1 //===---- CGBuiltin.cpp - Emit LLVM Code for builtins ---------------------===// 2 // 3 // The LLVM Compiler Infrastructure 4 // 5 // This file is distributed under the University of Illinois Open Source 6 // License. See LICENSE.TXT for details. 7 // 8 //===----------------------------------------------------------------------===// 9 // 10 // This contains code to emit Builtin calls as LLVM code. 11 // 12 //===----------------------------------------------------------------------===// 13 14 #include "CodeGenFunction.h" 15 #include "CGCXXABI.h" 16 #include "CGObjCRuntime.h" 17 #include "CodeGenModule.h" 18 #include "TargetInfo.h" 19 #include "clang/AST/ASTContext.h" 20 #include "clang/AST/Decl.h" 21 #include "clang/Basic/TargetBuiltins.h" 22 #include "clang/Basic/TargetInfo.h" 23 #include "clang/CodeGen/CGFunctionInfo.h" 24 #include "llvm/ADT/StringExtras.h" 25 #include "llvm/IR/CallSite.h" 26 #include "llvm/IR/DataLayout.h" 27 #include "llvm/IR/InlineAsm.h" 28 #include "llvm/IR/Intrinsics.h" 29 #include <sstream> 30 31 using namespace clang; 32 using namespace CodeGen; 33 using namespace llvm; 34 35 /// getBuiltinLibFunction - Given a builtin id for a function like 36 /// "__builtin_fabsf", return a Function* for "fabsf". 37 llvm::Value *CodeGenModule::getBuiltinLibFunction(const FunctionDecl *FD, 38 unsigned BuiltinID) { 39 assert(Context.BuiltinInfo.isLibFunction(BuiltinID)); 40 41 // Get the name, skip over the __builtin_ prefix (if necessary). 42 StringRef Name; 43 GlobalDecl D(FD); 44 45 // If the builtin has been declared explicitly with an assembler label, 46 // use the mangled name. This differs from the plain label on platforms 47 // that prefix labels. 48 if (FD->hasAttr<AsmLabelAttr>()) 49 Name = getMangledName(D); 50 else 51 Name = Context.BuiltinInfo.getName(BuiltinID) + 10; 52 53 llvm::FunctionType *Ty = 54 cast<llvm::FunctionType>(getTypes().ConvertType(FD->getType())); 55 56 return GetOrCreateLLVMFunction(Name, Ty, D, /*ForVTable=*/false); 57 } 58 59 /// Emit the conversions required to turn the given value into an 60 /// integer of the given size. 61 static Value *EmitToInt(CodeGenFunction &CGF, llvm::Value *V, 62 QualType T, llvm::IntegerType *IntType) { 63 V = CGF.EmitToMemory(V, T); 64 65 if (V->getType()->isPointerTy()) 66 return CGF.Builder.CreatePtrToInt(V, IntType); 67 68 assert(V->getType() == IntType); 69 return V; 70 } 71 72 static Value *EmitFromInt(CodeGenFunction &CGF, llvm::Value *V, 73 QualType T, llvm::Type *ResultType) { 74 V = CGF.EmitFromMemory(V, T); 75 76 if (ResultType->isPointerTy()) 77 return CGF.Builder.CreateIntToPtr(V, ResultType); 78 79 assert(V->getType() == ResultType); 80 return V; 81 } 82 83 /// Utility to insert an atomic instruction based on Instrinsic::ID 84 /// and the expression node. 85 static Value *MakeBinaryAtomicValue(CodeGenFunction &CGF, 86 llvm::AtomicRMWInst::BinOp Kind, 87 const CallExpr *E) { 88 QualType T = E->getType(); 89 assert(E->getArg(0)->getType()->isPointerType()); 90 assert(CGF.getContext().hasSameUnqualifiedType(T, 91 E->getArg(0)->getType()->getPointeeType())); 92 assert(CGF.getContext().hasSameUnqualifiedType(T, E->getArg(1)->getType())); 93 94 llvm::Value *DestPtr = CGF.EmitScalarExpr(E->getArg(0)); 95 unsigned AddrSpace = DestPtr->getType()->getPointerAddressSpace(); 96 97 llvm::IntegerType *IntType = 98 llvm::IntegerType::get(CGF.getLLVMContext(), 99 CGF.getContext().getTypeSize(T)); 100 llvm::Type *IntPtrType = IntType->getPointerTo(AddrSpace); 101 102 llvm::Value *Args[2]; 103 Args[0] = CGF.Builder.CreateBitCast(DestPtr, IntPtrType); 104 Args[1] = CGF.EmitScalarExpr(E->getArg(1)); 105 llvm::Type *ValueType = Args[1]->getType(); 106 Args[1] = EmitToInt(CGF, Args[1], T, IntType); 107 108 llvm::Value *Result = 109 CGF.Builder.CreateAtomicRMW(Kind, Args[0], Args[1], 110 llvm::SequentiallyConsistent); 111 return EmitFromInt(CGF, Result, T, ValueType); 112 } 113 114 static Value *EmitNontemporalStore(CodeGenFunction &CGF, const CallExpr *E) { 115 Value *Val = CGF.EmitScalarExpr(E->getArg(0)); 116 Value *Address = CGF.EmitScalarExpr(E->getArg(1)); 117 118 // Convert the type of the pointer to a pointer to the stored type. 119 Val = CGF.EmitToMemory(Val, E->getArg(0)->getType()); 120 Value *BC = CGF.Builder.CreateBitCast( 121 Address, llvm::PointerType::getUnqual(Val->getType()), "cast"); 122 LValue LV = CGF.MakeNaturalAlignAddrLValue(BC, E->getArg(0)->getType()); 123 LV.setNontemporal(true); 124 CGF.EmitStoreOfScalar(Val, LV, false); 125 return nullptr; 126 } 127 128 static Value *EmitNontemporalLoad(CodeGenFunction &CGF, const CallExpr *E) { 129 Value *Address = CGF.EmitScalarExpr(E->getArg(0)); 130 131 LValue LV = CGF.MakeNaturalAlignAddrLValue(Address, E->getType()); 132 LV.setNontemporal(true); 133 return CGF.EmitLoadOfScalar(LV, E->getExprLoc()); 134 } 135 136 static RValue EmitBinaryAtomic(CodeGenFunction &CGF, 137 llvm::AtomicRMWInst::BinOp Kind, 138 const CallExpr *E) { 139 return RValue::get(MakeBinaryAtomicValue(CGF, Kind, E)); 140 } 141 142 /// Utility to insert an atomic instruction based Instrinsic::ID and 143 /// the expression node, where the return value is the result of the 144 /// operation. 145 static RValue EmitBinaryAtomicPost(CodeGenFunction &CGF, 146 llvm::AtomicRMWInst::BinOp Kind, 147 const CallExpr *E, 148 Instruction::BinaryOps Op, 149 bool Invert = false) { 150 QualType T = E->getType(); 151 assert(E->getArg(0)->getType()->isPointerType()); 152 assert(CGF.getContext().hasSameUnqualifiedType(T, 153 E->getArg(0)->getType()->getPointeeType())); 154 assert(CGF.getContext().hasSameUnqualifiedType(T, E->getArg(1)->getType())); 155 156 llvm::Value *DestPtr = CGF.EmitScalarExpr(E->getArg(0)); 157 unsigned AddrSpace = DestPtr->getType()->getPointerAddressSpace(); 158 159 llvm::IntegerType *IntType = 160 llvm::IntegerType::get(CGF.getLLVMContext(), 161 CGF.getContext().getTypeSize(T)); 162 llvm::Type *IntPtrType = IntType->getPointerTo(AddrSpace); 163 164 llvm::Value *Args[2]; 165 Args[1] = CGF.EmitScalarExpr(E->getArg(1)); 166 llvm::Type *ValueType = Args[1]->getType(); 167 Args[1] = EmitToInt(CGF, Args[1], T, IntType); 168 Args[0] = CGF.Builder.CreateBitCast(DestPtr, IntPtrType); 169 170 llvm::Value *Result = 171 CGF.Builder.CreateAtomicRMW(Kind, Args[0], Args[1], 172 llvm::SequentiallyConsistent); 173 Result = CGF.Builder.CreateBinOp(Op, Result, Args[1]); 174 if (Invert) 175 Result = CGF.Builder.CreateBinOp(llvm::Instruction::Xor, Result, 176 llvm::ConstantInt::get(IntType, -1)); 177 Result = EmitFromInt(CGF, Result, T, ValueType); 178 return RValue::get(Result); 179 } 180 181 /// @brief Utility to insert an atomic cmpxchg instruction. 182 /// 183 /// @param CGF The current codegen function. 184 /// @param E Builtin call expression to convert to cmpxchg. 185 /// arg0 - address to operate on 186 /// arg1 - value to compare with 187 /// arg2 - new value 188 /// @param ReturnBool Specifies whether to return success flag of 189 /// cmpxchg result or the old value. 190 /// 191 /// @returns result of cmpxchg, according to ReturnBool 192 static Value *MakeAtomicCmpXchgValue(CodeGenFunction &CGF, const CallExpr *E, 193 bool ReturnBool) { 194 QualType T = ReturnBool ? E->getArg(1)->getType() : E->getType(); 195 llvm::Value *DestPtr = CGF.EmitScalarExpr(E->getArg(0)); 196 unsigned AddrSpace = DestPtr->getType()->getPointerAddressSpace(); 197 198 llvm::IntegerType *IntType = llvm::IntegerType::get( 199 CGF.getLLVMContext(), CGF.getContext().getTypeSize(T)); 200 llvm::Type *IntPtrType = IntType->getPointerTo(AddrSpace); 201 202 Value *Args[3]; 203 Args[0] = CGF.Builder.CreateBitCast(DestPtr, IntPtrType); 204 Args[1] = CGF.EmitScalarExpr(E->getArg(1)); 205 llvm::Type *ValueType = Args[1]->getType(); 206 Args[1] = EmitToInt(CGF, Args[1], T, IntType); 207 Args[2] = EmitToInt(CGF, CGF.EmitScalarExpr(E->getArg(2)), T, IntType); 208 209 Value *Pair = CGF.Builder.CreateAtomicCmpXchg(Args[0], Args[1], Args[2], 210 llvm::SequentiallyConsistent, 211 llvm::SequentiallyConsistent); 212 if (ReturnBool) 213 // Extract boolean success flag and zext it to int. 214 return CGF.Builder.CreateZExt(CGF.Builder.CreateExtractValue(Pair, 1), 215 CGF.ConvertType(E->getType())); 216 else 217 // Extract old value and emit it using the same type as compare value. 218 return EmitFromInt(CGF, CGF.Builder.CreateExtractValue(Pair, 0), T, 219 ValueType); 220 } 221 222 /// EmitFAbs - Emit a call to @llvm.fabs(). 223 static Value *EmitFAbs(CodeGenFunction &CGF, Value *V) { 224 Value *F = CGF.CGM.getIntrinsic(Intrinsic::fabs, V->getType()); 225 llvm::CallInst *Call = CGF.Builder.CreateCall(F, V); 226 Call->setDoesNotAccessMemory(); 227 return Call; 228 } 229 230 /// Emit the computation of the sign bit for a floating point value. Returns 231 /// the i1 sign bit value. 232 static Value *EmitSignBit(CodeGenFunction &CGF, Value *V) { 233 LLVMContext &C = CGF.CGM.getLLVMContext(); 234 235 llvm::Type *Ty = V->getType(); 236 int Width = Ty->getPrimitiveSizeInBits(); 237 llvm::Type *IntTy = llvm::IntegerType::get(C, Width); 238 V = CGF.Builder.CreateBitCast(V, IntTy); 239 if (Ty->isPPC_FP128Ty()) { 240 // We want the sign bit of the higher-order double. The bitcast we just 241 // did works as if the double-double was stored to memory and then 242 // read as an i128. The "store" will put the higher-order double in the 243 // lower address in both little- and big-Endian modes, but the "load" 244 // will treat those bits as a different part of the i128: the low bits in 245 // little-Endian, the high bits in big-Endian. Therefore, on big-Endian 246 // we need to shift the high bits down to the low before truncating. 247 Width >>= 1; 248 if (CGF.getTarget().isBigEndian()) { 249 Value *ShiftCst = llvm::ConstantInt::get(IntTy, Width); 250 V = CGF.Builder.CreateLShr(V, ShiftCst); 251 } 252 // We are truncating value in order to extract the higher-order 253 // double, which we will be using to extract the sign from. 254 IntTy = llvm::IntegerType::get(C, Width); 255 V = CGF.Builder.CreateTrunc(V, IntTy); 256 } 257 Value *Zero = llvm::Constant::getNullValue(IntTy); 258 return CGF.Builder.CreateICmpSLT(V, Zero); 259 } 260 261 static RValue emitLibraryCall(CodeGenFunction &CGF, const FunctionDecl *Fn, 262 const CallExpr *E, llvm::Value *calleeValue) { 263 return CGF.EmitCall(E->getCallee()->getType(), calleeValue, E, 264 ReturnValueSlot(), Fn); 265 } 266 267 /// \brief Emit a call to llvm.{sadd,uadd,ssub,usub,smul,umul}.with.overflow.* 268 /// depending on IntrinsicID. 269 /// 270 /// \arg CGF The current codegen function. 271 /// \arg IntrinsicID The ID for the Intrinsic we wish to generate. 272 /// \arg X The first argument to the llvm.*.with.overflow.*. 273 /// \arg Y The second argument to the llvm.*.with.overflow.*. 274 /// \arg Carry The carry returned by the llvm.*.with.overflow.*. 275 /// \returns The result (i.e. sum/product) returned by the intrinsic. 276 static llvm::Value *EmitOverflowIntrinsic(CodeGenFunction &CGF, 277 const llvm::Intrinsic::ID IntrinsicID, 278 llvm::Value *X, llvm::Value *Y, 279 llvm::Value *&Carry) { 280 // Make sure we have integers of the same width. 281 assert(X->getType() == Y->getType() && 282 "Arguments must be the same type. (Did you forget to make sure both " 283 "arguments have the same integer width?)"); 284 285 llvm::Value *Callee = CGF.CGM.getIntrinsic(IntrinsicID, X->getType()); 286 llvm::Value *Tmp = CGF.Builder.CreateCall(Callee, {X, Y}); 287 Carry = CGF.Builder.CreateExtractValue(Tmp, 1); 288 return CGF.Builder.CreateExtractValue(Tmp, 0); 289 } 290 291 namespace { 292 struct WidthAndSignedness { 293 unsigned Width; 294 bool Signed; 295 }; 296 } 297 298 static WidthAndSignedness 299 getIntegerWidthAndSignedness(const clang::ASTContext &context, 300 const clang::QualType Type) { 301 assert(Type->isIntegerType() && "Given type is not an integer."); 302 unsigned Width = Type->isBooleanType() ? 1 : context.getTypeInfo(Type).Width; 303 bool Signed = Type->isSignedIntegerType(); 304 return {Width, Signed}; 305 } 306 307 // Given one or more integer types, this function produces an integer type that 308 // encompasses them: any value in one of the given types could be expressed in 309 // the encompassing type. 310 static struct WidthAndSignedness 311 EncompassingIntegerType(ArrayRef<struct WidthAndSignedness> Types) { 312 assert(Types.size() > 0 && "Empty list of types."); 313 314 // If any of the given types is signed, we must return a signed type. 315 bool Signed = false; 316 for (const auto &Type : Types) { 317 Signed |= Type.Signed; 318 } 319 320 // The encompassing type must have a width greater than or equal to the width 321 // of the specified types. Aditionally, if the encompassing type is signed, 322 // its width must be strictly greater than the width of any unsigned types 323 // given. 324 unsigned Width = 0; 325 for (const auto &Type : Types) { 326 unsigned MinWidth = Type.Width + (Signed && !Type.Signed); 327 if (Width < MinWidth) { 328 Width = MinWidth; 329 } 330 } 331 332 return {Width, Signed}; 333 } 334 335 Value *CodeGenFunction::EmitVAStartEnd(Value *ArgValue, bool IsStart) { 336 llvm::Type *DestType = Int8PtrTy; 337 if (ArgValue->getType() != DestType) 338 ArgValue = 339 Builder.CreateBitCast(ArgValue, DestType, ArgValue->getName().data()); 340 341 Intrinsic::ID inst = IsStart ? Intrinsic::vastart : Intrinsic::vaend; 342 return Builder.CreateCall(CGM.getIntrinsic(inst), ArgValue); 343 } 344 345 RValue CodeGenFunction::EmitBuiltinExpr(const FunctionDecl *FD, 346 unsigned BuiltinID, const CallExpr *E, 347 ReturnValueSlot ReturnValue) { 348 // See if we can constant fold this builtin. If so, don't emit it at all. 349 Expr::EvalResult Result; 350 if (E->EvaluateAsRValue(Result, CGM.getContext()) && 351 !Result.hasSideEffects()) { 352 if (Result.Val.isInt()) 353 return RValue::get(llvm::ConstantInt::get(getLLVMContext(), 354 Result.Val.getInt())); 355 if (Result.Val.isFloat()) 356 return RValue::get(llvm::ConstantFP::get(getLLVMContext(), 357 Result.Val.getFloat())); 358 } 359 360 switch (BuiltinID) { 361 default: break; // Handle intrinsics and libm functions below. 362 case Builtin::BI__builtin___CFStringMakeConstantString: 363 case Builtin::BI__builtin___NSStringMakeConstantString: 364 return RValue::get(CGM.EmitConstantExpr(E, E->getType(), nullptr)); 365 case Builtin::BI__builtin_stdarg_start: 366 case Builtin::BI__builtin_va_start: 367 case Builtin::BI__va_start: 368 case Builtin::BI__builtin_va_end: 369 return RValue::get( 370 EmitVAStartEnd(BuiltinID == Builtin::BI__va_start 371 ? EmitScalarExpr(E->getArg(0)) 372 : EmitVAListRef(E->getArg(0)).getPointer(), 373 BuiltinID != Builtin::BI__builtin_va_end)); 374 case Builtin::BI__builtin_va_copy: { 375 Value *DstPtr = EmitVAListRef(E->getArg(0)).getPointer(); 376 Value *SrcPtr = EmitVAListRef(E->getArg(1)).getPointer(); 377 378 llvm::Type *Type = Int8PtrTy; 379 380 DstPtr = Builder.CreateBitCast(DstPtr, Type); 381 SrcPtr = Builder.CreateBitCast(SrcPtr, Type); 382 return RValue::get(Builder.CreateCall(CGM.getIntrinsic(Intrinsic::vacopy), 383 {DstPtr, SrcPtr})); 384 } 385 case Builtin::BI__builtin_abs: 386 case Builtin::BI__builtin_labs: 387 case Builtin::BI__builtin_llabs: { 388 Value *ArgValue = EmitScalarExpr(E->getArg(0)); 389 390 Value *NegOp = Builder.CreateNeg(ArgValue, "neg"); 391 Value *CmpResult = 392 Builder.CreateICmpSGE(ArgValue, 393 llvm::Constant::getNullValue(ArgValue->getType()), 394 "abscond"); 395 Value *Result = 396 Builder.CreateSelect(CmpResult, ArgValue, NegOp, "abs"); 397 398 return RValue::get(Result); 399 } 400 case Builtin::BI__builtin_fabs: 401 case Builtin::BI__builtin_fabsf: 402 case Builtin::BI__builtin_fabsl: { 403 Value *Arg1 = EmitScalarExpr(E->getArg(0)); 404 Value *Result = EmitFAbs(*this, Arg1); 405 return RValue::get(Result); 406 } 407 case Builtin::BI__builtin_fmod: 408 case Builtin::BI__builtin_fmodf: 409 case Builtin::BI__builtin_fmodl: { 410 Value *Arg1 = EmitScalarExpr(E->getArg(0)); 411 Value *Arg2 = EmitScalarExpr(E->getArg(1)); 412 Value *Result = Builder.CreateFRem(Arg1, Arg2, "fmod"); 413 return RValue::get(Result); 414 } 415 416 case Builtin::BI__builtin_conj: 417 case Builtin::BI__builtin_conjf: 418 case Builtin::BI__builtin_conjl: { 419 ComplexPairTy ComplexVal = EmitComplexExpr(E->getArg(0)); 420 Value *Real = ComplexVal.first; 421 Value *Imag = ComplexVal.second; 422 Value *Zero = 423 Imag->getType()->isFPOrFPVectorTy() 424 ? llvm::ConstantFP::getZeroValueForNegation(Imag->getType()) 425 : llvm::Constant::getNullValue(Imag->getType()); 426 427 Imag = Builder.CreateFSub(Zero, Imag, "sub"); 428 return RValue::getComplex(std::make_pair(Real, Imag)); 429 } 430 case Builtin::BI__builtin_creal: 431 case Builtin::BI__builtin_crealf: 432 case Builtin::BI__builtin_creall: 433 case Builtin::BIcreal: 434 case Builtin::BIcrealf: 435 case Builtin::BIcreall: { 436 ComplexPairTy ComplexVal = EmitComplexExpr(E->getArg(0)); 437 return RValue::get(ComplexVal.first); 438 } 439 440 case Builtin::BI__builtin_cimag: 441 case Builtin::BI__builtin_cimagf: 442 case Builtin::BI__builtin_cimagl: 443 case Builtin::BIcimag: 444 case Builtin::BIcimagf: 445 case Builtin::BIcimagl: { 446 ComplexPairTy ComplexVal = EmitComplexExpr(E->getArg(0)); 447 return RValue::get(ComplexVal.second); 448 } 449 450 case Builtin::BI__builtin_ctzs: 451 case Builtin::BI__builtin_ctz: 452 case Builtin::BI__builtin_ctzl: 453 case Builtin::BI__builtin_ctzll: { 454 Value *ArgValue = EmitScalarExpr(E->getArg(0)); 455 456 llvm::Type *ArgType = ArgValue->getType(); 457 Value *F = CGM.getIntrinsic(Intrinsic::cttz, ArgType); 458 459 llvm::Type *ResultType = ConvertType(E->getType()); 460 Value *ZeroUndef = Builder.getInt1(getTarget().isCLZForZeroUndef()); 461 Value *Result = Builder.CreateCall(F, {ArgValue, ZeroUndef}); 462 if (Result->getType() != ResultType) 463 Result = Builder.CreateIntCast(Result, ResultType, /*isSigned*/true, 464 "cast"); 465 return RValue::get(Result); 466 } 467 case Builtin::BI__builtin_clzs: 468 case Builtin::BI__builtin_clz: 469 case Builtin::BI__builtin_clzl: 470 case Builtin::BI__builtin_clzll: { 471 Value *ArgValue = EmitScalarExpr(E->getArg(0)); 472 473 llvm::Type *ArgType = ArgValue->getType(); 474 Value *F = CGM.getIntrinsic(Intrinsic::ctlz, ArgType); 475 476 llvm::Type *ResultType = ConvertType(E->getType()); 477 Value *ZeroUndef = Builder.getInt1(getTarget().isCLZForZeroUndef()); 478 Value *Result = Builder.CreateCall(F, {ArgValue, ZeroUndef}); 479 if (Result->getType() != ResultType) 480 Result = Builder.CreateIntCast(Result, ResultType, /*isSigned*/true, 481 "cast"); 482 return RValue::get(Result); 483 } 484 case Builtin::BI__builtin_ffs: 485 case Builtin::BI__builtin_ffsl: 486 case Builtin::BI__builtin_ffsll: { 487 // ffs(x) -> x ? cttz(x) + 1 : 0 488 Value *ArgValue = EmitScalarExpr(E->getArg(0)); 489 490 llvm::Type *ArgType = ArgValue->getType(); 491 Value *F = CGM.getIntrinsic(Intrinsic::cttz, ArgType); 492 493 llvm::Type *ResultType = ConvertType(E->getType()); 494 Value *Tmp = 495 Builder.CreateAdd(Builder.CreateCall(F, {ArgValue, Builder.getTrue()}), 496 llvm::ConstantInt::get(ArgType, 1)); 497 Value *Zero = llvm::Constant::getNullValue(ArgType); 498 Value *IsZero = Builder.CreateICmpEQ(ArgValue, Zero, "iszero"); 499 Value *Result = Builder.CreateSelect(IsZero, Zero, Tmp, "ffs"); 500 if (Result->getType() != ResultType) 501 Result = Builder.CreateIntCast(Result, ResultType, /*isSigned*/true, 502 "cast"); 503 return RValue::get(Result); 504 } 505 case Builtin::BI__builtin_parity: 506 case Builtin::BI__builtin_parityl: 507 case Builtin::BI__builtin_parityll: { 508 // parity(x) -> ctpop(x) & 1 509 Value *ArgValue = EmitScalarExpr(E->getArg(0)); 510 511 llvm::Type *ArgType = ArgValue->getType(); 512 Value *F = CGM.getIntrinsic(Intrinsic::ctpop, ArgType); 513 514 llvm::Type *ResultType = ConvertType(E->getType()); 515 Value *Tmp = Builder.CreateCall(F, ArgValue); 516 Value *Result = Builder.CreateAnd(Tmp, llvm::ConstantInt::get(ArgType, 1)); 517 if (Result->getType() != ResultType) 518 Result = Builder.CreateIntCast(Result, ResultType, /*isSigned*/true, 519 "cast"); 520 return RValue::get(Result); 521 } 522 case Builtin::BI__builtin_popcount: 523 case Builtin::BI__builtin_popcountl: 524 case Builtin::BI__builtin_popcountll: { 525 Value *ArgValue = EmitScalarExpr(E->getArg(0)); 526 527 llvm::Type *ArgType = ArgValue->getType(); 528 Value *F = CGM.getIntrinsic(Intrinsic::ctpop, ArgType); 529 530 llvm::Type *ResultType = ConvertType(E->getType()); 531 Value *Result = Builder.CreateCall(F, ArgValue); 532 if (Result->getType() != ResultType) 533 Result = Builder.CreateIntCast(Result, ResultType, /*isSigned*/true, 534 "cast"); 535 return RValue::get(Result); 536 } 537 case Builtin::BI__builtin_unpredictable: { 538 // Always return the argument of __builtin_unpredictable. LLVM does not 539 // handle this builtin. Metadata for this builtin should be added directly 540 // to instructions such as branches or switches that use it. 541 return RValue::get(EmitScalarExpr(E->getArg(0))); 542 } 543 case Builtin::BI__builtin_expect: { 544 Value *ArgValue = EmitScalarExpr(E->getArg(0)); 545 llvm::Type *ArgType = ArgValue->getType(); 546 547 Value *ExpectedValue = EmitScalarExpr(E->getArg(1)); 548 // Don't generate llvm.expect on -O0 as the backend won't use it for 549 // anything. 550 // Note, we still IRGen ExpectedValue because it could have side-effects. 551 if (CGM.getCodeGenOpts().OptimizationLevel == 0) 552 return RValue::get(ArgValue); 553 554 Value *FnExpect = CGM.getIntrinsic(Intrinsic::expect, ArgType); 555 Value *Result = 556 Builder.CreateCall(FnExpect, {ArgValue, ExpectedValue}, "expval"); 557 return RValue::get(Result); 558 } 559 case Builtin::BI__builtin_assume_aligned: { 560 Value *PtrValue = EmitScalarExpr(E->getArg(0)); 561 Value *OffsetValue = 562 (E->getNumArgs() > 2) ? EmitScalarExpr(E->getArg(2)) : nullptr; 563 564 Value *AlignmentValue = EmitScalarExpr(E->getArg(1)); 565 ConstantInt *AlignmentCI = cast<ConstantInt>(AlignmentValue); 566 unsigned Alignment = (unsigned) AlignmentCI->getZExtValue(); 567 568 EmitAlignmentAssumption(PtrValue, Alignment, OffsetValue); 569 return RValue::get(PtrValue); 570 } 571 case Builtin::BI__assume: 572 case Builtin::BI__builtin_assume: { 573 if (E->getArg(0)->HasSideEffects(getContext())) 574 return RValue::get(nullptr); 575 576 Value *ArgValue = EmitScalarExpr(E->getArg(0)); 577 Value *FnAssume = CGM.getIntrinsic(Intrinsic::assume); 578 return RValue::get(Builder.CreateCall(FnAssume, ArgValue)); 579 } 580 case Builtin::BI__builtin_bswap16: 581 case Builtin::BI__builtin_bswap32: 582 case Builtin::BI__builtin_bswap64: { 583 Value *ArgValue = EmitScalarExpr(E->getArg(0)); 584 llvm::Type *ArgType = ArgValue->getType(); 585 Value *F = CGM.getIntrinsic(Intrinsic::bswap, ArgType); 586 return RValue::get(Builder.CreateCall(F, ArgValue)); 587 } 588 case Builtin::BI__builtin_object_size: { 589 // We rely on constant folding to deal with expressions with side effects. 590 assert(!E->getArg(0)->HasSideEffects(getContext()) && 591 "should have been constant folded"); 592 593 // We pass this builtin onto the optimizer so that it can 594 // figure out the object size in more complex cases. 595 llvm::Type *ResType = ConvertType(E->getType()); 596 597 // LLVM only supports 0 and 2, make sure that we pass along that 598 // as a boolean. 599 Value *Ty = EmitScalarExpr(E->getArg(1)); 600 ConstantInt *CI = dyn_cast<ConstantInt>(Ty); 601 assert(CI); 602 uint64_t val = CI->getZExtValue(); 603 CI = ConstantInt::get(Builder.getInt1Ty(), (val & 0x2) >> 1); 604 // FIXME: Get right address space. 605 llvm::Type *Tys[] = { ResType, Builder.getInt8PtrTy(0) }; 606 Value *F = CGM.getIntrinsic(Intrinsic::objectsize, Tys); 607 return RValue::get( 608 Builder.CreateCall(F, {EmitScalarExpr(E->getArg(0)), CI})); 609 } 610 case Builtin::BI__builtin_prefetch: { 611 Value *Locality, *RW, *Address = EmitScalarExpr(E->getArg(0)); 612 // FIXME: Technically these constants should of type 'int', yes? 613 RW = (E->getNumArgs() > 1) ? EmitScalarExpr(E->getArg(1)) : 614 llvm::ConstantInt::get(Int32Ty, 0); 615 Locality = (E->getNumArgs() > 2) ? EmitScalarExpr(E->getArg(2)) : 616 llvm::ConstantInt::get(Int32Ty, 3); 617 Value *Data = llvm::ConstantInt::get(Int32Ty, 1); 618 Value *F = CGM.getIntrinsic(Intrinsic::prefetch); 619 return RValue::get(Builder.CreateCall(F, {Address, RW, Locality, Data})); 620 } 621 case Builtin::BI__builtin_readcyclecounter: { 622 Value *F = CGM.getIntrinsic(Intrinsic::readcyclecounter); 623 return RValue::get(Builder.CreateCall(F)); 624 } 625 case Builtin::BI__builtin___clear_cache: { 626 Value *Begin = EmitScalarExpr(E->getArg(0)); 627 Value *End = EmitScalarExpr(E->getArg(1)); 628 Value *F = CGM.getIntrinsic(Intrinsic::clear_cache); 629 return RValue::get(Builder.CreateCall(F, {Begin, End})); 630 } 631 case Builtin::BI__builtin_trap: 632 return RValue::get(EmitTrapCall(Intrinsic::trap)); 633 case Builtin::BI__debugbreak: 634 return RValue::get(EmitTrapCall(Intrinsic::debugtrap)); 635 case Builtin::BI__builtin_unreachable: { 636 if (SanOpts.has(SanitizerKind::Unreachable)) { 637 SanitizerScope SanScope(this); 638 EmitCheck(std::make_pair(static_cast<llvm::Value *>(Builder.getFalse()), 639 SanitizerKind::Unreachable), 640 "builtin_unreachable", EmitCheckSourceLocation(E->getExprLoc()), 641 None); 642 } else 643 Builder.CreateUnreachable(); 644 645 // We do need to preserve an insertion point. 646 EmitBlock(createBasicBlock("unreachable.cont")); 647 648 return RValue::get(nullptr); 649 } 650 651 case Builtin::BI__builtin_powi: 652 case Builtin::BI__builtin_powif: 653 case Builtin::BI__builtin_powil: { 654 Value *Base = EmitScalarExpr(E->getArg(0)); 655 Value *Exponent = EmitScalarExpr(E->getArg(1)); 656 llvm::Type *ArgType = Base->getType(); 657 Value *F = CGM.getIntrinsic(Intrinsic::powi, ArgType); 658 return RValue::get(Builder.CreateCall(F, {Base, Exponent})); 659 } 660 661 case Builtin::BI__builtin_isgreater: 662 case Builtin::BI__builtin_isgreaterequal: 663 case Builtin::BI__builtin_isless: 664 case Builtin::BI__builtin_islessequal: 665 case Builtin::BI__builtin_islessgreater: 666 case Builtin::BI__builtin_isunordered: { 667 // Ordered comparisons: we know the arguments to these are matching scalar 668 // floating point values. 669 Value *LHS = EmitScalarExpr(E->getArg(0)); 670 Value *RHS = EmitScalarExpr(E->getArg(1)); 671 672 switch (BuiltinID) { 673 default: llvm_unreachable("Unknown ordered comparison"); 674 case Builtin::BI__builtin_isgreater: 675 LHS = Builder.CreateFCmpOGT(LHS, RHS, "cmp"); 676 break; 677 case Builtin::BI__builtin_isgreaterequal: 678 LHS = Builder.CreateFCmpOGE(LHS, RHS, "cmp"); 679 break; 680 case Builtin::BI__builtin_isless: 681 LHS = Builder.CreateFCmpOLT(LHS, RHS, "cmp"); 682 break; 683 case Builtin::BI__builtin_islessequal: 684 LHS = Builder.CreateFCmpOLE(LHS, RHS, "cmp"); 685 break; 686 case Builtin::BI__builtin_islessgreater: 687 LHS = Builder.CreateFCmpONE(LHS, RHS, "cmp"); 688 break; 689 case Builtin::BI__builtin_isunordered: 690 LHS = Builder.CreateFCmpUNO(LHS, RHS, "cmp"); 691 break; 692 } 693 // ZExt bool to int type. 694 return RValue::get(Builder.CreateZExt(LHS, ConvertType(E->getType()))); 695 } 696 case Builtin::BI__builtin_isnan: { 697 Value *V = EmitScalarExpr(E->getArg(0)); 698 V = Builder.CreateFCmpUNO(V, V, "cmp"); 699 return RValue::get(Builder.CreateZExt(V, ConvertType(E->getType()))); 700 } 701 702 case Builtin::BI__builtin_isinf: { 703 // isinf(x) --> fabs(x) == infinity 704 Value *V = EmitScalarExpr(E->getArg(0)); 705 V = EmitFAbs(*this, V); 706 707 V = Builder.CreateFCmpOEQ(V, ConstantFP::getInfinity(V->getType()),"isinf"); 708 return RValue::get(Builder.CreateZExt(V, ConvertType(E->getType()))); 709 } 710 711 case Builtin::BI__builtin_isinf_sign: { 712 // isinf_sign(x) -> fabs(x) == infinity ? (signbit(x) ? -1 : 1) : 0 713 Value *Arg = EmitScalarExpr(E->getArg(0)); 714 Value *AbsArg = EmitFAbs(*this, Arg); 715 Value *IsInf = Builder.CreateFCmpOEQ( 716 AbsArg, ConstantFP::getInfinity(Arg->getType()), "isinf"); 717 Value *IsNeg = EmitSignBit(*this, Arg); 718 719 llvm::Type *IntTy = ConvertType(E->getType()); 720 Value *Zero = Constant::getNullValue(IntTy); 721 Value *One = ConstantInt::get(IntTy, 1); 722 Value *NegativeOne = ConstantInt::get(IntTy, -1); 723 Value *SignResult = Builder.CreateSelect(IsNeg, NegativeOne, One); 724 Value *Result = Builder.CreateSelect(IsInf, SignResult, Zero); 725 return RValue::get(Result); 726 } 727 728 case Builtin::BI__builtin_isnormal: { 729 // isnormal(x) --> x == x && fabsf(x) < infinity && fabsf(x) >= float_min 730 Value *V = EmitScalarExpr(E->getArg(0)); 731 Value *Eq = Builder.CreateFCmpOEQ(V, V, "iseq"); 732 733 Value *Abs = EmitFAbs(*this, V); 734 Value *IsLessThanInf = 735 Builder.CreateFCmpULT(Abs, ConstantFP::getInfinity(V->getType()),"isinf"); 736 APFloat Smallest = APFloat::getSmallestNormalized( 737 getContext().getFloatTypeSemantics(E->getArg(0)->getType())); 738 Value *IsNormal = 739 Builder.CreateFCmpUGE(Abs, ConstantFP::get(V->getContext(), Smallest), 740 "isnormal"); 741 V = Builder.CreateAnd(Eq, IsLessThanInf, "and"); 742 V = Builder.CreateAnd(V, IsNormal, "and"); 743 return RValue::get(Builder.CreateZExt(V, ConvertType(E->getType()))); 744 } 745 746 case Builtin::BI__builtin_isfinite: { 747 // isfinite(x) --> x == x && fabs(x) != infinity; 748 Value *V = EmitScalarExpr(E->getArg(0)); 749 Value *Eq = Builder.CreateFCmpOEQ(V, V, "iseq"); 750 751 Value *Abs = EmitFAbs(*this, V); 752 Value *IsNotInf = 753 Builder.CreateFCmpUNE(Abs, ConstantFP::getInfinity(V->getType()),"isinf"); 754 755 V = Builder.CreateAnd(Eq, IsNotInf, "and"); 756 return RValue::get(Builder.CreateZExt(V, ConvertType(E->getType()))); 757 } 758 759 case Builtin::BI__builtin_fpclassify: { 760 Value *V = EmitScalarExpr(E->getArg(5)); 761 llvm::Type *Ty = ConvertType(E->getArg(5)->getType()); 762 763 // Create Result 764 BasicBlock *Begin = Builder.GetInsertBlock(); 765 BasicBlock *End = createBasicBlock("fpclassify_end", this->CurFn); 766 Builder.SetInsertPoint(End); 767 PHINode *Result = 768 Builder.CreatePHI(ConvertType(E->getArg(0)->getType()), 4, 769 "fpclassify_result"); 770 771 // if (V==0) return FP_ZERO 772 Builder.SetInsertPoint(Begin); 773 Value *IsZero = Builder.CreateFCmpOEQ(V, Constant::getNullValue(Ty), 774 "iszero"); 775 Value *ZeroLiteral = EmitScalarExpr(E->getArg(4)); 776 BasicBlock *NotZero = createBasicBlock("fpclassify_not_zero", this->CurFn); 777 Builder.CreateCondBr(IsZero, End, NotZero); 778 Result->addIncoming(ZeroLiteral, Begin); 779 780 // if (V != V) return FP_NAN 781 Builder.SetInsertPoint(NotZero); 782 Value *IsNan = Builder.CreateFCmpUNO(V, V, "cmp"); 783 Value *NanLiteral = EmitScalarExpr(E->getArg(0)); 784 BasicBlock *NotNan = createBasicBlock("fpclassify_not_nan", this->CurFn); 785 Builder.CreateCondBr(IsNan, End, NotNan); 786 Result->addIncoming(NanLiteral, NotZero); 787 788 // if (fabs(V) == infinity) return FP_INFINITY 789 Builder.SetInsertPoint(NotNan); 790 Value *VAbs = EmitFAbs(*this, V); 791 Value *IsInf = 792 Builder.CreateFCmpOEQ(VAbs, ConstantFP::getInfinity(V->getType()), 793 "isinf"); 794 Value *InfLiteral = EmitScalarExpr(E->getArg(1)); 795 BasicBlock *NotInf = createBasicBlock("fpclassify_not_inf", this->CurFn); 796 Builder.CreateCondBr(IsInf, End, NotInf); 797 Result->addIncoming(InfLiteral, NotNan); 798 799 // if (fabs(V) >= MIN_NORMAL) return FP_NORMAL else FP_SUBNORMAL 800 Builder.SetInsertPoint(NotInf); 801 APFloat Smallest = APFloat::getSmallestNormalized( 802 getContext().getFloatTypeSemantics(E->getArg(5)->getType())); 803 Value *IsNormal = 804 Builder.CreateFCmpUGE(VAbs, ConstantFP::get(V->getContext(), Smallest), 805 "isnormal"); 806 Value *NormalResult = 807 Builder.CreateSelect(IsNormal, EmitScalarExpr(E->getArg(2)), 808 EmitScalarExpr(E->getArg(3))); 809 Builder.CreateBr(End); 810 Result->addIncoming(NormalResult, NotInf); 811 812 // return Result 813 Builder.SetInsertPoint(End); 814 return RValue::get(Result); 815 } 816 817 case Builtin::BIalloca: 818 case Builtin::BI_alloca: 819 case Builtin::BI__builtin_alloca: { 820 Value *Size = EmitScalarExpr(E->getArg(0)); 821 return RValue::get(Builder.CreateAlloca(Builder.getInt8Ty(), Size)); 822 } 823 case Builtin::BIbzero: 824 case Builtin::BI__builtin_bzero: { 825 Address Dest = EmitPointerWithAlignment(E->getArg(0)); 826 Value *SizeVal = EmitScalarExpr(E->getArg(1)); 827 EmitNonNullArgCheck(RValue::get(Dest.getPointer()), E->getArg(0)->getType(), 828 E->getArg(0)->getExprLoc(), FD, 0); 829 Builder.CreateMemSet(Dest, Builder.getInt8(0), SizeVal, false); 830 return RValue::get(Dest.getPointer()); 831 } 832 case Builtin::BImemcpy: 833 case Builtin::BI__builtin_memcpy: { 834 Address Dest = EmitPointerWithAlignment(E->getArg(0)); 835 Address Src = EmitPointerWithAlignment(E->getArg(1)); 836 Value *SizeVal = EmitScalarExpr(E->getArg(2)); 837 EmitNonNullArgCheck(RValue::get(Dest.getPointer()), E->getArg(0)->getType(), 838 E->getArg(0)->getExprLoc(), FD, 0); 839 EmitNonNullArgCheck(RValue::get(Src.getPointer()), E->getArg(1)->getType(), 840 E->getArg(1)->getExprLoc(), FD, 1); 841 Builder.CreateMemCpy(Dest, Src, SizeVal, false); 842 return RValue::get(Dest.getPointer()); 843 } 844 845 case Builtin::BI__builtin___memcpy_chk: { 846 // fold __builtin_memcpy_chk(x, y, cst1, cst2) to memcpy iff cst1<=cst2. 847 llvm::APSInt Size, DstSize; 848 if (!E->getArg(2)->EvaluateAsInt(Size, CGM.getContext()) || 849 !E->getArg(3)->EvaluateAsInt(DstSize, CGM.getContext())) 850 break; 851 if (Size.ugt(DstSize)) 852 break; 853 Address Dest = EmitPointerWithAlignment(E->getArg(0)); 854 Address Src = EmitPointerWithAlignment(E->getArg(1)); 855 Value *SizeVal = llvm::ConstantInt::get(Builder.getContext(), Size); 856 Builder.CreateMemCpy(Dest, Src, SizeVal, false); 857 return RValue::get(Dest.getPointer()); 858 } 859 860 case Builtin::BI__builtin_objc_memmove_collectable: { 861 Address DestAddr = EmitPointerWithAlignment(E->getArg(0)); 862 Address SrcAddr = EmitPointerWithAlignment(E->getArg(1)); 863 Value *SizeVal = EmitScalarExpr(E->getArg(2)); 864 CGM.getObjCRuntime().EmitGCMemmoveCollectable(*this, 865 DestAddr, SrcAddr, SizeVal); 866 return RValue::get(DestAddr.getPointer()); 867 } 868 869 case Builtin::BI__builtin___memmove_chk: { 870 // fold __builtin_memmove_chk(x, y, cst1, cst2) to memmove iff cst1<=cst2. 871 llvm::APSInt Size, DstSize; 872 if (!E->getArg(2)->EvaluateAsInt(Size, CGM.getContext()) || 873 !E->getArg(3)->EvaluateAsInt(DstSize, CGM.getContext())) 874 break; 875 if (Size.ugt(DstSize)) 876 break; 877 Address Dest = EmitPointerWithAlignment(E->getArg(0)); 878 Address Src = EmitPointerWithAlignment(E->getArg(1)); 879 Value *SizeVal = llvm::ConstantInt::get(Builder.getContext(), Size); 880 Builder.CreateMemMove(Dest, Src, SizeVal, false); 881 return RValue::get(Dest.getPointer()); 882 } 883 884 case Builtin::BImemmove: 885 case Builtin::BI__builtin_memmove: { 886 Address Dest = EmitPointerWithAlignment(E->getArg(0)); 887 Address Src = EmitPointerWithAlignment(E->getArg(1)); 888 Value *SizeVal = EmitScalarExpr(E->getArg(2)); 889 EmitNonNullArgCheck(RValue::get(Dest.getPointer()), E->getArg(0)->getType(), 890 E->getArg(0)->getExprLoc(), FD, 0); 891 EmitNonNullArgCheck(RValue::get(Src.getPointer()), E->getArg(1)->getType(), 892 E->getArg(1)->getExprLoc(), FD, 1); 893 Builder.CreateMemMove(Dest, Src, SizeVal, false); 894 return RValue::get(Dest.getPointer()); 895 } 896 case Builtin::BImemset: 897 case Builtin::BI__builtin_memset: { 898 Address Dest = EmitPointerWithAlignment(E->getArg(0)); 899 Value *ByteVal = Builder.CreateTrunc(EmitScalarExpr(E->getArg(1)), 900 Builder.getInt8Ty()); 901 Value *SizeVal = EmitScalarExpr(E->getArg(2)); 902 EmitNonNullArgCheck(RValue::get(Dest.getPointer()), E->getArg(0)->getType(), 903 E->getArg(0)->getExprLoc(), FD, 0); 904 Builder.CreateMemSet(Dest, ByteVal, SizeVal, false); 905 return RValue::get(Dest.getPointer()); 906 } 907 case Builtin::BI__builtin___memset_chk: { 908 // fold __builtin_memset_chk(x, y, cst1, cst2) to memset iff cst1<=cst2. 909 llvm::APSInt Size, DstSize; 910 if (!E->getArg(2)->EvaluateAsInt(Size, CGM.getContext()) || 911 !E->getArg(3)->EvaluateAsInt(DstSize, CGM.getContext())) 912 break; 913 if (Size.ugt(DstSize)) 914 break; 915 Address Dest = EmitPointerWithAlignment(E->getArg(0)); 916 Value *ByteVal = Builder.CreateTrunc(EmitScalarExpr(E->getArg(1)), 917 Builder.getInt8Ty()); 918 Value *SizeVal = llvm::ConstantInt::get(Builder.getContext(), Size); 919 Builder.CreateMemSet(Dest, ByteVal, SizeVal, false); 920 return RValue::get(Dest.getPointer()); 921 } 922 case Builtin::BI__builtin_dwarf_cfa: { 923 // The offset in bytes from the first argument to the CFA. 924 // 925 // Why on earth is this in the frontend? Is there any reason at 926 // all that the backend can't reasonably determine this while 927 // lowering llvm.eh.dwarf.cfa()? 928 // 929 // TODO: If there's a satisfactory reason, add a target hook for 930 // this instead of hard-coding 0, which is correct for most targets. 931 int32_t Offset = 0; 932 933 Value *F = CGM.getIntrinsic(Intrinsic::eh_dwarf_cfa); 934 return RValue::get(Builder.CreateCall(F, 935 llvm::ConstantInt::get(Int32Ty, Offset))); 936 } 937 case Builtin::BI__builtin_return_address: { 938 Value *Depth = 939 CGM.EmitConstantExpr(E->getArg(0), getContext().UnsignedIntTy, this); 940 Value *F = CGM.getIntrinsic(Intrinsic::returnaddress); 941 return RValue::get(Builder.CreateCall(F, Depth)); 942 } 943 case Builtin::BI__builtin_frame_address: { 944 Value *Depth = 945 CGM.EmitConstantExpr(E->getArg(0), getContext().UnsignedIntTy, this); 946 Value *F = CGM.getIntrinsic(Intrinsic::frameaddress); 947 return RValue::get(Builder.CreateCall(F, Depth)); 948 } 949 case Builtin::BI__builtin_extract_return_addr: { 950 Value *Address = EmitScalarExpr(E->getArg(0)); 951 Value *Result = getTargetHooks().decodeReturnAddress(*this, Address); 952 return RValue::get(Result); 953 } 954 case Builtin::BI__builtin_frob_return_addr: { 955 Value *Address = EmitScalarExpr(E->getArg(0)); 956 Value *Result = getTargetHooks().encodeReturnAddress(*this, Address); 957 return RValue::get(Result); 958 } 959 case Builtin::BI__builtin_dwarf_sp_column: { 960 llvm::IntegerType *Ty 961 = cast<llvm::IntegerType>(ConvertType(E->getType())); 962 int Column = getTargetHooks().getDwarfEHStackPointer(CGM); 963 if (Column == -1) { 964 CGM.ErrorUnsupported(E, "__builtin_dwarf_sp_column"); 965 return RValue::get(llvm::UndefValue::get(Ty)); 966 } 967 return RValue::get(llvm::ConstantInt::get(Ty, Column, true)); 968 } 969 case Builtin::BI__builtin_init_dwarf_reg_size_table: { 970 Value *Address = EmitScalarExpr(E->getArg(0)); 971 if (getTargetHooks().initDwarfEHRegSizeTable(*this, Address)) 972 CGM.ErrorUnsupported(E, "__builtin_init_dwarf_reg_size_table"); 973 return RValue::get(llvm::UndefValue::get(ConvertType(E->getType()))); 974 } 975 case Builtin::BI__builtin_eh_return: { 976 Value *Int = EmitScalarExpr(E->getArg(0)); 977 Value *Ptr = EmitScalarExpr(E->getArg(1)); 978 979 llvm::IntegerType *IntTy = cast<llvm::IntegerType>(Int->getType()); 980 assert((IntTy->getBitWidth() == 32 || IntTy->getBitWidth() == 64) && 981 "LLVM's __builtin_eh_return only supports 32- and 64-bit variants"); 982 Value *F = CGM.getIntrinsic(IntTy->getBitWidth() == 32 983 ? Intrinsic::eh_return_i32 984 : Intrinsic::eh_return_i64); 985 Builder.CreateCall(F, {Int, Ptr}); 986 Builder.CreateUnreachable(); 987 988 // We do need to preserve an insertion point. 