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 RValue EmitBinaryAtomic(CodeGenFunction &CGF, 115 llvm::AtomicRMWInst::BinOp Kind, 116 const CallExpr *E) { 117 return RValue::get(MakeBinaryAtomicValue(CGF, Kind, E)); 118 } 119 120 /// Utility to insert an atomic instruction based Instrinsic::ID and 121 /// the expression node, where the return value is the result of the 122 /// operation. 123 static RValue EmitBinaryAtomicPost(CodeGenFunction &CGF, 124 llvm::AtomicRMWInst::BinOp Kind, 125 const CallExpr *E, 126 Instruction::BinaryOps Op, 127 bool Invert = false) { 128 QualType T = E->getType(); 129 assert(E->getArg(0)->getType()->isPointerType()); 130 assert(CGF.getContext().hasSameUnqualifiedType(T, 131 E->getArg(0)->getType()->getPointeeType())); 132 assert(CGF.getContext().hasSameUnqualifiedType(T, E->getArg(1)->getType())); 133 134 llvm::Value *DestPtr = CGF.EmitScalarExpr(E->getArg(0)); 135 unsigned AddrSpace = DestPtr->getType()->getPointerAddressSpace(); 136 137 llvm::IntegerType *IntType = 138 llvm::IntegerType::get(CGF.getLLVMContext(), 139 CGF.getContext().getTypeSize(T)); 140 llvm::Type *IntPtrType = IntType->getPointerTo(AddrSpace); 141 142 llvm::Value *Args[2]; 143 Args[1] = CGF.EmitScalarExpr(E->getArg(1)); 144 llvm::Type *ValueType = Args[1]->getType(); 145 Args[1] = EmitToInt(CGF, Args[1], T, IntType); 146 Args[0] = CGF.Builder.CreateBitCast(DestPtr, IntPtrType); 147 148 llvm::Value *Result = 149 CGF.Builder.CreateAtomicRMW(Kind, Args[0], Args[1], 150 llvm::SequentiallyConsistent); 151 Result = CGF.Builder.CreateBinOp(Op, Result, Args[1]); 152 if (Invert) 153 Result = CGF.Builder.CreateBinOp(llvm::Instruction::Xor, Result, 154 llvm::ConstantInt::get(IntType, -1)); 155 Result = EmitFromInt(CGF, Result, T, ValueType); 156 return RValue::get(Result); 157 } 158 159 /// @brief Utility to insert an atomic cmpxchg instruction. 160 /// 161 /// @param CGF The current codegen function. 162 /// @param E Builtin call expression to convert to cmpxchg. 163 /// arg0 - address to operate on 164 /// arg1 - value to compare with 165 /// arg2 - new value 166 /// @param ReturnBool Specifies whether to return success flag of 167 /// cmpxchg result or the old value. 168 /// 169 /// @returns result of cmpxchg, according to ReturnBool 170 static Value *MakeAtomicCmpXchgValue(CodeGenFunction &CGF, const CallExpr *E, 171 bool ReturnBool) { 172 QualType T = ReturnBool ? E->getArg(1)->getType() : E->getType(); 173 llvm::Value *DestPtr = CGF.EmitScalarExpr(E->getArg(0)); 174 unsigned AddrSpace = DestPtr->getType()->getPointerAddressSpace(); 175 176 llvm::IntegerType *IntType = llvm::IntegerType::get( 177 CGF.getLLVMContext(), CGF.getContext().getTypeSize(T)); 178 llvm::Type *IntPtrType = IntType->getPointerTo(AddrSpace); 179 180 Value *Args[3]; 181 Args[0] = CGF.Builder.CreateBitCast(DestPtr, IntPtrType); 182 Args[1] = CGF.EmitScalarExpr(E->getArg(1)); 183 llvm::Type *ValueType = Args[1]->getType(); 184 Args[1] = EmitToInt(CGF, Args[1], T, IntType); 185 Args[2] = EmitToInt(CGF, CGF.EmitScalarExpr(E->getArg(2)), T, IntType); 186 187 Value *Pair = CGF.Builder.CreateAtomicCmpXchg(Args[0], Args[1], Args[2], 188 llvm::SequentiallyConsistent, 189 llvm::SequentiallyConsistent); 190 if (ReturnBool) 191 // Extract boolean success flag and zext it to int. 192 return CGF.Builder.CreateZExt(CGF.Builder.CreateExtractValue(Pair, 1), 193 CGF.ConvertType(E->getType())); 194 else 195 // Extract old value and emit it using the same type as compare value. 196 return EmitFromInt(CGF, CGF.Builder.CreateExtractValue(Pair, 0), T, 197 ValueType); 198 } 199 200 /// EmitFAbs - Emit a call to @llvm.fabs(). 201 static Value *EmitFAbs(CodeGenFunction &CGF, Value *V) { 202 Value *F = CGF.CGM.getIntrinsic(Intrinsic::fabs, V->getType()); 203 llvm::CallInst *Call = CGF.Builder.CreateCall(F, V); 204 Call->setDoesNotAccessMemory(); 205 return Call; 206 } 207 208 /// Emit the computation of the sign bit for a floating point value. Returns 209 /// the i1 sign bit value. 210 static Value *EmitSignBit(CodeGenFunction &CGF, Value *V) { 211 LLVMContext &C = CGF.CGM.getLLVMContext(); 212 213 llvm::Type *Ty = V->getType(); 214 int Width = Ty->getPrimitiveSizeInBits(); 215 llvm::Type *IntTy = llvm::IntegerType::get(C, Width); 216 V = CGF.Builder.CreateBitCast(V, IntTy); 217 if (Ty->isPPC_FP128Ty()) { 218 // The higher-order double comes first, and so we need to truncate the 219 // pair to extract the overall sign. The order of the pair is the same 220 // in both little- and big-Endian modes. 221 Width >>= 1; 222 IntTy = llvm::IntegerType::get(C, Width); 223 V = CGF.Builder.CreateTrunc(V, IntTy); 224 } 225 Value *Zero = llvm::Constant::getNullValue(IntTy); 226 return CGF.Builder.CreateICmpSLT(V, Zero); 227 } 228 229 static RValue emitLibraryCall(CodeGenFunction &CGF, const FunctionDecl *Fn, 230 const CallExpr *E, llvm::Value *calleeValue) { 231 return CGF.EmitCall(E->getCallee()->getType(), calleeValue, E, 232 ReturnValueSlot(), Fn); 233 } 234 235 /// \brief Emit a call to llvm.{sadd,uadd,ssub,usub,smul,umul}.with.overflow.* 236 /// depending on IntrinsicID. 237 /// 238 /// \arg CGF The current codegen function. 239 /// \arg IntrinsicID The ID for the Intrinsic we wish to generate. 240 /// \arg X The first argument to the llvm.*.with.overflow.*. 241 /// \arg Y The second argument to the llvm.*.with.overflow.*. 242 /// \arg Carry The carry returned by the llvm.*.with.overflow.*. 243 /// \returns The result (i.e. sum/product) returned by the intrinsic. 244 static llvm::Value *EmitOverflowIntrinsic(CodeGenFunction &CGF, 245 const llvm::Intrinsic::ID IntrinsicID, 246 llvm::Value *X, llvm::Value *Y, 247 llvm::Value *&Carry) { 248 // Make sure we have integers of the same width. 249 assert(X->getType() == Y->getType() && 250 "Arguments must be the same type. (Did you forget to make sure both " 251 "arguments have the same integer width?)"); 252 253 llvm::Value *Callee = CGF.CGM.getIntrinsic(IntrinsicID, X->getType()); 254 llvm::Value *Tmp = CGF.Builder.CreateCall(Callee, {X, Y}); 255 Carry = CGF.Builder.CreateExtractValue(Tmp, 1); 256 return CGF.Builder.CreateExtractValue(Tmp, 0); 257 } 258 259 RValue CodeGenFunction::EmitBuiltinExpr(const FunctionDecl *FD, 260 unsigned BuiltinID, const CallExpr *E, 261 ReturnValueSlot ReturnValue) { 262 // See if we can constant fold this builtin. If so, don't emit it at all. 263 Expr::EvalResult Result; 264 if (E->EvaluateAsRValue(Result, CGM.getContext()) && 265 !Result.hasSideEffects()) { 266 if (Result.Val.isInt()) 267 return RValue::get(llvm::ConstantInt::get(getLLVMContext(), 268 Result.Val.getInt())); 269 if (Result.Val.isFloat()) 270 return RValue::get(llvm::ConstantFP::get(getLLVMContext(), 271 Result.Val.getFloat())); 272 } 273 274 switch (BuiltinID) { 275 default: break; // Handle intrinsics and libm functions below. 276 case Builtin::BI__builtin___CFStringMakeConstantString: 277 case Builtin::BI__builtin___NSStringMakeConstantString: 278 return RValue::get(CGM.EmitConstantExpr(E, E->getType(), nullptr)); 279 case Builtin::BI__builtin_stdarg_start: 280 case Builtin::BI__builtin_va_start: 281 case Builtin::BI__va_start: 282 case Builtin::BI__builtin_va_end: { 283 Value *ArgValue = (BuiltinID == Builtin::BI__va_start) 284 ? EmitScalarExpr(E->getArg(0)) 285 : EmitVAListRef(E->getArg(0)); 286 llvm::Type *DestType = Int8PtrTy; 287 if (ArgValue->getType() != DestType) 288 ArgValue = Builder.CreateBitCast(ArgValue, DestType, 289 ArgValue->getName().data()); 290 291 Intrinsic::ID inst = (BuiltinID == Builtin::BI__builtin_va_end) ? 292 Intrinsic::vaend : Intrinsic::vastart; 293 return RValue::get(Builder.CreateCall(CGM.getIntrinsic(inst), ArgValue)); 294 } 295 case Builtin::BI__builtin_va_copy: { 296 Value *DstPtr = EmitVAListRef(E->getArg(0)); 297 Value *SrcPtr = EmitVAListRef(E->getArg(1)); 298 299 llvm::Type *Type = Int8PtrTy; 300 301 DstPtr = Builder.CreateBitCast(DstPtr, Type); 302 SrcPtr = Builder.CreateBitCast(SrcPtr, Type); 303 return RValue::get(Builder.CreateCall(CGM.getIntrinsic(Intrinsic::vacopy), 304 {DstPtr, SrcPtr})); 305 } 306 case Builtin::BI__builtin_abs: 307 case Builtin::BI__builtin_labs: 308 case Builtin::BI__builtin_llabs: { 309 Value *ArgValue = EmitScalarExpr(E->getArg(0)); 310 311 Value *NegOp = Builder.CreateNeg(ArgValue, "neg"); 312 Value *CmpResult = 313 Builder.CreateICmpSGE(ArgValue, 314 llvm::Constant::getNullValue(ArgValue->getType()), 315 "abscond"); 316 Value *Result = 317 Builder.CreateSelect(CmpResult, ArgValue, NegOp, "abs"); 318 319 return RValue::get(Result); 320 } 321 case Builtin::BI__builtin_fabs: 322 case Builtin::BI__builtin_fabsf: 323 case Builtin::BI__builtin_fabsl: { 324 Value *Arg1 = EmitScalarExpr(E->getArg(0)); 325 Value *Result = EmitFAbs(*this, Arg1); 326 return RValue::get(Result); 327 } 328 case Builtin::BI__builtin_fmod: 329 case Builtin::BI__builtin_fmodf: 330 case Builtin::BI__builtin_fmodl: { 331 Value *Arg1 = EmitScalarExpr(E->getArg(0)); 332 Value *Arg2 = EmitScalarExpr(E->getArg(1)); 333 Value *Result = Builder.CreateFRem(Arg1, Arg2, "fmod"); 334 return RValue::get(Result); 335 } 336 337 case Builtin::BI__builtin_conj: 338 case Builtin::BI__builtin_conjf: 339 case Builtin::BI__builtin_conjl: { 340 ComplexPairTy ComplexVal = EmitComplexExpr(E->getArg(0)); 341 Value *Real = ComplexVal.first; 342 Value *Imag = ComplexVal.second; 343 Value *Zero = 344 Imag->getType()->isFPOrFPVectorTy() 345 ? llvm::ConstantFP::getZeroValueForNegation(Imag->getType()) 346 : llvm::Constant::getNullValue(Imag->getType()); 347 348 Imag = Builder.CreateFSub(Zero, Imag, "sub"); 349 return RValue::getComplex(std::make_pair(Real, Imag)); 350 } 351 case Builtin::BI__builtin_creal: 352 case Builtin::BI__builtin_crealf: 353 case Builtin::BI__builtin_creall: 354 case Builtin::BIcreal: 355 case Builtin::BIcrealf: 356 case Builtin::BIcreall: { 357 ComplexPairTy ComplexVal = EmitComplexExpr(E->getArg(0)); 358 return RValue::get(ComplexVal.first); 359 } 360 361 case Builtin::BI__builtin_cimag: 362 case Builtin::BI__builtin_cimagf: 363 case Builtin::BI__builtin_cimagl: 364 case Builtin::BIcimag: 365 case Builtin::BIcimagf: 366 case Builtin::BIcimagl: { 367 ComplexPairTy ComplexVal = EmitComplexExpr(E->getArg(0)); 368 return RValue::get(ComplexVal.second); 369 } 370 371 case Builtin::BI__builtin_ctzs: 372 case Builtin::BI__builtin_ctz: 373 case Builtin::BI__builtin_ctzl: 374 case Builtin::BI__builtin_ctzll: { 375 Value *ArgValue = EmitScalarExpr(E->getArg(0)); 376 377 llvm::Type *ArgType = ArgValue->getType(); 378 Value *F = CGM.getIntrinsic(Intrinsic::cttz, ArgType); 379 380 llvm::Type *ResultType = ConvertType(E->getType()); 381 Value *ZeroUndef = Builder.getInt1(getTarget().isCLZForZeroUndef()); 382 Value *Result = Builder.CreateCall(F, {ArgValue, ZeroUndef}); 383 if (Result->getType() != ResultType) 384 Result = Builder.CreateIntCast(Result, ResultType, /*isSigned*/true, 385 "cast"); 386 return RValue::get(Result); 387 } 388 case Builtin::BI__builtin_clzs: 389 case Builtin::BI__builtin_clz: 390 case Builtin::BI__builtin_clzl: 391 case Builtin::BI__builtin_clzll: { 392 Value *ArgValue = EmitScalarExpr(E->getArg(0)); 393 394 llvm::Type *ArgType = ArgValue->getType(); 395 Value *F = CGM.getIntrinsic(Intrinsic::ctlz, ArgType); 396 397 llvm::Type *ResultType = ConvertType(E->getType()); 398 Value *ZeroUndef = Builder.getInt1(getTarget().isCLZForZeroUndef()); 399 Value *Result = Builder.CreateCall(F, {ArgValue, ZeroUndef}); 400 if (Result->getType() != ResultType) 401 Result = Builder.CreateIntCast(Result, ResultType, /*isSigned*/true, 402 "cast"); 403 return RValue::get(Result); 404 } 405 case Builtin::BI__builtin_ffs: 406 case Builtin::BI__builtin_ffsl: 407 case Builtin::BI__builtin_ffsll: { 408 // ffs(x) -> x ? cttz(x) + 1 : 0 409 Value *ArgValue = EmitScalarExpr(E->getArg(0)); 410 411 llvm::Type *ArgType = ArgValue->getType(); 412 Value *F = CGM.getIntrinsic(Intrinsic::cttz, ArgType); 413 414 llvm::Type *ResultType = ConvertType(E->getType()); 415 Value *Tmp = 416 Builder.CreateAdd(Builder.CreateCall(F, {ArgValue, Builder.getTrue()}), 417 llvm::ConstantInt::get(ArgType, 1)); 418 Value *Zero = llvm::Constant::getNullValue(ArgType); 419 Value *IsZero = Builder.CreateICmpEQ(ArgValue, Zero, "iszero"); 420 Value *Result = Builder.CreateSelect(IsZero, Zero, Tmp, "ffs"); 421 if (Result->getType() != ResultType) 422 Result = Builder.CreateIntCast(Result, ResultType, /*isSigned*/true, 423 "cast"); 424 return RValue::get(Result); 425 } 426 case Builtin::BI__builtin_parity: 427 case Builtin::BI__builtin_parityl: 428 case Builtin::BI__builtin_parityll: { 429 // parity(x) -> ctpop(x) & 1 430 Value *ArgValue = EmitScalarExpr(E->getArg(0)); 431 432 llvm::Type *ArgType = ArgValue->getType(); 433 Value *F = CGM.getIntrinsic(Intrinsic::ctpop, ArgType); 434 435 llvm::Type *ResultType = ConvertType(E->getType()); 436 Value *Tmp = Builder.CreateCall(F, ArgValue); 437 Value *Result = Builder.CreateAnd(Tmp, llvm::ConstantInt::get(ArgType, 1)); 438 if (Result->getType() != ResultType) 439 Result = Builder.CreateIntCast(Result, ResultType, /*isSigned*/true, 440 "cast"); 441 return RValue::get(Result); 442 } 443 case Builtin::BI__builtin_popcount: 444 case Builtin::BI__builtin_popcountl: 445 case Builtin::BI__builtin_popcountll: { 446 Value *ArgValue = EmitScalarExpr(E->getArg(0)); 447 448 llvm::Type *ArgType = ArgValue->getType(); 449 Value *F = CGM.getIntrinsic(Intrinsic::ctpop, ArgType); 450 451 llvm::Type *ResultType = ConvertType(E->getType()); 452 Value *Result = Builder.CreateCall(F, ArgValue); 453 if (Result->getType() != ResultType) 454 Result = Builder.CreateIntCast(Result, ResultType, /*isSigned*/true, 455 "cast"); 456 return RValue::get(Result); 457 } 458 case Builtin::BI__builtin_expect: { 459 Value *ArgValue = EmitScalarExpr(E->getArg(0)); 460 llvm::Type *ArgType = ArgValue->getType(); 461 462 Value *ExpectedValue = EmitScalarExpr(E->getArg(1)); 463 // Don't generate llvm.expect on -O0 as the backend won't use it for 464 // anything. 465 // Note, we still IRGen ExpectedValue because it could have side-effects. 466 if (CGM.getCodeGenOpts().OptimizationLevel == 0) 467 return RValue::get(ArgValue); 468 469 Value *FnExpect = CGM.getIntrinsic(Intrinsic::expect, ArgType); 470 Value *Result = 471 Builder.CreateCall(FnExpect, {ArgValue, ExpectedValue}, "expval"); 472 return RValue::get(Result); 473 } 474 case Builtin::BI__builtin_assume_aligned: { 475 Value *PtrValue = EmitScalarExpr(E->getArg(0)); 476 Value *OffsetValue = 477 (E->getNumArgs() > 2) ? EmitScalarExpr(E->getArg(2)) : nullptr; 478 479 Value *AlignmentValue = EmitScalarExpr(E->getArg(1)); 480 ConstantInt *AlignmentCI = cast<ConstantInt>(AlignmentValue); 481 unsigned Alignment = (unsigned) AlignmentCI->getZExtValue(); 482 483 EmitAlignmentAssumption(PtrValue, Alignment, OffsetValue); 484 return RValue::get(PtrValue); 485 } 486 case Builtin::BI__assume: 487 case Builtin::BI__builtin_assume: { 488 if (E->getArg(0)->HasSideEffects(getContext())) 489 return RValue::get(nullptr); 490 491 Value *ArgValue = EmitScalarExpr(E->getArg(0)); 492 Value *FnAssume = CGM.getIntrinsic(Intrinsic::assume); 493 return RValue::get(Builder.CreateCall(FnAssume, ArgValue)); 494 } 495 case Builtin::BI__builtin_bswap16: 496 case Builtin::BI__builtin_bswap32: 497 case Builtin::BI__builtin_bswap64: { 498 Value *ArgValue = EmitScalarExpr(E->getArg(0)); 499 llvm::Type *ArgType = ArgValue->getType(); 500 Value *F = CGM.getIntrinsic(Intrinsic::bswap, ArgType); 501 return RValue::get(Builder.CreateCall(F, ArgValue)); 502 } 503 case Builtin::BI__builtin_object_size: { 504 // We rely on constant folding to deal with expressions with side effects. 505 assert(!E->getArg(0)->HasSideEffects(getContext()) && 506 "should have been constant folded"); 507 508 // We pass this builtin onto the optimizer so that it can 509 // figure out the object size in more complex cases. 510 llvm::Type *ResType = ConvertType(E->getType()); 511 512 // LLVM only supports 0 and 2, make sure that we pass along that 513 // as a boolean. 514 Value *Ty = EmitScalarExpr(E->getArg(1)); 515 ConstantInt *CI = dyn_cast<ConstantInt>(Ty); 516 assert(CI); 517 uint64_t val = CI->getZExtValue(); 518 CI = ConstantInt::get(Builder.getInt1Ty(), (val & 0x2) >> 1); 519 // FIXME: Get right address space. 520 llvm::Type *Tys[] = { ResType, Builder.getInt8PtrTy(0) }; 521 Value *F = CGM.getIntrinsic(Intrinsic::objectsize, Tys); 522 return RValue::get( 523 Builder.CreateCall(F, {EmitScalarExpr(E->getArg(0)), CI})); 524 } 525 case Builtin::BI__builtin_prefetch: { 526 Value *Locality, *RW, *Address = EmitScalarExpr(E->getArg(0)); 527 // FIXME: Technically these constants should of type 'int', yes? 528 RW = (E->getNumArgs() > 1) ? EmitScalarExpr(E->getArg(1)) : 529 llvm::ConstantInt::get(Int32Ty, 0); 530 Locality = (E->getNumArgs() > 2) ? EmitScalarExpr(E->getArg(2)) : 531 llvm::ConstantInt::get(Int32Ty, 3); 532 Value *Data = llvm::ConstantInt::get(Int32Ty, 1); 533 Value *F = CGM.getIntrinsic(Intrinsic::prefetch); 534 return RValue::get(Builder.CreateCall(F, {Address, RW, Locality, Data})); 535 } 536 case Builtin::BI__builtin_readcyclecounter: { 537 Value *F = CGM.getIntrinsic(Intrinsic::readcyclecounter); 538 return RValue::get(Builder.CreateCall(F)); 539 } 540 case Builtin::BI__builtin___clear_cache: { 541 Value *Begin = EmitScalarExpr(E->getArg(0)); 542 Value *End = EmitScalarExpr(E->getArg(1)); 543 Value *F = CGM.getIntrinsic(Intrinsic::clear_cache); 544 return RValue::get(Builder.CreateCall(F, {Begin, End})); 545 } 546 case Builtin::BI__builtin_trap: 547 return RValue::get(EmitTrapCall(Intrinsic::trap)); 548 case Builtin::BI__debugbreak: 549 return RValue::get(EmitTrapCall(Intrinsic::debugtrap)); 550 case Builtin::BI__builtin_unreachable: { 551 if (SanOpts.has(SanitizerKind::Unreachable)) { 552 SanitizerScope SanScope(this); 553 EmitCheck(std::make_pair(static_cast<llvm::Value *>(Builder.getFalse()), 554 SanitizerKind::Unreachable), 555 "builtin_unreachable", EmitCheckSourceLocation(E->getExprLoc()), 556 None); 557 } else 558 Builder.CreateUnreachable(); 559 560 // We do need to preserve an insertion point. 561 EmitBlock(createBasicBlock("unreachable.cont")); 562 563 return RValue::get(nullptr); 564 } 565 566 case Builtin::BI__builtin_powi: 567 case Builtin::BI__builtin_powif: 568 case Builtin::BI__builtin_powil: { 569 Value *Base = EmitScalarExpr(E->getArg(0)); 570 Value *Exponent = EmitScalarExpr(E->getArg(1)); 571 llvm::Type *ArgType = Base->getType(); 572 Value *F = CGM.getIntrinsic(Intrinsic::powi, ArgType); 573 return RValue::get(Builder.CreateCall(F, {Base, Exponent})); 574 } 575 576 case Builtin::BI__builtin_isgreater: 577 case Builtin::BI__builtin_isgreaterequal: 578 case Builtin::BI__builtin_isless: 579 case Builtin::BI__builtin_islessequal: 580 case Builtin::BI__builtin_islessgreater: 581 case Builtin::BI__builtin_isunordered: { 582 // Ordered comparisons: we know the arguments to these are matching scalar 583 // floating point values. 584 Value *LHS = EmitScalarExpr(E->getArg(0)); 585 Value *RHS = EmitScalarExpr(E->getArg(1)); 586 587 switch (BuiltinID) { 588 default: llvm_unreachable("Unknown ordered comparison"); 589 case Builtin::BI__builtin_isgreater: 590 LHS = Builder.CreateFCmpOGT(LHS, RHS, "cmp"); 591 break; 592 case Builtin::BI__builtin_isgreaterequal: 593 LHS = Builder.CreateFCmpOGE(LHS, RHS, "cmp"); 594 break; 595 case Builtin::BI__builtin_isless: 596 LHS = Builder.CreateFCmpOLT(LHS, RHS, "cmp"); 597 break; 598 case Builtin::BI__builtin_islessequal: 599 LHS = Builder.CreateFCmpOLE(LHS, RHS, "cmp"); 600 break; 601 case Builtin::BI__builtin_islessgreater: 602 LHS = Builder.CreateFCmpONE(LHS, RHS, "cmp"); 603 break; 604 case Builtin::BI__builtin_isunordered: 605 LHS = Builder.CreateFCmpUNO(LHS, RHS, "cmp"); 606 break; 607 } 608 // ZExt bool to int type. 609 return RValue::get(Builder.CreateZExt(LHS, ConvertType(E->getType()))); 610 } 611 case Builtin::BI__builtin_isnan: { 612 Value *V = EmitScalarExpr(E->getArg(0)); 613 V = Builder.CreateFCmpUNO(V, V, "cmp"); 614 return RValue::get(Builder.CreateZExt(V, ConvertType(E->getType()))); 615 } 616 617 case Builtin::BI__builtin_isinf: { 618 // isinf(x) --> fabs(x) == infinity 619 Value *V = EmitScalarExpr(E->getArg(0)); 620 V = EmitFAbs(*this, V); 621 622 V = Builder.CreateFCmpOEQ(V, ConstantFP::getInfinity(V->getType()),"isinf"); 623 return RValue::get(Builder.CreateZExt(V, ConvertType(E->getType()))); 624 } 625 626 case Builtin::BI__builtin_isinf_sign: { 627 // isinf_sign(x) -> fabs(x) == infinity ? (signbit(x) ? -1 : 1) : 0 628 Value *Arg = EmitScalarExpr(E->getArg(0)); 629 Value *AbsArg = EmitFAbs(*this, Arg); 630 Value *IsInf = Builder.CreateFCmpOEQ( 631 AbsArg, ConstantFP::getInfinity(Arg->getType()), "isinf"); 632 Value *IsNeg = EmitSignBit(*this, Arg); 633 634 llvm::Type *IntTy = ConvertType(E->getType()); 635 Value *Zero = Constant::getNullValue(IntTy); 636 Value *One = ConstantInt::get(IntTy, 1); 637 Value *NegativeOne = ConstantInt::get(IntTy, -1); 638 Value *SignResult = Builder.CreateSelect(IsNeg, NegativeOne, One); 639 Value *Result = Builder.CreateSelect(IsInf, SignResult, Zero); 640 return RValue::get(Result); 641 } 642 643 case Builtin::BI__builtin_isnormal: { 644 // isnormal(x) --> x == x && fabsf(x) < infinity && fabsf(x) >= float_min 645 Value *V = EmitScalarExpr(E->getArg(0)); 646 Value *Eq = Builder.CreateFCmpOEQ(V, V, "iseq"); 647 648 Value *Abs = EmitFAbs(*this, V); 649 Value *IsLessThanInf = 650 Builder.CreateFCmpULT(Abs, ConstantFP::getInfinity(V->getType()),"isinf"); 651 APFloat Smallest = APFloat::getSmallestNormalized( 652 getContext().getFloatTypeSemantics(E->getArg(0)->getType())); 653 Value *IsNormal = 654 Builder.CreateFCmpUGE(Abs, ConstantFP::get(V->getContext(), Smallest), 655 "isnormal"); 656 V = Builder.CreateAnd(Eq, IsLessThanInf, "and"); 657 V = Builder.CreateAnd(V, IsNormal, "and"); 658 return RValue::get(Builder.CreateZExt(V, ConvertType(E->getType()))); 659 } 660 661 case Builtin::BI__builtin_isfinite: { 662 // isfinite(x) --> x == x && fabs(x) != infinity; 663 Value *V = EmitScalarExpr(E->getArg(0)); 664 Value *Eq = Builder.CreateFCmpOEQ(V, V, "iseq"); 665 666 Value *Abs = EmitFAbs(*this, V); 667 Value *IsNotInf = 668 Builder.CreateFCmpUNE(Abs, ConstantFP::getInfinity(V->getType()),"isinf"); 669 670 V = Builder.CreateAnd(Eq, IsNotInf, "and"); 671 return RValue::get(Builder.CreateZExt(V, ConvertType(E->getType()))); 672 } 673 674 case Builtin::BI__builtin_fpclassify: { 675 Value *V = EmitScalarExpr(E->getArg(5)); 676 llvm::Type *Ty = ConvertType(E->getArg(5)->getType()); 677 678 // Create Result 679 BasicBlock *Begin = Builder.GetInsertBlock(); 680 BasicBlock *End = createBasicBlock("fpclassify_end", this->CurFn); 681 Builder.SetInsertPoint(End); 682 PHINode *Result = 683 Builder.CreatePHI(ConvertType(E->getArg(0)->getType()), 4, 684 "fpclassify_result"); 685 686 // if (V==0) return FP_ZERO 687 Builder.SetInsertPoint(Begin); 688 Value *IsZero = Builder.CreateFCmpOEQ(V, Constant::getNullValue(Ty), 689 "iszero"); 690 Value *ZeroLiteral = EmitScalarExpr(E->getArg(4)); 691 BasicBlock *NotZero = createBasicBlock("fpclassify_not_zero", this->CurFn); 692 Builder.CreateCondBr(IsZero, End, NotZero); 693 Result->addIncoming(ZeroLiteral, Begin); 694 695 // if (V != V) return FP_NAN 696 Builder.SetInsertPoint(NotZero); 697 Value *IsNan = Builder.CreateFCmpUNO(V, V, "cmp"); 698 Value *NanLiteral = EmitScalarExpr(E->getArg(0)); 699 BasicBlock *NotNan = createBasicBlock("fpclassify_not_nan", this->CurFn); 700 Builder.CreateCondBr(IsNan, End, NotNan); 701 Result->addIncoming(NanLiteral, NotZero); 702 703 // if (fabs(V) == infinity) return FP_INFINITY 704 Builder.SetInsertPoint(NotNan); 705 Value *VAbs = EmitFAbs(*this, V); 706 Value *IsInf = 707 Builder.CreateFCmpOEQ(VAbs, ConstantFP::getInfinity(V->getType()), 708 "isinf"); 709 Value *InfLiteral = EmitScalarExpr(E->getArg(1)); 710 BasicBlock *NotInf = createBasicBlock("fpclassify_not_inf", this->CurFn); 711 Builder.CreateCondBr(IsInf, End, NotInf); 712 Result->addIncoming(InfLiteral, NotNan); 713 714 // if (fabs(V) >= MIN_NORMAL) return FP_NORMAL else FP_SUBNORMAL 715 Builder.SetInsertPoint(NotInf); 716 APFloat Smallest = APFloat::getSmallestNormalized( 717 getContext().getFloatTypeSemantics(E->getArg(5)->getType())); 718 Value *IsNormal = 719 Builder.CreateFCmpUGE(VAbs, ConstantFP::get(V->getContext(), Smallest), 720 "isnormal"); 721 Value *NormalResult = 722 Builder.CreateSelect(IsNormal, EmitScalarExpr(E->getArg(2)), 723 EmitScalarExpr(E->getArg(3))); 724 Builder.CreateBr(End); 725 Result->addIncoming(NormalResult, NotInf); 726 727 // return Result 728 Builder.SetInsertPoint(End); 729 return RValue::get(Result); 730 } 731 732 case Builtin::BIalloca: 733 case Builtin::BI_alloca: 734 case Builtin::BI__builtin_alloca: { 735 Value *Size = EmitScalarExpr(E->getArg(0)); 736 return RValue::get(Builder.CreateAlloca(Builder.getInt8Ty(), Size)); 737 } 738 case Builtin::BIbzero: 739 case Builtin::BI__builtin_bzero: { 740 std::pair<llvm::Value*, unsigned> Dest = 741 EmitPointerWithAlignment(E->getArg(0)); 742 Value *SizeVal = EmitScalarExpr(E->getArg(1)); 743 EmitNonNullArgCheck(RValue::get(Dest.first), E->getArg(0)->getType(), 744 E->getArg(0)->getExprLoc(), FD, 0); 745 Builder.CreateMemSet(Dest.first, Builder.getInt8(0), SizeVal, 746 Dest.second, false); 747 return RValue::get(Dest.first); 748 } 749 case Builtin::BImemcpy: 750 case Builtin::BI__builtin_memcpy: { 751 std::pair<llvm::Value*, unsigned> Dest = 752 EmitPointerWithAlignment(E->getArg(0)); 753 std::pair<llvm::Value*, unsigned> Src = 754 EmitPointerWithAlignment(E->getArg(1)); 755 Value *SizeVal = EmitScalarExpr(E->getArg(2)); 756 unsigned Align = std::min(Dest.second, Src.second); 757 EmitNonNullArgCheck(RValue::get(Dest.first), E->getArg(0)->getType(), 758 E->getArg(0)->getExprLoc(), FD, 0); 759 EmitNonNullArgCheck(RValue::get(Src.first), E->getArg(1)->getType(), 760 E->getArg(1)->getExprLoc(), FD, 1); 761 Builder.CreateMemCpy(Dest.first, Src.first, SizeVal, Align, false); 762 return RValue::get(Dest.first); 763 } 764 765 case Builtin::BI__builtin___memcpy_chk: { 766 // fold __builtin_memcpy_chk(x, y, cst1, cst2) to memcpy iff cst1<=cst2. 767 llvm::APSInt Size, DstSize; 768 if (!E->getArg(2)->EvaluateAsInt(Size, CGM.getContext()) || 769 !E->getArg(3)->EvaluateAsInt(DstSize, CGM.getContext())) 770 break; 771 if (Size.ugt(DstSize)) 772 break; 773 std::pair<llvm::Value*, unsigned> Dest = 774 EmitPointerWithAlignment(E->getArg(0)); 775 std::pair<llvm::Value*, unsigned> Src = 776 EmitPointerWithAlignment(E->getArg(1)); 777 Value *SizeVal = llvm::ConstantInt::get(Builder.getContext(), Size); 778 unsigned Align = std::min(Dest.second, Src.second); 779 Builder.CreateMemCpy(Dest.first, Src.first, SizeVal, Align, false); 780 return RValue::get(Dest.first); 781 } 782 783 case Builtin::BI__builtin_objc_memmove_collectable: { 784 Value *Address = EmitScalarExpr(E->getArg(0)); 785 Value *SrcAddr = EmitScalarExpr(E->getArg(1)); 786 Value *SizeVal = EmitScalarExpr(E->getArg(2)); 787 CGM.getObjCRuntime().EmitGCMemmoveCollectable(*this, 788 Address, SrcAddr, SizeVal); 789 return RValue::get(Address); 790 } 791 792 case Builtin::BI__builtin___memmove_chk: { 793 // fold __builtin_memmove_chk(x, y, cst1, cst2) to memmove iff cst1<=cst2. 794 llvm::APSInt Size, DstSize; 795 if (!E->getArg(2)->EvaluateAsInt(Size, CGM.getContext()) || 796 !E->getArg(3)->EvaluateAsInt(DstSize, CGM.getContext())) 797 break; 798 if (Size.ugt(DstSize)) 799 break; 800 std::pair<llvm::Value*, unsigned> Dest = 801 EmitPointerWithAlignment(E->getArg(0)); 802 std::pair<llvm::Value*, unsigned> Src = 803 EmitPointerWithAlignment(E->getArg(1)); 804 Value *SizeVal = llvm::ConstantInt::get(Builder.getContext(), Size); 805 unsigned Align = std::min(Dest.second, Src.second); 806 Builder.CreateMemMove(Dest.first, Src.first, SizeVal, Align, false); 807 return RValue::get(Dest.first); 808 } 809 810 case Builtin::BImemmove: 811 case Builtin::BI__builtin_memmove: { 812 std::pair<llvm::Value*, unsigned> Dest = 813 EmitPointerWithAlignment(E->getArg(0)); 814 std::pair<llvm::Value*, unsigned> Src = 815 EmitPointerWithAlignment(E->getArg(1)); 816 Value *SizeVal = EmitScalarExpr(E->getArg(2)); 817 unsigned Align = std::min(Dest.second, Src.second); 818 EmitNonNullArgCheck(RValue::get(Dest.first), E->getArg(0)->getType(), 819 E->getArg(0)->getExprLoc(), FD, 0); 820 EmitNonNullArgCheck(RValue::get(Src.first), E->getArg(1)->getType(), 821 E->getArg(1)->getExprLoc(), FD, 1); 822 Builder.CreateMemMove(Dest.first, Src.first, SizeVal, Align, false); 823 return RValue::get(Dest.first); 824 } 825 case Builtin::BImemset: 826 case Builtin::BI__builtin_memset: { 827 std::pair<llvm::Value*, unsigned> Dest = 828 EmitPointerWithAlignment(E->getArg(0)); 829 Value *ByteVal = Builder.CreateTrunc(EmitScalarExpr(E->getArg(1)), 830 Builder.getInt8Ty()); 831 Value *SizeVal = EmitScalarExpr(E->getArg(2)); 832 EmitNonNullArgCheck(RValue::get(Dest.first), E->getArg(0)->getType(), 833 E->getArg(0)->getExprLoc(), FD, 0); 834 Builder.CreateMemSet(Dest.first, ByteVal, SizeVal, Dest.second, false); 835 return RValue::get(Dest.first); 836 } 837 case Builtin::BI__builtin___memset_chk: { 838 // fold __builtin_memset_chk(x, y, cst1, cst2) to memset iff cst1<=cst2. 839 llvm::APSInt Size, DstSize; 840 if (!E->getArg(2)->EvaluateAsInt(Size, CGM.getContext()) || 841 !E->getArg(3)->EvaluateAsInt(DstSize, CGM.getContext())) 842 break; 843 if (Size.ugt(DstSize)) 844 break; 845 std::pair<llvm::Value*, unsigned> Dest = 846 EmitPointerWithAlignment(E->getArg(0)); 847 Value *ByteVal = Builder.CreateTrunc(EmitScalarExpr(E->getArg(1)), 848 Builder.getInt8Ty()); 849 Value *SizeVal = llvm::ConstantInt::get(Builder.getContext(), Size); 850 Builder.CreateMemSet(Dest.first, ByteVal, SizeVal, Dest.second, false); 851 return RValue::get(Dest.first); 852 } 853 case Builtin::BI__builtin_dwarf_cfa: { 854 // The offset in bytes from the first argument to the CFA. 855 // 856 // Why on earth is this in the frontend? Is there any reason at 857 // all that the backend can't reasonably determine this while 858 // lowering llvm.eh.dwarf.cfa()? 859 // 860 // TODO: If there's a satisfactory reason, add a target hook for 861 // this instead of hard-coding 0, which is correct for most targets. 862 int32_t Offset = 0; 863 864 Value *F = CGM.getIntrinsic(Intrinsic::eh_dwarf_cfa); 865 return RValue::get(Builder.CreateCall(F, 866 llvm::ConstantInt::get(Int32Ty, Offset))); 867 } 868 case Builtin::BI__builtin_return_address: { 869 Value *Depth = 870 CGM.EmitConstantExpr(E->getArg(0), getContext().UnsignedIntTy, this); 871 Value *F = CGM.getIntrinsic(Intrinsic::returnaddress); 872 return RValue::get(Builder.CreateCall(F, Depth)); 873 } 874 case Builtin::BI__builtin_frame_address: { 875 Value *Depth = 876 CGM.EmitConstantExpr(E->getArg(0), getContext().UnsignedIntTy, this); 877 Value *F = CGM.getIntrinsic(Intrinsic::frameaddress); 878 return RValue::get(Builder.CreateCall(F, Depth)); 879 } 880 case Builtin::BI__builtin_extract_return_addr: { 881 Value *Address = EmitScalarExpr(E->getArg(0)); 882 Value *Result = getTargetHooks().decodeReturnAddress(*this, Address); 883 return RValue::get(Result); 884 } 885 case Builtin::BI__builtin_frob_return_addr: { 886 Value *Address = EmitScalarExpr(E->getArg(0)); 887 Value *Result = getTargetHooks().encodeReturnAddress(*this, Address); 888 return RValue::get(Result); 889 } 890 case Builtin::BI__builtin_dwarf_sp_column: { 891 llvm::IntegerType *Ty 892 = cast<llvm::IntegerType>(ConvertType(E->getType())); 893 int Column = getTargetHooks().getDwarfEHStackPointer(CGM); 894 if (Column == -1) { 895 CGM.ErrorUnsupported(E, "__builtin_dwarf_sp_column"); 896 return RValue::get(llvm::UndefValue::get(Ty)); 897 } 898 return RValue::get(llvm::ConstantInt::get(Ty, Column, true)); 899 } 900 case Builtin::BI__builtin_init_dwarf_reg_size_table: { 901 Value *Address = EmitScalarExpr(E->getArg(0)); 902 if (getTargetHooks().initDwarfEHRegSizeTable(*this, Address)) 903 CGM.ErrorUnsupported(E, "__builtin_init_dwarf_reg_size_table"); 904 return RValue::get(llvm::UndefValue::get(ConvertType(E->getType()))); 905 } 906 case Builtin::BI__builtin_eh_return: { 907 Value *Int = EmitScalarExpr(E->getArg(0)); 908 Value *Ptr = EmitScalarExpr(E->getArg(1)); 909 910 llvm::IntegerType *IntTy = cast<llvm::IntegerType>(Int->getType()); 911 assert((IntTy->getBitWidth() == 32 || IntTy->getBitWidth() == 64) && 912 "LLVM's __builtin_eh_return only supports 32- and 64-bit variants"); 913 Value *F = CGM.getIntrinsic(IntTy->getBitWidth() == 32 914 ? Intrinsic::eh_return_i32 915 : Intrinsic::eh_return_i64); 916 Builder.CreateCall(F, {Int, Ptr}); 917 Builder.CreateUnreachable(); 918 919 // We do need to preserve an insertion point. 