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 // isinf(x) --> fabs(x) == infinity 791 Value *V = EmitScalarExpr(E->getArg(0)); 792 V = EmitFAbs(*this, V); 793 794 V = Builder.CreateFCmpOEQ(V, ConstantFP::getInfinity(V->getType()),"isinf"); 795 return RValue::get(Builder.CreateZExt(V, ConvertType(E->getType()))); 796 } 797 798 case Builtin::BI__builtin_isinf_sign: { 799 // isinf_sign(x) -> fabs(x) == infinity ? (signbit(x) ? -1 : 1) : 0 800 Value *Arg = EmitScalarExpr(E->getArg(0)); 801 Value *AbsArg = EmitFAbs(*this, Arg); 802 Value *IsInf = Builder.CreateFCmpOEQ( 803 AbsArg, ConstantFP::getInfinity(Arg->getType()), "isinf"); 804 Value *IsNeg = EmitSignBit(*this, Arg); 805 806 llvm::Type *IntTy = ConvertType(E->getType()); 807 Value *Zero = Constant::getNullValue(IntTy); 808 Value *One = ConstantInt::get(IntTy, 1); 809 Value *NegativeOne = ConstantInt::get(IntTy, -1); 810 Value *SignResult = Builder.CreateSelect(IsNeg, NegativeOne, One); 811 Value *Result = Builder.CreateSelect(IsInf, SignResult, Zero); 812 return RValue::get(Result); 813 } 814 815 case Builtin::BI__builtin_isnormal: { 816 // isnormal(x) --> x == x && fabsf(x) < infinity && fabsf(x) >= float_min 817 Value *V = EmitScalarExpr(E->getArg(0)); 818 Value *Eq = Builder.CreateFCmpOEQ(V, V, "iseq"); 819 820 Value *Abs = EmitFAbs(*this, V); 821 Value *IsLessThanInf = 822 Builder.CreateFCmpULT(Abs, ConstantFP::getInfinity(V->getType()),"isinf"); 823 APFloat Smallest = APFloat::getSmallestNormalized( 824 getContext().getFloatTypeSemantics(E->getArg(0)->getType())); 825 Value *IsNormal = 826 Builder.CreateFCmpUGE(Abs, ConstantFP::get(V->getContext(), Smallest), 827 "isnormal"); 828 V = Builder.CreateAnd(Eq, IsLessThanInf, "and"); 829 V = Builder.CreateAnd(V, IsNormal, "and"); 830 return RValue::get(Builder.CreateZExt(V, ConvertType(E->getType()))); 831 } 832 833 case Builtin::BI__builtin_isfinite: { 834 // isfinite(x) --> x == x && fabs(x) != infinity; 835 Value *V = EmitScalarExpr(E->getArg(0)); 836 Value *Eq = Builder.CreateFCmpOEQ(V, V, "iseq"); 837 838 Value *Abs = EmitFAbs(*this, V); 839 Value *IsNotInf = 840 Builder.CreateFCmpUNE(Abs, ConstantFP::getInfinity(V->getType()),"isinf"); 841 842 V = Builder.CreateAnd(Eq, IsNotInf, "and"); 843 return RValue::get(Builder.CreateZExt(V, ConvertType(E->getType()))); 844 } 845 846 case Builtin::BI__builtin_fpclassify: { 847 Value *V = EmitScalarExpr(E->getArg(5)); 848 llvm::Type *Ty = ConvertType(E->getArg(5)->getType()); 849 850 // Create Result 851 BasicBlock *Begin = Builder.GetInsertBlock(); 852 BasicBlock *End = createBasicBlock("fpclassify_end", this->CurFn); 853 Builder.SetInsertPoint(End); 854 PHINode *Result = 855 Builder.CreatePHI(ConvertType(E->getArg(0)->getType()), 4, 856 "fpclassify_result"); 857 858 // if (V==0) return FP_ZERO 859 Builder.SetInsertPoint(Begin); 860 Value *IsZero = Builder.CreateFCmpOEQ(V, Constant::getNullValue(Ty), 861 "iszero"); 862 Value *ZeroLiteral = EmitScalarExpr(E->getArg(4)); 863 BasicBlock *NotZero = createBasicBlock("fpclassify_not_zero", this->CurFn); 864 Builder.CreateCondBr(IsZero, End, NotZero); 865 Result->addIncoming(ZeroLiteral, Begin); 866 867 // if (V != V) return FP_NAN 868 Builder.SetInsertPoint(NotZero); 869 Value *IsNan = Builder.CreateFCmpUNO(V, V, "cmp"); 870 Value *NanLiteral = EmitScalarExpr(E->getArg(0)); 871 BasicBlock *NotNan = createBasicBlock("fpclassify_not_nan", this->CurFn); 872 Builder.CreateCondBr(IsNan, End, NotNan); 873 Result->addIncoming(NanLiteral, NotZero); 874 875 // if (fabs(V) == infinity) return FP_INFINITY 876 Builder.SetInsertPoint(NotNan); 877 Value *VAbs = EmitFAbs(*this, V); 878 Value *IsInf = 879 Builder.CreateFCmpOEQ(VAbs, ConstantFP::getInfinity(V->getType()), 880 "isinf"); 881 Value *InfLiteral = EmitScalarExpr(E->getArg(1)); 882 BasicBlock *NotInf = createBasicBlock("fpclassify_not_inf", this->CurFn); 883 Builder.CreateCondBr(IsInf, End, NotInf); 884 Result->addIncoming(InfLiteral, NotNan); 885 886 // if (fabs(V) >= MIN_NORMAL) return FP_NORMAL else FP_SUBNORMAL 887 Builder.SetInsertPoint(NotInf); 888 APFloat Smallest = APFloat::getSmallestNormalized( 889 getContext().getFloatTypeSemantics(E->getArg(5)->getType())); 890 Value *IsNormal = 891 Builder.CreateFCmpUGE(VAbs, ConstantFP::get(V->getContext(), Smallest), 892 "isnormal"); 893 Value *NormalResult = 894 Builder.CreateSelect(IsNormal, EmitScalarExpr(E->getArg(2)), 895 EmitScalarExpr(E->getArg(3))); 896 Builder.CreateBr(End); 897 Result->addIncoming(NormalResult, NotInf); 898 899 // return Result 900 Builder.SetInsertPoint(End); 901 return RValue::get(Result); 902 } 903 904 case Builtin::BIalloca: 905 case Builtin::BI_alloca: 906 case Builtin::BI__builtin_alloca: { 907 Value *Size = EmitScalarExpr(E->getArg(0)); 908 return RValue::get(Builder.CreateAlloca(Builder.getInt8Ty(), Size)); 909 } 910 case Builtin::BIbzero: 911 case Builtin::BI__builtin_bzero: { 912 Address Dest = EmitPointerWithAlignment(E->getArg(0)); 913 Value *SizeVal = EmitScalarExpr(E->getArg(1)); 914 EmitNonNullArgCheck(RValue::get(Dest.getPointer()), E->getArg(0)->getType(), 915 E->getArg(0)->getExprLoc(), FD, 0); 916 Builder.CreateMemSet(Dest, Builder.getInt8(0), SizeVal, false); 917 return RValue::get(Dest.getPointer()); 918 } 919 case Builtin::BImemcpy: 920 case Builtin::BI__builtin_memcpy: { 921 Address Dest = EmitPointerWithAlignment(E->getArg(0)); 922 Address Src = EmitPointerWithAlignment(E->getArg(1)); 923 Value *SizeVal = EmitScalarExpr(E->getArg(2)); 924 EmitNonNullArgCheck(RValue::get(Dest.getPointer()), E->getArg(0)->getType(), 925 E->getArg(0)->getExprLoc(), FD, 0); 926 EmitNonNullArgCheck(RValue::get(Src.getPointer()), E->getArg(1)->getType(), 927 E->getArg(1)->getExprLoc(), FD, 1); 928 Builder.CreateMemCpy(Dest, Src, SizeVal, false); 929 return RValue::get(Dest.getPointer()); 930 } 931 932 case Builtin::BI__builtin___memcpy_chk: { 933 // fold __builtin_memcpy_chk(x, y, cst1, cst2) to memcpy iff cst1<=cst2. 934 llvm::APSInt Size, DstSize; 935 if (!E->getArg(2)->EvaluateAsInt(Size, CGM.getContext()) || 936 !E->getArg(3)->EvaluateAsInt(DstSize, CGM.getContext())) 937 break; 938 if (Size.ugt(DstSize)) 939 break; 940 Address Dest = EmitPointerWithAlignment(E->getArg(0)); 941 Address Src = EmitPointerWithAlignment(E->getArg(1)); 942 Value *SizeVal = llvm::ConstantInt::get(Builder.getContext(), Size); 943 Builder.CreateMemCpy(Dest, Src, SizeVal, false); 944 return RValue::get(Dest.getPointer()); 945 } 946 947 case Builtin::BI__builtin_objc_memmove_collectable: { 948 Address DestAddr = EmitPointerWithAlignment(E->getArg(0)); 949 Address SrcAddr = EmitPointerWithAlignment(E->getArg(1)); 950 Value *SizeVal = EmitScalarExpr(E->getArg(2)); 951 CGM.getObjCRuntime().EmitGCMemmoveCollectable(*this, 952 DestAddr, SrcAddr, SizeVal); 953 return RValue::get(DestAddr.getPointer()); 954 } 955 956 case Builtin::BI__builtin___memmove_chk: { 957 // fold __builtin_memmove_chk(x, y, cst1, cst2) to memmove iff cst1<=cst2. 958 llvm::APSInt Size, DstSize; 959 if (!E->getArg(2)->EvaluateAsInt(Size, CGM.getContext()) || 960 !E->getArg(3)->EvaluateAsInt(DstSize, CGM.getContext())) 961 break; 962 if (Size.ugt(DstSize)) 963 break; 964 Address Dest = EmitPointerWithAlignment(E->getArg(0)); 965 Address Src = EmitPointerWithAlignment(E->getArg(1)); 966 Value *SizeVal = llvm::ConstantInt::get(Builder.getContext(), Size); 967 Builder.CreateMemMove(Dest, Src, SizeVal, false); 968 return RValue::get(Dest.getPointer()); 969 } 970 971 case Builtin::BImemmove: 972 case Builtin::BI__builtin_memmove: { 973 Address Dest = EmitPointerWithAlignment(E->getArg(0)); 974 Address Src = EmitPointerWithAlignment(E->getArg(1)); 975 Value *SizeVal = EmitScalarExpr(E->getArg(2)); 976 EmitNonNullArgCheck(RValue::get(Dest.getPointer()), E->getArg(0)->getType(), 977 E->getArg(0)->getExprLoc(), FD, 0); 978 EmitNonNullArgCheck(RValue::get(Src.getPointer()), E->getArg(1)->getType(), 979 E->getArg(1)->getExprLoc(), FD, 1); 980 Builder.CreateMemMove(Dest, Src, SizeVal, false); 981 return RValue::get(Dest.getPointer()); 982 } 983 case Builtin::BImemset: 984 case Builtin::BI__builtin_memset: { 985 Address Dest = EmitPointerWithAlignment(E->getArg(0)); 986 Value *ByteVal = Builder.CreateTrunc(EmitScalarExpr(E->getArg(1)), 987 Builder.getInt8Ty()); 988 Value *SizeVal = EmitScalarExpr(E->getArg(2)); 989 EmitNonNullArgCheck(RValue::get(Dest.getPointer()), E->getArg(0)->getType(), 990 E->getArg(0)->getExprLoc(), FD, 0); 991 Builder.CreateMemSet(Dest, ByteVal, SizeVal, false); 992 return RValue::get(Dest.getPointer()); 993 } 994 case Builtin::BI__builtin___memset_chk: { 995 // fold __builtin_memset_chk(x, y, cst1, cst2) to memset iff cst1<=cst2. 996 llvm::APSInt Size, DstSize; 997 if (!E->getArg(2)->EvaluateAsInt(Size, CGM.getContext()) || 998 !E->getArg(3)->EvaluateAsInt(DstSize, CGM.getContext())) 999 break; 1000 if (Size.ugt(DstSize)) 1001 break; 1002 Address Dest = EmitPointerWithAlignment(E->getArg(0)); 1003 Value *ByteVal = Builder.CreateTrunc(EmitScalarExpr(E->getArg(1)), 1004 Builder.getInt8Ty()); 1005 Value *SizeVal = llvm::ConstantInt::get(Builder.getContext(), Size); 1006 Builder.CreateMemSet(Dest, ByteVal, SizeVal, false); 1007 return RValue::get(Dest.getPointer()); 1008 } 1009 case Builtin::BI__builtin_dwarf_cfa: { 1010 // The offset in bytes from the first argument to the CFA. 1011 // 1012 // Why on earth is this in the frontend? Is there any reason at 1013 // all that the backend can't reasonably determine this while 1014 // lowering llvm.eh.dwarf.cfa()? 1015 // 1016 // TODO: If there's a satisfactory reason, add a target hook for 1017 // this instead of hard-coding 0, which is correct for most targets. 1018 int32_t Offset = 0; 1019 1020 Value *F = CGM.getIntrinsic(Intrinsic::eh_dwarf_cfa); 1021 return RValue::get(Builder.CreateCall(F, 1022 llvm::ConstantInt::get(Int32Ty, Offset))); 1023 } 1024 case Builtin::BI__builtin_return_address: { 1025 Value *Depth = 1026 CGM.EmitConstantExpr(E->getArg(0), getContext().UnsignedIntTy, this); 1027 Value *F = CGM.getIntrinsic(Intrinsic::returnaddress); 1028 return RValue::get(Builder.CreateCall(F, Depth)); 1029 } 1030 case Builtin::BI__builtin_frame_address: { 1031 Value *Depth = 1032 CGM.EmitConstantExpr(E->getArg(0), getContext().UnsignedIntTy, this); 1033 Value *F = CGM.getIntrinsic(Intrinsic::frameaddress); 1034 return RValue::get(Builder.CreateCall(F, Depth)); 1035 } 1036 case Builtin::BI__builtin_extract_return_addr: { 1037 Value *Address = EmitScalarExpr(E->getArg(0)); 1038 Value *Result = getTargetHooks().decodeReturnAddress(*this, Address); 1039 return RValue::get(Result); 1040 } 1041 case Builtin::BI__builtin_frob_return_addr: { 1042 Value *Address = EmitScalarExpr(E->getArg(0)); 1043 Value *Result = getTargetHooks().encodeReturnAddress(*this, Address); 1044 return RValue::get(Result); 1045 } 1046 case Builtin::BI__builtin_dwarf_sp_column: { 1047 llvm::IntegerType *Ty 1048 = cast<llvm::IntegerType>(ConvertType(E->getType())); 1049 int Column = getTargetHooks().getDwarfEHStackPointer(CGM); 1050 if (Column == -1) { 1051 CGM.ErrorUnsupported(E, "__builtin_dwarf_sp_column"); 1052 return RValue::get(llvm::UndefValue::get(Ty)); 1053 } 1054 return RValue::get(llvm::ConstantInt::get(Ty, Column, true)); 1055 } 1056 case Builtin::BI__builtin_init_dwarf_reg_size_table: { 1057 Value *Address = EmitScalarExpr(E->getArg(0)); 1058 if (getTargetHooks().initDwarfEHRegSizeTable(*this, Address)) 1059 CGM.ErrorUnsupported(E, "__builtin_init_dwarf_reg_size_table"); 1060 return RValue::get(llvm::UndefValue::get(ConvertType(E->getType()))); 1061 } 1062 case Builtin::BI__builtin_eh_return: { 1063 Value *Int = EmitScalarExpr(E->getArg(0)); 1064 Value *Ptr = EmitScalarExpr(E->getArg(1)); 1065 1066 llvm::IntegerType *IntTy = cast<llvm::IntegerType>(Int->getType()); 1067 assert((IntTy->getBitWidth() == 32 || IntTy->getBitWidth() == 64) && 1068 "LLVM's __builtin_eh_return only supports 32- and 64-bit variants"); 1069 Value *F = CGM.getIntrinsic(IntTy->getBitWidth() == 32 1070 ? Intrinsic::eh_return_i32 1071 : Intrinsic::eh_return_i64); 1072 Builder.CreateCall(F, {Int, Ptr}); 1073 Builder.CreateUnreachable(); 1074 1075 // We do need to preserve an insertion point. 1076 EmitBlock(createBasicBlock("builtin_eh_return.cont")); 1077 1078 return RValue::get(nullptr); 1079 } 1080 case Builtin::BI__builtin_unwind_init: { 1081 Value *F = CGM.getIntrinsic(Intrinsic::eh_unwind_init); 1082 return RValue::get(Builder.CreateCall(F)); 1083 } 1084 case Builtin::BI__builtin_extend_pointer: { 1085 // Extends a pointer to the size of an _Unwind_Word, which is 1086 // uint64_t on all platforms. Generally this gets poked into a 1087 // register and eventually used as an address, so if the 1088 // addressing registers are wider than pointers and the platform 1089 // doesn't implicitly ignore high-order bits when doing 1090 // addressing, we need to make sure we zext / sext based on 1091 // the platform's expectations. 1092 // 1093 // See: http://gcc.gnu.org/ml/gcc-bugs/2002-02/msg00237.html 1094 1095 // Cast the pointer to intptr_t. 1096 Value *Ptr = EmitScalarExpr(E->getArg(0)); 1097 Value *Result = Builder.CreatePtrToInt(Ptr, IntPtrTy, "extend.cast"); 1098 1099 // If that's 64 bits, we're done. 1100 if (IntPtrTy->getBitWidth() == 64) 1101 return RValue::get(Result); 1102 1103 // Otherwise, ask the codegen data what to do. 1104 if (getTargetHooks().extendPointerWithSExt()) 1105 return RValue::get(Builder.CreateSExt(Result, Int64Ty, "extend.sext")); 1106 else 1107 return RValue::get(Builder.CreateZExt(Result, Int64Ty, "extend.zext")); 1108 } 1109 case Builtin::BI__builtin_setjmp: { 1110 // Buffer is a void**. 1111 Address Buf = EmitPointerWithAlignment(E->getArg(0)); 1112 1113 // Store the frame pointer to the setjmp buffer. 1114 Value *FrameAddr = 1115 Builder.CreateCall(CGM.getIntrinsic(Intrinsic::frameaddress), 1116 ConstantInt::get(Int32Ty, 0)); 1117 Builder.CreateStore(FrameAddr, Buf); 1118 1119 // Store the stack pointer to the setjmp buffer. 1120 Value *StackAddr = 1121 Builder.CreateCall(CGM.getIntrinsic(Intrinsic::stacksave)); 1122 Address StackSaveSlot = 1123 Builder.CreateConstInBoundsGEP(Buf, 2, getPointerSize()); 1124 Builder.CreateStore(StackAddr, StackSaveSlot); 1125 1126 // Call LLVM's EH setjmp, which is lightweight. 1127 Value *F = CGM.getIntrinsic(Intrinsic::eh_sjlj_setjmp); 1128 Buf = Builder.CreateBitCast(Buf, Int8PtrTy); 1129 return RValue::get(Builder.CreateCall(F, Buf.getPointer())); 1130 } 1131 case Builtin::BI__builtin_longjmp: { 1132 Value *Buf = EmitScalarExpr(E->getArg(0)); 1133 Buf = Builder.CreateBitCast(Buf, Int8PtrTy); 1134 1135 // Call LLVM's EH longjmp, which is lightweight. 1136 Builder.CreateCall(CGM.getIntrinsic(Intrinsic::eh_sjlj_longjmp), Buf); 1137 1138 // longjmp doesn't return; mark this as unreachable. 1139 Builder.CreateUnreachable(); 1140 1141 // We do need to preserve an insertion point. 1142 EmitBlock(createBasicBlock("longjmp.cont")); 1143 1144 return RValue::get(nullptr); 1145 } 1146 case Builtin::BI__sync_fetch_and_add: 1147 case Builtin::BI__sync_fetch_and_sub: 1148 case Builtin::BI__sync_fetch_and_or: 1149 case Builtin::BI__sync_fetch_and_and: 1150 case Builtin::BI__sync_fetch_and_xor: 1151 case Builtin::BI__sync_fetch_and_nand: 1152 case Builtin::BI__sync_add_and_fetch: 1153 case Builtin::BI__sync_sub_and_fetch: 1154 case Builtin::BI__sync_and_and_fetch: 1155 case Builtin::BI__sync_or_and_fetch: 1156 case Builtin::BI__sync_xor_and_fetch: 1157 case Builtin::BI__sync_nand_and_fetch: 1158 case Builtin::BI__sync_val_compare_and_swap: 1159 case Builtin::BI__sync_bool_compare_and_swap: 1160 case Builtin::BI__sync_lock_test_and_set: 1161 case Builtin::BI__sync_lock_release: 1162 case Builtin::BI__sync_swap: 1163 llvm_unreachable("Shouldn't make it through sema"); 1164 case Builtin::BI__sync_fetch_and_add_1: 1165 case Builtin::BI__sync_fetch_and_add_2: 1166 case Builtin::BI__sync_fetch_and_add_4: 1167 case Builtin::BI__sync_fetch_and_add_8: 1168 case Builtin::BI__sync_fetch_and_add_16: 1169 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Add, E); 1170 case Builtin::BI__sync_fetch_and_sub_1: 1171 case Builtin::BI__sync_fetch_and_sub_2: 1172 case Builtin::BI__sync_fetch_and_sub_4: 1173 case Builtin::BI__sync_fetch_and_sub_8: 1174 case Builtin::BI__sync_fetch_and_sub_16: 1175 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Sub, E); 1176 case Builtin::BI__sync_fetch_and_or_1: 1177 case Builtin::BI__sync_fetch_and_or_2: 1178 case Builtin::BI__sync_fetch_and_or_4: 1179 case Builtin::BI__sync_fetch_and_or_8: 1180 case Builtin::BI__sync_fetch_and_or_16: 1181 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Or, E); 1182 case Builtin::BI__sync_fetch_and_and_1: 1183 case Builtin::BI__sync_fetch_and_and_2: 1184 case Builtin::BI__sync_fetch_and_and_4: 1185 case Builtin::BI__sync_fetch_and_and_8: 1186 case Builtin::BI__sync_fetch_and_and_16: 1187 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::And, E); 1188 case Builtin::BI__sync_fetch_and_xor_1: 1189 case Builtin::BI__sync_fetch_and_xor_2: 1190 case Builtin::BI__sync_fetch_and_xor_4: 1191 case Builtin::BI__sync_fetch_and_xor_8: 1192 case Builtin::BI__sync_fetch_and_xor_16: 1193 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Xor, E); 1194 case Builtin::BI__sync_fetch_and_nand_1: 1195 case Builtin::BI__sync_fetch_and_nand_2: 1196 case Builtin::BI__sync_fetch_and_nand_4: 1197 case Builtin::BI__sync_fetch_and_nand_8: 1198 case Builtin::BI__sync_fetch_and_nand_16: 1199 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Nand, E); 1200 1201 // Clang extensions: not overloaded yet. 1202 case Builtin::BI__sync_fetch_and_min: 1203 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Min, E); 1204 case Builtin::BI__sync_fetch_and_max: 1205 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Max, E); 1206 case Builtin::BI__sync_fetch_and_umin: 1207 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::UMin, E); 1208 case Builtin::BI__sync_fetch_and_umax: 1209 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::UMax, E); 1210 1211 case Builtin::BI__sync_add_and_fetch_1: 1212 case Builtin::BI__sync_add_and_fetch_2: 1213 case Builtin::BI__sync_add_and_fetch_4: 1214 case Builtin::BI__sync_add_and_fetch_8: 1215 case Builtin::BI__sync_add_and_fetch_16: 1216 return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::Add, E, 1217 llvm::Instruction::Add); 1218 case Builtin::BI__sync_sub_and_fetch_1: 1219 case Builtin::BI__sync_sub_and_fetch_2: 1220 case Builtin::BI__sync_sub_and_fetch_4: 1221 case Builtin::BI__sync_sub_and_fetch_8: 1222 case Builtin::BI__sync_sub_and_fetch_16: 1223 return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::Sub, E, 1224 llvm::Instruction::Sub); 1225 case Builtin::BI__sync_and_and_fetch_1: 1226 case Builtin::BI__sync_and_and_fetch_2: 1227 case Builtin::BI__sync_and_and_fetch_4: 1228 case Builtin::BI__sync_and_and_fetch_8: 1229 case Builtin::BI__sync_and_and_fetch_16: 1230 return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::And, E, 1231 llvm::Instruction::And); 1232 case Builtin::BI__sync_or_and_fetch_1: 1233 case Builtin::BI__sync_or_and_fetch_2: 1234 case Builtin::BI__sync_or_and_fetch_4: 1235 case Builtin::BI__sync_or_and_fetch_8: 1236 case Builtin::BI__sync_or_and_fetch_16: 1237 return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::Or, E, 1238 llvm::Instruction::Or); 1239 case Builtin::BI__sync_xor_and_fetch_1: 1240 case Builtin::BI__sync_xor_and_fetch_2: 1241 case Builtin::BI__sync_xor_and_fetch_4: 1242 case Builtin::BI__sync_xor_and_fetch_8: 1243 case Builtin::BI__sync_xor_and_fetch_16: 1244 return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::Xor, E, 1245 llvm::Instruction::Xor); 1246 case Builtin::BI__sync_nand_and_fetch_1: 1247 case Builtin::BI__sync_nand_and_fetch_2: 1248 case Builtin::BI__sync_nand_and_fetch_4: 1249 case Builtin::BI__sync_nand_and_fetch_8: 1250 case Builtin::BI__sync_nand_and_fetch_16: 1251 return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::Nand, E, 1252 llvm::Instruction::And, true); 1253 1254 case Builtin::BI__sync_val_compare_and_swap_1: 1255 case Builtin::BI__sync_val_compare_and_swap_2: 1256 case Builtin::BI__sync_val_compare_and_swap_4: 1257 case Builtin::BI__sync_val_compare_and_swap_8: 1258 case Builtin::BI__sync_val_compare_and_swap_16: 1259 return RValue::get(MakeAtomicCmpXchgValue(*this, E, false)); 1260 1261 case Builtin::BI__sync_bool_compare_and_swap_1: 1262 case Builtin::BI__sync_bool_compare_and_swap_2: 1263 case Builtin::BI__sync_bool_compare_and_swap_4: 1264 case Builtin::BI__sync_bool_compare_and_swap_8: 1265 case Builtin::BI__sync_bool_compare_and_swap_16: 1266 return RValue::get(MakeAtomicCmpXchgValue(*this, E, true)); 1267 1268 case Builtin::BI__sync_swap_1: 1269 case Builtin::BI__sync_swap_2: 1270 case Builtin::BI__sync_swap_4: 1271 case Builtin::BI__sync_swap_8: 1272 case Builtin::BI__sync_swap_16: 1273 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Xchg, E); 1274 1275 case Builtin::BI__sync_lock_test_and_set_1: 1276 case Builtin::BI__sync_lock_test_and_set_2: 1277 case Builtin::BI__sync_lock_test_and_set_4: 1278 case Builtin::BI__sync_lock_test_and_set_8: 1279 case Builtin::BI__sync_lock_test_and_set_16: 1280 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Xchg, E); 1281 1282 case Builtin::BI__sync_lock_release_1: 1283 case Builtin::BI__sync_lock_release_2: 1284 case Builtin::BI__sync_lock_release_4: 1285 case Builtin::BI__sync_lock_release_8: 1286 case Builtin::BI__sync_lock_release_16: { 1287 Value *Ptr = EmitScalarExpr(E->getArg(0)); 1288 QualType ElTy = E->getArg(0)->getType()->getPointeeType(); 1289 CharUnits StoreSize = getContext().getTypeSizeInChars(ElTy); 1290 llvm::Type *ITy = llvm::IntegerType::get(getLLVMContext(), 1291 StoreSize.getQuantity() * 8); 1292 Ptr = Builder.CreateBitCast(Ptr, ITy->getPointerTo()); 1293 llvm::StoreInst *Store = 1294 Builder.CreateAlignedStore(llvm::Constant::getNullValue(ITy), Ptr, 1295 StoreSize); 1296 Store->setAtomic(llvm::AtomicOrdering::Release); 1297 return RValue::get(nullptr); 1298 } 1299 1300 case Builtin::BI__sync_synchronize: { 1301 // We assume this is supposed to correspond to a C++0x-style 1302 // sequentially-consistent fence (i.e. this is only usable for 1303 // synchonization, not device I/O or anything like that). This intrinsic 1304 // is really badly designed in the sense that in theory, there isn't 1305 // any way to safely use it... but in practice, it mostly works 1306 // to use it with non-atomic loads and stores to get acquire/release 1307 // semantics. 1308 Builder.CreateFence(llvm::AtomicOrdering::SequentiallyConsistent); 1309 return RValue::get(nullptr); 1310 } 1311 1312 case Builtin::BI__builtin_nontemporal_load: 1313 return RValue::get(EmitNontemporalLoad(*this, E)); 1314 case Builtin::BI__builtin_nontemporal_store: 1315 return RValue::get(EmitNontemporalStore(*this, E)); 1316 case Builtin::BI__c11_atomic_is_lock_free: 1317 case Builtin::BI__atomic_is_lock_free: { 1318 // Call "bool __atomic_is_lock_free(size_t size, void *ptr)". For the 1319 // __c11 builtin, ptr is 0 (indicating a properly-aligned object), since 1320 // _Atomic(T) is always properly-aligned. 1321 const char *LibCallName = "__atomic_is_lock_free"; 1322 CallArgList Args; 1323 Args.add(RValue::get(EmitScalarExpr(E->getArg(0))), 1324 getContext().getSizeType()); 1325 if (BuiltinID == Builtin::BI__atomic_is_lock_free) 1326 Args.add(RValue::get(EmitScalarExpr(E->getArg(1))), 1327 getContext().VoidPtrTy); 1328 else 1329 Args.add(RValue::get(llvm::Constant::getNullValue(VoidPtrTy)), 1330 getContext().VoidPtrTy); 1331 const CGFunctionInfo &FuncInfo = 1332 CGM.getTypes().arrangeBuiltinFunctionCall(E->getType(), Args); 1333 llvm::FunctionType *FTy = CGM.getTypes().GetFunctionType(FuncInfo); 1334 llvm::Constant *Func = CGM.CreateRuntimeFunction(FTy, LibCallName); 1335 return EmitCall(FuncInfo, Func, ReturnValueSlot(), Args); 1336 } 1337 1338 case Builtin::BI__atomic_test_and_set: { 1339 // Look at the argument type to determine whether this is a volatile 1340 // operation. The parameter type is always volatile. 1341 QualType PtrTy = E->getArg(0)->IgnoreImpCasts()->getType(); 1342 bool Volatile = 1343 PtrTy->castAs<PointerType>()->getPointeeType().isVolatileQualified(); 1344 1345 Value *Ptr = EmitScalarExpr(E->getArg(0)); 1346 unsigned AddrSpace = Ptr->getType()->getPointerAddressSpace(); 1347 Ptr = Builder.CreateBitCast(Ptr, Int8Ty->getPointerTo(AddrSpace)); 1348 Value *NewVal = Builder.getInt8(1); 1349 Value *Order = EmitScalarExpr(E->getArg(1)); 1350 if (isa<llvm::ConstantInt>(Order)) { 1351 int ord = cast<llvm::ConstantInt>(Order)->getZExtValue(); 1352 AtomicRMWInst *Result = nullptr; 1353 switch (ord) { 1354 case 0: // memory_order_relaxed 1355 default: // invalid order 1356 Result = Builder.CreateAtomicRMW(llvm::AtomicRMWInst::Xchg, Ptr, NewVal, 1357 llvm::AtomicOrdering::Monotonic); 1358 break; 1359 case 1: // memory_order_consume 1360 case 2: // memory_order_acquire 1361 Result = Builder.CreateAtomicRMW(llvm::AtomicRMWInst::Xchg, Ptr, NewVal, 1362 llvm::AtomicOrdering::Acquire); 1363 break; 1364 case 3: // memory_order_release 1365 Result = Builder.CreateAtomicRMW(llvm::AtomicRMWInst::Xchg, Ptr, NewVal, 1366 llvm::AtomicOrdering::Release); 1367 break; 1368 case 4: // memory_order_acq_rel 1369 1370 Result = Builder.CreateAtomicRMW(llvm::AtomicRMWInst::Xchg, Ptr, NewVal, 1371 llvm::AtomicOrdering::AcquireRelease); 1372 break; 1373 case 5: // memory_order_seq_cst 1374 Result = Builder.CreateAtomicRMW( 1375 llvm::AtomicRMWInst::Xchg, Ptr, NewVal, 1376 llvm::AtomicOrdering::SequentiallyConsistent); 1377 break; 1378 } 1379 Result->setVolatile(Volatile); 1380 return RValue::get(Builder.CreateIsNotNull(Result, "tobool")); 1381 } 1382 1383 llvm::BasicBlock *ContBB = createBasicBlock("atomic.continue", CurFn); 1384 1385 llvm::BasicBlock *BBs[5] = { 1386 createBasicBlock("monotonic", CurFn), 1387 createBasicBlock("acquire", CurFn), 1388 createBasicBlock("release", CurFn), 1389 createBasicBlock("acqrel", CurFn), 1390 createBasicBlock("seqcst", CurFn) 1391 }; 1392 llvm::AtomicOrdering Orders[5] = { 1393 llvm::AtomicOrdering::Monotonic, llvm::AtomicOrdering::Acquire, 1394 llvm::AtomicOrdering::Release, llvm::AtomicOrdering::AcquireRelease, 1395 llvm::AtomicOrdering::SequentiallyConsistent}; 1396 1397 Order = Builder.CreateIntCast(Order, Builder.getInt32Ty(), false); 1398 llvm::SwitchInst *SI = Builder.CreateSwitch(Order, BBs[0]); 1399 1400 Builder.SetInsertPoint(ContBB); 1401 PHINode *Result = Builder.CreatePHI(Int8Ty, 5, "was_set"); 1402 1403 for (unsigned i = 0; i < 5; ++i) { 1404 Builder.SetInsertPoint(BBs[i]); 1405 AtomicRMWInst *RMW = Builder.CreateAtomicRMW(llvm::AtomicRMWInst::Xchg, 1406 Ptr, NewVal, Orders[i]); 1407 RMW->setVolatile(Volatile); 1408 Result->addIncoming(RMW, BBs[i]); 1409 Builder.CreateBr(ContBB); 1410 } 1411 1412 SI->addCase(Builder.getInt32(0), BBs[0]); 1413 SI->addCase(Builder.getInt32(1), BBs[1]); 1414 SI->addCase(Builder.getInt32(2), BBs[1]); 1415 SI->addCase(Builder.getInt32(3), BBs[2]); 1416 SI->addCase(Builder.getInt32(4), BBs[3]); 1417 SI->addCase(Builder.getInt32(5), BBs[4]); 1418 1419 Builder.SetInsertPoint(ContBB); 1420 return RValue::get(Builder.CreateIsNotNull(Result, "tobool")); 1421 } 1422 1423 case Builtin::BI__atomic_clear: { 1424 QualType PtrTy = E->getArg(0)->IgnoreImpCasts()->getType(); 1425 bool Volatile = 1426 PtrTy->castAs<PointerType>()->getPointeeType().isVolatileQualified(); 1427 1428 Address Ptr = EmitPointerWithAlignment(E->getArg(0)); 1429 unsigned AddrSpace = Ptr.getPointer()->getType()->getPointerAddressSpace(); 1430 Ptr = Builder.CreateBitCast(Ptr, Int8Ty->getPointerTo(AddrSpace)); 1431 Value *NewVal = Builder.getInt8(0); 1432 Value *Order = EmitScalarExpr(E->getArg(1)); 1433 if (isa<llvm::ConstantInt>(Order)) { 1434 int ord = cast<llvm::ConstantInt>(Order)->getZExtValue(); 1435 StoreInst *Store = Builder.CreateStore(NewVal, Ptr, Volatile); 1436 switch (ord) { 1437 case 0: // memory_order_relaxed 1438 default: // invalid order 1439 Store->setOrdering(llvm::AtomicOrdering::Monotonic); 1440 break; 1441 case 3: // memory_order_release 1442 Store->setOrdering(llvm::AtomicOrdering::Release); 1443 break; 1444 case 5: // memory_order_seq_cst 1445 Store->setOrdering(llvm::AtomicOrdering::SequentiallyConsistent); 1446 break; 1447 } 1448 return RValue::get(nullptr); 1449 } 1450 1451 llvm::BasicBlock *ContBB = createBasicBlock("atomic.continue", CurFn); 1452 1453 llvm::BasicBlock *BBs[3] = { 1454 createBasicBlock("monotonic", CurFn), 1455 createBasicBlock("release", CurFn), 1456 createBasicBlock("seqcst", CurFn) 1457 }; 1458 llvm::AtomicOrdering Orders[3] = { 1459 llvm::AtomicOrdering::Monotonic, llvm::AtomicOrdering::Release, 1460 llvm::AtomicOrdering::SequentiallyConsistent}; 1461 1462 Order = Builder.CreateIntCast(Order, Builder.getInt32Ty(), false); 1463 llvm::SwitchInst *SI = Builder.CreateSwitch(Order, BBs[0]); 1464 1465 for (unsigned i = 0; i < 3; ++i) { 1466 Builder.SetInsertPoint(BBs[i]); 1467 StoreInst *Store = Builder.CreateStore(NewVal, Ptr, Volatile); 1468 Store->setOrdering(Orders[i]); 1469 Builder.CreateBr(ContBB); 1470 } 1471 1472 SI->addCase(Builder.getInt32(0), BBs[0]); 1473 SI->addCase(Builder.getInt32(3), BBs[1]); 1474 SI->addCase(Builder.getInt32(5), BBs[2]); 1475 1476 Builder.SetInsertPoint(ContBB); 1477 return RValue::get(nullptr); 1478 } 1479 1480 case Builtin::BI__atomic_thread_fence: 1481 case Builtin::BI__atomic_signal_fence: 1482 case Builtin::BI__c11_atomic_thread_fence: 1483 case Builtin::BI__c11_atomic_signal_fence: { 1484 llvm::SynchronizationScope Scope; 1485 if (BuiltinID == Builtin::BI__atomic_signal_fence || 1486 BuiltinID == Builtin::BI__c11_atomic_signal_fence) 1487 Scope = llvm::SingleThread; 1488 else 1489 Scope = llvm::CrossThread; 1490 Value *Order = EmitScalarExpr(E->getArg(0)); 1491 if (isa<llvm::ConstantInt>(Order)) { 1492 int ord = cast<llvm::ConstantInt>(Order)->getZExtValue(); 1493 switch (ord) { 1494 case 0: // memory_order_relaxed 1495 default: // invalid order 1496 break; 1497 case 1: // memory_order_consume 1498 case 2: // memory_order_acquire 1499 Builder.CreateFence(llvm::AtomicOrdering::Acquire, Scope); 1500 break; 1501 case 3: // memory_order_release 1502 Builder.CreateFence(llvm::AtomicOrdering::Release, Scope); 1503 break; 1504 case 4: // memory_order_acq_rel 1505 Builder.CreateFence(llvm::AtomicOrdering::AcquireRelease, Scope); 1506 break; 1507 case 5: // memory_order_seq_cst 1508 Builder.CreateFence(llvm::AtomicOrdering::SequentiallyConsistent, 1509 Scope); 1510 break; 1511 } 1512 return RValue::get(nullptr); 1513 } 1514 1515 llvm::BasicBlock *AcquireBB, *ReleaseBB, *AcqRelBB, *SeqCstBB; 1516 AcquireBB = createBasicBlock("acquire", CurFn); 1517 ReleaseBB = createBasicBlock("release", CurFn); 1518 AcqRelBB = createBasicBlock("acqrel", CurFn); 1519 SeqCstBB = createBasicBlock("seqcst", CurFn); 1520 llvm::BasicBlock *ContBB = createBasicBlock("atomic.continue", CurFn); 1521 1522 Order = Builder.CreateIntCast(Order, Builder.getInt32Ty(), false); 1523 llvm::SwitchInst *SI = Builder.CreateSwitch(Order, ContBB); 1524 1525 Builder.SetInsertPoint(AcquireBB); 1526 Builder.CreateFence(llvm::AtomicOrdering::Acquire, Scope); 1527 Builder.CreateBr(ContBB); 1528 SI->addCase(Builder.getInt32(1), AcquireBB); 1529 SI->addCase(Builder.getInt32(2), AcquireBB); 1530 1531 Builder.SetInsertPoint(ReleaseBB); 1532 Builder.CreateFence(llvm::AtomicOrdering::Release, Scope); 1533 Builder.CreateBr(ContBB); 1534 SI->addCase(Builder.getInt32(3), ReleaseBB); 1535 1536 Builder.SetInsertPoint(AcqRelBB); 1537 Builder.CreateFence(llvm::AtomicOrdering::AcquireRelease, Scope); 1538 Builder.CreateBr(ContBB); 1539 SI->addCase(Builder.getInt32(4), AcqRelBB); 1540 1541 Builder.SetInsertPoint(SeqCstBB); 1542 Builder.CreateFence(llvm::AtomicOrdering::SequentiallyConsistent, Scope); 1543 Builder.CreateBr(ContBB); 1544 SI->addCase(Builder.getInt32(5), SeqCstBB); 1545 1546 Builder.SetInsertPoint(ContBB); 1547 return RValue::get(nullptr); 1548 } 1549 1550 // Library functions with special handling. 1551 case Builtin::BIsqrt: 1552 case Builtin::BIsqrtf: 1553 case Builtin::BIsqrtl: { 1554 // Transform a call to sqrt* into a @llvm.sqrt.* intrinsic call, but only 1555 // in finite- or unsafe-math mode (the intrinsic has different semantics 1556 // for handling negative numbers compared to the library function, so 1557 // -fmath-errno=0 is not enough). 1558 if (!FD->hasAttr<ConstAttr>()) 1559 break; 1560 if (!(CGM.getCodeGenOpts().UnsafeFPMath || 1561 CGM.getCodeGenOpts().NoNaNsFPMath)) 1562 break; 1563 Value *Arg0 = EmitScalarExpr(E->getArg(0)); 1564 llvm::Type *ArgType = Arg0->getType(); 1565 Value *F = CGM.getIntrinsic(Intrinsic::sqrt, ArgType); 1566 return RValue::get(Builder.CreateCall(F, Arg0)); 1567 } 1568 1569 case Builtin::BI__builtin_pow: 1570 case Builtin::BI__builtin_powf: 1571 case Builtin::BI__builtin_powl: 1572 case Builtin::BIpow: 1573 case Builtin::BIpowf: 1574 case Builtin::BIpowl: { 1575 // Transform a call to pow* into a @llvm.pow.* intrinsic call. 1576 if (!FD->hasAttr<ConstAttr>()) 1577 break; 1578 Value *Base = EmitScalarExpr(E->getArg(0)); 1579 Value *Exponent = EmitScalarExpr(E->getArg(1)); 1580 llvm::Type *ArgType = Base->getType(); 1581 Value *F = CGM.getIntrinsic(Intrinsic::pow, ArgType); 1582 return RValue::get(Builder.CreateCall(F, {Base, Exponent})); 1583 } 1584 1585 case Builtin::BIfma: 1586 case Builtin::BIfmaf: 1587 case Builtin::BIfmal: 1588 case Builtin::BI__builtin_fma: 1589 case Builtin::BI__builtin_fmaf: 1590 case Builtin::BI__builtin_fmal: { 1591 // Rewrite fma to intrinsic. 1592 Value *FirstArg = EmitScalarExpr(E->getArg(0)); 1593 llvm::Type *ArgType = FirstArg->getType(); 1594 Value *F = CGM.getIntrinsic(Intrinsic::fma, ArgType); 1595 return RValue::get( 1596 Builder.CreateCall(F, {FirstArg, EmitScalarExpr(E->getArg(1)), 1597 EmitScalarExpr(E->getArg(2))})); 1598 } 1599 1600 case Builtin::BI__builtin_signbit: 1601 case Builtin::BI__builtin_signbitf: 1602 case Builtin::BI__builtin_signbitl: { 1603 return RValue::get( 1604 Builder.CreateZExt(EmitSignBit(*this, EmitScalarExpr(E->getArg(0))), 1605 ConvertType(E->getType()))); 1606 } 1607 case Builtin::BI__builtin_annotation: { 1608 llvm::Value *AnnVal = EmitScalarExpr(E->getArg(0)); 1609 llvm::Value *F = CGM.getIntrinsic(llvm::Intrinsic::annotation, 1610 AnnVal->getType()); 1611 1612 // Get the annotation string, go through casts. Sema requires this to be a 1613 // non-wide string literal, potentially casted, so the cast<> is safe. 1614 const Expr *AnnotationStrExpr = E->getArg(1)->IgnoreParenCasts(); 1615 StringRef Str = cast<StringLiteral>(AnnotationStrExpr)->getString(); 1616 return RValue::get(EmitAnnotationCall(F, AnnVal, Str, E->getExprLoc())); 1617 } 1618 case Builtin::BI__builtin_addcb: 1619 case Builtin::BI__builtin_addcs: 1620 case Builtin::BI__builtin_addc: 1621 case Builtin::BI__builtin_addcl: 1622 case Builtin::BI__builtin_addcll: 1623 case Builtin::BI__builtin_subcb: 1624 case Builtin::BI__builtin_subcs: 1625 case Builtin::BI__builtin_subc: 1626 case Builtin::BI__builtin_subcl: 1627 case Builtin::BI__builtin_subcll: { 1628 1629 // We translate all of these builtins from expressions of the form: 1630 // int x = ..., y = ..., carryin = ..., carryout, result; 1631 // result = __builtin_addc(x, y, carryin, &carryout); 1632 // 1633 // to LLVM IR of the form: 1634 // 1635 // %tmp1 = call {i32, i1} @llvm.uadd.with.overflow.i32(i32 %x, i32 %y) 1636 // %tmpsum1 = extractvalue {i32, i1} %tmp1, 0 1637 // %carry1 = extractvalue {i32, i1} %tmp1, 1 1638 // %tmp2 = call {i32, i1} @llvm.uadd.with.overflow.i32(i32 %tmpsum1, 1639 // i32 %carryin) 1640 // %result = extractvalue {i32, i1} %tmp2, 0 1641 // %carry2 = extractvalue {i32, i1} %tmp2, 1 1642 // %tmp3 = or i1 %carry1, %carry2 1643 // %tmp4 = zext i1 %tmp3 to i32 1644 // store i32 %tmp4, i32* %carryout 1645 1646 // Scalarize our inputs. 