989 EmitBlock(createBasicBlock("builtin_eh_return.cont")); 990 991 return RValue::get(nullptr); 992 } 993 case Builtin::BI__builtin_unwind_init: { 994 Value *F = CGM.getIntrinsic(Intrinsic::eh_unwind_init); 995 return RValue::get(Builder.CreateCall(F)); 996 } 997 case Builtin::BI__builtin_extend_pointer: { 998 // Extends a pointer to the size of an _Unwind_Word, which is 999 // uint64_t on all platforms. Generally this gets poked into a 1000 // register and eventually used as an address, so if the 1001 // addressing registers are wider than pointers and the platform 1002 // doesn't implicitly ignore high-order bits when doing 1003 // addressing, we need to make sure we zext / sext based on 1004 // the platform's expectations. 1005 // 1006 // See: http://gcc.gnu.org/ml/gcc-bugs/2002-02/msg00237.html 1007 1008 // Cast the pointer to intptr_t. 1009 Value *Ptr = EmitScalarExpr(E->getArg(0)); 1010 Value *Result = Builder.CreatePtrToInt(Ptr, IntPtrTy, "extend.cast"); 1011 1012 // If that's 64 bits, we're done. 1013 if (IntPtrTy->getBitWidth() == 64) 1014 return RValue::get(Result); 1015 1016 // Otherwise, ask the codegen data what to do. 1017 if (getTargetHooks().extendPointerWithSExt()) 1018 return RValue::get(Builder.CreateSExt(Result, Int64Ty, "extend.sext")); 1019 else 1020 return RValue::get(Builder.CreateZExt(Result, Int64Ty, "extend.zext")); 1021 } 1022 case Builtin::BI__builtin_setjmp: { 1023 // Buffer is a void**. 1024 Address Buf = EmitPointerWithAlignment(E->getArg(0)); 1025 1026 // Store the frame pointer to the setjmp buffer. 1027 Value *FrameAddr = 1028 Builder.CreateCall(CGM.getIntrinsic(Intrinsic::frameaddress), 1029 ConstantInt::get(Int32Ty, 0)); 1030 Builder.CreateStore(FrameAddr, Buf); 1031 1032 // Store the stack pointer to the setjmp buffer. 1033 Value *StackAddr = 1034 Builder.CreateCall(CGM.getIntrinsic(Intrinsic::stacksave)); 1035 Address StackSaveSlot = 1036 Builder.CreateConstInBoundsGEP(Buf, 2, getPointerSize()); 1037 Builder.CreateStore(StackAddr, StackSaveSlot); 1038 1039 // Call LLVM's EH setjmp, which is lightweight. 1040 Value *F = CGM.getIntrinsic(Intrinsic::eh_sjlj_setjmp); 1041 Buf = Builder.CreateBitCast(Buf, Int8PtrTy); 1042 return RValue::get(Builder.CreateCall(F, Buf.getPointer())); 1043 } 1044 case Builtin::BI__builtin_longjmp: { 1045 Value *Buf = EmitScalarExpr(E->getArg(0)); 1046 Buf = Builder.CreateBitCast(Buf, Int8PtrTy); 1047 1048 // Call LLVM's EH longjmp, which is lightweight. 1049 Builder.CreateCall(CGM.getIntrinsic(Intrinsic::eh_sjlj_longjmp), Buf); 1050 1051 // longjmp doesn't return; mark this as unreachable. 1052 Builder.CreateUnreachable(); 1053 1054 // We do need to preserve an insertion point. 1055 EmitBlock(createBasicBlock("longjmp.cont")); 1056 1057 return RValue::get(nullptr); 1058 } 1059 case Builtin::BI__sync_fetch_and_add: 1060 case Builtin::BI__sync_fetch_and_sub: 1061 case Builtin::BI__sync_fetch_and_or: 1062 case Builtin::BI__sync_fetch_and_and: 1063 case Builtin::BI__sync_fetch_and_xor: 1064 case Builtin::BI__sync_fetch_and_nand: 1065 case Builtin::BI__sync_add_and_fetch: 1066 case Builtin::BI__sync_sub_and_fetch: 1067 case Builtin::BI__sync_and_and_fetch: 1068 case Builtin::BI__sync_or_and_fetch: 1069 case Builtin::BI__sync_xor_and_fetch: 1070 case Builtin::BI__sync_nand_and_fetch: 1071 case Builtin::BI__sync_val_compare_and_swap: 1072 case Builtin::BI__sync_bool_compare_and_swap: 1073 case Builtin::BI__sync_lock_test_and_set: 1074 case Builtin::BI__sync_lock_release: 1075 case Builtin::BI__sync_swap: 1076 llvm_unreachable("Shouldn't make it through sema"); 1077 case Builtin::BI__sync_fetch_and_add_1: 1078 case Builtin::BI__sync_fetch_and_add_2: 1079 case Builtin::BI__sync_fetch_and_add_4: 1080 case Builtin::BI__sync_fetch_and_add_8: 1081 case Builtin::BI__sync_fetch_and_add_16: 1082 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Add, E); 1083 case Builtin::BI__sync_fetch_and_sub_1: 1084 case Builtin::BI__sync_fetch_and_sub_2: 1085 case Builtin::BI__sync_fetch_and_sub_4: 1086 case Builtin::BI__sync_fetch_and_sub_8: 1087 case Builtin::BI__sync_fetch_and_sub_16: 1088 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Sub, E); 1089 case Builtin::BI__sync_fetch_and_or_1: 1090 case Builtin::BI__sync_fetch_and_or_2: 1091 case Builtin::BI__sync_fetch_and_or_4: 1092 case Builtin::BI__sync_fetch_and_or_8: 1093 case Builtin::BI__sync_fetch_and_or_16: 1094 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Or, E); 1095 case Builtin::BI__sync_fetch_and_and_1: 1096 case Builtin::BI__sync_fetch_and_and_2: 1097 case Builtin::BI__sync_fetch_and_and_4: 1098 case Builtin::BI__sync_fetch_and_and_8: 1099 case Builtin::BI__sync_fetch_and_and_16: 1100 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::And, E); 1101 case Builtin::BI__sync_fetch_and_xor_1: 1102 case Builtin::BI__sync_fetch_and_xor_2: 1103 case Builtin::BI__sync_fetch_and_xor_4: 1104 case Builtin::BI__sync_fetch_and_xor_8: 1105 case Builtin::BI__sync_fetch_and_xor_16: 1106 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Xor, E); 1107 case Builtin::BI__sync_fetch_and_nand_1: 1108 case Builtin::BI__sync_fetch_and_nand_2: 1109 case Builtin::BI__sync_fetch_and_nand_4: 1110 case Builtin::BI__sync_fetch_and_nand_8: 1111 case Builtin::BI__sync_fetch_and_nand_16: 1112 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Nand, E); 1113 1114 // Clang extensions: not overloaded yet. 1115 case Builtin::BI__sync_fetch_and_min: 1116 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Min, E); 1117 case Builtin::BI__sync_fetch_and_max: 1118 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Max, E); 1119 case Builtin::BI__sync_fetch_and_umin: 1120 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::UMin, E); 1121 case Builtin::BI__sync_fetch_and_umax: 1122 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::UMax, E); 1123 1124 case Builtin::BI__sync_add_and_fetch_1: 1125 case Builtin::BI__sync_add_and_fetch_2: 1126 case Builtin::BI__sync_add_and_fetch_4: 1127 case Builtin::BI__sync_add_and_fetch_8: 1128 case Builtin::BI__sync_add_and_fetch_16: 1129 return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::Add, E, 1130 llvm::Instruction::Add); 1131 case Builtin::BI__sync_sub_and_fetch_1: 1132 case Builtin::BI__sync_sub_and_fetch_2: 1133 case Builtin::BI__sync_sub_and_fetch_4: 1134 case Builtin::BI__sync_sub_and_fetch_8: 1135 case Builtin::BI__sync_sub_and_fetch_16: 1136 return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::Sub, E, 1137 llvm::Instruction::Sub); 1138 case Builtin::BI__sync_and_and_fetch_1: 1139 case Builtin::BI__sync_and_and_fetch_2: 1140 case Builtin::BI__sync_and_and_fetch_4: 1141 case Builtin::BI__sync_and_and_fetch_8: 1142 case Builtin::BI__sync_and_and_fetch_16: 1143 return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::And, E, 1144 llvm::Instruction::And); 1145 case Builtin::BI__sync_or_and_fetch_1: 1146 case Builtin::BI__sync_or_and_fetch_2: 1147 case Builtin::BI__sync_or_and_fetch_4: 1148 case Builtin::BI__sync_or_and_fetch_8: 1149 case Builtin::BI__sync_or_and_fetch_16: 1150 return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::Or, E, 1151 llvm::Instruction::Or); 1152 case Builtin::BI__sync_xor_and_fetch_1: 1153 case Builtin::BI__sync_xor_and_fetch_2: 1154 case Builtin::BI__sync_xor_and_fetch_4: 1155 case Builtin::BI__sync_xor_and_fetch_8: 1156 case Builtin::BI__sync_xor_and_fetch_16: 1157 return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::Xor, E, 1158 llvm::Instruction::Xor); 1159 case Builtin::BI__sync_nand_and_fetch_1: 1160 case Builtin::BI__sync_nand_and_fetch_2: 1161 case Builtin::BI__sync_nand_and_fetch_4: 1162 case Builtin::BI__sync_nand_and_fetch_8: 1163 case Builtin::BI__sync_nand_and_fetch_16: 1164 return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::Nand, E, 1165 llvm::Instruction::And, true); 1166 1167 case Builtin::BI__sync_val_compare_and_swap_1: 1168 case Builtin::BI__sync_val_compare_and_swap_2: 1169 case Builtin::BI__sync_val_compare_and_swap_4: 1170 case Builtin::BI__sync_val_compare_and_swap_8: 1171 case Builtin::BI__sync_val_compare_and_swap_16: 1172 return RValue::get(MakeAtomicCmpXchgValue(*this, E, false)); 1173 1174 case Builtin::BI__sync_bool_compare_and_swap_1: 1175 case Builtin::BI__sync_bool_compare_and_swap_2: 1176 case Builtin::BI__sync_bool_compare_and_swap_4: 1177 case Builtin::BI__sync_bool_compare_and_swap_8: 1178 case Builtin::BI__sync_bool_compare_and_swap_16: 1179 return RValue::get(MakeAtomicCmpXchgValue(*this, E, true)); 1180 1181 case Builtin::BI__sync_swap_1: 1182 case Builtin::BI__sync_swap_2: 1183 case Builtin::BI__sync_swap_4: 1184 case Builtin::BI__sync_swap_8: 1185 case Builtin::BI__sync_swap_16: 1186 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Xchg, E); 1187 1188 case Builtin::BI__sync_lock_test_and_set_1: 1189 case Builtin::BI__sync_lock_test_and_set_2: 1190 case Builtin::BI__sync_lock_test_and_set_4: 1191 case Builtin::BI__sync_lock_test_and_set_8: 1192 case Builtin::BI__sync_lock_test_and_set_16: 1193 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Xchg, E); 1194 1195 case Builtin::BI__sync_lock_release_1: 1196 case Builtin::BI__sync_lock_release_2: 1197 case Builtin::BI__sync_lock_release_4: 1198 case Builtin::BI__sync_lock_release_8: 1199 case Builtin::BI__sync_lock_release_16: { 1200 Value *Ptr = EmitScalarExpr(E->getArg(0)); 1201 QualType ElTy = E->getArg(0)->getType()->getPointeeType(); 1202 CharUnits StoreSize = getContext().getTypeSizeInChars(ElTy); 1203 llvm::Type *ITy = llvm::IntegerType::get(getLLVMContext(), 1204 StoreSize.getQuantity() * 8); 1205 Ptr = Builder.CreateBitCast(Ptr, ITy->getPointerTo()); 1206 llvm::StoreInst *Store = 1207 Builder.CreateAlignedStore(llvm::Constant::getNullValue(ITy), Ptr, 1208 StoreSize); 1209 Store->setAtomic(llvm::Release); 1210 return RValue::get(nullptr); 1211 } 1212 1213 case Builtin::BI__sync_synchronize: { 1214 // We assume this is supposed to correspond to a C++0x-style 1215 // sequentially-consistent fence (i.e. this is only usable for 1216 // synchonization, not device I/O or anything like that). This intrinsic 1217 // is really badly designed in the sense that in theory, there isn't 1218 // any way to safely use it... but in practice, it mostly works 1219 // to use it with non-atomic loads and stores to get acquire/release 1220 // semantics. 1221 Builder.CreateFence(llvm::SequentiallyConsistent); 1222 return RValue::get(nullptr); 1223 } 1224 1225 case Builtin::BI__builtin_nontemporal_load: 1226 return RValue::get(EmitNontemporalLoad(*this, E)); 1227 case Builtin::BI__builtin_nontemporal_store: 1228 return RValue::get(EmitNontemporalStore(*this, E)); 1229 case Builtin::BI__c11_atomic_is_lock_free: 1230 case Builtin::BI__atomic_is_lock_free: { 1231 // Call "bool __atomic_is_lock_free(size_t size, void *ptr)". For the 1232 // __c11 builtin, ptr is 0 (indicating a properly-aligned object), since 1233 // _Atomic(T) is always properly-aligned. 1234 const char *LibCallName = "__atomic_is_lock_free"; 1235 CallArgList Args; 1236 Args.add(RValue::get(EmitScalarExpr(E->getArg(0))), 1237 getContext().getSizeType()); 1238 if (BuiltinID == Builtin::BI__atomic_is_lock_free) 1239 Args.add(RValue::get(EmitScalarExpr(E->getArg(1))), 1240 getContext().VoidPtrTy); 1241 else 1242 Args.add(RValue::get(llvm::Constant::getNullValue(VoidPtrTy)), 1243 getContext().VoidPtrTy); 1244 const CGFunctionInfo &FuncInfo = 1245 CGM.getTypes().arrangeFreeFunctionCall(E->getType(), Args, 1246 FunctionType::ExtInfo(), 1247 RequiredArgs::All); 1248 llvm::FunctionType *FTy = CGM.getTypes().GetFunctionType(FuncInfo); 1249 llvm::Constant *Func = CGM.CreateRuntimeFunction(FTy, LibCallName); 1250 return EmitCall(FuncInfo, Func, ReturnValueSlot(), Args); 1251 } 1252 1253 case Builtin::BI__atomic_test_and_set: { 1254 // Look at the argument type to determine whether this is a volatile 1255 // operation. The parameter type is always volatile. 1256 QualType PtrTy = E->getArg(0)->IgnoreImpCasts()->getType(); 1257 bool Volatile = 1258 PtrTy->castAs<PointerType>()->getPointeeType().isVolatileQualified(); 1259 1260 Value *Ptr = EmitScalarExpr(E->getArg(0)); 1261 unsigned AddrSpace = Ptr->getType()->getPointerAddressSpace(); 1262 Ptr = Builder.CreateBitCast(Ptr, Int8Ty->getPointerTo(AddrSpace)); 1263 Value *NewVal = Builder.getInt8(1); 1264 Value *Order = EmitScalarExpr(E->getArg(1)); 1265 if (isa<llvm::ConstantInt>(Order)) { 1266 int ord = cast<llvm::ConstantInt>(Order)->getZExtValue(); 1267 AtomicRMWInst *Result = nullptr; 1268 switch (ord) { 1269 case 0: // memory_order_relaxed 1270 default: // invalid order 1271 Result = Builder.CreateAtomicRMW(llvm::AtomicRMWInst::Xchg, 1272 Ptr, NewVal, 1273 llvm::Monotonic); 1274 break; 1275 case 1: // memory_order_consume 1276 case 2: // memory_order_acquire 1277 Result = Builder.CreateAtomicRMW(llvm::AtomicRMWInst::Xchg, 1278 Ptr, NewVal, 1279 llvm::Acquire); 1280 break; 1281 case 3: // memory_order_release 1282 Result = Builder.CreateAtomicRMW(llvm::AtomicRMWInst::Xchg, 1283 Ptr, NewVal, 1284 llvm::Release); 1285 break; 1286 case 4: // memory_order_acq_rel 1287 Result = Builder.CreateAtomicRMW(llvm::AtomicRMWInst::Xchg, 1288 Ptr, NewVal, 1289 llvm::AcquireRelease); 1290 break; 1291 case 5: // memory_order_seq_cst 1292 Result = Builder.CreateAtomicRMW(llvm::AtomicRMWInst::Xchg, 1293 Ptr, NewVal, 1294 llvm::SequentiallyConsistent); 1295 break; 1296 } 1297 Result->setVolatile(Volatile); 1298 return RValue::get(Builder.CreateIsNotNull(Result, "tobool")); 1299 } 1300 1301 llvm::BasicBlock *ContBB = createBasicBlock("atomic.continue", CurFn); 1302 1303 llvm::BasicBlock *BBs[5] = { 1304 createBasicBlock("monotonic", CurFn), 1305 createBasicBlock("acquire", CurFn), 1306 createBasicBlock("release", CurFn), 1307 createBasicBlock("acqrel", CurFn), 1308 createBasicBlock("seqcst", CurFn) 1309 }; 1310 llvm::AtomicOrdering Orders[5] = { 1311 llvm::Monotonic, llvm::Acquire, llvm::Release, 1312 llvm::AcquireRelease, llvm::SequentiallyConsistent 1313 }; 1314 1315 Order = Builder.CreateIntCast(Order, Builder.getInt32Ty(), false); 1316 llvm::SwitchInst *SI = Builder.CreateSwitch(Order, BBs[0]); 1317 1318 Builder.SetInsertPoint(ContBB); 1319 PHINode *Result = Builder.CreatePHI(Int8Ty, 5, "was_set"); 1320 1321 for (unsigned i = 0; i < 5; ++i) { 1322 Builder.SetInsertPoint(BBs[i]); 1323 AtomicRMWInst *RMW = Builder.CreateAtomicRMW(llvm::AtomicRMWInst::Xchg, 1324 Ptr, NewVal, Orders[i]); 1325 RMW->setVolatile(Volatile); 1326 Result->addIncoming(RMW, BBs[i]); 1327 Builder.CreateBr(ContBB); 1328 } 1329 1330 SI->addCase(Builder.getInt32(0), BBs[0]); 1331 SI->addCase(Builder.getInt32(1), BBs[1]); 1332 SI->addCase(Builder.getInt32(2), BBs[1]); 1333 SI->addCase(Builder.getInt32(3), BBs[2]); 1334 SI->addCase(Builder.getInt32(4), BBs[3]); 1335 SI->addCase(Builder.getInt32(5), BBs[4]); 1336 1337 Builder.SetInsertPoint(ContBB); 1338 return RValue::get(Builder.CreateIsNotNull(Result, "tobool")); 1339 } 1340 1341 case Builtin::BI__atomic_clear: { 1342 QualType PtrTy = E->getArg(0)->IgnoreImpCasts()->getType(); 1343 bool Volatile = 1344 PtrTy->castAs<PointerType>()->getPointeeType().isVolatileQualified(); 1345 1346 Address Ptr = EmitPointerWithAlignment(E->getArg(0)); 1347 unsigned AddrSpace = Ptr.getPointer()->getType()->getPointerAddressSpace(); 1348 Ptr = Builder.CreateBitCast(Ptr, Int8Ty->getPointerTo(AddrSpace)); 1349 Value *NewVal = Builder.getInt8(0); 1350 Value *Order = EmitScalarExpr(E->getArg(1)); 1351 if (isa<llvm::ConstantInt>(Order)) { 1352 int ord = cast<llvm::ConstantInt>(Order)->getZExtValue(); 1353 StoreInst *Store = Builder.CreateStore(NewVal, Ptr, Volatile); 1354 switch (ord) { 1355 case 0: // memory_order_relaxed 1356 default: // invalid order 1357 Store->setOrdering(llvm::Monotonic); 1358 break; 1359 case 3: // memory_order_release 1360 Store->setOrdering(llvm::Release); 1361 break; 1362 case 5: // memory_order_seq_cst 1363 Store->setOrdering(llvm::SequentiallyConsistent); 1364 break; 1365 } 1366 return RValue::get(nullptr); 1367 } 1368 1369 llvm::BasicBlock *ContBB = createBasicBlock("atomic.continue", CurFn); 1370 1371 llvm::BasicBlock *BBs[3] = { 1372 createBasicBlock("monotonic", CurFn), 1373 createBasicBlock("release", CurFn), 1374 createBasicBlock("seqcst", CurFn) 1375 }; 1376 llvm::AtomicOrdering Orders[3] = { 1377 llvm::Monotonic, llvm::Release, llvm::SequentiallyConsistent 1378 }; 1379 1380 Order = Builder.CreateIntCast(Order, Builder.getInt32Ty(), false); 1381 llvm::SwitchInst *SI = Builder.CreateSwitch(Order, BBs[0]); 1382 1383 for (unsigned i = 0; i < 3; ++i) { 1384 Builder.SetInsertPoint(BBs[i]); 1385 StoreInst *Store = Builder.CreateStore(NewVal, Ptr, Volatile); 1386 Store->setOrdering(Orders[i]); 1387 Builder.CreateBr(ContBB); 1388 } 1389 1390 SI->addCase(Builder.getInt32(0), BBs[0]); 1391 SI->addCase(Builder.getInt32(3), BBs[1]); 1392 SI->addCase(Builder.getInt32(5), BBs[2]); 1393 1394 Builder.SetInsertPoint(ContBB); 1395 return RValue::get(nullptr); 1396 } 1397 1398 case Builtin::BI__atomic_thread_fence: 1399 case Builtin::BI__atomic_signal_fence: 1400 case Builtin::BI__c11_atomic_thread_fence: 1401 case Builtin::BI__c11_atomic_signal_fence: { 1402 llvm::SynchronizationScope Scope; 1403 if (BuiltinID == Builtin::BI__atomic_signal_fence || 1404 BuiltinID == Builtin::BI__c11_atomic_signal_fence) 1405 Scope = llvm::SingleThread; 1406 else 1407 Scope = llvm::CrossThread; 1408 Value *Order = EmitScalarExpr(E->getArg(0)); 1409 if (isa<llvm::ConstantInt>(Order)) { 1410 int ord = cast<llvm::ConstantInt>(Order)->getZExtValue(); 1411 switch (ord) { 1412 case 0: // memory_order_relaxed 1413 default: // invalid order 1414 break; 1415 case 1: // memory_order_consume 1416 case 2: // memory_order_acquire 1417 Builder.CreateFence(llvm::Acquire, Scope); 1418 break; 1419 case 3: // memory_order_release 1420 Builder.CreateFence(llvm::Release, Scope); 1421 break; 1422 case 4: // memory_order_acq_rel 1423 Builder.CreateFence(llvm::AcquireRelease, Scope); 1424 break; 1425 case 5: // memory_order_seq_cst 1426 Builder.CreateFence(llvm::SequentiallyConsistent, Scope); 1427 break; 1428 } 1429 return RValue::get(nullptr); 1430 } 1431 1432 llvm::BasicBlock *AcquireBB, *ReleaseBB, *AcqRelBB, *SeqCstBB; 1433 AcquireBB = createBasicBlock("acquire", CurFn); 1434 ReleaseBB = createBasicBlock("release", CurFn); 1435 AcqRelBB = createBasicBlock("acqrel", CurFn); 1436 SeqCstBB = createBasicBlock("seqcst", CurFn); 1437 llvm::BasicBlock *ContBB = createBasicBlock("atomic.continue", CurFn); 1438 1439 Order = Builder.CreateIntCast(Order, Builder.getInt32Ty(), false); 1440 llvm::SwitchInst *SI = Builder.CreateSwitch(Order, ContBB); 1441 1442 Builder.SetInsertPoint(AcquireBB); 1443 Builder.CreateFence(llvm::Acquire, Scope); 1444 Builder.CreateBr(ContBB); 1445 SI->addCase(Builder.getInt32(1), AcquireBB); 1446 SI->addCase(Builder.getInt32(2), AcquireBB); 1447 1448 Builder.SetInsertPoint(ReleaseBB); 1449 Builder.CreateFence(llvm::Release, Scope); 1450 Builder.CreateBr(ContBB); 1451 SI->addCase(Builder.getInt32(3), ReleaseBB); 1452 1453 Builder.SetInsertPoint(AcqRelBB); 1454 Builder.CreateFence(llvm::AcquireRelease, Scope); 1455 Builder.CreateBr(ContBB); 1456 SI->addCase(Builder.getInt32(4), AcqRelBB); 1457 1458 Builder.SetInsertPoint(SeqCstBB); 1459 Builder.CreateFence(llvm::SequentiallyConsistent, Scope); 1460 Builder.CreateBr(ContBB); 1461 SI->addCase(Builder.getInt32(5), SeqCstBB); 1462 1463 Builder.SetInsertPoint(ContBB); 1464 return RValue::get(nullptr); 1465 } 1466 1467 // Library functions with special handling. 1468 case Builtin::BIsqrt: 1469 case Builtin::BIsqrtf: 1470 case Builtin::BIsqrtl: { 1471 // Transform a call to sqrt* into a @llvm.sqrt.* intrinsic call, but only 1472 // in finite- or unsafe-math mode (the intrinsic has different semantics 1473 // for handling negative numbers compared to the library function, so 1474 // -fmath-errno=0 is not enough). 1475 if (!FD->hasAttr<ConstAttr>()) 1476 break; 1477 if (!(CGM.getCodeGenOpts().UnsafeFPMath || 1478 CGM.getCodeGenOpts().NoNaNsFPMath)) 1479 break; 1480 Value *Arg0 = EmitScalarExpr(E->getArg(0)); 1481 llvm::Type *ArgType = Arg0->getType(); 1482 Value *F = CGM.getIntrinsic(Intrinsic::sqrt, ArgType); 1483 return RValue::get(Builder.CreateCall(F, Arg0)); 1484 } 1485 1486 case Builtin::BI__builtin_pow: 1487 case Builtin::BI__builtin_powf: 1488 case Builtin::BI__builtin_powl: 1489 case Builtin::BIpow: 1490 case Builtin::BIpowf: 1491 case Builtin::BIpowl: { 1492 // Transform a call to pow* into a @llvm.pow.* intrinsic call. 1493 if (!FD->hasAttr<ConstAttr>()) 1494 break; 1495 Value *Base = EmitScalarExpr(E->getArg(0)); 1496 Value *Exponent = EmitScalarExpr(E->getArg(1)); 1497 llvm::Type *ArgType = Base->getType(); 1498 Value *F = CGM.getIntrinsic(Intrinsic::pow, ArgType); 1499 return RValue::get(Builder.CreateCall(F, {Base, Exponent})); 1500 } 1501 1502 case Builtin::BIfma: 1503 case Builtin::BIfmaf: 1504 case Builtin::BIfmal: 1505 case Builtin::BI__builtin_fma: 1506 case Builtin::BI__builtin_fmaf: 1507 case Builtin::BI__builtin_fmal: { 1508 // Rewrite fma to intrinsic. 1509 Value *FirstArg = EmitScalarExpr(E->getArg(0)); 1510 llvm::Type *ArgType = FirstArg->getType(); 1511 Value *F = CGM.getIntrinsic(Intrinsic::fma, ArgType); 1512 return RValue::get( 1513 Builder.CreateCall(F, {FirstArg, EmitScalarExpr(E->getArg(1)), 1514 EmitScalarExpr(E->getArg(2))})); 1515 } 1516 1517 case Builtin::BI__builtin_signbit: 1518 case Builtin::BI__builtin_signbitf: 1519 case Builtin::BI__builtin_signbitl: { 1520 return RValue::get( 1521 Builder.CreateZExt(EmitSignBit(*this, EmitScalarExpr(E->getArg(0))), 1522 ConvertType(E->getType()))); 1523 } 1524 case Builtin::BI__builtin_annotation: { 1525 llvm::Value *AnnVal = EmitScalarExpr(E->getArg(0)); 1526 llvm::Value *F = CGM.getIntrinsic(llvm::Intrinsic::annotation, 1527 AnnVal->getType()); 1528 1529 // Get the annotation string, go through casts. Sema requires this to be a 1530 // non-wide string literal, potentially casted, so the cast<> is safe. 1531 const Expr *AnnotationStrExpr = E->getArg(1)->IgnoreParenCasts(); 1532 StringRef Str = cast<StringLiteral>(AnnotationStrExpr)->getString(); 1533 return RValue::get(EmitAnnotationCall(F, AnnVal, Str, E->getExprLoc())); 1534 } 1535 case Builtin::BI__builtin_addcb: 1536 case Builtin::BI__builtin_addcs: 1537 case Builtin::BI__builtin_addc: 1538 case Builtin::BI__builtin_addcl: 1539 case Builtin::BI__builtin_addcll: 1540 case Builtin::BI__builtin_subcb: 1541 case Builtin::BI__builtin_subcs: 1542 case Builtin::BI__builtin_subc: 1543 case Builtin::BI__builtin_subcl: 1544 case Builtin::BI__builtin_subcll: { 1545 1546 // We translate all of these builtins from expressions of the form: 1547 // int x = ..., y = ..., carryin = ..., carryout, result; 1548 // result = __builtin_addc(x, y, carryin, &carryout); 1549 // 1550 // to LLVM IR of the form: 1551 // 1552 // %tmp1 = call {i32, i1} @llvm.uadd.with.overflow.i32(i32 %x, i32 %y) 1553 // %tmpsum1 = extractvalue {i32, i1} %tmp1, 0 1554 // %carry1 = extractvalue {i32, i1} %tmp1, 1 1555 // %tmp2 = call {i32, i1} @llvm.uadd.with.overflow.i32(i32 %tmpsum1, 1556 // i32 %carryin) 1557 // %result = extractvalue {i32, i1} %tmp2, 0 1558 // %carry2 = extractvalue {i32, i1} %tmp2, 1 1559 // %tmp3 = or i1 %carry1, %carry2 1560 // %tmp4 = zext i1 %tmp3 to i32 1561 // store i32 %tmp4, i32* %carryout 1562 1563 // Scalarize our inputs. 1564 llvm::Value *X = EmitScalarExpr(E->getArg(0)); 1565 llvm::Value *Y = EmitScalarExpr(E->getArg(1)); 1566 llvm::Value *Carryin = EmitScalarExpr(E->getArg(2)); 1567 Address CarryOutPtr = EmitPointerWithAlignment(E->getArg(3)); 1568 1569 // Decide if we are lowering to a uadd.with.overflow or usub.with.overflow. 1570 llvm::Intrinsic::ID IntrinsicId; 1571 switch (BuiltinID) { 1572 default: llvm_unreachable("Unknown multiprecision builtin id."); 1573 case Builtin::BI__builtin_addcb: 1574 case Builtin::BI__builtin_addcs: 1575 case Builtin::BI__builtin_addc: 1576 case Builtin::BI__builtin_addcl: 1577 case Builtin::BI__builtin_addcll: 1578 IntrinsicId = llvm::Intrinsic::uadd_with_overflow; 1579 break; 1580 case Builtin::BI__builtin_subcb: 1581 case Builtin::BI__builtin_subcs: 1582 case Builtin::BI__builtin_subc: 1583 case Builtin::BI__builtin_subcl: 1584 case Builtin::BI__builtin_subcll: 1585 IntrinsicId = llvm::Intrinsic::usub_with_overflow; 1586 break; 1587 } 1588 1589 // Construct our resulting LLVM IR expression. 1590 llvm::Value *Carry1; 1591 llvm::Value *Sum1 = EmitOverflowIntrinsic(*this, IntrinsicId, 1592 X, Y, Carry1); 1593 llvm::Value *Carry2; 1594 llvm::Value *Sum2 = EmitOverflowIntrinsic(*this, IntrinsicId, 1595 Sum1, Carryin, Carry2); 1596 llvm::Value *CarryOut = Builder.CreateZExt(Builder.CreateOr(Carry1, Carry2), 1597 X->getType()); 1598 Builder.CreateStore(CarryOut, CarryOutPtr); 1599 return RValue::get(Sum2); 1600 } 1601 1602 case Builtin::BI__builtin_add_overflow: 1603 case Builtin::BI__builtin_sub_overflow: 1604 case Builtin::BI__builtin_mul_overflow: { 1605 const clang::Expr *LeftArg = E->getArg(0); 1606 const clang::Expr *RightArg = E->getArg(1); 1607 const clang::Expr *ResultArg = E->getArg(2); 1608 1609 clang::QualType ResultQTy = 1610 ResultArg->getType()->castAs<PointerType>()->getPointeeType(); 1611 1612 WidthAndSignedness LeftInfo = 1613 getIntegerWidthAndSignedness(CGM.getContext(), LeftArg->getType()); 1614 WidthAndSignedness RightInfo = 1615 getIntegerWidthAndSignedness(CGM.getContext(), RightArg->getType()); 1616 WidthAndSignedness ResultInfo = 1617 getIntegerWidthAndSignedness(CGM.getContext(), ResultQTy); 1618 WidthAndSignedness EncompassingInfo = 1619 EncompassingIntegerType({LeftInfo, RightInfo, ResultInfo}); 1620 1621 llvm::Type *EncompassingLLVMTy = 1622 llvm::IntegerType::get(CGM.getLLVMContext(), EncompassingInfo.Width); 1623 1624 llvm::Type *ResultLLVMTy = CGM.getTypes().ConvertType(ResultQTy); 1625 1626 llvm::Intrinsic::ID IntrinsicId; 1627 switch (BuiltinID) { 1628 default: 1629 llvm_unreachable("Unknown overflow builtin id."); 1630 case Builtin::BI__builtin_add_overflow: 1631 IntrinsicId = EncompassingInfo.Signed 1632 ? llvm::Intrinsic::sadd_with_overflow 1633 : llvm::Intrinsic::uadd_with_overflow; 1634 break; 1635 case Builtin::BI__builtin_sub_overflow: 1636 IntrinsicId = EncompassingInfo.Signed 1637 ? llvm::Intrinsic::ssub_with_overflow 1638 : llvm::Intrinsic::usub_with_overflow; 1639 break; 1640 case Builtin::BI__builtin_mul_overflow: 1641 IntrinsicId = EncompassingInfo.Signed 1642 ? llvm::Intrinsic::smul_with_overflow 1643 : llvm::Intrinsic::umul_with_overflow; 1644 break; 1645 } 1646 1647 llvm::Value *Left = EmitScalarExpr(LeftArg); 1648 llvm::Value *Right = EmitScalarExpr(RightArg); 1649 Address ResultPtr = EmitPointerWithAlignment(ResultArg); 1650 1651 // Extend each operand to the encompassing type. 1652 Left = Builder.CreateIntCast(Left, EncompassingLLVMTy, LeftInfo.Signed); 1653 Right = Builder.CreateIntCast(Right, EncompassingLLVMTy, RightInfo.Signed); 1654 1655 // Perform the operation on the extended values. 1656 llvm::Value *Overflow, *Result; 1657 Result = EmitOverflowIntrinsic(*this, IntrinsicId, Left, Right, Overflow); 1658 1659 if (EncompassingInfo.Width > ResultInfo.Width) { 1660 // The encompassing type is wider than the result type, so we need to 1661 // truncate it. 1662 llvm::Value *ResultTrunc = Builder.CreateTrunc(Result, ResultLLVMTy); 1663 1664 // To see if the truncation caused an overflow, we will extend 1665 // the result and then compare it to the original result. 1666 llvm::Value *ResultTruncExt = Builder.CreateIntCast( 1667 ResultTrunc, EncompassingLLVMTy, ResultInfo.Signed); 1668 llvm::Value *TruncationOverflow = 1669 Builder.CreateICmpNE(Result, ResultTruncExt); 1670 1671 Overflow = Builder.CreateOr(Overflow, TruncationOverflow); 1672 Result = ResultTrunc; 1673 } 1674 1675 // Finally, store the result using the pointer. 1676 bool isVolatile = 1677 ResultArg->getType()->getPointeeType().isVolatileQualified(); 1678 Builder.CreateStore(EmitToMemory(Result, ResultQTy), ResultPtr, isVolatile); 1679 1680 return RValue::get(Overflow); 1681 } 1682 1683 case Builtin::BI__builtin_uadd_overflow: 1684 case Builtin::BI__builtin_uaddl_overflow: 1685 case Builtin::BI__builtin_uaddll_overflow: 1686 case Builtin::BI__builtin_usub_overflow: 1687 case Builtin::BI__builtin_usubl_overflow: 1688 case Builtin::BI__builtin_usubll_overflow: 1689 case Builtin::BI__builtin_umul_overflow: 1690 case Builtin::BI__builtin_umull_overflow: 1691 case Builtin::BI__builtin_umulll_overflow: 1692 case Builtin::BI__builtin_sadd_overflow: 1693 case Builtin::BI__builtin_saddl_overflow: 1694 case Builtin::BI__builtin_saddll_overflow: 1695 case Builtin::BI__builtin_ssub_overflow: 1696 case Builtin::BI__builtin_ssubl_overflow: 1697 case Builtin::BI__builtin_ssubll_overflow: 1698 case Builtin::BI__builtin_smul_overflow: 1699 case Builtin::BI__builtin_smull_overflow: 1700 case Builtin::BI__builtin_smulll_overflow: { 1701 1702 // We translate all of these builtins directly to the relevant llvm IR node. 1703 1704 // Scalarize our inputs. 1705 llvm::Value *X = EmitScalarExpr(E->getArg(0)); 1706 llvm::Value *Y = EmitScalarExpr(E->getArg(1)); 1707 Address SumOutPtr = EmitPointerWithAlignment(E->getArg(2)); 1708 1709 // Decide which of the overflow intrinsics we are lowering to: 1710 llvm::Intrinsic::ID IntrinsicId; 1711 switch (BuiltinID) { 1712 default: llvm_unreachable("Unknown overflow builtin id."); 1713 case Builtin::BI__builtin_uadd_overflow: 1714 case Builtin::BI__builtin_uaddl_overflow: 1715 case Builtin::BI__builtin_uaddll_overflow: 1716 IntrinsicId = llvm::Intrinsic::uadd_with_overflow; 1717 break; 1718 case Builtin::BI__builtin_usub_overflow: 1719 case Builtin::BI__builtin_usubl_overflow: 1720 case Builtin::BI__builtin_usubll_overflow: 1721 IntrinsicId = llvm::Intrinsic::usub_with_overflow; 1722 break; 1723 case Builtin::BI__builtin_umul_overflow: 1724 case Builtin::BI__builtin_umull_overflow: 1725 case Builtin::BI__builtin_umulll_overflow: 1726 IntrinsicId = llvm::Intrinsic::umul_with_overflow; 1727 break; 1728 case Builtin::BI__builtin_sadd_overflow: 1729 case Builtin::BI__builtin_saddl_overflow: 1730 case Builtin::BI__builtin_saddll_overflow: 1731 IntrinsicId = llvm::Intrinsic::sadd_with_overflow; 1732 break; 1733 case Builtin::BI__builtin_ssub_overflow: 1734 case Builtin::BI__builtin_ssubl_overflow: 1735 case Builtin::BI__builtin_ssubll_overflow: 1736 IntrinsicId = llvm::Intrinsic::ssub_with_overflow; 1737 break; 1738 case Builtin::BI__builtin_smul_overflow: 1739 case Builtin::BI__builtin_smull_overflow: 1740 case Builtin::BI__builtin_smulll_overflow: 1741 IntrinsicId = llvm::Intrinsic::smul_with_overflow; 1742 break; 1743 } 1744 1745 1746 llvm::Value *Carry; 1747 llvm::Value *Sum = EmitOverflowIntrinsic(*this, IntrinsicId, X, Y, Carry); 1748 Builder.CreateStore(Sum, SumOutPtr); 1749 1750 return RValue::get(Carry); 1751 } 1752 case Builtin::BI__builtin_addressof: 1753 return RValue::get(EmitLValue(E->getArg(0)).getPointer()); 1754 case Builtin::BI__builtin_operator_new: 1755 return EmitBuiltinNewDeleteCall(FD->getType()->castAs<FunctionProtoType>(), 1756 E->getArg(0), false); 1757 case Builtin::BI__builtin_operator_delete: 1758 return EmitBuiltinNewDeleteCall(FD->getType()->castAs<FunctionProtoType>(), 1759 E->getArg(0), true); 1760 case Builtin::BI__noop: 1761 // __noop always evaluates to an integer literal zero. 1762 return RValue::get(ConstantInt::get(IntTy, 0)); 1763 case Builtin::BI__builtin_call_with_static_chain: { 1764 const CallExpr *Call = cast<CallExpr>(E->getArg(0)); 1765 const Expr *Chain = E->getArg(1); 1766 return EmitCall(Call->getCallee()->getType(), 1767 EmitScalarExpr(Call->getCallee()), Call, ReturnValue, 1768 Call->getCalleeDecl(), EmitScalarExpr(Chain)); 1769 } 1770 case Builtin::BI_InterlockedExchange: 1771 case Builtin::BI_InterlockedExchangePointer: 1772 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Xchg, E); 1773 case Builtin::BI_InterlockedCompareExchangePointer: { 1774 llvm::Type *RTy; 1775 llvm::IntegerType *IntType = 1776 IntegerType::get(getLLVMContext(), 1777 getContext().getTypeSize(E->getType())); 1778 llvm::Type *IntPtrType = IntType->getPointerTo(); 1779 1780 llvm::Value *Destination = 1781 Builder.CreateBitCast(EmitScalarExpr(E->getArg(0)), IntPtrType); 1782 1783 llvm::Value *Exchange = EmitScalarExpr(E->getArg(1)); 1784 RTy = Exchange->getType(); 1785 Exchange = Builder.CreatePtrToInt(Exchange, IntType); 1786 1787 llvm::Value *Comparand = 1788 Builder.CreatePtrToInt(EmitScalarExpr(E->getArg(2)), IntType); 1789 1790 auto Result = Builder.CreateAtomicCmpXchg(Destination, Comparand, Exchange, 1791 SequentiallyConsistent, 1792 SequentiallyConsistent); 1793 Result->setVolatile(true); 1794 1795 return RValue::get(Builder.CreateIntToPtr(Builder.CreateExtractValue(Result, 1796 0), 1797 RTy)); 1798 } 1799 case Builtin::BI_InterlockedCompareExchange: { 1800 AtomicCmpXchgInst *CXI = Builder.CreateAtomicCmpXchg( 1801 EmitScalarExpr(E->getArg(0)), 1802 EmitScalarExpr(E->getArg(2)), 1803 EmitScalarExpr(E->getArg(1)), 1804 SequentiallyConsistent, 1805 SequentiallyConsistent); 1806 CXI->setVolatile(true); 1807 return RValue::get(Builder.CreateExtractValue(CXI, 0)); 1808 } 1809 case Builtin::BI_InterlockedIncrement: { 1810 AtomicRMWInst *RMWI = Builder.CreateAtomicRMW( 1811 AtomicRMWInst::Add, 1812 EmitScalarExpr(E->getArg(0)), 1813 ConstantInt::get(Int32Ty, 1), 1814 llvm::SequentiallyConsistent); 1815 RMWI->setVolatile(true); 1816 return RValue::get(Builder.CreateAdd(RMWI, ConstantInt::get(Int32Ty, 1))); 1817 } 1818 case Builtin::BI_InterlockedDecrement: { 1819 AtomicRMWInst *RMWI = Builder.CreateAtomicRMW( 1820 AtomicRMWInst::Sub, 1821 EmitScalarExpr(E->getArg(0)), 1822 ConstantInt::get(Int32Ty, 1), 1823 llvm::SequentiallyConsistent); 1824 RMWI->setVolatile(true); 1825 return RValue::get(Builder.CreateSub(RMWI, ConstantInt::get(Int32Ty, 1))); 1826 } 1827 case Builtin::BI_InterlockedExchangeAdd: { 1828 AtomicRMWInst *RMWI = Builder.CreateAtomicRMW( 1829 AtomicRMWInst::Add, 1830 EmitScalarExpr(E->getArg(0)), 1831 EmitScalarExpr(E->getArg(1)), 1832 llvm::SequentiallyConsistent); 1833 RMWI->setVolatile(true); 1834 return RValue::get(RMWI); 1835 } 1836 case Builtin::BI__readfsdword: { 1837 Value *IntToPtr = 1838 Builder.CreateIntToPtr(EmitScalarExpr(E->getArg(0)), 1839 llvm::PointerType::get(CGM.Int32Ty, 257)); 1840 LoadInst *Load = 1841 Builder.CreateAlignedLoad(IntToPtr, /*Align=*/4, /*isVolatile=*/true); 1842 return RValue::get(Load); 1843 } 1844 1845 case Builtin::BI__exception_code: 1846 case Builtin::BI_exception_code: 1847 return RValue::get(EmitSEHExceptionCode()); 1848 case Builtin::BI__exception_info: 1849 case Builtin::BI_exception_info: 1850 return RValue::get(EmitSEHExceptionInfo()); 1851 case Builtin::BI__abnormal_termination: 1852 case Builtin::BI_abnormal_termination: 1853 return RValue::get(EmitSEHAbnormalTermination()); 1854 case Builtin::BI_setjmpex: { 1855 if (getTarget().getTriple().isOSMSVCRT()) { 1856 llvm::Type *ArgTypes[] = {Int8PtrTy, Int8PtrTy}; 1857 llvm::AttributeSet ReturnsTwiceAttr = 1858 AttributeSet::get(getLLVMContext(), llvm::AttributeSet::FunctionIndex, 1859 llvm::Attribute::ReturnsTwice); 1860 llvm::Constant *SetJmpEx = CGM.CreateRuntimeFunction( 1861 llvm::FunctionType::get(IntTy, ArgTypes, /*isVarArg=*/false), 1862 "_setjmpex", ReturnsTwiceAttr); 1863 llvm::Value *Buf = Builder.CreateBitOrPointerCast( 1864 EmitScalarExpr(E->getArg(0)), Int8PtrTy); 1865 llvm::Value *FrameAddr = 1866 Builder.CreateCall(CGM.getIntrinsic(Intrinsic::frameaddress), 1867 ConstantInt::get(Int32Ty, 0)); 1868 llvm::Value *Args[] = {Buf, FrameAddr}; 1869 llvm::CallSite CS = EmitRuntimeCallOrInvoke(SetJmpEx, Args); 1870 CS.setAttributes(ReturnsTwiceAttr); 1871 return RValue::get(CS.getInstruction()); 1872 } 1873 break; 1874 } 1875 case Builtin::BI_setjmp: { 1876 if (getTarget().getTriple().isOSMSVCRT()) { 1877 llvm::AttributeSet ReturnsTwiceAttr = 1878 AttributeSet::get(getLLVMContext(), llvm::AttributeSet::FunctionIndex, 1879 llvm::Attribute::ReturnsTwice); 1880 llvm::Value *Buf = Builder.CreateBitOrPointerCast( 1881 EmitScalarExpr(E->getArg(0)), Int8PtrTy); 1882 llvm::CallSite CS; 1883 if (getTarget().getTriple().getArch() == llvm::Triple::x86) { 1884 llvm::Type *ArgTypes[] = {Int8PtrTy, IntTy}; 1885 llvm::Constant *SetJmp3 = CGM.CreateRuntimeFunction( 1886 llvm::FunctionType::get(IntTy, ArgTypes, /*isVarArg=*/true), 1887 "_setjmp3", ReturnsTwiceAttr); 1888 llvm::Value *Count = ConstantInt::get(IntTy, 0); 1889 llvm::Value *Args[] = {Buf, Count}; 1890 CS = EmitRuntimeCallOrInvoke(SetJmp3, Args); 1891 } else { 1892 llvm::Type *ArgTypes[] = {Int8PtrTy, Int8PtrTy}; 1893 llvm::Constant *SetJmp = CGM.CreateRuntimeFunction( 1894 llvm::FunctionType::get(IntTy, ArgTypes, /*isVarArg=*/false), 1895 "_setjmp", ReturnsTwiceAttr); 1896 llvm::Value *FrameAddr = 1897 Builder.CreateCall(CGM.getIntrinsic(Intrinsic::frameaddress), 1898 ConstantInt::get(Int32Ty, 0)); 1899 llvm::Value *Args[] = {Buf, FrameAddr}; 1900 CS = EmitRuntimeCallOrInvoke(SetJmp, Args); 1901 } 1902 CS.setAttributes(ReturnsTwiceAttr); 1903 return RValue::get(CS.getInstruction()); 1904 } 1905 break; 1906 } 1907 1908 case Builtin::BI__GetExceptionInfo: { 1909 if (llvm::GlobalVariable *GV = 1910 CGM.getCXXABI().getThrowInfo(FD->getParamDecl(0)->getType())) 1911 return RValue::get(llvm::ConstantExpr::getBitCast(GV, CGM.Int8PtrTy)); 1912 break; 1913 } 1914 } 1915 1916 // If this is an alias for a lib function (e.g. __builtin_sin), emit 1917 // the call using the normal call path, but using the unmangled 1918 // version of the function name. 1919 if (getContext().BuiltinInfo.isLibFunction(BuiltinID)) 1920 return emitLibraryCall(*this, FD, E, 1921 CGM.getBuiltinLibFunction(FD, BuiltinID)); 1922 1923 // If this is a predefined lib function (e.g. malloc), emit the call 1924 // using exactly the normal call path. 1925 if (getContext().BuiltinInfo.isPredefinedLibFunction(BuiltinID)) 1926 return emitLibraryCall(*this, FD, E, EmitScalarExpr(E->getCallee())); 1927 1928 // Check that a call to a target specific builtin has the correct target 1929 // features. 