920 EmitBlock(createBasicBlock("builtin_eh_return.cont")); 921 922 return RValue::get(nullptr); 923 } 924 case Builtin::BI__builtin_unwind_init: { 925 Value *F = CGM.getIntrinsic(Intrinsic::eh_unwind_init); 926 return RValue::get(Builder.CreateCall(F)); 927 } 928 case Builtin::BI__builtin_extend_pointer: { 929 // Extends a pointer to the size of an _Unwind_Word, which is 930 // uint64_t on all platforms. Generally this gets poked into a 931 // register and eventually used as an address, so if the 932 // addressing registers are wider than pointers and the platform 933 // doesn't implicitly ignore high-order bits when doing 934 // addressing, we need to make sure we zext / sext based on 935 // the platform's expectations. 936 // 937 // See: http://gcc.gnu.org/ml/gcc-bugs/2002-02/msg00237.html 938 939 // Cast the pointer to intptr_t. 940 Value *Ptr = EmitScalarExpr(E->getArg(0)); 941 Value *Result = Builder.CreatePtrToInt(Ptr, IntPtrTy, "extend.cast"); 942 943 // If that's 64 bits, we're done. 944 if (IntPtrTy->getBitWidth() == 64) 945 return RValue::get(Result); 946 947 // Otherwise, ask the codegen data what to do. 948 if (getTargetHooks().extendPointerWithSExt()) 949 return RValue::get(Builder.CreateSExt(Result, Int64Ty, "extend.sext")); 950 else 951 return RValue::get(Builder.CreateZExt(Result, Int64Ty, "extend.zext")); 952 } 953 case Builtin::BI__builtin_setjmp: { 954 // Buffer is a void**. 955 Value *Buf = EmitScalarExpr(E->getArg(0)); 956 957 // Store the frame pointer to the setjmp buffer. 958 Value *FrameAddr = 959 Builder.CreateCall(CGM.getIntrinsic(Intrinsic::frameaddress), 960 ConstantInt::get(Int32Ty, 0)); 961 Builder.CreateStore(FrameAddr, Buf); 962 963 // Store the stack pointer to the setjmp buffer. 964 Value *StackAddr = 965 Builder.CreateCall(CGM.getIntrinsic(Intrinsic::stacksave)); 966 Value *StackSaveSlot = 967 Builder.CreateGEP(Buf, ConstantInt::get(Int32Ty, 2)); 968 Builder.CreateStore(StackAddr, StackSaveSlot); 969 970 // Call LLVM's EH setjmp, which is lightweight. 971 Value *F = CGM.getIntrinsic(Intrinsic::eh_sjlj_setjmp); 972 Buf = Builder.CreateBitCast(Buf, Int8PtrTy); 973 return RValue::get(Builder.CreateCall(F, Buf)); 974 } 975 case Builtin::BI__builtin_longjmp: { 976 Value *Buf = EmitScalarExpr(E->getArg(0)); 977 Buf = Builder.CreateBitCast(Buf, Int8PtrTy); 978 979 // Call LLVM's EH longjmp, which is lightweight. 980 Builder.CreateCall(CGM.getIntrinsic(Intrinsic::eh_sjlj_longjmp), Buf); 981 982 // longjmp doesn't return; mark this as unreachable. 983 Builder.CreateUnreachable(); 984 985 // We do need to preserve an insertion point. 986 EmitBlock(createBasicBlock("longjmp.cont")); 987 988 return RValue::get(nullptr); 989 } 990 case Builtin::BI__sync_fetch_and_add: 991 case Builtin::BI__sync_fetch_and_sub: 992 case Builtin::BI__sync_fetch_and_or: 993 case Builtin::BI__sync_fetch_and_and: 994 case Builtin::BI__sync_fetch_and_xor: 995 case Builtin::BI__sync_fetch_and_nand: 996 case Builtin::BI__sync_add_and_fetch: 997 case Builtin::BI__sync_sub_and_fetch: 998 case Builtin::BI__sync_and_and_fetch: 999 case Builtin::BI__sync_or_and_fetch: 1000 case Builtin::BI__sync_xor_and_fetch: 1001 case Builtin::BI__sync_nand_and_fetch: 1002 case Builtin::BI__sync_val_compare_and_swap: 1003 case Builtin::BI__sync_bool_compare_and_swap: 1004 case Builtin::BI__sync_lock_test_and_set: 1005 case Builtin::BI__sync_lock_release: 1006 case Builtin::BI__sync_swap: 1007 llvm_unreachable("Shouldn't make it through sema"); 1008 case Builtin::BI__sync_fetch_and_add_1: 1009 case Builtin::BI__sync_fetch_and_add_2: 1010 case Builtin::BI__sync_fetch_and_add_4: 1011 case Builtin::BI__sync_fetch_and_add_8: 1012 case Builtin::BI__sync_fetch_and_add_16: 1013 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Add, E); 1014 case Builtin::BI__sync_fetch_and_sub_1: 1015 case Builtin::BI__sync_fetch_and_sub_2: 1016 case Builtin::BI__sync_fetch_and_sub_4: 1017 case Builtin::BI__sync_fetch_and_sub_8: 1018 case Builtin::BI__sync_fetch_and_sub_16: 1019 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Sub, E); 1020 case Builtin::BI__sync_fetch_and_or_1: 1021 case Builtin::BI__sync_fetch_and_or_2: 1022 case Builtin::BI__sync_fetch_and_or_4: 1023 case Builtin::BI__sync_fetch_and_or_8: 1024 case Builtin::BI__sync_fetch_and_or_16: 1025 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Or, E); 1026 case Builtin::BI__sync_fetch_and_and_1: 1027 case Builtin::BI__sync_fetch_and_and_2: 1028 case Builtin::BI__sync_fetch_and_and_4: 1029 case Builtin::BI__sync_fetch_and_and_8: 1030 case Builtin::BI__sync_fetch_and_and_16: 1031 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::And, E); 1032 case Builtin::BI__sync_fetch_and_xor_1: 1033 case Builtin::BI__sync_fetch_and_xor_2: 1034 case Builtin::BI__sync_fetch_and_xor_4: 1035 case Builtin::BI__sync_fetch_and_xor_8: 1036 case Builtin::BI__sync_fetch_and_xor_16: 1037 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Xor, E); 1038 case Builtin::BI__sync_fetch_and_nand_1: 1039 case Builtin::BI__sync_fetch_and_nand_2: 1040 case Builtin::BI__sync_fetch_and_nand_4: 1041 case Builtin::BI__sync_fetch_and_nand_8: 1042 case Builtin::BI__sync_fetch_and_nand_16: 1043 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Nand, E); 1044 1045 // Clang extensions: not overloaded yet. 1046 case Builtin::BI__sync_fetch_and_min: 1047 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Min, E); 1048 case Builtin::BI__sync_fetch_and_max: 1049 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Max, E); 1050 case Builtin::BI__sync_fetch_and_umin: 1051 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::UMin, E); 1052 case Builtin::BI__sync_fetch_and_umax: 1053 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::UMax, E); 1054 1055 case Builtin::BI__sync_add_and_fetch_1: 1056 case Builtin::BI__sync_add_and_fetch_2: 1057 case Builtin::BI__sync_add_and_fetch_4: 1058 case Builtin::BI__sync_add_and_fetch_8: 1059 case Builtin::BI__sync_add_and_fetch_16: 1060 return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::Add, E, 1061 llvm::Instruction::Add); 1062 case Builtin::BI__sync_sub_and_fetch_1: 1063 case Builtin::BI__sync_sub_and_fetch_2: 1064 case Builtin::BI__sync_sub_and_fetch_4: 1065 case Builtin::BI__sync_sub_and_fetch_8: 1066 case Builtin::BI__sync_sub_and_fetch_16: 1067 return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::Sub, E, 1068 llvm::Instruction::Sub); 1069 case Builtin::BI__sync_and_and_fetch_1: 1070 case Builtin::BI__sync_and_and_fetch_2: 1071 case Builtin::BI__sync_and_and_fetch_4: 1072 case Builtin::BI__sync_and_and_fetch_8: 1073 case Builtin::BI__sync_and_and_fetch_16: 1074 return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::And, E, 1075 llvm::Instruction::And); 1076 case Builtin::BI__sync_or_and_fetch_1: 1077 case Builtin::BI__sync_or_and_fetch_2: 1078 case Builtin::BI__sync_or_and_fetch_4: 1079 case Builtin::BI__sync_or_and_fetch_8: 1080 case Builtin::BI__sync_or_and_fetch_16: 1081 return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::Or, E, 1082 llvm::Instruction::Or); 1083 case Builtin::BI__sync_xor_and_fetch_1: 1084 case Builtin::BI__sync_xor_and_fetch_2: 1085 case Builtin::BI__sync_xor_and_fetch_4: 1086 case Builtin::BI__sync_xor_and_fetch_8: 1087 case Builtin::BI__sync_xor_and_fetch_16: 1088 return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::Xor, E, 1089 llvm::Instruction::Xor); 1090 case Builtin::BI__sync_nand_and_fetch_1: 1091 case Builtin::BI__sync_nand_and_fetch_2: 1092 case Builtin::BI__sync_nand_and_fetch_4: 1093 case Builtin::BI__sync_nand_and_fetch_8: 1094 case Builtin::BI__sync_nand_and_fetch_16: 1095 return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::Nand, E, 1096 llvm::Instruction::And, true); 1097 1098 case Builtin::BI__sync_val_compare_and_swap_1: 1099 case Builtin::BI__sync_val_compare_and_swap_2: 1100 case Builtin::BI__sync_val_compare_and_swap_4: 1101 case Builtin::BI__sync_val_compare_and_swap_8: 1102 case Builtin::BI__sync_val_compare_and_swap_16: 1103 return RValue::get(MakeAtomicCmpXchgValue(*this, E, false)); 1104 1105 case Builtin::BI__sync_bool_compare_and_swap_1: 1106 case Builtin::BI__sync_bool_compare_and_swap_2: 1107 case Builtin::BI__sync_bool_compare_and_swap_4: 1108 case Builtin::BI__sync_bool_compare_and_swap_8: 1109 case Builtin::BI__sync_bool_compare_and_swap_16: 1110 return RValue::get(MakeAtomicCmpXchgValue(*this, E, true)); 1111 1112 case Builtin::BI__sync_swap_1: 1113 case Builtin::BI__sync_swap_2: 1114 case Builtin::BI__sync_swap_4: 1115 case Builtin::BI__sync_swap_8: 1116 case Builtin::BI__sync_swap_16: 1117 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Xchg, E); 1118 1119 case Builtin::BI__sync_lock_test_and_set_1: 1120 case Builtin::BI__sync_lock_test_and_set_2: 1121 case Builtin::BI__sync_lock_test_and_set_4: 1122 case Builtin::BI__sync_lock_test_and_set_8: 1123 case Builtin::BI__sync_lock_test_and_set_16: 1124 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Xchg, E); 1125 1126 case Builtin::BI__sync_lock_release_1: 1127 case Builtin::BI__sync_lock_release_2: 1128 case Builtin::BI__sync_lock_release_4: 1129 case Builtin::BI__sync_lock_release_8: 1130 case Builtin::BI__sync_lock_release_16: { 1131 Value *Ptr = EmitScalarExpr(E->getArg(0)); 1132 QualType ElTy = E->getArg(0)->getType()->getPointeeType(); 1133 CharUnits StoreSize = getContext().getTypeSizeInChars(ElTy); 1134 llvm::Type *ITy = llvm::IntegerType::get(getLLVMContext(), 1135 StoreSize.getQuantity() * 8); 1136 Ptr = Builder.CreateBitCast(Ptr, ITy->getPointerTo()); 1137 llvm::StoreInst *Store = 1138 Builder.CreateStore(llvm::Constant::getNullValue(ITy), Ptr); 1139 Store->setAlignment(StoreSize.getQuantity()); 1140 Store->setAtomic(llvm::Release); 1141 return RValue::get(nullptr); 1142 } 1143 1144 case Builtin::BI__sync_synchronize: { 1145 // We assume this is supposed to correspond to a C++0x-style 1146 // sequentially-consistent fence (i.e. this is only usable for 1147 // synchonization, not device I/O or anything like that). This intrinsic 1148 // is really badly designed in the sense that in theory, there isn't 1149 // any way to safely use it... but in practice, it mostly works 1150 // to use it with non-atomic loads and stores to get acquire/release 1151 // semantics. 1152 Builder.CreateFence(llvm::SequentiallyConsistent); 1153 return RValue::get(nullptr); 1154 } 1155 1156 case Builtin::BI__c11_atomic_is_lock_free: 1157 case Builtin::BI__atomic_is_lock_free: { 1158 // Call "bool __atomic_is_lock_free(size_t size, void *ptr)". For the 1159 // __c11 builtin, ptr is 0 (indicating a properly-aligned object), since 1160 // _Atomic(T) is always properly-aligned. 1161 const char *LibCallName = "__atomic_is_lock_free"; 1162 CallArgList Args; 1163 Args.add(RValue::get(EmitScalarExpr(E->getArg(0))), 1164 getContext().getSizeType()); 1165 if (BuiltinID == Builtin::BI__atomic_is_lock_free) 1166 Args.add(RValue::get(EmitScalarExpr(E->getArg(1))), 1167 getContext().VoidPtrTy); 1168 else 1169 Args.add(RValue::get(llvm::Constant::getNullValue(VoidPtrTy)), 1170 getContext().VoidPtrTy); 1171 const CGFunctionInfo &FuncInfo = 1172 CGM.getTypes().arrangeFreeFunctionCall(E->getType(), Args, 1173 FunctionType::ExtInfo(), 1174 RequiredArgs::All); 1175 llvm::FunctionType *FTy = CGM.getTypes().GetFunctionType(FuncInfo); 1176 llvm::Constant *Func = CGM.CreateRuntimeFunction(FTy, LibCallName); 1177 return EmitCall(FuncInfo, Func, ReturnValueSlot(), Args); 1178 } 1179 1180 case Builtin::BI__atomic_test_and_set: { 1181 // Look at the argument type to determine whether this is a volatile 1182 // operation. The parameter type is always volatile. 1183 QualType PtrTy = E->getArg(0)->IgnoreImpCasts()->getType(); 1184 bool Volatile = 1185 PtrTy->castAs<PointerType>()->getPointeeType().isVolatileQualified(); 1186 1187 Value *Ptr = EmitScalarExpr(E->getArg(0)); 1188 unsigned AddrSpace = Ptr->getType()->getPointerAddressSpace(); 1189 Ptr = Builder.CreateBitCast(Ptr, Int8Ty->getPointerTo(AddrSpace)); 1190 Value *NewVal = Builder.getInt8(1); 1191 Value *Order = EmitScalarExpr(E->getArg(1)); 1192 if (isa<llvm::ConstantInt>(Order)) { 1193 int ord = cast<llvm::ConstantInt>(Order)->getZExtValue(); 1194 AtomicRMWInst *Result = nullptr; 1195 switch (ord) { 1196 case 0: // memory_order_relaxed 1197 default: // invalid order 1198 Result = Builder.CreateAtomicRMW(llvm::AtomicRMWInst::Xchg, 1199 Ptr, NewVal, 1200 llvm::Monotonic); 1201 break; 1202 case 1: // memory_order_consume 1203 case 2: // memory_order_acquire 1204 Result = Builder.CreateAtomicRMW(llvm::AtomicRMWInst::Xchg, 1205 Ptr, NewVal, 1206 llvm::Acquire); 1207 break; 1208 case 3: // memory_order_release 1209 Result = Builder.CreateAtomicRMW(llvm::AtomicRMWInst::Xchg, 1210 Ptr, NewVal, 1211 llvm::Release); 1212 break; 1213 case 4: // memory_order_acq_rel 1214 Result = Builder.CreateAtomicRMW(llvm::AtomicRMWInst::Xchg, 1215 Ptr, NewVal, 1216 llvm::AcquireRelease); 1217 break; 1218 case 5: // memory_order_seq_cst 1219 Result = Builder.CreateAtomicRMW(llvm::AtomicRMWInst::Xchg, 1220 Ptr, NewVal, 1221 llvm::SequentiallyConsistent); 1222 break; 1223 } 1224 Result->setVolatile(Volatile); 1225 return RValue::get(Builder.CreateIsNotNull(Result, "tobool")); 1226 } 1227 1228 llvm::BasicBlock *ContBB = createBasicBlock("atomic.continue", CurFn); 1229 1230 llvm::BasicBlock *BBs[5] = { 1231 createBasicBlock("monotonic", CurFn), 1232 createBasicBlock("acquire", CurFn), 1233 createBasicBlock("release", CurFn), 1234 createBasicBlock("acqrel", CurFn), 1235 createBasicBlock("seqcst", CurFn) 1236 }; 1237 llvm::AtomicOrdering Orders[5] = { 1238 llvm::Monotonic, llvm::Acquire, llvm::Release, 1239 llvm::AcquireRelease, llvm::SequentiallyConsistent 1240 }; 1241 1242 Order = Builder.CreateIntCast(Order, Builder.getInt32Ty(), false); 1243 llvm::SwitchInst *SI = Builder.CreateSwitch(Order, BBs[0]); 1244 1245 Builder.SetInsertPoint(ContBB); 1246 PHINode *Result = Builder.CreatePHI(Int8Ty, 5, "was_set"); 1247 1248 for (unsigned i = 0; i < 5; ++i) { 1249 Builder.SetInsertPoint(BBs[i]); 1250 AtomicRMWInst *RMW = Builder.CreateAtomicRMW(llvm::AtomicRMWInst::Xchg, 1251 Ptr, NewVal, Orders[i]); 1252 RMW->setVolatile(Volatile); 1253 Result->addIncoming(RMW, BBs[i]); 1254 Builder.CreateBr(ContBB); 1255 } 1256 1257 SI->addCase(Builder.getInt32(0), BBs[0]); 1258 SI->addCase(Builder.getInt32(1), BBs[1]); 1259 SI->addCase(Builder.getInt32(2), BBs[1]); 1260 SI->addCase(Builder.getInt32(3), BBs[2]); 1261 SI->addCase(Builder.getInt32(4), BBs[3]); 1262 SI->addCase(Builder.getInt32(5), BBs[4]); 1263 1264 Builder.SetInsertPoint(ContBB); 1265 return RValue::get(Builder.CreateIsNotNull(Result, "tobool")); 1266 } 1267 1268 case Builtin::BI__atomic_clear: { 1269 QualType PtrTy = E->getArg(0)->IgnoreImpCasts()->getType(); 1270 bool Volatile = 1271 PtrTy->castAs<PointerType>()->getPointeeType().isVolatileQualified(); 1272 1273 Value *Ptr = EmitScalarExpr(E->getArg(0)); 1274 unsigned AddrSpace = Ptr->getType()->getPointerAddressSpace(); 1275 Ptr = Builder.CreateBitCast(Ptr, Int8Ty->getPointerTo(AddrSpace)); 1276 Value *NewVal = Builder.getInt8(0); 1277 Value *Order = EmitScalarExpr(E->getArg(1)); 1278 if (isa<llvm::ConstantInt>(Order)) { 1279 int ord = cast<llvm::ConstantInt>(Order)->getZExtValue(); 1280 StoreInst *Store = Builder.CreateStore(NewVal, Ptr, Volatile); 1281 Store->setAlignment(1); 1282 switch (ord) { 1283 case 0: // memory_order_relaxed 1284 default: // invalid order 1285 Store->setOrdering(llvm::Monotonic); 1286 break; 1287 case 3: // memory_order_release 1288 Store->setOrdering(llvm::Release); 1289 break; 1290 case 5: // memory_order_seq_cst 1291 Store->setOrdering(llvm::SequentiallyConsistent); 1292 break; 1293 } 1294 return RValue::get(nullptr); 1295 } 1296 1297 llvm::BasicBlock *ContBB = createBasicBlock("atomic.continue", CurFn); 1298 1299 llvm::BasicBlock *BBs[3] = { 1300 createBasicBlock("monotonic", CurFn), 1301 createBasicBlock("release", CurFn), 1302 createBasicBlock("seqcst", CurFn) 1303 }; 1304 llvm::AtomicOrdering Orders[3] = { 1305 llvm::Monotonic, llvm::Release, llvm::SequentiallyConsistent 1306 }; 1307 1308 Order = Builder.CreateIntCast(Order, Builder.getInt32Ty(), false); 1309 llvm::SwitchInst *SI = Builder.CreateSwitch(Order, BBs[0]); 1310 1311 for (unsigned i = 0; i < 3; ++i) { 1312 Builder.SetInsertPoint(BBs[i]); 1313 StoreInst *Store = Builder.CreateStore(NewVal, Ptr, Volatile); 1314 Store->setAlignment(1); 1315 Store->setOrdering(Orders[i]); 1316 Builder.CreateBr(ContBB); 1317 } 1318 1319 SI->addCase(Builder.getInt32(0), BBs[0]); 1320 SI->addCase(Builder.getInt32(3), BBs[1]); 1321 SI->addCase(Builder.getInt32(5), BBs[2]); 1322 1323 Builder.SetInsertPoint(ContBB); 1324 return RValue::get(nullptr); 1325 } 1326 1327 case Builtin::BI__atomic_thread_fence: 1328 case Builtin::BI__atomic_signal_fence: 1329 case Builtin::BI__c11_atomic_thread_fence: 1330 case Builtin::BI__c11_atomic_signal_fence: { 1331 llvm::SynchronizationScope Scope; 1332 if (BuiltinID == Builtin::BI__atomic_signal_fence || 1333 BuiltinID == Builtin::BI__c11_atomic_signal_fence) 1334 Scope = llvm::SingleThread; 1335 else 1336 Scope = llvm::CrossThread; 1337 Value *Order = EmitScalarExpr(E->getArg(0)); 1338 if (isa<llvm::ConstantInt>(Order)) { 1339 int ord = cast<llvm::ConstantInt>(Order)->getZExtValue(); 1340 switch (ord) { 1341 case 0: // memory_order_relaxed 1342 default: // invalid order 1343 break; 1344 case 1: // memory_order_consume 1345 case 2: // memory_order_acquire 1346 Builder.CreateFence(llvm::Acquire, Scope); 1347 break; 1348 case 3: // memory_order_release 1349 Builder.CreateFence(llvm::Release, Scope); 1350 break; 1351 case 4: // memory_order_acq_rel 1352 Builder.CreateFence(llvm::AcquireRelease, Scope); 1353 break; 1354 case 5: // memory_order_seq_cst 1355 Builder.CreateFence(llvm::SequentiallyConsistent, Scope); 1356 break; 1357 } 1358 return RValue::get(nullptr); 1359 } 1360 1361 llvm::BasicBlock *AcquireBB, *ReleaseBB, *AcqRelBB, *SeqCstBB; 1362 AcquireBB = createBasicBlock("acquire", CurFn); 1363 ReleaseBB = createBasicBlock("release", CurFn); 1364 AcqRelBB = createBasicBlock("acqrel", CurFn); 1365 SeqCstBB = createBasicBlock("seqcst", CurFn); 1366 llvm::BasicBlock *ContBB = createBasicBlock("atomic.continue", CurFn); 1367 1368 Order = Builder.CreateIntCast(Order, Builder.getInt32Ty(), false); 1369 llvm::SwitchInst *SI = Builder.CreateSwitch(Order, ContBB); 1370 1371 Builder.SetInsertPoint(AcquireBB); 1372 Builder.CreateFence(llvm::Acquire, Scope); 1373 Builder.CreateBr(ContBB); 1374 SI->addCase(Builder.getInt32(1), AcquireBB); 1375 SI->addCase(Builder.getInt32(2), AcquireBB); 1376 1377 Builder.SetInsertPoint(ReleaseBB); 1378 Builder.CreateFence(llvm::Release, Scope); 1379 Builder.CreateBr(ContBB); 1380 SI->addCase(Builder.getInt32(3), ReleaseBB); 1381 1382 Builder.SetInsertPoint(AcqRelBB); 1383 Builder.CreateFence(llvm::AcquireRelease, Scope); 1384 Builder.CreateBr(ContBB); 1385 SI->addCase(Builder.getInt32(4), AcqRelBB); 1386 1387 Builder.SetInsertPoint(SeqCstBB); 1388 Builder.CreateFence(llvm::SequentiallyConsistent, Scope); 1389 Builder.CreateBr(ContBB); 1390 SI->addCase(Builder.getInt32(5), SeqCstBB); 1391 1392 Builder.SetInsertPoint(ContBB); 1393 return RValue::get(nullptr); 1394 } 1395 1396 // Library functions with special handling. 1397 case Builtin::BIsqrt: 1398 case Builtin::BIsqrtf: 1399 case Builtin::BIsqrtl: { 1400 // Transform a call to sqrt* into a @llvm.sqrt.* intrinsic call, but only 1401 // in finite- or unsafe-math mode (the intrinsic has different semantics 1402 // for handling negative numbers compared to the library function, so 1403 // -fmath-errno=0 is not enough). 1404 if (!FD->hasAttr<ConstAttr>()) 1405 break; 1406 if (!(CGM.getCodeGenOpts().UnsafeFPMath || 1407 CGM.getCodeGenOpts().NoNaNsFPMath)) 1408 break; 1409 Value *Arg0 = EmitScalarExpr(E->getArg(0)); 1410 llvm::Type *ArgType = Arg0->getType(); 1411 Value *F = CGM.getIntrinsic(Intrinsic::sqrt, ArgType); 1412 return RValue::get(Builder.CreateCall(F, Arg0)); 1413 } 1414 1415 case Builtin::BI__builtin_pow: 1416 case Builtin::BI__builtin_powf: 1417 case Builtin::BI__builtin_powl: 1418 case Builtin::BIpow: 1419 case Builtin::BIpowf: 1420 case Builtin::BIpowl: { 1421 // Transform a call to pow* into a @llvm.pow.* intrinsic call. 1422 if (!FD->hasAttr<ConstAttr>()) 1423 break; 1424 Value *Base = EmitScalarExpr(E->getArg(0)); 1425 Value *Exponent = EmitScalarExpr(E->getArg(1)); 1426 llvm::Type *ArgType = Base->getType(); 1427 Value *F = CGM.getIntrinsic(Intrinsic::pow, ArgType); 1428 return RValue::get(Builder.CreateCall(F, {Base, Exponent})); 1429 } 1430 1431 case Builtin::BIfma: 1432 case Builtin::BIfmaf: 1433 case Builtin::BIfmal: 1434 case Builtin::BI__builtin_fma: 1435 case Builtin::BI__builtin_fmaf: 1436 case Builtin::BI__builtin_fmal: { 1437 // Rewrite fma to intrinsic. 1438 Value *FirstArg = EmitScalarExpr(E->getArg(0)); 1439 llvm::Type *ArgType = FirstArg->getType(); 1440 Value *F = CGM.getIntrinsic(Intrinsic::fma, ArgType); 1441 return RValue::get( 1442 Builder.CreateCall(F, {FirstArg, EmitScalarExpr(E->getArg(1)), 1443 EmitScalarExpr(E->getArg(2))})); 1444 } 1445 1446 case Builtin::BI__builtin_signbit: 1447 case Builtin::BI__builtin_signbitf: 1448 case Builtin::BI__builtin_signbitl: { 1449 return RValue::get( 1450 Builder.CreateZExt(EmitSignBit(*this, EmitScalarExpr(E->getArg(0))), 1451 ConvertType(E->getType()))); 1452 } 1453 case Builtin::BI__builtin_annotation: { 1454 llvm::Value *AnnVal = EmitScalarExpr(E->getArg(0)); 1455 llvm::Value *F = CGM.getIntrinsic(llvm::Intrinsic::annotation, 1456 AnnVal->getType()); 1457 1458 // Get the annotation string, go through casts. Sema requires this to be a 1459 // non-wide string literal, potentially casted, so the cast<> is safe. 1460 const Expr *AnnotationStrExpr = E->getArg(1)->IgnoreParenCasts(); 1461 StringRef Str = cast<StringLiteral>(AnnotationStrExpr)->getString(); 1462 return RValue::get(EmitAnnotationCall(F, AnnVal, Str, E->getExprLoc())); 1463 } 1464 case Builtin::BI__builtin_addcb: 1465 case Builtin::BI__builtin_addcs: 1466 case Builtin::BI__builtin_addc: 1467 case Builtin::BI__builtin_addcl: 1468 case Builtin::BI__builtin_addcll: 1469 case Builtin::BI__builtin_subcb: 1470 case Builtin::BI__builtin_subcs: 1471 case Builtin::BI__builtin_subc: 1472 case Builtin::BI__builtin_subcl: 1473 case Builtin::BI__builtin_subcll: { 1474 1475 // We translate all of these builtins from expressions of the form: 1476 // int x = ..., y = ..., carryin = ..., carryout, result; 1477 // result = __builtin_addc(x, y, carryin, &carryout); 1478 // 1479 // to LLVM IR of the form: 1480 // 1481 // %tmp1 = call {i32, i1} @llvm.uadd.with.overflow.i32(i32 %x, i32 %y) 1482 // %tmpsum1 = extractvalue {i32, i1} %tmp1, 0 1483 // %carry1 = extractvalue {i32, i1} %tmp1, 1 1484 // %tmp2 = call {i32, i1} @llvm.uadd.with.overflow.i32(i32 %tmpsum1, 1485 // i32 %carryin) 1486 // %result = extractvalue {i32, i1} %tmp2, 0 1487 // %carry2 = extractvalue {i32, i1} %tmp2, 1 1488 // %tmp3 = or i1 %carry1, %carry2 1489 // %tmp4 = zext i1 %tmp3 to i32 1490 // store i32 %tmp4, i32* %carryout 1491 1492 // Scalarize our inputs. 1493 llvm::Value *X = EmitScalarExpr(E->getArg(0)); 1494 llvm::Value *Y = EmitScalarExpr(E->getArg(1)); 1495 llvm::Value *Carryin = EmitScalarExpr(E->getArg(2)); 1496 std::pair<llvm::Value*, unsigned> CarryOutPtr = 1497 EmitPointerWithAlignment(E->getArg(3)); 1498 1499 // Decide if we are lowering to a uadd.with.overflow or usub.with.overflow. 1500 llvm::Intrinsic::ID IntrinsicId; 1501 switch (BuiltinID) { 1502 default: llvm_unreachable("Unknown multiprecision builtin id."); 1503 case Builtin::BI__builtin_addcb: 1504 case Builtin::BI__builtin_addcs: 1505 case Builtin::BI__builtin_addc: 1506 case Builtin::BI__builtin_addcl: 1507 case Builtin::BI__builtin_addcll: 1508 IntrinsicId = llvm::Intrinsic::uadd_with_overflow; 1509 break; 1510 case Builtin::BI__builtin_subcb: 1511 case Builtin::BI__builtin_subcs: 1512 case Builtin::BI__builtin_subc: 1513 case Builtin::BI__builtin_subcl: 1514 case Builtin::BI__builtin_subcll: 1515 IntrinsicId = llvm::Intrinsic::usub_with_overflow; 1516 break; 1517 } 1518 1519 // Construct our resulting LLVM IR expression. 1520 llvm::Value *Carry1; 1521 llvm::Value *Sum1 = EmitOverflowIntrinsic(*this, IntrinsicId, 1522 X, Y, Carry1); 1523 llvm::Value *Carry2; 1524 llvm::Value *Sum2 = EmitOverflowIntrinsic(*this, IntrinsicId, 1525 Sum1, Carryin, Carry2); 1526 llvm::Value *CarryOut = Builder.CreateZExt(Builder.CreateOr(Carry1, Carry2), 1527 X->getType()); 1528 llvm::StoreInst *CarryOutStore = Builder.CreateStore(CarryOut, 1529 CarryOutPtr.first); 1530 CarryOutStore->setAlignment(CarryOutPtr.second); 1531 return RValue::get(Sum2); 1532 } 1533 case Builtin::BI__builtin_uadd_overflow: 1534 case Builtin::BI__builtin_uaddl_overflow: 1535 case Builtin::BI__builtin_uaddll_overflow: 1536 case Builtin::BI__builtin_usub_overflow: 1537 case Builtin::BI__builtin_usubl_overflow: 1538 case Builtin::BI__builtin_usubll_overflow: 1539 case Builtin::BI__builtin_umul_overflow: 1540 case Builtin::BI__builtin_umull_overflow: 1541 case Builtin::BI__builtin_umulll_overflow: 1542 case Builtin::BI__builtin_sadd_overflow: 1543 case Builtin::BI__builtin_saddl_overflow: 1544 case Builtin::BI__builtin_saddll_overflow: 1545 case Builtin::BI__builtin_ssub_overflow: 1546 case Builtin::BI__builtin_ssubl_overflow: 1547 case Builtin::BI__builtin_ssubll_overflow: 1548 case Builtin::BI__builtin_smul_overflow: 1549 case Builtin::BI__builtin_smull_overflow: 1550 case Builtin::BI__builtin_smulll_overflow: { 1551 1552 // We translate all of these builtins directly to the relevant llvm IR node. 1553 1554 // Scalarize our inputs. 1555 llvm::Value *X = EmitScalarExpr(E->getArg(0)); 1556 llvm::Value *Y = EmitScalarExpr(E->getArg(1)); 1557 std::pair<llvm::Value *, unsigned> SumOutPtr = 1558 EmitPointerWithAlignment(E->getArg(2)); 1559 1560 // Decide which of the overflow intrinsics we are lowering to: 1561 llvm::Intrinsic::ID IntrinsicId; 1562 switch (BuiltinID) { 1563 default: llvm_unreachable("Unknown security overflow builtin id."); 1564 case Builtin::BI__builtin_uadd_overflow: 1565 case Builtin::BI__builtin_uaddl_overflow: 1566 case Builtin::BI__builtin_uaddll_overflow: 1567 IntrinsicId = llvm::Intrinsic::uadd_with_overflow; 1568 break; 1569 case Builtin::BI__builtin_usub_overflow: 1570 case Builtin::BI__builtin_usubl_overflow: 1571 case Builtin::BI__builtin_usubll_overflow: 1572 IntrinsicId = llvm::Intrinsic::usub_with_overflow; 1573 break; 1574 case Builtin::BI__builtin_umul_overflow: 1575 case Builtin::BI__builtin_umull_overflow: 1576 case Builtin::BI__builtin_umulll_overflow: 1577 IntrinsicId = llvm::Intrinsic::umul_with_overflow; 1578 break; 1579 case Builtin::BI__builtin_sadd_overflow: 1580 case Builtin::BI__builtin_saddl_overflow: 1581 case Builtin::BI__builtin_saddll_overflow: 1582 IntrinsicId = llvm::Intrinsic::sadd_with_overflow; 1583 break; 1584 case Builtin::BI__builtin_ssub_overflow: 1585 case Builtin::BI__builtin_ssubl_overflow: 1586 case Builtin::BI__builtin_ssubll_overflow: 1587 IntrinsicId = llvm::Intrinsic::ssub_with_overflow; 1588 break; 1589 case Builtin::BI__builtin_smul_overflow: 1590 case Builtin::BI__builtin_smull_overflow: 1591 case Builtin::BI__builtin_smulll_overflow: 1592 IntrinsicId = llvm::Intrinsic::smul_with_overflow; 1593 break; 1594 } 1595 1596 1597 llvm::Value *Carry; 1598 llvm::Value *Sum = EmitOverflowIntrinsic(*this, IntrinsicId, X, Y, Carry); 1599 llvm::StoreInst *SumOutStore = Builder.CreateStore(Sum, SumOutPtr.first); 1600 SumOutStore->setAlignment(SumOutPtr.second); 1601 1602 return RValue::get(Carry); 1603 } 1604 case Builtin::BI__builtin_addressof: 1605 return RValue::get(EmitLValue(E->getArg(0)).getAddress()); 1606 case Builtin::BI__builtin_operator_new: 1607 return EmitBuiltinNewDeleteCall(FD->getType()->castAs<FunctionProtoType>(), 1608 E->getArg(0), false); 1609 case Builtin::BI__builtin_operator_delete: 1610 return EmitBuiltinNewDeleteCall(FD->getType()->castAs<FunctionProtoType>(), 1611 E->getArg(0), true); 1612 case Builtin::BI__noop: 1613 // __noop always evaluates to an integer literal zero. 1614 return RValue::get(ConstantInt::get(IntTy, 0)); 1615 case Builtin::BI__builtin_call_with_static_chain: { 1616 const CallExpr *Call = cast<CallExpr>(E->getArg(0)); 1617 const Expr *Chain = E->getArg(1); 1618 return EmitCall(Call->getCallee()->getType(), 1619 EmitScalarExpr(Call->getCallee()), Call, ReturnValue, 1620 Call->getCalleeDecl(), EmitScalarExpr(Chain)); 1621 } 1622 case Builtin::BI_InterlockedExchange: 1623 case Builtin::BI_InterlockedExchangePointer: 1624 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Xchg, E); 1625 case Builtin::BI_InterlockedCompareExchangePointer: { 1626 llvm::Type *RTy; 1627 llvm::IntegerType *IntType = 1628 IntegerType::get(getLLVMContext(), 1629 getContext().getTypeSize(E->getType())); 1630 llvm::Type *IntPtrType = IntType->getPointerTo(); 1631 1632 llvm::Value *Destination = 1633 Builder.CreateBitCast(EmitScalarExpr(E->getArg(0)), IntPtrType); 1634 1635 llvm::Value *Exchange = EmitScalarExpr(E->getArg(1)); 1636 RTy = Exchange->getType(); 1637 Exchange = Builder.CreatePtrToInt(Exchange, IntType); 1638 1639 llvm::Value *Comparand = 1640 Builder.CreatePtrToInt(EmitScalarExpr(E->getArg(2)), IntType); 1641 1642 auto Result = Builder.CreateAtomicCmpXchg(Destination, Comparand, Exchange, 1643 SequentiallyConsistent, 1644 SequentiallyConsistent); 1645 Result->setVolatile(true); 1646 1647 return RValue::get(Builder.CreateIntToPtr(Builder.CreateExtractValue(Result, 1648 0), 1649 RTy)); 1650 } 1651 case Builtin::BI_InterlockedCompareExchange: { 1652 AtomicCmpXchgInst *CXI = Builder.CreateAtomicCmpXchg( 1653 EmitScalarExpr(E->getArg(0)), 1654 EmitScalarExpr(E->getArg(2)), 1655 EmitScalarExpr(E->getArg(1)), 1656 SequentiallyConsistent, 1657 SequentiallyConsistent); 1658 CXI->setVolatile(true); 1659 return RValue::get(Builder.CreateExtractValue(CXI, 0)); 1660 } 1661 case Builtin::BI_InterlockedIncrement: { 1662 AtomicRMWInst *RMWI = Builder.CreateAtomicRMW( 1663 AtomicRMWInst::Add, 1664 EmitScalarExpr(E->getArg(0)), 1665 ConstantInt::get(Int32Ty, 1), 1666 llvm::SequentiallyConsistent); 1667 RMWI->setVolatile(true); 1668 return RValue::get(Builder.CreateAdd(RMWI, ConstantInt::get(Int32Ty, 1))); 1669 } 1670 case Builtin::BI_InterlockedDecrement: { 1671 AtomicRMWInst *RMWI = Builder.CreateAtomicRMW( 1672 AtomicRMWInst::Sub, 1673 EmitScalarExpr(E->getArg(0)), 1674 ConstantInt::get(Int32Ty, 1), 1675 llvm::SequentiallyConsistent); 1676 RMWI->setVolatile(true); 1677 return RValue::get(Builder.CreateSub(RMWI, ConstantInt::get(Int32Ty, 1))); 1678 } 1679 case Builtin::BI_InterlockedExchangeAdd: { 1680 AtomicRMWInst *RMWI = Builder.CreateAtomicRMW( 1681 AtomicRMWInst::Add, 1682 EmitScalarExpr(E->getArg(0)), 1683 EmitScalarExpr(E->getArg(1)), 1684 llvm::SequentiallyConsistent); 1685 RMWI->setVolatile(true); 1686 return RValue::get(RMWI); 1687 } 1688 case Builtin::BI__readfsdword: { 1689 Value *IntToPtr = 1690 Builder.CreateIntToPtr(EmitScalarExpr(E->getArg(0)), 1691 llvm::PointerType::get(CGM.Int32Ty, 257)); 1692 LoadInst *Load = 1693 Builder.CreateAlignedLoad(IntToPtr, /*Align=*/4, /*isVolatile=*/true); 1694 return RValue::get(Load); 1695 } 1696 1697 case Builtin::BI__exception_code: 1698 case Builtin::BI_exception_code: 1699 return RValue::get(EmitSEHExceptionCode()); 1700 case Builtin::BI__exception_info: 1701 case Builtin::BI_exception_info: 1702 return RValue::get(EmitSEHExceptionInfo()); 1703 case Builtin::BI__abnormal_termination: 1704 case Builtin::BI_abnormal_termination: 1705 return RValue::get(EmitSEHAbnormalTermination()); 1706 case Builtin::BI_setjmpex: { 1707 if (getTarget().getTriple().isOSMSVCRT()) { 1708 llvm::Type *ArgTypes[] = {Int8PtrTy, Int8PtrTy}; 1709 llvm::AttributeSet ReturnsTwiceAttr = 1710 AttributeSet::get(getLLVMContext(), llvm::AttributeSet::FunctionIndex, 1711 llvm::Attribute::ReturnsTwice); 1712 llvm::Constant *SetJmpEx = CGM.CreateRuntimeFunction( 1713 llvm::FunctionType::get(IntTy, ArgTypes, /*isVarArg=*/false), 1714 "_setjmpex", ReturnsTwiceAttr); 1715 llvm::Value *Buf = Builder.CreateBitOrPointerCast( 1716 EmitScalarExpr(E->getArg(0)), Int8PtrTy); 1717 llvm::Value *FrameAddr = 1718 Builder.CreateCall(CGM.getIntrinsic(Intrinsic::frameaddress), 1719 ConstantInt::get(Int32Ty, 0)); 1720 llvm::Value *Args[] = {Buf, FrameAddr}; 1721 llvm::CallSite CS = EmitRuntimeCallOrInvoke(SetJmpEx, Args); 1722 CS.setAttributes(ReturnsTwiceAttr); 1723 return RValue::get(CS.getInstruction()); 1724 } 1725 break; 1726 } 1727 case Builtin::BI_setjmp: { 1728 if (getTarget().getTriple().isOSMSVCRT()) { 1729 llvm::AttributeSet ReturnsTwiceAttr = 1730 AttributeSet::get(getLLVMContext(), llvm::AttributeSet::FunctionIndex, 1731 llvm::Attribute::ReturnsTwice); 1732 llvm::Value *Buf = Builder.CreateBitOrPointerCast( 1733 EmitScalarExpr(E->getArg(0)), Int8PtrTy); 1734 llvm::CallSite CS; 1735 if (getTarget().getTriple().getArch() == llvm::Triple::x86) { 1736 llvm::Type *ArgTypes[] = {Int8PtrTy, IntTy}; 1737 llvm::Constant *SetJmp3 = CGM.CreateRuntimeFunction( 1738 llvm::FunctionType::get(IntTy, ArgTypes, /*isVarArg=*/true), 1739 "_setjmp3", ReturnsTwiceAttr); 1740 llvm::Value *Count = ConstantInt::get(IntTy, 0); 1741 llvm::Value *Args[] = {Buf, Count}; 1742 CS = EmitRuntimeCallOrInvoke(SetJmp3, Args); 1743 } else { 1744 llvm::Type *ArgTypes[] = {Int8PtrTy, Int8PtrTy}; 1745 llvm::Constant *SetJmp = CGM.CreateRuntimeFunction( 1746 llvm::FunctionType::get(IntTy, ArgTypes, /*isVarArg=*/false), 1747 "_setjmp", ReturnsTwiceAttr); 1748 llvm::Value *FrameAddr = 1749 Builder.CreateCall(CGM.getIntrinsic(Intrinsic::frameaddress), 1750 ConstantInt::get(Int32Ty, 0)); 1751 llvm::Value *Args[] = {Buf, FrameAddr}; 1752 CS = EmitRuntimeCallOrInvoke(SetJmp, Args); 1753 } 1754 CS.setAttributes(ReturnsTwiceAttr); 1755 return RValue::get(CS.getInstruction()); 1756 } 1757 break; 1758 } 1759 1760 case Builtin::BI__GetExceptionInfo: { 1761 if (llvm::GlobalVariable *GV = 1762 CGM.getCXXABI().getThrowInfo(FD->getParamDecl(0)->getType())) 1763 return RValue::get(llvm::ConstantExpr::getBitCast(GV, CGM.Int8PtrTy)); 1764 break; 1765 } 1766 } 1767 1768 // If this is an alias for a lib function (e.g. __builtin_sin), emit 1769 // the call using the normal call path, but using the unmangled 1770 // version of the function name. 1771 if (getContext().BuiltinInfo.isLibFunction(BuiltinID)) 1772 return emitLibraryCall(*this, FD, E, 1773 CGM.getBuiltinLibFunction(FD, BuiltinID)); 1774 1775 // If this is a predefined lib function (e.g. malloc), emit the call 1776 // using exactly the normal call path. 1777 if (getContext().BuiltinInfo.isPredefinedLibFunction(BuiltinID)) 1778 return emitLibraryCall(*this, FD, E, EmitScalarExpr(E->getCallee())); 1779 1780 // See if we have a target specific intrinsic. 