1647 llvm::Value *X = EmitScalarExpr(E->getArg(0)); 1648 llvm::Value *Y = EmitScalarExpr(E->getArg(1)); 1649 llvm::Value *Carryin = EmitScalarExpr(E->getArg(2)); 1650 Address CarryOutPtr = EmitPointerWithAlignment(E->getArg(3)); 1651 1652 // Decide if we are lowering to a uadd.with.overflow or usub.with.overflow. 1653 llvm::Intrinsic::ID IntrinsicId; 1654 switch (BuiltinID) { 1655 default: llvm_unreachable("Unknown multiprecision builtin id."); 1656 case Builtin::BI__builtin_addcb: 1657 case Builtin::BI__builtin_addcs: 1658 case Builtin::BI__builtin_addc: 1659 case Builtin::BI__builtin_addcl: 1660 case Builtin::BI__builtin_addcll: 1661 IntrinsicId = llvm::Intrinsic::uadd_with_overflow; 1662 break; 1663 case Builtin::BI__builtin_subcb: 1664 case Builtin::BI__builtin_subcs: 1665 case Builtin::BI__builtin_subc: 1666 case Builtin::BI__builtin_subcl: 1667 case Builtin::BI__builtin_subcll: 1668 IntrinsicId = llvm::Intrinsic::usub_with_overflow; 1669 break; 1670 } 1671 1672 // Construct our resulting LLVM IR expression. 1673 llvm::Value *Carry1; 1674 llvm::Value *Sum1 = EmitOverflowIntrinsic(*this, IntrinsicId, 1675 X, Y, Carry1); 1676 llvm::Value *Carry2; 1677 llvm::Value *Sum2 = EmitOverflowIntrinsic(*this, IntrinsicId, 1678 Sum1, Carryin, Carry2); 1679 llvm::Value *CarryOut = Builder.CreateZExt(Builder.CreateOr(Carry1, Carry2), 1680 X->getType()); 1681 Builder.CreateStore(CarryOut, CarryOutPtr); 1682 return RValue::get(Sum2); 1683 } 1684 1685 case Builtin::BI__builtin_add_overflow: 1686 case Builtin::BI__builtin_sub_overflow: 1687 case Builtin::BI__builtin_mul_overflow: { 1688 const clang::Expr *LeftArg = E->getArg(0); 1689 const clang::Expr *RightArg = E->getArg(1); 1690 const clang::Expr *ResultArg = E->getArg(2); 1691 1692 clang::QualType ResultQTy = 1693 ResultArg->getType()->castAs<PointerType>()->getPointeeType(); 1694 1695 WidthAndSignedness LeftInfo = 1696 getIntegerWidthAndSignedness(CGM.getContext(), LeftArg->getType()); 1697 WidthAndSignedness RightInfo = 1698 getIntegerWidthAndSignedness(CGM.getContext(), RightArg->getType()); 1699 WidthAndSignedness ResultInfo = 1700 getIntegerWidthAndSignedness(CGM.getContext(), ResultQTy); 1701 WidthAndSignedness EncompassingInfo = 1702 EncompassingIntegerType({LeftInfo, RightInfo, ResultInfo}); 1703 1704 llvm::Type *EncompassingLLVMTy = 1705 llvm::IntegerType::get(CGM.getLLVMContext(), EncompassingInfo.Width); 1706 1707 llvm::Type *ResultLLVMTy = CGM.getTypes().ConvertType(ResultQTy); 1708 1709 llvm::Intrinsic::ID IntrinsicId; 1710 switch (BuiltinID) { 1711 default: 1712 llvm_unreachable("Unknown overflow builtin id."); 1713 case Builtin::BI__builtin_add_overflow: 1714 IntrinsicId = EncompassingInfo.Signed 1715 ? llvm::Intrinsic::sadd_with_overflow 1716 : llvm::Intrinsic::uadd_with_overflow; 1717 break; 1718 case Builtin::BI__builtin_sub_overflow: 1719 IntrinsicId = EncompassingInfo.Signed 1720 ? llvm::Intrinsic::ssub_with_overflow 1721 : llvm::Intrinsic::usub_with_overflow; 1722 break; 1723 case Builtin::BI__builtin_mul_overflow: 1724 IntrinsicId = EncompassingInfo.Signed 1725 ? llvm::Intrinsic::smul_with_overflow 1726 : llvm::Intrinsic::umul_with_overflow; 1727 break; 1728 } 1729 1730 llvm::Value *Left = EmitScalarExpr(LeftArg); 1731 llvm::Value *Right = EmitScalarExpr(RightArg); 1732 Address ResultPtr = EmitPointerWithAlignment(ResultArg); 1733 1734 // Extend each operand to the encompassing type. 1735 Left = Builder.CreateIntCast(Left, EncompassingLLVMTy, LeftInfo.Signed); 1736 Right = Builder.CreateIntCast(Right, EncompassingLLVMTy, RightInfo.Signed); 1737 1738 // Perform the operation on the extended values. 1739 llvm::Value *Overflow, *Result; 1740 Result = EmitOverflowIntrinsic(*this, IntrinsicId, Left, Right, Overflow); 1741 1742 if (EncompassingInfo.Width > ResultInfo.Width) { 1743 // The encompassing type is wider than the result type, so we need to 1744 // truncate it. 1745 llvm::Value *ResultTrunc = Builder.CreateTrunc(Result, ResultLLVMTy); 1746 1747 // To see if the truncation caused an overflow, we will extend 1748 // the result and then compare it to the original result. 1749 llvm::Value *ResultTruncExt = Builder.CreateIntCast( 1750 ResultTrunc, EncompassingLLVMTy, ResultInfo.Signed); 1751 llvm::Value *TruncationOverflow = 1752 Builder.CreateICmpNE(Result, ResultTruncExt); 1753 1754 Overflow = Builder.CreateOr(Overflow, TruncationOverflow); 1755 Result = ResultTrunc; 1756 } 1757 1758 // Finally, store the result using the pointer. 1759 bool isVolatile = 1760 ResultArg->getType()->getPointeeType().isVolatileQualified(); 1761 Builder.CreateStore(EmitToMemory(Result, ResultQTy), ResultPtr, isVolatile); 1762 1763 return RValue::get(Overflow); 1764 } 1765 1766 case Builtin::BI__builtin_uadd_overflow: 1767 case Builtin::BI__builtin_uaddl_overflow: 1768 case Builtin::BI__builtin_uaddll_overflow: 1769 case Builtin::BI__builtin_usub_overflow: 1770 case Builtin::BI__builtin_usubl_overflow: 1771 case Builtin::BI__builtin_usubll_overflow: 1772 case Builtin::BI__builtin_umul_overflow: 1773 case Builtin::BI__builtin_umull_overflow: 1774 case Builtin::BI__builtin_umulll_overflow: 1775 case Builtin::BI__builtin_sadd_overflow: 1776 case Builtin::BI__builtin_saddl_overflow: 1777 case Builtin::BI__builtin_saddll_overflow: 1778 case Builtin::BI__builtin_ssub_overflow: 1779 case Builtin::BI__builtin_ssubl_overflow: 1780 case Builtin::BI__builtin_ssubll_overflow: 1781 case Builtin::BI__builtin_smul_overflow: 1782 case Builtin::BI__builtin_smull_overflow: 1783 case Builtin::BI__builtin_smulll_overflow: { 1784 1785 // We translate all of these builtins directly to the relevant llvm IR node. 1786 1787 // Scalarize our inputs. 1788 llvm::Value *X = EmitScalarExpr(E->getArg(0)); 1789 llvm::Value *Y = EmitScalarExpr(E->getArg(1)); 1790 Address SumOutPtr = EmitPointerWithAlignment(E->getArg(2)); 1791 1792 // Decide which of the overflow intrinsics we are lowering to: 1793 llvm::Intrinsic::ID IntrinsicId; 1794 switch (BuiltinID) { 1795 default: llvm_unreachable("Unknown overflow builtin id."); 1796 case Builtin::BI__builtin_uadd_overflow: 1797 case Builtin::BI__builtin_uaddl_overflow: 1798 case Builtin::BI__builtin_uaddll_overflow: 1799 IntrinsicId = llvm::Intrinsic::uadd_with_overflow; 1800 break; 1801 case Builtin::BI__builtin_usub_overflow: 1802 case Builtin::BI__builtin_usubl_overflow: 1803 case Builtin::BI__builtin_usubll_overflow: 1804 IntrinsicId = llvm::Intrinsic::usub_with_overflow; 1805 break; 1806 case Builtin::BI__builtin_umul_overflow: 1807 case Builtin::BI__builtin_umull_overflow: 1808 case Builtin::BI__builtin_umulll_overflow: 1809 IntrinsicId = llvm::Intrinsic::umul_with_overflow; 1810 break; 1811 case Builtin::BI__builtin_sadd_overflow: 1812 case Builtin::BI__builtin_saddl_overflow: 1813 case Builtin::BI__builtin_saddll_overflow: 1814 IntrinsicId = llvm::Intrinsic::sadd_with_overflow; 1815 break; 1816 case Builtin::BI__builtin_ssub_overflow: 1817 case Builtin::BI__builtin_ssubl_overflow: 1818 case Builtin::BI__builtin_ssubll_overflow: 1819 IntrinsicId = llvm::Intrinsic::ssub_with_overflow; 1820 break; 1821 case Builtin::BI__builtin_smul_overflow: 1822 case Builtin::BI__builtin_smull_overflow: 1823 case Builtin::BI__builtin_smulll_overflow: 1824 IntrinsicId = llvm::Intrinsic::smul_with_overflow; 1825 break; 1826 } 1827 1828 1829 llvm::Value *Carry; 1830 llvm::Value *Sum = EmitOverflowIntrinsic(*this, IntrinsicId, X, Y, Carry); 1831 Builder.CreateStore(Sum, SumOutPtr); 1832 1833 return RValue::get(Carry); 1834 } 1835 case Builtin::BI__builtin_addressof: 1836 return RValue::get(EmitLValue(E->getArg(0)).getPointer()); 1837 case Builtin::BI__builtin_operator_new: 1838 return EmitBuiltinNewDeleteCall(FD->getType()->castAs<FunctionProtoType>(), 1839 E->getArg(0), false); 1840 case Builtin::BI__builtin_operator_delete: 1841 return EmitBuiltinNewDeleteCall(FD->getType()->castAs<FunctionProtoType>(), 1842 E->getArg(0), true); 1843 case Builtin::BI__noop: 1844 // __noop always evaluates to an integer literal zero. 1845 return RValue::get(ConstantInt::get(IntTy, 0)); 1846 case Builtin::BI__builtin_call_with_static_chain: { 1847 const CallExpr *Call = cast<CallExpr>(E->getArg(0)); 1848 const Expr *Chain = E->getArg(1); 1849 return EmitCall(Call->getCallee()->getType(), 1850 EmitScalarExpr(Call->getCallee()), Call, ReturnValue, 1851 Call->getCalleeDecl(), EmitScalarExpr(Chain)); 1852 } 1853 case Builtin::BI_InterlockedExchange: 1854 case Builtin::BI_InterlockedExchangePointer: 1855 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Xchg, E); 1856 case Builtin::BI_InterlockedCompareExchangePointer: { 1857 llvm::Type *RTy; 1858 llvm::IntegerType *IntType = 1859 IntegerType::get(getLLVMContext(), 1860 getContext().getTypeSize(E->getType())); 1861 llvm::Type *IntPtrType = IntType->getPointerTo(); 1862 1863 llvm::Value *Destination = 1864 Builder.CreateBitCast(EmitScalarExpr(E->getArg(0)), IntPtrType); 1865 1866 llvm::Value *Exchange = EmitScalarExpr(E->getArg(1)); 1867 RTy = Exchange->getType(); 1868 Exchange = Builder.CreatePtrToInt(Exchange, IntType); 1869 1870 llvm::Value *Comparand = 1871 Builder.CreatePtrToInt(EmitScalarExpr(E->getArg(2)), IntType); 1872 1873 auto Result = 1874 Builder.CreateAtomicCmpXchg(Destination, Comparand, Exchange, 1875 AtomicOrdering::SequentiallyConsistent, 1876 AtomicOrdering::SequentiallyConsistent); 1877 Result->setVolatile(true); 1878 1879 return RValue::get(Builder.CreateIntToPtr(Builder.CreateExtractValue(Result, 1880 0), 1881 RTy)); 1882 } 1883 case Builtin::BI_InterlockedCompareExchange: { 1884 AtomicCmpXchgInst *CXI = Builder.CreateAtomicCmpXchg( 1885 EmitScalarExpr(E->getArg(0)), 1886 EmitScalarExpr(E->getArg(2)), 1887 EmitScalarExpr(E->getArg(1)), 1888 AtomicOrdering::SequentiallyConsistent, 1889 AtomicOrdering::SequentiallyConsistent); 1890 CXI->setVolatile(true); 1891 return RValue::get(Builder.CreateExtractValue(CXI, 0)); 1892 } 1893 case Builtin::BI_InterlockedIncrement: { 1894 AtomicRMWInst *RMWI = Builder.CreateAtomicRMW( 1895 AtomicRMWInst::Add, 1896 EmitScalarExpr(E->getArg(0)), 1897 ConstantInt::get(Int32Ty, 1), 1898 llvm::AtomicOrdering::SequentiallyConsistent); 1899 RMWI->setVolatile(true); 1900 return RValue::get(Builder.CreateAdd(RMWI, ConstantInt::get(Int32Ty, 1))); 1901 } 1902 case Builtin::BI_InterlockedDecrement: { 1903 AtomicRMWInst *RMWI = Builder.CreateAtomicRMW( 1904 AtomicRMWInst::Sub, 1905 EmitScalarExpr(E->getArg(0)), 1906 ConstantInt::get(Int32Ty, 1), 1907 llvm::AtomicOrdering::SequentiallyConsistent); 1908 RMWI->setVolatile(true); 1909 return RValue::get(Builder.CreateSub(RMWI, ConstantInt::get(Int32Ty, 1))); 1910 } 1911 case Builtin::BI_InterlockedExchangeAdd: { 1912 AtomicRMWInst *RMWI = Builder.CreateAtomicRMW( 1913 AtomicRMWInst::Add, 1914 EmitScalarExpr(E->getArg(0)), 1915 EmitScalarExpr(E->getArg(1)), 1916 llvm::AtomicOrdering::SequentiallyConsistent); 1917 RMWI->setVolatile(true); 1918 return RValue::get(RMWI); 1919 } 1920 case Builtin::BI__readfsdword: { 1921 Value *IntToPtr = 1922 Builder.CreateIntToPtr(EmitScalarExpr(E->getArg(0)), 1923 llvm::PointerType::get(CGM.Int32Ty, 257)); 1924 LoadInst *Load = 1925 Builder.CreateAlignedLoad(IntToPtr, /*Align=*/4, /*isVolatile=*/true); 1926 return RValue::get(Load); 1927 } 1928 1929 case Builtin::BI__exception_code: 1930 case Builtin::BI_exception_code: 1931 return RValue::get(EmitSEHExceptionCode()); 1932 case Builtin::BI__exception_info: 1933 case Builtin::BI_exception_info: 1934 return RValue::get(EmitSEHExceptionInfo()); 1935 case Builtin::BI__abnormal_termination: 1936 case Builtin::BI_abnormal_termination: 1937 return RValue::get(EmitSEHAbnormalTermination()); 1938 case Builtin::BI_setjmpex: { 1939 if (getTarget().getTriple().isOSMSVCRT()) { 1940 llvm::Type *ArgTypes[] = {Int8PtrTy, Int8PtrTy}; 1941 llvm::AttributeSet ReturnsTwiceAttr = 1942 AttributeSet::get(getLLVMContext(), llvm::AttributeSet::FunctionIndex, 1943 llvm::Attribute::ReturnsTwice); 1944 llvm::Constant *SetJmpEx = CGM.CreateRuntimeFunction( 1945 llvm::FunctionType::get(IntTy, ArgTypes, /*isVarArg=*/false), 1946 "_setjmpex", ReturnsTwiceAttr); 1947 llvm::Value *Buf = Builder.CreateBitOrPointerCast( 1948 EmitScalarExpr(E->getArg(0)), Int8PtrTy); 1949 llvm::Value *FrameAddr = 1950 Builder.CreateCall(CGM.getIntrinsic(Intrinsic::frameaddress), 1951 ConstantInt::get(Int32Ty, 0)); 1952 llvm::Value *Args[] = {Buf, FrameAddr}; 1953 llvm::CallSite CS = EmitRuntimeCallOrInvoke(SetJmpEx, Args); 1954 CS.setAttributes(ReturnsTwiceAttr); 1955 return RValue::get(CS.getInstruction()); 1956 } 1957 break; 1958 } 1959 case Builtin::BI_setjmp: { 1960 if (getTarget().getTriple().isOSMSVCRT()) { 1961 llvm::AttributeSet ReturnsTwiceAttr = 1962 AttributeSet::get(getLLVMContext(), llvm::AttributeSet::FunctionIndex, 1963 llvm::Attribute::ReturnsTwice); 1964 llvm::Value *Buf = Builder.CreateBitOrPointerCast( 1965 EmitScalarExpr(E->getArg(0)), Int8PtrTy); 1966 llvm::CallSite CS; 1967 if (getTarget().getTriple().getArch() == llvm::Triple::x86) { 1968 llvm::Type *ArgTypes[] = {Int8PtrTy, IntTy}; 1969 llvm::Constant *SetJmp3 = CGM.CreateRuntimeFunction( 1970 llvm::FunctionType::get(IntTy, ArgTypes, /*isVarArg=*/true), 1971 "_setjmp3", ReturnsTwiceAttr); 1972 llvm::Value *Count = ConstantInt::get(IntTy, 0); 1973 llvm::Value *Args[] = {Buf, Count}; 1974 CS = EmitRuntimeCallOrInvoke(SetJmp3, Args); 1975 } else { 1976 llvm::Type *ArgTypes[] = {Int8PtrTy, Int8PtrTy}; 1977 llvm::Constant *SetJmp = CGM.CreateRuntimeFunction( 1978 llvm::FunctionType::get(IntTy, ArgTypes, /*isVarArg=*/false), 1979 "_setjmp", ReturnsTwiceAttr); 1980 llvm::Value *FrameAddr = 1981 Builder.CreateCall(CGM.getIntrinsic(Intrinsic::frameaddress), 1982 ConstantInt::get(Int32Ty, 0)); 1983 llvm::Value *Args[] = {Buf, FrameAddr}; 1984 CS = EmitRuntimeCallOrInvoke(SetJmp, Args); 1985 } 1986 CS.setAttributes(ReturnsTwiceAttr); 1987 return RValue::get(CS.getInstruction()); 1988 } 1989 break; 1990 } 1991 1992 case Builtin::BI__GetExceptionInfo: { 1993 if (llvm::GlobalVariable *GV = 1994 CGM.getCXXABI().getThrowInfo(FD->getParamDecl(0)->getType())) 1995 return RValue::get(llvm::ConstantExpr::getBitCast(GV, CGM.Int8PtrTy)); 1996 break; 1997 } 1998 1999 // OpenCL v2.0 s6.13.16.2, Built-in pipe read and write functions 2000 case Builtin::BIread_pipe: 2001 case Builtin::BIwrite_pipe: { 2002 Value *Arg0 = EmitScalarExpr(E->getArg(0)), 2003 *Arg1 = EmitScalarExpr(E->getArg(1)); 2004 2005 // Type of the generic packet parameter. 2006 unsigned GenericAS = 2007 getContext().getTargetAddressSpace(LangAS::opencl_generic); 2008 llvm::Type *I8PTy = llvm::PointerType::get( 2009 llvm::Type::getInt8Ty(getLLVMContext()), GenericAS); 2010 2011 // Testing which overloaded version we should generate the call for. 2012 if (2U == E->getNumArgs()) { 2013 const char *Name = (BuiltinID == Builtin::BIread_pipe) ? "__read_pipe_2" 2014 : "__write_pipe_2"; 2015 // Creating a generic function type to be able to call with any builtin or 2016 // user defined type. 2017 llvm::Type *ArgTys[] = {Arg0->getType(), I8PTy}; 2018 llvm::FunctionType *FTy = llvm::FunctionType::get( 2019 Int32Ty, llvm::ArrayRef<llvm::Type *>(ArgTys), false); 2020 Value *BCast = Builder.CreatePointerCast(Arg1, I8PTy); 2021 return RValue::get(Builder.CreateCall( 2022 CGM.CreateRuntimeFunction(FTy, Name), {Arg0, BCast})); 2023 } else { 2024 assert(4 == E->getNumArgs() && 2025 "Illegal number of parameters to pipe function"); 2026 const char *Name = (BuiltinID == Builtin::BIread_pipe) ? "__read_pipe_4" 2027 : "__write_pipe_4"; 2028 2029 llvm::Type *ArgTys[] = {Arg0->getType(), Arg1->getType(), Int32Ty, I8PTy}; 2030 Value *Arg2 = EmitScalarExpr(E->getArg(2)), 2031 *Arg3 = EmitScalarExpr(E->getArg(3)); 2032 llvm::FunctionType *FTy = llvm::FunctionType::get( 2033 Int32Ty, llvm::ArrayRef<llvm::Type *>(ArgTys), false); 2034 Value *BCast = Builder.CreatePointerCast(Arg3, I8PTy); 2035 // We know the third argument is an integer type, but we may need to cast 2036 // it to i32. 2037 if (Arg2->getType() != Int32Ty) 2038 Arg2 = Builder.CreateZExtOrTrunc(Arg2, Int32Ty); 2039 return RValue::get(Builder.CreateCall( 2040 CGM.CreateRuntimeFunction(FTy, Name), {Arg0, Arg1, Arg2, BCast})); 2041 } 2042 } 2043 // OpenCL v2.0 s6.13.16 ,s9.17.3.5 - Built-in pipe reserve read and write 2044 // functions 2045 case Builtin::BIreserve_read_pipe: 2046 case Builtin::BIreserve_write_pipe: 2047 case Builtin::BIwork_group_reserve_read_pipe: 2048 case Builtin::BIwork_group_reserve_write_pipe: 2049 case Builtin::BIsub_group_reserve_read_pipe: 2050 case Builtin::BIsub_group_reserve_write_pipe: { 2051 // Composing the mangled name for the function. 2052 const char *Name; 2053 if (BuiltinID == Builtin::BIreserve_read_pipe) 2054 Name = "__reserve_read_pipe"; 2055 else if (BuiltinID == Builtin::BIreserve_write_pipe) 2056 Name = "__reserve_write_pipe"; 2057 else if (BuiltinID == Builtin::BIwork_group_reserve_read_pipe) 2058 Name = "__work_group_reserve_read_pipe"; 2059 else if (BuiltinID == Builtin::BIwork_group_reserve_write_pipe) 2060 Name = "__work_group_reserve_write_pipe"; 2061 else if (BuiltinID == Builtin::BIsub_group_reserve_read_pipe) 2062 Name = "__sub_group_reserve_read_pipe"; 2063 else 2064 Name = "__sub_group_reserve_write_pipe"; 2065 2066 Value *Arg0 = EmitScalarExpr(E->getArg(0)), 2067 *Arg1 = EmitScalarExpr(E->getArg(1)); 2068 llvm::Type *ReservedIDTy = ConvertType(getContext().OCLReserveIDTy); 2069 2070 // Building the generic function prototype. 2071 llvm::Type *ArgTys[] = {Arg0->getType(), Int32Ty}; 2072 llvm::FunctionType *FTy = llvm::FunctionType::get( 2073 ReservedIDTy, llvm::ArrayRef<llvm::Type *>(ArgTys), false); 2074 // We know the second argument is an integer type, but we may need to cast 2075 // it to i32. 2076 if (Arg1->getType() != Int32Ty) 2077 Arg1 = Builder.CreateZExtOrTrunc(Arg1, Int32Ty); 2078 return RValue::get( 2079 Builder.CreateCall(CGM.CreateRuntimeFunction(FTy, Name), {Arg0, Arg1})); 2080 } 2081 // OpenCL v2.0 s6.13.16 ,s9.17.3.5 - Built-in pipe commit read and write 2082 // functions 2083 case Builtin::BIcommit_read_pipe: 2084 case Builtin::BIcommit_write_pipe: 2085 case Builtin::BIwork_group_commit_read_pipe: 2086 case Builtin::BIwork_group_commit_write_pipe: 2087 case Builtin::BIsub_group_commit_read_pipe: 2088 case Builtin::BIsub_group_commit_write_pipe: { 2089 const char *Name; 2090 if (BuiltinID == Builtin::BIcommit_read_pipe) 2091 Name = "__commit_read_pipe"; 2092 else if (BuiltinID == Builtin::BIcommit_write_pipe) 2093 Name = "__commit_write_pipe"; 2094 else if (BuiltinID == Builtin::BIwork_group_commit_read_pipe) 2095 Name = "__work_group_commit_read_pipe"; 2096 else if (BuiltinID == Builtin::BIwork_group_commit_write_pipe) 2097 Name = "__work_group_commit_write_pipe"; 2098 else if (BuiltinID == Builtin::BIsub_group_commit_read_pipe) 2099 Name = "__sub_group_commit_read_pipe"; 2100 else 2101 Name = "__sub_group_commit_write_pipe"; 2102 2103 Value *Arg0 = EmitScalarExpr(E->getArg(0)), 2104 *Arg1 = EmitScalarExpr(E->getArg(1)); 2105 2106 // Building the generic function prototype. 2107 llvm::Type *ArgTys[] = {Arg0->getType(), Arg1->getType()}; 2108 llvm::FunctionType *FTy = 2109 llvm::FunctionType::get(llvm::Type::getVoidTy(getLLVMContext()), 2110 llvm::ArrayRef<llvm::Type *>(ArgTys), false); 2111 2112 return RValue::get( 2113 Builder.CreateCall(CGM.CreateRuntimeFunction(FTy, Name), {Arg0, Arg1})); 2114 } 2115 // OpenCL v2.0 s6.13.16.4 Built-in pipe query functions 2116 case Builtin::BIget_pipe_num_packets: 2117 case Builtin::BIget_pipe_max_packets: { 2118 const char *Name; 2119 if (BuiltinID == Builtin::BIget_pipe_num_packets) 2120 Name = "__get_pipe_num_packets"; 2121 else 2122 Name = "__get_pipe_max_packets"; 2123 2124 // Building the generic function prototype. 2125 Value *Arg0 = EmitScalarExpr(E->getArg(0)); 2126 llvm::Type *ArgTys[] = {Arg0->getType()}; 2127 llvm::FunctionType *FTy = llvm::FunctionType::get( 2128 Int32Ty, llvm::ArrayRef<llvm::Type *>(ArgTys), false); 2129 2130 return RValue::get( 2131 Builder.CreateCall(CGM.CreateRuntimeFunction(FTy, Name), {Arg0})); 2132 } 2133 2134 case Builtin::BIprintf: 2135 if (getLangOpts().CUDA && getLangOpts().CUDAIsDevice) 2136 return EmitCUDADevicePrintfCallExpr(E, ReturnValue); 2137 break; 2138 case Builtin::BI__builtin_canonicalize: 2139 case Builtin::BI__builtin_canonicalizef: 2140 case Builtin::BI__builtin_canonicalizel: 2141 return RValue::get(emitUnaryBuiltin(*this, E, Intrinsic::canonicalize)); 2142 } 2143 2144 // If this is an alias for a lib function (e.g. __builtin_sin), emit 2145 // the call using the normal call path, but using the unmangled 2146 // version of the function name. 2147 if (getContext().BuiltinInfo.isLibFunction(BuiltinID)) 2148 return emitLibraryCall(*this, FD, E, 2149 CGM.getBuiltinLibFunction(FD, BuiltinID)); 2150 2151 // If this is a predefined lib function (e.g. malloc), emit the call 2152 // using exactly the normal call path. 2153 if (getContext().BuiltinInfo.isPredefinedLibFunction(BuiltinID)) 2154 return emitLibraryCall(*this, FD, E, EmitScalarExpr(E->getCallee())); 2155 2156 // Check that a call to a target specific builtin has the correct target 2157 // features. 2158 // This is down here to avoid non-target specific builtins, however, if 2159 // generic builtins start to require generic target features then we 2160 // can move this up to the beginning of the function. 2161 checkTargetFeatures(E, FD); 2162 2163 // See if we have a target specific intrinsic. 2164 const char *Name = getContext().BuiltinInfo.getName(BuiltinID); 2165 Intrinsic::ID IntrinsicID = Intrinsic::not_intrinsic; 2166 if (const char *Prefix = 2167 llvm::Triple::getArchTypePrefix(getTarget().getTriple().getArch())) { 2168 IntrinsicID = Intrinsic::getIntrinsicForGCCBuiltin(Prefix, Name); 2169 // NOTE we dont need to perform a compatibility flag check here since the 2170 // intrinsics are declared in Builtins*.def via LANGBUILTIN which filter the 2171 // MS builtins via ALL_MS_LANGUAGES and are filtered earlier. 2172 if (IntrinsicID == Intrinsic::not_intrinsic) 2173 IntrinsicID = Intrinsic::getIntrinsicForMSBuiltin(Prefix, Name); 2174 } 2175 2176 if (IntrinsicID != Intrinsic::not_intrinsic) { 2177 SmallVector<Value*, 16> Args; 2178 2179 // Find out if any arguments are required to be integer constant 2180 // expressions. 2181 unsigned ICEArguments = 0; 2182 ASTContext::GetBuiltinTypeError Error; 2183 getContext().GetBuiltinType(BuiltinID, Error, &ICEArguments); 2184 assert(Error == ASTContext::GE_None && "Should not codegen an error"); 2185 2186 Function *F = CGM.getIntrinsic(IntrinsicID); 2187 llvm::FunctionType *FTy = F->getFunctionType(); 2188 2189 for (unsigned i = 0, e = E->getNumArgs(); i != e; ++i) { 2190 Value *ArgValue; 2191 // If this is a normal argument, just emit it as a scalar. 2192 if ((ICEArguments & (1 << i)) == 0) { 2193 ArgValue = EmitScalarExpr(E->getArg(i)); 2194 } else { 2195 // If this is required to be a constant, constant fold it so that we 2196 // know that the generated intrinsic gets a ConstantInt. 2197 llvm::APSInt Result; 2198 bool IsConst = E->getArg(i)->isIntegerConstantExpr(Result,getContext()); 2199 assert(IsConst && "Constant arg isn't actually constant?"); 2200 (void)IsConst; 2201 ArgValue = llvm::ConstantInt::get(getLLVMContext(), Result); 2202 } 2203 2204 // If the intrinsic arg type is different from the builtin arg type 2205 // we need to do a bit cast. 2206 llvm::Type *PTy = FTy->getParamType(i); 2207 if (PTy != ArgValue->getType()) { 2208 assert(PTy->canLosslesslyBitCastTo(FTy->getParamType(i)) && 2209 "Must be able to losslessly bit cast to param"); 2210 ArgValue = Builder.CreateBitCast(ArgValue, PTy); 2211 } 2212 2213 Args.push_back(ArgValue); 2214 } 2215 2216 Value *V = Builder.CreateCall(F, Args); 2217 QualType BuiltinRetType = E->getType(); 2218 2219 llvm::Type *RetTy = VoidTy; 2220 if (!BuiltinRetType->isVoidType()) 2221 RetTy = ConvertType(BuiltinRetType); 2222 2223 if (RetTy != V->getType()) { 2224 assert(V->getType()->canLosslesslyBitCastTo(RetTy) && 2225 "Must be able to losslessly bit cast result type"); 2226 V = Builder.CreateBitCast(V, RetTy); 2227 } 2228 2229 return RValue::get(V); 2230 } 2231 2232 // See if we have a target specific builtin that needs to be lowered. 2233 if (Value *V = EmitTargetBuiltinExpr(BuiltinID, E)) 2234 return RValue::get(V); 2235 2236 ErrorUnsupported(E, "builtin function"); 2237 2238 // Unknown builtin, for now just dump it out and return undef. 2239 return GetUndefRValue(E->getType()); 2240 } 2241 2242 static Value *EmitTargetArchBuiltinExpr(CodeGenFunction *CGF, 2243 unsigned BuiltinID, const CallExpr *E, 2244 llvm::Triple::ArchType Arch) { 2245 switch (Arch) { 2246 case llvm::Triple::arm: 2247 case llvm::Triple::armeb: 2248 case llvm::Triple::thumb: 2249 case llvm::Triple::thumbeb: 2250 return CGF->EmitARMBuiltinExpr(BuiltinID, E); 2251 case llvm::Triple::aarch64: 2252 case llvm::Triple::aarch64_be: 2253 return CGF->EmitAArch64BuiltinExpr(BuiltinID, E); 2254 case llvm::Triple::x86: 2255 case llvm::Triple::x86_64: 2256 return CGF->EmitX86BuiltinExpr(BuiltinID, E); 2257 case llvm::Triple::ppc: 2258 case llvm::Triple::ppc64: 2259 case llvm::Triple::ppc64le: 2260 return CGF->EmitPPCBuiltinExpr(BuiltinID, E); 2261 case llvm::Triple::r600: 2262 case llvm::Triple::amdgcn: 2263 return CGF->EmitAMDGPUBuiltinExpr(BuiltinID, E); 2264 case llvm::Triple::systemz: 2265 return CGF->EmitSystemZBuiltinExpr(BuiltinID, E); 2266 case llvm::Triple::nvptx: 2267 case llvm::Triple::nvptx64: 2268 return CGF->EmitNVPTXBuiltinExpr(BuiltinID, E); 2269 case llvm::Triple::wasm32: 2270 case llvm::Triple::wasm64: 2271 return CGF->EmitWebAssemblyBuiltinExpr(BuiltinID, E); 2272 default: 2273 return nullptr; 2274 } 2275 } 2276 2277 Value *CodeGenFunction::EmitTargetBuiltinExpr(unsigned BuiltinID, 2278 const CallExpr *E) { 2279 if (getContext().BuiltinInfo.isAuxBuiltinID(BuiltinID)) { 2280 assert(getContext().getAuxTargetInfo() && "Missing aux target info"); 2281 return EmitTargetArchBuiltinExpr( 2282 this, getContext().BuiltinInfo.getAuxBuiltinID(BuiltinID), E, 2283 getContext().getAuxTargetInfo()->getTriple().getArch()); 2284 } 2285 2286 return EmitTargetArchBuiltinExpr(this, BuiltinID, E, 2287 getTarget().getTriple().getArch()); 2288 } 2289 2290 static llvm::VectorType *GetNeonType(CodeGenFunction *CGF, 2291 NeonTypeFlags TypeFlags, 2292 bool V1Ty=false) { 2293 int IsQuad = TypeFlags.isQuad(); 2294 switch (TypeFlags.getEltType()) { 2295 case NeonTypeFlags::Int8: 2296 case NeonTypeFlags::Poly8: 2297 return llvm::VectorType::get(CGF->Int8Ty, V1Ty ? 1 : (8 << IsQuad)); 2298 case NeonTypeFlags::Int16: 2299 case NeonTypeFlags::Poly16: 2300 case NeonTypeFlags::Float16: 2301 return llvm::VectorType::get(CGF->Int16Ty, V1Ty ? 1 : (4 << IsQuad)); 2302 case NeonTypeFlags::Int32: 2303 return llvm::VectorType::get(CGF->Int32Ty, V1Ty ? 1 : (2 << IsQuad)); 2304 case NeonTypeFlags::Int64: 2305 case NeonTypeFlags::Poly64: 2306 return llvm::VectorType::get(CGF->Int64Ty, V1Ty ? 1 : (1 << IsQuad)); 2307 case NeonTypeFlags::Poly128: 2308 // FIXME: i128 and f128 doesn't get fully support in Clang and llvm. 2309 // There is a lot of i128 and f128 API missing. 2310 // so we use v16i8 to represent poly128 and get pattern matched. 2311 return llvm::VectorType::get(CGF->Int8Ty, 16); 2312 case NeonTypeFlags::Float32: 2313 return llvm::VectorType::get(CGF->FloatTy, V1Ty ? 1 : (2 << IsQuad)); 2314 case NeonTypeFlags::Float64: 2315 return llvm::VectorType::get(CGF->DoubleTy, V1Ty ? 1 : (1 << IsQuad)); 2316 } 2317 llvm_unreachable("Unknown vector element type!"); 2318 } 2319 2320 static llvm::VectorType *GetFloatNeonType(CodeGenFunction *CGF, 2321 NeonTypeFlags IntTypeFlags) { 2322 int IsQuad = IntTypeFlags.isQuad(); 2323 switch (IntTypeFlags.getEltType()) { 2324 case NeonTypeFlags::Int32: 2325 return llvm::VectorType::get(CGF->FloatTy, (2 << IsQuad)); 2326 case NeonTypeFlags::Int64: 2327 return llvm::VectorType::get(CGF->DoubleTy, (1 << IsQuad)); 2328 default: 2329 llvm_unreachable("Type can't be converted to floating-point!"); 2330 } 2331 } 2332 2333 Value *CodeGenFunction::EmitNeonSplat(Value *V, Constant *C) { 2334 unsigned nElts = cast<llvm::VectorType>(V->getType())->getNumElements(); 2335 Value* SV = llvm::ConstantVector::getSplat(nElts, C); 2336 return Builder.CreateShuffleVector(V, V, SV, "lane"); 2337 } 2338 2339 Value *CodeGenFunction::EmitNeonCall(Function *F, SmallVectorImpl<Value*> &Ops, 2340 const char *name, 2341 unsigned shift, bool rightshift) { 2342 unsigned j = 0; 2343 for (Function::const_arg_iterator ai = F->arg_begin(), ae = F->arg_end(); 2344 ai != ae; ++ai, ++j) 2345 if (shift > 0 && shift == j) 2346 Ops[j] = EmitNeonShiftVector(Ops[j], ai->getType(), rightshift); 2347 else 2348 Ops[j] = Builder.CreateBitCast(Ops[j], ai->getType(), name); 2349 2350 return Builder.CreateCall(F, Ops, name); 2351 } 2352 2353 Value *CodeGenFunction::EmitNeonShiftVector(Value *V, llvm::Type *Ty, 2354 bool neg) { 2355 int SV = cast<ConstantInt>(V)->getSExtValue(); 2356 return ConstantInt::get(Ty, neg ? -SV : SV); 2357 } 2358 2359 // \brief Right-shift a vector by a constant. 2360 Value *CodeGenFunction::EmitNeonRShiftImm(Value *Vec, Value *Shift, 2361 llvm::Type *Ty, bool usgn, 2362 const char *name) { 2363 llvm::VectorType *VTy = cast<llvm::VectorType>(Ty); 2364 2365 int ShiftAmt = cast<ConstantInt>(Shift)->getSExtValue(); 2366 int EltSize = VTy->getScalarSizeInBits(); 2367 2368 Vec = Builder.CreateBitCast(Vec, Ty); 2369 2370 // lshr/ashr are undefined when the shift amount is equal to the vector 2371 // element size. 2372 if (ShiftAmt == EltSize) { 2373 if (usgn) { 2374 // Right-shifting an unsigned value by its size yields 0. 2375 return llvm::ConstantAggregateZero::get(VTy); 2376 } else { 2377 // Right-shifting a signed value by its size is equivalent 2378 // to a shift of size-1. 2379 --ShiftAmt; 2380 Shift = ConstantInt::get(VTy->getElementType(), ShiftAmt); 2381 } 2382 } 2383 2384 Shift = EmitNeonShiftVector(Shift, Ty, false); 2385 if (usgn) 2386 return Builder.CreateLShr(Vec, Shift, name); 2387 else 2388 return Builder.CreateAShr(Vec, Shift, name); 2389 } 2390 2391 enum { 2392 AddRetType = (1 << 0), 2393 Add1ArgType = (1 << 1), 2394 Add2ArgTypes = (1 << 2), 2395 2396 VectorizeRetType = (1 << 3), 2397 VectorizeArgTypes = (1 << 4), 2398 2399 InventFloatType = (1 << 5), 2400 UnsignedAlts = (1 << 6), 2401 2402 Use64BitVectors = (1 << 7), 2403 Use128BitVectors = (1 << 8), 2404 2405 Vectorize1ArgType = Add1ArgType | VectorizeArgTypes, 2406 VectorRet = AddRetType | VectorizeRetType, 2407 VectorRetGetArgs01 = 2408 AddRetType | Add2ArgTypes | VectorizeRetType | VectorizeArgTypes, 2409 FpCmpzModifiers = 2410 AddRetType | VectorizeRetType | Add1ArgType | InventFloatType 2411 }; 2412 2413 namespace { 2414 struct NeonIntrinsicInfo { 2415 const char *NameHint; 2416 unsigned BuiltinID; 2417 unsigned LLVMIntrinsic; 2418 unsigned AltLLVMIntrinsic; 2419 unsigned TypeModifier; 2420 2421 bool operator<(unsigned RHSBuiltinID) const { 2422 return BuiltinID < RHSBuiltinID; 2423 } 2424 bool operator<(const NeonIntrinsicInfo &TE) const { 2425 return BuiltinID < TE.BuiltinID; 2426 } 2427 }; 2428 } // end anonymous namespace 2429 2430 #define NEONMAP0(NameBase) \ 2431 { #NameBase, NEON::BI__builtin_neon_ ## NameBase, 0, 0, 0 } 2432 2433 #define NEONMAP1(NameBase, LLVMIntrinsic, TypeModifier) \ 2434 { #NameBase, NEON:: BI__builtin_neon_ ## NameBase, \ 2435 Intrinsic::LLVMIntrinsic, 0, TypeModifier } 2436 2437 #define NEONMAP2(NameBase, LLVMIntrinsic, AltLLVMIntrinsic, TypeModifier) \ 2438 { #NameBase, NEON:: BI__builtin_neon_ ## NameBase, \ 2439 Intrinsic::LLVMIntrinsic, Intrinsic::AltLLVMIntrinsic, \ 2440 TypeModifier } 2441 2442 static const NeonIntrinsicInfo ARMSIMDIntrinsicMap [] = { 2443 NEONMAP2(vabd_v, arm_neon_vabdu, arm_neon_vabds, Add1ArgType | UnsignedAlts), 2444 NEONMAP2(vabdq_v, arm_neon_vabdu, arm_neon_vabds, Add1ArgType | UnsignedAlts), 2445 NEONMAP1(vabs_v, arm_neon_vabs, 0), 2446 NEONMAP1(vabsq_v, arm_neon_vabs, 0), 2447 NEONMAP0(vaddhn_v), 2448 NEONMAP1(vaesdq_v, arm_neon_aesd, 0), 2449 NEONMAP1(vaeseq_v, arm_neon_aese, 0), 2450 NEONMAP1(vaesimcq_v, arm_neon_aesimc, 0), 2451 NEONMAP1(vaesmcq_v, arm_neon_aesmc, 0), 2452 NEONMAP1(vbsl_v, arm_neon_vbsl, AddRetType), 2453 NEONMAP1(vbslq_v, arm_neon_vbsl, AddRetType), 2454 NEONMAP1(vcage_v, arm_neon_vacge, 0), 2455 NEONMAP1(vcageq_v, arm_neon_vacge, 0), 2456 NEONMAP1(vcagt_v, arm_neon_vacgt, 0), 2457 NEONMAP1(vcagtq_v, arm_neon_vacgt, 0), 2458 NEONMAP1(vcale_v, arm_neon_vacge, 0), 2459 NEONMAP1(vcaleq_v, arm_neon_vacge, 0), 2460 NEONMAP1(vcalt_v, arm_neon_vacgt, 0), 2461 NEONMAP1(vcaltq_v, arm_neon_vacgt, 0), 2462 NEONMAP1(vcls_v, arm_neon_vcls, Add1ArgType), 2463 NEONMAP1(vclsq_v, arm_neon_vcls, Add1ArgType), 2464 NEONMAP1(vclz_v, ctlz, Add1ArgType), 2465 NEONMAP1(vclzq_v, ctlz, Add1ArgType), 2466 NEONMAP1(vcnt_v, ctpop, Add1ArgType), 2467 NEONMAP1(vcntq_v, ctpop, Add1ArgType), 2468 NEONMAP1(vcvt_f16_f32, arm_neon_vcvtfp2hf, 0), 2469 NEONMAP1(vcvt_f32_f16, arm_neon_vcvthf2fp, 0), 2470 NEONMAP0(vcvt_f32_v), 2471 NEONMAP2(vcvt_n_f32_v, arm_neon_vcvtfxu2fp, arm_neon_vcvtfxs2fp, 0), 2472 NEONMAP1(vcvt_n_s32_v, arm_neon_vcvtfp2fxs, 0), 2473 NEONMAP1(vcvt_n_s64_v, arm_neon_vcvtfp2fxs, 0), 2474 NEONMAP1(vcvt_n_u32_v, arm_neon_vcvtfp2fxu, 0), 2475 NEONMAP1(vcvt_n_u64_v, arm_neon_vcvtfp2fxu, 0), 2476 NEONMAP0(vcvt_s32_v), 2477 NEONMAP0(vcvt_s64_v), 2478 NEONMAP0(vcvt_u32_v), 2479 NEONMAP0(vcvt_u64_v), 2480 NEONMAP1(vcvta_s32_v, arm_neon_vcvtas, 0), 2481 NEONMAP1(vcvta_s64_v, arm_neon_vcvtas, 0), 2482 NEONMAP1(vcvta_u32_v, arm_neon_vcvtau, 0), 2483 NEONMAP1(vcvta_u64_v, arm_neon_vcvtau, 0), 2484 NEONMAP1(vcvtaq_s32_v, arm_neon_vcvtas, 0), 2485 NEONMAP1(vcvtaq_s64_v, arm_neon_vcvtas, 0), 2486 NEONMAP1(vcvtaq_u32_v, arm_neon_vcvtau, 0), 2487 NEONMAP1(vcvtaq_u64_v, arm_neon_vcvtau, 0), 2488 NEONMAP1(vcvtm_s32_v, arm_neon_vcvtms, 0), 2489 NEONMAP1(vcvtm_s64_v, arm_neon_vcvtms, 0), 2490 NEONMAP1(vcvtm_u32_v, arm_neon_vcvtmu, 0), 2491 NEONMAP1(vcvtm_u64_v, arm_neon_vcvtmu, 0), 2492 NEONMAP1(vcvtmq_s32_v, arm_neon_vcvtms, 0), 2493 NEONMAP1(vcvtmq_s64_v, arm_neon_vcvtms, 0), 2494 NEONMAP1(vcvtmq_u32_v, arm_neon_vcvtmu, 0), 2495 NEONMAP1(vcvtmq_u64_v, arm_neon_vcvtmu, 0), 2496 NEONMAP1(vcvtn_s32_v, arm_neon_vcvtns, 0), 2497 NEONMAP1(vcvtn_s64_v, arm_neon_vcvtns, 0), 2498 NEONMAP1(vcvtn_u32_v, arm_neon_vcvtnu, 0), 2499 NEONMAP1(vcvtn_u64_v, arm_neon_vcvtnu, 0), 2500 NEONMAP1(vcvtnq_s32_v, arm_neon_vcvtns, 0), 2501 NEONMAP1(vcvtnq_s64_v, arm_neon_vcvtns, 0), 2502 NEONMAP1(vcvtnq_u32_v, arm_neon_vcvtnu, 0), 2503 NEONMAP1(vcvtnq_u64_v, arm_neon_vcvtnu, 0), 2504 NEONMAP1(vcvtp_s32_v, arm_neon_vcvtps, 0), 2505 NEONMAP1(vcvtp_s64_v, arm_neon_vcvtps, 0), 2506 NEONMAP1(vcvtp_u32_v, arm_neon_vcvtpu, 0), 2507 NEONMAP1(vcvtp_u64_v, arm_neon_vcvtpu, 0), 2508 NEONMAP1(vcvtpq_s32_v, arm_neon_vcvtps, 0), 2509 NEONMAP1(vcvtpq_s64_v, arm_neon_vcvtps, 0), 2510 NEONMAP1(vcvtpq_u32_v, arm_neon_vcvtpu, 0), 2511 NEONMAP1(vcvtpq_u64_v, arm_neon_vcvtpu, 0), 2512 NEONMAP0(vcvtq_f32_v), 2513 NEONMAP2(vcvtq_n_f32_v, arm_neon_vcvtfxu2fp, arm_neon_vcvtfxs2fp, 0), 2514 NEONMAP1(vcvtq_n_s32_v, arm_neon_vcvtfp2fxs, 0), 2515 NEONMAP1(vcvtq_n_s64_v, arm_neon_vcvtfp2fxs, 0), 2516 NEONMAP1(vcvtq_n_u32_v, arm_neon_vcvtfp2fxu, 0), 2517 NEONMAP1(vcvtq_n_u64_v, arm_neon_vcvtfp2fxu, 0), 2518 NEONMAP0(vcvtq_s32_v), 2519 NEONMAP0(vcvtq_s64_v), 2520 NEONMAP0(vcvtq_u32_v), 2521 NEONMAP0(vcvtq_u64_v), 2522 NEONMAP0(vext_v), 2523 NEONMAP0(vextq_v), 2524 NEONMAP0(vfma_v), 2525 NEONMAP0(vfmaq_v), 2526 NEONMAP2(vhadd_v, arm_neon_vhaddu, arm_neon_vhadds, Add1ArgType | UnsignedAlts), 2527 NEONMAP2(vhaddq_v, arm_neon_vhaddu, arm_neon_vhadds, Add1ArgType | UnsignedAlts), 2528 NEONMAP2(vhsub_v, arm_neon_vhsubu, arm_neon_vhsubs, Add1ArgType | UnsignedAlts), 2529 NEONMAP2(vhsubq_v, arm_neon_vhsubu, arm_neon_vhsubs, Add1ArgType | UnsignedAlts), 2530 NEONMAP0(vld1_dup_v), 2531 NEONMAP1(vld1_v, arm_neon_vld1, 0), 2532 NEONMAP0(vld1q_dup_v), 