType), 2954 NEONMAP1(vsrid_n_s64, aarch64_neon_vsri, Vectorize1ArgType), 2955 NEONMAP1(vsrid_n_u64, aarch64_neon_vsri, Vectorize1ArgType), 2956 NEONMAP1(vuqaddb_s8, aarch64_neon_suqadd, Vectorize1ArgType | Use64BitVectors), 2957 NEONMAP1(vuqaddd_s64, aarch64_neon_suqadd, Add1ArgType), 2958 NEONMAP1(vuqaddh_s16, aarch64_neon_suqadd, Vectorize1ArgType | Use64BitVectors), 2959 NEONMAP1(vuqadds_s32, aarch64_neon_suqadd, Add1ArgType), 2960 }; 2961 2962 #undef NEONMAP0 2963 #undef NEONMAP1 2964 #undef NEONMAP2 2965 2966 static bool NEONSIMDIntrinsicsProvenSorted = false; 2967 2968 static bool AArch64SIMDIntrinsicsProvenSorted = false; 2969 static bool AArch64SISDIntrinsicsProvenSorted = false; 2970 2971 2972 static const NeonIntrinsicInfo * 2973 findNeonIntrinsicInMap(ArrayRef<NeonIntrinsicInfo> IntrinsicMap, 2974 unsigned BuiltinID, bool &MapProvenSorted) { 2975 2976 #ifndef NDEBUG 2977 if (!MapProvenSorted) { 2978 assert(std::is_sorted(std::begin(IntrinsicMap), std::end(IntrinsicMap))); 2979 MapProvenSorted = true; 2980 } 2981 #endif 2982 2983 const NeonIntrinsicInfo *Builtin = 2984 std::lower_bound(IntrinsicMap.begin(), IntrinsicMap.end(), BuiltinID); 2985 2986 if (Builtin != IntrinsicMap.end() && Builtin->BuiltinID == BuiltinID) 2987 return Builtin; 2988 2989 return nullptr; 2990 } 2991 2992 Function *CodeGenFunction::LookupNeonLLVMIntrinsic(unsigned IntrinsicID, 2993 unsigned Modifier, 2994 llvm::Type *ArgType, 2995 const CallExpr *E) { 2996 int VectorSize = 0; 2997 if (Modifier & Use64BitVectors) 2998 VectorSize = 64; 2999 else if (Modifier & Use128BitVectors) 3000 VectorSize = 128; 3001 3002 // Return type. 3003 SmallVector<llvm::Type *, 3> Tys; 3004 if (Modifier & AddRetType) { 3005 llvm::Type *Ty = ConvertType(E->getCallReturnType(getContext())); 3006 if (Modifier & VectorizeRetType) 3007 Ty = llvm::VectorType::get( 3008 Ty, VectorSize ? VectorSize / Ty->getPrimitiveSizeInBits() : 1); 3009 3010 Tys.push_back(Ty); 3011 } 3012 3013 // Arguments. 3014 if (Modifier & VectorizeArgTypes) { 3015 int Elts = VectorSize ? VectorSize / ArgType->getPrimitiveSizeInBits() : 1; 3016 ArgType = llvm::VectorType::get(ArgType, Elts); 3017 } 3018 3019 if (Modifier & (Add1ArgType | Add2ArgTypes)) 3020 Tys.push_back(ArgType); 3021 3022 if (Modifier & Add2ArgTypes) 3023 Tys.push_back(ArgType); 3024 3025 if (Modifier & InventFloatType) 3026 Tys.push_back(FloatTy); 3027 3028 return CGM.getIntrinsic(IntrinsicID, Tys); 3029 } 3030 3031 static Value *EmitCommonNeonSISDBuiltinExpr(CodeGenFunction &CGF, 3032 const NeonIntrinsicInfo &SISDInfo, 3033 SmallVectorImpl<Value *> &Ops, 3034 const CallExpr *E) { 3035 unsigned BuiltinID = SISDInfo.BuiltinID; 3036 unsigned int Int = SISDInfo.LLVMIntrinsic; 3037 unsigned Modifier = SISDInfo.TypeModifier; 3038 const char *s = SISDInfo.NameHint; 3039 3040 switch (BuiltinID) { 3041 case NEON::BI__builtin_neon_vcled_s64: 3042 case NEON::BI__builtin_neon_vcled_u64: 3043 case NEON::BI__builtin_neon_vcles_f32: 3044 case NEON::BI__builtin_neon_vcled_f64: 3045 case NEON::BI__builtin_neon_vcltd_s64: 3046 case NEON::BI__builtin_neon_vcltd_u64: 3047 case NEON::BI__builtin_neon_vclts_f32: 3048 case NEON::BI__builtin_neon_vcltd_f64: 3049 case NEON::BI__builtin_neon_vcales_f32: 3050 case NEON::BI__builtin_neon_vcaled_f64: 3051 case NEON::BI__builtin_neon_vcalts_f32: 3052 case NEON::BI__builtin_neon_vcaltd_f64: 3053 // Only one direction of comparisons actually exist, cmle is actually a cmge 3054 // with swapped operands. The table gives us the right intrinsic but we 3055 // still need to do the swap. 3056 std::swap(Ops[0], Ops[1]); 3057 break; 3058 } 3059 3060 assert(Int && "Generic code assumes a valid intrinsic"); 3061 3062 // Determine the type(s) of this overloaded AArch64 intrinsic. 3063 const Expr *Arg = E->getArg(0); 3064 llvm::Type *ArgTy = CGF.ConvertType(Arg->getType()); 3065 Function *F = CGF.LookupNeonLLVMIntrinsic(Int, Modifier, ArgTy, E); 3066 3067 int j = 0; 3068 ConstantInt *C0 = ConstantInt::get(CGF.SizeTy, 0); 3069 for (Function::const_arg_iterator ai = F->arg_begin(), ae = F->arg_end(); 3070 ai != ae; ++ai, ++j) { 3071 llvm::Type *ArgTy = ai->getType(); 3072 if (Ops[j]->getType()->getPrimitiveSizeInBits() == 3073 ArgTy->getPrimitiveSizeInBits()) 3074 continue; 3075 3076 assert(ArgTy->isVectorTy() && !Ops[j]->getType()->isVectorTy()); 3077 // The constant argument to an _n_ intrinsic always has Int32Ty, so truncate 3078 // it before inserting. 3079 Ops[j] = 3080 CGF.Builder.CreateTruncOrBitCast(Ops[j], ArgTy->getVectorElementType()); 3081 Ops[j] = 3082 CGF.Builder.CreateInsertElement(UndefValue::get(ArgTy), Ops[j], C0); 3083 } 3084 3085 Value *Result = CGF.EmitNeonCall(F, Ops, s); 3086 llvm::Type *ResultType = CGF.ConvertType(E->getType()); 3087 if (ResultType->getPrimitiveSizeInBits() < 3088 Result->getType()->getPrimitiveSizeInBits()) 3089 return CGF.Builder.CreateExtractElement(Result, C0); 3090 3091 return CGF.Builder.CreateBitCast(Result, ResultType, s); 3092 } 3093 3094 Value *CodeGenFunction::EmitCommonNeonBuiltinExpr( 3095 unsigned BuiltinID, unsigned LLVMIntrinsic, unsigned AltLLVMIntrinsic, 3096 const char *NameHint, unsigned Modifier, const CallExpr *E, 3097 SmallVectorImpl<llvm::Value *> &Ops, Address PtrOp0, Address PtrOp1) { 3098 // Get the last argument, which specifies the vector type. 3099 llvm::APSInt NeonTypeConst; 3100 const Expr *Arg = E->getArg(E->getNumArgs() - 1); 3101 if (!Arg->isIntegerConstantExpr(NeonTypeConst, getContext())) 3102 return nullptr; 3103 3104 // Determine the type of this overloaded NEON intrinsic. 3105 NeonTypeFlags Type(NeonTypeConst.getZExtValue()); 3106 bool Usgn = Type.isUnsigned(); 3107 bool Quad = Type.isQuad(); 3108 3109 llvm::VectorType *VTy = GetNeonType(this, Type); 3110 llvm::Type *Ty = VTy; 3111 if (!Ty) 3112 return nullptr; 3113 3114 auto getAlignmentValue32 = [&](Address addr) -> Value* { 3115 return Builder.getInt32(addr.getAlignment().getQuantity()); 3116 }; 3117 3118 unsigned Int = LLVMIntrinsic; 3119 if ((Modifier & UnsignedAlts) && !Usgn) 3120 Int = AltLLVMIntrinsic; 3121 3122 switch (BuiltinID) { 3123 default: break; 3124 case NEON::BI__builtin_neon_vabs_v: 3125 case NEON::BI__builtin_neon_vabsq_v: 3126 if (VTy->getElementType()->isFloatingPointTy()) 3127 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::fabs, Ty), Ops, "vabs"); 3128 return EmitNeonCall(CGM.getIntrinsic(LLVMIntrinsic, Ty), Ops, "vabs"); 3129 case NEON::BI__builtin_neon_vaddhn_v: { 3130 llvm::VectorType *SrcTy = 3131 llvm::VectorType::getExtendedElementVectorType(VTy); 3132 3133 // %sum = add <4 x i32> %lhs, %rhs 3134 Ops[0] = Builder.CreateBitCast(Ops[0], SrcTy); 3135 Ops[1] = Builder.CreateBitCast(Ops[1], SrcTy); 3136 Ops[0] = Builder.CreateAdd(Ops[0], Ops[1], "vaddhn"); 3137 3138 // %high = lshr <4 x i32> %sum, <i32 16, i32 16, i32 16, i32 16> 3139 Constant *ShiftAmt = 3140 ConstantInt::get(SrcTy, SrcTy->getScalarSizeInBits() / 2); 3141 Ops[0] = Builder.CreateLShr(Ops[0], ShiftAmt, "vaddhn"); 3142 3143 // %res = trunc <4 x i32> %high to <4 x i16> 3144 return Builder.CreateTrunc(Ops[0], VTy, "vaddhn"); 3145 } 3146 case NEON::BI__builtin_neon_vcale_v: 3147 case NEON::BI__builtin_neon_vcaleq_v: 3148 case NEON::BI__builtin_neon_vcalt_v: 3149 case NEON::BI__builtin_neon_vcaltq_v: 3150 std::swap(Ops[0], Ops[1]); 3151 case NEON::BI__builtin_neon_vcage_v: 3152 case NEON::BI__builtin_neon_vcageq_v: 3153 case NEON::BI__builtin_neon_vcagt_v: 3154 case NEON::BI__builtin_neon_vcagtq_v: { 3155 llvm::Type *VecFlt = llvm::VectorType::get( 3156 VTy->getScalarSizeInBits() == 32 ? FloatTy : DoubleTy, 3157 VTy->getNumElements()); 3158 llvm::Type *Tys[] = { VTy, VecFlt }; 3159 Function *F = CGM.getIntrinsic(LLVMIntrinsic, Tys); 3160 return EmitNeonCall(F, Ops, NameHint); 3161 } 3162 case NEON::BI__builtin_neon_vclz_v: 3163 case NEON::BI__builtin_neon_vclzq_v: 3164 // We generate target-independent intrinsic, which needs a second argument 3165 // for whether or not clz of zero is undefined; on ARM it isn't. 3166 Ops.push_back(Builder.getInt1(getTarget().isCLZForZeroUndef())); 3167 break; 3168 case NEON::BI__builtin_neon_vcvt_f32_v: 3169 case NEON::BI__builtin_neon_vcvtq_f32_v: 3170 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 3171 Ty = GetNeonType(this, NeonTypeFlags(NeonTypeFlags::Float32, false, Quad)); 3172 return Usgn ? Builder.CreateUIToFP(Ops[0], Ty, "vcvt") 3173 : Builder.CreateSIToFP(Ops[0], Ty, "vcvt"); 3174 case NEON::BI__builtin_neon_vcvt_n_f32_v: 3175 case NEON::BI__builtin_neon_vcvt_n_f64_v: 3176 case NEON::BI__builtin_neon_vcvtq_n_f32_v: 3177 case NEON::BI__builtin_neon_vcvtq_n_f64_v: { 3178 llvm::Type *Tys[2] = { GetFloatNeonType(this, Type), Ty }; 3179 Int = Usgn ? LLVMIntrinsic : AltLLVMIntrinsic; 3180 Function *F = CGM.getIntrinsic(Int, Tys); 3181 return EmitNeonCall(F, Ops, "vcvt_n"); 3182 } 3183 case NEON::BI__builtin_neon_vcvt_n_s32_v: 3184 case NEON::BI__builtin_neon_vcvt_n_u32_v: 3185 case NEON::BI__builtin_neon_vcvt_n_s64_v: 3186 case NEON::BI__builtin_neon_vcvt_n_u64_v: 3187 case NEON::BI__builtin_neon_vcvtq_n_s32_v: 3188 case NEON::BI__builtin_neon_vcvtq_n_u32_v: 3189 case NEON::BI__builtin_neon_vcvtq_n_s64_v: 3190 case NEON::BI__builtin_neon_vcvtq_n_u64_v: { 3191 llvm::Type *Tys[2] = { Ty, GetFloatNeonType(this, Type) }; 3192 Function *F = CGM.getIntrinsic(LLVMIntrinsic, Tys); 3193 return EmitNeonCall(F, Ops, "vcvt_n"); 3194 } 3195 case NEON::BI__builtin_neon_vcvt_s32_v: 3196 case NEON::BI__builtin_neon_vcvt_u32_v: 3197 case NEON::BI__builtin_neon_vcvt_s64_v: 3198 case NEON::BI__builtin_neon_vcvt_u64_v: 3199 case NEON::BI__builtin_neon_vcvtq_s32_v: 3200 case NEON::BI__builtin_neon_vcvtq_u32_v: 3201 case NEON::BI__builtin_neon_vcvtq_s64_v: 3202 case NEON::BI__builtin_neon_vcvtq_u64_v: { 3203 Ops[0] = Builder.CreateBitCast(Ops[0], GetFloatNeonType(this, Type)); 3204 return Usgn ? Builder.CreateFPToUI(Ops[0], Ty, "vcvt") 3205 : Builder.CreateFPToSI(Ops[0], Ty, "vcvt"); 3206 } 3207 case NEON::BI__builtin_neon_vcvta_s32_v: 3208 case NEON::BI__builtin_neon_vcvta_s64_v: 3209 case NEON::BI__builtin_neon_vcvta_u32_v: 3210 case NEON::BI__builtin_neon_vcvta_u64_v: 3211 case NEON::BI__builtin_neon_vcvtaq_s32_v: 3212 case NEON::BI__builtin_neon_vcvtaq_s64_v: 3213 case NEON::BI__builtin_neon_vcvtaq_u32_v: 3214 case NEON::BI__builtin_neon_vcvtaq_u64_v: 3215 case NEON::BI__builtin_neon_vcvtn_s32_v: 3216 case NEON::BI__builtin_neon_vcvtn_s64_v: 3217 case NEON::BI__builtin_neon_vcvtn_u32_v: 3218 case NEON::BI__builtin_neon_vcvtn_u64_v: 3219 case NEON::BI__builtin_neon_vcvtnq_s32_v: 3220 case NEON::BI__builtin_neon_vcvtnq_s64_v: 3221 case NEON::BI__builtin_neon_vcvtnq_u32_v: 3222 case NEON::BI__builtin_neon_vcvtnq_u64_v: 3223 case NEON::BI__builtin_neon_vcvtp_s32_v: 3224 case NEON::BI__builtin_neon_vcvtp_s64_v: 3225 case NEON::BI__builtin_neon_vcvtp_u32_v: 3226 case NEON::BI__builtin_neon_vcvtp_u64_v: 3227 case NEON::BI__builtin_neon_vcvtpq_s32_v: 3228 case NEON::BI__builtin_neon_vcvtpq_s64_v: 3229 case NEON::BI__builtin_neon_vcvtpq_u32_v: 3230 case NEON::BI__builtin_neon_vcvtpq_u64_v: 3231 case NEON::BI__builtin_neon_vcvtm_s32_v: 3232 case NEON::BI__builtin_neon_vcvtm_s64_v: 3233 case NEON::BI__builtin_neon_vcvtm_u32_v: 3234 case NEON::BI__builtin_neon_vcvtm_u64_v: 3235 case NEON::BI__builtin_neon_vcvtmq_s32_v: 3236 case NEON::BI__builtin_neon_vcvtmq_s64_v: 3237 case NEON::BI__builtin_neon_vcvtmq_u32_v: 3238 case NEON::BI__builtin_neon_vcvtmq_u64_v: { 3239 llvm::Type *Tys[2] = { Ty, GetFloatNeonType(this, Type) }; 3240 return EmitNeonCall(CGM.getIntrinsic(LLVMIntrinsic, Tys), Ops, NameHint); 3241 } 3242 case NEON::BI__builtin_neon_vext_v: 3243 case NEON::BI__builtin_neon_vextq_v: { 3244 int CV = cast<ConstantInt>(Ops[2])->getSExtValue(); 3245 SmallVector<Constant*, 16> Indices; 3246 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) 3247 Indices.push_back(ConstantInt::get(Int32Ty, i+CV)); 3248 3249 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 3250 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 3251 Value *SV = llvm::ConstantVector::get(Indices); 3252 return Builder.CreateShuffleVector(Ops[0], Ops[1], SV, "vext"); 3253 } 3254 case NEON::BI__builtin_neon_vfma_v: 3255 case NEON::BI__builtin_neon_vfmaq_v: { 3256 Value *F = CGM.getIntrinsic(Intrinsic::fma, Ty); 3257 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 3258 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 3259 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 3260 3261 // NEON intrinsic puts accumulator first, unlike the LLVM fma. 3262 return Builder.CreateCall(F, {Ops[1], Ops[2], Ops[0]}); 3263 } 3264 case NEON::BI__builtin_neon_vld1_v: 3265 case NEON::BI__builtin_neon_vld1q_v: { 3266 llvm::Type *Tys[] = {Ty, Int8PtrTy}; 3267 Ops.push_back(getAlignmentValue32(PtrOp0)); 3268 return EmitNeonCall(CGM.getIntrinsic(LLVMIntrinsic, Tys), Ops, "vld1"); 3269 } 3270 case NEON::BI__builtin_neon_vld2_v: 3271 case NEON::BI__builtin_neon_vld2q_v: 3272 case NEON::BI__builtin_neon_vld3_v: 3273 case NEON::BI__builtin_neon_vld3q_v: 3274 case NEON::BI__builtin_neon_vld4_v: 3275 case NEON::BI__builtin_neon_vld4q_v: { 3276 llvm::Type *Tys[] = {Ty, Int8PtrTy}; 3277 Function *F = CGM.getIntrinsic(LLVMIntrinsic, Tys); 3278 Value *Align = getAlignmentValue32(PtrOp1); 3279 Ops[1] = Builder.CreateCall(F, {Ops[1], Align}, NameHint); 3280 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 3281 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 3282 return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]); 3283 } 3284 case NEON::BI__builtin_neon_vld1_dup_v: 3285 case NEON::BI__builtin_neon_vld1q_dup_v: { 3286 Value *V = UndefValue::get(Ty); 3287 Ty = llvm::PointerType::getUnqual(VTy->getElementType()); 3288 PtrOp0 = Builder.CreateBitCast(PtrOp0, Ty); 3289 LoadInst *Ld = Builder.CreateLoad(PtrOp0); 3290 llvm::Constant *CI = ConstantInt::get(SizeTy, 0); 3291 Ops[0] = Builder.CreateInsertElement(V, Ld, CI); 3292 return EmitNeonSplat(Ops[0], CI); 3293 } 3294 case NEON::BI__builtin_neon_vld2_lane_v: 3295 case NEON::BI__builtin_neon_vld2q_lane_v: 3296 case NEON::BI__builtin_neon_vld3_lane_v: 3297 case NEON::BI__builtin_neon_vld3q_lane_v: 3298 case NEON::BI__builtin_neon_vld4_lane_v: 3299 case NEON::BI__builtin_neon_vld4q_lane_v: { 3300 llvm::Type *Tys[] = {Ty, Int8PtrTy}; 3301 Function *F = CGM.getIntrinsic(LLVMIntrinsic, Tys); 3302 for (unsigned I = 2; I < Ops.size() - 1; ++I) 3303 Ops[I] = Builder.CreateBitCast(Ops[I], Ty); 3304 Ops.push_back(getAlignmentValue32(PtrOp1)); 3305 Ops[1] = Builder.CreateCall(F, makeArrayRef(Ops).slice(1), NameHint); 3306 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 3307 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 3308 return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]); 3309 } 3310 case NEON::BI__builtin_neon_vmovl_v: { 3311 llvm::Type *DTy =llvm::VectorType::getTruncatedElementVectorType(VTy); 3312 Ops[0] = Builder.CreateBitCast(Ops[0], DTy); 3313 if (Usgn) 3314 return Builder.CreateZExt(Ops[0], Ty, "vmovl"); 3315 return Builder.CreateSExt(Ops[0], Ty, "vmovl"); 3316 } 3317 case NEON::BI__builtin_neon_vmovn_v: { 3318 llvm::Type *QTy = llvm::VectorType::getExtendedElementVectorType(VTy); 3319 Ops[0] = Builder.CreateBitCast(Ops[0], QTy); 3320 return Builder.CreateTrunc(Ops[0], Ty, "vmovn"); 3321 } 3322 case NEON::BI__builtin_neon_vmull_v: 3323 // FIXME: the integer vmull operations could be emitted in terms of pure 3324 // LLVM IR (2 exts followed by a mul). Unfortunately LLVM has a habit of 3325 // hoisting the exts outside loops. Until global ISel comes along that can 3326 // see through such movement this leads to bad CodeGen. So we need an 3327 // intrinsic for now. 3328 Int = Usgn ? Intrinsic::arm_neon_vmullu : Intrinsic::arm_neon_vmulls; 3329 Int = Type.isPoly() ? (unsigned)Intrinsic::arm_neon_vmullp : Int; 3330 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmull"); 3331 case NEON::BI__builtin_neon_vpadal_v: 3332 case NEON::BI__builtin_neon_vpadalq_v: { 3333 // The source operand type has twice as many elements of half the size. 3334 unsigned EltBits = VTy->getElementType()->getPrimitiveSizeInBits(); 3335 llvm::Type *EltTy = 3336 llvm::IntegerType::get(getLLVMContext(), EltBits / 2); 3337 llvm::Type *NarrowTy = 3338 llvm::VectorType::get(EltTy, VTy->getNumElements() * 2); 3339 llvm::Type *Tys[2] = { Ty, NarrowTy }; 3340 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, NameHint); 3341 } 3342 case NEON::BI__builtin_neon_vpaddl_v: 3343 case NEON::BI__builtin_neon_vpaddlq_v: { 3344 // The source operand type has twice as many elements of half the size. 3345 unsigned EltBits = VTy->getElementType()->getPrimitiveSizeInBits(); 3346 llvm::Type *EltTy = 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, "vpaddl"); 3351 } 3352 case NEON::BI__builtin_neon_vqdmlal_v: 3353 case NEON::BI__builtin_neon_vqdmlsl_v: { 3354 SmallVector<Value *, 2> MulOps(Ops.begin() + 1, Ops.end()); 3355 Ops[1] = 3356 EmitNeonCall(CGM.getIntrinsic(LLVMIntrinsic, Ty), MulOps, "vqdmlal"); 3357 Ops.resize(2); 3358 return EmitNeonCall(CGM.getIntrinsic(AltLLVMIntrinsic, Ty), Ops, NameHint); 3359 } 3360 case NEON::BI__builtin_neon_vqshl_n_v: 3361 case NEON::BI__builtin_neon_vqshlq_n_v: 3362 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqshl_n", 3363 1, false); 3364 case NEON::BI__builtin_neon_vqshlu_n_v: 3365 case NEON::BI__builtin_neon_vqshluq_n_v: 3366 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqshlu_n", 3367 1, false); 3368 case NEON::BI__builtin_neon_vrecpe_v: 3369 case NEON::BI__builtin_neon_vrecpeq_v: 3370 case NEON::BI__builtin_neon_vrsqrte_v: 3371 case NEON::BI__builtin_neon_vrsqrteq_v: 3372 Int = Ty->isFPOrFPVectorTy() ? LLVMIntrinsic : AltLLVMIntrinsic; 3373 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, NameHint); 3374 3375 case NEON::BI__builtin_neon_vrshr_n_v: 3376 case NEON::BI__builtin_neon_vrshrq_n_v: 3377 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrshr_n", 3378 1, true); 3379 case NEON::BI__builtin_neon_vshl_n_v: 3380 case NEON::BI__builtin_neon_vshlq_n_v: 3381 Ops[1] = EmitNeonShiftVector(Ops[1], Ty, false); 3382 return Builder.CreateShl(Builder.CreateBitCast(Ops[0],Ty), Ops[1], 3383 "vshl_n"); 3384 case NEON::BI__builtin_neon_vshll_n_v: { 3385 llvm::Type *SrcTy = llvm::VectorType::getTruncatedElementVectorType(VTy); 3386 Ops[0] = Builder.CreateBitCast(Ops[0], SrcTy); 3387 if (Usgn) 3388 Ops[0] = Builder.CreateZExt(Ops[0], VTy); 3389 else 3390 Ops[0] = Builder.CreateSExt(Ops[0], VTy); 3391 Ops[1] = EmitNeonShiftVector(Ops[1], VTy, false); 3392 return Builder.CreateShl(Ops[0], Ops[1], "vshll_n"); 3393 } 3394 case NEON::BI__builtin_neon_vshrn_n_v: { 3395 llvm::Type *SrcTy = llvm::VectorType::getExtendedElementVectorType(VTy); 3396 Ops[0] = Builder.CreateBitCast(Ops[0], SrcTy); 3397 Ops[1] = EmitNeonShiftVector(Ops[1], SrcTy, false); 3398 if (Usgn) 3399 Ops[0] = Builder.CreateLShr(Ops[0], Ops[1]); 3400 else 3401 Ops[0] = Builder.CreateAShr(Ops[0], Ops[1]); 3402 return Builder.CreateTrunc(Ops[0], Ty, "vshrn_n"); 3403 } 3404 case NEON::BI__builtin_neon_vshr_n_v: 3405 case NEON::BI__builtin_neon_vshrq_n_v: 3406 return EmitNeonRShiftImm(Ops[0], Ops[1], Ty, Usgn, "vshr_n"); 3407 case NEON::BI__builtin_neon_vst1_v: 3408 case NEON::BI__builtin_neon_vst1q_v: 3409 case NEON::BI__builtin_neon_vst2_v: 3410 case NEON::BI__builtin_neon_vst2q_v: 3411 case NEON::BI__builtin_neon_vst3_v: 3412 case NEON::BI__builtin_neon_vst3q_v: 3413 case NEON::BI__builtin_neon_vst4_v: 3414 case NEON::BI__builtin_neon_vst4q_v: 3415 case NEON::BI__builtin_neon_vst2_lane_v: 3416 case NEON::BI__builtin_neon_vst2q_lane_v: 3417 case NEON::BI__builtin_neon_vst3_lane_v: 3418 case NEON::BI__builtin_neon_vst3q_lane_v: 3419 case NEON::BI__builtin_neon_vst4_lane_v: 3420 case NEON::BI__builtin_neon_vst4q_lane_v: { 3421 llvm::Type *Tys[] = {Int8PtrTy, Ty}; 3422 Ops.push_back(getAlignmentValue32(PtrOp0)); 3423 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, ""); 3424 } 3425 case NEON::BI__builtin_neon_vsubhn_v: { 3426 llvm::VectorType *SrcTy = 3427 llvm::VectorType::getExtendedElementVectorType(VTy); 3428 3429 // %sum = add <4 x i32> %lhs, %rhs 3430 Ops[0] = Builder.CreateBitCast(Ops[0], SrcTy); 3431 Ops[1] = Builder.CreateBitCast(Ops[1], SrcTy); 3432 Ops[0] = Builder.CreateSub(Ops[0], Ops[1], "vsubhn"); 3433 3434 // %high = lshr <4 x i32> %sum, <i32 16, i32 16, i32 16, i32 16> 3435 Constant *ShiftAmt = 3436 ConstantInt::get(SrcTy, SrcTy->getScalarSizeInBits() / 2); 3437 Ops[0] = Builder.CreateLShr(Ops[0], ShiftAmt, "vsubhn"); 3438 3439 // %res = trunc <4 x i32> %high to <4 x i16> 3440 return Builder.CreateTrunc(Ops[0], VTy, "vsubhn"); 3441 } 3442 case NEON::BI__builtin_neon_vtrn_v: 3443 case NEON::BI__builtin_neon_vtrnq_v: { 3444 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty)); 3445 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 3446 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 3447 Value *SV = nullptr; 3448 3449 for (unsigned vi = 0; vi != 2; ++vi) { 3450 SmallVector<Constant*, 16> Indices; 3451 for (unsigned i = 0, e = VTy->getNumElements(); i != e; i += 2) { 3452 Indices.push_back(Builder.getInt32(i+vi)); 3453 Indices.push_back(Builder.getInt32(i+e+vi)); 3454 } 3455 Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ty, Ops[0], vi); 3456 SV = llvm::ConstantVector::get(Indices); 3457 SV = Builder.CreateShuffleVector(Ops[1], Ops[2], SV, "vtrn"); 3458 SV = Builder.CreateDefaultAlignedStore(SV, Addr); 3459 } 3460 return SV; 3461 } 3462 case NEON::BI__builtin_neon_vtst_v: 3463 case NEON::BI__builtin_neon_vtstq_v: { 3464 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 3465 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 3466 Ops[0] = Builder.CreateAnd(Ops[0], Ops[1]); 3467 Ops[0] = Builder.CreateICmp(ICmpInst::ICMP_NE, Ops[0], 3468 ConstantAggregateZero::get(Ty)); 3469 return Builder.CreateSExt(Ops[0], Ty, "vtst"); 3470 } 3471 case NEON::BI__builtin_neon_vuzp_v: 3472 case NEON::BI__builtin_neon_vuzpq_v: { 3473 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty)); 3474 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 3475 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 3476 Value *SV = nullptr; 3477 3478 for (unsigned vi = 0; vi != 2; ++vi) { 3479 SmallVector<Constant*, 16> Indices; 3480 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) 3481 Indices.push_back(ConstantInt::get(Int32Ty, 2*i+vi)); 3482 3483 Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ty, Ops[0], vi); 3484 SV = llvm::ConstantVector::get(Indices); 3485 SV = Builder.CreateShuffleVector(Ops[1], Ops[2], SV, "vuzp"); 3486 SV = Builder.CreateDefaultAlignedStore(SV, Addr); 3487 } 3488 return SV; 3489 } 3490 case NEON::BI__builtin_neon_vzip_v: 3491 case NEON::BI__builtin_neon_vzipq_v: { 3492 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty)); 3493 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 3494 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 3495 Value *SV = nullptr; 3496 3497 for (unsigned vi = 0; vi != 2; ++vi) { 3498 SmallVector<Constant*, 16> Indices; 3499 for (unsigned i = 0, e = VTy->getNumElements(); i != e; i += 2) { 3500 Indices.push_back(ConstantInt::get(Int32Ty, (i + vi*e) >> 1)); 3501 Indices.push_back(ConstantInt::get(Int32Ty, ((i + vi*e) >> 1)+e)); 3502 } 3503 Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ty, Ops[0], vi); 3504 SV = llvm::ConstantVector::get(Indices); 3505 SV = Builder.CreateShuffleVector(Ops[1], Ops[2], SV, "vzip"); 3506 SV = Builder.CreateDefaultAlignedStore(SV, Addr); 3507 } 3508 return SV; 3509 } 3510 } 3511 3512 assert(Int && "Expected valid intrinsic number"); 3513 3514 // Determine the type(s) of this overloaded AArch64 intrinsic. 