1930 // This is down here to avoid non-target specific builtins, however, if 1931 // generic builtins start to require generic target features then we 1932 // can move this up to the beginning of the function. 1933 checkTargetFeatures(E, FD); 1934 1935 // See if we have a target specific intrinsic. 1936 const char *Name = getContext().BuiltinInfo.getName(BuiltinID); 1937 Intrinsic::ID IntrinsicID = Intrinsic::not_intrinsic; 1938 if (const char *Prefix = 1939 llvm::Triple::getArchTypePrefix(getTarget().getTriple().getArch())) { 1940 IntrinsicID = Intrinsic::getIntrinsicForGCCBuiltin(Prefix, Name); 1941 // NOTE we dont need to perform a compatibility flag check here since the 1942 // intrinsics are declared in Builtins*.def via LANGBUILTIN which filter the 1943 // MS builtins via ALL_MS_LANGUAGES and are filtered earlier. 1944 if (IntrinsicID == Intrinsic::not_intrinsic) 1945 IntrinsicID = Intrinsic::getIntrinsicForMSBuiltin(Prefix, Name); 1946 } 1947 1948 if (IntrinsicID != Intrinsic::not_intrinsic) { 1949 SmallVector<Value*, 16> Args; 1950 1951 // Find out if any arguments are required to be integer constant 1952 // expressions. 1953 unsigned ICEArguments = 0; 1954 ASTContext::GetBuiltinTypeError Error; 1955 getContext().GetBuiltinType(BuiltinID, Error, &ICEArguments); 1956 assert(Error == ASTContext::GE_None && "Should not codegen an error"); 1957 1958 Function *F = CGM.getIntrinsic(IntrinsicID); 1959 llvm::FunctionType *FTy = F->getFunctionType(); 1960 1961 for (unsigned i = 0, e = E->getNumArgs(); i != e; ++i) { 1962 Value *ArgValue; 1963 // If this is a normal argument, just emit it as a scalar. 1964 if ((ICEArguments & (1 << i)) == 0) { 1965 ArgValue = EmitScalarExpr(E->getArg(i)); 1966 } else { 1967 // If this is required to be a constant, constant fold it so that we 1968 // know that the generated intrinsic gets a ConstantInt. 1969 llvm::APSInt Result; 1970 bool IsConst = E->getArg(i)->isIntegerConstantExpr(Result,getContext()); 1971 assert(IsConst && "Constant arg isn't actually constant?"); 1972 (void)IsConst; 1973 ArgValue = llvm::ConstantInt::get(getLLVMContext(), Result); 1974 } 1975 1976 // If the intrinsic arg type is different from the builtin arg type 1977 // we need to do a bit cast. 1978 llvm::Type *PTy = FTy->getParamType(i); 1979 if (PTy != ArgValue->getType()) { 1980 assert(PTy->canLosslesslyBitCastTo(FTy->getParamType(i)) && 1981 "Must be able to losslessly bit cast to param"); 1982 ArgValue = Builder.CreateBitCast(ArgValue, PTy); 1983 } 1984 1985 Args.push_back(ArgValue); 1986 } 1987 1988 Value *V = Builder.CreateCall(F, Args); 1989 QualType BuiltinRetType = E->getType(); 1990 1991 llvm::Type *RetTy = VoidTy; 1992 if (!BuiltinRetType->isVoidType()) 1993 RetTy = ConvertType(BuiltinRetType); 1994 1995 if (RetTy != V->getType()) { 1996 assert(V->getType()->canLosslesslyBitCastTo(RetTy) && 1997 "Must be able to losslessly bit cast result type"); 1998 V = Builder.CreateBitCast(V, RetTy); 1999 } 2000 2001 return RValue::get(V); 2002 } 2003 2004 // See if we have a target specific builtin that needs to be lowered. 2005 if (Value *V = EmitTargetBuiltinExpr(BuiltinID, E)) 2006 return RValue::get(V); 2007 2008 ErrorUnsupported(E, "builtin function"); 2009 2010 // Unknown builtin, for now just dump it out and return undef. 2011 return GetUndefRValue(E->getType()); 2012 } 2013 2014 static Value *EmitTargetArchBuiltinExpr(CodeGenFunction *CGF, 2015 unsigned BuiltinID, const CallExpr *E, 2016 llvm::Triple::ArchType Arch) { 2017 switch (Arch) { 2018 case llvm::Triple::arm: 2019 case llvm::Triple::armeb: 2020 case llvm::Triple::thumb: 2021 case llvm::Triple::thumbeb: 2022 return CGF->EmitARMBuiltinExpr(BuiltinID, E); 2023 case llvm::Triple::aarch64: 2024 case llvm::Triple::aarch64_be: 2025 return CGF->EmitAArch64BuiltinExpr(BuiltinID, E); 2026 case llvm::Triple::x86: 2027 case llvm::Triple::x86_64: 2028 return CGF->EmitX86BuiltinExpr(BuiltinID, E); 2029 case llvm::Triple::ppc: 2030 case llvm::Triple::ppc64: 2031 case llvm::Triple::ppc64le: 2032 return CGF->EmitPPCBuiltinExpr(BuiltinID, E); 2033 case llvm::Triple::r600: 2034 case llvm::Triple::amdgcn: 2035 return CGF->EmitAMDGPUBuiltinExpr(BuiltinID, E); 2036 case llvm::Triple::systemz: 2037 return CGF->EmitSystemZBuiltinExpr(BuiltinID, E); 2038 case llvm::Triple::nvptx: 2039 case llvm::Triple::nvptx64: 2040 return CGF->EmitNVPTXBuiltinExpr(BuiltinID, E); 2041 case llvm::Triple::wasm32: 2042 case llvm::Triple::wasm64: 2043 return CGF->EmitWebAssemblyBuiltinExpr(BuiltinID, E); 2044 default: 2045 return nullptr; 2046 } 2047 } 2048 2049 Value *CodeGenFunction::EmitTargetBuiltinExpr(unsigned BuiltinID, 2050 const CallExpr *E) { 2051 if (getContext().BuiltinInfo.isAuxBuiltinID(BuiltinID)) { 2052 assert(getContext().getAuxTargetInfo() && "Missing aux target info"); 2053 return EmitTargetArchBuiltinExpr( 2054 this, getContext().BuiltinInfo.getAuxBuiltinID(BuiltinID), E, 2055 getContext().getAuxTargetInfo()->getTriple().getArch()); 2056 } 2057 2058 return EmitTargetArchBuiltinExpr(this, BuiltinID, E, 2059 getTarget().getTriple().getArch()); 2060 } 2061 2062 static llvm::VectorType *GetNeonType(CodeGenFunction *CGF, 2063 NeonTypeFlags TypeFlags, 2064 bool V1Ty=false) { 2065 int IsQuad = TypeFlags.isQuad(); 2066 switch (TypeFlags.getEltType()) { 2067 case NeonTypeFlags::Int8: 2068 case NeonTypeFlags::Poly8: 2069 return llvm::VectorType::get(CGF->Int8Ty, V1Ty ? 1 : (8 << IsQuad)); 2070 case NeonTypeFlags::Int16: 2071 case NeonTypeFlags::Poly16: 2072 case NeonTypeFlags::Float16: 2073 return llvm::VectorType::get(CGF->Int16Ty, V1Ty ? 1 : (4 << IsQuad)); 2074 case NeonTypeFlags::Int32: 2075 return llvm::VectorType::get(CGF->Int32Ty, V1Ty ? 1 : (2 << IsQuad)); 2076 case NeonTypeFlags::Int64: 2077 case NeonTypeFlags::Poly64: 2078 return llvm::VectorType::get(CGF->Int64Ty, V1Ty ? 1 : (1 << IsQuad)); 2079 case NeonTypeFlags::Poly128: 2080 // FIXME: i128 and f128 doesn't get fully support in Clang and llvm. 2081 // There is a lot of i128 and f128 API missing. 2082 // so we use v16i8 to represent poly128 and get pattern matched. 2083 return llvm::VectorType::get(CGF->Int8Ty, 16); 2084 case NeonTypeFlags::Float32: 2085 return llvm::VectorType::get(CGF->FloatTy, V1Ty ? 1 : (2 << IsQuad)); 2086 case NeonTypeFlags::Float64: 2087 return llvm::VectorType::get(CGF->DoubleTy, V1Ty ? 1 : (1 << IsQuad)); 2088 } 2089 llvm_unreachable("Unknown vector element type!"); 2090 } 2091 2092 static llvm::VectorType *GetFloatNeonType(CodeGenFunction *CGF, 2093 NeonTypeFlags IntTypeFlags) { 2094 int IsQuad = IntTypeFlags.isQuad(); 2095 switch (IntTypeFlags.getEltType()) { 2096 case NeonTypeFlags::Int32: 2097 return llvm::VectorType::get(CGF->FloatTy, (2 << IsQuad)); 2098 case NeonTypeFlags::Int64: 2099 return llvm::VectorType::get(CGF->DoubleTy, (1 << IsQuad)); 2100 default: 2101 llvm_unreachable("Type can't be converted to floating-point!"); 2102 } 2103 } 2104 2105 Value *CodeGenFunction::EmitNeonSplat(Value *V, Constant *C) { 2106 unsigned nElts = cast<llvm::VectorType>(V->getType())->getNumElements(); 2107 Value* SV = llvm::ConstantVector::getSplat(nElts, C); 2108 return Builder.CreateShuffleVector(V, V, SV, "lane"); 2109 } 2110 2111 Value *CodeGenFunction::EmitNeonCall(Function *F, SmallVectorImpl<Value*> &Ops, 2112 const char *name, 2113 unsigned shift, bool rightshift) { 2114 unsigned j = 0; 2115 for (Function::const_arg_iterator ai = F->arg_begin(), ae = F->arg_end(); 2116 ai != ae; ++ai, ++j) 2117 if (shift > 0 && shift == j) 2118 Ops[j] = EmitNeonShiftVector(Ops[j], ai->getType(), rightshift); 2119 else 2120 Ops[j] = Builder.CreateBitCast(Ops[j], ai->getType(), name); 2121 2122 return Builder.CreateCall(F, Ops, name); 2123 } 2124 2125 Value *CodeGenFunction::EmitNeonShiftVector(Value *V, llvm::Type *Ty, 2126 bool neg) { 2127 int SV = cast<ConstantInt>(V)->getSExtValue(); 2128 return ConstantInt::get(Ty, neg ? -SV : SV); 2129 } 2130 2131 // \brief Right-shift a vector by a constant. 2132 Value *CodeGenFunction::EmitNeonRShiftImm(Value *Vec, Value *Shift, 2133 llvm::Type *Ty, bool usgn, 2134 const char *name) { 2135 llvm::VectorType *VTy = cast<llvm::VectorType>(Ty); 2136 2137 int ShiftAmt = cast<ConstantInt>(Shift)->getSExtValue(); 2138 int EltSize = VTy->getScalarSizeInBits(); 2139 2140 Vec = Builder.CreateBitCast(Vec, Ty); 2141 2142 // lshr/ashr are undefined when the shift amount is equal to the vector 2143 // element size. 2144 if (ShiftAmt == EltSize) { 2145 if (usgn) { 2146 // Right-shifting an unsigned value by its size yields 0. 2147 return llvm::ConstantAggregateZero::get(VTy); 2148 } else { 2149 // Right-shifting a signed value by its size is equivalent 2150 // to a shift of size-1. 2151 --ShiftAmt; 2152 Shift = ConstantInt::get(VTy->getElementType(), ShiftAmt); 2153 } 2154 } 2155 2156 Shift = EmitNeonShiftVector(Shift, Ty, false); 2157 if (usgn) 2158 return Builder.CreateLShr(Vec, Shift, name); 2159 else 2160 return Builder.CreateAShr(Vec, Shift, name); 2161 } 2162 2163 enum { 2164 AddRetType = (1 << 0), 2165 Add1ArgType = (1 << 1), 2166 Add2ArgTypes = (1 << 2), 2167 2168 VectorizeRetType = (1 << 3), 2169 VectorizeArgTypes = (1 << 4), 2170 2171 InventFloatType = (1 << 5), 2172 UnsignedAlts = (1 << 6), 2173 2174 Use64BitVectors = (1 << 7), 2175 Use128BitVectors = (1 << 8), 2176 2177 Vectorize1ArgType = Add1ArgType | VectorizeArgTypes, 2178 VectorRet = AddRetType | VectorizeRetType, 2179 VectorRetGetArgs01 = 2180 AddRetType | Add2ArgTypes | VectorizeRetType | VectorizeArgTypes, 2181 FpCmpzModifiers = 2182 AddRetType | VectorizeRetType | Add1ArgType | InventFloatType 2183 }; 2184 2185 namespace { 2186 struct NeonIntrinsicInfo { 2187 unsigned BuiltinID; 2188 unsigned LLVMIntrinsic; 2189 unsigned AltLLVMIntrinsic; 2190 const char *NameHint; 2191 unsigned TypeModifier; 2192 2193 bool operator<(unsigned RHSBuiltinID) const { 2194 return BuiltinID < RHSBuiltinID; 2195 } 2196 bool operator<(const NeonIntrinsicInfo &TE) const { 2197 return BuiltinID < TE.BuiltinID; 2198 } 2199 }; 2200 } // end anonymous namespace 2201 2202 #define NEONMAP0(NameBase) \ 2203 { NEON::BI__builtin_neon_ ## NameBase, 0, 0, #NameBase, 0 } 2204 2205 #define NEONMAP1(NameBase, LLVMIntrinsic, TypeModifier) \ 2206 { NEON:: BI__builtin_neon_ ## NameBase, \ 2207 Intrinsic::LLVMIntrinsic, 0, #NameBase, TypeModifier } 2208 2209 #define NEONMAP2(NameBase, LLVMIntrinsic, AltLLVMIntrinsic, TypeModifier) \ 2210 { NEON:: BI__builtin_neon_ ## NameBase, \ 2211 Intrinsic::LLVMIntrinsic, Intrinsic::AltLLVMIntrinsic, \ 2212 #NameBase, TypeModifier } 2213 2214 static const NeonIntrinsicInfo ARMSIMDIntrinsicMap [] = { 2215 NEONMAP2(vabd_v, arm_neon_vabdu, arm_neon_vabds, Add1ArgType | UnsignedAlts), 2216 NEONMAP2(vabdq_v, arm_neon_vabdu, arm_neon_vabds, Add1ArgType | UnsignedAlts), 2217 NEONMAP1(vabs_v, arm_neon_vabs, 0), 2218 NEONMAP1(vabsq_v, arm_neon_vabs, 0), 2219 NEONMAP0(vaddhn_v), 2220 NEONMAP1(vaesdq_v, arm_neon_aesd, 0), 2221 NEONMAP1(vaeseq_v, arm_neon_aese, 0), 2222 NEONMAP1(vaesimcq_v, arm_neon_aesimc, 0), 2223 NEONMAP1(vaesmcq_v, arm_neon_aesmc, 0), 2224 NEONMAP1(vbsl_v, arm_neon_vbsl, AddRetType), 2225 NEONMAP1(vbslq_v, arm_neon_vbsl, AddRetType), 2226 NEONMAP1(vcage_v, arm_neon_vacge, 0), 2227 NEONMAP1(vcageq_v, arm_neon_vacge, 0), 2228 NEONMAP1(vcagt_v, arm_neon_vacgt, 0), 2229 NEONMAP1(vcagtq_v, arm_neon_vacgt, 0), 2230 NEONMAP1(vcale_v, arm_neon_vacge, 0), 2231 NEONMAP1(vcaleq_v, arm_neon_vacge, 0), 2232 NEONMAP1(vcalt_v, arm_neon_vacgt, 0), 2233 NEONMAP1(vcaltq_v, arm_neon_vacgt, 0), 2234 NEONMAP1(vcls_v, arm_neon_vcls, Add1ArgType), 2235 NEONMAP1(vclsq_v, arm_neon_vcls, Add1ArgType), 2236 NEONMAP1(vclz_v, ctlz, Add1ArgType), 2237 NEONMAP1(vclzq_v, ctlz, Add1ArgType), 2238 NEONMAP1(vcnt_v, ctpop, Add1ArgType), 2239 NEONMAP1(vcntq_v, ctpop, Add1ArgType), 2240 NEONMAP1(vcvt_f16_f32, arm_neon_vcvtfp2hf, 0), 2241 NEONMAP1(vcvt_f32_f16, arm_neon_vcvthf2fp, 0), 2242 NEONMAP0(vcvt_f32_v), 2243 NEONMAP2(vcvt_n_f32_v, arm_neon_vcvtfxu2fp, arm_neon_vcvtfxs2fp, 0), 2244 NEONMAP1(vcvt_n_s32_v, arm_neon_vcvtfp2fxs, 0), 2245 NEONMAP1(vcvt_n_s64_v, arm_neon_vcvtfp2fxs, 0), 2246 NEONMAP1(vcvt_n_u32_v, arm_neon_vcvtfp2fxu, 0), 2247 NEONMAP1(vcvt_n_u64_v, arm_neon_vcvtfp2fxu, 0), 2248 NEONMAP0(vcvt_s32_v), 2249 NEONMAP0(vcvt_s64_v), 2250 NEONMAP0(vcvt_u32_v), 2251 NEONMAP0(vcvt_u64_v), 2252 NEONMAP1(vcvta_s32_v, arm_neon_vcvtas, 0), 2253 NEONMAP1(vcvta_s64_v, arm_neon_vcvtas, 0), 2254 NEONMAP1(vcvta_u32_v, arm_neon_vcvtau, 0), 2255 NEONMAP1(vcvta_u64_v, arm_neon_vcvtau, 0), 2256 NEONMAP1(vcvtaq_s32_v, arm_neon_vcvtas, 0), 2257 NEONMAP1(vcvtaq_s64_v, arm_neon_vcvtas, 0), 2258 NEONMAP1(vcvtaq_u32_v, arm_neon_vcvtau, 0), 2259 NEONMAP1(vcvtaq_u64_v, arm_neon_vcvtau, 0), 2260 NEONMAP1(vcvtm_s32_v, arm_neon_vcvtms, 0), 2261 NEONMAP1(vcvtm_s64_v, arm_neon_vcvtms, 0), 2262 NEONMAP1(vcvtm_u32_v, arm_neon_vcvtmu, 0), 2263 NEONMAP1(vcvtm_u64_v, arm_neon_vcvtmu, 0), 2264 NEONMAP1(vcvtmq_s32_v, arm_neon_vcvtms, 0), 2265 NEONMAP1(vcvtmq_s64_v, arm_neon_vcvtms, 0), 2266 NEONMAP1(vcvtmq_u32_v, arm_neon_vcvtmu, 0), 2267 NEONMAP1(vcvtmq_u64_v, arm_neon_vcvtmu, 0), 2268 NEONMAP1(vcvtn_s32_v, arm_neon_vcvtns, 0), 2269 NEONMAP1(vcvtn_s64_v, arm_neon_vcvtns, 0), 2270 NEONMAP1(vcvtn_u32_v, arm_neon_vcvtnu, 0), 2271 NEONMAP1(vcvtn_u64_v, arm_neon_vcvtnu, 0), 2272 NEONMAP1(vcvtnq_s32_v, arm_neon_vcvtns, 0), 2273 NEONMAP1(vcvtnq_s64_v, arm_neon_vcvtns, 0), 2274 NEONMAP1(vcvtnq_u32_v, arm_neon_vcvtnu, 0), 2275 NEONMAP1(vcvtnq_u64_v, arm_neon_vcvtnu, 0), 2276 NEONMAP1(vcvtp_s32_v, arm_neon_vcvtps, 0), 2277 NEONMAP1(vcvtp_s64_v, arm_neon_vcvtps, 0), 2278 NEONMAP1(vcvtp_u32_v, arm_neon_vcvtpu, 0), 2279 NEONMAP1(vcvtp_u64_v, arm_neon_vcvtpu, 0), 2280 NEONMAP1(vcvtpq_s32_v, arm_neon_vcvtps, 0), 2281 NEONMAP1(vcvtpq_s64_v, arm_neon_vcvtps, 0), 2282 NEONMAP1(vcvtpq_u32_v, arm_neon_vcvtpu, 0), 2283 NEONMAP1(vcvtpq_u64_v, arm_neon_vcvtpu, 0), 2284 NEONMAP0(vcvtq_f32_v), 2285 NEONMAP2(vcvtq_n_f32_v, arm_neon_vcvtfxu2fp, arm_neon_vcvtfxs2fp, 0), 2286 NEONMAP1(vcvtq_n_s32_v, arm_neon_vcvtfp2fxs, 0), 2287 NEONMAP1(vcvtq_n_s64_v, arm_neon_vcvtfp2fxs, 0), 2288 NEONMAP1(vcvtq_n_u32_v, arm_neon_vcvtfp2fxu, 0), 2289 NEONMAP1(vcvtq_n_u64_v, arm_neon_vcvtfp2fxu, 0), 2290 NEONMAP0(vcvtq_s32_v), 2291 NEONMAP0(vcvtq_s64_v), 2292 NEONMAP0(vcvtq_u32_v), 2293 NEONMAP0(vcvtq_u64_v), 2294 NEONMAP0(vext_v), 2295 NEONMAP0(vextq_v), 2296 NEONMAP0(vfma_v), 2297 NEONMAP0(vfmaq_v), 2298 NEONMAP2(vhadd_v, arm_neon_vhaddu, arm_neon_vhadds, Add1ArgType | UnsignedAlts), 2299 NEONMAP2(vhaddq_v, arm_neon_vhaddu, arm_neon_vhadds, Add1ArgType | UnsignedAlts), 2300 NEONMAP2(vhsub_v, arm_neon_vhsubu, arm_neon_vhsubs, Add1ArgType | UnsignedAlts), 2301 NEONMAP2(vhsubq_v, arm_neon_vhsubu, arm_neon_vhsubs, Add1ArgType | UnsignedAlts), 2302 NEONMAP0(vld1_dup_v), 2303 NEONMAP1(vld1_v, arm_neon_vld1, 0), 2304 NEONMAP0(vld1q_dup_v), 2305 NEONMAP1(vld1q_v, arm_neon_vld1, 0), 2306 NEONMAP1(vld2_lane_v, arm_neon_vld2lane, 0), 2307 NEONMAP1(vld2_v, arm_neon_vld2, 0), 2308 NEONMAP1(vld2q_lane_v, arm_neon_vld2lane, 0), 2309 NEONMAP1(vld2q_v, arm_neon_vld2, 0), 2310 NEONMAP1(vld3_lane_v, arm_neon_vld3lane, 0), 2311 NEONMAP1(vld3_v, arm_neon_vld3, 0), 2312 NEONMAP1(vld3q_lane_v, arm_neon_vld3lane, 0), 2313 NEONMAP1(vld3q_v, arm_neon_vld3, 0), 2314 NEONMAP1(vld4_lane_v, arm_neon_vld4lane, 0), 2315 NEONMAP1(vld4_v, arm_neon_vld4, 0), 2316 NEONMAP1(vld4q_lane_v, arm_neon_vld4lane, 0), 2317 NEONMAP1(vld4q_v, arm_neon_vld4, 0), 2318 NEONMAP2(vmax_v, arm_neon_vmaxu, arm_neon_vmaxs, Add1ArgType | UnsignedAlts), 2319 NEONMAP1(vmaxnm_v, arm_neon_vmaxnm, Add1ArgType), 2320 NEONMAP1(vmaxnmq_v, arm_neon_vmaxnm, Add1ArgType), 2321 NEONMAP2(vmaxq_v, arm_neon_vmaxu, arm_neon_vmaxs, Add1ArgType | UnsignedAlts), 2322 NEONMAP2(vmin_v, arm_neon_vminu, arm_neon_vmins, Add1ArgType | UnsignedAlts), 2323 NEONMAP1(vminnm_v, arm_neon_vminnm, Add1ArgType), 2324 NEONMAP1(vminnmq_v, arm_neon_vminnm, Add1ArgType), 2325 NEONMAP2(vminq_v, arm_neon_vminu, arm_neon_vmins, Add1ArgType | UnsignedAlts), 2326 NEONMAP0(vmovl_v), 2327 NEONMAP0(vmovn_v), 2328 NEONMAP1(vmul_v, arm_neon_vmulp, Add1ArgType), 2329 NEONMAP0(vmull_v), 2330 NEONMAP1(vmulq_v, arm_neon_vmulp, Add1ArgType), 2331 NEONMAP2(vpadal_v, arm_neon_vpadalu, arm_neon_vpadals, UnsignedAlts), 2332 NEONMAP2(vpadalq_v, arm_neon_vpadalu, arm_neon_vpadals, UnsignedAlts), 2333 NEONMAP1(vpadd_v, arm_neon_vpadd, Add1ArgType), 2334 NEONMAP2(vpaddl_v, arm_neon_vpaddlu, arm_neon_vpaddls, UnsignedAlts), 2335 NEONMAP2(vpaddlq_v, arm_neon_vpaddlu, arm_neon_vpaddls, UnsignedAlts), 2336 NEONMAP1(vpaddq_v, arm_neon_vpadd, Add1ArgType), 2337 NEONMAP2(vpmax_v, arm_neon_vpmaxu, arm_neon_vpmaxs, Add1ArgType | UnsignedAlts), 2338 NEONMAP2(vpmin_v, arm_neon_vpminu, arm_neon_vpmins, Add1ArgType | UnsignedAlts), 2339 NEONMAP1(vqabs_v, arm_neon_vqabs, Add1ArgType), 2340 NEONMAP1(vqabsq_v, arm_neon_vqabs, Add1ArgType), 2341 NEONMAP2(vqadd_v, arm_neon_vqaddu, arm_neon_vqadds, Add1ArgType | UnsignedAlts), 2342 NEONMAP2(vqaddq_v, arm_neon_vqaddu, arm_neon_vqadds, Add1ArgType | UnsignedAlts), 2343 NEONMAP2(vqdmlal_v, arm_neon_vqdmull, arm_neon_vqadds, 0), 2344 NEONMAP2(vqdmlsl_v, arm_neon_vqdmull, arm_neon_vqsubs, 0), 2345 NEONMAP1(vqdmulh_v, arm_neon_vqdmulh, Add1ArgType), 2346 NEONMAP1(vqdmulhq_v, arm_neon_vqdmulh, Add1ArgType), 2347 NEONMAP1(vqdmull_v, arm_neon_vqdmull, Add1ArgType), 2348 NEONMAP2(vqmovn_v, arm_neon_vqmovnu, arm_neon_vqmovns, Add1ArgType | UnsignedAlts), 2349 NEONMAP1(vqmovun_v, arm_neon_vqmovnsu, Add1ArgType), 2350 NEONMAP1(vqneg_v, arm_neon_vqneg, Add1ArgType), 2351 NEONMAP1(vqnegq_v, arm_neon_vqneg, Add1ArgType), 2352 NEONMAP1(vqrdmulh_v, arm_neon_vqrdmulh, Add1ArgType), 2353 NEONMAP1(vqrdmulhq_v, arm_neon_vqrdmulh, Add1ArgType), 2354 NEONMAP2(vqrshl_v, arm_neon_vqrshiftu, arm_neon_vqrshifts, Add1ArgType | UnsignedAlts), 2355 NEONMAP2(vqrshlq_v, arm_neon_vqrshiftu, arm_neon_vqrshifts, Add1ArgType | UnsignedAlts), 2356 NEONMAP2(vqshl_n_v, arm_neon_vqshiftu, arm_neon_vqshifts, UnsignedAlts), 2357 NEONMAP2(vqshl_v, arm_neon_vqshiftu, arm_neon_vqshifts, Add1ArgType | UnsignedAlts), 2358 NEONMAP2(vqshlq_n_v, arm_neon_vqshiftu, arm_neon_vqshifts, UnsignedAlts), 2359 NEONMAP2(vqshlq_v, arm_neon_vqshiftu, arm_neon_vqshifts, Add1ArgType | UnsignedAlts), 2360 NEONMAP1(vqshlu_n_v, arm_neon_vqshiftsu, 0), 2361 NEONMAP1(vqshluq_n_v, arm_neon_vqshiftsu, 0), 2362 NEONMAP2(vqsub_v, arm_neon_vqsubu, arm_neon_vqsubs, Add1ArgType | UnsignedAlts), 2363 NEONMAP2(vqsubq_v, arm_neon_vqsubu, arm_neon_vqsubs, Add1ArgType | UnsignedAlts), 2364 NEONMAP1(vraddhn_v, arm_neon_vraddhn, Add1ArgType), 2365 NEONMAP2(vrecpe_v, arm_neon_vrecpe, arm_neon_vrecpe, 0), 2366 NEONMAP2(vrecpeq_v, arm_neon_vrecpe, arm_neon_vrecpe, 0), 2367 NEONMAP1(vrecps_v, arm_neon_vrecps, Add1ArgType), 2368 NEONMAP1(vrecpsq_v, arm_neon_vrecps, Add1ArgType), 2369 NEONMAP2(vrhadd_v, arm_neon_vrhaddu, arm_neon_vrhadds, Add1ArgType | UnsignedAlts), 2370 NEONMAP2(vrhaddq_v, arm_neon_vrhaddu, arm_neon_vrhadds, Add1ArgType | UnsignedAlts), 2371 NEONMAP1(vrnd_v, arm_neon_vrintz, Add1ArgType), 2372 NEONMAP1(vrnda_v, arm_neon_vrinta, Add1ArgType), 2373 NEONMAP1(vrndaq_v, arm_neon_vrinta, Add1ArgType), 2374 NEONMAP1(vrndm_v, arm_neon_vrintm, Add1ArgType), 2375 NEONMAP1(vrndmq_v, arm_neon_vrintm, Add1ArgType), 2376 NEONMAP1(vrndn_v, arm_neon_vrintn, Add1ArgType), 2377 NEONMAP1(vrndnq_v, arm_neon_vrintn, Add1ArgType), 2378 NEONMAP1(vrndp_v, arm_neon_vrintp, Add1ArgType), 2379 NEONMAP1(vrndpq_v, arm_neon_vrintp, Add1ArgType), 2380 NEONMAP1(vrndq_v, arm_neon_vrintz, Add1ArgType), 2381 NEONMAP1(vrndx_v, arm_neon_vrintx, Add1ArgType), 2382 NEONMAP1(vrndxq_v, arm_neon_vrintx, Add1ArgType), 2383 NEONMAP2(vrshl_v, arm_neon_vrshiftu, arm_neon_vrshifts, Add1ArgType | UnsignedAlts), 2384 NEONMAP2(vrshlq_v, arm_neon_vrshiftu, arm_neon_vrshifts, Add1ArgType | UnsignedAlts), 2385 NEONMAP2(vrshr_n_v, arm_neon_vrshiftu, arm_neon_vrshifts, UnsignedAlts), 2386 NEONMAP2(vrshrq_n_v, arm_neon_vrshiftu, arm_neon_vrshifts, UnsignedAlts), 2387 NEONMAP2(vrsqrte_v, arm_neon_vrsqrte, arm_neon_vrsqrte, 0), 2388 NEONMAP2(vrsqrteq_v, arm_neon_vrsqrte, arm_neon_vrsqrte, 0), 2389 NEONMAP1(vrsqrts_v, arm_neon_vrsqrts, Add1ArgType), 2390 NEONMAP1(vrsqrtsq_v, arm_neon_vrsqrts, Add1ArgType), 2391 NEONMAP1(vrsubhn_v, arm_neon_vrsubhn, Add1ArgType), 2392 NEONMAP1(vsha1su0q_v, arm_neon_sha1su0, 0), 2393 NEONMAP1(vsha1su1q_v, arm_neon_sha1su1, 0), 2394 NEONMAP1(vsha256h2q_v, arm_neon_sha256h2, 0), 2395 NEONMAP1(vsha256hq_v, arm_neon_sha256h, 0), 2396 NEONMAP1(vsha256su0q_v, arm_neon_sha256su0, 0), 2397 NEONMAP1(vsha256su1q_v, arm_neon_sha256su1, 0), 2398 NEONMAP0(vshl_n_v), 2399 NEONMAP2(vshl_v, arm_neon_vshiftu, arm_neon_vshifts, Add1ArgType | UnsignedAlts), 2400 NEONMAP0(vshll_n_v), 2401 NEONMAP0(vshlq_n_v), 2402 NEONMAP2(vshlq_v, arm_neon_vshiftu, arm_neon_vshifts, Add1ArgType | UnsignedAlts), 2403 NEONMAP0(vshr_n_v), 2404 NEONMAP0(vshrn_n_v), 2405 NEONMAP0(vshrq_n_v), 2406 NEONMAP1(vst1_v, arm_neon_vst1, 0), 2407 NEONMAP1(vst1q_v, arm_neon_vst1, 0), 2408 NEONMAP1(vst2_lane_v, arm_neon_vst2lane, 0), 2409 NEONMAP1(vst2_v, arm_neon_vst2, 0), 2410 NEONMAP1(vst2q_lane_v, arm_neon_vst2lane, 0), 2411 NEONMAP1(vst2q_v, arm_neon_vst2, 0), 2412 NEONMAP1(vst3_lane_v, arm_neon_vst3lane, 0), 2413 NEONMAP1(vst3_v, arm_neon_vst3, 0), 2414 NEONMAP1(vst3q_lane_v, arm_neon_vst3lane, 0), 2415 NEONMAP1(vst3q_v, arm_neon_vst3, 0), 2416 NEONMAP1(vst4_lane_v, arm_neon_vst4lane, 0), 2417 NEONMAP1(vst4_v, arm_neon_vst4, 0), 2418 NEONMAP1(vst4q_lane_v, arm_neon_vst4lane, 0), 2419 NEONMAP1(vst4q_v, arm_neon_vst4, 0), 2420 NEONMAP0(vsubhn_v), 2421 NEONMAP0(vtrn_v), 2422 NEONMAP0(vtrnq_v), 2423 NEONMAP0(vtst_v), 2424 NEONMAP0(vtstq_v), 2425 NEONMAP0(vuzp_v), 2426 NEONMAP0(vuzpq_v), 2427 NEONMAP0(vzip_v), 2428 NEONMAP0(vzipq_v) 2429 }; 2430 2431 static const NeonIntrinsicInfo AArch64SIMDIntrinsicMap[] = { 2432 NEONMAP1(vabs_v, aarch64_neon_abs, 0), 2433 NEONMAP1(vabsq_v, aarch64_neon_abs, 0), 2434 NEONMAP0(vaddhn_v), 2435 NEONMAP1(vaesdq_v, aarch64_crypto_aesd, 0), 2436 NEONMAP1(vaeseq_v, aarch64_crypto_aese, 0), 2437 NEONMAP1(vaesimcq_v, aarch64_crypto_aesimc, 0), 2438 NEONMAP1(vaesmcq_v, aarch64_crypto_aesmc, 0), 2439 NEONMAP1(vcage_v, aarch64_neon_facge, 0), 2440 NEONMAP1(vcageq_v, aarch64_neon_facge, 0), 2441 NEONMAP1(vcagt_v, aarch64_neon_facgt, 0), 2442 NEONMAP1(vcagtq_v, aarch64_neon_facgt, 0), 2443 NEONMAP1(vcale_v, aarch64_neon_facge, 0), 2444 NEONMAP1(vcaleq_v, aarch64_neon_facge, 0), 2445 NEONMAP1(vcalt_v, aarch64_neon_facgt, 0), 2446 NEONMAP1(vcaltq_v, aarch64_neon_facgt, 0), 2447 NEONMAP1(vcls_v, aarch64_neon_cls, Add1ArgType), 2448 NEONMAP1(vclsq_v, aarch64_neon_cls, Add1ArgType), 2449 NEONMAP1(vclz_v, ctlz, Add1ArgType), 2450 NEONMAP1(vclzq_v, ctlz, Add1ArgType), 2451 NEONMAP1(vcnt_v, ctpop, Add1ArgType), 2452 NEONMAP1(vcntq_v, ctpop, Add1ArgType), 2453 NEONMAP1(vcvt_f16_f32, aarch64_neon_vcvtfp2hf, 0), 2454 NEONMAP1(vcvt_f32_f16, aarch64_neon_vcvthf2fp, 0), 2455 NEONMAP0(vcvt_f32_v), 2456 NEONMAP2(vcvt_n_f32_v, aarch64_neon_vcvtfxu2fp, aarch64_neon_vcvtfxs2fp, 0), 2457 NEONMAP2(vcvt_n_f64_v, aarch64_neon_vcvtfxu2fp, aarch64_neon_vcvtfxs2fp, 0), 2458 NEONMAP1(vcvt_n_s32_v, aarch64_neon_vcvtfp2fxs, 0), 2459 NEONMAP1(vcvt_n_s64_v, aarch64_neon_vcvtfp2fxs, 0), 2460 NEONMAP1(vcvt_n_u32_v, aarch64_neon_vcvtfp2fxu, 0), 2461 NEONMAP1(vcvt_n_u64_v, aarch64_neon_vcvtfp2fxu, 0), 2462 NEONMAP0(vcvtq_f32_v), 2463 NEONMAP2(vcvtq_n_f32_v, aarch64_neon_vcvtfxu2fp, aarch64_neon_vcvtfxs2fp, 0), 2464 NEONMAP2(vcvtq_n_f64_v, aarch64_neon_vcvtfxu2fp, aarch64_neon_vcvtfxs2fp, 0), 2465 NEONMAP1(vcvtq_n_s32_v, aarch64_neon_vcvtfp2fxs, 0), 2466 NEONMAP1(vcvtq_n_s64_v, aarch64_neon_vcvtfp2fxs, 0), 2467 NEONMAP1(vcvtq_n_u32_v, aarch64_neon_vcvtfp2fxu, 0), 2468 NEONMAP1(vcvtq_n_u64_v, aarch64_neon_vcvtfp2fxu, 0), 2469 NEONMAP1(vcvtx_f32_v, aarch64_neon_fcvtxn, AddRetType | Add1ArgType), 2470 NEONMAP0(vext_v), 2471 NEONMAP0(vextq_v), 2472 NEONMAP0(vfma_v), 2473 NEONMAP0(vfmaq_v), 2474 NEONMAP2(vhadd_v, aarch64_neon_uhadd, aarch64_neon_shadd, Add1ArgType | UnsignedAlts), 2475 NEONMAP2(vhaddq_v, aarch64_neon_uhadd, aarch64_neon_shadd, Add1ArgType | UnsignedAlts), 2476 NEONMAP2(vhsub_v, aarch64_neon_uhsub, aarch64_neon_shsub, Add1ArgType | UnsignedAlts), 2477 NEONMAP2(vhsubq_v, aarch64_neon_uhsub, aarch64_neon_shsub, Add1ArgType | UnsignedAlts), 2478 NEONMAP0(vmovl_v), 2479 NEONMAP0(vmovn_v), 2480 NEONMAP1(vmul_v, aarch64_neon_pmul, Add1ArgType), 2481 NEONMAP1(vmulq_v, aarch64_neon_pmul, Add1ArgType), 2482 NEONMAP1(vpadd_v, aarch64_neon_addp, Add1ArgType), 2483 NEONMAP2(vpaddl_v, aarch64_neon_uaddlp, aarch64_neon_saddlp, UnsignedAlts), 2484 NEONMAP2(vpaddlq_v, aarch64_neon_uaddlp, aarch64_neon_saddlp, UnsignedAlts), 2485 NEONMAP1(vpaddq_v, aarch64_neon_addp, Add1ArgType), 2486 NEONMAP1(vqabs_v, aarch64_neon_sqabs, Add1ArgType), 2487 NEONMAP1(vqabsq_v, aarch64_neon_sqabs, Add1ArgType), 2488 NEONMAP2(vqadd_v, aarch64_neon_uqadd, aarch64_neon_sqadd, Add1ArgType | UnsignedAlts), 2489 NEONMAP2(vqaddq_v, aarch64_neon_uqadd, aarch64_neon_sqadd, Add1ArgType | UnsignedAlts), 2490 NEONMAP2(vqdmlal_v, aarch64_neon_sqdmull, aarch64_neon_sqadd, 0), 2491 NEONMAP2(vqdmlsl_v, aarch64_neon_sqdmull, aarch64_neon_sqsub, 0), 2492 NEONMAP1(vqdmulh_v, aarch64_neon_sqdmulh, Add1ArgType), 2493 NEONMAP1(vqdmulhq_v, aarch64_neon_sqdmulh, Add1ArgType), 2494 NEONMAP1(vqdmull_v, aarch64_neon_sqdmull, Add1ArgType), 2495 NEONMAP2(vqmovn_v, aarch64_neon_uqxtn, aarch64_neon_sqxtn, Add1ArgType | UnsignedAlts), 2496 NEONMAP1(vqmovun_v, aarch64_neon_sqxtun, Add1ArgType), 2497 NEONMAP1(vqneg_v, aarch64_neon_sqneg, Add1ArgType), 2498 NEONMAP1(vqnegq_v, aarch64_neon_sqneg, Add1ArgType), 2499 NEONMAP1(vqrdmulh_v, aarch64_neon_sqrdmulh, Add1ArgType), 2500 NEONMAP1(vqrdmulhq_v, aarch64_neon_sqrdmulh, Add1ArgType), 2501 NEONMAP2(vqrshl_v, aarch64_neon_uqrshl, aarch64_neon_sqrshl, Add1ArgType | UnsignedAlts), 2502 NEONMAP2(vqrshlq_v, aarch64_neon_uqrshl, aarch64_neon_sqrshl, Add1ArgType | UnsignedAlts), 2503 NEONMAP2(vqshl_n_v, aarch64_neon_uqshl, aarch64_neon_sqshl, UnsignedAlts), 2504 NEONMAP2(vqshl_v, aarch64_neon_uqshl, aarch64_neon_sqshl, Add1ArgType | UnsignedAlts), 2505 NEONMAP2(vqshlq_n_v, aarch64_neon_uqshl, aarch64_neon_sqshl,UnsignedAlts), 2506 NEONMAP2(vqshlq_v, aarch64_neon_uqshl, aarch64_neon_sqshl, Add1ArgType | UnsignedAlts), 2507 NEONMAP1(vqshlu_n_v, aarch64_neon_sqshlu, 0), 2508 NEONMAP1(vqshluq_n_v, aarch64_neon_sqshlu, 0), 2509 NEONMAP2(vqsub_v, aarch64_neon_uqsub, aarch64_neon_sqsub, Add1ArgType | UnsignedAlts), 2510 NEONMAP2(vqsubq_v, aarch64_neon_uqsub, aarch64_neon_sqsub, Add1ArgType | UnsignedAlts), 2511 NEONMAP1(vraddhn_v, aarch64_neon_raddhn, Add1ArgType), 2512 NEONMAP2(vrecpe_v, aarch64_neon_frecpe, aarch64_neon_urecpe, 0), 2513 NEONMAP2(vrecpeq_v, aarch64_neon_frecpe, aarch64_neon_urecpe, 0), 2514 NEONMAP1(vrecps_v, aarch64_neon_frecps, Add1ArgType), 2515 NEONMAP1(vrecpsq_v, aarch64_neon_frecps, Add1ArgType), 2516 NEONMAP2(vrhadd_v, aarch64_neon_urhadd, aarch64_neon_srhadd, Add1ArgType | UnsignedAlts), 2517 NEONMAP2(vrhaddq_v, aarch64_neon_urhadd, aarch64_neon_srhadd, Add1ArgType | UnsignedAlts), 2518 NEONMAP2(vrshl_v, aarch64_neon_urshl, aarch64_neon_srshl, Add1ArgType | UnsignedAlts), 2519 NEONMAP2(vrshlq_v, aarch64_neon_urshl, aarch64_neon_srshl, Add1ArgType | UnsignedAlts), 2520 NEONMAP2(vrshr_n_v, aarch64_neon_urshl, aarch64_neon_srshl, UnsignedAlts), 2521 NEONMAP2(vrshrq_n_v, aarch64_neon_urshl, aarch64_neon_srshl, UnsignedAlts), 2522 NEONMAP2(vrsqrte_v, aarch64_neon_frsqrte, aarch64_neon_ursqrte, 0), 2523 NEONMAP2(vrsqrteq_v, aarch64_neon_frsqrte, aarch64_neon_ursqrte, 0), 2524 NEONMAP1(vrsqrts_v, aarch64_neon_frsqrts, Add1ArgType), 2525 NEONMAP1(vrsqrtsq_v, aarch64_neon_frsqrts, Add1ArgType), 2526 NEONMAP1(vrsubhn_v, aarch64_neon_rsubhn, Add1ArgType), 2527 NEONMAP1(vsha1su0q_v, aarch64_crypto_sha1su0, 0), 2528 NEONMAP1(vsha1su1q_v, aarch64_crypto_sha1su1, 0), 2529 NEONMAP1(vsha256h2q_v, aarch64_crypto_sha256h2, 0), 2530 NEONMAP1(vsha256hq_v, aarch64_crypto_sha256h, 0), 2531 NEONMAP1(vsha256su0q_v, aarch64_crypto_sha256su0, 0), 2532 NEONMAP1(vsha256su1q_v, aarch64_crypto_sha256su1, 0), 2533 NEONMAP0(vshl_n_v), 2534 NEONMAP2(vshl_v, aarch64_neon_ushl, aarch64_neon_sshl, Add1ArgType | UnsignedAlts), 2535 NEONMAP0(vshll_n_v), 2536 NEONMAP0(vshlq_n_v), 2537 NEONMAP2(vshlq_v, aarch64_neon_ushl, aarch64_neon_sshl, Add1ArgType | UnsignedAlts), 2538 NEONMAP0(vshr_n_v), 2539 NEONMAP0(vshrn_n_v), 2540 NEONMAP0(vshrq_n_v), 2541 NEONMAP0(vsubhn_v), 2542 NEONMAP0(vtst_v), 2543 NEONMAP0(vtstq_v), 2544 }; 2545 2546 static const NeonIntrinsicInfo AArch64SISDIntrinsicMap[] = { 2547 NEONMAP1(vabdd_f64, aarch64_sisd_fabd, Add1ArgType), 2548 NEONMAP1(vabds_f32, aarch64_sisd_fabd, Add1ArgType), 2549 NEONMAP1(vabsd_s64, aarch64_neon_abs, Add1ArgType), 2550 NEONMAP1(vaddlv_s32, aarch64_neon_saddlv, AddRetType | Add1ArgType), 2551 NEONMAP1(vaddlv_u32, aarch64_neon_uaddlv, AddRetType | Add1ArgType), 2552 NEONMAP1(vaddlvq_s32, aarch64_neon_saddlv, AddRetType | Add1ArgType), 2553 NEONMAP1(vaddlvq_u32, aarch64_neon_uaddlv, AddRetType | Add1ArgType), 2554 NEONMAP1(vaddv_f32, aarch64_neon_faddv, AddRetType | Add1ArgType), 2555 NEONMAP1(vaddv_s32, aarch64_neon_saddv, AddRetType | Add1ArgType), 2556 NEONMAP1(vaddv_u32, aarch64_neon_uaddv, AddRetType | Add1ArgType), 2557 NEONMAP1(vaddvq_f32, aarch64_neon_faddv, AddRetType | Add1ArgType), 2558 NEONMAP1(vaddvq_f64, aarch64_neon_faddv, AddRetType | Add1ArgType), 2559 NEONMAP1(vaddvq_s32, aarch64_neon_saddv, AddRetType | Add1ArgType), 2560 NEONMAP1(vaddvq_s64, aarch64_neon_saddv, AddRetType | Add1ArgType), 2561 NEONMAP1(vaddvq_u32, aarch64_neon_uaddv, AddRetType | Add1ArgType), 2562 NEONMAP1(vaddvq_u64, aarch64_neon_uaddv, AddRetType | Add1ArgType), 2563 NEONMAP1(vcaged_f64, aarch64_neon_facge, AddRetType | Add1ArgType), 2564 NEONMAP1(vcages_f32, aarch64_neon_facge, AddRetType | Add1ArgType), 2565 NEONMAP1(vcagtd_f64, aarch64_neon_facgt, AddRetType | Add1ArgType), 2566 NEONMAP1(vcagts_f32, aarch64_neon_facgt, AddRetType | Add1ArgType), 2567 NEONMAP1(vcaled_f64, aarch64_neon_facge, AddRetType | Add1ArgType), 2568 NEONMAP1(vcales_f32, aarch64_neon_facge, AddRetType | Add1ArgType), 2569 NEONMAP1(vcaltd_f64, aarch64_neon_facgt, AddRetType | Add1ArgType), 2570 NEONMAP1(vcalts_f32, aarch64_neon_facgt, AddRetType | Add1ArgType), 2571 NEONMAP1(vcvtad_s64_f64, aarch64_neon_fcvtas, AddRetType | Add1ArgType), 2572 NEONMAP1(vcvtad_u64_f64, aarch64_neon_fcvtau, AddRetType | Add1ArgType), 2573 NEONMAP1(vcvtas_s32_f32, aarch64_neon_fcvtas, AddRetType | Add1ArgType), 2574 NEONMAP1(vcvtas_u32_f32, aarch64_neon_fcvtau, AddRetType | Add1ArgType), 2575 NEONMAP1(vcvtd_n_f64_s64, aarch64_neon_vcvtfxs2fp, AddRetType | Add1ArgType), 2576 NEONMAP1(vcvtd_n_f64_u64, aarch64_neon_vcvtfxu2fp, AddRetType | Add1ArgType), 2577 NEONMAP1(vcvtd_n_s64_f64, aarch64_neon_vcvtfp2fxs, AddRetType | Add1ArgType), 2578 NEONMAP1(vcvtd_n_u64_f64, aarch64_neon_vcvtfp2fxu, AddRetType | Add1ArgType), 2579 NEONMAP1(vcvtmd_s64_f64, aarch64_neon_fcvtms, AddRetType | Add1ArgType), 2580 NEONMAP1(vcvtmd_u64_f64, aarch64_neon_fcvtmu, AddRetType | Add1ArgType), 2581 NEONMAP1(vcvtms_s32_f32, aarch64_neon_fcvtms, AddRetType | Add1ArgType), 2582 NEONMAP1(vcvtms_u32_f32, aarch64_neon_fcvtmu, AddRetType | Add1ArgType), 2583 NEONMAP1(vcvtnd_s64_f64, aarch64_neon_fcvtns, AddRetType | Add1ArgType), 2584 NEONMAP1(vcvtnd_u64_f64, aarch64_neon_fcvtnu, AddRetType | Add1ArgType), 2585 NEONMAP1(vcvtns_s32_f32, aarch64_neon_fcvtns, AddRetType | Add1ArgType), 2586 NEONMAP1(vcvtns_u32_f32, aarch64_neon_fcvtnu, AddRetType | Add1ArgType), 2587 NEONMAP1(vcvtpd_s64_f64, aarch64_neon_fcvtps, AddRetType | Add1ArgType), 2588 NEONMAP1(vcvtpd_u64_f64, aarch64_neon_fcvtpu, AddRetType | Add1ArgType), 2589 NEONMAP1(vcvtps_s32_f32, aarch64_neon_fcvtps, AddRetType | Add1ArgType), 2590 NEONMAP1(vcvtps_u32_f32, aarch64_neon_fcvtpu, AddRetType | Add1ArgType), 2591 NEONMAP1(vcvts_n_f32_s32, aarch64_neon_vcvtfxs2fp, AddRetType | Add1ArgType), 2592 NEONMAP1(vcvts_n_f32_u32, aarch64_neon_vcvtfxu2fp, AddRetType | Add1ArgType), 2593 NEONMAP1(vcvts_n_s32_f32, aarch64_neon_vcvtfp2fxs, AddRetType | Add1ArgType), 2594 NEONMAP1(vcvts_n_u32_f32, aarch64_neon_vcvtfp2fxu, AddRetType | Add1ArgType), 2595 NEONMAP1(vcvtxd_f32_f64, aarch64_sisd_fcvtxn, 0), 2596 NEONMAP1(vmaxnmv_f32, aarch64_neon_fmaxnmv, AddRetType | Add1ArgType), 2597 NEONMAP1(vmaxnmvq_f32, aarch64_neon_fmaxnmv, AddRetType | Add1ArgType), 2598 NEONMAP1(vmaxnmvq_f64, aarch64_neon_fmaxnmv, AddRetType | Add1ArgType), 2599 NEONMAP1(vmaxv_f32, aarch64_neon_fmaxv, AddRetType | Add1ArgType), 2600 NEONMAP1(vmaxv_s32, aarch64_neon_smaxv, AddRetType | Add1ArgType), 2601 NEONMAP1(vmaxv_u32, aarch64_neon_umaxv, AddRetType | Add1ArgType), 2602 NEONMAP1(vmaxvq_f32, aarch64_neon_fmaxv, AddRetType | Add1ArgType), 2603 NEONMAP1(vmaxvq_f64, aarch64_neon_fmaxv, AddRetType | Add1ArgType), 2604 NEONMAP1(vmaxvq_s32, aarch64_neon_smaxv, AddRetType | Add1ArgType), 2605 NEONMAP1(vmaxvq_u32, aarch64_neon_umaxv, AddRetType | Add1ArgType), 2606 NEONMAP1(vminnmv_f32, aarch64_neon_fminnmv, AddRetType | Add1ArgType), 2607 NEONMAP1(vminnmvq_f32, aarch64_neon_fminnmv, AddRetType | Add1ArgType), 2608 NEONMAP1(vminnmvq_f64, aarch64_neon_fminnmv, AddRetType | Add1ArgType), 2609 NEONMAP1(vminv_f32, aarch64_neon_fminv, AddRetType | Add1ArgType), 2610 NEONMAP1(vminv_s32, aarch64_neon_sminv, AddRetType | Add1ArgType), 2611 NEONMAP1(vminv_u32, aarch64_neon_uminv, AddRetType | Add1ArgType), 2612 NEONMAP1(vminvq_f32, aarch64_neon_fminv, AddRetType | Add1ArgType), 2613 NEONMAP1(vminvq_f64, aarch64_neon_fminv, AddRetType | Add1ArgType), 2614 NEONMAP1(vminvq_s32, aarch64_neon_sminv, AddRetType | Add1ArgType), 2615 NEONMAP1(vminvq_u32, aarch64_neon_uminv, AddRetType | Add1ArgType), 2616 NEONMAP1(vmull_p64, aarch64_neon_pmull64, 0), 2617 NEONMAP1(vmulxd_f64, aarch64_neon_fmulx, Add1ArgType), 2618 NEONMAP1(vmulxs_f32, aarch64_neon_fmulx, Add1ArgType), 2619 NEONMAP1(vpaddd_s64, aarch64_neon_uaddv, AddRetType | Add1ArgType), 2620 NEONMAP1(vpaddd_u64, aarch64_neon_uaddv, AddRetType | Add1ArgType), 2621 NEONMAP1(vpmaxnmqd_f64, aarch64_neon_fmaxnmv, AddRetType | Add1ArgType), 2622 NEONMAP1(vpmaxnms_f32, aarch64_neon_fmaxnmv, AddRetType | Add1ArgType), 2623 NEONMAP1(vpmaxqd_f64, aarch64_neon_fmaxv, AddRetType | Add1ArgType), 2624 NEONMAP1(vpmaxs_f32, aarch64_neon_fmaxv, AddRetType | Add1ArgType), 2625 NEONMAP1(vpminnmqd_f64, aarch64_neon_fminnmv, AddRetType | Add1ArgType), 2626 NEONMAP1(vpminnms_f32, aarch64_neon_fminnmv, AddRetType | Add1ArgType), 2627 NEONMAP1(vpminqd_f64, aarch64_neon_fminv, AddRetType | Add1ArgType), 2628 NEONMAP1(vpmins_f32, aarch64_neon_fminv, AddRetType | Add1ArgType), 2629 NEONMAP1(vqabsb_s8, aarch64_neon_sqabs, Vectorize1ArgType | Use64BitVectors), 2630 NEONMAP1(vqabsd_s64, aarch64_neon_sqabs, Add1ArgType), 2631 NEONMAP1(vqabsh_s16, aarch64_neon_sqabs, Vectorize1ArgType | Use64BitVectors), 2632 NEONMAP1(vqabss_s32, aarch64_neon_sqabs, Add1ArgType), 2633 NEONMAP1(vqaddb_s8, aarch64_neon_sqadd, Vectorize1ArgType | Use64BitVectors), 2634 NEONMAP1(vqaddb_u8, aarch64_neon_uqadd, Vectorize1ArgType | Use64BitVectors), 2635 NEONMAP1(vqaddd_s64, aarch64_neon_sqadd, Add1ArgType), 2636 NEONMAP1(vqaddd_u64, aarch64_neon_uqadd, Add1ArgType), 2637 NEONMAP1(vqaddh_s16, aarch64_neon_sqadd, Vectorize1ArgType | Use64BitVectors), 2638 NEONMAP1(vqaddh_u16, aarch64_neon_uqadd, Vectorize1ArgType | Use64BitVectors), 2639 NEONMAP1(vqadds_s32, aarch64_neon_sqadd, Add1ArgType), 2640 NEONMAP1(vqadds_u32, aarch64_neon_uqadd, Add1ArgType), 2641 NEONMAP1(vqdmulhh_s16, aarch64_neon_sqdmulh, Vectorize1ArgType | Use64BitVectors), 2642 NEONMAP1(vqdmulhs_s32, aarch64_neon_sqdmulh, Add1ArgType), 2643 NEONMAP1(vqdmullh_s16, aarch64_neon_sqdmull, VectorRet | Use128BitVectors), 2644 NEONMAP1(vqdmulls_s32, aarch64_neon_sqdmulls_scalar, 0), 2645 NEONMAP1(vqmovnd_s64, aarch64_neon_scalar_sqxtn, AddRetType | Add1ArgType), 2646 NEONMAP1(vqmovnd_u64, aarch64_neon_scalar_uqxtn, AddRetType | Add1ArgType), 2647 NEONMAP1(vqmovnh_s16, aarch64_neon_sqxtn, VectorRet | Use64BitVectors), 2648 NEONMAP1(vqmovnh_u16, aarch64_neon_uqxtn, VectorRet | Use64BitVectors), 2649 NEONMAP1(vqmovns_s32, aarch64_neon_sqxtn, VectorRet | Use64BitVectors), 2650 NEONMAP1(vqmovns_u32, aarch64_neon_uqxtn, VectorRet | Use64BitVectors), 2651 NEONMAP1(vqmovund_s64, aarch64_neon_scalar_sqxtun, AddRetType | Add1ArgType), 2652 NEONMAP1(vqmovunh_s16, aarch64_neon_sqxtun, VectorRet | Use64BitVectors), 2653 NEONMAP1(vqmovuns_s32, aarch64_neon_sqxtun, VectorRet | Use64BitVectors), 2654 NEONMAP1(vqnegb_s8, aarch64_neon_sqneg, Vectorize1ArgType | Use64BitVectors), 2655 NEONMAP1(vqnegd_s64, aarch64_neon_sqneg, Add1ArgType), 2656 NEONMAP1(vqnegh_s16, aarch64_neon_sqneg, Vectorize1ArgType | Use64BitVectors), 2657 NEONMAP1(vqnegs_s32, aarch64_neon_sqneg, Add1ArgType), 2658 NEONMAP1(vqrdmulhh_s16, aarch64_neon_sqrdmulh, Vectorize1ArgType | Use64BitVectors), 2659 NEONMAP1(vqrdmulhs_s32, aarch64_neon_sqrdmulh, Add1ArgType), 2660 NEONMAP1(vqrshlb_s8, aarch64_neon_sqrshl, Vectorize1ArgType | Use64BitVectors), 2661 NEONMAP1(vqrshlb_u8, aarch64_neon_uqrshl, Vectorize1ArgType | Use64BitVectors), 2662 NEONMAP1(vqrshld_s64, aarch64_neon_sqrshl, Add1ArgType), 2663 NEONMAP1(vqrshld_u64, aarch64_neon_uqrshl, Add1ArgType), 2664 NEONMAP1(vqrshlh_s16, aarch64_neon_sqrshl, Vectorize1ArgType | Use64BitVectors), 2665 NEONMAP1(vqrshlh_u16, aarch64_neon_uqrshl, Vectorize1ArgType | Use64BitVectors), 2666 NEONMAP1(vqrshls_s32, aarch64_neon_sqrshl, Add1ArgType), 2667 NEONMAP1(vqrshls_u32, aarch64_neon_uqrshl, Add1ArgType), 