2533 NEONMAP1(vld1q_v, arm_neon_vld1, 0), 2534 NEONMAP1(vld2_lane_v, arm_neon_vld2lane, 0), 2535 NEONMAP1(vld2_v, arm_neon_vld2, 0), 2536 NEONMAP1(vld2q_lane_v, arm_neon_vld2lane, 0), 2537 NEONMAP1(vld2q_v, arm_neon_vld2, 0), 2538 NEONMAP1(vld3_lane_v, arm_neon_vld3lane, 0), 2539 NEONMAP1(vld3_v, arm_neon_vld3, 0), 2540 NEONMAP1(vld3q_lane_v, arm_neon_vld3lane, 0), 2541 NEONMAP1(vld3q_v, arm_neon_vld3, 0), 2542 NEONMAP1(vld4_lane_v, arm_neon_vld4lane, 0), 2543 NEONMAP1(vld4_v, arm_neon_vld4, 0), 2544 NEONMAP1(vld4q_lane_v, arm_neon_vld4lane, 0), 2545 NEONMAP1(vld4q_v, arm_neon_vld4, 0), 2546 NEONMAP2(vmax_v, arm_neon_vmaxu, arm_neon_vmaxs, Add1ArgType | UnsignedAlts), 2547 NEONMAP1(vmaxnm_v, arm_neon_vmaxnm, Add1ArgType), 2548 NEONMAP1(vmaxnmq_v, arm_neon_vmaxnm, Add1ArgType), 2549 NEONMAP2(vmaxq_v, arm_neon_vmaxu, arm_neon_vmaxs, Add1ArgType | UnsignedAlts), 2550 NEONMAP2(vmin_v, arm_neon_vminu, arm_neon_vmins, Add1ArgType | UnsignedAlts), 2551 NEONMAP1(vminnm_v, arm_neon_vminnm, Add1ArgType), 2552 NEONMAP1(vminnmq_v, arm_neon_vminnm, Add1ArgType), 2553 NEONMAP2(vminq_v, arm_neon_vminu, arm_neon_vmins, Add1ArgType | UnsignedAlts), 2554 NEONMAP0(vmovl_v), 2555 NEONMAP0(vmovn_v), 2556 NEONMAP1(vmul_v, arm_neon_vmulp, Add1ArgType), 2557 NEONMAP0(vmull_v), 2558 NEONMAP1(vmulq_v, arm_neon_vmulp, Add1ArgType), 2559 NEONMAP2(vpadal_v, arm_neon_vpadalu, arm_neon_vpadals, UnsignedAlts), 2560 NEONMAP2(vpadalq_v, arm_neon_vpadalu, arm_neon_vpadals, UnsignedAlts), 2561 NEONMAP1(vpadd_v, arm_neon_vpadd, Add1ArgType), 2562 NEONMAP2(vpaddl_v, arm_neon_vpaddlu, arm_neon_vpaddls, UnsignedAlts), 2563 NEONMAP2(vpaddlq_v, arm_neon_vpaddlu, arm_neon_vpaddls, UnsignedAlts), 2564 NEONMAP1(vpaddq_v, arm_neon_vpadd, Add1ArgType), 2565 NEONMAP2(vpmax_v, arm_neon_vpmaxu, arm_neon_vpmaxs, Add1ArgType | UnsignedAlts), 2566 NEONMAP2(vpmin_v, arm_neon_vpminu, arm_neon_vpmins, Add1ArgType | UnsignedAlts), 2567 NEONMAP1(vqabs_v, arm_neon_vqabs, Add1ArgType), 2568 NEONMAP1(vqabsq_v, arm_neon_vqabs, Add1ArgType), 2569 NEONMAP2(vqadd_v, arm_neon_vqaddu, arm_neon_vqadds, Add1ArgType | UnsignedAlts), 2570 NEONMAP2(vqaddq_v, arm_neon_vqaddu, arm_neon_vqadds, Add1ArgType | UnsignedAlts), 2571 NEONMAP2(vqdmlal_v, arm_neon_vqdmull, arm_neon_vqadds, 0), 2572 NEONMAP2(vqdmlsl_v, arm_neon_vqdmull, arm_neon_vqsubs, 0), 2573 NEONMAP1(vqdmulh_v, arm_neon_vqdmulh, Add1ArgType), 2574 NEONMAP1(vqdmulhq_v, arm_neon_vqdmulh, Add1ArgType), 2575 NEONMAP1(vqdmull_v, arm_neon_vqdmull, Add1ArgType), 2576 NEONMAP2(vqmovn_v, arm_neon_vqmovnu, arm_neon_vqmovns, Add1ArgType | UnsignedAlts), 2577 NEONMAP1(vqmovun_v, arm_neon_vqmovnsu, Add1ArgType), 2578 NEONMAP1(vqneg_v, arm_neon_vqneg, Add1ArgType), 2579 NEONMAP1(vqnegq_v, arm_neon_vqneg, Add1ArgType), 2580 NEONMAP1(vqrdmulh_v, arm_neon_vqrdmulh, Add1ArgType), 2581 NEONMAP1(vqrdmulhq_v, arm_neon_vqrdmulh, Add1ArgType), 2582 NEONMAP2(vqrshl_v, arm_neon_vqrshiftu, arm_neon_vqrshifts, Add1ArgType | UnsignedAlts), 2583 NEONMAP2(vqrshlq_v, arm_neon_vqrshiftu, arm_neon_vqrshifts, Add1ArgType | UnsignedAlts), 2584 NEONMAP2(vqshl_n_v, arm_neon_vqshiftu, arm_neon_vqshifts, UnsignedAlts), 2585 NEONMAP2(vqshl_v, arm_neon_vqshiftu, arm_neon_vqshifts, Add1ArgType | UnsignedAlts), 2586 NEONMAP2(vqshlq_n_v, arm_neon_vqshiftu, arm_neon_vqshifts, UnsignedAlts), 2587 NEONMAP2(vqshlq_v, arm_neon_vqshiftu, arm_neon_vqshifts, Add1ArgType | UnsignedAlts), 2588 NEONMAP1(vqshlu_n_v, arm_neon_vqshiftsu, 0), 2589 NEONMAP1(vqshluq_n_v, arm_neon_vqshiftsu, 0), 2590 NEONMAP2(vqsub_v, arm_neon_vqsubu, arm_neon_vqsubs, Add1ArgType | UnsignedAlts), 2591 NEONMAP2(vqsubq_v, arm_neon_vqsubu, arm_neon_vqsubs, Add1ArgType | UnsignedAlts), 2592 NEONMAP1(vraddhn_v, arm_neon_vraddhn, Add1ArgType), 2593 NEONMAP2(vrecpe_v, arm_neon_vrecpe, arm_neon_vrecpe, 0), 2594 NEONMAP2(vrecpeq_v, arm_neon_vrecpe, arm_neon_vrecpe, 0), 2595 NEONMAP1(vrecps_v, arm_neon_vrecps, Add1ArgType), 2596 NEONMAP1(vrecpsq_v, arm_neon_vrecps, Add1ArgType), 2597 NEONMAP2(vrhadd_v, arm_neon_vrhaddu, arm_neon_vrhadds, Add1ArgType | UnsignedAlts), 2598 NEONMAP2(vrhaddq_v, arm_neon_vrhaddu, arm_neon_vrhadds, Add1ArgType | UnsignedAlts), 2599 NEONMAP1(vrnd_v, arm_neon_vrintz, Add1ArgType), 2600 NEONMAP1(vrnda_v, arm_neon_vrinta, Add1ArgType), 2601 NEONMAP1(vrndaq_v, arm_neon_vrinta, Add1ArgType), 2602 NEONMAP1(vrndm_v, arm_neon_vrintm, Add1ArgType), 2603 NEONMAP1(vrndmq_v, arm_neon_vrintm, Add1ArgType), 2604 NEONMAP1(vrndn_v, arm_neon_vrintn, Add1ArgType), 2605 NEONMAP1(vrndnq_v, arm_neon_vrintn, Add1ArgType), 2606 NEONMAP1(vrndp_v, arm_neon_vrintp, Add1ArgType), 2607 NEONMAP1(vrndpq_v, arm_neon_vrintp, Add1ArgType), 2608 NEONMAP1(vrndq_v, arm_neon_vrintz, Add1ArgType), 2609 NEONMAP1(vrndx_v, arm_neon_vrintx, Add1ArgType), 2610 NEONMAP1(vrndxq_v, arm_neon_vrintx, Add1ArgType), 2611 NEONMAP2(vrshl_v, arm_neon_vrshiftu, arm_neon_vrshifts, Add1ArgType | UnsignedAlts), 2612 NEONMAP2(vrshlq_v, arm_neon_vrshiftu, arm_neon_vrshifts, Add1ArgType | UnsignedAlts), 2613 NEONMAP2(vrshr_n_v, arm_neon_vrshiftu, arm_neon_vrshifts, UnsignedAlts), 2614 NEONMAP2(vrshrq_n_v, arm_neon_vrshiftu, arm_neon_vrshifts, UnsignedAlts), 2615 NEONMAP2(vrsqrte_v, arm_neon_vrsqrte, arm_neon_vrsqrte, 0), 2616 NEONMAP2(vrsqrteq_v, arm_neon_vrsqrte, arm_neon_vrsqrte, 0), 2617 NEONMAP1(vrsqrts_v, arm_neon_vrsqrts, Add1ArgType), 2618 NEONMAP1(vrsqrtsq_v, arm_neon_vrsqrts, Add1ArgType), 2619 NEONMAP1(vrsubhn_v, arm_neon_vrsubhn, Add1ArgType), 2620 NEONMAP1(vsha1su0q_v, arm_neon_sha1su0, 0), 2621 NEONMAP1(vsha1su1q_v, arm_neon_sha1su1, 0), 2622 NEONMAP1(vsha256h2q_v, arm_neon_sha256h2, 0), 2623 NEONMAP1(vsha256hq_v, arm_neon_sha256h, 0), 2624 NEONMAP1(vsha256su0q_v, arm_neon_sha256su0, 0), 2625 NEONMAP1(vsha256su1q_v, arm_neon_sha256su1, 0), 2626 NEONMAP0(vshl_n_v), 2627 NEONMAP2(vshl_v, arm_neon_vshiftu, arm_neon_vshifts, Add1ArgType | UnsignedAlts), 2628 NEONMAP0(vshll_n_v), 2629 NEONMAP0(vshlq_n_v), 2630 NEONMAP2(vshlq_v, arm_neon_vshiftu, arm_neon_vshifts, Add1ArgType | UnsignedAlts), 2631 NEONMAP0(vshr_n_v), 2632 NEONMAP0(vshrn_n_v), 2633 NEONMAP0(vshrq_n_v), 2634 NEONMAP1(vst1_v, arm_neon_vst1, 0), 2635 NEONMAP1(vst1q_v, arm_neon_vst1, 0), 2636 NEONMAP1(vst2_lane_v, arm_neon_vst2lane, 0), 2637 NEONMAP1(vst2_v, arm_neon_vst2, 0), 2638 NEONMAP1(vst2q_lane_v, arm_neon_vst2lane, 0), 2639 NEONMAP1(vst2q_v, arm_neon_vst2, 0), 2640 NEONMAP1(vst3_lane_v, arm_neon_vst3lane, 0), 2641 NEONMAP1(vst3_v, arm_neon_vst3, 0), 2642 NEONMAP1(vst3q_lane_v, arm_neon_vst3lane, 0), 2643 NEONMAP1(vst3q_v, arm_neon_vst3, 0), 2644 NEONMAP1(vst4_lane_v, arm_neon_vst4lane, 0), 2645 NEONMAP1(vst4_v, arm_neon_vst4, 0), 2646 NEONMAP1(vst4q_lane_v, arm_neon_vst4lane, 0), 2647 NEONMAP1(vst4q_v, arm_neon_vst4, 0), 2648 NEONMAP0(vsubhn_v), 2649 NEONMAP0(vtrn_v), 2650 NEONMAP0(vtrnq_v), 2651 NEONMAP0(vtst_v), 2652 NEONMAP0(vtstq_v), 2653 NEONMAP0(vuzp_v), 2654 NEONMAP0(vuzpq_v), 2655 NEONMAP0(vzip_v), 2656 NEONMAP0(vzipq_v) 2657 }; 2658 2659 static const NeonIntrinsicInfo AArch64SIMDIntrinsicMap[] = { 2660 NEONMAP1(vabs_v, aarch64_neon_abs, 0), 2661 NEONMAP1(vabsq_v, aarch64_neon_abs, 0), 2662 NEONMAP0(vaddhn_v), 2663 NEONMAP1(vaesdq_v, aarch64_crypto_aesd, 0), 2664 NEONMAP1(vaeseq_v, aarch64_crypto_aese, 0), 2665 NEONMAP1(vaesimcq_v, aarch64_crypto_aesimc, 0), 2666 NEONMAP1(vaesmcq_v, aarch64_crypto_aesmc, 0), 2667 NEONMAP1(vcage_v, aarch64_neon_facge, 0), 2668 NEONMAP1(vcageq_v, aarch64_neon_facge, 0), 2669 NEONMAP1(vcagt_v, aarch64_neon_facgt, 0), 2670 NEONMAP1(vcagtq_v, aarch64_neon_facgt, 0), 2671 NEONMAP1(vcale_v, aarch64_neon_facge, 0), 2672 NEONMAP1(vcaleq_v, aarch64_neon_facge, 0), 2673 NEONMAP1(vcalt_v, aarch64_neon_facgt, 0), 2674 NEONMAP1(vcaltq_v, aarch64_neon_facgt, 0), 2675 NEONMAP1(vcls_v, aarch64_neon_cls, Add1ArgType), 2676 NEONMAP1(vclsq_v, aarch64_neon_cls, Add1ArgType), 2677 NEONMAP1(vclz_v, ctlz, Add1ArgType), 2678 NEONMAP1(vclzq_v, ctlz, Add1ArgType), 2679 NEONMAP1(vcnt_v, ctpop, Add1ArgType), 2680 NEONMAP1(vcntq_v, ctpop, Add1ArgType), 2681 NEONMAP1(vcvt_f16_f32, aarch64_neon_vcvtfp2hf, 0), 2682 NEONMAP1(vcvt_f32_f16, aarch64_neon_vcvthf2fp, 0), 2683 NEONMAP0(vcvt_f32_v), 2684 NEONMAP2(vcvt_n_f32_v, aarch64_neon_vcvtfxu2fp, aarch64_neon_vcvtfxs2fp, 0), 2685 NEONMAP2(vcvt_n_f64_v, aarch64_neon_vcvtfxu2fp, aarch64_neon_vcvtfxs2fp, 0), 2686 NEONMAP1(vcvt_n_s32_v, aarch64_neon_vcvtfp2fxs, 0), 2687 NEONMAP1(vcvt_n_s64_v, aarch64_neon_vcvtfp2fxs, 0), 2688 NEONMAP1(vcvt_n_u32_v, aarch64_neon_vcvtfp2fxu, 0), 2689 NEONMAP1(vcvt_n_u64_v, aarch64_neon_vcvtfp2fxu, 0), 2690 NEONMAP0(vcvtq_f32_v), 2691 NEONMAP2(vcvtq_n_f32_v, aarch64_neon_vcvtfxu2fp, aarch64_neon_vcvtfxs2fp, 0), 2692 NEONMAP2(vcvtq_n_f64_v, aarch64_neon_vcvtfxu2fp, aarch64_neon_vcvtfxs2fp, 0), 2693 NEONMAP1(vcvtq_n_s32_v, aarch64_neon_vcvtfp2fxs, 0), 2694 NEONMAP1(vcvtq_n_s64_v, aarch64_neon_vcvtfp2fxs, 0), 2695 NEONMAP1(vcvtq_n_u32_v, aarch64_neon_vcvtfp2fxu, 0), 2696 NEONMAP1(vcvtq_n_u64_v, aarch64_neon_vcvtfp2fxu, 0), 2697 NEONMAP1(vcvtx_f32_v, aarch64_neon_fcvtxn, AddRetType | Add1ArgType), 2698 NEONMAP0(vext_v), 2699 NEONMAP0(vextq_v), 2700 NEONMAP0(vfma_v), 2701 NEONMAP0(vfmaq_v), 2702 NEONMAP2(vhadd_v, aarch64_neon_uhadd, aarch64_neon_shadd, Add1ArgType | UnsignedAlts), 2703 NEONMAP2(vhaddq_v, aarch64_neon_uhadd, aarch64_neon_shadd, Add1ArgType | UnsignedAlts), 2704 NEONMAP2(vhsub_v, aarch64_neon_uhsub, aarch64_neon_shsub, Add1ArgType | UnsignedAlts), 2705 NEONMAP2(vhsubq_v, aarch64_neon_uhsub, aarch64_neon_shsub, Add1ArgType | UnsignedAlts), 2706 NEONMAP0(vmovl_v), 2707 NEONMAP0(vmovn_v), 2708 NEONMAP1(vmul_v, aarch64_neon_pmul, Add1ArgType), 2709 NEONMAP1(vmulq_v, aarch64_neon_pmul, Add1ArgType), 2710 NEONMAP1(vpadd_v, aarch64_neon_addp, Add1ArgType), 2711 NEONMAP2(vpaddl_v, aarch64_neon_uaddlp, aarch64_neon_saddlp, UnsignedAlts), 2712 NEONMAP2(vpaddlq_v, aarch64_neon_uaddlp, aarch64_neon_saddlp, UnsignedAlts), 2713 NEONMAP1(vpaddq_v, aarch64_neon_addp, Add1ArgType), 2714 NEONMAP1(vqabs_v, aarch64_neon_sqabs, Add1ArgType), 2715 NEONMAP1(vqabsq_v, aarch64_neon_sqabs, Add1ArgType), 2716 NEONMAP2(vqadd_v, aarch64_neon_uqadd, aarch64_neon_sqadd, Add1ArgType | UnsignedAlts), 2717 NEONMAP2(vqaddq_v, aarch64_neon_uqadd, aarch64_neon_sqadd, Add1ArgType | UnsignedAlts), 2718 NEONMAP2(vqdmlal_v, aarch64_neon_sqdmull, aarch64_neon_sqadd, 0), 2719 NEONMAP2(vqdmlsl_v, aarch64_neon_sqdmull, aarch64_neon_sqsub, 0), 2720 NEONMAP1(vqdmulh_v, aarch64_neon_sqdmulh, Add1ArgType), 2721 NEONMAP1(vqdmulhq_v, aarch64_neon_sqdmulh, Add1ArgType), 2722 NEONMAP1(vqdmull_v, aarch64_neon_sqdmull, Add1ArgType), 2723 NEONMAP2(vqmovn_v, aarch64_neon_uqxtn, aarch64_neon_sqxtn, Add1ArgType | UnsignedAlts), 2724 NEONMAP1(vqmovun_v, aarch64_neon_sqxtun, Add1ArgType), 2725 NEONMAP1(vqneg_v, aarch64_neon_sqneg, Add1ArgType), 2726 NEONMAP1(vqnegq_v, aarch64_neon_sqneg, Add1ArgType), 2727 NEONMAP1(vqrdmulh_v, aarch64_neon_sqrdmulh, Add1ArgType), 2728 NEONMAP1(vqrdmulhq_v, aarch64_neon_sqrdmulh, Add1ArgType), 2729 NEONMAP2(vqrshl_v, aarch64_neon_uqrshl, aarch64_neon_sqrshl, Add1ArgType | UnsignedAlts), 2730 NEONMAP2(vqrshlq_v, aarch64_neon_uqrshl, aarch64_neon_sqrshl, Add1ArgType | UnsignedAlts), 2731 NEONMAP2(vqshl_n_v, aarch64_neon_uqshl, aarch64_neon_sqshl, UnsignedAlts), 2732 NEONMAP2(vqshl_v, aarch64_neon_uqshl, aarch64_neon_sqshl, Add1ArgType | UnsignedAlts), 2733 NEONMAP2(vqshlq_n_v, aarch64_neon_uqshl, aarch64_neon_sqshl,UnsignedAlts), 2734 NEONMAP2(vqshlq_v, aarch64_neon_uqshl, aarch64_neon_sqshl, Add1ArgType | UnsignedAlts), 2735 NEONMAP1(vqshlu_n_v, aarch64_neon_sqshlu, 0), 2736 NEONMAP1(vqshluq_n_v, aarch64_neon_sqshlu, 0), 2737 NEONMAP2(vqsub_v, aarch64_neon_uqsub, aarch64_neon_sqsub, Add1ArgType | UnsignedAlts), 2738 NEONMAP2(vqsubq_v, aarch64_neon_uqsub, aarch64_neon_sqsub, Add1ArgType | UnsignedAlts), 2739 NEONMAP1(vraddhn_v, aarch64_neon_raddhn, Add1ArgType), 2740 NEONMAP2(vrecpe_v, aarch64_neon_frecpe, aarch64_neon_urecpe, 0), 2741 NEONMAP2(vrecpeq_v, aarch64_neon_frecpe, aarch64_neon_urecpe, 0), 2742 NEONMAP1(vrecps_v, aarch64_neon_frecps, Add1ArgType), 2743 NEONMAP1(vrecpsq_v, aarch64_neon_frecps, Add1ArgType), 2744 NEONMAP2(vrhadd_v, aarch64_neon_urhadd, aarch64_neon_srhadd, Add1ArgType | UnsignedAlts), 2745 NEONMAP2(vrhaddq_v, aarch64_neon_urhadd, aarch64_neon_srhadd, Add1ArgType | UnsignedAlts), 2746 NEONMAP2(vrshl_v, aarch64_neon_urshl, aarch64_neon_srshl, Add1ArgType | UnsignedAlts), 2747 NEONMAP2(vrshlq_v, aarch64_neon_urshl, aarch64_neon_srshl, Add1ArgType | UnsignedAlts), 2748 NEONMAP2(vrshr_n_v, aarch64_neon_urshl, aarch64_neon_srshl, UnsignedAlts), 2749 NEONMAP2(vrshrq_n_v, aarch64_neon_urshl, aarch64_neon_srshl, UnsignedAlts), 2750 NEONMAP2(vrsqrte_v, aarch64_neon_frsqrte, aarch64_neon_ursqrte, 0), 2751 NEONMAP2(vrsqrteq_v, aarch64_neon_frsqrte, aarch64_neon_ursqrte, 0), 2752 NEONMAP1(vrsqrts_v, aarch64_neon_frsqrts, Add1ArgType), 2753 NEONMAP1(vrsqrtsq_v, aarch64_neon_frsqrts, Add1ArgType), 2754 NEONMAP1(vrsubhn_v, aarch64_neon_rsubhn, Add1ArgType), 2755 NEONMAP1(vsha1su0q_v, aarch64_crypto_sha1su0, 0), 2756 NEONMAP1(vsha1su1q_v, aarch64_crypto_sha1su1, 0), 2757 NEONMAP1(vsha256h2q_v, aarch64_crypto_sha256h2, 0), 2758 NEONMAP1(vsha256hq_v, aarch64_crypto_sha256h, 0), 2759 NEONMAP1(vsha256su0q_v, aarch64_crypto_sha256su0, 0), 2760 NEONMAP1(vsha256su1q_v, aarch64_crypto_sha256su1, 0), 2761 NEONMAP0(vshl_n_v), 2762 NEONMAP2(vshl_v, aarch64_neon_ushl, aarch64_neon_sshl, Add1ArgType | UnsignedAlts), 2763 NEONMAP0(vshll_n_v), 2764 NEONMAP0(vshlq_n_v), 2765 NEONMAP2(vshlq_v, aarch64_neon_ushl, aarch64_neon_sshl, Add1ArgType | UnsignedAlts), 2766 NEONMAP0(vshr_n_v), 2767 NEONMAP0(vshrn_n_v), 2768 NEONMAP0(vshrq_n_v), 2769 NEONMAP0(vsubhn_v), 2770 NEONMAP0(vtst_v), 2771 NEONMAP0(vtstq_v), 2772 }; 2773 2774 static const NeonIntrinsicInfo AArch64SISDIntrinsicMap[] = { 2775 NEONMAP1(vabdd_f64, aarch64_sisd_fabd, Add1ArgType), 2776 NEONMAP1(vabds_f32, aarch64_sisd_fabd, Add1ArgType), 2777 NEONMAP1(vabsd_s64, aarch64_neon_abs, Add1ArgType), 2778 NEONMAP1(vaddlv_s32, aarch64_neon_saddlv, AddRetType | Add1ArgType), 2779 NEONMAP1(vaddlv_u32, aarch64_neon_uaddlv, AddRetType | Add1ArgType), 2780 NEONMAP1(vaddlvq_s32, aarch64_neon_saddlv, AddRetType | Add1ArgType), 2781 NEONMAP1(vaddlvq_u32, aarch64_neon_uaddlv, AddRetType | Add1ArgType), 2782 NEONMAP1(vaddv_f32, aarch64_neon_faddv, AddRetType | Add1ArgType), 2783 NEONMAP1(vaddv_s32, aarch64_neon_saddv, AddRetType | Add1ArgType), 2784 NEONMAP1(vaddv_u32, aarch64_neon_uaddv, AddRetType | Add1ArgType), 2785 NEONMAP1(vaddvq_f32, aarch64_neon_faddv, AddRetType | Add1ArgType), 2786 NEONMAP1(vaddvq_f64, aarch64_neon_faddv, AddRetType | Add1ArgType), 2787 NEONMAP1(vaddvq_s32, aarch64_neon_saddv, AddRetType | Add1ArgType), 2788 NEONMAP1(vaddvq_s64, aarch64_neon_saddv, AddRetType | Add1ArgType), 2789 NEONMAP1(vaddvq_u32, aarch64_neon_uaddv, AddRetType | Add1ArgType), 2790 NEONMAP1(vaddvq_u64, aarch64_neon_uaddv, AddRetType | Add1ArgType), 2791 NEONMAP1(vcaged_f64, aarch64_neon_facge, AddRetType | Add1ArgType), 2792 NEONMAP1(vcages_f32, aarch64_neon_facge, AddRetType | Add1ArgType), 2793 NEONMAP1(vcagtd_f64, aarch64_neon_facgt, AddRetType | Add1ArgType), 2794 NEONMAP1(vcagts_f32, aarch64_neon_facgt, AddRetType | Add1ArgType), 2795 NEONMAP1(vcaled_f64, aarch64_neon_facge, AddRetType | Add1ArgType), 2796 NEONMAP1(vcales_f32, aarch64_neon_facge, AddRetType | Add1ArgType), 2797 NEONMAP1(vcaltd_f64, aarch64_neon_facgt, AddRetType | Add1ArgType), 2798 NEONMAP1(vcalts_f32, aarch64_neon_facgt, AddRetType | Add1ArgType), 2799 NEONMAP1(vcvtad_s64_f64, aarch64_neon_fcvtas, AddRetType | Add1ArgType), 2800 NEONMAP1(vcvtad_u64_f64, aarch64_neon_fcvtau, AddRetType | Add1ArgType), 2801 NEONMAP1(vcvtas_s32_f32, aarch64_neon_fcvtas, AddRetType | Add1ArgType), 2802 NEONMAP1(vcvtas_u32_f32, aarch64_neon_fcvtau, AddRetType | Add1ArgType), 2803 NEONMAP1(vcvtd_n_f64_s64, aarch64_neon_vcvtfxs2fp, AddRetType | Add1ArgType), 2804 NEONMAP1(vcvtd_n_f64_u64, aarch64_neon_vcvtfxu2fp, AddRetType | Add1ArgType), 2805 NEONMAP1(vcvtd_n_s64_f64, aarch64_neon_vcvtfp2fxs, AddRetType | Add1ArgType), 2806 NEONMAP1(vcvtd_n_u64_f64, aarch64_neon_vcvtfp2fxu, AddRetType | Add1ArgType), 2807 NEONMAP1(vcvtmd_s64_f64, aarch64_neon_fcvtms, AddRetType | Add1ArgType), 2808 NEONMAP1(vcvtmd_u64_f64, aarch64_neon_fcvtmu, AddRetType | Add1ArgType), 2809 NEONMAP1(vcvtms_s32_f32, aarch64_neon_fcvtms, AddRetType | Add1ArgType), 2810 NEONMAP1(vcvtms_u32_f32, aarch64_neon_fcvtmu, AddRetType | Add1ArgType), 2811 NEONMAP1(vcvtnd_s64_f64, aarch64_neon_fcvtns, AddRetType | Add1ArgType), 2812 NEONMAP1(vcvtnd_u64_f64, aarch64_neon_fcvtnu, AddRetType | Add1ArgType), 2813 NEONMAP1(vcvtns_s32_f32, aarch64_neon_fcvtns, AddRetType | Add1ArgType), 2814 NEONMAP1(vcvtns_u32_f32, aarch64_neon_fcvtnu, AddRetType | Add1ArgType), 2815 NEONMAP1(vcvtpd_s64_f64, aarch64_neon_fcvtps, AddRetType | Add1ArgType), 2816 NEONMAP1(vcvtpd_u64_f64, aarch64_neon_fcvtpu, AddRetType | Add1ArgType), 2817 NEONMAP1(vcvtps_s32_f32, aarch64_neon_fcvtps, AddRetType | Add1ArgType), 2818 NEONMAP1(vcvtps_u32_f32, aarch64_neon_fcvtpu, AddRetType | Add1ArgType), 2819 NEONMAP1(vcvts_n_f32_s32, aarch64_neon_vcvtfxs2fp, AddRetType | Add1ArgType), 2820 NEONMAP1(vcvts_n_f32_u32, aarch64_neon_vcvtfxu2fp, AddRetType | Add1ArgType), 2821 NEONMAP1(vcvts_n_s32_f32, aarch64_neon_vcvtfp2fxs, AddRetType | Add1ArgType), 2822 NEONMAP1(vcvts_n_u32_f32, aarch64_neon_vcvtfp2fxu, AddRetType | Add1ArgType), 2823 NEONMAP1(vcvtxd_f32_f64, aarch64_sisd_fcvtxn, 0), 2824 NEONMAP1(vmaxnmv_f32, aarch64_neon_fmaxnmv, AddRetType | Add1ArgType), 2825 NEONMAP1(vmaxnmvq_f32, aarch64_neon_fmaxnmv, AddRetType | Add1ArgType), 2826 NEONMAP1(vmaxnmvq_f64, aarch64_neon_fmaxnmv, AddRetType | Add1ArgType), 2827 NEONMAP1(vmaxv_f32, aarch64_neon_fmaxv, AddRetType | Add1ArgType), 2828 NEONMAP1(vmaxv_s32, aarch64_neon_smaxv, AddRetType | Add1ArgType), 2829 NEONMAP1(vmaxv_u32, aarch64_neon_umaxv, AddRetType | Add1ArgType), 2830 NEONMAP1(vmaxvq_f32, aarch64_neon_fmaxv, AddRetType | Add1ArgType), 2831 NEONMAP1(vmaxvq_f64, aarch64_neon_fmaxv, AddRetType | Add1ArgType), 2832 NEONMAP1(vmaxvq_s32, aarch64_neon_smaxv, AddRetType | Add1ArgType), 2833 NEONMAP1(vmaxvq_u32, aarch64_neon_umaxv, AddRetType | Add1ArgType), 2834 NEONMAP1(vminnmv_f32, aarch64_neon_fminnmv, AddRetType | Add1ArgType), 2835 NEONMAP1(vminnmvq_f32, aarch64_neon_fminnmv, AddRetType | Add1ArgType), 2836 NEONMAP1(vminnmvq_f64, aarch64_neon_fminnmv, AddRetType | Add1ArgType), 2837 NEONMAP1(vminv_f32, aarch64_neon_fminv, AddRetType | Add1ArgType), 2838 NEONMAP1(vminv_s32, aarch64_neon_sminv, AddRetType | Add1ArgType), 2839 NEONMAP1(vminv_u32, aarch64_neon_uminv, AddRetType | Add1ArgType), 2840 NEONMAP1(vminvq_f32, aarch64_neon_fminv, AddRetType | Add1ArgType), 2841 NEONMAP1(vminvq_f64, aarch64_neon_fminv, AddRetType | Add1ArgType), 2842 NEONMAP1(vminvq_s32, aarch64_neon_sminv, AddRetType | Add1ArgType), 2843 NEONMAP1(vminvq_u32, aarch64_neon_uminv, AddRetType | Add1ArgType), 2844 NEONMAP1(vmull_p64, aarch64_neon_pmull64, 0), 2845 NEONMAP1(vmulxd_f64, aarch64_neon_fmulx, Add1ArgType), 2846 NEONMAP1(vmulxs_f32, aarch64_neon_fmulx, Add1ArgType), 2847 NEONMAP1(vpaddd_s64, aarch64_neon_uaddv, AddRetType | Add1ArgType), 2848 NEONMAP1(vpaddd_u64, aarch64_neon_uaddv, AddRetType | Add1ArgType), 2849 NEONMAP1(vpmaxnmqd_f64, aarch64_neon_fmaxnmv, AddRetType | Add1ArgType), 2850 NEONMAP1(vpmaxnms_f32, aarch64_neon_fmaxnmv, AddRetType | Add1ArgType), 2851 NEONMAP1(vpmaxqd_f64, aarch64_neon_fmaxv, AddRetType | Add1ArgType), 2852 NEONMAP1(vpmaxs_f32, aarch64_neon_fmaxv, AddRetType | Add1ArgType), 2853 NEONMAP1(vpminnmqd_f64, aarch64_neon_fminnmv, AddRetType | Add1ArgType), 2854 NEONMAP1(vpminnms_f32, aarch64_neon_fminnmv, AddRetType | Add1ArgType), 2855 NEONMAP1(vpminqd_f64, aarch64_neon_fminv, AddRetType | Add1ArgType), 2856 NEONMAP1(vpmins_f32, aarch64_neon_fminv, AddRetType | Add1ArgType), 2857 NEONMAP1(vqabsb_s8, aarch64_neon_sqabs, Vectorize1ArgType | Use64BitVectors), 2858 NEONMAP1(vqabsd_s64, aarch64_neon_sqabs, Add1ArgType), 2859 NEONMAP1(vqabsh_s16, aarch64_neon_sqabs, Vectorize1ArgType | Use64BitVectors), 2860 NEONMAP1(vqabss_s32, aarch64_neon_sqabs, Add1ArgType), 2861 NEONMAP1(vqaddb_s8, aarch64_neon_sqadd, Vectorize1ArgType | Use64BitVectors), 2862 NEONMAP1(vqaddb_u8, aarch64_neon_uqadd, Vectorize1ArgType | Use64BitVectors), 2863 NEONMAP1(vqaddd_s64, aarch64_neon_sqadd, Add1ArgType), 2864 NEONMAP1(vqaddd_u64, aarch64_neon_uqadd, Add1ArgType), 2865 NEONMAP1(vqaddh_s16, aarch64_neon_sqadd, Vectorize1ArgType | Use64BitVectors), 2866 NEONMAP1(vqaddh_u16, aarch64_neon_uqadd, Vectorize1ArgType | Use64BitVectors), 2867 NEONMAP1(vqadds_s32, aarch64_neon_sqadd, Add1ArgType), 2868 NEONMAP1(vqadds_u32, aarch64_neon_uqadd, Add1ArgType), 2869 NEONMAP1(vqdmulhh_s16, aarch64_neon_sqdmulh, Vectorize1ArgType | Use64BitVectors), 2870 NEONMAP1(vqdmulhs_s32, aarch64_neon_sqdmulh, Add1ArgType), 2871 NEONMAP1(vqdmullh_s16, aarch64_neon_sqdmull, VectorRet | Use128BitVectors), 2872 NEONMAP1(vqdmulls_s32, aarch64_neon_sqdmulls_scalar, 0), 2873 NEONMAP1(vqmovnd_s64, aarch64_neon_scalar_sqxtn, AddRetType | Add1ArgType), 2874 NEONMAP1(vqmovnd_u64, aarch64_neon_scalar_uqxtn, AddRetType | Add1ArgType), 2875 NEONMAP1(vqmovnh_s16, aarch64_neon_sqxtn, VectorRet | Use64BitVectors), 2876 NEONMAP1(vqmovnh_u16, aarch64_neon_uqxtn, VectorRet | Use64BitVectors), 2877 NEONMAP1(vqmovns_s32, aarch64_neon_sqxtn, VectorRet | Use64BitVectors), 2878 NEONMAP1(vqmovns_u32, aarch64_neon_uqxtn, VectorRet | Use64BitVectors), 2879 NEONMAP1(vqmovund_s64, aarch64_neon_scalar_sqxtun, AddRetType | Add1ArgType), 2880 NEONMAP1(vqmovunh_s16, aarch64_neon_sqxtun, VectorRet | Use64BitVectors), 2881 NEONMAP1(vqmovuns_s32, aarch64_neon_sqxtun, VectorRet | Use64BitVectors), 2882 NEONMAP1(vqnegb_s8, aarch64_neon_sqneg, Vectorize1ArgType | Use64BitVectors), 2883 NEONMAP1(vqnegd_s64, aarch64_neon_sqneg, Add1ArgType), 2884 NEONMAP1(vqnegh_s16, aarch64_neon_sqneg, Vectorize1ArgType | Use64BitVectors), 2885 NEONMAP1(vqnegs_s32, aarch64_neon_sqneg, Add1ArgType), 2886 NEONMAP1(vqrdmulhh_s16, aarch64_neon_sqrdmulh, Vectorize1ArgType | Use64BitVectors), 2887 NEONMAP1(vqrdmulhs_s32, aarch64_neon_sqrdmulh, Add1ArgType), 2888 NEONMAP1(vqrshlb_s8, aarch64_neon_sqrshl, Vectorize1ArgType | Use64BitVectors), 2889 NEONMAP1(vqrshlb_u8, aarch64_neon_uqrshl, Vectorize1ArgType | Use64BitVectors), 2890 NEONMAP1(vqrshld_s64, aarch64_neon_sqrshl, Add1ArgType), 2891 NEONMAP1(vqrshld_u64, aarch64_neon_uqrshl, Add1ArgType), 2892 NEONMAP1(vqrshlh_s16, aarch64_neon_sqrshl, Vectorize1ArgType | Use64BitVectors), 2893 NEONMAP1(vqrshlh_u16, aarch64_neon_uqrshl, Vectorize1ArgType | Use64BitVectors), 2894 NEONMAP1(vqrshls_s32, aarch64_neon_sqrshl, Add1ArgType), 2895 NEONMAP1(vqrshls_u32, aarch64_neon_uqrshl, Add1ArgType), 2896 NEONMAP1(vqrshrnd_n_s64, aarch64_neon_sqrshrn, AddRetType), 2897 NEONMAP1(vqrshrnd_n_u64, aarch64_neon_uqrshrn, AddRetType), 2898 NEONMAP1(vqrshrnh_n_s16, aarch64_neon_sqrshrn, VectorRet | Use64BitVectors), 2899 NEONMAP1(vqrshrnh_n_u16, aarch64_neon_uqrshrn, VectorRet | Use64BitVectors), 2900 NEONMAP1(vqrshrns_n_s32, aarch64_neon_sqrshrn, VectorRet | Use64BitVectors), 2901 NEONMAP1(vqrshrns_n_u32, aarch64_neon_uqrshrn, VectorRet | Use64BitVectors), 2902 NEONMAP1(vqrshrund_n_s64, aarch64_neon_sqrshrun, AddRetType), 2903 NEONMAP1(vqrshrunh_n_s16, aarch64_neon_sqrshrun, VectorRet | Use64BitVectors), 2904 NEONMAP1(vqrshruns_n_s32, aarch64_neon_sqrshrun, VectorRet | Use64BitVectors), 2905 NEONMAP1(vqshlb_n_s8, aarch64_neon_sqshl, Vectorize1ArgType | Use64BitVectors), 2906 NEONMAP1(vqshlb_n_u8, aarch64_neon_uqshl, Vectorize1ArgType | Use64BitVectors), 2907 NEONMAP1(vqshlb_s8, aarch64_neon_sqshl, Vectorize1ArgType | Use64BitVectors), 2908 NEONMAP1(vqshlb_u8, aarch64_neon_uqshl, Vectorize1ArgType | Use64BitVectors), 2909 NEONMAP1(vqshld_s64, aarch64_neon_sqshl, Add1ArgType), 2910 NEONMAP1(vqshld_u64, aarch64_neon_uqshl, Add1ArgType), 2911 NEONMAP1(vqshlh_n_s16, aarch64_neon_sqshl, Vectorize1ArgType | Use64BitVectors), 2912 NEONMAP1(vqshlh_n_u16, aarch64_neon_uqshl, Vectorize1ArgType | Use64BitVectors), 2913 NEONMAP1(vqshlh_s16, aarch64_neon_sqshl, Vectorize1ArgType | Use64BitVectors), 2914 NEONMAP1(vqshlh_u16, aarch64_neon_uqshl, Vectorize1ArgType | Use64BitVectors), 2915 NEONMAP1(vqshls_n_s32, aarch64_neon_sqshl, Add1ArgType), 2916 NEONMAP1(vqshls_n_u32, aarch64_neon_uqshl, Add1ArgType), 2917 NEONMAP1(vqshls_s32, aarch64_neon_sqshl, Add1ArgType), 2918 NEONMAP1(vqshls_u32, aarch64_neon_uqshl, Add1ArgType), 2919 NEONMAP1(vqshlub_n_s8, aarch64_neon_sqshlu, Vectorize1ArgType | Use64BitVectors), 2920 NEONMAP1(vqshluh_n_s16, aarch64_neon_sqshlu, Vectorize1ArgType | Use64BitVectors), 2921 NEONMAP1(vqshlus_n_s32, aarch64_neon_sqshlu, Add1ArgType), 2922 NEONMAP1(vqshrnd_n_s64, aarch64_neon_sqshrn, AddRetType), 2923 NEONMAP1(vqshrnd_n_u64, aarch64_neon_uqshrn, AddRetType), 2924 NEONMAP1(vqshrnh_n_s16, aarch64_neon_sqshrn, VectorRet | Use64BitVectors), 2925 NEONMAP1(vqshrnh_n_u16, aarch64_neon_uqshrn, VectorRet | Use64BitVectors), 2926 NEONMAP1(vqshrns_n_s32, aarch64_neon_sqshrn, VectorRet | Use64BitVectors), 2927 NEONMAP1(vqshrns_n_u32, aarch64_neon_uqshrn, VectorRet | Use64BitVectors), 2928 NEONMAP1(vqshrund_n_s64, aarch64_neon_sqshrun, AddRetType), 2929 NEONMAP1(vqshrunh_n_s16, aarch64_neon_sqshrun, VectorRet | Use64BitVectors), 2930 NEONMAP1(vqshruns_n_s32, aarch64_neon_sqshrun, VectorRet | Use64BitVectors), 2931 NEONMAP1(vqsubb_s8, aarch64_neon_sqsub, Vectorize1ArgType | Use64BitVectors), 2932 NEONMAP1(vqsubb_u8, aarch64_neon_uqsub, Vectorize1ArgType | Use64BitVectors), 2933 NEONMAP1(vqsubd_s64, aarch64_neon_sqsub, Add1ArgType), 2934 NEONMAP1(vqsubd_u64, aarch64_neon_uqsub, Add1ArgType), 2935 NEONMAP1(vqsubh_s16, aarch64_neon_sqsub, Vectorize1ArgType | Use64BitVectors), 2936 NEONMAP1(vqsubh_u16, aarch64_neon_uqsub, Vectorize1ArgType | Use64BitVectors), 2937 NEONMAP1(vqsubs_s32, aarch64_neon_sqsub, Add1ArgType), 2938 NEONMAP1(vqsubs_u32, aarch64_neon_uqsub, Add1ArgType), 2939 NEONMAP1(vrecped_f64, aarch64_neon_frecpe, Add1ArgType), 2940 NEONMAP1(vrecpes_f32, aarch64_neon_frecpe, Add1ArgType), 2941 NEONMAP1(vrecpxd_f64, aarch64_neon_frecpx, Add1ArgType), 2942 NEONMAP1(vrecpxs_f32, aarch64_neon_frecpx, Add1ArgType), 2943 NEONMAP1(vrshld_s64, aarch64_neon_srshl, Add1ArgType), 2944 NEONMAP1(vrshld_u64, aarch64_neon_urshl, Add1ArgType), 2945 NEONMAP1(vrsqrted_f64, aarch64_neon_frsqrte, Add1ArgType), 2946 NEONMAP1(vrsqrtes_f32, aarch64_neon_frsqrte, Add1ArgType), 2947 NEONMAP1(vrsqrtsd_f64, aarch64_neon_frsqrts, Add1ArgType), 2948 NEONMAP1(vrsqrtss_f32, aarch64_neon_frsqrts, Add1ArgType), 2949 NEONMAP1(vsha1cq_u32, aarch64_crypto_sha1c, 0), 2950 NEONMAP1(vsha1h_u32, aarch64_crypto_sha1h, 0), 2951 NEONMAP1(vsha1mq_u32, aarch64_crypto_sha1m, 0), 2952 NEONMAP1(vsha1pq_u32, aarch64_crypto_sha1p, 0), 2953 NEONMAP1(vshld_s64, aarch64_neon_sshl, Add1ArgType), 2954 NEONMAP1(vshld_u64, aarch64_neon_ushl, Add1ArgType), 2955 NEONMAP1(vslid_n_s64, aarch64_neon_vsli, Vectorize1ArgType), 2956 NEONMAP1(vslid_n_u64, aarch64_neon_vsli, Vectorize1ArgType), 2957 NEONMAP1(vsqaddb_u8, aarch64_neon_usqadd, Vectorize1ArgType | Use64BitVectors), 2958 NEONMAP1(vsqaddd_u64, aarch64_neon_usqadd, Add1ArgType), 2959 NEONMAP1(vsqaddh_u16, aarch64_neon_usqadd, Vectorize1ArgType | Use64BitVectors), 2960 NEONMAP1(vsqadds_u32, aarch64_neon_usqadd, Add1ArgType), 2961 NEONMAP1(vsrid_n_s64, aarch64_neon_vsri, Vectorize1ArgType), 2962 NEONMAP1(vsrid_n_u64, aarch64_neon_vsri, Vectorize1ArgType), 2963 NEONMAP1(vuqaddb_s8, aarch64_neon_suqadd, Vectorize1ArgType | Use64BitVectors), 2964 NEONMAP1(vuqaddd_s64, aarch64_neon_suqadd, Add1ArgType), 2965 NEONMAP1(vuqaddh_s16, aarch64_neon_suqadd, Vectorize1ArgType | Use64BitVectors), 2966 NEONMAP1(vuqadds_s32, aarch64_neon_suqadd, Add1ArgType), 2967 }; 2968 2969 #undef NEONMAP0 2970 #undef NEONMAP1 2971 #undef NEONMAP2 2972 2973 static bool NEONSIMDIntrinsicsProvenSorted = false; 2974 2975 static bool AArch64SIMDIntrinsicsProvenSorted = false; 2976 static bool AArch64SISDIntrinsicsProvenSorted = false; 2977 2978 2979 static const NeonIntrinsicInfo * 2980 findNeonIntrinsicInMap(ArrayRef<NeonIntrinsicInfo> IntrinsicMap, 2981 unsigned BuiltinID, bool &MapProvenSorted) { 2982 2983 #ifndef NDEBUG 2984 if (!MapProvenSorted) { 2985 assert(std::is_sorted(std::begin(IntrinsicMap), std::end(IntrinsicMap))); 2986 MapProvenSorted = true; 2987 } 2988 #endif 2989 2990 const NeonIntrinsicInfo *Builtin = 2991 std::lower_bound(IntrinsicMap.begin(), IntrinsicMap.end(), BuiltinID); 2992 2993 if (Builtin != IntrinsicMap.end() && Builtin->BuiltinID == BuiltinID) 2994 return Builtin; 2995 2996 return nullptr; 2997 } 2998 2999 Function *CodeGenFunction::LookupNeonLLVMIntrinsic(unsigned IntrinsicID, 3000 unsigned Modifier, 3001 llvm::Type *ArgType, 3002 const CallExpr *E) { 3003 int VectorSize = 0; 3004 if (Modifier & Use64BitVectors) 3005 VectorSize = 64; 3006 else if (Modifier & Use128BitVectors) 3007 VectorSize = 128; 3008 3009 // Return type. 3010 SmallVector<llvm::Type *, 3> Tys; 3011 if (Modifier & AddRetType) { 3012 llvm::Type *Ty = ConvertType(E->getCallReturnType(getContext())); 3013 if (Modifier & VectorizeRetType) 3014 Ty = llvm::VectorType::get( 3015 Ty, VectorSize ? VectorSize / Ty->getPrimitiveSizeInBits() : 1); 3016 3017 Tys.push_back(Ty); 3018 } 3019 3020 // Arguments. 3021 if (Modifier & VectorizeArgTypes) { 3022 int Elts = VectorSize ? VectorSize / ArgType->getPrimitiveSizeInBits() : 1; 3023 ArgType = llvm::VectorType::get(ArgType, Elts); 3024 } 3025 3026 if (Modifier & (Add1ArgType | Add2ArgTypes)) 3027 Tys.push_back(ArgType); 3028 3029 if (Modifier & Add2ArgTypes) 3030 Tys.push_back(ArgType); 3031 3032 if (Modifier & InventFloatType) 3033 Tys.push_back(FloatTy); 3034 3035 return CGM.getIntrinsic(IntrinsicID, Tys); 3036 } 3037 3038 static Value *EmitCommonNeonSISDBuiltinExpr(CodeGenFunction &CGF, 3039 const NeonIntrinsicInfo &SISDInfo, 3040 SmallVectorImpl<Value *> &Ops, 3041 const CallExpr *E) { 3042 unsigned BuiltinID = SISDInfo.BuiltinID; 3043 unsigned int Int = SISDInfo.LLVMIntrinsic; 3044 unsigned Modifier = SISDInfo.TypeModifier; 3045 const char *s = SISDInfo.NameHint; 3046 3047 switch (BuiltinID) { 3048 case NEON::BI__builtin_neon_vcled_s64: 3049 case NEON::BI__builtin_neon_vcled_u64: 3050 case NEON::BI__builtin_neon_vcles_f32: 3051 case NEON::BI__builtin_neon_vcled_f64: 3052 case NEON::BI__builtin_neon_vcltd_s64: 3053 case NEON::BI__builtin_neon_vcltd_u64: 3054 case NEON::BI__builtin_neon_vclts_f32: 3055 case NEON::BI__builtin_neon_vcltd_f64: 3056 case NEON::BI__builtin_neon_vcales_f32: 3057 case NEON::BI__builtin_neon_vcaled_f64: 3058 case NEON::BI__builtin_neon_vcalts_f32: 3059 case NEON::BI__builtin_neon_vcaltd_f64: 3060 // Only one direction of comparisons actually exist, cmle is actually a cmge 3061 // with swapped operands. The table gives us the right intrinsic but we 3062 // still need to do the swap. 3063 std::swap(Ops[0], Ops[1]); 3064 break; 3065 } 3066 3067 assert(Int && "Generic code assumes a valid intrinsic"); 3068 3069 // Determine the type(s) of this overloaded AArch64 intrinsic. 3070 const Expr *Arg = E->getArg(0); 3071 llvm::Type *ArgTy = CGF.ConvertType(Arg->getType()); 3072 Function *F = CGF.LookupNeonLLVMIntrinsic(Int, Modifier, ArgTy, E); 3073 3074 int j = 0; 3075 ConstantInt *C0 = ConstantInt::get(CGF.SizeTy, 0); 3076 for (Function::const_arg_iterator ai = F->arg_begin(), ae = F->arg_end(); 3077 ai != ae; ++ai, ++j) { 3078 llvm::Type *ArgTy = ai->getType(); 3079 if (Ops[j]->getType()->getPrimitiveSizeInBits() == 3080 ArgTy->getPrimitiveSizeInBits()) 3081 continue; 3082 3083 assert(ArgTy->isVectorTy() && !Ops[j]->getType()->isVectorTy()); 3084 // The constant argument to an _n_ intrinsic always has Int32Ty, so truncate 3085 // it before inserting. 3086 Ops[j] = 3087 CGF.Builder.CreateTruncOrBitCast(Ops[j], ArgTy->getVectorElementType()); 3088 Ops[j] = 3089 CGF.Builder.CreateInsertElement(UndefValue::get(ArgTy), Ops[j], C0); 3090 } 3091 3092 Value *Result = CGF.EmitNeonCall(F, Ops, s); 3093 llvm::Type *ResultType = CGF.ConvertType(E->getType()); 3094 if (ResultType->getPrimitiveSizeInBits() < 3095 Result->getType()->getPrimitiveSizeInBits()) 3096 return CGF.Builder.CreateExtractElement(Result, C0); 3097 3098 return CGF.Builder.CreateBitCast(Result, ResultType, s); 3099 } 3100 3101 Value *CodeGenFunction::EmitCommonNeonBuiltinExpr( 3102 unsigned BuiltinID, unsigned LLVMIntrinsic, unsigned AltLLVMIntrinsic, 3103 const char *NameHint, unsigned Modifier, const CallExpr *E, 3104 SmallVectorImpl<llvm::Value *> &Ops, Address PtrOp0, Address PtrOp1) { 3105 // Get the last argument, which specifies the vector type. 3106 llvm::APSInt NeonTypeConst; 3107 const Expr *Arg = E->getArg(E->getNumArgs() - 1); 3108 if (!Arg->isIntegerConstantExpr(NeonTypeConst, getContext())) 3109 return nullptr; 3110 3111 // Determine the type of this overloaded NEON intrinsic. 