1781 const char *Name = getContext().BuiltinInfo.GetName(BuiltinID); 1782 Intrinsic::ID IntrinsicID = Intrinsic::not_intrinsic; 1783 if (const char *Prefix = 1784 llvm::Triple::getArchTypePrefix(getTarget().getTriple().getArch())) { 1785 IntrinsicID = Intrinsic::getIntrinsicForGCCBuiltin(Prefix, Name); 1786 // NOTE we dont need to perform a compatibility flag check here since the 1787 // intrinsics are declared in Builtins*.def via LANGBUILTIN which filter the 1788 // MS builtins via ALL_MS_LANGUAGES and are filtered earlier. 1789 if (IntrinsicID == Intrinsic::not_intrinsic) 1790 IntrinsicID = Intrinsic::getIntrinsicForMSBuiltin(Prefix, Name); 1791 } 1792 1793 if (IntrinsicID != Intrinsic::not_intrinsic) { 1794 SmallVector<Value*, 16> Args; 1795 1796 // Find out if any arguments are required to be integer constant 1797 // expressions. 1798 unsigned ICEArguments = 0; 1799 ASTContext::GetBuiltinTypeError Error; 1800 getContext().GetBuiltinType(BuiltinID, Error, &ICEArguments); 1801 assert(Error == ASTContext::GE_None && "Should not codegen an error"); 1802 1803 Function *F = CGM.getIntrinsic(IntrinsicID); 1804 llvm::FunctionType *FTy = F->getFunctionType(); 1805 1806 for (unsigned i = 0, e = E->getNumArgs(); i != e; ++i) { 1807 Value *ArgValue; 1808 // If this is a normal argument, just emit it as a scalar. 1809 if ((ICEArguments & (1 << i)) == 0) { 1810 ArgValue = EmitScalarExpr(E->getArg(i)); 1811 } else { 1812 // If this is required to be a constant, constant fold it so that we 1813 // know that the generated intrinsic gets a ConstantInt. 1814 llvm::APSInt Result; 1815 bool IsConst = E->getArg(i)->isIntegerConstantExpr(Result,getContext()); 1816 assert(IsConst && "Constant arg isn't actually constant?"); 1817 (void)IsConst; 1818 ArgValue = llvm::ConstantInt::get(getLLVMContext(), Result); 1819 } 1820 1821 // If the intrinsic arg type is different from the builtin arg type 1822 // we need to do a bit cast. 1823 llvm::Type *PTy = FTy->getParamType(i); 1824 if (PTy != ArgValue->getType()) { 1825 assert(PTy->canLosslesslyBitCastTo(FTy->getParamType(i)) && 1826 "Must be able to losslessly bit cast to param"); 1827 ArgValue = Builder.CreateBitCast(ArgValue, PTy); 1828 } 1829 1830 Args.push_back(ArgValue); 1831 } 1832 1833 Value *V = Builder.CreateCall(F, Args); 1834 QualType BuiltinRetType = E->getType(); 1835 1836 llvm::Type *RetTy = VoidTy; 1837 if (!BuiltinRetType->isVoidType()) 1838 RetTy = ConvertType(BuiltinRetType); 1839 1840 if (RetTy != V->getType()) { 1841 assert(V->getType()->canLosslesslyBitCastTo(RetTy) && 1842 "Must be able to losslessly bit cast result type"); 1843 V = Builder.CreateBitCast(V, RetTy); 1844 } 1845 1846 return RValue::get(V); 1847 } 1848 1849 // See if we have a target specific builtin that needs to be lowered. 1850 if (Value *V = EmitTargetBuiltinExpr(BuiltinID, E)) 1851 return RValue::get(V); 1852 1853 ErrorUnsupported(E, "builtin function"); 1854 1855 // Unknown builtin, for now just dump it out and return undef. 1856 return GetUndefRValue(E->getType()); 1857 } 1858 1859 Value *CodeGenFunction::EmitTargetBuiltinExpr(unsigned BuiltinID, 1860 const CallExpr *E) { 1861 switch (getTarget().getTriple().getArch()) { 1862 case llvm::Triple::arm: 1863 case llvm::Triple::armeb: 1864 case llvm::Triple::thumb: 1865 case llvm::Triple::thumbeb: 1866 return EmitARMBuiltinExpr(BuiltinID, E); 1867 case llvm::Triple::aarch64: 1868 case llvm::Triple::aarch64_be: 1869 return EmitAArch64BuiltinExpr(BuiltinID, E); 1870 case llvm::Triple::x86: 1871 case llvm::Triple::x86_64: 1872 return EmitX86BuiltinExpr(BuiltinID, E); 1873 case llvm::Triple::ppc: 1874 case llvm::Triple::ppc64: 1875 case llvm::Triple::ppc64le: 1876 return EmitPPCBuiltinExpr(BuiltinID, E); 1877 case llvm::Triple::r600: 1878 case llvm::Triple::amdgcn: 1879 return EmitAMDGPUBuiltinExpr(BuiltinID, E); 1880 case llvm::Triple::systemz: 1881 return EmitSystemZBuiltinExpr(BuiltinID, E); 1882 case llvm::Triple::nvptx: 1883 case llvm::Triple::nvptx64: 1884 return EmitNVPTXBuiltinExpr(BuiltinID, E); 1885 default: 1886 return nullptr; 1887 } 1888 } 1889 1890 static llvm::VectorType *GetNeonType(CodeGenFunction *CGF, 1891 NeonTypeFlags TypeFlags, 1892 bool V1Ty=false) { 1893 int IsQuad = TypeFlags.isQuad(); 1894 switch (TypeFlags.getEltType()) { 1895 case NeonTypeFlags::Int8: 1896 case NeonTypeFlags::Poly8: 1897 return llvm::VectorType::get(CGF->Int8Ty, V1Ty ? 1 : (8 << IsQuad)); 1898 case NeonTypeFlags::Int16: 1899 case NeonTypeFlags::Poly16: 1900 case NeonTypeFlags::Float16: 1901 return llvm::VectorType::get(CGF->Int16Ty, V1Ty ? 1 : (4 << IsQuad)); 1902 case NeonTypeFlags::Int32: 1903 return llvm::VectorType::get(CGF->Int32Ty, V1Ty ? 1 : (2 << IsQuad)); 1904 case NeonTypeFlags::Int64: 1905 case NeonTypeFlags::Poly64: 1906 return llvm::VectorType::get(CGF->Int64Ty, V1Ty ? 1 : (1 << IsQuad)); 1907 case NeonTypeFlags::Poly128: 1908 // FIXME: i128 and f128 doesn't get fully support in Clang and llvm. 1909 // There is a lot of i128 and f128 API missing. 1910 // so we use v16i8 to represent poly128 and get pattern matched. 1911 return llvm::VectorType::get(CGF->Int8Ty, 16); 1912 case NeonTypeFlags::Float32: 1913 return llvm::VectorType::get(CGF->FloatTy, V1Ty ? 1 : (2 << IsQuad)); 1914 case NeonTypeFlags::Float64: 1915 return llvm::VectorType::get(CGF->DoubleTy, V1Ty ? 1 : (1 << IsQuad)); 1916 } 1917 llvm_unreachable("Unknown vector element type!"); 1918 } 1919 1920 Value *CodeGenFunction::EmitNeonSplat(Value *V, Constant *C) { 1921 unsigned nElts = cast<llvm::VectorType>(V->getType())->getNumElements(); 1922 Value* SV = llvm::ConstantVector::getSplat(nElts, C); 1923 return Builder.CreateShuffleVector(V, V, SV, "lane"); 1924 } 1925 1926 Value *CodeGenFunction::EmitNeonCall(Function *F, SmallVectorImpl<Value*> &Ops, 1927 const char *name, 1928 unsigned shift, bool rightshift) { 1929 unsigned j = 0; 1930 for (Function::const_arg_iterator ai = F->arg_begin(), ae = F->arg_end(); 1931 ai != ae; ++ai, ++j) 1932 if (shift > 0 && shift == j) 1933 Ops[j] = EmitNeonShiftVector(Ops[j], ai->getType(), rightshift); 1934 else 1935 Ops[j] = Builder.CreateBitCast(Ops[j], ai->getType(), name); 1936 1937 return Builder.CreateCall(F, Ops, name); 1938 } 1939 1940 Value *CodeGenFunction::EmitNeonShiftVector(Value *V, llvm::Type *Ty, 1941 bool neg) { 1942 int SV = cast<ConstantInt>(V)->getSExtValue(); 1943 return ConstantInt::get(Ty, neg ? -SV : SV); 1944 } 1945 1946 // \brief Right-shift a vector by a constant. 1947 Value *CodeGenFunction::EmitNeonRShiftImm(Value *Vec, Value *Shift, 1948 llvm::Type *Ty, bool usgn, 1949 const char *name) { 1950 llvm::VectorType *VTy = cast<llvm::VectorType>(Ty); 1951 1952 int ShiftAmt = cast<ConstantInt>(Shift)->getSExtValue(); 1953 int EltSize = VTy->getScalarSizeInBits(); 1954 1955 Vec = Builder.CreateBitCast(Vec, Ty); 1956 1957 // lshr/ashr are undefined when the shift amount is equal to the vector 1958 // element size. 1959 if (ShiftAmt == EltSize) { 1960 if (usgn) { 1961 // Right-shifting an unsigned value by its size yields 0. 1962 return llvm::ConstantAggregateZero::get(VTy); 1963 } else { 1964 // Right-shifting a signed value by its size is equivalent 1965 // to a shift of size-1. 1966 --ShiftAmt; 1967 Shift = ConstantInt::get(VTy->getElementType(), ShiftAmt); 1968 } 1969 } 1970 1971 Shift = EmitNeonShiftVector(Shift, Ty, false); 1972 if (usgn) 1973 return Builder.CreateLShr(Vec, Shift, name); 1974 else 1975 return Builder.CreateAShr(Vec, Shift, name); 1976 } 1977 1978 /// GetPointeeAlignment - Given an expression with a pointer type, find the 1979 /// alignment of the type referenced by the pointer. Skip over implicit 1980 /// casts. 1981 std::pair<llvm::Value*, unsigned> 1982 CodeGenFunction::EmitPointerWithAlignment(const Expr *Addr) { 1983 assert(Addr->getType()->isPointerType()); 1984 Addr = Addr->IgnoreParens(); 1985 if (const ImplicitCastExpr *ICE = dyn_cast<ImplicitCastExpr>(Addr)) { 1986 if ((ICE->getCastKind() == CK_BitCast || ICE->getCastKind() == CK_NoOp) && 1987 ICE->getSubExpr()->getType()->isPointerType()) { 1988 std::pair<llvm::Value*, unsigned> Ptr = 1989 EmitPointerWithAlignment(ICE->getSubExpr()); 1990 Ptr.first = Builder.CreateBitCast(Ptr.first, 1991 ConvertType(Addr->getType())); 1992 return Ptr; 1993 } else if (ICE->getCastKind() == CK_ArrayToPointerDecay) { 1994 LValue LV = EmitLValue(ICE->getSubExpr()); 1995 unsigned Align = LV.getAlignment().getQuantity(); 1996 if (!Align) { 1997 // FIXME: Once LValues are fixed to always set alignment, 1998 // zap this code. 1999 QualType PtTy = ICE->getSubExpr()->getType(); 2000 if (!PtTy->isIncompleteType()) 2001 Align = getContext().getTypeAlignInChars(PtTy).getQuantity(); 2002 else 2003 Align = 1; 2004 } 2005 return std::make_pair(LV.getAddress(), Align); 2006 } 2007 } 2008 if (const UnaryOperator *UO = dyn_cast<UnaryOperator>(Addr)) { 2009 if (UO->getOpcode() == UO_AddrOf) { 2010 LValue LV = EmitLValue(UO->getSubExpr()); 2011 unsigned Align = LV.getAlignment().getQuantity(); 2012 if (!Align) { 2013 // FIXME: Once LValues are fixed to always set alignment, 2014 // zap this code. 2015 QualType PtTy = UO->getSubExpr()->getType(); 2016 if (!PtTy->isIncompleteType()) 2017 Align = getContext().getTypeAlignInChars(PtTy).getQuantity(); 2018 else 2019 Align = 1; 2020 } 2021 return std::make_pair(LV.getAddress(), Align); 2022 } 2023 } 2024 2025 unsigned Align = 1; 2026 QualType PtTy = Addr->getType()->getPointeeType(); 2027 if (!PtTy->isIncompleteType()) 2028 Align = getContext().getTypeAlignInChars(PtTy).getQuantity(); 2029 2030 return std::make_pair(EmitScalarExpr(Addr), Align); 2031 } 2032 2033 enum { 2034 AddRetType = (1 << 0), 2035 Add1ArgType = (1 << 1), 2036 Add2ArgTypes = (1 << 2), 2037 2038 VectorizeRetType = (1 << 3), 2039 VectorizeArgTypes = (1 << 4), 2040 2041 InventFloatType = (1 << 5), 2042 UnsignedAlts = (1 << 6), 2043 2044 Use64BitVectors = (1 << 7), 2045 Use128BitVectors = (1 << 8), 2046 2047 Vectorize1ArgType = Add1ArgType | VectorizeArgTypes, 2048 VectorRet = AddRetType | VectorizeRetType, 2049 VectorRetGetArgs01 = 2050 AddRetType | Add2ArgTypes | VectorizeRetType | VectorizeArgTypes, 2051 FpCmpzModifiers = 2052 AddRetType | VectorizeRetType | Add1ArgType | InventFloatType 2053 }; 2054 2055 struct NeonIntrinsicInfo { 2056 unsigned BuiltinID; 2057 unsigned LLVMIntrinsic; 2058 unsigned AltLLVMIntrinsic; 2059 const char *NameHint; 2060 unsigned TypeModifier; 2061 2062 bool operator<(unsigned RHSBuiltinID) const { 2063 return BuiltinID < RHSBuiltinID; 2064 } 2065 }; 2066 2067 #define NEONMAP0(NameBase) \ 2068 { NEON::BI__builtin_neon_ ## NameBase, 0, 0, #NameBase, 0 } 2069 2070 #define NEONMAP1(NameBase, LLVMIntrinsic, TypeModifier) \ 2071 { NEON:: BI__builtin_neon_ ## NameBase, \ 2072 Intrinsic::LLVMIntrinsic, 0, #NameBase, TypeModifier } 2073 2074 #define NEONMAP2(NameBase, LLVMIntrinsic, AltLLVMIntrinsic, TypeModifier) \ 2075 { NEON:: BI__builtin_neon_ ## NameBase, \ 2076 Intrinsic::LLVMIntrinsic, Intrinsic::AltLLVMIntrinsic, \ 2077 #NameBase, TypeModifier } 2078 2079 static NeonIntrinsicInfo ARMSIMDIntrinsicMap [] = { 2080 NEONMAP2(vabd_v, arm_neon_vabdu, arm_neon_vabds, Add1ArgType | UnsignedAlts), 2081 NEONMAP2(vabdq_v, arm_neon_vabdu, arm_neon_vabds, Add1ArgType | UnsignedAlts), 2082 NEONMAP1(vabs_v, arm_neon_vabs, 0), 2083 NEONMAP1(vabsq_v, arm_neon_vabs, 0), 2084 NEONMAP0(vaddhn_v), 2085 NEONMAP1(vaesdq_v, arm_neon_aesd, 0), 2086 NEONMAP1(vaeseq_v, arm_neon_aese, 0), 2087 NEONMAP1(vaesimcq_v, arm_neon_aesimc, 0), 2088 NEONMAP1(vaesmcq_v, arm_neon_aesmc, 0), 2089 NEONMAP1(vbsl_v, arm_neon_vbsl, AddRetType), 2090 NEONMAP1(vbslq_v, arm_neon_vbsl, AddRetType), 2091 NEONMAP1(vcage_v, arm_neon_vacge, 0), 2092 NEONMAP1(vcageq_v, arm_neon_vacge, 0), 2093 NEONMAP1(vcagt_v, arm_neon_vacgt, 0), 2094 NEONMAP1(vcagtq_v, arm_neon_vacgt, 0), 2095 NEONMAP1(vcale_v, arm_neon_vacge, 0), 2096 NEONMAP1(vcaleq_v, arm_neon_vacge, 0), 2097 NEONMAP1(vcalt_v, arm_neon_vacgt, 0), 2098 NEONMAP1(vcaltq_v, arm_neon_vacgt, 0), 2099 NEONMAP1(vcls_v, arm_neon_vcls, Add1ArgType), 2100 NEONMAP1(vclsq_v, arm_neon_vcls, Add1ArgType), 2101 NEONMAP1(vclz_v, ctlz, Add1ArgType), 2102 NEONMAP1(vclzq_v, ctlz, Add1ArgType), 2103 NEONMAP1(vcnt_v, ctpop, Add1ArgType), 2104 NEONMAP1(vcntq_v, ctpop, Add1ArgType), 2105 NEONMAP1(vcvt_f16_v, arm_neon_vcvtfp2hf, 0), 2106 NEONMAP1(vcvt_f32_f16, arm_neon_vcvthf2fp, 0), 2107 NEONMAP0(vcvt_f32_v), 2108 NEONMAP2(vcvt_n_f32_v, arm_neon_vcvtfxu2fp, arm_neon_vcvtfxs2fp, 0), 2109 NEONMAP1(vcvt_n_s32_v, arm_neon_vcvtfp2fxs, 0), 2110 NEONMAP1(vcvt_n_s64_v, arm_neon_vcvtfp2fxs, 0), 2111 NEONMAP1(vcvt_n_u32_v, arm_neon_vcvtfp2fxu, 0), 2112 NEONMAP1(vcvt_n_u64_v, arm_neon_vcvtfp2fxu, 0), 2113 NEONMAP0(vcvt_s32_v), 2114 NEONMAP0(vcvt_s64_v), 2115 NEONMAP0(vcvt_u32_v), 2116 NEONMAP0(vcvt_u64_v), 2117 NEONMAP1(vcvta_s32_v, arm_neon_vcvtas, 0), 2118 NEONMAP1(vcvta_s64_v, arm_neon_vcvtas, 0), 2119 NEONMAP1(vcvta_u32_v, arm_neon_vcvtau, 0), 2120 NEONMAP1(vcvta_u64_v, arm_neon_vcvtau, 0), 2121 NEONMAP1(vcvtaq_s32_v, arm_neon_vcvtas, 0), 2122 NEONMAP1(vcvtaq_s64_v, arm_neon_vcvtas, 0), 2123 NEONMAP1(vcvtaq_u32_v, arm_neon_vcvtau, 0), 2124 NEONMAP1(vcvtaq_u64_v, arm_neon_vcvtau, 0), 2125 NEONMAP1(vcvtm_s32_v, arm_neon_vcvtms, 0), 2126 NEONMAP1(vcvtm_s64_v, arm_neon_vcvtms, 0), 2127 NEONMAP1(vcvtm_u32_v, arm_neon_vcvtmu, 0), 2128 NEONMAP1(vcvtm_u64_v, arm_neon_vcvtmu, 0), 2129 NEONMAP1(vcvtmq_s32_v, arm_neon_vcvtms, 0), 2130 NEONMAP1(vcvtmq_s64_v, arm_neon_vcvtms, 0), 2131 NEONMAP1(vcvtmq_u32_v, arm_neon_vcvtmu, 0), 2132 NEONMAP1(vcvtmq_u64_v, arm_neon_vcvtmu, 0), 2133 NEONMAP1(vcvtn_s32_v, arm_neon_vcvtns, 0), 2134 NEONMAP1(vcvtn_s64_v, arm_neon_vcvtns, 0), 2135 NEONMAP1(vcvtn_u32_v, arm_neon_vcvtnu, 0), 2136 NEONMAP1(vcvtn_u64_v, arm_neon_vcvtnu, 0), 2137 NEONMAP1(vcvtnq_s32_v, arm_neon_vcvtns, 0), 2138 NEONMAP1(vcvtnq_s64_v, arm_neon_vcvtns, 0), 2139 NEONMAP1(vcvtnq_u32_v, arm_neon_vcvtnu, 0), 2140 NEONMAP1(vcvtnq_u64_v, arm_neon_vcvtnu, 0), 2141 NEONMAP1(vcvtp_s32_v, arm_neon_vcvtps, 0), 2142 NEONMAP1(vcvtp_s64_v, arm_neon_vcvtps, 0), 2143 NEONMAP1(vcvtp_u32_v, arm_neon_vcvtpu, 0), 2144 NEONMAP1(vcvtp_u64_v, arm_neon_vcvtpu, 0), 2145 NEONMAP1(vcvtpq_s32_v, arm_neon_vcvtps, 0), 2146 NEONMAP1(vcvtpq_s64_v, arm_neon_vcvtps, 0), 2147 NEONMAP1(vcvtpq_u32_v, arm_neon_vcvtpu, 0), 2148 NEONMAP1(vcvtpq_u64_v, arm_neon_vcvtpu, 0), 2149 NEONMAP0(vcvtq_f32_v), 2150 NEONMAP2(vcvtq_n_f32_v, arm_neon_vcvtfxu2fp, arm_neon_vcvtfxs2fp, 0), 2151 NEONMAP1(vcvtq_n_s32_v, arm_neon_vcvtfp2fxs, 0), 2152 NEONMAP1(vcvtq_n_s64_v, arm_neon_vcvtfp2fxs, 0), 2153 NEONMAP1(vcvtq_n_u32_v, arm_neon_vcvtfp2fxu, 0), 2154 NEONMAP1(vcvtq_n_u64_v, arm_neon_vcvtfp2fxu, 0), 2155 NEONMAP0(vcvtq_s32_v), 2156 NEONMAP0(vcvtq_s64_v), 2157 NEONMAP0(vcvtq_u32_v), 2158 NEONMAP0(vcvtq_u64_v), 2159 NEONMAP0(vext_v), 2160 NEONMAP0(vextq_v), 2161 NEONMAP0(vfma_v), 2162 NEONMAP0(vfmaq_v), 2163 NEONMAP2(vhadd_v, arm_neon_vhaddu, arm_neon_vhadds, Add1ArgType | UnsignedAlts), 2164 NEONMAP2(vhaddq_v, arm_neon_vhaddu, arm_neon_vhadds, Add1ArgType | UnsignedAlts), 2165 NEONMAP2(vhsub_v, arm_neon_vhsubu, arm_neon_vhsubs, Add1ArgType | UnsignedAlts), 2166 NEONMAP2(vhsubq_v, arm_neon_vhsubu, arm_neon_vhsubs, Add1ArgType | UnsignedAlts), 2167 NEONMAP0(vld1_dup_v), 2168 NEONMAP1(vld1_v, arm_neon_vld1, 0), 2169 NEONMAP0(vld1q_dup_v), 2170 NEONMAP1(vld1q_v, arm_neon_vld1, 0), 2171 NEONMAP1(vld2_lane_v, arm_neon_vld2lane, 0), 2172 NEONMAP1(vld2_v, arm_neon_vld2, 0), 2173 NEONMAP1(vld2q_lane_v, arm_neon_vld2lane, 0), 2174 NEONMAP1(vld2q_v, arm_neon_vld2, 0), 2175 NEONMAP1(vld3_lane_v, arm_neon_vld3lane, 0), 2176 NEONMAP1(vld3_v, arm_neon_vld3, 0), 2177 NEONMAP1(vld3q_lane_v, arm_neon_vld3lane, 0), 2178 NEONMAP1(vld3q_v, arm_neon_vld3, 0), 2179 NEONMAP1(vld4_lane_v, arm_neon_vld4lane, 0), 2180 NEONMAP1(vld4_v, arm_neon_vld4, 0), 2181 NEONMAP1(vld4q_lane_v, arm_neon_vld4lane, 0), 2182 NEONMAP1(vld4q_v, arm_neon_vld4, 0), 2183 NEONMAP2(vmax_v, arm_neon_vmaxu, arm_neon_vmaxs, Add1ArgType | UnsignedAlts), 2184 NEONMAP1(vmaxnm_v, arm_neon_vmaxnm, Add1ArgType), 2185 NEONMAP1(vmaxnmq_v, arm_neon_vmaxnm, Add1ArgType), 2186 NEONMAP2(vmaxq_v, arm_neon_vmaxu, arm_neon_vmaxs, Add1ArgType | UnsignedAlts), 2187 NEONMAP2(vmin_v, arm_neon_vminu, arm_neon_vmins, Add1ArgType | UnsignedAlts), 2188 NEONMAP1(vminnm_v, arm_neon_vminnm, Add1ArgType), 2189 NEONMAP1(vminnmq_v, arm_neon_vminnm, Add1ArgType), 2190 NEONMAP2(vminq_v, arm_neon_vminu, arm_neon_vmins, Add1ArgType | UnsignedAlts), 2191 NEONMAP0(vmovl_v), 2192 NEONMAP0(vmovn_v), 2193 NEONMAP1(vmul_v, arm_neon_vmulp, Add1ArgType), 2194 NEONMAP0(vmull_v), 2195 NEONMAP1(vmulq_v, arm_neon_vmulp, Add1ArgType), 2196 NEONMAP2(vpadal_v, arm_neon_vpadalu, arm_neon_vpadals, UnsignedAlts), 2197 NEONMAP2(vpadalq_v, arm_neon_vpadalu, arm_neon_vpadals, UnsignedAlts), 2198 NEONMAP1(vpadd_v, arm_neon_vpadd, Add1ArgType), 2199 NEONMAP2(vpaddl_v, arm_neon_vpaddlu, arm_neon_vpaddls, UnsignedAlts), 2200 NEONMAP2(vpaddlq_v, arm_neon_vpaddlu, arm_neon_vpaddls, UnsignedAlts), 2201 NEONMAP1(vpaddq_v, arm_neon_vpadd, Add1ArgType), 2202 NEONMAP2(vpmax_v, arm_neon_vpmaxu, arm_neon_vpmaxs, Add1ArgType | UnsignedAlts), 2203 NEONMAP2(vpmin_v, arm_neon_vpminu, arm_neon_vpmins, Add1ArgType | UnsignedAlts), 2204 NEONMAP1(vqabs_v, arm_neon_vqabs, Add1ArgType), 2205 NEONMAP1(vqabsq_v, arm_neon_vqabs, Add1ArgType), 2206 NEONMAP2(vqadd_v, arm_neon_vqaddu, arm_neon_vqadds, Add1ArgType | UnsignedAlts), 2207 NEONMAP2(vqaddq_v, arm_neon_vqaddu, arm_neon_vqadds, Add1ArgType | UnsignedAlts), 2208 NEONMAP2(vqdmlal_v, arm_neon_vqdmull, arm_neon_vqadds, 0), 2209 NEONMAP2(vqdmlsl_v, arm_neon_vqdmull, arm_neon_vqsubs, 0), 2210 NEONMAP1(vqdmulh_v, arm_neon_vqdmulh, Add1ArgType), 2211 NEONMAP1(vqdmulhq_v, arm_neon_vqdmulh, Add1ArgType), 2212 NEONMAP1(vqdmull_v, arm_neon_vqdmull, Add1ArgType), 2213 NEONMAP2(vqmovn_v, arm_neon_vqmovnu, arm_neon_vqmovns, Add1ArgType | UnsignedAlts), 2214 NEONMAP1(vqmovun_v, arm_neon_vqmovnsu, Add1ArgType), 2215 NEONMAP1(vqneg_v, arm_neon_vqneg, Add1ArgType), 2216 NEONMAP1(vqnegq_v, arm_neon_vqneg, Add1ArgType), 2217 NEONMAP1(vqrdmulh_v, arm_neon_vqrdmulh, Add1ArgType), 2218 NEONMAP1(vqrdmulhq_v, arm_neon_vqrdmulh, Add1ArgType), 2219 NEONMAP2(vqrshl_v, arm_neon_vqrshiftu, arm_neon_vqrshifts, Add1ArgType | UnsignedAlts), 2220 NEONMAP2(vqrshlq_v, arm_neon_vqrshiftu, arm_neon_vqrshifts, Add1ArgType | UnsignedAlts), 2221 NEONMAP2(vqshl_n_v, arm_neon_vqshiftu, arm_neon_vqshifts, UnsignedAlts), 2222 NEONMAP2(vqshl_v, arm_neon_vqshiftu, arm_neon_vqshifts, Add1ArgType | UnsignedAlts), 2223 NEONMAP2(vqshlq_n_v, arm_neon_vqshiftu, arm_neon_vqshifts, UnsignedAlts), 2224 NEONMAP2(vqshlq_v, arm_neon_vqshiftu, arm_neon_vqshifts, Add1ArgType | UnsignedAlts), 2225 NEONMAP1(vqshlu_n_v, arm_neon_vqshiftsu, 0), 2226 NEONMAP1(vqshluq_n_v, arm_neon_vqshiftsu, 0), 2227 NEONMAP2(vqsub_v, arm_neon_vqsubu, arm_neon_vqsubs, Add1ArgType | UnsignedAlts), 2228 NEONMAP2(vqsubq_v, arm_neon_vqsubu, arm_neon_vqsubs, Add1ArgType | UnsignedAlts), 2229 NEONMAP1(vraddhn_v, arm_neon_vraddhn, Add1ArgType), 2230 NEONMAP2(vrecpe_v, arm_neon_vrecpe, arm_neon_vrecpe, 0), 2231 NEONMAP2(vrecpeq_v, arm_neon_vrecpe, arm_neon_vrecpe, 0), 2232 NEONMAP1(vrecps_v, arm_neon_vrecps, Add1ArgType), 2233 NEONMAP1(vrecpsq_v, arm_neon_vrecps, Add1ArgType), 2234 NEONMAP2(vrhadd_v, arm_neon_vrhaddu, arm_neon_vrhadds, Add1ArgType | UnsignedAlts), 2235 NEONMAP2(vrhaddq_v, arm_neon_vrhaddu, arm_neon_vrhadds, Add1ArgType | UnsignedAlts), 2236 NEONMAP1(vrnd_v, arm_neon_vrintz, Add1ArgType), 2237 NEONMAP1(vrnda_v, arm_neon_vrinta, Add1ArgType), 2238 NEONMAP1(vrndaq_v, arm_neon_vrinta, Add1ArgType), 2239 NEONMAP1(vrndm_v, arm_neon_vrintm, Add1ArgType), 2240 NEONMAP1(vrndmq_v, arm_neon_vrintm, Add1ArgType), 2241 NEONMAP1(vrndn_v, arm_neon_vrintn, Add1ArgType), 2242 NEONMAP1(vrndnq_v, arm_neon_vrintn, Add1ArgType), 2243 NEONMAP1(vrndp_v, arm_neon_vrintp, Add1ArgType), 2244 NEONMAP1(vrndpq_v, arm_neon_vrintp, Add1ArgType), 2245 NEONMAP1(vrndq_v, arm_neon_vrintz, Add1ArgType), 2246 NEONMAP1(vrndx_v, arm_neon_vrintx, Add1ArgType), 2247 NEONMAP1(vrndxq_v, arm_neon_vrintx, Add1ArgType), 2248 NEONMAP2(vrshl_v, arm_neon_vrshiftu, arm_neon_vrshifts, Add1ArgType | UnsignedAlts), 2249 NEONMAP2(vrshlq_v, arm_neon_vrshiftu, arm_neon_vrshifts, Add1ArgType | UnsignedAlts), 2250 NEONMAP2(vrshr_n_v, arm_neon_vrshiftu, arm_neon_vrshifts, UnsignedAlts), 2251 NEONMAP2(vrshrq_n_v, arm_neon_vrshiftu, arm_neon_vrshifts, UnsignedAlts), 2252 NEONMAP2(vrsqrte_v, arm_neon_vrsqrte, arm_neon_vrsqrte, 0), 2253 NEONMAP2(vrsqrteq_v, arm_neon_vrsqrte, arm_neon_vrsqrte, 0), 2254 NEONMAP1(vrsqrts_v, arm_neon_vrsqrts, Add1ArgType), 2255 NEONMAP1(vrsqrtsq_v, arm_neon_vrsqrts, Add1ArgType), 2256 NEONMAP1(vrsubhn_v, arm_neon_vrsubhn, Add1ArgType), 2257 NEONMAP1(vsha1su0q_v, arm_neon_sha1su0, 0), 2258 NEONMAP1(vsha1su1q_v, arm_neon_sha1su1, 0), 2259 NEONMAP1(vsha256h2q_v, arm_neon_sha256h2, 0), 2260 NEONMAP1(vsha256hq_v, arm_neon_sha256h, 0), 2261 NEONMAP1(vsha256su0q_v, arm_neon_sha256su0, 0), 2262 NEONMAP1(vsha256su1q_v, arm_neon_sha256su1, 0), 2263 NEONMAP0(vshl_n_v), 2264 NEONMAP2(vshl_v, arm_neon_vshiftu, arm_neon_vshifts, Add1ArgType | UnsignedAlts), 2265 NEONMAP0(vshll_n_v), 2266 NEONMAP0(vshlq_n_v), 2267 NEONMAP2(vshlq_v, arm_neon_vshiftu, arm_neon_vshifts, Add1ArgType | UnsignedAlts), 2268 NEONMAP0(vshr_n_v), 2269 NEONMAP0(vshrn_n_v), 2270 NEONMAP0(vshrq_n_v), 2271 NEONMAP1(vst1_v, arm_neon_vst1, 0), 2272 NEONMAP1(vst1q_v, arm_neon_vst1, 0), 2273 NEONMAP1(vst2_lane_v, arm_neon_vst2lane, 0), 2274 NEONMAP1(vst2_v, arm_neon_vst2, 0), 2275 NEONMAP1(vst2q_lane_v, arm_neon_vst2lane, 0), 2276 NEONMAP1(vst2q_v, arm_neon_vst2, 0), 2277 NEONMAP1(vst3_lane_v, arm_neon_vst3lane, 0), 2278 NEONMAP1(vst3_v, arm_neon_vst3, 0), 2279 NEONMAP1(vst3q_lane_v, arm_neon_vst3lane, 0), 2280 NEONMAP1(vst3q_v, arm_neon_vst3, 0), 2281 NEONMAP1(vst4_lane_v, arm_neon_vst4lane, 0), 2282 NEONMAP1(vst4_v, arm_neon_vst4, 0), 2283 NEONMAP1(vst4q_lane_v, arm_neon_vst4lane, 0), 2284 NEONMAP1(vst4q_v, arm_neon_vst4, 0), 2285 NEONMAP0(vsubhn_v), 2286 NEONMAP0(vtrn_v), 2287 NEONMAP0(vtrnq_v), 2288 NEONMAP0(vtst_v), 2289 NEONMAP0(vtstq_v), 2290 NEONMAP0(vuzp_v), 2291 NEONMAP0(vuzpq_v), 2292 NEONMAP0(vzip_v), 2293 NEONMAP0(vzipq_v) 2294 }; 2295 2296 static NeonIntrinsicInfo AArch64SIMDIntrinsicMap[] = { 2297 NEONMAP1(vabs_v, aarch64_neon_abs, 0), 2298 NEONMAP1(vabsq_v, aarch64_neon_abs, 0), 2299 NEONMAP0(vaddhn_v), 2300 NEONMAP1(vaesdq_v, aarch64_crypto_aesd, 0), 2301 NEONMAP1(vaeseq_v, aarch64_crypto_aese, 0), 2302 NEONMAP1(vaesimcq_v, aarch64_crypto_aesimc, 0), 2303 NEONMAP1(vaesmcq_v, aarch64_crypto_aesmc, 0), 2304 NEONMAP1(vcage_v, aarch64_neon_facge, 0), 2305 NEONMAP1(vcageq_v, aarch64_neon_facge, 0), 2306 NEONMAP1(vcagt_v, aarch64_neon_facgt, 0), 2307 NEONMAP1(vcagtq_v, aarch64_neon_facgt, 0), 2308 NEONMAP1(vcale_v, aarch64_neon_facge, 0), 2309 NEONMAP1(vcaleq_v, aarch64_neon_facge, 0), 2310 NEONMAP1(vcalt_v, aarch64_neon_facgt, 0), 2311 NEONMAP1(vcaltq_v, aarch64_neon_facgt, 0), 2312 NEONMAP1(vcls_v, aarch64_neon_cls, Add1ArgType), 2313 NEONMAP1(vclsq_v, aarch64_neon_cls, Add1ArgType), 2314 NEONMAP1(vclz_v, ctlz, Add1ArgType), 2315 NEONMAP1(vclzq_v, ctlz, Add1ArgType), 2316 NEONMAP1(vcnt_v, ctpop, Add1ArgType), 2317 NEONMAP1(vcntq_v, ctpop, Add1ArgType), 2318 NEONMAP1(vcvt_f16_v, aarch64_neon_vcvtfp2hf, 0), 2319 NEONMAP1(vcvt_f32_f16, aarch64_neon_vcvthf2fp, 0), 2320 NEONMAP0(vcvt_f32_v), 2321 NEONMAP2(vcvt_n_f32_v, aarch64_neon_vcvtfxu2fp, aarch64_neon_vcvtfxs2fp, 0), 2322 NEONMAP2(vcvt_n_f64_v, aarch64_neon_vcvtfxu2fp, aarch64_neon_vcvtfxs2fp, 0), 2323 NEONMAP1(vcvt_n_s32_v, aarch64_neon_vcvtfp2fxs, 0), 2324 NEONMAP1(vcvt_n_s64_v, aarch64_neon_vcvtfp2fxs, 0), 2325 NEONMAP1(vcvt_n_u32_v, aarch64_neon_vcvtfp2fxu, 0), 2326 NEONMAP1(vcvt_n_u64_v, aarch64_neon_vcvtfp2fxu, 0), 2327 NEONMAP0(vcvtq_f32_v), 2328 NEONMAP2(vcvtq_n_f32_v, aarch64_neon_vcvtfxu2fp, aarch64_neon_vcvtfxs2fp, 0), 2329 NEONMAP2(vcvtq_n_f64_v, aarch64_neon_vcvtfxu2fp, aarch64_neon_vcvtfxs2fp, 0), 2330 NEONMAP1(vcvtq_n_s32_v, aarch64_neon_vcvtfp2fxs, 0), 2331 NEONMAP1(vcvtq_n_s64_v, aarch64_neon_vcvtfp2fxs, 0), 2332 NEONMAP1(vcvtq_n_u32_v, aarch64_neon_vcvtfp2fxu, 0), 2333 NEONMAP1(vcvtq_n_u64_v, aarch64_neon_vcvtfp2fxu, 0), 2334 NEONMAP1(vcvtx_f32_v, aarch64_neon_fcvtxn, AddRetType | Add1ArgType), 2335 NEONMAP0(vext_v), 2336 NEONMAP0(vextq_v), 2337 NEONMAP0(vfma_v), 2338 NEONMAP0(vfmaq_v), 2339 NEONMAP2(vhadd_v, aarch64_neon_uhadd, aarch64_neon_shadd, Add1ArgType | UnsignedAlts), 2340 NEONMAP2(vhaddq_v, aarch64_neon_uhadd, aarch64_neon_shadd, Add1ArgType | UnsignedAlts), 2341 NEONMAP2(vhsub_v, aarch64_neon_uhsub, aarch64_neon_shsub, Add1ArgType | UnsignedAlts), 2342 NEONMAP2(vhsubq_v, aarch64_neon_uhsub, aarch64_neon_shsub, Add1ArgType | UnsignedAlts), 2343 NEONMAP0(vmovl_v), 2344 NEONMAP0(vmovn_v), 2345 NEONMAP1(vmul_v, aarch64_neon_pmul, Add1ArgType), 2346 NEONMAP1(vmulq_v, aarch64_neon_pmul, Add1ArgType), 2347 NEONMAP1(vpadd_v, aarch64_neon_addp, Add1ArgType), 2348 NEONMAP2(vpaddl_v, aarch64_neon_uaddlp, aarch64_neon_saddlp, UnsignedAlts), 2349 NEONMAP2(vpaddlq_v, aarch64_neon_uaddlp, aarch64_neon_saddlp, UnsignedAlts), 2350 NEONMAP1(vpaddq_v, aarch64_neon_addp, Add1ArgType), 2351 NEONMAP1(vqabs_v, aarch64_neon_sqabs, Add1ArgType), 2352 NEONMAP1(vqabsq_v, aarch64_neon_sqabs, Add1ArgType), 2353 NEONMAP2(vqadd_v, aarch64_neon_uqadd, aarch64_neon_sqadd, Add1ArgType | UnsignedAlts), 2354 NEONMAP2(vqaddq_v, aarch64_neon_uqadd, aarch64_neon_sqadd, Add1ArgType | UnsignedAlts), 2355 NEONMAP2(vqdmlal_v, aarch64_neon_sqdmull, aarch64_neon_sqadd, 0), 2356 NEONMAP2(vqdmlsl_v, aarch64_neon_sqdmull, aarch64_neon_sqsub, 0), 2357 NEONMAP1(vqdmulh_v, aarch64_neon_sqdmulh, Add1ArgType), 2358 NEONMAP1(vqdmulhq_v, aarch64_neon_sqdmulh, Add1ArgType), 2359 NEONMAP1(vqdmull_v, aarch64_neon_sqdmull, Add1ArgType), 2360 NEONMAP2(vqmovn_v, aarch64_neon_uqxtn, aarch64_neon_sqxtn, Add1ArgType | UnsignedAlts), 2361 NEONMAP1(vqmovun_v, aarch64_neon_sqxtun, Add1ArgType), 2362 NEONMAP1(vqneg_v, aarch64_neon_sqneg, Add1ArgType), 2363 NEONMAP1(vqnegq_v, aarch64_neon_sqneg, Add1ArgType), 2364 NEONMAP1(vqrdmulh_v, aarch64_neon_sqrdmulh, Add1ArgType), 2365 NEONMAP1(vqrdmulhq_v, aarch64_neon_sqrdmulh, Add1ArgType), 2366 NEONMAP2(vqrshl_v, aarch64_neon_uqrshl, aarch64_neon_sqrshl, Add1ArgType | UnsignedAlts), 2367 NEONMAP2(vqrshlq_v, aarch64_neon_uqrshl, aarch64_neon_sqrshl, Add1ArgType | UnsignedAlts), 2368 NEONMAP2(vqshl_n_v, aarch64_neon_uqshl, aarch64_neon_sqshl, UnsignedAlts), 2369 NEONMAP2(vqshl_v, aarch64_neon_uqshl, aarch64_neon_sqshl, Add1ArgType | UnsignedAlts), 2370 NEONMAP2(vqshlq_n_v, aarch64_neon_uqshl, aarch64_neon_sqshl,UnsignedAlts), 2371 NEONMAP2(vqshlq_v, aarch64_neon_uqshl, aarch64_neon_sqshl, Add1ArgType | UnsignedAlts), 2372 NEONMAP1(vqshlu_n_v, aarch64_neon_sqshlu, 0), 2373 NEONMAP1(vqshluq_n_v, aarch64_neon_sqshlu, 0), 2374 NEONMAP2(vqsub_v, aarch64_neon_uqsub, aarch64_neon_sqsub, Add1ArgType | UnsignedAlts), 2375 NEONMAP2(vqsubq_v, aarch64_neon_uqsub, aarch64_neon_sqsub, Add1ArgType | UnsignedAlts), 2376 NEONMAP1(vraddhn_v, aarch64_neon_raddhn, Add1ArgType), 2377 NEONMAP2(vrecpe_v, aarch64_neon_frecpe, aarch64_neon_urecpe, 0), 2378 NEONMAP2(vrecpeq_v, aarch64_neon_frecpe, aarch64_neon_urecpe, 0), 2379 NEONMAP1(vrecps_v, aarch64_neon_frecps, Add1ArgType), 2380 NEONMAP1(vrecpsq_v, aarch64_neon_frecps, Add1ArgType), 2381 NEONMAP2(vrhadd_v, aarch64_neon_urhadd, aarch64_neon_srhadd, Add1ArgType | UnsignedAlts), 2382 NEONMAP2(vrhaddq_v, aarch64_neon_urhadd, aarch64_neon_srhadd, Add1ArgType | UnsignedAlts), 2383 NEONMAP2(vrshl_v, aarch64_neon_urshl, aarch64_neon_srshl, Add1ArgType | UnsignedAlts), 2384 NEONMAP2(vrshlq_v, aarch64_neon_urshl, aarch64_neon_srshl, Add1ArgType | UnsignedAlts), 2385 NEONMAP2(vrshr_n_v, aarch64_neon_urshl, aarch64_neon_srshl, UnsignedAlts), 2386 NEONMAP2(vrshrq_n_v, aarch64_neon_urshl, aarch64_neon_srshl, UnsignedAlts), 2387 NEONMAP2(vrsqrte_v, aarch64_neon_frsqrte, aarch64_neon_ursqrte, 0), 2388 NEONMAP2(vrsqrteq_v, aarch64_neon_frsqrte, aarch64_neon_ursqrte, 0), 2389 NEONMAP1(vrsqrts_v, aarch64_neon_frsqrts, Add1ArgType), 2390 NEONMAP1(vrsqrtsq_v, aarch64_neon_frsqrts, Add1ArgType), 2391 NEONMAP1(vrsubhn_v, aarch64_neon_rsubhn, Add1ArgType), 2392 NEONMAP1(vsha1su0q_v, aarch64_crypto_sha1su0, 0), 2393 NEONMAP1(vsha1su1q_v, aarch64_crypto_sha1su1, 0), 2394 NEONMAP1(vsha256h2q_v, aarch64_crypto_sha256h2, 0), 2395 NEONMAP1(vsha256hq_v, aarch64_crypto_sha256h, 0), 2396 NEONMAP1(vsha256su0q_v, aarch64_crypto_sha256su0, 0), 2397 NEONMAP1(vsha256su1q_v, aarch64_crypto_sha256su1, 0), 2398 NEONMAP0(vshl_n_v), 2399 NEONMAP2(vshl_v, aarch64_neon_ushl, aarch64_neon_sshl, Add1ArgType | UnsignedAlts), 2400 NEONMAP0(vshll_n_v), 2401 NEONMAP0(vshlq_n_v), 2402 NEONMAP2(vshlq_v, aarch64_neon_ushl, aarch64_neon_sshl, Add1ArgType | UnsignedAlts), 2403 NEONMAP0(vshr_n_v), 2404 NEONMAP0(vshrn_n_v), 2405 NEONMAP0(vshrq_n_v), 2406 NEONMAP0(vsubhn_v), 2407 NEONMAP0(vtst_v), 2408 NEONMAP0(vtstq_v), 2409 }; 2410 2411 static NeonIntrinsicInfo AArch64SISDIntrinsicMap[] = { 2412 NEONMAP1(vabdd_f64, aarch64_sisd_fabd, Add1ArgType), 2413 NEONMAP1(vabds_f32, aarch64_sisd_fabd, Add1ArgType), 2414 NEONMAP1(vabsd_s64, aarch64_neon_abs, Add1ArgType), 2415 NEONMAP1(vaddlv_s32, aarch64_neon_saddlv, AddRetType | Add1ArgType), 2416 NEONMAP1(vaddlv_u32, aarch64_neon_uaddlv, AddRetType | Add1ArgType), 2417 NEONMAP1(vaddlvq_s32, aarch64_neon_saddlv, AddRetType | Add1ArgType), 2418 NEONMAP1(vaddlvq_u32, aarch64_neon_uaddlv, AddRetType | Add1ArgType), 2419 NEONMAP1(vaddv_f32, aarch64_neon_faddv, AddRetType | Add1ArgType), 2420 NEONMAP1(vaddv_s32, aarch64_neon_saddv, AddRetType | Add1ArgType), 2421 NEONMAP1(vaddv_u32, aarch64_neon_uaddv, AddRetType | Add1ArgType), 2422 NEONMAP1(vaddvq_f32, aarch64_neon_faddv, AddRetType | Add1ArgType), 2423 NEONMAP1(vaddvq_f64, aarch64_neon_faddv, AddRetType | Add1ArgType), 2424 NEONMAP1(vaddvq_s32, aarch64_neon_saddv, AddRetType | Add1ArgType), 2425 NEONMAP1(vaddvq_s64, aarch64_neon_saddv, AddRetType | Add1ArgType), 2426 NEONMAP1(vaddvq_u32, aarch64_neon_uaddv, AddRetType | Add1ArgType), 2427 NEONMAP1(vaddvq_u64, aarch64_neon_uaddv, AddRetType | Add1ArgType), 2428 NEONMAP1(vcaged_f64, aarch64_neon_facge, AddRetType | Add1ArgType), 2429 NEONMAP1(vcages_f32, aarch64_neon_facge, AddRetType | Add1ArgType), 2430 NEONMAP1(vcagtd_f64, aarch64_neon_facgt, AddRetType | Add1ArgType), 2431 NEONMAP1(vcagts_f32, aarch64_neon_facgt, AddRetType | Add1ArgType), 2432 NEONMAP1(vcaled_f64, aarch64_neon_facge, AddRetType | Add1ArgType), 2433 NEONMAP1(vcales_f32, aarch64_neon_facge, AddRetType | Add1ArgType), 2434 NEONMAP1(vcaltd_f64, aarch64_neon_facgt, AddRetType | Add1ArgType), 2435 NEONMAP1(vcalts_f32, aarch64_neon_facgt, AddRetType | Add1ArgType), 2436 NEONMAP1(vcvtad_s64_f64, aarch64_neon_fcvtas, AddRetType | Add1ArgType), 2437 NEONMAP1(vcvtad_u64_f64, aarch64_neon_fcvtau, AddRetType | Add1ArgType), 2438 NEONMAP1(vcvtas_s32_f32, aarch64_neon_fcvtas, AddRetType | Add1ArgType), 2439 NEONMAP1(vcvtas_u32_f32, aarch64_neon_fcvtau, AddRetType | Add1ArgType), 2440 NEONMAP1(vcvtd_n_f64_s64, aarch64_neon_vcvtfxs2fp, AddRetType | Add1ArgType), 2441 NEONMAP1(vcvtd_n_f64_u64, aarch64_neon_vcvtfxu2fp, AddRetType | Add1ArgType), 2442 NEONMAP1(vcvtd_n_s64_f64, aarch64_neon_vcvtfp2fxs, AddRetType | Add1ArgType), 2443 NEONMAP1(vcvtd_n_u64_f64, aarch64_neon_vcvtfp2fxu, AddRetType | Add1ArgType), 2444 NEONMAP1(vcvtmd_s64_f64, aarch64_neon_fcvtms, AddRetType | Add1ArgType), 2445 NEONMAP1(vcvtmd_u64_f64, aarch64_neon_fcvtmu, AddRetType | Add1ArgType), 2446 NEONMAP1(vcvtms_s32_f32, aarch64_neon_fcvtms, AddRetType | Add1ArgType), 2447 NEONMAP1(vcvtms_u32_f32, aarch64_neon_fcvtmu, AddRetType | Add1ArgType), 2448 NEONMAP1(vcvtnd_s64_f64, aarch64_neon_fcvtns, AddRetType | Add1ArgType), 2449 NEONMAP1(vcvtnd_u64_f64, aarch64_neon_fcvtnu, AddRetType | Add1ArgType), 2450 NEONMAP1(vcvtns_s32_f32, aarch64_neon_fcvtns, AddRetType | Add1ArgType), 2451 NEONMAP1(vcvtns_u32_f32, aarch64_neon_fcvtnu, AddRetType | Add1ArgType), 2452 NEONMAP1(vcvtpd_s64_f64, aarch64_neon_fcvtps, AddRetType | Add1ArgType), 2453 NEONMAP1(vcvtpd_u64_f64, aarch64_neon_fcvtpu, AddRetType | Add1ArgType), 2454 NEONMAP1(vcvtps_s32_f32, aarch64_neon_fcvtps, AddRetType | Add1ArgType), 2455 NEONMAP1(vcvtps_u32_f32, aarch64_neon_fcvtpu, AddRetType | Add1ArgType), 2456 NEONMAP1(vcvts_n_f32_s32, aarch64_neon_vcvtfxs2fp, AddRetType | Add1ArgType), 2457 NEONMAP1(vcvts_n_f32_u32, aarch64_neon_vcvtfxu2fp, AddRetType | Add1ArgType), 2458 NEONMAP1(vcvts_n_s32_f32, aarch64_neon_vcvtfp2fxs, AddRetType | Add1ArgType), 2459 NEONMAP1(vcvts_n_u32_f32, aarch64_neon_vcvtfp2fxu, AddRetType | Add1ArgType), 2460 NEONMAP1(vcvtxd_f32_f64, aarch64_sisd_fcvtxn, 0), 2461 NEONMAP1(vmaxnmv_f32, aarch64_neon_fmaxnmv, AddRetType | Add1ArgType), 2462 NEONMAP1(vmaxnmvq_f32, aarch64_neon_fmaxnmv, AddRetType | Add1ArgType), 2463 NEONMAP1(vmaxnmvq_f64, aarch64_neon_fmaxnmv, AddRetType | Add1ArgType), 2464 NEONMAP1(vmaxv_f32, aarch64_neon_fmaxv, AddRetType | Add1ArgType), 2465 NEONMAP1(vmaxv_s32, aarch64_neon_smaxv, AddRetType | Add1ArgType), 2466 NEONMAP1(vmaxv_u32, aarch64_neon_umaxv, AddRetType | Add1ArgType), 2467 NEONMAP1(vmaxvq_f32, aarch64_neon_fmaxv, AddRetType | Add1ArgType), 2468 NEONMAP1(vmaxvq_f64, aarch64_neon_fmaxv, AddRetType | Add1ArgType), 2469 NEONMAP1(vmaxvq_s32, aarch64_neon_smaxv, AddRetType | Add1ArgType), 2470 NEONMAP1(vmaxvq_u32, aarch64_neon_umaxv, AddRetType | Add1ArgType), 2471 NEONMAP1(vminnmv_f32, aarch64_neon_fminnmv, AddRetType | Add1ArgType), 2472 NEONMAP1(vminnmvq_f32, aarch64_neon_fminnmv, AddRetType | Add1ArgType), 2473 NEONMAP1(vminnmvq_f64, aarch64_neon_fminnmv, AddRetType | Add1ArgType), 2474 NEONMAP1(vminv_f32, aarch64_neon_fminv, AddRetType | Add1ArgType), 2475 NEONMAP1(vminv_s32, aarch64_neon_sminv, AddRetType | Add1ArgType), 2476 NEONMAP1(vminv_u32, aarch64_neon_uminv, AddRetType | Add1ArgType), 2477 NEONMAP1(vminvq_f32, aarch64_neon_fminv, AddRetType | Add1ArgType), 2478 NEONMAP1(vminvq_f64, aarch64_neon_fminv, AddRetType | Add1ArgType), 2479 NEONMAP1(vminvq_s32, aarch64_neon_sminv, AddRetType | Add1ArgType), 2480 NEONMAP1(vminvq_u32, aarch64_neon_uminv, AddRetType | Add1ArgType), 2481 NEONMAP1(vmull_p64, aarch64_neon_pmull64, 0), 2482 NEONMAP1(vmulxd_f64, aarch64_neon_fmulx, Add1ArgType), 2483 NEONMAP1(vmulxs_f32, aarch64_neon_fmulx, Add1ArgType), 2484 NEONMAP1(vpaddd_s64, aarch64_neon_uaddv, AddRetType | Add1ArgType), 2485 NEONMAP1(vpaddd_u64, aarch64_neon_uaddv, AddRetType | Add1ArgType), 2486 NEONMAP1(vpmaxnmqd_f64, aarch64_neon_fmaxnmv, AddRetType | Add1ArgType), 2487 NEONMAP1(vpmaxnms_f32, aarch64_neon_fmaxnmv, AddRetType | Add1ArgType), 2488 NEONMAP1(vpmaxqd_f64, aarch64_neon_fmaxv, AddRetType | Add1ArgType), 2489 NEONMAP1(vpmaxs_f32, aarch64_neon_fmaxv, AddRetType | Add1ArgType), 2490 NEONMAP1(vpminnmqd_f64, aarch64_neon_fminnmv, AddRetType | Add1ArgType), 2491 NEONMAP1(vpminnms_f32, aarch64_neon_fminnmv, AddRetType | Add1ArgType), 2492 NEONMAP1(vpminqd_f64, aarch64_neon_fminv, AddRetType | Add1ArgType), 2493 NEONMAP1(vpmins_f32, aarch64_neon_fminv, AddRetType | Add1ArgType), 2494 NEONMAP1(vqabsb_s8, aarch64_neon_sqabs, Vectorize1ArgType | Use64BitVectors), 2495 NEONMAP1(vqabsd_s64, aarch64_neon_sqabs, Add1ArgType), 2496 NEONMAP1(vqabsh_s16, aarch64_neon_sqabs, Vectorize1ArgType | Use64BitVectors), 2497 NEONMAP1(vqabss_s32, aarch64_neon_sqabs, Add1ArgType), 2498 NEONMAP1(vqaddb_s8, aarch64_neon_sqadd, Vectorize1ArgType | Use64BitVectors), 2499 NEONMAP1(vqaddb_u8, aarch64_neon_uqadd, Vectorize1ArgType | Use64BitVectors), 2500 NEONMAP1(vqaddd_s64, aarch64_neon_sqadd, Add1ArgType), 2501 NEONMAP1(vqaddd_u64, aarch64_neon_uqadd, Add1ArgType), 2502 NEONMAP1(vqaddh_s16, aarch64_neon_sqadd, Vectorize1ArgType | Use64BitVectors), 2503 NEONMAP1(vqaddh_u16, aarch64_neon_uqadd, Vectorize1ArgType | Use64BitVectors), 2504 NEONMAP1(vqadds_s32, aarch64_neon_sqadd, Add1ArgType), 2505 NEONMAP1(vqadds_u32, aarch64_neon_uqadd, Add1ArgType), 2506 NEONMAP1(vqdmulhh_s16, aarch64_neon_sqdmulh, Vectorize1ArgType | Use64BitVectors), 2507 NEONMAP1(vqdmulhs_s32, aarch64_neon_sqdmulh, Add1ArgType), 2508 NEONMAP1(vqdmullh_s16, aarch64_neon_sqdmull, VectorRet | Use128BitVectors), 2509 NEONMAP1(vqdmulls_s32, aarch64_neon_sqdmulls_scalar, 0), 2510 NEONMAP1(vqmovnd_s64, aarch64_neon_scalar_sqxtn, AddRetType | Add1ArgType), 2511 NEONMAP1(vqmovnd_u64, aarch64_neon_scalar_uqxtn, AddRetType | Add1ArgType), 2512 NEONMAP1(vqmovnh_s16, aarch64_neon_sqxtn, VectorRet | Use64BitVectors), 2513 NEONMAP1(vqmovnh_u16, aarch64_neon_uqxtn, VectorRet | Use64BitVectors), 2514 NEONMAP1(vqmovns_s32, aarch64_neon_sqxtn, VectorRet | Use64BitVectors), 2515 NEONMAP1(vqmovns_u32, aarch64_neon_uqxtn, VectorRet | Use64BitVectors), 2516 NEONMAP1(vqmovund_s64, aarch64_neon_scalar_sqxtun, AddRetType | Add1ArgType), 2517 NEONMAP1(vqmovunh_s16, aarch64_neon_sqxtun, VectorRet | Use64BitVectors), 2518 NEONMAP1(vqmovuns_s32, aarch64_neon_sqxtun, VectorRet | Use64BitVectors), 2519 NEONMAP1(vqnegb_s8, aarch64_neon_sqneg, Vectorize1ArgType | Use64BitVectors), 2520 NEONMAP1(vqnegd_s64, aarch64_neon_sqneg, Add1ArgType), 2521 NEONMAP1(vqnegh_s16, aarch64_neon_sqneg, Vectorize1ArgType | Use64BitVectors), 2522 NEONMAP1(vqnegs_s32, aarch64_neon_sqneg, Add1ArgType), 2523 NEONMAP1(vqrdmulhh_s16, aarch64_neon_sqrdmulh, Vectorize1ArgType | Use64BitVectors), 2524 NEONMAP1(vqrdmulhs_s32, aarch64_neon_sqrdmulh, Add1ArgType), 2525 NEONMAP1(vqrshlb_s8, aarch64_neon_sqrshl, Vectorize1ArgType | Use64BitVectors), 2526 NEONMAP1(vqrshlb_u8, aarch64_neon_uqrshl, Vectorize1ArgType | Use64BitVectors), 2527 NEONMAP1(vqrshld_s64, aarch64_neon_sqrshl, Add1ArgType), 2528 NEONMAP1(vqrshld_u64, aarch64_neon_uqrshl, Add1ArgType), 2529 NEONMAP1(vqrshlh_s16, aarch64_neon_sqrshl, Vectorize1ArgType | Use64BitVectors), 2530 NEONMAP1(vqrshlh_u16, aarch64_neon_uqrshl, Vectorize1ArgType | Use64BitVectors), 2531 NEONMAP1(vqrshls_s32, aarch64_neon_sqrshl, Add1ArgType), 2532 NEONMAP1(vqrshls_u32, aarch64_neon_uqrshl, Add1ArgType), 2533 NEONMAP1(vqrshrnd_n_s64, aarch64_neon_sqrshrn, AddRetType), 2534 NEONMAP1(vqrshrnd_n_u64, aarch64_neon_uqrshrn, AddRetType), 2535 NEONMAP1(vqrshrnh_n_s16, aarch64_neon_sqrshrn, VectorRet | Use64BitVectors), 2536 NEONMAP1(vqrshrnh_n_u16, aarch64_neon_uqrshrn, VectorRet | Use64BitVectors), 2537 NEONMAP1(vqrshrns_n_s32, aarch64_neon_sqrshrn, VectorRet | Use64BitVectors), 2538 NEONMAP1(vqrshrns_n_u32, aarch64_neon_uqrshrn, VectorRet | Use64BitVectors), 2539 NEONMAP1(vqrshrund_n_s64, aarch64_neon_sqrshrun, AddRetType), 2540 NEONMAP1(vqrshrunh_n_s16, aarch64_neon_sqrshrun, VectorRet | Use64BitVectors), 2541 NEONMAP1(vqrshruns_n_s32, aarch64_neon_sqrshrun, VectorRet | Use64BitVectors), 2542 NEONMAP1(vqshlb_n_s8, aarch64_neon_sqshl, Vectorize1ArgType | Use64BitVectors), 2543 NEONMAP1(vqshlb_n_u8, aarch64_neon_uqshl, Vectorize1ArgType | Use64BitVectors), 2544 NEONMAP1(vqshlb_s8, aarch64_neon_sqshl, Vectorize1ArgType | Use64BitVectors), 2545 NEONMAP1(vqshlb_u8, aarch64_neon_uqshl, Vectorize1ArgType | Use64BitVectors), 2546 NEONMAP1(vqshld_s64, aarch64_neon_sqshl, Add1ArgType), 2547 NEONMAP1(vqshld_u64, aarch64_neon_uqshl, Add1ArgType), 2548 NEONMAP1(vqshlh_n_s16, aarch64_neon_sqshl, Vectorize1ArgType | Use64BitVectors), 2549 NEONMAP1(vqshlh_n_u16, aarch64_neon_uqshl, Vectorize1ArgType | Use64BitVectors), 2550 NEONMAP1(vqshlh_s16, aarch64_neon_sqshl, Vectorize1ArgType | Use64BitVectors), 2551 NEONMAP1(vqshlh_u16, aarch64_neon_uqshl, Vectorize1ArgType | Use64BitVectors), 2552 NEONMAP1(vqshls_n_s32, aarch64_neon_sqshl, Add1ArgType), 2553 NEONMAP1(vqshls_n_u32, aarch64_neon_uqshl, Add1ArgType), 2554 NEONMAP1(vqshls_s32, aarch64_neon_sqshl, Add1ArgType), 2555 NEONMAP1(vqshls_u32, aarch64_neon_uqshl, Add1ArgType), 2556 NEONMAP1(vqshlub_n_s8, aarch64_neon_sqshlu, Vectorize1ArgType | Use64BitVectors), 2557 NEONMAP1(vqshluh_n_s16, aarch64_neon_sqshlu, Vectorize1ArgType | Use64BitVectors), 2558 NEONMAP1(vqshlus_n_s32, aarch64_neon_sqshlu, Add1ArgType), 2559 NEONMAP1(vqshrnd_n_s64, aarch64_neon_sqshrn, AddRetType), 2560 NEONMAP1(vqshrnd_n_u64, aarch64_neon_uqshrn, AddRetType), 2561 NEONMAP1(vqshrnh_n_s16, aarch64_neon_sqshrn, VectorRet | Use64BitVectors), 2562 NEONMAP1(vqshrnh_n_u16, aarch64_neon_uqshrn, VectorRet | Use64BitVectors), 2563 NEONMAP1(vqshrns_n_s32, aarch64_neon_sqshrn, VectorRet | Use64BitVectors), 2564 NEONMAP1(vqshrns_n_u32, aarch64_neon_uqshrn, VectorRet | Use64BitVectors), 2565 NEONMAP1(vqshrund_n_s64, aarch64_neon_sqshrun, AddRetType), 2566 NEONMAP1(vqshrunh_n_s16, aarch64_neon_sqshrun, VectorRet | Use64BitVectors), 2567 NEONMAP1(vqshruns_n_s32, aarch64_neon_sqshrun, VectorRet | Use64BitVectors), 2568 NEONMAP1(vqsubb_s8, aarch64_neon_sqsub, Vectorize1ArgType | Use64BitVectors), 2569 NEONMAP1(vqsubb_u8, aarch64_neon_uqsub, Vectorize1ArgType | Use64BitVectors), 2570 NEONMAP1(vqsubd_s64, aarch64_neon_sqsub, Add1ArgType), 2571 NEONMAP1(vqsubd_u64, aarch64_neon_uqsub, Add1ArgType), 2572 NEONMAP1(vqsubh_s16, aarch64_neon_sqsub, Vectorize1ArgType | Use64BitVectors), 2573 NEONMAP1(vqsubh_u16, aarch64_neon_uqsub, Vectorize1ArgType | Use64BitVectors), 2574 NEONMAP1(vqsubs_s32, aarch64_neon_sqsub, Add1ArgType), 2575 NEONMAP1(vqsubs_u32, aarch64_neon_uqsub, Add1ArgType), 2576 NEONMAP1(vrecped_f64, aarch64_neon_frecpe, Add1ArgType), 2577 NEONMAP1(vrecpes_f32, aarch64_neon_frecpe, Add1ArgType), 2578 NEONMAP1(vrecpxd_f64, aarch64_neon_frecpx, Add1ArgType), 2579 NEONMAP1(vrecpxs_f32, aarch64_neon_frecpx, Add1ArgType), 2580 NEONMAP1(vrshld_s64, aarch64_neon_srshl, Add1ArgType), 2581 NEONMAP1(vrshld_u64, aarch64_neon_urshl, Add1ArgType), 2582 NEONMAP1(vrsqrted_f64, aarch64_neon_frsqrte, Add1ArgType), 2583 NEONMAP1(vrsqrtes_f32, aarch64_neon_frsqrte, Add1ArgType), 2584 NEONMAP1(vrsqrtsd_f64, aarch64_neon_frsqrts, Add1ArgType), 2585 NEONMAP1(vrsqrtss_f32, aarch64_neon_frsqrts, Add1ArgType), 2586 NEONMAP1(vsha1cq_u32, aarch64_crypto_sha1c, 0), 2587 NEONMAP1(vsha1h_u32, aarch64_crypto_sha1h, 0), 2588 NEONMAP1(vsha1mq_u32, aarch64_crypto_sha1m, 0), 2589 NEONMAP1(vsha1pq_u32, aarch64_crypto_sha1p, 0), 2590 NEONMAP1(vshld_s64, aarch64_neon_sshl, Add1ArgType), 2591 NEONMAP1(vshld_u64, aarch64_neon_ushl, Add1ArgType), 2592 NEONMAP1(vslid_n_s64, aarch64_neon_vsli, Vectorize1ArgType), 2593 NEONMAP1(vslid_n_u64, aarch64_neon_vsli, Vectorize1ArgType), 2594 NEONMAP1(vsqaddb_u8, aarch64_neon_usqadd, Vectorize1ArgType | Use64BitVectors), 2595 NEONMAP1(vsqaddd_u64, aarch64_neon_usqadd, Add1ArgType), 2596 NEONMAP1(vsqaddh_u16, aarch64_neon_usqadd, Vectorize1ArgType | Use64BitVectors), 2597 NEONMAP1(vsqadds_u32, aarch64_neon_usqadd, Add1ArgType), 2598 NEONMAP1(vsrid_n_s64, aarch64_neon_vsri, Vectorize1ArgType), 2599 NEONMAP1(vsrid_n_u64, aarch64_neon_vsri, Vectorize1ArgType), 2600 NEONMAP1(vuqaddb_s8, aarch64_neon_suqadd, Vectorize1ArgType | Use64BitVectors), 2601 NEONMAP1(vuqaddd_s64, aarch64_neon_suqadd, Add1ArgType), 2602 NEONMAP1(vuqaddh_s16, aarch64_neon_suqadd, Vectorize1ArgType | Use64BitVectors), 2603 NEONMAP1(vuqadds_s32, aarch64_neon_suqadd, Add1ArgType), 2604 }; 2605 2606 #undef NEONMAP0 2607 #undef NEONMAP1 2608 #undef NEONMAP2 2609 2610 static bool NEONSIMDIntrinsicsProvenSorted = false; 2611 2612 static bool AArch64SIMDIntrinsicsProvenSorted = false; 2613 static bool AArch64SISDIntrinsicsProvenSorted = false; 2614 2615 2616 static const NeonIntrinsicInfo * 2617 findNeonIntrinsicInMap(ArrayRef<NeonIntrinsicInfo> IntrinsicMap, 2618 unsigned BuiltinID, bool &MapProvenSorted) { 2619 2620 #ifndef NDEBUG 2621 if (!MapProvenSorted) { 2622 // FIXME: use std::is_sorted once C++11 is allowed 2623 for (unsigned i = 0; i < IntrinsicMap.size() - 1; ++i) 2624 assert(IntrinsicMap[i].BuiltinID <= IntrinsicMap[i + 1].BuiltinID); 2625 MapProvenSorted = true; 2626 } 2627 #endif 2628 2629 const NeonIntrinsicInfo *Builtin = 2630 std::lower_bound(IntrinsicMap.begin(), IntrinsicMap.end(), BuiltinID); 2631 2632 if (Builtin != IntrinsicMap.end() && Builtin->BuiltinID == BuiltinID) 2633 return Builtin; 2634 2635 return nullptr; 2636 } 2637 2638 Function *CodeGenFunction::LookupNeonLLVMIntrinsic(unsigned IntrinsicID, 2639 unsigned Modifier, 2640 llvm::Type *ArgType, 2641 const CallExpr *E) { 2642 int VectorSize = 0; 2643 if (Modifier & Use64BitVectors) 2644 VectorSize = 64; 2645 else if (Modifier & Use128BitVectors) 2646 VectorSize = 128; 2647 2648 // Return type. 2649 SmallVector<llvm::Type *, 3> Tys; 2650 if (Modifier & AddRetType) { 2651 llvm::Type *Ty = ConvertType(E->getCallReturnType(getContext())); 2652 if (Modifier & VectorizeRetType) 2653 Ty = llvm::VectorType::get( 2654 Ty, VectorSize ? VectorSize / Ty->getPrimitiveSizeInBits() : 1); 2655 2656 Tys.push_back(Ty); 2657 } 2658 2659 // Arguments. 2660 if (Modifier & VectorizeArgTypes) { 2661 int Elts = VectorSize ? VectorSize / ArgType->getPrimitiveSizeInBits() : 1; 2662 ArgType = llvm::VectorType::get(ArgType, Elts); 2663 } 2664 2665 if (Modifier & (Add1ArgType | Add2ArgTypes)) 2666 Tys.push_back(ArgType); 2667 2668 if (Modifier & Add2ArgTypes) 2669 Tys.push_back(ArgType); 2670 2671 if (Modifier & InventFloatType) 2672 Tys.push_back(FloatTy); 2673 2674 return CGM.getIntrinsic(IntrinsicID, Tys); 2675 } 2676 2677 static Value *EmitCommonNeonSISDBuiltinExpr(CodeGenFunction &CGF, 2678 const NeonIntrinsicInfo &SISDInfo, 2679 SmallVectorImpl<Value *> &Ops, 2680 const CallExpr *E) { 2681 unsigned BuiltinID = SISDInfo.BuiltinID; 2682 unsigned int Int = SISDInfo.LLVMIntrinsic; 2683 unsigned Modifier = SISDInfo.TypeModifier; 2684 const char *s = SISDInfo.NameHint; 2685 2686 switch (BuiltinID) { 2687 case NEON::BI__builtin_neon_vcled_s64: 2688 case NEON::BI__builtin_neon_vcled_u64: 2689 case NEON::BI__builtin_neon_vcles_f32: 2690 case NEON::BI__builtin_neon_vcled_f64: 2691 case NEON::BI__builtin_neon_vcltd_s64: 2692 case NEON::BI__builtin_neon_vcltd_u64: 2693 case NEON::BI__builtin_neon_vclts_f32: 2694 case NEON::BI__builtin_neon_vcltd_f64: 2695 case NEON::BI__builtin_neon_vcales_f32: 2696 case NEON::BI__builtin_neon_vcaled_f64: 2697 case NEON::BI__builtin_neon_vcalts_f32: 2698 case NEON::BI__builtin_neon_vcaltd_f64: 2699 // Only one direction of comparisons actually exist, cmle is actually a cmge 2700 // with swapped operands. The table gives us the right intrinsic but we 2701 // still need to do the swap. 2702 std::swap(Ops[0], Ops[1]); 2703 break; 2704 } 2705 2706 assert(Int && "Generic code assumes a valid intrinsic"); 2707 2708 // Determine the type(s) of this overloaded AArch64 intrinsic. 2709 const Expr *Arg = E->getArg(0); 2710 llvm::Type *ArgTy = CGF.ConvertType(Arg->getType()); 2711 Function *F = CGF.LookupNeonLLVMIntrinsic(Int, Modifier, ArgTy, E); 2712 2713 int j = 0; 2714 ConstantInt *C0 = ConstantInt::get(CGF.SizeTy, 0); 2715 for (Function::const_arg_iterator ai = F->arg_begin(), ae = F->arg_end(); 2716 ai != ae; ++ai, ++j) { 2717 llvm::Type *ArgTy = ai->getType(); 2718 if (Ops[j]->getType()->getPrimitiveSizeInBits() == 2719 ArgTy->getPrimitiveSizeInBits()) 2720 continue; 2721 2722 assert(ArgTy->isVectorTy() && !Ops[j]->getType()->isVectorTy()); 2723 // The constant argument to an _n_ intrinsic always has Int32Ty, so truncate 2724 // it before inserting. 2725 Ops[j] = 2726 CGF.Builder.CreateTruncOrBitCast(Ops[j], ArgTy->getVectorElementType()); 2727 Ops[j] = 2728 CGF.Builder.CreateInsertElement(UndefValue::get(ArgTy), Ops[j], C0); 2729 } 2730 2731 Value *Result = CGF.EmitNeonCall(F, Ops, s); 2732 llvm::Type *ResultType = CGF.ConvertType(E->getType()); 2733 if (ResultType->getPrimitiveSizeInBits() < 2734 Result->getType()->getPrimitiveSizeInBits()) 2735 return CGF.Builder.CreateExtractElement(Result, C0); 2736 2737 return CGF.Builder.CreateBitCast(Result, ResultType, s); 2738 } 2739 2740 Value *CodeGenFunction::EmitCommonNeonBuiltinExpr( 2741 unsigned BuiltinID, unsigned LLVMIntrinsic, unsigned AltLLVMIntrinsic, 2742 const char *NameHint, unsigned Modifier, const CallExpr *E, 2743 SmallVectorImpl<llvm::Value *> &Ops, llvm::Value *Align) { 2744 // Get the last argument, which specifies the vector type. 2745 llvm::APSInt NeonTypeConst; 2746 const Expr *Arg = E->getArg(E->getNumArgs() - 1); 2747 if (!Arg->isIntegerConstantExpr(NeonTypeConst, getContext())) 2748 return nullptr; 2749 2750 // Determine the type of this overloaded NEON intrinsic. 2751 NeonTypeFlags Type(NeonTypeConst.getZExtValue()); 2752 bool Usgn = Type.isUnsigned(); 2753 bool Quad = Type.isQuad(); 2754 2755 llvm::VectorType *VTy = GetNeonType(this, Type); 2756 llvm::Type *Ty = VTy; 2757 if (!Ty) 2758 return nullptr; 2759 2760 unsigned Int = LLVMIntrinsic; 2761 if ((Modifier & UnsignedAlts) && !Usgn) 2762 Int = AltLLVMIntrinsic; 2763 2764 switch (BuiltinID) { 2765 default: break; 2766 case NEON::BI__builtin_neon_vabs_v: 2767 case NEON::BI__builtin_neon_vabsq_v: 2768 if (VTy->getElementType()->isFloatingPointTy()) 2769 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::fabs, Ty), Ops, "vabs"); 2770 return EmitNeonCall(CGM.getIntrinsic(LLVMIntrinsic, Ty), Ops, "vabs"); 2771 case NEON::BI__builtin_neon_vaddhn_v: { 2772 llvm::VectorType *SrcTy = 2773 llvm::VectorType::getExtendedElementVectorType(VTy); 2774 2775 // %sum = add <4 x i32> %lhs, %rhs 2776 Ops[0] = Builder.CreateBitCast(Ops[0], SrcTy); 2777 Ops[1] = Builder.CreateBitCast(Ops[1], SrcTy); 2778 Ops[0] = Builder.CreateAdd(Ops[0], Ops[1], "vaddhn"); 2779 2780 // %high = lshr <4 x i32> %sum, <i32 16, i32 16, i32 16, i32 16> 2781 Constant *ShiftAmt = 2782 ConstantInt::get(SrcTy, SrcTy->getScalarSizeInBits() / 2); 2783 Ops[0] = Builder.CreateLShr(Ops[0], ShiftAmt, "vaddhn"); 2784 2785 // %res = trunc <4 x i32> %high to <4 x i16> 2786 return Builder.CreateTrunc(Ops[0], VTy, "vaddhn"); 2787 } 2788 case NEON::BI__builtin_neon_vcale_v: 2789 case NEON::BI__builtin_neon_vcaleq_v: 2790 case NEON::BI__builtin_neon_vcalt_v: 2791 case NEON::BI__builtin_neon_vcaltq_v: 2792 std::swap(Ops[0], Ops[1]); 2793 case NEON::BI__builtin_neon_vcage_v: 2794 case NEON::BI__builtin_neon_vcageq_v: 2795 case NEON::BI__builtin_neon_vcagt_v: 2796 case NEON::BI__builtin_neon_vcagtq_v: { 2797 llvm::Type *VecFlt = llvm::VectorType::get( 2798 VTy->getScalarSizeInBits() == 32 ? FloatTy : DoubleTy, 2799 VTy->getNumElements()); 2800 llvm::Type *Tys[] = { VTy, VecFlt }; 2801 Function *F = CGM.getIntrinsic(LLVMIntrinsic, Tys); 2802 return EmitNeonCall(F, Ops, NameHint); 2803 } 2804 case NEON::BI__builtin_neon_vclz_v: 2805 case NEON::BI__builtin_neon_vclzq_v: 2806 // We generate target-independent intrinsic, which needs a second argument 2807 // for whether or not clz of zero is undefined; on ARM it isn't. 2808 Ops.push_back(Builder.getInt1(getTarget().isCLZForZeroUndef())); 2809 break; 2810 case NEON::BI__builtin_neon_vcvt_f32_v: 2811 case NEON::BI__builtin_neon_vcvtq_f32_v: 2812 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 2813 Ty = GetNeonType(this, NeonTypeFlags(NeonTypeFlags::Float32, false, Quad)); 2814 return Usgn ? Builder.CreateUIToFP(Ops[0], Ty, "vcvt") 2815 : Builder.CreateSIToFP(Ops[0], Ty, "vcvt"); 2816 case NEON::BI__builtin_neon_vcvt_n_f32_v: 2817 case NEON::BI__builtin_neon_vcvt_n_f64_v: 2818 case NEON::BI__builtin_neon_vcvtq_n_f32_v: 2819 case NEON::BI__builtin_neon_vcvtq_n_f64_v: { 2820 bool Double = 2821 (cast<llvm::IntegerType>(VTy->getElementType())->getBitWidth() == 64); 2822 llvm::Type *FloatTy = 2823 GetNeonType(this, NeonTypeFlags(Double ? NeonTypeFlags::Float64 2824 : NeonTypeFlags::Float32, 2825 false, Quad)); 2826 llvm::Type *Tys[2] = { FloatTy, Ty }; 2827 Int = Usgn ? LLVMIntrinsic : AltLLVMIntrinsic; 2828 Function *F = CGM.getIntrinsic(Int, Tys); 2829 return EmitNeonCall(F, Ops, "vcvt_n"); 2830 } 2831 case NEON::BI__builtin_neon_vcvt_n_s32_v: 2832 case NEON::BI__builtin_neon_vcvt_n_u32_v: 2833 case NEON::BI__builtin_neon_vcvt_n_s64_v: 2834 case NEON::BI__builtin_neon_vcvt_n_u64_v: 2835 case NEON::BI__builtin_neon_vcvtq_n_s32_v: 2836 case NEON::BI__builtin_neon_vcvtq_n_u32_v: 2837 case NEON::BI__builtin_neon_vcvtq_n_s64_v: 2838 case NEON::BI__builtin_neon_vcvtq_n_u64_v: { 2839 bool Double = 2840 (cast<llvm::IntegerType>(VTy->getElementType())->getBitWidth() == 64); 2841 llvm::Type *FloatTy = 2842 GetNeonType(this, NeonTypeFlags(Double ? NeonTypeFlags::Float64 2843 : NeonTypeFlags::Float32, 2844 false, Quad)); 2845 llvm::Type *Tys[2] = { Ty, FloatTy }; 2846 Function *F = CGM.getIntrinsic(LLVMIntrinsic, Tys); 2847 return EmitNeonCall(F, Ops, "vcvt_n"); 2848 } 2849 case NEON::BI__builtin_neon_vcvt_s32_v: 2850 case NEON::BI__builtin_neon_vcvt_u32_v: 2851 case NEON::BI__builtin_neon_vcvt_s64_v: 2852 case NEON::BI__builtin_neon_vcvt_u64_v: 2853 case NEON::BI__builtin_neon_vcvtq_s32_v: 2854 case NEON::BI__builtin_neon_vcvtq_u32_v: 2855 case NEON::BI__builtin_neon_vcvtq_s64_v: 2856 case NEON::BI__builtin_neon_vcvtq_u64_v: { 2857 bool Double = 2858 (cast<llvm::IntegerType>(VTy->getElementType())->getBitWidth() == 64); 2859 llvm::Type *FloatTy = 2860 GetNeonType(this, NeonTypeFlags(Double ? NeonTypeFlags::Float64 2861 : NeonTypeFlags::Float32, 2862 false, Quad)); 2863 Ops[0] = Builder.CreateBitCast(Ops[0], FloatTy); 2864 return Usgn ? Builder.CreateFPToUI(Ops[0], Ty, "vcvt") 2865 : Builder.CreateFPToSI(Ops[0], Ty, "vcvt"); 2866 } 2867 case NEON::BI__builtin_neon_vcvta_s32_v: 2868 case NEON::BI__builtin_neon_vcvta_s64_v: 2869 case NEON::BI__builtin_neon_vcvta_u32_v: 2870 case NEON::BI__builtin_neon_vcvta_u64_v: 2871 case NEON::BI__builtin_neon_vcvtaq_s32_v: 2872 case NEON::BI__builtin_neon_vcvtaq_s64_v: 2873 case NEON::BI__builtin_neon_vcvtaq_u32_v: 2874 case NEON::BI__builtin_neon_vcvtaq_u64_v: 2875 case NEON::BI__builtin_neon_vcvtn_s32_v: 2876 case NEON::BI__builtin_neon_vcvtn_s64_v: 2877 case NEON::BI__builtin_neon_vcvtn_u32_v: 2878 case NEON::BI__builtin_neon_vcvtn_u64_v: 2879 case NEON::BI__builtin_neon_vcvtnq_s32_v: 2880 case NEON::BI__builtin_neon_vcvtnq_s64_v: 2881 case NEON::BI__builtin_neon_vcvtnq_u32_v: 2882 case NEON::BI__builtin_neon_vcvtnq_u64_v: 2883 case NEON::BI__builtin_neon_vcvtp_s32_v: 2884 case NEON::BI__builtin_neon_vcvtp_s64_v: 2885 case NEON::BI__builtin_neon_vcvtp_u32_v: 2886 case NEON::BI__builtin_neon_vcvtp_u64_v: 2887 case NEON::BI__builtin_neon_vcvtpq_s32_v: 2888 case NEON::BI__builtin_neon_vcvtpq_s64_v: 2889 case NEON::BI__builtin_neon_vcvtpq_u32_v: 2890 case NEON::BI__builtin_neon_vcvtpq_u64_v: 2891 case NEON::BI__builtin_neon_vcvtm_s32_v: 2892 case NEON::BI__builtin_neon_vcvtm_s64_v: 2893 case NEON::BI__builtin_neon_vcvtm_u32_v: 2894 case NEON::BI__builtin_neon_vcvtm_u64_v: 2895 case NEON::BI__builtin_neon_vcvtmq_s32_v: 2896 case NEON::BI__builtin_neon_vcvtmq_s64_v: 2897 case NEON::BI__builtin_neon_vcvtmq_u32_v: 2898 case NEON::BI__builtin_neon_vcvtmq_u64_v: { 2899 bool Double = 2900 (cast<llvm::IntegerType>(VTy->getElementType())->getBitWidth() == 64); 2901 llvm::Type *InTy = 2902 GetNeonType(this, 2903 NeonTypeFlags(Double ? NeonTypeFlags::Float64 2904 : NeonTypeFlags::Float32, false, Quad)); 2905 llvm::Type *Tys[2] = { Ty, InTy }; 2906 return EmitNeonCall(CGM.getIntrinsic(LLVMIntrinsic, Tys), Ops, NameHint); 2907 } 2908 case NEON::BI__builtin_neon_vext_v: 2909 case NEON::BI__builtin_neon_vextq_v: { 2910 int CV = cast<ConstantInt>(Ops[2])->getSExtValue(); 2911 SmallVector<Constant*, 16> Indices; 2912 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) 2913 Indices.push_back(ConstantInt::get(Int32Ty, i+CV)); 2914 2915 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 2916 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 2917 Value *SV = llvm::ConstantVector::get(Indices); 2918 return Builder.CreateShuffleVector(Ops[0], Ops[1], SV, "vext"); 2919 } 2920 case NEON::BI__builtin_neon_vfma_v: 2921 case NEON::BI__builtin_neon_vfmaq_v: { 2922 Value *F = CGM.getIntrinsic(Intrinsic::fma, Ty); 2923 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 2924 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 2925 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 2926 2927 // NEON intrinsic puts accumulator first, unlike the LLVM fma. 2928 return Builder.CreateCall(F, {Ops[1], Ops[2], Ops[0]}); 2929 } 2930 case NEON::BI__builtin_neon_vld1_v: 2931 case NEON::BI__builtin_neon_vld1q_v: 2932 Ops.push_back(Align); 2933 return EmitNeonCall(CGM.getIntrinsic(LLVMIntrinsic, Ty), Ops, "vld1"); 2934 case NEON::BI__builtin_neon_vld2_v: 2935 case NEON::BI__builtin_neon_vld2q_v: 2936 case NEON::BI__builtin_neon_vld3_v: 2937 case NEON::BI__builtin_neon_vld3q_v: 2938 case NEON::BI__builtin_neon_vld4_v: 2939 case NEON::BI__builtin_neon_vld4q_v: { 2940 Function *F = CGM.getIntrinsic(LLVMIntrinsic, Ty); 2941 Ops[1] = Builder.CreateCall(F, {Ops[1], Align}, NameHint); 2942 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 2943 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 2944 return Builder.CreateStore(Ops[1], Ops[0]); 2945 } 2946 case NEON::BI__builtin_neon_vld1_dup_v: 2947 case NEON::BI__builtin_neon_vld1q_dup_v: { 2948 Value *V = UndefValue::get(Ty); 2949 Ty = llvm::PointerType::getUnqual(VTy->getElementType()); 2950 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 2951 LoadInst *Ld = Builder.CreateLoad(Ops[0]); 2952 Ld->setAlignment(cast<ConstantInt>(Align)->getZExtValue()); 2953 llvm::Constant *CI = ConstantInt::get(SizeTy, 0); 2954 Ops[0] = Builder.CreateInsertElement(V, Ld, CI); 2955 return EmitNeonSplat(Ops[0], CI); 2956 } 2957 case NEON::BI__builtin_neon_vld2_lane_v: 2958 case NEON::BI__builtin_neon_vld2q_lane_v: 2959 case NEON::BI__builtin_neon_vld3_lane_v: 2960 case NEON::BI__builtin_neon_vld3q_lane_v: 2961 case NEON::BI__builtin_neon_vld4_lane_v: 2962 case NEON::BI__builtin_neon_vld4q_lane_v: { 2963 Function *F = CGM.getIntrinsic(LLVMIntrinsic, Ty); 2964 for (unsigned I = 2; I < Ops.size() - 1; ++I) 2965 Ops[I] = Builder.CreateBitCast(Ops[I], Ty); 2966 Ops.push_back(Align); 2967 Ops[1] = Builder.CreateCall(F, makeArrayRef(Ops).slice(1), NameHint); 2968 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 2969 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 2970 return Builder.CreateStore(Ops[1], Ops[0]); 2971 } 2972 case NEON::BI__builtin_neon_vmovl_v: { 2973 llvm::Type *DTy =llvm::VectorType::getTruncatedElementVectorType(VTy); 2974 Ops[0] = Builder.CreateBitCast(Ops[0], DTy); 2975 if (Usgn) 2976 return Builder.CreateZExt(Ops[0], Ty, "vmovl"); 2977 return Builder.CreateSExt(Ops[0], Ty, "vmovl"); 2978 } 2979 case NEON::BI__builtin_neon_vmovn_v: { 2980 llvm::Type *QTy = llvm::VectorType::getExtendedElementVectorType(VTy); 2981 Ops[0] = Builder.CreateBitCast(Ops[0], QTy); 2982 return Builder.CreateTrunc(Ops[0], Ty, "vmovn"); 2983 } 2984 case NEON::BI__builtin_neon_vmull_v: 2985 // FIXME: the integer vmull operations could be emitted in terms of pure 2986 // LLVM IR (2 exts followed by a mul). Unfortunately LLVM has a habit of 2987 // hoisting the exts outside loops. Until global ISel comes along that can 2988 // see through such movement this leads to bad CodeGen. So we need an 2989 // intrinsic for now. 2990 Int = Usgn ? Intrinsic::arm_neon_vmullu : Intrinsic::arm_neon_vmulls; 2991 Int = Type.isPoly() ? (unsigned)Intrinsic::arm_neon_vmullp : Int; 2992 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmull"); 2993 case NEON::BI__builtin_neon_vpadal_v: 2994 case NEON::BI__builtin_neon_vpadalq_v: { 2995 // The source operand type has twice as many elements of half the size. 2996 unsigned EltBits = VTy->getElementType()->getPrimitiveSizeInBits(); 2997 llvm::Type *EltTy = 2998 llvm::IntegerType::get(getLLVMContext(), EltBits / 2); 2999 llvm::Type *NarrowTy = 3000 llvm::VectorType::get(EltTy, VTy->getNumElements() * 2); 3001 llvm::Type *Tys[2] = { Ty, NarrowTy }; 3002 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, NameHint); 3003 } 3004 case NEON::BI__builtin_neon_vpaddl_v: 3005 case NEON::BI__builtin_neon_vpaddlq_v: { 3006 // The source operand type has twice as many elements of half the size. 3007 unsigned EltBits = VTy->getElementType()->getPrimitiveSizeInBits(); 3008 llvm::Type *EltTy = llvm::IntegerType::get(getLLVMContext(), EltBits / 2); 3009 llvm::Type *NarrowTy = 3010 llvm::VectorType::get(EltTy, VTy->getNumElements() * 2); 3011 llvm::Type *Tys[2] = { Ty, NarrowTy }; 3012 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vpaddl"); 3013 } 3014 case NEON::BI__builtin_neon_vqdmlal_v: 3015 case NEON::BI__builtin_neon_vqdmlsl_v: { 3016 SmallVector<Value *, 2> MulOps(Ops.begin() + 1, Ops.end()); 3017 Ops[1] = 3018 EmitNeonCall(CGM.getIntrinsic(LLVMIntrinsic, Ty), MulOps, "vqdmlal"); 3019 Ops.resize(2); 3020 return EmitNeonCall(CGM.getIntrinsic(AltLLVMIntrinsic, Ty), Ops, NameHint); 3021 } 3022 case NEON::BI__builtin_neon_vqshl_n_v: 3023 case NEON::BI__builtin_neon_vqshlq_n_v: 3024 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqshl_n", 3025 1, false); 3026 case NEON::BI__builtin_neon_vqshlu_n_v: 3027 case NEON::BI__builtin_neon_vqshluq_n_v: 3028 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqshlu_n", 3029 1, false); 3030 case NEON::BI__builtin_neon_vrecpe_v: 3031 case NEON::BI__builtin_neon_vrecpeq_v: 3032 case NEON::BI__builtin_neon_vrsqrte_v: 3033 case NEON::BI__builtin_neon_vrsqrteq_v: 3034 Int = Ty->isFPOrFPVectorTy() ? LLVMIntrinsic : AltLLVMIntrinsic; 3035 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, NameHint); 3036 3037 case NEON::BI__builtin_neon_vrshr_n_v: 3038 case NEON::BI__builtin_neon_vrshrq_n_v: 3039 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrshr_n", 3040 1, true); 3041 case NEON::BI__builtin_neon_vshl_n_v: 3042 case NEON::BI__builtin_neon_vshlq_n_v: 3043 Ops[1] = EmitNeonShiftVector(Ops[1], Ty, false); 3044 return Builder.CreateShl(Builder.CreateBitCast(Ops[0],Ty), Ops[1], 