2668 NEONMAP1(vqrshrnd_n_s64, aarch64_neon_sqrshrn, AddRetType), 2669 NEONMAP1(vqrshrnd_n_u64, aarch64_neon_uqrshrn, AddRetType), 2670 NEONMAP1(vqrshrnh_n_s16, aarch64_neon_sqrshrn, VectorRet | Use64BitVectors), 2671 NEONMAP1(vqrshrnh_n_u16, aarch64_neon_uqrshrn, VectorRet | Use64BitVectors), 2672 NEONMAP1(vqrshrns_n_s32, aarch64_neon_sqrshrn, VectorRet | Use64BitVectors), 2673 NEONMAP1(vqrshrns_n_u32, aarch64_neon_uqrshrn, VectorRet | Use64BitVectors), 2674 NEONMAP1(vqrshrund_n_s64, aarch64_neon_sqrshrun, AddRetType), 2675 NEONMAP1(vqrshrunh_n_s16, aarch64_neon_sqrshrun, VectorRet | Use64BitVectors), 2676 NEONMAP1(vqrshruns_n_s32, aarch64_neon_sqrshrun, VectorRet | Use64BitVectors), 2677 NEONMAP1(vqshlb_n_s8, aarch64_neon_sqshl, Vectorize1ArgType | Use64BitVectors), 2678 NEONMAP1(vqshlb_n_u8, aarch64_neon_uqshl, Vectorize1ArgType | Use64BitVectors), 2679 NEONMAP1(vqshlb_s8, aarch64_neon_sqshl, Vectorize1ArgType | Use64BitVectors), 2680 NEONMAP1(vqshlb_u8, aarch64_neon_uqshl, Vectorize1ArgType | Use64BitVectors), 2681 NEONMAP1(vqshld_s64, aarch64_neon_sqshl, Add1ArgType), 2682 NEONMAP1(vqshld_u64, aarch64_neon_uqshl, Add1ArgType), 2683 NEONMAP1(vqshlh_n_s16, aarch64_neon_sqshl, Vectorize1ArgType | Use64BitVectors), 2684 NEONMAP1(vqshlh_n_u16, aarch64_neon_uqshl, Vectorize1ArgType | Use64BitVectors), 2685 NEONMAP1(vqshlh_s16, aarch64_neon_sqshl, Vectorize1ArgType | Use64BitVectors), 2686 NEONMAP1(vqshlh_u16, aarch64_neon_uqshl, Vectorize1ArgType | Use64BitVectors), 2687 NEONMAP1(vqshls_n_s32, aarch64_neon_sqshl, Add1ArgType), 2688 NEONMAP1(vqshls_n_u32, aarch64_neon_uqshl, Add1ArgType), 2689 NEONMAP1(vqshls_s32, aarch64_neon_sqshl, Add1ArgType), 2690 NEONMAP1(vqshls_u32, aarch64_neon_uqshl, Add1ArgType), 2691 NEONMAP1(vqshlub_n_s8, aarch64_neon_sqshlu, Vectorize1ArgType | Use64BitVectors), 2692 NEONMAP1(vqshluh_n_s16, aarch64_neon_sqshlu, Vectorize1ArgType | Use64BitVectors), 2693 NEONMAP1(vqshlus_n_s32, aarch64_neon_sqshlu, Add1ArgType), 2694 NEONMAP1(vqshrnd_n_s64, aarch64_neon_sqshrn, AddRetType), 2695 NEONMAP1(vqshrnd_n_u64, aarch64_neon_uqshrn, AddRetType), 2696 NEONMAP1(vqshrnh_n_s16, aarch64_neon_sqshrn, VectorRet | Use64BitVectors), 2697 NEONMAP1(vqshrnh_n_u16, aarch64_neon_uqshrn, VectorRet | Use64BitVectors), 2698 NEONMAP1(vqshrns_n_s32, aarch64_neon_sqshrn, VectorRet | Use64BitVectors), 2699 NEONMAP1(vqshrns_n_u32, aarch64_neon_uqshrn, VectorRet | Use64BitVectors), 2700 NEONMAP1(vqshrund_n_s64, aarch64_neon_sqshrun, AddRetType), 2701 NEONMAP1(vqshrunh_n_s16, aarch64_neon_sqshrun, VectorRet | Use64BitVectors), 2702 NEONMAP1(vqshruns_n_s32, aarch64_neon_sqshrun, VectorRet | Use64BitVectors), 2703 NEONMAP1(vqsubb_s8, aarch64_neon_sqsub, Vectorize1ArgType | Use64BitVectors), 2704 NEONMAP1(vqsubb_u8, aarch64_neon_uqsub, Vectorize1ArgType | Use64BitVectors), 2705 NEONMAP1(vqsubd_s64, aarch64_neon_sqsub, Add1ArgType), 2706 NEONMAP1(vqsubd_u64, aarch64_neon_uqsub, Add1ArgType), 2707 NEONMAP1(vqsubh_s16, aarch64_neon_sqsub, Vectorize1ArgType | Use64BitVectors), 2708 NEONMAP1(vqsubh_u16, aarch64_neon_uqsub, Vectorize1ArgType | Use64BitVectors), 2709 NEONMAP1(vqsubs_s32, aarch64_neon_sqsub, Add1ArgType), 2710 NEONMAP1(vqsubs_u32, aarch64_neon_uqsub, Add1ArgType), 2711 NEONMAP1(vrecped_f64, aarch64_neon_frecpe, Add1ArgType), 2712 NEONMAP1(vrecpes_f32, aarch64_neon_frecpe, Add1ArgType), 2713 NEONMAP1(vrecpxd_f64, aarch64_neon_frecpx, Add1ArgType), 2714 NEONMAP1(vrecpxs_f32, aarch64_neon_frecpx, Add1ArgType), 2715 NEONMAP1(vrshld_s64, aarch64_neon_srshl, Add1ArgType), 2716 NEONMAP1(vrshld_u64, aarch64_neon_urshl, Add1ArgType), 2717 NEONMAP1(vrsqrted_f64, aarch64_neon_frsqrte, Add1ArgType), 2718 NEONMAP1(vrsqrtes_f32, aarch64_neon_frsqrte, Add1ArgType), 2719 NEONMAP1(vrsqrtsd_f64, aarch64_neon_frsqrts, Add1ArgType), 2720 NEONMAP1(vrsqrtss_f32, aarch64_neon_frsqrts, Add1ArgType), 2721 NEONMAP1(vsha1cq_u32, aarch64_crypto_sha1c, 0), 2722 NEONMAP1(vsha1h_u32, aarch64_crypto_sha1h, 0), 2723 NEONMAP1(vsha1mq_u32, aarch64_crypto_sha1m, 0), 2724 NEONMAP1(vsha1pq_u32, aarch64_crypto_sha1p, 0), 2725 NEONMAP1(vshld_s64, aarch64_neon_sshl, Add1ArgType), 2726 NEONMAP1(vshld_u64, aarch64_neon_ushl, Add1ArgType), 2727 NEONMAP1(vslid_n_s64, aarch64_neon_vsli, Vectorize1ArgType), 2728 NEONMAP1(vslid_n_u64, aarch64_neon_vsli, Vectorize1ArgType), 2729 NEONMAP1(vsqaddb_u8, aarch64_neon_usqadd, Vectorize1ArgType | Use64BitVectors), 2730 NEONMAP1(vsqaddd_u64, aarch64_neon_usqadd, Add1ArgType), 2731 NEONMAP1(vsqaddh_u16, aarch64_neon_usqadd, Vectorize1ArgType | Use64BitVectors), 2732 NEONMAP1(vsqadds_u32, aarch64_neon_usqadd, Add1ArgType), 2733 NEONMAP1(vsrid_n_s64, aarch64_neon_vsri, Vectorize1ArgType), 2734 NEONMAP1(vsrid_n_u64, aarch64_neon_vsri, Vectorize1ArgType), 2735 NEONMAP1(vuqaddb_s8, aarch64_neon_suqadd, Vectorize1ArgType | Use64BitVectors), 2736 NEONMAP1(vuqaddd_s64, aarch64_neon_suqadd, Add1ArgType), 2737 NEONMAP1(vuqaddh_s16, aarch64_neon_suqadd, Vectorize1ArgType | Use64BitVectors), 2738 NEONMAP1(vuqadds_s32, aarch64_neon_suqadd, Add1ArgType), 2739 }; 2740 2741 #undef NEONMAP0 2742 #undef NEONMAP1 2743 #undef NEONMAP2 2744 2745 static bool NEONSIMDIntrinsicsProvenSorted = false; 2746 2747 static bool AArch64SIMDIntrinsicsProvenSorted = false; 2748 static bool AArch64SISDIntrinsicsProvenSorted = false; 2749 2750 2751 static const NeonIntrinsicInfo * 2752 findNeonIntrinsicInMap(ArrayRef<NeonIntrinsicInfo> IntrinsicMap, 2753 unsigned BuiltinID, bool &MapProvenSorted) { 2754 2755 #ifndef NDEBUG 2756 if (!MapProvenSorted) { 2757 assert(std::is_sorted(std::begin(IntrinsicMap), std::end(IntrinsicMap))); 2758 MapProvenSorted = true; 2759 } 2760 #endif 2761 2762 const NeonIntrinsicInfo *Builtin = 2763 std::lower_bound(IntrinsicMap.begin(), IntrinsicMap.end(), BuiltinID); 2764 2765 if (Builtin != IntrinsicMap.end() && Builtin->BuiltinID == BuiltinID) 2766 return Builtin; 2767 2768 return nullptr; 2769 } 2770 2771 Function *CodeGenFunction::LookupNeonLLVMIntrinsic(unsigned IntrinsicID, 2772 unsigned Modifier, 2773 llvm::Type *ArgType, 2774 const CallExpr *E) { 2775 int VectorSize = 0; 2776 if (Modifier & Use64BitVectors) 2777 VectorSize = 64; 2778 else if (Modifier & Use128BitVectors) 2779 VectorSize = 128; 2780 2781 // Return type. 2782 SmallVector<llvm::Type *, 3> Tys; 2783 if (Modifier & AddRetType) { 2784 llvm::Type *Ty = ConvertType(E->getCallReturnType(getContext())); 2785 if (Modifier & VectorizeRetType) 2786 Ty = llvm::VectorType::get( 2787 Ty, VectorSize ? VectorSize / Ty->getPrimitiveSizeInBits() : 1); 2788 2789 Tys.push_back(Ty); 2790 } 2791 2792 // Arguments. 2793 if (Modifier & VectorizeArgTypes) { 2794 int Elts = VectorSize ? VectorSize / ArgType->getPrimitiveSizeInBits() : 1; 2795 ArgType = llvm::VectorType::get(ArgType, Elts); 2796 } 2797 2798 if (Modifier & (Add1ArgType | Add2ArgTypes)) 2799 Tys.push_back(ArgType); 2800 2801 if (Modifier & Add2ArgTypes) 2802 Tys.push_back(ArgType); 2803 2804 if (Modifier & InventFloatType) 2805 Tys.push_back(FloatTy); 2806 2807 return CGM.getIntrinsic(IntrinsicID, Tys); 2808 } 2809 2810 static Value *EmitCommonNeonSISDBuiltinExpr(CodeGenFunction &CGF, 2811 const NeonIntrinsicInfo &SISDInfo, 2812 SmallVectorImpl<Value *> &Ops, 2813 const CallExpr *E) { 2814 unsigned BuiltinID = SISDInfo.BuiltinID; 2815 unsigned int Int = SISDInfo.LLVMIntrinsic; 2816 unsigned Modifier = SISDInfo.TypeModifier; 2817 const char *s = SISDInfo.NameHint; 2818 2819 switch (BuiltinID) { 2820 case NEON::BI__builtin_neon_vcled_s64: 2821 case NEON::BI__builtin_neon_vcled_u64: 2822 case NEON::BI__builtin_neon_vcles_f32: 2823 case NEON::BI__builtin_neon_vcled_f64: 2824 case NEON::BI__builtin_neon_vcltd_s64: 2825 case NEON::BI__builtin_neon_vcltd_u64: 2826 case NEON::BI__builtin_neon_vclts_f32: 2827 case NEON::BI__builtin_neon_vcltd_f64: 2828 case NEON::BI__builtin_neon_vcales_f32: 2829 case NEON::BI__builtin_neon_vcaled_f64: 2830 case NEON::BI__builtin_neon_vcalts_f32: 2831 case NEON::BI__builtin_neon_vcaltd_f64: 2832 // Only one direction of comparisons actually exist, cmle is actually a cmge 2833 // with swapped operands. The table gives us the right intrinsic but we 2834 // still need to do the swap. 2835 std::swap(Ops[0], Ops[1]); 2836 break; 2837 } 2838 2839 assert(Int && "Generic code assumes a valid intrinsic"); 2840 2841 // Determine the type(s) of this overloaded AArch64 intrinsic. 2842 const Expr *Arg = E->getArg(0); 2843 llvm::Type *ArgTy = CGF.ConvertType(Arg->getType()); 2844 Function *F = CGF.LookupNeonLLVMIntrinsic(Int, Modifier, ArgTy, E); 2845 2846 int j = 0; 2847 ConstantInt *C0 = ConstantInt::get(CGF.SizeTy, 0); 2848 for (Function::const_arg_iterator ai = F->arg_begin(), ae = F->arg_end(); 2849 ai != ae; ++ai, ++j) { 2850 llvm::Type *ArgTy = ai->getType(); 2851 if (Ops[j]->getType()->getPrimitiveSizeInBits() == 2852 ArgTy->getPrimitiveSizeInBits()) 2853 continue; 2854 2855 assert(ArgTy->isVectorTy() && !Ops[j]->getType()->isVectorTy()); 2856 // The constant argument to an _n_ intrinsic always has Int32Ty, so truncate 2857 // it before inserting. 2858 Ops[j] = 2859 CGF.Builder.CreateTruncOrBitCast(Ops[j], ArgTy->getVectorElementType()); 2860 Ops[j] = 2861 CGF.Builder.CreateInsertElement(UndefValue::get(ArgTy), Ops[j], C0); 2862 } 2863 2864 Value *Result = CGF.EmitNeonCall(F, Ops, s); 2865 llvm::Type *ResultType = CGF.ConvertType(E->getType()); 2866 if (ResultType->getPrimitiveSizeInBits() < 2867 Result->getType()->getPrimitiveSizeInBits()) 2868 return CGF.Builder.CreateExtractElement(Result, C0); 2869 2870 return CGF.Builder.CreateBitCast(Result, ResultType, s); 2871 } 2872 2873 Value *CodeGenFunction::EmitCommonNeonBuiltinExpr( 2874 unsigned BuiltinID, unsigned LLVMIntrinsic, unsigned AltLLVMIntrinsic, 2875 const char *NameHint, unsigned Modifier, const CallExpr *E, 2876 SmallVectorImpl<llvm::Value *> &Ops, Address PtrOp0, Address PtrOp1) { 2877 // Get the last argument, which specifies the vector type. 2878 llvm::APSInt NeonTypeConst; 2879 const Expr *Arg = E->getArg(E->getNumArgs() - 1); 2880 if (!Arg->isIntegerConstantExpr(NeonTypeConst, getContext())) 2881 return nullptr; 2882 2883 // Determine the type of this overloaded NEON intrinsic. 2884 NeonTypeFlags Type(NeonTypeConst.getZExtValue()); 2885 bool Usgn = Type.isUnsigned(); 2886 bool Quad = Type.isQuad(); 2887 2888 llvm::VectorType *VTy = GetNeonType(this, Type); 2889 llvm::Type *Ty = VTy; 2890 if (!Ty) 2891 return nullptr; 2892 2893 auto getAlignmentValue32 = [&](Address addr) -> Value* { 2894 return Builder.getInt32(addr.getAlignment().getQuantity()); 2895 }; 2896 2897 unsigned Int = LLVMIntrinsic; 2898 if ((Modifier & UnsignedAlts) && !Usgn) 2899 Int = AltLLVMIntrinsic; 2900 2901 switch (BuiltinID) { 2902 default: break; 2903 case NEON::BI__builtin_neon_vabs_v: 2904 case NEON::BI__builtin_neon_vabsq_v: 2905 if (VTy->getElementType()->isFloatingPointTy()) 2906 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::fabs, Ty), Ops, "vabs"); 2907 return EmitNeonCall(CGM.getIntrinsic(LLVMIntrinsic, Ty), Ops, "vabs"); 2908 case NEON::BI__builtin_neon_vaddhn_v: { 2909 llvm::VectorType *SrcTy = 2910 llvm::VectorType::getExtendedElementVectorType(VTy); 2911 2912 // %sum = add <4 x i32> %lhs, %rhs 2913 Ops[0] = Builder.CreateBitCast(Ops[0], SrcTy); 2914 Ops[1] = Builder.CreateBitCast(Ops[1], SrcTy); 2915 Ops[0] = Builder.CreateAdd(Ops[0], Ops[1], "vaddhn"); 2916 2917 // %high = lshr <4 x i32> %sum, <i32 16, i32 16, i32 16, i32 16> 2918 Constant *ShiftAmt = 2919 ConstantInt::get(SrcTy, SrcTy->getScalarSizeInBits() / 2); 2920 Ops[0] = Builder.CreateLShr(Ops[0], ShiftAmt, "vaddhn"); 2921 2922 // %res = trunc <4 x i32> %high to <4 x i16> 2923 return Builder.CreateTrunc(Ops[0], VTy, "vaddhn"); 2924 } 2925 case NEON::BI__builtin_neon_vcale_v: 2926 case NEON::BI__builtin_neon_vcaleq_v: 2927 case NEON::BI__builtin_neon_vcalt_v: 2928 case NEON::BI__builtin_neon_vcaltq_v: 2929 std::swap(Ops[0], Ops[1]); 2930 case NEON::BI__builtin_neon_vcage_v: 2931 case NEON::BI__builtin_neon_vcageq_v: 2932 case NEON::BI__builtin_neon_vcagt_v: 2933 case NEON::BI__builtin_neon_vcagtq_v: { 2934 llvm::Type *VecFlt = llvm::VectorType::get( 2935 VTy->getScalarSizeInBits() == 32 ? FloatTy : DoubleTy, 2936 VTy->getNumElements()); 2937 llvm::Type *Tys[] = { VTy, VecFlt }; 2938 Function *F = CGM.getIntrinsic(LLVMIntrinsic, Tys); 2939 return EmitNeonCall(F, Ops, NameHint); 2940 } 2941 case NEON::BI__builtin_neon_vclz_v: 2942 case NEON::BI__builtin_neon_vclzq_v: 2943 // We generate target-independent intrinsic, which needs a second argument 2944 // for whether or not clz of zero is undefined; on ARM it isn't. 2945 Ops.push_back(Builder.getInt1(getTarget().isCLZForZeroUndef())); 2946 break; 2947 case NEON::BI__builtin_neon_vcvt_f32_v: 2948 case NEON::BI__builtin_neon_vcvtq_f32_v: 2949 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 2950 Ty = GetNeonType(this, NeonTypeFlags(NeonTypeFlags::Float32, false, Quad)); 2951 return Usgn ? Builder.CreateUIToFP(Ops[0], Ty, "vcvt") 2952 : Builder.CreateSIToFP(Ops[0], Ty, "vcvt"); 2953 case NEON::BI__builtin_neon_vcvt_n_f32_v: 2954 case NEON::BI__builtin_neon_vcvt_n_f64_v: 2955 case NEON::BI__builtin_neon_vcvtq_n_f32_v: 2956 case NEON::BI__builtin_neon_vcvtq_n_f64_v: { 2957 llvm::Type *Tys[2] = { GetFloatNeonType(this, Type), Ty }; 2958 Int = Usgn ? LLVMIntrinsic : AltLLVMIntrinsic; 2959 Function *F = CGM.getIntrinsic(Int, Tys); 2960 return EmitNeonCall(F, Ops, "vcvt_n"); 2961 } 2962 case NEON::BI__builtin_neon_vcvt_n_s32_v: 2963 case NEON::BI__builtin_neon_vcvt_n_u32_v: 2964 case NEON::BI__builtin_neon_vcvt_n_s64_v: 2965 case NEON::BI__builtin_neon_vcvt_n_u64_v: 2966 case NEON::BI__builtin_neon_vcvtq_n_s32_v: 2967 case NEON::BI__builtin_neon_vcvtq_n_u32_v: 2968 case NEON::BI__builtin_neon_vcvtq_n_s64_v: 2969 case NEON::BI__builtin_neon_vcvtq_n_u64_v: { 2970 llvm::Type *Tys[2] = { Ty, GetFloatNeonType(this, Type) }; 2971 Function *F = CGM.getIntrinsic(LLVMIntrinsic, Tys); 2972 return EmitNeonCall(F, Ops, "vcvt_n"); 2973 } 2974 case NEON::BI__builtin_neon_vcvt_s32_v: 2975 case NEON::BI__builtin_neon_vcvt_u32_v: 2976 case NEON::BI__builtin_neon_vcvt_s64_v: 2977 case NEON::BI__builtin_neon_vcvt_u64_v: 2978 case NEON::BI__builtin_neon_vcvtq_s32_v: 2979 case NEON::BI__builtin_neon_vcvtq_u32_v: 2980 case NEON::BI__builtin_neon_vcvtq_s64_v: 2981 case NEON::BI__builtin_neon_vcvtq_u64_v: { 2982 Ops[0] = Builder.CreateBitCast(Ops[0], GetFloatNeonType(this, Type)); 2983 return Usgn ? Builder.CreateFPToUI(Ops[0], Ty, "vcvt") 2984 : Builder.CreateFPToSI(Ops[0], Ty, "vcvt"); 2985 } 2986 case NEON::BI__builtin_neon_vcvta_s32_v: 2987 case NEON::BI__builtin_neon_vcvta_s64_v: 2988 case NEON::BI__builtin_neon_vcvta_u32_v: 2989 case NEON::BI__builtin_neon_vcvta_u64_v: 2990 case NEON::BI__builtin_neon_vcvtaq_s32_v: 2991 case NEON::BI__builtin_neon_vcvtaq_s64_v: 2992 case NEON::BI__builtin_neon_vcvtaq_u32_v: 2993 case NEON::BI__builtin_neon_vcvtaq_u64_v: 2994 case NEON::BI__builtin_neon_vcvtn_s32_v: 2995 case NEON::BI__builtin_neon_vcvtn_s64_v: 2996 case NEON::BI__builtin_neon_vcvtn_u32_v: 2997 case NEON::BI__builtin_neon_vcvtn_u64_v: 2998 case NEON::BI__builtin_neon_vcvtnq_s32_v: 2999 case NEON::BI__builtin_neon_vcvtnq_s64_v: 3000 case NEON::BI__builtin_neon_vcvtnq_u32_v: 3001 case NEON::BI__builtin_neon_vcvtnq_u64_v: 3002 case NEON::BI__builtin_neon_vcvtp_s32_v: 3003 case NEON::BI__builtin_neon_vcvtp_s64_v: 3004 case NEON::BI__builtin_neon_vcvtp_u32_v: 3005 case NEON::BI__builtin_neon_vcvtp_u64_v: 3006 case NEON::BI__builtin_neon_vcvtpq_s32_v: 3007 case NEON::BI__builtin_neon_vcvtpq_s64_v: 3008 case NEON::BI__builtin_neon_vcvtpq_u32_v: 3009 case NEON::BI__builtin_neon_vcvtpq_u64_v: 3010 case NEON::BI__builtin_neon_vcvtm_s32_v: 3011 case NEON::BI__builtin_neon_vcvtm_s64_v: 3012 case NEON::BI__builtin_neon_vcvtm_u32_v: 3013 case NEON::BI__builtin_neon_vcvtm_u64_v: 3014 case NEON::BI__builtin_neon_vcvtmq_s32_v: 3015 case NEON::BI__builtin_neon_vcvtmq_s64_v: 3016 case NEON::BI__builtin_neon_vcvtmq_u32_v: 3017 case NEON::BI__builtin_neon_vcvtmq_u64_v: { 3018 llvm::Type *Tys[2] = { Ty, GetFloatNeonType(this, Type) }; 3019 return EmitNeonCall(CGM.getIntrinsic(LLVMIntrinsic, Tys), Ops, NameHint); 3020 } 3021 case NEON::BI__builtin_neon_vext_v: 3022 case NEON::BI__builtin_neon_vextq_v: { 3023 int CV = cast<ConstantInt>(Ops[2])->getSExtValue(); 3024 SmallVector<Constant*, 16> Indices; 3025 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) 3026 Indices.push_back(ConstantInt::get(Int32Ty, i+CV)); 3027 3028 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 3029 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 3030 Value *SV = llvm::ConstantVector::get(Indices); 3031 return Builder.CreateShuffleVector(Ops[0], Ops[1], SV, "vext"); 3032 } 3033 case NEON::BI__builtin_neon_vfma_v: 3034 case NEON::BI__builtin_neon_vfmaq_v: { 3035 Value *F = CGM.getIntrinsic(Intrinsic::fma, Ty); 3036 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 3037 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 3038 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 3039 3040 // NEON intrinsic puts accumulator first, unlike the LLVM fma. 3041 return Builder.CreateCall(F, {Ops[1], Ops[2], Ops[0]}); 3042 } 3043 case NEON::BI__builtin_neon_vld1_v: 3044 case NEON::BI__builtin_neon_vld1q_v: { 3045 llvm::Type *Tys[] = {Ty, Int8PtrTy}; 3046 Ops.push_back(getAlignmentValue32(PtrOp0)); 3047 return EmitNeonCall(CGM.getIntrinsic(LLVMIntrinsic, Tys), Ops, "vld1"); 3048 } 3049 case NEON::BI__builtin_neon_vld2_v: 3050 case NEON::BI__builtin_neon_vld2q_v: 3051 case NEON::BI__builtin_neon_vld3_v: 3052 case NEON::BI__builtin_neon_vld3q_v: 3053 case NEON::BI__builtin_neon_vld4_v: 3054 case NEON::BI__builtin_neon_vld4q_v: { 3055 llvm::Type *Tys[] = {Ty, Int8PtrTy}; 3056 Function *F = CGM.getIntrinsic(LLVMIntrinsic, Tys); 3057 Value *Align = getAlignmentValue32(PtrOp1); 3058 Ops[1] = Builder.CreateCall(F, {Ops[1], Align}, NameHint); 3059 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 3060 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 3061 return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]); 3062 } 3063 case NEON::BI__builtin_neon_vld1_dup_v: 3064 case NEON::BI__builtin_neon_vld1q_dup_v: { 3065 Value *V = UndefValue::get(Ty); 3066 Ty = llvm::PointerType::getUnqual(VTy->getElementType()); 3067 PtrOp0 = Builder.CreateBitCast(PtrOp0, Ty); 3068 LoadInst *Ld = Builder.CreateLoad(PtrOp0); 3069 llvm::Constant *CI = ConstantInt::get(SizeTy, 0); 3070 Ops[0] = Builder.CreateInsertElement(V, Ld, CI); 3071 return EmitNeonSplat(Ops[0], CI); 3072 } 3073 case NEON::BI__builtin_neon_vld2_lane_v: 3074 case NEON::BI__builtin_neon_vld2q_lane_v: 3075 case NEON::BI__builtin_neon_vld3_lane_v: 3076 case NEON::BI__builtin_neon_vld3q_lane_v: 3077 case NEON::BI__builtin_neon_vld4_lane_v: 3078 case NEON::BI__builtin_neon_vld4q_lane_v: { 3079 llvm::Type *Tys[] = {Ty, Int8PtrTy}; 3080 Function *F = CGM.getIntrinsic(LLVMIntrinsic, Tys); 3081 for (unsigned I = 2; I < Ops.size() - 1; ++I) 3082 Ops[I] = Builder.CreateBitCast(Ops[I], Ty); 3083 Ops.push_back(getAlignmentValue32(PtrOp1)); 3084 Ops[1] = Builder.CreateCall(F, makeArrayRef(Ops).slice(1), NameHint); 3085 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 3086 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 3087 return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]); 3088 } 3089 case NEON::BI__builtin_neon_vmovl_v: { 3090 llvm::Type *DTy =llvm::VectorType::getTruncatedElementVectorType(VTy); 3091 Ops[0] = Builder.CreateBitCast(Ops[0], DTy); 3092 if (Usgn) 3093 return Builder.CreateZExt(Ops[0], Ty, "vmovl"); 3094 return Builder.CreateSExt(Ops[0], Ty, "vmovl"); 3095 } 3096 case NEON::BI__builtin_neon_vmovn_v: { 3097 llvm::Type *QTy = llvm::VectorType::getExtendedElementVectorType(VTy); 3098 Ops[0] = Builder.CreateBitCast(Ops[0], QTy); 3099 return Builder.CreateTrunc(Ops[0], Ty, "vmovn"); 3100 } 3101 case NEON::BI__builtin_neon_vmull_v: 3102 // FIXME: the integer vmull operations could be emitted in terms of pure 3103 // LLVM IR (2 exts followed by a mul). Unfortunately LLVM has a habit of 3104 // hoisting the exts outside loops. Until global ISel comes along that can 3105 // see through such movement this leads to bad CodeGen. So we need an 3106 // intrinsic for now. 3107 Int = Usgn ? Intrinsic::arm_neon_vmullu : Intrinsic::arm_neon_vmulls; 3108 Int = Type.isPoly() ? (unsigned)Intrinsic::arm_neon_vmullp : Int; 3109 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmull"); 3110 case NEON::BI__builtin_neon_vpadal_v: 3111 case NEON::BI__builtin_neon_vpadalq_v: { 3112 // The source operand type has twice as many elements of half the size. 3113 unsigned EltBits = VTy->getElementType()->getPrimitiveSizeInBits(); 3114 llvm::Type *EltTy = 3115 llvm::IntegerType::get(getLLVMContext(), EltBits / 2); 3116 llvm::Type *NarrowTy = 3117 llvm::VectorType::get(EltTy, VTy->getNumElements() * 2); 3118 llvm::Type *Tys[2] = { Ty, NarrowTy }; 3119 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, NameHint); 3120 } 3121 case NEON::BI__builtin_neon_vpaddl_v: 3122 case NEON::BI__builtin_neon_vpaddlq_v: { 3123 // The source operand type has twice as many elements of half the size. 3124 unsigned EltBits = VTy->getElementType()->getPrimitiveSizeInBits(); 3125 llvm::Type *EltTy = llvm::IntegerType::get(getLLVMContext(), EltBits / 2); 3126 llvm::Type *NarrowTy = 3127 llvm::VectorType::get(EltTy, VTy->getNumElements() * 2); 3128 llvm::Type *Tys[2] = { Ty, NarrowTy }; 3129 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vpaddl"); 3130 } 3131 case NEON::BI__builtin_neon_vqdmlal_v: 3132 case NEON::BI__builtin_neon_vqdmlsl_v: { 3133 SmallVector<Value *, 2> MulOps(Ops.begin() + 1, Ops.end()); 3134 Ops[1] = 3135 EmitNeonCall(CGM.getIntrinsic(LLVMIntrinsic, Ty), MulOps, "vqdmlal"); 3136 Ops.resize(2); 3137 return EmitNeonCall(CGM.getIntrinsic(AltLLVMIntrinsic, Ty), Ops, NameHint); 3138 } 3139 case NEON::BI__builtin_neon_vqshl_n_v: 3140 case NEON::BI__builtin_neon_vqshlq_n_v: 3141 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqshl_n", 3142 1, false); 3143 case NEON::BI__builtin_neon_vqshlu_n_v: 3144 case NEON::BI__builtin_neon_vqshluq_n_v: 3145 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqshlu_n", 3146 1, false); 3147 case NEON::BI__builtin_neon_vrecpe_v: 3148 case NEON::BI__builtin_neon_vrecpeq_v: 3149 case NEON::BI__builtin_neon_vrsqrte_v: 3150 case NEON::BI__builtin_neon_vrsqrteq_v: 3151 Int = Ty->isFPOrFPVectorTy() ? LLVMIntrinsic : AltLLVMIntrinsic; 3152 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, NameHint); 3153 3154 case NEON::BI__builtin_neon_vrshr_n_v: 3155 case NEON::BI__builtin_neon_vrshrq_n_v: 3156 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrshr_n", 3157 1, true); 3158 case NEON::BI__builtin_neon_vshl_n_v: 3159 case NEON::BI__builtin_neon_vshlq_n_v: 3160 Ops[1] = EmitNeonShiftVector(Ops[1], Ty, false); 3161 return Builder.CreateShl(Builder.CreateBitCast(Ops[0],Ty), Ops[1], 3162 "vshl_n"); 3163 case NEON::BI__builtin_neon_vshll_n_v: { 3164 llvm::Type *SrcTy = llvm::VectorType::getTruncatedElementVectorType(VTy); 3165 Ops[0] = Builder.CreateBitCast(Ops[0], SrcTy); 3166 if (Usgn) 3167 Ops[0] = Builder.CreateZExt(Ops[0], VTy); 3168 else 3169 Ops[0] = Builder.CreateSExt(Ops[0], VTy); 3170 Ops[1] = EmitNeonShiftVector(Ops[1], VTy, false); 3171 return Builder.CreateShl(Ops[0], Ops[1], "vshll_n"); 3172 } 3173 case NEON::BI__builtin_neon_vshrn_n_v: { 3174 llvm::Type *SrcTy = llvm::VectorType::getExtendedElementVectorType(VTy); 3175 Ops[0] = Builder.CreateBitCast(Ops[0], SrcTy); 3176 Ops[1] = EmitNeonShiftVector(Ops[1], SrcTy, false); 3177 if (Usgn) 3178 Ops[0] = Builder.CreateLShr(Ops[0], Ops[1]); 3179 else 3180 Ops[0] = Builder.CreateAShr(Ops[0], Ops[1]); 3181 return Builder.CreateTrunc(Ops[0], Ty, "vshrn_n"); 3182 } 3183 case NEON::BI__builtin_neon_vshr_n_v: 3184 case NEON::BI__builtin_neon_vshrq_n_v: 3185 return EmitNeonRShiftImm(Ops[0], Ops[1], Ty, Usgn, "vshr_n"); 3186 case NEON::BI__builtin_neon_vst1_v: 3187 case NEON::BI__builtin_neon_vst1q_v: 3188 case NEON::BI__builtin_neon_vst2_v: 3189 case NEON::BI__builtin_neon_vst2q_v: 3190 case NEON::BI__builtin_neon_vst3_v: 3191 case NEON::BI__builtin_neon_vst3q_v: 3192 case NEON::BI__builtin_neon_vst4_v: 3193 case NEON::BI__builtin_neon_vst4q_v: 3194 case NEON::BI__builtin_neon_vst2_lane_v: 3195 case NEON::BI__builtin_neon_vst2q_lane_v: 3196 case NEON::BI__builtin_neon_vst3_lane_v: 3197 case NEON::BI__builtin_neon_vst3q_lane_v: 3198 case NEON::BI__builtin_neon_vst4_lane_v: 3199 case NEON::BI__builtin_neon_vst4q_lane_v: { 3200 llvm::Type *Tys[] = {Int8PtrTy, Ty}; 3201 Ops.push_back(getAlignmentValue32(PtrOp0)); 3202 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, ""); 3203 } 3204 case NEON::BI__builtin_neon_vsubhn_v: { 3205 llvm::VectorType *SrcTy = 3206 llvm::VectorType::getExtendedElementVectorType(VTy); 3207 3208 // %sum = add <4 x i32> %lhs, %rhs 3209 Ops[0] = Builder.CreateBitCast(Ops[0], SrcTy); 3210 Ops[1] = Builder.CreateBitCast(Ops[1], SrcTy); 3211 Ops[0] = Builder.CreateSub(Ops[0], Ops[1], "vsubhn"); 3212 3213 // %high = lshr <4 x i32> %sum, <i32 16, i32 16, i32 16, i32 16> 3214 Constant *ShiftAmt = 3215 ConstantInt::get(SrcTy, SrcTy->getScalarSizeInBits() / 2); 3216 Ops[0] = Builder.CreateLShr(Ops[0], ShiftAmt, "vsubhn"); 3217 3218 // %res = trunc <4 x i32> %high to <4 x i16> 3219 return Builder.CreateTrunc(Ops[0], VTy, "vsubhn"); 3220 } 3221 case NEON::BI__builtin_neon_vtrn_v: 3222 case NEON::BI__builtin_neon_vtrnq_v: { 3223 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty)); 3224 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 3225 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 3226 Value *SV = nullptr; 3227 3228 for (unsigned vi = 0; vi != 2; ++vi) { 3229 SmallVector<Constant*, 16> Indices; 3230 for (unsigned i = 0, e = VTy->getNumElements(); i != e; i += 2) { 3231 Indices.push_back(Builder.getInt32(i+vi)); 3232 Indices.push_back(Builder.getInt32(i+e+vi)); 3233 } 3234 Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ty, Ops[0], vi); 3235 SV = llvm::ConstantVector::get(Indices); 3236 SV = Builder.CreateShuffleVector(Ops[1], Ops[2], SV, "vtrn"); 3237 SV = Builder.CreateDefaultAlignedStore(SV, Addr); 3238 } 3239 return SV; 3240 } 3241 case NEON::BI__builtin_neon_vtst_v: 3242 case NEON::BI__builtin_neon_vtstq_v: { 3243 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 3244 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 3245 Ops[0] = Builder.CreateAnd(Ops[0], Ops[1]); 3246 Ops[0] = Builder.CreateICmp(ICmpInst::ICMP_NE, Ops[0], 3247 ConstantAggregateZero::get(Ty)); 3248 return Builder.CreateSExt(Ops[0], Ty, "vtst"); 3249 } 3250 case NEON::BI__builtin_neon_vuzp_v: 3251 case NEON::BI__builtin_neon_vuzpq_v: { 3252 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty)); 3253 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 3254 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 3255 Value *SV = nullptr; 3256 3257 for (unsigned vi = 0; vi != 2; ++vi) { 3258 SmallVector<Constant*, 16> Indices; 3259 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) 3260 Indices.push_back(ConstantInt::get(Int32Ty, 2*i+vi)); 3261 3262 Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ty, Ops[0], vi); 3263 SV = llvm::ConstantVector::get(Indices); 3264 SV = Builder.CreateShuffleVector(Ops[1], Ops[2], SV, "vuzp"); 3265 SV = Builder.CreateDefaultAlignedStore(SV, Addr); 3266 } 3267 return SV; 3268 } 3269 case NEON::BI__builtin_neon_vzip_v: 3270 case NEON::BI__builtin_neon_vzipq_v: { 3271 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty)); 3272 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 3273 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 3274 Value *SV = nullptr; 3275 3276 for (unsigned vi = 0; vi != 2; ++vi) { 3277 SmallVector<Constant*, 16> Indices; 3278 for (unsigned i = 0, e = VTy->getNumElements(); i != e; i += 2) { 3279 Indices.push_back(ConstantInt::get(Int32Ty, (i + vi*e) >> 1)); 3280 Indices.push_back(ConstantInt::get(Int32Ty, ((i + vi*e) >> 1)+e)); 3281 } 3282 Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ty, Ops[0], vi); 3283 SV = llvm::ConstantVector::get(Indices); 3284 SV = Builder.CreateShuffleVector(Ops[1], Ops[2], SV, "vzip"); 3285 SV = Builder.CreateDefaultAlignedStore(SV, Addr); 3286 } 3287 return SV; 3288 } 3289 } 3290 3291 assert(Int && "Expected valid intrinsic number"); 3292 3293 // Determine the type(s) of this overloaded AArch64 intrinsic. 3294 Function *F = LookupNeonLLVMIntrinsic(Int, Modifier, Ty, E); 3295 3296 Value *Result = EmitNeonCall(F, Ops, NameHint); 3297 llvm::Type *ResultType = ConvertType(E->getType()); 3298 // AArch64 intrinsic one-element vector type cast to 3299 // scalar type expected by the builtin 3300 return Builder.CreateBitCast(Result, ResultType, NameHint); 3301 } 3302 3303 Value *CodeGenFunction::EmitAArch64CompareBuiltinExpr( 3304 Value *Op, llvm::Type *Ty, const CmpInst::Predicate Fp, 3305 const CmpInst::Predicate Ip, const Twine &Name) { 3306 llvm::Type *OTy = Op->getType(); 3307 3308 // FIXME: this is utterly horrific. We should not be looking at previous 3309 // codegen context to find out what needs doing. Unfortunately TableGen 3310 // currently gives us exactly the same calls for vceqz_f32 and vceqz_s32 3311 // (etc). 3312 if (BitCastInst *BI = dyn_cast<BitCastInst>(Op)) 3313 OTy = BI->getOperand(0)->getType(); 3314 3315 Op = Builder.CreateBitCast(Op, OTy); 3316 if (OTy->getScalarType()->isFloatingPointTy()) { 3317 Op = Builder.CreateFCmp(Fp, Op, Constant::getNullValue(OTy)); 3318 } else { 3319 Op = Builder.CreateICmp(Ip, Op, Constant::getNullValue(OTy)); 3320 } 3321 return Builder.CreateSExt(Op, Ty, Name); 3322 } 3323 3324 static Value *packTBLDVectorList(CodeGenFunction &CGF, ArrayRef<Value *> Ops, 3325 Value *ExtOp, Value *IndexOp, 3326 llvm::Type *ResTy, unsigned IntID, 3327 const char *Name) { 3328 SmallVector<Value *, 2> TblOps; 3329 if (ExtOp) 3330 TblOps.push_back(ExtOp); 3331 3332 // Build a vector containing sequential number like (0, 1, 2, ..., 15) 3333 SmallVector<Constant*, 16> Indices; 3334 llvm::VectorType *TblTy = cast<llvm::VectorType>(Ops[0]->getType()); 3335 for (unsigned i = 0, e = TblTy->getNumElements(); i != e; ++i) { 3336 Indices.push_back(ConstantInt::get(CGF.Int32Ty, 2*i)); 3337 Indices.push_back(ConstantInt::get(CGF.Int32Ty, 2*i+1)); 3338 } 3339 Value *SV = llvm::ConstantVector::get(Indices); 3340 3341 int PairPos = 0, End = Ops.size() - 1; 3342 while (PairPos < End) { 3343 TblOps.push_back(CGF.Builder.CreateShuffleVector(Ops[PairPos], 3344 Ops[PairPos+1], SV, Name)); 3345 PairPos += 2; 3346 } 3347 3348 // If there's an odd number of 64-bit lookup table, fill the high 64-bit 3349 // of the 128-bit lookup table with zero. 3350 if (PairPos == End) { 3351 Value *ZeroTbl = ConstantAggregateZero::get(TblTy); 3352 TblOps.push_back(CGF.Builder.CreateShuffleVector(Ops[PairPos], 3353 ZeroTbl, SV, Name)); 3354 } 3355 3356 Function *TblF; 3357 TblOps.push_back(IndexOp); 3358 TblF = CGF.CGM.getIntrinsic(IntID, ResTy); 3359 3360 return CGF.EmitNeonCall(TblF, TblOps, Name); 3361 } 3362 3363 Value *CodeGenFunction::GetValueForARMHint(unsigned BuiltinID) { 3364 unsigned Value; 3365 switch (BuiltinID) { 3366 default: 3367 return nullptr; 3368 case ARM::BI__builtin_arm_nop: 3369 Value = 0; 3370 break; 3371 case ARM::BI__builtin_arm_yield: 3372 case ARM::BI__yield: 3373 Value = 1; 3374 break; 3375 case ARM::BI__builtin_arm_wfe: 3376 case ARM::BI__wfe: 3377 Value = 2; 3378 break; 3379 case ARM::BI__builtin_arm_wfi: 3380 case ARM::BI__wfi: 3381 Value = 3; 3382 break; 3383 case ARM::BI__builtin_arm_sev: 3384 case ARM::BI__sev: 3385 Value = 4; 3386 break; 3387 case ARM::BI__builtin_arm_sevl: 3388 case ARM::BI__sevl: 3389 Value = 5; 3390 break; 3391 } 3392 3393 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::arm_hint), 3394 llvm::ConstantInt::get(Int32Ty, Value)); 3395 } 3396 3397 // Generates the IR for the read/write special register builtin, 3398 // ValueType is the type of the value that is to be written or read, 3399 // RegisterType is the type of the register being written to or read from. 3400 static Value *EmitSpecialRegisterBuiltin(CodeGenFunction &CGF, 3401 const CallExpr *E, 3402 llvm::Type *RegisterType, 3403 llvm::Type *ValueType, bool IsRead) { 3404 // write and register intrinsics only support 32 and 64 bit operations. 3405 assert((RegisterType->isIntegerTy(32) || RegisterType->isIntegerTy(64)) 3406 && "Unsupported size for register."); 3407 3408 CodeGen::CGBuilderTy &Builder = CGF.Builder; 3409 CodeGen::CodeGenModule &CGM = CGF.CGM; 3410 LLVMContext &Context = CGM.getLLVMContext(); 3411 3412 const Expr *SysRegStrExpr = E->getArg(0)->IgnoreParenCasts(); 3413 StringRef SysReg = cast<StringLiteral>(SysRegStrExpr)->getString(); 3414 3415 llvm::Metadata *Ops[] = { llvm::MDString::get(Context, SysReg) }; 3416 llvm::MDNode *RegName = llvm::MDNode::get(Context, Ops); 3417 llvm::Value *Metadata = llvm::MetadataAsValue::get(Context, RegName); 3418 3419 llvm::Type *Types[] = { RegisterType }; 3420 3421 bool MixedTypes = RegisterType->isIntegerTy(64) && ValueType->isIntegerTy(32); 3422 assert(!(RegisterType->isIntegerTy(32) && ValueType->isIntegerTy(64)) 3423 && "Can't fit 64-bit value in 32-bit register"); 3424 3425 if (IsRead) { 3426 llvm::Value *F = CGM.getIntrinsic(llvm::Intrinsic::read_register, Types); 3427 llvm::Value *Call = Builder.CreateCall(F, Metadata); 3428 3429 if (MixedTypes) 3430 // Read into 64 bit register and then truncate result to 32 bit. 3431 return Builder.CreateTrunc(Call, ValueType); 3432 3433 if (ValueType->isPointerTy()) 3434 // Have i32/i64 result (Call) but want to return a VoidPtrTy (i8*). 3435 return Builder.CreateIntToPtr(Call, ValueType); 3436 3437 return Call; 3438 } 3439 3440 llvm::Value *F = CGM.getIntrinsic(llvm::Intrinsic::write_register, Types); 3441 llvm::Value *ArgValue = CGF.EmitScalarExpr(E->getArg(1)); 3442 if (MixedTypes) { 3443 // Extend 32 bit write value to 64 bit to pass to write. 3444 ArgValue = Builder.CreateZExt(ArgValue, RegisterType); 3445 return Builder.CreateCall(F, { Metadata, ArgValue }); 3446 } 3447 3448 if (ValueType->isPointerTy()) { 3449 // Have VoidPtrTy ArgValue but want to return an i32/i64. 3450 ArgValue = Builder.CreatePtrToInt(ArgValue, RegisterType); 3451 return Builder.CreateCall(F, { Metadata, ArgValue }); 3452 } 3453 3454 return Builder.CreateCall(F, { Metadata, ArgValue }); 3455 } 3456 3457 /// Return true if BuiltinID is an overloaded Neon intrinsic with an extra 3458 /// argument that specifies the vector type. 3459 static bool HasExtraNeonArgument(unsigned BuiltinID) { 3460 switch (BuiltinID) { 3461 default: break; 3462 case NEON::BI__builtin_neon_vget_lane_i8: 3463 case NEON::BI__builtin_neon_vget_lane_i16: 3464 case NEON::BI__builtin_neon_vget_lane_i32: 3465 case NEON::BI__builtin_neon_vget_lane_i64: 3466 case NEON::BI__builtin_neon_vget_lane_f32: 3467 case NEON::BI__builtin_neon_vgetq_lane_i8: 3468 case NEON::BI__builtin_neon_vgetq_lane_i16: 3469 case NEON::BI__builtin_neon_vgetq_lane_i32: 3470 case NEON::BI__builtin_neon_vgetq_lane_i64: 3471 case NEON::BI__builtin_neon_vgetq_lane_f32: 3472 case NEON::BI__builtin_neon_vset_lane_i8: 3473 case NEON::BI__builtin_neon_vset_lane_i16: 3474 case NEON::BI__builtin_neon_vset_lane_i32: 3475 case NEON::BI__builtin_neon_vset_lane_i64: 3476 case NEON::BI__builtin_neon_vset_lane_f32: 3477 case NEON::BI__builtin_neon_vsetq_lane_i8: 3478 case NEON::BI__builtin_neon_vsetq_lane_i16: 3479 case NEON::BI__builtin_neon_vsetq_lane_i32: 3480 case NEON::BI__builtin_neon_vsetq_lane_i64: 3481 case NEON::BI__builtin_neon_vsetq_lane_f32: 3482 case NEON::BI__builtin_neon_vsha1h_u32: 3483 case NEON::BI__builtin_neon_vsha1cq_u32: 3484 case NEON::BI__builtin_neon_vsha1pq_u32: 3485 case NEON::BI__builtin_neon_vsha1mq_u32: 3486 case ARM::BI_MoveToCoprocessor: 3487 case ARM::BI_MoveToCoprocessor2: 3488 return false; 3489 } 3490 return true; 3491 } 3492 3493 Value *CodeGenFunction::EmitARMBuiltinExpr(unsigned BuiltinID, 3494 const CallExpr *E) { 3495 if (auto Hint = GetValueForARMHint(BuiltinID)) 3496 return Hint; 3497 3498 if (BuiltinID == ARM::BI__emit) { 3499 bool IsThumb = getTarget().getTriple().getArch() == llvm::Triple::thumb; 3500 llvm::FunctionType *FTy = 3501 llvm::FunctionType::get(VoidTy, /*Variadic=*/false); 3502 3503 APSInt Value; 3504 if (!E->getArg(0)->EvaluateAsInt(Value, CGM.getContext())) 3505 llvm_unreachable("Sema will ensure that the parameter is constant"); 3506 3507 uint64_t ZExtValue = Value.zextOrTrunc(IsThumb ? 