3515 Function *F = LookupNeonLLVMIntrinsic(Int, Modifier, Ty, E); 3516 3517 Value *Result = EmitNeonCall(F, Ops, NameHint); 3518 llvm::Type *ResultType = ConvertType(E->getType()); 3519 // AArch64 intrinsic one-element vector type cast to 3520 // scalar type expected by the builtin 3521 return Builder.CreateBitCast(Result, ResultType, NameHint); 3522 } 3523 3524 Value *CodeGenFunction::EmitAArch64CompareBuiltinExpr( 3525 Value *Op, llvm::Type *Ty, const CmpInst::Predicate Fp, 3526 const CmpInst::Predicate Ip, const Twine &Name) { 3527 llvm::Type *OTy = Op->getType(); 3528 3529 // FIXME: this is utterly horrific. We should not be looking at previous 3530 // codegen context to find out what needs doing. Unfortunately TableGen 3531 // currently gives us exactly the same calls for vceqz_f32 and vceqz_s32 3532 // (etc). 3533 if (BitCastInst *BI = dyn_cast<BitCastInst>(Op)) 3534 OTy = BI->getOperand(0)->getType(); 3535 3536 Op = Builder.CreateBitCast(Op, OTy); 3537 if (OTy->getScalarType()->isFloatingPointTy()) { 3538 Op = Builder.CreateFCmp(Fp, Op, Constant::getNullValue(OTy)); 3539 } else { 3540 Op = Builder.CreateICmp(Ip, Op, Constant::getNullValue(OTy)); 3541 } 3542 return Builder.CreateSExt(Op, Ty, Name); 3543 } 3544 3545 static Value *packTBLDVectorList(CodeGenFunction &CGF, ArrayRef<Value *> Ops, 3546 Value *ExtOp, Value *IndexOp, 3547 llvm::Type *ResTy, unsigned IntID, 3548 const char *Name) { 3549 SmallVector<Value *, 2> TblOps; 3550 if (ExtOp) 3551 TblOps.push_back(ExtOp); 3552 3553 // Build a vector containing sequential number like (0, 1, 2, ..., 15) 3554 SmallVector<Constant*, 16> Indices; 3555 llvm::VectorType *TblTy = cast<llvm::VectorType>(Ops[0]->getType()); 3556 for (unsigned i = 0, e = TblTy->getNumElements(); i != e; ++i) { 3557 Indices.push_back(ConstantInt::get(CGF.Int32Ty, 2*i)); 3558 Indices.push_back(ConstantInt::get(CGF.Int32Ty, 2*i+1)); 3559 } 3560 Value *SV = llvm::ConstantVector::get(Indices); 3561 3562 int PairPos = 0, End = Ops.size() - 1; 3563 while (PairPos < End) { 3564 TblOps.push_back(CGF.Builder.CreateShuffleVector(Ops[PairPos], 3565 Ops[PairPos+1], SV, Name)); 3566 PairPos += 2; 3567 } 3568 3569 // If there's an odd number of 64-bit lookup table, fill the high 64-bit 3570 // of the 128-bit lookup table with zero. 3571 if (PairPos == End) { 3572 Value *ZeroTbl = ConstantAggregateZero::get(TblTy); 3573 TblOps.push_back(CGF.Builder.CreateShuffleVector(Ops[PairPos], 3574 ZeroTbl, SV, Name)); 3575 } 3576 3577 Function *TblF; 3578 TblOps.push_back(IndexOp); 3579 TblF = CGF.CGM.getIntrinsic(IntID, ResTy); 3580 3581 return CGF.EmitNeonCall(TblF, TblOps, Name); 3582 } 3583 3584 Value *CodeGenFunction::GetValueForARMHint(unsigned BuiltinID) { 3585 unsigned Value; 3586 switch (BuiltinID) { 3587 default: 3588 return nullptr; 3589 case ARM::BI__builtin_arm_nop: 3590 Value = 0; 3591 break; 3592 case ARM::BI__builtin_arm_yield: 3593 case ARM::BI__yield: 3594 Value = 1; 3595 break; 3596 case ARM::BI__builtin_arm_wfe: 3597 case ARM::BI__wfe: 3598 Value = 2; 3599 break; 3600 case ARM::BI__builtin_arm_wfi: 3601 case ARM::BI__wfi: 3602 Value = 3; 3603 break; 3604 case ARM::BI__builtin_arm_sev: 3605 case ARM::BI__sev: 3606 Value = 4; 3607 break; 3608 case ARM::BI__builtin_arm_sevl: 3609 case ARM::BI__sevl: 3610 Value = 5; 3611 break; 3612 } 3613 3614 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::arm_hint), 3615 llvm::ConstantInt::get(Int32Ty, Value)); 3616 } 3617 3618 // Generates the IR for the read/write special register builtin, 3619 // ValueType is the type of the value that is to be written or read, 3620 // RegisterType is the type of the register being written to or read from. 3621 static Value *EmitSpecialRegisterBuiltin(CodeGenFunction &CGF, 3622 const CallExpr *E, 3623 llvm::Type *RegisterType, 3624 llvm::Type *ValueType, bool IsRead) { 3625 // write and register intrinsics only support 32 and 64 bit operations. 3626 assert((RegisterType->isIntegerTy(32) || RegisterType->isIntegerTy(64)) 3627 && "Unsupported size for register."); 3628 3629 CodeGen::CGBuilderTy &Builder = CGF.Builder; 3630 CodeGen::CodeGenModule &CGM = CGF.CGM; 3631 LLVMContext &Context = CGM.getLLVMContext(); 3632 3633 const Expr *SysRegStrExpr = E->getArg(0)->IgnoreParenCasts(); 3634 StringRef SysReg = cast<StringLiteral>(SysRegStrExpr)->getString(); 3635 3636 llvm::Metadata *Ops[] = { llvm::MDString::get(Context, SysReg) }; 3637 llvm::MDNode *RegName = llvm::MDNode::get(Context, Ops); 3638 llvm::Value *Metadata = llvm::MetadataAsValue::get(Context, RegName); 3639 3640 llvm::Type *Types[] = { RegisterType }; 3641 3642 bool MixedTypes = RegisterType->isIntegerTy(64) && ValueType->isIntegerTy(32); 3643 assert(!(RegisterType->isIntegerTy(32) && ValueType->isIntegerTy(64)) 3644 && "Can't fit 64-bit value in 32-bit register"); 3645 3646 if (IsRead) { 3647 llvm::Value *F = CGM.getIntrinsic(llvm::Intrinsic::read_register, Types); 3648 llvm::Value *Call = Builder.CreateCall(F, Metadata); 3649 3650 if (MixedTypes) 3651 // Read into 64 bit register and then truncate result to 32 bit. 3652 return Builder.CreateTrunc(Call, ValueType); 3653 3654 if (ValueType->isPointerTy()) 3655 // Have i32/i64 result (Call) but want to return a VoidPtrTy (i8*). 3656 return Builder.CreateIntToPtr(Call, ValueType); 3657 3658 return Call; 3659 } 3660 3661 llvm::Value *F = CGM.getIntrinsic(llvm::Intrinsic::write_register, Types); 3662 llvm::Value *ArgValue = CGF.EmitScalarExpr(E->getArg(1)); 3663 if (MixedTypes) { 3664 // Extend 32 bit write value to 64 bit to pass to write. 3665 ArgValue = Builder.CreateZExt(ArgValue, RegisterType); 3666 return Builder.CreateCall(F, { Metadata, ArgValue }); 3667 } 3668 3669 if (ValueType->isPointerTy()) { 3670 // Have VoidPtrTy ArgValue but want to return an i32/i64. 3671 ArgValue = Builder.CreatePtrToInt(ArgValue, RegisterType); 3672 return Builder.CreateCall(F, { Metadata, ArgValue }); 3673 } 3674 3675 return Builder.CreateCall(F, { Metadata, ArgValue }); 3676 } 3677 3678 /// Return true if BuiltinID is an overloaded Neon intrinsic with an extra 3679 /// argument that specifies the vector type. 3680 static bool HasExtraNeonArgument(unsigned BuiltinID) { 3681 switch (BuiltinID) { 3682 default: break; 3683 case NEON::BI__builtin_neon_vget_lane_i8: 3684 case NEON::BI__builtin_neon_vget_lane_i16: 3685 case NEON::BI__builtin_neon_vget_lane_i32: 3686 case NEON::BI__builtin_neon_vget_lane_i64: 3687 case NEON::BI__builtin_neon_vget_lane_f32: 3688 case NEON::BI__builtin_neon_vgetq_lane_i8: 3689 case NEON::BI__builtin_neon_vgetq_lane_i16: 3690 case NEON::BI__builtin_neon_vgetq_lane_i32: 3691 case NEON::BI__builtin_neon_vgetq_lane_i64: 3692 case NEON::BI__builtin_neon_vgetq_lane_f32: 3693 case NEON::BI__builtin_neon_vset_lane_i8: 3694 case NEON::BI__builtin_neon_vset_lane_i16: 3695 case NEON::BI__builtin_neon_vset_lane_i32: 3696 case NEON::BI__builtin_neon_vset_lane_i64: 3697 case NEON::BI__builtin_neon_vset_lane_f32: 3698 case NEON::BI__builtin_neon_vsetq_lane_i8: 3699 case NEON::BI__builtin_neon_vsetq_lane_i16: 3700 case NEON::BI__builtin_neon_vsetq_lane_i32: 3701 case NEON::BI__builtin_neon_vsetq_lane_i64: 3702 case NEON::BI__builtin_neon_vsetq_lane_f32: 3703 case NEON::BI__builtin_neon_vsha1h_u32: 3704 case NEON::BI__builtin_neon_vsha1cq_u32: 3705 case NEON::BI__builtin_neon_vsha1pq_u32: 3706 case NEON::BI__builtin_neon_vsha1mq_u32: 3707 case ARM::BI_MoveToCoprocessor: 3708 case ARM::BI_MoveToCoprocessor2: 3709 return false; 3710 } 3711 return true; 3712 } 3713 3714 Value *CodeGenFunction::EmitARMBuiltinExpr(unsigned BuiltinID, 3715 const CallExpr *E) { 3716 if (auto Hint = GetValueForARMHint(BuiltinID)) 3717 return Hint; 3718 3719 if (BuiltinID == ARM::BI__emit) { 3720 bool IsThumb = getTarget().getTriple().getArch() == llvm::Triple::thumb; 3721 llvm::FunctionType *FTy = 3722 llvm::FunctionType::get(VoidTy, /*Variadic=*/false); 3723 3724 APSInt Value; 3725 if (!E->getArg(0)->EvaluateAsInt(Value, CGM.getContext())) 3726 llvm_unreachable("Sema will ensure that the parameter is constant"); 3727 3728 uint64_t ZExtValue = Value.zextOrTrunc(IsThumb ? 16 : 32).getZExtValue(); 3729 3730 llvm::InlineAsm *Emit = 3731 IsThumb ? InlineAsm::get(FTy, ".inst.n 0x" + utohexstr(ZExtValue), "", 3732 /*SideEffects=*/true) 3733 : InlineAsm::get(FTy, ".inst 0x" + utohexstr(ZExtValue), "", 3734 /*SideEffects=*/true); 3735 3736 return Builder.CreateCall(Emit); 3737 } 3738 3739 if (BuiltinID == ARM::BI__builtin_arm_dbg) { 3740 Value *Option = EmitScalarExpr(E->getArg(0)); 3741 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::arm_dbg), Option); 3742 } 3743 3744 if (BuiltinID == ARM::BI__builtin_arm_prefetch) { 3745 Value *Address = EmitScalarExpr(E->getArg(0)); 3746 Value *RW = EmitScalarExpr(E->getArg(1)); 3747 Value *IsData = EmitScalarExpr(E->getArg(2)); 3748 3749 // Locality is not supported on ARM target 3750 Value *Locality = llvm::ConstantInt::get(Int32Ty, 3); 3751 3752 Value *F = CGM.getIntrinsic(Intrinsic::prefetch); 3753 return Builder.CreateCall(F, {Address, RW, Locality, IsData}); 3754 } 3755 3756 if (BuiltinID == ARM::BI__builtin_arm_rbit) { 3757 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::arm_rbit), 3758 EmitScalarExpr(E->getArg(0)), 3759 "rbit"); 3760 } 3761 3762 if (BuiltinID == ARM::BI__clear_cache) { 3763 assert(E->getNumArgs() == 2 && "__clear_cache takes 2 arguments"); 3764 const FunctionDecl *FD = E->getDirectCallee(); 3765 Value *Ops[2]; 3766 for (unsigned i = 0; i < 2; i++) 3767 Ops[i] = EmitScalarExpr(E->getArg(i)); 3768 llvm::Type *Ty = CGM.getTypes().ConvertType(FD->getType()); 3769 llvm::FunctionType *FTy = cast<llvm::FunctionType>(Ty); 3770 StringRef Name = FD->getName(); 3771 return EmitNounwindRuntimeCall(CGM.CreateRuntimeFunction(FTy, Name), Ops); 3772 } 3773 3774 if (BuiltinID == ARM::BI__builtin_arm_ldrexd || 3775 ((BuiltinID == ARM::BI__builtin_arm_ldrex || 3776 BuiltinID == ARM::BI__builtin_arm_ldaex) && 3777 getContext().getTypeSize(E->getType()) == 64) || 3778 BuiltinID == ARM::BI__ldrexd) { 3779 Function *F; 3780 3781 switch (BuiltinID) { 3782 default: llvm_unreachable("unexpected builtin"); 3783 case ARM::BI__builtin_arm_ldaex: 3784 F = CGM.getIntrinsic(Intrinsic::arm_ldaexd); 3785 break; 3786 case ARM::BI__builtin_arm_ldrexd: 3787 case ARM::BI__builtin_arm_ldrex: 3788 case ARM::BI__ldrexd: 3789 F = CGM.getIntrinsic(Intrinsic::arm_ldrexd); 3790 break; 3791 } 3792 3793 Value *LdPtr = EmitScalarExpr(E->getArg(0)); 3794 Value *Val = Builder.CreateCall(F, Builder.CreateBitCast(LdPtr, Int8PtrTy), 3795 "ldrexd"); 3796 3797 Value *Val0 = Builder.CreateExtractValue(Val, 1); 3798 Value *Val1 = Builder.CreateExtractValue(Val, 0); 3799 Val0 = Builder.CreateZExt(Val0, Int64Ty); 3800 Val1 = Builder.CreateZExt(Val1, Int64Ty); 3801 3802 Value *ShiftCst = llvm::ConstantInt::get(Int64Ty, 32); 3803 Val = Builder.CreateShl(Val0, ShiftCst, "shl", true /* nuw */); 3804 Val = Builder.CreateOr(Val, Val1); 3805 return Builder.CreateBitCast(Val, ConvertType(E->getType())); 3806 } 3807 3808 if (BuiltinID == ARM::BI__builtin_arm_ldrex || 3809 BuiltinID == ARM::BI__builtin_arm_ldaex) { 3810 Value *LoadAddr = EmitScalarExpr(E->getArg(0)); 3811 3812 QualType Ty = E->getType(); 3813 llvm::Type *RealResTy = ConvertType(Ty); 3814 llvm::Type *IntResTy = llvm::IntegerType::get(getLLVMContext(), 3815 getContext().getTypeSize(Ty)); 3816 LoadAddr = Builder.CreateBitCast(LoadAddr, IntResTy->getPointerTo()); 3817 3818 Function *F = CGM.getIntrinsic(BuiltinID == ARM::BI__builtin_arm_ldaex 3819 ? Intrinsic::arm_ldaex 3820 : Intrinsic::arm_ldrex, 3821 LoadAddr->getType()); 3822 Value *Val = Builder.CreateCall(F, LoadAddr, "ldrex"); 3823 3824 if (RealResTy->isPointerTy()) 3825 return Builder.CreateIntToPtr(Val, RealResTy); 3826 else { 3827 Val = Builder.CreateTruncOrBitCast(Val, IntResTy); 3828 return Builder.CreateBitCast(Val, RealResTy); 3829 } 3830 } 3831 3832 if (BuiltinID == ARM::BI__builtin_arm_strexd || 3833 ((BuiltinID == ARM::BI__builtin_arm_stlex || 3834 BuiltinID == ARM::BI__builtin_arm_strex) && 3835 getContext().getTypeSize(E->getArg(0)->getType()) == 64)) { 3836 Function *F = CGM.getIntrinsic(BuiltinID == ARM::BI__builtin_arm_stlex 3837 ? Intrinsic::arm_stlexd 3838 : Intrinsic::arm_strexd); 3839 llvm::Type *STy = llvm::StructType::get(Int32Ty, Int32Ty, nullptr); 3840 3841 Address Tmp = CreateMemTemp(E->getArg(0)->getType()); 3842 Value *Val = EmitScalarExpr(E->getArg(0)); 3843 Builder.CreateStore(Val, Tmp); 3844 3845 Address LdPtr = Builder.CreateBitCast(Tmp,llvm::PointerType::getUnqual(STy)); 3846 Val = Builder.CreateLoad(LdPtr); 3847 3848 Value *Arg0 = Builder.CreateExtractValue(Val, 0); 3849 Value *Arg1 = Builder.CreateExtractValue(Val, 1); 3850 Value *StPtr = Builder.CreateBitCast(EmitScalarExpr(E->getArg(1)), Int8PtrTy); 3851 return Builder.CreateCall(F, {Arg0, Arg1, StPtr}, "strexd"); 3852 } 3853 3854 if (BuiltinID == ARM::BI__builtin_arm_strex || 3855 BuiltinID == ARM::BI__builtin_arm_stlex) { 3856 Value *StoreVal = EmitScalarExpr(E->getArg(0)); 3857 Value *StoreAddr = EmitScalarExpr(E->getArg(1)); 3858 3859 QualType Ty = E->getArg(0)->getType(); 3860 llvm::Type *StoreTy = llvm::IntegerType::get(getLLVMContext(), 3861 getContext().getTypeSize(Ty)); 3862 StoreAddr = Builder.CreateBitCast(StoreAddr, StoreTy->getPointerTo()); 3863 3864 if (StoreVal->getType()->isPointerTy()) 3865 StoreVal = Builder.CreatePtrToInt(StoreVal, Int32Ty); 3866 else { 3867 StoreVal = Builder.CreateBitCast(StoreVal, StoreTy); 3868 StoreVal = Builder.CreateZExtOrBitCast(StoreVal, Int32Ty); 3869 } 3870 3871 Function *F = CGM.getIntrinsic(BuiltinID == ARM::BI__builtin_arm_stlex 3872 ? Intrinsic::arm_stlex 3873 : Intrinsic::arm_strex, 3874 StoreAddr->getType()); 3875 return Builder.CreateCall(F, {StoreVal, StoreAddr}, "strex"); 3876 } 3877 3878 if (BuiltinID == ARM::BI__builtin_arm_clrex) { 3879 Function *F = CGM.getIntrinsic(Intrinsic::arm_clrex); 3880 return Builder.CreateCall(F); 3881 } 3882 3883 // CRC32 3884 Intrinsic::ID CRCIntrinsicID = Intrinsic::not_intrinsic; 3885 switch (BuiltinID) { 3886 case ARM::BI__builtin_arm_crc32b: 3887 CRCIntrinsicID = Intrinsic::arm_crc32b; break; 3888 case ARM::BI__builtin_arm_crc32cb: 3889 CRCIntrinsicID = Intrinsic::arm_crc32cb; break; 3890 case ARM::BI__builtin_arm_crc32h: 3891 CRCIntrinsicID = Intrinsic::arm_crc32h; break; 3892 case ARM::BI__builtin_arm_crc32ch: 3893 CRCIntrinsicID = Intrinsic::arm_crc32ch; break; 3894 case ARM::BI__builtin_arm_crc32w: 3895 case ARM::BI__builtin_arm_crc32d: 3896 CRCIntrinsicID = Intrinsic::arm_crc32w; break; 3897 case ARM::BI__builtin_arm_crc32cw: 3898 case ARM::BI__builtin_arm_crc32cd: 3899 CRCIntrinsicID = Intrinsic::arm_crc32cw; break; 3900 } 3901 3902 if (CRCIntrinsicID != Intrinsic::not_intrinsic) { 3903 Value *Arg0 = EmitScalarExpr(E->getArg(0)); 3904 Value *Arg1 = EmitScalarExpr(E->getArg(1)); 3905 3906 // crc32{c,}d intrinsics are implemnted as two calls to crc32{c,}w 3907 // intrinsics, hence we need different codegen for these cases. 3908 if (BuiltinID == ARM::BI__builtin_arm_crc32d || 3909 BuiltinID == ARM::BI__builtin_arm_crc32cd) { 3910 Value *C1 = llvm::ConstantInt::get(Int64Ty, 32); 3911 Value *Arg1a = Builder.CreateTruncOrBitCast(Arg1, Int32Ty); 3912 Value *Arg1b = Builder.CreateLShr(Arg1, C1); 3913 Arg1b = Builder.CreateTruncOrBitCast(Arg1b, Int32Ty); 3914 3915 Function *F = CGM.getIntrinsic(CRCIntrinsicID); 3916 Value *Res = Builder.CreateCall(F, {Arg0, Arg1a}); 3917 return Builder.CreateCall(F, {Res, Arg1b}); 3918 } else { 3919 Arg1 = Builder.CreateZExtOrBitCast(Arg1, Int32Ty); 3920 3921 Function *F = CGM.getIntrinsic(CRCIntrinsicID); 3922 return Builder.CreateCall(F, {Arg0, Arg1}); 3923 } 3924 } 3925 3926 if (BuiltinID == ARM::BI__builtin_arm_rsr || 3927 BuiltinID == ARM::BI__builtin_arm_rsr64 || 3928 BuiltinID == ARM::BI__builtin_arm_rsrp || 3929 BuiltinID == ARM::BI__builtin_arm_wsr || 3930 BuiltinID == ARM::BI__builtin_arm_wsr64 || 3931 BuiltinID == ARM::BI__builtin_arm_wsrp) { 3932 3933 bool IsRead = BuiltinID == ARM::BI__builtin_arm_rsr || 3934 BuiltinID == ARM::BI__builtin_arm_rsr64 || 3935 BuiltinID == ARM::BI__builtin_arm_rsrp; 3936 3937 bool IsPointerBuiltin = BuiltinID == ARM::BI__builtin_arm_rsrp || 3938 BuiltinID == ARM::BI__builtin_arm_wsrp; 3939 3940 bool Is64Bit = BuiltinID == ARM::BI__builtin_arm_rsr64 || 3941 BuiltinID == ARM::BI__builtin_arm_wsr64; 3942 3943 llvm::Type *ValueType; 3944 llvm::Type *RegisterType; 3945 if (IsPointerBuiltin) { 3946 ValueType = VoidPtrTy; 3947 RegisterType = Int32Ty; 3948 } else if (Is64Bit) { 3949 ValueType = RegisterType = Int64Ty; 3950 } else { 3951 ValueType = RegisterType = Int32Ty; 3952 } 3953 3954 return EmitSpecialRegisterBuiltin(*this, E, RegisterType, ValueType, IsRead); 3955 } 3956 3957 // Find out if any arguments are required to be integer constant 3958 // expressions. 3959 unsigned ICEArguments = 0; 3960 ASTContext::GetBuiltinTypeError Error; 3961 getContext().GetBuiltinType(BuiltinID, Error, &ICEArguments); 3962 assert(Error == ASTContext::GE_None && "Should not codegen an error"); 3963 3964 auto getAlignmentValue32 = [&](Address addr) -> Value* { 3965 return Builder.getInt32(addr.getAlignment().getQuantity()); 3966 }; 3967 3968 Address PtrOp0 = Address::invalid(); 3969 Address PtrOp1 = Address::invalid(); 3970 SmallVector<Value*, 4> Ops; 3971 bool HasExtraArg = HasExtraNeonArgument(BuiltinID); 3972 unsigned NumArgs = E->getNumArgs() - (HasExtraArg ? 1 : 0); 3973 for (unsigned i = 0, e = NumArgs; i != e; i++) { 3974 if (i == 0) { 3975 switch (BuiltinID) { 3976 case NEON::BI__builtin_neon_vld1_v: 3977 case NEON::BI__builtin_neon_vld1q_v: 3978 case NEON::BI__builtin_neon_vld1q_lane_v: 3979 case NEON::BI__builtin_neon_vld1_lane_v: 3980 case NEON::BI__builtin_neon_vld1_dup_v: 3981 case NEON::BI__builtin_neon_vld1q_dup_v: 3982 case NEON::BI__builtin_neon_vst1_v: 3983 case NEON::BI__builtin_neon_vst1q_v: 3984 case NEON::BI__builtin_neon_vst1q_lane_v: 3985 case NEON::BI__builtin_neon_vst1_lane_v: 3986 case NEON::BI__builtin_neon_vst2_v: 3987 case NEON::BI__builtin_neon_vst2q_v: 3988 case NEON::BI__builtin_neon_vst2_lane_v: 3989 case NEON::BI__builtin_neon_vst2q_lane_v: 3990 case NEON::BI__builtin_neon_vst3_v: 3991 case NEON::BI__builtin_neon_vst3q_v: 3992 case NEON::BI__builtin_neon_vst3_lane_v: 3993 case NEON::BI__builtin_neon_vst3q_lane_v: 3994 case NEON::BI__builtin_neon_vst4_v: 3995 case NEON::BI__builtin_neon_vst4q_v: 3996 case NEON::BI__builtin_neon_vst4_lane_v: 3997 case NEON::BI__builtin_neon_vst4q_lane_v: 3998 // Get the alignment for the argument in addition to the value; 3999 // we'll use it later. 4000 PtrOp0 = EmitPointerWithAlignment(E->getArg(0)); 4001 Ops.push_back(PtrOp0.getPointer()); 4002 continue; 4003 } 4004 } 4005 if (i == 1) { 4006 switch (BuiltinID) { 4007 case NEON::BI__builtin_neon_vld2_v: 4008 case NEON::BI__builtin_neon_vld2q_v: 4009 case NEON::BI__builtin_neon_vld3_v: 4010 case NEON::BI__builtin_neon_vld3q_v: 4011 case NEON::BI__builtin_neon_vld4_v: 4012 case NEON::BI__builtin_neon_vld4q_v: 4013 case NEON::BI__builtin_neon_vld2_lane_v: 4014 case NEON::BI__builtin_neon_vld2q_lane_v: 4015 case NEON::BI__builtin_neon_vld3_lane_v: 4016 case NEON::BI__builtin_neon_vld3q_lane_v: 4017 case NEON::BI__builtin_neon_vld4_lane_v: 4018 case NEON::BI__builtin_neon_vld4q_lane_v: 4019 case NEON::BI__builtin_neon_vld2_dup_v: 4020 case NEON::BI__builtin_neon_vld3_dup_v: 4021 case NEON::BI__builtin_neon_vld4_dup_v: 4022 // Get the alignment for the argument in addition to the value; 4023 // we'll use it later. 4024 PtrOp1 = EmitPointerWithAlignment(E->getArg(1)); 4025 Ops.push_back(PtrOp1.getPointer()); 4026 continue; 4027 } 4028 } 4029 4030 if ((ICEArguments & (1 << i)) == 0) { 4031 Ops.push_back(EmitScalarExpr(E->getArg(i))); 4032 } else { 4033 // If this is required to be a constant, constant fold it so that we know 4034 // that the generated intrinsic gets a ConstantInt. 4035 llvm::APSInt Result; 4036 bool IsConst = E->getArg(i)->isIntegerConstantExpr(Result, getContext()); 4037 assert(IsConst && "Constant arg isn't actually constant?"); (void)IsConst; 4038 Ops.push_back(llvm::ConstantInt::get(getLLVMContext(), Result)); 4039 } 4040 } 4041 4042 switch (BuiltinID) { 4043 default: break; 4044 4045 case NEON::BI__builtin_neon_vget_lane_i8: 4046 case NEON::BI__builtin_neon_vget_lane_i16: 4047 case NEON::BI__builtin_neon_vget_lane_i32: 4048 case NEON::BI__builtin_neon_vget_lane_i64: 4049 case NEON::BI__builtin_neon_vget_lane_f32: 4050 case NEON::BI__builtin_neon_vgetq_lane_i8: 4051 case NEON::BI__builtin_neon_vgetq_lane_i16: 4052 case NEON::BI__builtin_neon_vgetq_lane_i32: 4053 case NEON::BI__builtin_neon_vgetq_lane_i64: 4054 case NEON::BI__builtin_neon_vgetq_lane_f32: 4055 return Builder.CreateExtractElement(Ops[0], Ops[1], "vget_lane"); 4056 4057 case NEON::BI__builtin_neon_vset_lane_i8: 4058 case NEON::BI__builtin_neon_vset_lane_i16: 4059 case NEON::BI__builtin_neon_vset_lane_i32: 4060 case NEON::BI__builtin_neon_vset_lane_i64: 4061 case NEON::BI__builtin_neon_vset_lane_f32: 4062 case NEON::BI__builtin_neon_vsetq_lane_i8: 4063 case NEON::BI__builtin_neon_vsetq_lane_i16: 4064 case NEON::BI__builtin_neon_vsetq_lane_i32: 4065 case NEON::BI__builtin_neon_vsetq_lane_i64: 4066 case NEON::BI__builtin_neon_vsetq_lane_f32: 4067 return Builder.CreateInsertElement(Ops[1], Ops[0], Ops[2], "vset_lane"); 4068 4069 case NEON::BI__builtin_neon_vsha1h_u32: 4070 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_sha1h), Ops, 4071 "vsha1h"); 4072 case NEON::BI__builtin_neon_vsha1cq_u32: 4073 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_sha1c), Ops, 4074 "vsha1h"); 4075 case NEON::BI__builtin_neon_vsha1pq_u32: 4076 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_sha1p), Ops, 4077 "vsha1h"); 4078 case NEON::BI__builtin_neon_vsha1mq_u32: 4079 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_sha1m), Ops, 4080 "vsha1h"); 4081 4082 // The ARM _MoveToCoprocessor builtins put the input register value as 4083 // the first argument, but the LLVM intrinsic expects it as the third one. 4084 case ARM::BI_MoveToCoprocessor: 4085 case ARM::BI_MoveToCoprocessor2: { 4086 Function *F = CGM.getIntrinsic(BuiltinID == ARM::BI_MoveToCoprocessor ? 4087 Intrinsic::arm_mcr : Intrinsic::arm_mcr2); 4088 return Builder.CreateCall(F, {Ops[1], Ops[2], Ops[0], 4089 Ops[3], Ops[4], Ops[5]}); 4090 } 4091 } 4092 4093 // Get the last argument, which specifies the vector type. 4094 assert(HasExtraArg); 4095 llvm::APSInt Result; 4096 const Expr *Arg = E->getArg(E->getNumArgs()-1); 4097 if (!Arg->isIntegerConstantExpr(Result, getContext())) 4098 return nullptr; 4099 4100 if (BuiltinID == ARM::BI__builtin_arm_vcvtr_f || 4101 BuiltinID == ARM::BI__builtin_arm_vcvtr_d) { 4102 // Determine the overloaded type of this builtin. 4103 llvm::Type *Ty; 4104 if (BuiltinID == ARM::BI__builtin_arm_vcvtr_f) 4105 Ty = FloatTy; 4106 else 4107 Ty = DoubleTy; 4108 4109 // Determine whether this is an unsigned conversion or not. 4110 bool usgn = Result.getZExtValue() == 1; 4111 unsigned Int = usgn ? Intrinsic::arm_vcvtru : Intrinsic::arm_vcvtr; 4112 4113 // Call the appropriate intrinsic. 