3045 "vshl_n"); 3046 case NEON::BI__builtin_neon_vshll_n_v: { 3047 llvm::Type *SrcTy = llvm::VectorType::getTruncatedElementVectorType(VTy); 3048 Ops[0] = Builder.CreateBitCast(Ops[0], SrcTy); 3049 if (Usgn) 3050 Ops[0] = Builder.CreateZExt(Ops[0], VTy); 3051 else 3052 Ops[0] = Builder.CreateSExt(Ops[0], VTy); 3053 Ops[1] = EmitNeonShiftVector(Ops[1], VTy, false); 3054 return Builder.CreateShl(Ops[0], Ops[1], "vshll_n"); 3055 } 3056 case NEON::BI__builtin_neon_vshrn_n_v: { 3057 llvm::Type *SrcTy = llvm::VectorType::getExtendedElementVectorType(VTy); 3058 Ops[0] = Builder.CreateBitCast(Ops[0], SrcTy); 3059 Ops[1] = EmitNeonShiftVector(Ops[1], SrcTy, false); 3060 if (Usgn) 3061 Ops[0] = Builder.CreateLShr(Ops[0], Ops[1]); 3062 else 3063 Ops[0] = Builder.CreateAShr(Ops[0], Ops[1]); 3064 return Builder.CreateTrunc(Ops[0], Ty, "vshrn_n"); 3065 } 3066 case NEON::BI__builtin_neon_vshr_n_v: 3067 case NEON::BI__builtin_neon_vshrq_n_v: 3068 return EmitNeonRShiftImm(Ops[0], Ops[1], Ty, Usgn, "vshr_n"); 3069 case NEON::BI__builtin_neon_vst1_v: 3070 case NEON::BI__builtin_neon_vst1q_v: 3071 case NEON::BI__builtin_neon_vst2_v: 3072 case NEON::BI__builtin_neon_vst2q_v: 3073 case NEON::BI__builtin_neon_vst3_v: 3074 case NEON::BI__builtin_neon_vst3q_v: 3075 case NEON::BI__builtin_neon_vst4_v: 3076 case NEON::BI__builtin_neon_vst4q_v: 3077 case NEON::BI__builtin_neon_vst2_lane_v: 3078 case NEON::BI__builtin_neon_vst2q_lane_v: 3079 case NEON::BI__builtin_neon_vst3_lane_v: 3080 case NEON::BI__builtin_neon_vst3q_lane_v: 3081 case NEON::BI__builtin_neon_vst4_lane_v: 3082 case NEON::BI__builtin_neon_vst4q_lane_v: 3083 Ops.push_back(Align); 3084 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, ""); 3085 case NEON::BI__builtin_neon_vsubhn_v: { 3086 llvm::VectorType *SrcTy = 3087 llvm::VectorType::getExtendedElementVectorType(VTy); 3088 3089 // %sum = add <4 x i32> %lhs, %rhs 3090 Ops[0] = Builder.CreateBitCast(Ops[0], SrcTy); 3091 Ops[1] = Builder.CreateBitCast(Ops[1], SrcTy); 3092 Ops[0] = Builder.CreateSub(Ops[0], Ops[1], "vsubhn"); 3093 3094 // %high = lshr <4 x i32> %sum, <i32 16, i32 16, i32 16, i32 16> 3095 Constant *ShiftAmt = 3096 ConstantInt::get(SrcTy, SrcTy->getScalarSizeInBits() / 2); 3097 Ops[0] = Builder.CreateLShr(Ops[0], ShiftAmt, "vsubhn"); 3098 3099 // %res = trunc <4 x i32> %high to <4 x i16> 3100 return Builder.CreateTrunc(Ops[0], VTy, "vsubhn"); 3101 } 3102 case NEON::BI__builtin_neon_vtrn_v: 3103 case NEON::BI__builtin_neon_vtrnq_v: { 3104 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty)); 3105 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 3106 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 3107 Value *SV = nullptr; 3108 3109 for (unsigned vi = 0; vi != 2; ++vi) { 3110 SmallVector<Constant*, 16> Indices; 3111 for (unsigned i = 0, e = VTy->getNumElements(); i != e; i += 2) { 3112 Indices.push_back(Builder.getInt32(i+vi)); 3113 Indices.push_back(Builder.getInt32(i+e+vi)); 3114 } 3115 Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ty, Ops[0], vi); 3116 SV = llvm::ConstantVector::get(Indices); 3117 SV = Builder.CreateShuffleVector(Ops[1], Ops[2], SV, "vtrn"); 3118 SV = Builder.CreateStore(SV, Addr); 3119 } 3120 return SV; 3121 } 3122 case NEON::BI__builtin_neon_vtst_v: 3123 case NEON::BI__builtin_neon_vtstq_v: { 3124 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 3125 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 3126 Ops[0] = Builder.CreateAnd(Ops[0], Ops[1]); 3127 Ops[0] = Builder.CreateICmp(ICmpInst::ICMP_NE, Ops[0], 3128 ConstantAggregateZero::get(Ty)); 3129 return Builder.CreateSExt(Ops[0], Ty, "vtst"); 3130 } 3131 case NEON::BI__builtin_neon_vuzp_v: 3132 case NEON::BI__builtin_neon_vuzpq_v: { 3133 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty)); 3134 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 3135 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 3136 Value *SV = nullptr; 3137 3138 for (unsigned vi = 0; vi != 2; ++vi) { 3139 SmallVector<Constant*, 16> Indices; 3140 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) 3141 Indices.push_back(ConstantInt::get(Int32Ty, 2*i+vi)); 3142 3143 Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ty, Ops[0], vi); 3144 SV = llvm::ConstantVector::get(Indices); 3145 SV = Builder.CreateShuffleVector(Ops[1], Ops[2], SV, "vuzp"); 3146 SV = Builder.CreateStore(SV, Addr); 3147 } 3148 return SV; 3149 } 3150 case NEON::BI__builtin_neon_vzip_v: 3151 case NEON::BI__builtin_neon_vzipq_v: { 3152 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty)); 3153 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 3154 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 3155 Value *SV = nullptr; 3156 3157 for (unsigned vi = 0; vi != 2; ++vi) { 3158 SmallVector<Constant*, 16> Indices; 3159 for (unsigned i = 0, e = VTy->getNumElements(); i != e; i += 2) { 3160 Indices.push_back(ConstantInt::get(Int32Ty, (i + vi*e) >> 1)); 3161 Indices.push_back(ConstantInt::get(Int32Ty, ((i + vi*e) >> 1)+e)); 3162 } 3163 Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ty, Ops[0], vi); 3164 SV = llvm::ConstantVector::get(Indices); 3165 SV = Builder.CreateShuffleVector(Ops[1], Ops[2], SV, "vzip"); 3166 SV = Builder.CreateStore(SV, Addr); 3167 } 3168 return SV; 3169 } 3170 } 3171 3172 assert(Int && "Expected valid intrinsic number"); 3173 3174 // Determine the type(s) of this overloaded AArch64 intrinsic. 3175 Function *F = LookupNeonLLVMIntrinsic(Int, Modifier, Ty, E); 3176 3177 Value *Result = EmitNeonCall(F, Ops, NameHint); 3178 llvm::Type *ResultType = ConvertType(E->getType()); 3179 // AArch64 intrinsic one-element vector type cast to 3180 // scalar type expected by the builtin 3181 return Builder.CreateBitCast(Result, ResultType, NameHint); 3182 } 3183 3184 Value *CodeGenFunction::EmitAArch64CompareBuiltinExpr( 3185 Value *Op, llvm::Type *Ty, const CmpInst::Predicate Fp, 3186 const CmpInst::Predicate Ip, const Twine &Name) { 3187 llvm::Type *OTy = Op->getType(); 3188 3189 // FIXME: this is utterly horrific. We should not be looking at previous 3190 // codegen context to find out what needs doing. Unfortunately TableGen 3191 // currently gives us exactly the same calls for vceqz_f32 and vceqz_s32 3192 // (etc). 3193 if (BitCastInst *BI = dyn_cast<BitCastInst>(Op)) 3194 OTy = BI->getOperand(0)->getType(); 3195 3196 Op = Builder.CreateBitCast(Op, OTy); 3197 if (OTy->getScalarType()->isFloatingPointTy()) { 3198 Op = Builder.CreateFCmp(Fp, Op, Constant::getNullValue(OTy)); 3199 } else { 3200 Op = Builder.CreateICmp(Ip, Op, Constant::getNullValue(OTy)); 3201 } 3202 return Builder.CreateSExt(Op, Ty, Name); 3203 } 3204 3205 static Value *packTBLDVectorList(CodeGenFunction &CGF, ArrayRef<Value *> Ops, 3206 Value *ExtOp, Value *IndexOp, 3207 llvm::Type *ResTy, unsigned IntID, 3208 const char *Name) { 3209 SmallVector<Value *, 2> TblOps; 3210 if (ExtOp) 3211 TblOps.push_back(ExtOp); 3212 3213 // Build a vector containing sequential number like (0, 1, 2, ..., 15) 3214 SmallVector<Constant*, 16> Indices; 3215 llvm::VectorType *TblTy = cast<llvm::VectorType>(Ops[0]->getType()); 3216 for (unsigned i = 0, e = TblTy->getNumElements(); i != e; ++i) { 3217 Indices.push_back(ConstantInt::get(CGF.Int32Ty, 2*i)); 3218 Indices.push_back(ConstantInt::get(CGF.Int32Ty, 2*i+1)); 3219 } 3220 Value *SV = llvm::ConstantVector::get(Indices); 3221 3222 int PairPos = 0, End = Ops.size() - 1; 3223 while (PairPos < End) { 3224 TblOps.push_back(CGF.Builder.CreateShuffleVector(Ops[PairPos], 3225 Ops[PairPos+1], SV, Name)); 3226 PairPos += 2; 3227 } 3228 3229 // If there's an odd number of 64-bit lookup table, fill the high 64-bit 3230 // of the 128-bit lookup table with zero. 3231 if (PairPos == End) { 3232 Value *ZeroTbl = ConstantAggregateZero::get(TblTy); 3233 TblOps.push_back(CGF.Builder.CreateShuffleVector(Ops[PairPos], 3234 ZeroTbl, SV, Name)); 3235 } 3236 3237 Function *TblF; 3238 TblOps.push_back(IndexOp); 3239 TblF = CGF.CGM.getIntrinsic(IntID, ResTy); 3240 3241 return CGF.EmitNeonCall(TblF, TblOps, Name); 3242 } 3243 3244 Value *CodeGenFunction::GetValueForARMHint(unsigned BuiltinID) { 3245 unsigned Value; 3246 switch (BuiltinID) { 3247 default: 3248 return nullptr; 3249 case ARM::BI__builtin_arm_nop: 3250 Value = 0; 3251 break; 3252 case ARM::BI__builtin_arm_yield: 3253 case ARM::BI__yield: 3254 Value = 1; 3255 break; 3256 case ARM::BI__builtin_arm_wfe: 3257 case ARM::BI__wfe: 3258 Value = 2; 3259 break; 3260 case ARM::BI__builtin_arm_wfi: 3261 case ARM::BI__wfi: 3262 Value = 3; 3263 break; 3264 case ARM::BI__builtin_arm_sev: 3265 case ARM::BI__sev: 3266 Value = 4; 3267 break; 3268 case ARM::BI__builtin_arm_sevl: 3269 case ARM::BI__sevl: 3270 Value = 5; 3271 break; 3272 } 3273 3274 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::arm_hint), 3275 llvm::ConstantInt::get(Int32Ty, Value)); 3276 } 3277 3278 // Generates the IR for the read/write special register builtin, 3279 // ValueType is the type of the value that is to be written or read, 3280 // RegisterType is the type of the register being written to or read from. 3281 static Value *EmitSpecialRegisterBuiltin(CodeGenFunction &CGF, 3282 const CallExpr *E, 3283 llvm::Type *RegisterType, 3284 llvm::Type *ValueType, bool IsRead) { 3285 // write and register intrinsics only support 32 and 64 bit operations. 3286 assert((RegisterType->isIntegerTy(32) || RegisterType->isIntegerTy(64)) 3287 && "Unsupported size for register."); 3288 3289 CodeGen::CGBuilderTy &Builder = CGF.Builder; 3290 CodeGen::CodeGenModule &CGM = CGF.CGM; 3291 LLVMContext &Context = CGM.getLLVMContext(); 3292 3293 const Expr *SysRegStrExpr = E->getArg(0)->IgnoreParenCasts(); 3294 StringRef SysReg = cast<StringLiteral>(SysRegStrExpr)->getString(); 3295 3296 llvm::Metadata *Ops[] = { llvm::MDString::get(Context, SysReg) }; 3297 llvm::MDNode *RegName = llvm::MDNode::get(Context, Ops); 3298 llvm::Value *Metadata = llvm::MetadataAsValue::get(Context, RegName); 3299 3300 llvm::Type *Types[] = { RegisterType }; 3301 3302 bool MixedTypes = RegisterType->isIntegerTy(64) && ValueType->isIntegerTy(32); 3303 assert(!(RegisterType->isIntegerTy(32) && ValueType->isIntegerTy(64)) 3304 && "Can't fit 64-bit value in 32-bit register"); 3305 3306 if (IsRead) { 3307 llvm::Value *F = CGM.getIntrinsic(llvm::Intrinsic::read_register, Types); 3308 llvm::Value *Call = Builder.CreateCall(F, Metadata); 3309 3310 if (MixedTypes) 3311 // Read into 64 bit register and then truncate result to 32 bit. 3312 return Builder.CreateTrunc(Call, ValueType); 3313 3314 if (ValueType->isPointerTy()) 3315 // Have i32/i64 result (Call) but want to return a VoidPtrTy (i8*). 3316 return Builder.CreateIntToPtr(Call, ValueType); 3317 3318 return Call; 3319 } 3320 3321 llvm::Value *F = CGM.getIntrinsic(llvm::Intrinsic::write_register, Types); 3322 llvm::Value *ArgValue = CGF.EmitScalarExpr(E->getArg(1)); 3323 if (MixedTypes) { 3324 // Extend 32 bit write value to 64 bit to pass to write. 3325 ArgValue = Builder.CreateZExt(ArgValue, RegisterType); 3326 return Builder.CreateCall(F, { Metadata, ArgValue }); 3327 } 3328 3329 if (ValueType->isPointerTy()) { 3330 // Have VoidPtrTy ArgValue but want to return an i32/i64. 3331 ArgValue = Builder.CreatePtrToInt(ArgValue, RegisterType); 3332 return Builder.CreateCall(F, { Metadata, ArgValue }); 3333 } 3334 3335 return Builder.CreateCall(F, { Metadata, ArgValue }); 3336 } 3337 3338 /// Return true if BuiltinID is an overloaded Neon intrinsic with an extra 3339 /// argument that specifies the vector type. 3340 static bool HasExtraNeonArgument(unsigned BuiltinID) { 3341 switch (BuiltinID) { 3342 default: break; 3343 case NEON::BI__builtin_neon_vget_lane_i8: 3344 case NEON::BI__builtin_neon_vget_lane_i16: 3345 case NEON::BI__builtin_neon_vget_lane_i32: 3346 case NEON::BI__builtin_neon_vget_lane_i64: 3347 case NEON::BI__builtin_neon_vget_lane_f32: 3348 case NEON::BI__builtin_neon_vgetq_lane_i8: 3349 case NEON::BI__builtin_neon_vgetq_lane_i16: 3350 case NEON::BI__builtin_neon_vgetq_lane_i32: 3351 case NEON::BI__builtin_neon_vgetq_lane_i64: 3352 case NEON::BI__builtin_neon_vgetq_lane_f32: 3353 case NEON::BI__builtin_neon_vset_lane_i8: 3354 case NEON::BI__builtin_neon_vset_lane_i16: 3355 case NEON::BI__builtin_neon_vset_lane_i32: 3356 case NEON::BI__builtin_neon_vset_lane_i64: 3357 case NEON::BI__builtin_neon_vset_lane_f32: 3358 case NEON::BI__builtin_neon_vsetq_lane_i8: 3359 case NEON::BI__builtin_neon_vsetq_lane_i16: 3360 case NEON::BI__builtin_neon_vsetq_lane_i32: 3361 case NEON::BI__builtin_neon_vsetq_lane_i64: 3362 case NEON::BI__builtin_neon_vsetq_lane_f32: 3363 case NEON::BI__builtin_neon_vsha1h_u32: 3364 case NEON::BI__builtin_neon_vsha1cq_u32: 3365 case NEON::BI__builtin_neon_vsha1pq_u32: 3366 case NEON::BI__builtin_neon_vsha1mq_u32: 3367 case ARM::BI_MoveToCoprocessor: 3368 case ARM::BI_MoveToCoprocessor2: 3369 return false; 3370 } 3371 return true; 3372 } 3373 3374 Value *CodeGenFunction::EmitARMBuiltinExpr(unsigned BuiltinID, 3375 const CallExpr *E) { 3376 if (auto Hint = GetValueForARMHint(BuiltinID)) 3377 return Hint; 3378 3379 if (BuiltinID == ARM::BI__emit) { 3380 bool IsThumb = getTarget().getTriple().getArch() == llvm::Triple::thumb; 3381 llvm::FunctionType *FTy = 3382 llvm::FunctionType::get(VoidTy, /*Variadic=*/false); 3383 3384 APSInt Value; 3385 if (!E->getArg(0)->EvaluateAsInt(Value, CGM.getContext())) 3386 llvm_unreachable("Sema will ensure that the parameter is constant"); 3387 3388 uint64_t ZExtValue = Value.zextOrTrunc(IsThumb ? 16 : 32).getZExtValue(); 3389 3390 llvm::InlineAsm *Emit = 3391 IsThumb ? InlineAsm::get(FTy, ".inst.n 0x" + utohexstr(ZExtValue), "", 3392 /*SideEffects=*/true) 3393 : InlineAsm::get(FTy, ".inst 0x" + utohexstr(ZExtValue), "", 3394 /*SideEffects=*/true); 3395 3396 return Builder.CreateCall(Emit); 3397 } 3398 3399 if (BuiltinID == ARM::BI__builtin_arm_dbg) { 3400 Value *Option = EmitScalarExpr(E->getArg(0)); 3401 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::arm_dbg), Option); 3402 } 3403 3404 if (BuiltinID == ARM::BI__builtin_arm_prefetch) { 3405 Value *Address = EmitScalarExpr(E->getArg(0)); 3406 Value *RW = EmitScalarExpr(E->getArg(1)); 3407 Value *IsData = EmitScalarExpr(E->getArg(2)); 3408 3409 // Locality is not supported on ARM target 3410 Value *Locality = llvm::ConstantInt::get(Int32Ty, 3); 3411 3412 Value *F = CGM.getIntrinsic(Intrinsic::prefetch); 3413 return Builder.CreateCall(F, {Address, RW, Locality, IsData}); 3414 } 3415 3416 if (BuiltinID == ARM::BI__builtin_arm_rbit) { 3417 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::arm_rbit), 3418 EmitScalarExpr(E->getArg(0)), 3419 "rbit"); 3420 } 3421 3422 if (BuiltinID == ARM::BI__clear_cache) { 3423 assert(E->getNumArgs() == 2 && "__clear_cache takes 2 arguments"); 3424 const FunctionDecl *FD = E->getDirectCallee(); 3425 Value *Ops[2]; 3426 for (unsigned i = 0; i < 2; i++) 3427 Ops[i] = EmitScalarExpr(E->getArg(i)); 3428 llvm::Type *Ty = CGM.getTypes().ConvertType(FD->getType()); 3429 llvm::FunctionType *FTy = cast<llvm::FunctionType>(Ty); 3430 StringRef Name = FD->getName(); 3431 return EmitNounwindRuntimeCall(CGM.CreateRuntimeFunction(FTy, Name), Ops); 3432 } 3433 3434 if (BuiltinID == ARM::BI__builtin_arm_ldrexd || 3435 ((BuiltinID == ARM::BI__builtin_arm_ldrex || 3436 BuiltinID == ARM::BI__builtin_arm_ldaex) && 3437 getContext().getTypeSize(E->getType()) == 64) || 3438 BuiltinID == ARM::BI__ldrexd) { 3439 Function *F; 3440 3441 switch (BuiltinID) { 3442 default: llvm_unreachable("unexpected builtin"); 3443 case ARM::BI__builtin_arm_ldaex: 3444 F = CGM.getIntrinsic(Intrinsic::arm_ldaexd); 3445 break; 3446 case ARM::BI__builtin_arm_ldrexd: 3447 case ARM::BI__builtin_arm_ldrex: 3448 case ARM::BI__ldrexd: 3449 F = CGM.getIntrinsic(Intrinsic::arm_ldrexd); 3450 break; 3451 } 3452 3453 Value *LdPtr = EmitScalarExpr(E->getArg(0)); 3454 Value *Val = Builder.CreateCall(F, Builder.CreateBitCast(LdPtr, Int8PtrTy), 3455 "ldrexd"); 3456 3457 Value *Val0 = Builder.CreateExtractValue(Val, 1); 3458 Value *Val1 = Builder.CreateExtractValue(Val, 0); 3459 Val0 = Builder.CreateZExt(Val0, Int64Ty); 3460 Val1 = Builder.CreateZExt(Val1, Int64Ty); 3461 3462 Value *ShiftCst = llvm::ConstantInt::get(Int64Ty, 32); 3463 Val = Builder.CreateShl(Val0, ShiftCst, "shl", true /* nuw */); 3464 Val = Builder.CreateOr(Val, Val1); 3465 return Builder.CreateBitCast(Val, ConvertType(E->getType())); 3466 } 3467 3468 if (BuiltinID == ARM::BI__builtin_arm_ldrex || 3469 BuiltinID == ARM::BI__builtin_arm_ldaex) { 3470 Value *LoadAddr = EmitScalarExpr(E->getArg(0)); 3471 3472 QualType Ty = E->getType(); 3473 llvm::Type *RealResTy = ConvertType(Ty); 3474 llvm::Type *IntResTy = llvm::IntegerType::get(getLLVMContext(), 3475 getContext().getTypeSize(Ty)); 3476 LoadAddr = Builder.CreateBitCast(LoadAddr, IntResTy->getPointerTo()); 3477 3478 Function *F = CGM.getIntrinsic(BuiltinID == ARM::BI__builtin_arm_ldaex 3479 ? Intrinsic::arm_ldaex 3480 : Intrinsic::arm_ldrex, 3481 LoadAddr->getType()); 3482 Value *Val = Builder.CreateCall(F, LoadAddr, "ldrex"); 3483 3484 if (RealResTy->isPointerTy()) 3485 return Builder.CreateIntToPtr(Val, RealResTy); 3486 else { 3487 Val = Builder.CreateTruncOrBitCast(Val, IntResTy); 3488 return Builder.CreateBitCast(Val, RealResTy); 3489 } 3490 } 3491 3492 if (BuiltinID == ARM::BI__builtin_arm_strexd || 3493 ((BuiltinID == ARM::BI__builtin_arm_stlex || 3494 BuiltinID == ARM::BI__builtin_arm_strex) && 3495 getContext().getTypeSize(E->getArg(0)->getType()) == 64)) { 3496 Function *F = CGM.getIntrinsic(BuiltinID == ARM::BI__builtin_arm_stlex 3497 ? Intrinsic::arm_stlexd 3498 : Intrinsic::arm_strexd); 3499 llvm::Type *STy = llvm::StructType::get(Int32Ty, Int32Ty, nullptr); 3500 3501 Value *Tmp = CreateMemTemp(E->getArg(0)->getType()); 3502 Value *Val = EmitScalarExpr(E->getArg(0)); 3503 Builder.CreateStore(Val, Tmp); 3504 3505 Value *LdPtr = Builder.CreateBitCast(Tmp,llvm::PointerType::getUnqual(STy)); 3506 Val = Builder.CreateLoad(LdPtr); 3507 3508 Value *Arg0 = Builder.CreateExtractValue(Val, 0); 3509 Value *Arg1 = Builder.CreateExtractValue(Val, 1); 3510 Value *StPtr = Builder.CreateBitCast(EmitScalarExpr(E->getArg(1)), Int8PtrTy); 3511 return Builder.CreateCall(F, {Arg0, Arg1, StPtr}, "strexd"); 3512 } 3513 3514 if (BuiltinID == ARM::BI__builtin_arm_strex || 3515 BuiltinID == ARM::BI__builtin_arm_stlex) { 3516 Value *StoreVal = EmitScalarExpr(E->getArg(0)); 3517 Value *StoreAddr = EmitScalarExpr(E->getArg(1)); 3518 3519 QualType Ty = E->getArg(0)->getType(); 3520 llvm::Type *StoreTy = llvm::IntegerType::get(getLLVMContext(), 3521 getContext().getTypeSize(Ty)); 3522 StoreAddr = Builder.CreateBitCast(StoreAddr, StoreTy->getPointerTo()); 3523 3524 if (StoreVal->getType()->isPointerTy()) 3525 StoreVal = Builder.CreatePtrToInt(StoreVal, Int32Ty); 3526 else { 3527 StoreVal = Builder.CreateBitCast(StoreVal, StoreTy); 3528 StoreVal = Builder.CreateZExtOrBitCast(StoreVal, Int32Ty); 3529 } 3530 3531 Function *F = CGM.getIntrinsic(BuiltinID == ARM::BI__builtin_arm_stlex 3532 ? Intrinsic::arm_stlex 3533 : Intrinsic::arm_strex, 3534 StoreAddr->getType()); 3535 return Builder.CreateCall(F, {StoreVal, StoreAddr}, "strex"); 3536 } 3537 3538 if (BuiltinID == ARM::BI__builtin_arm_clrex) { 3539 Function *F = CGM.getIntrinsic(Intrinsic::arm_clrex); 3540 return Builder.CreateCall(F); 3541 } 3542 3543 // CRC32 3544 Intrinsic::ID CRCIntrinsicID = Intrinsic::not_intrinsic; 3545 switch (BuiltinID) { 3546 case ARM::BI__builtin_arm_crc32b: 3547 CRCIntrinsicID = Intrinsic::arm_crc32b; break; 3548 case ARM::BI__builtin_arm_crc32cb: 3549 CRCIntrinsicID = Intrinsic::arm_crc32cb; break; 3550 case ARM::BI__builtin_arm_crc32h: 3551 CRCIntrinsicID = Intrinsic::arm_crc32h; break; 3552 case ARM::BI__builtin_arm_crc32ch: 3553 CRCIntrinsicID = Intrinsic::arm_crc32ch; break; 3554 case ARM::BI__builtin_arm_crc32w: 3555 case ARM::BI__builtin_arm_crc32d: 3556 CRCIntrinsicID = Intrinsic::arm_crc32w; break; 3557 case ARM::BI__builtin_arm_crc32cw: 3558 case ARM::BI__builtin_arm_crc32cd: 3559 CRCIntrinsicID = Intrinsic::arm_crc32cw; break; 3560 } 3561 3562 if (CRCIntrinsicID != Intrinsic::not_intrinsic) { 3563 Value *Arg0 = EmitScalarExpr(E->getArg(0)); 3564 Value *Arg1 = EmitScalarExpr(E->getArg(1)); 3565 3566 // crc32{c,}d intrinsics are implemnted as two calls to crc32{c,}w 3567 // intrinsics, hence we need different codegen for these cases. 3568 if (BuiltinID == ARM::BI__builtin_arm_crc32d || 3569 BuiltinID == ARM::BI__builtin_arm_crc32cd) { 3570 Value *C1 = llvm::ConstantInt::get(Int64Ty, 32); 3571 Value *Arg1a = Builder.CreateTruncOrBitCast(Arg1, Int32Ty); 3572 Value *Arg1b = Builder.CreateLShr(Arg1, C1); 3573 Arg1b = Builder.CreateTruncOrBitCast(Arg1b, Int32Ty); 3574 3575 Function *F = CGM.getIntrinsic(CRCIntrinsicID); 3576 Value *Res = Builder.CreateCall(F, {Arg0, Arg1a}); 3577 return Builder.CreateCall(F, {Res, Arg1b}); 3578 } else { 3579 Arg1 = Builder.CreateZExtOrBitCast(Arg1, Int32Ty); 3580 3581 Function *F = CGM.getIntrinsic(CRCIntrinsicID); 3582 return Builder.CreateCall(F, {Arg0, Arg1}); 3583 } 3584 } 3585 3586 if (BuiltinID == ARM::BI__builtin_arm_rsr || 3587 BuiltinID == ARM::BI__builtin_arm_rsr64 || 3588 BuiltinID == ARM::BI__builtin_arm_rsrp || 3589 BuiltinID == ARM::BI__builtin_arm_wsr || 3590 BuiltinID == ARM::BI__builtin_arm_wsr64 || 3591 BuiltinID == ARM::BI__builtin_arm_wsrp) { 3592 3593 bool IsRead = BuiltinID == ARM::BI__builtin_arm_rsr || 3594 BuiltinID == ARM::BI__builtin_arm_rsr64 || 3595 BuiltinID == ARM::BI__builtin_arm_rsrp; 3596 3597 bool IsPointerBuiltin = BuiltinID == ARM::BI__builtin_arm_rsrp || 3598 BuiltinID == ARM::BI__builtin_arm_wsrp; 3599 3600 bool Is64Bit = BuiltinID == ARM::BI__builtin_arm_rsr64 || 3601 BuiltinID == ARM::BI__builtin_arm_wsr64; 3602 3603 llvm::Type *ValueType; 3604 llvm::Type *RegisterType; 3605 if (IsPointerBuiltin) { 3606 ValueType = VoidPtrTy; 3607 RegisterType = Int32Ty; 3608 } else if (Is64Bit) { 3609 ValueType = RegisterType = Int64Ty; 3610 } else { 3611 ValueType = RegisterType = Int32Ty; 3612 } 3613 3614 return EmitSpecialRegisterBuiltin(*this, E, RegisterType, ValueType, IsRead); 3615 } 3616 3617 // Find out if any arguments are required to be integer constant 3618 // expressions. 3619 unsigned ICEArguments = 0; 3620 ASTContext::GetBuiltinTypeError Error; 3621 getContext().GetBuiltinType(BuiltinID, Error, &ICEArguments); 3622 assert(Error == ASTContext::GE_None && "Should not codegen an error"); 3623 3624 SmallVector<Value*, 4> Ops; 3625 llvm::Value *Align = nullptr; 3626 bool HasExtraArg = HasExtraNeonArgument(BuiltinID); 3627 unsigned NumArgs = E->getNumArgs() - (HasExtraArg ? 1 : 0); 3628 for (unsigned i = 0, e = NumArgs; i != e; i++) { 3629 if (i == 0) { 3630 switch (BuiltinID) { 3631 case NEON::BI__builtin_neon_vld1_v: 3632 case NEON::BI__builtin_neon_vld1q_v: 3633 case NEON::BI__builtin_neon_vld1q_lane_v: 3634 case NEON::BI__builtin_neon_vld1_lane_v: 3635 case NEON::BI__builtin_neon_vld1_dup_v: 3636 case NEON::BI__builtin_neon_vld1q_dup_v: 3637 case NEON::BI__builtin_neon_vst1_v: 3638 case NEON::BI__builtin_neon_vst1q_v: 3639 case NEON::BI__builtin_neon_vst1q_lane_v: 3640 case NEON::BI__builtin_neon_vst1_lane_v: 3641 case NEON::BI__builtin_neon_vst2_v: 3642 case NEON::BI__builtin_neon_vst2q_v: 3643 case NEON::BI__builtin_neon_vst2_lane_v: 3644 case NEON::BI__builtin_neon_vst2q_lane_v: 3645 case NEON::BI__builtin_neon_vst3_v: 3646 case NEON::BI__builtin_neon_vst3q_v: 3647 case NEON::BI__builtin_neon_vst3_lane_v: 3648 case NEON::BI__builtin_neon_vst3q_lane_v: 3649 case NEON::BI__builtin_neon_vst4_v: 3650 case NEON::BI__builtin_neon_vst4q_v: 3651 case NEON::BI__builtin_neon_vst4_lane_v: 3652 case NEON::BI__builtin_neon_vst4q_lane_v: 3653 // Get the alignment for the argument in addition to the value; 3654 // we'll use it later. 3655 std::pair<llvm::Value*, unsigned> Src = 3656 EmitPointerWithAlignment(E->getArg(0)); 3657 Ops.push_back(Src.first); 3658 Align = Builder.getInt32(Src.second); 3659 continue; 3660 } 3661 } 3662 if (i == 1) { 3663 switch (BuiltinID) { 3664 case NEON::BI__builtin_neon_vld2_v: 3665 case NEON::BI__builtin_neon_vld2q_v: 3666 case NEON::BI__builtin_neon_vld3_v: 3667 case NEON::BI__builtin_neon_vld3q_v: 3668 case NEON::BI__builtin_neon_vld4_v: 3669 case NEON::BI__builtin_neon_vld4q_v: 3670 case NEON::BI__builtin_neon_vld2_lane_v: 3671 case NEON::BI__builtin_neon_vld2q_lane_v: 3672 case NEON::BI__builtin_neon_vld3_lane_v: 3673 case NEON::BI__builtin_neon_vld3q_lane_v: 3674 case NEON::BI__builtin_neon_vld4_lane_v: 3675 case NEON::BI__builtin_neon_vld4q_lane_v: 3676 case NEON::BI__builtin_neon_vld2_dup_v: 3677 case NEON::BI__builtin_neon_vld3_dup_v: 3678 case NEON::BI__builtin_neon_vld4_dup_v: 3679 // Get the alignment for the argument in addition to the value; 3680 // we'll use it later. 3681 std::pair<llvm::Value*, unsigned> Src = 3682 EmitPointerWithAlignment(E->getArg(1)); 3683 Ops.push_back(Src.first); 3684 Align = Builder.getInt32(Src.second); 3685 continue; 3686 } 3687 } 3688 3689 if ((ICEArguments & (1 << i)) == 0) { 3690 Ops.push_back(EmitScalarExpr(E->getArg(i))); 3691 } else { 3692 // If this is required to be a constant, constant fold it so that we know 3693 // that the generated intrinsic gets a ConstantInt. 3694 llvm::APSInt Result; 3695 bool IsConst = E->getArg(i)->isIntegerConstantExpr(Result, getContext()); 3696 assert(IsConst && "Constant arg isn't actually constant?"); (void)IsConst; 3697 Ops.push_back(llvm::ConstantInt::get(getLLVMContext(), Result)); 3698 } 3699 } 3700 3701 switch (BuiltinID) { 3702 default: break; 3703 3704 case NEON::BI__builtin_neon_vget_lane_i8: 3705 case NEON::BI__builtin_neon_vget_lane_i16: 3706 case NEON::BI__builtin_neon_vget_lane_i32: 3707 case NEON::BI__builtin_neon_vget_lane_i64: 3708 case NEON::BI__builtin_neon_vget_lane_f32: 3709 case NEON::BI__builtin_neon_vgetq_lane_i8: 3710 case NEON::BI__builtin_neon_vgetq_lane_i16: 3711 case NEON::BI__builtin_neon_vgetq_lane_i32: 3712 case NEON::BI__builtin_neon_vgetq_lane_i64: 3713 case NEON::BI__builtin_neon_vgetq_lane_f32: 3714 return Builder.CreateExtractElement(Ops[0], Ops[1], "vget_lane"); 3715 3716 case NEON::BI__builtin_neon_vset_lane_i8: 3717 case NEON::BI__builtin_neon_vset_lane_i16: 3718 case NEON::BI__builtin_neon_vset_lane_i32: 3719 case NEON::BI__builtin_neon_vset_lane_i64: 3720 case NEON::BI__builtin_neon_vset_lane_f32: 3721 case NEON::BI__builtin_neon_vsetq_lane_i8: 3722 case NEON::BI__builtin_neon_vsetq_lane_i16: 3723 case NEON::BI__builtin_neon_vsetq_lane_i32: 3724 case NEON::BI__builtin_neon_vsetq_lane_i64: 3725 case NEON::BI__builtin_neon_vsetq_lane_f32: 3726 return Builder.CreateInsertElement(Ops[1], Ops[0], Ops[2], "vset_lane"); 3727 3728 case NEON::BI__builtin_neon_vsha1h_u32: 3729 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_sha1h), Ops, 3730 "vsha1h"); 3731 case NEON::BI__builtin_neon_vsha1cq_u32: 3732 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_sha1c), Ops, 3733 "vsha1h"); 3734 case NEON::BI__builtin_neon_vsha1pq_u32: 3735 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_sha1p), Ops, 3736 "vsha1h"); 3737 case NEON::BI__builtin_neon_vsha1mq_u32: 3738 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_sha1m), Ops, 3739 "vsha1h"); 3740 3741 // The ARM _MoveToCoprocessor builtins put the input register value as 3742 // the first argument, but the LLVM intrinsic expects it as the third one. 3743 case ARM::BI_MoveToCoprocessor: 3744 case ARM::BI_MoveToCoprocessor2: { 3745 Function *F = CGM.getIntrinsic(BuiltinID == ARM::BI_MoveToCoprocessor ? 3746 Intrinsic::arm_mcr : Intrinsic::arm_mcr2); 3747 return Builder.CreateCall(F, {Ops[1], Ops[2], Ops[0], 3748 Ops[3], Ops[4], Ops[5]}); 3749 } 3750 } 3751 3752 // Get the last argument, which specifies the vector type. 3753 assert(HasExtraArg); 3754 llvm::APSInt Result; 3755 const Expr *Arg = E->getArg(E->getNumArgs()-1); 3756 if (!Arg->isIntegerConstantExpr(Result, getContext())) 3757 return nullptr; 3758 3759 if (BuiltinID == ARM::BI__builtin_arm_vcvtr_f || 3760 BuiltinID == ARM::BI__builtin_arm_vcvtr_d) { 3761 // Determine the overloaded type of this builtin. 3762 llvm::Type *Ty; 3763 if (BuiltinID == ARM::BI__builtin_arm_vcvtr_f) 3764 Ty = FloatTy; 3765 else 3766 Ty = DoubleTy; 3767 3768 // Determine whether this is an unsigned conversion or not. 3769 bool usgn = Result.getZExtValue() == 1; 3770 unsigned Int = usgn ? Intrinsic::arm_vcvtru : Intrinsic::arm_vcvtr; 3771 3772 // Call the appropriate intrinsic. 3773 Function *F = CGM.getIntrinsic(Int, Ty); 3774 return Builder.CreateCall(F, Ops, "vcvtr"); 3775 } 3776 3777 // Determine the type of this overloaded NEON intrinsic. 3778 NeonTypeFlags Type(Result.getZExtValue()); 3779 bool usgn = Type.isUnsigned(); 3780 bool rightShift = false; 3781 3782 llvm::VectorType *VTy = GetNeonType(this, Type); 3783 llvm::Type *Ty = VTy; 3784 if (!Ty) 3785 return nullptr; 3786 3787 // Many NEON builtins have identical semantics and uses in ARM and 3788 // AArch64. Emit these in a single function. 3789 auto IntrinsicMap = makeArrayRef(ARMSIMDIntrinsicMap); 3790 const NeonIntrinsicInfo *Builtin = findNeonIntrinsicInMap( 3791 IntrinsicMap, BuiltinID, NEONSIMDIntrinsicsProvenSorted); 3792 if (Builtin) 3793 return EmitCommonNeonBuiltinExpr( 3794 Builtin->BuiltinID, Builtin->LLVMIntrinsic, Builtin->AltLLVMIntrinsic, 3795 Builtin->NameHint, Builtin->TypeModifier, E, Ops, Align); 3796 3797 unsigned Int; 3798 switch (BuiltinID) { 3799 default: return nullptr; 3800 case NEON::BI__builtin_neon_vld1q_lane_v: 3801 // Handle 64-bit integer elements as a special case. Use shuffles of 3802 // one-element vectors to avoid poor code for i64 in the backend. 3803 if (VTy->getElementType()->isIntegerTy(64)) { 3804 // Extract the other lane. 3805 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 3806 uint32_t Lane = cast<ConstantInt>(Ops[2])->getZExtValue(); 3807 Value *SV = llvm::ConstantVector::get(ConstantInt::get(Int32Ty, 1-Lane)); 3808 Ops[1] = Builder.CreateShuffleVector(Ops[1], Ops[1], SV); 3809 // Load the value as a one-element vector. 3810 Ty = llvm::VectorType::get(VTy->getElementType(), 1); 3811 Function *F = CGM.getIntrinsic(Intrinsic::arm_neon_vld1, Ty); 3812 Value *Ld = Builder.CreateCall(F, {Ops[0], Align}); 3813 // Combine them. 3814 uint32_t Indices[] = {1 - Lane, Lane}; 3815 SV = llvm::ConstantDataVector::get(getLLVMContext(), Indices); 3816 return Builder.CreateShuffleVector(Ops[1], Ld, SV, "vld1q_lane"); 3817 } 3818 // fall through 3819 case NEON::BI__builtin_neon_vld1_lane_v: { 3820 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 3821 Ty = llvm::PointerType::getUnqual(VTy->getElementType()); 3822 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 3823 LoadInst *Ld = Builder.CreateLoad(Ops[0]); 3824 Ld->setAlignment(cast<ConstantInt>(Align)->getZExtValue()); 3825 return Builder.CreateInsertElement(Ops[1], Ld, Ops[2], "vld1_lane"); 3826 } 3827 case NEON::BI__builtin_neon_vld2_dup_v: 3828 case NEON::BI__builtin_neon_vld3_dup_v: 3829 case NEON::BI__builtin_neon_vld4_dup_v: { 3830 // Handle 64-bit elements as a special-case. There is no "dup" needed. 3831 if (VTy->getElementType()->getPrimitiveSizeInBits() == 64) { 3832 switch (BuiltinID) { 3833 case NEON::BI__builtin_neon_vld2_dup_v: 3834 Int = Intrinsic::arm_neon_vld2; 3835 break; 3836 case NEON::BI__builtin_neon_vld3_dup_v: 3837 Int = Intrinsic::arm_neon_vld3; 3838 break; 3839 case NEON::BI__builtin_neon_vld4_dup_v: 3840 Int = Intrinsic::arm_neon_vld4; 3841 break; 3842 default: llvm_unreachable("unknown vld_dup intrinsic?"); 3843 } 3844 Function *F = CGM.getIntrinsic(Int, Ty); 3845 Ops[1] = Builder.CreateCall(F, {Ops[1], Align}, "vld_dup"); 3846 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 3847 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 3848 return Builder.CreateStore(Ops[1], Ops[0]); 3849 } 3850 switch (BuiltinID) { 3851 case NEON::BI__builtin_neon_vld2_dup_v: 3852 Int = Intrinsic::arm_neon_vld2lane; 3853 break; 3854 case NEON::BI__builtin_neon_vld3_dup_v: 3855 Int = Intrinsic::arm_neon_vld3lane; 3856 break; 3857 case NEON::BI__builtin_neon_vld4_dup_v: 3858 Int = Intrinsic::arm_neon_vld4lane; 3859 break; 3860 default: llvm_unreachable("unknown vld_dup intrinsic?"); 3861 } 3862 Function *F = CGM.getIntrinsic(Int, Ty); 3863 llvm::StructType *STy = cast<llvm::StructType>(F->getReturnType()); 3864 3865 SmallVector<Value*, 6> Args; 3866 Args.push_back(Ops[1]); 3867 Args.append(STy->getNumElements(), UndefValue::get(Ty)); 3868 3869 llvm::Constant *CI = ConstantInt::get(Int32Ty, 0); 3870 Args.push_back(CI); 3871 Args.push_back(Align); 3872 3873 Ops[1] = Builder.CreateCall(F, Args, "vld_dup"); 3874 // splat lane 0 to all elts in each vector of the result. 