3112 NeonTypeFlags Type(NeonTypeConst.getZExtValue()); 3113 bool Usgn = Type.isUnsigned(); 3114 bool Quad = Type.isQuad(); 3115 3116 llvm::VectorType *VTy = GetNeonType(this, Type); 3117 llvm::Type *Ty = VTy; 3118 if (!Ty) 3119 return nullptr; 3120 3121 auto getAlignmentValue32 = [&](Address addr) -> Value* { 3122 return Builder.getInt32(addr.getAlignment().getQuantity()); 3123 }; 3124 3125 unsigned Int = LLVMIntrinsic; 3126 if ((Modifier & UnsignedAlts) && !Usgn) 3127 Int = AltLLVMIntrinsic; 3128 3129 switch (BuiltinID) { 3130 default: break; 3131 case NEON::BI__builtin_neon_vabs_v: 3132 case NEON::BI__builtin_neon_vabsq_v: 3133 if (VTy->getElementType()->isFloatingPointTy()) 3134 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::fabs, Ty), Ops, "vabs"); 3135 return EmitNeonCall(CGM.getIntrinsic(LLVMIntrinsic, Ty), Ops, "vabs"); 3136 case NEON::BI__builtin_neon_vaddhn_v: { 3137 llvm::VectorType *SrcTy = 3138 llvm::VectorType::getExtendedElementVectorType(VTy); 3139 3140 // %sum = add <4 x i32> %lhs, %rhs 3141 Ops[0] = Builder.CreateBitCast(Ops[0], SrcTy); 3142 Ops[1] = Builder.CreateBitCast(Ops[1], SrcTy); 3143 Ops[0] = Builder.CreateAdd(Ops[0], Ops[1], "vaddhn"); 3144 3145 // %high = lshr <4 x i32> %sum, <i32 16, i32 16, i32 16, i32 16> 3146 Constant *ShiftAmt = 3147 ConstantInt::get(SrcTy, SrcTy->getScalarSizeInBits() / 2); 3148 Ops[0] = Builder.CreateLShr(Ops[0], ShiftAmt, "vaddhn"); 3149 3150 // %res = trunc <4 x i32> %high to <4 x i16> 3151 return Builder.CreateTrunc(Ops[0], VTy, "vaddhn"); 3152 } 3153 case NEON::BI__builtin_neon_vcale_v: 3154 case NEON::BI__builtin_neon_vcaleq_v: 3155 case NEON::BI__builtin_neon_vcalt_v: 3156 case NEON::BI__builtin_neon_vcaltq_v: 3157 std::swap(Ops[0], Ops[1]); 3158 case NEON::BI__builtin_neon_vcage_v: 3159 case NEON::BI__builtin_neon_vcageq_v: 3160 case NEON::BI__builtin_neon_vcagt_v: 3161 case NEON::BI__builtin_neon_vcagtq_v: { 3162 llvm::Type *VecFlt = llvm::VectorType::get( 3163 VTy->getScalarSizeInBits() == 32 ? FloatTy : DoubleTy, 3164 VTy->getNumElements()); 3165 llvm::Type *Tys[] = { VTy, VecFlt }; 3166 Function *F = CGM.getIntrinsic(LLVMIntrinsic, Tys); 3167 return EmitNeonCall(F, Ops, NameHint); 3168 } 3169 case NEON::BI__builtin_neon_vclz_v: 3170 case NEON::BI__builtin_neon_vclzq_v: 3171 // We generate target-independent intrinsic, which needs a second argument 3172 // for whether or not clz of zero is undefined; on ARM it isn't. 3173 Ops.push_back(Builder.getInt1(getTarget().isCLZForZeroUndef())); 3174 break; 3175 case NEON::BI__builtin_neon_vcvt_f32_v: 3176 case NEON::BI__builtin_neon_vcvtq_f32_v: 3177 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 3178 Ty = GetNeonType(this, NeonTypeFlags(NeonTypeFlags::Float32, false, Quad)); 3179 return Usgn ? Builder.CreateUIToFP(Ops[0], Ty, "vcvt") 3180 : Builder.CreateSIToFP(Ops[0], Ty, "vcvt"); 3181 case NEON::BI__builtin_neon_vcvt_n_f32_v: 3182 case NEON::BI__builtin_neon_vcvt_n_f64_v: 3183 case NEON::BI__builtin_neon_vcvtq_n_f32_v: 3184 case NEON::BI__builtin_neon_vcvtq_n_f64_v: { 3185 llvm::Type *Tys[2] = { GetFloatNeonType(this, Type), Ty }; 3186 Int = Usgn ? LLVMIntrinsic : AltLLVMIntrinsic; 3187 Function *F = CGM.getIntrinsic(Int, Tys); 3188 return EmitNeonCall(F, Ops, "vcvt_n"); 3189 } 3190 case NEON::BI__builtin_neon_vcvt_n_s32_v: 3191 case NEON::BI__builtin_neon_vcvt_n_u32_v: 3192 case NEON::BI__builtin_neon_vcvt_n_s64_v: 3193 case NEON::BI__builtin_neon_vcvt_n_u64_v: 3194 case NEON::BI__builtin_neon_vcvtq_n_s32_v: 3195 case NEON::BI__builtin_neon_vcvtq_n_u32_v: 3196 case NEON::BI__builtin_neon_vcvtq_n_s64_v: 3197 case NEON::BI__builtin_neon_vcvtq_n_u64_v: { 3198 llvm::Type *Tys[2] = { Ty, GetFloatNeonType(this, Type) }; 3199 Function *F = CGM.getIntrinsic(LLVMIntrinsic, Tys); 3200 return EmitNeonCall(F, Ops, "vcvt_n"); 3201 } 3202 case NEON::BI__builtin_neon_vcvt_s32_v: 3203 case NEON::BI__builtin_neon_vcvt_u32_v: 3204 case NEON::BI__builtin_neon_vcvt_s64_v: 3205 case NEON::BI__builtin_neon_vcvt_u64_v: 3206 case NEON::BI__builtin_neon_vcvtq_s32_v: 3207 case NEON::BI__builtin_neon_vcvtq_u32_v: 3208 case NEON::BI__builtin_neon_vcvtq_s64_v: 3209 case NEON::BI__builtin_neon_vcvtq_u64_v: { 3210 Ops[0] = Builder.CreateBitCast(Ops[0], GetFloatNeonType(this, Type)); 3211 return Usgn ? Builder.CreateFPToUI(Ops[0], Ty, "vcvt") 3212 : Builder.CreateFPToSI(Ops[0], Ty, "vcvt"); 3213 } 3214 case NEON::BI__builtin_neon_vcvta_s32_v: 3215 case NEON::BI__builtin_neon_vcvta_s64_v: 3216 case NEON::BI__builtin_neon_vcvta_u32_v: 3217 case NEON::BI__builtin_neon_vcvta_u64_v: 3218 case NEON::BI__builtin_neon_vcvtaq_s32_v: 3219 case NEON::BI__builtin_neon_vcvtaq_s64_v: 3220 case NEON::BI__builtin_neon_vcvtaq_u32_v: 3221 case NEON::BI__builtin_neon_vcvtaq_u64_v: 3222 case NEON::BI__builtin_neon_vcvtn_s32_v: 3223 case NEON::BI__builtin_neon_vcvtn_s64_v: 3224 case NEON::BI__builtin_neon_vcvtn_u32_v: 3225 case NEON::BI__builtin_neon_vcvtn_u64_v: 3226 case NEON::BI__builtin_neon_vcvtnq_s32_v: 3227 case NEON::BI__builtin_neon_vcvtnq_s64_v: 3228 case NEON::BI__builtin_neon_vcvtnq_u32_v: 3229 case NEON::BI__builtin_neon_vcvtnq_u64_v: 3230 case NEON::BI__builtin_neon_vcvtp_s32_v: 3231 case NEON::BI__builtin_neon_vcvtp_s64_v: 3232 case NEON::BI__builtin_neon_vcvtp_u32_v: 3233 case NEON::BI__builtin_neon_vcvtp_u64_v: 3234 case NEON::BI__builtin_neon_vcvtpq_s32_v: 3235 case NEON::BI__builtin_neon_vcvtpq_s64_v: 3236 case NEON::BI__builtin_neon_vcvtpq_u32_v: 3237 case NEON::BI__builtin_neon_vcvtpq_u64_v: 3238 case NEON::BI__builtin_neon_vcvtm_s32_v: 3239 case NEON::BI__builtin_neon_vcvtm_s64_v: 3240 case NEON::BI__builtin_neon_vcvtm_u32_v: 3241 case NEON::BI__builtin_neon_vcvtm_u64_v: 3242 case NEON::BI__builtin_neon_vcvtmq_s32_v: 3243 case NEON::BI__builtin_neon_vcvtmq_s64_v: 3244 case NEON::BI__builtin_neon_vcvtmq_u32_v: 3245 case NEON::BI__builtin_neon_vcvtmq_u64_v: { 3246 llvm::Type *Tys[2] = { Ty, GetFloatNeonType(this, Type) }; 3247 return EmitNeonCall(CGM.getIntrinsic(LLVMIntrinsic, Tys), Ops, NameHint); 3248 } 3249 case NEON::BI__builtin_neon_vext_v: 3250 case NEON::BI__builtin_neon_vextq_v: { 3251 int CV = cast<ConstantInt>(Ops[2])->getSExtValue(); 3252 SmallVector<Constant*, 16> Indices; 3253 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) 3254 Indices.push_back(ConstantInt::get(Int32Ty, i+CV)); 3255 3256 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 3257 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 3258 Value *SV = llvm::ConstantVector::get(Indices); 3259 return Builder.CreateShuffleVector(Ops[0], Ops[1], SV, "vext"); 3260 } 3261 case NEON::BI__builtin_neon_vfma_v: 3262 case NEON::BI__builtin_neon_vfmaq_v: { 3263 Value *F = CGM.getIntrinsic(Intrinsic::fma, Ty); 3264 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 3265 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 3266 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 3267 3268 // NEON intrinsic puts accumulator first, unlike the LLVM fma. 3269 return Builder.CreateCall(F, {Ops[1], Ops[2], Ops[0]}); 3270 } 3271 case NEON::BI__builtin_neon_vld1_v: 3272 case NEON::BI__builtin_neon_vld1q_v: { 3273 llvm::Type *Tys[] = {Ty, Int8PtrTy}; 3274 Ops.push_back(getAlignmentValue32(PtrOp0)); 3275 return EmitNeonCall(CGM.getIntrinsic(LLVMIntrinsic, Tys), Ops, "vld1"); 3276 } 3277 case NEON::BI__builtin_neon_vld2_v: 3278 case NEON::BI__builtin_neon_vld2q_v: 3279 case NEON::BI__builtin_neon_vld3_v: 3280 case NEON::BI__builtin_neon_vld3q_v: 3281 case NEON::BI__builtin_neon_vld4_v: 3282 case NEON::BI__builtin_neon_vld4q_v: { 3283 llvm::Type *Tys[] = {Ty, Int8PtrTy}; 3284 Function *F = CGM.getIntrinsic(LLVMIntrinsic, Tys); 3285 Value *Align = getAlignmentValue32(PtrOp1); 3286 Ops[1] = Builder.CreateCall(F, {Ops[1], Align}, NameHint); 3287 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 3288 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 3289 return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]); 3290 } 3291 case NEON::BI__builtin_neon_vld1_dup_v: 3292 case NEON::BI__builtin_neon_vld1q_dup_v: { 3293 Value *V = UndefValue::get(Ty); 3294 Ty = llvm::PointerType::getUnqual(VTy->getElementType()); 3295 PtrOp0 = Builder.CreateBitCast(PtrOp0, Ty); 3296 LoadInst *Ld = Builder.CreateLoad(PtrOp0); 3297 llvm::Constant *CI = ConstantInt::get(SizeTy, 0); 3298 Ops[0] = Builder.CreateInsertElement(V, Ld, CI); 3299 return EmitNeonSplat(Ops[0], CI); 3300 } 3301 case NEON::BI__builtin_neon_vld2_lane_v: 3302 case NEON::BI__builtin_neon_vld2q_lane_v: 3303 case NEON::BI__builtin_neon_vld3_lane_v: 3304 case NEON::BI__builtin_neon_vld3q_lane_v: 3305 case NEON::BI__builtin_neon_vld4_lane_v: 3306 case NEON::BI__builtin_neon_vld4q_lane_v: { 3307 llvm::Type *Tys[] = {Ty, Int8PtrTy}; 3308 Function *F = CGM.getIntrinsic(LLVMIntrinsic, Tys); 3309 for (unsigned I = 2; I < Ops.size() - 1; ++I) 3310 Ops[I] = Builder.CreateBitCast(Ops[I], Ty); 3311 Ops.push_back(getAlignmentValue32(PtrOp1)); 3312 Ops[1] = Builder.CreateCall(F, makeArrayRef(Ops).slice(1), NameHint); 3313 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 3314 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 3315 return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]); 3316 } 3317 case NEON::BI__builtin_neon_vmovl_v: { 3318 llvm::Type *DTy =llvm::VectorType::getTruncatedElementVectorType(VTy); 3319 Ops[0] = Builder.CreateBitCast(Ops[0], DTy); 3320 if (Usgn) 3321 return Builder.CreateZExt(Ops[0], Ty, "vmovl"); 3322 return Builder.CreateSExt(Ops[0], Ty, "vmovl"); 3323 } 3324 case NEON::BI__builtin_neon_vmovn_v: { 3325 llvm::Type *QTy = llvm::VectorType::getExtendedElementVectorType(VTy); 3326 Ops[0] = Builder.CreateBitCast(Ops[0], QTy); 3327 return Builder.CreateTrunc(Ops[0], Ty, "vmovn"); 3328 } 3329 case NEON::BI__builtin_neon_vmull_v: 3330 // FIXME: the integer vmull operations could be emitted in terms of pure 3331 // LLVM IR (2 exts followed by a mul). Unfortunately LLVM has a habit of 3332 // hoisting the exts outside loops. Until global ISel comes along that can 3333 // see through such movement this leads to bad CodeGen. So we need an 3334 // intrinsic for now. 3335 Int = Usgn ? Intrinsic::arm_neon_vmullu : Intrinsic::arm_neon_vmulls; 3336 Int = Type.isPoly() ? (unsigned)Intrinsic::arm_neon_vmullp : Int; 3337 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmull"); 3338 case NEON::BI__builtin_neon_vpadal_v: 3339 case NEON::BI__builtin_neon_vpadalq_v: { 3340 // The source operand type has twice as many elements of half the size. 3341 unsigned EltBits = VTy->getElementType()->getPrimitiveSizeInBits(); 3342 llvm::Type *EltTy = 3343 llvm::IntegerType::get(getLLVMContext(), EltBits / 2); 3344 llvm::Type *NarrowTy = 3345 llvm::VectorType::get(EltTy, VTy->getNumElements() * 2); 3346 llvm::Type *Tys[2] = { Ty, NarrowTy }; 3347 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, NameHint); 3348 } 3349 case NEON::BI__builtin_neon_vpaddl_v: 3350 case NEON::BI__builtin_neon_vpaddlq_v: { 3351 // The source operand type has twice as many elements of half the size. 3352 unsigned EltBits = VTy->getElementType()->getPrimitiveSizeInBits(); 3353 llvm::Type *EltTy = llvm::IntegerType::get(getLLVMContext(), EltBits / 2); 3354 llvm::Type *NarrowTy = 3355 llvm::VectorType::get(EltTy, VTy->getNumElements() * 2); 3356 llvm::Type *Tys[2] = { Ty, NarrowTy }; 3357 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vpaddl"); 3358 } 3359 case NEON::BI__builtin_neon_vqdmlal_v: 3360 case NEON::BI__builtin_neon_vqdmlsl_v: { 3361 SmallVector<Value *, 2> MulOps(Ops.begin() + 1, Ops.end()); 3362 Ops[1] = 3363 EmitNeonCall(CGM.getIntrinsic(LLVMIntrinsic, Ty), MulOps, "vqdmlal"); 3364 Ops.resize(2); 3365 return EmitNeonCall(CGM.getIntrinsic(AltLLVMIntrinsic, Ty), Ops, NameHint); 3366 } 3367 case NEON::BI__builtin_neon_vqshl_n_v: 3368 case NEON::BI__builtin_neon_vqshlq_n_v: 3369 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqshl_n", 3370 1, false); 3371 case NEON::BI__builtin_neon_vqshlu_n_v: 3372 case NEON::BI__builtin_neon_vqshluq_n_v: 3373 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqshlu_n", 3374 1, false); 3375 case NEON::BI__builtin_neon_vrecpe_v: 3376 case NEON::BI__builtin_neon_vrecpeq_v: 3377 case NEON::BI__builtin_neon_vrsqrte_v: 3378 case NEON::BI__builtin_neon_vrsqrteq_v: 3379 Int = Ty->isFPOrFPVectorTy() ? LLVMIntrinsic : AltLLVMIntrinsic; 3380 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, NameHint); 3381 3382 case NEON::BI__builtin_neon_vrshr_n_v: 3383 case NEON::BI__builtin_neon_vrshrq_n_v: 3384 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrshr_n", 3385 1, true); 3386 case NEON::BI__builtin_neon_vshl_n_v: 3387 case NEON::BI__builtin_neon_vshlq_n_v: 3388 Ops[1] = EmitNeonShiftVector(Ops[1], Ty, false); 3389 return Builder.CreateShl(Builder.CreateBitCast(Ops[0],Ty), Ops[1], 3390 "vshl_n"); 3391 case NEON::BI__builtin_neon_vshll_n_v: { 3392 llvm::Type *SrcTy = llvm::VectorType::getTruncatedElementVectorType(VTy); 3393 Ops[0] = Builder.CreateBitCast(Ops[0], SrcTy); 3394 if (Usgn) 3395 Ops[0] = Builder.CreateZExt(Ops[0], VTy); 3396 else 3397 Ops[0] = Builder.CreateSExt(Ops[0], VTy); 3398 Ops[1] = EmitNeonShiftVector(Ops[1], VTy, false); 3399 return Builder.CreateShl(Ops[0], Ops[1], "vshll_n"); 3400 } 3401 case NEON::BI__builtin_neon_vshrn_n_v: { 3402 llvm::Type *SrcTy = llvm::VectorType::getExtendedElementVectorType(VTy); 3403 Ops[0] = Builder.CreateBitCast(Ops[0], SrcTy); 3404 Ops[1] = EmitNeonShiftVector(Ops[1], SrcTy, false); 3405 if (Usgn) 3406 Ops[0] = Builder.CreateLShr(Ops[0], Ops[1]); 3407 else 3408 Ops[0] = Builder.CreateAShr(Ops[0], Ops[1]); 3409 return Builder.CreateTrunc(Ops[0], Ty, "vshrn_n"); 3410 } 3411 case NEON::BI__builtin_neon_vshr_n_v: 3412 case NEON::BI__builtin_neon_vshrq_n_v: 3413 return EmitNeonRShiftImm(Ops[0], Ops[1], Ty, Usgn, "vshr_n"); 3414 case NEON::BI__builtin_neon_vst1_v: 3415 case NEON::BI__builtin_neon_vst1q_v: 3416 case NEON::BI__builtin_neon_vst2_v: 3417 case NEON::BI__builtin_neon_vst2q_v: 3418 case NEON::BI__builtin_neon_vst3_v: 3419 case NEON::BI__builtin_neon_vst3q_v: 3420 case NEON::BI__builtin_neon_vst4_v: 3421 case NEON::BI__builtin_neon_vst4q_v: 3422 case NEON::BI__builtin_neon_vst2_lane_v: 3423 case NEON::BI__builtin_neon_vst2q_lane_v: 3424 case NEON::BI__builtin_neon_vst3_lane_v: 3425 case NEON::BI__builtin_neon_vst3q_lane_v: 3426 case NEON::BI__builtin_neon_vst4_lane_v: 3427 case NEON::BI__builtin_neon_vst4q_lane_v: { 3428 llvm::Type *Tys[] = {Int8PtrTy, Ty}; 3429 Ops.push_back(getAlignmentValue32(PtrOp0)); 3430 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, ""); 3431 } 3432 case NEON::BI__builtin_neon_vsubhn_v: { 3433 llvm::VectorType *SrcTy = 3434 llvm::VectorType::getExtendedElementVectorType(VTy); 3435 3436 // %sum = add <4 x i32> %lhs, %rhs 3437 Ops[0] = Builder.CreateBitCast(Ops[0], SrcTy); 3438 Ops[1] = Builder.CreateBitCast(Ops[1], SrcTy); 3439 Ops[0] = Builder.CreateSub(Ops[0], Ops[1], "vsubhn"); 3440 3441 // %high = lshr <4 x i32> %sum, <i32 16, i32 16, i32 16, i32 16> 3442 Constant *ShiftAmt = 3443 ConstantInt::get(SrcTy, SrcTy->getScalarSizeInBits() / 2); 3444 Ops[0] = Builder.CreateLShr(Ops[0], ShiftAmt, "vsubhn"); 3445 3446 // %res = trunc <4 x i32> %high to <4 x i16> 3447 return Builder.CreateTrunc(Ops[0], VTy, "vsubhn"); 3448 } 3449 case NEON::BI__builtin_neon_vtrn_v: 3450 case NEON::BI__builtin_neon_vtrnq_v: { 3451 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty)); 3452 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 3453 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 3454 Value *SV = nullptr; 3455 3456 for (unsigned vi = 0; vi != 2; ++vi) { 3457 SmallVector<Constant*, 16> Indices; 3458 for (unsigned i = 0, e = VTy->getNumElements(); i != e; i += 2) { 3459 Indices.push_back(Builder.getInt32(i+vi)); 3460 Indices.push_back(Builder.getInt32(i+e+vi)); 3461 } 3462 Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ty, Ops[0], vi); 3463 SV = llvm::ConstantVector::get(Indices); 3464 SV = Builder.CreateShuffleVector(Ops[1], Ops[2], SV, "vtrn"); 3465 SV = Builder.CreateDefaultAlignedStore(SV, Addr); 3466 } 3467 return SV; 3468 } 3469 case NEON::BI__builtin_neon_vtst_v: 3470 case NEON::BI__builtin_neon_vtstq_v: { 3471 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 3472 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 3473 Ops[0] = Builder.CreateAnd(Ops[0], Ops[1]); 3474 Ops[0] = Builder.CreateICmp(ICmpInst::ICMP_NE, Ops[0], 3475 ConstantAggregateZero::get(Ty)); 3476 return Builder.CreateSExt(Ops[0], Ty, "vtst"); 3477 } 3478 case NEON::BI__builtin_neon_vuzp_v: 3479 case NEON::BI__builtin_neon_vuzpq_v: { 3480 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty)); 3481 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 3482 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 3483 Value *SV = nullptr; 3484 3485 for (unsigned vi = 0; vi != 2; ++vi) { 3486 SmallVector<Constant*, 16> Indices; 3487 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) 3488 Indices.push_back(ConstantInt::get(Int32Ty, 2*i+vi)); 3489 3490 Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ty, Ops[0], vi); 3491 SV = llvm::ConstantVector::get(Indices); 3492 SV = Builder.CreateShuffleVector(Ops[1], Ops[2], SV, "vuzp"); 3493 SV = Builder.CreateDefaultAlignedStore(SV, Addr); 3494 } 3495 return SV; 3496 } 3497 case NEON::BI__builtin_neon_vzip_v: 3498 case NEON::BI__builtin_neon_vzipq_v: { 3499 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty)); 3500 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 3501 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 3502 Value *SV = nullptr; 3503 3504 for (unsigned vi = 0; vi != 2; ++vi) { 3505 SmallVector<Constant*, 16> Indices; 3506 for (unsigned i = 0, e = VTy->getNumElements(); i != e; i += 2) { 3507 Indices.push_back(ConstantInt::get(Int32Ty, (i + vi*e) >> 1)); 3508 Indices.push_back(ConstantInt::get(Int32Ty, ((i + vi*e) >> 1)+e)); 3509 } 3510 Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ty, Ops[0], vi); 3511 SV = llvm::ConstantVector::get(Indices); 3512 SV = Builder.CreateShuffleVector(Ops[1], Ops[2], SV, "vzip"); 3513 SV = Builder.CreateDefaultAlignedStore(SV, Addr); 3514 } 3515 return SV; 3516 } 3517 } 3518 3519 assert(Int && "Expected valid intrinsic number"); 3520 3521 // Determine the type(s) of this overloaded AArch64 intrinsic. 3522 Function *F = LookupNeonLLVMIntrinsic(Int, Modifier, Ty, E); 3523 3524 Value *Result = EmitNeonCall(F, Ops, NameHint); 3525 llvm::Type *ResultType = ConvertType(E->getType()); 3526 // AArch64 intrinsic one-element vector type cast to 3527 // scalar type expected by the builtin 3528 return Builder.CreateBitCast(Result, ResultType, NameHint); 3529 } 3530 3531 Value *CodeGenFunction::EmitAArch64CompareBuiltinExpr( 3532 Value *Op, llvm::Type *Ty, const CmpInst::Predicate Fp, 3533 const CmpInst::Predicate Ip, const Twine &Name) { 3534 llvm::Type *OTy = Op->getType(); 3535 3536 // FIXME: this is utterly horrific. We should not be looking at previous 3537 // codegen context to find out what needs doing. Unfortunately TableGen 3538 // currently gives us exactly the same calls for vceqz_f32 and vceqz_s32 3539 // (etc). 3540 if (BitCastInst *BI = dyn_cast<BitCastInst>(Op)) 3541 OTy = BI->getOperand(0)->getType(); 3542 3543 Op = Builder.CreateBitCast(Op, OTy); 3544 if (OTy->getScalarType()->isFloatingPointTy()) { 3545 Op = Builder.CreateFCmp(Fp, Op, Constant::getNullValue(OTy)); 3546 } else { 3547 Op = Builder.CreateICmp(Ip, Op, Constant::getNullValue(OTy)); 3548 } 3549 return Builder.CreateSExt(Op, Ty, Name); 3550 } 3551 3552 static Value *packTBLDVectorList(CodeGenFunction &CGF, ArrayRef<Value *> Ops, 3553 Value *ExtOp, Value *IndexOp, 3554 llvm::Type *ResTy, unsigned IntID, 3555 const char *Name) { 3556 SmallVector<Value *, 2> TblOps; 3557 if (ExtOp) 3558 TblOps.push_back(ExtOp); 3559 3560 // Build a vector containing sequential number like (0, 1, 2, ..., 15) 3561 SmallVector<Constant*, 16> Indices; 3562 llvm::VectorType *TblTy = cast<llvm::VectorType>(Ops[0]->getType()); 3563 for (unsigned i = 0, e = TblTy->getNumElements(); i != e; ++i) { 3564 Indices.push_back(ConstantInt::get(CGF.Int32Ty, 2*i)); 3565 Indices.push_back(ConstantInt::get(CGF.Int32Ty, 2*i+1)); 3566 } 3567 Value *SV = llvm::ConstantVector::get(Indices); 3568 3569 int PairPos = 0, End = Ops.size() - 1; 3570 while (PairPos < End) { 3571 TblOps.push_back(CGF.Builder.CreateShuffleVector(Ops[PairPos], 3572 Ops[PairPos+1], SV, Name)); 3573 PairPos += 2; 3574 } 3575 3576 // If there's an odd number of 64-bit lookup table, fill the high 64-bit 3577 // of the 128-bit lookup table with zero. 3578 if (PairPos == End) { 3579 Value *ZeroTbl = ConstantAggregateZero::get(TblTy); 3580 TblOps.push_back(CGF.Builder.CreateShuffleVector(Ops[PairPos], 3581 ZeroTbl, SV, Name)); 3582 } 3583 3584 Function *TblF; 3585 TblOps.push_back(IndexOp); 3586 TblF = CGF.CGM.getIntrinsic(IntID, ResTy); 3587 3588 return CGF.EmitNeonCall(TblF, TblOps, Name); 3589 } 3590 3591 Value *CodeGenFunction::GetValueForARMHint(unsigned BuiltinID) { 3592 unsigned Value; 3593 switch (BuiltinID) { 3594 default: 3595 return nullptr; 3596 case ARM::BI__builtin_arm_nop: 3597 Value = 0; 3598 break; 3599 case ARM::BI__builtin_arm_yield: 3600 case ARM::BI__yield: 3601 Value = 1; 3602 break; 3603 case ARM::BI__builtin_arm_wfe: 3604 case ARM::BI__wfe: 3605 Value = 2; 3606 break; 3607 case ARM::BI__builtin_arm_wfi: 3608 case ARM::BI__wfi: 3609 Value = 3; 3610 break; 3611 case ARM::BI__builtin_arm_sev: 3612 case ARM::BI__sev: 3613 Value = 4; 3614 break; 3615 case ARM::BI__builtin_arm_sevl: 3616 case ARM::BI__sevl: 3617 Value = 5; 3618 break; 3619 } 3620 3621 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::arm_hint), 3622 llvm::ConstantInt::get(Int32Ty, Value)); 3623 } 3624 3625 // Generates the IR for the read/write special register builtin, 3626 // ValueType is the type of the value that is to be written or read, 3627 // RegisterType is the type of the register being written to or read from. 3628 static Value *EmitSpecialRegisterBuiltin(CodeGenFunction &CGF, 3629 const CallExpr *E, 3630 llvm::Type *RegisterType, 3631 llvm::Type *ValueType, bool IsRead) { 3632 // write and register intrinsics only support 32 and 64 bit operations. 3633 assert((RegisterType->isIntegerTy(32) || RegisterType->isIntegerTy(64)) 3634 && "Unsupported size for register."); 3635 3636 CodeGen::CGBuilderTy &Builder = CGF.Builder; 3637 CodeGen::CodeGenModule &CGM = CGF.CGM; 3638 LLVMContext &Context = CGM.getLLVMContext(); 3639 3640 const Expr *SysRegStrExpr = E->getArg(0)->IgnoreParenCasts(); 3641 StringRef SysReg = cast<StringLiteral>(SysRegStrExpr)->getString(); 3642 3643 llvm::Metadata *Ops[] = { llvm::MDString::get(Context, SysReg) }; 3644 llvm::MDNode *RegName = llvm::MDNode::get(Context, Ops); 3645 llvm::Value *Metadata = llvm::MetadataAsValue::get(Context, RegName); 3646 3647 llvm::Type *Types[] = { RegisterType }; 3648 3649 bool MixedTypes = RegisterType->isIntegerTy(64) && ValueType->isIntegerTy(32); 3650 assert(!(RegisterType->isIntegerTy(32) && ValueType->isIntegerTy(64)) 3651 && "Can't fit 64-bit value in 32-bit register"); 3652 3653 if (IsRead) { 3654 llvm::Value *F = CGM.getIntrinsic(llvm::Intrinsic::read_register, Types); 3655 llvm::Value *Call = Builder.CreateCall(F, Metadata); 3656 3657 if (MixedTypes) 3658 // Read into 64 bit register and then truncate result to 32 bit. 3659 return Builder.CreateTrunc(Call, ValueType); 3660 3661 if (ValueType->isPointerTy()) 3662 // Have i32/i64 result (Call) but want to return a VoidPtrTy (i8*). 3663 return Builder.CreateIntToPtr(Call, ValueType); 3664 3665 return Call; 3666 } 3667 3668 llvm::Value *F = CGM.getIntrinsic(llvm::Intrinsic::write_register, Types); 3669 llvm::Value *ArgValue = CGF.EmitScalarExpr(E->getArg(1)); 3670 if (MixedTypes) { 3671 // Extend 32 bit write value to 64 bit to pass to write. 3672 ArgValue = Builder.CreateZExt(ArgValue, RegisterType); 3673 return Builder.CreateCall(F, { Metadata, ArgValue }); 3674 } 3675 3676 if (ValueType->isPointerTy()) { 3677 // Have VoidPtrTy ArgValue but want to return an i32/i64. 3678 ArgValue = Builder.CreatePtrToInt(ArgValue, RegisterType); 3679 return Builder.CreateCall(F, { Metadata, ArgValue }); 3680 } 3681 3682 return Builder.CreateCall(F, { Metadata, ArgValue }); 3683 } 3684 3685 /// Return true if BuiltinID is an overloaded Neon intrinsic with an extra 3686 /// argument that specifies the vector type. 3687 static bool HasExtraNeonArgument(unsigned BuiltinID) { 3688 switch (BuiltinID) { 3689 default: break; 3690 case NEON::BI__builtin_neon_vget_lane_i8: 3691 case NEON::BI__builtin_neon_vget_lane_i16: 3692 case NEON::BI__builtin_neon_vget_lane_i32: 3693 case NEON::BI__builtin_neon_vget_lane_i64: 3694 case NEON::BI__builtin_neon_vget_lane_f32: 3695 case NEON::BI__builtin_neon_vgetq_lane_i8: 3696 case NEON::BI__builtin_neon_vgetq_lane_i16: 3697 case NEON::BI__builtin_neon_vgetq_lane_i32: 3698 case NEON::BI__builtin_neon_vgetq_lane_i64: 3699 case NEON::BI__builtin_neon_vgetq_lane_f32: 3700 case NEON::BI__builtin_neon_vset_lane_i8: 3701 case NEON::BI__builtin_neon_vset_lane_i16: 3702 case NEON::BI__builtin_neon_vset_lane_i32: 3703 case NEON::BI__builtin_neon_vset_lane_i64: 3704 case NEON::BI__builtin_neon_vset_lane_f32: 3705 case NEON::BI__builtin_neon_vsetq_lane_i8: 3706 case NEON::BI__builtin_neon_vsetq_lane_i16: 3707 case NEON::BI__builtin_neon_vsetq_lane_i32: 3708 case NEON::BI__builtin_neon_vsetq_lane_i64: 3709 case NEON::BI__builtin_neon_vsetq_lane_f32: 3710 case NEON::BI__builtin_neon_vsha1h_u32: 3711 case NEON::BI__builtin_neon_vsha1cq_u32: 3712 case NEON::BI__builtin_neon_vsha1pq_u32: 3713 case NEON::BI__builtin_neon_vsha1mq_u32: 3714 case ARM::BI_MoveToCoprocessor: 3715 case ARM::BI_MoveToCoprocessor2: 3716 return false; 3717 } 3718 return true; 3719 } 3720 3721 Value *CodeGenFunction::EmitARMBuiltinExpr(unsigned BuiltinID, 3722 const CallExpr *E) { 3723 if (auto Hint = GetValueForARMHint(BuiltinID)) 3724 return Hint; 3725 3726 if (BuiltinID == ARM::BI__emit) { 3727 bool IsThumb = getTarget().getTriple().getArch() == llvm::Triple::thumb; 3728 llvm::FunctionType *FTy = 3729 llvm::FunctionType::get(VoidTy, /*Variadic=*/false); 3730 3731 APSInt Value; 3732 if (!E->getArg(0)->EvaluateAsInt(Value, CGM.getContext())) 3733 llvm_unreachable("Sema will ensure that the parameter is constant"); 3734 3735 uint64_t ZExtValue = Value.zextOrTrunc(IsThumb ? 16 : 32).getZExtValue(); 3736 3737 llvm::InlineAsm *Emit = 3738 IsThumb ? InlineAsm::get(FTy, ".inst.n 0x" + utohexstr(ZExtValue), "", 3739 /*SideEffects=*/true) 3740 : InlineAsm::get(FTy, ".inst 0x" + utohexstr(ZExtValue), "", 3741 /*SideEffects=*/true); 3742 3743 return Builder.CreateCall(Emit); 3744 } 3745 3746 if (BuiltinID == ARM::BI__builtin_arm_dbg) { 3747 Value *Option = EmitScalarExpr(E->getArg(0)); 3748 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::arm_dbg), Option); 3749 } 3750 3751 if (BuiltinID == ARM::BI__builtin_arm_prefetch) { 3752 Value *Address = EmitScalarExpr(E->getArg(0)); 3753 Value *RW = EmitScalarExpr(E->getArg(1)); 3754 Value *IsData = EmitScalarExpr(E->getArg(2)); 3755 3756 // Locality is not supported on ARM target 3757 Value *Locality = llvm::ConstantInt::get(Int32Ty, 3); 3758 3759 Value *F = CGM.getIntrinsic(Intrinsic::prefetch); 3760 return Builder.CreateCall(F, {Address, RW, Locality, IsData}); 3761 } 3762 3763 if (BuiltinID == ARM::BI__builtin_arm_rbit) { 3764 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::arm_rbit), 3765 EmitScalarExpr(E->getArg(0)), 3766 "rbit"); 3767 } 3768 3769 if (BuiltinID == ARM::BI__clear_cache) { 3770 assert(E->getNumArgs() == 2 && "__clear_cache takes 2 arguments"); 3771 const FunctionDecl *FD = E->getDirectCallee(); 3772 Value *Ops[2]; 3773 for (unsigned i = 0; i < 2; i++) 3774 Ops[i] = EmitScalarExpr(E->getArg(i)); 3775 llvm::Type *Ty = CGM.getTypes().ConvertType(FD->getType()); 3776 llvm::FunctionType *FTy = cast<llvm::FunctionType>(Ty); 3777 StringRef Name = FD->getName(); 3778 return EmitNounwindRuntimeCall(CGM.CreateRuntimeFunction(FTy, Name), Ops); 3779 } 3780 3781 if (BuiltinID == ARM::BI__builtin_arm_ldrexd || 3782 ((BuiltinID == ARM::BI__builtin_arm_ldrex || 3783 BuiltinID == ARM::BI__builtin_arm_ldaex) && 3784 getContext().getTypeSize(E->getType()) == 64) || 3785 BuiltinID == ARM::BI__ldrexd) { 3786 Function *F; 3787 3788 switch (BuiltinID) { 3789 default: llvm_unreachable("unexpected builtin"); 3790 case ARM::BI__builtin_arm_ldaex: 3791 F = CGM.getIntrinsic(Intrinsic::arm_ldaexd); 3792 break; 3793 case ARM::BI__builtin_arm_ldrexd: 3794 case ARM::BI__builtin_arm_ldrex: 3795 case ARM::BI__ldrexd: 3796 F = CGM.getIntrinsic(Intrinsic::arm_ldrexd); 3797 break; 3798 } 3799 3800 Value *LdPtr = EmitScalarExpr(E->getArg(0)); 3801 Value *Val = Builder.CreateCall(F, Builder.CreateBitCast(LdPtr, Int8PtrTy), 3802 "ldrexd"); 3803 3804 Value *Val0 = Builder.CreateExtractValue(Val, 1); 3805 Value *Val1 = Builder.CreateExtractValue(Val, 0); 3806 Val0 = Builder.CreateZExt(Val0, Int64Ty); 3807 Val1 = Builder.CreateZExt(Val1, Int64Ty); 3808 3809 Value *ShiftCst = llvm::ConstantInt::get(Int64Ty, 32); 3810 Val = Builder.CreateShl(Val0, ShiftCst, "shl", true /* nuw */); 3811 Val = Builder.CreateOr(Val, Val1); 3812 return Builder.CreateBitCast(Val, ConvertType(E->getType())); 3813 } 3814 3815 if (BuiltinID == ARM::BI__builtin_arm_ldrex || 3816 BuiltinID == ARM::BI__builtin_arm_ldaex) { 3817 Value *LoadAddr = EmitScalarExpr(E->getArg(0)); 3818 3819 QualType Ty = E->getType(); 3820 llvm::Type *RealResTy = ConvertType(Ty); 3821 llvm::Type *IntResTy = llvm::IntegerType::get(getLLVMContext(), 3822 getContext().getTypeSize(Ty)); 3823 LoadAddr = Builder.CreateBitCast(LoadAddr, IntResTy->getPointerTo()); 3824 3825 Function *F = CGM.getIntrinsic(BuiltinID == ARM::BI__builtin_arm_ldaex 3826 ? Intrinsic::arm_ldaex 3827 : Intrinsic::arm_ldrex, 3828 LoadAddr->getType()); 3829 Value *Val = Builder.CreateCall(F, LoadAddr, "ldrex"); 3830 3831 if (RealResTy->isPointerTy()) 3832 return Builder.CreateIntToPtr(Val, RealResTy); 3833 else { 3834 Val = Builder.CreateTruncOrBitCast(Val, IntResTy); 3835 return Builder.CreateBitCast(Val, RealResTy); 3836 } 3837 } 3838 3839 if (BuiltinID == ARM::BI__builtin_arm_strexd || 3840 ((BuiltinID == ARM::BI__builtin_arm_stlex || 3841 BuiltinID == ARM::BI__builtin_arm_strex) && 3842 getContext().getTypeSize(E->getArg(0)->getType()) == 64)) { 3843 Function *F = CGM.getIntrinsic(BuiltinID == ARM::BI__builtin_arm_stlex 3844 ? Intrinsic::arm_stlexd 3845 : Intrinsic::arm_strexd); 3846 llvm::Type *STy = llvm::StructType::get(Int32Ty, Int32Ty, nullptr); 3847 3848 Address Tmp = CreateMemTemp(E->getArg(0)->getType()); 3849 Value *Val = EmitScalarExpr(E->getArg(0)); 3850 Builder.CreateStore(Val, Tmp); 3851 3852 Address LdPtr = Builder.CreateBitCast(Tmp,llvm::PointerType::getUnqual(STy)); 3853 Val = Builder.CreateLoad(LdPtr); 3854 3855 Value *Arg0 = Builder.CreateExtractValue(Val, 0); 3856 Value *Arg1 = Builder.CreateExtractValue(Val, 1); 3857 Value *StPtr = Builder.CreateBitCast(EmitScalarExpr(E->getArg(1)), Int8PtrTy); 3858 return Builder.CreateCall(F, {Arg0, Arg1, StPtr}, "strexd"); 3859 } 3860 3861 if (BuiltinID == ARM::BI__builtin_arm_strex || 3862 BuiltinID == ARM::BI__builtin_arm_stlex) { 3863 Value *StoreVal = EmitScalarExpr(E->getArg(0)); 3864 Value *StoreAddr = EmitScalarExpr(E->getArg(1)); 3865 3866 QualType Ty = E->getArg(0)->getType(); 3867 llvm::Type *StoreTy = llvm::IntegerType::get(getLLVMContext(), 3868 getContext().getTypeSize(Ty)); 3869 StoreAddr = Builder.CreateBitCast(StoreAddr, StoreTy->getPointerTo()); 3870 3871 if (StoreVal->getType()->isPointerTy()) 3872 StoreVal = Builder.CreatePtrToInt(StoreVal, Int32Ty); 3873 else { 3874 StoreVal = Builder.CreateBitCast(StoreVal, StoreTy); 3875 StoreVal = Builder.CreateZExtOrBitCast(StoreVal, Int32Ty); 3876 } 3877 3878 Function *F = CGM.getIntrinsic(BuiltinID == ARM::BI__builtin_arm_stlex 3879 ? Intrinsic::arm_stlex 3880 : Intrinsic::arm_strex, 3881 StoreAddr->getType()); 3882 return Builder.CreateCall(F, {StoreVal, StoreAddr}, "strex"); 3883 } 3884 3885 if (BuiltinID == ARM::BI__builtin_arm_clrex) { 3886 Function *F = CGM.getIntrinsic(Intrinsic::arm_clrex); 3887 return Builder.CreateCall(F); 3888 } 3889 3890 // CRC32 3891 Intrinsic::ID CRCIntrinsicID = Intrinsic::not_intrinsic; 3892 switch (BuiltinID) { 3893 case ARM::BI__builtin_arm_crc32b: 3894 CRCIntrinsicID = Intrinsic::arm_crc32b; break; 3895 case ARM::BI__builtin_arm_crc32cb: 3896 CRCIntrinsicID = Intrinsic::arm_crc32cb; break; 3897 case ARM::BI__builtin_arm_crc32h: 3898 CRCIntrinsicID = Intrinsic::arm_crc32h; break; 3899 case ARM::BI__builtin_arm_crc32ch: 3900 CRCIntrinsicID = Intrinsic::arm_crc32ch; break; 3901 case ARM::BI__builtin_arm_crc32w: 3902 case ARM::BI__builtin_arm_crc32d: 3903 CRCIntrinsicID = Intrinsic::arm_crc32w; break; 3904 case ARM::BI__builtin_arm_crc32cw: 3905 case ARM::BI__builtin_arm_crc32cd: 3906 CRCIntrinsicID = Intrinsic::arm_crc32cw; break; 3907 } 3908 3909 if (CRCIntrinsicID != Intrinsic::not_intrinsic) { 3910 Value *Arg0 = EmitScalarExpr(E->getArg(0)); 3911 Value *Arg1 = EmitScalarExpr(E->getArg(1)); 3912 3913 // crc32{c,}d intrinsics are implemnted as two calls to crc32{c,}w 3914 // intrinsics, hence we need different codegen for these cases. 3915 if (BuiltinID == ARM::BI__builtin_arm_crc32d || 3916 BuiltinID == ARM::BI__builtin_arm_crc32cd) { 3917 Value *C1 = llvm::ConstantInt::get(Int64Ty, 32); 3918 Value *Arg1a = Builder.CreateTruncOrBitCast(Arg1, Int32Ty); 3919 Value *Arg1b = Builder.CreateLShr(Arg1, C1); 3920 Arg1b = Builder.CreateTruncOrBitCast(Arg1b, Int32Ty); 3921 3922 Function *F = CGM.getIntrinsic(CRCIntrinsicID); 3923 Value *Res = Builder.CreateCall(F, {Arg0, Arg1a}); 3924 return Builder.CreateCall(F, {Res, Arg1b}); 3925 } else { 3926 Arg1 = Builder.CreateZExtOrBitCast(Arg1, Int32Ty); 3927 3928 Function *F = CGM.getIntrinsic(CRCIntrinsicID); 3929 return Builder.CreateCall(F, {Arg0, Arg1}); 3930 } 3931 } 3932 3933 if (BuiltinID == ARM::BI__builtin_arm_rsr || 3934 BuiltinID == ARM::BI__builtin_arm_rsr64 || 3935 BuiltinID == ARM::BI__builtin_arm_rsrp || 3936 BuiltinID == ARM::BI__builtin_arm_wsr || 3937 BuiltinID == ARM::BI__builtin_arm_wsr64 || 3938 BuiltinID == ARM::BI__builtin_arm_wsrp) { 3939 3940 bool IsRead = BuiltinID == ARM::BI__builtin_arm_rsr || 3941 BuiltinID == ARM::BI__builtin_arm_rsr64 || 3942 BuiltinID == ARM::BI__builtin_arm_rsrp; 3943 3944 bool IsPointerBuiltin = BuiltinID == ARM::BI__builtin_arm_rsrp || 3945 BuiltinID == ARM::BI__builtin_arm_wsrp; 3946 3947 bool Is64Bit = BuiltinID == ARM::BI__builtin_arm_rsr64 || 3948 BuiltinID == ARM::BI__builtin_arm_wsr64; 3949 3950 llvm::Type *ValueType; 3951 llvm::Type *RegisterType; 3952 if (IsPointerBuiltin) { 3953 ValueType = VoidPtrTy; 3954 RegisterType = Int32Ty; 3955 } else if (Is64Bit) { 3956 ValueType = RegisterType = Int64Ty; 3957 } else { 3958 ValueType = RegisterType = Int32Ty; 3959 } 3960 3961 return EmitSpecialRegisterBuiltin(*this, E, RegisterType, ValueType, IsRead); 3962 } 3963 3964 // Find out if any arguments are required to be integer constant 3965 // expressions. 3966 unsigned ICEArguments = 0; 3967 ASTContext::GetBuiltinTypeError Error; 3968 getContext().GetBuiltinType(BuiltinID, Error, &ICEArguments); 3969 assert(Error == ASTContext::GE_None && "Should not codegen an error"); 3970 3971 auto getAlignmentValue32 = [&](Address addr) -> Value* { 3972 return Builder.getInt32(addr.getAlignment().getQuantity()); 3973 }; 3974 3975 Address PtrOp0 = Address::invalid(); 3976 Address PtrOp1 = Address::invalid(); 3977 SmallVector<Value*, 4> Ops; 3978 bool HasExtraArg = HasExtraNeonArgument(BuiltinID); 3979 unsigned NumArgs = E->getNumArgs() - (HasExtraArg ? 1 : 0); 3980 for (unsigned i = 0, e = NumArgs; i != e; i++) { 3981 if (i == 0) { 3982 switch (BuiltinID) { 3983 case NEON::BI__builtin_neon_vld1_v: 3984 case NEON::BI__builtin_neon_vld1q_v: 3985 case NEON::BI__builtin_neon_vld1q_lane_v: 3986 case NEON::BI__builtin_neon_vld1_lane_v: 3987 case NEON::BI__builtin_neon_vld1_dup_v: 3988 case NEON::BI__builtin_neon_vld1q_dup_v: 3989 case NEON::BI__builtin_neon_vst1_v: 3990 case NEON::BI__builtin_neon_vst1q_v: 3991 case NEON::BI__builtin_neon_vst1q_lane_v: 3992 case NEON::BI__builtin_neon_vst1_lane_v: 3993 case NEON::BI__builtin_neon_vst2_v: 3994 case NEON::BI__builtin_neon_vst2q_v: 3995 case NEON::BI__builtin_neon_vst2_lane_v: 3996 case NEON::BI__builtin_neon_vst2q_lane_v: 3997 case NEON::BI__builtin_neon_vst3_v: 3998 case NEON::BI__builtin_neon_vst3q_v: 3999 case NEON::BI__builtin_neon_vst3_lane_v: 4000 case NEON::BI__builtin_neon_vst3q_lane_v: 4001 case NEON::BI__builtin_neon_vst4_v: 4002 case NEON::BI__builtin_neon_vst4q_v: 4003 case NEON::BI__builtin_neon_vst4_lane_v: 4004 case NEON::BI__builtin_neon_vst4q_lane_v: 4005 // Get the alignment for the argument in addition to the value; 4006 // we'll use it later. 4007 PtrOp0 = EmitPointerWithAlignment(E->getArg(0)); 4008 Ops.push_back(PtrOp0.getPointer()); 4009 continue; 4010 } 4011 } 4012 if (i == 1) { 4013 switch (BuiltinID) { 4014 case NEON::BI__builtin_neon_vld2_v: 4015 case NEON::BI__builtin_neon_vld2q_v: 4016 case NEON::BI__builtin_neon_vld3_v: 4017 case NEON::BI__builtin_neon_vld3q_v: 4018 case NEON::BI__builtin_neon_vld4_v: 4019 case NEON::BI__builtin_neon_vld4q_v: 4020 case NEON::BI__builtin_neon_vld2_lane_v: 4021 case NEON::BI__builtin_neon_vld2q_lane_v: 4022 case NEON::BI__builtin_neon_vld3_lane_v: 4023 case NEON::BI__builtin_neon_vld3q_lane_v: 4024 case NEON::BI__builtin_neon_vld4_lane_v: 4025 case NEON::BI__builtin_neon_vld4q_lane_v: 4026 case NEON::BI__builtin_neon_vld2_dup_v: 4027 case NEON::BI__builtin_neon_vld3_dup_v: 4028 case NEON::BI__builtin_neon_vld4_dup_v: 4029 // Get the alignment for the argument in addition to the value; 4030 // we'll use it later. 4031 PtrOp1 = EmitPointerWithAlignment(E->getArg(1)); 4032 Ops.push_back(PtrOp1.getPointer()); 4033 continue; 4034 } 4035 } 4036 4037 if ((ICEArguments & (1 << i)) == 0) { 4038 Ops.push_back(EmitScalarExpr(E->getArg(i))); 4039 } else { 4040 // If this is required to be a constant, constant fold it so that we know 4041 // that the generated intrinsic gets a ConstantInt. 