4114 Function *F = CGM.getIntrinsic(Int, Ty); 4115 return Builder.CreateCall(F, Ops, "vcvtr"); 4116 } 4117 4118 // Determine the type of this overloaded NEON intrinsic. 4119 NeonTypeFlags Type(Result.getZExtValue()); 4120 bool usgn = Type.isUnsigned(); 4121 bool rightShift = false; 4122 4123 llvm::VectorType *VTy = GetNeonType(this, Type); 4124 llvm::Type *Ty = VTy; 4125 if (!Ty) 4126 return nullptr; 4127 4128 // Many NEON builtins have identical semantics and uses in ARM and 4129 // AArch64. Emit these in a single function. 4130 auto IntrinsicMap = makeArrayRef(ARMSIMDIntrinsicMap); 4131 const NeonIntrinsicInfo *Builtin = findNeonIntrinsicInMap( 4132 IntrinsicMap, BuiltinID, NEONSIMDIntrinsicsProvenSorted); 4133 if (Builtin) 4134 return EmitCommonNeonBuiltinExpr( 4135 Builtin->BuiltinID, Builtin->LLVMIntrinsic, Builtin->AltLLVMIntrinsic, 4136 Builtin->NameHint, Builtin->TypeModifier, E, Ops, PtrOp0, PtrOp1); 4137 4138 unsigned Int; 4139 switch (BuiltinID) { 4140 default: return nullptr; 4141 case NEON::BI__builtin_neon_vld1q_lane_v: 4142 // Handle 64-bit integer elements as a special case. Use shuffles of 4143 // one-element vectors to avoid poor code for i64 in the backend. 4144 if (VTy->getElementType()->isIntegerTy(64)) { 4145 // Extract the other lane. 4146 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 4147 uint32_t Lane = cast<ConstantInt>(Ops[2])->getZExtValue(); 4148 Value *SV = llvm::ConstantVector::get(ConstantInt::get(Int32Ty, 1-Lane)); 4149 Ops[1] = Builder.CreateShuffleVector(Ops[1], Ops[1], SV); 4150 // Load the value as a one-element vector. 4151 Ty = llvm::VectorType::get(VTy->getElementType(), 1); 4152 llvm::Type *Tys[] = {Ty, Int8PtrTy}; 4153 Function *F = CGM.getIntrinsic(Intrinsic::arm_neon_vld1, Tys); 4154 Value *Align = getAlignmentValue32(PtrOp0); 4155 Value *Ld = Builder.CreateCall(F, {Ops[0], Align}); 4156 // Combine them. 4157 uint32_t Indices[] = {1 - Lane, Lane}; 4158 SV = llvm::ConstantDataVector::get(getLLVMContext(), Indices); 4159 return Builder.CreateShuffleVector(Ops[1], Ld, SV, "vld1q_lane"); 4160 } 4161 // fall through 4162 case NEON::BI__builtin_neon_vld1_lane_v: { 4163 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 4164 PtrOp0 = Builder.CreateElementBitCast(PtrOp0, VTy->getElementType()); 4165 Value *Ld = Builder.CreateLoad(PtrOp0); 4166 return Builder.CreateInsertElement(Ops[1], Ld, Ops[2], "vld1_lane"); 4167 } 4168 case NEON::BI__builtin_neon_vld2_dup_v: 4169 case NEON::BI__builtin_neon_vld3_dup_v: 4170 case NEON::BI__builtin_neon_vld4_dup_v: { 4171 // Handle 64-bit elements as a special-case. There is no "dup" needed. 4172 if (VTy->getElementType()->getPrimitiveSizeInBits() == 64) { 4173 switch (BuiltinID) { 4174 case NEON::BI__builtin_neon_vld2_dup_v: 4175 Int = Intrinsic::arm_neon_vld2; 4176 break; 4177 case NEON::BI__builtin_neon_vld3_dup_v: 4178 Int = Intrinsic::arm_neon_vld3; 4179 break; 4180 case NEON::BI__builtin_neon_vld4_dup_v: 4181 Int = Intrinsic::arm_neon_vld4; 4182 break; 4183 default: llvm_unreachable("unknown vld_dup intrinsic?"); 4184 } 4185 llvm::Type *Tys[] = {Ty, Int8PtrTy}; 4186 Function *F = CGM.getIntrinsic(Int, Tys); 4187 llvm::Value *Align = getAlignmentValue32(PtrOp1); 4188 Ops[1] = Builder.CreateCall(F, {Ops[1], Align}, "vld_dup"); 4189 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 4190 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 4191 return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]); 4192 } 4193 switch (BuiltinID) { 4194 case NEON::BI__builtin_neon_vld2_dup_v: 4195 Int = Intrinsic::arm_neon_vld2lane; 4196 break; 4197 case NEON::BI__builtin_neon_vld3_dup_v: 4198 Int = Intrinsic::arm_neon_vld3lane; 4199 break; 4200 case NEON::BI__builtin_neon_vld4_dup_v: 4201 Int = Intrinsic::arm_neon_vld4lane; 4202 break; 4203 default: llvm_unreachable("unknown vld_dup intrinsic?"); 4204 } 4205 llvm::Type *Tys[] = {Ty, Int8PtrTy}; 4206 Function *F = CGM.getIntrinsic(Int, Tys); 4207 llvm::StructType *STy = cast<llvm::StructType>(F->getReturnType()); 4208 4209 SmallVector<Value*, 6> Args; 4210 Args.push_back(Ops[1]); 4211 Args.append(STy->getNumElements(), UndefValue::get(Ty)); 4212 4213 llvm::Constant *CI = ConstantInt::get(Int32Ty, 0); 4214 Args.push_back(CI); 4215 Args.push_back(getAlignmentValue32(PtrOp1)); 4216 4217 Ops[1] = Builder.CreateCall(F, Args, "vld_dup"); 4218 // splat lane 0 to all elts in each vector of the result. 4219 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 4220 Value *Val = Builder.CreateExtractValue(Ops[1], i); 4221 Value *Elt = Builder.CreateBitCast(Val, Ty); 4222 Elt = EmitNeonSplat(Elt, CI); 4223 Elt = Builder.CreateBitCast(Elt, Val->getType()); 4224 Ops[1] = Builder.CreateInsertValue(Ops[1], Elt, i); 4225 } 4226 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 4227 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 4228 return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]); 4229 } 4230 case NEON::BI__builtin_neon_vqrshrn_n_v: 4231 Int = 4232 usgn ? Intrinsic::arm_neon_vqrshiftnu : Intrinsic::arm_neon_vqrshiftns; 4233 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqrshrn_n", 4234 1, true); 4235 case NEON::BI__builtin_neon_vqrshrun_n_v: 4236 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vqrshiftnsu, Ty), 4237 Ops, "vqrshrun_n", 1, true); 4238 case NEON::BI__builtin_neon_vqshrn_n_v: 4239 Int = usgn ? Intrinsic::arm_neon_vqshiftnu : Intrinsic::arm_neon_vqshiftns; 4240 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqshrn_n", 4241 1, true); 4242 case NEON::BI__builtin_neon_vqshrun_n_v: 4243 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vqshiftnsu, Ty), 4244 Ops, "vqshrun_n", 1, true); 4245 case NEON::BI__builtin_neon_vrecpe_v: 4246 case NEON::BI__builtin_neon_vrecpeq_v: 4247 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vrecpe, Ty), 4248 Ops, "vrecpe"); 4249 case NEON::BI__builtin_neon_vrshrn_n_v: 4250 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vrshiftn, Ty), 4251 Ops, "vrshrn_n", 1, true); 4252 case NEON::BI__builtin_neon_vrsra_n_v: 4253 case NEON::BI__builtin_neon_vrsraq_n_v: 4254 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 4255 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 4256 Ops[2] = EmitNeonShiftVector(Ops[2], Ty, true); 4257 Int = usgn ? Intrinsic::arm_neon_vrshiftu : Intrinsic::arm_neon_vrshifts; 4258 Ops[1] = Builder.CreateCall(CGM.getIntrinsic(Int, Ty), {Ops[1], Ops[2]}); 4259 return Builder.CreateAdd(Ops[0], Ops[1], "vrsra_n"); 4260 case NEON::BI__builtin_neon_vsri_n_v: 4261 case NEON::BI__builtin_neon_vsriq_n_v: 4262 rightShift = true; 4263 case NEON::BI__builtin_neon_vsli_n_v: 4264 case NEON::BI__builtin_neon_vsliq_n_v: 4265 Ops[2] = EmitNeonShiftVector(Ops[2], Ty, rightShift); 4266 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vshiftins, Ty), 4267 Ops, "vsli_n"); 4268 case NEON::BI__builtin_neon_vsra_n_v: 4269 case NEON::BI__builtin_neon_vsraq_n_v: 4270 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 4271 Ops[1] = EmitNeonRShiftImm(Ops[1], Ops[2], Ty, usgn, "vsra_n"); 4272 return Builder.CreateAdd(Ops[0], Ops[1]); 4273 case NEON::BI__builtin_neon_vst1q_lane_v: 4274 // Handle 64-bit integer elements as a special case. Use a shuffle to get 4275 // a one-element vector and avoid poor code for i64 in the backend. 4276 if (VTy->getElementType()->isIntegerTy(64)) { 4277 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 4278 Value *SV = llvm::ConstantVector::get(cast<llvm::Constant>(Ops[2])); 4279 Ops[1] = Builder.CreateShuffleVector(Ops[1], Ops[1], SV); 4280 Ops[2] = getAlignmentValue32(PtrOp0); 4281 llvm::Type *Tys[] = {Int8PtrTy, Ops[1]->getType()}; 4282 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::arm_neon_vst1, 4283 Tys), Ops); 4284 } 4285 // fall through 4286 case NEON::BI__builtin_neon_vst1_lane_v: { 4287 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 4288 Ops[1] = Builder.CreateExtractElement(Ops[1], Ops[2]); 4289 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 4290 auto St = Builder.CreateStore(Ops[1], Builder.CreateBitCast(PtrOp0, Ty)); 4291 return St; 4292 } 4293 case NEON::BI__builtin_neon_vtbl1_v: 4294 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbl1), 4295 Ops, "vtbl1"); 4296 case NEON::BI__builtin_neon_vtbl2_v: 4297 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbl2), 4298 Ops, "vtbl2"); 4299 case NEON::BI__builtin_neon_vtbl3_v: 4300 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbl3), 4301 Ops, "vtbl3"); 4302 case NEON::BI__builtin_neon_vtbl4_v: 4303 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbl4), 4304 Ops, "vtbl4"); 4305 case NEON::BI__builtin_neon_vtbx1_v: 4306 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbx1), 4307 Ops, "vtbx1"); 4308 case NEON::BI__builtin_neon_vtbx2_v: 4309 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbx2), 4310 Ops, "vtbx2"); 4311 case NEON::BI__builtin_neon_vtbx3_v: 4312 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbx3), 4313 Ops, "vtbx3"); 4314 case NEON::BI__builtin_neon_vtbx4_v: 4315 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbx4), 4316 Ops, "vtbx4"); 4317 } 4318 } 4319 4320 static Value *EmitAArch64TblBuiltinExpr(CodeGenFunction &CGF, unsigned BuiltinID, 4321 const CallExpr *E, 4322 SmallVectorImpl<Value *> &Ops) { 4323 unsigned int Int = 0; 4324 const char *s = nullptr; 4325 4326 switch (BuiltinID) { 4327 default: 4328 return nullptr; 4329 case NEON::BI__builtin_neon_vtbl1_v: 4330 case NEON::BI__builtin_neon_vqtbl1_v: 4331 case NEON::BI__builtin_neon_vqtbl1q_v: 4332 case NEON::BI__builtin_neon_vtbl2_v: 4333 case NEON::BI__builtin_neon_vqtbl2_v: 4334 case NEON::BI__builtin_neon_vqtbl2q_v: 4335 case NEON::BI__builtin_neon_vtbl3_v: 4336 case NEON::BI__builtin_neon_vqtbl3_v: 4337 case NEON::BI__builtin_neon_vqtbl3q_v: 4338 case NEON::BI__builtin_neon_vtbl4_v: 4339 case NEON::BI__builtin_neon_vqtbl4_v: 4340 case NEON::BI__builtin_neon_vqtbl4q_v: 4341 break; 4342 case NEON::BI__builtin_neon_vtbx1_v: 4343 case NEON::BI__builtin_neon_vqtbx1_v: 4344 case NEON::BI__builtin_neon_vqtbx1q_v: 4345 case NEON::BI__builtin_neon_vtbx2_v: 4346 case NEON::BI__builtin_neon_vqtbx2_v: 4347 case NEON::BI__builtin_neon_vqtbx2q_v: 4348 case NEON::BI__builtin_neon_vtbx3_v: 4349 case NEON::BI__builtin_neon_vqtbx3_v: 4350 case NEON::BI__builtin_neon_vqtbx3q_v: 4351 case NEON::BI__builtin_neon_vtbx4_v: 4352 case NEON::BI__builtin_neon_vqtbx4_v: 4353 case NEON::BI__builtin_neon_vqtbx4q_v: 4354 break; 4355 } 4356 4357 assert(E->getNumArgs() >= 3); 4358 4359 // Get the last argument, which specifies the vector type. 4360 llvm::APSInt Result; 4361 const Expr *Arg = E->getArg(E->getNumArgs() - 1); 4362 if (!Arg->isIntegerConstantExpr(Result, CGF.getContext())) 4363 return nullptr; 4364 4365 // Determine the type of this overloaded NEON intrinsic. 4366 NeonTypeFlags Type(Result.getZExtValue()); 4367 llvm::VectorType *Ty = GetNeonType(&CGF, Type); 4368 if (!Ty) 4369 return nullptr; 4370 4371 CodeGen::CGBuilderTy &Builder = CGF.Builder; 4372 4373 // AArch64 scalar builtins are not overloaded, they do not have an extra 4374 // argument that specifies the vector type, need to handle each case. 4375 switch (BuiltinID) { 4376 case NEON::BI__builtin_neon_vtbl1_v: { 4377 return packTBLDVectorList(CGF, makeArrayRef(Ops).slice(0, 1), nullptr, 4378 Ops[1], Ty, Intrinsic::aarch64_neon_tbl1, 4379 "vtbl1"); 4380 } 4381 case NEON::BI__builtin_neon_vtbl2_v: { 4382 return packTBLDVectorList(CGF, makeArrayRef(Ops).slice(0, 2), nullptr, 4383 Ops[2], Ty, Intrinsic::aarch64_neon_tbl1, 4384 "vtbl1"); 4385 } 4386 case NEON::BI__builtin_neon_vtbl3_v: { 4387 return packTBLDVectorList(CGF, makeArrayRef(Ops).slice(0, 3), nullptr, 4388 Ops[3], Ty, Intrinsic::aarch64_neon_tbl2, 4389 "vtbl2"); 4390 } 4391 case NEON::BI__builtin_neon_vtbl4_v: { 4392 return packTBLDVectorList(CGF, makeArrayRef(Ops).slice(0, 4), nullptr, 4393 Ops[4], Ty, Intrinsic::aarch64_neon_tbl2, 4394 "vtbl2"); 4395 } 4396 case NEON::BI__builtin_neon_vtbx1_v: { 4397 Value *TblRes = 4398 packTBLDVectorList(CGF, makeArrayRef(Ops).slice(1, 1), nullptr, Ops[2], 4399 Ty, Intrinsic::aarch64_neon_tbl1, "vtbl1"); 4400 4401 llvm::Constant *EightV = ConstantInt::get(Ty, 8); 4402 Value *CmpRes = Builder.CreateICmp(ICmpInst::ICMP_UGE, Ops[2], EightV); 4403 CmpRes = Builder.CreateSExt(CmpRes, Ty); 4404 4405 Value *EltsFromInput = Builder.CreateAnd(CmpRes, Ops[0]); 4406 Value *EltsFromTbl = Builder.CreateAnd(Builder.CreateNot(CmpRes), TblRes); 4407 return Builder.CreateOr(EltsFromInput, EltsFromTbl, "vtbx"); 4408 } 4409 case NEON::BI__builtin_neon_vtbx2_v: { 4410 return packTBLDVectorList(CGF, makeArrayRef(Ops).slice(1, 2), Ops[0], 4411 Ops[3], Ty, Intrinsic::aarch64_neon_tbx1, 4412 "vtbx1"); 4413 } 4414 case NEON::BI__builtin_neon_vtbx3_v: { 4415 Value *TblRes = 4416 packTBLDVectorList(CGF, makeArrayRef(Ops).slice(1, 3), nullptr, Ops[4], 4417 Ty, Intrinsic::aarch64_neon_tbl2, "vtbl2"); 4418 4419 llvm::Constant *TwentyFourV = ConstantInt::get(Ty, 24); 4420 Value *CmpRes = Builder.CreateICmp(ICmpInst::ICMP_UGE, Ops[4], 4421 TwentyFourV); 4422 CmpRes = Builder.CreateSExt(CmpRes, Ty); 4423 4424 Value *EltsFromInput = Builder.CreateAnd(CmpRes, Ops[0]); 4425 Value *EltsFromTbl = Builder.CreateAnd(Builder.CreateNot(CmpRes), TblRes); 4426 return Builder.CreateOr(EltsFromInput, EltsFromTbl, "vtbx"); 4427 } 4428 case NEON::BI__builtin_neon_vtbx4_v: { 4429 return packTBLDVectorList(CGF, makeArrayRef(Ops).slice(1, 4), Ops[0], 4430 Ops[5], Ty, Intrinsic::aarch64_neon_tbx2, 4431 "vtbx2"); 4432 } 4433 case NEON::BI__builtin_neon_vqtbl1_v: 4434 case NEON::BI__builtin_neon_vqtbl1q_v: 4435 Int = Intrinsic::aarch64_neon_tbl1; s = "vtbl1"; break; 4436 case NEON::BI__builtin_neon_vqtbl2_v: 4437 case NEON::BI__builtin_neon_vqtbl2q_v: { 4438 Int = Intrinsic::aarch64_neon_tbl2; s = "vtbl2"; break; 4439 case NEON::BI__builtin_neon_vqtbl3_v: 4440 case NEON::BI__builtin_neon_vqtbl3q_v: 4441 Int = Intrinsic::aarch64_neon_tbl3; s = "vtbl3"; break; 4442 case NEON::BI__builtin_neon_vqtbl4_v: 4443 case NEON::BI__builtin_neon_vqtbl4q_v: 4444 Int = Intrinsic::aarch64_neon_tbl4; s = "vtbl4"; break; 4445 case NEON::BI__builtin_neon_vqtbx1_v: 4446 case NEON::BI__builtin_neon_vqtbx1q_v: 4447 Int = Intrinsic::aarch64_neon_tbx1; s = "vtbx1"; break; 4448 case NEON::BI__builtin_neon_vqtbx2_v: 4449 case NEON::BI__builtin_neon_vqtbx2q_v: 4450 Int = Intrinsic::aarch64_neon_tbx2; s = "vtbx2"; break; 4451 case NEON::BI__builtin_neon_vqtbx3_v: 4452 case NEON::BI__builtin_neon_vqtbx3q_v: 4453 Int = Intrinsic::aarch64_neon_tbx3; s = "vtbx3"; break; 4454 case NEON::BI__builtin_neon_vqtbx4_v: 4455 case NEON::BI__builtin_neon_vqtbx4q_v: 4456 Int = Intrinsic::aarch64_neon_tbx4; s = "vtbx4"; break; 4457 } 4458 } 4459 4460 if (!Int) 4461 return nullptr; 4462 4463 Function *F = CGF.CGM.getIntrinsic(Int, Ty); 4464 return CGF.EmitNeonCall(F, Ops, s); 4465 } 4466 4467 Value *CodeGenFunction::vectorWrapScalar16(Value *Op) { 4468 llvm::Type *VTy = llvm::VectorType::get(Int16Ty, 4); 4469 Op = Builder.CreateBitCast(Op, Int16Ty); 4470 Value *V = UndefValue::get(VTy); 4471 llvm::Constant *CI = ConstantInt::get(SizeTy, 0); 4472 Op = Builder.CreateInsertElement(V, Op, CI); 4473 return Op; 4474 } 4475 4476 Value *CodeGenFunction::EmitAArch64BuiltinExpr(unsigned BuiltinID, 4477 const CallExpr *E) { 4478 unsigned HintID = static_cast<unsigned>(-1); 4479 switch (BuiltinID) { 4480 default: break; 4481 case AArch64::BI__builtin_arm_nop: 4482 HintID = 0; 4483 break; 4484 case AArch64::BI__builtin_arm_yield: 4485 HintID = 1; 4486 break; 4487 case AArch64::BI__builtin_arm_wfe: 4488 HintID = 2; 4489 break; 4490 case AArch64::BI__builtin_arm_wfi: 4491 HintID = 3; 4492 break; 4493 case AArch64::BI__builtin_arm_sev: 4494 HintID = 4; 4495 break; 4496 case AArch64::BI__builtin_arm_sevl: 4497 HintID = 5; 4498 break; 4499 } 4500 4501 if (HintID != static_cast<unsigned>(-1)) { 4502 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_hint); 4503 return Builder.CreateCall(F, llvm::ConstantInt::get(Int32Ty, HintID)); 4504 } 4505 4506 if (BuiltinID == AArch64::BI__builtin_arm_prefetch) { 4507 Value *Address = EmitScalarExpr(E->getArg(0)); 4508 Value *RW = EmitScalarExpr(E->getArg(1)); 4509 Value *CacheLevel = EmitScalarExpr(E->getArg(2)); 4510 Value *RetentionPolicy = EmitScalarExpr(E->getArg(3)); 4511 Value *IsData = EmitScalarExpr(E->getArg(4)); 4512 4513 Value *Locality = nullptr; 4514 if (cast<llvm::ConstantInt>(RetentionPolicy)->isZero()) { 4515 // Temporal fetch, needs to convert cache level to locality. 4516 Locality = llvm::ConstantInt::get(Int32Ty, 4517 -cast<llvm::ConstantInt>(CacheLevel)->getValue() + 3); 4518 } else { 4519 // Streaming fetch. 4520 Locality = llvm::ConstantInt::get(Int32Ty, 0); 4521 } 4522 4523 // FIXME: We need AArch64 specific LLVM intrinsic if we want to specify 4524 // PLDL3STRM or PLDL2STRM. 4525 Value *F = CGM.getIntrinsic(Intrinsic::prefetch); 4526 return Builder.CreateCall(F, {Address, RW, Locality, IsData}); 4527 } 4528 4529 if (BuiltinID == AArch64::BI__builtin_arm_rbit) { 4530 assert((getContext().getTypeSize(E->getType()) == 32) && 4531 "rbit of unusual size!"); 4532 llvm::Value *Arg = EmitScalarExpr(E->getArg(0)); 4533 return Builder.CreateCall( 4534 CGM.getIntrinsic(Intrinsic::aarch64_rbit, Arg->getType()), Arg, "rbit"); 4535 } 4536 if (BuiltinID == AArch64::BI__builtin_arm_rbit64) { 4537 assert((getContext().getTypeSize(E->getType()) == 64) && 4538 "rbit of unusual size!"); 4539 llvm::Value *Arg = EmitScalarExpr(E->getArg(0)); 4540 return Builder.CreateCall( 4541 CGM.getIntrinsic(Intrinsic::aarch64_rbit, Arg->getType()), Arg, "rbit"); 4542 } 4543 4544 if (BuiltinID == AArch64::BI__clear_cache) { 4545 assert(E->getNumArgs() == 2 && "__clear_cache takes 2 arguments"); 4546 const FunctionDecl *FD = E->getDirectCallee(); 4547 Value *Ops[2]; 4548 for (unsigned i = 0; i < 2; i++) 4549 Ops[i] = EmitScalarExpr(E->getArg(i)); 4550 llvm::Type *Ty = CGM.getTypes().ConvertType(FD->getType()); 4551 llvm::FunctionType *FTy = cast<llvm::FunctionType>(Ty); 4552 StringRef Name = FD->getName(); 4553 return EmitNounwindRuntimeCall(CGM.CreateRuntimeFunction(FTy, Name), Ops); 4554 } 4555 4556 if ((BuiltinID == AArch64::BI__builtin_arm_ldrex || 4557 BuiltinID == AArch64::BI__builtin_arm_ldaex) && 4558 getContext().getTypeSize(E->getType()) == 128) { 4559 Function *F = CGM.getIntrinsic(BuiltinID == AArch64::BI__builtin_arm_ldaex 4560 ? Intrinsic::aarch64_ldaxp 4561 : Intrinsic::aarch64_ldxp); 4562 4563 Value *LdPtr = EmitScalarExpr(E->getArg(0)); 4564 Value *Val = Builder.CreateCall(F, Builder.CreateBitCast(LdPtr, Int8PtrTy), 4565 "ldxp"); 4566 4567 Value *Val0 = Builder.CreateExtractValue(Val, 1); 4568 Value *Val1 = Builder.CreateExtractValue(Val, 0); 4569 llvm::Type *Int128Ty = llvm::IntegerType::get(getLLVMContext(), 128); 4570 Val0 = Builder.CreateZExt(Val0, Int128Ty); 4571 Val1 = Builder.CreateZExt(Val1, Int128Ty); 4572 4573 Value *ShiftCst = llvm::ConstantInt::get(Int128Ty, 64); 4574 Val = Builder.CreateShl(Val0, ShiftCst, "shl", true /* nuw */); 4575 Val = Builder.CreateOr(Val, Val1); 4576 return Builder.CreateBitCast(Val, ConvertType(E->getType())); 4577 } else if (BuiltinID == AArch64::BI__builtin_arm_ldrex || 4578 BuiltinID == AArch64::BI__builtin_arm_ldaex) { 4579 Value *LoadAddr = EmitScalarExpr(E->getArg(0)); 4580 4581 QualType Ty = E->getType(); 4582 llvm::Type *RealResTy = ConvertType(Ty); 4583 llvm::Type *IntResTy = llvm::IntegerType::get(getLLVMContext(), 4584 getContext().getTypeSize(Ty)); 4585 LoadAddr = Builder.CreateBitCast(LoadAddr, IntResTy->getPointerTo()); 4586 4587 Function *F = CGM.getIntrinsic(BuiltinID == AArch64::BI__builtin_arm_ldaex 4588 ? Intrinsic::aarch64_ldaxr 4589 : Intrinsic::aarch64_ldxr, 4590 LoadAddr->getType()); 4591 Value *Val = Builder.CreateCall(F, LoadAddr, "ldxr"); 4592 4593 if (RealResTy->isPointerTy()) 4594 return Builder.CreateIntToPtr(Val, RealResTy); 4595 4596 Val = Builder.CreateTruncOrBitCast(Val, IntResTy); 4597 return Builder.CreateBitCast(Val, RealResTy); 4598 } 4599 4600 if ((BuiltinID == AArch64::BI__builtin_arm_strex || 4601 BuiltinID == AArch64::BI__builtin_arm_stlex) && 4602 getContext().getTypeSize(E->getArg(0)->getType()) == 128) { 4603 Function *F = CGM.getIntrinsic(BuiltinID == AArch64::BI__builtin_arm_stlex 4604 ? Intrinsic::aarch64_stlxp 4605 : Intrinsic::aarch64_stxp); 4606 llvm::Type *STy = llvm::StructType::get(Int64Ty, Int64Ty, nullptr); 4607 4608 Address Tmp = CreateMemTemp(E->getArg(0)->getType()); 4609 EmitAnyExprToMem(E->getArg(0), Tmp, Qualifiers(), /*init*/ true); 4610 4611 Tmp = Builder.CreateBitCast(Tmp, llvm::PointerType::getUnqual(STy)); 4612 llvm::Value *Val = Builder.CreateLoad(Tmp); 4613 4614 Value *Arg0 = Builder.CreateExtractValue(Val, 0); 4615 Value *Arg1 = Builder.CreateExtractValue(Val, 1); 4616 Value *StPtr = Builder.CreateBitCast(EmitScalarExpr(E->getArg(1)), 4617 Int8PtrTy); 4618 return Builder.CreateCall(F, {Arg0, Arg1, StPtr}, "stxp"); 4619 } 4620 4621 if (BuiltinID == AArch64::BI__builtin_arm_strex || 4622 BuiltinID == AArch64::BI__builtin_arm_stlex) { 4623 Value *StoreVal = EmitScalarExpr(E->getArg(0)); 4624 Value *StoreAddr = EmitScalarExpr(E->getArg(1)); 4625 4626 QualType Ty = E->getArg(0)->getType(); 4627 llvm::Type *StoreTy = llvm::IntegerType::get(getLLVMContext(), 4628 getContext().getTypeSize(Ty)); 4629 StoreAddr = Builder.CreateBitCast(StoreAddr, StoreTy->getPointerTo()); 4630 4631 if (StoreVal->getType()->isPointerTy()) 4632 StoreVal = Builder.CreatePtrToInt(StoreVal, Int64Ty); 4633 else { 4634 StoreVal = Builder.CreateBitCast(StoreVal, StoreTy); 4635 StoreVal = Builder.CreateZExtOrBitCast(StoreVal, Int64Ty); 4636 } 4637 4638 Function *F = CGM.getIntrinsic(BuiltinID == AArch64::BI__builtin_arm_stlex 4639 ? Intrinsic::aarch64_stlxr 4640 : Intrinsic::aarch64_stxr, 4641 StoreAddr->getType()); 4642 return Builder.CreateCall(F, {StoreVal, StoreAddr}, "stxr"); 4643 } 4644 4645 if (BuiltinID == AArch64::BI__builtin_arm_clrex) { 4646 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_clrex); 4647 return Builder.CreateCall(F); 4648 } 4649 4650 // CRC32 4651 Intrinsic::ID CRCIntrinsicID = Intrinsic::not_intrinsic; 4652 switch (BuiltinID) { 4653 case AArch64::BI__builtin_arm_crc32b: 4654 CRCIntrinsicID = Intrinsic::aarch64_crc32b; break; 4655 case AArch64::BI__builtin_arm_crc32cb: 4656 CRCIntrinsicID = Intrinsic::aarch64_crc32cb; break; 4657 case AArch64::BI__builtin_arm_crc32h: 4658 CRCIntrinsicID = Intrinsic::aarch64_crc32h; break; 4659 case AArch64::BI__builtin_arm_crc32ch: 4660 CRCIntrinsicID = Intrinsic::aarch64_crc32ch; break; 4661 case AArch64::BI__builtin_arm_crc32w: 4662 CRCIntrinsicID = Intrinsic::aarch64_crc32w; break; 4663 case AArch64::BI__builtin_arm_crc32cw: 4664 CRCIntrinsicID = Intrinsic::aarch64_crc32cw; break; 4665 case AArch64::BI__builtin_arm_crc32d: 4666 CRCIntrinsicID = Intrinsic::aarch64_crc32x; break; 4667 case AArch64::BI__builtin_arm_crc32cd: 4668 CRCIntrinsicID = Intrinsic::aarch64_crc32cx; break; 4669 } 4670 4671 if (CRCIntrinsicID != Intrinsic::not_intrinsic) { 4672 Value *Arg0 = EmitScalarExpr(E->getArg(0)); 4673 Value *Arg1 = EmitScalarExpr(E->getArg(1)); 4674 Function *F = CGM.getIntrinsic(CRCIntrinsicID); 4675 4676 llvm::Type *DataTy = F->getFunctionType()->getParamType(1); 4677 Arg1 = Builder.CreateZExtOrBitCast(Arg1, DataTy); 4678 4679 return Builder.CreateCall(F, {Arg0, Arg1}); 4680 } 4681 4682 if (BuiltinID == AArch64::BI__builtin_arm_rsr || 4683 BuiltinID == AArch64::BI__builtin_arm_rsr64 || 4684 BuiltinID == AArch64::BI__builtin_arm_rsrp || 4685 BuiltinID == AArch64::BI__builtin_arm_wsr || 4686 BuiltinID == AArch64::BI__builtin_arm_wsr64 || 4687 BuiltinID == AArch64::BI__builtin_arm_wsrp) { 4688 4689 bool IsRead = BuiltinID == AArch64::BI__builtin_arm_rsr || 4690 BuiltinID == AArch64::BI__builtin_arm_rsr64 || 4691 BuiltinID == AArch64::BI__builtin_arm_rsrp; 4692 4693 bool IsPointerBuiltin = BuiltinID == AArch64::BI__builtin_arm_rsrp || 4694 BuiltinID == AArch64::BI__builtin_arm_wsrp; 4695 4696 bool Is64Bit = BuiltinID != AArch64::BI__builtin_arm_rsr && 4697 BuiltinID != AArch64::BI__builtin_arm_wsr; 4698 4699 llvm::Type *ValueType; 4700 llvm::Type *RegisterType = Int64Ty; 4701 if (IsPointerBuiltin) { 4702 ValueType = VoidPtrTy; 4703 } else if (Is64Bit) { 4704 ValueType = Int64Ty; 4705 } else { 4706 ValueType = Int32Ty; 4707 } 4708 4709 return EmitSpecialRegisterBuiltin(*this, E, RegisterType, ValueType, IsRead); 4710 } 4711 4712 // Find out if any arguments are required to be integer constant 4713 // expressions. 