16 : 32).getZExtValue(); 3508 3509 llvm::InlineAsm *Emit = 3510 IsThumb ? InlineAsm::get(FTy, ".inst.n 0x" + utohexstr(ZExtValue), "", 3511 /*SideEffects=*/true) 3512 : InlineAsm::get(FTy, ".inst 0x" + utohexstr(ZExtValue), "", 3513 /*SideEffects=*/true); 3514 3515 return Builder.CreateCall(Emit); 3516 } 3517 3518 if (BuiltinID == ARM::BI__builtin_arm_dbg) { 3519 Value *Option = EmitScalarExpr(E->getArg(0)); 3520 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::arm_dbg), Option); 3521 } 3522 3523 if (BuiltinID == ARM::BI__builtin_arm_prefetch) { 3524 Value *Address = EmitScalarExpr(E->getArg(0)); 3525 Value *RW = EmitScalarExpr(E->getArg(1)); 3526 Value *IsData = EmitScalarExpr(E->getArg(2)); 3527 3528 // Locality is not supported on ARM target 3529 Value *Locality = llvm::ConstantInt::get(Int32Ty, 3); 3530 3531 Value *F = CGM.getIntrinsic(Intrinsic::prefetch); 3532 return Builder.CreateCall(F, {Address, RW, Locality, IsData}); 3533 } 3534 3535 if (BuiltinID == ARM::BI__builtin_arm_rbit) { 3536 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::arm_rbit), 3537 EmitScalarExpr(E->getArg(0)), 3538 "rbit"); 3539 } 3540 3541 if (BuiltinID == ARM::BI__clear_cache) { 3542 assert(E->getNumArgs() == 2 && "__clear_cache takes 2 arguments"); 3543 const FunctionDecl *FD = E->getDirectCallee(); 3544 Value *Ops[2]; 3545 for (unsigned i = 0; i < 2; i++) 3546 Ops[i] = EmitScalarExpr(E->getArg(i)); 3547 llvm::Type *Ty = CGM.getTypes().ConvertType(FD->getType()); 3548 llvm::FunctionType *FTy = cast<llvm::FunctionType>(Ty); 3549 StringRef Name = FD->getName(); 3550 return EmitNounwindRuntimeCall(CGM.CreateRuntimeFunction(FTy, Name), Ops); 3551 } 3552 3553 if (BuiltinID == ARM::BI__builtin_arm_ldrexd || 3554 ((BuiltinID == ARM::BI__builtin_arm_ldrex || 3555 BuiltinID == ARM::BI__builtin_arm_ldaex) && 3556 getContext().getTypeSize(E->getType()) == 64) || 3557 BuiltinID == ARM::BI__ldrexd) { 3558 Function *F; 3559 3560 switch (BuiltinID) { 3561 default: llvm_unreachable("unexpected builtin"); 3562 case ARM::BI__builtin_arm_ldaex: 3563 F = CGM.getIntrinsic(Intrinsic::arm_ldaexd); 3564 break; 3565 case ARM::BI__builtin_arm_ldrexd: 3566 case ARM::BI__builtin_arm_ldrex: 3567 case ARM::BI__ldrexd: 3568 F = CGM.getIntrinsic(Intrinsic::arm_ldrexd); 3569 break; 3570 } 3571 3572 Value *LdPtr = EmitScalarExpr(E->getArg(0)); 3573 Value *Val = Builder.CreateCall(F, Builder.CreateBitCast(LdPtr, Int8PtrTy), 3574 "ldrexd"); 3575 3576 Value *Val0 = Builder.CreateExtractValue(Val, 1); 3577 Value *Val1 = Builder.CreateExtractValue(Val, 0); 3578 Val0 = Builder.CreateZExt(Val0, Int64Ty); 3579 Val1 = Builder.CreateZExt(Val1, Int64Ty); 3580 3581 Value *ShiftCst = llvm::ConstantInt::get(Int64Ty, 32); 3582 Val = Builder.CreateShl(Val0, ShiftCst, "shl", true /* nuw */); 3583 Val = Builder.CreateOr(Val, Val1); 3584 return Builder.CreateBitCast(Val, ConvertType(E->getType())); 3585 } 3586 3587 if (BuiltinID == ARM::BI__builtin_arm_ldrex || 3588 BuiltinID == ARM::BI__builtin_arm_ldaex) { 3589 Value *LoadAddr = EmitScalarExpr(E->getArg(0)); 3590 3591 QualType Ty = E->getType(); 3592 llvm::Type *RealResTy = ConvertType(Ty); 3593 llvm::Type *IntResTy = llvm::IntegerType::get(getLLVMContext(), 3594 getContext().getTypeSize(Ty)); 3595 LoadAddr = Builder.CreateBitCast(LoadAddr, IntResTy->getPointerTo()); 3596 3597 Function *F = CGM.getIntrinsic(BuiltinID == ARM::BI__builtin_arm_ldaex 3598 ? Intrinsic::arm_ldaex 3599 : Intrinsic::arm_ldrex, 3600 LoadAddr->getType()); 3601 Value *Val = Builder.CreateCall(F, LoadAddr, "ldrex"); 3602 3603 if (RealResTy->isPointerTy()) 3604 return Builder.CreateIntToPtr(Val, RealResTy); 3605 else { 3606 Val = Builder.CreateTruncOrBitCast(Val, IntResTy); 3607 return Builder.CreateBitCast(Val, RealResTy); 3608 } 3609 } 3610 3611 if (BuiltinID == ARM::BI__builtin_arm_strexd || 3612 ((BuiltinID == ARM::BI__builtin_arm_stlex || 3613 BuiltinID == ARM::BI__builtin_arm_strex) && 3614 getContext().getTypeSize(E->getArg(0)->getType()) == 64)) { 3615 Function *F = CGM.getIntrinsic(BuiltinID == ARM::BI__builtin_arm_stlex 3616 ? Intrinsic::arm_stlexd 3617 : Intrinsic::arm_strexd); 3618 llvm::Type *STy = llvm::StructType::get(Int32Ty, Int32Ty, nullptr); 3619 3620 Address Tmp = CreateMemTemp(E->getArg(0)->getType()); 3621 Value *Val = EmitScalarExpr(E->getArg(0)); 3622 Builder.CreateStore(Val, Tmp); 3623 3624 Address LdPtr = Builder.CreateBitCast(Tmp,llvm::PointerType::getUnqual(STy)); 3625 Val = Builder.CreateLoad(LdPtr); 3626 3627 Value *Arg0 = Builder.CreateExtractValue(Val, 0); 3628 Value *Arg1 = Builder.CreateExtractValue(Val, 1); 3629 Value *StPtr = Builder.CreateBitCast(EmitScalarExpr(E->getArg(1)), Int8PtrTy); 3630 return Builder.CreateCall(F, {Arg0, Arg1, StPtr}, "strexd"); 3631 } 3632 3633 if (BuiltinID == ARM::BI__builtin_arm_strex || 3634 BuiltinID == ARM::BI__builtin_arm_stlex) { 3635 Value *StoreVal = EmitScalarExpr(E->getArg(0)); 3636 Value *StoreAddr = EmitScalarExpr(E->getArg(1)); 3637 3638 QualType Ty = E->getArg(0)->getType(); 3639 llvm::Type *StoreTy = llvm::IntegerType::get(getLLVMContext(), 3640 getContext().getTypeSize(Ty)); 3641 StoreAddr = Builder.CreateBitCast(StoreAddr, StoreTy->getPointerTo()); 3642 3643 if (StoreVal->getType()->isPointerTy()) 3644 StoreVal = Builder.CreatePtrToInt(StoreVal, Int32Ty); 3645 else { 3646 StoreVal = Builder.CreateBitCast(StoreVal, StoreTy); 3647 StoreVal = Builder.CreateZExtOrBitCast(StoreVal, Int32Ty); 3648 } 3649 3650 Function *F = CGM.getIntrinsic(BuiltinID == ARM::BI__builtin_arm_stlex 3651 ? Intrinsic::arm_stlex 3652 : Intrinsic::arm_strex, 3653 StoreAddr->getType()); 3654 return Builder.CreateCall(F, {StoreVal, StoreAddr}, "strex"); 3655 } 3656 3657 if (BuiltinID == ARM::BI__builtin_arm_clrex) { 3658 Function *F = CGM.getIntrinsic(Intrinsic::arm_clrex); 3659 return Builder.CreateCall(F); 3660 } 3661 3662 // CRC32 3663 Intrinsic::ID CRCIntrinsicID = Intrinsic::not_intrinsic; 3664 switch (BuiltinID) { 3665 case ARM::BI__builtin_arm_crc32b: 3666 CRCIntrinsicID = Intrinsic::arm_crc32b; break; 3667 case ARM::BI__builtin_arm_crc32cb: 3668 CRCIntrinsicID = Intrinsic::arm_crc32cb; break; 3669 case ARM::BI__builtin_arm_crc32h: 3670 CRCIntrinsicID = Intrinsic::arm_crc32h; break; 3671 case ARM::BI__builtin_arm_crc32ch: 3672 CRCIntrinsicID = Intrinsic::arm_crc32ch; break; 3673 case ARM::BI__builtin_arm_crc32w: 3674 case ARM::BI__builtin_arm_crc32d: 3675 CRCIntrinsicID = Intrinsic::arm_crc32w; break; 3676 case ARM::BI__builtin_arm_crc32cw: 3677 case ARM::BI__builtin_arm_crc32cd: 3678 CRCIntrinsicID = Intrinsic::arm_crc32cw; break; 3679 } 3680 3681 if (CRCIntrinsicID != Intrinsic::not_intrinsic) { 3682 Value *Arg0 = EmitScalarExpr(E->getArg(0)); 3683 Value *Arg1 = EmitScalarExpr(E->getArg(1)); 3684 3685 // crc32{c,}d intrinsics are implemnted as two calls to crc32{c,}w 3686 // intrinsics, hence we need different codegen for these cases. 3687 if (BuiltinID == ARM::BI__builtin_arm_crc32d || 3688 BuiltinID == ARM::BI__builtin_arm_crc32cd) { 3689 Value *C1 = llvm::ConstantInt::get(Int64Ty, 32); 3690 Value *Arg1a = Builder.CreateTruncOrBitCast(Arg1, Int32Ty); 3691 Value *Arg1b = Builder.CreateLShr(Arg1, C1); 3692 Arg1b = Builder.CreateTruncOrBitCast(Arg1b, Int32Ty); 3693 3694 Function *F = CGM.getIntrinsic(CRCIntrinsicID); 3695 Value *Res = Builder.CreateCall(F, {Arg0, Arg1a}); 3696 return Builder.CreateCall(F, {Res, Arg1b}); 3697 } else { 3698 Arg1 = Builder.CreateZExtOrBitCast(Arg1, Int32Ty); 3699 3700 Function *F = CGM.getIntrinsic(CRCIntrinsicID); 3701 return Builder.CreateCall(F, {Arg0, Arg1}); 3702 } 3703 } 3704 3705 if (BuiltinID == ARM::BI__builtin_arm_rsr || 3706 BuiltinID == ARM::BI__builtin_arm_rsr64 || 3707 BuiltinID == ARM::BI__builtin_arm_rsrp || 3708 BuiltinID == ARM::BI__builtin_arm_wsr || 3709 BuiltinID == ARM::BI__builtin_arm_wsr64 || 3710 BuiltinID == ARM::BI__builtin_arm_wsrp) { 3711 3712 bool IsRead = BuiltinID == ARM::BI__builtin_arm_rsr || 3713 BuiltinID == ARM::BI__builtin_arm_rsr64 || 3714 BuiltinID == ARM::BI__builtin_arm_rsrp; 3715 3716 bool IsPointerBuiltin = BuiltinID == ARM::BI__builtin_arm_rsrp || 3717 BuiltinID == ARM::BI__builtin_arm_wsrp; 3718 3719 bool Is64Bit = BuiltinID == ARM::BI__builtin_arm_rsr64 || 3720 BuiltinID == ARM::BI__builtin_arm_wsr64; 3721 3722 llvm::Type *ValueType; 3723 llvm::Type *RegisterType; 3724 if (IsPointerBuiltin) { 3725 ValueType = VoidPtrTy; 3726 RegisterType = Int32Ty; 3727 } else if (Is64Bit) { 3728 ValueType = RegisterType = Int64Ty; 3729 } else { 3730 ValueType = RegisterType = Int32Ty; 3731 } 3732 3733 return EmitSpecialRegisterBuiltin(*this, E, RegisterType, ValueType, IsRead); 3734 } 3735 3736 // Find out if any arguments are required to be integer constant 3737 // expressions. 3738 unsigned ICEArguments = 0; 3739 ASTContext::GetBuiltinTypeError Error; 3740 getContext().GetBuiltinType(BuiltinID, Error, &ICEArguments); 3741 assert(Error == ASTContext::GE_None && "Should not codegen an error"); 3742 3743 auto getAlignmentValue32 = [&](Address addr) -> Value* { 3744 return Builder.getInt32(addr.getAlignment().getQuantity()); 3745 }; 3746 3747 Address PtrOp0 = Address::invalid(); 3748 Address PtrOp1 = Address::invalid(); 3749 SmallVector<Value*, 4> Ops; 3750 bool HasExtraArg = HasExtraNeonArgument(BuiltinID); 3751 unsigned NumArgs = E->getNumArgs() - (HasExtraArg ? 1 : 0); 3752 for (unsigned i = 0, e = NumArgs; i != e; i++) { 3753 if (i == 0) { 3754 switch (BuiltinID) { 3755 case NEON::BI__builtin_neon_vld1_v: 3756 case NEON::BI__builtin_neon_vld1q_v: 3757 case NEON::BI__builtin_neon_vld1q_lane_v: 3758 case NEON::BI__builtin_neon_vld1_lane_v: 3759 case NEON::BI__builtin_neon_vld1_dup_v: 3760 case NEON::BI__builtin_neon_vld1q_dup_v: 3761 case NEON::BI__builtin_neon_vst1_v: 3762 case NEON::BI__builtin_neon_vst1q_v: 3763 case NEON::BI__builtin_neon_vst1q_lane_v: 3764 case NEON::BI__builtin_neon_vst1_lane_v: 3765 case NEON::BI__builtin_neon_vst2_v: 3766 case NEON::BI__builtin_neon_vst2q_v: 3767 case NEON::BI__builtin_neon_vst2_lane_v: 3768 case NEON::BI__builtin_neon_vst2q_lane_v: 3769 case NEON::BI__builtin_neon_vst3_v: 3770 case NEON::BI__builtin_neon_vst3q_v: 3771 case NEON::BI__builtin_neon_vst3_lane_v: 3772 case NEON::BI__builtin_neon_vst3q_lane_v: 3773 case NEON::BI__builtin_neon_vst4_v: 3774 case NEON::BI__builtin_neon_vst4q_v: 3775 case NEON::BI__builtin_neon_vst4_lane_v: 3776 case NEON::BI__builtin_neon_vst4q_lane_v: 3777 // Get the alignment for the argument in addition to the value; 3778 // we'll use it later. 3779 PtrOp0 = EmitPointerWithAlignment(E->getArg(0)); 3780 Ops.push_back(PtrOp0.getPointer()); 3781 continue; 3782 } 3783 } 3784 if (i == 1) { 3785 switch (BuiltinID) { 3786 case NEON::BI__builtin_neon_vld2_v: 3787 case NEON::BI__builtin_neon_vld2q_v: 3788 case NEON::BI__builtin_neon_vld3_v: 3789 case NEON::BI__builtin_neon_vld3q_v: 3790 case NEON::BI__builtin_neon_vld4_v: 3791 case NEON::BI__builtin_neon_vld4q_v: 3792 case NEON::BI__builtin_neon_vld2_lane_v: 3793 case NEON::BI__builtin_neon_vld2q_lane_v: 3794 case NEON::BI__builtin_neon_vld3_lane_v: 3795 case NEON::BI__builtin_neon_vld3q_lane_v: 3796 case NEON::BI__builtin_neon_vld4_lane_v: 3797 case NEON::BI__builtin_neon_vld4q_lane_v: 3798 case NEON::BI__builtin_neon_vld2_dup_v: 3799 case NEON::BI__builtin_neon_vld3_dup_v: 3800 case NEON::BI__builtin_neon_vld4_dup_v: 3801 // Get the alignment for the argument in addition to the value; 3802 // we'll use it later. 3803 PtrOp1 = EmitPointerWithAlignment(E->getArg(1)); 3804 Ops.push_back(PtrOp1.getPointer()); 3805 continue; 3806 } 3807 } 3808 3809 if ((ICEArguments & (1 << i)) == 0) { 3810 Ops.push_back(EmitScalarExpr(E->getArg(i))); 3811 } else { 3812 // If this is required to be a constant, constant fold it so that we know 3813 // that the generated intrinsic gets a ConstantInt. 3814 llvm::APSInt Result; 3815 bool IsConst = E->getArg(i)->isIntegerConstantExpr(Result, getContext()); 3816 assert(IsConst && "Constant arg isn't actually constant?"); (void)IsConst; 3817 Ops.push_back(llvm::ConstantInt::get(getLLVMContext(), Result)); 3818 } 3819 } 3820 3821 switch (BuiltinID) { 3822 default: break; 3823 3824 case NEON::BI__builtin_neon_vget_lane_i8: 3825 case NEON::BI__builtin_neon_vget_lane_i16: 3826 case NEON::BI__builtin_neon_vget_lane_i32: 3827 case NEON::BI__builtin_neon_vget_lane_i64: 3828 case NEON::BI__builtin_neon_vget_lane_f32: 3829 case NEON::BI__builtin_neon_vgetq_lane_i8: 3830 case NEON::BI__builtin_neon_vgetq_lane_i16: 3831 case NEON::BI__builtin_neon_vgetq_lane_i32: 3832 case NEON::BI__builtin_neon_vgetq_lane_i64: 3833 case NEON::BI__builtin_neon_vgetq_lane_f32: 3834 return Builder.CreateExtractElement(Ops[0], Ops[1], "vget_lane"); 3835 3836 case NEON::BI__builtin_neon_vset_lane_i8: 3837 case NEON::BI__builtin_neon_vset_lane_i16: 3838 case NEON::BI__builtin_neon_vset_lane_i32: 3839 case NEON::BI__builtin_neon_vset_lane_i64: 3840 case NEON::BI__builtin_neon_vset_lane_f32: 3841 case NEON::BI__builtin_neon_vsetq_lane_i8: 3842 case NEON::BI__builtin_neon_vsetq_lane_i16: 3843 case NEON::BI__builtin_neon_vsetq_lane_i32: 3844 case NEON::BI__builtin_neon_vsetq_lane_i64: 3845 case NEON::BI__builtin_neon_vsetq_lane_f32: 3846 return Builder.CreateInsertElement(Ops[1], Ops[0], Ops[2], "vset_lane"); 3847 3848 case NEON::BI__builtin_neon_vsha1h_u32: 3849 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_sha1h), Ops, 3850 "vsha1h"); 3851 case NEON::BI__builtin_neon_vsha1cq_u32: 3852 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_sha1c), Ops, 3853 "vsha1h"); 3854 case NEON::BI__builtin_neon_vsha1pq_u32: 3855 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_sha1p), Ops, 3856 "vsha1h"); 3857 case NEON::BI__builtin_neon_vsha1mq_u32: 3858 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_sha1m), Ops, 3859 "vsha1h"); 3860 3861 // The ARM _MoveToCoprocessor builtins put the input register value as 3862 // the first argument, but the LLVM intrinsic expects it as the third one. 3863 case ARM::BI_MoveToCoprocessor: 3864 case ARM::BI_MoveToCoprocessor2: { 3865 Function *F = CGM.getIntrinsic(BuiltinID == ARM::BI_MoveToCoprocessor ? 3866 Intrinsic::arm_mcr : Intrinsic::arm_mcr2); 3867 return Builder.CreateCall(F, {Ops[1], Ops[2], Ops[0], 3868 Ops[3], Ops[4], Ops[5]}); 3869 } 3870 } 3871 3872 // Get the last argument, which specifies the vector type. 3873 assert(HasExtraArg); 3874 llvm::APSInt Result; 3875 const Expr *Arg = E->getArg(E->getNumArgs()-1); 3876 if (!Arg->isIntegerConstantExpr(Result, getContext())) 3877 return nullptr; 3878 3879 if (BuiltinID == ARM::BI__builtin_arm_vcvtr_f || 3880 BuiltinID == ARM::BI__builtin_arm_vcvtr_d) { 3881 // Determine the overloaded type of this builtin. 3882 llvm::Type *Ty; 3883 if (BuiltinID == ARM::BI__builtin_arm_vcvtr_f) 3884 Ty = FloatTy; 3885 else 3886 Ty = DoubleTy; 3887 3888 // Determine whether this is an unsigned conversion or not. 3889 bool usgn = Result.getZExtValue() == 1; 3890 unsigned Int = usgn ? Intrinsic::arm_vcvtru : Intrinsic::arm_vcvtr; 3891 3892 // Call the appropriate intrinsic. 3893 Function *F = CGM.getIntrinsic(Int, Ty); 3894 return Builder.CreateCall(F, Ops, "vcvtr"); 3895 } 3896 3897 // Determine the type of this overloaded NEON intrinsic. 3898 NeonTypeFlags Type(Result.getZExtValue()); 3899 bool usgn = Type.isUnsigned(); 3900 bool rightShift = false; 3901 3902 llvm::VectorType *VTy = GetNeonType(this, Type); 3903 llvm::Type *Ty = VTy; 3904 if (!Ty) 3905 return nullptr; 3906 3907 // Many NEON builtins have identical semantics and uses in ARM and 3908 // AArch64. Emit these in a single function. 3909 auto IntrinsicMap = makeArrayRef(ARMSIMDIntrinsicMap); 3910 const NeonIntrinsicInfo *Builtin = findNeonIntrinsicInMap( 3911 IntrinsicMap, BuiltinID, NEONSIMDIntrinsicsProvenSorted); 3912 if (Builtin) 3913 return EmitCommonNeonBuiltinExpr( 3914 Builtin->BuiltinID, Builtin->LLVMIntrinsic, Builtin->AltLLVMIntrinsic, 3915 Builtin->NameHint, Builtin->TypeModifier, E, Ops, PtrOp0, PtrOp1); 3916 3917 unsigned Int; 3918 switch (BuiltinID) { 3919 default: return nullptr; 3920 case NEON::BI__builtin_neon_vld1q_lane_v: 3921 // Handle 64-bit integer elements as a special case. Use shuffles of 3922 // one-element vectors to avoid poor code for i64 in the backend. 3923 if (VTy->getElementType()->isIntegerTy(64)) { 3924 // Extract the other lane. 3925 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 3926 uint32_t Lane = cast<ConstantInt>(Ops[2])->getZExtValue(); 3927 Value *SV = llvm::ConstantVector::get(ConstantInt::get(Int32Ty, 1-Lane)); 3928 Ops[1] = Builder.CreateShuffleVector(Ops[1], Ops[1], SV); 3929 // Load the value as a one-element vector. 3930 Ty = llvm::VectorType::get(VTy->getElementType(), 1); 3931 llvm::Type *Tys[] = {Ty, Int8PtrTy}; 3932 Function *F = CGM.getIntrinsic(Intrinsic::arm_neon_vld1, Tys); 3933 Value *Align = getAlignmentValue32(PtrOp0); 3934 Value *Ld = Builder.CreateCall(F, {Ops[0], Align}); 3935 // Combine them. 3936 uint32_t Indices[] = {1 - Lane, Lane}; 3937 SV = llvm::ConstantDataVector::get(getLLVMContext(), Indices); 3938 return Builder.CreateShuffleVector(Ops[1], Ld, SV, "vld1q_lane"); 3939 } 3940 // fall through 3941 case NEON::BI__builtin_neon_vld1_lane_v: { 3942 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 3943 PtrOp0 = Builder.CreateElementBitCast(PtrOp0, VTy->getElementType()); 3944 Value *Ld = Builder.CreateLoad(PtrOp0); 3945 return Builder.CreateInsertElement(Ops[1], Ld, Ops[2], "vld1_lane"); 3946 } 3947 case NEON::BI__builtin_neon_vld2_dup_v: 3948 case NEON::BI__builtin_neon_vld3_dup_v: 3949 case NEON::BI__builtin_neon_vld4_dup_v: { 3950 // Handle 64-bit elements as a special-case. There is no "dup" needed. 3951 if (VTy->getElementType()->getPrimitiveSizeInBits() == 64) { 3952 switch (BuiltinID) { 3953 case NEON::BI__builtin_neon_vld2_dup_v: 3954 Int = Intrinsic::arm_neon_vld2; 3955 break; 3956 case NEON::BI__builtin_neon_vld3_dup_v: 3957 Int = Intrinsic::arm_neon_vld3; 3958 break; 3959 case NEON::BI__builtin_neon_vld4_dup_v: 3960 Int = Intrinsic::arm_neon_vld4; 3961 break; 3962 default: llvm_unreachable("unknown vld_dup intrinsic?"); 3963 } 3964 llvm::Type *Tys[] = {Ty, Int8PtrTy}; 3965 Function *F = CGM.getIntrinsic(Int, Tys); 3966 llvm::Value *Align = getAlignmentValue32(PtrOp1); 3967 Ops[1] = Builder.CreateCall(F, {Ops[1], Align}, "vld_dup"); 3968 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 3969 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 3970 return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]); 3971 } 3972 switch (BuiltinID) { 3973 case NEON::BI__builtin_neon_vld2_dup_v: 3974 Int = Intrinsic::arm_neon_vld2lane; 3975 break; 3976 case NEON::BI__builtin_neon_vld3_dup_v: 3977 Int = Intrinsic::arm_neon_vld3lane; 3978 break; 3979 case NEON::BI__builtin_neon_vld4_dup_v: 3980 Int = Intrinsic::arm_neon_vld4lane; 3981 break; 3982 default: llvm_unreachable("unknown vld_dup intrinsic?"); 3983 } 3984 llvm::Type *Tys[] = {Ty, Int8PtrTy}; 3985 Function *F = CGM.getIntrinsic(Int, Tys); 3986 llvm::StructType *STy = cast<llvm::StructType>(F->getReturnType()); 3987 3988 SmallVector<Value*, 6> Args; 3989 Args.push_back(Ops[1]); 3990 Args.append(STy->getNumElements(), UndefValue::get(Ty)); 3991 3992 llvm::Constant *CI = ConstantInt::get(Int32Ty, 0); 3993 Args.push_back(CI); 3994 Args.push_back(getAlignmentValue32(PtrOp1)); 3995 3996 Ops[1] = Builder.CreateCall(F, Args, "vld_dup"); 3997 // splat lane 0 to all elts in each vector of the result. 3998 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 3999 Value *Val = Builder.CreateExtractValue(Ops[1], i); 4000 Value *Elt = Builder.CreateBitCast(Val, Ty); 4001 Elt = EmitNeonSplat(Elt, CI); 4002 Elt = Builder.CreateBitCast(Elt, Val->getType()); 4003 Ops[1] = Builder.CreateInsertValue(Ops[1], Elt, i); 4004 } 4005 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 4006 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 4007 return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]); 4008 } 4009 case NEON::BI__builtin_neon_vqrshrn_n_v: 4010 Int = 4011 usgn ? Intrinsic::arm_neon_vqrshiftnu : Intrinsic::arm_neon_vqrshiftns; 4012 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqrshrn_n", 4013 1, true); 4014 case NEON::BI__builtin_neon_vqrshrun_n_v: 4015 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vqrshiftnsu, Ty), 4016 Ops, "vqrshrun_n", 1, true); 4017 case NEON::BI__builtin_neon_vqshrn_n_v: 4018 Int = usgn ? Intrinsic::arm_neon_vqshiftnu : Intrinsic::arm_neon_vqshiftns; 4019 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqshrn_n", 4020 1, true); 4021 case NEON::BI__builtin_neon_vqshrun_n_v: 4022 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vqshiftnsu, Ty), 4023 Ops, "vqshrun_n", 1, true); 4024 case NEON::BI__builtin_neon_vrecpe_v: 4025 case NEON::BI__builtin_neon_vrecpeq_v: 4026 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vrecpe, Ty), 4027 Ops, "vrecpe"); 4028 case NEON::BI__builtin_neon_vrshrn_n_v: 4029 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vrshiftn, Ty), 4030 Ops, "vrshrn_n", 1, true); 4031 case NEON::BI__builtin_neon_vrsra_n_v: 4032 case NEON::BI__builtin_neon_vrsraq_n_v: 4033 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 4034 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 4035 Ops[2] = EmitNeonShiftVector(Ops[2], Ty, true); 4036 Int = usgn ? Intrinsic::arm_neon_vrshiftu : Intrinsic::arm_neon_vrshifts; 4037 Ops[1] = Builder.CreateCall(CGM.getIntrinsic(Int, Ty), {Ops[1], Ops[2]}); 4038 return Builder.CreateAdd(Ops[0], Ops[1], "vrsra_n"); 4039 case NEON::BI__builtin_neon_vsri_n_v: 4040 case NEON::BI__builtin_neon_vsriq_n_v: 4041 rightShift = true; 4042 case NEON::BI__builtin_neon_vsli_n_v: 4043 case NEON::BI__builtin_neon_vsliq_n_v: 4044 Ops[2] = EmitNeonShiftVector(Ops[2], Ty, rightShift); 4045 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vshiftins, Ty), 4046 Ops, "vsli_n"); 4047 case NEON::BI__builtin_neon_vsra_n_v: 4048 case NEON::BI__builtin_neon_vsraq_n_v: 4049 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 4050 Ops[1] = EmitNeonRShiftImm(Ops[1], Ops[2], Ty, usgn, "vsra_n"); 4051 return Builder.CreateAdd(Ops[0], Ops[1]); 4052 case NEON::BI__builtin_neon_vst1q_lane_v: 4053 // Handle 64-bit integer elements as a special case. Use a shuffle to get 4054 // a one-element vector and avoid poor code for i64 in the backend. 4055 if (VTy->getElementType()->isIntegerTy(64)) { 4056 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 4057 Value *SV = llvm::ConstantVector::get(cast<llvm::Constant>(Ops[2])); 4058 Ops[1] = Builder.CreateShuffleVector(Ops[1], Ops[1], SV); 4059 Ops[2] = getAlignmentValue32(PtrOp0); 4060 llvm::Type *Tys[] = {Int8PtrTy, Ops[1]->getType()}; 4061 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::arm_neon_vst1, 4062 Tys), Ops); 4063 } 4064 // fall through 4065 case NEON::BI__builtin_neon_vst1_lane_v: { 4066 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 4067 Ops[1] = Builder.CreateExtractElement(Ops[1], Ops[2]); 4068 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 4069 auto St = Builder.CreateStore(Ops[1], Builder.CreateBitCast(PtrOp0, Ty)); 4070 return St; 4071 } 4072 case NEON::BI__builtin_neon_vtbl1_v: 4073 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbl1), 4074 Ops, "vtbl1"); 4075 case NEON::BI__builtin_neon_vtbl2_v: 4076 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbl2), 4077 Ops, "vtbl2"); 4078 case NEON::BI__builtin_neon_vtbl3_v: 4079 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbl3), 4080 Ops, "vtbl3"); 4081 case NEON::BI__builtin_neon_vtbl4_v: 4082 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbl4), 4083 Ops, "vtbl4"); 4084 case NEON::BI__builtin_neon_vtbx1_v: 4085 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbx1), 4086 Ops, "vtbx1"); 4087 case NEON::BI__builtin_neon_vtbx2_v: 4088 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbx2), 4089 Ops, "vtbx2"); 4090 case NEON::BI__builtin_neon_vtbx3_v: 4091 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbx3), 4092 Ops, "vtbx3"); 4093 case NEON::BI__builtin_neon_vtbx4_v: 4094 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbx4), 4095 Ops, "vtbx4"); 4096 } 4097 } 4098 4099 static Value *EmitAArch64TblBuiltinExpr(CodeGenFunction &CGF, unsigned BuiltinID, 4100 const CallExpr *E, 4101 SmallVectorImpl<Value *> &Ops) { 4102 unsigned int Int = 0; 4103 const char *s = nullptr; 4104 4105 switch (BuiltinID) { 4106 default: 4107 return nullptr; 4108 case NEON::BI__builtin_neon_vtbl1_v: 4109 case NEON::BI__builtin_neon_vqtbl1_v: 4110 case NEON::BI__builtin_neon_vqtbl1q_v: 4111 case NEON::BI__builtin_neon_vtbl2_v: 4112 case NEON::BI__builtin_neon_vqtbl2_v: 4113 case NEON::BI__builtin_neon_vqtbl2q_v: 4114 case NEON::BI__builtin_neon_vtbl3_v: 4115 case NEON::BI__builtin_neon_vqtbl3_v: 4116 case NEON::BI__builtin_neon_vqtbl3q_v: 4117 case NEON::BI__builtin_neon_vtbl4_v: 4118 case NEON::BI__builtin_neon_vqtbl4_v: 4119 case NEON::BI__builtin_neon_vqtbl4q_v: 4120 break; 4121 case NEON::BI__builtin_neon_vtbx1_v: 4122 case NEON::BI__builtin_neon_vqtbx1_v: 4123 case NEON::BI__builtin_neon_vqtbx1q_v: 4124 case NEON::BI__builtin_neon_vtbx2_v: 4125 case NEON::BI__builtin_neon_vqtbx2_v: 4126 case NEON::BI__builtin_neon_vqtbx2q_v: 4127 case NEON::BI__builtin_neon_vtbx3_v: 4128 case NEON::BI__builtin_neon_vqtbx3_v: 4129 case NEON::BI__builtin_neon_vqtbx3q_v: 4130 case NEON::BI__builtin_neon_vtbx4_v: 4131 case NEON::BI__builtin_neon_vqtbx4_v: 4132 case NEON::BI__builtin_neon_vqtbx4q_v: 4133 break; 4134 } 4135 4136 assert(E->getNumArgs() >= 3); 4137 4138 // Get the last argument, which specifies the vector type. 4139 llvm::APSInt Result; 4140 const Expr *Arg = E->getArg(E->getNumArgs() - 1); 4141 if (!Arg->isIntegerConstantExpr(Result, CGF.getContext())) 4142 return nullptr; 4143 4144 // Determine the type of this overloaded NEON intrinsic. 4145 NeonTypeFlags Type(Result.getZExtValue()); 4146 llvm::VectorType *Ty = GetNeonType(&CGF, Type); 4147 if (!Ty) 4148 return nullptr; 4149 4150 CodeGen::CGBuilderTy &Builder = CGF.Builder; 4151 4152 // AArch64 scalar builtins are not overloaded, they do not have an extra 4153 // argument that specifies the vector type, need to handle each case. 4154 switch (BuiltinID) { 4155 case NEON::BI__builtin_neon_vtbl1_v: { 4156 return packTBLDVectorList(CGF, makeArrayRef(Ops).slice(0, 1), nullptr, 4157 Ops[1], Ty, Intrinsic::aarch64_neon_tbl1, 4158 "vtbl1"); 4159 } 4160 case NEON::BI__builtin_neon_vtbl2_v: { 4161 return packTBLDVectorList(CGF, makeArrayRef(Ops).slice(0, 2), nullptr, 4162 Ops[2], Ty, Intrinsic::aarch64_neon_tbl1, 4163 "vtbl1"); 4164 } 4165 case NEON::BI__builtin_neon_vtbl3_v: { 4166 return packTBLDVectorList(CGF, makeArrayRef(Ops).slice(0, 3), nullptr, 4167 Ops[3], Ty, Intrinsic::aarch64_neon_tbl2, 4168 "vtbl2"); 4169 } 4170 case NEON::BI__builtin_neon_vtbl4_v: { 4171 return packTBLDVectorList(CGF, makeArrayRef(Ops).slice(0, 4), nullptr, 4172 Ops[4], Ty, Intrinsic::aarch64_neon_tbl2, 4173 "vtbl2"); 4174 } 4175 case NEON::BI__builtin_neon_vtbx1_v: { 4176 Value *TblRes = 4177 packTBLDVectorList(CGF, makeArrayRef(Ops).slice(1, 1), nullptr, Ops[2], 4178 Ty, Intrinsic::aarch64_neon_tbl1, "vtbl1"); 4179 4180 llvm::Constant *EightV = ConstantInt::get(Ty, 8); 4181 Value *CmpRes = Builder.CreateICmp(ICmpInst::ICMP_UGE, Ops[2], EightV); 4182 CmpRes = Builder.CreateSExt(CmpRes, Ty); 4183 4184 Value *EltsFromInput = Builder.CreateAnd(CmpRes, Ops[0]); 4185 Value *EltsFromTbl = Builder.CreateAnd(Builder.CreateNot(CmpRes), TblRes); 4186 return Builder.CreateOr(EltsFromInput, EltsFromTbl, "vtbx"); 4187 } 4188 case NEON::BI__builtin_neon_vtbx2_v: { 4189 return packTBLDVectorList(CGF, makeArrayRef(Ops).slice(1, 2), Ops[0], 4190 Ops[3], Ty, Intrinsic::aarch64_neon_tbx1, 4191 "vtbx1"); 4192 } 4193 case NEON::BI__builtin_neon_vtbx3_v: { 4194 Value *TblRes = 4195 packTBLDVectorList(CGF, makeArrayRef(Ops).slice(1, 3), nullptr, Ops[4], 4196 Ty, Intrinsic::aarch64_neon_tbl2, "vtbl2"); 4197 4198 llvm::Constant *TwentyFourV = ConstantInt::get(Ty, 24); 4199 Value *CmpRes = Builder.CreateICmp(ICmpInst::ICMP_UGE, Ops[4], 4200 TwentyFourV); 4201 CmpRes = Builder.CreateSExt(CmpRes, Ty); 4202 4203 Value *EltsFromInput = Builder.CreateAnd(CmpRes, Ops[0]); 4204 Value *EltsFromTbl = Builder.CreateAnd(Builder.CreateNot(CmpRes), TblRes); 4205 return Builder.CreateOr(EltsFromInput, EltsFromTbl, "vtbx"); 4206 } 4207 case NEON::BI__builtin_neon_vtbx4_v: { 4208 return packTBLDVectorList(CGF, makeArrayRef(Ops).slice(1, 4), Ops[0], 4209 Ops[5], Ty, Intrinsic::aarch64_neon_tbx2, 4210 "vtbx2"); 4211 } 4212 case NEON::BI__builtin_neon_vqtbl1_v: 4213 case NEON::BI__builtin_neon_vqtbl1q_v: 4214 Int = Intrinsic::aarch64_neon_tbl1; s = "vtbl1"; break; 4215 case NEON::BI__builtin_neon_vqtbl2_v: 4216 case NEON::BI__builtin_neon_vqtbl2q_v: { 4217 Int = Intrinsic::aarch64_neon_tbl2; s = "vtbl2"; break; 4218 case NEON::BI__builtin_neon_vqtbl3_v: 4219 case NEON::BI__builtin_neon_vqtbl3q_v: 4220 Int = Intrinsic::aarch64_neon_tbl3; s = "vtbl3"; break; 4221 case NEON::BI__builtin_neon_vqtbl4_v: 4222 case NEON::BI__builtin_neon_vqtbl4q_v: 4223 Int = Intrinsic::aarch64_neon_tbl4; s = "vtbl4"; break; 4224 case NEON::BI__builtin_neon_vqtbx1_v: 4225 case NEON::BI__builtin_neon_vqtbx1q_v: 4226 Int = Intrinsic::aarch64_neon_tbx1; s = "vtbx1"; break; 4227 case NEON::BI__builtin_neon_vqtbx2_v: 4228 case NEON::BI__builtin_neon_vqtbx2q_v: 4229 Int = Intrinsic::aarch64_neon_tbx2; s = "vtbx2"; break; 4230 case NEON::BI__builtin_neon_vqtbx3_v: 4231 case NEON::BI__builtin_neon_vqtbx3q_v: 4232 Int = Intrinsic::aarch64_neon_tbx3; s = "vtbx3"; break; 4233 case NEON::BI__builtin_neon_vqtbx4_v: 4234 case NEON::BI__builtin_neon_vqtbx4q_v: 4235 Int = Intrinsic::aarch64_neon_tbx4; s = "vtbx4"; break; 4236 } 4237 } 4238 4239 if (!Int) 4240 return nullptr; 4241 4242 Function *F = CGF.CGM.getIntrinsic(Int, Ty); 4243 return CGF.EmitNeonCall(F, Ops, s); 4244 } 4245 4246 Value *CodeGenFunction::vectorWrapScalar16(Value *Op) { 4247 llvm::Type *VTy = llvm::VectorType::get(Int16Ty, 4); 4248 Op = Builder.CreateBitCast(Op, Int16Ty); 4249 Value *V = UndefValue::get(VTy); 4250 llvm::Constant *CI = ConstantInt::get(SizeTy, 0); 4251 Op = Builder.CreateInsertElement(V, Op, CI); 4252 return Op; 4253 } 4254 4255 Value *CodeGenFunction::EmitAArch64BuiltinExpr(unsigned BuiltinID, 4256 const CallExpr *E) { 4257 unsigned HintID = static_cast<unsigned>(-1); 4258 switch (BuiltinID) { 4259 default: break; 4260 case AArch64::BI__builtin_arm_nop: 4261 HintID = 0; 4262 break; 4263 case AArch64::BI__builtin_arm_yield: 4264 HintID = 1; 4265 break; 4266 case AArch64::BI__builtin_arm_wfe: 4267 HintID = 2; 4268 break; 4269 case AArch64::BI__builtin_arm_wfi: 4270 HintID = 3; 4271 break; 4272 case AArch64::BI__builtin_arm_sev: 4273 HintID = 4; 4274 break; 4275 case AArch64::BI__builtin_arm_sevl: 4276 HintID = 5; 4277 break; 4278 } 4279 4280 if (HintID != static_cast<unsigned>(-1)) { 4281 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_hint); 4282 return Builder.CreateCall(F, llvm::ConstantInt::get(Int32Ty, HintID)); 4283 } 4284 4285 if (BuiltinID == AArch64::BI__builtin_arm_prefetch) { 4286 Value *Address = EmitScalarExpr(E->getArg(0)); 4287 Value *RW = EmitScalarExpr(E->getArg(1)); 4288 Value *CacheLevel = EmitScalarExpr(E->getArg(2)); 4289 Value *RetentionPolicy = EmitScalarExpr(E->getArg(3)); 4290 Value *IsData = EmitScalarExpr(E->getArg(4)); 4291 4292 Value *Locality = nullptr; 4293 if (cast<llvm::ConstantInt>(RetentionPolicy)->isZero()) { 4294 // Temporal fetch, needs to convert cache level to locality. 4295 Locality = llvm::ConstantInt::get(Int32Ty, 4296 -cast<llvm::ConstantInt>(CacheLevel)->getValue() + 3); 4297 } else { 4298 // Streaming fetch. 4299 Locality = llvm::ConstantInt::get(Int32Ty, 0); 4300 } 4301 4302 // FIXME: We need AArch64 specific LLVM intrinsic if we want to specify 4303 // PLDL3STRM or PLDL2STRM. 4304 Value *F = CGM.getIntrinsic(Intrinsic::prefetch); 4305 return Builder.CreateCall(F, {Address, RW, Locality, IsData}); 4306 } 4307 4308 if (BuiltinID == AArch64::BI__builtin_arm_rbit) { 4309 assert((getContext().getTypeSize(E->getType()) == 32) && 4310 "rbit of unusual size!"); 4311 llvm::Value *Arg = EmitScalarExpr(E->getArg(0)); 4312 return Builder.CreateCall( 4313 CGM.getIntrinsic(Intrinsic::aarch64_rbit, Arg->getType()), Arg, "rbit"); 4314 } 4315 if (BuiltinID == AArch64::BI__builtin_arm_rbit64) { 4316 assert((getContext().getTypeSize(E->getType()) == 64) && 4317 "rbit of unusual size!"); 4318 llvm::Value *Arg = EmitScalarExpr(E->getArg(0)); 4319 return Builder.CreateCall( 4320 CGM.getIntrinsic(Intrinsic::aarch64_rbit, Arg->getType()), Arg, "rbit"); 4321 } 4322 4323 if (BuiltinID == AArch64::BI__clear_cache) { 4324 assert(E->getNumArgs() == 2 && "__clear_cache takes 2 arguments"); 4325 const FunctionDecl *FD = E->getDirectCallee(); 4326 Value *Ops[2]; 4327 for (unsigned i = 0; i < 2; i++) 4328 Ops[i] = EmitScalarExpr(E->getArg(i)); 4329 llvm::Type *Ty = CGM.getTypes().ConvertType(FD->getType()); 4330 llvm::FunctionType *FTy = cast<llvm::FunctionType>(Ty); 4331 StringRef Name = FD->getName(); 4332 return EmitNounwindRuntimeCall(CGM.CreateRuntimeFunction(FTy, Name), Ops); 4333 } 4334 4335 if ((BuiltinID == AArch64::BI__builtin_arm_ldrex || 4336 BuiltinID == AArch64::BI__builtin_arm_ldaex) && 4337 getContext().getTypeSize(E->getType()) == 128) { 4338 Function *F = CGM.getIntrinsic(BuiltinID == AArch64::BI__builtin_arm_ldaex 4339 ? Intrinsic::aarch64_ldaxp 4340 : Intrinsic::aarch64_ldxp); 4341 4342 Value *LdPtr = EmitScalarExpr(E->getArg(0)); 4343 Value *Val = Builder.CreateCall(F, Builder.CreateBitCast(LdPtr, Int8PtrTy), 4344 "ldxp"); 4345 4346 Value *Val0 = Builder.CreateExtractValue(Val, 1); 4347 Value *Val1 = Builder.CreateExtractValue(Val, 0); 4348 llvm::Type *Int128Ty = llvm::IntegerType::get(getLLVMContext(), 128); 4349 Val0 = Builder.CreateZExt(Val0, Int128Ty); 4350 Val1 = Builder.CreateZExt(Val1, Int128Ty); 4351 4352 Value *ShiftCst = llvm::ConstantInt::get(Int128Ty, 64); 4353 Val = Builder.CreateShl(Val0, ShiftCst, "shl", true /* nuw */); 4354 Val = Builder.CreateOr(Val, Val1); 4355 return Builder.CreateBitCast(Val, ConvertType(E->getType())); 4356 } else if (BuiltinID == AArch64::BI__builtin_arm_ldrex || 4357 BuiltinID == AArch64::BI__builtin_arm_ldaex) { 4358 Value *LoadAddr = EmitScalarExpr(E->getArg(0)); 4359 4360 QualType Ty = E->getType(); 4361 llvm::Type *RealResTy = ConvertType(Ty); 4362 llvm::Type *IntResTy = llvm::IntegerType::get(getLLVMContext(), 4363 getContext().getTypeSize(Ty)); 4364 LoadAddr = Builder.CreateBitCast(LoadAddr, IntResTy->getPointerTo()); 4365 4366 Function *F = CGM.getIntrinsic(BuiltinID == AArch64::BI__builtin_arm_ldaex 4367 ? Intrinsic::aarch64_ldaxr 4368 : Intrinsic::aarch64_ldxr, 4369 LoadAddr->getType()); 4370 Value *Val = Builder.CreateCall(F, LoadAddr, "ldxr"); 4371 4372 if (RealResTy->isPointerTy()) 4373 return Builder.CreateIntToPtr(Val, RealResTy); 4374 4375 Val = Builder.CreateTruncOrBitCast(Val, IntResTy); 4376 return Builder.CreateBitCast(Val, RealResTy); 4377 } 4378 4379 if ((BuiltinID == AArch64::BI__builtin_arm_strex || 4380 BuiltinID == AArch64::BI__builtin_arm_stlex) && 4381 getContext().getTypeSize(E->getArg(0)->getType()) == 128) { 4382 Function *F = CGM.getIntrinsic(BuiltinID == AArch64::BI__builtin_arm_stlex 4383 ? Intrinsic::aarch64_stlxp 4384 : Intrinsic::aarch64_stxp); 4385 llvm::Type *STy = llvm::StructType::get(Int64Ty, Int64Ty, nullptr); 4386 4387 Address Tmp = CreateMemTemp(E->getArg(0)->getType()); 4388 EmitAnyExprToMem(E->getArg(0), Tmp, Qualifiers(), /*init*/ true); 4389 4390 Tmp = Builder.CreateBitCast(Tmp, llvm::PointerType::getUnqual(STy)); 4391 llvm::Value *Val = Builder.CreateLoad(Tmp); 4392 4393 Value *Arg0 = Builder.CreateExtractValue(Val, 0); 4394 Value *Arg1 = Builder.CreateExtractValue(Val, 1); 4395 Value *StPtr = Builder.CreateBitCast(EmitScalarExpr(E->getArg(1)), 4396 Int8PtrTy); 4397 return Builder.CreateCall(F, {Arg0, Arg1, StPtr}, "stxp"); 4398 } 4399 4400 if (BuiltinID == AArch64::BI__builtin_arm_strex || 4401 BuiltinID == AArch64::BI__builtin_arm_stlex) { 4402 Value *StoreVal = EmitScalarExpr(E->getArg(0)); 4403 Value *StoreAddr = EmitScalarExpr(E->getArg(1)); 4404 4405 QualType Ty = E->getArg(0)->getType(); 4406 llvm::Type *StoreTy = llvm::IntegerType::get(getLLVMContext(), 4407 getContext().getTypeSize(Ty)); 4408 StoreAddr = Builder.CreateBitCast(StoreAddr, StoreTy->getPointerTo()); 4409 4410 if (StoreVal->getType()->isPointerTy()) 4411 StoreVal = Builder.CreatePtrToInt(StoreVal, Int64Ty); 4412 else { 4413 StoreVal = Builder.CreateBitCast(StoreVal, StoreTy); 4414 StoreVal = Builder.CreateZExtOrBitCast(StoreVal, Int64Ty); 4415 } 4416 4417 Function *F = CGM.getIntrinsic(BuiltinID == AArch64::BI__builtin_arm_stlex 4418 ? Intrinsic::aarch64_stlxr 4419 : Intrinsic::aarch64_stxr, 4420 StoreAddr->getType()); 4421 return Builder.CreateCall(F, {StoreVal, StoreAddr}, "stxr"); 4422 } 4423 4424 if (BuiltinID == AArch64::BI__builtin_arm_clrex) { 4425 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_clrex); 4426 return Builder.CreateCall(F); 4427 } 4428 4429 if (BuiltinID == AArch64::BI__builtin_thread_pointer) { 4430 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_thread_pointer); 4431 return Builder.CreateCall(F); 4432 } 4433 4434 // CRC32 4435 Intrinsic::ID CRCIntrinsicID = Intrinsic::not_intrinsic; 4436 switch (BuiltinID) { 4437 case AArch64::BI__builtin_arm_crc32b: 4438 CRCIntrinsicID = Intrinsic::aarch64_crc32b; break; 4439 case AArch64::BI__builtin_arm_crc32cb: 4440 CRCIntrinsicID = Intrinsic::aarch64_crc32cb; break; 4441 case AArch64::BI__builtin_arm_crc32h: 4442 CRCIntrinsicID = Intrinsic::aarch64_crc32h; break; 4443 case AArch64::BI__builtin_arm_crc32ch: 4444 CRCIntrinsicID = Intrinsic::aarch64_crc32ch; break; 4445 case AArch64::BI__builtin_arm_crc32w: 4446 CRCIntrinsicID = Intrinsic::aarch64_crc32w; break; 4447 case AArch64::BI__builtin_arm_crc32cw: 4448 CRCIntrinsicID = Intrinsic::aarch64_crc32cw; break; 4449 case AArch64::BI__builtin_arm_crc32d: 4450 CRCIntrinsicID = Intrinsic::aarch64_crc32x; break; 4451 case AArch64::BI__builtin_arm_crc32cd: 4452 CRCIntrinsicID = Intrinsic::aarch64_crc32cx; break; 4453 } 4454 4455 if (CRCIntrinsicID != Intrinsic::not_intrinsic) { 4456 Value *Arg0 = EmitScalarExpr(E->getArg(0)); 4457 Value *Arg1 = EmitScalarExpr(E->getArg(1)); 4458 Function *F = CGM.getIntrinsic(CRCIntrinsicID); 4459 4460 llvm::Type *DataTy = F->getFunctionType()->getParamType(1); 4461 Arg1 = Builder.CreateZExtOrBitCast(Arg1, DataTy); 4462 4463 return Builder.CreateCall(F, {Arg0, Arg1}); 4464 } 4465 4466 if (BuiltinID == AArch64::BI__builtin_arm_rsr || 4467 BuiltinID == AArch64::BI__builtin_arm_rsr64 || 4468 BuiltinID == AArch64::BI__builtin_arm_rsrp || 4469 BuiltinID == AArch64::BI__builtin_arm_wsr || 4470 BuiltinID == AArch64::BI__builtin_arm_wsr64 || 4471 BuiltinID == AArch64::BI__builtin_arm_wsrp) { 4472 4473 bool IsRead = BuiltinID == AArch64::BI__builtin_arm_rsr || 4474 BuiltinID == AArch64::BI__builtin_arm_rsr64 || 4475 BuiltinID == AArch64::BI__builtin_arm_rsrp; 4476 4477 bool IsPointerBuiltin = BuiltinID == AArch64::BI__builtin_arm_rsrp || 4478 BuiltinID == AArch64::BI__builtin_arm_wsrp; 4479 4480 bool Is64Bit = BuiltinID != AArch64::BI__builtin_arm_rsr && 4481 BuiltinID != AArch64::BI__builtin_arm_wsr; 4482 4483 llvm::Type *ValueType; 4484 llvm::Type *RegisterType = Int64Ty; 4485 if (IsPointerBuiltin) { 4486 ValueType = VoidPtrTy; 4487 } else if (Is64Bit) { 4488 ValueType = Int64Ty; 4489 } else { 4490 ValueType = Int32Ty; 4491 } 4492 4493 return EmitSpecialRegisterBuiltin(*this, E, RegisterType, ValueType, IsRead); 4494 } 4495 4496 // Find out if any arguments are required to be integer constant 4497 // expressions. 4498 unsigned ICEArguments = 0; 4499 ASTContext::GetBuiltinTypeError Error; 4500 getContext().GetBuiltinType(BuiltinID, Error, &ICEArguments); 4501 assert(Error == ASTContext::GE_None && "Should not codegen an error"); 4502 4503 llvm::SmallVector<Value*, 4> Ops; 4504 for (unsigned i = 0, e = E->getNumArgs() - 1; i != e; i++) { 4505 if ((ICEArguments & (1 << i)) == 0) { 4506 Ops.push_back(EmitScalarExpr(E->getArg(i))); 4507 } else { 4508 // If this is required to be a constant, constant fold it so that we know 4509 // that the generated intrinsic gets a ConstantInt. 