4042 llvm::APSInt Result; 4043 bool IsConst = E->getArg(i)->isIntegerConstantExpr(Result, getContext()); 4044 assert(IsConst && "Constant arg isn't actually constant?"); (void)IsConst; 4045 Ops.push_back(llvm::ConstantInt::get(getLLVMContext(), Result)); 4046 } 4047 } 4048 4049 switch (BuiltinID) { 4050 default: break; 4051 4052 case NEON::BI__builtin_neon_vget_lane_i8: 4053 case NEON::BI__builtin_neon_vget_lane_i16: 4054 case NEON::BI__builtin_neon_vget_lane_i32: 4055 case NEON::BI__builtin_neon_vget_lane_i64: 4056 case NEON::BI__builtin_neon_vget_lane_f32: 4057 case NEON::BI__builtin_neon_vgetq_lane_i8: 4058 case NEON::BI__builtin_neon_vgetq_lane_i16: 4059 case NEON::BI__builtin_neon_vgetq_lane_i32: 4060 case NEON::BI__builtin_neon_vgetq_lane_i64: 4061 case NEON::BI__builtin_neon_vgetq_lane_f32: 4062 return Builder.CreateExtractElement(Ops[0], Ops[1], "vget_lane"); 4063 4064 case NEON::BI__builtin_neon_vset_lane_i8: 4065 case NEON::BI__builtin_neon_vset_lane_i16: 4066 case NEON::BI__builtin_neon_vset_lane_i32: 4067 case NEON::BI__builtin_neon_vset_lane_i64: 4068 case NEON::BI__builtin_neon_vset_lane_f32: 4069 case NEON::BI__builtin_neon_vsetq_lane_i8: 4070 case NEON::BI__builtin_neon_vsetq_lane_i16: 4071 case NEON::BI__builtin_neon_vsetq_lane_i32: 4072 case NEON::BI__builtin_neon_vsetq_lane_i64: 4073 case NEON::BI__builtin_neon_vsetq_lane_f32: 4074 return Builder.CreateInsertElement(Ops[1], Ops[0], Ops[2], "vset_lane"); 4075 4076 case NEON::BI__builtin_neon_vsha1h_u32: 4077 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_sha1h), Ops, 4078 "vsha1h"); 4079 case NEON::BI__builtin_neon_vsha1cq_u32: 4080 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_sha1c), Ops, 4081 "vsha1h"); 4082 case NEON::BI__builtin_neon_vsha1pq_u32: 4083 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_sha1p), Ops, 4084 "vsha1h"); 4085 case NEON::BI__builtin_neon_vsha1mq_u32: 4086 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_sha1m), Ops, 4087 "vsha1h"); 4088 4089 // The ARM _MoveToCoprocessor builtins put the input register value as 4090 // the first argument, but the LLVM intrinsic expects it as the third one. 4091 case ARM::BI_MoveToCoprocessor: 4092 case ARM::BI_MoveToCoprocessor2: { 4093 Function *F = CGM.getIntrinsic(BuiltinID == ARM::BI_MoveToCoprocessor ? 4094 Intrinsic::arm_mcr : Intrinsic::arm_mcr2); 4095 return Builder.CreateCall(F, {Ops[1], Ops[2], Ops[0], 4096 Ops[3], Ops[4], Ops[5]}); 4097 } 4098 } 4099 4100 // Get the last argument, which specifies the vector type. 4101 assert(HasExtraArg); 4102 llvm::APSInt Result; 4103 const Expr *Arg = E->getArg(E->getNumArgs()-1); 4104 if (!Arg->isIntegerConstantExpr(Result, getContext())) 4105 return nullptr; 4106 4107 if (BuiltinID == ARM::BI__builtin_arm_vcvtr_f || 4108 BuiltinID == ARM::BI__builtin_arm_vcvtr_d) { 4109 // Determine the overloaded type of this builtin. 4110 llvm::Type *Ty; 4111 if (BuiltinID == ARM::BI__builtin_arm_vcvtr_f) 4112 Ty = FloatTy; 4113 else 4114 Ty = DoubleTy; 4115 4116 // Determine whether this is an unsigned conversion or not. 4117 bool usgn = Result.getZExtValue() == 1; 4118 unsigned Int = usgn ? Intrinsic::arm_vcvtru : Intrinsic::arm_vcvtr; 4119 4120 // Call the appropriate intrinsic. 4121 Function *F = CGM.getIntrinsic(Int, Ty); 4122 return Builder.CreateCall(F, Ops, "vcvtr"); 4123 } 4124 4125 // Determine the type of this overloaded NEON intrinsic. 4126 NeonTypeFlags Type(Result.getZExtValue()); 4127 bool usgn = Type.isUnsigned(); 4128 bool rightShift = false; 4129 4130 llvm::VectorType *VTy = GetNeonType(this, Type); 4131 llvm::Type *Ty = VTy; 4132 if (!Ty) 4133 return nullptr; 4134 4135 // Many NEON builtins have identical semantics and uses in ARM and 4136 // AArch64. Emit these in a single function. 4137 auto IntrinsicMap = makeArrayRef(ARMSIMDIntrinsicMap); 4138 const NeonIntrinsicInfo *Builtin = findNeonIntrinsicInMap( 4139 IntrinsicMap, BuiltinID, NEONSIMDIntrinsicsProvenSorted); 4140 if (Builtin) 4141 return EmitCommonNeonBuiltinExpr( 4142 Builtin->BuiltinID, Builtin->LLVMIntrinsic, Builtin->AltLLVMIntrinsic, 4143 Builtin->NameHint, Builtin->TypeModifier, E, Ops, PtrOp0, PtrOp1); 4144 4145 unsigned Int; 4146 switch (BuiltinID) { 4147 default: return nullptr; 4148 case NEON::BI__builtin_neon_vld1q_lane_v: 4149 // Handle 64-bit integer elements as a special case. Use shuffles of 4150 // one-element vectors to avoid poor code for i64 in the backend. 4151 if (VTy->getElementType()->isIntegerTy(64)) { 4152 // Extract the other lane. 4153 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 4154 uint32_t Lane = cast<ConstantInt>(Ops[2])->getZExtValue(); 4155 Value *SV = llvm::ConstantVector::get(ConstantInt::get(Int32Ty, 1-Lane)); 4156 Ops[1] = Builder.CreateShuffleVector(Ops[1], Ops[1], SV); 4157 // Load the value as a one-element vector. 4158 Ty = llvm::VectorType::get(VTy->getElementType(), 1); 4159 llvm::Type *Tys[] = {Ty, Int8PtrTy}; 4160 Function *F = CGM.getIntrinsic(Intrinsic::arm_neon_vld1, Tys); 4161 Value *Align = getAlignmentValue32(PtrOp0); 4162 Value *Ld = Builder.CreateCall(F, {Ops[0], Align}); 4163 // Combine them. 4164 uint32_t Indices[] = {1 - Lane, Lane}; 4165 SV = llvm::ConstantDataVector::get(getLLVMContext(), Indices); 4166 return Builder.CreateShuffleVector(Ops[1], Ld, SV, "vld1q_lane"); 4167 } 4168 // fall through 4169 case NEON::BI__builtin_neon_vld1_lane_v: { 4170 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 4171 PtrOp0 = Builder.CreateElementBitCast(PtrOp0, VTy->getElementType()); 4172 Value *Ld = Builder.CreateLoad(PtrOp0); 4173 return Builder.CreateInsertElement(Ops[1], Ld, Ops[2], "vld1_lane"); 4174 } 4175 case NEON::BI__builtin_neon_vld2_dup_v: 4176 case NEON::BI__builtin_neon_vld3_dup_v: 4177 case NEON::BI__builtin_neon_vld4_dup_v: { 4178 // Handle 64-bit elements as a special-case. There is no "dup" needed. 4179 if (VTy->getElementType()->getPrimitiveSizeInBits() == 64) { 4180 switch (BuiltinID) { 4181 case NEON::BI__builtin_neon_vld2_dup_v: 4182 Int = Intrinsic::arm_neon_vld2; 4183 break; 4184 case NEON::BI__builtin_neon_vld3_dup_v: 4185 Int = Intrinsic::arm_neon_vld3; 4186 break; 4187 case NEON::BI__builtin_neon_vld4_dup_v: 4188 Int = Intrinsic::arm_neon_vld4; 4189 break; 4190 default: llvm_unreachable("unknown vld_dup intrinsic?"); 4191 } 4192 llvm::Type *Tys[] = {Ty, Int8PtrTy}; 4193 Function *F = CGM.getIntrinsic(Int, Tys); 4194 llvm::Value *Align = getAlignmentValue32(PtrOp1); 4195 Ops[1] = Builder.CreateCall(F, {Ops[1], Align}, "vld_dup"); 4196 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 4197 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 4198 return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]); 4199 } 4200 switch (BuiltinID) { 4201 case NEON::BI__builtin_neon_vld2_dup_v: 4202 Int = Intrinsic::arm_neon_vld2lane; 4203 break; 4204 case NEON::BI__builtin_neon_vld3_dup_v: 4205 Int = Intrinsic::arm_neon_vld3lane; 4206 break; 4207 case NEON::BI__builtin_neon_vld4_dup_v: 4208 Int = Intrinsic::arm_neon_vld4lane; 4209 break; 4210 default: llvm_unreachable("unknown vld_dup intrinsic?"); 4211 } 4212 llvm::Type *Tys[] = {Ty, Int8PtrTy}; 4213 Function *F = CGM.getIntrinsic(Int, Tys); 4214 llvm::StructType *STy = cast<llvm::StructType>(F->getReturnType()); 4215 4216 SmallVector<Value*, 6> Args; 4217 Args.push_back(Ops[1]); 4218 Args.append(STy->getNumElements(), UndefValue::get(Ty)); 4219 4220 llvm::Constant *CI = ConstantInt::get(Int32Ty, 0); 4221 Args.push_back(CI); 4222 Args.push_back(getAlignmentValue32(PtrOp1)); 4223 4224 Ops[1] = Builder.CreateCall(F, Args, "vld_dup"); 4225 // splat lane 0 to all elts in each vector of the result. 4226 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 4227 Value *Val = Builder.CreateExtractValue(Ops[1], i); 4228 Value *Elt = Builder.CreateBitCast(Val, Ty); 4229 Elt = EmitNeonSplat(Elt, CI); 4230 Elt = Builder.CreateBitCast(Elt, Val->getType()); 4231 Ops[1] = Builder.CreateInsertValue(Ops[1], Elt, i); 4232 } 4233 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 4234 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 4235 return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]); 4236 } 4237 case NEON::BI__builtin_neon_vqrshrn_n_v: 4238 Int = 4239 usgn ? Intrinsic::arm_neon_vqrshiftnu : Intrinsic::arm_neon_vqrshiftns; 4240 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqrshrn_n", 4241 1, true); 4242 case NEON::BI__builtin_neon_vqrshrun_n_v: 4243 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vqrshiftnsu, Ty), 4244 Ops, "vqrshrun_n", 1, true); 4245 case NEON::BI__builtin_neon_vqshrn_n_v: 4246 Int = usgn ? Intrinsic::arm_neon_vqshiftnu : Intrinsic::arm_neon_vqshiftns; 4247 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqshrn_n", 4248 1, true); 4249 case NEON::BI__builtin_neon_vqshrun_n_v: 4250 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vqshiftnsu, Ty), 4251 Ops, "vqshrun_n", 1, true); 4252 case NEON::BI__builtin_neon_vrecpe_v: 4253 case NEON::BI__builtin_neon_vrecpeq_v: 4254 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vrecpe, Ty), 4255 Ops, "vrecpe"); 4256 case NEON::BI__builtin_neon_vrshrn_n_v: 4257 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vrshiftn, Ty), 4258 Ops, "vrshrn_n", 1, true); 4259 case NEON::BI__builtin_neon_vrsra_n_v: 4260 case NEON::BI__builtin_neon_vrsraq_n_v: 4261 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 4262 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 4263 Ops[2] = EmitNeonShiftVector(Ops[2], Ty, true); 4264 Int = usgn ? Intrinsic::arm_neon_vrshiftu : Intrinsic::arm_neon_vrshifts; 4265 Ops[1] = Builder.CreateCall(CGM.getIntrinsic(Int, Ty), {Ops[1], Ops[2]}); 4266 return Builder.CreateAdd(Ops[0], Ops[1], "vrsra_n"); 4267 case NEON::BI__builtin_neon_vsri_n_v: 4268 case NEON::BI__builtin_neon_vsriq_n_v: 4269 rightShift = true; 4270 case NEON::BI__builtin_neon_vsli_n_v: 4271 case NEON::BI__builtin_neon_vsliq_n_v: 4272 Ops[2] = EmitNeonShiftVector(Ops[2], Ty, rightShift); 4273 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vshiftins, Ty), 4274 Ops, "vsli_n"); 4275 case NEON::BI__builtin_neon_vsra_n_v: 4276 case NEON::BI__builtin_neon_vsraq_n_v: 4277 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 4278 Ops[1] = EmitNeonRShiftImm(Ops[1], Ops[2], Ty, usgn, "vsra_n"); 4279 return Builder.CreateAdd(Ops[0], Ops[1]); 4280 case NEON::BI__builtin_neon_vst1q_lane_v: 4281 // Handle 64-bit integer elements as a special case. Use a shuffle to get 4282 // a one-element vector and avoid poor code for i64 in the backend. 4283 if (VTy->getElementType()->isIntegerTy(64)) { 4284 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 4285 Value *SV = llvm::ConstantVector::get(cast<llvm::Constant>(Ops[2])); 4286 Ops[1] = Builder.CreateShuffleVector(Ops[1], Ops[1], SV); 4287 Ops[2] = getAlignmentValue32(PtrOp0); 4288 llvm::Type *Tys[] = {Int8PtrTy, Ops[1]->getType()}; 4289 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::arm_neon_vst1, 4290 Tys), Ops); 4291 } 4292 // fall through 4293 case NEON::BI__builtin_neon_vst1_lane_v: { 4294 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 4295 Ops[1] = Builder.CreateExtractElement(Ops[1], Ops[2]); 4296 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 4297 auto St = Builder.CreateStore(Ops[1], Builder.CreateBitCast(PtrOp0, Ty)); 4298 return St; 4299 } 4300 case NEON::BI__builtin_neon_vtbl1_v: 4301 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbl1), 4302 Ops, "vtbl1"); 4303 case NEON::BI__builtin_neon_vtbl2_v: 4304 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbl2), 4305 Ops, "vtbl2"); 4306 case NEON::BI__builtin_neon_vtbl3_v: 4307 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbl3), 4308 Ops, "vtbl3"); 4309 case NEON::BI__builtin_neon_vtbl4_v: 4310 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbl4), 4311 Ops, "vtbl4"); 4312 case NEON::BI__builtin_neon_vtbx1_v: 4313 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbx1), 4314 Ops, "vtbx1"); 4315 case NEON::BI__builtin_neon_vtbx2_v: 4316 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbx2), 4317 Ops, "vtbx2"); 4318 case NEON::BI__builtin_neon_vtbx3_v: 4319 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbx3), 4320 Ops, "vtbx3"); 4321 case NEON::BI__builtin_neon_vtbx4_v: 4322 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbx4), 4323 Ops, "vtbx4"); 4324 } 4325 } 4326 4327 static Value *EmitAArch64TblBuiltinExpr(CodeGenFunction &CGF, unsigned BuiltinID, 4328 const CallExpr *E, 4329 SmallVectorImpl<Value *> &Ops) { 4330 unsigned int Int = 0; 4331 const char *s = nullptr; 4332 4333 switch (BuiltinID) { 4334 default: 4335 return nullptr; 4336 case NEON::BI__builtin_neon_vtbl1_v: 4337 case NEON::BI__builtin_neon_vqtbl1_v: 4338 case NEON::BI__builtin_neon_vqtbl1q_v: 4339 case NEON::BI__builtin_neon_vtbl2_v: 4340 case NEON::BI__builtin_neon_vqtbl2_v: 4341 case NEON::BI__builtin_neon_vqtbl2q_v: 4342 case NEON::BI__builtin_neon_vtbl3_v: 4343 case NEON::BI__builtin_neon_vqtbl3_v: 4344 case NEON::BI__builtin_neon_vqtbl3q_v: 4345 case NEON::BI__builtin_neon_vtbl4_v: 4346 case NEON::BI__builtin_neon_vqtbl4_v: 4347 case NEON::BI__builtin_neon_vqtbl4q_v: 4348 break; 4349 case NEON::BI__builtin_neon_vtbx1_v: 4350 case NEON::BI__builtin_neon_vqtbx1_v: 4351 case NEON::BI__builtin_neon_vqtbx1q_v: 4352 case NEON::BI__builtin_neon_vtbx2_v: 4353 case NEON::BI__builtin_neon_vqtbx2_v: 4354 case NEON::BI__builtin_neon_vqtbx2q_v: 4355 case NEON::BI__builtin_neon_vtbx3_v: 4356 case NEON::BI__builtin_neon_vqtbx3_v: 4357 case NEON::BI__builtin_neon_vqtbx3q_v: 4358 case NEON::BI__builtin_neon_vtbx4_v: 4359 case NEON::BI__builtin_neon_vqtbx4_v: 4360 case NEON::BI__builtin_neon_vqtbx4q_v: 4361 break; 4362 } 4363 4364 assert(E->getNumArgs() >= 3); 4365 4366 // Get the last argument, which specifies the vector type. 4367 llvm::APSInt Result; 4368 const Expr *Arg = E->getArg(E->getNumArgs() - 1); 4369 if (!Arg->isIntegerConstantExpr(Result, CGF.getContext())) 4370 return nullptr; 4371 4372 // Determine the type of this overloaded NEON intrinsic. 4373 NeonTypeFlags Type(Result.getZExtValue()); 4374 llvm::VectorType *Ty = GetNeonType(&CGF, Type); 4375 if (!Ty) 4376 return nullptr; 4377 4378 CodeGen::CGBuilderTy &Builder = CGF.Builder; 4379 4380 // AArch64 scalar builtins are not overloaded, they do not have an extra 4381 // argument that specifies the vector type, need to handle each case. 4382 switch (BuiltinID) { 4383 case NEON::BI__builtin_neon_vtbl1_v: { 4384 return packTBLDVectorList(CGF, makeArrayRef(Ops).slice(0, 1), nullptr, 4385 Ops[1], Ty, Intrinsic::aarch64_neon_tbl1, 4386 "vtbl1"); 4387 } 4388 case NEON::BI__builtin_neon_vtbl2_v: { 4389 return packTBLDVectorList(CGF, makeArrayRef(Ops).slice(0, 2), nullptr, 4390 Ops[2], Ty, Intrinsic::aarch64_neon_tbl1, 4391 "vtbl1"); 4392 } 4393 case NEON::BI__builtin_neon_vtbl3_v: { 4394 return packTBLDVectorList(CGF, makeArrayRef(Ops).slice(0, 3), nullptr, 4395 Ops[3], Ty, Intrinsic::aarch64_neon_tbl2, 4396 "vtbl2"); 4397 } 4398 case NEON::BI__builtin_neon_vtbl4_v: { 4399 return packTBLDVectorList(CGF, makeArrayRef(Ops).slice(0, 4), nullptr, 4400 Ops[4], Ty, Intrinsic::aarch64_neon_tbl2, 4401 "vtbl2"); 4402 } 4403 case NEON::BI__builtin_neon_vtbx1_v: { 4404 Value *TblRes = 4405 packTBLDVectorList(CGF, makeArrayRef(Ops).slice(1, 1), nullptr, Ops[2], 4406 Ty, Intrinsic::aarch64_neon_tbl1, "vtbl1"); 4407 4408 llvm::Constant *EightV = ConstantInt::get(Ty, 8); 4409 Value *CmpRes = Builder.CreateICmp(ICmpInst::ICMP_UGE, Ops[2], EightV); 4410 CmpRes = Builder.CreateSExt(CmpRes, Ty); 4411 4412 Value *EltsFromInput = Builder.CreateAnd(CmpRes, Ops[0]); 4413 Value *EltsFromTbl = Builder.CreateAnd(Builder.CreateNot(CmpRes), TblRes); 4414 return Builder.CreateOr(EltsFromInput, EltsFromTbl, "vtbx"); 4415 } 4416 case NEON::BI__builtin_neon_vtbx2_v: { 4417 return packTBLDVectorList(CGF, makeArrayRef(Ops).slice(1, 2), Ops[0], 4418 Ops[3], Ty, Intrinsic::aarch64_neon_tbx1, 4419 "vtbx1"); 4420 } 4421 case NEON::BI__builtin_neon_vtbx3_v: { 4422 Value *TblRes = 4423 packTBLDVectorList(CGF, makeArrayRef(Ops).slice(1, 3), nullptr, Ops[4], 4424 Ty, Intrinsic::aarch64_neon_tbl2, "vtbl2"); 4425 4426 llvm::Constant *TwentyFourV = ConstantInt::get(Ty, 24); 4427 Value *CmpRes = Builder.CreateICmp(ICmpInst::ICMP_UGE, Ops[4], 4428 TwentyFourV); 4429 CmpRes = Builder.CreateSExt(CmpRes, Ty); 4430 4431 Value *EltsFromInput = Builder.CreateAnd(CmpRes, Ops[0]); 4432 Value *EltsFromTbl = Builder.CreateAnd(Builder.CreateNot(CmpRes), TblRes); 4433 return Builder.CreateOr(EltsFromInput, EltsFromTbl, "vtbx"); 4434 } 4435 case NEON::BI__builtin_neon_vtbx4_v: { 4436 return packTBLDVectorList(CGF, makeArrayRef(Ops).slice(1, 4), Ops[0], 4437 Ops[5], Ty, Intrinsic::aarch64_neon_tbx2, 4438 "vtbx2"); 4439 } 4440 case NEON::BI__builtin_neon_vqtbl1_v: 4441 case NEON::BI__builtin_neon_vqtbl1q_v: 4442 Int = Intrinsic::aarch64_neon_tbl1; s = "vtbl1"; break; 4443 case NEON::BI__builtin_neon_vqtbl2_v: 4444 case NEON::BI__builtin_neon_vqtbl2q_v: { 4445 Int = Intrinsic::aarch64_neon_tbl2; s = "vtbl2"; break; 4446 case NEON::BI__builtin_neon_vqtbl3_v: 4447 case NEON::BI__builtin_neon_vqtbl3q_v: 4448 Int = Intrinsic::aarch64_neon_tbl3; s = "vtbl3"; break; 4449 case NEON::BI__builtin_neon_vqtbl4_v: 4450 case NEON::BI__builtin_neon_vqtbl4q_v: 4451 Int = Intrinsic::aarch64_neon_tbl4; s = "vtbl4"; break; 4452 case NEON::BI__builtin_neon_vqtbx1_v: 4453 case NEON::BI__builtin_neon_vqtbx1q_v: 4454 Int = Intrinsic::aarch64_neon_tbx1; s = "vtbx1"; break; 4455 case NEON::BI__builtin_neon_vqtbx2_v: 4456 case NEON::BI__builtin_neon_vqtbx2q_v: 4457 Int = Intrinsic::aarch64_neon_tbx2; s = "vtbx2"; break; 4458 case NEON::BI__builtin_neon_vqtbx3_v: 4459 case NEON::BI__builtin_neon_vqtbx3q_v: 4460 Int = Intrinsic::aarch64_neon_tbx3; s = "vtbx3"; break; 4461 case NEON::BI__builtin_neon_vqtbx4_v: 4462 case NEON::BI__builtin_neon_vqtbx4q_v: 4463 Int = Intrinsic::aarch64_neon_tbx4; s = "vtbx4"; break; 4464 } 4465 } 4466 4467 if (!Int) 4468 return nullptr; 4469 4470 Function *F = CGF.CGM.getIntrinsic(Int, Ty); 4471 return CGF.EmitNeonCall(F, Ops, s); 4472 } 4473 4474 Value *CodeGenFunction::vectorWrapScalar16(Value *Op) { 4475 llvm::Type *VTy = llvm::VectorType::get(Int16Ty, 4); 4476 Op = Builder.CreateBitCast(Op, Int16Ty); 4477 Value *V = UndefValue::get(VTy); 4478 llvm::Constant *CI = ConstantInt::get(SizeTy, 0); 4479 Op = Builder.CreateInsertElement(V, Op, CI); 4480 return Op; 4481 } 4482 4483 Value *CodeGenFunction::EmitAArch64BuiltinExpr(unsigned BuiltinID, 4484 const CallExpr *E) { 4485 unsigned HintID = static_cast<unsigned>(-1); 4486 switch (BuiltinID) { 4487 default: break; 4488 case AArch64::BI__builtin_arm_nop: 4489 HintID = 0; 4490 break; 4491 case AArch64::BI__builtin_arm_yield: 4492 HintID = 1; 4493 break; 4494 case AArch64::BI__builtin_arm_wfe: 4495 HintID = 2; 4496 break; 4497 case AArch64::BI__builtin_arm_wfi: 4498 HintID = 3; 4499 break; 4500 case AArch64::BI__builtin_arm_sev: 4501 HintID = 4; 4502 break; 4503 case AArch64::BI__builtin_arm_sevl: 4504 HintID = 5; 4505 break; 4506 } 4507 4508 if (HintID != static_cast<unsigned>(-1)) { 4509 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_hint); 4510 return Builder.CreateCall(F, llvm::ConstantInt::get(Int32Ty, HintID)); 4511 } 4512 4513 if (BuiltinID == AArch64::BI__builtin_arm_prefetch) { 4514 Value *Address = EmitScalarExpr(E->getArg(0)); 4515 Value *RW = EmitScalarExpr(E->getArg(1)); 4516 Value *CacheLevel = EmitScalarExpr(E->getArg(2)); 4517 Value *RetentionPolicy = EmitScalarExpr(E->getArg(3)); 4518 Value *IsData = EmitScalarExpr(E->getArg(4)); 4519 4520 Value *Locality = nullptr; 4521 if (cast<llvm::ConstantInt>(RetentionPolicy)->isZero()) { 4522 // Temporal fetch, needs to convert cache level to locality. 4523 Locality = llvm::ConstantInt::get(Int32Ty, 4524 -cast<llvm::ConstantInt>(CacheLevel)->getValue() + 3); 4525 } else { 4526 // Streaming fetch. 4527 Locality = llvm::ConstantInt::get(Int32Ty, 0); 4528 } 4529 4530 // FIXME: We need AArch64 specific LLVM intrinsic if we want to specify 4531 // PLDL3STRM or PLDL2STRM. 4532 Value *F = CGM.getIntrinsic(Intrinsic::prefetch); 4533 return Builder.CreateCall(F, {Address, RW, Locality, IsData}); 4534 } 4535 4536 if (BuiltinID == AArch64::BI__builtin_arm_rbit) { 4537 assert((getContext().getTypeSize(E->getType()) == 32) && 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 if (BuiltinID == AArch64::BI__builtin_arm_rbit64) { 4544 assert((getContext().getTypeSize(E->getType()) == 64) && 4545 "rbit of unusual size!"); 4546 llvm::Value *Arg = EmitScalarExpr(E->getArg(0)); 4547 return Builder.CreateCall( 4548 CGM.getIntrinsic(Intrinsic::aarch64_rbit, Arg->getType()), Arg, "rbit"); 4549 } 4550 4551 if (BuiltinID == AArch64::BI__clear_cache) { 4552 assert(E->getNumArgs() == 2 && "__clear_cache takes 2 arguments"); 4553 const FunctionDecl *FD = E->getDirectCallee(); 4554 Value *Ops[2]; 4555 for (unsigned i = 0; i < 2; i++) 4556 Ops[i] = EmitScalarExpr(E->getArg(i)); 4557 llvm::Type *Ty = CGM.getTypes().ConvertType(FD->getType()); 4558 llvm::FunctionType *FTy = cast<llvm::FunctionType>(Ty); 4559 StringRef Name = FD->getName(); 4560 return EmitNounwindRuntimeCall(CGM.CreateRuntimeFunction(FTy, Name), Ops); 4561 } 4562 4563 if ((BuiltinID == AArch64::BI__builtin_arm_ldrex || 4564 BuiltinID == AArch64::BI__builtin_arm_ldaex) && 4565 getContext().getTypeSize(E->getType()) == 128) { 4566 Function *F = CGM.getIntrinsic(BuiltinID == AArch64::BI__builtin_arm_ldaex 4567 ? Intrinsic::aarch64_ldaxp 4568 : Intrinsic::aarch64_ldxp); 4569 4570 Value *LdPtr = EmitScalarExpr(E->getArg(0)); 4571 Value *Val = Builder.CreateCall(F, Builder.CreateBitCast(LdPtr, Int8PtrTy), 4572 "ldxp"); 4573 4574 Value *Val0 = Builder.CreateExtractValue(Val, 1); 4575 Value *Val1 = Builder.CreateExtractValue(Val, 0); 4576 llvm::Type *Int128Ty = llvm::IntegerType::get(getLLVMContext(), 128); 4577 Val0 = Builder.CreateZExt(Val0, Int128Ty); 4578 Val1 = Builder.CreateZExt(Val1, Int128Ty); 4579 4580 Value *ShiftCst = llvm::ConstantInt::get(Int128Ty, 64); 4581 Val = Builder.CreateShl(Val0, ShiftCst, "shl", true /* nuw */); 4582 Val = Builder.CreateOr(Val, Val1); 4583 return Builder.CreateBitCast(Val, ConvertType(E->getType())); 4584 } else if (BuiltinID == AArch64::BI__builtin_arm_ldrex || 4585 BuiltinID == AArch64::BI__builtin_arm_ldaex) { 4586 Value *LoadAddr = EmitScalarExpr(E->getArg(0)); 4587 4588 QualType Ty = E->getType(); 4589 llvm::Type *RealResTy = ConvertType(Ty); 4590 llvm::Type *IntResTy = llvm::IntegerType::get(getLLVMContext(), 4591 getContext().getTypeSize(Ty)); 4592 LoadAddr = Builder.CreateBitCast(LoadAddr, IntResTy->getPointerTo()); 4593 4594 Function *F = CGM.getIntrinsic(BuiltinID == AArch64::BI__builtin_arm_ldaex 4595 ? Intrinsic::aarch64_ldaxr 4596 : Intrinsic::aarch64_ldxr, 4597 LoadAddr->getType()); 4598 Value *Val = Builder.CreateCall(F, LoadAddr, "ldxr"); 4599 4600 if (RealResTy->isPointerTy()) 4601 return Builder.CreateIntToPtr(Val, RealResTy); 4602 4603 Val = Builder.CreateTruncOrBitCast(Val, IntResTy); 4604 return Builder.CreateBitCast(Val, RealResTy); 4605 } 4606 4607 if ((BuiltinID == AArch64::BI__builtin_arm_strex || 4608 BuiltinID == AArch64::BI__builtin_arm_stlex) && 4609 getContext().getTypeSize(E->getArg(0)->getType()) == 128) { 4610 Function *F = CGM.getIntrinsic(BuiltinID == AArch64::BI__builtin_arm_stlex 4611 ? Intrinsic::aarch64_stlxp 4612 : Intrinsic::aarch64_stxp); 4613 llvm::Type *STy = llvm::StructType::get(Int64Ty, Int64Ty, nullptr); 4614 4615 Address Tmp = CreateMemTemp(E->getArg(0)->getType()); 4616 EmitAnyExprToMem(E->getArg(0), Tmp, Qualifiers(), /*init*/ true); 4617 4618 Tmp = Builder.CreateBitCast(Tmp, llvm::PointerType::getUnqual(STy)); 4619 llvm::Value *Val = Builder.CreateLoad(Tmp); 4620 4621 Value *Arg0 = Builder.CreateExtractValue(Val, 0); 4622 Value *Arg1 = Builder.CreateExtractValue(Val, 1); 4623 Value *StPtr = Builder.CreateBitCast(EmitScalarExpr(E->getArg(1)), 4624 Int8PtrTy); 4625 return Builder.CreateCall(F, {Arg0, Arg1, StPtr}, "stxp"); 4626 } 4627 4628 if (BuiltinID == AArch64::BI__builtin_arm_strex || 4629 BuiltinID == AArch64::BI__builtin_arm_stlex) { 4630 Value *StoreVal = EmitScalarExpr(E->getArg(0)); 4631 Value *StoreAddr = EmitScalarExpr(E->getArg(1)); 4632 4633 QualType Ty = E->getArg(0)->getType(); 4634 llvm::Type *StoreTy = llvm::IntegerType::get(getLLVMContext(), 4635 getContext().getTypeSize(Ty)); 4636 StoreAddr = Builder.CreateBitCast(StoreAddr, StoreTy->getPointerTo()); 4637 4638 if (StoreVal->getType()->isPointerTy()) 4639 StoreVal = Builder.CreatePtrToInt(StoreVal, Int64Ty); 4640 else { 4641 StoreVal = Builder.CreateBitCast(StoreVal, StoreTy); 4642 StoreVal = Builder.CreateZExtOrBitCast(StoreVal, Int64Ty); 4643 } 4644 4645 Function *F = CGM.getIntrinsic(BuiltinID == AArch64::BI__builtin_arm_stlex 4646 ? Intrinsic::aarch64_stlxr 4647 : Intrinsic::aarch64_stxr, 4648 StoreAddr->getType()); 4649 return Builder.CreateCall(F, {StoreVal, StoreAddr}, "stxr"); 4650 } 4651 4652 if (BuiltinID == AArch64::BI__builtin_arm_clrex) { 4653 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_clrex); 4654 return Builder.CreateCall(F); 4655 } 4656 4657 if (BuiltinID == AArch64::BI__builtin_thread_pointer) { 4658 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_thread_pointer); 4659 return Builder.CreateCall(F); 4660 } 4661 4662 // CRC32 4663 Intrinsic::ID CRCIntrinsicID = Intrinsic::not_intrinsic; 4664 switch (BuiltinID) { 4665 case AArch64::BI__builtin_arm_crc32b: 4666 CRCIntrinsicID = Intrinsic::aarch64_crc32b; break; 4667 case AArch64::BI__builtin_arm_crc32cb: 4668 CRCIntrinsicID = Intrinsic::aarch64_crc32cb; break; 4669 case AArch64::BI__builtin_arm_crc32h: 4670 CRCIntrinsicID = Intrinsic::aarch64_crc32h; break; 4671 case AArch64::BI__builtin_arm_crc32ch: 4672 CRCIntrinsicID = Intrinsic::aarch64_crc32ch; break; 4673 case AArch64::BI__builtin_arm_crc32w: 4674 CRCIntrinsicID = Intrinsic::aarch64_crc32w; break; 4675 case AArch64::BI__builtin_arm_crc32cw: 4676 CRCIntrinsicID = Intrinsic::aarch64_crc32cw; break; 4677 case AArch64::BI__builtin_arm_crc32d: 4678 CRCIntrinsicID = Intrinsic::aarch64_crc32x; break; 4679 case AArch64::BI__builtin_arm_crc32cd: 4680 CRCIntrinsicID = Intrinsic::aarch64_crc32cx; break; 4681 } 4682 4683 if (CRCIntrinsicID != Intrinsic::not_intrinsic) { 4684 Value *Arg0 = EmitScalarExpr(E->getArg(0)); 4685 Value *Arg1 = EmitScalarExpr(E->getArg(1)); 4686 Function *F = CGM.getIntrinsic(CRCIntrinsicID); 4687 4688 llvm::Type *DataTy = F->getFunctionType()->getParamType(1); 4689 Arg1 = Builder.CreateZExtOrBitCast(Arg1, DataTy); 4690 4691 return Builder.CreateCall(F, {Arg0, Arg1}); 4692 } 4693 4694 if (BuiltinID == AArch64::BI__builtin_arm_rsr || 4695 BuiltinID == AArch64::BI__builtin_arm_rsr64 || 4696 BuiltinID == AArch64::BI__builtin_arm_rsrp || 4697 BuiltinID == AArch64::BI__builtin_arm_wsr || 4698 BuiltinID == AArch64::BI__builtin_arm_wsr64 || 4699 BuiltinID == AArch64::BI__builtin_arm_wsrp) { 4700 4701 bool IsRead = BuiltinID == AArch64::BI__builtin_arm_rsr || 4702 BuiltinID == AArch64::BI__builtin_arm_rsr64 || 4703 BuiltinID == AArch64::BI__builtin_arm_rsrp; 4704 4705 bool IsPointerBuiltin = BuiltinID == AArch64::BI__builtin_arm_rsrp || 4706 BuiltinID == AArch64::BI__builtin_arm_wsrp; 4707 4708 bool Is64Bit = BuiltinID != AArch64::BI__builtin_arm_rsr && 4709 BuiltinID != AArch64::BI__builtin_arm_wsr; 4710 4711 llvm::Type *ValueType; 4712 llvm::Type *RegisterType = Int64Ty; 4713 if (IsPointerBuiltin) { 4714 ValueType = VoidPtrTy; 4715 } else if (Is64Bit) { 4716 ValueType = Int64Ty; 4717 } else { 4718 ValueType = Int32Ty; 4719 } 4720 4721 return EmitSpecialRegisterBuiltin(*this, E, RegisterType, ValueType, IsRead); 4722 } 4723 4724 // Find out if any arguments are required to be integer constant 4725 // expressions. 4726 unsigned ICEArguments = 0; 4727 ASTContext::GetBuiltinTypeError Error; 4728 getContext().GetBuiltinType(BuiltinID, Error, &ICEArguments); 4729 assert(Error == ASTContext::GE_None && "Should not codegen an error"); 4730 4731 llvm::SmallVector<Value*, 4> Ops; 4732 for (unsigned i = 0, e = E->getNumArgs() - 1; i != e; i++) { 4733 if ((ICEArguments & (1 << i)) == 0) { 4734 Ops.push_back(EmitScalarExpr(E->getArg(i))); 4735 } else { 4736 // If this is required to be a constant, constant fold it so that we know 4737 // that the generated intrinsic gets a ConstantInt. 4738 llvm::APSInt Result; 4739 bool IsConst = E->getArg(i)->isIntegerConstantExpr(Result, getContext()); 4740 assert(IsConst && "Constant arg isn't actually constant?"); 4741 (void)IsConst; 4742 Ops.push_back(llvm::ConstantInt::get(getLLVMContext(), Result)); 4743 } 4744 } 4745 4746 auto SISDMap = makeArrayRef(AArch64SISDIntrinsicMap); 4747 const NeonIntrinsicInfo *Builtin = findNeonIntrinsicInMap( 4748 SISDMap, BuiltinID, AArch64SISDIntrinsicsProvenSorted); 4749 4750 if (Builtin) { 4751 Ops.push_back(EmitScalarExpr(E->getArg(E->getNumArgs() - 1))); 4752 Value *Result = EmitCommonNeonSISDBuiltinExpr(*this, *Builtin, Ops, E); 4753 assert(Result && "SISD intrinsic should have been handled"); 4754 return Result; 4755 } 4756 4757 llvm::APSInt Result; 4758 const Expr *Arg = E->getArg(E->getNumArgs()-1); 4759 NeonTypeFlags Type(0); 4760 if (Arg->isIntegerConstantExpr(Result, getContext())) 4761 // Determine the type of this overloaded NEON intrinsic. 4762 Type = NeonTypeFlags(Result.getZExtValue()); 4763 4764 bool usgn = Type.isUnsigned(); 4765 bool quad = Type.isQuad(); 4766 4767 // Handle non-overloaded intrinsics first. 4768 switch (BuiltinID) { 4769 default: break; 4770 case NEON::BI__builtin_neon_vldrq_p128: { 4771 llvm::Type *Int128PTy = llvm::Type::getIntNPtrTy(getLLVMContext(), 128); 4772 Value *Ptr = Builder.CreateBitCast(EmitScalarExpr(E->getArg(0)), Int128PTy); 4773 return Builder.CreateDefaultAlignedLoad(Ptr); 4774 } 4775 case NEON::BI__builtin_neon_vstrq_p128: { 4776 llvm::Type *Int128PTy = llvm::Type::getIntNPtrTy(getLLVMContext(), 128); 4777 Value *Ptr = Builder.CreateBitCast(Ops[0], Int128PTy); 4778 return Builder.CreateDefaultAlignedStore(EmitScalarExpr(E->getArg(1)), Ptr); 4779 } 4780 case NEON::BI__builtin_neon_vcvts_u32_f32: 4781 case NEON::BI__builtin_neon_vcvtd_u64_f64: 4782 usgn = true; 4783 // FALL THROUGH 4784 case NEON::BI__builtin_neon_vcvts_s32_f32: 4785 case NEON::BI__builtin_neon_vcvtd_s64_f64: { 4786 Ops.push_back(EmitScalarExpr(E->getArg(0))); 4787 bool Is64 = Ops[0]->getType()->getPrimitiveSizeInBits() == 64; 4788 llvm::Type *InTy = Is64 ? Int64Ty : Int32Ty; 4789 llvm::Type *FTy = Is64 ? DoubleTy : FloatTy; 4790 Ops[0] = Builder.CreateBitCast(Ops[0], FTy); 4791 if (usgn) 4792 return Builder.CreateFPToUI(Ops[0], InTy); 4793 return Builder.CreateFPToSI(Ops[0], InTy); 4794 } 4795 case NEON::BI__builtin_neon_vcvts_f32_u32: 4796 case NEON::BI__builtin_neon_vcvtd_f64_u64: 4797 usgn = true; 4798 // FALL THROUGH 4799 case NEON::BI__builtin_neon_vcvts_f32_s32: 4800 case NEON::BI__builtin_neon_vcvtd_f64_s64: { 4801 Ops.push_back(EmitScalarExpr(E->getArg(0))); 4802 bool Is64 = Ops[0]->getType()->getPrimitiveSizeInBits() == 64; 4803 llvm::Type *InTy = Is64 ? Int64Ty : Int32Ty; 4804 llvm::Type *FTy = Is64 ? DoubleTy : FloatTy; 4805 Ops[0] = Builder.CreateBitCast(Ops[0], InTy); 4806 if (usgn) 4807 return Builder.CreateUIToFP(Ops[0], FTy); 4808 return Builder.CreateSIToFP(Ops[0], FTy); 4809 } 4810 case NEON::BI__builtin_neon_vpaddd_s64: { 4811 llvm::Type *Ty = llvm::VectorType::get(Int64Ty, 2); 4812 Value *Vec = EmitScalarExpr(E->getArg(0)); 4813 // The vector is v2f64, so make sure it's bitcast to that. 4814 Vec = Builder.CreateBitCast(Vec, Ty, "v2i64"); 4815 llvm::Value *Idx0 = llvm::ConstantInt::get(SizeTy, 0); 4816 llvm::Value *Idx1 = llvm::ConstantInt::get(SizeTy, 1); 4817 Value *Op0 = Builder.CreateExtractElement(Vec, Idx0, "lane0"); 4818 Value *Op1 = Builder.CreateExtractElement(Vec, Idx1, "lane1"); 4819 // Pairwise addition of a v2f64 into a scalar f64. 4820 return Builder.CreateAdd(Op0, Op1, "vpaddd"); 4821 } 4822 case NEON::BI__builtin_neon_vpaddd_f64: { 4823 llvm::Type *Ty = 4824 llvm::VectorType::get(DoubleTy, 2); 4825 Value *Vec = EmitScalarExpr(E->getArg(0)); 4826 // The vector is v2f64, so make sure it's bitcast to that. 4827 Vec = Builder.CreateBitCast(Vec, Ty, "v2f64"); 4828 llvm::Value *Idx0 = llvm::ConstantInt::get(SizeTy, 0); 4829 llvm::Value *Idx1 = llvm::ConstantInt::get(SizeTy, 1); 4830 Value *Op0 = Builder.CreateExtractElement(Vec, Idx0, "lane0"); 4831 Value *Op1 = Builder.CreateExtractElement(Vec, Idx1, "lane1"); 4832 // Pairwise addition of a v2f64 into a scalar f64. 4833 return Builder.CreateFAdd(Op0, Op1, "vpaddd"); 4834 } 4835 case NEON::BI__builtin_neon_vpadds_f32: { 4836 llvm::Type *Ty = 4837 llvm::VectorType::get(FloatTy, 2); 4838 Value *Vec = EmitScalarExpr(E->getArg(0)); 4839 // The vector is v2f32, so make sure it's bitcast to that. 4840 Vec = Builder.CreateBitCast(Vec, Ty, "v2f32"); 4841 llvm::Value *Idx0 = llvm::ConstantInt::get(SizeTy, 0); 4842 llvm::Value *Idx1 = llvm::ConstantInt::get(SizeTy, 1); 4843 Value *Op0 = Builder.CreateExtractElement(Vec, Idx0, "lane0"); 4844 Value *Op1 = Builder.CreateExtractElement(Vec, Idx1, "lane1"); 4845 // Pairwise addition of a v2f32 into a scalar f32. 4846 return Builder.CreateFAdd(Op0, Op1, "vpaddd"); 4847 } 4848 case NEON::BI__builtin_neon_vceqzd_s64: 4849 case NEON::BI__builtin_neon_vceqzd_f64: 4850 case NEON::BI__builtin_neon_vceqzs_f32: 4851 Ops.push_back(EmitScalarExpr(E->getArg(0))); 4852 return EmitAArch64CompareBuiltinExpr( 4853 Ops[0], ConvertType(E->getCallReturnType(getContext())), 4854 ICmpInst::FCMP_OEQ, ICmpInst::ICMP_EQ, "vceqz"); 4855 case NEON::BI__builtin_neon_vcgezd_s64: 4856 case NEON::BI__builtin_neon_vcgezd_f64: 4857 case NEON::BI__builtin_neon_vcgezs_f32: 4858 Ops.push_back(EmitScalarExpr(E->getArg(0))); 4859 return EmitAArch64CompareBuiltinExpr( 4860 Ops[0], ConvertType(E->getCallReturnType(getContext())), 4861 ICmpInst::FCMP_OGE, ICmpInst::ICMP_SGE, "vcgez"); 4862 case NEON::BI__builtin_neon_vclezd_s64: 4863 case NEON::BI__builtin_neon_vclezd_f64: 4864 case NEON::BI__builtin_neon_vclezs_f32: 4865 Ops.push_back(EmitScalarExpr(E->getArg(0))); 4866 return EmitAArch64CompareBuiltinExpr( 4867 Ops[0], ConvertType(E->getCallReturnType(getContext())), 4868 ICmpInst::FCMP_OLE, ICmpInst::ICMP_SLE, "vclez"); 4869 case NEON::BI__builtin_neon_vcgtzd_s64: 4870 case NEON::BI__builtin_neon_vcgtzd_f64: 4871 case NEON::BI__builtin_neon_vcgtzs_f32: 4872 Ops.push_back(EmitScalarExpr(E->getArg(0))); 4873 return EmitAArch64CompareBuiltinExpr( 4874 Ops[0], ConvertType(E->getCallReturnType(getContext())), 4875 ICmpInst::FCMP_OGT, ICmpInst::ICMP_SGT, "vcgtz"); 4876 case NEON::BI__builtin_neon_vcltzd_s64: 4877 case NEON::BI__builtin_neon_vcltzd_f64: 4878 case NEON::BI__builtin_neon_vcltzs_f32: 4879 Ops.push_back(EmitScalarExpr(E->getArg(0))); 4880 return EmitAArch64CompareBuiltinExpr( 4881 Ops[0], ConvertType(E->getCallReturnType(getContext())), 4882 ICmpInst::FCMP_OLT, ICmpInst::ICMP_SLT, "vcltz"); 4883 4884 case NEON::BI__builtin_neon_vceqzd_u64: { 4885 Ops.push_back(EmitScalarExpr(E->getArg(0))); 4886 Ops[0] = Builder.CreateBitCast(Ops[0], Int64Ty); 4887 Ops[0] = 4888 Builder.CreateICmpEQ(Ops[0], llvm::Constant::getNullValue(Int64Ty)); 4889 return Builder.CreateSExt(Ops[0], Int64Ty, "vceqzd"); 4890 } 4891 case NEON::BI__builtin_neon_vceqd_f64: 4892 case NEON::BI__builtin_neon_vcled_f64: 4893 case NEON::BI__builtin_neon_vcltd_f64: 4894 case NEON::BI__builtin_neon_vcged_f64: 4895 case NEON::BI__builtin_neon_vcgtd_f64: { 4896 llvm::CmpInst::Predicate P; 4897 switch (BuiltinID) { 4898 default: llvm_unreachable("missing builtin ID in switch!"); 4899 case NEON::BI__builtin_neon_vceqd_f64: P = llvm::FCmpInst::FCMP_OEQ; break; 4900 case NEON::BI__builtin_neon_vcled_f64: P = llvm::FCmpInst::FCMP_OLE; break; 4901 case NEON::BI__builtin_neon_vcltd_f64: P = llvm::FCmpInst::FCMP_OLT; break; 4902 case NEON::BI__builtin_neon_vcged_f64: P = llvm::FCmpInst::FCMP_OGE; break; 4903 case NEON::BI__builtin_neon_vcgtd_f64: P = llvm::FCmpInst::FCMP_OGT; break; 4904 } 4905 Ops.push_back(EmitScalarExpr(E->getArg(1))); 4906 Ops[0] = Builder.CreateBitCast(Ops[0], DoubleTy); 4907 Ops[1] = Builder.CreateBitCast(Ops[1], DoubleTy); 4908 Ops[0] = Builder.CreateFCmp(P, Ops[0], Ops[1]); 4909 return Builder.CreateSExt(Ops[0], Int64Ty, "vcmpd"); 4910 } 4911 case NEON::BI__builtin_neon_vceqs_f32: 4912 case NEON::BI__builtin_neon_vcles_f32: 4913 case NEON::BI__builtin_neon_vclts_f32: 4914 case NEON::BI__builtin_neon_vcges_f32: 4915 case NEON::BI__builtin_neon_vcgts_f32: { 4916 llvm::CmpInst::Predicate P; 4917 switch (BuiltinID) { 4918 default: llvm_unreachable("missing builtin ID in switch!"); 4919 case NEON::BI__builtin_neon_vceqs_f32: P = llvm::FCmpInst::FCMP_OEQ; break; 4920 case NEON::BI__builtin_neon_vcles_f32: P = llvm::FCmpInst::FCMP_OLE; break; 4921 case NEON::BI__builtin_neon_vclts_f32: P = llvm::FCmpInst::FCMP_OLT; break; 4922 case NEON::BI__builtin_neon_vcges_f32: P = llvm::FCmpInst::FCMP_OGE; break; 4923 case NEON::BI__builtin_neon_vcgts_f32: P = llvm::FCmpInst::FCMP_OGT; break; 4924 } 4925 Ops.push_back(EmitScalarExpr(E->getArg(1))); 4926 Ops[0] = Builder.CreateBitCast(Ops[0], FloatTy); 4927 Ops[1] = Builder.CreateBitCast(Ops[1], FloatTy); 4928 Ops[0] = Builder.CreateFCmp(P, Ops[0], Ops[1]); 4929 return Builder.CreateSExt(Ops[0], Int32Ty, "vcmpd"); 4930 } 4931 case NEON::BI__builtin_neon_vceqd_s64: 4932 case NEON::BI__builtin_neon_vceqd_u64: 4933 case NEON::BI__builtin_neon_vcgtd_s64: 4934 case NEON::BI__builtin_neon_vcgtd_u64: 4935 case NEON::BI__builtin_neon_vcltd_s64: 4936 case NEON::BI__builtin_neon_vcltd_u64: 4937 case NEON::BI__builtin_neon_vcged_u64: 4938 case NEON::BI__builtin_neon_vcged_s64: 4939 case NEON::BI__builtin_neon_vcled_u64: 4940 case NEON::BI__builtin_neon_vcled_s64: { 4941 llvm::CmpInst::Predicate P; 4942 switch (BuiltinID) { 4943 default: llvm_unreachable("missing builtin ID in switch!"); 4944 case NEON::BI__builtin_neon_vceqd_s64: 4945 case NEON::BI__builtin_neon_vceqd_u64:P = llvm::ICmpInst::ICMP_EQ;break; 4946 case NEON::BI__builtin_neon_vcgtd_s64:P = llvm::ICmpInst::ICMP_SGT;break; 4947 case NEON::BI__builtin_neon_vcgtd_u64:P = llvm::ICmpInst::ICMP_UGT;break; 4948 case NEON::BI__builtin_neon_vcltd_s64:P = llvm::ICmpInst::ICMP_SLT;break; 4949 case NEON::BI__builtin_neon_vcltd_u64:P = llvm::ICmpInst::ICMP_ULT;break; 4950 case NEON::BI__builtin_neon_vcged_u64:P = llvm::ICmpInst::ICMP_UGE;break; 4951 case NEON::BI__builtin_neon_vcged_s64:P = llvm::ICmpInst::ICMP_SGE;break; 4952 case NEON::BI__builtin_neon_vcled_u64:P = llvm::ICmpInst::ICMP_ULE;break; 4953 case NEON::BI__builtin_neon_vcled_s64:P = llvm::ICmpInst::ICMP_SLE;break; 4954 } 4955 Ops.push_back(EmitScalarExpr(E->getArg(1))); 4956 Ops[0] = Builder.CreateBitCast(Ops[0], Int64Ty); 4957 Ops[1] = Builder.CreateBitCast(Ops[1], Int64Ty); 4958 Ops[0] = Builder.CreateICmp(P, Ops[0], Ops[1]); 4959 return Builder.CreateSExt(Ops[0], Int64Ty, "vceqd"); 4960 } 4961 case NEON::BI__builtin_neon_vtstd_s64: 4962 case NEON::BI__builtin_neon_vtstd_u64: { 4963 Ops.push_back(EmitScalarExpr(E->getArg(1))); 4964 Ops[0] = Builder.CreateBitCast(Ops[0], Int64Ty); 4965 Ops[1] = Builder.CreateBitCast(Ops[1], Int64Ty); 4966 Ops[0] = Builder.CreateAnd(Ops[0], Ops[1]); 4967 Ops[0] = Builder.CreateICmp(ICmpInst::ICMP_NE, Ops[0], 4968 llvm::Constant::getNullValue(Int64Ty)); 4969 return Builder.CreateSExt(Ops[0], Int64Ty, "vtstd"); 4970 } 4971 case NEON::BI__builtin_neon_vset_lane_i8: 4972 case NEON::BI__builtin_neon_vset_lane_i16: 4973 case NEON::BI__builtin_neon_vset_lane_i32: 4974 case NEON::BI__builtin_neon_vset_lane_i64: 4975 case NEON::BI__builtin_neon_vset_lane_f32: 4976 case NEON::BI__builtin_neon_vsetq_lane_i8: 4977 case NEON::BI__builtin_neon_vsetq_lane_i16: 4978 case NEON::BI__builtin_neon_vsetq_lane_i32: 4979 case NEON::BI__builtin_neon_vsetq_lane_i64: 4980 case NEON::BI__builtin_neon_vsetq_lane_f32: 4981 Ops.push_back(EmitScalarExpr(E->getArg(2))); 4982 return Builder.CreateInsertElement(Ops[1], Ops[0], Ops[2], "vset_lane"); 4983 case NEON::BI__builtin_neon_vset_lane_f64: 4984 // The vector type needs a cast for the v1f64 variant. 