4714 unsigned ICEArguments = 0; 4715 ASTContext::GetBuiltinTypeError Error; 4716 getContext().GetBuiltinType(BuiltinID, Error, &ICEArguments); 4717 assert(Error == ASTContext::GE_None && "Should not codegen an error"); 4718 4719 llvm::SmallVector<Value*, 4> Ops; 4720 for (unsigned i = 0, e = E->getNumArgs() - 1; i != e; i++) { 4721 if ((ICEArguments & (1 << i)) == 0) { 4722 Ops.push_back(EmitScalarExpr(E->getArg(i))); 4723 } else { 4724 // If this is required to be a constant, constant fold it so that we know 4725 // that the generated intrinsic gets a ConstantInt. 4726 llvm::APSInt Result; 4727 bool IsConst = E->getArg(i)->isIntegerConstantExpr(Result, getContext()); 4728 assert(IsConst && "Constant arg isn't actually constant?"); 4729 (void)IsConst; 4730 Ops.push_back(llvm::ConstantInt::get(getLLVMContext(), Result)); 4731 } 4732 } 4733 4734 auto SISDMap = makeArrayRef(AArch64SISDIntrinsicMap); 4735 const NeonIntrinsicInfo *Builtin = findNeonIntrinsicInMap( 4736 SISDMap, BuiltinID, AArch64SISDIntrinsicsProvenSorted); 4737 4738 if (Builtin) { 4739 Ops.push_back(EmitScalarExpr(E->getArg(E->getNumArgs() - 1))); 4740 Value *Result = EmitCommonNeonSISDBuiltinExpr(*this, *Builtin, Ops, E); 4741 assert(Result && "SISD intrinsic should have been handled"); 4742 return Result; 4743 } 4744 4745 llvm::APSInt Result; 4746 const Expr *Arg = E->getArg(E->getNumArgs()-1); 4747 NeonTypeFlags Type(0); 4748 if (Arg->isIntegerConstantExpr(Result, getContext())) 4749 // Determine the type of this overloaded NEON intrinsic. 4750 Type = NeonTypeFlags(Result.getZExtValue()); 4751 4752 bool usgn = Type.isUnsigned(); 4753 bool quad = Type.isQuad(); 4754 4755 // Handle non-overloaded intrinsics first. 4756 switch (BuiltinID) { 4757 default: break; 4758 case NEON::BI__builtin_neon_vldrq_p128: { 4759 llvm::Type *Int128PTy = llvm::Type::getIntNPtrTy(getLLVMContext(), 128); 4760 Value *Ptr = Builder.CreateBitCast(EmitScalarExpr(E->getArg(0)), Int128PTy); 4761 return Builder.CreateDefaultAlignedLoad(Ptr); 4762 } 4763 case NEON::BI__builtin_neon_vstrq_p128: { 4764 llvm::Type *Int128PTy = llvm::Type::getIntNPtrTy(getLLVMContext(), 128); 4765 Value *Ptr = Builder.CreateBitCast(Ops[0], Int128PTy); 4766 return Builder.CreateDefaultAlignedStore(EmitScalarExpr(E->getArg(1)), Ptr); 4767 } 4768 case NEON::BI__builtin_neon_vcvts_u32_f32: 4769 case NEON::BI__builtin_neon_vcvtd_u64_f64: 4770 usgn = true; 4771 // FALL THROUGH 4772 case NEON::BI__builtin_neon_vcvts_s32_f32: 4773 case NEON::BI__builtin_neon_vcvtd_s64_f64: { 4774 Ops.push_back(EmitScalarExpr(E->getArg(0))); 4775 bool Is64 = Ops[0]->getType()->getPrimitiveSizeInBits() == 64; 4776 llvm::Type *InTy = Is64 ? Int64Ty : Int32Ty; 4777 llvm::Type *FTy = Is64 ? DoubleTy : FloatTy; 4778 Ops[0] = Builder.CreateBitCast(Ops[0], FTy); 4779 if (usgn) 4780 return Builder.CreateFPToUI(Ops[0], InTy); 4781 return Builder.CreateFPToSI(Ops[0], InTy); 4782 } 4783 case NEON::BI__builtin_neon_vcvts_f32_u32: 4784 case NEON::BI__builtin_neon_vcvtd_f64_u64: 4785 usgn = true; 4786 // FALL THROUGH 4787 case NEON::BI__builtin_neon_vcvts_f32_s32: 4788 case NEON::BI__builtin_neon_vcvtd_f64_s64: { 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], InTy); 4794 if (usgn) 4795 return Builder.CreateUIToFP(Ops[0], FTy); 4796 return Builder.CreateSIToFP(Ops[0], FTy); 4797 } 4798 case NEON::BI__builtin_neon_vpaddd_s64: { 4799 llvm::Type *Ty = llvm::VectorType::get(Int64Ty, 2); 4800 Value *Vec = EmitScalarExpr(E->getArg(0)); 4801 // The vector is v2f64, so make sure it's bitcast to that. 4802 Vec = Builder.CreateBitCast(Vec, Ty, "v2i64"); 4803 llvm::Value *Idx0 = llvm::ConstantInt::get(SizeTy, 0); 4804 llvm::Value *Idx1 = llvm::ConstantInt::get(SizeTy, 1); 4805 Value *Op0 = Builder.CreateExtractElement(Vec, Idx0, "lane0"); 4806 Value *Op1 = Builder.CreateExtractElement(Vec, Idx1, "lane1"); 4807 // Pairwise addition of a v2f64 into a scalar f64. 4808 return Builder.CreateAdd(Op0, Op1, "vpaddd"); 4809 } 4810 case NEON::BI__builtin_neon_vpaddd_f64: { 4811 llvm::Type *Ty = 4812 llvm::VectorType::get(DoubleTy, 2); 4813 Value *Vec = EmitScalarExpr(E->getArg(0)); 4814 // The vector is v2f64, so make sure it's bitcast to that. 4815 Vec = Builder.CreateBitCast(Vec, Ty, "v2f64"); 4816 llvm::Value *Idx0 = llvm::ConstantInt::get(SizeTy, 0); 4817 llvm::Value *Idx1 = llvm::ConstantInt::get(SizeTy, 1); 4818 Value *Op0 = Builder.CreateExtractElement(Vec, Idx0, "lane0"); 4819 Value *Op1 = Builder.CreateExtractElement(Vec, Idx1, "lane1"); 4820 // Pairwise addition of a v2f64 into a scalar f64. 4821 return Builder.CreateFAdd(Op0, Op1, "vpaddd"); 4822 } 4823 case NEON::BI__builtin_neon_vpadds_f32: { 4824 llvm::Type *Ty = 4825 llvm::VectorType::get(FloatTy, 2); 4826 Value *Vec = EmitScalarExpr(E->getArg(0)); 4827 // The vector is v2f32, so make sure it's bitcast to that. 4828 Vec = Builder.CreateBitCast(Vec, Ty, "v2f32"); 4829 llvm::Value *Idx0 = llvm::ConstantInt::get(SizeTy, 0); 4830 llvm::Value *Idx1 = llvm::ConstantInt::get(SizeTy, 1); 4831 Value *Op0 = Builder.CreateExtractElement(Vec, Idx0, "lane0"); 4832 Value *Op1 = Builder.CreateExtractElement(Vec, Idx1, "lane1"); 4833 // Pairwise addition of a v2f32 into a scalar f32. 4834 return Builder.CreateFAdd(Op0, Op1, "vpaddd"); 4835 } 4836 case NEON::BI__builtin_neon_vceqzd_s64: 4837 case NEON::BI__builtin_neon_vceqzd_f64: 4838 case NEON::BI__builtin_neon_vceqzs_f32: 4839 Ops.push_back(EmitScalarExpr(E->getArg(0))); 4840 return EmitAArch64CompareBuiltinExpr( 4841 Ops[0], ConvertType(E->getCallReturnType(getContext())), 4842 ICmpInst::FCMP_OEQ, ICmpInst::ICMP_EQ, "vceqz"); 4843 case NEON::BI__builtin_neon_vcgezd_s64: 4844 case NEON::BI__builtin_neon_vcgezd_f64: 4845 case NEON::BI__builtin_neon_vcgezs_f32: 4846 Ops.push_back(EmitScalarExpr(E->getArg(0))); 4847 return EmitAArch64CompareBuiltinExpr( 4848 Ops[0], ConvertType(E->getCallReturnType(getContext())), 4849 ICmpInst::FCMP_OGE, ICmpInst::ICMP_SGE, "vcgez"); 4850 case NEON::BI__builtin_neon_vclezd_s64: 4851 case NEON::BI__builtin_neon_vclezd_f64: 4852 case NEON::BI__builtin_neon_vclezs_f32: 4853 Ops.push_back(EmitScalarExpr(E->getArg(0))); 4854 return EmitAArch64CompareBuiltinExpr( 4855 Ops[0], ConvertType(E->getCallReturnType(getContext())), 4856 ICmpInst::FCMP_OLE, ICmpInst::ICMP_SLE, "vclez"); 4857 case NEON::BI__builtin_neon_vcgtzd_s64: 4858 case NEON::BI__builtin_neon_vcgtzd_f64: 4859 case NEON::BI__builtin_neon_vcgtzs_f32: 4860 Ops.push_back(EmitScalarExpr(E->getArg(0))); 4861 return EmitAArch64CompareBuiltinExpr( 4862 Ops[0], ConvertType(E->getCallReturnType(getContext())), 4863 ICmpInst::FCMP_OGT, ICmpInst::ICMP_SGT, "vcgtz"); 4864 case NEON::BI__builtin_neon_vcltzd_s64: 4865 case NEON::BI__builtin_neon_vcltzd_f64: 4866 case NEON::BI__builtin_neon_vcltzs_f32: 4867 Ops.push_back(EmitScalarExpr(E->getArg(0))); 4868 return EmitAArch64CompareBuiltinExpr( 4869 Ops[0], ConvertType(E->getCallReturnType(getContext())), 4870 ICmpInst::FCMP_OLT, ICmpInst::ICMP_SLT, "vcltz"); 4871 4872 case NEON::BI__builtin_neon_vceqzd_u64: { 4873 Ops.push_back(EmitScalarExpr(E->getArg(0))); 4874 Ops[0] = Builder.CreateBitCast(Ops[0], Int64Ty); 4875 Ops[0] = 4876 Builder.CreateICmpEQ(Ops[0], llvm::Constant::getNullValue(Int64Ty)); 4877 return Builder.CreateSExt(Ops[0], Int64Ty, "vceqzd"); 4878 } 4879 case NEON::BI__builtin_neon_vceqd_f64: 4880 case NEON::BI__builtin_neon_vcled_f64: 4881 case NEON::BI__builtin_neon_vcltd_f64: 4882 case NEON::BI__builtin_neon_vcged_f64: 4883 case NEON::BI__builtin_neon_vcgtd_f64: { 4884 llvm::CmpInst::Predicate P; 4885 switch (BuiltinID) { 4886 default: llvm_unreachable("missing builtin ID in switch!"); 4887 case NEON::BI__builtin_neon_vceqd_f64: P = llvm::FCmpInst::FCMP_OEQ; break; 4888 case NEON::BI__builtin_neon_vcled_f64: P = llvm::FCmpInst::FCMP_OLE; break; 4889 case NEON::BI__builtin_neon_vcltd_f64: P = llvm::FCmpInst::FCMP_OLT; break; 4890 case NEON::BI__builtin_neon_vcged_f64: P = llvm::FCmpInst::FCMP_OGE; break; 4891 case NEON::BI__builtin_neon_vcgtd_f64: P = llvm::FCmpInst::FCMP_OGT; break; 4892 } 4893 Ops.push_back(EmitScalarExpr(E->getArg(1))); 4894 Ops[0] = Builder.CreateBitCast(Ops[0], DoubleTy); 4895 Ops[1] = Builder.CreateBitCast(Ops[1], DoubleTy); 4896 Ops[0] = Builder.CreateFCmp(P, Ops[0], Ops[1]); 4897 return Builder.CreateSExt(Ops[0], Int64Ty, "vcmpd"); 4898 } 4899 case NEON::BI__builtin_neon_vceqs_f32: 4900 case NEON::BI__builtin_neon_vcles_f32: 4901 case NEON::BI__builtin_neon_vclts_f32: 4902 case NEON::BI__builtin_neon_vcges_f32: 4903 case NEON::BI__builtin_neon_vcgts_f32: { 4904 llvm::CmpInst::Predicate P; 4905 switch (BuiltinID) { 4906 default: llvm_unreachable("missing builtin ID in switch!"); 4907 case NEON::BI__builtin_neon_vceqs_f32: P = llvm::FCmpInst::FCMP_OEQ; break; 4908 case NEON::BI__builtin_neon_vcles_f32: P = llvm::FCmpInst::FCMP_OLE; break; 4909 case NEON::BI__builtin_neon_vclts_f32: P = llvm::FCmpInst::FCMP_OLT; break; 4910 case NEON::BI__builtin_neon_vcges_f32: P = llvm::FCmpInst::FCMP_OGE; break; 4911 case NEON::BI__builtin_neon_vcgts_f32: P = llvm::FCmpInst::FCMP_OGT; break; 4912 } 4913 Ops.push_back(EmitScalarExpr(E->getArg(1))); 4914 Ops[0] = Builder.CreateBitCast(Ops[0], FloatTy); 4915 Ops[1] = Builder.CreateBitCast(Ops[1], FloatTy); 4916 Ops[0] = Builder.CreateFCmp(P, Ops[0], Ops[1]); 4917 return Builder.CreateSExt(Ops[0], Int32Ty, "vcmpd"); 4918 } 4919 case NEON::BI__builtin_neon_vceqd_s64: 4920 case NEON::BI__builtin_neon_vceqd_u64: 4921 case NEON::BI__builtin_neon_vcgtd_s64: 4922 case NEON::BI__builtin_neon_vcgtd_u64: 4923 case NEON::BI__builtin_neon_vcltd_s64: 4924 case NEON::BI__builtin_neon_vcltd_u64: 4925 case NEON::BI__builtin_neon_vcged_u64: 4926 case NEON::BI__builtin_neon_vcged_s64: 4927 case NEON::BI__builtin_neon_vcled_u64: 4928 case NEON::BI__builtin_neon_vcled_s64: { 4929 llvm::CmpInst::Predicate P; 4930 switch (BuiltinID) { 4931 default: llvm_unreachable("missing builtin ID in switch!"); 4932 case NEON::BI__builtin_neon_vceqd_s64: 4933 case NEON::BI__builtin_neon_vceqd_u64:P = llvm::ICmpInst::ICMP_EQ;break; 4934 case NEON::BI__builtin_neon_vcgtd_s64:P = llvm::ICmpInst::ICMP_SGT;break; 4935 case NEON::BI__builtin_neon_vcgtd_u64:P = llvm::ICmpInst::ICMP_UGT;break; 4936 case NEON::BI__builtin_neon_vcltd_s64:P = llvm::ICmpInst::ICMP_SLT;break; 4937 case NEON::BI__builtin_neon_vcltd_u64:P = llvm::ICmpInst::ICMP_ULT;break; 4938 case NEON::BI__builtin_neon_vcged_u64:P = llvm::ICmpInst::ICMP_UGE;break; 4939 case NEON::BI__builtin_neon_vcged_s64:P = llvm::ICmpInst::ICMP_SGE;break; 4940 case NEON::BI__builtin_neon_vcled_u64:P = llvm::ICmpInst::ICMP_ULE;break; 4941 case NEON::BI__builtin_neon_vcled_s64:P = llvm::ICmpInst::ICMP_SLE;break; 4942 } 4943 Ops.push_back(EmitScalarExpr(E->getArg(1))); 4944 Ops[0] = Builder.CreateBitCast(Ops[0], Int64Ty); 4945 Ops[1] = Builder.CreateBitCast(Ops[1], Int64Ty); 4946 Ops[0] = Builder.CreateICmp(P, Ops[0], Ops[1]); 4947 return Builder.CreateSExt(Ops[0], Int64Ty, "vceqd"); 4948 } 4949 case NEON::BI__builtin_neon_vtstd_s64: 4950 case NEON::BI__builtin_neon_vtstd_u64: { 4951 Ops.push_back(EmitScalarExpr(E->getArg(1))); 4952 Ops[0] = Builder.CreateBitCast(Ops[0], Int64Ty); 4953 Ops[1] = Builder.CreateBitCast(Ops[1], Int64Ty); 4954 Ops[0] = Builder.CreateAnd(Ops[0], Ops[1]); 4955 Ops[0] = Builder.CreateICmp(ICmpInst::ICMP_NE, Ops[0], 4956 llvm::Constant::getNullValue(Int64Ty)); 4957 return Builder.CreateSExt(Ops[0], Int64Ty, "vtstd"); 4958 } 4959 case NEON::BI__builtin_neon_vset_lane_i8: 4960 case NEON::BI__builtin_neon_vset_lane_i16: 4961 case NEON::BI__builtin_neon_vset_lane_i32: 4962 case NEON::BI__builtin_neon_vset_lane_i64: 4963 case NEON::BI__builtin_neon_vset_lane_f32: 4964 case NEON::BI__builtin_neon_vsetq_lane_i8: 4965 case NEON::BI__builtin_neon_vsetq_lane_i16: 4966 case NEON::BI__builtin_neon_vsetq_lane_i32: 4967 case NEON::BI__builtin_neon_vsetq_lane_i64: 4968 case NEON::BI__builtin_neon_vsetq_lane_f32: 4969 Ops.push_back(EmitScalarExpr(E->getArg(2))); 4970 return Builder.CreateInsertElement(Ops[1], Ops[0], Ops[2], "vset_lane"); 4971 case NEON::BI__builtin_neon_vset_lane_f64: 4972 // The vector type needs a cast for the v1f64 variant. 4973 Ops[1] = Builder.CreateBitCast(Ops[1], 4974 llvm::VectorType::get(DoubleTy, 1)); 4975 Ops.push_back(EmitScalarExpr(E->getArg(2))); 4976 return Builder.CreateInsertElement(Ops[1], Ops[0], Ops[2], "vset_lane"); 4977 case NEON::BI__builtin_neon_vsetq_lane_f64: 4978 // The vector type needs a cast for the v2f64 variant. 4979 Ops[1] = Builder.CreateBitCast(Ops[1], 4980 llvm::VectorType::get(DoubleTy, 2)); 4981 Ops.push_back(EmitScalarExpr(E->getArg(2))); 4982 return Builder.CreateInsertElement(Ops[1], Ops[0], Ops[2], "vset_lane"); 4983 4984 case NEON::BI__builtin_neon_vget_lane_i8: 4985 case NEON::BI__builtin_neon_vdupb_lane_i8: 4986 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::VectorType::get(Int8Ty, 8)); 4987 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4988 "vget_lane"); 4989 case NEON::BI__builtin_neon_vgetq_lane_i8: 4990 case NEON::BI__builtin_neon_vdupb_laneq_i8: 4991 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::VectorType::get(Int8Ty, 16)); 4992 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4993 "vgetq_lane"); 4994 case NEON::BI__builtin_neon_vget_lane_i16: 4995 case NEON::BI__builtin_neon_vduph_lane_i16: 4996 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::VectorType::get(Int16Ty, 4)); 4997 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4998 "vget_lane"); 4999 case NEON::BI__builtin_neon_vgetq_lane_i16: 5000 case NEON::BI__builtin_neon_vduph_laneq_i16: 5001 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::VectorType::get(Int16Ty, 8)); 5002 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 5003 "vgetq_lane"); 5004 case NEON::BI__builtin_neon_vget_lane_i32: 5005 case NEON::BI__builtin_neon_vdups_lane_i32: 5006 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::VectorType::get(Int32Ty, 2)); 5007 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 5008 "vget_lane"); 5009 case NEON::BI__builtin_neon_vdups_lane_f32: 5010 Ops[0] = Builder.CreateBitCast(Ops[0], 5011 llvm::VectorType::get(FloatTy, 2)); 5012 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 5013 "vdups_lane"); 5014 case NEON::BI__builtin_neon_vgetq_lane_i32: 5015 case NEON::BI__builtin_neon_vdups_laneq_i32: 5016 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::VectorType::get(Int32Ty, 4)); 5017 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 5018 "vgetq_lane"); 5019 case NEON::BI__builtin_neon_vget_lane_i64: 5020 case NEON::BI__builtin_neon_vdupd_lane_i64: 5021 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::VectorType::get(Int64Ty, 1)); 5022 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 5023 "vget_lane"); 5024 case NEON::BI__builtin_neon_vdupd_lane_f64: 5025 Ops[0] = Builder.CreateBitCast(Ops[0], 5026 llvm::VectorType::get(DoubleTy, 1)); 5027 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 5028 "vdupd_lane"); 5029 case NEON::BI__builtin_neon_vgetq_lane_i64: 5030 case NEON::BI__builtin_neon_vdupd_laneq_i64: 5031 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::VectorType::get(Int64Ty, 2)); 5032 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 5033 "vgetq_lane"); 5034 case NEON::BI__builtin_neon_vget_lane_f32: 5035 Ops[0] = Builder.CreateBitCast(Ops[0], 5036 llvm::VectorType::get(FloatTy, 2)); 5037 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 5038 "vget_lane"); 5039 case NEON::BI__builtin_neon_vget_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 "vget_lane"); 5044 case NEON::BI__builtin_neon_vgetq_lane_f32: 5045 case NEON::BI__builtin_neon_vdups_laneq_f32: 5046 Ops[0] = Builder.CreateBitCast(Ops[0], 5047 llvm::VectorType::get(FloatTy, 4)); 5048 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 5049 "vgetq_lane"); 5050 case NEON::BI__builtin_neon_vgetq_lane_f64: 5051 case NEON::BI__builtin_neon_vdupd_laneq_f64: 5052 Ops[0] = Builder.CreateBitCast(Ops[0], 5053 llvm::VectorType::get(DoubleTy, 2)); 5054 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 5055 "vgetq_lane"); 5056 case NEON::BI__builtin_neon_vaddd_s64: 5057 case NEON::BI__builtin_neon_vaddd_u64: 5058 return Builder.CreateAdd(Ops[0], EmitScalarExpr(E->getArg(1)), "vaddd"); 5059 case NEON::BI__builtin_neon_vsubd_s64: 5060 case NEON::BI__builtin_neon_vsubd_u64: 5061 return Builder.CreateSub(Ops[0], EmitScalarExpr(E->getArg(1)), "vsubd"); 5062 case NEON::BI__builtin_neon_vqdmlalh_s16: 5063 case NEON::BI__builtin_neon_vqdmlslh_s16: { 5064 SmallVector<Value *, 2> ProductOps; 5065 ProductOps.push_back(vectorWrapScalar16(Ops[1])); 5066 ProductOps.push_back(vectorWrapScalar16(EmitScalarExpr(E->getArg(2)))); 5067 llvm::Type *VTy = llvm::VectorType::get(Int32Ty, 4); 5068 Ops[1] = EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_sqdmull, VTy), 5069 ProductOps, "vqdmlXl"); 5070 Constant *CI = ConstantInt::get(SizeTy, 0); 5071 Ops[1] = Builder.CreateExtractElement(Ops[1], CI, "lane0"); 5072 5073 unsigned AccumInt = BuiltinID == NEON::BI__builtin_neon_vqdmlalh_s16 5074 ? Intrinsic::aarch64_neon_sqadd 5075 : Intrinsic::aarch64_neon_sqsub; 5076 return EmitNeonCall(CGM.getIntrinsic(AccumInt, Int32Ty), Ops, "vqdmlXl"); 5077 } 5078 case NEON::BI__builtin_neon_vqshlud_n_s64: { 5079 Ops.push_back(EmitScalarExpr(E->getArg(1))); 5080 Ops[1] = Builder.CreateZExt(Ops[1], Int64Ty); 5081 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_sqshlu, Int64Ty), 5082 Ops, "vqshlu_n"); 5083 } 5084 case NEON::BI__builtin_neon_vqshld_n_u64: 5085 case NEON::BI__builtin_neon_vqshld_n_s64: { 5086 unsigned Int = BuiltinID == NEON::BI__builtin_neon_vqshld_n_u64 5087 ? Intrinsic::aarch64_neon_uqshl 5088 : Intrinsic::aarch64_neon_sqshl; 5089 Ops.push_back(EmitScalarExpr(E->getArg(1))); 5090 Ops[1] = Builder.CreateZExt(Ops[1], Int64Ty); 5091 return EmitNeonCall(CGM.getIntrinsic(Int, Int64Ty), Ops, "vqshl_n"); 5092 } 5093 case NEON::BI__builtin_neon_vrshrd_n_u64: 5094 case NEON::BI__builtin_neon_vrshrd_n_s64: { 5095 unsigned Int = BuiltinID == NEON::BI__builtin_neon_vrshrd_n_u64 5096 ? Intrinsic::aarch64_neon_urshl 5097 : Intrinsic::aarch64_neon_srshl; 5098 Ops.push_back(EmitScalarExpr(E->getArg(1))); 5099 int SV = cast<ConstantInt>(Ops[1])->getSExtValue(); 5100 Ops[1] = ConstantInt::get(Int64Ty, -SV); 5101 return EmitNeonCall(CGM.getIntrinsic(Int, Int64Ty), Ops, "vrshr_n"); 5102 } 5103 case NEON::BI__builtin_neon_vrsrad_n_u64: 5104 case NEON::BI__builtin_neon_vrsrad_n_s64: { 5105 unsigned Int = BuiltinID == NEON::BI__builtin_neon_vrsrad_n_u64 5106 ? Intrinsic::aarch64_neon_urshl 5107 : Intrinsic::aarch64_neon_srshl; 5108 Ops[1] = Builder.CreateBitCast(Ops[1], Int64Ty); 5109 Ops.push_back(Builder.CreateNeg(EmitScalarExpr(E->getArg(2)))); 5110 Ops[1] = Builder.CreateCall(CGM.getIntrinsic(Int, Int64Ty), 5111 {Ops[1], Builder.CreateSExt(Ops[2], Int64Ty)}); 5112 return Builder.CreateAdd(Ops[0], Builder.CreateBitCast(Ops[1], Int64Ty)); 5113 } 5114 case NEON::BI__builtin_neon_vshld_n_s64: 5115 case NEON::BI__builtin_neon_vshld_n_u64: { 5116 llvm::ConstantInt *Amt = cast<ConstantInt>(EmitScalarExpr(E->getArg(1))); 5117 return Builder.CreateShl( 5118 Ops[0], ConstantInt::get(Int64Ty, Amt->getZExtValue()), "shld_n"); 5119 } 5120 case NEON::BI__builtin_neon_vshrd_n_s64: { 5121 llvm::ConstantInt *Amt = cast<ConstantInt>(EmitScalarExpr(E->getArg(1))); 5122 return Builder.CreateAShr( 5123 Ops[0], ConstantInt::get(Int64Ty, std::min(static_cast<uint64_t>(63), 5124 Amt->getZExtValue())), 5125 "shrd_n"); 5126 } 5127 case NEON::BI__builtin_neon_vshrd_n_u64: { 5128 llvm::ConstantInt *Amt = cast<ConstantInt>(EmitScalarExpr(E->getArg(1))); 5129 uint64_t ShiftAmt = Amt->getZExtValue(); 5130 // Right-shifting an unsigned value by its size yields 0. 5131 if (ShiftAmt == 64) 5132 return ConstantInt::get(Int64Ty, 0); 5133 return Builder.CreateLShr(Ops[0], ConstantInt::get(Int64Ty, ShiftAmt), 5134 "shrd_n"); 5135 } 5136 case NEON::BI__builtin_neon_vsrad_n_s64: { 5137 llvm::ConstantInt *Amt = cast<ConstantInt>(EmitScalarExpr(E->getArg(2))); 5138 Ops[1] = Builder.CreateAShr( 5139 Ops[1], ConstantInt::get(Int64Ty, std::min(static_cast<uint64_t>(63), 5140 Amt->getZExtValue())), 5141 "shrd_n"); 5142 return Builder.CreateAdd(Ops[0], Ops[1]); 5143 } 5144 case NEON::BI__builtin_neon_vsrad_n_u64: { 5145 llvm::ConstantInt *Amt = cast<ConstantInt>(EmitScalarExpr(E->getArg(2))); 5146 uint64_t ShiftAmt = Amt->getZExtValue(); 5147 // Right-shifting an unsigned value by its size yields 0. 5148 // As Op + 0 = Op, return Ops[0] directly. 