4510 llvm::APSInt Result; 4511 bool IsConst = E->getArg(i)->isIntegerConstantExpr(Result, getContext()); 4512 assert(IsConst && "Constant arg isn't actually constant?"); 4513 (void)IsConst; 4514 Ops.push_back(llvm::ConstantInt::get(getLLVMContext(), Result)); 4515 } 4516 } 4517 4518 auto SISDMap = makeArrayRef(AArch64SISDIntrinsicMap); 4519 const NeonIntrinsicInfo *Builtin = findNeonIntrinsicInMap( 4520 SISDMap, BuiltinID, AArch64SISDIntrinsicsProvenSorted); 4521 4522 if (Builtin) { 4523 Ops.push_back(EmitScalarExpr(E->getArg(E->getNumArgs() - 1))); 4524 Value *Result = EmitCommonNeonSISDBuiltinExpr(*this, *Builtin, Ops, E); 4525 assert(Result && "SISD intrinsic should have been handled"); 4526 return Result; 4527 } 4528 4529 llvm::APSInt Result; 4530 const Expr *Arg = E->getArg(E->getNumArgs()-1); 4531 NeonTypeFlags Type(0); 4532 if (Arg->isIntegerConstantExpr(Result, getContext())) 4533 // Determine the type of this overloaded NEON intrinsic. 4534 Type = NeonTypeFlags(Result.getZExtValue()); 4535 4536 bool usgn = Type.isUnsigned(); 4537 bool quad = Type.isQuad(); 4538 4539 // Handle non-overloaded intrinsics first. 4540 switch (BuiltinID) { 4541 default: break; 4542 case NEON::BI__builtin_neon_vldrq_p128: { 4543 llvm::Type *Int128PTy = llvm::Type::getIntNPtrTy(getLLVMContext(), 128); 4544 Value *Ptr = Builder.CreateBitCast(EmitScalarExpr(E->getArg(0)), Int128PTy); 4545 return Builder.CreateDefaultAlignedLoad(Ptr); 4546 } 4547 case NEON::BI__builtin_neon_vstrq_p128: { 4548 llvm::Type *Int128PTy = llvm::Type::getIntNPtrTy(getLLVMContext(), 128); 4549 Value *Ptr = Builder.CreateBitCast(Ops[0], Int128PTy); 4550 return Builder.CreateDefaultAlignedStore(EmitScalarExpr(E->getArg(1)), Ptr); 4551 } 4552 case NEON::BI__builtin_neon_vcvts_u32_f32: 4553 case NEON::BI__builtin_neon_vcvtd_u64_f64: 4554 usgn = true; 4555 // FALL THROUGH 4556 case NEON::BI__builtin_neon_vcvts_s32_f32: 4557 case NEON::BI__builtin_neon_vcvtd_s64_f64: { 4558 Ops.push_back(EmitScalarExpr(E->getArg(0))); 4559 bool Is64 = Ops[0]->getType()->getPrimitiveSizeInBits() == 64; 4560 llvm::Type *InTy = Is64 ? Int64Ty : Int32Ty; 4561 llvm::Type *FTy = Is64 ? DoubleTy : FloatTy; 4562 Ops[0] = Builder.CreateBitCast(Ops[0], FTy); 4563 if (usgn) 4564 return Builder.CreateFPToUI(Ops[0], InTy); 4565 return Builder.CreateFPToSI(Ops[0], InTy); 4566 } 4567 case NEON::BI__builtin_neon_vcvts_f32_u32: 4568 case NEON::BI__builtin_neon_vcvtd_f64_u64: 4569 usgn = true; 4570 // FALL THROUGH 4571 case NEON::BI__builtin_neon_vcvts_f32_s32: 4572 case NEON::BI__builtin_neon_vcvtd_f64_s64: { 4573 Ops.push_back(EmitScalarExpr(E->getArg(0))); 4574 bool Is64 = Ops[0]->getType()->getPrimitiveSizeInBits() == 64; 4575 llvm::Type *InTy = Is64 ? Int64Ty : Int32Ty; 4576 llvm::Type *FTy = Is64 ? DoubleTy : FloatTy; 4577 Ops[0] = Builder.CreateBitCast(Ops[0], InTy); 4578 if (usgn) 4579 return Builder.CreateUIToFP(Ops[0], FTy); 4580 return Builder.CreateSIToFP(Ops[0], FTy); 4581 } 4582 case NEON::BI__builtin_neon_vpaddd_s64: { 4583 llvm::Type *Ty = llvm::VectorType::get(Int64Ty, 2); 4584 Value *Vec = EmitScalarExpr(E->getArg(0)); 4585 // The vector is v2f64, so make sure it's bitcast to that. 4586 Vec = Builder.CreateBitCast(Vec, Ty, "v2i64"); 4587 llvm::Value *Idx0 = llvm::ConstantInt::get(SizeTy, 0); 4588 llvm::Value *Idx1 = llvm::ConstantInt::get(SizeTy, 1); 4589 Value *Op0 = Builder.CreateExtractElement(Vec, Idx0, "lane0"); 4590 Value *Op1 = Builder.CreateExtractElement(Vec, Idx1, "lane1"); 4591 // Pairwise addition of a v2f64 into a scalar f64. 4592 return Builder.CreateAdd(Op0, Op1, "vpaddd"); 4593 } 4594 case NEON::BI__builtin_neon_vpaddd_f64: { 4595 llvm::Type *Ty = 4596 llvm::VectorType::get(DoubleTy, 2); 4597 Value *Vec = EmitScalarExpr(E->getArg(0)); 4598 // The vector is v2f64, so make sure it's bitcast to that. 4599 Vec = Builder.CreateBitCast(Vec, Ty, "v2f64"); 4600 llvm::Value *Idx0 = llvm::ConstantInt::get(SizeTy, 0); 4601 llvm::Value *Idx1 = llvm::ConstantInt::get(SizeTy, 1); 4602 Value *Op0 = Builder.CreateExtractElement(Vec, Idx0, "lane0"); 4603 Value *Op1 = Builder.CreateExtractElement(Vec, Idx1, "lane1"); 4604 // Pairwise addition of a v2f64 into a scalar f64. 4605 return Builder.CreateFAdd(Op0, Op1, "vpaddd"); 4606 } 4607 case NEON::BI__builtin_neon_vpadds_f32: { 4608 llvm::Type *Ty = 4609 llvm::VectorType::get(FloatTy, 2); 4610 Value *Vec = EmitScalarExpr(E->getArg(0)); 4611 // The vector is v2f32, so make sure it's bitcast to that. 4612 Vec = Builder.CreateBitCast(Vec, Ty, "v2f32"); 4613 llvm::Value *Idx0 = llvm::ConstantInt::get(SizeTy, 0); 4614 llvm::Value *Idx1 = llvm::ConstantInt::get(SizeTy, 1); 4615 Value *Op0 = Builder.CreateExtractElement(Vec, Idx0, "lane0"); 4616 Value *Op1 = Builder.CreateExtractElement(Vec, Idx1, "lane1"); 4617 // Pairwise addition of a v2f32 into a scalar f32. 4618 return Builder.CreateFAdd(Op0, Op1, "vpaddd"); 4619 } 4620 case NEON::BI__builtin_neon_vceqzd_s64: 4621 case NEON::BI__builtin_neon_vceqzd_f64: 4622 case NEON::BI__builtin_neon_vceqzs_f32: 4623 Ops.push_back(EmitScalarExpr(E->getArg(0))); 4624 return EmitAArch64CompareBuiltinExpr( 4625 Ops[0], ConvertType(E->getCallReturnType(getContext())), 4626 ICmpInst::FCMP_OEQ, ICmpInst::ICMP_EQ, "vceqz"); 4627 case NEON::BI__builtin_neon_vcgezd_s64: 4628 case NEON::BI__builtin_neon_vcgezd_f64: 4629 case NEON::BI__builtin_neon_vcgezs_f32: 4630 Ops.push_back(EmitScalarExpr(E->getArg(0))); 4631 return EmitAArch64CompareBuiltinExpr( 4632 Ops[0], ConvertType(E->getCallReturnType(getContext())), 4633 ICmpInst::FCMP_OGE, ICmpInst::ICMP_SGE, "vcgez"); 4634 case NEON::BI__builtin_neon_vclezd_s64: 4635 case NEON::BI__builtin_neon_vclezd_f64: 4636 case NEON::BI__builtin_neon_vclezs_f32: 4637 Ops.push_back(EmitScalarExpr(E->getArg(0))); 4638 return EmitAArch64CompareBuiltinExpr( 4639 Ops[0], ConvertType(E->getCallReturnType(getContext())), 4640 ICmpInst::FCMP_OLE, ICmpInst::ICMP_SLE, "vclez"); 4641 case NEON::BI__builtin_neon_vcgtzd_s64: 4642 case NEON::BI__builtin_neon_vcgtzd_f64: 4643 case NEON::BI__builtin_neon_vcgtzs_f32: 4644 Ops.push_back(EmitScalarExpr(E->getArg(0))); 4645 return EmitAArch64CompareBuiltinExpr( 4646 Ops[0], ConvertType(E->getCallReturnType(getContext())), 4647 ICmpInst::FCMP_OGT, ICmpInst::ICMP_SGT, "vcgtz"); 4648 case NEON::BI__builtin_neon_vcltzd_s64: 4649 case NEON::BI__builtin_neon_vcltzd_f64: 4650 case NEON::BI__builtin_neon_vcltzs_f32: 4651 Ops.push_back(EmitScalarExpr(E->getArg(0))); 4652 return EmitAArch64CompareBuiltinExpr( 4653 Ops[0], ConvertType(E->getCallReturnType(getContext())), 4654 ICmpInst::FCMP_OLT, ICmpInst::ICMP_SLT, "vcltz"); 4655 4656 case NEON::BI__builtin_neon_vceqzd_u64: { 4657 Ops.push_back(EmitScalarExpr(E->getArg(0))); 4658 Ops[0] = Builder.CreateBitCast(Ops[0], Int64Ty); 4659 Ops[0] = 4660 Builder.CreateICmpEQ(Ops[0], llvm::Constant::getNullValue(Int64Ty)); 4661 return Builder.CreateSExt(Ops[0], Int64Ty, "vceqzd"); 4662 } 4663 case NEON::BI__builtin_neon_vceqd_f64: 4664 case NEON::BI__builtin_neon_vcled_f64: 4665 case NEON::BI__builtin_neon_vcltd_f64: 4666 case NEON::BI__builtin_neon_vcged_f64: 4667 case NEON::BI__builtin_neon_vcgtd_f64: { 4668 llvm::CmpInst::Predicate P; 4669 switch (BuiltinID) { 4670 default: llvm_unreachable("missing builtin ID in switch!"); 4671 case NEON::BI__builtin_neon_vceqd_f64: P = llvm::FCmpInst::FCMP_OEQ; break; 4672 case NEON::BI__builtin_neon_vcled_f64: P = llvm::FCmpInst::FCMP_OLE; break; 4673 case NEON::BI__builtin_neon_vcltd_f64: P = llvm::FCmpInst::FCMP_OLT; break; 4674 case NEON::BI__builtin_neon_vcged_f64: P = llvm::FCmpInst::FCMP_OGE; break; 4675 case NEON::BI__builtin_neon_vcgtd_f64: P = llvm::FCmpInst::FCMP_OGT; break; 4676 } 4677 Ops.push_back(EmitScalarExpr(E->getArg(1))); 4678 Ops[0] = Builder.CreateBitCast(Ops[0], DoubleTy); 4679 Ops[1] = Builder.CreateBitCast(Ops[1], DoubleTy); 4680 Ops[0] = Builder.CreateFCmp(P, Ops[0], Ops[1]); 4681 return Builder.CreateSExt(Ops[0], Int64Ty, "vcmpd"); 4682 } 4683 case NEON::BI__builtin_neon_vceqs_f32: 4684 case NEON::BI__builtin_neon_vcles_f32: 4685 case NEON::BI__builtin_neon_vclts_f32: 4686 case NEON::BI__builtin_neon_vcges_f32: 4687 case NEON::BI__builtin_neon_vcgts_f32: { 4688 llvm::CmpInst::Predicate P; 4689 switch (BuiltinID) { 4690 default: llvm_unreachable("missing builtin ID in switch!"); 4691 case NEON::BI__builtin_neon_vceqs_f32: P = llvm::FCmpInst::FCMP_OEQ; break; 4692 case NEON::BI__builtin_neon_vcles_f32: P = llvm::FCmpInst::FCMP_OLE; break; 4693 case NEON::BI__builtin_neon_vclts_f32: P = llvm::FCmpInst::FCMP_OLT; break; 4694 case NEON::BI__builtin_neon_vcges_f32: P = llvm::FCmpInst::FCMP_OGE; break; 4695 case NEON::BI__builtin_neon_vcgts_f32: P = llvm::FCmpInst::FCMP_OGT; break; 4696 } 4697 Ops.push_back(EmitScalarExpr(E->getArg(1))); 4698 Ops[0] = Builder.CreateBitCast(Ops[0], FloatTy); 4699 Ops[1] = Builder.CreateBitCast(Ops[1], FloatTy); 4700 Ops[0] = Builder.CreateFCmp(P, Ops[0], Ops[1]); 4701 return Builder.CreateSExt(Ops[0], Int32Ty, "vcmpd"); 4702 } 4703 case NEON::BI__builtin_neon_vceqd_s64: 4704 case NEON::BI__builtin_neon_vceqd_u64: 4705 case NEON::BI__builtin_neon_vcgtd_s64: 4706 case NEON::BI__builtin_neon_vcgtd_u64: 4707 case NEON::BI__builtin_neon_vcltd_s64: 4708 case NEON::BI__builtin_neon_vcltd_u64: 4709 case NEON::BI__builtin_neon_vcged_u64: 4710 case NEON::BI__builtin_neon_vcged_s64: 4711 case NEON::BI__builtin_neon_vcled_u64: 4712 case NEON::BI__builtin_neon_vcled_s64: { 4713 llvm::CmpInst::Predicate P; 4714 switch (BuiltinID) { 4715 default: llvm_unreachable("missing builtin ID in switch!"); 4716 case NEON::BI__builtin_neon_vceqd_s64: 4717 case NEON::BI__builtin_neon_vceqd_u64:P = llvm::ICmpInst::ICMP_EQ;break; 4718 case NEON::BI__builtin_neon_vcgtd_s64:P = llvm::ICmpInst::ICMP_SGT;break; 4719 case NEON::BI__builtin_neon_vcgtd_u64:P = llvm::ICmpInst::ICMP_UGT;break; 4720 case NEON::BI__builtin_neon_vcltd_s64:P = llvm::ICmpInst::ICMP_SLT;break; 4721 case NEON::BI__builtin_neon_vcltd_u64:P = llvm::ICmpInst::ICMP_ULT;break; 4722 case NEON::BI__builtin_neon_vcged_u64:P = llvm::ICmpInst::ICMP_UGE;break; 4723 case NEON::BI__builtin_neon_vcged_s64:P = llvm::ICmpInst::ICMP_SGE;break; 4724 case NEON::BI__builtin_neon_vcled_u64:P = llvm::ICmpInst::ICMP_ULE;break; 4725 case NEON::BI__builtin_neon_vcled_s64:P = llvm::ICmpInst::ICMP_SLE;break; 4726 } 4727 Ops.push_back(EmitScalarExpr(E->getArg(1))); 4728 Ops[0] = Builder.CreateBitCast(Ops[0], Int64Ty); 4729 Ops[1] = Builder.CreateBitCast(Ops[1], Int64Ty); 4730 Ops[0] = Builder.CreateICmp(P, Ops[0], Ops[1]); 4731 return Builder.CreateSExt(Ops[0], Int64Ty, "vceqd"); 4732 } 4733 case NEON::BI__builtin_neon_vtstd_s64: 4734 case NEON::BI__builtin_neon_vtstd_u64: { 4735 Ops.push_back(EmitScalarExpr(E->getArg(1))); 4736 Ops[0] = Builder.CreateBitCast(Ops[0], Int64Ty); 4737 Ops[1] = Builder.CreateBitCast(Ops[1], Int64Ty); 4738 Ops[0] = Builder.CreateAnd(Ops[0], Ops[1]); 4739 Ops[0] = Builder.CreateICmp(ICmpInst::ICMP_NE, Ops[0], 4740 llvm::Constant::getNullValue(Int64Ty)); 4741 return Builder.CreateSExt(Ops[0], Int64Ty, "vtstd"); 4742 } 4743 case NEON::BI__builtin_neon_vset_lane_i8: 4744 case NEON::BI__builtin_neon_vset_lane_i16: 4745 case NEON::BI__builtin_neon_vset_lane_i32: 4746 case NEON::BI__builtin_neon_vset_lane_i64: 4747 case NEON::BI__builtin_neon_vset_lane_f32: 4748 case NEON::BI__builtin_neon_vsetq_lane_i8: 4749 case NEON::BI__builtin_neon_vsetq_lane_i16: 4750 case NEON::BI__builtin_neon_vsetq_lane_i32: 4751 case NEON::BI__builtin_neon_vsetq_lane_i64: 4752 case NEON::BI__builtin_neon_vsetq_lane_f32: 4753 Ops.push_back(EmitScalarExpr(E->getArg(2))); 4754 return Builder.CreateInsertElement(Ops[1], Ops[0], Ops[2], "vset_lane"); 4755 case NEON::BI__builtin_neon_vset_lane_f64: 4756 // The vector type needs a cast for the v1f64 variant. 4757 Ops[1] = Builder.CreateBitCast(Ops[1], 4758 llvm::VectorType::get(DoubleTy, 1)); 4759 Ops.push_back(EmitScalarExpr(E->getArg(2))); 4760 return Builder.CreateInsertElement(Ops[1], Ops[0], Ops[2], "vset_lane"); 4761 case NEON::BI__builtin_neon_vsetq_lane_f64: 4762 // The vector type needs a cast for the v2f64 variant. 4763 Ops[1] = Builder.CreateBitCast(Ops[1], 4764 llvm::VectorType::get(DoubleTy, 2)); 4765 Ops.push_back(EmitScalarExpr(E->getArg(2))); 4766 return Builder.CreateInsertElement(Ops[1], Ops[0], Ops[2], "vset_lane"); 4767 4768 case NEON::BI__builtin_neon_vget_lane_i8: 4769 case NEON::BI__builtin_neon_vdupb_lane_i8: 4770 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::VectorType::get(Int8Ty, 8)); 4771 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4772 "vget_lane"); 4773 case NEON::BI__builtin_neon_vgetq_lane_i8: 4774 case NEON::BI__builtin_neon_vdupb_laneq_i8: 4775 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::VectorType::get(Int8Ty, 16)); 4776 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4777 "vgetq_lane"); 4778 case NEON::BI__builtin_neon_vget_lane_i16: 4779 case NEON::BI__builtin_neon_vduph_lane_i16: 4780 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::VectorType::get(Int16Ty, 4)); 4781 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4782 "vget_lane"); 4783 case NEON::BI__builtin_neon_vgetq_lane_i16: 4784 case NEON::BI__builtin_neon_vduph_laneq_i16: 4785 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::VectorType::get(Int16Ty, 8)); 4786 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4787 "vgetq_lane"); 4788 case NEON::BI__builtin_neon_vget_lane_i32: 4789 case NEON::BI__builtin_neon_vdups_lane_i32: 4790 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::VectorType::get(Int32Ty, 2)); 4791 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4792 "vget_lane"); 4793 case NEON::BI__builtin_neon_vdups_lane_f32: 4794 Ops[0] = Builder.CreateBitCast(Ops[0], 4795 llvm::VectorType::get(FloatTy, 2)); 4796 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4797 "vdups_lane"); 4798 case NEON::BI__builtin_neon_vgetq_lane_i32: 4799 case NEON::BI__builtin_neon_vdups_laneq_i32: 4800 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::VectorType::get(Int32Ty, 4)); 4801 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4802 "vgetq_lane"); 4803 case NEON::BI__builtin_neon_vget_lane_i64: 4804 case NEON::BI__builtin_neon_vdupd_lane_i64: 4805 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::VectorType::get(Int64Ty, 1)); 4806 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4807 "vget_lane"); 4808 case NEON::BI__builtin_neon_vdupd_lane_f64: 4809 Ops[0] = Builder.CreateBitCast(Ops[0], 4810 llvm::VectorType::get(DoubleTy, 1)); 4811 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4812 "vdupd_lane"); 4813 case NEON::BI__builtin_neon_vgetq_lane_i64: 4814 case NEON::BI__builtin_neon_vdupd_laneq_i64: 4815 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::VectorType::get(Int64Ty, 2)); 4816 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4817 "vgetq_lane"); 4818 case NEON::BI__builtin_neon_vget_lane_f32: 4819 Ops[0] = Builder.CreateBitCast(Ops[0], 4820 llvm::VectorType::get(FloatTy, 2)); 4821 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4822 "vget_lane"); 4823 case NEON::BI__builtin_neon_vget_lane_f64: 4824 Ops[0] = Builder.CreateBitCast(Ops[0], 4825 llvm::VectorType::get(DoubleTy, 1)); 4826 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4827 "vget_lane"); 4828 case NEON::BI__builtin_neon_vgetq_lane_f32: 4829 case NEON::BI__builtin_neon_vdups_laneq_f32: 4830 Ops[0] = Builder.CreateBitCast(Ops[0], 4831 llvm::VectorType::get(FloatTy, 4)); 4832 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4833 "vgetq_lane"); 4834 case NEON::BI__builtin_neon_vgetq_lane_f64: 4835 case NEON::BI__builtin_neon_vdupd_laneq_f64: 4836 Ops[0] = Builder.CreateBitCast(Ops[0], 4837 llvm::VectorType::get(DoubleTy, 2)); 4838 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4839 "vgetq_lane"); 4840 case NEON::BI__builtin_neon_vaddd_s64: 4841 case NEON::BI__builtin_neon_vaddd_u64: 4842 return Builder.CreateAdd(Ops[0], EmitScalarExpr(E->getArg(1)), "vaddd"); 4843 case NEON::BI__builtin_neon_vsubd_s64: 4844 case NEON::BI__builtin_neon_vsubd_u64: 4845 return Builder.CreateSub(Ops[0], EmitScalarExpr(E->getArg(1)), "vsubd"); 4846 case NEON::BI__builtin_neon_vqdmlalh_s16: 4847 case NEON::BI__builtin_neon_vqdmlslh_s16: { 4848 SmallVector<Value *, 2> ProductOps; 4849 ProductOps.push_back(vectorWrapScalar16(Ops[1])); 4850 ProductOps.push_back(vectorWrapScalar16(EmitScalarExpr(E->getArg(2)))); 4851 llvm::Type *VTy = llvm::VectorType::get(Int32Ty, 4); 4852 Ops[1] = EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_sqdmull, VTy), 4853 ProductOps, "vqdmlXl"); 4854 Constant *CI = ConstantInt::get(SizeTy, 0); 4855 Ops[1] = Builder.CreateExtractElement(Ops[1], CI, "lane0"); 4856 4857 unsigned AccumInt = BuiltinID == NEON::BI__builtin_neon_vqdmlalh_s16 4858 ? Intrinsic::aarch64_neon_sqadd 4859 : Intrinsic::aarch64_neon_sqsub; 4860 return EmitNeonCall(CGM.getIntrinsic(AccumInt, Int32Ty), Ops, "vqdmlXl"); 4861 } 4862 case NEON::BI__builtin_neon_vqshlud_n_s64: { 4863 Ops.push_back(EmitScalarExpr(E->getArg(1))); 4864 Ops[1] = Builder.CreateZExt(Ops[1], Int64Ty); 4865 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_sqshlu, Int64Ty), 4866 Ops, "vqshlu_n"); 4867 } 4868 case NEON::BI__builtin_neon_vqshld_n_u64: 4869 case NEON::BI__builtin_neon_vqshld_n_s64: { 4870 unsigned Int = BuiltinID == NEON::BI__builtin_neon_vqshld_n_u64 4871 ? Intrinsic::aarch64_neon_uqshl 4872 : Intrinsic::aarch64_neon_sqshl; 4873 Ops.push_back(EmitScalarExpr(E->getArg(1))); 4874 Ops[1] = Builder.CreateZExt(Ops[1], Int64Ty); 4875 return EmitNeonCall(CGM.getIntrinsic(Int, Int64Ty), Ops, "vqshl_n"); 4876 } 4877 case NEON::BI__builtin_neon_vrshrd_n_u64: 4878 case NEON::BI__builtin_neon_vrshrd_n_s64: { 4879 unsigned Int = BuiltinID == NEON::BI__builtin_neon_vrshrd_n_u64 4880 ? Intrinsic::aarch64_neon_urshl 4881 : Intrinsic::aarch64_neon_srshl; 4882 Ops.push_back(EmitScalarExpr(E->getArg(1))); 4883 int SV = cast<ConstantInt>(Ops[1])->getSExtValue(); 4884 Ops[1] = ConstantInt::get(Int64Ty, -SV); 4885 return EmitNeonCall(CGM.getIntrinsic(Int, Int64Ty), Ops, "vrshr_n"); 4886 } 4887 case NEON::BI__builtin_neon_vrsrad_n_u64: 4888 case NEON::BI__builtin_neon_vrsrad_n_s64: { 4889 unsigned Int = BuiltinID == NEON::BI__builtin_neon_vrsrad_n_u64 4890 ? Intrinsic::aarch64_neon_urshl 4891 : Intrinsic::aarch64_neon_srshl; 4892 Ops[1] = Builder.CreateBitCast(Ops[1], Int64Ty); 4893 Ops.push_back(Builder.CreateNeg(EmitScalarExpr(E->getArg(2)))); 4894 Ops[1] = Builder.CreateCall(CGM.getIntrinsic(Int, Int64Ty), 4895 {Ops[1], Builder.CreateSExt(Ops[2], Int64Ty)}); 4896 return Builder.CreateAdd(Ops[0], Builder.CreateBitCast(Ops[1], Int64Ty)); 4897 } 4898 case NEON::BI__builtin_neon_vshld_n_s64: 4899 case NEON::BI__builtin_neon_vshld_n_u64: { 4900 llvm::ConstantInt *Amt = cast<ConstantInt>(EmitScalarExpr(E->getArg(1))); 4901 return Builder.CreateShl( 4902 Ops[0], ConstantInt::get(Int64Ty, Amt->getZExtValue()), "shld_n"); 4903 } 4904 case NEON::BI__builtin_neon_vshrd_n_s64: { 4905 llvm::ConstantInt *Amt = cast<ConstantInt>(EmitScalarExpr(E->getArg(1))); 4906 return Builder.CreateAShr( 4907 Ops[0], ConstantInt::get(Int64Ty, std::min(static_cast<uint64_t>(63), 4908 Amt->getZExtValue())), 4909 "shrd_n"); 4910 } 4911 case NEON::BI__builtin_neon_vshrd_n_u64: { 4912 llvm::ConstantInt *Amt = cast<ConstantInt>(EmitScalarExpr(E->getArg(1))); 4913 uint64_t ShiftAmt = Amt->getZExtValue(); 4914 // Right-shifting an unsigned value by its size yields 0. 4915 if (ShiftAmt == 64) 4916 return ConstantInt::get(Int64Ty, 0); 4917 return Builder.CreateLShr(Ops[0], ConstantInt::get(Int64Ty, ShiftAmt), 4918 "shrd_n"); 4919 } 4920 case NEON::BI__builtin_neon_vsrad_n_s64: { 4921 llvm::ConstantInt *Amt = cast<ConstantInt>(EmitScalarExpr(E->getArg(2))); 4922 Ops[1] = Builder.CreateAShr( 4923 Ops[1], ConstantInt::get(Int64Ty, std::min(static_cast<uint64_t>(63), 4924 Amt->getZExtValue())), 4925 "shrd_n"); 4926 return Builder.CreateAdd(Ops[0], Ops[1]); 4927 } 4928 case NEON::BI__builtin_neon_vsrad_n_u64: { 4929 llvm::ConstantInt *Amt = cast<ConstantInt>(EmitScalarExpr(E->getArg(2))); 4930 uint64_t ShiftAmt = Amt->getZExtValue(); 4931 // Right-shifting an unsigned value by its size yields 0. 4932 // As Op + 0 = Op, return Ops[0] directly. 4933 if (ShiftAmt == 64) 4934 return Ops[0]; 4935 Ops[1] = Builder.CreateLShr(Ops[1], ConstantInt::get(Int64Ty, ShiftAmt), 4936 "shrd_n"); 4937 return Builder.CreateAdd(Ops[0], Ops[1]); 4938 } 4939 case NEON::BI__builtin_neon_vqdmlalh_lane_s16: 4940 case NEON::BI__builtin_neon_vqdmlalh_laneq_s16: 4941 case NEON::BI__builtin_neon_vqdmlslh_lane_s16: 4942 case NEON::BI__builtin_neon_vqdmlslh_laneq_s16: { 4943 Ops[2] = Builder.CreateExtractElement(Ops[2], EmitScalarExpr(E->getArg(3)), 4944 "lane"); 4945 SmallVector<Value *, 2> ProductOps; 4946 ProductOps.push_back(vectorWrapScalar16(Ops[1])); 4947 ProductOps.push_back(vectorWrapScalar16(Ops[2])); 4948 llvm::Type *VTy = llvm::VectorType::get(Int32Ty, 4); 4949 Ops[1] = EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_sqdmull, VTy), 4950 ProductOps, "vqdmlXl"); 4951 Constant *CI = ConstantInt::get(SizeTy, 0); 4952 Ops[1] = Builder.CreateExtractElement(Ops[1], CI, "lane0"); 4953 Ops.pop_back(); 4954 4955 unsigned AccInt = (BuiltinID == NEON::BI__builtin_neon_vqdmlalh_lane_s16 || 4956 BuiltinID == NEON::BI__builtin_neon_vqdmlalh_laneq_s16) 4957 ? Intrinsic::aarch64_neon_sqadd 4958 : Intrinsic::aarch64_neon_sqsub; 4959 return EmitNeonCall(CGM.getIntrinsic(AccInt, Int32Ty), Ops, "vqdmlXl"); 4960 } 4961 case NEON::BI__builtin_neon_vqdmlals_s32: 4962 case NEON::BI__builtin_neon_vqdmlsls_s32: { 4963 SmallVector<Value *, 2> ProductOps; 4964 ProductOps.push_back(Ops[1]); 4965 ProductOps.push_back(EmitScalarExpr(E->getArg(2))); 4966 Ops[1] = 4967 EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_sqdmulls_scalar), 4968 ProductOps, "vqdmlXl"); 4969 4970 unsigned AccumInt = BuiltinID == NEON::BI__builtin_neon_vqdmlals_s32 4971 ? Intrinsic::aarch64_neon_sqadd 4972 : Intrinsic::aarch64_neon_sqsub; 4973 return EmitNeonCall(CGM.getIntrinsic(AccumInt, Int64Ty), Ops, "vqdmlXl"); 4974 } 4975 case NEON::BI__builtin_neon_vqdmlals_lane_s32: 4976 case NEON::BI__builtin_neon_vqdmlals_laneq_s32: 4977 case NEON::BI__builtin_neon_vqdmlsls_lane_s32: 4978 case NEON::BI__builtin_neon_vqdmlsls_laneq_s32: { 4979 Ops[2] = Builder.CreateExtractElement(Ops[2], EmitScalarExpr(E->getArg(3)), 4980 "lane"); 4981 SmallVector<Value *, 2> ProductOps; 4982 ProductOps.push_back(Ops[1]); 4983 ProductOps.push_back(Ops[2]); 4984 Ops[1] = 4985 EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_sqdmulls_scalar), 4986 ProductOps, "vqdmlXl"); 4987 Ops.pop_back(); 4988 4989 unsigned AccInt = (BuiltinID == NEON::BI__builtin_neon_vqdmlals_lane_s32 || 4990 BuiltinID == NEON::BI__builtin_neon_vqdmlals_laneq_s32) 4991 ? Intrinsic::aarch64_neon_sqadd 4992 : Intrinsic::aarch64_neon_sqsub; 4993 return EmitNeonCall(CGM.getIntrinsic(AccInt, Int64Ty), Ops, "vqdmlXl"); 4994 } 4995 } 4996 4997 llvm::VectorType *VTy = GetNeonType(this, Type); 4998 llvm::Type *Ty = VTy; 4999 if (!Ty) 5000 return nullptr; 5001 5002 // Not all intrinsics handled by the common case work for AArch64 yet, so only 5003 // defer to common code if it's been added to our special map. 5004 Builtin = findNeonIntrinsicInMap(AArch64SIMDIntrinsicMap, BuiltinID, 5005 AArch64SIMDIntrinsicsProvenSorted); 5006 5007 if (Builtin) 5008 return EmitCommonNeonBuiltinExpr( 5009 Builtin->BuiltinID, Builtin->LLVMIntrinsic, Builtin->AltLLVMIntrinsic, 5010 Builtin->NameHint, Builtin->TypeModifier, E, Ops, 5011 /*never use addresses*/ Address::invalid(), Address::invalid()); 5012 5013 if (Value *V = EmitAArch64TblBuiltinExpr(*this, BuiltinID, E, Ops)) 5014 return V; 5015 5016 unsigned Int; 5017 switch (BuiltinID) { 5018 default: return nullptr; 5019 case NEON::BI__builtin_neon_vbsl_v: 5020 case NEON::BI__builtin_neon_vbslq_v: { 5021 llvm::Type *BitTy = llvm::VectorType::getInteger(VTy); 5022 Ops[0] = Builder.CreateBitCast(Ops[0], BitTy, "vbsl"); 5023 Ops[1] = Builder.CreateBitCast(Ops[1], BitTy, "vbsl"); 5024 Ops[2] = Builder.CreateBitCast(Ops[2], BitTy, "vbsl"); 5025 5026 Ops[1] = Builder.CreateAnd(Ops[0], Ops[1], "vbsl"); 5027 Ops[2] = Builder.CreateAnd(Builder.CreateNot(Ops[0]), Ops[2], "vbsl"); 5028 Ops[0] = Builder.CreateOr(Ops[1], Ops[2], "vbsl"); 5029 return Builder.CreateBitCast(Ops[0], Ty); 5030 } 5031 case NEON::BI__builtin_neon_vfma_lane_v: 5032 case NEON::BI__builtin_neon_vfmaq_lane_v: { // Only used for FP types 5033 // The ARM builtins (and instructions) have the addend as the first 5034 // operand, but the 'fma' intrinsics have it last. Swap it around here. 5035 Value *Addend = Ops[0]; 5036 Value *Multiplicand = Ops[1]; 5037 Value *LaneSource = Ops[2]; 5038 Ops[0] = Multiplicand; 5039 Ops[1] = LaneSource; 5040 Ops[2] = Addend; 5041 5042 // Now adjust things to handle the lane access. 5043 llvm::Type *SourceTy = BuiltinID == NEON::BI__builtin_neon_vfmaq_lane_v ? 5044 llvm::VectorType::get(VTy->getElementType(), VTy->getNumElements() / 2) : 5045 VTy; 5046 llvm::Constant *cst = cast<Constant>(Ops[3]); 5047 Value *SV = llvm::ConstantVector::getSplat(VTy->getNumElements(), cst); 5048 Ops[1] = Builder.CreateBitCast(Ops[1], SourceTy); 5049 Ops[1] = Builder.CreateShuffleVector(Ops[1], Ops[1], SV, "lane"); 5050 5051 Ops.pop_back(); 5052 Int = Intrinsic::fma; 5053 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "fmla"); 5054 } 5055 case NEON::BI__builtin_neon_vfma_laneq_v: { 5056 llvm::VectorType *VTy = cast<llvm::VectorType>(Ty); 5057 // v1f64 fma should be mapped to Neon scalar f64 fma 5058 if (VTy && VTy->getElementType() == DoubleTy) { 5059 Ops[0] = Builder.CreateBitCast(Ops[0], DoubleTy); 5060 Ops[1] = Builder.CreateBitCast(Ops[1], DoubleTy); 5061 llvm::Type *VTy = GetNeonType(this, 5062 NeonTypeFlags(NeonTypeFlags::Float64, false, true)); 5063 Ops[2] = Builder.CreateBitCast(Ops[2], VTy); 5064 Ops[2] = Builder.CreateExtractElement(Ops[2], Ops[3], "extract"); 5065 Value *F = CGM.getIntrinsic(Intrinsic::fma, DoubleTy); 5066 Value *Result = Builder.CreateCall(F, {Ops[1], Ops[2], Ops[0]}); 5067 return Builder.CreateBitCast(Result, Ty); 5068 } 5069 Value *F = CGM.getIntrinsic(Intrinsic::fma, Ty); 5070 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 5071 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 5072 5073 llvm::Type *STy = llvm::VectorType::get(VTy->getElementType(), 5074 VTy->getNumElements() * 2); 5075 Ops[2] = Builder.CreateBitCast(Ops[2], STy); 5076 Value* SV = llvm::ConstantVector::getSplat(VTy->getNumElements(), 5077 cast<ConstantInt>(Ops[3])); 5078 Ops[2] = Builder.CreateShuffleVector(Ops[2], Ops[2], SV, "lane"); 5079 5080 return Builder.CreateCall(F, {Ops[2], Ops[1], Ops[0]}); 5081 } 5082 case NEON::BI__builtin_neon_vfmaq_laneq_v: { 5083 Value *F = CGM.getIntrinsic(Intrinsic::fma, Ty); 5084 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 5085 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 5086 5087 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 5088 Ops[2] = EmitNeonSplat(Ops[2], cast<ConstantInt>(Ops[3])); 5089 return Builder.CreateCall(F, {Ops[2], Ops[1], Ops[0]}); 5090 } 5091 case NEON::BI__builtin_neon_vfmas_lane_f32: 5092 case NEON::BI__builtin_neon_vfmas_laneq_f32: 5093 case NEON::BI__builtin_neon_vfmad_lane_f64: 5094 case NEON::BI__builtin_neon_vfmad_laneq_f64: { 5095 Ops.push_back(EmitScalarExpr(E->getArg(3))); 5096 llvm::Type *Ty = ConvertType(E->getCallReturnType(getContext())); 5097 Value *F = CGM.getIntrinsic(Intrinsic::fma, Ty); 5098 Ops[2] = Builder.CreateExtractElement(Ops[2], Ops[3], "extract"); 5099 return Builder.CreateCall(F, {Ops[1], Ops[2], Ops[0]}); 5100 } 5101 case NEON::BI__builtin_neon_vfms_v: 5102 case NEON::BI__builtin_neon_vfmsq_v: { // Only used for FP types 5103 // FIXME: probably remove when we no longer support aarch64_simd.h 5104 // (arm_neon.h delegates to vfma). 5105 5106 // The ARM builtins (and instructions) have the addend as the first 5107 // operand, but the 'fma' intrinsics have it last. Swap it around here. 5108 Value *Subtrahend = Ops[0]; 5109 Value *Multiplicand = Ops[2]; 5110 Ops[0] = Multiplicand; 5111 Ops[2] = Subtrahend; 5112 Ops[1] = Builder.CreateBitCast(Ops[1], VTy); 5113 Ops[1] = Builder.CreateFNeg(Ops[1]); 5114 Int = Intrinsic::fma; 5115 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "fmls"); 5116 } 5117 case NEON::BI__builtin_neon_vmull_v: 5118 // FIXME: improve sharing scheme to cope with 3 alternative LLVM intrinsics. 5119 Int = usgn ? Intrinsic::aarch64_neon_umull : Intrinsic::aarch64_neon_smull; 5120 if (Type.isPoly()) Int = Intrinsic::aarch64_neon_pmull; 5121 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmull"); 5122 case NEON::BI__builtin_neon_vmax_v: 5123 case NEON::BI__builtin_neon_vmaxq_v: 5124 // FIXME: improve sharing scheme to cope with 3 alternative LLVM intrinsics. 5125 Int = usgn ? Intrinsic::aarch64_neon_umax : Intrinsic::aarch64_neon_smax; 5126 if (Ty->isFPOrFPVectorTy()) Int = Intrinsic::aarch64_neon_fmax; 5127 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmax"); 5128 case NEON::BI__builtin_neon_vmin_v: 5129 case NEON::BI__builtin_neon_vminq_v: 5130 // FIXME: improve sharing scheme to cope with 3 alternative LLVM intrinsics. 5131 Int = usgn ? Intrinsic::aarch64_neon_umin : Intrinsic::aarch64_neon_smin; 5132 if (Ty->isFPOrFPVectorTy()) Int = Intrinsic::aarch64_neon_fmin; 5133 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmin"); 5134 case NEON::BI__builtin_neon_vabd_v: 5135 case NEON::BI__builtin_neon_vabdq_v: 5136 // FIXME: improve sharing scheme to cope with 3 alternative LLVM intrinsics. 5137 Int = usgn ? Intrinsic::aarch64_neon_uabd : Intrinsic::aarch64_neon_sabd; 5138 if (Ty->isFPOrFPVectorTy()) Int = Intrinsic::aarch64_neon_fabd; 5139 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vabd"); 5140 case NEON::BI__builtin_neon_vpadal_v: 5141 case NEON::BI__builtin_neon_vpadalq_v: { 5142 unsigned ArgElts = VTy->getNumElements(); 5143 llvm::IntegerType *EltTy = cast<IntegerType>(VTy->getElementType()); 5144 unsigned BitWidth = EltTy->getBitWidth(); 5145 llvm::Type *ArgTy = llvm::VectorType::get( 5146 llvm::IntegerType::get(getLLVMContext(), BitWidth/2), 2*ArgElts); 5147 llvm::Type* Tys[2] = { VTy, ArgTy }; 5148 Int = usgn ? Intrinsic::aarch64_neon_uaddlp : Intrinsic::aarch64_neon_saddlp; 5149 SmallVector<llvm::Value*, 1> TmpOps; 5150 TmpOps.push_back(Ops[1]); 5151 Function *F = CGM.getIntrinsic(Int, Tys); 5152 llvm::Value *tmp = EmitNeonCall(F, TmpOps, "vpadal"); 5153 llvm::Value *addend = Builder.CreateBitCast(Ops[0], tmp->getType()); 5154 return Builder.CreateAdd(tmp, addend); 5155 } 5156 case NEON::BI__builtin_neon_vpmin_v: 5157 case NEON::BI__builtin_neon_vpminq_v: 5158 // FIXME: improve sharing scheme to cope with 3 alternative LLVM intrinsics. 5159 Int = usgn ? Intrinsic::aarch64_neon_uminp : Intrinsic::aarch64_neon_sminp; 5160 if (Ty->isFPOrFPVectorTy()) Int = Intrinsic::aarch64_neon_fminp; 5161 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vpmin"); 5162 case NEON::BI__builtin_neon_vpmax_v: 5163 case NEON::BI__builtin_neon_vpmaxq_v: 5164 // FIXME: improve sharing scheme to cope with 3 alternative LLVM intrinsics. 5165 Int = usgn ? Intrinsic::aarch64_neon_umaxp : Intrinsic::aarch64_neon_smaxp; 5166 if (Ty->isFPOrFPVectorTy()) Int = Intrinsic::aarch64_neon_fmaxp; 5167 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vpmax"); 5168 case NEON::BI__builtin_neon_vminnm_v: 5169 case NEON::BI__builtin_neon_vminnmq_v: 5170 Int = Intrinsic::aarch64_neon_fminnm; 5171 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vminnm"); 5172 case NEON::BI__builtin_neon_vmaxnm_v: 5173 case NEON::BI__builtin_neon_vmaxnmq_v: 5174 Int = Intrinsic::aarch64_neon_fmaxnm; 5175 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmaxnm"); 5176 case NEON::BI__builtin_neon_vrecpss_f32: { 5177 Ops.push_back(EmitScalarExpr(E->getArg(1))); 5178 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_frecps, FloatTy), 5179 Ops, "vrecps"); 5180 } 5181 case NEON::BI__builtin_neon_vrecpsd_f64: { 5182 Ops.push_back(EmitScalarExpr(E->getArg(1))); 5183 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_frecps, DoubleTy), 5184 Ops, "vrecps"); 5185 } 5186 case NEON::BI__builtin_neon_vqshrun_n_v: 5187 Int = Intrinsic::aarch64_neon_sqshrun; 5188 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqshrun_n"); 5189 case NEON::BI__builtin_neon_vqrshrun_n_v: 5190 Int = Intrinsic::aarch64_neon_sqrshrun; 5191 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqrshrun_n"); 5192 case NEON::BI__builtin_neon_vqshrn_n_v: 5193 Int = usgn ? Intrinsic::aarch64_neon_uqshrn : Intrinsic::aarch64_neon_sqshrn; 5194 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqshrn_n"); 5195 case NEON::BI__builtin_neon_vrshrn_n_v: 5196 Int = Intrinsic::aarch64_neon_rshrn; 5197 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrshrn_n"); 5198 case NEON::BI__builtin_neon_vqrshrn_n_v: 5199 Int = usgn ? Intrinsic::aarch64_neon_uqrshrn : Intrinsic::aarch64_neon_sqrshrn; 5200 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqrshrn_n"); 5201 case NEON::BI__builtin_neon_vrnda_v: 5202 case NEON::BI__builtin_neon_vrndaq_v: { 5203 Int = Intrinsic::round; 5204 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrnda"); 5205 } 5206 case NEON::BI__builtin_neon_vrndi_v: 5207 case NEON::BI__builtin_neon_vrndiq_v: { 5208 Int = Intrinsic::nearbyint; 5209 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrndi"); 5210 } 5211 case NEON::BI__builtin_neon_vrndm_v: 5212 case NEON::BI__builtin_neon_vrndmq_v: { 5213 Int = Intrinsic::floor; 5214 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrndm"); 5215 } 5216 case NEON::BI__builtin_neon_vrndn_v: 5217 case NEON::BI__builtin_neon_vrndnq_v: { 5218 Int = Intrinsic::aarch64_neon_frintn; 5219 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrndn"); 5220 } 5221 case NEON::BI__builtin_neon_vrndp_v: 5222 case NEON::BI__builtin_neon_vrndpq_v: { 5223 Int = Intrinsic::ceil; 5224 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrndp"); 5225 } 5226 case NEON::BI__builtin_neon_vrndx_v: 5227 case NEON::BI__builtin_neon_vrndxq_v: { 5228 Int = Intrinsic::rint; 5229 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrndx"); 5230 } 5231 case NEON::BI__builtin_neon_vrnd_v: 5232 case NEON::BI__builtin_neon_vrndq_v: { 5233 Int = Intrinsic::trunc; 5234 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrndz"); 5235 } 5236 case NEON::BI__builtin_neon_vceqz_v: 5237 case NEON::BI__builtin_neon_vceqzq_v: 5238 return EmitAArch64CompareBuiltinExpr(Ops[0], Ty, ICmpInst::FCMP_OEQ, 5239 ICmpInst::ICMP_EQ, "vceqz"); 5240 case NEON::BI__builtin_neon_vcgez_v: 5241 case NEON::BI__builtin_neon_vcgezq_v: 5242 return EmitAArch64CompareBuiltinExpr(Ops[0], Ty, ICmpInst::FCMP_OGE, 5243 ICmpInst::ICMP_SGE, "vcgez"); 5244 case NEON::BI__builtin_neon_vclez_v: 5245 case NEON::BI__builtin_neon_vclezq_v: 5246 return EmitAArch64CompareBuiltinExpr(Ops[0], Ty, ICmpInst::FCMP_OLE, 5247 ICmpInst::ICMP_SLE, "vclez"); 5248 case NEON::BI__builtin_neon_vcgtz_v: 5249 case NEON::BI__builtin_neon_vcgtzq_v: 5250 return EmitAArch64CompareBuiltinExpr(Ops[0], Ty, ICmpInst::FCMP_OGT, 5251 ICmpInst::ICMP_SGT, "vcgtz"); 5252 case NEON::BI__builtin_neon_vcltz_v: 5253 case NEON::BI__builtin_neon_vcltzq_v: 5254 return EmitAArch64CompareBuiltinExpr(Ops[0], Ty, ICmpInst::FCMP_OLT, 5255 ICmpInst::ICMP_SLT, "vcltz"); 5256 case NEON::BI__builtin_neon_vcvt_f64_v: 5257 case NEON::BI__builtin_neon_vcvtq_f64_v: 5258 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 5259 Ty = GetNeonType(this, NeonTypeFlags(NeonTypeFlags::Float64, false, quad)); 5260 return usgn ? Builder.CreateUIToFP(Ops[0], Ty, "vcvt") 5261 : Builder.CreateSIToFP(Ops[0], Ty, "vcvt"); 5262 case NEON::BI__builtin_neon_vcvt_f64_f32: { 5263 assert(Type.getEltType() == NeonTypeFlags::Float64 && quad && 5264 "unexpected vcvt_f64_f32 builtin"); 5265 NeonTypeFlags SrcFlag = NeonTypeFlags(NeonTypeFlags::Float32, false, false); 5266 Ops[0] = Builder.CreateBitCast(Ops[0], GetNeonType(this, SrcFlag)); 5267 5268 return Builder.CreateFPExt(Ops[0], Ty, "vcvt"); 5269 } 5270 case NEON::BI__builtin_neon_vcvt_f32_f64: { 5271 assert(Type.getEltType() == NeonTypeFlags::Float32 && 5272 "unexpected vcvt_f32_f64 builtin"); 5273 NeonTypeFlags SrcFlag = NeonTypeFlags(NeonTypeFlags::Float64, false, true); 5274 Ops[0] = Builder.CreateBitCast(Ops[0], GetNeonType(this, SrcFlag)); 5275 5276 return Builder.CreateFPTrunc(Ops[0], Ty, "vcvt"); 5277 } 5278 case NEON::BI__builtin_neon_vcvt_s32_v: 5279 case NEON::BI__builtin_neon_vcvt_u32_v: 5280 case NEON::BI__builtin_neon_vcvt_s64_v: 5281 case NEON::BI__builtin_neon_vcvt_u64_v: 5282 case NEON::BI__builtin_neon_vcvtq_s32_v: 5283 case NEON::BI__builtin_neon_vcvtq_u32_v: 5284 case NEON::BI__builtin_neon_vcvtq_s64_v: 5285 case NEON::BI__builtin_neon_vcvtq_u64_v: { 5286 Ops[0] = Builder.CreateBitCast(Ops[0], GetFloatNeonType(this, Type)); 5287 if (usgn) 5288 return Builder.CreateFPToUI(Ops[0], Ty); 5289 return Builder.CreateFPToSI(Ops[0], Ty); 5290 } 5291 case NEON::BI__builtin_neon_vcvta_s32_v: 5292 case NEON::BI__builtin_neon_vcvtaq_s32_v: 5293 case NEON::BI__builtin_neon_vcvta_u32_v: 5294 case NEON::BI__builtin_neon_vcvtaq_u32_v: 5295 case NEON::BI__builtin_neon_vcvta_s64_v: 5296 case NEON::BI__builtin_neon_vcvtaq_s64_v: 5297 case NEON::BI__builtin_neon_vcvta_u64_v: 5298 case NEON::BI__builtin_neon_vcvtaq_u64_v: { 5299 Int = usgn ? Intrinsic::aarch64_neon_fcvtau : Intrinsic::aarch64_neon_fcvtas; 5300 llvm::Type *Tys[2] = { Ty, GetFloatNeonType(this, Type) }; 5301 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vcvta"); 5302 } 5303 case NEON::BI__builtin_neon_vcvtm_s32_v: 5304 case NEON::BI__builtin_neon_vcvtmq_s32_v: 5305 case NEON::BI__builtin_neon_vcvtm_u32_v: 5306 case NEON::BI__builtin_neon_vcvtmq_u32_v: 5307 case NEON::BI__builtin_neon_vcvtm_s64_v: 5308 case NEON::BI__builtin_neon_vcvtmq_s64_v: 5309 case NEON::BI__builtin_neon_vcvtm_u64_v: 5310 case NEON::BI__builtin_neon_vcvtmq_u64_v: { 5311 Int = usgn ? Intrinsic::aarch64_neon_fcvtmu : Intrinsic::aarch64_neon_fcvtms; 5312 llvm::Type *Tys[2] = { Ty, GetFloatNeonType(this, Type) }; 5313 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vcvtm"); 5314 } 5315 case NEON::BI__builtin_neon_vcvtn_s32_v: 5316 case NEON::BI__builtin_neon_vcvtnq_s32_v: 5317 case NEON::BI__builtin_neon_vcvtn_u32_v: 5318 case NEON::BI__builtin_neon_vcvtnq_u32_v: 5319 case