4985 Ops[1] = Builder.CreateBitCast(Ops[1], 4986 llvm::VectorType::get(DoubleTy, 1)); 4987 Ops.push_back(EmitScalarExpr(E->getArg(2))); 4988 return Builder.CreateInsertElement(Ops[1], Ops[0], Ops[2], "vset_lane"); 4989 case NEON::BI__builtin_neon_vsetq_lane_f64: 4990 // The vector type needs a cast for the v2f64 variant. 4991 Ops[1] = Builder.CreateBitCast(Ops[1], 4992 llvm::VectorType::get(DoubleTy, 2)); 4993 Ops.push_back(EmitScalarExpr(E->getArg(2))); 4994 return Builder.CreateInsertElement(Ops[1], Ops[0], Ops[2], "vset_lane"); 4995 4996 case NEON::BI__builtin_neon_vget_lane_i8: 4997 case NEON::BI__builtin_neon_vdupb_lane_i8: 4998 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::VectorType::get(Int8Ty, 8)); 4999 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 5000 "vget_lane"); 5001 case NEON::BI__builtin_neon_vgetq_lane_i8: 5002 case NEON::BI__builtin_neon_vdupb_laneq_i8: 5003 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::VectorType::get(Int8Ty, 16)); 5004 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 5005 "vgetq_lane"); 5006 case NEON::BI__builtin_neon_vget_lane_i16: 5007 case NEON::BI__builtin_neon_vduph_lane_i16: 5008 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::VectorType::get(Int16Ty, 4)); 5009 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 5010 "vget_lane"); 5011 case NEON::BI__builtin_neon_vgetq_lane_i16: 5012 case NEON::BI__builtin_neon_vduph_laneq_i16: 5013 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::VectorType::get(Int16Ty, 8)); 5014 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 5015 "vgetq_lane"); 5016 case NEON::BI__builtin_neon_vget_lane_i32: 5017 case NEON::BI__builtin_neon_vdups_lane_i32: 5018 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::VectorType::get(Int32Ty, 2)); 5019 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 5020 "vget_lane"); 5021 case NEON::BI__builtin_neon_vdups_lane_f32: 5022 Ops[0] = Builder.CreateBitCast(Ops[0], 5023 llvm::VectorType::get(FloatTy, 2)); 5024 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 5025 "vdups_lane"); 5026 case NEON::BI__builtin_neon_vgetq_lane_i32: 5027 case NEON::BI__builtin_neon_vdups_laneq_i32: 5028 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::VectorType::get(Int32Ty, 4)); 5029 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 5030 "vgetq_lane"); 5031 case NEON::BI__builtin_neon_vget_lane_i64: 5032 case NEON::BI__builtin_neon_vdupd_lane_i64: 5033 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::VectorType::get(Int64Ty, 1)); 5034 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 5035 "vget_lane"); 5036 case NEON::BI__builtin_neon_vdupd_lane_f64: 5037 Ops[0] = Builder.CreateBitCast(Ops[0], 5038 llvm::VectorType::get(DoubleTy, 1)); 5039 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 5040 "vdupd_lane"); 5041 case NEON::BI__builtin_neon_vgetq_lane_i64: 5042 case NEON::BI__builtin_neon_vdupd_laneq_i64: 5043 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::VectorType::get(Int64Ty, 2)); 5044 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 5045 "vgetq_lane"); 5046 case NEON::BI__builtin_neon_vget_lane_f32: 5047 Ops[0] = Builder.CreateBitCast(Ops[0], 5048 llvm::VectorType::get(FloatTy, 2)); 5049 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 5050 "vget_lane"); 5051 case NEON::BI__builtin_neon_vget_lane_f64: 5052 Ops[0] = Builder.CreateBitCast(Ops[0], 5053 llvm::VectorType::get(DoubleTy, 1)); 5054 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 5055 "vget_lane"); 5056 case NEON::BI__builtin_neon_vgetq_lane_f32: 5057 case NEON::BI__builtin_neon_vdups_laneq_f32: 5058 Ops[0] = Builder.CreateBitCast(Ops[0], 5059 llvm::VectorType::get(FloatTy, 4)); 5060 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 5061 "vgetq_lane"); 5062 case NEON::BI__builtin_neon_vgetq_lane_f64: 5063 case NEON::BI__builtin_neon_vdupd_laneq_f64: 5064 Ops[0] = Builder.CreateBitCast(Ops[0], 5065 llvm::VectorType::get(DoubleTy, 2)); 5066 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 5067 "vgetq_lane"); 5068 case NEON::BI__builtin_neon_vaddd_s64: 5069 case NEON::BI__builtin_neon_vaddd_u64: 5070 return Builder.CreateAdd(Ops[0], EmitScalarExpr(E->getArg(1)), "vaddd"); 5071 case NEON::BI__builtin_neon_vsubd_s64: 5072 case NEON::BI__builtin_neon_vsubd_u64: 5073 return Builder.CreateSub(Ops[0], EmitScalarExpr(E->getArg(1)), "vsubd"); 5074 case NEON::BI__builtin_neon_vqdmlalh_s16: 5075 case NEON::BI__builtin_neon_vqdmlslh_s16: { 5076 SmallVector<Value *, 2> ProductOps; 5077 ProductOps.push_back(vectorWrapScalar16(Ops[1])); 5078 ProductOps.push_back(vectorWrapScalar16(EmitScalarExpr(E->getArg(2)))); 5079 llvm::Type *VTy = llvm::VectorType::get(Int32Ty, 4); 5080 Ops[1] = EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_sqdmull, VTy), 5081 ProductOps, "vqdmlXl"); 5082 Constant *CI = ConstantInt::get(SizeTy, 0); 5083 Ops[1] = Builder.CreateExtractElement(Ops[1], CI, "lane0"); 5084 5085 unsigned AccumInt = BuiltinID == NEON::BI__builtin_neon_vqdmlalh_s16 5086 ? Intrinsic::aarch64_neon_sqadd 5087 : Intrinsic::aarch64_neon_sqsub; 5088 return EmitNeonCall(CGM.getIntrinsic(AccumInt, Int32Ty), Ops, "vqdmlXl"); 5089 } 5090 case NEON::BI__builtin_neon_vqshlud_n_s64: { 5091 Ops.push_back(EmitScalarExpr(E->getArg(1))); 5092 Ops[1] = Builder.CreateZExt(Ops[1], Int64Ty); 5093 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_sqshlu, Int64Ty), 5094 Ops, "vqshlu_n"); 5095 } 5096 case NEON::BI__builtin_neon_vqshld_n_u64: 5097 case NEON::BI__builtin_neon_vqshld_n_s64: { 5098 unsigned Int = BuiltinID == NEON::BI__builtin_neon_vqshld_n_u64 5099 ? Intrinsic::aarch64_neon_uqshl 5100 : Intrinsic::aarch64_neon_sqshl; 5101 Ops.push_back(EmitScalarExpr(E->getArg(1))); 5102 Ops[1] = Builder.CreateZExt(Ops[1], Int64Ty); 5103 return EmitNeonCall(CGM.getIntrinsic(Int, Int64Ty), Ops, "vqshl_n"); 5104 } 5105 case NEON::BI__builtin_neon_vrshrd_n_u64: 5106 case NEON::BI__builtin_neon_vrshrd_n_s64: { 5107 unsigned Int = BuiltinID == NEON::BI__builtin_neon_vrshrd_n_u64 5108 ? Intrinsic::aarch64_neon_urshl 5109 : Intrinsic::aarch64_neon_srshl; 5110 Ops.push_back(EmitScalarExpr(E->getArg(1))); 5111 int SV = cast<ConstantInt>(Ops[1])->getSExtValue(); 5112 Ops[1] = ConstantInt::get(Int64Ty, -SV); 5113 return EmitNeonCall(CGM.getIntrinsic(Int, Int64Ty), Ops, "vrshr_n"); 5114 } 5115 case NEON::BI__builtin_neon_vrsrad_n_u64: 5116 case NEON::BI__builtin_neon_vrsrad_n_s64: { 5117 unsigned Int = BuiltinID == NEON::BI__builtin_neon_vrsrad_n_u64 5118 ? Intrinsic::aarch64_neon_urshl 5119 : Intrinsic::aarch64_neon_srshl; 5120 Ops[1] = Builder.CreateBitCast(Ops[1], Int64Ty); 5121 Ops.push_back(Builder.CreateNeg(EmitScalarExpr(E->getArg(2)))); 5122 Ops[1] = Builder.CreateCall(CGM.getIntrinsic(Int, Int64Ty), 5123 {Ops[1], Builder.CreateSExt(Ops[2], Int64Ty)}); 5124 return Builder.CreateAdd(Ops[0], Builder.CreateBitCast(Ops[1], Int64Ty)); 5125 } 5126 case NEON::BI__builtin_neon_vshld_n_s64: 5127 case NEON::BI__builtin_neon_vshld_n_u64: { 5128 llvm::ConstantInt *Amt = cast<ConstantInt>(EmitScalarExpr(E->getArg(1))); 5129 return Builder.CreateShl( 5130 Ops[0], ConstantInt::get(Int64Ty, Amt->getZExtValue()), "shld_n"); 5131 } 5132 case NEON::BI__builtin_neon_vshrd_n_s64: { 5133 llvm::ConstantInt *Amt = cast<ConstantInt>(EmitScalarExpr(E->getArg(1))); 5134 return Builder.CreateAShr( 5135 Ops[0], ConstantInt::get(Int64Ty, std::min(static_cast<uint64_t>(63), 5136 Amt->getZExtValue())), 5137 "shrd_n"); 5138 } 5139 case NEON::BI__builtin_neon_vshrd_n_u64: { 5140 llvm::ConstantInt *Amt = cast<ConstantInt>(EmitScalarExpr(E->getArg(1))); 5141 uint64_t ShiftAmt = Amt->getZExtValue(); 5142 // Right-shifting an unsigned value by its size yields 0. 5143 if (ShiftAmt == 64) 5144 return ConstantInt::get(Int64Ty, 0); 5145 return Builder.CreateLShr(Ops[0], ConstantInt::get(Int64Ty, ShiftAmt), 5146 "shrd_n"); 5147 } 5148 case NEON::BI__builtin_neon_vsrad_n_s64: { 5149 llvm::ConstantInt *Amt = cast<ConstantInt>(EmitScalarExpr(E->getArg(2))); 5150 Ops[1] = Builder.CreateAShr( 5151 Ops[1], ConstantInt::get(Int64Ty, std::min(static_cast<uint64_t>(63), 5152 Amt->getZExtValue())), 5153 "shrd_n"); 5154 return Builder.CreateAdd(Ops[0], Ops[1]); 5155 } 5156 case NEON::BI__builtin_neon_vsrad_n_u64: { 5157 llvm::ConstantInt *Amt = cast<ConstantInt>(EmitScalarExpr(E->getArg(2))); 5158 uint64_t ShiftAmt = Amt->getZExtValue(); 5159 // Right-shifting an unsigned value by its size yields 0. 5160 // As Op + 0 = Op, return Ops[0] directly. 5161 if (ShiftAmt == 64) 5162 return Ops[0]; 5163 Ops[1] = Builder.CreateLShr(Ops[1], ConstantInt::get(Int64Ty, ShiftAmt), 5164 "shrd_n"); 5165 return Builder.CreateAdd(Ops[0], Ops[1]); 5166 } 5167 case NEON::BI__builtin_neon_vqdmlalh_lane_s16: 5168 case NEON::BI__builtin_neon_vqdmlalh_laneq_s16: 5169 case NEON::BI__builtin_neon_vqdmlslh_lane_s16: 5170 case NEON::BI__builtin_neon_vqdmlslh_laneq_s16: { 5171 Ops[2] = Builder.CreateExtractElement(Ops[2], EmitScalarExpr(E->getArg(3)), 5172 "lane"); 5173 SmallVector<Value *, 2> ProductOps; 5174 ProductOps.push_back(vectorWrapScalar16(Ops[1])); 5175 ProductOps.push_back(vectorWrapScalar16(Ops[2])); 5176 llvm::Type *VTy = llvm::VectorType::get(Int32Ty, 4); 5177 Ops[1] = EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_sqdmull, VTy), 5178 ProductOps, "vqdmlXl"); 5179 Constant *CI = ConstantInt::get(SizeTy, 0); 5180 Ops[1] = Builder.CreateExtractElement(Ops[1], CI, "lane0"); 5181 Ops.pop_back(); 5182 5183 unsigned AccInt = (BuiltinID == NEON::BI__builtin_neon_vqdmlalh_lane_s16 || 5184 BuiltinID == NEON::BI__builtin_neon_vqdmlalh_laneq_s16) 5185 ? Intrinsic::aarch64_neon_sqadd 5186 : Intrinsic::aarch64_neon_sqsub; 5187 return EmitNeonCall(CGM.getIntrinsic(AccInt, Int32Ty), Ops, "vqdmlXl"); 5188 } 5189 case NEON::BI__builtin_neon_vqdmlals_s32: 5190 case NEON::BI__builtin_neon_vqdmlsls_s32: { 5191 SmallVector<Value *, 2> ProductOps; 5192 ProductOps.push_back(Ops[1]); 5193 ProductOps.push_back(EmitScalarExpr(E->getArg(2))); 5194 Ops[1] = 5195 EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_sqdmulls_scalar), 5196 ProductOps, "vqdmlXl"); 5197 5198 unsigned AccumInt = BuiltinID == NEON::BI__builtin_neon_vqdmlals_s32 5199 ? Intrinsic::aarch64_neon_sqadd 5200 : Intrinsic::aarch64_neon_sqsub; 5201 return EmitNeonCall(CGM.getIntrinsic(AccumInt, Int64Ty), Ops, "vqdmlXl"); 5202 } 5203 case NEON::BI__builtin_neon_vqdmlals_lane_s32: 5204 case NEON::BI__builtin_neon_vqdmlals_laneq_s32: 5205 case NEON::BI__builtin_neon_vqdmlsls_lane_s32: 5206 case NEON::BI__builtin_neon_vqdmlsls_laneq_s32: { 5207 Ops[2] = Builder.CreateExtractElement(Ops[2], EmitScalarExpr(E->getArg(3)), 5208 "lane"); 5209 SmallVector<Value *, 2> ProductOps; 5210 ProductOps.push_back(Ops[1]); 5211 ProductOps.push_back(Ops[2]); 5212 Ops[1] = 5213 EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_sqdmulls_scalar), 5214 ProductOps, "vqdmlXl"); 5215 Ops.pop_back(); 5216 5217 unsigned AccInt = (BuiltinID == NEON::BI__builtin_neon_vqdmlals_lane_s32 || 5218 BuiltinID == NEON::BI__builtin_neon_vqdmlals_laneq_s32) 5219 ? Intrinsic::aarch64_neon_sqadd 5220 : Intrinsic::aarch64_neon_sqsub; 5221 return EmitNeonCall(CGM.getIntrinsic(AccInt, Int64Ty), Ops, "vqdmlXl"); 5222 } 5223 } 5224 5225 llvm::VectorType *VTy = GetNeonType(this, Type); 5226 llvm::Type *Ty = VTy; 5227 if (!Ty) 5228 return nullptr; 5229 5230 // Not all intrinsics handled by the common case work for AArch64 yet, so only 5231 // defer to common code if it's been added to our special map. 5232 Builtin = findNeonIntrinsicInMap(AArch64SIMDIntrinsicMap, BuiltinID, 5233 AArch64SIMDIntrinsicsProvenSorted); 5234 5235 if (Builtin) 5236 return EmitCommonNeonBuiltinExpr( 5237 Builtin->BuiltinID, Builtin->LLVMIntrinsic, Builtin->AltLLVMIntrinsic, 5238 Builtin->NameHint, Builtin->TypeModifier, E, Ops, 5239 /*never use addresses*/ Address::invalid(), Address::invalid()); 5240 5241 if (Value *V = EmitAArch64TblBuiltinExpr(*this, BuiltinID, E, Ops)) 5242 return V; 5243 5244 unsigned Int; 5245 switch (BuiltinID) { 5246 default: return nullptr; 5247 case NEON::BI__builtin_neon_vbsl_v: 5248 case NEON::BI__builtin_neon_vbslq_v: { 5249 llvm::Type *BitTy = llvm::VectorType::getInteger(VTy); 5250 Ops[0] = Builder.CreateBitCast(Ops[0], BitTy, "vbsl"); 5251 Ops[1] = Builder.CreateBitCast(Ops[1], BitTy, "vbsl"); 5252 Ops[2] = Builder.CreateBitCast(Ops[2], BitTy, "vbsl"); 5253 5254 Ops[1] = Builder.CreateAnd(Ops[0], Ops[1], "vbsl"); 5255 Ops[2] = Builder.CreateAnd(Builder.CreateNot(Ops[0]), Ops[2], "vbsl"); 5256 Ops[0] = Builder.CreateOr(Ops[1], Ops[2], "vbsl"); 5257 return Builder.CreateBitCast(Ops[0], Ty); 5258 } 5259 case NEON::BI__builtin_neon_vfma_lane_v: 5260 case NEON::BI__builtin_neon_vfmaq_lane_v: { // Only used for FP types 5261 // The ARM builtins (and instructions) have the addend as the first 5262 // operand, but the 'fma' intrinsics have it last. Swap it around here. 5263 Value *Addend = Ops[0]; 5264 Value *Multiplicand = Ops[1]; 5265 Value *LaneSource = Ops[2]; 5266 Ops[0] = Multiplicand; 5267 Ops[1] = LaneSource; 5268 Ops[2] = Addend; 5269 5270 // Now adjust things to handle the lane access. 5271 llvm::Type *SourceTy = BuiltinID == NEON::BI__builtin_neon_vfmaq_lane_v ? 5272 llvm::VectorType::get(VTy->getElementType(), VTy->getNumElements() / 2) : 5273 VTy; 5274 llvm::Constant *cst = cast<Constant>(Ops[3]); 5275 Value *SV = llvm::ConstantVector::getSplat(VTy->getNumElements(), cst); 5276 Ops[1] = Builder.CreateBitCast(Ops[1], SourceTy); 5277 Ops[1] = Builder.CreateShuffleVector(Ops[1], Ops[1], SV, "lane"); 5278 5279 Ops.pop_back(); 5280 Int = Intrinsic::fma; 5281 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "fmla"); 5282 } 5283 case NEON::BI__builtin_neon_vfma_laneq_v: { 5284 llvm::VectorType *VTy = cast<llvm::VectorType>(Ty); 5285 // v1f64 fma should be mapped to Neon scalar f64 fma 5286 if (VTy && VTy->getElementType() == DoubleTy) { 5287 Ops[0] = Builder.CreateBitCast(Ops[0], DoubleTy); 5288 Ops[1] = Builder.CreateBitCast(Ops[1], DoubleTy); 5289 llvm::Type *VTy = GetNeonType(this, 5290 NeonTypeFlags(NeonTypeFlags::Float64, false, true)); 5291 Ops[2] = Builder.CreateBitCast(Ops[2], VTy); 5292 Ops[2] = Builder.CreateExtractElement(Ops[2], Ops[3], "extract"); 5293 Value *F = CGM.getIntrinsic(Intrinsic::fma, DoubleTy); 5294 Value *Result = Builder.CreateCall(F, {Ops[1], Ops[2], Ops[0]}); 5295 return Builder.CreateBitCast(Result, Ty); 5296 } 5297 Value *F = CGM.getIntrinsic(Intrinsic::fma, Ty); 5298 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 5299 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 5300 5301 llvm::Type *STy = llvm::VectorType::get(VTy->getElementType(), 5302 VTy->getNumElements() * 2); 5303 Ops[2] = Builder.CreateBitCast(Ops[2], STy); 5304 Value* SV = llvm::ConstantVector::getSplat(VTy->getNumElements(), 5305 cast<ConstantInt>(Ops[3])); 5306 Ops[2] = Builder.CreateShuffleVector(Ops[2], Ops[2], SV, "lane"); 5307 5308 return Builder.CreateCall(F, {Ops[2], Ops[1], Ops[0]}); 5309 } 5310 case NEON::BI__builtin_neon_vfmaq_laneq_v: { 5311 Value *F = CGM.getIntrinsic(Intrinsic::fma, Ty); 5312 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 5313 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 5314 5315 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 5316 Ops[2] = EmitNeonSplat(Ops[2], cast<ConstantInt>(Ops[3])); 5317 return Builder.CreateCall(F, {Ops[2], Ops[1], Ops[0]}); 5318 } 5319 case NEON::BI__builtin_neon_vfmas_lane_f32: 5320 case NEON::BI__builtin_neon_vfmas_laneq_f32: 5321 case NEON::BI__builtin_neon_vfmad_lane_f64: 5322 case NEON::BI__builtin_neon_vfmad_laneq_f64: { 5323 Ops.push_back(EmitScalarExpr(E->getArg(3))); 5324 llvm::Type *Ty = ConvertType(E->getCallReturnType(getContext())); 5325 Value *F = CGM.getIntrinsic(Intrinsic::fma, Ty); 5326 Ops[2] = Builder.CreateExtractElement(Ops[2], Ops[3], "extract"); 5327 return Builder.CreateCall(F, {Ops[1], Ops[2], Ops[0]}); 5328 } 5329 case NEON::BI__builtin_neon_vfms_v: 5330 case NEON::BI__builtin_neon_vfmsq_v: { // Only used for FP types 5331 // FIXME: probably remove when we no longer support aarch64_simd.h 5332 // (arm_neon.h delegates to vfma). 5333 5334 // The ARM builtins (and instructions) have the addend as the first 5335 // operand, but the 'fma' intrinsics have it last. Swap it around here. 5336 Value *Subtrahend = Ops[0]; 5337 Value *Multiplicand = Ops[2]; 5338 Ops[0] = Multiplicand; 5339 Ops[2] = Subtrahend; 5340 Ops[1] = Builder.CreateBitCast(Ops[1], VTy); 5341 Ops[1] = Builder.CreateFNeg(Ops[1]); 5342 Int = Intrinsic::fma; 5343 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "fmls"); 5344 } 5345 case NEON::BI__builtin_neon_vmull_v: 5346 // FIXME: improve sharing scheme to cope with 3 alternative LLVM intrinsics. 5347 Int = usgn ? Intrinsic::aarch64_neon_umull : Intrinsic::aarch64_neon_smull; 5348 if (Type.isPoly()) Int = Intrinsic::aarch64_neon_pmull; 5349 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmull"); 5350 case NEON::BI__builtin_neon_vmax_v: 5351 case NEON::BI__builtin_neon_vmaxq_v: 5352 // FIXME: improve sharing scheme to cope with 3 alternative LLVM intrinsics. 5353 Int = usgn ? Intrinsic::aarch64_neon_umax : Intrinsic::aarch64_neon_smax; 5354 if (Ty->isFPOrFPVectorTy()) Int = Intrinsic::aarch64_neon_fmax; 5355 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmax"); 5356 case NEON::BI__builtin_neon_vmin_v: 5357 case NEON::BI__builtin_neon_vminq_v: 5358 // FIXME: improve sharing scheme to cope with 3 alternative LLVM intrinsics. 5359 Int = usgn ? Intrinsic::aarch64_neon_umin : Intrinsic::aarch64_neon_smin; 5360 if (Ty->isFPOrFPVectorTy()) Int = Intrinsic::aarch64_neon_fmin; 5361 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmin"); 5362 case NEON::BI__builtin_neon_vabd_v: 5363 case NEON::BI__builtin_neon_vabdq_v: 5364 // FIXME: improve sharing scheme to cope with 3 alternative LLVM intrinsics. 5365 Int = usgn ? Intrinsic::aarch64_neon_uabd : Intrinsic::aarch64_neon_sabd; 5366 if (Ty->isFPOrFPVectorTy()) Int = Intrinsic::aarch64_neon_fabd; 5367 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vabd"); 5368 case NEON::BI__builtin_neon_vpadal_v: 5369 case NEON::BI__builtin_neon_vpadalq_v: { 5370 unsigned ArgElts = VTy->getNumElements(); 5371 llvm::IntegerType *EltTy = cast<IntegerType>(VTy->getElementType()); 5372 unsigned BitWidth = EltTy->getBitWidth(); 5373 llvm::Type *ArgTy = llvm::VectorType::get( 5374 llvm::IntegerType::get(getLLVMContext(), BitWidth/2), 2*ArgElts); 5375 llvm::Type* Tys[2] = { VTy, ArgTy }; 5376 Int = usgn ? Intrinsic::aarch64_neon_uaddlp : Intrinsic::aarch64_neon_saddlp; 5377 SmallVector<llvm::Value*, 1> TmpOps; 5378 TmpOps.push_back(Ops[1]); 5379 Function *F = CGM.getIntrinsic(Int, Tys); 5380 llvm::Value *tmp = EmitNeonCall(F, TmpOps, "vpadal"); 5381 llvm::Value *addend = Builder.CreateBitCast(Ops[0], tmp->getType()); 5382 return Builder.CreateAdd(tmp, addend); 5383 } 5384 case NEON::BI__builtin_neon_vpmin_v: 5385 case NEON::BI__builtin_neon_vpminq_v: 5386 // FIXME: improve sharing scheme to cope with 3 alternative LLVM intrinsics. 5387 Int = usgn ? Intrinsic::aarch64_neon_uminp : Intrinsic::aarch64_neon_sminp; 5388 if (Ty->isFPOrFPVectorTy()) Int = Intrinsic::aarch64_neon_fminp; 5389 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vpmin"); 5390 case NEON::BI__builtin_neon_vpmax_v: 5391 case NEON::BI__builtin_neon_vpmaxq_v: 5392 // FIXME: improve sharing scheme to cope with 3 alternative LLVM intrinsics. 5393 Int = usgn ? Intrinsic::aarch64_neon_umaxp : Intrinsic::aarch64_neon_smaxp; 5394 if (Ty->isFPOrFPVectorTy()) Int = Intrinsic::aarch64_neon_fmaxp; 5395 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vpmax"); 5396 case NEON::BI__builtin_neon_vminnm_v: 5397 case NEON::BI__builtin_neon_vminnmq_v: 5398 Int = Intrinsic::aarch64_neon_fminnm; 5399 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vminnm"); 5400 case NEON::BI__builtin_neon_vmaxnm_v: 5401 case NEON::BI__builtin_neon_vmaxnmq_v: 5402 Int = Intrinsic::aarch64_neon_fmaxnm; 5403 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmaxnm"); 5404 case NEON::BI__builtin_neon_vrecpss_f32: { 5405 Ops.push_back(EmitScalarExpr(E->getArg(1))); 5406 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_frecps, FloatTy), 5407 Ops, "vrecps"); 5408 } 5409 case NEON::BI__builtin_neon_vrecpsd_f64: { 5410 Ops.push_back(EmitScalarExpr(E->getArg(1))); 5411 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_frecps, DoubleTy), 5412 Ops, "vrecps"); 5413 } 5414 case NEON::BI__builtin_neon_vqshrun_n_v: 5415 Int = Intrinsic::aarch64_neon_sqshrun; 5416 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqshrun_n"); 5417 case NEON::BI__builtin_neon_vqrshrun_n_v: 5418 Int = Intrinsic::aarch64_neon_sqrshrun; 5419 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqrshrun_n"); 5420 case NEON::BI__builtin_neon_vqshrn_n_v: 5421 Int = usgn ? Intrinsic::aarch64_neon_uqshrn : Intrinsic::aarch64_neon_sqshrn; 5422 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqshrn_n"); 5423 case NEON::BI__builtin_neon_vrshrn_n_v: 5424 Int = Intrinsic::aarch64_neon_rshrn; 5425 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrshrn_n"); 5426 case NEON::BI__builtin_neon_vqrshrn_n_v: 5427 Int = usgn ? Intrinsic::aarch64_neon_uqrshrn : Intrinsic::aarch64_neon_sqrshrn; 5428 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqrshrn_n"); 5429 case NEON::BI__builtin_neon_vrnda_v: 5430 case NEON::BI__builtin_neon_vrndaq_v: { 5431 Int = Intrinsic::round; 5432 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrnda"); 5433 } 5434 case NEON::BI__builtin_neon_vrndi_v: 5435 case NEON::BI__builtin_neon_vrndiq_v: { 5436 Int = Intrinsic::nearbyint; 5437 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrndi"); 5438 } 5439 case NEON::BI__builtin_neon_vrndm_v: 5440 case NEON::BI__builtin_neon_vrndmq_v: { 5441 Int = Intrinsic::floor; 5442 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrndm"); 5443 } 5444 case NEON::BI__builtin_neon_vrndn_v: 5445 case NEON::BI__builtin_neon_vrndnq_v: { 5446 Int = Intrinsic::aarch64_neon_frintn; 5447 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrndn"); 5448 } 5449 case NEON::BI__builtin_neon_vrndp_v: 5450 case NEON::BI__builtin_neon_vrndpq_v: { 5451 Int = Intrinsic::ceil; 5452 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrndp"); 5453 } 5454 case NEON::BI__builtin_neon_vrndx_v: 5455 case NEON::BI__builtin_neon_vrndxq_v: { 5456 Int = Intrinsic::rint; 5457 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrndx"); 5458 } 5459 case NEON::BI__builtin_neon_vrnd_v: 5460 case NEON::BI__builtin_neon_vrndq_v: { 5461 Int = Intrinsic::trunc; 5462 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrndz"); 5463 } 5464 case NEON::BI__builtin_neon_vceqz_v: 5465 case NEON::BI__builtin_neon_vceqzq_v: 5466 return EmitAArch64CompareBuiltinExpr(Ops[0], Ty, ICmpInst::FCMP_OEQ, 5467 ICmpInst::ICMP_EQ, "vceqz"); 5468 case NEON::BI__builtin_neon_vcgez_v: 5469 case NEON::BI__builtin_neon_vcgezq_v: 5470 return EmitAArch64CompareBuiltinExpr(Ops[0], Ty, ICmpInst::FCMP_OGE, 5471 ICmpInst::ICMP_SGE, "vcgez"); 5472 case NEON::BI__builtin_neon_vclez_v: 5473 case NEON::BI__builtin_neon_vclezq_v: 5474 return EmitAArch64CompareBuiltinExpr(Ops[0], Ty, ICmpInst::FCMP_OLE, 5475 ICmpInst::ICMP_SLE, "vclez"); 5476 case NEON::BI__builtin_neon_vcgtz_v: 5477 case NEON::BI__builtin_neon_vcgtzq_v: 5478 return EmitAArch64CompareBuiltinExpr(Ops[0], Ty, ICmpInst::FCMP_OGT, 5479 ICmpInst::ICMP_SGT, "vcgtz"); 5480 case NEON::BI__builtin_neon_vcltz_v: 5481 case NEON::BI__builtin_neon_vcltzq_v: 5482 return EmitAArch64CompareBuiltinExpr(Ops[0], Ty, ICmpInst::FCMP_OLT, 5483 ICmpInst::ICMP_SLT, "vcltz"); 5484 case NEON::BI__builtin_neon_vcvt_f64_v: 5485 case NEON::BI__builtin_neon_vcvtq_f64_v: 5486 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 5487 Ty = GetNeonType(this, NeonTypeFlags(NeonTypeFlags::Float64, false, quad)); 5488 return usgn ? Builder.CreateUIToFP(Ops[0], Ty, "vcvt") 5489 : Builder.CreateSIToFP(Ops[0], Ty, "vcvt"); 5490 case NEON::BI__builtin_neon_vcvt_f64_f32: { 5491 assert(Type.getEltType() == NeonTypeFlags::Float64 && quad && 5492 "unexpected vcvt_f64_f32 builtin"); 5493 NeonTypeFlags SrcFlag = NeonTypeFlags(NeonTypeFlags::Float32, false, false); 5494 Ops[0] = Builder.CreateBitCast(Ops[0], GetNeonType(this, SrcFlag)); 5495 5496 return Builder.CreateFPExt(Ops[0], Ty, "vcvt"); 5497 } 5498 case NEON::BI__builtin_neon_vcvt_f32_f64: { 5499 assert(Type.getEltType() == NeonTypeFlags::Float32 && 5500 "unexpected vcvt_f32_f64 builtin"); 5501 NeonTypeFlags SrcFlag = NeonTypeFlags(NeonTypeFlags::Float64, false, true); 5502 Ops[0] = Builder.CreateBitCast(Ops[0], GetNeonType(this, SrcFlag)); 5503 5504 return Builder.CreateFPTrunc(Ops[0], Ty, "vcvt"); 5505 } 5506 case NEON::BI__builtin_neon_vcvt_s32_v: 5507 case NEON::BI__builtin_neon_vcvt_u32_v: 5508 case NEON::BI__builtin_neon_vcvt_s64_v: 5509 case NEON::BI__builtin_neon_vcvt_u64_v: 5510 case NEON::BI__builtin_neon_vcvtq_s32_v: 5511 case NEON::BI__builtin_neon_vcvtq_u32_v: 5512 case NEON::BI__builtin_neon_vcvtq_s64_v: 5513 case NEON::BI__builtin_neon_vcvtq_u64_v: { 5514 Ops[0] = Builder.CreateBitCast(Ops[0], GetFloatNeonType(this, Type)); 5515 if (usgn) 5516 return Builder.CreateFPToUI(Ops[0], Ty); 5517 return Builder.CreateFPToSI(Ops[0], Ty); 5518 } 5519 case NEON::BI__builtin_neon_vcvta_s32_v: 5520 case NEON::BI__builtin_neon_vcvtaq_s32_v: 5521 case NEON::BI__builtin_neon_vcvta_u32_v: 5522 case NEON::BI__builtin_neon_vcvtaq_u32_v: 5523 case NEON::BI__builtin_neon_vcvta_s64_v: 5524 case NEON::BI__builtin_neon_vcvtaq_s64_v: 5525 case NEON::BI__builtin_neon_vcvta_u64_v: 5526 case NEON::BI__builtin_neon_vcvtaq_u64_v: { 5527 Int = usgn ? Intrinsic::aarch64_neon_fcvtau : Intrinsic::aarch64_neon_fcvtas; 5528 llvm::Type *Tys[2] = { Ty, GetFloatNeonType(this, Type) }; 5529 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vcvta"); 5530 } 5531 case NEON::BI__builtin_neon_vcvtm_s32_v: 5532 case NEON::BI__builtin_neon_vcvtmq_s32_v: 5533 case NEON::BI__builtin_neon_vcvtm_u32_v: 5534 case NEON::BI__builtin_neon_vcvtmq_u32_v: 5535 case NEON::BI__builtin_neon_vcvtm_s64_v: 5536 case NEON::BI__builtin_neon_vcvtmq_s64_v: 5537 case NEON::BI__builtin_neon_vcvtm_u64_v: 5538 case NEON::BI__builtin_neon_vcvtmq_u64_v: { 5539 Int = usgn ? Intrinsic::aarch64_neon_fcvtmu : Intrinsic::aarch64_neon_fcvtms; 5540 llvm::Type *Tys[2] = { Ty, GetFloatNeonType(this, Type) }; 5541 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vcvtm"); 5542 } 5543 case NEON::BI__builtin_neon_vcvtn_s32_v: 5544 case NEON::BI__builtin_neon_vcvtnq_s32_v: 5545 case NEON::BI__builtin_neon_vcvtn_u32_v: 5546 case NEON::BI__builtin_neon_vcvtnq_u32_v: 5547 case NEON::BI__builtin_neon_vcvtn_s64_v: 5548 case NEON::BI__builtin_neon_vcvtnq_s64_v: 5549 case NEON::BI__builtin_neon_vcvtn_u64_v: 5550 case NEON::BI__builtin_neon_vcvtnq_u64_v: { 5551 Int = usgn ? Intrinsic::aarch64_neon_fcvtnu : Intrinsic::aarch64_neon_fcvtns; 5552 llvm::Type *Tys[2] = { Ty, GetFloatNeonType(this, Type) }; 5553 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vcvtn"); 5554 } 5555 case NEON::BI__builtin_neon_vcvtp_s32_v: 5556 case NEON::BI__builtin_neon_vcvtpq_s32_v: 5557 case NEON::BI__builtin_neon_vcvtp_u32_v: 5558 case NEON::BI__builtin_neon_vcvtpq_u32_v: 5559 case NEON::BI__builtin_neon_vcvtp_s64_v: 5560 case NEON::BI__builtin_neon_vcvtpq_s64_v: 5561 case NEON::BI__builtin_neon_vcvtp_u64_v: 5562 case NEON::BI__builtin_neon_vcvtpq_u64_v: { 5563 Int = usgn ? Intrinsic::aarch64_neon_fcvtpu : Intrinsic::aarch64_neon_fcvtps; 5564 llvm::Type *Tys[2] = { Ty, GetFloatNeonType(this, Type) }; 5565 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vcvtp"); 5566 } 5567 case NEON::BI__builtin_neon_vmulx_v: 5568 case NEON::BI__builtin_neon_vmulxq_v: { 5569 Int = Intrinsic::aarch64_neon_fmulx; 5570 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmulx"); 5571 } 5572 case NEON::BI__builtin_neon_vmul_lane_v: 5573 case NEON::BI__builtin_neon_vmul_laneq_v: { 5574 // v1f64 vmul_lane should be mapped to Neon scalar mul lane 5575 bool Quad = false; 5576 if (BuiltinID == NEON::BI__builtin_neon_vmul_laneq_v) 5577 Quad = true; 5578 Ops[0] = Builder.CreateBitCast(Ops[0], DoubleTy); 5579 llvm::Type *VTy = GetNeonType(this, 5580 NeonTypeFlags(NeonTypeFlags::Float64, false, Quad)); 5581 Ops[1] = Builder.CreateBitCast(Ops[1], VTy); 5582 Ops[1] = Builder.CreateExtractElement(Ops[1], Ops[2], "extract"); 5583 Value *Result = Builder.CreateFMul(Ops[0], Ops[1]); 5584 return Builder.CreateBitCast(Result, Ty); 5585 } 5586 case NEON::BI__builtin_neon_vnegd_s64: 5587 return Builder.CreateNeg(EmitScalarExpr(E->getArg(0)), "vnegd"); 5588 case NEON::BI__builtin_neon_vpmaxnm_v: 5589 case NEON::BI__builtin_neon_vpmaxnmq_v: { 5590 Int = Intrinsic::aarch64_neon_fmaxnmp; 5591 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vpmaxnm"); 5592 } 5593 case NEON::BI__builtin_neon_vpminnm_v: 5594 case NEON::BI__builtin_neon_vpminnmq_v: { 5595 Int = Intrinsic::aarch64_neon_fminnmp; 5596 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vpminnm"); 5597 } 5598 case NEON::BI__builtin_neon_vsqrt_v: 5599 case NEON::BI__builtin_neon_vsqrtq_v: { 5600 Int = Intrinsic::sqrt; 5601 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 5602 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vsqrt"); 5603 } 5604 case NEON::BI__builtin_neon_vrbit_v: 5605 case NEON::BI__builtin_neon_vrbitq_v: { 5606 Int = Intrinsic::aarch64_neon_rbit; 5607 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrbit"); 5608 } 5609 case NEON::BI__builtin_neon_vaddv_u8: 5610 // FIXME: These are handled by the AArch64 scalar code. 5611 usgn = true; 5612 // FALLTHROUGH 5613 case NEON::BI__builtin_neon_vaddv_s8: { 5614 Int = usgn ? Intrinsic::aarch64_neon_uaddv : Intrinsic::aarch64_neon_saddv; 5615 Ty = Int32Ty; 5616 VTy = llvm::VectorType::get(Int8Ty, 8); 5617 llvm::Type *Tys[2] = { Ty, VTy }; 5618 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5619 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddv"); 5620 return Builder.CreateTrunc(Ops[0], Int8Ty); 5621 } 5622 case NEON::BI__builtin_neon_vaddv_u16: 5623 usgn = true; 5624 // FALLTHROUGH 5625 case NEON::BI__builtin_neon_vaddv_s16: { 5626 Int = usgn ? Intrinsic::aarch64_neon_uaddv : Intrinsic::aarch64_neon_saddv; 5627 Ty = Int32Ty; 5628 VTy = llvm::VectorType::get(Int16Ty, 4); 5629 llvm::Type *Tys[2] = { Ty, VTy }; 5630 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5631 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddv"); 5632 return Builder.CreateTrunc(Ops[0], Int16Ty); 5633 } 5634 case NEON::BI__builtin_neon_vaddvq_u8: 5635 usgn = true; 5636 // FALLTHROUGH 5637 case NEON::BI__builtin_neon_vaddvq_s8: { 5638 Int = usgn ? Intrinsic::aarch64_neon_uaddv : Intrinsic::aarch64_neon_saddv; 5639 Ty = Int32Ty; 5640 VTy = llvm::VectorType::get(Int8Ty, 16); 5641 llvm::Type *Tys[2] = { Ty, VTy }; 5642 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5643 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddv"); 5644 return Builder.CreateTrunc(Ops[0], Int8Ty); 5645 } 5646 case NEON::BI__builtin_neon_vaddvq_u16: 5647 usgn = true; 5648 // FALLTHROUGH 5649 case NEON::BI__builtin_neon_vaddvq_s16: { 5650 Int = usgn ? Intrinsic::aarch64_neon_uaddv : Intrinsic::aarch64_neon_saddv; 5651 Ty = Int32Ty; 5652 VTy = llvm::VectorType::get(Int16Ty, 8); 5653 llvm::Type *Tys[2] = { Ty, VTy }; 5654 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5655 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddv"); 5656 return Builder.CreateTrunc(Ops[0], Int16Ty); 5657 } 5658 case NEON::BI__builtin_neon_vmaxv_u8: { 5659 Int = Intrinsic::aarch64_neon_umaxv; 5660 Ty = Int32Ty; 5661 VTy = llvm::VectorType::get(Int8Ty, 8); 5662 llvm::Type *Tys[2] = { Ty, VTy }; 5663 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5664 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv"); 5665 return Builder.CreateTrunc(Ops[0], Int8Ty); 5666 } 5667 case NEON::BI__builtin_neon_vmaxv_u16: { 5668 Int = Intrinsic::aarch64_neon_umaxv; 5669 Ty = Int32Ty; 5670 VTy = llvm::VectorType::get(Int16Ty, 4); 5671 llvm::Type *Tys[2] = { Ty, VTy }; 5672 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5673 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv"); 5674 return Builder.CreateTrunc(Ops[0], Int16Ty); 5675 } 5676 case NEON::BI__builtin_neon_vmaxvq_u8: { 5677 Int = Intrinsic::aarch64_neon_umaxv; 5678 Ty = Int32Ty; 5679 VTy = llvm::VectorType::get(Int8Ty, 16); 5680 llvm::Type *Tys[2] = { Ty, VTy }; 5681 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5682 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv"); 5683 return Builder.CreateTrunc(Ops[0], Int8Ty); 5684 } 5685 case NEON::BI__builtin_neon_vmaxvq_u16: { 5686 Int = Intrinsic::aarch64_neon_umaxv; 5687 Ty = Int32Ty; 5688 VTy = llvm::VectorType::get(Int16Ty, 8); 5689 llvm::Type *Tys[2] = { Ty, VTy }; 5690 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5691 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv"); 5692 return Builder.CreateTrunc(Ops[0], Int16Ty); 5693 } 5694 case NEON::BI__builtin_neon_vmaxv_s8: { 5695 Int = Intrinsic::aarch64_neon_smaxv; 5696 Ty = Int32Ty; 5697 VTy = llvm::VectorType::get(Int8Ty, 8); 5698 llvm::Type *Tys[2] = { Ty, VTy }; 5699 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5700 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv"); 5701 return Builder.CreateTrunc(Ops[0], Int8Ty); 5702 } 5703 case NEON::BI__builtin_neon_vmaxv_s16: { 5704 Int = Intrinsic::aarch64_neon_smaxv; 5705 Ty = Int32Ty; 5706 VTy = llvm::VectorType::get(Int16Ty, 4); 5707 llvm::Type *Tys[2] = { Ty, VTy }; 5708 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5709 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv"); 5710 return Builder.CreateTrunc(Ops[0], Int16Ty); 5711 } 5712 case NEON::BI__builtin_neon_vmaxvq_s8: { 5713 Int = Intrinsic::aarch64_neon_smaxv; 5714 Ty = Int32Ty; 5715 VTy = llvm::VectorType::get(Int8Ty, 16); 5716 llvm::Type *Tys[2] = { Ty, VTy }; 5717 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5718 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv"); 5719 return Builder.CreateTrunc(Ops[0], Int8Ty); 5720 } 5721 case NEON::BI__builtin_neon_vmaxvq_s16: { 5722 Int = Intrinsic::aarch64_neon_smaxv; 5723 Ty = Int32Ty; 5724 VTy = llvm::VectorType::get(Int16Ty, 8); 5725 llvm::Type *Tys[2] = { Ty, VTy }; 5726 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5727 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv"); 5728 return Builder.CreateTrunc(Ops[0], Int16Ty); 5729 } 5730 case NEON::BI__builtin_neon_vminv_u8: { 5731 Int = Intrinsic::aarch64_neon_uminv; 5732 Ty = Int32Ty; 5733 VTy = llvm::VectorType::get(Int8Ty, 8); 5734 llvm::Type *Tys[2] = { Ty, VTy }; 5735 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5736 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv"); 5737 return Builder.CreateTrunc(Ops[0], Int8Ty); 5738 } 5739 case NEON::BI__builtin_neon_vminv_u16: { 5740 Int = Intrinsic::aarch64_neon_uminv; 5741 Ty = Int32Ty; 5742 VTy = llvm::VectorType::get(Int16Ty, 4); 5743 llvm::Type *Tys[2] = { Ty, VTy }; 5744 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5745 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv"); 5746 return Builder.CreateTrunc(Ops[0], Int16Ty); 5747 } 5748 case NEON::BI__builtin_neon_vminvq_u8: { 5749 Int = Intrinsic::aarch64_neon_uminv; 5750 Ty = Int32Ty; 5751 VTy = llvm::VectorType::get(Int8Ty, 16); 5752 llvm::Type *Tys[2] = { Ty, VTy }; 5753 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5754 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv"); 5755 return Builder.CreateTrunc(Ops[0], Int8Ty); 5756 } 5757 case NEON::BI__builtin_neon_vminvq_u16: { 5758 Int = Intrinsic::aarch64_neon_uminv; 5759 Ty = Int32Ty; 5760 VTy = llvm::VectorType::get(Int16Ty, 8); 5761 llvm::Type *Tys[2] = { Ty, VTy }; 5762 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5763 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv"); 5764 return Builder.CreateTrunc(Ops[0], Int16Ty); 5765 } 5766 case NEON::BI__builtin_neon_vminv_s8: { 5767 Int = Intrinsic::aarch64_neon_sminv; 5768 Ty = Int32Ty; 5769 VTy = llvm::VectorType::get(Int8Ty, 8); 5770 llvm::Type *Tys[2] = { Ty, VTy }; 5771 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5772 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv"); 5773 return Builder.CreateTrunc(Ops[0], Int8Ty); 5774 } 5775 case NEON::BI__builtin_neon_vminv_s16: { 5776 Int = Intrinsic::aarch64_neon_sminv; 5777 Ty = Int32Ty; 5778 VTy = llvm::VectorType::get(Int16Ty, 4); 5779 llvm::Type *Tys[2] = { Ty, VTy }; 5780 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5781 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv"); 5782 return Builder.CreateTrunc(Ops[0], Int16Ty); 5783 } 5784 case NEON::BI__builtin_neon_vminvq_s8: { 5785 Int = Intrinsic::aarch64_neon_sminv; 5786 Ty = Int32Ty; 5787 VTy = llvm::VectorType::get(Int8Ty, 16); 5788 llvm::Type *Tys[2] = { Ty, VTy }; 5789 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5790 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv"); 5791 return Builder.CreateTrunc(Ops[0], Int8Ty); 5792 } 5793 case NEON::BI__builtin_neon_vminvq_s16: { 5794 Int = Intrinsic::aarch64_neon_sminv; 5795 Ty = Int32Ty; 5796 VTy = llvm::VectorType::get(Int16Ty, 8); 5797 llvm::Type *Tys[2] = { Ty, VTy }; 5798 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5799 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv"); 5800 return Builder.CreateTrunc(Ops[0], Int16Ty); 5801 } 5802 case NEON::BI__builtin_neon_vmul_n_f64: { 5803 Ops[0] = Builder.CreateBitCast(Ops[0], DoubleTy); 5804 Value *RHS = Builder.CreateBitCast(EmitScalarExpr(E->getArg(1)), DoubleTy); 5805 return Builder.CreateFMul(Ops[0], RHS); 5806 } 5807 case NEON::BI__builtin_neon_vaddlv_u8: { 5808 Int = Intrinsic::aarch64_neon_uaddlv; 5809 Ty = Int32Ty; 5810 VTy = llvm::VectorType::get(Int8Ty, 8); 5811 llvm::Type *Tys[2] = { Ty, VTy }; 5812 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5813 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv"); 5814 return Builder.CreateTrunc(Ops[0], Int16Ty); 5815 } 5816 case NEON::BI__builtin_neon_vaddlv_u16: { 5817 Int = Intrinsic::aarch64_neon_uaddlv; 5818 Ty = Int32Ty; 5819 VTy = llvm::VectorType::get(Int16Ty, 4); 5820 llvm::Type *Tys[2] = { Ty, VTy }; 5821 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5822 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv"); 5823 } 5824 case NEON::BI__builtin_neon_vaddlvq_u8: { 5825 Int = Intrinsic::aarch64_neon_uaddlv; 5826 Ty = Int32Ty; 5827 VTy = llvm::VectorType::get(Int8Ty, 16); 5828 llvm::Type *Tys[2] = { Ty, VTy }; 5829 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5830 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv"); 5831 return Builder.CreateTrunc(Ops[0], Int16Ty); 5832 } 5833 case NEON::BI__builtin_neon_vaddlvq_u16: { 5834 Int = Intrinsic::aarch64_neon_uaddlv; 5835 Ty = Int32Ty; 5836 VTy = llvm::VectorType::get(Int16Ty, 8); 5837 llvm::Type *Tys[2] = { Ty, VTy }; 5838 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5839 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv"); 5840 } 5841 case NEON::BI__builtin_neon_vaddlv_s8: { 5842 Int = Intrinsic::aarch64_neon_saddlv; 5843 Ty = Int32Ty; 5844 VTy = llvm::VectorType::get(Int8Ty, 8); 5845 llvm::Type *Tys[2] = { Ty, VTy }; 5846 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5847 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv"); 5848 return Builder.CreateTrunc(Ops[0], Int16Ty); 5849 } 5850 case NEON::BI__builtin_neon_vaddlv_s16: { 5851 Int = Intrinsic::aarch64_neon_saddlv; 5852 Ty = Int32Ty; 5853 VTy = llvm::VectorType::get(Int16Ty, 4); 5854 llvm::Type *Tys[2] = { Ty, VTy }; 5855 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5856 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv"); 5857 } 5858 case NEON::BI__builtin_neon_vaddlvq_s8: { 5859 Int = Intrinsic::aarch64_neon_saddlv; 5860 Ty = Int32Ty; 5861 VTy = llvm::VectorType::get(Int8Ty, 16); 5862 llvm::Type *Tys[2] = { Ty, VTy }; 5863 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5864 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv"); 5865 return Builder.CreateTrunc(Ops[0], Int16Ty); 5866 } 5867 case NEON::BI__builtin_neon_vaddlvq_s16: { 5868 Int = Intrinsic::aarch64_neon_saddlv; 5869 Ty = Int32Ty; 5870 VTy = llvm::VectorType::get(Int16Ty, 8); 5871 llvm::Type *Tys[2] = { Ty, VTy }; 5872 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5873 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv"); 5874 } 5875 case NEON::BI__builtin_neon_vsri_n_v: 5876 case NEON::BI__builtin_neon_vsriq_n_v: { 5877 Int = Intrinsic::aarch64_neon_vsri; 5878 llvm::Function *Intrin = CGM.getIntrinsic(Int, Ty); 5879 return EmitNeonCall(Intrin, Ops, "vsri_n"); 5880 } 5881 case NEON::BI__builtin_neon_vsli_n_v: 5882 case NEON::BI__builtin_neon_vsliq_n_v: { 5883 Int = Intrinsic::aarch64_neon_vsli; 5884 llvm::Function *Intrin = CGM.getIntrinsic(Int, Ty); 5885 return EmitNeonCall(Intrin, Ops, "vsli_n"); 5886 } 5887 case NEON::BI__builtin_neon_vsra_n_v: 5888 case NEON::BI__builtin_neon_vsraq_n_v: 5889 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 5890 Ops[1] = EmitNeonRShiftImm(Ops[1], Ops[2], Ty, usgn, "vsra_n"); 5891 return Builder.CreateAdd(Ops[0], Ops[1]); 5892 case NEON::BI__builtin_neon_vrsra_n_v: 5893 case NEON::BI__builtin_neon_vrsraq_n_v: { 5894 Int = usgn ? Intrinsic::aarch64_neon_urshl : Intrinsic::aarch64_neon_srshl; 5895 SmallVector<llvm::Value*,2> TmpOps; 5896 TmpOps.push_back(Ops[1]); 5897 TmpOps.push_back(Ops[2]); 5898 Function* F = CGM.getIntrinsic(Int, Ty); 5899 llvm::Value *tmp = EmitNeonCall(F, TmpOps, "vrshr_n", 1, true); 5900 Ops[0] = Builder.CreateBitCast(Ops[0], VTy); 5901 return Builder.CreateAdd(Ops[0], tmp); 5902 } 5903 // FIXME: Sharing loads & stores with 32-bit is complicated by the absence 5904 // of an Align parameter here. 5905 case NEON::BI__builtin_neon_vld1_x2_v: 5906 case NEON::BI__builtin_neon_vld1q_x2_v: 5907 case NEON::BI__builtin_neon_vld1_x3_v: 5908 case NEON::BI__builtin_neon_vld1q_x3_v: 5909 case NEON::BI__builtin_neon_vld1_x4_v: 5910 case NEON::BI__builtin_neon_vld1q_x4_v: { 5911 llvm::Type *PTy = llvm::PointerType::getUnqual(VTy->getVectorElementType()); 5912 Ops[1] = Builder.CreateBitCast(Ops[1], PTy); 5913 llvm::Type *Tys[2] = { VTy, PTy }; 5914 unsigned Int; 5915 switch (BuiltinID) { 5916 case NEON::BI__builtin_neon_vld1_x2_v: 5917 case NEON::BI__builtin_neon_vld1q_x2_v: 5918 Int = Intrinsic::aarch64_neon_ld1x2; 5919 break; 5920 case NEON::BI__builtin_neon_vld1_x3_v: 5921 case NEON::BI__builtin_neon_vld1q_x3_v: 5922 Int = Intrinsic::aarch64_neon_ld1x3; 5923 break; 5924 case NEON::BI__builtin_neon_vld1_x4_v: 5925 case NEON::BI__builtin_neon_vld1q_x4_v: 5926 Int = Intrinsic::aarch64_neon_ld1x4; 5927 break; 5928 } 5929 Function *F = CGM.getIntrinsic(Int, Tys); 5930 Ops[1] = Builder.CreateCall(F, Ops[1], "vld1xN"); 5931 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 5932 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 5933 return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]); 5934 } 5935 case NEON::BI__builtin_neon_vst1_x2_v: 5936 case NEON::BI__builtin_neon_vst1q_x2_v: 5937 case NEON::BI__builtin_neon_vst1_x3_v: 5938 case NEON::BI__builtin_neon_vst1q_x3_v: 5939 case NEON::BI__builtin_neon_vst1_x4_v: 5940 case NEON::BI__builtin_neon_vst1q_x4_v: { 5941 llvm::Type *PTy = llvm::PointerType::getUnqual(VTy->getVectorElementType()); 5942 llvm::Type *Tys[2] = { VTy, PTy }; 5943 unsigned Int; 5944 switch (BuiltinID) { 5945 case NEON::BI__builtin_neon_vst1_x2_v: 5946 case NEON::BI__builtin_neon_vst1q_x2_v: 5947 Int = Intrinsic::aarch64_neon_st1x2; 5948 break; 5949 case NEON::BI__builtin_neon_vst1_x3_v: 5950 case NEON::BI__builtin_neon_vst1q_x3_v: 5951 Int = Intrinsic::aarch64_neon_st1x3; 5952 break; 5953 case NEON::BI__builtin_neon_vst1_x4_v: 5954 case NEON::BI__builtin_neon_vst1q_x4_v: 5955 Int = Intrinsic::aarch64_neon_st1x4; 5956 break; 5957 } 5958 std::rotate(Ops.begin(), Ops.begin() + 1, Ops.end()); 5959 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, ""); 5960 } 5961 case NEON::BI__builtin_neon_vld1_v: 5962 case NEON::BI__builtin_neon_vld1q_v: 5963 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(VTy)); 5964 return Builder.CreateDefaultAlignedLoad(Ops[0]); 5965 case NEON::BI__builtin_neon_vst1_v: 5966 case NEON::BI__builtin_neon_vst1q_v: 5967 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(VTy)); 5968 Ops[1] = Builder.CreateBitCast(Ops[1], VTy); 5969 return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]); 5970 case NEON::BI__builtin_neon_vld1_lane_v: 5971 case NEON::BI__builtin_neon_vld1q_lane_v: 5972 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 5973 Ty = llvm::PointerType::getUnqual(VTy->getElementType()); 5974 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 5975 Ops[0] = Builder.CreateDefaultAlignedLoad(Ops[0]); 5976 return Builder.CreateInsertElement(Ops[1], Ops[0], Ops[2], "vld1_lane"); 5977 case NEON::BI__builtin_neon_vld1_dup_v: 5978 case NEON::BI__builtin_neon_vld1q_dup_v: { 5979 Value *V = UndefValue::get(Ty); 5980 Ty = llvm::PointerType::getUnqual(VTy->getElementType()); 5981 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 5982 Ops[0] = Builder.CreateDefaultAlignedLoad(Ops[0]); 5983 llvm::Constant *CI = ConstantInt::get(Int32Ty, 0); 5984 Ops[0] = Builder.CreateInsertElement(V, Ops[0], CI); 5985 return EmitNeonSplat(Ops[0], CI); 5986 } 5987 case NEON::BI__builtin_neon_vst1_lane_v: 5988 case NEON::BI__builtin_neon_vst1q_lane_v: 5989 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 5990 Ops[1] = Builder.CreateExtractElement(Ops[1], Ops[2]); 5991 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 5992 return Builder.CreateDefaultAlignedStore(Ops[1], 5993 Builder.CreateBitCast(Ops[0], Ty)); 5994 case NEON::BI__builtin_neon_vld2_v: 5995 case NEON::BI__builtin_neon_vld2q_v: { 5996 llvm::Type *PTy = llvm::PointerType::getUnqual(VTy); 5997 Ops[1] = Builder.CreateBitCast(Ops[1], PTy); 5998 llvm::Type *Tys[2] = { VTy, PTy }; 5999 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld2, Tys); 6000 Ops[1] = Builder.CreateCall(F, Ops[1], "vld2"); 6001 Ops[0] = Builder.CreateBitCast(Ops[0], 6002 llvm::PointerType::getUnqual(Ops[1]->getType())); 6003 return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]); 6004 } 6005 case NEON::BI__builtin_neon_vld3_v: 6006 case NEON::BI__builtin_neon_vld3q_v: { 6007 llvm::Type *PTy = llvm::PointerType::getUnqual(VTy); 6008 Ops[1] = Builder.CreateBitCast(Ops[1], PTy); 6009 llvm::Type *Tys[2] = { VTy, PTy }; 6010 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld3, Tys); 6011 Ops[1] = Builder.CreateCall(F, Ops[1], "vld3"); 6012 Ops[0] = Builder.CreateBitCast(Ops[0], 6013 llvm::PointerType::getUnqual(Ops[1]->getType())); 6014 return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]); 6015 } 6016 case NEON::BI__builtin_neon_vld4_v: 6017 case NEON::BI__builtin_neon_vld4q_v: { 6018 llvm::Type *PTy = llvm::PointerType::getUnqual(VTy); 6019 Ops[1] = Builder.CreateBitCast(Ops[1], PTy); 6020 llvm::Type *Tys[2] = { VTy, PTy }; 6021 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld4, Tys); 6022 Ops[1] = Builder.CreateCall(F, Ops[1], "vld4"); 6023 Ops[0] = Builder.CreateBitCast(Ops[0], 6024 llvm::PointerType::getUnqual(Ops[1]->getType())); 6025 return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]); 6026 } 6027 case NEON::BI__builtin_neon_vld2_dup_v: 6028 case NEON::BI__builtin_neon_vld2q_dup_v: { 6029 llvm::Type *PTy = 6030 llvm::PointerType::getUnqual(VTy->getElementType()); 6031 Ops[1] = Builder.CreateBitCast(Ops[1], PTy); 6032 llvm::Type *Tys[2] = { VTy, PTy }; 6033 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld2r, Tys); 6034 Ops[1] = Builder.CreateCall(F, Ops[1], "vld2"); 6035 Ops[0] = Builder.CreateBitCast(Ops[0], 6036 llvm::PointerType::getUnqual(Ops[1]->getType())); 6037 return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]); 6038 } 6039 case NEON::BI__builtin_neon_vld3_dup_v: 6040 case NEON::BI__builtin_neon_vld3q_dup_v: { 6041 llvm::Type *PTy = 6042 llvm::PointerType::getUnqual(VTy->getElementType()); 6043 Ops[1] = Builder.CreateBitCast(Ops[1], PTy); 6044 llvm::Type *Tys[2] = { VTy, PTy }; 6045 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld3r, Tys); 6046 Ops[1] = Builder.CreateCall(F, Ops[1], "vld3"); 6047 Ops[0] = Builder.CreateBitCast(Ops[0], 6048 llvm::PointerType::getUnqual(Ops[1]->getType())); 6049 return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]); 6050 } 6051 case NEON::BI__builtin_neon_vld4_dup_v: 6052 case NEON::BI__builtin_neon_vld4q_dup_v: { 6053 llvm::Type *PTy = 6054 llvm::PointerType::getUnqual(VTy->getElementType()); 6055 Ops[1] = Builder.CreateBitCast(Ops[1], PTy); 6056 llvm::Type *Tys[2] = { VTy, PTy }; 6057 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld4r, Tys); 6058 Ops[1] = Builder.CreateCall(F, Ops[1], "vld4"); 6059 Ops[0] = Builder.CreateBitCast(Ops[0], 6060 llvm::PointerType::getUnqual(Ops[1]->getType())); 6061 return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]); 6062 } 6063 case NEON::BI__builtin_neon_vld2_lane_v: 6064 case NEON::BI__builtin_neon_vld2q_lane_v: { 6065 llvm::Type *Tys[2] = { VTy, Ops[1]->getType() }; 6066 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld2lane, Tys); 6067 Ops.push_back(Ops[1]); 6068 Ops.erase(Ops.begin()+1); 6069 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 6070 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 6071 Ops[3] = Builder.CreateZExt(Ops[3], Int64Ty); 6072 Ops[1] = Builder.CreateCall(F, makeArrayRef(Ops).slice(1), "vld2_lane"); 6073 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 6074 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 6075 return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]); 6076 } 6077 case NEON::BI__builtin_neon_vld3_lane_v: 6078 case NEON::BI__builtin_neon_vld3q_lane_v: { 6079 llvm::Type *Tys[2] = { VTy, Ops[1]->getType() }; 6080 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld3lane, Tys); 6081 Ops.push_back(Ops[1]); 6082 Ops.erase(Ops.begin()+1); 6083 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 6084 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 6085 Ops[3] = Builder.CreateBitCast(Ops[3], Ty); 6086 Ops[4] = Builder.CreateZExt(Ops[4], Int64Ty); 6087 Ops[1] = Builder.CreateCall(F, makeArrayRef(Ops).slice(1), "vld3_lane"); 6088 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 6089 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 6090 return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]); 6091 } 6092 case NEON::BI__builtin_neon_vld4_lane_v: 6093 case NEON::BI__builtin_neon_vld4q_lane_v: { 6094 llvm::Type *Tys[2] = { VTy, Ops[1]->getType() }; 6095 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld4lane, Tys); 6096 Ops.push_back(Ops[1]); 6097 Ops.erase(Ops.begin()+1); 6098 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 6099 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 6100 Ops[3] = Builder.CreateBitCast(Ops[3], Ty); 6101 Ops[4] = Builder.CreateBitCast(Ops[4], Ty); 6102 Ops[5] = Builder.CreateZExt(Ops[5], Int64Ty); 6103 Ops[1] = Builder.CreateCall(F, makeArrayRef(Ops).slice(1), "vld4_lane"); 6104 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 6105 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 6106 return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]); 6107 } 6108 case NEON::BI__builtin_neon_vst2_v: 6109 case NEON::BI__builtin_neon_vst2q_v: { 6110 Ops.push_back(Ops[0]); 6111 Ops.erase(Ops.begin()); 6112 llvm::Type *Tys[2] = { VTy, Ops[2]->getType() }; 6113 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_st2, Tys), 6114 Ops, ""); 6115 } 6116 case NEON::BI__builtin_neon_vst2_lane_v: 6117 case NEON::BI__builtin_neon_vst2q_lane_v: { 6118 Ops.push_back(Ops[0]); 6119 Ops.erase(Ops.begin()); 6120 Ops[2] = Builder.CreateZExt(Ops[2], Int64Ty); 6121 llvm::Type *Tys[2] = { VTy, Ops[3]->getType() }; 6122 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_st2lane, Tys), 6123 Ops, ""); 6124 } 6125 case NEON::BI__builtin_neon_vst3_v: 6126 case NEON::BI__builtin_neon_vst3q_v: { 6127 Ops.push_back(Ops[0]); 6128 Ops.erase(Ops.begin()); 6129 llvm::Type *Tys[2] = { VTy, Ops[3]->getType() }; 6130 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_st3, Tys), 6131 Ops, ""); 6132 } 6133 case NEON::BI__builtin_neon_vst3_lane_v: 6134 case NEON::BI__builtin_neon_vst3q_lane_v: { 6135 Ops.push_back(Ops[0]); 6136 Ops.erase(Ops.begin()); 6137 Ops[3] = Builder.CreateZExt(Ops[3], Int64Ty); 6138 llvm::Type *Tys[2] = { VTy, Ops[4]->getType() }; 6139 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_st3lane, Tys), 6140 Ops, ""); 6141 } 6142 case NEON::BI__builtin_neon_vst4_v: 6143 case NEON::BI__builtin_neon_vst4q_v: { 6144 Ops.push_back(Ops[0]); 6145 Ops.erase(Ops.begin()); 6146 llvm::Type *Tys[2] = { VTy, Ops[4]->getType() }; 6147 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_st4, Tys), 6148 Ops, ""); 6149 } 6150 case NEON::BI__builtin_neon_vst4_lane_v: 6151 case NEON::BI__builtin_neon_vst4q_lane_v: { 6152 Ops.push_back(Ops[0]); 6153 Ops.erase(Ops.begin()); 6154 Ops[4] = Builder.CreateZExt(Ops[4], Int64Ty); 6155 llvm::Type *Tys[2] = { VTy, Ops[5]->getType() }; 6156 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_st4lane, Tys), 6157 Ops, ""); 6158 } 6159 case NEON::BI__builtin_neon_vtrn_v: 6160 case NEON::BI__builtin_neon_vtrnq_v: { 6161 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty)); 6162 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 6163 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 6164 Value *SV = nullptr; 6165 6166 for (unsigned vi = 0; vi != 2; ++vi) { 6167 SmallVector<Constant*, 16> Indices; 6168 for (unsigned i = 0, e = VTy->getNumElements(); i != e; i += 2) { 6169 Indices.push_back(ConstantInt::get(Int32Ty, i+vi)); 6170 Indices.push_back(ConstantInt::get(Int32Ty, i+e+vi)); 6171 } 6172 Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ty, Ops[0], vi); 6173 SV = llvm::ConstantVector::get(Indices); 6174 SV = Builder.CreateShuffleVector(Ops[1], Ops[2], SV, "vtrn"); 6175 SV = Builder.CreateDefaultAlignedStore(SV, Addr); 6176 } 6177 return SV; 6178 } 6179 case NEON::BI__builtin_neon_vuzp_v: 6180 case NEON::BI__builtin_neon_vuzpq_v: { 6181 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty)); 6182 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 6183 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 6184 Value *SV = nullptr; 6185 6186 for (unsigned vi = 0; vi != 2; ++vi) { 6187 SmallVector<Constant*, 16> Indices; 6188 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) 6189 Indices.push_back(ConstantInt::get(Int32Ty, 2*i+vi)); 6190 6191 Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ty, Ops[0], vi); 6192 SV = llvm::ConstantVector::get(Indices); 6193 SV = Builder.CreateShuffleVector(Ops[1], Ops[2], SV, "vuzp"); 6194 SV = Builder.CreateDefaultAlignedStore(SV, Addr); 6195 } 6196 return SV; 6197 } 6198 case NEON::BI__builtin_neon_vzip_v: 6199 case NEON::BI__builtin_neon_vzipq_v: { 6200 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty)); 6201 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 6202 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 6203 Value *SV = nullptr; 6204 6205 for (unsigned vi = 0; vi != 2; ++vi) { 6206 SmallVector<Constant*, 16> Indices; 6207 for (unsigned i = 0, e = VTy->getNumElements(); i != e; i += 2) { 6208 Indices.push_back(ConstantInt::get(Int32Ty, (i + vi*e) >> 1)); 6209 Indices.push_back(ConstantInt::get(Int32Ty, ((i + vi*e) >> 1)+e)); 6210 } 6211 Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ty, Ops[0], vi); 6212 SV = llvm::ConstantVector::get(Indices); 6213 SV = Builder.CreateShuffleVector(Ops[1], Ops[2], SV, "vzip"); 6214 SV = Builder.CreateDefaultAlignedStore(SV, Addr); 6215 } 6216 return SV; 6217 } 6218 case NEON::BI__builtin_neon_vqtbl1q_v: { 6219 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbl1, Ty), 6220 Ops, "vtbl1"); 6221 } 6222 case NEON::BI__builtin_neon_vqtbl2q_v: { 6223 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbl2, Ty), 6224 Ops, "vtbl2"); 6225 } 6226 case NEON::BI__builtin_neon_vqtbl3q_v: { 6227 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbl3, Ty), 6228 Ops, "vtbl3"); 6229 } 6230 case NEON::BI__builtin_neon_vqtbl4q_v: { 6231 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbl4, Ty), 6232 Ops, "vtbl4"); 6233 } 6234 case NEON::BI__builtin_neon_vqtbx1q_v: { 6235 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbx1, Ty), 6236 Ops, "vtbx1"); 6237 } 6238 case NEON::BI__builtin_neon_vqtbx2q_v: { 6239 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbx2, Ty), 6240 Ops, "vtbx2"); 6241 } 6242 case NEON::BI__builtin_neon_vqtbx3q_v: { 6243 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbx3, Ty), 6244 Ops, "vtbx3"); 6245 } 6246 case NEON::BI__builtin_neon_vqtbx4q_v: { 6247 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbx4, Ty), 6248 Ops, "vtbx4"); 6249 } 6250 case NEON::BI__builtin_neon_vsqadd_v: 6251 case NEON::BI__builtin_neon_vsqaddq_v: { 6252 Int = Intrinsic::aarch64_neon_usqadd; 6253 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vsqadd"); 6254 } 6255 case NEON::BI__builtin_neon_vuqadd_v: 6256 case NEON::BI__builtin_neon_vuqaddq_v: { 6257 Int = Intrinsic::aarch64_neon_suqadd; 6258 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vuqadd"); 6259 } 6260 } 6261 } 6262 6263 llvm::Value *CodeGenFunction:: 6264 BuildVector(ArrayRef<llvm::Value*> Ops) { 6265 assert((Ops.size() & (Ops.size() - 1)) == 0 && 6266 "Not a power-of-two sized vector!"); 6267 bool AllConstants = true; 6268 for (unsigned i = 0, e = Ops.size(); i != e && AllConstants; ++i) 6269 AllConstants &= isa<Constant>(Ops[i]); 6270 6271 // If this is a constant vector, create a ConstantVector. 6272 if (AllConstants) { 6273 SmallVector<llvm::Constant*, 16> CstOps; 6274 for (unsigned i = 0, e = Ops.size(); i != e; ++i) 6275 CstOps.push_back(cast<Constant>(Ops[i])); 6276 return llvm::ConstantVector::get(CstOps); 6277 } 6278 6279 // Otherwise, insertelement the values to build the vector. 6280 Value *Result = 6281 llvm::UndefValue::get(llvm::VectorType::get(Ops[0]->getType(), Ops.size())); 6282 6283 for (unsigned i = 0, e = Ops.size(); i != e; ++i) 6284 Result = Builder.CreateInsertElement(Result, Ops[i], Builder.getInt32(i)); 6285 6286 return Result; 6287 } 6288 6289 Value *CodeGenFunction::EmitX86BuiltinExpr(unsigned BuiltinID, 6290 const CallExpr *E) { 6291 if (BuiltinID == X86::BI__builtin_ms_va_start || 6292 BuiltinID == X86::BI__builtin_ms_va_end) 6293 return EmitVAStartEnd(EmitMSVAListRef(E->getArg(0)).getPointer(), 6294 BuiltinID == X86::BI__builtin_ms_va_start); 6295 if (BuiltinID == X86::BI__builtin_ms_va_copy) { 6296 // Lower this manually. We can't reliably determine whether or not any 6297 // given va_copy() is for a Win64 va_list from the calling convention 6298 // alone, because it's legal to do this from a System V ABI function. 6299 // With opaque pointer types, we won't have enough information in LLVM 6300 // IR to determine this from the argument types, either. Best to do it 6301 // now, while we have enough information. 6302 Address DestAddr = EmitMSVAListRef(E->getArg(0)); 6303 Address SrcAddr = EmitMSVAListRef(E->getArg(1)); 6304 6305 llvm::Type *BPP = Int8PtrPtrTy; 6306 6307 DestAddr = Address(Builder.CreateBitCast(DestAddr.getPointer(), BPP, "cp"), 6308 DestAddr.getAlignment()); 6309 SrcAddr = Address(Builder.CreateBitCast(SrcAddr.getPointer(), BPP, "ap"), 6310 SrcAddr.getAlignment()); 6311 6312 Value *ArgPtr = Builder.CreateLoad(SrcAddr, "ap.val"); 6313 return Builder.CreateStore(ArgPtr, DestAddr); 6314 } 6315 6316 SmallVector<Value*, 4> Ops; 6317 6318 // Find out if any arguments are required to be integer constant expressions. 6319 unsigned ICEArguments = 0; 6320 ASTContext::GetBuiltinTypeError Error; 6321 getContext().GetBuiltinType(BuiltinID, Error, &ICEArguments); 6322 assert(Error == ASTContext::GE_None && "Should not codegen an error"); 6323 6324 for (unsigned i = 0, e = E->getNumArgs(); i != e; i++) { 6325 // If this is a normal argument, just emit it as a scalar. 6326 if ((ICEArguments & (1 << i)) == 0) { 6327 Ops.push_back(EmitScalarExpr(E->getArg(i))); 6328 continue; 6329 } 6330 6331 // If this is required to be a constant, constant fold it so that we know 6332 // that the generated intrinsic gets a ConstantInt. 6333 llvm::APSInt Result; 6334 bool IsConst = E->getArg(i)->isIntegerConstantExpr(Result, getContext()); 6335 assert(IsConst && "Constant arg isn't actually constant?"); (void)IsConst; 6336 Ops.push_back(llvm::ConstantInt::get(getLLVMContext(), Result)); 6337 } 6338 6339 switch (BuiltinID) { 6340 default: return nullptr; 6341 case X86::BI__builtin_cpu_supports: { 6342 const Expr *FeatureExpr = E->getArg(0)->IgnoreParenCasts(); 6343 StringRef FeatureStr = cast<StringLiteral>(FeatureExpr)->getString(); 6344 6345 // TODO: When/if this becomes more than x86 specific then use a TargetInfo 6346 // based mapping. 6347 // Processor features and mapping to processor feature value. 6348 enum X86Features { 6349 CMOV = 0, 6350 MMX, 6351 POPCNT, 6352 SSE, 6353 SSE2, 6354 SSE3, 6355 SSSE3, 6356 SSE4_1, 6357 SSE4_2, 6358 AVX, 6359 AVX2, 6360 SSE4_A, 6361 FMA4, 6362 XOP, 6363 FMA, 6364 AVX512F, 6365 BMI, 6366 BMI2, 6367 MAX 6368 }; 6369 6370 X86Features Feature = StringSwitch<X86Features>(FeatureStr) 6371 .Case("cmov", X86Features::CMOV) 6372 .Case("mmx", X86Features::MMX) 6373 .Case("popcnt", X86Features::POPCNT) 6374 .Case("sse", X86Features::SSE) 6375 .Case("sse2", X86Features::SSE2) 6376 .Case("sse3", X86Features::SSE3) 6377 .Case("sse4.1", X86Features::SSE4_1) 6378 .Case("sse4.2", X86Features::SSE4_2) 6379 .Case("avx", X86Features::AVX) 6380 .Case("avx2", X86Features::AVX2) 6381 .Case("sse4a", X86Features::SSE4_A) 6382 .Case("fma4", X86Features::FMA4) 6383 .Case("xop", X86Features::XOP) 6384 .Case("fma", X86Features::FMA) 6385 .Case("avx512f", X86Features::AVX512F) 6386 .Case("bmi", X86Features::BMI) 6387 .Case("bmi2", X86Features::BMI2) 6388 .Default(X86Features::MAX); 6389 assert(Feature != X86Features::MAX && "Invalid feature!"); 6390 6391 // Matching the struct layout from the compiler-rt/libgcc structure that is 6392 // filled in: 6393 // unsigned int __cpu_vendor; 6394 // unsigned int __cpu_type; 6395 // unsigned int __cpu_subtype; 6396 // unsigned int __cpu_features[1]; 6397 llvm::Type *STy = llvm::StructType::get( 6398 Int32Ty, Int32Ty, Int32Ty, llvm::ArrayType::get(Int32Ty, 1), nullptr); 6399 6400 // Grab the global __cpu_model. 6401 llvm::Constant *CpuModel = CGM.CreateRuntimeVariable(STy, "__cpu_model"); 6402 6403 // Grab the first (0th) element from the field __cpu_features off of the 6404 // global in the struct STy. 6405 Value *Idxs[] = { 6406 ConstantInt::get(Int32Ty, 0), 6407 ConstantInt::get(Int32Ty, 3), 6408 ConstantInt::get(Int32Ty, 0) 6409 }; 6410 Value *CpuFeatures = Builder.CreateGEP(STy, CpuModel, Idxs); 6411 Value *Features = Builder.CreateAlignedLoad(CpuFeatures, 6412 CharUnits::fromQuantity(4)); 6413 6414 // Check the value of the bit corresponding to the feature requested. 6415 Value *Bitset = Builder.CreateAnd( 6416 Features, llvm::ConstantInt::get(Int32Ty, 1ULL << Feature)); 6417 return Builder.CreateICmpNE(Bitset, llvm::ConstantInt::get(Int32Ty, 0)); 6418 } 6419 case X86::BI_mm_prefetch: { 6420 Value *Address = Ops[0]; 6421 Value *RW = ConstantInt::get(Int32Ty, 0); 6422 Value *Locality = Ops[1]; 6423 Value *Data = ConstantInt::get(Int32Ty, 1); 6424 Value *F = CGM.getIntrinsic(Intrinsic::prefetch); 6425 return Builder.CreateCall(F, {Address, RW, Locality, Data}); 6426 } 6427 case X86::BI__builtin_ia32_undef128: 6428 case X86::BI__builtin_ia32_undef256: 6429 case X86::BI__builtin_ia32_undef512: 6430 return UndefValue::get(ConvertType(E->getType())); 6431 case X86::BI__builtin_ia32_vec_init_v8qi: 6432 case X86::BI__builtin_ia32_vec_init_v4hi: 6433 case X86::BI__builtin_ia32_vec_init_v2si: 6434 return Builder.CreateBitCast(BuildVector(Ops), 6435 llvm::Type::getX86_MMXTy(getLLVMContext())); 6436 case X86::BI__builtin_ia32_vec_ext_v2si: 6437 return Builder.CreateExtractElement(Ops[0], 6438 llvm::ConstantInt::get(Ops[1]->getType(), 0)); 6439 case X86::BI__builtin_ia32_ldmxcsr: { 6440 Address Tmp = CreateMemTemp(E->getArg(0)->getType()); 6441 Builder.CreateStore(Ops[0], Tmp); 6442 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::x86_sse_ldmxcsr), 6443 Builder.CreateBitCast(Tmp.getPointer(), Int8PtrTy)); 6444 } 6445 case X86::BI__builtin_ia32_stmxcsr: { 6446 Address Tmp = CreateMemTemp(E->getType()); 6447 Builder.CreateCall(CGM.getIntrinsic(Intrinsic::x86_sse_stmxcsr), 6448 Builder.CreateBitCast(Tmp.getPointer(), Int8PtrTy)); 6449 return Builder.CreateLoad(Tmp, "stmxcsr"); 6450 } 6451 case X86::BI__builtin_ia32_xsave: 6452 case X86::BI__builtin_ia32_xsave64: 6453 case X86::BI__builtin_ia32_xrstor: 6454 case X86::BI__builtin_ia32_xrstor64: 6455 case X86::BI__builtin_ia32_xsaveopt: 6456 case X86::BI__builtin_ia32_xsaveopt64: 6457 case X86::BI__builtin_ia32_xrstors: 6458 case X86::BI__builtin_ia32_xrstors64: 6459 case X86::BI__builtin_ia32_xsavec: 6460 case X86::BI__builtin_ia32_xsavec64: 6461 case X86::BI__builtin_ia32_xsaves: 6462 case X86::BI__builtin_ia32_xsaves64: { 6463 Intrinsic::ID ID; 6464 #define INTRINSIC_X86_XSAVE_ID(NAME) \ 6465 case X86::BI__builtin_ia32_##NAME: \ 6466 ID = Intrinsic::x86_##NAME; \ 6467 break 6468 switch (BuiltinID) { 6469 default: llvm_unreachable("Unsupported intrinsic!"); 6470 INTRINSIC_X86_XSAVE_ID(xsave); 6471 INTRINSIC_X86_XSAVE_ID(xsave64); 6472 INTRINSIC_X86_XSAVE_ID(xrstor); 6473 INTRINSIC_X86_XSAVE_ID(xrstor64); 6474 INTRINSIC_X86_XSAVE_ID(xsaveopt); 6475 INTRINSIC_X86_XSAVE_ID(xsaveopt64); 6476 INTRINSIC_X86_XSAVE_ID(xrstors); 6477 INTRINSIC_X86_XSAVE_ID(xrstors64); 6478 INTRINSIC_X86_XSAVE_ID(xsavec); 6479 INTRINSIC_X86_XSAVE_ID(xsavec64); 6480 INTRINSIC_X86_XSAVE_ID(xsaves); 6481 INTRINSIC_X86_XSAVE_ID(xsaves64); 6482 } 6483 #undef INTRINSIC_X86_XSAVE_ID 6484 Value *Mhi = Builder.CreateTrunc( 6485 Builder.CreateLShr(Ops[1], ConstantInt::get(Int64Ty, 32)), Int32Ty); 6486 Value *Mlo = Builder.CreateTrunc(Ops[1], Int32Ty); 6487 Ops[1] = Mhi; 6488 Ops.push_back(Mlo); 6489 return Builder.CreateCall(CGM.getIntrinsic(ID), Ops); 6490 } 6491 case X86::BI__builtin_ia32_storehps: 6492 case X86::BI__builtin_ia32_storelps: { 6493 llvm::Type *PtrTy = llvm::PointerType::getUnqual(Int64Ty); 6494 llvm::Type *VecTy = llvm::VectorType::get(Int64Ty, 2); 6495 6496 // cast val v2i64 6497 Ops[1] = Builder.CreateBitCast(Ops[1], VecTy, "cast"); 6498 6499 // extract (0, 1) 6500 unsigned Index = BuiltinID == X86::BI__builtin_ia32_storelps ? 0 : 1; 6501 llvm::Value *Idx = llvm::ConstantInt::get(SizeTy, Index); 6502 Ops[1] = Builder.CreateExtractElement(Ops[1], Idx, "extract"); 6503 6504 // cast pointer to i64 & store 6505 Ops[0] = Builder.CreateBitCast(Ops[0], PtrTy); 6506 return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]); 6507 } 6508 case X86::BI__builtin_ia32_palignr128: 6509 case X86::BI__builtin_ia32_palignr256: { 6510 unsigned ShiftVal = cast<llvm::ConstantInt>(Ops[2])->getZExtValue(); 6511 6512 unsigned NumElts = 6513 cast<llvm::VectorType>(Ops[0]->getType())->getNumElements(); 6514 assert(NumElts % 16 == 0); 6515 unsigned NumLanes = NumElts / 16; 6516 unsigned NumLaneElts = NumElts / NumLanes; 6517 6518 // If palignr is shifting the pair of vectors more than the size of two 6519 // lanes, emit zero. 6520 if (ShiftVal >= (2 * NumLaneElts)) 6521 return llvm::Constant::getNullValue(ConvertType(E->getType())); 6522 6523 // If palignr is shifting the pair of input vectors more than one lane, 6524 // but less than two lanes, convert to shifting in zeroes. 6525 if (ShiftVal > NumLaneElts) { 6526 ShiftVal -= NumLaneElts; 6527 Ops[1] = Ops[0]; 6528 Ops[0] = llvm::Constant::getNullValue(Ops[0]->getType()); 6529 } 6530 6531 uint32_t Indices[32]; 6532 // 256-bit palignr operates on 128-bit lanes so we need to handle that 6533 for (unsigned l = 0; l != NumElts; l += NumLaneElts) { 6534 for (unsigned i = 0; i != NumLaneElts; ++i) { 6535 unsigned Idx = ShiftVal + i; 6536 if (Idx >= NumLaneElts) 6537 Idx += NumElts - NumLaneElts; // End of lane, switch operand. 6538 Indices[l + i] = Idx + l; 6539 } 6540 } 6541 6542 Value *SV = llvm::ConstantDataVector::get(getLLVMContext(), 6543 makeArrayRef(Indices, NumElts)); 6544 return Builder.CreateShuffleVector(Ops[1], Ops[0], SV, "palignr"); 6545 } 6546 case X86::BI__builtin_ia32_pslldqi256: { 6547 // Shift value is in bits so divide by 8. 