5149 if (ShiftAmt == 64) 5150 return Ops[0]; 5151 Ops[1] = Builder.CreateLShr(Ops[1], ConstantInt::get(Int64Ty, ShiftAmt), 5152 "shrd_n"); 5153 return Builder.CreateAdd(Ops[0], Ops[1]); 5154 } 5155 case NEON::BI__builtin_neon_vqdmlalh_lane_s16: 5156 case NEON::BI__builtin_neon_vqdmlalh_laneq_s16: 5157 case NEON::BI__builtin_neon_vqdmlslh_lane_s16: 5158 case NEON::BI__builtin_neon_vqdmlslh_laneq_s16: { 5159 Ops[2] = Builder.CreateExtractElement(Ops[2], EmitScalarExpr(E->getArg(3)), 5160 "lane"); 5161 SmallVector<Value *, 2> ProductOps; 5162 ProductOps.push_back(vectorWrapScalar16(Ops[1])); 5163 ProductOps.push_back(vectorWrapScalar16(Ops[2])); 5164 llvm::Type *VTy = llvm::VectorType::get(Int32Ty, 4); 5165 Ops[1] = EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_sqdmull, VTy), 5166 ProductOps, "vqdmlXl"); 5167 Constant *CI = ConstantInt::get(SizeTy, 0); 5168 Ops[1] = Builder.CreateExtractElement(Ops[1], CI, "lane0"); 5169 Ops.pop_back(); 5170 5171 unsigned AccInt = (BuiltinID == NEON::BI__builtin_neon_vqdmlalh_lane_s16 || 5172 BuiltinID == NEON::BI__builtin_neon_vqdmlalh_laneq_s16) 5173 ? Intrinsic::aarch64_neon_sqadd 5174 : Intrinsic::aarch64_neon_sqsub; 5175 return EmitNeonCall(CGM.getIntrinsic(AccInt, Int32Ty), Ops, "vqdmlXl"); 5176 } 5177 case NEON::BI__builtin_neon_vqdmlals_s32: 5178 case NEON::BI__builtin_neon_vqdmlsls_s32: { 5179 SmallVector<Value *, 2> ProductOps; 5180 ProductOps.push_back(Ops[1]); 5181 ProductOps.push_back(EmitScalarExpr(E->getArg(2))); 5182 Ops[1] = 5183 EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_sqdmulls_scalar), 5184 ProductOps, "vqdmlXl"); 5185 5186 unsigned AccumInt = BuiltinID == NEON::BI__builtin_neon_vqdmlals_s32 5187 ? Intrinsic::aarch64_neon_sqadd 5188 : Intrinsic::aarch64_neon_sqsub; 5189 return EmitNeonCall(CGM.getIntrinsic(AccumInt, Int64Ty), Ops, "vqdmlXl"); 5190 } 5191 case NEON::BI__builtin_neon_vqdmlals_lane_s32: 5192 case NEON::BI__builtin_neon_vqdmlals_laneq_s32: 5193 case NEON::BI__builtin_neon_vqdmlsls_lane_s32: 5194 case NEON::BI__builtin_neon_vqdmlsls_laneq_s32: { 5195 Ops[2] = Builder.CreateExtractElement(Ops[2], EmitScalarExpr(E->getArg(3)), 5196 "lane"); 5197 SmallVector<Value *, 2> ProductOps; 5198 ProductOps.push_back(Ops[1]); 5199 ProductOps.push_back(Ops[2]); 5200 Ops[1] = 5201 EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_sqdmulls_scalar), 5202 ProductOps, "vqdmlXl"); 5203 Ops.pop_back(); 5204 5205 unsigned AccInt = (BuiltinID == NEON::BI__builtin_neon_vqdmlals_lane_s32 || 5206 BuiltinID == NEON::BI__builtin_neon_vqdmlals_laneq_s32) 5207 ? Intrinsic::aarch64_neon_sqadd 5208 : Intrinsic::aarch64_neon_sqsub; 5209 return EmitNeonCall(CGM.getIntrinsic(AccInt, Int64Ty), Ops, "vqdmlXl"); 5210 } 5211 } 5212 5213 llvm::VectorType *VTy = GetNeonType(this, Type); 5214 llvm::Type *Ty = VTy; 5215 if (!Ty) 5216 return nullptr; 5217 5218 // Not all intrinsics handled by the common case work for AArch64 yet, so only 5219 // defer to common code if it's been added to our special map. 5220 Builtin = findNeonIntrinsicInMap(AArch64SIMDIntrinsicMap, BuiltinID, 5221 AArch64SIMDIntrinsicsProvenSorted); 5222 5223 if (Builtin) 5224 return EmitCommonNeonBuiltinExpr( 5225 Builtin->BuiltinID, Builtin->LLVMIntrinsic, Builtin->AltLLVMIntrinsic, 5226 Builtin->NameHint, Builtin->TypeModifier, E, Ops, 5227 /*never use addresses*/ Address::invalid(), Address::invalid()); 5228 5229 if (Value *V = EmitAArch64TblBuiltinExpr(*this, BuiltinID, E, Ops)) 5230 return V; 5231 5232 unsigned Int; 5233 switch (BuiltinID) { 5234 default: return nullptr; 5235 case NEON::BI__builtin_neon_vbsl_v: 5236 case NEON::BI__builtin_neon_vbslq_v: { 5237 llvm::Type *BitTy = llvm::VectorType::getInteger(VTy); 5238 Ops[0] = Builder.CreateBitCast(Ops[0], BitTy, "vbsl"); 5239 Ops[1] = Builder.CreateBitCast(Ops[1], BitTy, "vbsl"); 5240 Ops[2] = Builder.CreateBitCast(Ops[2], BitTy, "vbsl"); 5241 5242 Ops[1] = Builder.CreateAnd(Ops[0], Ops[1], "vbsl"); 5243 Ops[2] = Builder.CreateAnd(Builder.CreateNot(Ops[0]), Ops[2], "vbsl"); 5244 Ops[0] = Builder.CreateOr(Ops[1], Ops[2], "vbsl"); 5245 return Builder.CreateBitCast(Ops[0], Ty); 5246 } 5247 case NEON::BI__builtin_neon_vfma_lane_v: 5248 case NEON::BI__builtin_neon_vfmaq_lane_v: { // Only used for FP types 5249 // The ARM builtins (and instructions) have the addend as the first 5250 // operand, but the 'fma' intrinsics have it last. Swap it around here. 5251 Value *Addend = Ops[0]; 5252 Value *Multiplicand = Ops[1]; 5253 Value *LaneSource = Ops[2]; 5254 Ops[0] = Multiplicand; 5255 Ops[1] = LaneSource; 5256 Ops[2] = Addend; 5257 5258 // Now adjust things to handle the lane access. 5259 llvm::Type *SourceTy = BuiltinID == NEON::BI__builtin_neon_vfmaq_lane_v ? 5260 llvm::VectorType::get(VTy->getElementType(), VTy->getNumElements() / 2) : 5261 VTy; 5262 llvm::Constant *cst = cast<Constant>(Ops[3]); 5263 Value *SV = llvm::ConstantVector::getSplat(VTy->getNumElements(), cst); 5264 Ops[1] = Builder.CreateBitCast(Ops[1], SourceTy); 5265 Ops[1] = Builder.CreateShuffleVector(Ops[1], Ops[1], SV, "lane"); 5266 5267 Ops.pop_back(); 5268 Int = Intrinsic::fma; 5269 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "fmla"); 5270 } 5271 case NEON::BI__builtin_neon_vfma_laneq_v: { 5272 llvm::VectorType *VTy = cast<llvm::VectorType>(Ty); 5273 // v1f64 fma should be mapped to Neon scalar f64 fma 5274 if (VTy && VTy->getElementType() == DoubleTy) { 5275 Ops[0] = Builder.CreateBitCast(Ops[0], DoubleTy); 5276 Ops[1] = Builder.CreateBitCast(Ops[1], DoubleTy); 5277 llvm::Type *VTy = GetNeonType(this, 5278 NeonTypeFlags(NeonTypeFlags::Float64, false, true)); 5279 Ops[2] = Builder.CreateBitCast(Ops[2], VTy); 5280 Ops[2] = Builder.CreateExtractElement(Ops[2], Ops[3], "extract"); 5281 Value *F = CGM.getIntrinsic(Intrinsic::fma, DoubleTy); 5282 Value *Result = Builder.CreateCall(F, {Ops[1], Ops[2], Ops[0]}); 5283 return Builder.CreateBitCast(Result, Ty); 5284 } 5285 Value *F = CGM.getIntrinsic(Intrinsic::fma, Ty); 5286 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 5287 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 5288 5289 llvm::Type *STy = llvm::VectorType::get(VTy->getElementType(), 5290 VTy->getNumElements() * 2); 5291 Ops[2] = Builder.CreateBitCast(Ops[2], STy); 5292 Value* SV = llvm::ConstantVector::getSplat(VTy->getNumElements(), 5293 cast<ConstantInt>(Ops[3])); 5294 Ops[2] = Builder.CreateShuffleVector(Ops[2], Ops[2], SV, "lane"); 5295 5296 return Builder.CreateCall(F, {Ops[2], Ops[1], Ops[0]}); 5297 } 5298 case NEON::BI__builtin_neon_vfmaq_laneq_v: { 5299 Value *F = CGM.getIntrinsic(Intrinsic::fma, Ty); 5300 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 5301 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 5302 5303 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 5304 Ops[2] = EmitNeonSplat(Ops[2], cast<ConstantInt>(Ops[3])); 5305 return Builder.CreateCall(F, {Ops[2], Ops[1], Ops[0]}); 5306 } 5307 case NEON::BI__builtin_neon_vfmas_lane_f32: 5308 case NEON::BI__builtin_neon_vfmas_laneq_f32: 5309 case NEON::BI__builtin_neon_vfmad_lane_f64: 5310 case NEON::BI__builtin_neon_vfmad_laneq_f64: { 5311 Ops.push_back(EmitScalarExpr(E->getArg(3))); 5312 llvm::Type *Ty = ConvertType(E->getCallReturnType(getContext())); 5313 Value *F = CGM.getIntrinsic(Intrinsic::fma, Ty); 5314 Ops[2] = Builder.CreateExtractElement(Ops[2], Ops[3], "extract"); 5315 return Builder.CreateCall(F, {Ops[1], Ops[2], Ops[0]}); 5316 } 5317 case NEON::BI__builtin_neon_vmull_v: 5318 // FIXME: improve sharing scheme to cope with 3 alternative LLVM intrinsics. 5319 Int = usgn ? Intrinsic::aarch64_neon_umull : Intrinsic::aarch64_neon_smull; 5320 if (Type.isPoly()) Int = Intrinsic::aarch64_neon_pmull; 5321 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmull"); 5322 case NEON::BI__builtin_neon_vmax_v: 5323 case NEON::BI__builtin_neon_vmaxq_v: 5324 // FIXME: improve sharing scheme to cope with 3 alternative LLVM intrinsics. 5325 Int = usgn ? Intrinsic::aarch64_neon_umax : Intrinsic::aarch64_neon_smax; 5326 if (Ty->isFPOrFPVectorTy()) Int = Intrinsic::aarch64_neon_fmax; 5327 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmax"); 5328 case NEON::BI__builtin_neon_vmin_v: 5329 case NEON::BI__builtin_neon_vminq_v: 5330 // FIXME: improve sharing scheme to cope with 3 alternative LLVM intrinsics. 5331 Int = usgn ? Intrinsic::aarch64_neon_umin : Intrinsic::aarch64_neon_smin; 5332 if (Ty->isFPOrFPVectorTy()) Int = Intrinsic::aarch64_neon_fmin; 5333 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmin"); 5334 case NEON::BI__builtin_neon_vabd_v: 5335 case NEON::BI__builtin_neon_vabdq_v: 5336 // FIXME: improve sharing scheme to cope with 3 alternative LLVM intrinsics. 5337 Int = usgn ? Intrinsic::aarch64_neon_uabd : Intrinsic::aarch64_neon_sabd; 5338 if (Ty->isFPOrFPVectorTy()) Int = Intrinsic::aarch64_neon_fabd; 5339 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vabd"); 5340 case NEON::BI__builtin_neon_vpadal_v: 5341 case NEON::BI__builtin_neon_vpadalq_v: { 5342 unsigned ArgElts = VTy->getNumElements(); 5343 llvm::IntegerType *EltTy = cast<IntegerType>(VTy->getElementType()); 5344 unsigned BitWidth = EltTy->getBitWidth(); 5345 llvm::Type *ArgTy = llvm::VectorType::get( 5346 llvm::IntegerType::get(getLLVMContext(), BitWidth/2), 2*ArgElts); 5347 llvm::Type* Tys[2] = { VTy, ArgTy }; 5348 Int = usgn ? Intrinsic::aarch64_neon_uaddlp : Intrinsic::aarch64_neon_saddlp; 5349 SmallVector<llvm::Value*, 1> TmpOps; 5350 TmpOps.push_back(Ops[1]); 5351 Function *F = CGM.getIntrinsic(Int, Tys); 5352 llvm::Value *tmp = EmitNeonCall(F, TmpOps, "vpadal"); 5353 llvm::Value *addend = Builder.CreateBitCast(Ops[0], tmp->getType()); 5354 return Builder.CreateAdd(tmp, addend); 5355 } 5356 case NEON::BI__builtin_neon_vpmin_v: 5357 case NEON::BI__builtin_neon_vpminq_v: 5358 // FIXME: improve sharing scheme to cope with 3 alternative LLVM intrinsics. 5359 Int = usgn ? Intrinsic::aarch64_neon_uminp : Intrinsic::aarch64_neon_sminp; 5360 if (Ty->isFPOrFPVectorTy()) Int = Intrinsic::aarch64_neon_fminp; 5361 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vpmin"); 5362 case NEON::BI__builtin_neon_vpmax_v: 5363 case NEON::BI__builtin_neon_vpmaxq_v: 5364 // FIXME: improve sharing scheme to cope with 3 alternative LLVM intrinsics. 5365 Int = usgn ? Intrinsic::aarch64_neon_umaxp : Intrinsic::aarch64_neon_smaxp; 5366 if (Ty->isFPOrFPVectorTy()) Int = Intrinsic::aarch64_neon_fmaxp; 5367 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vpmax"); 5368 case NEON::BI__builtin_neon_vminnm_v: 5369 case NEON::BI__builtin_neon_vminnmq_v: 5370 Int = Intrinsic::aarch64_neon_fminnm; 5371 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vminnm"); 5372 case NEON::BI__builtin_neon_vmaxnm_v: 5373 case NEON::BI__builtin_neon_vmaxnmq_v: 5374 Int = Intrinsic::aarch64_neon_fmaxnm; 5375 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmaxnm"); 5376 case NEON::BI__builtin_neon_vrecpss_f32: { 5377 Ops.push_back(EmitScalarExpr(E->getArg(1))); 5378 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_frecps, FloatTy), 5379 Ops, "vrecps"); 5380 } 5381 case NEON::BI__builtin_neon_vrecpsd_f64: { 5382 Ops.push_back(EmitScalarExpr(E->getArg(1))); 5383 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_frecps, DoubleTy), 5384 Ops, "vrecps"); 5385 } 5386 case NEON::BI__builtin_neon_vqshrun_n_v: 5387 Int = Intrinsic::aarch64_neon_sqshrun; 5388 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqshrun_n"); 5389 case NEON::BI__builtin_neon_vqrshrun_n_v: 5390 Int = Intrinsic::aarch64_neon_sqrshrun; 5391 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqrshrun_n"); 5392 case NEON::BI__builtin_neon_vqshrn_n_v: 5393 Int = usgn ? Intrinsic::aarch64_neon_uqshrn : Intrinsic::aarch64_neon_sqshrn; 5394 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqshrn_n"); 5395 case NEON::BI__builtin_neon_vrshrn_n_v: 5396 Int = Intrinsic::aarch64_neon_rshrn; 5397 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrshrn_n"); 5398 case NEON::BI__builtin_neon_vqrshrn_n_v: 5399 Int = usgn ? Intrinsic::aarch64_neon_uqrshrn : Intrinsic::aarch64_neon_sqrshrn; 5400 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqrshrn_n"); 5401 case NEON::BI__builtin_neon_vrnda_v: 5402 case NEON::BI__builtin_neon_vrndaq_v: { 5403 Int = Intrinsic::round; 5404 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrnda"); 5405 } 5406 case NEON::BI__builtin_neon_vrndi_v: 5407 case NEON::BI__builtin_neon_vrndiq_v: { 5408 Int = Intrinsic::nearbyint; 5409 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrndi"); 5410 } 5411 case NEON::BI__builtin_neon_vrndm_v: 5412 case NEON::BI__builtin_neon_vrndmq_v: { 5413 Int = Intrinsic::floor; 5414 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrndm"); 5415 } 5416 case NEON::BI__builtin_neon_vrndn_v: 5417 case NEON::BI__builtin_neon_vrndnq_v: { 5418 Int = Intrinsic::aarch64_neon_frintn; 5419 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrndn"); 5420 } 5421 case NEON::BI__builtin_neon_vrndp_v: 5422 case NEON::BI__builtin_neon_vrndpq_v: { 5423 Int = Intrinsic::ceil; 5424 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrndp"); 5425 } 5426 case NEON::BI__builtin_neon_vrndx_v: 5427 case NEON::BI__builtin_neon_vrndxq_v: { 5428 Int = Intrinsic::rint; 5429 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrndx"); 5430 } 5431 case NEON::BI__builtin_neon_vrnd_v: 5432 case NEON::BI__builtin_neon_vrndq_v: { 5433 Int = Intrinsic::trunc; 5434 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrndz"); 5435 } 5436 case NEON::BI__builtin_neon_vceqz_v: 5437 case NEON::BI__builtin_neon_vceqzq_v: 5438 return EmitAArch64CompareBuiltinExpr(Ops[0], Ty, ICmpInst::FCMP_OEQ, 5439 ICmpInst::ICMP_EQ, "vceqz"); 5440 case NEON::BI__builtin_neon_vcgez_v: 5441 case NEON::BI__builtin_neon_vcgezq_v: 5442 return EmitAArch64CompareBuiltinExpr(Ops[0], Ty, ICmpInst::FCMP_OGE, 5443 ICmpInst::ICMP_SGE, "vcgez"); 5444 case NEON::BI__builtin_neon_vclez_v: 5445 case NEON::BI__builtin_neon_vclezq_v: 5446 return EmitAArch64CompareBuiltinExpr(Ops[0], Ty, ICmpInst::FCMP_OLE, 5447 ICmpInst::ICMP_SLE, "vclez"); 5448 case NEON::BI__builtin_neon_vcgtz_v: 5449 case NEON::BI__builtin_neon_vcgtzq_v: 5450 return EmitAArch64CompareBuiltinExpr(Ops[0], Ty, ICmpInst::FCMP_OGT, 5451 ICmpInst::ICMP_SGT, "vcgtz"); 5452 case NEON::BI__builtin_neon_vcltz_v: 5453 case NEON::BI__builtin_neon_vcltzq_v: 5454 return EmitAArch64CompareBuiltinExpr(Ops[0], Ty, ICmpInst::FCMP_OLT, 5455 ICmpInst::ICMP_SLT, "vcltz"); 5456 case NEON::BI__builtin_neon_vcvt_f64_v: 5457 case NEON::BI__builtin_neon_vcvtq_f64_v: 5458 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 5459 Ty = GetNeonType(this, NeonTypeFlags(NeonTypeFlags::Float64, false, quad)); 5460 return usgn ? Builder.CreateUIToFP(Ops[0], Ty, "vcvt") 5461 : Builder.CreateSIToFP(Ops[0], Ty, "vcvt"); 5462 case NEON::BI__builtin_neon_vcvt_f64_f32: { 5463 assert(Type.getEltType() == NeonTypeFlags::Float64 && quad && 5464 "unexpected vcvt_f64_f32 builtin"); 5465 NeonTypeFlags SrcFlag = NeonTypeFlags(NeonTypeFlags::Float32, false, false); 5466 Ops[0] = Builder.CreateBitCast(Ops[0], GetNeonType(this, SrcFlag)); 5467 5468 return Builder.CreateFPExt(Ops[0], Ty, "vcvt"); 5469 } 5470 case NEON::BI__builtin_neon_vcvt_f32_f64: { 5471 assert(Type.getEltType() == NeonTypeFlags::Float32 && 5472 "unexpected vcvt_f32_f64 builtin"); 5473 NeonTypeFlags SrcFlag = NeonTypeFlags(NeonTypeFlags::Float64, false, true); 5474 Ops[0] = Builder.CreateBitCast(Ops[0], GetNeonType(this, SrcFlag)); 5475 5476 return Builder.CreateFPTrunc(Ops[0], Ty, "vcvt"); 5477 } 5478 case NEON::BI__builtin_neon_vcvt_s32_v: 5479 case NEON::BI__builtin_neon_vcvt_u32_v: 5480 case NEON::BI__builtin_neon_vcvt_s64_v: 5481 case NEON::BI__builtin_neon_vcvt_u64_v: 5482 case NEON::BI__builtin_neon_vcvtq_s32_v: 5483 case NEON::BI__builtin_neon_vcvtq_u32_v: 5484 case NEON::BI__builtin_neon_vcvtq_s64_v: 5485 case NEON::BI__builtin_neon_vcvtq_u64_v: { 5486 Ops[0] = Builder.CreateBitCast(Ops[0], GetFloatNeonType(this, Type)); 5487 if (usgn) 5488 return Builder.CreateFPToUI(Ops[0], Ty); 5489 return Builder.CreateFPToSI(Ops[0], Ty); 5490 } 5491 case NEON::BI__builtin_neon_vcvta_s32_v: 5492 case NEON::BI__builtin_neon_vcvtaq_s32_v: 5493 case NEON::BI__builtin_neon_vcvta_u32_v: 5494 case NEON::BI__builtin_neon_vcvtaq_u32_v: 5495 case NEON::BI__builtin_neon_vcvta_s64_v: 5496 case NEON::BI__builtin_neon_vcvtaq_s64_v: 5497 case NEON::BI__builtin_neon_vcvta_u64_v: 5498 case NEON::BI__builtin_neon_vcvtaq_u64_v: { 5499 Int = usgn ? Intrinsic::aarch64_neon_fcvtau : Intrinsic::aarch64_neon_fcvtas; 5500 llvm::Type *Tys[2] = { Ty, GetFloatNeonType(this, Type) }; 5501 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vcvta"); 5502 } 5503 case NEON::BI__builtin_neon_vcvtm_s32_v: 5504 case NEON::BI__builtin_neon_vcvtmq_s32_v: 5505 case NEON::BI__builtin_neon_vcvtm_u32_v: 5506 case NEON::BI__builtin_neon_vcvtmq_u32_v: 5507 case NEON::BI__builtin_neon_vcvtm_s64_v: 5508 case NEON::BI__builtin_neon_vcvtmq_s64_v: 5509 case NEON::BI__builtin_neon_vcvtm_u64_v: 5510 case NEON::BI__builtin_neon_vcvtmq_u64_v: { 5511 Int = usgn ? Intrinsic::aarch64_neon_fcvtmu : Intrinsic::aarch64_neon_fcvtms; 5512 llvm::Type *Tys[2] = { Ty, GetFloatNeonType(this, Type) }; 5513 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vcvtm"); 5514 } 5515 case NEON::BI__builtin_neon_vcvtn_s32_v: 5516 case NEON::BI__builtin_neon_vcvtnq_s32_v: 5517 case NEON::BI__builtin_neon_vcvtn_u32_v: 5518 case NEON::BI__builtin_neon_vcvtnq_u32_v: 5519 case NEON::BI__builtin_neon_vcvtn_s64_v: 5520 case NEON::BI__builtin_neon_vcvtnq_s64_v: 5521 case NEON::BI__builtin_neon_vcvtn_u64_v: 5522 case NEON::BI__builtin_neon_vcvtnq_u64_v: { 5523 Int = usgn ? Intrinsic::aarch64_neon_fcvtnu : Intrinsic::aarch64_neon_fcvtns; 5524 llvm::Type *Tys[2] = { Ty, GetFloatNeonType(this, Type) }; 5525 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vcvtn"); 5526 } 5527 case NEON::BI__builtin_neon_vcvtp_s32_v: 5528 case NEON::BI__builtin_neon_vcvtpq_s32_v: 5529 case NEON::BI__builtin_neon_vcvtp_u32_v: 5530 case NEON::BI__builtin_neon_vcvtpq_u32_v: 5531 case NEON::BI__builtin_neon_vcvtp_s64_v: 5532 case NEON::BI__builtin_neon_vcvtpq_s64_v: 5533 case NEON::BI__builtin_neon_vcvtp_u64_v: 5534 case NEON::BI__builtin_neon_vcvtpq_u64_v: { 5535 Int = usgn ? Intrinsic::aarch64_neon_fcvtpu : Intrinsic::aarch64_neon_fcvtps; 5536 llvm::Type *Tys[2] = { Ty, GetFloatNeonType(this, Type) }; 5537 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vcvtp"); 5538 } 5539 case NEON::BI__builtin_neon_vmulx_v: 5540 case NEON::BI__builtin_neon_vmulxq_v: { 5541 Int = Intrinsic::aarch64_neon_fmulx; 5542 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmulx"); 5543 } 5544 case NEON::BI__builtin_neon_vmul_lane_v: 5545 case NEON::BI__builtin_neon_vmul_laneq_v: { 5546 // v1f64 vmul_lane should be mapped to Neon scalar mul lane 5547 bool Quad = false; 5548 if (BuiltinID == NEON::BI__builtin_neon_vmul_laneq_v) 5549 Quad = true; 5550 Ops[0] = Builder.CreateBitCast(Ops[0], DoubleTy); 5551 llvm::Type *VTy = GetNeonType(this, 5552 NeonTypeFlags(NeonTypeFlags::Float64, false, Quad)); 5553 Ops[1] = Builder.CreateBitCast(Ops[1], VTy); 5554 Ops[1] = Builder.CreateExtractElement(Ops[1], Ops[2], "extract"); 5555 Value *Result = Builder.CreateFMul(Ops[0], Ops[1]); 5556 return Builder.CreateBitCast(Result, Ty); 5557 } 5558 case NEON::BI__builtin_neon_vnegd_s64: 5559 return Builder.CreateNeg(EmitScalarExpr(E->getArg(0)), "vnegd"); 5560 case NEON::BI__builtin_neon_vpmaxnm_v: 5561 case NEON::BI__builtin_neon_vpmaxnmq_v: { 5562 Int = Intrinsic::aarch64_neon_fmaxnmp; 5563 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vpmaxnm"); 5564 } 5565 case NEON::BI__builtin_neon_vpminnm_v: 5566 case NEON::BI__builtin_neon_vpminnmq_v: { 5567 Int = Intrinsic::aarch64_neon_fminnmp; 5568 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vpminnm"); 5569 } 5570 case NEON::BI__builtin_neon_vsqrt_v: 5571 case NEON::BI__builtin_neon_vsqrtq_v: { 5572 Int = Intrinsic::sqrt; 5573 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 5574 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vsqrt"); 5575 } 5576 case NEON::BI__builtin_neon_vrbit_v: 5577 case NEON::BI__builtin_neon_vrbitq_v: { 5578 Int = Intrinsic::aarch64_neon_rbit; 5579 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrbit"); 5580 } 5581 case NEON::BI__builtin_neon_vaddv_u8: 5582 // FIXME: These are handled by the AArch64 scalar code. 5583 usgn = true; 5584 // FALLTHROUGH 5585 case NEON::BI__builtin_neon_vaddv_s8: { 5586 Int = usgn ? Intrinsic::aarch64_neon_uaddv : Intrinsic::aarch64_neon_saddv; 5587 Ty = Int32Ty; 5588 VTy = llvm::VectorType::get(Int8Ty, 8); 5589 llvm::Type *Tys[2] = { Ty, VTy }; 5590 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5591 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddv"); 5592 return Builder.CreateTrunc(Ops[0], Int8Ty); 5593 } 5594 case NEON::BI__builtin_neon_vaddv_u16: 5595 usgn = true; 5596 // FALLTHROUGH 5597 case NEON::BI__builtin_neon_vaddv_s16: { 5598 Int = usgn ? Intrinsic::aarch64_neon_uaddv : Intrinsic::aarch64_neon_saddv; 5599 Ty = Int32Ty; 5600 VTy = llvm::VectorType::get(Int16Ty, 4); 5601 llvm::Type *Tys[2] = { Ty, VTy }; 5602 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5603 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddv"); 5604 return Builder.CreateTrunc(Ops[0], Int16Ty); 5605 } 5606 case NEON::BI__builtin_neon_vaddvq_u8: 5607 usgn = true; 5608 // FALLTHROUGH 5609 case NEON::BI__builtin_neon_vaddvq_s8: { 5610 Int = usgn ? Intrinsic::aarch64_neon_uaddv : Intrinsic::aarch64_neon_saddv; 5611 Ty = Int32Ty; 5612 VTy = llvm::VectorType::get(Int8Ty, 16); 5613 llvm::Type *Tys[2] = { Ty, VTy }; 5614 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5615 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddv"); 5616 return Builder.CreateTrunc(Ops[0], Int8Ty); 5617 } 5618 case NEON::BI__builtin_neon_vaddvq_u16: 5619 usgn = true; 5620 // FALLTHROUGH 5621 case NEON::BI__builtin_neon_vaddvq_s16: { 5622 Int = usgn ? Intrinsic::aarch64_neon_uaddv : Intrinsic::aarch64_neon_saddv; 5623 Ty = Int32Ty; 5624 VTy = llvm::VectorType::get(Int16Ty, 8); 5625 llvm::Type *Tys[2] = { Ty, VTy }; 5626 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5627 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddv"); 5628 return Builder.CreateTrunc(Ops[0], Int16Ty); 5629 } 5630 case NEON::BI__builtin_neon_vmaxv_u8: { 5631 Int = Intrinsic::aarch64_neon_umaxv; 5632 Ty = Int32Ty; 5633 VTy = llvm::VectorType::get(Int8Ty, 8); 5634 llvm::Type *Tys[2] = { Ty, VTy }; 5635 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5636 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv"); 5637 return Builder.CreateTrunc(Ops[0], Int8Ty); 5638 } 5639 case NEON::BI__builtin_neon_vmaxv_u16: { 5640 Int = Intrinsic::aarch64_neon_umaxv; 5641 Ty = Int32Ty; 5642 VTy = llvm::VectorType::get(Int16Ty, 4); 5643 llvm::Type *Tys[2] = { Ty, VTy }; 5644 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5645 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv"); 5646 return Builder.CreateTrunc(Ops[0], Int16Ty); 5647 } 5648 case NEON::BI__builtin_neon_vmaxvq_u8: { 5649 Int = Intrinsic::aarch64_neon_umaxv; 5650 Ty = Int32Ty; 5651 VTy = llvm::VectorType::get(Int8Ty, 16); 5652 llvm::Type *Tys[2] = { Ty, VTy }; 5653 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5654 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv"); 5655 return Builder.CreateTrunc(Ops[0], Int8Ty); 5656 } 5657 case NEON::BI__builtin_neon_vmaxvq_u16: { 5658 Int = Intrinsic::aarch64_neon_umaxv; 5659 Ty = Int32Ty; 5660 VTy = llvm::VectorType::get(Int16Ty, 8); 5661 llvm::Type *Tys[2] = { Ty, VTy }; 5662 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5663 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv"); 5664 return Builder.CreateTrunc(Ops[0], Int16Ty); 5665 } 5666 case NEON::BI__builtin_neon_vmaxv_s8: { 5667 Int = Intrinsic::aarch64_neon_smaxv; 5668 Ty = Int32Ty; 5669 VTy = llvm::VectorType::get(Int8Ty, 8); 5670 llvm::Type *Tys[2] = { Ty, VTy }; 5671 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5672 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv"); 5673 return Builder.CreateTrunc(Ops[0], Int8Ty); 5674 } 5675 case NEON::BI__builtin_neon_vmaxv_s16: { 5676 Int = Intrinsic::aarch64_neon_smaxv; 5677 Ty = Int32Ty; 5678 VTy = llvm::VectorType::get(Int16Ty, 4); 5679 llvm::Type *Tys[2] = { Ty, VTy }; 5680 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5681 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv"); 5682 return Builder.CreateTrunc(Ops[0], Int16Ty); 5683 } 5684 case NEON::BI__builtin_neon_vmaxvq_s8: { 5685 Int = Intrinsic::aarch64_neon_smaxv; 5686 Ty = Int32Ty; 5687 VTy = llvm::VectorType::get(Int8Ty, 16); 5688 llvm::Type *Tys[2] = { Ty, VTy }; 5689 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5690 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv"); 5691 return Builder.CreateTrunc(Ops[0], Int8Ty); 5692 } 5693 case NEON::BI__builtin_neon_vmaxvq_s16: { 5694 Int = Intrinsic::aarch64_neon_smaxv; 5695 Ty = Int32Ty; 5696 VTy = llvm::VectorType::get(Int16Ty, 8); 5697 llvm::Type *Tys[2] = { Ty, VTy }; 5698 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5699 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv"); 5700 return Builder.CreateTrunc(Ops[0], Int16Ty); 5701 } 5702 case NEON::BI__builtin_neon_vminv_u8: { 5703 Int = Intrinsic::aarch64_neon_uminv; 5704 Ty = Int32Ty; 5705 VTy = llvm::VectorType::get(Int8Ty, 8); 5706 llvm::Type *Tys[2] = { Ty, VTy }; 5707 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5708 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv"); 5709 return Builder.CreateTrunc(Ops[0], Int8Ty); 5710 } 5711 case NEON::BI__builtin_neon_vminv_u16: { 5712 Int = Intrinsic::aarch64_neon_uminv; 5713 Ty = Int32Ty; 5714 VTy = llvm::VectorType::get(Int16Ty, 4); 5715 llvm::Type *Tys[2] = { Ty, VTy }; 5716 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5717 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv"); 5718 return Builder.CreateTrunc(Ops[0], Int16Ty); 5719 } 5720 case NEON::BI__builtin_neon_vminvq_u8: { 5721 Int = Intrinsic::aarch64_neon_uminv; 5722 Ty = Int32Ty; 5723 VTy = llvm::VectorType::get(Int8Ty, 16); 5724 llvm::Type *Tys[2] = { Ty, VTy }; 5725 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5726 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv"); 5727 return Builder.CreateTrunc(Ops[0], Int8Ty); 5728 } 5729 case NEON::BI__builtin_neon_vminvq_u16: { 5730 Int = Intrinsic::aarch64_neon_uminv; 5731 Ty = Int32Ty; 5732 VTy = llvm::VectorType::get(Int16Ty, 8); 5733 llvm::Type *Tys[2] = { Ty, VTy }; 5734 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5735 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv"); 5736 return Builder.CreateTrunc(Ops[0], Int16Ty); 5737 } 5738 case NEON::BI__builtin_neon_vminv_s8: { 5739 Int = Intrinsic::aarch64_neon_sminv; 5740 Ty = Int32Ty; 5741 VTy = llvm::VectorType::get(Int8Ty, 8); 5742 llvm::Type *Tys[2] = { Ty, VTy }; 5743 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5744 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv"); 5745 return Builder.CreateTrunc(Ops[0], Int8Ty); 5746 } 5747 case NEON::BI__builtin_neon_vminv_s16: { 5748 Int = Intrinsic::aarch64_neon_sminv; 5749 Ty = Int32Ty; 5750 VTy = llvm::VectorType::get(Int16Ty, 4); 5751 llvm::Type *Tys[2] = { Ty, VTy }; 5752 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5753 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv"); 5754 return Builder.CreateTrunc(Ops[0], Int16Ty); 5755 } 5756 case NEON::BI__builtin_neon_vminvq_s8: { 5757 Int = Intrinsic::aarch64_neon_sminv; 5758 Ty = Int32Ty; 5759 VTy = llvm::VectorType::get(Int8Ty, 16); 5760 llvm::Type *Tys[2] = { Ty, VTy }; 5761 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5762 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv"); 5763 return Builder.CreateTrunc(Ops[0], Int8Ty); 5764 } 5765 case NEON::BI__builtin_neon_vminvq_s16: { 5766 Int = Intrinsic::aarch64_neon_sminv; 5767 Ty = Int32Ty; 5768 VTy = llvm::VectorType::get(Int16Ty, 8); 5769 llvm::Type *Tys[2] = { Ty, VTy }; 5770 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5771 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv"); 5772 return Builder.CreateTrunc(Ops[0], Int16Ty); 5773 } 5774 case NEON::BI__builtin_neon_vmul_n_f64: { 5775 Ops[0] = Builder.CreateBitCast(Ops[0], DoubleTy); 5776 Value *RHS = Builder.CreateBitCast(EmitScalarExpr(E->getArg(1)), DoubleTy); 5777 return Builder.CreateFMul(Ops[0], RHS); 5778 } 5779 case NEON::BI__builtin_neon_vaddlv_u8: { 5780 Int = Intrinsic::aarch64_neon_uaddlv; 5781 Ty = Int32Ty; 5782 VTy = llvm::VectorType::get(Int8Ty, 8); 5783 llvm::Type *Tys[2] = { Ty, VTy }; 5784 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5785 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv"); 5786 return Builder.CreateTrunc(Ops[0], Int16Ty); 5787 } 5788 case NEON::BI__builtin_neon_vaddlv_u16: { 5789 Int = Intrinsic::aarch64_neon_uaddlv; 5790 Ty = Int32Ty; 5791 VTy = llvm::VectorType::get(Int16Ty, 4); 5792 llvm::Type *Tys[2] = { Ty, VTy }; 5793 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5794 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv"); 5795 } 5796 case NEON::BI__builtin_neon_vaddlvq_u8: { 5797 Int = Intrinsic::aarch64_neon_uaddlv; 5798 Ty = Int32Ty; 5799 VTy = llvm::VectorType::get(Int8Ty, 16); 5800 llvm::Type *Tys[2] = { Ty, VTy }; 5801 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5802 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv"); 5803 return Builder.CreateTrunc(Ops[0], Int16Ty); 5804 } 5805 case NEON::BI__builtin_neon_vaddlvq_u16: { 5806 Int = Intrinsic::aarch64_neon_uaddlv; 5807 Ty = Int32Ty; 5808 VTy = llvm::VectorType::get(Int16Ty, 8); 5809 llvm::Type *Tys[2] = { Ty, VTy }; 5810 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5811 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv"); 5812 } 5813 case NEON::BI__builtin_neon_vaddlv_s8: { 5814 Int = Intrinsic::aarch64_neon_saddlv; 5815 Ty = Int32Ty; 5816 VTy = llvm::VectorType::get(Int8Ty, 8); 5817 llvm::Type *Tys[2] = { Ty, VTy }; 5818 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5819 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv"); 5820 return Builder.CreateTrunc(Ops[0], Int16Ty); 5821 } 5822 case NEON::BI__builtin_neon_vaddlv_s16: { 5823 Int = Intrinsic::aarch64_neon_saddlv; 5824 Ty = Int32Ty; 5825 VTy = llvm::VectorType::get(Int16Ty, 4); 5826 llvm::Type *Tys[2] = { Ty, VTy }; 5827 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5828 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv"); 5829 } 5830 case NEON::BI__builtin_neon_vaddlvq_s8: { 5831 Int = Intrinsic::aarch64_neon_saddlv; 5832 Ty = Int32Ty; 5833 VTy = llvm::VectorType::get(Int8Ty, 16); 5834 llvm::Type *Tys[2] = { Ty, VTy }; 5835 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5836 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv"); 5837 return Builder.CreateTrunc(Ops[0], Int16Ty); 5838 } 5839 case NEON::BI__builtin_neon_vaddlvq_s16: { 5840 Int = Intrinsic::aarch64_neon_saddlv; 5841 Ty = Int32Ty; 5842 VTy = llvm::VectorType::get(Int16Ty, 8); 5843 llvm::Type *Tys[2] = { Ty, VTy }; 5844 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5845 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv"); 5846 } 5847 case NEON::BI__builtin_neon_vsri_n_v: 5848 case NEON::BI__builtin_neon_vsriq_n_v: { 5849 Int = Intrinsic::aarch64_neon_vsri; 5850 llvm::Function *Intrin = CGM.getIntrinsic(Int, Ty); 5851 return EmitNeonCall(Intrin, Ops, "vsri_n"); 5852 } 5853 case NEON::BI__builtin_neon_vsli_n_v: 5854 case NEON::BI__builtin_neon_vsliq_n_v: { 5855 Int = Intrinsic::aarch64_neon_vsli; 5856 llvm::Function *Intrin = CGM.getIntrinsic(Int, Ty); 5857 return EmitNeonCall(Intrin, Ops, "vsli_n"); 5858 } 5859 case NEON::BI__builtin_neon_vsra_n_v: 5860 case NEON::BI__builtin_neon_vsraq_n_v: 5861 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 5862 Ops[1] = EmitNeonRShiftImm(Ops[1], Ops[2], Ty, usgn, "vsra_n"); 5863 return Builder.CreateAdd(Ops[0], Ops[1]); 5864 case NEON::BI__builtin_neon_vrsra_n_v: 5865 case NEON::BI__builtin_neon_vrsraq_n_v: { 5866 Int = usgn ? Intrinsic::aarch64_neon_urshl : Intrinsic::aarch64_neon_srshl; 5867 SmallVector<llvm::Value*,2> TmpOps; 5868 TmpOps.push_back(Ops[1]); 5869 TmpOps.push_back(Ops[2]); 5870 Function* F = CGM.getIntrinsic(Int, Ty); 5871 llvm::Value *tmp = EmitNeonCall(F, TmpOps, "vrshr_n", 1, true); 5872 Ops[0] = Builder.CreateBitCast(Ops[0], VTy); 5873 return Builder.CreateAdd(Ops[0], tmp); 5874 } 5875 // FIXME: Sharing loads & stores with 32-bit is complicated by the absence 5876 // of an Align parameter here. 5877 case NEON::BI__builtin_neon_vld1_x2_v: 5878 case NEON::BI__builtin_neon_vld1q_x2_v: 5879 case NEON::BI__builtin_neon_vld1_x3_v: 5880 case NEON::BI__builtin_neon_vld1q_x3_v: 5881 case NEON::BI__builtin_neon_vld1_x4_v: 5882 case NEON::BI__builtin_neon_vld1q_x4_v: { 5883 llvm::Type *PTy = llvm::PointerType::getUnqual(VTy->getVectorElementType()); 5884 Ops[1] = Builder.CreateBitCast(Ops[1], PTy); 5885 llvm::Type *Tys[2] = { VTy, PTy }; 5886 unsigned Int; 5887 switch (BuiltinID) { 5888 case NEON::BI__builtin_neon_vld1_x2_v: 5889 case NEON::BI__builtin_neon_vld1q_x2_v: 5890 Int = Intrinsic::aarch64_neon_ld1x2; 5891 break; 5892 case NEON::BI__builtin_neon_vld1_x3_v: 5893 case NEON::BI__builtin_neon_vld1q_x3_v: 5894 Int = Intrinsic::aarch64_neon_ld1x3; 5895 break; 5896 case NEON::BI__builtin_neon_vld1_x4_v: 5897 case NEON::BI__builtin_neon_vld1q_x4_v: 5898 Int = Intrinsic::aarch64_neon_ld1x4; 5899 break; 5900 } 5901 Function *F = CGM.getIntrinsic(Int, Tys); 5902 Ops[1] = Builder.CreateCall(F, Ops[1], "vld1xN"); 5903 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 5904 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 5905 return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]); 5906 } 5907 case NEON::BI__builtin_neon_vst1_x2_v: 5908 case NEON::BI__builtin_neon_vst1q_x2_v: 5909 case NEON::BI__builtin_neon_vst1_x3_v: 5910 case NEON::BI__builtin_neon_vst1q_x3_v: 5911 case NEON::BI__builtin_neon_vst1_x4_v: 5912 case NEON::BI__builtin_neon_vst1q_x4_v: { 5913 llvm::Type *PTy = llvm::PointerType::getUnqual(VTy->getVectorElementType()); 5914 llvm::Type *Tys[2] = { VTy, PTy }; 5915 unsigned Int; 5916 switch (BuiltinID) { 5917 case NEON::BI__builtin_neon_vst1_x2_v: 5918 case NEON::BI__builtin_neon_vst1q_x2_v: 5919 Int = Intrinsic::aarch64_neon_st1x2; 5920 break; 5921 case NEON::BI__builtin_neon_vst1_x3_v: 5922 case NEON::BI__builtin_neon_vst1q_x3_v: 5923 Int = Intrinsic::aarch64_neon_st1x3; 5924 break; 5925 case NEON::BI__builtin_neon_vst1_x4_v: 5926 case NEON::BI__builtin_neon_vst1q_x4_v: 5927 Int = Intrinsic::aarch64_neon_st1x4; 5928 break; 5929 } 5930 std::rotate(Ops.begin(), Ops.begin() + 1, Ops.end()); 5931 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, ""); 5932 } 5933 case NEON::BI__builtin_neon_vld1_v: 5934 case NEON::BI__builtin_neon_vld1q_v: 5935 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(VTy)); 5936 return Builder.CreateDefaultAlignedLoad(Ops[0]); 5937 case NEON::BI__builtin_neon_vst1_v: 5938 case NEON::BI__builtin_neon_vst1q_v: 5939 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(VTy)); 5940 Ops[1] = Builder.CreateBitCast(Ops[1], VTy); 5941 return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]); 5942 case NEON::BI__builtin_neon_vld1_lane_v: 5943 case NEON::BI__builtin_neon_vld1q_lane_v: 5944 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 5945 Ty = llvm::PointerType::getUnqual(VTy->getElementType()); 5946 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 5947 Ops[0] = Builder.CreateDefaultAlignedLoad(Ops[0]); 5948 return Builder.CreateInsertElement(Ops[1], Ops[0], Ops[2], "vld1_lane"); 5949 case NEON::BI__builtin_neon_vld1_dup_v: 5950 case NEON::BI__builtin_neon_vld1q_dup_v: { 5951 Value *V = UndefValue::get(Ty); 5952 Ty = llvm::PointerType::getUnqual(VTy->getElementType()); 5953 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 5954 Ops[0] = Builder.CreateDefaultAlignedLoad(Ops[0]); 5955 llvm::Constant *CI = ConstantInt::get(Int32Ty, 0); 5956 Ops[0] = Builder.CreateInsertElement(V, Ops[0], CI); 5957 return EmitNeonSplat(Ops[0], CI); 5958 } 5959 case NEON::BI__builtin_neon_vst1_lane_v: 5960 case NEON::BI__builtin_neon_vst1q_lane_v: 5961 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 5962 Ops[1] = Builder.CreateExtractElement(Ops[1], Ops[2]); 5963 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 5964 return Builder.CreateDefaultAlignedStore(Ops[1], 5965 Builder.CreateBitCast(Ops[0], Ty)); 5966 case NEON::BI__builtin_neon_vld2_v: 5967 case NEON::BI__builtin_neon_vld2q_v: { 5968 llvm::Type *PTy = llvm::PointerType::getUnqual(VTy); 5969 Ops[1] = Builder.CreateBitCast(Ops[1], PTy); 5970 llvm::Type *Tys[2] = { VTy, PTy }; 5971 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld2, Tys); 5972 Ops[1] = Builder.CreateCall(F, Ops[1], "vld2"); 5973 Ops[0] = Builder.CreateBitCast(Ops[0], 5974 llvm::PointerType::getUnqual(Ops[1]->getType())); 5975 return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]); 5976 } 5977 case NEON::BI__builtin_neon_vld3_v: 5978 case NEON::BI__builtin_neon_vld3q_v: { 5979 llvm::Type *PTy = llvm::PointerType::getUnqual(VTy); 5980 Ops[1] = Builder.CreateBitCast(Ops[1], PTy); 5981 llvm::Type *Tys[2] = { VTy, PTy }; 5982 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld3, Tys); 5983 Ops[1] = Builder.CreateCall(F, Ops[1], "vld3"); 5984 Ops[0] = Builder.CreateBitCast(Ops[0], 5985 llvm::PointerType::getUnqual(Ops[1]->getType())); 5986 return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]); 5987 } 5988 case NEON::BI__builtin_neon_vld4_v: 5989 case NEON::BI__builtin_neon_vld4q_v: { 5990 llvm::Type *PTy = llvm::PointerType::getUnqual(VTy); 5991 Ops[1] = Builder.CreateBitCast(Ops[1], PTy); 5992 llvm::Type *Tys[2] = { VTy, PTy }; 5993 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld4, Tys); 5994 Ops[1] = Builder.CreateCall(F, Ops[1], "vld4"); 5995 Ops[0] = Builder.CreateBitCast(Ops[0], 5996 llvm::PointerType::getUnqual(Ops[1]->getType())); 5997 return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]); 5998 } 5999 case NEON::BI__builtin_neon_vld2_dup_v: 6000 case NEON::BI__builtin_neon_vld2q_dup_v: { 6001 llvm::Type *PTy = 6002 llvm::PointerType::getUnqual(VTy->getElementType()); 6003 Ops[1] = Builder.CreateBitCast(Ops[1], PTy); 6004 llvm::Type *Tys[2] = { VTy, PTy }; 6005 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld2r, Tys); 6006 Ops[1] = Builder.CreateCall(F, Ops[1], "vld2"); 6007 Ops[0] = Builder.CreateBitCast(Ops[0], 6008 llvm::PointerType::getUnqual(Ops[1]->getType())); 6009 return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]); 6010 } 6011 case NEON::BI__builtin_neon_vld3_dup_v: 6012 case NEON::BI__builtin_neon_vld3q_dup_v: { 6013 llvm::Type *PTy = 6014 llvm::PointerType::getUnqual(VTy->getElementType()); 6015 Ops[1] = Builder.CreateBitCast(Ops[1], PTy); 6016 llvm::Type *Tys[2] = { VTy, PTy }; 6017 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld3r, Tys); 6018 Ops[1] = Builder.CreateCall(F, Ops[1], "vld3"); 6019 Ops[0] = Builder.CreateBitCast(Ops[0], 6020 llvm::PointerType::getUnqual(Ops[1]->getType())); 6021 return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]); 6022 } 6023 case NEON::BI__builtin_neon_vld4_dup_v: 6024 case NEON::BI__builtin_neon_vld4q_dup_v: { 6025 llvm::Type *PTy = 6026 llvm::PointerType::getUnqual(VTy->getElementType()); 6027 Ops[1] = Builder.CreateBitCast(Ops[1], PTy); 6028 llvm::Type *Tys[2] = { VTy, PTy }; 6029 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld4r, Tys); 6030 Ops[1] = Builder.CreateCall(F, Ops[1], "vld4"); 6031 Ops[0] = Builder.CreateBitCast(Ops[0], 6032 llvm::PointerType::getUnqual(Ops[1]->getType())); 6033 return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]); 6034 } 6035 case NEON::BI__builtin_neon_vld2_lane_v: 6036 case NEON::BI__builtin_neon_vld2q_lane_v: { 6037 llvm::Type *Tys[2] = { VTy, Ops[1]->getType() }; 6038 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld2lane, Tys); 6039 Ops.push_back(Ops[1]); 6040 Ops.erase(Ops.begin()+1); 6041 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 6042 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 6043 Ops[3] = Builder.CreateZExt(Ops[3], Int64Ty); 6044 Ops[1] = Builder.CreateCall(F, makeArrayRef(Ops).slice(1), "vld2_lane"); 6045 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 6046 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 6047 return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]); 6048 } 6049 case NEON::BI__builtin_neon_vld3_lane_v: 6050 case NEON::BI__builtin_neon_vld3q_lane_v: { 6051 llvm::Type *Tys[2] = { VTy, Ops[1]->getType() }; 6052 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld3lane, Tys); 6053 Ops.push_back(Ops[1]); 6054 Ops.erase(Ops.begin()+1); 6055 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 6056 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 6057 Ops[3] = Builder.CreateBitCast(Ops[3], Ty); 6058 Ops[4] = Builder.CreateZExt(Ops[4], Int64Ty); 6059 Ops[1] = Builder.CreateCall(F, makeArrayRef(Ops).slice(1), "vld3_lane"); 6060 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 6061 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 6062 return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]); 6063 } 6064 case NEON::BI__builtin_neon_vld4_lane_v: 6065 case NEON::BI__builtin_neon_vld4q_lane_v: { 6066 llvm::Type *Tys[2] = { VTy, Ops[1]->getType() }; 6067 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld4lane, Tys); 6068 Ops.push_back(Ops[1]); 6069 Ops.erase(Ops.begin()+1); 6070 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 6071 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 6072 Ops[3] = Builder.CreateBitCast(Ops[3], Ty); 6073 Ops[4] = Builder.CreateBitCast(Ops[4], Ty); 6074 Ops[5] = Builder.CreateZExt(Ops[5], Int64Ty); 6075 Ops[1] = Builder.CreateCall(F, makeArrayRef(Ops).slice(1), "vld4_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_vst2_v: 6081 case NEON::BI__builtin_neon_vst2q_v: { 6082 Ops.push_back(Ops[0]); 6083 Ops.erase(Ops.begin()); 6084 llvm::Type *Tys[2] = { VTy, Ops[2]->getType() }; 6085 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_st2, Tys), 6086 Ops, ""); 6087 } 6088 case NEON::BI__builtin_neon_vst2_lane_v: 6089 case NEON::BI__builtin_neon_vst2q_lane_v: { 6090 Ops.push_back(Ops[0]); 6091 Ops.erase(Ops.begin()); 6092 Ops[2] = Builder.CreateZExt(Ops[2], Int64Ty); 6093 llvm::Type *Tys[2] = { VTy, Ops[3]->getType() }; 6094 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_st2lane, Tys), 6095 Ops, ""); 6096 } 6097 case NEON::BI__builtin_neon_vst3_v: 6098 case NEON::BI__builtin_neon_vst3q_v: { 6099 Ops.push_back(Ops[0]); 6100 Ops.erase(Ops.begin()); 6101 llvm::Type *Tys[2] = { VTy, Ops[3]->getType() }; 6102 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_st3, Tys), 6103 Ops, ""); 6104 } 6105 case NEON::BI__builtin_neon_vst3_lane_v: 6106 case NEON::BI__builtin_neon_vst3q_lane_v: { 6107 Ops.push_back(Ops[0]); 6108 Ops.erase(Ops.begin()); 6109 Ops[3] = Builder.CreateZExt(Ops[3], Int64Ty); 6110 llvm::Type *Tys[2] = { VTy, Ops[4]->getType() }; 6111 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_st3lane, Tys), 6112 Ops, ""); 6113 } 6114 case NEON::BI__builtin_neon_vst4_v: 6115 case NEON::BI__builtin_neon_vst4q_v: { 6116 Ops.push_back(Ops[0]); 6117 Ops.erase(Ops.begin()); 6118 llvm::Type *Tys[2] = { VTy, Ops[4]->getType() }; 6119 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_st4, Tys), 6120 Ops, ""); 6121 } 6122 case NEON::BI__builtin_neon_vst4_lane_v: 6123 case NEON::BI__builtin_neon_vst4q_lane_v: { 6124 Ops.push_back(Ops[0]); 6125 Ops.erase(Ops.begin()); 6126 Ops[4] = Builder.CreateZExt(Ops[4], Int64Ty); 6127 llvm::Type *Tys[2] = { VTy, Ops[5]->getType() }; 6128 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_st4lane, Tys), 6129 Ops, ""); 6130 } 6131 case NEON::BI__builtin_neon_vtrn_v: 6132 case NEON::BI__builtin_neon_vtrnq_v: { 6133 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty)); 6134 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 6135 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 6136 Value *SV = nullptr; 6137 6138 for (unsigned vi = 0; vi != 2; ++vi) { 6139 SmallVector<Constant*, 16> Indices; 6140 for (unsigned i = 0, e = VTy->getNumElements(); i != e; i += 2) { 6141 Indices.push_back(ConstantInt::get(Int32Ty, i+vi)); 6142 Indices.push_back(ConstantInt::get(Int32Ty, i+e+vi)); 6143 } 6144 Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ty, Ops[0], vi); 6145 SV = llvm::ConstantVector::get(Indices); 6146 SV = Builder.CreateShuffleVector(Ops[1], Ops[2], SV, "vtrn"); 6147 SV = Builder.CreateDefaultAlignedStore(SV, Addr); 6148 } 6149 return SV; 6150 } 6151 case NEON::BI__builtin_neon_vuzp_v: 6152 case NEON::BI__builtin_neon_vuzpq_v: { 6153 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty)); 6154 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 6155 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 6156 Value *SV = nullptr; 6157 6158 for (unsigned vi = 0; vi != 2; ++vi) { 6159 SmallVector<Constant*, 16> Indices; 6160 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) 6161 Indices.push_back(ConstantInt::get(Int32Ty, 2*i+vi)); 6162 6163 Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ty, Ops[0], vi); 6164 SV = llvm::ConstantVector::get(Indices); 6165 SV = Builder.CreateShuffleVector(Ops[1], Ops[2], SV, "vuzp"); 6166 SV = Builder.CreateDefaultAlignedStore(SV, Addr); 6167 } 6168 return SV; 6169 } 6170 case NEON::BI__builtin_neon_vzip_v: 6171 case NEON::BI__builtin_neon_vzipq_v: { 6172 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty)); 6173 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 6174 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 6175 Value *SV = nullptr; 6176 6177 for (unsigned vi = 0; vi != 2; ++vi) { 6178 SmallVector<Constant*, 16> Indices; 6179 for (unsigned i = 0, e = VTy->getNumElements(); i != e; i += 2) { 6180 Indices.push_back(ConstantInt::get(Int32Ty, (i + vi*e) >> 1)); 6181 Indices.push_back(ConstantInt::get(Int32Ty, ((i + vi*e) >> 1)+e)); 6182 } 6183 Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ty, Ops[0], vi); 6184 SV = llvm::ConstantVector::get(Indices); 6185 SV = Builder.CreateShuffleVector(Ops[1], Ops[2], SV, "vzip"); 6186 SV = Builder.CreateDefaultAlignedStore(SV, Addr); 6187 } 6188 return SV; 6189 } 6190 case NEON::BI__builtin_neon_vqtbl1q_v: { 6191 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbl1, Ty), 6192 Ops, "vtbl1"); 6193 } 6194 case NEON::BI__builtin_neon_vqtbl2q_v: { 6195 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbl2, Ty), 6196 Ops, "vtbl2"); 6197 } 6198 case NEON::BI__builtin_neon_vqtbl3q_v: { 6199 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbl3, Ty), 6200 Ops, "vtbl3"); 6201 } 6202 case NEON::BI__builtin_neon_vqtbl4q_v: { 6203 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbl4, Ty), 6204 Ops, "vtbl4"); 6205 } 6206 case NEON::BI__builtin_neon_vqtbx1q_v: { 6207 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbx1, Ty), 6208 Ops, "vtbx1"); 6209 } 6210 case NEON::BI__builtin_neon_vqtbx2q_v: { 6211 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbx2, Ty), 6212 Ops, "vtbx2"); 6213 } 6214 case NEON::BI__builtin_neon_vqtbx3q_v: { 6215 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbx3, Ty), 6216 Ops, "vtbx3"); 6217 } 6218 case NEON::BI__builtin_neon_vqtbx4q_v: { 6219 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbx4, Ty), 6220 Ops, "vtbx4"); 6221 } 6222 case NEON::BI__builtin_neon_vsqadd_v: 6223 case NEON::BI__builtin_neon_vsqaddq_v: { 6224 Int = Intrinsic::aarch64_neon_usqadd; 6225 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vsqadd"); 6226 } 6227 case NEON::BI__builtin_neon_vuqadd_v: 6228 case NEON::BI__builtin_neon_vuqaddq_v: { 6229 Int = Intrinsic::aarch64_neon_suqadd; 6230 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vuqadd"); 6231 } 6232 } 6233 } 6234 6235 llvm::Value *CodeGenFunction:: 6236 BuildVector(ArrayRef<llvm::Value*> Ops) { 6237 assert((Ops.size() & (Ops.size() - 1)) == 0 && 6238 "Not a power-of-two sized vector!"); 6239 bool AllConstants = true; 6240 for (unsigned i = 0, e = Ops.size(); i != e && AllConstants; ++i) 6241 AllConstants &= isa<Constant>(Ops[i]); 6242 6243 // If this is a constant vector, create a ConstantVector. 