NEON::BI__builtin_neon_vcvtn_s64_v: 5320 case NEON::BI__builtin_neon_vcvtnq_s64_v: 5321 case NEON::BI__builtin_neon_vcvtn_u64_v: 5322 case NEON::BI__builtin_neon_vcvtnq_u64_v: { 5323 Int = usgn ? Intrinsic::aarch64_neon_fcvtnu : Intrinsic::aarch64_neon_fcvtns; 5324 llvm::Type *Tys[2] = { Ty, GetFloatNeonType(this, Type) }; 5325 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vcvtn"); 5326 } 5327 case NEON::BI__builtin_neon_vcvtp_s32_v: 5328 case NEON::BI__builtin_neon_vcvtpq_s32_v: 5329 case NEON::BI__builtin_neon_vcvtp_u32_v: 5330 case NEON::BI__builtin_neon_vcvtpq_u32_v: 5331 case NEON::BI__builtin_neon_vcvtp_s64_v: 5332 case NEON::BI__builtin_neon_vcvtpq_s64_v: 5333 case NEON::BI__builtin_neon_vcvtp_u64_v: 5334 case NEON::BI__builtin_neon_vcvtpq_u64_v: { 5335 Int = usgn ? Intrinsic::aarch64_neon_fcvtpu : Intrinsic::aarch64_neon_fcvtps; 5336 llvm::Type *Tys[2] = { Ty, GetFloatNeonType(this, Type) }; 5337 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vcvtp"); 5338 } 5339 case NEON::BI__builtin_neon_vmulx_v: 5340 case NEON::BI__builtin_neon_vmulxq_v: { 5341 Int = Intrinsic::aarch64_neon_fmulx; 5342 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmulx"); 5343 } 5344 case NEON::BI__builtin_neon_vmul_lane_v: 5345 case NEON::BI__builtin_neon_vmul_laneq_v: { 5346 // v1f64 vmul_lane should be mapped to Neon scalar mul lane 5347 bool Quad = false; 5348 if (BuiltinID == NEON::BI__builtin_neon_vmul_laneq_v) 5349 Quad = true; 5350 Ops[0] = Builder.CreateBitCast(Ops[0], DoubleTy); 5351 llvm::Type *VTy = GetNeonType(this, 5352 NeonTypeFlags(NeonTypeFlags::Float64, false, Quad)); 5353 Ops[1] = Builder.CreateBitCast(Ops[1], VTy); 5354 Ops[1] = Builder.CreateExtractElement(Ops[1], Ops[2], "extract"); 5355 Value *Result = Builder.CreateFMul(Ops[0], Ops[1]); 5356 return Builder.CreateBitCast(Result, Ty); 5357 } 5358 case NEON::BI__builtin_neon_vnegd_s64: 5359 return Builder.CreateNeg(EmitScalarExpr(E->getArg(0)), "vnegd"); 5360 case NEON::BI__builtin_neon_vpmaxnm_v: 5361 case NEON::BI__builtin_neon_vpmaxnmq_v: { 5362 Int = Intrinsic::aarch64_neon_fmaxnmp; 5363 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vpmaxnm"); 5364 } 5365 case NEON::BI__builtin_neon_vpminnm_v: 5366 case NEON::BI__builtin_neon_vpminnmq_v: { 5367 Int = Intrinsic::aarch64_neon_fminnmp; 5368 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vpminnm"); 5369 } 5370 case NEON::BI__builtin_neon_vsqrt_v: 5371 case NEON::BI__builtin_neon_vsqrtq_v: { 5372 Int = Intrinsic::sqrt; 5373 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 5374 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vsqrt"); 5375 } 5376 case NEON::BI__builtin_neon_vrbit_v: 5377 case NEON::BI__builtin_neon_vrbitq_v: { 5378 Int = Intrinsic::aarch64_neon_rbit; 5379 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrbit"); 5380 } 5381 case NEON::BI__builtin_neon_vaddv_u8: 5382 // FIXME: These are handled by the AArch64 scalar code. 5383 usgn = true; 5384 // FALLTHROUGH 5385 case NEON::BI__builtin_neon_vaddv_s8: { 5386 Int = usgn ? Intrinsic::aarch64_neon_uaddv : Intrinsic::aarch64_neon_saddv; 5387 Ty = Int32Ty; 5388 VTy = llvm::VectorType::get(Int8Ty, 8); 5389 llvm::Type *Tys[2] = { Ty, VTy }; 5390 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5391 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddv"); 5392 return Builder.CreateTrunc(Ops[0], Int8Ty); 5393 } 5394 case NEON::BI__builtin_neon_vaddv_u16: 5395 usgn = true; 5396 // FALLTHROUGH 5397 case NEON::BI__builtin_neon_vaddv_s16: { 5398 Int = usgn ? Intrinsic::aarch64_neon_uaddv : Intrinsic::aarch64_neon_saddv; 5399 Ty = Int32Ty; 5400 VTy = llvm::VectorType::get(Int16Ty, 4); 5401 llvm::Type *Tys[2] = { Ty, VTy }; 5402 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5403 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddv"); 5404 return Builder.CreateTrunc(Ops[0], Int16Ty); 5405 } 5406 case NEON::BI__builtin_neon_vaddvq_u8: 5407 usgn = true; 5408 // FALLTHROUGH 5409 case NEON::BI__builtin_neon_vaddvq_s8: { 5410 Int = usgn ? Intrinsic::aarch64_neon_uaddv : Intrinsic::aarch64_neon_saddv; 5411 Ty = Int32Ty; 5412 VTy = llvm::VectorType::get(Int8Ty, 16); 5413 llvm::Type *Tys[2] = { Ty, VTy }; 5414 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5415 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddv"); 5416 return Builder.CreateTrunc(Ops[0], Int8Ty); 5417 } 5418 case NEON::BI__builtin_neon_vaddvq_u16: 5419 usgn = true; 5420 // FALLTHROUGH 5421 case NEON::BI__builtin_neon_vaddvq_s16: { 5422 Int = usgn ? Intrinsic::aarch64_neon_uaddv : Intrinsic::aarch64_neon_saddv; 5423 Ty = Int32Ty; 5424 VTy = llvm::VectorType::get(Int16Ty, 8); 5425 llvm::Type *Tys[2] = { Ty, VTy }; 5426 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5427 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddv"); 5428 return Builder.CreateTrunc(Ops[0], Int16Ty); 5429 } 5430 case NEON::BI__builtin_neon_vmaxv_u8: { 5431 Int = Intrinsic::aarch64_neon_umaxv; 5432 Ty = Int32Ty; 5433 VTy = llvm::VectorType::get(Int8Ty, 8); 5434 llvm::Type *Tys[2] = { Ty, VTy }; 5435 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5436 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv"); 5437 return Builder.CreateTrunc(Ops[0], Int8Ty); 5438 } 5439 case NEON::BI__builtin_neon_vmaxv_u16: { 5440 Int = Intrinsic::aarch64_neon_umaxv; 5441 Ty = Int32Ty; 5442 VTy = llvm::VectorType::get(Int16Ty, 4); 5443 llvm::Type *Tys[2] = { Ty, VTy }; 5444 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5445 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv"); 5446 return Builder.CreateTrunc(Ops[0], Int16Ty); 5447 } 5448 case NEON::BI__builtin_neon_vmaxvq_u8: { 5449 Int = Intrinsic::aarch64_neon_umaxv; 5450 Ty = Int32Ty; 5451 VTy = llvm::VectorType::get(Int8Ty, 16); 5452 llvm::Type *Tys[2] = { Ty, VTy }; 5453 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5454 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv"); 5455 return Builder.CreateTrunc(Ops[0], Int8Ty); 5456 } 5457 case NEON::BI__builtin_neon_vmaxvq_u16: { 5458 Int = Intrinsic::aarch64_neon_umaxv; 5459 Ty = Int32Ty; 5460 VTy = llvm::VectorType::get(Int16Ty, 8); 5461 llvm::Type *Tys[2] = { Ty, VTy }; 5462 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5463 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv"); 5464 return Builder.CreateTrunc(Ops[0], Int16Ty); 5465 } 5466 case NEON::BI__builtin_neon_vmaxv_s8: { 5467 Int = Intrinsic::aarch64_neon_smaxv; 5468 Ty = Int32Ty; 5469 VTy = llvm::VectorType::get(Int8Ty, 8); 5470 llvm::Type *Tys[2] = { Ty, VTy }; 5471 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5472 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv"); 5473 return Builder.CreateTrunc(Ops[0], Int8Ty); 5474 } 5475 case NEON::BI__builtin_neon_vmaxv_s16: { 5476 Int = Intrinsic::aarch64_neon_smaxv; 5477 Ty = Int32Ty; 5478 VTy = llvm::VectorType::get(Int16Ty, 4); 5479 llvm::Type *Tys[2] = { Ty, VTy }; 5480 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5481 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv"); 5482 return Builder.CreateTrunc(Ops[0], Int16Ty); 5483 } 5484 case NEON::BI__builtin_neon_vmaxvq_s8: { 5485 Int = Intrinsic::aarch64_neon_smaxv; 5486 Ty = Int32Ty; 5487 VTy = llvm::VectorType::get(Int8Ty, 16); 5488 llvm::Type *Tys[2] = { Ty, VTy }; 5489 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5490 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv"); 5491 return Builder.CreateTrunc(Ops[0], Int8Ty); 5492 } 5493 case NEON::BI__builtin_neon_vmaxvq_s16: { 5494 Int = Intrinsic::aarch64_neon_smaxv; 5495 Ty = Int32Ty; 5496 VTy = llvm::VectorType::get(Int16Ty, 8); 5497 llvm::Type *Tys[2] = { Ty, VTy }; 5498 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5499 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv"); 5500 return Builder.CreateTrunc(Ops[0], Int16Ty); 5501 } 5502 case NEON::BI__builtin_neon_vminv_u8: { 5503 Int = Intrinsic::aarch64_neon_uminv; 5504 Ty = Int32Ty; 5505 VTy = llvm::VectorType::get(Int8Ty, 8); 5506 llvm::Type *Tys[2] = { Ty, VTy }; 5507 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5508 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv"); 5509 return Builder.CreateTrunc(Ops[0], Int8Ty); 5510 } 5511 case NEON::BI__builtin_neon_vminv_u16: { 5512 Int = Intrinsic::aarch64_neon_uminv; 5513 Ty = Int32Ty; 5514 VTy = llvm::VectorType::get(Int16Ty, 4); 5515 llvm::Type *Tys[2] = { Ty, VTy }; 5516 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5517 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv"); 5518 return Builder.CreateTrunc(Ops[0], Int16Ty); 5519 } 5520 case NEON::BI__builtin_neon_vminvq_u8: { 5521 Int = Intrinsic::aarch64_neon_uminv; 5522 Ty = Int32Ty; 5523 VTy = llvm::VectorType::get(Int8Ty, 16); 5524 llvm::Type *Tys[2] = { Ty, VTy }; 5525 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5526 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv"); 5527 return Builder.CreateTrunc(Ops[0], Int8Ty); 5528 } 5529 case NEON::BI__builtin_neon_vminvq_u16: { 5530 Int = Intrinsic::aarch64_neon_uminv; 5531 Ty = Int32Ty; 5532 VTy = llvm::VectorType::get(Int16Ty, 8); 5533 llvm::Type *Tys[2] = { Ty, VTy }; 5534 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5535 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv"); 5536 return Builder.CreateTrunc(Ops[0], Int16Ty); 5537 } 5538 case NEON::BI__builtin_neon_vminv_s8: { 5539 Int = Intrinsic::aarch64_neon_sminv; 5540 Ty = Int32Ty; 5541 VTy = llvm::VectorType::get(Int8Ty, 8); 5542 llvm::Type *Tys[2] = { Ty, VTy }; 5543 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5544 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv"); 5545 return Builder.CreateTrunc(Ops[0], Int8Ty); 5546 } 5547 case NEON::BI__builtin_neon_vminv_s16: { 5548 Int = Intrinsic::aarch64_neon_sminv; 5549 Ty = Int32Ty; 5550 VTy = llvm::VectorType::get(Int16Ty, 4); 5551 llvm::Type *Tys[2] = { Ty, VTy }; 5552 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5553 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv"); 5554 return Builder.CreateTrunc(Ops[0], Int16Ty); 5555 } 5556 case NEON::BI__builtin_neon_vminvq_s8: { 5557 Int = Intrinsic::aarch64_neon_sminv; 5558 Ty = Int32Ty; 5559 VTy = llvm::VectorType::get(Int8Ty, 16); 5560 llvm::Type *Tys[2] = { Ty, VTy }; 5561 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5562 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv"); 5563 return Builder.CreateTrunc(Ops[0], Int8Ty); 5564 } 5565 case NEON::BI__builtin_neon_vminvq_s16: { 5566 Int = Intrinsic::aarch64_neon_sminv; 5567 Ty = Int32Ty; 5568 VTy = llvm::VectorType::get(Int16Ty, 8); 5569 llvm::Type *Tys[2] = { Ty, VTy }; 5570 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5571 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv"); 5572 return Builder.CreateTrunc(Ops[0], Int16Ty); 5573 } 5574 case NEON::BI__builtin_neon_vmul_n_f64: { 5575 Ops[0] = Builder.CreateBitCast(Ops[0], DoubleTy); 5576 Value *RHS = Builder.CreateBitCast(EmitScalarExpr(E->getArg(1)), DoubleTy); 5577 return Builder.CreateFMul(Ops[0], RHS); 5578 } 5579 case NEON::BI__builtin_neon_vaddlv_u8: { 5580 Int = Intrinsic::aarch64_neon_uaddlv; 5581 Ty = Int32Ty; 5582 VTy = llvm::VectorType::get(Int8Ty, 8); 5583 llvm::Type *Tys[2] = { Ty, VTy }; 5584 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5585 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv"); 5586 return Builder.CreateTrunc(Ops[0], Int16Ty); 5587 } 5588 case NEON::BI__builtin_neon_vaddlv_u16: { 5589 Int = Intrinsic::aarch64_neon_uaddlv; 5590 Ty = Int32Ty; 5591 VTy = llvm::VectorType::get(Int16Ty, 4); 5592 llvm::Type *Tys[2] = { Ty, VTy }; 5593 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5594 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv"); 5595 } 5596 case NEON::BI__builtin_neon_vaddlvq_u8: { 5597 Int = Intrinsic::aarch64_neon_uaddlv; 5598 Ty = Int32Ty; 5599 VTy = llvm::VectorType::get(Int8Ty, 16); 5600 llvm::Type *Tys[2] = { Ty, VTy }; 5601 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5602 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv"); 5603 return Builder.CreateTrunc(Ops[0], Int16Ty); 5604 } 5605 case NEON::BI__builtin_neon_vaddlvq_u16: { 5606 Int = Intrinsic::aarch64_neon_uaddlv; 5607 Ty = Int32Ty; 5608 VTy = llvm::VectorType::get(Int16Ty, 8); 5609 llvm::Type *Tys[2] = { Ty, VTy }; 5610 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5611 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv"); 5612 } 5613 case NEON::BI__builtin_neon_vaddlv_s8: { 5614 Int = Intrinsic::aarch64_neon_saddlv; 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, "vaddlv"); 5620 return Builder.CreateTrunc(Ops[0], Int16Ty); 5621 } 5622 case NEON::BI__builtin_neon_vaddlv_s16: { 5623 Int = Intrinsic::aarch64_neon_saddlv; 5624 Ty = Int32Ty; 5625 VTy = llvm::VectorType::get(Int16Ty, 4); 5626 llvm::Type *Tys[2] = { Ty, VTy }; 5627 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5628 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv"); 5629 } 5630 case NEON::BI__builtin_neon_vaddlvq_s8: { 5631 Int = Intrinsic::aarch64_neon_saddlv; 5632 Ty = Int32Ty; 5633 VTy = llvm::VectorType::get(Int8Ty, 16); 5634 llvm::Type *Tys[2] = { Ty, VTy }; 5635 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5636 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv"); 5637 return Builder.CreateTrunc(Ops[0], Int16Ty); 5638 } 5639 case NEON::BI__builtin_neon_vaddlvq_s16: { 5640 Int = Intrinsic::aarch64_neon_saddlv; 5641 Ty = Int32Ty; 5642 VTy = llvm::VectorType::get(Int16Ty, 8); 5643 llvm::Type *Tys[2] = { Ty, VTy }; 5644 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5645 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv"); 5646 } 5647 case NEON::BI__builtin_neon_vsri_n_v: 5648 case NEON::BI__builtin_neon_vsriq_n_v: { 5649 Int = Intrinsic::aarch64_neon_vsri; 5650 llvm::Function *Intrin = CGM.getIntrinsic(Int, Ty); 5651 return EmitNeonCall(Intrin, Ops, "vsri_n"); 5652 } 5653 case NEON::BI__builtin_neon_vsli_n_v: 5654 case NEON::BI__builtin_neon_vsliq_n_v: { 5655 Int = Intrinsic::aarch64_neon_vsli; 5656 llvm::Function *Intrin = CGM.getIntrinsic(Int, Ty); 5657 return EmitNeonCall(Intrin, Ops, "vsli_n"); 5658 } 5659 case NEON::BI__builtin_neon_vsra_n_v: 5660 case NEON::BI__builtin_neon_vsraq_n_v: 5661 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 5662 Ops[1] = EmitNeonRShiftImm(Ops[1], Ops[2], Ty, usgn, "vsra_n"); 5663 return Builder.CreateAdd(Ops[0], Ops[1]); 5664 case NEON::BI__builtin_neon_vrsra_n_v: 5665 case NEON::BI__builtin_neon_vrsraq_n_v: { 5666 Int = usgn ? Intrinsic::aarch64_neon_urshl : Intrinsic::aarch64_neon_srshl; 5667 SmallVector<llvm::Value*,2> TmpOps; 5668 TmpOps.push_back(Ops[1]); 5669 TmpOps.push_back(Ops[2]); 5670 Function* F = CGM.getIntrinsic(Int, Ty); 5671 llvm::Value *tmp = EmitNeonCall(F, TmpOps, "vrshr_n", 1, true); 5672 Ops[0] = Builder.CreateBitCast(Ops[0], VTy); 5673 return Builder.CreateAdd(Ops[0], tmp); 5674 } 5675 // FIXME: Sharing loads & stores with 32-bit is complicated by the absence 5676 // of an Align parameter here. 5677 case NEON::BI__builtin_neon_vld1_x2_v: 5678 case NEON::BI__builtin_neon_vld1q_x2_v: 5679 case NEON::BI__builtin_neon_vld1_x3_v: 5680 case NEON::BI__builtin_neon_vld1q_x3_v: 5681 case NEON::BI__builtin_neon_vld1_x4_v: 5682 case NEON::BI__builtin_neon_vld1q_x4_v: { 5683 llvm::Type *PTy = llvm::PointerType::getUnqual(VTy->getVectorElementType()); 5684 Ops[1] = Builder.CreateBitCast(Ops[1], PTy); 5685 llvm::Type *Tys[2] = { VTy, PTy }; 5686 unsigned Int; 5687 switch (BuiltinID) { 5688 case NEON::BI__builtin_neon_vld1_x2_v: 5689 case NEON::BI__builtin_neon_vld1q_x2_v: 5690 Int = Intrinsic::aarch64_neon_ld1x2; 5691 break; 5692 case NEON::BI__builtin_neon_vld1_x3_v: 5693 case NEON::BI__builtin_neon_vld1q_x3_v: 5694 Int = Intrinsic::aarch64_neon_ld1x3; 5695 break; 5696 case NEON::BI__builtin_neon_vld1_x4_v: 5697 case NEON::BI__builtin_neon_vld1q_x4_v: 5698 Int = Intrinsic::aarch64_neon_ld1x4; 5699 break; 5700 } 5701 Function *F = CGM.getIntrinsic(Int, Tys); 5702 Ops[1] = Builder.CreateCall(F, Ops[1], "vld1xN"); 5703 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 5704 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 5705 return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]); 5706 } 5707 case NEON::BI__builtin_neon_vst1_x2_v: 5708 case NEON::BI__builtin_neon_vst1q_x2_v: 5709 case NEON::BI__builtin_neon_vst1_x3_v: 5710 case NEON::BI__builtin_neon_vst1q_x3_v: 5711 case NEON::BI__builtin_neon_vst1_x4_v: 5712 case NEON::BI__builtin_neon_vst1q_x4_v: { 5713 llvm::Type *PTy = llvm::PointerType::getUnqual(VTy->getVectorElementType()); 5714 llvm::Type *Tys[2] = { VTy, PTy }; 5715 unsigned Int; 5716 switch (BuiltinID) { 5717 case NEON::BI__builtin_neon_vst1_x2_v: 5718 case NEON::BI__builtin_neon_vst1q_x2_v: 5719 Int = Intrinsic::aarch64_neon_st1x2; 5720 break; 5721 case NEON::BI__builtin_neon_vst1_x3_v: 5722 case NEON::BI__builtin_neon_vst1q_x3_v: 5723 Int = Intrinsic::aarch64_neon_st1x3; 5724 break; 5725 case NEON::BI__builtin_neon_vst1_x4_v: 5726 case NEON::BI__builtin_neon_vst1q_x4_v: 5727 Int = Intrinsic::aarch64_neon_st1x4; 5728 break; 5729 } 5730 std::rotate(Ops.begin(), Ops.begin() + 1, Ops.end()); 5731 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, ""); 5732 } 5733 case NEON::BI__builtin_neon_vld1_v: 5734 case NEON::BI__builtin_neon_vld1q_v: 5735 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(VTy)); 5736 return Builder.CreateDefaultAlignedLoad(Ops[0]); 5737 case NEON::BI__builtin_neon_vst1_v: 5738 case NEON::BI__builtin_neon_vst1q_v: 5739 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(VTy)); 5740 Ops[1] = Builder.CreateBitCast(Ops[1], VTy); 5741 return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]); 5742 case NEON::BI__builtin_neon_vld1_lane_v: 5743 case NEON::BI__builtin_neon_vld1q_lane_v: 5744 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 5745 Ty = llvm::PointerType::getUnqual(VTy->getElementType()); 5746 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 5747 Ops[0] = Builder.CreateDefaultAlignedLoad(Ops[0]); 5748 return Builder.CreateInsertElement(Ops[1], Ops[0], Ops[2], "vld1_lane"); 5749 case NEON::BI__builtin_neon_vld1_dup_v: 5750 case NEON::BI__builtin_neon_vld1q_dup_v: { 5751 Value *V = UndefValue::get(Ty); 5752 Ty = llvm::PointerType::getUnqual(VTy->getElementType()); 5753 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 5754 Ops[0] = Builder.CreateDefaultAlignedLoad(Ops[0]); 5755 llvm::Constant *CI = ConstantInt::get(Int32Ty, 0); 5756 Ops[0] = Builder.CreateInsertElement(V, Ops[0], CI); 5757 return EmitNeonSplat(Ops[0], CI); 5758 } 5759 case NEON::BI__builtin_neon_vst1_lane_v: 5760 case NEON::BI__builtin_neon_vst1q_lane_v: 5761 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 5762 Ops[1] = Builder.CreateExtractElement(Ops[1], Ops[2]); 5763 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 5764 return Builder.CreateDefaultAlignedStore(Ops[1], 5765 Builder.CreateBitCast(Ops[0], Ty)); 5766 case NEON::BI__builtin_neon_vld2_v: 5767 case NEON::BI__builtin_neon_vld2q_v: { 5768 llvm::Type *PTy = llvm::PointerType::getUnqual(VTy); 5769 Ops[1] = Builder.CreateBitCast(Ops[1], PTy); 5770 llvm::Type *Tys[2] = { VTy, PTy }; 5771 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld2, Tys); 5772 Ops[1] = Builder.CreateCall(F, Ops[1], "vld2"); 5773 Ops[0] = Builder.CreateBitCast(Ops[0], 5774 llvm::PointerType::getUnqual(Ops[1]->getType())); 5775 return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]); 5776 } 5777 case NEON::BI__builtin_neon_vld3_v: 5778 case NEON::BI__builtin_neon_vld3q_v: { 5779 llvm::Type *PTy = llvm::PointerType::getUnqual(VTy); 5780 Ops[1] = Builder.CreateBitCast(Ops[1], PTy); 5781 llvm::Type *Tys[2] = { VTy, PTy }; 5782 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld3, Tys); 5783 Ops[1] = Builder.CreateCall(F, Ops[1], "vld3"); 5784 Ops[0] = Builder.CreateBitCast(Ops[0], 5785 llvm::PointerType::getUnqual(Ops[1]->getType())); 5786 return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]); 5787 } 5788 case NEON::BI__builtin_neon_vld4_v: 5789 case NEON::BI__builtin_neon_vld4q_v: { 5790 llvm::Type *PTy = llvm::PointerType::getUnqual(VTy); 5791 Ops[1] = Builder.CreateBitCast(Ops[1], PTy); 5792 llvm::Type *Tys[2] = { VTy, PTy }; 5793 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld4, Tys); 5794 Ops[1] = Builder.CreateCall(F, Ops[1], "vld4"); 5795 Ops[0] = Builder.CreateBitCast(Ops[0], 5796 llvm::PointerType::getUnqual(Ops[1]->getType())); 5797 return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]); 5798 } 5799 case NEON::BI__builtin_neon_vld2_dup_v: 5800 case NEON::BI__builtin_neon_vld2q_dup_v: { 5801 llvm::Type *PTy = 5802 llvm::PointerType::getUnqual(VTy->getElementType()); 5803 Ops[1] = Builder.CreateBitCast(Ops[1], PTy); 5804 llvm::Type *Tys[2] = { VTy, PTy }; 5805 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld2r, Tys); 5806 Ops[1] = Builder.CreateCall(F, Ops[1], "vld2"); 5807 Ops[0] = Builder.CreateBitCast(Ops[0], 5808 llvm::PointerType::getUnqual(Ops[1]->getType())); 5809 return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]); 5810 } 5811 case NEON::BI__builtin_neon_vld3_dup_v: 5812 case NEON::BI__builtin_neon_vld3q_dup_v: { 5813 llvm::Type *PTy = 5814 llvm::PointerType::getUnqual(VTy->getElementType()); 5815 Ops[1] = Builder.CreateBitCast(Ops[1], PTy); 5816 llvm::Type *Tys[2] = { VTy, PTy }; 5817 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld3r, Tys); 5818 Ops[1] = Builder.CreateCall(F, Ops[1], "vld3"); 5819 Ops[0] = Builder.CreateBitCast(Ops[0], 5820 llvm::PointerType::getUnqual(Ops[1]->getType())); 5821 return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]); 5822 } 5823 case NEON::BI__builtin_neon_vld4_dup_v: 5824 case NEON::BI__builtin_neon_vld4q_dup_v: { 5825 llvm::Type *PTy = 5826 llvm::PointerType::getUnqual(VTy->getElementType()); 5827 Ops[1] = Builder.CreateBitCast(Ops[1], PTy); 5828 llvm::Type *Tys[2] = { VTy, PTy }; 5829 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld4r, Tys); 5830 Ops[1] = Builder.CreateCall(F, Ops[1], "vld4"); 5831 Ops[0] = Builder.CreateBitCast(Ops[0], 5832 llvm::PointerType::getUnqual(Ops[1]->getType())); 5833 return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]); 5834 } 5835 case NEON::BI__builtin_neon_vld2_lane_v: 5836 case NEON::BI__builtin_neon_vld2q_lane_v: { 5837 llvm::Type *Tys[2] = { VTy, Ops[1]->getType() }; 5838 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld2lane, Tys); 5839 Ops.push_back(Ops[1]); 5840 Ops.erase(Ops.begin()+1); 5841 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 5842 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 5843 Ops[3] = Builder.CreateZExt(Ops[3], Int64Ty); 5844 Ops[1] = Builder.CreateCall(F, makeArrayRef(Ops).slice(1), "vld2_lane"); 5845 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 5846 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 5847 return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]); 5848 } 5849 case NEON::BI__builtin_neon_vld3_lane_v: 5850 case NEON::BI__builtin_neon_vld3q_lane_v: { 5851 llvm::Type *Tys[2] = { VTy, Ops[1]->getType() }; 5852 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld3lane, Tys); 5853 Ops.push_back(Ops[1]); 5854 Ops.erase(Ops.begin()+1); 5855 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 5856 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 5857 Ops[3] = Builder.CreateBitCast(Ops[3], Ty); 5858 Ops[4] = Builder.CreateZExt(Ops[4], Int64Ty); 5859 Ops[1] = Builder.CreateCall(F, makeArrayRef(Ops).slice(1), "vld3_lane"); 5860 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 5861 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 5862 return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]); 5863 } 5864 case NEON::BI__builtin_neon_vld4_lane_v: 5865 case NEON::BI__builtin_neon_vld4q_lane_v: { 5866 llvm::Type *Tys[2] = { VTy, Ops[1]->getType() }; 5867 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld4lane, Tys); 5868 Ops.push_back(Ops[1]); 5869 Ops.erase(Ops.begin()+1); 5870 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 5871 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 5872 Ops[3] = Builder.CreateBitCast(Ops[3], Ty); 5873 Ops[4] = Builder.CreateBitCast(Ops[4], Ty); 5874 Ops[5] = Builder.CreateZExt(Ops[5], Int64Ty); 5875 Ops[1] = Builder.CreateCall(F, makeArrayRef(Ops).slice(1), "vld4_lane"); 5876 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 5877 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 5878 return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]); 5879 } 5880 case NEON::BI__builtin_neon_vst2_v: 5881 case NEON::BI__builtin_neon_vst2q_v: { 5882 Ops.push_back(Ops[0]); 5883 Ops.erase(Ops.begin()); 5884 llvm::Type *Tys[2] = { VTy, Ops[2]->getType() }; 5885 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_st2, Tys), 5886 Ops, ""); 5887 } 5888 case NEON::BI__builtin_neon_vst2_lane_v: 5889 case NEON::BI__builtin_neon_vst2q_lane_v: { 5890 Ops.push_back(Ops[0]); 5891 Ops.erase(Ops.begin()); 5892 Ops[2] = Builder.CreateZExt(Ops[2], Int64Ty); 5893 llvm::Type *Tys[2] = { VTy, Ops[3]->getType() }; 5894 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_st2lane, Tys), 5895 Ops, ""); 5896 } 5897 case NEON::BI__builtin_neon_vst3_v: 5898 case NEON::BI__builtin_neon_vst3q_v: { 5899 Ops.push_back(Ops[0]); 5900 Ops.erase(Ops.begin()); 5901 llvm::Type *Tys[2] = { VTy, Ops[3]->getType() }; 5902 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_st3, Tys), 5903 Ops, ""); 5904 } 5905 case NEON::BI__builtin_neon_vst3_lane_v: 5906 case NEON::BI__builtin_neon_vst3q_lane_v: { 5907 Ops.push_back(Ops[0]); 5908 Ops.erase(Ops.begin()); 5909 Ops[3] = Builder.CreateZExt(Ops[3], Int64Ty); 5910 llvm::Type *Tys[2] = { VTy, Ops[4]->getType() }; 5911 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_st3lane, Tys), 5912 Ops, ""); 5913 } 5914 case NEON::BI__builtin_neon_vst4_v: 5915 case NEON::BI__builtin_neon_vst4q_v: { 5916 Ops.push_back(Ops[0]); 5917 Ops.erase(Ops.begin()); 5918 llvm::Type *Tys[2] = { VTy, Ops[4]->getType() }; 5919 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_st4, Tys), 5920 Ops, ""); 5921 } 5922 case NEON::BI__builtin_neon_vst4_lane_v: 5923 case NEON::BI__builtin_neon_vst4q_lane_v: { 5924 Ops.push_back(Ops[0]); 5925 Ops.erase(Ops.begin()); 5926 Ops[4] = Builder.CreateZExt(Ops[4], Int64Ty); 5927 llvm::Type *Tys[2] = { VTy, Ops[5]->getType() }; 5928 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_st4lane, Tys), 5929 Ops, ""); 5930 } 5931 case NEON::BI__builtin_neon_vtrn_v: 5932 case NEON::BI__builtin_neon_vtrnq_v: { 5933 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty)); 5934 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 5935 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 5936 Value *SV = nullptr; 5937 5938 for (unsigned vi = 0; vi != 2; ++vi) { 5939 SmallVector<Constant*, 16> Indices; 5940 for (unsigned i = 0, e = VTy->getNumElements(); i != e; i += 2) { 5941 Indices.push_back(ConstantInt::get(Int32Ty, i+vi)); 5942 Indices.push_back(ConstantInt::get(Int32Ty, i+e+vi)); 5943 } 5944 Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ty, Ops[0], vi); 5945 SV = llvm::ConstantVector::get(Indices); 5946 SV = Builder.CreateShuffleVector(Ops[1], Ops[2], SV, "vtrn"); 5947 SV = Builder.CreateDefaultAlignedStore(SV, Addr); 5948 } 5949 return SV; 5950 } 5951 case NEON::BI__builtin_neon_vuzp_v: 5952 case NEON::BI__builtin_neon_vuzpq_v: { 5953 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty)); 5954 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 5955 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 5956 Value *SV = nullptr; 5957 5958 for (unsigned vi = 0; vi != 2; ++vi) { 5959 SmallVector<Constant*, 16> Indices; 5960 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) 5961 Indices.push_back(ConstantInt::get(Int32Ty, 2*i+vi)); 5962 5963 Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ty, Ops[0], vi); 5964 SV = llvm::ConstantVector::get(Indices); 5965 SV = Builder.CreateShuffleVector(Ops[1], Ops[2], SV, "vuzp"); 5966 SV = Builder.CreateDefaultAlignedStore(SV, Addr); 5967 } 5968 return SV; 5969 } 5970 case NEON::BI__builtin_neon_vzip_v: 5971 case NEON::BI__builtin_neon_vzipq_v: { 5972 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty)); 5973 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 5974 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 5975 Value *SV = nullptr; 5976 5977 for (unsigned vi = 0; vi != 2; ++vi) { 5978 SmallVector<Constant*, 16> Indices; 5979 for (unsigned i = 0, e = VTy->getNumElements(); i != e; i += 2) { 5980 Indices.push_back(ConstantInt::get(Int32Ty, (i + vi*e) >> 1)); 5981 Indices.push_back(ConstantInt::get(Int32Ty, ((i + vi*e) >> 1)+e)); 5982 } 5983 Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ty, Ops[0], vi); 5984 SV = llvm::ConstantVector::get(Indices); 5985 SV = Builder.CreateShuffleVector(Ops[1], Ops[2], SV, "vzip"); 5986 SV = Builder.CreateDefaultAlignedStore(SV, Addr); 5987 } 5988 return SV; 5989 } 5990 case NEON::BI__builtin_neon_vqtbl1q_v: { 5991 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbl1, Ty), 5992 Ops, "vtbl1"); 5993 } 5994 case NEON::BI__builtin_neon_vqtbl2q_v: { 5995 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbl2, Ty), 5996 Ops, "vtbl2"); 5997 } 5998 case NEON::BI__builtin_neon_vqtbl3q_v: { 5999 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbl3, Ty), 6000 Ops, "vtbl3"); 6001 } 6002 case NEON::BI__builtin_neon_vqtbl4q_v: { 6003 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbl4, Ty), 6004 Ops, "vtbl4"); 6005 } 6006 case NEON::BI__builtin_neon_vqtbx1q_v: { 6007 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbx1, Ty), 6008 Ops, "vtbx1"); 6009 } 6010 case NEON::BI__builtin_neon_vqtbx2q_v: { 6011 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbx2, Ty), 6012 Ops, "vtbx2"); 6013 } 6014 case NEON::BI__builtin_neon_vqtbx3q_v: { 6015 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbx3, Ty), 6016 Ops, "vtbx3"); 6017 } 6018 case NEON::BI__builtin_neon_vqtbx4q_v: { 6019 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbx4, Ty), 6020 Ops, "vtbx4"); 6021 } 6022 case NEON::BI__builtin_neon_vsqadd_v: 6023 case NEON::BI__builtin_neon_vsqaddq_v: { 6024 Int = Intrinsic::aarch64_neon_usqadd; 6025 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vsqadd"); 6026 } 6027 case NEON::BI__builtin_neon_vuqadd_v: 6028 case NEON::BI__builtin_neon_vuqaddq_v: { 6029 Int = Intrinsic::aarch64_neon_suqadd; 6030 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vuqadd"); 6031 } 6032 } 6033 } 6034 6035 llvm::Value *CodeGenFunction:: 6036 BuildVector(ArrayRef<llvm::Value*> Ops) { 6037 assert((Ops.size() & (Ops.size() - 1)) == 0 && 6038 "Not a power-of-two sized vector!"); 6039 bool AllConstants = true; 6040 for (unsigned i = 0, e = Ops.size(); i != e && AllConstants; ++i) 6041 AllConstants &= isa<Constant>(Ops[i]); 6042 6043 // If this is a constant vector, create a ConstantVector. 6044 if (AllConstants) { 6045 SmallVector<llvm::Constant*, 16> CstOps; 6046 for (unsigned i = 0, e = Ops.size(); i != e; ++i) 6047 CstOps.push_back(cast<Constant>(Ops[i])); 6048 return llvm::ConstantVector::get(CstOps); 6049 } 6050 6051 // Otherwise, insertelement the values to build the vector. 6052 Value *Result = 6053 llvm::UndefValue::get(llvm::VectorType::get(Ops[0]->getType(), Ops.size())); 6054 6055 for (unsigned i = 0, e = Ops.size(); i != e; ++i) 6056 Result = Builder.CreateInsertElement(Result, Ops[i], Builder.getInt32(i)); 6057 6058 return Result; 6059 } 6060 6061 Value *CodeGenFunction::EmitX86BuiltinExpr(unsigned BuiltinID, 6062 const CallExpr *E) { 6063 if (BuiltinID == X86::BI__builtin_ms_va_start || 6064 BuiltinID == X86::BI__builtin_ms_va_end) 6065 return EmitVAStartEnd(EmitMSVAListRef(E->getArg(0)).getPointer(), 6066 BuiltinID == X86::BI__builtin_ms_va_start); 6067 if (BuiltinID == X86::BI__builtin_ms_va_copy) { 6068 // Lower this manually. We can't reliably determine whether or not any 6069 // given va_copy() is for a Win64 va_list from the calling convention 6070 // alone, because it's legal to do this from a System V ABI function. 6071 // With opaque pointer types, we won't have enough information in LLVM 6072 // IR to determine this from the argument types, either. Best to do it 6073 // now, while we have enough information. 6074 Address DestAddr = EmitMSVAListRef(E->getArg(0)); 6075 Address SrcAddr = EmitMSVAListRef(E->getArg(1)); 6076 6077 llvm::Type *BPP = Int8PtrPtrTy; 6078 6079 DestAddr = Address(Builder.CreateBitCast(DestAddr.getPointer(), BPP, "cp"), 6080 DestAddr.getAlignment()); 6081 SrcAddr = Address(Builder.CreateBitCast(SrcAddr.getPointer(), BPP, "ap"), 6082 SrcAddr.getAlignment()); 6083 6084 Value *ArgPtr = Builder.CreateLoad(SrcAddr, "ap.val"); 6085 return Builder.CreateStore(ArgPtr, DestAddr); 6086 } 6087 6088 SmallVector<Value*, 4> Ops; 6089 6090 // Find out if any arguments are required to be integer constant expressions. 6091 unsigned ICEArguments = 0; 6092 ASTContext::GetBuiltinTypeError Error; 6093 getContext().GetBuiltinType(BuiltinID, Error, &ICEArguments); 6094 assert(Error == ASTContext::GE_None && "Should not codegen an error"); 6095 6096 for (unsigned i = 0, e = E->getNumArgs(); i != e; i++) { 6097 // If this is a normal argument, just emit it as a scalar. 6098 if ((ICEArguments & (1 << i)) == 0) { 6099 Ops.push_back(EmitScalarExpr(E->getArg(i))); 6100 continue; 6101 } 6102 6103 // If this is required to be a constant, constant fold it so that we know 6104 // that the generated intrinsic gets a ConstantInt. 6105 llvm::APSInt Result; 6106 bool IsConst = E->getArg(i)->isIntegerConstantExpr(Result, getContext()); 6107 assert(IsConst && "Constant arg isn't actually constant?"); (void)IsConst; 6108 Ops.push_back(llvm::ConstantInt::get(getLLVMContext(), Result)); 6109 } 6110 6111 switch (BuiltinID) { 6112 default: return nullptr; 6113 case X86::BI__builtin_cpu_supports: { 6114 const Expr *FeatureExpr = E->getArg(0)->IgnoreParenCasts(); 6115 StringRef FeatureStr = cast<StringLiteral>(FeatureExpr)->getString(); 6116 6117 // TODO: When/if this becomes more than x86 specific then use a TargetInfo 6118 // based mapping. 6119 // Processor features and mapping to processor feature value. 6120 enum X86Features { 6121 CMOV = 0, 6122 MMX, 6123 POPCNT, 6124 SSE, 6125 SSE2, 6126 SSE3, 6127 SSSE3, 6128 SSE4_1, 6129 SSE4_2, 6130 AVX, 6131 AVX2, 6132 SSE4_A, 6133 FMA4, 6134 XOP, 6135 FMA, 6136 AVX512F, 6137 BMI, 6138 BMI2, 6139 MAX 6140 }; 6141 6142 X86Features Feature = StringSwitch<X86Features>(FeatureStr) 6143 .Case("cmov", X86Features::CMOV) 6144 .Case("mmx", X86Features::MMX) 6145 .Case("popcnt", X86Features::POPCNT) 6146 .Case("sse", X86Features::SSE) 6147 .Case("sse2", X86Features::SSE2) 6148 .Case("sse3", X86Features::SSE3) 6149 .Case("sse4.1", X86Features::SSE4_1) 6150 .Case("sse4.2", X86Features::SSE4_2) 6151 .Case("avx", X86Features::AVX) 6152 .Case("avx2", X86Features::AVX2) 6153 .Case("sse4a", X86Features::SSE4_A) 6154 .Case("fma4", X86Features::FMA4) 6155 .Case("xop", X86Features::XOP) 6156 .Case("fma", X86Features::FMA) 6157 .Case("avx512f", X86Features::AVX512F) 6158 .Case("bmi", X86Features::BMI) 6159 .Case("bmi2", X86Features::BMI2) 6160 .Default(X86Features::MAX); 6161 assert(Feature != X86Features::MAX && "Invalid feature!"); 6162 6163 // Matching the struct layout from the compiler-rt/libgcc structure that is 6164 // filled in: 6165 // unsigned int __cpu_vendor; 6166 // unsigned int __cpu_type; 6167 // unsigned int __cpu_subtype; 6168 // unsigned int __cpu_features[1]; 6169 llvm::Type *STy = llvm::StructType::get( 6170 Int32Ty, Int32Ty, Int32Ty, llvm::ArrayType::get(Int32Ty, 1), nullptr); 6171 6172 // Grab the global __cpu_model. 6173 llvm::Constant *CpuModel = CGM.CreateRuntimeVariable(STy, "__cpu_model"); 6174 6175 // Grab the first (0th) element from the field __cpu_features off of the 6176 // global in the struct STy. 6177 Value *Idxs[] = { 6178 ConstantInt::get(Int32Ty, 0), 6179 ConstantInt::get(Int32Ty, 3), 6180 ConstantInt::get(Int32Ty, 0) 6181 }; 6182 Value *CpuFeatures = Builder.CreateGEP(STy, CpuModel, Idxs); 6183 Value *Features = Builder.CreateAlignedLoad(CpuFeatures, 6184 CharUnits::fromQuantity(4)); 6185 6186 // Check the value of the bit corresponding to the feature requested. 6187 Value *Bitset = Builder.CreateAnd( 6188 Features, llvm::ConstantInt::get(Int32Ty, 1 << Feature)); 6189 return Builder.CreateICmpNE(Bitset, llvm::ConstantInt::get(Int32Ty, 0)); 6190 } 6191 case X86::BI_mm_prefetch: { 6192 Value *Address = Ops[0]; 6193 Value *RW = ConstantInt::get(Int32Ty, 0); 6194 Value *Locality = Ops[1]; 6195 Value *Data = ConstantInt::get(Int32Ty, 1); 6196 Value *F = CGM.getIntrinsic(Intrinsic::prefetch); 6197 return Builder.CreateCall(F, {Address, RW, Locality, Data}); 6198 } 6199 case X86::BI__builtin_ia32_undef128: 6200 case X86::BI__builtin_ia32_undef256: 6201 case X86::BI__builtin_ia32_undef512: 6202 return UndefValue::get(ConvertType(E->getType())); 6203 case X86::BI__builtin_ia32_vec_init_v8qi: 6204 case X86::BI__builtin_ia32_vec_init_v4hi: 6205 case X86::BI__builtin_ia32_vec_init_v2si: 6206 return Builder.CreateBitCast(BuildVector(Ops), 6207 llvm::Type::getX86_MMXTy(getLLVMContext())); 6208 case X86::BI__builtin_ia32_vec_ext_v2si: 6209 return Builder.CreateExtractElement(Ops[0], 6210 llvm::ConstantInt::get(Ops[1]->getType(), 0)); 6211 case X86::BI__builtin_ia32_ldmxcsr: { 6212 Address Tmp = CreateMemTemp(E->getArg(0)->getType()); 6213 Builder.CreateStore(Ops[0], Tmp); 6214 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::x86_sse_ldmxcsr), 6215 Builder.CreateBitCast(Tmp.getPointer(), Int8PtrTy)); 6216 } 6217 case X86::BI__builtin_ia32_stmxcsr: { 6218 Address Tmp = CreateMemTemp(E->getType()); 6219 Builder.CreateCall(CGM.getIntrinsic(Intrinsic::x86_sse_stmxcsr), 6220 Builder.CreateBitCast(Tmp.getPointer(), Int8PtrTy)); 6221 return Builder.CreateLoad(Tmp, "stmxcsr"); 6222 } 6223 case X86::BI__builtin_ia32_xsave: 6224 case X86::BI__builtin_ia32_xsave64: 6225 case X86::BI__builtin_ia32_xrstor: 6226 case X86::BI__builtin_ia32_xrstor64: 6227 case X86::BI__builtin_ia32_xsaveopt: 6228 case X86::BI__builtin_ia32_xsaveopt64: 6229 case X86::BI__builtin_ia32_xrstors: 6230 case X86::BI__builtin_ia32_xrstors64: 6231 case X86::BI__builtin_ia32_xsavec: 6232 case X86::BI__builtin_ia32_xsavec64: 6233 case X86::BI__builtin_ia32_xsaves: 6234 case X86::BI__builtin_ia32_xsaves64: { 6235 Intrinsic::ID ID; 6236 #define INTRINSIC_X86_XSAVE_ID(NAME) \ 6237 case X86::BI__builtin_ia32_##NAME: \ 6238 ID = Intrinsic::x86_##NAME; \ 6239 break 6240 switch (BuiltinID) { 6241 default: llvm_unreachable("Unsupported intrinsic!"); 6242 INTRINSIC_X86_XSAVE_ID(xsave); 6243 INTRINSIC_X86_XSAVE_ID(xsave64); 6244 INTRINSIC_X86_XSAVE_ID(xrstor); 6245 INTRINSIC_X86_XSAVE_ID(xrstor64); 6246 INTRINSIC_X86_XSAVE_ID(xsaveopt); 6247 INTRINSIC_X86_XSAVE_ID(xsaveopt64); 6248 INTRINSIC_X86_XSAVE_ID(xrstors); 6249 INTRINSIC_X86_XSAVE_ID(xrstors64); 6250 INTRINSIC_X86_XSAVE_ID(xsavec); 6251 INTRINSIC_X86_XSAVE_ID(xsavec64); 6252 INTRINSIC_X86_XSAVE_ID(xsaves); 6253 INTRINSIC_X86_XSAVE_ID(xsaves64); 6254 } 6255 #undef INTRINSIC_X86_XSAVE_ID 6256 Value *Mhi = Builder.CreateTrunc( 6257 Builder.CreateLShr(Ops[1], ConstantInt::get(Int64Ty, 32)), Int32Ty); 6258 Value *Mlo = Builder.CreateTrunc(Ops[1], Int32Ty); 6259 Ops[1] = Mhi; 6260 Ops.push_back(Mlo); 6261 return Builder.CreateCall(CGM.getIntrinsic(ID), Ops); 6262 } 6263 case X86::BI__builtin_ia32_storehps: 6264 case X86::BI__builtin_ia32_storelps: { 6265 llvm::Type *PtrTy = llvm::PointerType::getUnqual(Int64Ty); 6266 llvm::Type *VecTy = llvm::VectorType::get(Int64Ty, 2); 6267 6268 // cast val v2i64 6269 Ops[1] = Builder.CreateBitCast(Ops[1], VecTy, "cast"); 6270 6271 // extract (0, 1) 6272 unsigned Index = BuiltinID == X86::BI__builtin_ia32_storelps ? 