3875 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 3876 Value *Val = Builder.CreateExtractValue(Ops[1], i); 3877 Value *Elt = Builder.CreateBitCast(Val, Ty); 3878 Elt = EmitNeonSplat(Elt, CI); 3879 Elt = Builder.CreateBitCast(Elt, Val->getType()); 3880 Ops[1] = Builder.CreateInsertValue(Ops[1], Elt, i); 3881 } 3882 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 3883 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 3884 return Builder.CreateStore(Ops[1], Ops[0]); 3885 } 3886 case NEON::BI__builtin_neon_vqrshrn_n_v: 3887 Int = 3888 usgn ? Intrinsic::arm_neon_vqrshiftnu : Intrinsic::arm_neon_vqrshiftns; 3889 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqrshrn_n", 3890 1, true); 3891 case NEON::BI__builtin_neon_vqrshrun_n_v: 3892 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vqrshiftnsu, Ty), 3893 Ops, "vqrshrun_n", 1, true); 3894 case NEON::BI__builtin_neon_vqshrn_n_v: 3895 Int = usgn ? Intrinsic::arm_neon_vqshiftnu : Intrinsic::arm_neon_vqshiftns; 3896 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqshrn_n", 3897 1, true); 3898 case NEON::BI__builtin_neon_vqshrun_n_v: 3899 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vqshiftnsu, Ty), 3900 Ops, "vqshrun_n", 1, true); 3901 case NEON::BI__builtin_neon_vrecpe_v: 3902 case NEON::BI__builtin_neon_vrecpeq_v: 3903 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vrecpe, Ty), 3904 Ops, "vrecpe"); 3905 case NEON::BI__builtin_neon_vrshrn_n_v: 3906 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vrshiftn, Ty), 3907 Ops, "vrshrn_n", 1, true); 3908 case NEON::BI__builtin_neon_vrsra_n_v: 3909 case NEON::BI__builtin_neon_vrsraq_n_v: 3910 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 3911 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 3912 Ops[2] = EmitNeonShiftVector(Ops[2], Ty, true); 3913 Int = usgn ? Intrinsic::arm_neon_vrshiftu : Intrinsic::arm_neon_vrshifts; 3914 Ops[1] = Builder.CreateCall(CGM.getIntrinsic(Int, Ty), {Ops[1], Ops[2]}); 3915 return Builder.CreateAdd(Ops[0], Ops[1], "vrsra_n"); 3916 case NEON::BI__builtin_neon_vsri_n_v: 3917 case NEON::BI__builtin_neon_vsriq_n_v: 3918 rightShift = true; 3919 case NEON::BI__builtin_neon_vsli_n_v: 3920 case NEON::BI__builtin_neon_vsliq_n_v: 3921 Ops[2] = EmitNeonShiftVector(Ops[2], Ty, rightShift); 3922 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vshiftins, Ty), 3923 Ops, "vsli_n"); 3924 case NEON::BI__builtin_neon_vsra_n_v: 3925 case NEON::BI__builtin_neon_vsraq_n_v: 3926 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 3927 Ops[1] = EmitNeonRShiftImm(Ops[1], Ops[2], Ty, usgn, "vsra_n"); 3928 return Builder.CreateAdd(Ops[0], Ops[1]); 3929 case NEON::BI__builtin_neon_vst1q_lane_v: 3930 // Handle 64-bit integer elements as a special case. Use a shuffle to get 3931 // a one-element vector and avoid poor code for i64 in the backend. 3932 if (VTy->getElementType()->isIntegerTy(64)) { 3933 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 3934 Value *SV = llvm::ConstantVector::get(cast<llvm::Constant>(Ops[2])); 3935 Ops[1] = Builder.CreateShuffleVector(Ops[1], Ops[1], SV); 3936 Ops[2] = Align; 3937 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::arm_neon_vst1, 3938 Ops[1]->getType()), Ops); 3939 } 3940 // fall through 3941 case NEON::BI__builtin_neon_vst1_lane_v: { 3942 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 3943 Ops[1] = Builder.CreateExtractElement(Ops[1], Ops[2]); 3944 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 3945 StoreInst *St = Builder.CreateStore(Ops[1], 3946 Builder.CreateBitCast(Ops[0], Ty)); 3947 St->setAlignment(cast<ConstantInt>(Align)->getZExtValue()); 3948 return St; 3949 } 3950 case NEON::BI__builtin_neon_vtbl1_v: 3951 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbl1), 3952 Ops, "vtbl1"); 3953 case NEON::BI__builtin_neon_vtbl2_v: 3954 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbl2), 3955 Ops, "vtbl2"); 3956 case NEON::BI__builtin_neon_vtbl3_v: 3957 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbl3), 3958 Ops, "vtbl3"); 3959 case NEON::BI__builtin_neon_vtbl4_v: 3960 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbl4), 3961 Ops, "vtbl4"); 3962 case NEON::BI__builtin_neon_vtbx1_v: 3963 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbx1), 3964 Ops, "vtbx1"); 3965 case NEON::BI__builtin_neon_vtbx2_v: 3966 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbx2), 3967 Ops, "vtbx2"); 3968 case NEON::BI__builtin_neon_vtbx3_v: 3969 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbx3), 3970 Ops, "vtbx3"); 3971 case NEON::BI__builtin_neon_vtbx4_v: 3972 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbx4), 3973 Ops, "vtbx4"); 3974 } 3975 } 3976 3977 static Value *EmitAArch64TblBuiltinExpr(CodeGenFunction &CGF, unsigned BuiltinID, 3978 const CallExpr *E, 3979 SmallVectorImpl<Value *> &Ops) { 3980 unsigned int Int = 0; 3981 const char *s = nullptr; 3982 3983 switch (BuiltinID) { 3984 default: 3985 return nullptr; 3986 case NEON::BI__builtin_neon_vtbl1_v: 3987 case NEON::BI__builtin_neon_vqtbl1_v: 3988 case NEON::BI__builtin_neon_vqtbl1q_v: 3989 case NEON::BI__builtin_neon_vtbl2_v: 3990 case NEON::BI__builtin_neon_vqtbl2_v: 3991 case NEON::BI__builtin_neon_vqtbl2q_v: 3992 case NEON::BI__builtin_neon_vtbl3_v: 3993 case NEON::BI__builtin_neon_vqtbl3_v: 3994 case NEON::BI__builtin_neon_vqtbl3q_v: 3995 case NEON::BI__builtin_neon_vtbl4_v: 3996 case NEON::BI__builtin_neon_vqtbl4_v: 3997 case NEON::BI__builtin_neon_vqtbl4q_v: 3998 break; 3999 case NEON::BI__builtin_neon_vtbx1_v: 4000 case NEON::BI__builtin_neon_vqtbx1_v: 4001 case NEON::BI__builtin_neon_vqtbx1q_v: 4002 case NEON::BI__builtin_neon_vtbx2_v: 4003 case NEON::BI__builtin_neon_vqtbx2_v: 4004 case NEON::BI__builtin_neon_vqtbx2q_v: 4005 case NEON::BI__builtin_neon_vtbx3_v: 4006 case NEON::BI__builtin_neon_vqtbx3_v: 4007 case NEON::BI__builtin_neon_vqtbx3q_v: 4008 case NEON::BI__builtin_neon_vtbx4_v: 4009 case NEON::BI__builtin_neon_vqtbx4_v: 4010 case NEON::BI__builtin_neon_vqtbx4q_v: 4011 break; 4012 } 4013 4014 assert(E->getNumArgs() >= 3); 4015 4016 // Get the last argument, which specifies the vector type. 4017 llvm::APSInt Result; 4018 const Expr *Arg = E->getArg(E->getNumArgs() - 1); 4019 if (!Arg->isIntegerConstantExpr(Result, CGF.getContext())) 4020 return nullptr; 4021 4022 // Determine the type of this overloaded NEON intrinsic. 4023 NeonTypeFlags Type(Result.getZExtValue()); 4024 llvm::VectorType *Ty = GetNeonType(&CGF, Type); 4025 if (!Ty) 4026 return nullptr; 4027 4028 CodeGen::CGBuilderTy &Builder = CGF.Builder; 4029 4030 // AArch64 scalar builtins are not overloaded, they do not have an extra 4031 // argument that specifies the vector type, need to handle each case. 4032 switch (BuiltinID) { 4033 case NEON::BI__builtin_neon_vtbl1_v: { 4034 return packTBLDVectorList(CGF, makeArrayRef(Ops).slice(0, 1), nullptr, 4035 Ops[1], Ty, Intrinsic::aarch64_neon_tbl1, 4036 "vtbl1"); 4037 } 4038 case NEON::BI__builtin_neon_vtbl2_v: { 4039 return packTBLDVectorList(CGF, makeArrayRef(Ops).slice(0, 2), nullptr, 4040 Ops[2], Ty, Intrinsic::aarch64_neon_tbl1, 4041 "vtbl1"); 4042 } 4043 case NEON::BI__builtin_neon_vtbl3_v: { 4044 return packTBLDVectorList(CGF, makeArrayRef(Ops).slice(0, 3), nullptr, 4045 Ops[3], Ty, Intrinsic::aarch64_neon_tbl2, 4046 "vtbl2"); 4047 } 4048 case NEON::BI__builtin_neon_vtbl4_v: { 4049 return packTBLDVectorList(CGF, makeArrayRef(Ops).slice(0, 4), nullptr, 4050 Ops[4], Ty, Intrinsic::aarch64_neon_tbl2, 4051 "vtbl2"); 4052 } 4053 case NEON::BI__builtin_neon_vtbx1_v: { 4054 Value *TblRes = 4055 packTBLDVectorList(CGF, makeArrayRef(Ops).slice(1, 1), nullptr, Ops[2], 4056 Ty, Intrinsic::aarch64_neon_tbl1, "vtbl1"); 4057 4058 llvm::Constant *EightV = ConstantInt::get(Ty, 8); 4059 Value *CmpRes = Builder.CreateICmp(ICmpInst::ICMP_UGE, Ops[2], EightV); 4060 CmpRes = Builder.CreateSExt(CmpRes, Ty); 4061 4062 Value *EltsFromInput = Builder.CreateAnd(CmpRes, Ops[0]); 4063 Value *EltsFromTbl = Builder.CreateAnd(Builder.CreateNot(CmpRes), TblRes); 4064 return Builder.CreateOr(EltsFromInput, EltsFromTbl, "vtbx"); 4065 } 4066 case NEON::BI__builtin_neon_vtbx2_v: { 4067 return packTBLDVectorList(CGF, makeArrayRef(Ops).slice(1, 2), Ops[0], 4068 Ops[3], Ty, Intrinsic::aarch64_neon_tbx1, 4069 "vtbx1"); 4070 } 4071 case NEON::BI__builtin_neon_vtbx3_v: { 4072 Value *TblRes = 4073 packTBLDVectorList(CGF, makeArrayRef(Ops).slice(1, 3), nullptr, Ops[4], 4074 Ty, Intrinsic::aarch64_neon_tbl2, "vtbl2"); 4075 4076 llvm::Constant *TwentyFourV = ConstantInt::get(Ty, 24); 4077 Value *CmpRes = Builder.CreateICmp(ICmpInst::ICMP_UGE, Ops[4], 4078 TwentyFourV); 4079 CmpRes = Builder.CreateSExt(CmpRes, Ty); 4080 4081 Value *EltsFromInput = Builder.CreateAnd(CmpRes, Ops[0]); 4082 Value *EltsFromTbl = Builder.CreateAnd(Builder.CreateNot(CmpRes), TblRes); 4083 return Builder.CreateOr(EltsFromInput, EltsFromTbl, "vtbx"); 4084 } 4085 case NEON::BI__builtin_neon_vtbx4_v: { 4086 return packTBLDVectorList(CGF, makeArrayRef(Ops).slice(1, 4), Ops[0], 4087 Ops[5], Ty, Intrinsic::aarch64_neon_tbx2, 4088 "vtbx2"); 4089 } 4090 case NEON::BI__builtin_neon_vqtbl1_v: 4091 case NEON::BI__builtin_neon_vqtbl1q_v: 4092 Int = Intrinsic::aarch64_neon_tbl1; s = "vtbl1"; break; 4093 case NEON::BI__builtin_neon_vqtbl2_v: 4094 case NEON::BI__builtin_neon_vqtbl2q_v: { 4095 Int = Intrinsic::aarch64_neon_tbl2; s = "vtbl2"; break; 4096 case NEON::BI__builtin_neon_vqtbl3_v: 4097 case NEON::BI__builtin_neon_vqtbl3q_v: 4098 Int = Intrinsic::aarch64_neon_tbl3; s = "vtbl3"; break; 4099 case NEON::BI__builtin_neon_vqtbl4_v: 4100 case NEON::BI__builtin_neon_vqtbl4q_v: 4101 Int = Intrinsic::aarch64_neon_tbl4; s = "vtbl4"; break; 4102 case NEON::BI__builtin_neon_vqtbx1_v: 4103 case NEON::BI__builtin_neon_vqtbx1q_v: 4104 Int = Intrinsic::aarch64_neon_tbx1; s = "vtbx1"; break; 4105 case NEON::BI__builtin_neon_vqtbx2_v: 4106 case NEON::BI__builtin_neon_vqtbx2q_v: 4107 Int = Intrinsic::aarch64_neon_tbx2; s = "vtbx2"; break; 4108 case NEON::BI__builtin_neon_vqtbx3_v: 4109 case NEON::BI__builtin_neon_vqtbx3q_v: 4110 Int = Intrinsic::aarch64_neon_tbx3; s = "vtbx3"; break; 4111 case NEON::BI__builtin_neon_vqtbx4_v: 4112 case NEON::BI__builtin_neon_vqtbx4q_v: 4113 Int = Intrinsic::aarch64_neon_tbx4; s = "vtbx4"; break; 4114 } 4115 } 4116 4117 if (!Int) 4118 return nullptr; 4119 4120 Function *F = CGF.CGM.getIntrinsic(Int, Ty); 4121 return CGF.EmitNeonCall(F, Ops, s); 4122 } 4123 4124 Value *CodeGenFunction::vectorWrapScalar16(Value *Op) { 4125 llvm::Type *VTy = llvm::VectorType::get(Int16Ty, 4); 4126 Op = Builder.CreateBitCast(Op, Int16Ty); 4127 Value *V = UndefValue::get(VTy); 4128 llvm::Constant *CI = ConstantInt::get(SizeTy, 0); 4129 Op = Builder.CreateInsertElement(V, Op, CI); 4130 return Op; 4131 } 4132 4133 Value *CodeGenFunction::vectorWrapScalar8(Value *Op) { 4134 llvm::Type *VTy = llvm::VectorType::get(Int8Ty, 8); 4135 Op = Builder.CreateBitCast(Op, Int8Ty); 4136 Value *V = UndefValue::get(VTy); 4137 llvm::Constant *CI = ConstantInt::get(SizeTy, 0); 4138 Op = Builder.CreateInsertElement(V, Op, CI); 4139 return Op; 4140 } 4141 4142 Value *CodeGenFunction::EmitAArch64BuiltinExpr(unsigned BuiltinID, 4143 const CallExpr *E) { 4144 unsigned HintID = static_cast<unsigned>(-1); 4145 switch (BuiltinID) { 4146 default: break; 4147 case AArch64::BI__builtin_arm_nop: 4148 HintID = 0; 4149 break; 4150 case AArch64::BI__builtin_arm_yield: 4151 HintID = 1; 4152 break; 4153 case AArch64::BI__builtin_arm_wfe: 4154 HintID = 2; 4155 break; 4156 case AArch64::BI__builtin_arm_wfi: 4157 HintID = 3; 4158 break; 4159 case AArch64::BI__builtin_arm_sev: 4160 HintID = 4; 4161 break; 4162 case AArch64::BI__builtin_arm_sevl: 4163 HintID = 5; 4164 break; 4165 } 4166 4167 if (HintID != static_cast<unsigned>(-1)) { 4168 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_hint); 4169 return Builder.CreateCall(F, llvm::ConstantInt::get(Int32Ty, HintID)); 4170 } 4171 4172 if (BuiltinID == AArch64::BI__builtin_arm_prefetch) { 4173 Value *Address = EmitScalarExpr(E->getArg(0)); 4174 Value *RW = EmitScalarExpr(E->getArg(1)); 4175 Value *CacheLevel = EmitScalarExpr(E->getArg(2)); 4176 Value *RetentionPolicy = EmitScalarExpr(E->getArg(3)); 4177 Value *IsData = EmitScalarExpr(E->getArg(4)); 4178 4179 Value *Locality = nullptr; 4180 if (cast<llvm::ConstantInt>(RetentionPolicy)->isZero()) { 4181 // Temporal fetch, needs to convert cache level to locality. 4182 Locality = llvm::ConstantInt::get(Int32Ty, 4183 -cast<llvm::ConstantInt>(CacheLevel)->getValue() + 3); 4184 } else { 4185 // Streaming fetch. 4186 Locality = llvm::ConstantInt::get(Int32Ty, 0); 4187 } 4188 4189 // FIXME: We need AArch64 specific LLVM intrinsic if we want to specify 4190 // PLDL3STRM or PLDL2STRM. 4191 Value *F = CGM.getIntrinsic(Intrinsic::prefetch); 4192 return Builder.CreateCall(F, {Address, RW, Locality, IsData}); 4193 } 4194 4195 if (BuiltinID == AArch64::BI__builtin_arm_rbit) { 4196 assert((getContext().getTypeSize(E->getType()) == 32) && 4197 "rbit of unusual size!"); 4198 llvm::Value *Arg = EmitScalarExpr(E->getArg(0)); 4199 return Builder.CreateCall( 4200 CGM.getIntrinsic(Intrinsic::aarch64_rbit, Arg->getType()), Arg, "rbit"); 4201 } 4202 if (BuiltinID == AArch64::BI__builtin_arm_rbit64) { 4203 assert((getContext().getTypeSize(E->getType()) == 64) && 4204 "rbit of unusual size!"); 4205 llvm::Value *Arg = EmitScalarExpr(E->getArg(0)); 4206 return Builder.CreateCall( 4207 CGM.getIntrinsic(Intrinsic::aarch64_rbit, Arg->getType()), Arg, "rbit"); 4208 } 4209 4210 if (BuiltinID == AArch64::BI__clear_cache) { 4211 assert(E->getNumArgs() == 2 && "__clear_cache takes 2 arguments"); 4212 const FunctionDecl *FD = E->getDirectCallee(); 4213 Value *Ops[2]; 4214 for (unsigned i = 0; i < 2; i++) 4215 Ops[i] = EmitScalarExpr(E->getArg(i)); 4216 llvm::Type *Ty = CGM.getTypes().ConvertType(FD->getType()); 4217 llvm::FunctionType *FTy = cast<llvm::FunctionType>(Ty); 4218 StringRef Name = FD->getName(); 4219 return EmitNounwindRuntimeCall(CGM.CreateRuntimeFunction(FTy, Name), Ops); 4220 } 4221 4222 if ((BuiltinID == AArch64::BI__builtin_arm_ldrex || 4223 BuiltinID == AArch64::BI__builtin_arm_ldaex) && 4224 getContext().getTypeSize(E->getType()) == 128) { 4225 Function *F = CGM.getIntrinsic(BuiltinID == AArch64::BI__builtin_arm_ldaex 4226 ? Intrinsic::aarch64_ldaxp 4227 : Intrinsic::aarch64_ldxp); 4228 4229 Value *LdPtr = EmitScalarExpr(E->getArg(0)); 4230 Value *Val = Builder.CreateCall(F, Builder.CreateBitCast(LdPtr, Int8PtrTy), 4231 "ldxp"); 4232 4233 Value *Val0 = Builder.CreateExtractValue(Val, 1); 4234 Value *Val1 = Builder.CreateExtractValue(Val, 0); 4235 llvm::Type *Int128Ty = llvm::IntegerType::get(getLLVMContext(), 128); 4236 Val0 = Builder.CreateZExt(Val0, Int128Ty); 4237 Val1 = Builder.CreateZExt(Val1, Int128Ty); 4238 4239 Value *ShiftCst = llvm::ConstantInt::get(Int128Ty, 64); 4240 Val = Builder.CreateShl(Val0, ShiftCst, "shl", true /* nuw */); 4241 Val = Builder.CreateOr(Val, Val1); 4242 return Builder.CreateBitCast(Val, ConvertType(E->getType())); 4243 } else if (BuiltinID == AArch64::BI__builtin_arm_ldrex || 4244 BuiltinID == AArch64::BI__builtin_arm_ldaex) { 4245 Value *LoadAddr = EmitScalarExpr(E->getArg(0)); 4246 4247 QualType Ty = E->getType(); 4248 llvm::Type *RealResTy = ConvertType(Ty); 4249 llvm::Type *IntResTy = llvm::IntegerType::get(getLLVMContext(), 4250 getContext().getTypeSize(Ty)); 4251 LoadAddr = Builder.CreateBitCast(LoadAddr, IntResTy->getPointerTo()); 4252 4253 Function *F = CGM.getIntrinsic(BuiltinID == AArch64::BI__builtin_arm_ldaex 4254 ? Intrinsic::aarch64_ldaxr 4255 : Intrinsic::aarch64_ldxr, 4256 LoadAddr->getType()); 4257 Value *Val = Builder.CreateCall(F, LoadAddr, "ldxr"); 4258 4259 if (RealResTy->isPointerTy()) 4260 return Builder.CreateIntToPtr(Val, RealResTy); 4261 4262 Val = Builder.CreateTruncOrBitCast(Val, IntResTy); 4263 return Builder.CreateBitCast(Val, RealResTy); 4264 } 4265 4266 if ((BuiltinID == AArch64::BI__builtin_arm_strex || 4267 BuiltinID == AArch64::BI__builtin_arm_stlex) && 4268 getContext().getTypeSize(E->getArg(0)->getType()) == 128) { 4269 Function *F = CGM.getIntrinsic(BuiltinID == AArch64::BI__builtin_arm_stlex 4270 ? Intrinsic::aarch64_stlxp 4271 : Intrinsic::aarch64_stxp); 4272 llvm::Type *STy = llvm::StructType::get(Int64Ty, Int64Ty, nullptr); 4273 4274 Value *One = llvm::ConstantInt::get(Int32Ty, 1); 4275 Value *Tmp = Builder.CreateAlloca(ConvertType(E->getArg(0)->getType()), 4276 One); 4277 Value *Val = EmitScalarExpr(E->getArg(0)); 4278 Builder.CreateStore(Val, Tmp); 4279 4280 Value *LdPtr = Builder.CreateBitCast(Tmp,llvm::PointerType::getUnqual(STy)); 4281 Val = Builder.CreateLoad(LdPtr); 4282 4283 Value *Arg0 = Builder.CreateExtractValue(Val, 0); 4284 Value *Arg1 = Builder.CreateExtractValue(Val, 1); 4285 Value *StPtr = Builder.CreateBitCast(EmitScalarExpr(E->getArg(1)), 4286 Int8PtrTy); 4287 return Builder.CreateCall(F, {Arg0, Arg1, StPtr}, "stxp"); 4288 } 4289 4290 if (BuiltinID == AArch64::BI__builtin_arm_strex || 4291 BuiltinID == AArch64::BI__builtin_arm_stlex) { 4292 Value *StoreVal = EmitScalarExpr(E->getArg(0)); 4293 Value *StoreAddr = EmitScalarExpr(E->getArg(1)); 4294 4295 QualType Ty = E->getArg(0)->getType(); 4296 llvm::Type *StoreTy = llvm::IntegerType::get(getLLVMContext(), 4297 getContext().getTypeSize(Ty)); 4298 StoreAddr = Builder.CreateBitCast(StoreAddr, StoreTy->getPointerTo()); 4299 4300 if (StoreVal->getType()->isPointerTy()) 4301 StoreVal = Builder.CreatePtrToInt(StoreVal, Int64Ty); 4302 else { 4303 StoreVal = Builder.CreateBitCast(StoreVal, StoreTy); 4304 StoreVal = Builder.CreateZExtOrBitCast(StoreVal, Int64Ty); 4305 } 4306 4307 Function *F = CGM.getIntrinsic(BuiltinID == AArch64::BI__builtin_arm_stlex 4308 ? Intrinsic::aarch64_stlxr 4309 : Intrinsic::aarch64_stxr, 4310 StoreAddr->getType()); 4311 return Builder.CreateCall(F, {StoreVal, StoreAddr}, "stxr"); 4312 } 4313 4314 if (BuiltinID == AArch64::BI__builtin_arm_clrex) { 4315 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_clrex); 4316 return Builder.CreateCall(F); 4317 } 4318 4319 if (BuiltinID == AArch64::BI__builtin_thread_pointer) { 4320 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_thread_pointer); 4321 return Builder.CreateCall(F); 4322 } 4323 4324 // CRC32 4325 Intrinsic::ID CRCIntrinsicID = Intrinsic::not_intrinsic; 4326 switch (BuiltinID) { 4327 case AArch64::BI__builtin_arm_crc32b: 4328 CRCIntrinsicID = Intrinsic::aarch64_crc32b; break; 4329 case AArch64::BI__builtin_arm_crc32cb: 4330 CRCIntrinsicID = Intrinsic::aarch64_crc32cb; break; 4331 case AArch64::BI__builtin_arm_crc32h: 4332 CRCIntrinsicID = Intrinsic::aarch64_crc32h; break; 4333 case AArch64::BI__builtin_arm_crc32ch: 4334 CRCIntrinsicID = Intrinsic::aarch64_crc32ch; break; 4335 case AArch64::BI__builtin_arm_crc32w: 4336 CRCIntrinsicID = Intrinsic::aarch64_crc32w; break; 4337 case AArch64::BI__builtin_arm_crc32cw: 4338 CRCIntrinsicID = Intrinsic::aarch64_crc32cw; break; 4339 case AArch64::BI__builtin_arm_crc32d: 4340 CRCIntrinsicID = Intrinsic::aarch64_crc32x; break; 4341 case AArch64::BI__builtin_arm_crc32cd: 4342 CRCIntrinsicID = Intrinsic::aarch64_crc32cx; break; 4343 } 4344 4345 if (CRCIntrinsicID != Intrinsic::not_intrinsic) { 4346 Value *Arg0 = EmitScalarExpr(E->getArg(0)); 4347 Value *Arg1 = EmitScalarExpr(E->getArg(1)); 4348 Function *F = CGM.getIntrinsic(CRCIntrinsicID); 4349 4350 llvm::Type *DataTy = F->getFunctionType()->getParamType(1); 4351 Arg1 = Builder.CreateZExtOrBitCast(Arg1, DataTy); 4352 4353 return Builder.CreateCall(F, {Arg0, Arg1}); 4354 } 4355 4356 if (BuiltinID == AArch64::BI__builtin_arm_rsr || 4357 BuiltinID == AArch64::BI__builtin_arm_rsr64 || 4358 BuiltinID == AArch64::BI__builtin_arm_rsrp || 4359 BuiltinID == AArch64::BI__builtin_arm_wsr || 4360 BuiltinID == AArch64::BI__builtin_arm_wsr64 || 4361 BuiltinID == AArch64::BI__builtin_arm_wsrp) { 4362 4363 bool IsRead = BuiltinID == AArch64::BI__builtin_arm_rsr || 4364 BuiltinID == AArch64::BI__builtin_arm_rsr64 || 4365 BuiltinID == AArch64::BI__builtin_arm_rsrp; 4366 4367 bool IsPointerBuiltin = BuiltinID == AArch64::BI__builtin_arm_rsrp || 4368 BuiltinID == AArch64::BI__builtin_arm_wsrp; 4369 4370 bool Is64Bit = BuiltinID != AArch64::BI__builtin_arm_rsr && 4371 BuiltinID != AArch64::BI__builtin_arm_wsr; 4372 4373 llvm::Type *ValueType; 4374 llvm::Type *RegisterType = Int64Ty; 4375 if (IsPointerBuiltin) { 4376 ValueType = VoidPtrTy; 4377 } else if (Is64Bit) { 4378 ValueType = Int64Ty; 4379 } else { 4380 ValueType = Int32Ty; 4381 } 4382 4383 return EmitSpecialRegisterBuiltin(*this, E, RegisterType, ValueType, IsRead); 4384 } 4385 4386 // Find out if any arguments are required to be integer constant 4387 // expressions. 4388 unsigned ICEArguments = 0; 4389 ASTContext::GetBuiltinTypeError Error; 4390 getContext().GetBuiltinType(BuiltinID, Error, &ICEArguments); 4391 assert(Error == ASTContext::GE_None && "Should not codegen an error"); 4392 4393 llvm::SmallVector<Value*, 4> Ops; 4394 for (unsigned i = 0, e = E->getNumArgs() - 1; i != e; i++) { 4395 if ((ICEArguments & (1 << i)) == 0) { 4396 Ops.push_back(EmitScalarExpr(E->getArg(i))); 4397 } else { 4398 // If this is required to be a constant, constant fold it so that we know 4399 // that the generated intrinsic gets a ConstantInt. 4400 llvm::APSInt Result; 4401 bool IsConst = E->getArg(i)->isIntegerConstantExpr(Result, getContext()); 4402 assert(IsConst && "Constant arg isn't actually constant?"); 4403 (void)IsConst; 4404 Ops.push_back(llvm::ConstantInt::get(getLLVMContext(), Result)); 4405 } 4406 } 4407 4408 auto SISDMap = makeArrayRef(AArch64SISDIntrinsicMap); 4409 const NeonIntrinsicInfo *Builtin = findNeonIntrinsicInMap( 4410 SISDMap, BuiltinID, AArch64SISDIntrinsicsProvenSorted); 4411 4412 if (Builtin) { 4413 Ops.push_back(EmitScalarExpr(E->getArg(E->getNumArgs() - 1))); 4414 Value *Result = EmitCommonNeonSISDBuiltinExpr(*this, *Builtin, Ops, E); 4415 assert(Result && "SISD intrinsic should have been handled"); 4416 return Result; 4417 } 4418 4419 llvm::APSInt Result; 4420 const Expr *Arg = E->getArg(E->getNumArgs()-1); 4421 NeonTypeFlags Type(0); 4422 if (Arg->isIntegerConstantExpr(Result, getContext())) 4423 // Determine the type of this overloaded NEON intrinsic. 4424 Type = NeonTypeFlags(Result.getZExtValue()); 4425 4426 bool usgn = Type.isUnsigned(); 4427 bool quad = Type.isQuad(); 4428 4429 // Handle non-overloaded intrinsics first. 4430 switch (BuiltinID) { 4431 default: break; 4432 case NEON::BI__builtin_neon_vldrq_p128: { 4433 llvm::Type *Int128PTy = llvm::Type::getIntNPtrTy(getLLVMContext(), 128); 4434 Value *Ptr = Builder.CreateBitCast(EmitScalarExpr(E->getArg(0)), Int128PTy); 4435 return Builder.CreateLoad(Ptr); 4436 } 4437 case NEON::BI__builtin_neon_vstrq_p128: { 4438 llvm::Type *Int128PTy = llvm::Type::getIntNPtrTy(getLLVMContext(), 128); 4439 Value *Ptr = Builder.CreateBitCast(Ops[0], Int128PTy); 4440 return Builder.CreateStore(EmitScalarExpr(E->getArg(1)), Ptr); 4441 } 4442 case NEON::BI__builtin_neon_vcvts_u32_f32: 4443 case NEON::BI__builtin_neon_vcvtd_u64_f64: 4444 usgn = true; 4445 // FALL THROUGH 4446 case NEON::BI__builtin_neon_vcvts_s32_f32: 4447 case NEON::BI__builtin_neon_vcvtd_s64_f64: { 4448 Ops.push_back(EmitScalarExpr(E->getArg(0))); 4449 bool Is64 = Ops[0]->getType()->getPrimitiveSizeInBits() == 64; 4450 llvm::Type *InTy = Is64 ? Int64Ty : Int32Ty; 4451 llvm::Type *FTy = Is64 ? DoubleTy : FloatTy; 4452 Ops[0] = Builder.CreateBitCast(Ops[0], FTy); 4453 if (usgn) 4454 return Builder.CreateFPToUI(Ops[0], InTy); 4455 return Builder.CreateFPToSI(Ops[0], InTy); 4456 } 4457 case NEON::BI__builtin_neon_vcvts_f32_u32: 4458 case NEON::BI__builtin_neon_vcvtd_f64_u64: 4459 usgn = true; 4460 // FALL THROUGH 4461 case NEON::BI__builtin_neon_vcvts_f32_s32: 4462 case NEON::BI__builtin_neon_vcvtd_f64_s64: { 4463 Ops.push_back(EmitScalarExpr(E->getArg(0))); 4464 bool Is64 = Ops[0]->getType()->getPrimitiveSizeInBits() == 64; 4465 llvm::Type *InTy = Is64 ? Int64Ty : Int32Ty; 4466 llvm::Type *FTy = Is64 ? DoubleTy : FloatTy; 4467 Ops[0] = Builder.CreateBitCast(Ops[0], InTy); 4468 if (usgn) 4469 return Builder.CreateUIToFP(Ops[0], FTy); 4470 return Builder.CreateSIToFP(Ops[0], FTy); 4471 } 4472 case NEON::BI__builtin_neon_vpaddd_s64: { 4473 llvm::Type *Ty = llvm::VectorType::get(Int64Ty, 2); 4474 Value *Vec = EmitScalarExpr(E->getArg(0)); 4475 // The vector is v2f64, so make sure it's bitcast to that. 4476 Vec = Builder.CreateBitCast(Vec, Ty, "v2i64"); 4477 llvm::Value *Idx0 = llvm::ConstantInt::get(SizeTy, 0); 4478 llvm::Value *Idx1 = llvm::ConstantInt::get(SizeTy, 1); 4479 Value *Op0 = Builder.CreateExtractElement(Vec, Idx0, "lane0"); 4480 Value *Op1 = Builder.CreateExtractElement(Vec, Idx1, "lane1"); 4481 // Pairwise addition of a v2f64 into a scalar f64. 4482 return Builder.CreateAdd(Op0, Op1, "vpaddd"); 4483 } 4484 case NEON::BI__builtin_neon_vpaddd_f64: { 4485 llvm::Type *Ty = 4486 llvm::VectorType::get(llvm::Type::getDoubleTy(getLLVMContext()), 2); 4487 Value *Vec = EmitScalarExpr(E->getArg(0)); 4488 // The vector is v2f64, so make sure it's bitcast to that. 4489 Vec = Builder.CreateBitCast(Vec, Ty, "v2f64"); 4490 llvm::Value *Idx0 = llvm::ConstantInt::get(SizeTy, 0); 4491 llvm::Value *Idx1 = llvm::ConstantInt::get(SizeTy, 1); 4492 Value *Op0 = Builder.CreateExtractElement(Vec, Idx0, "lane0"); 4493 Value *Op1 = Builder.CreateExtractElement(Vec, Idx1, "lane1"); 4494 // Pairwise addition of a v2f64 into a scalar f64. 4495 return Builder.CreateFAdd(Op0, Op1, "vpaddd"); 4496 } 4497 case NEON::BI__builtin_neon_vpadds_f32: { 4498 llvm::Type *Ty = 4499 llvm::VectorType::get(llvm::Type::getFloatTy(getLLVMContext()), 2); 4500 Value *Vec = EmitScalarExpr(E->getArg(0)); 4501 // The vector is v2f32, so make sure it's bitcast to that. 4502 Vec = Builder.CreateBitCast(Vec, Ty, "v2f32"); 4503 llvm::Value *Idx0 = llvm::ConstantInt::get(SizeTy, 0); 4504 llvm::Value *Idx1 = llvm::ConstantInt::get(SizeTy, 1); 4505 Value *Op0 = Builder.CreateExtractElement(Vec, Idx0, "lane0"); 4506 Value *Op1 = Builder.CreateExtractElement(Vec, Idx1, "lane1"); 4507 // Pairwise addition of a v2f32 into a scalar f32. 4508 return Builder.CreateFAdd(Op0, Op1, "vpaddd"); 4509 } 4510 case NEON::BI__builtin_neon_vceqzd_s64: 4511 case NEON::BI__builtin_neon_vceqzd_f64: 4512 case NEON::BI__builtin_neon_vceqzs_f32: 4513 Ops.push_back(EmitScalarExpr(E->getArg(0))); 4514 return EmitAArch64CompareBuiltinExpr( 4515 Ops[0], ConvertType(E->getCallReturnType(getContext())), 4516 ICmpInst::FCMP_OEQ, ICmpInst::ICMP_EQ, "vceqz"); 4517 case NEON::BI__builtin_neon_vcgezd_s64: 4518 case NEON::BI__builtin_neon_vcgezd_f64: 4519 case NEON::BI__builtin_neon_vcgezs_f32: 4520 Ops.push_back(EmitScalarExpr(E->getArg(0))); 4521 return EmitAArch64CompareBuiltinExpr( 4522 Ops[0], ConvertType(E->getCallReturnType(getContext())), 4523 ICmpInst::FCMP_OGE, ICmpInst::ICMP_SGE, "vcgez"); 4524 case NEON::BI__builtin_neon_vclezd_s64: 4525 case NEON::BI__builtin_neon_vclezd_f64: 4526 case NEON::BI__builtin_neon_vclezs_f32: 4527 Ops.push_back(EmitScalarExpr(E->getArg(0))); 4528 return EmitAArch64CompareBuiltinExpr( 4529 Ops[0], ConvertType(E->getCallReturnType(getContext())), 4530 ICmpInst::FCMP_OLE, ICmpInst::ICMP_SLE, "vclez"); 4531 case NEON::BI__builtin_neon_vcgtzd_s64: 4532 case NEON::BI__builtin_neon_vcgtzd_f64: 4533 case NEON::BI__builtin_neon_vcgtzs_f32: 4534 Ops.push_back(EmitScalarExpr(E->getArg(0))); 4535 return EmitAArch64CompareBuiltinExpr( 4536 Ops[0], ConvertType(E->getCallReturnType(getContext())), 4537 ICmpInst::FCMP_OGT, ICmpInst::ICMP_SGT, "vcgtz"); 4538 case NEON::BI__builtin_neon_vcltzd_s64: 4539 case NEON::BI__builtin_neon_vcltzd_f64: 4540 case NEON::BI__builtin_neon_vcltzs_f32: 4541 Ops.push_back(EmitScalarExpr(E->getArg(0))); 4542 return EmitAArch64CompareBuiltinExpr( 4543 Ops[0], ConvertType(E->getCallReturnType(getContext())), 4544 ICmpInst::FCMP_OLT, ICmpInst::ICMP_SLT, "vcltz"); 4545 4546 case NEON::BI__builtin_neon_vceqzd_u64: { 4547 Ops.push_back(EmitScalarExpr(E->getArg(0))); 4548 Ops[0] = Builder.CreateBitCast(Ops[0], Int64Ty); 4549 Ops[0] = 4550 Builder.CreateICmpEQ(Ops[0], llvm::Constant::getNullValue(Int64Ty)); 4551 return Builder.CreateSExt(Ops[0], Int64Ty, "vceqzd"); 4552 } 4553 case NEON::BI__builtin_neon_vceqd_f64: 4554 case NEON::BI__builtin_neon_vcled_f64: 4555 case NEON::BI__builtin_neon_vcltd_f64: 4556 case NEON::BI__builtin_neon_vcged_f64: 4557 case NEON::BI__builtin_neon_vcgtd_f64: { 4558 llvm::CmpInst::Predicate P; 4559 switch (BuiltinID) { 4560 default: llvm_unreachable("missing builtin ID in switch!"); 4561 case NEON::BI__builtin_neon_vceqd_f64: P = llvm::FCmpInst::FCMP_OEQ; break; 4562 case NEON::BI__builtin_neon_vcled_f64: P = llvm::FCmpInst::FCMP_OLE; break; 4563 case NEON::BI__builtin_neon_vcltd_f64: P = llvm::FCmpInst::FCMP_OLT; break; 4564 case NEON::BI__builtin_neon_vcged_f64: P = llvm::FCmpInst::FCMP_OGE; break; 4565 case NEON::BI__builtin_neon_vcgtd_f64: P = llvm::FCmpInst::FCMP_OGT; break; 4566 } 4567 Ops.push_back(EmitScalarExpr(E->getArg(1))); 4568 Ops[0] = Builder.CreateBitCast(Ops[0], DoubleTy); 4569 Ops[1] = Builder.CreateBitCast(Ops[1], DoubleTy); 4570 Ops[0] = Builder.CreateFCmp(P, Ops[0], Ops[1]); 4571 return Builder.CreateSExt(Ops[0], Int64Ty, "vcmpd"); 4572 } 4573 case NEON::BI__builtin_neon_vceqs_f32: 4574 case NEON::BI__builtin_neon_vcles_f32: 4575 case NEON::BI__builtin_neon_vclts_f32: 4576 case NEON::BI__builtin_neon_vcges_f32: 4577 case NEON::BI__builtin_neon_vcgts_f32: { 4578 llvm::CmpInst::Predicate P; 4579 switch (BuiltinID) { 4580 default: llvm_unreachable("missing builtin ID in switch!"); 4581 case NEON::BI__builtin_neon_vceqs_f32: P = llvm::FCmpInst::FCMP_OEQ; break; 4582 case NEON::BI__builtin_neon_vcles_f32: P = llvm::FCmpInst::FCMP_OLE; break; 4583 case NEON::BI__builtin_neon_vclts_f32: P = llvm::FCmpInst::FCMP_OLT; break; 4584 case NEON::BI__builtin_neon_vcges_f32: P = llvm::FCmpInst::FCMP_OGE; break; 4585 case NEON::BI__builtin_neon_vcgts_f32: P = llvm::FCmpInst::FCMP_OGT; break; 4586 } 4587 Ops.push_back(EmitScalarExpr(E->getArg(1))); 4588 Ops[0] = Builder.CreateBitCast(Ops[0], FloatTy); 4589 Ops[1] = Builder.CreateBitCast(Ops[1], FloatTy); 4590 Ops[0] = Builder.CreateFCmp(P, Ops[0], Ops[1]); 4591 return Builder.CreateSExt(Ops[0], Int32Ty, "vcmpd"); 4592 } 4593 case NEON::BI__builtin_neon_vceqd_s64: 4594 case NEON::BI__builtin_neon_vceqd_u64: 4595 case NEON::BI__builtin_neon_vcgtd_s64: 4596 case NEON::BI__builtin_neon_vcgtd_u64: 4597 case NEON::BI__builtin_neon_vcltd_s64: 4598 case NEON::BI__builtin_neon_vcltd_u64: 4599 case NEON::BI__builtin_neon_vcged_u64: 4600 case NEON::BI__builtin_neon_vcged_s64: 4601 case NEON::BI__builtin_neon_vcled_u64: 4602 case NEON::BI__builtin_neon_vcled_s64: { 4603 llvm::CmpInst::Predicate P; 4604 switch (BuiltinID) { 4605 default: llvm_unreachable("missing builtin ID in switch!"); 4606 case NEON::BI__builtin_neon_vceqd_s64: 4607 case NEON::BI__builtin_neon_vceqd_u64:P = llvm::ICmpInst::ICMP_EQ;break; 4608 case NEON::BI__builtin_neon_vcgtd_s64:P = llvm::ICmpInst::ICMP_SGT;break; 4609 case NEON::BI__builtin_neon_vcgtd_u64:P = llvm::ICmpInst::ICMP_UGT;break; 4610 case NEON::BI__builtin_neon_vcltd_s64:P = llvm::ICmpInst::ICMP_SLT;break; 4611 case NEON::BI__builtin_neon_vcltd_u64:P = llvm::ICmpInst::ICMP_ULT;break; 4612 case NEON::BI__builtin_neon_vcged_u64:P = llvm::ICmpInst::ICMP_UGE;break; 4613 case NEON::BI__builtin_neon_vcged_s64:P = llvm::ICmpInst::ICMP_SGE;break; 4614 case NEON::BI__builtin_neon_vcled_u64:P = llvm::ICmpInst::ICMP_ULE;break; 4615 case NEON::BI__builtin_neon_vcled_s64:P = llvm::ICmpInst::ICMP_SLE;break; 4616 } 4617 Ops.push_back(EmitScalarExpr(E->getArg(1))); 4618 Ops[0] = Builder.CreateBitCast(Ops[0], Int64Ty); 4619 Ops[1] = Builder.CreateBitCast(Ops[1], Int64Ty); 4620 Ops[0] = Builder.CreateICmp(P, Ops[0], Ops[1]); 4621 return Builder.CreateSExt(Ops[0], Int64Ty, "vceqd"); 4622 } 4623 case NEON::BI__builtin_neon_vtstd_s64: 4624 case NEON::BI__builtin_neon_vtstd_u64: { 4625 Ops.push_back(EmitScalarExpr(E->getArg(1))); 4626 Ops[0] = Builder.CreateBitCast(Ops[0], Int64Ty); 4627 Ops[1] = Builder.CreateBitCast(Ops[1], Int64Ty); 4628 Ops[0] = Builder.CreateAnd(Ops[0], Ops[1]); 4629 Ops[0] = Builder.CreateICmp(ICmpInst::ICMP_NE, Ops[0], 4630 llvm::Constant::getNullValue(Int64Ty)); 4631 return Builder.CreateSExt(Ops[0], Int64Ty, "vtstd"); 4632 } 4633 case NEON::BI__builtin_neon_vset_lane_i8: 4634 case NEON::BI__builtin_neon_vset_lane_i16: 4635 case NEON::BI__builtin_neon_vset_lane_i32: 4636 case NEON::BI__builtin_neon_vset_lane_i64: 4637 case NEON::BI__builtin_neon_vset_lane_f32: 4638 case NEON::BI__builtin_neon_vsetq_lane_i8: 4639 case NEON::BI__builtin_neon_vsetq_lane_i16: 4640 case NEON::BI__builtin_neon_vsetq_lane_i32: 4641 case NEON::BI__builtin_neon_vsetq_lane_i64: 4642 case NEON::BI__builtin_neon_vsetq_lane_f32: 4643 Ops.push_back(EmitScalarExpr(E->getArg(2))); 4644 return Builder.CreateInsertElement(Ops[1], Ops[0], Ops[2], "vset_lane"); 4645 case NEON::BI__builtin_neon_vset_lane_f64: 4646 // The vector type needs a cast for the v1f64 variant. 