6244 if (AllConstants) { 6245 SmallVector<llvm::Constant*, 16> CstOps; 6246 for (unsigned i = 0, e = Ops.size(); i != e; ++i) 6247 CstOps.push_back(cast<Constant>(Ops[i])); 6248 return llvm::ConstantVector::get(CstOps); 6249 } 6250 6251 // Otherwise, insertelement the values to build the vector. 6252 Value *Result = 6253 llvm::UndefValue::get(llvm::VectorType::get(Ops[0]->getType(), Ops.size())); 6254 6255 for (unsigned i = 0, e = Ops.size(); i != e; ++i) 6256 Result = Builder.CreateInsertElement(Result, Ops[i], Builder.getInt32(i)); 6257 6258 return Result; 6259 } 6260 6261 Value *CodeGenFunction::EmitX86BuiltinExpr(unsigned BuiltinID, 6262 const CallExpr *E) { 6263 if (BuiltinID == X86::BI__builtin_ms_va_start || 6264 BuiltinID == X86::BI__builtin_ms_va_end) 6265 return EmitVAStartEnd(EmitMSVAListRef(E->getArg(0)).getPointer(), 6266 BuiltinID == X86::BI__builtin_ms_va_start); 6267 if (BuiltinID == X86::BI__builtin_ms_va_copy) { 6268 // Lower this manually. We can't reliably determine whether or not any 6269 // given va_copy() is for a Win64 va_list from the calling convention 6270 // alone, because it's legal to do this from a System V ABI function. 6271 // With opaque pointer types, we won't have enough information in LLVM 6272 // IR to determine this from the argument types, either. Best to do it 6273 // now, while we have enough information. 6274 Address DestAddr = EmitMSVAListRef(E->getArg(0)); 6275 Address SrcAddr = EmitMSVAListRef(E->getArg(1)); 6276 6277 llvm::Type *BPP = Int8PtrPtrTy; 6278 6279 DestAddr = Address(Builder.CreateBitCast(DestAddr.getPointer(), BPP, "cp"), 6280 DestAddr.getAlignment()); 6281 SrcAddr = Address(Builder.CreateBitCast(SrcAddr.getPointer(), BPP, "ap"), 6282 SrcAddr.getAlignment()); 6283 6284 Value *ArgPtr = Builder.CreateLoad(SrcAddr, "ap.val"); 6285 return Builder.CreateStore(ArgPtr, DestAddr); 6286 } 6287 6288 SmallVector<Value*, 4> Ops; 6289 6290 // Find out if any arguments are required to be integer constant expressions. 6291 unsigned ICEArguments = 0; 6292 ASTContext::GetBuiltinTypeError Error; 6293 getContext().GetBuiltinType(BuiltinID, Error, &ICEArguments); 6294 assert(Error == ASTContext::GE_None && "Should not codegen an error"); 6295 6296 for (unsigned i = 0, e = E->getNumArgs(); i != e; i++) { 6297 // If this is a normal argument, just emit it as a scalar. 6298 if ((ICEArguments & (1 << i)) == 0) { 6299 Ops.push_back(EmitScalarExpr(E->getArg(i))); 6300 continue; 6301 } 6302 6303 // If this is required to be a constant, constant fold it so that we know 6304 // that the generated intrinsic gets a ConstantInt. 6305 llvm::APSInt Result; 6306 bool IsConst = E->getArg(i)->isIntegerConstantExpr(Result, getContext()); 6307 assert(IsConst && "Constant arg isn't actually constant?"); (void)IsConst; 6308 Ops.push_back(llvm::ConstantInt::get(getLLVMContext(), Result)); 6309 } 6310 6311 switch (BuiltinID) { 6312 default: return nullptr; 6313 case X86::BI__builtin_cpu_supports: { 6314 const Expr *FeatureExpr = E->getArg(0)->IgnoreParenCasts(); 6315 StringRef FeatureStr = cast<StringLiteral>(FeatureExpr)->getString(); 6316 6317 // TODO: When/if this becomes more than x86 specific then use a TargetInfo 6318 // based mapping. 6319 // Processor features and mapping to processor feature value. 6320 enum X86Features { 6321 CMOV = 0, 6322 MMX, 6323 POPCNT, 6324 SSE, 6325 SSE2, 6326 SSE3, 6327 SSSE3, 6328 SSE4_1, 6329 SSE4_2, 6330 AVX, 6331 AVX2, 6332 SSE4_A, 6333 FMA4, 6334 XOP, 6335 FMA, 6336 AVX512F, 6337 BMI, 6338 BMI2, 6339 MAX 6340 }; 6341 6342 X86Features Feature = StringSwitch<X86Features>(FeatureStr) 6343 .Case("cmov", X86Features::CMOV) 6344 .Case("mmx", X86Features::MMX) 6345 .Case("popcnt", X86Features::POPCNT) 6346 .Case("sse", X86Features::SSE) 6347 .Case("sse2", X86Features::SSE2) 6348 .Case("sse3", X86Features::SSE3) 6349 .Case("sse4.1", X86Features::SSE4_1) 6350 .Case("sse4.2", X86Features::SSE4_2) 6351 .Case("avx", X86Features::AVX) 6352 .Case("avx2", X86Features::AVX2) 6353 .Case("sse4a", X86Features::SSE4_A) 6354 .Case("fma4", X86Features::FMA4) 6355 .Case("xop", X86Features::XOP) 6356 .Case("fma", X86Features::FMA) 6357 .Case("avx512f", X86Features::AVX512F) 6358 .Case("bmi", X86Features::BMI) 6359 .Case("bmi2", X86Features::BMI2) 6360 .Default(X86Features::MAX); 6361 assert(Feature != X86Features::MAX && "Invalid feature!"); 6362 6363 // Matching the struct layout from the compiler-rt/libgcc structure that is 6364 // filled in: 6365 // unsigned int __cpu_vendor; 6366 // unsigned int __cpu_type; 6367 // unsigned int __cpu_subtype; 6368 // unsigned int __cpu_features[1]; 6369 llvm::Type *STy = llvm::StructType::get( 6370 Int32Ty, Int32Ty, Int32Ty, llvm::ArrayType::get(Int32Ty, 1), nullptr); 6371 6372 // Grab the global __cpu_model. 6373 llvm::Constant *CpuModel = CGM.CreateRuntimeVariable(STy, "__cpu_model"); 6374 6375 // Grab the first (0th) element from the field __cpu_features off of the 6376 // global in the struct STy. 6377 Value *Idxs[] = { 6378 ConstantInt::get(Int32Ty, 0), 6379 ConstantInt::get(Int32Ty, 3), 6380 ConstantInt::get(Int32Ty, 0) 6381 }; 6382 Value *CpuFeatures = Builder.CreateGEP(STy, CpuModel, Idxs); 6383 Value *Features = Builder.CreateAlignedLoad(CpuFeatures, 6384 CharUnits::fromQuantity(4)); 6385 6386 // Check the value of the bit corresponding to the feature requested. 6387 Value *Bitset = Builder.CreateAnd( 6388 Features, llvm::ConstantInt::get(Int32Ty, 1ULL << Feature)); 6389 return Builder.CreateICmpNE(Bitset, llvm::ConstantInt::get(Int32Ty, 0)); 6390 } 6391 case X86::BI_mm_prefetch: { 6392 Value *Address = Ops[0]; 6393 Value *RW = ConstantInt::get(Int32Ty, 0); 6394 Value *Locality = Ops[1]; 6395 Value *Data = ConstantInt::get(Int32Ty, 1); 6396 Value *F = CGM.getIntrinsic(Intrinsic::prefetch); 6397 return Builder.CreateCall(F, {Address, RW, Locality, Data}); 6398 } 6399 case X86::BI__builtin_ia32_undef128: 6400 case X86::BI__builtin_ia32_undef256: 6401 case X86::BI__builtin_ia32_undef512: 6402 return UndefValue::get(ConvertType(E->getType())); 6403 case X86::BI__builtin_ia32_vec_init_v8qi: 6404 case X86::BI__builtin_ia32_vec_init_v4hi: 6405 case X86::BI__builtin_ia32_vec_init_v2si: 6406 return Builder.CreateBitCast(BuildVector(Ops), 6407 llvm::Type::getX86_MMXTy(getLLVMContext())); 6408 case X86::BI__builtin_ia32_vec_ext_v2si: 6409 return Builder.CreateExtractElement(Ops[0], 6410 llvm::ConstantInt::get(Ops[1]->getType(), 0)); 6411 case X86::BI__builtin_ia32_ldmxcsr: { 6412 Address Tmp = CreateMemTemp(E->getArg(0)->getType()); 6413 Builder.CreateStore(Ops[0], Tmp); 6414 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::x86_sse_ldmxcsr), 6415 Builder.CreateBitCast(Tmp.getPointer(), Int8PtrTy)); 6416 } 6417 case X86::BI__builtin_ia32_stmxcsr: { 6418 Address Tmp = CreateMemTemp(E->getType()); 6419 Builder.CreateCall(CGM.getIntrinsic(Intrinsic::x86_sse_stmxcsr), 6420 Builder.CreateBitCast(Tmp.getPointer(), Int8PtrTy)); 6421 return Builder.CreateLoad(Tmp, "stmxcsr"); 6422 } 6423 case X86::BI__builtin_ia32_xsave: 6424 case X86::BI__builtin_ia32_xsave64: 6425 case X86::BI__builtin_ia32_xrstor: 6426 case X86::BI__builtin_ia32_xrstor64: 6427 case X86::BI__builtin_ia32_xsaveopt: 6428 case X86::BI__builtin_ia32_xsaveopt64: 6429 case X86::BI__builtin_ia32_xrstors: 6430 case X86::BI__builtin_ia32_xrstors64: 6431 case X86::BI__builtin_ia32_xsavec: 6432 case X86::BI__builtin_ia32_xsavec64: 6433 case X86::BI__builtin_ia32_xsaves: 6434 case X86::BI__builtin_ia32_xsaves64: { 6435 Intrinsic::ID ID; 6436 #define INTRINSIC_X86_XSAVE_ID(NAME) \ 6437 case X86::BI__builtin_ia32_##NAME: \ 6438 ID = Intrinsic::x86_##NAME; \ 6439 break 6440 switch (BuiltinID) { 6441 default: llvm_unreachable("Unsupported intrinsic!"); 6442 INTRINSIC_X86_XSAVE_ID(xsave); 6443 INTRINSIC_X86_XSAVE_ID(xsave64); 6444 INTRINSIC_X86_XSAVE_ID(xrstor); 6445 INTRINSIC_X86_XSAVE_ID(xrstor64); 6446 INTRINSIC_X86_XSAVE_ID(xsaveopt); 6447 INTRINSIC_X86_XSAVE_ID(xsaveopt64); 6448 INTRINSIC_X86_XSAVE_ID(xrstors); 6449 INTRINSIC_X86_XSAVE_ID(xrstors64); 6450 INTRINSIC_X86_XSAVE_ID(xsavec); 6451 INTRINSIC_X86_XSAVE_ID(xsavec64); 6452 INTRINSIC_X86_XSAVE_ID(xsaves); 6453 INTRINSIC_X86_XSAVE_ID(xsaves64); 6454 } 6455 #undef INTRINSIC_X86_XSAVE_ID 6456 Value *Mhi = Builder.CreateTrunc( 6457 Builder.CreateLShr(Ops[1], ConstantInt::get(Int64Ty, 32)), Int32Ty); 6458 Value *Mlo = Builder.CreateTrunc(Ops[1], Int32Ty); 6459 Ops[1] = Mhi; 6460 Ops.push_back(Mlo); 6461 return Builder.CreateCall(CGM.getIntrinsic(ID), Ops); 6462 } 6463 case X86::BI__builtin_ia32_storehps: 6464 case X86::BI__builtin_ia32_storelps: { 6465 llvm::Type *PtrTy = llvm::PointerType::getUnqual(Int64Ty); 6466 llvm::Type *VecTy = llvm::VectorType::get(Int64Ty, 2); 6467 6468 // cast val v2i64 6469 Ops[1] = Builder.CreateBitCast(Ops[1], VecTy, "cast"); 6470 6471 // extract (0, 1) 6472 unsigned Index = BuiltinID == X86::BI__builtin_ia32_storelps ? 0 : 1; 6473 llvm::Value *Idx = llvm::ConstantInt::get(SizeTy, Index); 6474 Ops[1] = Builder.CreateExtractElement(Ops[1], Idx, "extract"); 6475 6476 // cast pointer to i64 & store 6477 Ops[0] = Builder.CreateBitCast(Ops[0], PtrTy); 6478 return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]); 6479 } 6480 case X86::BI__builtin_ia32_palignr128: 6481 case X86::BI__builtin_ia32_palignr256: { 6482 unsigned ShiftVal = cast<llvm::ConstantInt>(Ops[2])->getZExtValue(); 6483 6484 unsigned NumElts = 6485 cast<llvm::VectorType>(Ops[0]->getType())->getNumElements(); 6486 assert(NumElts % 16 == 0); 6487 unsigned NumLanes = NumElts / 16; 6488 unsigned NumLaneElts = NumElts / NumLanes; 6489 6490 // If palignr is shifting the pair of vectors more than the size of two 6491 // lanes, emit zero. 6492 if (ShiftVal >= (2 * NumLaneElts)) 6493 return llvm::Constant::getNullValue(ConvertType(E->getType())); 6494 6495 // If palignr is shifting the pair of input vectors more than one lane, 6496 // but less than two lanes, convert to shifting in zeroes. 6497 if (ShiftVal > NumLaneElts) { 6498 ShiftVal -= NumLaneElts; 6499 Ops[1] = Ops[0]; 6500 Ops[0] = llvm::Constant::getNullValue(Ops[0]->getType()); 6501 } 6502 6503 uint32_t Indices[32]; 6504 // 256-bit palignr operates on 128-bit lanes so we need to handle that 6505 for (unsigned l = 0; l != NumElts; l += NumLaneElts) { 6506 for (unsigned i = 0; i != NumLaneElts; ++i) { 6507 unsigned Idx = ShiftVal + i; 6508 if (Idx >= NumLaneElts) 6509 Idx += NumElts - NumLaneElts; // End of lane, switch operand. 6510 Indices[l + i] = Idx + l; 6511 } 6512 } 6513 6514 Value *SV = llvm::ConstantDataVector::get(getLLVMContext(), 6515 makeArrayRef(Indices, NumElts)); 6516 return Builder.CreateShuffleVector(Ops[1], Ops[0], SV, "palignr"); 6517 } 6518 case X86::BI__builtin_ia32_pslldqi256: { 6519 // Shift value is in bits so divide by 8. 6520 unsigned shiftVal = cast<llvm::ConstantInt>(Ops[1])->getZExtValue() >> 3; 6521 6522 // If pslldq is shifting the vector more than 15 bytes, emit zero. 6523 if (shiftVal >= 16) 6524 return llvm::Constant::getNullValue(ConvertType(E->getType())); 6525 6526 uint32_t Indices[32]; 6527 // 256-bit pslldq operates on 128-bit lanes so we need to handle that 6528 for (unsigned l = 0; l != 32; l += 16) { 6529 for (unsigned i = 0; i != 16; ++i) { 6530 unsigned Idx = 32 + i - shiftVal; 6531 if (Idx < 32) Idx -= 16; // end of lane, switch operand. 6532 Indices[l + i] = Idx + l; 6533 } 6534 } 6535 6536 llvm::Type *VecTy = llvm::VectorType::get(Int8Ty, 32); 6537 Ops[0] = Builder.CreateBitCast(Ops[0], VecTy, "cast"); 6538 Value *Zero = llvm::Constant::getNullValue(VecTy); 6539 6540 Value *SV = llvm::ConstantDataVector::get(getLLVMContext(), Indices); 6541 SV = Builder.CreateShuffleVector(Zero, Ops[0], SV, "pslldq"); 6542 llvm::Type *ResultType = ConvertType(E->getType()); 6543 return Builder.CreateBitCast(SV, ResultType, "cast"); 6544 } 6545 case X86::BI__builtin_ia32_psrldqi256: { 6546 // Shift value is in bits so divide by 8. 6547 unsigned shiftVal = cast<llvm::ConstantInt>(Ops[1])->getZExtValue() >> 3; 6548 6549 // If psrldq is shifting the vector more than 15 bytes, emit zero. 6550 if (shiftVal >= 16) 6551 return llvm::Constant::getNullValue(ConvertType(E->getType())); 6552 6553 uint32_t Indices[32]; 6554 // 256-bit psrldq operates on 128-bit lanes so we need to handle that 6555 for (unsigned l = 0; l != 32; l += 16) { 6556 for (unsigned i = 0; i != 16; ++i) { 6557 unsigned Idx = i + shiftVal; 6558 if (Idx >= 16) Idx += 16; // end of lane, switch operand. 6559 Indices[l + i] = Idx + l; 6560 } 6561 } 6562 6563 llvm::Type *VecTy = llvm::VectorType::get(Int8Ty, 32); 6564 Ops[0] = Builder.CreateBitCast(Ops[0], VecTy, "cast"); 6565 Value *Zero = llvm::Constant::getNullValue(VecTy); 6566 6567 Value *SV = llvm::ConstantDataVector::get(getLLVMContext(), Indices); 6568 SV = Builder.CreateShuffleVector(Ops[0], Zero, SV, "psrldq"); 6569 llvm::Type *ResultType = ConvertType(E->getType()); 6570 return Builder.CreateBitCast(SV, ResultType, "cast"); 6571 } 6572 case X86::BI__builtin_ia32_movntps: 6573 case X86::BI__builtin_ia32_movntps256: 6574 case X86::BI__builtin_ia32_movntpd: 6575 case X86::BI__builtin_ia32_movntpd256: 6576 case X86::BI__builtin_ia32_movntdq: 6577 case X86::BI__builtin_ia32_movntdq256: 6578 case X86::BI__builtin_ia32_movnti: 6579 case X86::BI__builtin_ia32_movnti64: { 6580 llvm::MDNode *Node = llvm::MDNode::get( 6581 getLLVMContext(), llvm::ConstantAsMetadata::get(Builder.getInt32(1))); 6582 6583 // Convert the type of the pointer to a pointer to the stored type. 6584 Value *BC = Builder.CreateBitCast(Ops[0], 6585 llvm::PointerType::getUnqual(Ops[1]->getType()), 6586 "cast"); 6587 StoreInst *SI = Builder.CreateDefaultAlignedStore(Ops[1], BC); 6588 SI->setMetadata(CGM.getModule().getMDKindID("nontemporal"), Node); 6589 6590 // If the operand is an integer, we can't assume alignment. Otherwise, 6591 // assume natural alignment. 6592 QualType ArgTy = E->getArg(1)->getType(); 6593 unsigned Align; 6594 if (ArgTy->isIntegerType()) 6595 Align = 1; 6596 else 6597 Align = getContext().getTypeSizeInChars(ArgTy).getQuantity(); 6598 SI->setAlignment(Align); 6599 return SI; 6600 } 6601 // 3DNow! 6602 case X86::BI__builtin_ia32_pswapdsf: 6603 case X86::BI__builtin_ia32_pswapdsi: { 6604 llvm::Type *MMXTy = llvm::Type::getX86_MMXTy(getLLVMContext()); 6605 Ops[0] = Builder.CreateBitCast(Ops[0], MMXTy, "cast"); 6606 llvm::Function *F = CGM.getIntrinsic(Intrinsic::x86_3dnowa_pswapd); 6607 return Builder.CreateCall(F, Ops, "pswapd"); 6608 } 6609 case X86::BI__builtin_ia32_rdrand16_step: 6610 case X86::BI__builtin_ia32_rdrand32_step: 6611 case X86::BI__builtin_ia32_rdrand64_step: 6612 case X86::BI__builtin_ia32_rdseed16_step: 6613 case X86::BI__builtin_ia32_rdseed32_step: 6614 case X86::BI__builtin_ia32_rdseed64_step: { 6615 Intrinsic::ID ID; 6616 switch (BuiltinID) { 6617 default: llvm_unreachable("Unsupported intrinsic!"); 6618 case X86::BI__builtin_ia32_rdrand16_step: 6619 ID = Intrinsic::x86_rdrand_16; 6620 break; 6621 case X86::BI__builtin_ia32_rdrand32_step: 6622 ID = Intrinsic::x86_rdrand_32; 6623 break; 6624 case X86::BI__builtin_ia32_rdrand64_step: 6625 ID = Intrinsic::x86_rdrand_64; 6626 break; 6627 case X86::BI__builtin_ia32_rdseed16_step: 6628 ID = Intrinsic::x86_rdseed_16; 6629 break; 6630 case X86::BI__builtin_ia32_rdseed32_step: 6631 ID = Intrinsic::x86_rdseed_32; 6632 break; 6633 case X86::BI__builtin_ia32_rdseed64_step: 6634 ID = Intrinsic::x86_rdseed_64; 6635 break; 6636 } 6637 6638 Value *Call = Builder.CreateCall(CGM.getIntrinsic(ID)); 6639 Builder.CreateDefaultAlignedStore(Builder.CreateExtractValue(Call, 0), 6640 Ops[0]); 6641 return Builder.CreateExtractValue(Call, 1); 6642 } 6643 // SSE comparison intrisics 6644 case X86::BI__builtin_ia32_cmpeqps: 6645 case X86::BI__builtin_ia32_cmpltps: 6646 case X86::BI__builtin_ia32_cmpleps: 6647 case X86::BI__builtin_ia32_cmpunordps: 6648 case X86::BI__builtin_ia32_cmpneqps: 6649 case X86::BI__builtin_ia32_cmpnltps: 6650 case X86::BI__builtin_ia32_cmpnleps: 6651 case X86::BI__builtin_ia32_cmpordps: 6652 case X86::BI__builtin_ia32_cmpeqss: 6653 case X86::BI__builtin_ia32_cmpltss: 6654 case X86::BI__builtin_ia32_cmpless: 6655 case X86::BI__builtin_ia32_cmpunordss: 6656 case X86::BI__builtin_ia32_cmpneqss: 6657 case X86::BI__builtin_ia32_cmpnltss: 6658 case X86::BI__builtin_ia32_cmpnless: 6659 case X86::BI__builtin_ia32_cmpordss: 6660 case X86::BI__builtin_ia32_cmpeqpd: 6661 case X86::BI__builtin_ia32_cmpltpd: 6662 case X86::BI__builtin_ia32_cmplepd: 6663 case X86::BI__builtin_ia32_cmpunordpd: 6664 case X86::BI__builtin_ia32_cmpneqpd: 6665 case X86::BI__builtin_ia32_cmpnltpd: 6666 case X86::BI__builtin_ia32_cmpnlepd: 6667 case X86::BI__builtin_ia32_cmpordpd: 6668 case X86::BI__builtin_ia32_cmpeqsd: 6669 case X86::BI__builtin_ia32_cmpltsd: 6670 case X86::BI__builtin_ia32_cmplesd: 6671 case X86::BI__builtin_ia32_cmpunordsd: 6672 case X86::BI__builtin_ia32_cmpneqsd: 6673 case X86::BI__builtin_ia32_cmpnltsd: 6674 case X86::BI__builtin_ia32_cmpnlesd: 6675 case X86::BI__builtin_ia32_cmpordsd: 6676 // These exist so that the builtin that takes an immediate can be bounds 6677 // checked by clang to avoid passing bad immediates to the backend. Since 6678 // AVX has a larger immediate than SSE we would need separate builtins to 6679 // do the different bounds checking. Rather than create a clang specific 6680 // SSE only builtin, this implements eight separate builtins to match gcc 6681 // implementation. 6682 6683 // Choose the immediate. 6684 unsigned Imm; 6685 switch (BuiltinID) { 6686 default: llvm_unreachable("Unsupported intrinsic!"); 6687 case X86::BI__builtin_ia32_cmpeqps: 6688 case X86::BI__builtin_ia32_cmpeqss: 6689 case X86::BI__builtin_ia32_cmpeqpd: 6690 case X86::BI__builtin_ia32_cmpeqsd: 6691 Imm = 0; 6692 break; 6693 case X86::BI__builtin_ia32_cmpltps: 6694 case X86::BI__builtin_ia32_cmpltss: 6695 case X86::BI__builtin_ia32_cmpltpd: 6696 case X86::BI__builtin_ia32_cmpltsd: 6697 Imm = 1; 6698 break; 6699 case X86::BI__builtin_ia32_cmpleps: 6700 case X86::BI__builtin_ia32_cmpless: 6701 case X86::BI__builtin_ia32_cmplepd: 6702 case X86::BI__builtin_ia32_cmplesd: 6703 Imm = 2; 6704 break; 6705 case X86::BI__builtin_ia32_cmpunordps: 6706 case X86::BI__builtin_ia32_cmpunordss: 6707 case X86::BI__builtin_ia32_cmpunordpd: 6708 case X86::BI__builtin_ia32_cmpunordsd: 6709 Imm = 3; 6710 break; 6711 case X86::BI__builtin_ia32_cmpneqps: 6712 case X86::BI__builtin_ia32_cmpneqss: 6713 case X86::BI__builtin_ia32_cmpneqpd: 6714 case X86::BI__builtin_ia32_cmpneqsd: 6715 Imm = 4; 6716 break; 6717 case X86::BI__builtin_ia32_cmpnltps: 6718 case X86::BI__builtin_ia32_cmpnltss: 6719 case X86::BI__builtin_ia32_cmpnltpd: 6720 case X86::BI__builtin_ia32_cmpnltsd: 6721 Imm = 5; 6722 break; 6723 case X86::BI__builtin_ia32_cmpnleps: 6724 case X86::BI__builtin_ia32_cmpnless: 6725 case X86::BI__builtin_ia32_cmpnlepd: 6726 case X86::BI__builtin_ia32_cmpnlesd: 6727 Imm = 6; 6728 break; 6729 case X86::BI__builtin_ia32_cmpordps: 6730 case X86::BI__builtin_ia32_cmpordss: 6731 case X86::BI__builtin_ia32_cmpordpd: 6732 case X86::BI__builtin_ia32_cmpordsd: 6733 Imm = 7; 6734 break; 6735 } 6736 6737 // Choose the intrinsic ID. 6738 const char *name; 6739 Intrinsic::ID ID; 6740 switch (BuiltinID) { 6741 default: llvm_unreachable("Unsupported intrinsic!"); 6742 case X86::BI__builtin_ia32_cmpeqps: 6743 case X86::BI__builtin_ia32_cmpltps: 6744 case X86::BI__builtin_ia32_cmpleps: 6745 case X86::BI__builtin_ia32_cmpunordps: 6746 case X86::BI__builtin_ia32_cmpneqps: 6747 case X86::BI__builtin_ia32_cmpnltps: 6748 case X86::BI__builtin_ia32_cmpnleps: 6749 case X86::BI__builtin_ia32_cmpordps: 6750 name = "cmpps"; 6751 ID = Intrinsic::x86_sse_cmp_ps; 6752 break; 6753 case X86::BI__builtin_ia32_cmpeqss: 6754 case X86::BI__builtin_ia32_cmpltss: 6755 case X86::BI__builtin_ia32_cmpless: 6756 case X86::BI__builtin_ia32_cmpunordss: 6757 case X86::BI__builtin_ia32_cmpneqss: 6758 case X86::BI__builtin_ia32_cmpnltss: 6759 case X86::BI__builtin_ia32_cmpnless: 6760 case X86::BI__builtin_ia32_cmpordss: 6761 name = "cmpss"; 6762 ID = Intrinsic::x86_sse_cmp_ss; 6763 break; 6764 case X86::BI__builtin_ia32_cmpeqpd: 6765 case X86::BI__builtin_ia32_cmpltpd: 6766 case X86::BI__builtin_ia32_cmplepd: 6767 case X86::BI__builtin_ia32_cmpunordpd: 6768 case X86::BI__builtin_ia32_cmpneqpd: 6769 case X86::BI__builtin_ia32_cmpnltpd: 6770 case X86::BI__builtin_ia32_cmpnlepd: 6771 case X86::BI__builtin_ia32_cmpordpd: 6772 name = "cmppd"; 6773 ID = Intrinsic::x86_sse2_cmp_pd; 6774 break; 6775 case X86::BI__builtin_ia32_cmpeqsd: 6776 case X86::BI__builtin_ia32_cmpltsd: 6777 case X86::BI__builtin_ia32_cmplesd: 6778 case X86::BI__builtin_ia32_cmpunordsd: 6779 case X86::BI__builtin_ia32_cmpneqsd: 6780 case X86::BI__builtin_ia32_cmpnltsd: 6781 case X86::BI__builtin_ia32_cmpnlesd: 6782 case X86::BI__builtin_ia32_cmpordsd: 6783 name = "cmpsd"; 6784 ID = Intrinsic::x86_sse2_cmp_sd; 6785 break; 6786 } 6787 6788 Ops.push_back(llvm::ConstantInt::get(Int8Ty, Imm)); 6789 llvm::Function *F = CGM.getIntrinsic(ID); 6790 return Builder.CreateCall(F, Ops, name); 6791 } 6792 } 6793 6794 6795 Value *CodeGenFunction::EmitPPCBuiltinExpr(unsigned BuiltinID, 6796 const CallExpr *E) { 6797 SmallVector<Value*, 4> Ops; 6798 6799 for (unsigned i = 0, e = E->getNumArgs(); i != e; i++) 6800 Ops.push_back(EmitScalarExpr(E->getArg(i))); 6801 6802 Intrinsic::ID ID = Intrinsic::not_intrinsic; 6803 6804 switch (BuiltinID) { 6805 default: return nullptr; 6806 6807 // __builtin_ppc_get_timebase is GCC 4.8+'s PowerPC-specific name for what we 6808 // call __builtin_readcyclecounter. 