6548 unsigned shiftVal = cast<llvm::ConstantInt>(Ops[1])->getZExtValue() >> 3; 6549 6550 // If pslldq is shifting the vector more than 15 bytes, emit zero. 6551 if (shiftVal >= 16) 6552 return llvm::Constant::getNullValue(ConvertType(E->getType())); 6553 6554 uint32_t Indices[32]; 6555 // 256-bit pslldq operates on 128-bit lanes so we need to handle that 6556 for (unsigned l = 0; l != 32; l += 16) { 6557 for (unsigned i = 0; i != 16; ++i) { 6558 unsigned Idx = 32 + i - shiftVal; 6559 if (Idx < 32) Idx -= 16; // end of lane, switch operand. 6560 Indices[l + i] = Idx + l; 6561 } 6562 } 6563 6564 llvm::Type *VecTy = llvm::VectorType::get(Int8Ty, 32); 6565 Ops[0] = Builder.CreateBitCast(Ops[0], VecTy, "cast"); 6566 Value *Zero = llvm::Constant::getNullValue(VecTy); 6567 6568 Value *SV = llvm::ConstantDataVector::get(getLLVMContext(), Indices); 6569 SV = Builder.CreateShuffleVector(Zero, Ops[0], SV, "pslldq"); 6570 llvm::Type *ResultType = ConvertType(E->getType()); 6571 return Builder.CreateBitCast(SV, ResultType, "cast"); 6572 } 6573 case X86::BI__builtin_ia32_psrldqi256: { 6574 // Shift value is in bits so divide by 8. 6575 unsigned shiftVal = cast<llvm::ConstantInt>(Ops[1])->getZExtValue() >> 3; 6576 6577 // If psrldq is shifting the vector more than 15 bytes, emit zero. 6578 if (shiftVal >= 16) 6579 return llvm::Constant::getNullValue(ConvertType(E->getType())); 6580 6581 uint32_t Indices[32]; 6582 // 256-bit psrldq operates on 128-bit lanes so we need to handle that 6583 for (unsigned l = 0; l != 32; l += 16) { 6584 for (unsigned i = 0; i != 16; ++i) { 6585 unsigned Idx = i + shiftVal; 6586 if (Idx >= 16) Idx += 16; // end of lane, switch operand. 6587 Indices[l + i] = Idx + l; 6588 } 6589 } 6590 6591 llvm::Type *VecTy = llvm::VectorType::get(Int8Ty, 32); 6592 Ops[0] = Builder.CreateBitCast(Ops[0], VecTy, "cast"); 6593 Value *Zero = llvm::Constant::getNullValue(VecTy); 6594 6595 Value *SV = llvm::ConstantDataVector::get(getLLVMContext(), Indices); 6596 SV = Builder.CreateShuffleVector(Ops[0], Zero, SV, "psrldq"); 6597 llvm::Type *ResultType = ConvertType(E->getType()); 6598 return Builder.CreateBitCast(SV, ResultType, "cast"); 6599 } 6600 case X86::BI__builtin_ia32_movntps: 6601 case X86::BI__builtin_ia32_movntps256: 6602 case X86::BI__builtin_ia32_movntpd: 6603 case X86::BI__builtin_ia32_movntpd256: 6604 case X86::BI__builtin_ia32_movntdq: 6605 case X86::BI__builtin_ia32_movntdq256: 6606 case X86::BI__builtin_ia32_movnti: 6607 case X86::BI__builtin_ia32_movnti64: { 6608 llvm::MDNode *Node = llvm::MDNode::get( 6609 getLLVMContext(), llvm::ConstantAsMetadata::get(Builder.getInt32(1))); 6610 6611 // Convert the type of the pointer to a pointer to the stored type. 6612 Value *BC = Builder.CreateBitCast(Ops[0], 6613 llvm::PointerType::getUnqual(Ops[1]->getType()), 6614 "cast"); 6615 StoreInst *SI = Builder.CreateDefaultAlignedStore(Ops[1], BC); 6616 SI->setMetadata(CGM.getModule().getMDKindID("nontemporal"), Node); 6617 6618 // If the operand is an integer, we can't assume alignment. Otherwise, 6619 // assume natural alignment. 6620 QualType ArgTy = E->getArg(1)->getType(); 6621 unsigned Align; 6622 if (ArgTy->isIntegerType()) 6623 Align = 1; 6624 else 6625 Align = getContext().getTypeSizeInChars(ArgTy).getQuantity(); 6626 SI->setAlignment(Align); 6627 return SI; 6628 } 6629 // 3DNow! 6630 case X86::BI__builtin_ia32_pswapdsf: 6631 case X86::BI__builtin_ia32_pswapdsi: { 6632 llvm::Type *MMXTy = llvm::Type::getX86_MMXTy(getLLVMContext()); 6633 Ops[0] = Builder.CreateBitCast(Ops[0], MMXTy, "cast"); 6634 llvm::Function *F = CGM.getIntrinsic(Intrinsic::x86_3dnowa_pswapd); 6635 return Builder.CreateCall(F, Ops, "pswapd"); 6636 } 6637 case X86::BI__builtin_ia32_rdrand16_step: 6638 case X86::BI__builtin_ia32_rdrand32_step: 6639 case X86::BI__builtin_ia32_rdrand64_step: 6640 case X86::BI__builtin_ia32_rdseed16_step: 6641 case X86::BI__builtin_ia32_rdseed32_step: 6642 case X86::BI__builtin_ia32_rdseed64_step: { 6643 Intrinsic::ID ID; 6644 switch (BuiltinID) { 6645 default: llvm_unreachable("Unsupported intrinsic!"); 6646 case X86::BI__builtin_ia32_rdrand16_step: 6647 ID = Intrinsic::x86_rdrand_16; 6648 break; 6649 case X86::BI__builtin_ia32_rdrand32_step: 6650 ID = Intrinsic::x86_rdrand_32; 6651 break; 6652 case X86::BI__builtin_ia32_rdrand64_step: 6653 ID = Intrinsic::x86_rdrand_64; 6654 break; 6655 case X86::BI__builtin_ia32_rdseed16_step: 6656 ID = Intrinsic::x86_rdseed_16; 6657 break; 6658 case X86::BI__builtin_ia32_rdseed32_step: 6659 ID = Intrinsic::x86_rdseed_32; 6660 break; 6661 case X86::BI__builtin_ia32_rdseed64_step: 6662 ID = Intrinsic::x86_rdseed_64; 6663 break; 6664 } 6665 6666 Value *Call = Builder.CreateCall(CGM.getIntrinsic(ID)); 6667 Builder.CreateDefaultAlignedStore(Builder.CreateExtractValue(Call, 0), 6668 Ops[0]); 6669 return Builder.CreateExtractValue(Call, 1); 6670 } 6671 // SSE comparison intrisics 6672 case X86::BI__builtin_ia32_cmpeqps: 6673 case X86::BI__builtin_ia32_cmpltps: 6674 case X86::BI__builtin_ia32_cmpleps: 6675 case X86::BI__builtin_ia32_cmpunordps: 6676 case X86::BI__builtin_ia32_cmpneqps: 6677 case X86::BI__builtin_ia32_cmpnltps: 6678 case X86::BI__builtin_ia32_cmpnleps: 6679 case X86::BI__builtin_ia32_cmpordps: 6680 case X86::BI__builtin_ia32_cmpeqss: 6681 case X86::BI__builtin_ia32_cmpltss: 6682 case X86::BI__builtin_ia32_cmpless: 6683 case X86::BI__builtin_ia32_cmpunordss: 6684 case X86::BI__builtin_ia32_cmpneqss: 6685 case X86::BI__builtin_ia32_cmpnltss: 6686 case X86::BI__builtin_ia32_cmpnless: 6687 case X86::BI__builtin_ia32_cmpordss: 6688 case X86::BI__builtin_ia32_cmpeqpd: 6689 case X86::BI__builtin_ia32_cmpltpd: 6690 case X86::BI__builtin_ia32_cmplepd: 6691 case X86::BI__builtin_ia32_cmpunordpd: 6692 case X86::BI__builtin_ia32_cmpneqpd: 6693 case X86::BI__builtin_ia32_cmpnltpd: 6694 case X86::BI__builtin_ia32_cmpnlepd: 6695 case X86::BI__builtin_ia32_cmpordpd: 6696 case X86::BI__builtin_ia32_cmpeqsd: 6697 case X86::BI__builtin_ia32_cmpltsd: 6698 case X86::BI__builtin_ia32_cmplesd: 6699 case X86::BI__builtin_ia32_cmpunordsd: 6700 case X86::BI__builtin_ia32_cmpneqsd: 6701 case X86::BI__builtin_ia32_cmpnltsd: 6702 case X86::BI__builtin_ia32_cmpnlesd: 6703 case X86::BI__builtin_ia32_cmpordsd: 6704 // These exist so that the builtin that takes an immediate can be bounds 6705 // checked by clang to avoid passing bad immediates to the backend. Since 6706 // AVX has a larger immediate than SSE we would need separate builtins to 6707 // do the different bounds checking. Rather than create a clang specific 6708 // SSE only builtin, this implements eight separate builtins to match gcc 6709 // implementation. 6710 6711 // Choose the immediate. 6712 unsigned Imm; 6713 switch (BuiltinID) { 6714 default: llvm_unreachable("Unsupported intrinsic!"); 6715 case X86::BI__builtin_ia32_cmpeqps: 6716 case X86::BI__builtin_ia32_cmpeqss: 6717 case X86::BI__builtin_ia32_cmpeqpd: 6718 case X86::BI__builtin_ia32_cmpeqsd: 6719 Imm = 0; 6720 break; 6721 case X86::BI__builtin_ia32_cmpltps: 6722 case X86::BI__builtin_ia32_cmpltss: 6723 case X86::BI__builtin_ia32_cmpltpd: 6724 case X86::BI__builtin_ia32_cmpltsd: 6725 Imm = 1; 6726 break; 6727 case X86::BI__builtin_ia32_cmpleps: 6728 case X86::BI__builtin_ia32_cmpless: 6729 case X86::BI__builtin_ia32_cmplepd: 6730 case X86::BI__builtin_ia32_cmplesd: 6731 Imm = 2; 6732 break; 6733 case X86::BI__builtin_ia32_cmpunordps: 6734 case X86::BI__builtin_ia32_cmpunordss: 6735 case X86::BI__builtin_ia32_cmpunordpd: 6736 case X86::BI__builtin_ia32_cmpunordsd: 6737 Imm = 3; 6738 break; 6739 case X86::BI__builtin_ia32_cmpneqps: 6740 case X86::BI__builtin_ia32_cmpneqss: 6741 case X86::BI__builtin_ia32_cmpneqpd: 6742 case X86::BI__builtin_ia32_cmpneqsd: 6743 Imm = 4; 6744 break; 6745 case X86::BI__builtin_ia32_cmpnltps: 6746 case X86::BI__builtin_ia32_cmpnltss: 6747 case X86::BI__builtin_ia32_cmpnltpd: 6748 case X86::BI__builtin_ia32_cmpnltsd: 6749 Imm = 5; 6750 break; 6751 case X86::BI__builtin_ia32_cmpnleps: 6752 case X86::BI__builtin_ia32_cmpnless: 6753 case X86::BI__builtin_ia32_cmpnlepd: 6754 case X86::BI__builtin_ia32_cmpnlesd: 6755 Imm = 6; 6756 break; 6757 case X86::BI__builtin_ia32_cmpordps: 6758 case X86::BI__builtin_ia32_cmpordss: 6759 case X86::BI__builtin_ia32_cmpordpd: 6760 case X86::BI__builtin_ia32_cmpordsd: 6761 Imm = 7; 6762 break; 6763 } 6764 6765 // Choose the intrinsic ID. 6766 const char *name; 6767 Intrinsic::ID ID; 6768 switch (BuiltinID) { 6769 default: llvm_unreachable("Unsupported intrinsic!"); 6770 case X86::BI__builtin_ia32_cmpeqps: 6771 case X86::BI__builtin_ia32_cmpltps: 6772 case X86::BI__builtin_ia32_cmpleps: 6773 case X86::BI__builtin_ia32_cmpunordps: 6774 case X86::BI__builtin_ia32_cmpneqps: 6775 case X86::BI__builtin_ia32_cmpnltps: 6776 case X86::BI__builtin_ia32_cmpnleps: 6777 case X86::BI__builtin_ia32_cmpordps: 6778 name = "cmpps"; 6779 ID = Intrinsic::x86_sse_cmp_ps; 6780 break; 6781 case X86::BI__builtin_ia32_cmpeqss: 6782 case X86::BI__builtin_ia32_cmpltss: 6783 case X86::BI__builtin_ia32_cmpless: 6784 case X86::BI__builtin_ia32_cmpunordss: 6785 case X86::BI__builtin_ia32_cmpneqss: 6786 case X86::BI__builtin_ia32_cmpnltss: 6787 case X86::BI__builtin_ia32_cmpnless: 6788 case X86::BI__builtin_ia32_cmpordss: 6789 name = "cmpss"; 6790 ID = Intrinsic::x86_sse_cmp_ss; 6791 break; 6792 case X86::BI__builtin_ia32_cmpeqpd: 6793 case X86::BI__builtin_ia32_cmpltpd: 6794 case X86::BI__builtin_ia32_cmplepd: 6795 case X86::BI__builtin_ia32_cmpunordpd: 6796 case X86::BI__builtin_ia32_cmpneqpd: 6797 case X86::BI__builtin_ia32_cmpnltpd: 6798 case X86::BI__builtin_ia32_cmpnlepd: 6799 case X86::BI__builtin_ia32_cmpordpd: 6800 name = "cmppd"; 6801 ID = Intrinsic::x86_sse2_cmp_pd; 6802 break; 6803 case X86::BI__builtin_ia32_cmpeqsd: 6804 case X86::BI__builtin_ia32_cmpltsd: 6805 case X86::BI__builtin_ia32_cmplesd: 6806 case X86::BI__builtin_ia32_cmpunordsd: 6807 case X86::BI__builtin_ia32_cmpneqsd: 6808 case X86::BI__builtin_ia32_cmpnltsd: 6809 case X86::BI__builtin_ia32_cmpnlesd: 6810 case X86::BI__builtin_ia32_cmpordsd: 6811 name = "cmpsd"; 6812 ID = Intrinsic::x86_sse2_cmp_sd; 6813 break; 6814 } 6815 6816 Ops.push_back(llvm::ConstantInt::get(Int8Ty, Imm)); 6817 llvm::Function *F = CGM.getIntrinsic(ID); 6818 return Builder.CreateCall(F, Ops, name); 6819 } 6820 } 6821 6822 6823 Value *CodeGenFunction::EmitPPCBuiltinExpr(unsigned BuiltinID, 6824 const CallExpr *E) { 6825 SmallVector<Value*, 4> Ops; 6826 6827 for (unsigned i = 0, e = E->getNumArgs(); i != e; i++) 6828 Ops.push_back(EmitScalarExpr(E->getArg(i))); 6829 6830 Intrinsic::ID ID = Intrinsic::not_intrinsic; 6831 6832 switch (BuiltinID) { 6833 default: return nullptr; 6834 6835 // __builtin_ppc_get_timebase is GCC 4.8+'s PowerPC-specific name for what we 6836 // call __builtin_readcyclecounter. 6837 case PPC::BI__builtin_ppc_get_timebase: 6838 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::readcyclecounter)); 6839 6840 // vec_ld, vec_lvsl, vec_lvsr 6841 case PPC::BI__builtin_altivec_lvx: 6842 case PPC::BI__builtin_altivec_lvxl: 6843 case PPC::BI__builtin_altivec_lvebx: 6844 case PPC::BI__builtin_altivec_lvehx: 6845 case PPC::BI__builtin_altivec_lvewx: 6846 case PPC::BI__builtin_altivec_lvsl: 6847 case PPC::BI__builtin_altivec_lvsr: 6848 case PPC::BI__builtin_vsx_lxvd2x: 6849 case PPC::BI__builtin_vsx_lxvw4x: 6850 { 6851 Ops[1] = Builder.CreateBitCast(Ops[1], Int8PtrTy); 6852 6853 Ops[0] = Builder.CreateGEP(Ops[1], Ops[0]); 6854 Ops.pop_back(); 6855 6856 switch (BuiltinID) { 6857 default: llvm_unreachable("Unsupported ld/lvsl/lvsr intrinsic!"); 6858 case PPC::BI__builtin_altivec_lvx: 6859 ID = Intrinsic::ppc_altivec_lvx; 6860 break; 6861 case PPC::BI__builtin_altivec_lvxl: 6862 ID = Intrinsic::ppc_altivec_lvxl; 6863 break; 6864 case PPC::BI__builtin_altivec_lvebx: 6865 ID = Intrinsic::ppc_altivec_lvebx; 6866 break; 6867 case PPC::BI__builtin_altivec_lvehx: 6868 ID = Intrinsic::ppc_altivec_lvehx; 6869 break; 6870 case PPC::BI__builtin_altivec_lvewx: 6871 ID = Intrinsic::ppc_altivec_lvewx; 6872 break; 6873 case PPC::BI__builtin_altivec_lvsl: 6874 ID = Intrinsic::ppc_altivec_lvsl; 6875 break; 6876 case PPC::BI__builtin_altivec_lvsr: 6877 ID = Intrinsic::ppc_altivec_lvsr; 6878 break; 6879 case PPC::BI__builtin_vsx_lxvd2x: 6880 ID = Intrinsic::ppc_vsx_lxvd2x; 6881 break; 6882 case PPC::BI__builtin_vsx_lxvw4x: 6883 ID = Intrinsic::ppc_vsx_lxvw4x; 6884 break; 6885 } 6886 llvm::Function *F = CGM.getIntrinsic(ID); 6887 return Builder.CreateCall(F, Ops, ""); 6888 } 6889 6890 // vec_st 6891 case PPC::BI__builtin_altivec_stvx: 6892 case PPC::BI__builtin_altivec_stvxl: 6893 case PPC::BI__builtin_altivec_stvebx: 6894 case PPC::BI__builtin_altivec_stvehx: 6895 case PPC::BI__builtin_altivec_stvewx: 6896 case PPC::BI__builtin_vsx_stxvd2x: 6897 case PPC::BI__builtin_vsx_stxvw4x: 6898 { 6899 Ops[2] = Builder.CreateBitCast(Ops[2], Int8PtrTy); 6900 Ops[1] = Builder.CreateGEP(Ops[2], Ops[1]); 6901 Ops.pop_back(); 6902 6903 switch (BuiltinID) { 6904 default: llvm_unreachable("Unsupported st intrinsic!"); 6905 case PPC::BI__builtin_altivec_stvx: 6906 ID = Intrinsic::ppc_altivec_stvx; 6907 break; 6908 case PPC::BI__builtin_altivec_stvxl: 6909 ID = Intrinsic::ppc_altivec_stvxl; 6910 break; 6911 case PPC::BI__builtin_altivec_stvebx: 6912 ID = Intrinsic::ppc_altivec_stvebx; 6913 break; 6914 case PPC::BI__builtin_altivec_stvehx: 6915 ID = Intrinsic::ppc_altivec_stvehx; 6916 break; 6917 case PPC::BI__builtin_altivec_stvewx: 6918 ID = Intrinsic::ppc_altivec_stvewx; 6919 break; 6920 case PPC::BI__builtin_vsx_stxvd2x: 6921 ID = Intrinsic::ppc_vsx_stxvd2x; 6922 break; 6923 case PPC::BI__builtin_vsx_stxvw4x: 6924 ID = Intrinsic::ppc_vsx_stxvw4x; 6925 break; 6926 } 6927 llvm::Function *F = CGM.getIntrinsic(ID); 6928 return Builder.CreateCall(F, Ops, ""); 6929 } 6930 // Square root 6931 case PPC::BI__builtin_vsx_xvsqrtsp: 6932 case PPC::BI__builtin_vsx_xvsqrtdp: { 6933 llvm::Type *ResultType = ConvertType(E->getType()); 6934 Value *X = EmitScalarExpr(E->getArg(0)); 6935 ID = Intrinsic::sqrt; 6936 llvm::Function *F = CGM.getIntrinsic(ID, ResultType); 6937 return Builder.CreateCall(F, X); 6938 } 6939 // Count leading zeros 6940 case PPC::BI__builtin_altivec_vclzb: 6941 case PPC::BI__builtin_altivec_vclzh: 6942 case PPC::BI__builtin_altivec_vclzw: 6943 case PPC::BI__builtin_altivec_vclzd: { 6944 llvm::Type *ResultType = ConvertType(E->getType()); 6945 Value *X = EmitScalarExpr(E->getArg(0)); 6946 Value *Undef = ConstantInt::get(Builder.getInt1Ty(), false); 6947 Function *F = CGM.getIntrinsic(Intrinsic::ctlz, ResultType); 6948 return Builder.CreateCall(F, {X, Undef}); 6949 } 6950 // Copy sign 6951 case PPC::BI__builtin_vsx_xvcpsgnsp: 6952 case PPC::BI__builtin_vsx_xvcpsgndp: { 6953 llvm::Type *ResultType = ConvertType(E->getType()); 6954 Value *X = EmitScalarExpr(E->getArg(0)); 6955 Value *Y = EmitScalarExpr(E->getArg(1)); 6956 ID = Intrinsic::copysign; 6957 llvm::Function *F = CGM.getIntrinsic(ID, ResultType); 6958 return Builder.CreateCall(F, {X, Y}); 6959 } 6960 // Rounding/truncation 6961 case PPC::BI__builtin_vsx_xvrspip: 6962 case PPC::BI__builtin_vsx_xvrdpip: 6963 case PPC::BI__builtin_vsx_xvrdpim: 6964 case PPC::BI__builtin_vsx_xvrspim: 6965 case PPC::BI__builtin_vsx_xvrdpi: 6966 case PPC::BI__builtin_vsx_xvrspi: 6967 case PPC::BI__builtin_vsx_xvrdpic: 6968 case PPC::BI__builtin_vsx_xvrspic: 6969 case PPC::BI__builtin_vsx_xvrdpiz: 6970 case PPC::BI__builtin_vsx_xvrspiz: { 6971 llvm::Type *ResultType = ConvertType(E->getType()); 6972 Value *X = EmitScalarExpr(E->getArg(0)); 6973 if (BuiltinID == PPC::BI__builtin_vsx_xvrdpim || 6974 BuiltinID == PPC::BI__builtin_vsx_xvrspim) 6975 ID = Intrinsic::floor; 6976 else if (BuiltinID == PPC::BI__builtin_vsx_xvrdpi || 6977 BuiltinID == PPC::BI__builtin_vsx_xvrspi) 6978 ID = Intrinsic::round; 6979 else if (BuiltinID == PPC::BI__builtin_vsx_xvrdpic || 6980 BuiltinID == PPC::BI__builtin_vsx_xvrspic) 6981 ID = Intrinsic::nearbyint; 6982 else if (BuiltinID == PPC::BI__builtin_vsx_xvrdpip || 6983 BuiltinID == PPC::BI__builtin_vsx_xvrspip) 6984 ID = Intrinsic::ceil; 6985 else if (BuiltinID == PPC::BI__builtin_vsx_xvrdpiz || 6986 BuiltinID == PPC::BI__builtin_vsx_xvrspiz) 6987 ID = Intrinsic::trunc; 6988 llvm::Function *F = CGM.getIntrinsic(ID, ResultType); 6989 return Builder.CreateCall(F, X); 6990 } 6991 6992 // Absolute value 6993 case PPC::BI__builtin_vsx_xvabsdp: 6994 case PPC::BI__builtin_vsx_xvabssp: { 6995 llvm::Type *ResultType = ConvertType(E->getType()); 6996 Value *X = EmitScalarExpr(E->getArg(0)); 6997 llvm::Function *F = CGM.getIntrinsic(Intrinsic::fabs, ResultType); 6998 return Builder.CreateCall(F, X); 6999 } 7000 7001 // FMA variations 7002 case PPC::BI__builtin_vsx_xvmaddadp: 7003 case PPC::BI__builtin_vsx_xvmaddasp: 7004 case PPC::BI__builtin_vsx_xvnmaddadp: 7005 case PPC::BI__builtin_vsx_xvnmaddasp: 7006 case PPC::BI__builtin_vsx_xvmsubadp: 7007 case PPC::BI__builtin_vsx_xvmsubasp: 7008 case PPC::BI__builtin_vsx_xvnmsubadp: 7009 case PPC::BI__builtin_vsx_xvnmsubasp: { 7010 llvm::Type *ResultType = ConvertType(E->getType()); 7011 Value *X = EmitScalarExpr(E->getArg(0)); 7012 Value *Y = EmitScalarExpr(E->getArg(1)); 7013 Value *Z = EmitScalarExpr(E->getArg(2)); 7014 Value *Zero = llvm::ConstantFP::getZeroValueForNegation(ResultType); 7015 llvm::Function *F = CGM.getIntrinsic(Intrinsic::fma, ResultType); 7016 switch (BuiltinID) { 7017 case PPC::BI__builtin_vsx_xvmaddadp: 7018 case PPC::BI__builtin_vsx_xvmaddasp: 7019 return Builder.CreateCall(F, {X, Y, Z}); 7020 case PPC::BI__builtin_vsx_xvnmaddadp: 7021 case PPC::BI__builtin_vsx_xvnmaddasp: 7022 return Builder.CreateFSub(Zero, 7023 Builder.CreateCall(F, {X, Y, Z}), "sub"); 7024 case PPC::BI__builtin_vsx_xvmsubadp: 7025 case PPC::BI__builtin_vsx_xvmsubasp: 7026 return Builder.CreateCall(F, 7027 {X, Y, Builder.CreateFSub(Zero, Z, "sub")}); 7028 case PPC::BI__builtin_vsx_xvnmsubadp: 7029 case PPC::BI__builtin_vsx_xvnmsubasp: 7030 Value *FsubRes = 7031 Builder.CreateCall(F, {X, Y, Builder.CreateFSub(Zero, Z, "sub")}); 7032 return Builder.CreateFSub(Zero, FsubRes, "sub"); 7033 } 7034 llvm_unreachable("Unknown FMA operation"); 7035 return nullptr; // Suppress no-return warning 7036 } 7037 } 7038 } 7039 7040 Value *CodeGenFunction::EmitAMDGPUBuiltinExpr(unsigned BuiltinID, 7041 const CallExpr *E) { 7042 switch (BuiltinID) { 7043 case AMDGPU::BI__builtin_amdgcn_div_scale: 7044 case AMDGPU::BI__builtin_amdgcn_div_scalef: { 7045 // Translate from the intrinsics's struct return to the builtin's out 7046 // argument. 7047 7048 Address FlagOutPtr = EmitPointerWithAlignment(E->getArg(3)); 7049 7050 llvm::Value *X = EmitScalarExpr(E->getArg(0)); 7051 llvm::Value *Y = EmitScalarExpr(E->getArg(1)); 7052 llvm::Value *Z = EmitScalarExpr(E->getArg(2)); 7053 7054 llvm::Value *Callee = CGM.getIntrinsic(Intrinsic::amdgcn_div_scale, 7055 X->getType()); 7056 7057 llvm::Value *Tmp = Builder.CreateCall(Callee, {X, Y, Z}); 7058 7059 llvm::Value *Result = Builder.CreateExtractValue(Tmp, 0); 7060 llvm::Value *Flag = Builder.CreateExtractValue(Tmp, 1); 7061 7062 llvm::Type *RealFlagType 7063 = FlagOutPtr.getPointer()->getType()->getPointerElementType(); 7064 7065 llvm::Value *FlagExt = Builder.CreateZExt(Flag, RealFlagType); 7066 Builder.CreateStore(FlagExt, FlagOutPtr); 7067 return Result; 7068 } 7069 case AMDGPU::BI__builtin_amdgcn_div_fmas: 7070 case AMDGPU::BI__builtin_amdgcn_div_fmasf: { 7071 llvm::Value *Src0 = EmitScalarExpr(E->getArg(0)); 7072 llvm::Value *Src1 = EmitScalarExpr(E->getArg(1)); 7073 llvm::Value *Src2 = EmitScalarExpr(E->getArg(2)); 7074 llvm::Value *Src3 = EmitScalarExpr(E->getArg(3)); 7075 7076 llvm::Value *F = CGM.getIntrinsic(Intrinsic::amdgcn_div_fmas, 7077 Src0->getType()); 7078 llvm::Value *Src3ToBool = Builder.CreateIsNotNull(Src3); 7079 return Builder.CreateCall(F, {Src0, Src1, Src2, Src3ToBool}); 7080 } 7081 case AMDGPU::BI__builtin_amdgcn_div_fixup: 7082 case AMDGPU::BI__builtin_amdgcn_div_fixupf: 7083 return emitTernaryFPBuiltin(*this, E, Intrinsic::amdgcn_div_fixup); 7084 case AMDGPU::BI__builtin_amdgcn_trig_preop: 7085 case AMDGPU::BI__builtin_amdgcn_trig_preopf: 7086 return emitFPIntBuiltin(*this, E, Intrinsic::amdgcn_trig_preop); 7087 case AMDGPU::BI__builtin_amdgcn_rcp: 7088 case AMDGPU::BI__builtin_amdgcn_rcpf: 7089 return emitUnaryBuiltin(*this, E, Intrinsic::amdgcn_rcp); 7090 case AMDGPU::BI__builtin_amdgcn_rsq: 7091 case AMDGPU::BI__builtin_amdgcn_rsqf: 7092 return emitUnaryBuiltin(*this, E, Intrinsic::amdgcn_rsq); 7093 case AMDGPU::BI__builtin_amdgcn_rsq_clamp: 7094 case AMDGPU::BI__builtin_amdgcn_rsq_clampf: 7095 return emitUnaryBuiltin(*this, E, Intrinsic::amdgcn_rsq_clamp); 7096 case AMDGPU::BI__builtin_amdgcn_sinf: 7097 return emitUnaryBuiltin(*this, E, Intrinsic::amdgcn_sin); 7098 case AMDGPU::BI__builtin_amdgcn_cosf: 7099 return emitUnaryBuiltin(*this, E, Intrinsic::amdgcn_cos); 7100 case AMDGPU::BI__builtin_amdgcn_log_clampf: 7101 return emitUnaryBuiltin(*this, E, Intrinsic::amdgcn_log_clamp); 7102 case AMDGPU::BI__builtin_amdgcn_ldexp: 7103 case AMDGPU::BI__builtin_amdgcn_ldexpf: 7104 return emitFPIntBuiltin(*this, E, Intrinsic::amdgcn_ldexp); 7105 case AMDGPU::BI__builtin_amdgcn_frexp_mant: 7106 case AMDGPU::BI__builtin_amdgcn_frexp_mantf: { 7107 return emitUnaryBuiltin(*this, E, Intrinsic::amdgcn_frexp_mant); 7108 } 7109 case AMDGPU::BI__builtin_amdgcn_frexp_exp: 7110 case AMDGPU::BI__builtin_amdgcn_frexp_expf: { 7111 return emitUnaryBuiltin(*this, E, Intrinsic::amdgcn_frexp_exp); 7112 } 7113 case AMDGPU::BI__builtin_amdgcn_class: 7114 case AMDGPU::BI__builtin_amdgcn_classf: 7115 return emitFPIntBuiltin(*this, E, Intrinsic::amdgcn_class); 7116 7117 // Legacy amdgpu prefix 7118 case AMDGPU::BI__builtin_amdgpu_rsq: 7119 case AMDGPU::BI__builtin_amdgpu_rsqf: { 7120 if (getTarget().getTriple().getArch() == Triple::amdgcn) 7121 return emitUnaryBuiltin(*this, E, Intrinsic::amdgcn_rsq); 7122 return emitUnaryBuiltin(*this, E, Intrinsic::r600_rsq); 7123 } 7124 case AMDGPU::BI__builtin_amdgpu_ldexp: 7125 case AMDGPU::BI__builtin_amdgpu_ldexpf: { 7126 if (getTarget().getTriple().getArch() == Triple::amdgcn) 7127 return emitFPIntBuiltin(*this, E, Intrinsic::amdgcn_ldexp); 7128 return emitFPIntBuiltin(*this, E, Intrinsic::AMDGPU_ldexp); 7129 } 7130 default: 7131 return nullptr; 7132 } 7133 } 7134 7135 /// Handle a SystemZ function in which the final argument is a pointer 7136 /// to an int that receives the post-instruction CC value. At the LLVM level 7137 /// this is represented as a function that returns a {result, cc} pair. 7138 static Value *EmitSystemZIntrinsicWithCC(CodeGenFunction &CGF, 7139 unsigned IntrinsicID, 7140 const CallExpr *E) { 7141 unsigned NumArgs = E->getNumArgs() - 1; 7142 SmallVector<Value *, 8> Args(NumArgs); 7143 for (unsigned I = 0; I < NumArgs; ++I) 7144 Args[I] = CGF.EmitScalarExpr(E->getArg(I)); 7145 Address CCPtr = CGF.EmitPointerWithAlignment(E->getArg(NumArgs)); 7146 Value *F = CGF.CGM.getIntrinsic(IntrinsicID); 7147 Value *Call = CGF.Builder.CreateCall(F, Args); 7148 Value *CC = CGF.Builder.CreateExtractValue(Call, 1); 7149 CGF.Builder.CreateStore(CC, CCPtr); 7150 return CGF.Builder.CreateExtractValue(Call, 0); 7151 } 7152 7153 Value *CodeGenFunction::EmitSystemZBuiltinExpr(unsigned BuiltinID, 7154 const CallExpr *E) { 7155 switch (BuiltinID) { 7156 case SystemZ::BI__builtin_tbegin: { 7157 Value *TDB = EmitScalarExpr(E->getArg(0)); 7158 Value *Control = llvm::ConstantInt::get(Int32Ty, 0xff0c); 7159 Value *F = CGM.getIntrinsic(Intrinsic::s390_tbegin); 7160 return Builder.CreateCall(F, {TDB, Control}); 7161 } 7162 case SystemZ::BI__builtin_tbegin_nofloat: { 7163 Value *TDB = EmitScalarExpr(E->getArg(0)); 7164 Value *Control = llvm::ConstantInt::get(Int32Ty, 0xff0c); 7165 Value *F = CGM.getIntrinsic(Intrinsic::s390_tbegin_nofloat); 7166 return Builder.CreateCall(F, {TDB, Control}); 7167 } 7168 case SystemZ::BI__builtin_tbeginc: { 7169 Value *TDB = llvm::ConstantPointerNull::get(Int8PtrTy); 7170 Value *Control = llvm::ConstantInt::get(Int32Ty, 0xff08); 7171 Value *F = CGM.getIntrinsic(Intrinsic::s390_tbeginc); 7172 return Builder.CreateCall(F, {TDB, Control}); 7173 } 7174 case SystemZ::BI__builtin_tabort: { 7175 Value *Data = EmitScalarExpr(E->getArg(0)); 7176 Value *F = CGM.getIntrinsic(Intrinsic::s390_tabort); 7177 return Builder.CreateCall(F, Builder.CreateSExt(Data, Int64Ty, "tabort")); 7178 } 7179 case SystemZ::BI__builtin_non_tx_store: { 7180 Value *Address = EmitScalarExpr(E->getArg(0)); 7181 Value *Data = EmitScalarExpr(E->getArg(1)); 7182 Value *F = CGM.getIntrinsic(Intrinsic::s390_ntstg); 7183 return Builder.CreateCall(F, {Data, Address}); 7184 } 7185 7186 // Vector builtins. Note that most vector builtins are mapped automatically 7187 // to target-specific LLVM intrinsics. The ones handled specially here can 7188 // be represented via standard LLVM IR, which is preferable to enable common 7189 // LLVM optimizations. 7190 7191 case SystemZ::BI__builtin_s390_vpopctb: 7192 case SystemZ::BI__builtin_s390_vpopcth: 7193 case SystemZ::BI__builtin_s390_vpopctf: 7194 case SystemZ::BI__builtin_s390_vpopctg: { 7195 llvm::Type *ResultType = ConvertType(E->getType()); 7196 Value *X = EmitScalarExpr(E->getArg(0)); 7197 Function *F = CGM.getIntrinsic(Intrinsic::ctpop, ResultType); 7198 return Builder.CreateCall(F, X); 7199 } 7200 7201 case SystemZ::BI__builtin_s390_vclzb: 7202 case SystemZ::BI__builtin_s390_vclzh: 7203 case SystemZ::BI__builtin_s390_vclzf: 7204 case SystemZ::BI__builtin_s390_vclzg: { 7205 llvm::Type *ResultType = ConvertType(E->getType()); 7206 Value *X = EmitScalarExpr(E->getArg(0)); 7207 Value *Undef = ConstantInt::get(Builder.getInt1Ty(), false); 7208 Function *F = CGM.getIntrinsic(Intrinsic::ctlz, ResultType); 7209 return Builder.CreateCall(F, {X, Undef}); 7210 } 7211 7212 case SystemZ::BI__builtin_s390_vctzb: 7213 case SystemZ::BI__builtin_s390_vctzh: 7214 case SystemZ::BI__builtin_s390_vctzf: 7215 case SystemZ::BI__builtin_s390_vctzg: { 7216 llvm::Type *ResultType = ConvertType(E->getType()); 7217 Value *X = EmitScalarExpr(E->getArg(0)); 7218 Value *Undef = ConstantInt::get(Builder.getInt1Ty(), false); 7219 Function *F = CGM.getIntrinsic(Intrinsic::cttz, ResultType); 7220 return Builder.CreateCall(F, {X, Undef}); 7221 } 7222 7223 case SystemZ::BI__builtin_s390_vfsqdb: { 7224 llvm::Type *ResultType = ConvertType(E->getType()); 7225 Value *X = EmitScalarExpr(E->getArg(0)); 7226 Function *F = CGM.getIntrinsic(Intrinsic::sqrt, ResultType); 7227 return Builder.CreateCall(F, X); 7228 } 7229 case SystemZ::BI__builtin_s390_vfmadb: { 7230 llvm::Type *ResultType = ConvertType(E->getType()); 7231 Value *X = EmitScalarExpr(E->getArg(0)); 7232 Value *Y = EmitScalarExpr(E->getArg(1)); 7233 Value *Z = EmitScalarExpr(E->getArg(2)); 7234 Function *F = CGM.getIntrinsic(Intrinsic::fma, ResultType); 7235 return Builder.CreateCall(F, {X, Y, Z}); 7236 } 7237 case SystemZ::BI__builtin_s390_vfmsdb: { 7238 llvm::Type *ResultType = ConvertType(E->getType()); 7239 Value *X = EmitScalarExpr(E->getArg(0)); 7240 Value *Y = EmitScalarExpr(E->getArg(1)); 7241 Value *Z = EmitScalarExpr(E->getArg(2)); 7242 Value *Zero = llvm::ConstantFP::getZeroValueForNegation(ResultType); 7243 Function *F = CGM.getIntrinsic(Intrinsic::fma, ResultType); 7244 return Builder.CreateCall(F, {X, Y, Builder.CreateFSub(Zero, Z, "sub")}); 7245 } 7246 case SystemZ::BI__builtin_s390_vflpdb: { 7247 llvm::Type *ResultType = ConvertType(E->getType()); 7248 Value *X = EmitScalarExpr(E->getArg(0)); 7249 Function *F = CGM.getIntrinsic(Intrinsic::fabs, ResultType); 7250 return Builder.CreateCall(F, X); 7251 } 7252 case SystemZ::BI__builtin_s390_vflndb: { 7253 llvm::Type *ResultType = ConvertType(E->getType()); 7254 Value *X = EmitScalarExpr(E->getArg(0)); 7255 Value *Zero = llvm::ConstantFP::getZeroValueForNegation(ResultType); 7256 Function *F = CGM.getIntrinsic(Intrinsic::fabs, ResultType); 7257 return Builder.CreateFSub(Zero, Builder.CreateCall(F, X), "sub"); 7258 } 7259 case SystemZ::BI__builtin_s390_vfidb: { 7260 llvm::Type *ResultType = ConvertType(E->getType()); 7261 Value *X = EmitScalarExpr(E->getArg(0)); 7262 // Constant-fold the M4 and M5 mask arguments. 7263 llvm::APSInt M4, M5; 7264 bool IsConstM4 = E->getArg(1)->isIntegerConstantExpr(M4, getContext()); 7265 bool IsConstM5 = E->getArg(2)->isIntegerConstantExpr(M5, getContext()); 7266 assert(IsConstM4 && IsConstM5 && "Constant arg isn't actually constant?"); 7267 (void)IsConstM4; (void)IsConstM5; 7268 // Check whether this instance of vfidb can be represented via a LLVM 7269 // standard intrinsic. We only support some combinations of M4 and M5. 7270 Intrinsic::ID ID = Intrinsic::not_intrinsic; 7271 switch (M4.getZExtValue()) { 7272 default: break; 7273 case 0: // IEEE-inexact exception allowed 7274 switch (M5.getZExtValue()) { 7275 default: break; 7276 case 0: ID = Intrinsic::rint; break; 7277 } 7278 break; 7279 case 4: // IEEE-inexact exception suppressed 7280 switch (M5.getZExtValue()) { 7281 default: break; 7282 case 0: ID = Intrinsic::nearbyint; break; 7283 case 1: ID = Intrinsic::round; break; 7284 case 5: ID = Intrinsic::trunc; break; 7285 case 6: ID = Intrinsic::ceil; break; 7286 case 7: ID = Intrinsic::floor; break; 7287 } 7288 break; 7289 } 7290 if (ID != Intrinsic::not_intrinsic) { 7291 Function *F = CGM.getIntrinsic(ID, ResultType); 7292 return Builder.CreateCall(F, X); 7293 } 7294 Function *F = CGM.getIntrinsic(Intrinsic::s390_vfidb); 7295 Value *M4Value = llvm::ConstantInt::get(getLLVMContext(), M4); 7296 Value *M5Value = llvm::ConstantInt::get(getLLVMContext(), M5); 7297 return Builder.CreateCall(F, {X, M4Value, M5Value}); 7298 } 7299 7300 // Vector intrisincs that output the post-instruction CC value. 7301 7302 #define INTRINSIC_WITH_CC(NAME) \ 7303 case SystemZ::BI__builtin_##NAME: \ 7304 return EmitSystemZIntrinsicWithCC(*this, Intrinsic::NAME, E) 7305 7306 INTRINSIC_WITH_CC(s390_vpkshs); 7307 INTRINSIC_WITH_CC(s390_vpksfs); 7308 INTRINSIC_WITH_CC(s390_vpksgs); 7309 7310 INTRINSIC_WITH_CC(s390_vpklshs); 7311 INTRINSIC_WITH_CC(s390_vpklsfs); 7312 INTRINSIC_WITH_CC(s390_vpklsgs); 7313 7314 INTRINSIC_WITH_CC(s390_vceqbs); 7315 INTRINSIC_WITH_CC(s390_vceqhs); 7316 INTRINSIC_WITH_CC(s390_vceqfs); 7317 INTRINSIC_WITH_CC(s390_vceqgs); 7318 7319 INTRINSIC_WITH_CC(s390_vchbs); 7320 INTRINSIC_WITH_CC(s390_vchhs); 7321 INTRINSIC_WITH_CC(s390_vchfs); 7322 INTRINSIC_WITH_CC(s390_vchgs); 7323 7324 INTRINSIC_WITH_CC(s390_vchlbs); 7325 INTRINSIC_WITH_CC(s390_vchlhs); 7326 INTRINSIC_WITH_CC(s390_vchlfs); 7327 INTRINSIC_WITH_CC(s390_vchlgs); 7328 7329 INTRINSIC_WITH_CC(s390_vfaebs); 7330 INTRINSIC_WITH_CC(s390_vfaehs); 7331 INTRINSIC_WITH_CC(s390_vfaefs); 7332 7333 INTRINSIC_WITH_CC(s390_vfaezbs); 7334 INTRINSIC_WITH_CC(s390_vfaezhs); 7335 INTRINSIC_WITH_CC(s390_vfaezfs); 7336 7337 INTRINSIC_WITH_CC(s390_vfeebs); 7338 INTRINSIC_WITH_CC(s390_vfeehs); 7339 INTRINSIC_WITH_CC(s390_vfeefs); 7340 7341 INTRINSIC_WITH_CC(s390_vfeezbs); 7342 INTRINSIC_WITH_CC(s390_vfeezhs); 7343 INTRINSIC_WITH_CC(s390_vfeezfs); 7344 7345 INTRINSIC_WITH_CC(s390_vfenebs); 7346 INTRINSIC_WITH_CC(s390_vfenehs); 7347 INTRINSIC_WITH_CC(s390_vfenefs); 7348 7349 INTRINSIC_WITH_CC(s390_vfenezbs); 7350 INTRINSIC_WITH_CC(s390_vfenezhs); 7351 INTRINSIC_WITH_CC(s390_vfenezfs); 7352 7353 INTRINSIC_WITH_CC(s390_vistrbs); 7354 INTRINSIC_WITH_CC(s390_vistrhs); 7355 INTRINSIC_WITH_CC(s390_vistrfs); 7356 7357 INTRINSIC_WITH_CC(s390_vstrcbs); 7358 INTRINSIC_WITH_CC(s390_vstrchs); 7359 INTRINSIC_WITH_CC(s390_vstrcfs); 7360 7361 INTRINSIC_WITH_CC(s390_vstrczbs); 7362 INTRINSIC_WITH_CC(s390_vstrczhs); 7363 INTRINSIC_WITH_CC(s390_vstrczfs); 7364 7365 INTRINSIC_WITH_CC(s390_vfcedbs); 7366 INTRINSIC_WITH_CC(s390_vfchdbs); 7367 INTRINSIC_WITH_CC(s390_vfchedbs); 7368 7369 INTRINSIC_WITH_CC(s390_vftcidb); 7370 7371 #undef INTRINSIC_WITH_CC 7372 7373 default: 7374 return nullptr; 7375 } 7376 } 7377 7378 Value *CodeGenFunction::EmitNVPTXBuiltinExpr(unsigned BuiltinID, 7379 const CallExpr *E) { 7380 switch (BuiltinID) { 7381 case NVPTX::BI__nvvm_atom_add_gen_i: 7382 case NVPTX::BI__nvvm_atom_add_gen_l: 7383 case NVPTX::BI__nvvm_atom_add_gen_ll: 7384 return MakeBinaryAtomicValue(*this, llvm::AtomicRMWInst::Add, E); 7385 7386 case NVPTX::BI__nvvm_atom_sub_gen_i: 7387 case NVPTX::BI__nvvm_atom_sub_gen_l: 7388 case NVPTX::BI__nvvm_atom_sub_gen_ll: 7389 return MakeBinaryAtomicValue(*this, llvm::AtomicRMWInst::Sub, E); 7390 7391 case NVPTX::BI__nvvm_atom_and_gen_i: 7392 case NVPTX::BI__nvvm_atom_and_gen_l: 7393 case NVPTX::BI__nvvm_atom_and_gen_ll: 7394 return MakeBinaryAtomicValue(*this, llvm::AtomicRMWInst::And, E); 7395 7396 case NVPTX::BI__nvvm_atom_or_gen_i: 7397 case NVPTX::BI__nvvm_atom_or_gen_l: 7398 case NVPTX::BI__nvvm_atom_or_gen_ll: 7399 return MakeBinaryAtomicValue(*this, llvm::AtomicRMWInst::Or, E); 7400 7401 case NVPTX::BI__nvvm_atom_xor_gen_i: 7402 case NVPTX::BI__nvvm_atom_xor_gen_l: 7403 case NVPTX::BI__nvvm_atom_xor_gen_ll: 7404 return MakeBinaryAtomicValue(*this, llvm::AtomicRMWInst::Xor, E); 7405 7406 case NVPTX::BI__nvvm_atom_xchg_gen_i: 7407 case NVPTX::BI__nvvm_atom_xchg_gen_l: 7408 case NVPTX::BI__nvvm_atom_xchg_gen_ll: 7409 return MakeBinaryAtomicValue(*this, llvm::AtomicRMWInst::Xchg, E); 7410 7411 case NVPTX::BI__nvvm_atom_max_gen_i: 7412 case NVPTX::BI__nvvm_atom_max_gen_l: 7413 case NVPTX::BI__nvvm_atom_max_gen_ll: 7414 return MakeBinaryAtomicValue(*this, llvm::AtomicRMWInst::Max, E); 7415 7416 case NVPTX::BI__nvvm_atom_max_gen_ui: 7417 case NVPTX::BI__nvvm_atom_max_gen_ul: 7418 case NVPTX::BI__nvvm_atom_max_gen_ull: 7419 return MakeBinaryAtomicValue(*this, llvm::AtomicRMWInst::UMax, E); 7420 7421 case NVPTX::BI__nvvm_atom_min_gen_i: 7422 case NVPTX::BI__nvvm_atom_min_gen_l: 7423 case NVPTX::BI__nvvm_atom_min_gen_ll: 7424 return MakeBinaryAtomicValue(*this, llvm::AtomicRMWInst::Min, E); 7425 7426 case NVPTX::BI__nvvm_atom_min_gen_ui: 7427 case NVPTX::BI__nvvm_atom_min_gen_ul: 7428 case NVPTX::BI__nvvm_atom_min_gen_ull: 7429 return MakeBinaryAtomicValue(*this, llvm::AtomicRMWInst::UMin, E); 7430 7431 case NVPTX::BI__nvvm_atom_cas_gen_i: 7432 case NVPTX::BI__nvvm_atom_cas_gen_l: 7433 case NVPTX::BI__nvvm_atom_cas_gen_ll: 7434 // __nvvm_atom_cas_gen_* should return the old value rather than the 7435 // success flag. 7436 return MakeAtomicCmpXchgValue(*this, E, /*ReturnBool=*/false); 7437 7438 case NVPTX::BI__nvvm_atom_add_gen_f: { 7439 Value *Ptr = EmitScalarExpr(E->getArg(0)); 7440 Value *Val = EmitScalarExpr(E->getArg(1)); 7441 // atomicrmw only deals with integer arguments so we need to use 7442 // LLVM's nvvm_atomic_load_add_f32 intrinsic for that. 7443 Value *FnALAF32 = 7444 CGM.getIntrinsic(Intrinsic::nvvm_atomic_load_add_f32, Ptr->getType()); 7445 return Builder.CreateCall(FnALAF32, {Ptr, Val}); 7446 } 7447 7448 case NVPTX::BI__nvvm_atom_inc_gen_ui: { 7449 Value *Ptr = EmitScalarExpr(E->getArg(0)); 7450 Value *Val = EmitScalarExpr(E->getArg(1)); 7451 Value *FnALI32 = 7452 CGM.getIntrinsic(Intrinsic::nvvm_atomic_load_inc_32, Ptr->getType()); 7453 return Builder.CreateCall(FnALI32, {Ptr, Val}); 7454 } 7455 7456 case NVPTX::BI__nvvm_atom_dec_gen_ui: { 7457 Value *Ptr = EmitScalarExpr(E->getArg(0)); 7458 Value *Val = EmitScalarExpr(E->getArg(1)); 7459 Value *FnALD32 = 7460 CGM.getIntrinsic(Intrinsic::nvvm_atomic_load_dec_32, Ptr->getType()); 7461 return Builder.CreateCall(FnALD32, {Ptr, Val}); 7462 } 7463 7464 default: 7465 return nullptr; 7466 } 7467 } 7468 7469 Value *CodeGenFunction::EmitWebAssemblyBuiltinExpr(unsigned BuiltinID, 7470 const CallExpr *E) { 7471 switch (BuiltinID) { 7472 case WebAssembly::BI__builtin_wasm_memory_size: { 7473 llvm::Type *ResultType = ConvertType(E->getType()); 7474 Value *Callee = CGM.getIntrinsic(Intrinsic::wasm_memory_size, ResultType); 7475 return Builder.CreateCall(Callee); 7476 } 7477 case WebAssembly::BI__builtin_wasm_grow_memory: { 7478 Value *X = EmitScalarExpr(E->getArg(0)); 7479 Value *Callee = CGM.getIntrinsic(Intrinsic::wasm_grow_memory, X->getType()); 7480 return Builder.CreateCall(Callee, X); 7481 } 7482 7483 default: 7484 return nullptr; 7485 } 7486 } 7487