0 : 1; 6273 llvm::Value *Idx = llvm::ConstantInt::get(SizeTy, Index); 6274 Ops[1] = Builder.CreateExtractElement(Ops[1], Idx, "extract"); 6275 6276 // cast pointer to i64 & store 6277 Ops[0] = Builder.CreateBitCast(Ops[0], PtrTy); 6278 return Builder.CreateDefaultAlignedStore(Ops[1], Ops[0]); 6279 } 6280 case X86::BI__builtin_ia32_palignr128: 6281 case X86::BI__builtin_ia32_palignr256: { 6282 unsigned ShiftVal = cast<llvm::ConstantInt>(Ops[2])->getZExtValue(); 6283 6284 unsigned NumElts = 6285 cast<llvm::VectorType>(Ops[0]->getType())->getNumElements(); 6286 assert(NumElts % 16 == 0); 6287 unsigned NumLanes = NumElts / 16; 6288 unsigned NumLaneElts = NumElts / NumLanes; 6289 6290 // If palignr is shifting the pair of vectors more than the size of two 6291 // lanes, emit zero. 6292 if (ShiftVal >= (2 * NumLaneElts)) 6293 return llvm::Constant::getNullValue(ConvertType(E->getType())); 6294 6295 // If palignr is shifting the pair of input vectors more than one lane, 6296 // but less than two lanes, convert to shifting in zeroes. 6297 if (ShiftVal > NumLaneElts) { 6298 ShiftVal -= NumLaneElts; 6299 Ops[1] = Ops[0]; 6300 Ops[0] = llvm::Constant::getNullValue(Ops[0]->getType()); 6301 } 6302 6303 uint32_t Indices[32]; 6304 // 256-bit palignr operates on 128-bit lanes so we need to handle that 6305 for (unsigned l = 0; l != NumElts; l += NumLaneElts) { 6306 for (unsigned i = 0; i != NumLaneElts; ++i) { 6307 unsigned Idx = ShiftVal + i; 6308 if (Idx >= NumLaneElts) 6309 Idx += NumElts - NumLaneElts; // End of lane, switch operand. 6310 Indices[l + i] = Idx + l; 6311 } 6312 } 6313 6314 Value *SV = llvm::ConstantDataVector::get(getLLVMContext(), 6315 makeArrayRef(Indices, NumElts)); 6316 return Builder.CreateShuffleVector(Ops[1], Ops[0], SV, "palignr"); 6317 } 6318 case X86::BI__builtin_ia32_pslldqi256: { 6319 // Shift value is in bits so divide by 8. 6320 unsigned shiftVal = cast<llvm::ConstantInt>(Ops[1])->getZExtValue() >> 3; 6321 6322 // If pslldq is shifting the vector more than 15 bytes, emit zero. 6323 if (shiftVal >= 16) 6324 return llvm::Constant::getNullValue(ConvertType(E->getType())); 6325 6326 uint32_t Indices[32]; 6327 // 256-bit pslldq operates on 128-bit lanes so we need to handle that 6328 for (unsigned l = 0; l != 32; l += 16) { 6329 for (unsigned i = 0; i != 16; ++i) { 6330 unsigned Idx = 32 + i - shiftVal; 6331 if (Idx < 32) Idx -= 16; // end of lane, switch operand. 6332 Indices[l + i] = Idx + l; 6333 } 6334 } 6335 6336 llvm::Type *VecTy = llvm::VectorType::get(Int8Ty, 32); 6337 Ops[0] = Builder.CreateBitCast(Ops[0], VecTy, "cast"); 6338 Value *Zero = llvm::Constant::getNullValue(VecTy); 6339 6340 Value *SV = llvm::ConstantDataVector::get(getLLVMContext(), Indices); 6341 SV = Builder.CreateShuffleVector(Zero, Ops[0], SV, "pslldq"); 6342 llvm::Type *ResultType = ConvertType(E->getType()); 6343 return Builder.CreateBitCast(SV, ResultType, "cast"); 6344 } 6345 case X86::BI__builtin_ia32_psrldqi256: { 6346 // Shift value is in bits so divide by 8. 6347 unsigned shiftVal = cast<llvm::ConstantInt>(Ops[1])->getZExtValue() >> 3; 6348 6349 // If psrldq is shifting the vector more than 15 bytes, emit zero. 6350 if (shiftVal >= 16) 6351 return llvm::Constant::getNullValue(ConvertType(E->getType())); 6352 6353 uint32_t Indices[32]; 6354 // 256-bit psrldq operates on 128-bit lanes so we need to handle that 6355 for (unsigned l = 0; l != 32; l += 16) { 6356 for (unsigned i = 0; i != 16; ++i) { 6357 unsigned Idx = i + shiftVal; 6358 if (Idx >= 16) Idx += 16; // end of lane, switch operand. 6359 Indices[l + i] = Idx + l; 6360 } 6361 } 6362 6363 llvm::Type *VecTy = llvm::VectorType::get(Int8Ty, 32); 6364 Ops[0] = Builder.CreateBitCast(Ops[0], VecTy, "cast"); 6365 Value *Zero = llvm::Constant::getNullValue(VecTy); 6366 6367 Value *SV = llvm::ConstantDataVector::get(getLLVMContext(), Indices); 6368 SV = Builder.CreateShuffleVector(Ops[0], Zero, SV, "psrldq"); 6369 llvm::Type *ResultType = ConvertType(E->getType()); 6370 return Builder.CreateBitCast(SV, ResultType, "cast"); 6371 } 6372 case X86::BI__builtin_ia32_movntps: 6373 case X86::BI__builtin_ia32_movntps256: 6374 case X86::BI__builtin_ia32_movntpd: 6375 case X86::BI__builtin_ia32_movntpd256: 6376 case X86::BI__builtin_ia32_movntdq: 6377 case X86::BI__builtin_ia32_movntdq256: 6378 case X86::BI__builtin_ia32_movnti: 6379 case X86::BI__builtin_ia32_movnti64: { 6380 llvm::MDNode *Node = llvm::MDNode::get( 6381 getLLVMContext(), llvm::ConstantAsMetadata::get(Builder.getInt32(1))); 6382 6383 // Convert the type of the pointer to a pointer to the stored type. 6384 Value *BC = Builder.CreateBitCast(Ops[0], 6385 llvm::PointerType::getUnqual(Ops[1]->getType()), 6386 "cast"); 6387 StoreInst *SI = Builder.CreateDefaultAlignedStore(Ops[1], BC); 6388 SI->setMetadata(CGM.getModule().getMDKindID("nontemporal"), Node); 6389 6390 // If the operand is an integer, we can't assume alignment. Otherwise, 6391 // assume natural alignment. 6392 QualType ArgTy = E->getArg(1)->getType(); 6393 unsigned Align; 6394 if (ArgTy->isIntegerType()) 6395 Align = 1; 6396 else 6397 Align = getContext().getTypeSizeInChars(ArgTy).getQuantity(); 6398 SI->setAlignment(Align); 6399 return SI; 6400 } 6401 // 3DNow! 6402 case X86::BI__builtin_ia32_pswapdsf: 6403 case X86::BI__builtin_ia32_pswapdsi: { 6404 llvm::Type *MMXTy = llvm::Type::getX86_MMXTy(getLLVMContext()); 6405 Ops[0] = Builder.CreateBitCast(Ops[0], MMXTy, "cast"); 6406 llvm::Function *F = CGM.getIntrinsic(Intrinsic::x86_3dnowa_pswapd); 6407 return Builder.CreateCall(F, Ops, "pswapd"); 6408 } 6409 case X86::BI__builtin_ia32_rdrand16_step: 6410 case X86::BI__builtin_ia32_rdrand32_step: 6411 case X86::BI__builtin_ia32_rdrand64_step: 6412 case X86::BI__builtin_ia32_rdseed16_step: 6413 case X86::BI__builtin_ia32_rdseed32_step: 6414 case X86::BI__builtin_ia32_rdseed64_step: { 6415 Intrinsic::ID ID; 6416 switch (BuiltinID) { 6417 default: llvm_unreachable("Unsupported intrinsic!"); 6418 case X86::BI__builtin_ia32_rdrand16_step: 6419 ID = Intrinsic::x86_rdrand_16; 6420 break; 6421 case X86::BI__builtin_ia32_rdrand32_step: 6422 ID = Intrinsic::x86_rdrand_32; 6423 break; 6424 case X86::BI__builtin_ia32_rdrand64_step: 6425 ID = Intrinsic::x86_rdrand_64; 6426 break; 6427 case X86::BI__builtin_ia32_rdseed16_step: 6428 ID = Intrinsic::x86_rdseed_16; 6429 break; 6430 case X86::BI__builtin_ia32_rdseed32_step: 6431 ID = Intrinsic::x86_rdseed_32; 6432 break; 6433 case X86::BI__builtin_ia32_rdseed64_step: 6434 ID = Intrinsic::x86_rdseed_64; 6435 break; 6436 } 6437 6438 Value *Call = Builder.CreateCall(CGM.getIntrinsic(ID)); 6439 Builder.CreateDefaultAlignedStore(Builder.CreateExtractValue(Call, 0), 6440 Ops[0]); 6441 return Builder.CreateExtractValue(Call, 1); 6442 } 6443 // SSE comparison intrisics 6444 case X86::BI__builtin_ia32_cmpeqps: 6445 case X86::BI__builtin_ia32_cmpltps: 6446 case X86::BI__builtin_ia32_cmpleps: 6447 case X86::BI__builtin_ia32_cmpunordps: 6448 case X86::BI__builtin_ia32_cmpneqps: 6449 case X86::BI__builtin_ia32_cmpnltps: 6450 case X86::BI__builtin_ia32_cmpnleps: 6451 case X86::BI__builtin_ia32_cmpordps: 6452 case X86::BI__builtin_ia32_cmpeqss: 6453 case X86::BI__builtin_ia32_cmpltss: 6454 case X86::BI__builtin_ia32_cmpless: 6455 case X86::BI__builtin_ia32_cmpunordss: 6456 case X86::BI__builtin_ia32_cmpneqss: 6457 case X86::BI__builtin_ia32_cmpnltss: 6458 case X86::BI__builtin_ia32_cmpnless: 6459 case X86::BI__builtin_ia32_cmpordss: 6460 case X86::BI__builtin_ia32_cmpeqpd: 6461 case X86::BI__builtin_ia32_cmpltpd: 6462 case X86::BI__builtin_ia32_cmplepd: 6463 case X86::BI__builtin_ia32_cmpunordpd: 6464 case X86::BI__builtin_ia32_cmpneqpd: 6465 case X86::BI__builtin_ia32_cmpnltpd: 6466 case X86::BI__builtin_ia32_cmpnlepd: 6467 case X86::BI__builtin_ia32_cmpordpd: 6468 case X86::BI__builtin_ia32_cmpeqsd: 6469 case X86::BI__builtin_ia32_cmpltsd: 6470 case X86::BI__builtin_ia32_cmplesd: 6471 case X86::BI__builtin_ia32_cmpunordsd: 6472 case X86::BI__builtin_ia32_cmpneqsd: 6473 case X86::BI__builtin_ia32_cmpnltsd: 6474 case X86::BI__builtin_ia32_cmpnlesd: 6475 case X86::BI__builtin_ia32_cmpordsd: 6476 // These exist so that the builtin that takes an immediate can be bounds 6477 // checked by clang to avoid passing bad immediates to the backend. Since 6478 // AVX has a larger immediate than SSE we would need separate builtins to 6479 // do the different bounds checking. Rather than create a clang specific 6480 // SSE only builtin, this implements eight separate builtins to match gcc 6481 // implementation. 6482 6483 // Choose the immediate. 6484 unsigned Imm; 6485 switch (BuiltinID) { 6486 default: llvm_unreachable("Unsupported intrinsic!"); 6487 case X86::BI__builtin_ia32_cmpeqps: 6488 case X86::BI__builtin_ia32_cmpeqss: 6489 case X86::BI__builtin_ia32_cmpeqpd: 6490 case X86::BI__builtin_ia32_cmpeqsd: 6491 Imm = 0; 6492 break; 6493 case X86::BI__builtin_ia32_cmpltps: 6494 case X86::BI__builtin_ia32_cmpltss: 6495 case X86::BI__builtin_ia32_cmpltpd: 6496 case X86::BI__builtin_ia32_cmpltsd: 6497 Imm = 1; 6498 break; 6499 case X86::BI__builtin_ia32_cmpleps: 6500 case X86::BI__builtin_ia32_cmpless: 6501 case X86::BI__builtin_ia32_cmplepd: 6502 case X86::BI__builtin_ia32_cmplesd: 6503 Imm = 2; 6504 break; 6505 case X86::BI__builtin_ia32_cmpunordps: 6506 case X86::BI__builtin_ia32_cmpunordss: 6507 case X86::BI__builtin_ia32_cmpunordpd: 6508 case X86::BI__builtin_ia32_cmpunordsd: 6509 Imm = 3; 6510 break; 6511 case X86::BI__builtin_ia32_cmpneqps: 6512 case X86::BI__builtin_ia32_cmpneqss: 6513 case X86::BI__builtin_ia32_cmpneqpd: 6514 case X86::BI__builtin_ia32_cmpneqsd: 6515 Imm = 4; 6516 break; 6517 case X86::BI__builtin_ia32_cmpnltps: 6518 case X86::BI__builtin_ia32_cmpnltss: 6519 case X86::BI__builtin_ia32_cmpnltpd: 6520 case X86::BI__builtin_ia32_cmpnltsd: 6521 Imm = 5; 6522 break; 6523 case X86::BI__builtin_ia32_cmpnleps: 6524 case X86::BI__builtin_ia32_cmpnless: 6525 case X86::BI__builtin_ia32_cmpnlepd: 6526 case X86::BI__builtin_ia32_cmpnlesd: 6527 Imm = 6; 6528 break; 6529 case X86::BI__builtin_ia32_cmpordps: 6530 case X86::BI__builtin_ia32_cmpordss: 6531 case X86::BI__builtin_ia32_cmpordpd: 6532 case X86::BI__builtin_ia32_cmpordsd: 6533 Imm = 7; 6534 break; 6535 } 6536 6537 // Choose the intrinsic ID. 6538 const char *name; 6539 Intrinsic::ID ID; 6540 switch (BuiltinID) { 6541 default: llvm_unreachable("Unsupported intrinsic!"); 6542 case X86::BI__builtin_ia32_cmpeqps: 6543 case X86::BI__builtin_ia32_cmpltps: 6544 case X86::BI__builtin_ia32_cmpleps: 6545 case X86::BI__builtin_ia32_cmpunordps: 6546 case X86::BI__builtin_ia32_cmpneqps: 6547 case X86::BI__builtin_ia32_cmpnltps: 6548 case X86::BI__builtin_ia32_cmpnleps: 6549 case X86::BI__builtin_ia32_cmpordps: 6550 name = "cmpps"; 6551 ID = Intrinsic::x86_sse_cmp_ps; 6552 break; 6553 case X86::BI__builtin_ia32_cmpeqss: 6554 case X86::BI__builtin_ia32_cmpltss: 6555 case X86::BI__builtin_ia32_cmpless: 6556 case X86::BI__builtin_ia32_cmpunordss: 6557 case X86::BI__builtin_ia32_cmpneqss: 6558 case X86::BI__builtin_ia32_cmpnltss: 6559 case X86::BI__builtin_ia32_cmpnless: 6560 case X86::BI__builtin_ia32_cmpordss: 6561 name = "cmpss"; 6562 ID = Intrinsic::x86_sse_cmp_ss; 6563 break; 6564 case X86::BI__builtin_ia32_cmpeqpd: 6565 case X86::BI__builtin_ia32_cmpltpd: 6566 case X86::BI__builtin_ia32_cmplepd: 6567 case X86::BI__builtin_ia32_cmpunordpd: 6568 case X86::BI__builtin_ia32_cmpneqpd: 6569 case X86::BI__builtin_ia32_cmpnltpd: 6570 case X86::BI__builtin_ia32_cmpnlepd: 6571 case X86::BI__builtin_ia32_cmpordpd: 6572 name = "cmppd"; 6573 ID = Intrinsic::x86_sse2_cmp_pd; 6574 break; 6575 case X86::BI__builtin_ia32_cmpeqsd: 6576 case X86::BI__builtin_ia32_cmpltsd: 6577 case X86::BI__builtin_ia32_cmplesd: 6578 case X86::BI__builtin_ia32_cmpunordsd: 6579 case X86::BI__builtin_ia32_cmpneqsd: 6580 case X86::BI__builtin_ia32_cmpnltsd: 6581 case X86::BI__builtin_ia32_cmpnlesd: 6582 case X86::BI__builtin_ia32_cmpordsd: 6583 name = "cmpsd"; 6584 ID = Intrinsic::x86_sse2_cmp_sd; 6585 break; 6586 } 6587 6588 Ops.push_back(llvm::ConstantInt::get(Int8Ty, Imm)); 6589 llvm::Function *F = CGM.getIntrinsic(ID); 6590 return Builder.CreateCall(F, Ops, name); 6591 } 6592 } 6593 6594 6595 Value *CodeGenFunction::EmitPPCBuiltinExpr(unsigned BuiltinID, 6596 const CallExpr *E) { 6597 SmallVector<Value*, 4> Ops; 6598 6599 for (unsigned i = 0, e = E->getNumArgs(); i != e; i++) 6600 Ops.push_back(EmitScalarExpr(E->getArg(i))); 6601 6602 Intrinsic::ID ID = Intrinsic::not_intrinsic; 6603 6604 switch (BuiltinID) { 6605 default: return nullptr; 6606 6607 // __builtin_ppc_get_timebase is GCC 4.8+'s PowerPC-specific name for what we 6608 // call __builtin_readcyclecounter. 6609 case PPC::BI__builtin_ppc_get_timebase: 6610 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::readcyclecounter)); 6611 6612 // vec_ld, vec_lvsl, vec_lvsr 6613 case PPC::BI__builtin_altivec_lvx: 6614 case PPC::BI__builtin_altivec_lvxl: 6615 case PPC::BI__builtin_altivec_lvebx: 6616 case PPC::BI__builtin_altivec_lvehx: 6617 case PPC::BI__builtin_altivec_lvewx: 6618 case PPC::BI__builtin_altivec_lvsl: 6619 case PPC::BI__builtin_altivec_lvsr: 6620 case PPC::BI__builtin_vsx_lxvd2x: 6621 case PPC::BI__builtin_vsx_lxvw4x: 6622 { 6623 Ops[1] = Builder.CreateBitCast(Ops[1], Int8PtrTy); 6624 6625 Ops[0] = Builder.CreateGEP(Ops[1], Ops[0]); 6626 Ops.pop_back(); 6627 6628 switch (BuiltinID) { 6629 default: llvm_unreachable("Unsupported ld/lvsl/lvsr intrinsic!"); 6630 case PPC::BI__builtin_altivec_lvx: 6631 ID = Intrinsic::ppc_altivec_lvx; 6632 break; 6633 case PPC::BI__builtin_altivec_lvxl: 6634 ID = Intrinsic::ppc_altivec_lvxl; 6635 break; 6636 case PPC::BI__builtin_altivec_lvebx: 6637 ID = Intrinsic::ppc_altivec_lvebx; 6638 break; 6639 case PPC::BI__builtin_altivec_lvehx: 6640 ID = Intrinsic::ppc_altivec_lvehx; 6641 break; 6642 case PPC::BI__builtin_altivec_lvewx: 6643 ID = Intrinsic::ppc_altivec_lvewx; 6644 break; 6645 case PPC::BI__builtin_altivec_lvsl: 6646 ID = Intrinsic::ppc_altivec_lvsl; 6647 break; 6648 case PPC::BI__builtin_altivec_lvsr: 6649 ID = Intrinsic::ppc_altivec_lvsr; 6650 break; 6651 case PPC::BI__builtin_vsx_lxvd2x: 6652 ID = Intrinsic::ppc_vsx_lxvd2x; 6653 break; 6654 case PPC::BI__builtin_vsx_lxvw4x: 6655 ID = Intrinsic::ppc_vsx_lxvw4x; 6656 break; 6657 } 6658 llvm::Function *F = CGM.getIntrinsic(ID); 6659 return Builder.CreateCall(F, Ops, ""); 6660 } 6661 6662 // vec_st 6663 case PPC::BI__builtin_altivec_stvx: 6664 case PPC::BI__builtin_altivec_stvxl: 6665 case PPC::BI__builtin_altivec_stvebx: 6666 case PPC::BI__builtin_altivec_stvehx: 6667 case PPC::BI__builtin_altivec_stvewx: 6668 case PPC::BI__builtin_vsx_stxvd2x: 6669 case PPC::BI__builtin_vsx_stxvw4x: 6670 { 6671 Ops[2] = Builder.CreateBitCast(Ops[2], Int8PtrTy); 6672 Ops[1] = Builder.CreateGEP(Ops[2], Ops[1]); 6673 Ops.pop_back(); 6674 6675 switch (BuiltinID) { 6676 default: llvm_unreachable("Unsupported st intrinsic!"); 6677 case PPC::BI__builtin_altivec_stvx: 6678 ID = Intrinsic::ppc_altivec_stvx; 6679 break; 6680 case PPC::BI__builtin_altivec_stvxl: 6681 ID = Intrinsic::ppc_altivec_stvxl; 6682 break; 6683 case PPC::BI__builtin_altivec_stvebx: 6684 ID = Intrinsic::ppc_altivec_stvebx; 6685 break; 6686 case PPC::BI__builtin_altivec_stvehx: 6687 ID = Intrinsic::ppc_altivec_stvehx; 6688 break; 6689 case PPC::BI__builtin_altivec_stvewx: 6690 ID = Intrinsic::ppc_altivec_stvewx; 6691 break; 6692 case PPC::BI__builtin_vsx_stxvd2x: 6693 ID = Intrinsic::ppc_vsx_stxvd2x; 6694 break; 6695 case PPC::BI__builtin_vsx_stxvw4x: 6696 ID = Intrinsic::ppc_vsx_stxvw4x; 6697 break; 6698 } 6699 llvm::Function *F = CGM.getIntrinsic(ID); 6700 return Builder.CreateCall(F, Ops, ""); 6701 } 6702 // Square root 6703 case PPC::BI__builtin_vsx_xvsqrtsp: 6704 case PPC::BI__builtin_vsx_xvsqrtdp: { 6705 llvm::Type *ResultType = ConvertType(E->getType()); 6706 Value *X = EmitScalarExpr(E->getArg(0)); 6707 ID = Intrinsic::sqrt; 6708 llvm::Function *F = CGM.getIntrinsic(ID, ResultType); 6709 return Builder.CreateCall(F, X); 6710 } 6711 // Count leading zeros 6712 case PPC::BI__builtin_altivec_vclzb: 6713 case PPC::BI__builtin_altivec_vclzh: 6714 case PPC::BI__builtin_altivec_vclzw: 6715 case PPC::BI__builtin_altivec_vclzd: { 6716 llvm::Type *ResultType = ConvertType(E->getType()); 6717 Value *X = EmitScalarExpr(E->getArg(0)); 6718 Value *Undef = ConstantInt::get(Builder.getInt1Ty(), false); 6719 Function *F = CGM.getIntrinsic(Intrinsic::ctlz, ResultType); 6720 return Builder.CreateCall(F, {X, Undef}); 6721 } 6722 // Copy sign 6723 case PPC::BI__builtin_vsx_xvcpsgnsp: 6724 case PPC::BI__builtin_vsx_xvcpsgndp: { 6725 llvm::Type *ResultType = ConvertType(E->getType()); 6726 Value *X = EmitScalarExpr(E->getArg(0)); 6727 Value *Y = EmitScalarExpr(E->getArg(1)); 6728 ID = Intrinsic::copysign; 6729 llvm::Function *F = CGM.getIntrinsic(ID, ResultType); 6730 return Builder.CreateCall(F, {X, Y}); 6731 } 6732 // Rounding/truncation 6733 case PPC::BI__builtin_vsx_xvrspip: 6734 case PPC::BI__builtin_vsx_xvrdpip: 6735 case PPC::BI__builtin_vsx_xvrdpim: 6736 case PPC::BI__builtin_vsx_xvrspim: 6737 case PPC::BI__builtin_vsx_xvrdpi: 6738 case PPC::BI__builtin_vsx_xvrspi: 6739 case PPC::BI__builtin_vsx_xvrdpic: 6740 case PPC::BI__builtin_vsx_xvrspic: 6741 case PPC::BI__builtin_vsx_xvrdpiz: 6742 case PPC::BI__builtin_vsx_xvrspiz: { 6743 llvm::Type *ResultType = ConvertType(E->getType()); 6744 Value *X = EmitScalarExpr(E->getArg(0)); 6745 if (BuiltinID == PPC::BI__builtin_vsx_xvrdpim || 6746 BuiltinID == PPC::BI__builtin_vsx_xvrspim) 6747 ID = Intrinsic::floor; 6748 else if (BuiltinID == PPC::BI__builtin_vsx_xvrdpi || 6749 BuiltinID == PPC::BI__builtin_vsx_xvrspi) 6750 ID = Intrinsic::round; 6751 else if (BuiltinID == PPC::BI__builtin_vsx_xvrdpic || 6752 BuiltinID == PPC::BI__builtin_vsx_xvrspic) 6753 ID = Intrinsic::nearbyint; 6754 else if (BuiltinID == PPC::BI__builtin_vsx_xvrdpip || 6755 BuiltinID == PPC::BI__builtin_vsx_xvrspip) 6756 ID = Intrinsic::ceil; 6757 else if (BuiltinID == PPC::BI__builtin_vsx_xvrdpiz || 6758 BuiltinID == PPC::BI__builtin_vsx_xvrspiz) 6759 ID = Intrinsic::trunc; 6760 llvm::Function *F = CGM.getIntrinsic(ID, ResultType); 6761 return Builder.CreateCall(F, X); 6762 } 6763 // FMA variations 6764 case PPC::BI__builtin_vsx_xvmaddadp: 6765 case PPC::BI__builtin_vsx_xvmaddasp: 6766 case PPC::BI__builtin_vsx_xvnmaddadp: 6767 case PPC::BI__builtin_vsx_xvnmaddasp: 6768 case PPC::BI__builtin_vsx_xvmsubadp: 6769 case PPC::BI__builtin_vsx_xvmsubasp: 6770 case PPC::BI__builtin_vsx_xvnmsubadp: 6771 case PPC::BI__builtin_vsx_xvnmsubasp: { 6772 llvm::Type *ResultType = ConvertType(E->getType()); 6773 Value *X = EmitScalarExpr(E->getArg(0)); 6774 Value *Y = EmitScalarExpr(E->getArg(1)); 6775 Value *Z = EmitScalarExpr(E->getArg(2)); 6776 Value *Zero = llvm::ConstantFP::getZeroValueForNegation(ResultType); 6777 llvm::Function *F = CGM.getIntrinsic(Intrinsic::fma, ResultType); 6778 switch (BuiltinID) { 6779 case PPC::BI__builtin_vsx_xvmaddadp: 6780 case PPC::BI__builtin_vsx_xvmaddasp: 6781 return Builder.CreateCall(F, {X, Y, Z}); 6782 case PPC::BI__builtin_vsx_xvnmaddadp: 6783 case PPC::BI__builtin_vsx_xvnmaddasp: 6784 return Builder.CreateFSub(Zero, 6785 Builder.CreateCall(F, {X, Y, Z}), "sub"); 6786 case PPC::BI__builtin_vsx_xvmsubadp: 6787 case PPC::BI__builtin_vsx_xvmsubasp: 6788 return Builder.CreateCall(F, 6789 {X, Y, Builder.CreateFSub(Zero, Z, "sub")}); 6790 case PPC::BI__builtin_vsx_xvnmsubadp: 6791 case PPC::BI__builtin_vsx_xvnmsubasp: 6792 Value *FsubRes = 6793 Builder.CreateCall(F, {X, Y, Builder.CreateFSub(Zero, Z, "sub")}); 6794 return Builder.CreateFSub(Zero, FsubRes, "sub"); 6795 } 6796 llvm_unreachable("Unknown FMA operation"); 6797 return nullptr; // Suppress no-return warning 6798 } 6799 } 6800 } 6801 6802 // Emit an intrinsic that has 1 float or double. 6803 static Value *emitUnaryFPBuiltin(CodeGenFunction &CGF, 6804 const CallExpr *E, 6805 unsigned IntrinsicID) { 6806 llvm::Value *Src0 = CGF.EmitScalarExpr(E->getArg(0)); 6807 6808 Value *F = CGF.CGM.getIntrinsic(IntrinsicID, Src0->getType()); 6809 return CGF.Builder.CreateCall(F, Src0); 6810 } 6811 6812 // Emit an intrinsic that has 3 float or double operands. 6813 static Value *emitTernaryFPBuiltin(CodeGenFunction &CGF, 6814 const CallExpr *E, 6815 unsigned IntrinsicID) { 6816 llvm::Value *Src0 = CGF.EmitScalarExpr(E->getArg(0)); 6817 llvm::Value *Src1 = CGF.EmitScalarExpr(E->getArg(1)); 6818 llvm::Value *Src2 = CGF.EmitScalarExpr(E->getArg(2)); 6819 6820 Value *F = CGF.CGM.getIntrinsic(IntrinsicID, Src0->getType()); 6821 return CGF.Builder.CreateCall(F, {Src0, Src1, Src2}); 6822 } 6823 6824 // Emit an intrinsic that has 1 float or double operand, and 1 integer. 6825 static Value *emitFPIntBuiltin(CodeGenFunction &CGF, 6826 const CallExpr *E, 6827 unsigned IntrinsicID) { 6828 llvm::Value *Src0 = CGF.EmitScalarExpr(E->getArg(0)); 6829 llvm::Value *Src1 = CGF.EmitScalarExpr(E->getArg(1)); 6830 6831 Value *F = CGF.CGM.getIntrinsic(IntrinsicID, Src0->getType()); 6832 return CGF.Builder.CreateCall(F, {Src0, Src1}); 6833 } 6834 6835 Value *CodeGenFunction::EmitAMDGPUBuiltinExpr(unsigned BuiltinID, 6836 const CallExpr *E) { 6837 switch (BuiltinID) { 6838 case AMDGPU::BI__builtin_amdgpu_div_scale: 6839 case AMDGPU::BI__builtin_amdgpu_div_scalef: { 6840 // Translate from the intrinsics's struct return to the builtin's out 6841 // argument. 6842 6843 Address FlagOutPtr = EmitPointerWithAlignment(E->getArg(3)); 6844 6845 llvm::Value *X = EmitScalarExpr(E->getArg(0)); 6846 llvm::Value *Y = EmitScalarExpr(E->getArg(1)); 6847 llvm::Value *Z = EmitScalarExpr(E->getArg(2)); 6848 6849 llvm::Value *Callee = CGM.getIntrinsic(Intrinsic::AMDGPU_div_scale, 6850 X->getType()); 6851 6852 llvm::Value *Tmp = Builder.CreateCall(Callee, {X, Y, Z}); 6853 6854 llvm::Value *Result = Builder.CreateExtractValue(Tmp, 0); 6855 llvm::Value *Flag = Builder.CreateExtractValue(Tmp, 1); 6856 6857 llvm::Type *RealFlagType 6858 = FlagOutPtr.getPointer()->getType()->getPointerElementType(); 6859 6860 llvm::Value *FlagExt = Builder.CreateZExt(Flag, RealFlagType); 6861 Builder.CreateStore(FlagExt, FlagOutPtr); 6862 return Result; 6863 } 6864 case AMDGPU::BI__builtin_amdgpu_div_fmas: 6865 case AMDGPU::BI__builtin_amdgpu_div_fmasf: { 6866 llvm::Value *Src0 = EmitScalarExpr(E->getArg(0)); 6867 llvm::Value *Src1 = EmitScalarExpr(E->getArg(1)); 6868 llvm::Value *Src2 = EmitScalarExpr(E->getArg(2)); 6869 llvm::Value *Src3 = EmitScalarExpr(E->getArg(3)); 6870 6871 llvm::Value *F = CGM.getIntrinsic(Intrinsic::AMDGPU_div_fmas, 6872 Src0->getType()); 6873 llvm::Value *Src3ToBool = Builder.CreateIsNotNull(Src3); 6874 return Builder.CreateCall(F, {Src0, Src1, Src2, Src3ToBool}); 6875 } 6876 case AMDGPU::BI__builtin_amdgpu_div_fixup: 6877 case AMDGPU::BI__builtin_amdgpu_div_fixupf: 6878 return emitTernaryFPBuiltin(*this, E, Intrinsic::AMDGPU_div_fixup); 6879 case AMDGPU::BI__builtin_amdgpu_trig_preop: 6880 case AMDGPU::BI__builtin_amdgpu_trig_preopf: 6881 return emitFPIntBuiltin(*this, E, Intrinsic::AMDGPU_trig_preop); 6882 case AMDGPU::BI__builtin_amdgpu_rcp: 6883 case AMDGPU::BI__builtin_amdgpu_rcpf: 6884 return emitUnaryFPBuiltin(*this, E, Intrinsic::AMDGPU_rcp); 6885 case AMDGPU::BI__builtin_amdgpu_rsq: 6886 case AMDGPU::BI__builtin_amdgpu_rsqf: 6887 return emitUnaryFPBuiltin(*this, E, Intrinsic::AMDGPU_rsq); 6888 case AMDGPU::BI__builtin_amdgpu_rsq_clamped: 6889 case AMDGPU::BI__builtin_amdgpu_rsq_clampedf: 6890 return emitUnaryFPBuiltin(*this, E, Intrinsic::AMDGPU_rsq_clamped); 6891 case AMDGPU::BI__builtin_amdgpu_ldexp: 6892 case AMDGPU::BI__builtin_amdgpu_ldexpf: 6893 return emitFPIntBuiltin(*this, E, Intrinsic::AMDGPU_ldexp); 6894 case AMDGPU::BI__builtin_amdgpu_class: 6895 case AMDGPU::BI__builtin_amdgpu_classf: 6896 return emitFPIntBuiltin(*this, E, Intrinsic::AMDGPU_class); 6897 default: 6898 return nullptr; 6899 } 6900 } 6901 6902 /// Handle a SystemZ function in which the final argument is a pointer 6903 /// to an int that receives the post-instruction CC value. At the LLVM level 6904 /// this is represented as a function that returns a {result, cc} pair. 6905 static Value *EmitSystemZIntrinsicWithCC(CodeGenFunction &CGF, 6906 unsigned IntrinsicID, 6907 const CallExpr *E) { 6908 unsigned NumArgs = E->getNumArgs() - 1; 6909 SmallVector<Value *, 8> Args(NumArgs); 6910 for (unsigned I = 0; I < NumArgs; ++I) 6911 Args[I] = CGF.EmitScalarExpr(E->getArg(I)); 6912 Address CCPtr = CGF.EmitPointerWithAlignment(E->getArg(NumArgs)); 6913 Value *F = CGF.CGM.getIntrinsic(IntrinsicID); 6914 Value *Call = CGF.Builder.CreateCall(F, Args); 6915 Value *CC = CGF.Builder.CreateExtractValue(Call, 1); 6916 CGF.Builder.CreateStore(CC, CCPtr); 6917 return CGF.Builder.CreateExtractValue(Call, 0); 6918 } 6919 6920 Value *CodeGenFunction::EmitSystemZBuiltinExpr(unsigned BuiltinID, 6921 const CallExpr *E) { 6922 switch (BuiltinID) { 6923 case SystemZ::BI__builtin_tbegin: { 6924 Value *TDB = EmitScalarExpr(E->getArg(0)); 6925 Value *Control = llvm::ConstantInt::get(Int32Ty, 0xff0c); 6926 Value *F = CGM.getIntrinsic(Intrinsic::s390_tbegin); 6927 return Builder.CreateCall(F, {TDB, Control}); 6928 } 6929 case SystemZ::BI__builtin_tbegin_nofloat: { 6930 Value *TDB = EmitScalarExpr(E->getArg(0)); 6931 Value *Control = llvm::ConstantInt::get(Int32Ty, 0xff0c); 6932 Value *F = CGM.getIntrinsic(Intrinsic::s390_tbegin_nofloat); 6933 return Builder.CreateCall(F, {TDB, Control}); 6934 } 6935 case SystemZ::BI__builtin_tbeginc: { 6936 Value *TDB = llvm::ConstantPointerNull::get(Int8PtrTy); 6937 Value *Control = llvm::ConstantInt::get(Int32Ty, 0xff08); 6938 Value *F = CGM.getIntrinsic(Intrinsic::s390_tbeginc); 6939 return Builder.CreateCall(F, {TDB, Control}); 6940 } 6941 case SystemZ::BI__builtin_tabort: { 6942 Value *Data = EmitScalarExpr(E->getArg(0)); 6943 Value *F = CGM.getIntrinsic(Intrinsic::s390_tabort); 6944 return Builder.CreateCall(F, Builder.CreateSExt(Data, Int64Ty, "tabort")); 6945 } 6946 case SystemZ::BI__builtin_non_tx_store: { 6947 Value *Address = EmitScalarExpr(E->getArg(0)); 6948 Value *Data = EmitScalarExpr(E->getArg(1)); 6949 Value *F = CGM.getIntrinsic(Intrinsic::s390_ntstg); 6950 return Builder.CreateCall(F, {Data, Address}); 6951 } 6952 6953 // Vector builtins. Note that most vector builtins are mapped automatically 6954 // to target-specific LLVM intrinsics. The ones handled specially here can 6955 // be represented via standard LLVM IR, which is preferable to enable common 6956 // LLVM optimizations. 6957 6958 case SystemZ::BI__builtin_s390_vpopctb: 6959 case SystemZ::BI__builtin_s390_vpopcth: 6960 case SystemZ::BI__builtin_s390_vpopctf: 6961 case SystemZ::BI__builtin_s390_vpopctg: { 6962 llvm::Type *ResultType = ConvertType(E->getType()); 6963 Value *X = EmitScalarExpr(E->getArg(0)); 6964 Function *F = CGM.getIntrinsic(Intrinsic::ctpop, ResultType); 6965 return Builder.CreateCall(F, X); 6966 } 6967 6968 case SystemZ::BI__builtin_s390_vclzb: 6969 case SystemZ::BI__builtin_s390_vclzh: 6970 case SystemZ::BI__builtin_s390_vclzf: 6971 case SystemZ::BI__builtin_s390_vclzg: { 6972 llvm::Type *ResultType = ConvertType(E->getType()); 6973 Value *X = EmitScalarExpr(E->getArg(0)); 6974 Value *Undef = ConstantInt::get(Builder.getInt1Ty(), false); 6975 Function *F = CGM.getIntrinsic(Intrinsic::ctlz, ResultType); 6976 return Builder.CreateCall(F, {X, Undef}); 6977 } 6978 6979 case SystemZ::BI__builtin_s390_vctzb: 6980 case SystemZ::BI__builtin_s390_vctzh: 6981 case SystemZ::BI__builtin_s390_vctzf: 6982 case SystemZ::BI__builtin_s390_vctzg: { 6983 llvm::Type *ResultType = ConvertType(E->getType()); 6984 Value *X = EmitScalarExpr(E->getArg(0)); 6985 Value *Undef = ConstantInt::get(Builder.getInt1Ty(), false); 6986 Function *F = CGM.getIntrinsic(Intrinsic::cttz, ResultType); 6987 return Builder.CreateCall(F, {X, Undef}); 6988 } 6989 6990 case SystemZ::BI__builtin_s390_vfsqdb: { 6991 llvm::Type *ResultType = ConvertType(E->getType()); 6992 Value *X = EmitScalarExpr(E->getArg(0)); 6993 Function *F = CGM.getIntrinsic(Intrinsic::sqrt, ResultType); 6994 return Builder.CreateCall(F, X); 6995 } 6996 case SystemZ::BI__builtin_s390_vfmadb: { 6997 llvm::Type *ResultType = ConvertType(E->getType()); 6998 Value *X = EmitScalarExpr(E->getArg(0)); 6999 Value *Y = EmitScalarExpr(E->getArg(1)); 7000 Value *Z = EmitScalarExpr(E->getArg(2)); 7001 Function *F = CGM.getIntrinsic(Intrinsic::fma, ResultType); 7002 return Builder.CreateCall(F, {X, Y, Z}); 7003 } 7004 case SystemZ::BI__builtin_s390_vfmsdb: { 7005 llvm::Type *ResultType = ConvertType(E->getType()); 7006 Value *X = EmitScalarExpr(E->getArg(0)); 7007 Value *Y = EmitScalarExpr(E->getArg(1)); 7008 Value *Z = EmitScalarExpr(E->getArg(2)); 7009 Value *Zero = llvm::ConstantFP::getZeroValueForNegation(ResultType); 7010 Function *F = CGM.getIntrinsic(Intrinsic::fma, ResultType); 7011 return Builder.CreateCall(F, {X, Y, Builder.CreateFSub(Zero, Z, "sub")}); 7012 } 7013 case SystemZ::BI__builtin_s390_vflpdb: { 7014 llvm::Type *ResultType = ConvertType(E->getType()); 7015 Value *X = EmitScalarExpr(E->getArg(0)); 7016 Function *F = CGM.getIntrinsic(Intrinsic::fabs, ResultType); 7017 return Builder.CreateCall(F, X); 7018 } 7019 case SystemZ::BI__builtin_s390_vflndb: { 7020 llvm::Type *ResultType = ConvertType(E->getType()); 7021 Value *X = EmitScalarExpr(E->getArg(0)); 7022 Value *Zero = llvm::ConstantFP::getZeroValueForNegation(ResultType); 7023 Function *F = CGM.getIntrinsic(Intrinsic::fabs, ResultType); 7024 return Builder.CreateFSub(Zero, Builder.CreateCall(F, X), "sub"); 7025 } 7026 case SystemZ::BI__builtin_s390_vfidb: { 7027 llvm::Type *ResultType = ConvertType(E->getType()); 7028 Value *X = EmitScalarExpr(E->getArg(0)); 7029 // Constant-fold the M4 and M5 mask arguments. 7030 llvm::APSInt M4, M5; 7031 bool IsConstM4 = E->getArg(1)->isIntegerConstantExpr(M4, getContext()); 7032 bool IsConstM5 = E->getArg(2)->isIntegerConstantExpr(M5, getContext()); 7033 assert(IsConstM4 && IsConstM5 && "Constant arg isn't actually constant?"); 7034 (void)IsConstM4; (void)IsConstM5; 7035 // Check whether this instance of vfidb can be represented via a LLVM 7036 // standard intrinsic. We only support some combinations of M4 and M5. 7037 Intrinsic::ID ID = Intrinsic::not_intrinsic; 7038 switch (M4.getZExtValue()) { 7039 default: break; 7040 case 0: // IEEE-inexact exception allowed 7041 switch (M5.getZExtValue()) { 7042 default: break; 7043 case 0: ID = Intrinsic::rint; break; 7044 } 7045 break; 7046 case 4: // IEEE-inexact exception suppressed 7047 switch (M5.getZExtValue()) { 7048 default: break; 7049 case 0: ID = Intrinsic::nearbyint; break; 7050 case 1: ID = Intrinsic::round; break; 7051 case 5: ID = Intrinsic::trunc; break; 7052 case 6: ID = Intrinsic::ceil; break; 7053 case 7: ID = Intrinsic::floor; break; 7054 } 7055 break; 7056 } 7057 if (ID != Intrinsic::not_intrinsic) { 7058 Function *F = CGM.getIntrinsic(ID, ResultType); 7059 return Builder.CreateCall(F, X); 7060 } 7061 Function *F = CGM.getIntrinsic(Intrinsic::s390_vfidb); 7062 Value *M4Value = llvm::ConstantInt::get(getLLVMContext(), M4); 7063 Value *M5Value = llvm::ConstantInt::get(getLLVMContext(), M5); 7064 return Builder.CreateCall(F, {X, M4Value, M5Value}); 7065 } 7066 7067 // Vector intrisincs that output the post-instruction CC value. 7068 7069 #define INTRINSIC_WITH_CC(NAME) \ 7070 case SystemZ::BI__builtin_##NAME: \ 7071 return EmitSystemZIntrinsicWithCC(*this, Intrinsic::NAME, E) 7072 7073 INTRINSIC_WITH_CC(s390_vpkshs); 7074 INTRINSIC_WITH_CC(s390_vpksfs); 7075 INTRINSIC_WITH_CC(s390_vpksgs); 7076 7077 INTRINSIC_WITH_CC(s390_vpklshs); 7078 INTRINSIC_WITH_CC(s390_vpklsfs); 7079 INTRINSIC_WITH_CC(s390_vpklsgs); 7080 7081 INTRINSIC_WITH_CC(s390_vceqbs); 7082 INTRINSIC_WITH_CC(s390_vceqhs); 7083 INTRINSIC_WITH_CC(s390_vceqfs); 7084 INTRINSIC_WITH_CC(s390_vceqgs); 7085 7086 INTRINSIC_WITH_CC(s390_vchbs); 7087 INTRINSIC_WITH_CC(s390_vchhs); 7088 INTRINSIC_WITH_CC(s390_vchfs); 7089 INTRINSIC_WITH_CC(s390_vchgs); 7090 7091 INTRINSIC_WITH_CC(s390_vchlbs); 7092 INTRINSIC_WITH_CC(s390_vchlhs); 7093 INTRINSIC_WITH_CC(s390_vchlfs); 7094 INTRINSIC_WITH_CC(s390_vchlgs); 7095 7096 INTRINSIC_WITH_CC(s390_vfaebs); 7097 INTRINSIC_WITH_CC(s390_vfaehs); 7098 INTRINSIC_WITH_CC(s390_vfaefs); 7099 7100 INTRINSIC_WITH_CC(s390_vfaezbs); 7101 INTRINSIC_WITH_CC(s390_vfaezhs); 7102 INTRINSIC_WITH_CC(s390_vfaezfs); 7103 7104 INTRINSIC_WITH_CC(s390_vfeebs); 7105 INTRINSIC_WITH_CC(s390_vfeehs); 7106 INTRINSIC_WITH_CC(s390_vfeefs); 7107 7108 INTRINSIC_WITH_CC(s390_vfeezbs); 7109 INTRINSIC_WITH_CC(s390_vfeezhs); 7110 INTRINSIC_WITH_CC(s390_vfeezfs); 7111 7112 INTRINSIC_WITH_CC(s390_vfenebs); 7113 INTRINSIC_WITH_CC(s390_vfenehs); 7114 INTRINSIC_WITH_CC(s390_vfenefs); 7115 7116 INTRINSIC_WITH_CC(s390_vfenezbs); 7117 INTRINSIC_WITH_CC(s390_vfenezhs); 7118 INTRINSIC_WITH_CC(s390_vfenezfs); 7119 7120 INTRINSIC_WITH_CC(s390_vistrbs); 7121 INTRINSIC_WITH_CC(s390_vistrhs); 7122 INTRINSIC_WITH_CC(s390_vistrfs); 7123 7124 INTRINSIC_WITH_CC(s390_vstrcbs); 7125 INTRINSIC_WITH_CC(s390_vstrchs); 7126 INTRINSIC_WITH_CC(s390_vstrcfs); 7127 7128 INTRINSIC_WITH_CC(s390_vstrczbs); 7129 INTRINSIC_WITH_CC(s390_vstrczhs); 7130 INTRINSIC_WITH_CC(s390_vstrczfs); 7131 7132 INTRINSIC_WITH_CC(s390_vfcedbs); 7133 INTRINSIC_WITH_CC(s390_vfchdbs); 7134 INTRINSIC_WITH_CC(s390_vfchedbs); 7135 7136 INTRINSIC_WITH_CC(s390_vftcidb); 7137 7138 #undef INTRINSIC_WITH_CC 7139 7140 default: 7141 return nullptr; 7142 } 7143 } 7144 7145 Value *CodeGenFunction::EmitNVPTXBuiltinExpr(unsigned BuiltinID, 7146 const CallExpr *E) { 7147 switch (BuiltinID) { 7148 case NVPTX::BI__nvvm_atom_add_gen_i: 7149 case NVPTX::BI__nvvm_atom_add_gen_l: 7150 case NVPTX::BI__nvvm_atom_add_gen_ll: 7151 return MakeBinaryAtomicValue(*this, llvm::AtomicRMWInst::Add, E); 7152 7153 case NVPTX::BI__nvvm_atom_sub_gen_i: 7154 case NVPTX::BI__nvvm_atom_sub_gen_l: 7155 case NVPTX::BI__nvvm_atom_sub_gen_ll: 7156 return MakeBinaryAtomicValue(*this, llvm::AtomicRMWInst::Sub, E); 7157 7158 case NVPTX::BI__nvvm_atom_and_gen_i: 7159 case NVPTX::BI__nvvm_atom_and_gen_l: 7160 case NVPTX::BI__nvvm_atom_and_gen_ll: 7161 return MakeBinaryAtomicValue(*this, llvm::AtomicRMWInst::And, E); 7162 7163 case NVPTX::BI__nvvm_atom_or_gen_i: 7164 case NVPTX::BI__nvvm_atom_or_gen_l: 7165 case NVPTX::BI__nvvm_atom_or_gen_ll: 7166 return MakeBinaryAtomicValue(*this, llvm::AtomicRMWInst::Or, E); 7167 7168 case NVPTX::BI__nvvm_atom_xor_gen_i: 7169 case NVPTX::BI__nvvm_atom_xor_gen_l: 7170 case NVPTX::BI__nvvm_atom_xor_gen_ll: 7171 return MakeBinaryAtomicValue(*this, llvm::AtomicRMWInst::Xor, E); 7172 7173 case NVPTX::BI__nvvm_atom_xchg_gen_i: 7174 case NVPTX::BI__nvvm_atom_xchg_gen_l: 7175 case NVPTX::BI__nvvm_atom_xchg_gen_ll: 7176 return MakeBinaryAtomicValue(*this, llvm::AtomicRMWInst::Xchg, E); 7177 7178 case NVPTX::BI__nvvm_atom_max_gen_i: 7179 case NVPTX::BI__nvvm_atom_max_gen_l: 7180 case NVPTX::BI__nvvm_atom_max_gen_ll: 7181 return MakeBinaryAtomicValue(*this, llvm::AtomicRMWInst::Max, E); 7182 7183 case NVPTX::BI__nvvm_atom_max_gen_ui: 7184 case NVPTX::BI__nvvm_atom_max_gen_ul: 7185 case NVPTX::BI__nvvm_atom_max_gen_ull: 7186 return MakeBinaryAtomicValue(*this, llvm::AtomicRMWInst::UMax, E); 7187 7188 case NVPTX::BI__nvvm_atom_min_gen_i: 7189 case NVPTX::BI__nvvm_atom_min_gen_l: 7190 case NVPTX::BI__nvvm_atom_min_gen_ll: 7191 return MakeBinaryAtomicValue(*this, llvm::AtomicRMWInst::Min, E); 7192 7193 case NVPTX::BI__nvvm_atom_min_gen_ui: 7194 case NVPTX::BI__nvvm_atom_min_gen_ul: 7195 case NVPTX::BI__nvvm_atom_min_gen_ull: 7196 return MakeBinaryAtomicValue(*this, llvm::AtomicRMWInst::UMin, E); 7197 7198 case NVPTX::BI__nvvm_atom_cas_gen_i: 7199 case NVPTX::BI__nvvm_atom_cas_gen_l: 7200 case NVPTX::BI__nvvm_atom_cas_gen_ll: 7201 // __nvvm_atom_cas_gen_* should return the old value rather than the 7202 // success flag. 7203 return MakeAtomicCmpXchgValue(*this, E, /*ReturnBool=*/false); 7204 7205 case NVPTX::BI__nvvm_atom_add_gen_f: { 7206 Value *Ptr = EmitScalarExpr(E->getArg(0)); 7207 Value *Val = EmitScalarExpr(E->getArg(1)); 7208 // atomicrmw only deals with integer arguments so we need to use 7209 // LLVM's nvvm_atomic_load_add_f32 intrinsic for that. 7210 Value *FnALAF32 = 7211 CGM.getIntrinsic(Intrinsic::nvvm_atomic_load_add_f32, Ptr->getType()); 7212 return Builder.CreateCall(FnALAF32, {Ptr, Val}); 7213 } 7214 7215 default: 7216 return nullptr; 7217 } 7218 } 7219 7220 Value *CodeGenFunction::EmitWebAssemblyBuiltinExpr(unsigned BuiltinID, 7221 const CallExpr *E) { 7222 switch (BuiltinID) { 7223 case WebAssembly::BI__builtin_wasm_memory_size: { 7224 llvm::Type *ResultType = ConvertType(E->getType()); 7225 Value *Callee = CGM.getIntrinsic(Intrinsic::wasm_memory_size, ResultType); 7226 return Builder.CreateCall(Callee); 7227 } 7228 case WebAssembly::BI__builtin_wasm_grow_memory: { 7229 Value *X = EmitScalarExpr(E->getArg(0)); 7230 Value *Callee = CGM.getIntrinsic(Intrinsic::wasm_grow_memory, X->getType()); 7231 return Builder.CreateCall(Callee, X); 7232 } 7233 7234 default: 7235 return nullptr; 7236 } 7237 } 7238