6809 case PPC::BI__builtin_ppc_get_timebase: 6810 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::readcyclecounter)); 6811 6812 // vec_ld, vec_lvsl, vec_lvsr 6813 case PPC::BI__builtin_altivec_lvx: 6814 case PPC::BI__builtin_altivec_lvxl: 6815 case PPC::BI__builtin_altivec_lvebx: 6816 case PPC::BI__builtin_altivec_lvehx: 6817 case PPC::BI__builtin_altivec_lvewx: 6818 case PPC::BI__builtin_altivec_lvsl: 6819 case PPC::BI__builtin_altivec_lvsr: 6820 case PPC::BI__builtin_vsx_lxvd2x: 6821 case PPC::BI__builtin_vsx_lxvw4x: 6822 { 6823 Ops[1] = Builder.CreateBitCast(Ops[1], Int8PtrTy); 6824 6825 Ops[0] = Builder.CreateGEP(Ops[1], Ops[0]); 6826 Ops.pop_back(); 6827 6828 switch (BuiltinID) { 6829 default: llvm_unreachable("Unsupported ld/lvsl/lvsr intrinsic!"); 6830 case PPC::BI__builtin_altivec_lvx: 6831 ID = Intrinsic::ppc_altivec_lvx; 6832 break; 6833 case PPC::BI__builtin_altivec_lvxl: 6834 ID = Intrinsic::ppc_altivec_lvxl; 6835 break; 6836 case PPC::BI__builtin_altivec_lvebx: 6837 ID = Intrinsic::ppc_altivec_lvebx; 6838 break; 6839 case PPC::BI__builtin_altivec_lvehx: 6840 ID = Intrinsic::ppc_altivec_lvehx; 6841 break; 6842 case PPC::BI__builtin_altivec_lvewx: 6843 ID = Intrinsic::ppc_altivec_lvewx; 6844 break; 6845 case PPC::BI__builtin_altivec_lvsl: 6846 ID = Intrinsic::ppc_altivec_lvsl; 6847 break; 6848 case PPC::BI__builtin_altivec_lvsr: 6849 ID = Intrinsic::ppc_altivec_lvsr; 6850 break; 6851 case PPC::BI__builtin_vsx_lxvd2x: 6852 ID = Intrinsic::ppc_vsx_lxvd2x; 6853 break; 6854 case PPC::BI__builtin_vsx_lxvw4x: 6855 ID = Intrinsic::ppc_vsx_lxvw4x; 6856 break; 6857 } 6858 llvm::Function *F = CGM.getIntrinsic(ID); 6859 return Builder.CreateCall(F, Ops, ""); 6860 } 6861 6862 // vec_st 6863 case PPC::BI__builtin_altivec_stvx: 6864 case PPC::BI__builtin_altivec_stvxl: 6865 case PPC::BI__builtin_altivec_stvebx: 6866 case PPC::BI__builtin_altivec_stvehx: 6867 case PPC::BI__builtin_altivec_stvewx: 6868 case PPC::BI__builtin_vsx_stxvd2x: 6869 case PPC::BI__builtin_vsx_stxvw4x: 6870 { 6871 Ops[2] = Builder.CreateBitCast(Ops[2], Int8PtrTy); 6872 Ops[1] = Builder.CreateGEP(Ops[2], Ops[1]); 6873 Ops.pop_back(); 6874 6875 switch (BuiltinID) { 6876 default: llvm_unreachable("Unsupported st intrinsic!"); 6877 case PPC::BI__builtin_altivec_stvx: 6878 ID = Intrinsic::ppc_altivec_stvx; 6879 break; 6880 case PPC::BI__builtin_altivec_stvxl: 6881 ID = Intrinsic::ppc_altivec_stvxl; 6882 break; 6883 case PPC::BI__builtin_altivec_stvebx: 6884 ID = Intrinsic::ppc_altivec_stvebx; 6885 break; 6886 case PPC::BI__builtin_altivec_stvehx: 6887 ID = Intrinsic::ppc_altivec_stvehx; 6888 break; 6889 case PPC::BI__builtin_altivec_stvewx: 6890 ID = Intrinsic::ppc_altivec_stvewx; 6891 break; 6892 case PPC::BI__builtin_vsx_stxvd2x: 6893 ID = Intrinsic::ppc_vsx_stxvd2x; 6894 break; 6895 case PPC::BI__builtin_vsx_stxvw4x: 6896 ID = Intrinsic::ppc_vsx_stxvw4x; 6897 break; 6898 } 6899 llvm::Function *F = CGM.getIntrinsic(ID); 6900 return Builder.CreateCall(F, Ops, ""); 6901 } 6902 // Square root 6903 case PPC::BI__builtin_vsx_xvsqrtsp: 6904 case PPC::BI__builtin_vsx_xvsqrtdp: { 6905 llvm::Type *ResultType = ConvertType(E->getType()); 6906 Value *X = EmitScalarExpr(E->getArg(0)); 6907 ID = Intrinsic::sqrt; 6908 llvm::Function *F = CGM.getIntrinsic(ID, ResultType); 6909 return Builder.CreateCall(F, X); 6910 } 6911 // Count leading zeros 6912 case PPC::BI__builtin_altivec_vclzb: 6913 case PPC::BI__builtin_altivec_vclzh: 6914 case PPC::BI__builtin_altivec_vclzw: 6915 case PPC::BI__builtin_altivec_vclzd: { 6916 llvm::Type *ResultType = ConvertType(E->getType()); 6917 Value *X = EmitScalarExpr(E->getArg(0)); 6918 Value *Undef = ConstantInt::get(Builder.getInt1Ty(), false); 6919 Function *F = CGM.getIntrinsic(Intrinsic::ctlz, ResultType); 6920 return Builder.CreateCall(F, {X, Undef}); 6921 } 6922 // Copy sign 6923 case PPC::BI__builtin_vsx_xvcpsgnsp: 6924 case PPC::BI__builtin_vsx_xvcpsgndp: { 6925 llvm::Type *ResultType = ConvertType(E->getType()); 6926 Value *X = EmitScalarExpr(E->getArg(0)); 6927 Value *Y = EmitScalarExpr(E->getArg(1)); 6928 ID = Intrinsic::copysign; 6929 llvm::Function *F = CGM.getIntrinsic(ID, ResultType); 6930 return Builder.CreateCall(F, {X, Y}); 6931 } 6932 // Rounding/truncation 6933 case PPC::BI__builtin_vsx_xvrspip: 6934 case PPC::BI__builtin_vsx_xvrdpip: 6935 case PPC::BI__builtin_vsx_xvrdpim: 6936 case PPC::BI__builtin_vsx_xvrspim: 6937 case PPC::BI__builtin_vsx_xvrdpi: 6938 case PPC::BI__builtin_vsx_xvrspi: 6939 case PPC::BI__builtin_vsx_xvrdpic: 6940 case PPC::BI__builtin_vsx_xvrspic: 6941 case PPC::BI__builtin_vsx_xvrdpiz: 6942 case PPC::BI__builtin_vsx_xvrspiz: { 6943 llvm::Type *ResultType = ConvertType(E->getType()); 6944 Value *X = EmitScalarExpr(E->getArg(0)); 6945 if (BuiltinID == PPC::BI__builtin_vsx_xvrdpim || 6946 BuiltinID == PPC::BI__builtin_vsx_xvrspim) 6947 ID = Intrinsic::floor; 6948 else if (BuiltinID == PPC::BI__builtin_vsx_xvrdpi || 6949 BuiltinID == PPC::BI__builtin_vsx_xvrspi) 6950 ID = Intrinsic::round; 6951 else if (BuiltinID == PPC::BI__builtin_vsx_xvrdpic || 6952 BuiltinID == PPC::BI__builtin_vsx_xvrspic) 6953 ID = Intrinsic::nearbyint; 6954 else if (BuiltinID == PPC::BI__builtin_vsx_xvrdpip || 6955 BuiltinID == PPC::BI__builtin_vsx_xvrspip) 6956 ID = Intrinsic::ceil; 6957 else if (BuiltinID == PPC::BI__builtin_vsx_xvrdpiz || 6958 BuiltinID == PPC::BI__builtin_vsx_xvrspiz) 6959 ID = Intrinsic::trunc; 6960 llvm::Function *F = CGM.getIntrinsic(ID, ResultType); 6961 return Builder.CreateCall(F, X); 6962 } 6963 6964 // Absolute value 6965 case PPC::BI__builtin_vsx_xvabsdp: 6966 case PPC::BI__builtin_vsx_xvabssp: { 6967 llvm::Type *ResultType = ConvertType(E->getType()); 6968 Value *X = EmitScalarExpr(E->getArg(0)); 6969 llvm::Function *F = CGM.getIntrinsic(Intrinsic::fabs, ResultType); 6970 return Builder.CreateCall(F, X); 6971 } 6972 6973 // FMA variations 6974 case PPC::BI__builtin_vsx_xvmaddadp: 6975 case PPC::BI__builtin_vsx_xvmaddasp: 6976 case PPC::BI__builtin_vsx_xvnmaddadp: 6977 case PPC::BI__builtin_vsx_xvnmaddasp: 6978 case PPC::BI__builtin_vsx_xvmsubadp: 6979 case PPC::BI__builtin_vsx_xvmsubasp: 6980 case PPC::BI__builtin_vsx_xvnmsubadp: 6981 case PPC::BI__builtin_vsx_xvnmsubasp: { 6982 llvm::Type *ResultType = ConvertType(E->getType()); 6983 Value *X = EmitScalarExpr(E->getArg(0)); 6984 Value *Y = EmitScalarExpr(E->getArg(1)); 6985 Value *Z = EmitScalarExpr(E->getArg(2)); 6986 Value *Zero = llvm::ConstantFP::getZeroValueForNegation(ResultType); 6987 llvm::Function *F = CGM.getIntrinsic(Intrinsic::fma, ResultType); 6988 switch (BuiltinID) { 6989 case PPC::BI__builtin_vsx_xvmaddadp: 6990 case PPC::BI__builtin_vsx_xvmaddasp: 6991 return Builder.CreateCall(F, {X, Y, Z}); 6992 case PPC::BI__builtin_vsx_xvnmaddadp: 6993 case PPC::BI__builtin_vsx_xvnmaddasp: 6994 return Builder.CreateFSub(Zero, 6995 Builder.CreateCall(F, {X, Y, Z}), "sub"); 6996 case PPC::BI__builtin_vsx_xvmsubadp: 6997 case PPC::BI__builtin_vsx_xvmsubasp: 6998 return Builder.CreateCall(F, 6999 {X, Y, Builder.CreateFSub(Zero, Z, "sub")}); 7000 case PPC::BI__builtin_vsx_xvnmsubadp: 7001 case PPC::BI__builtin_vsx_xvnmsubasp: 7002 Value *FsubRes = 7003 Builder.CreateCall(F, {X, Y, Builder.CreateFSub(Zero, Z, "sub")}); 7004 return Builder.CreateFSub(Zero, FsubRes, "sub"); 7005 } 7006 llvm_unreachable("Unknown FMA operation"); 7007 return nullptr; // Suppress no-return warning 7008 } 7009 } 7010 } 7011 7012 Value *CodeGenFunction::EmitAMDGPUBuiltinExpr(unsigned BuiltinID, 7013 const CallExpr *E) { 7014 switch (BuiltinID) { 7015 case AMDGPU::BI__builtin_amdgcn_div_scale: 7016 case AMDGPU::BI__builtin_amdgcn_div_scalef: { 7017 // Translate from the intrinsics's struct return to the builtin's out 7018 // argument. 7019 7020 Address FlagOutPtr = EmitPointerWithAlignment(E->getArg(3)); 7021 7022 llvm::Value *X = EmitScalarExpr(E->getArg(0)); 7023 llvm::Value *Y = EmitScalarExpr(E->getArg(1)); 7024 llvm::Value *Z = EmitScalarExpr(E->getArg(2)); 7025 7026 llvm::Value *Callee = CGM.getIntrinsic(Intrinsic::amdgcn_div_scale, 7027 X->getType()); 7028 7029 llvm::Value *Tmp = Builder.CreateCall(Callee, {X, Y, Z}); 7030 7031 llvm::Value *Result = Builder.CreateExtractValue(Tmp, 0); 7032 llvm::Value *Flag = Builder.CreateExtractValue(Tmp, 1); 7033 7034 llvm::Type *RealFlagType 7035 = FlagOutPtr.getPointer()->getType()->getPointerElementType(); 7036 7037 llvm::Value *FlagExt = Builder.CreateZExt(Flag, RealFlagType); 7038 Builder.CreateStore(FlagExt, FlagOutPtr); 7039 return Result; 7040 } 7041 case AMDGPU::BI__builtin_amdgcn_div_fmas: 7042 case AMDGPU::BI__builtin_amdgcn_div_fmasf: { 7043 llvm::Value *Src0 = EmitScalarExpr(E->getArg(0)); 7044 llvm::Value *Src1 = EmitScalarExpr(E->getArg(1)); 7045 llvm::Value *Src2 = EmitScalarExpr(E->getArg(2)); 7046 llvm::Value *Src3 = EmitScalarExpr(E->getArg(3)); 7047 7048 llvm::Value *F = CGM.getIntrinsic(Intrinsic::amdgcn_div_fmas, 7049 Src0->getType()); 7050 llvm::Value *Src3ToBool = Builder.CreateIsNotNull(Src3); 7051 return Builder.CreateCall(F, {Src0, Src1, Src2, Src3ToBool}); 7052 } 7053 case AMDGPU::BI__builtin_amdgcn_div_fixup: 7054 case AMDGPU::BI__builtin_amdgcn_div_fixupf: 7055 return emitTernaryFPBuiltin(*this, E, Intrinsic::amdgcn_div_fixup); 7056 case AMDGPU::BI__builtin_amdgcn_trig_preop: 7057 case AMDGPU::BI__builtin_amdgcn_trig_preopf: 7058 return emitFPIntBuiltin(*this, E, Intrinsic::amdgcn_trig_preop); 7059 case AMDGPU::BI__builtin_amdgcn_rcp: 7060 case AMDGPU::BI__builtin_amdgcn_rcpf: 7061 return emitUnaryBuiltin(*this, E, Intrinsic::amdgcn_rcp); 7062 case AMDGPU::BI__builtin_amdgcn_rsq: 7063 case AMDGPU::BI__builtin_amdgcn_rsqf: 7064 return emitUnaryBuiltin(*this, E, Intrinsic::amdgcn_rsq); 7065 case AMDGPU::BI__builtin_amdgcn_rsq_clamp: 7066 case AMDGPU::BI__builtin_amdgcn_rsq_clampf: 7067 return emitUnaryBuiltin(*this, E, Intrinsic::amdgcn_rsq_clamp); 7068 case AMDGPU::BI__builtin_amdgcn_sinf: 7069 return emitUnaryBuiltin(*this, E, Intrinsic::amdgcn_sin); 7070 case AMDGPU::BI__builtin_amdgcn_cosf: 7071 return emitUnaryBuiltin(*this, E, Intrinsic::amdgcn_cos); 7072 case AMDGPU::BI__builtin_amdgcn_log_clampf: 7073 return emitUnaryBuiltin(*this, E, Intrinsic::amdgcn_log_clamp); 7074 case AMDGPU::BI__builtin_amdgcn_ldexp: 7075 case AMDGPU::BI__builtin_amdgcn_ldexpf: 7076 return emitFPIntBuiltin(*this, E, Intrinsic::amdgcn_ldexp); 7077 case AMDGPU::BI__builtin_amdgcn_frexp_mant: 7078 case AMDGPU::BI__builtin_amdgcn_frexp_mantf: { 7079 return emitUnaryBuiltin(*this, E, Intrinsic::amdgcn_frexp_mant); 7080 } 7081 case AMDGPU::BI__builtin_amdgcn_frexp_exp: 7082 case AMDGPU::BI__builtin_amdgcn_frexp_expf: { 7083 return emitUnaryBuiltin(*this, E, Intrinsic::amdgcn_frexp_exp); 7084 } 7085 case AMDGPU::BI__builtin_amdgcn_class: 7086 case AMDGPU::BI__builtin_amdgcn_classf: 7087 return emitFPIntBuiltin(*this, E, Intrinsic::amdgcn_class); 7088 7089 // Legacy amdgpu prefix 7090 case AMDGPU::BI__builtin_amdgpu_rsq: 7091 case AMDGPU::BI__builtin_amdgpu_rsqf: { 7092 if (getTarget().getTriple().getArch() == Triple::amdgcn) 7093 return emitUnaryBuiltin(*this, E, Intrinsic::amdgcn_rsq); 7094 return emitUnaryBuiltin(*this, E, Intrinsic::r600_rsq); 7095 } 7096 case AMDGPU::BI__builtin_amdgpu_ldexp: 7097 case AMDGPU::BI__builtin_amdgpu_ldexpf: { 7098 if (getTarget().getTriple().getArch() == Triple::amdgcn) 7099 return emitFPIntBuiltin(*this, E, Intrinsic::amdgcn_ldexp); 7100 return emitFPIntBuiltin(*this, E, Intrinsic::AMDGPU_ldexp); 7101 } 7102 default: 7103 return nullptr; 7104 } 7105 } 7106 7107 /// Handle a SystemZ function in which the final argument is a pointer 7108 /// to an int that receives the post-instruction CC value. At the LLVM level 7109 /// this is represented as a function that returns a {result, cc} pair. 7110 static Value *EmitSystemZIntrinsicWithCC(CodeGenFunction &CGF, 7111 unsigned IntrinsicID, 7112 const CallExpr *E) { 7113 unsigned NumArgs = E->getNumArgs() - 1; 7114 SmallVector<Value *, 8> Args(NumArgs); 7115 for (unsigned I = 0; I < NumArgs; ++I) 7116 Args[I] = CGF.EmitScalarExpr(E->getArg(I)); 7117 Address CCPtr = CGF.EmitPointerWithAlignment(E->getArg(NumArgs)); 7118 Value *F = CGF.CGM.getIntrinsic(IntrinsicID); 7119 Value *Call = CGF.Builder.CreateCall(F, Args); 7120 Value *CC = CGF.Builder.CreateExtractValue(Call, 1); 7121 CGF.Builder.CreateStore(CC, CCPtr); 7122 return CGF.Builder.CreateExtractValue(Call, 0); 7123 } 7124 7125 Value *CodeGenFunction::EmitSystemZBuiltinExpr(unsigned BuiltinID, 7126 const CallExpr *E) { 7127 switch (BuiltinID) { 7128 case SystemZ::BI__builtin_tbegin: { 7129 Value *TDB = EmitScalarExpr(E->getArg(0)); 7130 Value *Control = llvm::ConstantInt::get(Int32Ty, 0xff0c); 7131 Value *F = CGM.getIntrinsic(Intrinsic::s390_tbegin); 7132 return Builder.CreateCall(F, {TDB, Control}); 7133 } 7134 case SystemZ::BI__builtin_tbegin_nofloat: { 7135 Value *TDB = EmitScalarExpr(E->getArg(0)); 7136 Value *Control = llvm::ConstantInt::get(Int32Ty, 0xff0c); 7137 Value *F = CGM.getIntrinsic(Intrinsic::s390_tbegin_nofloat); 7138 return Builder.CreateCall(F, {TDB, Control}); 7139 } 7140 case SystemZ::BI__builtin_tbeginc: { 7141 Value *TDB = llvm::ConstantPointerNull::get(Int8PtrTy); 7142 Value *Control = llvm::ConstantInt::get(Int32Ty, 0xff08); 7143 Value *F = CGM.getIntrinsic(Intrinsic::s390_tbeginc); 7144 return Builder.CreateCall(F, {TDB, Control}); 7145 } 7146 case SystemZ::BI__builtin_tabort: { 7147 Value *Data = EmitScalarExpr(E->getArg(0)); 7148 Value *F = CGM.getIntrinsic(Intrinsic::s390_tabort); 7149 return Builder.CreateCall(F, Builder.CreateSExt(Data, Int64Ty, "tabort")); 7150 } 7151 case SystemZ::BI__builtin_non_tx_store: { 7152 Value *Address = EmitScalarExpr(E->getArg(0)); 7153 Value *Data = EmitScalarExpr(E->getArg(1)); 7154 Value *F = CGM.getIntrinsic(Intrinsic::s390_ntstg); 7155 return Builder.CreateCall(F, {Data, Address}); 7156 } 7157 7158 // Vector builtins. Note that most vector builtins are mapped automatically 7159 // to target-specific LLVM intrinsics. The ones handled specially here can 7160 // be represented via standard LLVM IR, which is preferable to enable common 7161 // LLVM optimizations. 7162 7163 case SystemZ::BI__builtin_s390_vpopctb: 7164 case SystemZ::BI__builtin_s390_vpopcth: 7165 case SystemZ::BI__builtin_s390_vpopctf: 7166 case SystemZ::BI__builtin_s390_vpopctg: { 7167 llvm::Type *ResultType = ConvertType(E->getType()); 7168 Value *X = EmitScalarExpr(E->getArg(0)); 7169 Function *F = CGM.getIntrinsic(Intrinsic::ctpop, ResultType); 7170 return Builder.CreateCall(F, X); 7171 } 7172 7173 case SystemZ::BI__builtin_s390_vclzb: 7174 case SystemZ::BI__builtin_s390_vclzh: 7175 case SystemZ::BI__builtin_s390_vclzf: 7176 case SystemZ::BI__builtin_s390_vclzg: { 7177 llvm::Type *ResultType = ConvertType(E->getType()); 7178 Value *X = EmitScalarExpr(E->getArg(0)); 7179 Value *Undef = ConstantInt::get(Builder.getInt1Ty(), false); 7180 Function *F = CGM.getIntrinsic(Intrinsic::ctlz, ResultType); 7181 return Builder.CreateCall(F, {X, Undef}); 7182 } 7183 7184 case SystemZ::BI__builtin_s390_vctzb: 7185 case SystemZ::BI__builtin_s390_vctzh: 7186 case SystemZ::BI__builtin_s390_vctzf: 7187 case SystemZ::BI__builtin_s390_vctzg: { 7188 llvm::Type *ResultType = ConvertType(E->getType()); 7189 Value *X = EmitScalarExpr(E->getArg(0)); 7190 Value *Undef = ConstantInt::get(Builder.getInt1Ty(), false); 7191 Function *F = CGM.getIntrinsic(Intrinsic::cttz, ResultType); 7192 return Builder.CreateCall(F, {X, Undef}); 7193 } 7194 7195 case SystemZ::BI__builtin_s390_vfsqdb: { 7196 llvm::Type *ResultType = ConvertType(E->getType()); 7197 Value *X = EmitScalarExpr(E->getArg(0)); 7198 Function *F = CGM.getIntrinsic(Intrinsic::sqrt, ResultType); 7199 return Builder.CreateCall(F, X); 7200 } 7201 case SystemZ::BI__builtin_s390_vfmadb: { 7202 llvm::Type *ResultType = ConvertType(E->getType()); 7203 Value *X = EmitScalarExpr(E->getArg(0)); 7204 Value *Y = EmitScalarExpr(E->getArg(1)); 7205 Value *Z = EmitScalarExpr(E->getArg(2)); 7206 Function *F = CGM.getIntrinsic(Intrinsic::fma, ResultType); 7207 return Builder.CreateCall(F, {X, Y, Z}); 7208 } 7209 case SystemZ::BI__builtin_s390_vfmsdb: { 7210 llvm::Type *ResultType = ConvertType(E->getType()); 7211 Value *X = EmitScalarExpr(E->getArg(0)); 7212 Value *Y = EmitScalarExpr(E->getArg(1)); 7213 Value *Z = EmitScalarExpr(E->getArg(2)); 7214 Value *Zero = llvm::ConstantFP::getZeroValueForNegation(ResultType); 7215 Function *F = CGM.getIntrinsic(Intrinsic::fma, ResultType); 7216 return Builder.CreateCall(F, {X, Y, Builder.CreateFSub(Zero, Z, "sub")}); 7217 } 7218 case SystemZ::BI__builtin_s390_vflpdb: { 7219 llvm::Type *ResultType = ConvertType(E->getType()); 7220 Value *X = EmitScalarExpr(E->getArg(0)); 7221 Function *F = CGM.getIntrinsic(Intrinsic::fabs, ResultType); 7222 return Builder.CreateCall(F, X); 7223 } 7224 case SystemZ::BI__builtin_s390_vflndb: { 7225 llvm::Type *ResultType = ConvertType(E->getType()); 7226 Value *X = EmitScalarExpr(E->getArg(0)); 7227 Value *Zero = llvm::ConstantFP::getZeroValueForNegation(ResultType); 7228 Function *F = CGM.getIntrinsic(Intrinsic::fabs, ResultType); 7229 return Builder.CreateFSub(Zero, Builder.CreateCall(F, X), "sub"); 7230 } 7231 case SystemZ::BI__builtin_s390_vfidb: { 7232 llvm::Type *ResultType = ConvertType(E->getType()); 7233 Value *X = EmitScalarExpr(E->getArg(0)); 7234 // Constant-fold the M4 and M5 mask arguments. 7235 llvm::APSInt M4, M5; 7236 bool IsConstM4 = E->getArg(1)->isIntegerConstantExpr(M4, getContext()); 7237 bool IsConstM5 = E->getArg(2)->isIntegerConstantExpr(M5, getContext()); 7238 assert(IsConstM4 && IsConstM5 && "Constant arg isn't actually constant?"); 7239 (void)IsConstM4; (void)IsConstM5; 7240 // Check whether this instance of vfidb can be represented via a LLVM 7241 // standard intrinsic. We only support some combinations of M4 and M5. 7242 Intrinsic::ID ID = Intrinsic::not_intrinsic; 7243 switch (M4.getZExtValue()) { 7244 default: break; 7245 case 0: // IEEE-inexact exception allowed 7246 switch (M5.getZExtValue()) { 7247 default: break; 7248 case 0: ID = Intrinsic::rint; break; 7249 } 7250 break; 7251 case 4: // IEEE-inexact exception suppressed 7252 switch (M5.getZExtValue()) { 7253 default: break; 7254 case 0: ID = Intrinsic::nearbyint; break; 7255 case 1: ID = Intrinsic::round; break; 7256 case 5: ID = Intrinsic::trunc; break; 7257 case 6: ID = Intrinsic::ceil; break; 7258 case 7: ID = Intrinsic::floor; break; 7259 } 7260 break; 7261 } 7262 if (ID != Intrinsic::not_intrinsic) { 7263 Function *F = CGM.getIntrinsic(ID, ResultType); 7264 return Builder.CreateCall(F, X); 7265 } 7266 Function *F = CGM.getIntrinsic(Intrinsic::s390_vfidb); 7267 Value *M4Value = llvm::ConstantInt::get(getLLVMContext(), M4); 7268 Value *M5Value = llvm::ConstantInt::get(getLLVMContext(), M5); 7269 return Builder.CreateCall(F, {X, M4Value, M5Value}); 7270 } 7271 7272 // Vector intrisincs that output the post-instruction CC value. 7273 7274 #define INTRINSIC_WITH_CC(NAME) \ 7275 case SystemZ::BI__builtin_##NAME: \ 7276 return EmitSystemZIntrinsicWithCC(*this, Intrinsic::NAME, E) 7277 7278 INTRINSIC_WITH_CC(s390_vpkshs); 7279 INTRINSIC_WITH_CC(s390_vpksfs); 7280 INTRINSIC_WITH_CC(s390_vpksgs); 7281 7282 INTRINSIC_WITH_CC(s390_vpklshs); 7283 INTRINSIC_WITH_CC(s390_vpklsfs); 7284 INTRINSIC_WITH_CC(s390_vpklsgs); 7285 7286 INTRINSIC_WITH_CC(s390_vceqbs); 7287 INTRINSIC_WITH_CC(s390_vceqhs); 7288 INTRINSIC_WITH_CC(s390_vceqfs); 7289 INTRINSIC_WITH_CC(s390_vceqgs); 7290 7291 INTRINSIC_WITH_CC(s390_vchbs); 7292 INTRINSIC_WITH_CC(s390_vchhs); 7293 INTRINSIC_WITH_CC(s390_vchfs); 7294 INTRINSIC_WITH_CC(s390_vchgs); 7295 7296 INTRINSIC_WITH_CC(s390_vchlbs); 7297 INTRINSIC_WITH_CC(s390_vchlhs); 7298 INTRINSIC_WITH_CC(s390_vchlfs); 7299 INTRINSIC_WITH_CC(s390_vchlgs); 7300 7301 INTRINSIC_WITH_CC(s390_vfaebs); 7302 INTRINSIC_WITH_CC(s390_vfaehs); 7303 INTRINSIC_WITH_CC(s390_vfaefs); 7304 7305 INTRINSIC_WITH_CC(s390_vfaezbs); 7306 INTRINSIC_WITH_CC(s390_vfaezhs); 7307 INTRINSIC_WITH_CC(s390_vfaezfs); 7308 7309 INTRINSIC_WITH_CC(s390_vfeebs); 7310 INTRINSIC_WITH_CC(s390_vfeehs); 7311 INTRINSIC_WITH_CC(s390_vfeefs); 7312 7313 INTRINSIC_WITH_CC(s390_vfeezbs); 7314 INTRINSIC_WITH_CC(s390_vfeezhs); 7315 INTRINSIC_WITH_CC(s390_vfeezfs); 7316 7317 INTRINSIC_WITH_CC(s390_vfenebs); 7318 INTRINSIC_WITH_CC(s390_vfenehs); 7319 INTRINSIC_WITH_CC(s390_vfenefs); 7320 7321 INTRINSIC_WITH_CC(s390_vfenezbs); 7322 INTRINSIC_WITH_CC(s390_vfenezhs); 7323 INTRINSIC_WITH_CC(s390_vfenezfs); 7324 7325 INTRINSIC_WITH_CC(s390_vistrbs); 7326 INTRINSIC_WITH_CC(s390_vistrhs); 7327 INTRINSIC_WITH_CC(s390_vistrfs); 7328 7329 INTRINSIC_WITH_CC(s390_vstrcbs); 7330 INTRINSIC_WITH_CC(s390_vstrchs); 7331 INTRINSIC_WITH_CC(s390_vstrcfs); 7332 7333 INTRINSIC_WITH_CC(s390_vstrczbs); 7334 INTRINSIC_WITH_CC(s390_vstrczhs); 7335 INTRINSIC_WITH_CC(s390_vstrczfs); 7336 7337 INTRINSIC_WITH_CC(s390_vfcedbs); 7338 INTRINSIC_WITH_CC(s390_vfchdbs); 7339 INTRINSIC_WITH_CC(s390_vfchedbs); 7340 7341 INTRINSIC_WITH_CC(s390_vftcidb); 7342 7343 #undef INTRINSIC_WITH_CC 7344 7345 default: 7346 return nullptr; 7347 } 7348 } 7349 7350 Value *CodeGenFunction::EmitNVPTXBuiltinExpr(unsigned BuiltinID, 7351 const CallExpr *E) { 7352 switch (BuiltinID) { 7353 case NVPTX::BI__nvvm_atom_add_gen_i: 7354 case NVPTX::BI__nvvm_atom_add_gen_l: 7355 case NVPTX::BI__nvvm_atom_add_gen_ll: 7356 return MakeBinaryAtomicValue(*this, llvm::AtomicRMWInst::Add, E); 7357 7358 case NVPTX::BI__nvvm_atom_sub_gen_i: 7359 case NVPTX::BI__nvvm_atom_sub_gen_l: 7360 case NVPTX::BI__nvvm_atom_sub_gen_ll: 7361 return MakeBinaryAtomicValue(*this, llvm::AtomicRMWInst::Sub, E); 7362 7363 case NVPTX::BI__nvvm_atom_and_gen_i: 7364 case NVPTX::BI__nvvm_atom_and_gen_l: 7365 case NVPTX::BI__nvvm_atom_and_gen_ll: 7366 return MakeBinaryAtomicValue(*this, llvm::AtomicRMWInst::And, E); 7367 7368 case NVPTX::BI__nvvm_atom_or_gen_i: 7369 case NVPTX::BI__nvvm_atom_or_gen_l: 7370 case NVPTX::BI__nvvm_atom_or_gen_ll: 7371 return MakeBinaryAtomicValue(*this, llvm::AtomicRMWInst::Or, E); 7372 7373 case NVPTX::BI__nvvm_atom_xor_gen_i: 7374 case NVPTX::BI__nvvm_atom_xor_gen_l: 7375 case NVPTX::BI__nvvm_atom_xor_gen_ll: 7376 return MakeBinaryAtomicValue(*this, llvm::AtomicRMWInst::Xor, E); 7377 7378 case NVPTX::BI__nvvm_atom_xchg_gen_i: 7379 case NVPTX::BI__nvvm_atom_xchg_gen_l: 7380 case NVPTX::BI__nvvm_atom_xchg_gen_ll: 7381 return MakeBinaryAtomicValue(*this, llvm::AtomicRMWInst::Xchg, E); 7382 7383 case NVPTX::BI__nvvm_atom_max_gen_i: 7384 case NVPTX::BI__nvvm_atom_max_gen_l: 7385 case NVPTX::BI__nvvm_atom_max_gen_ll: 7386 return MakeBinaryAtomicValue(*this, llvm::AtomicRMWInst::Max, E); 7387 7388 case NVPTX::BI__nvvm_atom_max_gen_ui: 7389 case NVPTX::BI__nvvm_atom_max_gen_ul: 7390 case NVPTX::BI__nvvm_atom_max_gen_ull: 7391 return MakeBinaryAtomicValue(*this, llvm::AtomicRMWInst::UMax, E); 7392 7393 case NVPTX::BI__nvvm_atom_min_gen_i: 7394 case NVPTX::BI__nvvm_atom_min_gen_l: 7395 case NVPTX::BI__nvvm_atom_min_gen_ll: 7396 return MakeBinaryAtomicValue(*this, llvm::AtomicRMWInst::Min, E); 7397 7398 case NVPTX::BI__nvvm_atom_min_gen_ui: 7399 case NVPTX::BI__nvvm_atom_min_gen_ul: 7400 case NVPTX::BI__nvvm_atom_min_gen_ull: 7401 return MakeBinaryAtomicValue(*this, llvm::AtomicRMWInst::UMin, E); 7402 7403 case NVPTX::BI__nvvm_atom_cas_gen_i: 7404 case NVPTX::BI__nvvm_atom_cas_gen_l: 7405 case NVPTX::BI__nvvm_atom_cas_gen_ll: 7406 // __nvvm_atom_cas_gen_* should return the old value rather than the 7407 // success flag. 7408 return MakeAtomicCmpXchgValue(*this, E, /*ReturnBool=*/false); 7409 7410 case NVPTX::BI__nvvm_atom_add_gen_f: { 7411 Value *Ptr = EmitScalarExpr(E->getArg(0)); 7412 Value *Val = EmitScalarExpr(E->getArg(1)); 7413 // atomicrmw only deals with integer arguments so we need to use 7414 // LLVM's nvvm_atomic_load_add_f32 intrinsic for that. 7415 Value *FnALAF32 = 7416 CGM.getIntrinsic(Intrinsic::nvvm_atomic_load_add_f32, Ptr->getType()); 7417 return Builder.CreateCall(FnALAF32, {Ptr, Val}); 7418 } 7419 7420 case NVPTX::BI__nvvm_atom_inc_gen_ui: { 7421 Value *Ptr = EmitScalarExpr(E->getArg(0)); 7422 Value *Val = EmitScalarExpr(E->getArg(1)); 7423 Value *FnALI32 = 7424 CGM.getIntrinsic(Intrinsic::nvvm_atomic_load_inc_32, Ptr->getType()); 7425 return Builder.CreateCall(FnALI32, {Ptr, Val}); 7426 } 7427 7428 case NVPTX::BI__nvvm_atom_dec_gen_ui: { 7429 Value *Ptr = EmitScalarExpr(E->getArg(0)); 7430 Value *Val = EmitScalarExpr(E->getArg(1)); 7431 Value *FnALD32 = 7432 CGM.getIntrinsic(Intrinsic::nvvm_atomic_load_dec_32, Ptr->getType()); 7433 return Builder.CreateCall(FnALD32, {Ptr, Val}); 7434 } 7435 7436 default: 7437 return nullptr; 7438 } 7439 } 7440 7441 Value *CodeGenFunction::EmitWebAssemblyBuiltinExpr(unsigned BuiltinID, 7442 const CallExpr *E) { 7443 switch (BuiltinID) { 7444 case WebAssembly::BI__builtin_wasm_current_memory: { 7445 llvm::Type *ResultType = ConvertType(E->getType()); 7446 Value *Callee = CGM.getIntrinsic(Intrinsic::wasm_current_memory, ResultType); 7447 return Builder.CreateCall(Callee); 7448 } 7449 case WebAssembly::BI__builtin_wasm_grow_memory: { 7450 Value *X = EmitScalarExpr(E->getArg(0)); 7451 Value *Callee = CGM.getIntrinsic(Intrinsic::wasm_grow_memory, X->getType()); 7452 return Builder.CreateCall(Callee, X); 7453 } 7454 7455 default: 7456 return nullptr; 7457 } 7458 } 7459