4647 Ops[1] = Builder.CreateBitCast(Ops[1], 4648 llvm::VectorType::get(DoubleTy, 1)); 4649 Ops.push_back(EmitScalarExpr(E->getArg(2))); 4650 return Builder.CreateInsertElement(Ops[1], Ops[0], Ops[2], "vset_lane"); 4651 case NEON::BI__builtin_neon_vsetq_lane_f64: 4652 // The vector type needs a cast for the v2f64 variant. 4653 Ops[1] = Builder.CreateBitCast(Ops[1], 4654 llvm::VectorType::get(llvm::Type::getDoubleTy(getLLVMContext()), 2)); 4655 Ops.push_back(EmitScalarExpr(E->getArg(2))); 4656 return Builder.CreateInsertElement(Ops[1], Ops[0], Ops[2], "vset_lane"); 4657 4658 case NEON::BI__builtin_neon_vget_lane_i8: 4659 case NEON::BI__builtin_neon_vdupb_lane_i8: 4660 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::VectorType::get(Int8Ty, 8)); 4661 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4662 "vget_lane"); 4663 case NEON::BI__builtin_neon_vgetq_lane_i8: 4664 case NEON::BI__builtin_neon_vdupb_laneq_i8: 4665 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::VectorType::get(Int8Ty, 16)); 4666 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4667 "vgetq_lane"); 4668 case NEON::BI__builtin_neon_vget_lane_i16: 4669 case NEON::BI__builtin_neon_vduph_lane_i16: 4670 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::VectorType::get(Int16Ty, 4)); 4671 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4672 "vget_lane"); 4673 case NEON::BI__builtin_neon_vgetq_lane_i16: 4674 case NEON::BI__builtin_neon_vduph_laneq_i16: 4675 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::VectorType::get(Int16Ty, 8)); 4676 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4677 "vgetq_lane"); 4678 case NEON::BI__builtin_neon_vget_lane_i32: 4679 case NEON::BI__builtin_neon_vdups_lane_i32: 4680 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::VectorType::get(Int32Ty, 2)); 4681 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4682 "vget_lane"); 4683 case NEON::BI__builtin_neon_vdups_lane_f32: 4684 Ops[0] = Builder.CreateBitCast(Ops[0], 4685 llvm::VectorType::get(llvm::Type::getFloatTy(getLLVMContext()), 2)); 4686 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4687 "vdups_lane"); 4688 case NEON::BI__builtin_neon_vgetq_lane_i32: 4689 case NEON::BI__builtin_neon_vdups_laneq_i32: 4690 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::VectorType::get(Int32Ty, 4)); 4691 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4692 "vgetq_lane"); 4693 case NEON::BI__builtin_neon_vget_lane_i64: 4694 case NEON::BI__builtin_neon_vdupd_lane_i64: 4695 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::VectorType::get(Int64Ty, 1)); 4696 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4697 "vget_lane"); 4698 case NEON::BI__builtin_neon_vdupd_lane_f64: 4699 Ops[0] = Builder.CreateBitCast(Ops[0], 4700 llvm::VectorType::get(llvm::Type::getDoubleTy(getLLVMContext()), 1)); 4701 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4702 "vdupd_lane"); 4703 case NEON::BI__builtin_neon_vgetq_lane_i64: 4704 case NEON::BI__builtin_neon_vdupd_laneq_i64: 4705 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::VectorType::get(Int64Ty, 2)); 4706 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4707 "vgetq_lane"); 4708 case NEON::BI__builtin_neon_vget_lane_f32: 4709 Ops[0] = Builder.CreateBitCast(Ops[0], 4710 llvm::VectorType::get(llvm::Type::getFloatTy(getLLVMContext()), 2)); 4711 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4712 "vget_lane"); 4713 case NEON::BI__builtin_neon_vget_lane_f64: 4714 Ops[0] = Builder.CreateBitCast(Ops[0], 4715 llvm::VectorType::get(llvm::Type::getDoubleTy(getLLVMContext()), 1)); 4716 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4717 "vget_lane"); 4718 case NEON::BI__builtin_neon_vgetq_lane_f32: 4719 case NEON::BI__builtin_neon_vdups_laneq_f32: 4720 Ops[0] = Builder.CreateBitCast(Ops[0], 4721 llvm::VectorType::get(llvm::Type::getFloatTy(getLLVMContext()), 4)); 4722 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4723 "vgetq_lane"); 4724 case NEON::BI__builtin_neon_vgetq_lane_f64: 4725 case NEON::BI__builtin_neon_vdupd_laneq_f64: 4726 Ops[0] = Builder.CreateBitCast(Ops[0], 4727 llvm::VectorType::get(llvm::Type::getDoubleTy(getLLVMContext()), 2)); 4728 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4729 "vgetq_lane"); 4730 case NEON::BI__builtin_neon_vaddd_s64: 4731 case NEON::BI__builtin_neon_vaddd_u64: 4732 return Builder.CreateAdd(Ops[0], EmitScalarExpr(E->getArg(1)), "vaddd"); 4733 case NEON::BI__builtin_neon_vsubd_s64: 4734 case NEON::BI__builtin_neon_vsubd_u64: 4735 return Builder.CreateSub(Ops[0], EmitScalarExpr(E->getArg(1)), "vsubd"); 4736 case NEON::BI__builtin_neon_vqdmlalh_s16: 4737 case NEON::BI__builtin_neon_vqdmlslh_s16: { 4738 SmallVector<Value *, 2> ProductOps; 4739 ProductOps.push_back(vectorWrapScalar16(Ops[1])); 4740 ProductOps.push_back(vectorWrapScalar16(EmitScalarExpr(E->getArg(2)))); 4741 llvm::Type *VTy = llvm::VectorType::get(Int32Ty, 4); 4742 Ops[1] = EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_sqdmull, VTy), 4743 ProductOps, "vqdmlXl"); 4744 Constant *CI = ConstantInt::get(SizeTy, 0); 4745 Ops[1] = Builder.CreateExtractElement(Ops[1], CI, "lane0"); 4746 4747 unsigned AccumInt = BuiltinID == NEON::BI__builtin_neon_vqdmlalh_s16 4748 ? Intrinsic::aarch64_neon_sqadd 4749 : Intrinsic::aarch64_neon_sqsub; 4750 return EmitNeonCall(CGM.getIntrinsic(AccumInt, Int32Ty), Ops, "vqdmlXl"); 4751 } 4752 case NEON::BI__builtin_neon_vqshlud_n_s64: { 4753 Ops.push_back(EmitScalarExpr(E->getArg(1))); 4754 Ops[1] = Builder.CreateZExt(Ops[1], Int64Ty); 4755 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_sqshlu, Int64Ty), 4756 Ops, "vqshlu_n"); 4757 } 4758 case NEON::BI__builtin_neon_vqshld_n_u64: 4759 case NEON::BI__builtin_neon_vqshld_n_s64: { 4760 unsigned Int = BuiltinID == NEON::BI__builtin_neon_vqshld_n_u64 4761 ? Intrinsic::aarch64_neon_uqshl 4762 : Intrinsic::aarch64_neon_sqshl; 4763 Ops.push_back(EmitScalarExpr(E->getArg(1))); 4764 Ops[1] = Builder.CreateZExt(Ops[1], Int64Ty); 4765 return EmitNeonCall(CGM.getIntrinsic(Int, Int64Ty), Ops, "vqshl_n"); 4766 } 4767 case NEON::BI__builtin_neon_vrshrd_n_u64: 4768 case NEON::BI__builtin_neon_vrshrd_n_s64: { 4769 unsigned Int = BuiltinID == NEON::BI__builtin_neon_vrshrd_n_u64 4770 ? Intrinsic::aarch64_neon_urshl 4771 : Intrinsic::aarch64_neon_srshl; 4772 Ops.push_back(EmitScalarExpr(E->getArg(1))); 4773 int SV = cast<ConstantInt>(Ops[1])->getSExtValue(); 4774 Ops[1] = ConstantInt::get(Int64Ty, -SV); 4775 return EmitNeonCall(CGM.getIntrinsic(Int, Int64Ty), Ops, "vrshr_n"); 4776 } 4777 case NEON::BI__builtin_neon_vrsrad_n_u64: 4778 case NEON::BI__builtin_neon_vrsrad_n_s64: { 4779 unsigned Int = BuiltinID == NEON::BI__builtin_neon_vrsrad_n_u64 4780 ? Intrinsic::aarch64_neon_urshl 4781 : Intrinsic::aarch64_neon_srshl; 4782 Ops[1] = Builder.CreateBitCast(Ops[1], Int64Ty); 4783 Ops.push_back(Builder.CreateNeg(EmitScalarExpr(E->getArg(2)))); 4784 Ops[1] = Builder.CreateCall(CGM.getIntrinsic(Int, Int64Ty), 4785 {Ops[1], Builder.CreateSExt(Ops[2], Int64Ty)}); 4786 return Builder.CreateAdd(Ops[0], Builder.CreateBitCast(Ops[1], Int64Ty)); 4787 } 4788 case NEON::BI__builtin_neon_vshld_n_s64: 4789 case NEON::BI__builtin_neon_vshld_n_u64: { 4790 llvm::ConstantInt *Amt = cast<ConstantInt>(EmitScalarExpr(E->getArg(1))); 4791 return Builder.CreateShl( 4792 Ops[0], ConstantInt::get(Int64Ty, Amt->getZExtValue()), "shld_n"); 4793 } 4794 case NEON::BI__builtin_neon_vshrd_n_s64: { 4795 llvm::ConstantInt *Amt = cast<ConstantInt>(EmitScalarExpr(E->getArg(1))); 4796 return Builder.CreateAShr( 4797 Ops[0], ConstantInt::get(Int64Ty, std::min(static_cast<uint64_t>(63), 4798 Amt->getZExtValue())), 4799 "shrd_n"); 4800 } 4801 case NEON::BI__builtin_neon_vshrd_n_u64: { 4802 llvm::ConstantInt *Amt = cast<ConstantInt>(EmitScalarExpr(E->getArg(1))); 4803 uint64_t ShiftAmt = Amt->getZExtValue(); 4804 // Right-shifting an unsigned value by its size yields 0. 4805 if (ShiftAmt == 64) 4806 return ConstantInt::get(Int64Ty, 0); 4807 return Builder.CreateLShr(Ops[0], ConstantInt::get(Int64Ty, ShiftAmt), 4808 "shrd_n"); 4809 } 4810 case NEON::BI__builtin_neon_vsrad_n_s64: { 4811 llvm::ConstantInt *Amt = cast<ConstantInt>(EmitScalarExpr(E->getArg(2))); 4812 Ops[1] = Builder.CreateAShr( 4813 Ops[1], ConstantInt::get(Int64Ty, std::min(static_cast<uint64_t>(63), 4814 Amt->getZExtValue())), 4815 "shrd_n"); 4816 return Builder.CreateAdd(Ops[0], Ops[1]); 4817 } 4818 case NEON::BI__builtin_neon_vsrad_n_u64: { 4819 llvm::ConstantInt *Amt = cast<ConstantInt>(EmitScalarExpr(E->getArg(2))); 4820 uint64_t ShiftAmt = Amt->getZExtValue(); 4821 // Right-shifting an unsigned value by its size yields 0. 4822 // As Op + 0 = Op, return Ops[0] directly. 4823 if (ShiftAmt == 64) 4824 return Ops[0]; 4825 Ops[1] = Builder.CreateLShr(Ops[1], ConstantInt::get(Int64Ty, ShiftAmt), 4826 "shrd_n"); 4827 return Builder.CreateAdd(Ops[0], Ops[1]); 4828 } 4829 case NEON::BI__builtin_neon_vqdmlalh_lane_s16: 4830 case NEON::BI__builtin_neon_vqdmlalh_laneq_s16: 4831 case NEON::BI__builtin_neon_vqdmlslh_lane_s16: 4832 case NEON::BI__builtin_neon_vqdmlslh_laneq_s16: { 4833 Ops[2] = Builder.CreateExtractElement(Ops[2], EmitScalarExpr(E->getArg(3)), 4834 "lane"); 4835 SmallVector<Value *, 2> ProductOps; 4836 ProductOps.push_back(vectorWrapScalar16(Ops[1])); 4837 ProductOps.push_back(vectorWrapScalar16(Ops[2])); 4838 llvm::Type *VTy = llvm::VectorType::get(Int32Ty, 4); 4839 Ops[1] = EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_sqdmull, VTy), 4840 ProductOps, "vqdmlXl"); 4841 Constant *CI = ConstantInt::get(SizeTy, 0); 4842 Ops[1] = Builder.CreateExtractElement(Ops[1], CI, "lane0"); 4843 Ops.pop_back(); 4844 4845 unsigned AccInt = (BuiltinID == NEON::BI__builtin_neon_vqdmlalh_lane_s16 || 4846 BuiltinID == NEON::BI__builtin_neon_vqdmlalh_laneq_s16) 4847 ? Intrinsic::aarch64_neon_sqadd 4848 : Intrinsic::aarch64_neon_sqsub; 4849 return EmitNeonCall(CGM.getIntrinsic(AccInt, Int32Ty), Ops, "vqdmlXl"); 4850 } 4851 case NEON::BI__builtin_neon_vqdmlals_s32: 4852 case NEON::BI__builtin_neon_vqdmlsls_s32: { 4853 SmallVector<Value *, 2> ProductOps; 4854 ProductOps.push_back(Ops[1]); 4855 ProductOps.push_back(EmitScalarExpr(E->getArg(2))); 4856 Ops[1] = 4857 EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_sqdmulls_scalar), 4858 ProductOps, "vqdmlXl"); 4859 4860 unsigned AccumInt = BuiltinID == NEON::BI__builtin_neon_vqdmlals_s32 4861 ? Intrinsic::aarch64_neon_sqadd 4862 : Intrinsic::aarch64_neon_sqsub; 4863 return EmitNeonCall(CGM.getIntrinsic(AccumInt, Int64Ty), Ops, "vqdmlXl"); 4864 } 4865 case NEON::BI__builtin_neon_vqdmlals_lane_s32: 4866 case NEON::BI__builtin_neon_vqdmlals_laneq_s32: 4867 case NEON::BI__builtin_neon_vqdmlsls_lane_s32: 4868 case NEON::BI__builtin_neon_vqdmlsls_laneq_s32: { 4869 Ops[2] = Builder.CreateExtractElement(Ops[2], EmitScalarExpr(E->getArg(3)), 4870 "lane"); 4871 SmallVector<Value *, 2> ProductOps; 4872 ProductOps.push_back(Ops[1]); 4873 ProductOps.push_back(Ops[2]); 4874 Ops[1] = 4875 EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_sqdmulls_scalar), 4876 ProductOps, "vqdmlXl"); 4877 Ops.pop_back(); 4878 4879 unsigned AccInt = (BuiltinID == NEON::BI__builtin_neon_vqdmlals_lane_s32 || 4880 BuiltinID == NEON::BI__builtin_neon_vqdmlals_laneq_s32) 4881 ? Intrinsic::aarch64_neon_sqadd 4882 : Intrinsic::aarch64_neon_sqsub; 4883 return EmitNeonCall(CGM.getIntrinsic(AccInt, Int64Ty), Ops, "vqdmlXl"); 4884 } 4885 } 4886 4887 llvm::VectorType *VTy = GetNeonType(this, Type); 4888 llvm::Type *Ty = VTy; 4889 if (!Ty) 4890 return nullptr; 4891 4892 // Not all intrinsics handled by the common case work for AArch64 yet, so only 4893 // defer to common code if it's been added to our special map. 4894 Builtin = findNeonIntrinsicInMap(AArch64SIMDIntrinsicMap, BuiltinID, 4895 AArch64SIMDIntrinsicsProvenSorted); 4896 4897 if (Builtin) 4898 return EmitCommonNeonBuiltinExpr( 4899 Builtin->BuiltinID, Builtin->LLVMIntrinsic, Builtin->AltLLVMIntrinsic, 4900 Builtin->NameHint, Builtin->TypeModifier, E, Ops, nullptr); 4901 4902 if (Value *V = EmitAArch64TblBuiltinExpr(*this, BuiltinID, E, Ops)) 4903 return V; 4904 4905 unsigned Int; 4906 switch (BuiltinID) { 4907 default: return nullptr; 4908 case NEON::BI__builtin_neon_vbsl_v: 4909 case NEON::BI__builtin_neon_vbslq_v: { 4910 llvm::Type *BitTy = llvm::VectorType::getInteger(VTy); 4911 Ops[0] = Builder.CreateBitCast(Ops[0], BitTy, "vbsl"); 4912 Ops[1] = Builder.CreateBitCast(Ops[1], BitTy, "vbsl"); 4913 Ops[2] = Builder.CreateBitCast(Ops[2], BitTy, "vbsl"); 4914 4915 Ops[1] = Builder.CreateAnd(Ops[0], Ops[1], "vbsl"); 4916 Ops[2] = Builder.CreateAnd(Builder.CreateNot(Ops[0]), Ops[2], "vbsl"); 4917 Ops[0] = Builder.CreateOr(Ops[1], Ops[2], "vbsl"); 4918 return Builder.CreateBitCast(Ops[0], Ty); 4919 } 4920 case NEON::BI__builtin_neon_vfma_lane_v: 4921 case NEON::BI__builtin_neon_vfmaq_lane_v: { // Only used for FP types 4922 // The ARM builtins (and instructions) have the addend as the first 4923 // operand, but the 'fma' intrinsics have it last. Swap it around here. 4924 Value *Addend = Ops[0]; 4925 Value *Multiplicand = Ops[1]; 4926 Value *LaneSource = Ops[2]; 4927 Ops[0] = Multiplicand; 4928 Ops[1] = LaneSource; 4929 Ops[2] = Addend; 4930 4931 // Now adjust things to handle the lane access. 4932 llvm::Type *SourceTy = BuiltinID == NEON::BI__builtin_neon_vfmaq_lane_v ? 4933 llvm::VectorType::get(VTy->getElementType(), VTy->getNumElements() / 2) : 4934 VTy; 4935 llvm::Constant *cst = cast<Constant>(Ops[3]); 4936 Value *SV = llvm::ConstantVector::getSplat(VTy->getNumElements(), cst); 4937 Ops[1] = Builder.CreateBitCast(Ops[1], SourceTy); 4938 Ops[1] = Builder.CreateShuffleVector(Ops[1], Ops[1], SV, "lane"); 4939 4940 Ops.pop_back(); 4941 Int = Intrinsic::fma; 4942 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "fmla"); 4943 } 4944 case NEON::BI__builtin_neon_vfma_laneq_v: { 4945 llvm::VectorType *VTy = cast<llvm::VectorType>(Ty); 4946 // v1f64 fma should be mapped to Neon scalar f64 fma 4947 if (VTy && VTy->getElementType() == DoubleTy) { 4948 Ops[0] = Builder.CreateBitCast(Ops[0], DoubleTy); 4949 Ops[1] = Builder.CreateBitCast(Ops[1], DoubleTy); 4950 llvm::Type *VTy = GetNeonType(this, 4951 NeonTypeFlags(NeonTypeFlags::Float64, false, true)); 4952 Ops[2] = Builder.CreateBitCast(Ops[2], VTy); 4953 Ops[2] = Builder.CreateExtractElement(Ops[2], Ops[3], "extract"); 4954 Value *F = CGM.getIntrinsic(Intrinsic::fma, DoubleTy); 4955 Value *Result = Builder.CreateCall(F, {Ops[1], Ops[2], Ops[0]}); 4956 return Builder.CreateBitCast(Result, Ty); 4957 } 4958 Value *F = CGM.getIntrinsic(Intrinsic::fma, Ty); 4959 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 4960 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 4961 4962 llvm::Type *STy = llvm::VectorType::get(VTy->getElementType(), 4963 VTy->getNumElements() * 2); 4964 Ops[2] = Builder.CreateBitCast(Ops[2], STy); 4965 Value* SV = llvm::ConstantVector::getSplat(VTy->getNumElements(), 4966 cast<ConstantInt>(Ops[3])); 4967 Ops[2] = Builder.CreateShuffleVector(Ops[2], Ops[2], SV, "lane"); 4968 4969 return Builder.CreateCall(F, {Ops[2], Ops[1], Ops[0]}); 4970 } 4971 case NEON::BI__builtin_neon_vfmaq_laneq_v: { 4972 Value *F = CGM.getIntrinsic(Intrinsic::fma, Ty); 4973 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 4974 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 4975 4976 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 4977 Ops[2] = EmitNeonSplat(Ops[2], cast<ConstantInt>(Ops[3])); 4978 return Builder.CreateCall(F, {Ops[2], Ops[1], Ops[0]}); 4979 } 4980 case NEON::BI__builtin_neon_vfmas_lane_f32: 4981 case NEON::BI__builtin_neon_vfmas_laneq_f32: 4982 case NEON::BI__builtin_neon_vfmad_lane_f64: 4983 case NEON::BI__builtin_neon_vfmad_laneq_f64: { 4984 Ops.push_back(EmitScalarExpr(E->getArg(3))); 4985 llvm::Type *Ty = ConvertType(E->getCallReturnType(getContext())); 4986 Value *F = CGM.getIntrinsic(Intrinsic::fma, Ty); 4987 Ops[2] = Builder.CreateExtractElement(Ops[2], Ops[3], "extract"); 4988 return Builder.CreateCall(F, {Ops[1], Ops[2], Ops[0]}); 4989 } 4990 case NEON::BI__builtin_neon_vfms_v: 4991 case NEON::BI__builtin_neon_vfmsq_v: { // Only used for FP types 4992 // FIXME: probably remove when we no longer support aarch64_simd.h 4993 // (arm_neon.h delegates to vfma). 4994 4995 // The ARM builtins (and instructions) have the addend as the first 4996 // operand, but the 'fma' intrinsics have it last. Swap it around here. 4997 Value *Subtrahend = Ops[0]; 4998 Value *Multiplicand = Ops[2]; 4999 Ops[0] = Multiplicand; 5000 Ops[2] = Subtrahend; 5001 Ops[1] = Builder.CreateBitCast(Ops[1], VTy); 5002 Ops[1] = Builder.CreateFNeg(Ops[1]); 5003 Int = Intrinsic::fma; 5004 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "fmls"); 5005 } 5006 case NEON::BI__builtin_neon_vmull_v: 5007 // FIXME: improve sharing scheme to cope with 3 alternative LLVM intrinsics. 5008 Int = usgn ? Intrinsic::aarch64_neon_umull : Intrinsic::aarch64_neon_smull; 5009 if (Type.isPoly()) Int = Intrinsic::aarch64_neon_pmull; 5010 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmull"); 5011 case NEON::BI__builtin_neon_vmax_v: 5012 case NEON::BI__builtin_neon_vmaxq_v: 5013 // FIXME: improve sharing scheme to cope with 3 alternative LLVM intrinsics. 5014 Int = usgn ? Intrinsic::aarch64_neon_umax : Intrinsic::aarch64_neon_smax; 5015 if (Ty->isFPOrFPVectorTy()) Int = Intrinsic::aarch64_neon_fmax; 5016 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmax"); 5017 case NEON::BI__builtin_neon_vmin_v: 5018 case NEON::BI__builtin_neon_vminq_v: 5019 // FIXME: improve sharing scheme to cope with 3 alternative LLVM intrinsics. 5020 Int = usgn ? Intrinsic::aarch64_neon_umin : Intrinsic::aarch64_neon_smin; 5021 if (Ty->isFPOrFPVectorTy()) Int = Intrinsic::aarch64_neon_fmin; 5022 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmin"); 5023 case NEON::BI__builtin_neon_vabd_v: 5024 case NEON::BI__builtin_neon_vabdq_v: 5025 // FIXME: improve sharing scheme to cope with 3 alternative LLVM intrinsics. 5026 Int = usgn ? Intrinsic::aarch64_neon_uabd : Intrinsic::aarch64_neon_sabd; 5027 if (Ty->isFPOrFPVectorTy()) Int = Intrinsic::aarch64_neon_fabd; 5028 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vabd"); 5029 case NEON::BI__builtin_neon_vpadal_v: 5030 case NEON::BI__builtin_neon_vpadalq_v: { 5031 unsigned ArgElts = VTy->getNumElements(); 5032 llvm::IntegerType *EltTy = cast<IntegerType>(VTy->getElementType()); 5033 unsigned BitWidth = EltTy->getBitWidth(); 5034 llvm::Type *ArgTy = llvm::VectorType::get( 5035 llvm::IntegerType::get(getLLVMContext(), BitWidth/2), 2*ArgElts); 5036 llvm::Type* Tys[2] = { VTy, ArgTy }; 5037 Int = usgn ? Intrinsic::aarch64_neon_uaddlp : Intrinsic::aarch64_neon_saddlp; 5038 SmallVector<llvm::Value*, 1> TmpOps; 5039 TmpOps.push_back(Ops[1]); 5040 Function *F = CGM.getIntrinsic(Int, Tys); 5041 llvm::Value *tmp = EmitNeonCall(F, TmpOps, "vpadal"); 5042 llvm::Value *addend = Builder.CreateBitCast(Ops[0], tmp->getType()); 5043 return Builder.CreateAdd(tmp, addend); 5044 } 5045 case NEON::BI__builtin_neon_vpmin_v: 5046 case NEON::BI__builtin_neon_vpminq_v: 5047 // FIXME: improve sharing scheme to cope with 3 alternative LLVM intrinsics. 5048 Int = usgn ? Intrinsic::aarch64_neon_uminp : Intrinsic::aarch64_neon_sminp; 5049 if (Ty->isFPOrFPVectorTy()) Int = Intrinsic::aarch64_neon_fminp; 5050 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vpmin"); 5051 case NEON::BI__builtin_neon_vpmax_v: 5052 case NEON::BI__builtin_neon_vpmaxq_v: 5053 // FIXME: improve sharing scheme to cope with 3 alternative LLVM intrinsics. 5054 Int = usgn ? Intrinsic::aarch64_neon_umaxp : Intrinsic::aarch64_neon_smaxp; 5055 if (Ty->isFPOrFPVectorTy()) Int = Intrinsic::aarch64_neon_fmaxp; 5056 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vpmax"); 5057 case NEON::BI__builtin_neon_vminnm_v: 5058 case NEON::BI__builtin_neon_vminnmq_v: 5059 Int = Intrinsic::aarch64_neon_fminnm; 5060 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vminnm"); 5061 case NEON::BI__builtin_neon_vmaxnm_v: 5062 case NEON::BI__builtin_neon_vmaxnmq_v: 5063 Int = Intrinsic::aarch64_neon_fmaxnm; 5064 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmaxnm"); 5065 case NEON::BI__builtin_neon_vrecpss_f32: { 5066 llvm::Type *f32Type = llvm::Type::getFloatTy(getLLVMContext()); 5067 Ops.push_back(EmitScalarExpr(E->getArg(1))); 5068 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_frecps, f32Type), 5069 Ops, "vrecps"); 5070 } 5071 case NEON::BI__builtin_neon_vrecpsd_f64: { 5072 llvm::Type *f64Type = llvm::Type::getDoubleTy(getLLVMContext()); 5073 Ops.push_back(EmitScalarExpr(E->getArg(1))); 5074 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_frecps, f64Type), 5075 Ops, "vrecps"); 5076 } 5077 case NEON::BI__builtin_neon_vqshrun_n_v: 5078 Int = Intrinsic::aarch64_neon_sqshrun; 5079 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqshrun_n"); 5080 case NEON::BI__builtin_neon_vqrshrun_n_v: 5081 Int = Intrinsic::aarch64_neon_sqrshrun; 5082 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqrshrun_n"); 5083 case NEON::BI__builtin_neon_vqshrn_n_v: 5084 Int = usgn ? Intrinsic::aarch64_neon_uqshrn : Intrinsic::aarch64_neon_sqshrn; 5085 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqshrn_n"); 5086 case NEON::BI__builtin_neon_vrshrn_n_v: 5087 Int = Intrinsic::aarch64_neon_rshrn; 5088 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrshrn_n"); 5089 case NEON::BI__builtin_neon_vqrshrn_n_v: 5090 Int = usgn ? Intrinsic::aarch64_neon_uqrshrn : Intrinsic::aarch64_neon_sqrshrn; 5091 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqrshrn_n"); 5092 case NEON::BI__builtin_neon_vrnda_v: 5093 case NEON::BI__builtin_neon_vrndaq_v: { 5094 Int = Intrinsic::round; 5095 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrnda"); 5096 } 5097 case NEON::BI__builtin_neon_vrndi_v: 5098 case NEON::BI__builtin_neon_vrndiq_v: { 5099 Int = Intrinsic::nearbyint; 5100 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrndi"); 5101 } 5102 case NEON::BI__builtin_neon_vrndm_v: 5103 case NEON::BI__builtin_neon_vrndmq_v: { 5104 Int = Intrinsic::floor; 5105 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrndm"); 5106 } 5107 case NEON::BI__builtin_neon_vrndn_v: 5108 case NEON::BI__builtin_neon_vrndnq_v: { 5109 Int = Intrinsic::aarch64_neon_frintn; 5110 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrndn"); 5111 } 5112 case NEON::BI__builtin_neon_vrndp_v: 5113 case NEON::BI__builtin_neon_vrndpq_v: { 5114 Int = Intrinsic::ceil; 5115 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrndp"); 5116 } 5117 case NEON::BI__builtin_neon_vrndx_v: 5118 case NEON::BI__builtin_neon_vrndxq_v: { 5119 Int = Intrinsic::rint; 5120 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrndx"); 5121 } 5122 case NEON::BI__builtin_neon_vrnd_v: 5123 case NEON::BI__builtin_neon_vrndq_v: { 5124 Int = Intrinsic::trunc; 5125 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrndz"); 5126 } 5127 case NEON::BI__builtin_neon_vceqz_v: 5128 case NEON::BI__builtin_neon_vceqzq_v: 5129 return EmitAArch64CompareBuiltinExpr(Ops[0], Ty, ICmpInst::FCMP_OEQ, 5130 ICmpInst::ICMP_EQ, "vceqz"); 5131 case NEON::BI__builtin_neon_vcgez_v: 5132 case NEON::BI__builtin_neon_vcgezq_v: 5133 return EmitAArch64CompareBuiltinExpr(Ops[0], Ty, ICmpInst::FCMP_OGE, 5134 ICmpInst::ICMP_SGE, "vcgez"); 5135 case NEON::BI__builtin_neon_vclez_v: 5136 case NEON::BI__builtin_neon_vclezq_v: 5137 return EmitAArch64CompareBuiltinExpr(Ops[0], Ty, ICmpInst::FCMP_OLE, 5138 ICmpInst::ICMP_SLE, "vclez"); 5139 case NEON::BI__builtin_neon_vcgtz_v: 5140 case NEON::BI__builtin_neon_vcgtzq_v: 5141 return EmitAArch64CompareBuiltinExpr(Ops[0], Ty, ICmpInst::FCMP_OGT, 5142 ICmpInst::ICMP_SGT, "vcgtz"); 5143 case NEON::BI__builtin_neon_vcltz_v: 5144 case NEON::BI__builtin_neon_vcltzq_v: 5145 return EmitAArch64CompareBuiltinExpr(Ops[0], Ty, ICmpInst::FCMP_OLT, 5146 ICmpInst::ICMP_SLT, "vcltz"); 5147 case NEON::BI__builtin_neon_vcvt_f64_v: 5148 case NEON::BI__builtin_neon_vcvtq_f64_v: 5149 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 5150 Ty = GetNeonType(this, NeonTypeFlags(NeonTypeFlags::Float64, false, quad)); 5151 return usgn ? Builder.CreateUIToFP(Ops[0], Ty, "vcvt") 5152 : Builder.CreateSIToFP(Ops[0], Ty, "vcvt"); 5153 case NEON::BI__builtin_neon_vcvt_f64_f32: { 5154 assert(Type.getEltType() == NeonTypeFlags::Float64 && quad && 5155 "unexpected vcvt_f64_f32 builtin"); 5156 NeonTypeFlags SrcFlag = NeonTypeFlags(NeonTypeFlags::Float32, false, false); 5157 Ops[0] = Builder.CreateBitCast(Ops[0], GetNeonType(this, SrcFlag)); 5158 5159 return Builder.CreateFPExt(Ops[0], Ty, "vcvt"); 5160 } 5161 case NEON::BI__builtin_neon_vcvt_f32_f64: { 5162 assert(Type.getEltType() == NeonTypeFlags::Float32 && 5163 "unexpected vcvt_f32_f64 builtin"); 5164 NeonTypeFlags SrcFlag = NeonTypeFlags(NeonTypeFlags::Float64, false, true); 5165 Ops[0] = Builder.CreateBitCast(Ops[0], GetNeonType(this, SrcFlag)); 5166 5167 return Builder.CreateFPTrunc(Ops[0], Ty, "vcvt"); 5168 } 5169 case NEON::BI__builtin_neon_vcvt_s32_v: 5170 case NEON::BI__builtin_neon_vcvt_u32_v: 5171 case NEON::BI__builtin_neon_vcvt_s64_v: 5172 case NEON::BI__builtin_neon_vcvt_u64_v: 5173 case NEON::BI__builtin_neon_vcvtq_s32_v: 5174 case NEON::BI__builtin_neon_vcvtq_u32_v: 5175 case NEON::BI__builtin_neon_vcvtq_s64_v: 5176 case NEON::BI__builtin_neon_vcvtq_u64_v: { 5177 bool Double = 5178 (cast<llvm::IntegerType>(VTy->getElementType())->getBitWidth() == 64); 5179 llvm::Type *InTy = 5180 GetNeonType(this, 5181 NeonTypeFlags(Double ? NeonTypeFlags::Float64 5182 : NeonTypeFlags::Float32, false, quad)); 5183 Ops[0] = Builder.CreateBitCast(Ops[0], InTy); 5184 if (usgn) 5185 return Builder.CreateFPToUI(Ops[0], Ty); 5186 return Builder.CreateFPToSI(Ops[0], Ty); 5187 } 5188 case NEON::BI__builtin_neon_vcvta_s32_v: 5189 case NEON::BI__builtin_neon_vcvtaq_s32_v: 5190 case NEON::BI__builtin_neon_vcvta_u32_v: 5191 case NEON::BI__builtin_neon_vcvtaq_u32_v: 5192 case NEON::BI__builtin_neon_vcvta_s64_v: 5193 case NEON::BI__builtin_neon_vcvtaq_s64_v: 5194 case NEON::BI__builtin_neon_vcvta_u64_v: 5195 case NEON::BI__builtin_neon_vcvtaq_u64_v: { 5196 Int = usgn ? Intrinsic::aarch64_neon_fcvtau : Intrinsic::aarch64_neon_fcvtas; 5197 bool Double = 5198 (cast<llvm::IntegerType>(VTy->getElementType())->getBitWidth() == 64); 5199 llvm::Type *InTy = 5200 GetNeonType(this, 5201 NeonTypeFlags(Double ? NeonTypeFlags::Float64 5202 : NeonTypeFlags::Float32, false, quad)); 5203 llvm::Type *Tys[2] = { Ty, InTy }; 5204 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vcvta"); 5205 } 5206 case NEON::BI__builtin_neon_vcvtm_s32_v: 5207 case NEON::BI__builtin_neon_vcvtmq_s32_v: 5208 case NEON::BI__builtin_neon_vcvtm_u32_v: 5209 case NEON::BI__builtin_neon_vcvtmq_u32_v: 5210 case NEON::BI__builtin_neon_vcvtm_s64_v: 5211 case NEON::BI__builtin_neon_vcvtmq_s64_v: 5212 case NEON::BI__builtin_neon_vcvtm_u64_v: 5213 case NEON::BI__builtin_neon_vcvtmq_u64_v: { 5214 Int = usgn ? Intrinsic::aarch64_neon_fcvtmu : Intrinsic::aarch64_neon_fcvtms; 5215 bool Double = 5216 (cast<llvm::IntegerType>(VTy->getElementType())->getBitWidth() == 64); 5217 llvm::Type *InTy = 5218 GetNeonType(this, 5219 NeonTypeFlags(Double ? NeonTypeFlags::Float64 5220 : NeonTypeFlags::Float32, false, quad)); 5221 llvm::Type *Tys[2] = { Ty, InTy }; 5222 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vcvtm"); 5223 } 5224 case NEON::BI__builtin_neon_vcvtn_s32_v: 5225 case NEON::BI__builtin_neon_vcvtnq_s32_v: 5226 case NEON::BI__builtin_neon_vcvtn_u32_v: 5227 case NEON::BI__builtin_neon_vcvtnq_u32_v: 5228 case NEON::BI__builtin_neon_vcvtn_s64_v: 5229 case NEON::BI__builtin_neon_vcvtnq_s64_v: 5230 case NEON::BI__builtin_neon_vcvtn_u64_v: 5231 case NEON::BI__builtin_neon_vcvtnq_u64_v: { 5232 Int = usgn ? Intrinsic::aarch64_neon_fcvtnu : Intrinsic::aarch64_neon_fcvtns; 5233 bool Double = 5234 (cast<llvm::IntegerType>(VTy->getElementType())->getBitWidth() == 64); 5235 llvm::Type *InTy = 5236 GetNeonType(this, 5237 NeonTypeFlags(Double ? NeonTypeFlags::Float64 5238 : NeonTypeFlags::Float32, false, quad)); 5239 llvm::Type *Tys[2] = { Ty, InTy }; 5240 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vcvtn"); 5241 } 5242 case NEON::BI__builtin_neon_vcvtp_s32_v: 5243 case NEON::BI__builtin_neon_vcvtpq_s32_v: 5244 case NEON::BI__builtin_neon_vcvtp_u32_v: 5245 case NEON::BI__builtin_neon_vcvtpq_u32_v: 5246 case NEON::BI__builtin_neon_vcvtp_s64_v: 5247 case NEON::BI__builtin_neon_vcvtpq_s64_v: 5248 case NEON::BI__builtin_neon_vcvtp_u64_v: 5249 case NEON::BI__builtin_neon_vcvtpq_u64_v: { 5250 Int = usgn ? Intrinsic::aarch64_neon_fcvtpu : Intrinsic::aarch64_neon_fcvtps; 5251 bool Double = 5252 (cast<llvm::IntegerType>(VTy->getElementType())->getBitWidth() == 64); 5253 llvm::Type *InTy = 5254 GetNeonType(this, 5255 NeonTypeFlags(Double ? NeonTypeFlags::Float64 5256 : NeonTypeFlags::Float32, false, quad)); 5257 llvm::Type *Tys[2] = { Ty, InTy }; 5258 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vcvtp"); 5259 } 5260 case NEON::BI__builtin_neon_vmulx_v: 5261 case NEON::BI__builtin_neon_vmulxq_v: { 5262 Int = Intrinsic::aarch64_neon_fmulx; 5263 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmulx"); 5264 } 5265 case NEON::BI__builtin_neon_vmul_lane_v: 5266 case NEON::BI__builtin_neon_vmul_laneq_v: { 5267 // v1f64 vmul_lane should be mapped to Neon scalar mul lane 5268 bool Quad = false; 5269 if (BuiltinID == NEON::BI__builtin_neon_vmul_laneq_v) 5270 Quad = true; 5271 Ops[0] = Builder.CreateBitCast(Ops[0], DoubleTy); 5272 llvm::Type *VTy = GetNeonType(this, 5273 NeonTypeFlags(NeonTypeFlags::Float64, false, Quad)); 5274 Ops[1] = Builder.CreateBitCast(Ops[1], VTy); 5275 Ops[1] = Builder.CreateExtractElement(Ops[1], Ops[2], "extract"); 5276 Value *Result = Builder.CreateFMul(Ops[0], Ops[1]); 5277 return Builder.CreateBitCast(Result, Ty); 5278 } 5279 case NEON::BI__builtin_neon_vnegd_s64: 5280 return Builder.CreateNeg(EmitScalarExpr(E->getArg(0)), "vnegd"); 5281 case NEON::BI__builtin_neon_vpmaxnm_v: 5282 case NEON::BI__builtin_neon_vpmaxnmq_v: { 5283 Int = Intrinsic::aarch64_neon_fmaxnmp; 5284 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vpmaxnm"); 5285 } 5286 case NEON::BI__builtin_neon_vpminnm_v: 5287 case NEON::BI__builtin_neon_vpminnmq_v: { 5288 Int = Intrinsic::aarch64_neon_fminnmp; 5289 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vpminnm"); 5290 } 5291 case NEON::BI__builtin_neon_vsqrt_v: 5292 case NEON::BI__builtin_neon_vsqrtq_v: { 5293 Int = Intrinsic::sqrt; 5294 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 5295 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vsqrt"); 5296 } 5297 case NEON::BI__builtin_neon_vrbit_v: 5298 case NEON::BI__builtin_neon_vrbitq_v: { 5299 Int = Intrinsic::aarch64_neon_rbit; 5300 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrbit"); 5301 } 5302 case NEON::BI__builtin_neon_vaddv_u8: 5303 // FIXME: These are handled by the AArch64 scalar code. 5304 usgn = true; 5305 // FALLTHROUGH 5306 case NEON::BI__builtin_neon_vaddv_s8: { 5307 Int = usgn ? Intrinsic::aarch64_neon_uaddv : Intrinsic::aarch64_neon_saddv; 5308 Ty = Int32Ty; 5309 VTy = llvm::VectorType::get(Int8Ty, 8); 5310 llvm::Type *Tys[2] = { Ty, VTy }; 5311 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5312 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddv"); 5313 return Builder.CreateTrunc(Ops[0], Int8Ty); 5314 } 5315 case NEON::BI__builtin_neon_vaddv_u16: 5316 usgn = true; 5317 // FALLTHROUGH 5318 case NEON::BI__builtin_neon_vaddv_s16: { 5319 Int = usgn ? Intrinsic::aarch64_neon_uaddv : Intrinsic::aarch64_neon_saddv; 5320 Ty = Int32Ty; 5321 VTy = llvm::VectorType::get(Int16Ty, 4); 5322 llvm::Type *Tys[2] = { Ty, VTy }; 5323 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5324 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddv"); 5325 return Builder.CreateTrunc(Ops[0], Int16Ty); 5326 } 5327 case NEON::BI__builtin_neon_vaddvq_u8: 5328 usgn = true; 5329 // FALLTHROUGH 5330 case NEON::BI__builtin_neon_vaddvq_s8: { 5331 Int = usgn ? Intrinsic::aarch64_neon_uaddv : Intrinsic::aarch64_neon_saddv; 5332 Ty = Int32Ty; 5333 VTy = llvm::VectorType::get(Int8Ty, 16); 5334 llvm::Type *Tys[2] = { Ty, VTy }; 5335 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5336 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddv"); 5337 return Builder.CreateTrunc(Ops[0], Int8Ty); 5338 } 5339 case NEON::BI__builtin_neon_vaddvq_u16: 5340 usgn = true; 5341 // FALLTHROUGH 5342 case NEON::BI__builtin_neon_vaddvq_s16: { 5343 Int = usgn ? Intrinsic::aarch64_neon_uaddv : Intrinsic::aarch64_neon_saddv; 5344 Ty = Int32Ty; 5345 VTy = llvm::VectorType::get(Int16Ty, 8); 5346 llvm::Type *Tys[2] = { Ty, VTy }; 5347 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5348 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddv"); 5349 return Builder.CreateTrunc(Ops[0], Int16Ty); 5350 } 5351 case NEON::BI__builtin_neon_vmaxv_u8: { 5352 Int = Intrinsic::aarch64_neon_umaxv; 5353 Ty = Int32Ty; 5354 VTy = llvm::VectorType::get(Int8Ty, 8); 5355 llvm::Type *Tys[2] = { Ty, VTy }; 5356 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5357 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv"); 5358 return Builder.CreateTrunc(Ops[0], Int8Ty); 5359 } 5360 case NEON::BI__builtin_neon_vmaxv_u16: { 5361 Int = Intrinsic::aarch64_neon_umaxv; 5362 Ty = Int32Ty; 5363 VTy = llvm::VectorType::get(Int16Ty, 4); 5364 llvm::Type *Tys[2] = { Ty, VTy }; 5365 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5366 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv"); 5367 return Builder.CreateTrunc(Ops[0], Int16Ty); 5368 } 5369 case NEON::BI__builtin_neon_vmaxvq_u8: { 5370 Int = Intrinsic::aarch64_neon_umaxv; 5371 Ty = Int32Ty; 5372 VTy = llvm::VectorType::get(Int8Ty, 16); 5373 llvm::Type *Tys[2] = { Ty, VTy }; 5374 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5375 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv"); 5376 return Builder.CreateTrunc(Ops[0], Int8Ty); 5377 } 5378 case NEON::BI__builtin_neon_vmaxvq_u16: { 5379 Int = Intrinsic::aarch64_neon_umaxv; 5380 Ty = Int32Ty; 5381 VTy = llvm::VectorType::get(Int16Ty, 8); 5382 llvm::Type *Tys[2] = { Ty, VTy }; 5383 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5384 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv"); 5385 return Builder.CreateTrunc(Ops[0], Int16Ty); 5386 } 5387 case NEON::BI__builtin_neon_vmaxv_s8: { 5388 Int = Intrinsic::aarch64_neon_smaxv; 5389 Ty = Int32Ty; 5390 VTy = llvm::VectorType::get(Int8Ty, 8); 5391 llvm::Type *Tys[2] = { Ty, VTy }; 5392 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5393 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv"); 5394 return Builder.CreateTrunc(Ops[0], Int8Ty); 5395 } 5396 case NEON::BI__builtin_neon_vmaxv_s16: { 5397 Int = Intrinsic::aarch64_neon_smaxv; 5398 Ty = Int32Ty; 5399 VTy = llvm::VectorType::get(Int16Ty, 4); 5400 llvm::Type *Tys[2] = { Ty, VTy }; 5401 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5402 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv"); 5403 return Builder.CreateTrunc(Ops[0], Int16Ty); 5404 } 5405 case NEON::BI__builtin_neon_vmaxvq_s8: { 5406 Int = Intrinsic::aarch64_neon_smaxv; 5407 Ty = Int32Ty; 5408 VTy = llvm::VectorType::get(Int8Ty, 16); 5409 llvm::Type *Tys[2] = { Ty, VTy }; 5410 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5411 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv"); 5412 return Builder.CreateTrunc(Ops[0], Int8Ty); 5413 } 5414 case NEON::BI__builtin_neon_vmaxvq_s16: { 5415 Int = Intrinsic::aarch64_neon_smaxv; 5416 Ty = Int32Ty; 5417 VTy = llvm::VectorType::get(Int16Ty, 8); 5418 llvm::Type *Tys[2] = { Ty, VTy }; 5419 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5420 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vmaxv"); 5421 return Builder.CreateTrunc(Ops[0], Int16Ty); 5422 } 5423 case NEON::BI__builtin_neon_vminv_u8: { 5424 Int = Intrinsic::aarch64_neon_uminv; 5425 Ty = Int32Ty; 5426 VTy = llvm::VectorType::get(Int8Ty, 8); 5427 llvm::Type *Tys[2] = { Ty, VTy }; 5428 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5429 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv"); 5430 return Builder.CreateTrunc(Ops[0], Int8Ty); 5431 } 5432 case NEON::BI__builtin_neon_vminv_u16: { 5433 Int = Intrinsic::aarch64_neon_uminv; 5434 Ty = Int32Ty; 5435 VTy = llvm::VectorType::get(Int16Ty, 4); 5436 llvm::Type *Tys[2] = { Ty, VTy }; 5437 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5438 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv"); 5439 return Builder.CreateTrunc(Ops[0], Int16Ty); 5440 } 5441 case NEON::BI__builtin_neon_vminvq_u8: { 5442 Int = Intrinsic::aarch64_neon_uminv; 5443 Ty = Int32Ty; 5444 VTy = llvm::VectorType::get(Int8Ty, 16); 5445 llvm::Type *Tys[2] = { Ty, VTy }; 5446 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5447 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv"); 5448 return Builder.CreateTrunc(Ops[0], Int8Ty); 5449 } 5450 case NEON::BI__builtin_neon_vminvq_u16: { 5451 Int = Intrinsic::aarch64_neon_uminv; 5452 Ty = Int32Ty; 5453 VTy = llvm::VectorType::get(Int16Ty, 8); 5454 llvm::Type *Tys[2] = { Ty, VTy }; 5455 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5456 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv"); 5457 return Builder.CreateTrunc(Ops[0], Int16Ty); 5458 } 5459 case NEON::BI__builtin_neon_vminv_s8: { 5460 Int = Intrinsic::aarch64_neon_sminv; 5461 Ty = Int32Ty; 5462 VTy = llvm::VectorType::get(Int8Ty, 8); 5463 llvm::Type *Tys[2] = { Ty, VTy }; 5464 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5465 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv"); 5466 return Builder.CreateTrunc(Ops[0], Int8Ty); 5467 } 5468 case NEON::BI__builtin_neon_vminv_s16: { 5469 Int = Intrinsic::aarch64_neon_sminv; 5470 Ty = Int32Ty; 5471 VTy = llvm::VectorType::get(Int16Ty, 4); 5472 llvm::Type *Tys[2] = { Ty, VTy }; 5473 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5474 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv"); 5475 return Builder.CreateTrunc(Ops[0], Int16Ty); 5476 } 5477 case NEON::BI__builtin_neon_vminvq_s8: { 5478 Int = Intrinsic::aarch64_neon_sminv; 5479 Ty = Int32Ty; 5480 VTy = llvm::VectorType::get(Int8Ty, 16); 5481 llvm::Type *Tys[2] = { Ty, VTy }; 5482 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5483 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv"); 5484 return Builder.CreateTrunc(Ops[0], Int8Ty); 5485 } 5486 case NEON::BI__builtin_neon_vminvq_s16: { 5487 Int = Intrinsic::aarch64_neon_sminv; 5488 Ty = Int32Ty; 5489 VTy = llvm::VectorType::get(Int16Ty, 8); 5490 llvm::Type *Tys[2] = { Ty, VTy }; 5491 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5492 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vminv"); 5493 return Builder.CreateTrunc(Ops[0], Int16Ty); 5494 } 5495 case NEON::BI__builtin_neon_vmul_n_f64: { 5496 Ops[0] = Builder.CreateBitCast(Ops[0], DoubleTy); 5497 Value *RHS = Builder.CreateBitCast(EmitScalarExpr(E->getArg(1)), DoubleTy); 5498 return Builder.CreateFMul(Ops[0], RHS); 5499 } 5500 case NEON::BI__builtin_neon_vaddlv_u8: { 5501 Int = Intrinsic::aarch64_neon_uaddlv; 5502 Ty = Int32Ty; 5503 VTy = llvm::VectorType::get(Int8Ty, 8); 5504 llvm::Type *Tys[2] = { Ty, VTy }; 5505 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5506 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv"); 5507 return Builder.CreateTrunc(Ops[0], Int16Ty); 5508 } 5509 case NEON::BI__builtin_neon_vaddlv_u16: { 5510 Int = Intrinsic::aarch64_neon_uaddlv; 5511 Ty = Int32Ty; 5512 VTy = llvm::VectorType::get(Int16Ty, 4); 5513 llvm::Type *Tys[2] = { Ty, VTy }; 5514 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5515 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv"); 5516 } 5517 case NEON::BI__builtin_neon_vaddlvq_u8: { 5518 Int = Intrinsic::aarch64_neon_uaddlv; 5519 Ty = Int32Ty; 5520 VTy = llvm::VectorType::get(Int8Ty, 16); 5521 llvm::Type *Tys[2] = { Ty, VTy }; 5522 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5523 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv"); 5524 return Builder.CreateTrunc(Ops[0], Int16Ty); 5525 } 5526 case NEON::BI__builtin_neon_vaddlvq_u16: { 5527 Int = Intrinsic::aarch64_neon_uaddlv; 5528 Ty = Int32Ty; 5529 VTy = llvm::VectorType::get(Int16Ty, 8); 5530 llvm::Type *Tys[2] = { Ty, VTy }; 5531 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5532 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv"); 5533 } 5534 case NEON::BI__builtin_neon_vaddlv_s8: { 5535 Int = Intrinsic::aarch64_neon_saddlv; 5536 Ty = Int32Ty; 5537 VTy = llvm::VectorType::get(Int8Ty, 8); 5538 llvm::Type *Tys[2] = { Ty, VTy }; 5539 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5540 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv"); 5541 return Builder.CreateTrunc(Ops[0], Int16Ty); 5542 } 5543 case NEON::BI__builtin_neon_vaddlv_s16: { 5544 Int = Intrinsic::aarch64_neon_saddlv; 5545 Ty = Int32Ty; 5546 VTy = llvm::VectorType::get(Int16Ty, 4); 5547 llvm::Type *Tys[2] = { Ty, VTy }; 5548 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5549 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv"); 5550 } 5551 case NEON::BI__builtin_neon_vaddlvq_s8: { 5552 Int = Intrinsic::aarch64_neon_saddlv; 5553 Ty = Int32Ty; 5554 VTy = llvm::VectorType::get(Int8Ty, 16); 5555 llvm::Type *Tys[2] = { Ty, VTy }; 5556 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5557 Ops[0] = EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv"); 5558 return Builder.CreateTrunc(Ops[0], Int16Ty); 5559 } 5560 case NEON::BI__builtin_neon_vaddlvq_s16: { 5561 Int = Intrinsic::aarch64_neon_saddlv; 5562 Ty = Int32Ty; 5563 VTy = llvm::VectorType::get(Int16Ty, 8); 5564 llvm::Type *Tys[2] = { Ty, VTy }; 5565 Ops.push_back(EmitScalarExpr(E->getArg(0))); 5566 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vaddlv"); 5567 } 5568 case NEON::BI__builtin_neon_vsri_n_v: 5569 case NEON::BI__builtin_neon_vsriq_n_v: { 5570 Int = Intrinsic::aarch64_neon_vsri; 5571 llvm::Function *Intrin = CGM.getIntrinsic(Int, Ty); 5572 return EmitNeonCall(Intrin, Ops, "vsri_n"); 5573 } 5574 case NEON::BI__builtin_neon_vsli_n_v: 5575 case NEON::BI__builtin_neon_vsliq_n_v: { 5576 Int = Intrinsic::aarch64_neon_vsli; 5577 llvm::Function *Intrin = CGM.getIntrinsic(Int, Ty); 5578 return EmitNeonCall(Intrin, Ops, "vsli_n"); 5579 } 5580 case NEON::BI__builtin_neon_vsra_n_v: 5581 case NEON::BI__builtin_neon_vsraq_n_v: 5582 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 5583 Ops[1] = EmitNeonRShiftImm(Ops[1], Ops[2], Ty, usgn, "vsra_n"); 5584 return Builder.CreateAdd(Ops[0], Ops[1]); 5585 case NEON::BI__builtin_neon_vrsra_n_v: 5586 case NEON::BI__builtin_neon_vrsraq_n_v: { 5587 Int = usgn ? Intrinsic::aarch64_neon_urshl : Intrinsic::aarch64_neon_srshl; 5588 SmallVector<llvm::Value*,2> TmpOps; 5589 TmpOps.push_back(Ops[1]); 5590 TmpOps.push_back(Ops[2]); 5591 Function* F = CGM.getIntrinsic(Int, Ty); 5592 llvm::Value *tmp = EmitNeonCall(F, TmpOps, "vrshr_n", 1, true); 5593 Ops[0] = Builder.CreateBitCast(Ops[0], VTy); 5594 return Builder.CreateAdd(Ops[0], tmp); 5595 } 5596 // FIXME: Sharing loads & stores with 32-bit is complicated by the absence 5597 // of an Align parameter here. 5598 case NEON::BI__builtin_neon_vld1_x2_v: 5599 case NEON::BI__builtin_neon_vld1q_x2_v: 5600 case NEON::BI__builtin_neon_vld1_x3_v: 5601 case NEON::BI__builtin_neon_vld1q_x3_v: 5602 case NEON::BI__builtin_neon_vld1_x4_v: 5603 case NEON::BI__builtin_neon_vld1q_x4_v: { 5604 llvm::Type *PTy = llvm::PointerType::getUnqual(VTy->getVectorElementType()); 5605 Ops[1] = Builder.CreateBitCast(Ops[1], PTy); 5606 llvm::Type *Tys[2] = { VTy, PTy }; 5607 unsigned Int; 5608 switch (BuiltinID) { 5609 case NEON::BI__builtin_neon_vld1_x2_v: 5610 case NEON::BI__builtin_neon_vld1q_x2_v: 5611 Int = Intrinsic::aarch64_neon_ld1x2; 5612 break; 5613 case NEON::BI__builtin_neon_vld1_x3_v: 5614 case NEON::BI__builtin_neon_vld1q_x3_v: 5615 Int = Intrinsic::aarch64_neon_ld1x3; 5616 break; 5617 case NEON::BI__builtin_neon_vld1_x4_v: 5618 case NEON::BI__builtin_neon_vld1q_x4_v: 5619 Int = Intrinsic::aarch64_neon_ld1x4; 5620 break; 5621 } 5622 Function *F = CGM.getIntrinsic(Int, Tys); 5623 Ops[1] = Builder.CreateCall(F, Ops[1], "vld1xN"); 5624 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 5625 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 5626 return Builder.CreateStore(Ops[1], Ops[0]); 5627 } 5628 case NEON::BI__builtin_neon_vst1_x2_v: 5629 case NEON::BI__builtin_neon_vst1q_x2_v: 5630 case NEON::BI__builtin_neon_vst1_x3_v: 5631 case NEON::BI__builtin_neon_vst1q_x3_v: 5632 case NEON::BI__builtin_neon_vst1_x4_v: 5633 case NEON::BI__builtin_neon_vst1q_x4_v: { 5634 llvm::Type *PTy = llvm::PointerType::getUnqual(VTy->getVectorElementType()); 5635 llvm::Type *Tys[2] = { VTy, PTy }; 5636 unsigned Int; 5637 switch (BuiltinID) { 5638 case NEON::BI__builtin_neon_vst1_x2_v: 5639 case NEON::BI__builtin_neon_vst1q_x2_v: 5640 Int = Intrinsic::aarch64_neon_st1x2; 5641 break; 5642 case NEON::BI__builtin_neon_vst1_x3_v: 5643 case NEON::BI__builtin_neon_vst1q_x3_v: 5644 Int = Intrinsic::aarch64_neon_st1x3; 5645 break; 5646 case NEON::BI__builtin_neon_vst1_x4_v: 5647 case NEON::BI__builtin_neon_vst1q_x4_v: 5648 Int = Intrinsic::aarch64_neon_st1x4; 5649 break; 5650 } 5651 std::rotate(Ops.begin(), Ops.begin() + 1, Ops.end()); 5652 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, ""); 5653 } 5654 case NEON::BI__builtin_neon_vld1_v: 5655 case NEON::BI__builtin_neon_vld1q_v: 5656 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(VTy)); 5657 return Builder.CreateLoad(Ops[0]); 5658 case NEON::BI__builtin_neon_vst1_v: 5659 case NEON::BI__builtin_neon_vst1q_v: 5660 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(VTy)); 5661 Ops[1] = Builder.CreateBitCast(Ops[1], VTy); 5662 return Builder.CreateStore(Ops[1], Ops[0]); 5663 case NEON::BI__builtin_neon_vld1_lane_v: 5664 case NEON::BI__builtin_neon_vld1q_lane_v: 5665 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 5666 Ty = llvm::PointerType::getUnqual(VTy->getElementType()); 5667 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 5668 Ops[0] = Builder.CreateLoad(Ops[0]); 5669 return Builder.CreateInsertElement(Ops[1], Ops[0], Ops[2], "vld1_lane"); 5670 case NEON::BI__builtin_neon_vld1_dup_v: 5671 case NEON::BI__builtin_neon_vld1q_dup_v: { 5672 Value *V = UndefValue::get(Ty); 5673 Ty = llvm::PointerType::getUnqual(VTy->getElementType()); 5674 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 5675 Ops[0] = Builder.CreateLoad(Ops[0]); 5676 llvm::Constant *CI = ConstantInt::get(Int32Ty, 0); 5677 Ops[0] = Builder.CreateInsertElement(V, Ops[0], CI); 5678 return EmitNeonSplat(Ops[0], CI); 5679 } 5680 case NEON::BI__builtin_neon_vst1_lane_v: 5681 case NEON::BI__builtin_neon_vst1q_lane_v: 5682 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 5683 Ops[1] = Builder.CreateExtractElement(Ops[1], Ops[2]); 5684 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 5685 return Builder.CreateStore(Ops[1], Builder.CreateBitCast(Ops[0], Ty)); 5686 case NEON::BI__builtin_neon_vld2_v: 5687 case NEON::BI__builtin_neon_vld2q_v: { 5688 llvm::Type *PTy = llvm::PointerType::getUnqual(VTy); 5689 Ops[1] = Builder.CreateBitCast(Ops[1], PTy); 5690 llvm::Type *Tys[2] = { VTy, PTy }; 5691 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld2, Tys); 5692 Ops[1] = Builder.CreateCall(F, Ops[1], "vld2"); 5693 Ops[0] = Builder.CreateBitCast(Ops[0], 5694 llvm::PointerType::getUnqual(Ops[1]->getType())); 5695 return Builder.CreateStore(Ops[1], Ops[0]); 5696 } 5697 case NEON::BI__builtin_neon_vld3_v: 5698 case NEON::BI__builtin_neon_vld3q_v: { 5699 llvm::Type *PTy = llvm::PointerType::getUnqual(VTy); 5700 Ops[1] = Builder.CreateBitCast(Ops[1], PTy); 5701 llvm::Type *Tys[2] = { VTy, PTy }; 5702 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld3, Tys); 5703 Ops[1] = Builder.CreateCall(F, Ops[1], "vld3"); 5704 Ops[0] = Builder.CreateBitCast(Ops[0], 5705 llvm::PointerType::getUnqual(Ops[1]->getType())); 5706 return Builder.CreateStore(Ops[1], Ops[0]); 5707 } 5708 case NEON::BI__builtin_neon_vld4_v: 5709 case NEON::BI__builtin_neon_vld4q_v: { 5710 llvm::Type *PTy = llvm::PointerType::getUnqual(VTy); 5711 Ops[1] = Builder.CreateBitCast(Ops[1], PTy); 5712 llvm::Type *Tys[2] = { VTy, PTy }; 5713 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld4, Tys); 5714 Ops[1] = Builder.CreateCall(F, Ops[1], "vld4"); 5715 Ops[0] = Builder.CreateBitCast(Ops[0], 5716 llvm::PointerType::getUnqual(Ops[1]->getType())); 5717 return Builder.CreateStore(Ops[1], Ops[0]); 5718 } 5719 case NEON::BI__builtin_neon_vld2_dup_v: 5720 case NEON::BI__builtin_neon_vld2q_dup_v: { 5721 llvm::Type *PTy = 5722 llvm::PointerType::getUnqual(VTy->getElementType()); 5723 Ops[1] = Builder.CreateBitCast(Ops[1], PTy); 5724 llvm::Type *Tys[2] = { VTy, PTy }; 5725 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld2r, Tys); 5726 Ops[1] = Builder.CreateCall(F, Ops[1], "vld2"); 5727 Ops[0] = Builder.CreateBitCast(Ops[0], 5728 llvm::PointerType::getUnqual(Ops[1]->getType())); 5729 return Builder.CreateStore(Ops[1], Ops[0]); 5730 } 5731 case NEON::BI__builtin_neon_vld3_dup_v: 5732 case NEON::BI__builtin_neon_vld3q_dup_v: { 5733 llvm::Type *PTy = 5734 llvm::PointerType::getUnqual(VTy->getElementType()); 5735 Ops[1] = Builder.CreateBitCast(Ops[1], PTy); 5736 llvm::Type *Tys[2] = { VTy, PTy }; 5737 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld3r, Tys); 5738 Ops[1] = Builder.CreateCall(F, Ops[1], "vld3"); 5739 Ops[0] = Builder.CreateBitCast(Ops[0], 5740 llvm::PointerType::getUnqual(Ops[1]->getType())); 5741 return Builder.CreateStore(Ops[1], Ops[0]); 5742 } 5743 case NEON::BI__builtin_neon_vld4_dup_v: 5744 case NEON::BI__builtin_neon_vld4q_dup_v: { 5745 llvm::Type *PTy = 5746 llvm::PointerType::getUnqual(VTy->getElementType()); 5747 Ops[1] = Builder.CreateBitCast(Ops[1], PTy); 5748 llvm::Type *Tys[2] = { VTy, PTy }; 5749 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld4r, Tys); 5750 Ops[1] = Builder.CreateCall(F, Ops[1], "vld4"); 5751 Ops[0] = Builder.CreateBitCast(Ops[0], 5752 llvm::PointerType::getUnqual(Ops[1]->getType())); 5753 return Builder.CreateStore(Ops[1], Ops[0]); 5754 } 5755 case NEON::BI__builtin_neon_vld2_lane_v: 5756 case NEON::BI__builtin_neon_vld2q_lane_v: { 5757 llvm::Type *Tys[2] = { VTy, Ops[1]->getType() }; 5758 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld2lane, Tys); 5759 Ops.push_back(Ops[1]); 5760 Ops.erase(Ops.begin()+1); 5761 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 5762 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 5763 Ops[3] = Builder.CreateZExt(Ops[3], Int64Ty); 5764 Ops[1] = Builder.CreateCall(F, makeArrayRef(Ops).slice(1), "vld2_lane"); 5765 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 5766 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 5767 return Builder.CreateStore(Ops[1], Ops[0]); 5768 } 5769 case NEON::BI__builtin_neon_vld3_lane_v: 5770 case NEON::BI__builtin_neon_vld3q_lane_v: { 5771 llvm::Type *Tys[2] = { VTy, Ops[1]->getType() }; 5772 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld3lane, Tys); 5773 Ops.push_back(Ops[1]); 5774 Ops.erase(Ops.begin()+1); 5775 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 5776 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 5777 Ops[3] = Builder.CreateBitCast(Ops[3], Ty); 5778 Ops[4] = Builder.CreateZExt(Ops[4], Int64Ty); 5779 Ops[1] = Builder.CreateCall(F, makeArrayRef(Ops).slice(1), "vld3_lane"); 5780 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 5781 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 5782 return Builder.CreateStore(Ops[1], Ops[0]); 5783 } 5784 case NEON::BI__builtin_neon_vld4_lane_v: 5785 case NEON::BI__builtin_neon_vld4q_lane_v: { 5786 llvm::Type *Tys[2] = { VTy, Ops[1]->getType() }; 5787 Function *F = CGM.getIntrinsic(Intrinsic::aarch64_neon_ld4lane, Tys); 5788 Ops.push_back(Ops[1]); 5789 Ops.erase(Ops.begin()+1); 5790 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 5791 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 5792 Ops[3] = Builder.CreateBitCast(Ops[3], Ty); 5793 Ops[4] = Builder.CreateBitCast(Ops[4], Ty); 5794 Ops[5] = Builder.CreateZExt(Ops[5], Int64Ty); 5795 Ops[1] = Builder.CreateCall(F, makeArrayRef(Ops).slice(1), "vld4_lane"); 5796 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 5797 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 5798 return Builder.CreateStore(Ops[1], Ops[0]); 5799 } 5800 case NEON::BI__builtin_neon_vst2_v: 5801 case NEON::BI__builtin_neon_vst2q_v: { 5802 Ops.push_back(Ops[0]); 5803 Ops.erase(Ops.begin()); 5804 llvm::Type *Tys[2] = { VTy, Ops[2]->getType() }; 5805 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_st2, Tys), 5806 Ops, ""); 5807 } 5808 case NEON::BI__builtin_neon_vst2_lane_v: 5809 case NEON::BI__builtin_neon_vst2q_lane_v: { 5810 Ops.push_back(Ops[0]); 5811 Ops.erase(Ops.begin()); 5812 Ops[2] = Builder.CreateZExt(Ops[2], Int64Ty); 5813 llvm::Type *Tys[2] = { VTy, Ops[3]->getType() }; 5814 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_st2lane, Tys), 5815 Ops, ""); 5816 } 5817 case NEON::BI__builtin_neon_vst3_v: 5818 case NEON::BI__builtin_neon_vst3q_v: { 5819 Ops.push_back(Ops[0]); 5820 Ops.erase(Ops.begin()); 5821 llvm::Type *Tys[2] = { VTy, Ops[3]->getType() }; 5822 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_st3, Tys), 5823 Ops, ""); 5824 } 5825 case NEON::BI__builtin_neon_vst3_lane_v: 5826 case NEON::BI__builtin_neon_vst3q_lane_v: { 5827 Ops.push_back(Ops[0]); 5828 Ops.erase(Ops.begin()); 5829 Ops[3] = Builder.CreateZExt(Ops[3], Int64Ty); 5830 llvm::Type *Tys[2] = { VTy, Ops[4]->getType() }; 5831 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_st3lane, Tys), 5832 Ops, ""); 5833 } 5834 case NEON::BI__builtin_neon_vst4_v: 5835 case NEON::BI__builtin_neon_vst4q_v: { 5836 Ops.push_back(Ops[0]); 5837 Ops.erase(Ops.begin()); 5838 llvm::Type *Tys[2] = { VTy, Ops[4]->getType() }; 5839 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_st4, Tys), 5840 Ops, ""); 5841 } 5842 case NEON::BI__builtin_neon_vst4_lane_v: 5843 case NEON::BI__builtin_neon_vst4q_lane_v: { 5844 Ops.push_back(Ops[0]); 5845 Ops.erase(Ops.begin()); 5846 Ops[4] = Builder.CreateZExt(Ops[4], Int64Ty); 5847 llvm::Type *Tys[2] = { VTy, Ops[5]->getType() }; 5848 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_st4lane, Tys), 5849 Ops, ""); 5850 } 5851 case NEON::BI__builtin_neon_vtrn_v: 5852 case NEON::BI__builtin_neon_vtrnq_v: { 5853 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty)); 5854 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 5855 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 5856 Value *SV = nullptr; 5857 5858 for (unsigned vi = 0; vi != 2; ++vi) { 5859 SmallVector<Constant*, 16> Indices; 5860 for (unsigned i = 0, e = VTy->getNumElements(); i != e; i += 2) { 5861 Indices.push_back(ConstantInt::get(Int32Ty, i+vi)); 5862 Indices.push_back(ConstantInt::get(Int32Ty, i+e+vi)); 5863 } 5864 Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ty, Ops[0], vi); 5865 SV = llvm::ConstantVector::get(Indices); 5866 SV = Builder.CreateShuffleVector(Ops[1], Ops[2], SV, "vtrn"); 5867 SV = Builder.CreateStore(SV, Addr); 5868 } 5869 return SV; 5870 } 5871 case NEON::BI__builtin_neon_vuzp_v: 5872 case NEON::BI__builtin_neon_vuzpq_v: { 5873 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty)); 5874 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 5875 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 5876 Value *SV = nullptr; 5877 5878 for (unsigned vi = 0; vi != 2; ++vi) { 5879 SmallVector<Constant*, 16> Indices; 5880 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) 5881 Indices.push_back(ConstantInt::get(Int32Ty, 2*i+vi)); 5882 5883 Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ty, Ops[0], vi); 5884 SV = llvm::ConstantVector::get(Indices); 5885 SV = Builder.CreateShuffleVector(Ops[1], Ops[2], SV, "vuzp"); 5886 SV = Builder.CreateStore(SV, Addr); 5887 } 5888 return SV; 5889 } 5890 case NEON::BI__builtin_neon_vzip_v: 5891 case NEON::BI__builtin_neon_vzipq_v: { 5892 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty)); 5893 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 5894 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 5895 Value *SV = nullptr; 5896 5897 for (unsigned vi = 0; vi != 2; ++vi) { 5898 SmallVector<Constant*, 16> Indices; 5899 for (unsigned i = 0, e = VTy->getNumElements(); i != e; i += 2) { 5900 Indices.push_back(ConstantInt::get(Int32Ty, (i + vi*e) >> 1)); 5901 Indices.push_back(ConstantInt::get(Int32Ty, ((i + vi*e) >> 1)+e)); 5902 } 5903 Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ty, Ops[0], vi); 5904 SV = llvm::ConstantVector::get(Indices); 5905 SV = Builder.CreateShuffleVector(Ops[1], Ops[2], SV, "vzip"); 5906 SV = Builder.CreateStore(SV, Addr); 5907 } 5908 return SV; 5909 } 5910 case NEON::BI__builtin_neon_vqtbl1q_v: { 5911 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbl1, Ty), 5912 Ops, "vtbl1"); 5913 } 5914 case NEON::BI__builtin_neon_vqtbl2q_v: { 5915 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbl2, Ty), 5916 Ops, "vtbl2"); 5917 } 5918 case NEON::BI__builtin_neon_vqtbl3q_v: { 5919 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbl3, Ty), 5920 Ops, "vtbl3"); 5921 } 5922 case NEON::BI__builtin_neon_vqtbl4q_v: { 5923 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbl4, Ty), 5924 Ops, "vtbl4"); 5925 } 5926 case NEON::BI__builtin_neon_vqtbx1q_v: { 5927 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbx1, Ty), 5928 Ops, "vtbx1"); 5929 } 5930 case NEON::BI__builtin_neon_vqtbx2q_v: { 5931 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbx2, Ty), 5932 Ops, "vtbx2"); 5933 } 5934 case NEON::BI__builtin_neon_vqtbx3q_v: { 5935 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbx3, Ty), 5936 Ops, "vtbx3"); 5937 } 5938 case NEON::BI__builtin_neon_vqtbx4q_v: { 5939 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::aarch64_neon_tbx4, Ty), 5940 Ops, "vtbx4"); 5941 } 5942 case NEON::BI__builtin_neon_vsqadd_v: 5943 case NEON::BI__builtin_neon_vsqaddq_v: { 5944 Int = Intrinsic::aarch64_neon_usqadd; 5945 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vsqadd"); 5946 } 5947 case NEON::BI__builtin_neon_vuqadd_v: 5948 case NEON::BI__builtin_neon_vuqaddq_v: { 5949 Int = Intrinsic::aarch64_neon_suqadd; 5950 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vuqadd"); 5951 } 5952 } 5953 } 5954 5955 llvm::Value *CodeGenFunction:: 5956 BuildVector(ArrayRef<llvm::Value*> Ops) { 5957 assert((Ops.size() & (Ops.size() - 1)) == 0 && 5958 "Not a power-of-two sized vector!"); 5959 bool AllConstants = true; 5960 for (unsigned i = 0, e = Ops.size(); i != e && AllConstants; ++i) 5961 AllConstants &= isa<Constant>(Ops[i]); 5962 5963 // If this is a constant vector, create a ConstantVector. 5964 if (AllConstants) { 5965 SmallVector<llvm::Constant*, 16> CstOps; 5966 for (unsigned i = 0, e = Ops.size(); i != e; ++i) 5967 CstOps.push_back(cast<Constant>(Ops[i])); 5968 return llvm::ConstantVector::get(CstOps); 5969 } 5970 5971 // Otherwise, insertelement the values to build the vector. 5972 Value *Result = 5973 llvm::UndefValue::get(llvm::VectorType::get(Ops[0]->getType(), Ops.size())); 5974 5975 for (unsigned i = 0, e = Ops.size(); i != e; ++i) 5976 Result = Builder.CreateInsertElement(Result, Ops[i], Builder.getInt32(i)); 5977 5978 return Result; 5979 } 5980 5981 Value *CodeGenFunction::EmitX86BuiltinExpr(unsigned BuiltinID, 5982 const CallExpr *E) { 5983 SmallVector<Value*, 4> Ops; 5984 5985 // Find out if any arguments are required to be integer constant expressions. 5986 unsigned ICEArguments = 0; 5987 ASTContext::GetBuiltinTypeError Error; 5988 getContext().GetBuiltinType(BuiltinID, Error, &ICEArguments); 5989 assert(Error == ASTContext::GE_None && "Should not codegen an error"); 5990 5991 for (unsigned i = 0, e = E->getNumArgs(); i != e; i++) { 5992 // If this is a normal argument, just emit it as a scalar. 5993 if ((ICEArguments & (1 << i)) == 0) { 5994 Ops.push_back(EmitScalarExpr(E->getArg(i))); 5995 continue; 5996 } 5997 5998 // If this is required to be a constant, constant fold it so that we know 5999 // that the generated intrinsic gets a ConstantInt. 6000 llvm::APSInt Result; 6001 bool IsConst = E->getArg(i)->isIntegerConstantExpr(Result, getContext()); 6002 assert(IsConst && "Constant arg isn't actually constant?"); (void)IsConst; 6003 Ops.push_back(llvm::ConstantInt::get(getLLVMContext(), Result)); 6004 } 6005 6006 switch (BuiltinID) { 6007 default: return nullptr; 6008 case X86::BI__builtin_cpu_supports: { 6009 const Expr *FeatureExpr = E->getArg(0)->IgnoreParenCasts(); 6010 StringRef FeatureStr = cast<StringLiteral>(FeatureExpr)->getString(); 6011 6012 // TODO: When/if this becomes more than x86 specific then use a TargetInfo 6013 // based mapping. 6014 // Processor features and mapping to processor feature value. 6015 enum X86Features { 6016 CMOV = 0, 6017 MMX, 6018 POPCNT, 6019 SSE, 6020 SSE2, 6021 SSE3, 6022 SSSE3, 6023 SSE4_1, 6024 SSE4_2, 6025 AVX, 6026 AVX2, 6027 SSE4_A, 6028 FMA4, 6029 XOP, 6030 FMA, 6031 AVX512F, 6032 BMI, 6033 BMI2, 6034 MAX 6035 }; 6036 6037 X86Features Feature = StringSwitch<X86Features>(FeatureStr) 6038 .Case("cmov", X86Features::CMOV) 6039 .Case("mmx", X86Features::MMX) 6040 .Case("popcnt", X86Features::POPCNT) 6041 .Case("sse", X86Features::SSE) 6042 .Case("sse2", X86Features::SSE2) 6043 .Case("sse3", X86Features::SSE3) 6044 .Case("sse4.1", X86Features::SSE4_1) 6045 .Case("sse4.2", X86Features::SSE4_2) 6046 .Case("avx", X86Features::AVX) 6047 .Case("avx2", X86Features::AVX2) 6048 .Case("sse4a", X86Features::SSE4_A) 6049 .Case("fma4", X86Features::FMA4) 6050 .Case("xop", X86Features::XOP) 6051 .Case("fma", X86Features::FMA) 6052 .Case("avx512f", X86Features::AVX512F) 6053 .Case("bmi", X86Features::BMI) 6054 .Case("bmi2", X86Features::BMI2) 6055 .Default(X86Features::MAX); 6056 assert(Feature != X86Features::MAX && "Invalid feature!"); 6057 6058 // Matching the struct layout from the compiler-rt/libgcc structure that is 6059 // filled in: 6060 // unsigned int __cpu_vendor; 6061 // unsigned int __cpu_type; 6062 // unsigned int __cpu_subtype; 6063 // unsigned int __cpu_features[1]; 6064 llvm::Type *STy = llvm::StructType::get( 6065 Int32Ty, Int32Ty, Int32Ty, llvm::ArrayType::get(Int32Ty, 1), nullptr); 6066 6067 // Grab the global __cpu_model. 6068 llvm::Constant *CpuModel = CGM.CreateRuntimeVariable(STy, "__cpu_model"); 6069 6070 // Grab the first (0th) element from the field __cpu_features off of the 6071 // global in the struct STy. 6072 Value *Idxs[] = { 6073 ConstantInt::get(Int32Ty, 0), 6074 ConstantInt::get(Int32Ty, 3), 6075 ConstantInt::get(Int32Ty, 0) 6076 }; 6077 Value *CpuFeatures = Builder.CreateGEP(STy, CpuModel, Idxs); 6078 Value *Features = Builder.CreateLoad(CpuFeatures); 6079 6080 // Check the value of the bit corresponding to the feature requested. 6081 Value *Bitset = Builder.CreateAnd( 6082 Features, llvm::ConstantInt::get(Int32Ty, 1 << Feature)); 6083 return Builder.CreateICmpNE(Bitset, llvm::ConstantInt::get(Int32Ty, 0)); 6084 } 6085 case X86::BI_mm_prefetch: { 6086 Value *Address = EmitScalarExpr(E->getArg(0)); 6087 Value *RW = ConstantInt::get(Int32Ty, 0); 6088 Value *Locality = EmitScalarExpr(E->getArg(1)); 6089 Value *Data = ConstantInt::get(Int32Ty, 1); 6090 Value *F = CGM.getIntrinsic(Intrinsic::prefetch); 6091 return Builder.CreateCall(F, {Address, RW, Locality, Data}); 6092 } 6093 case X86::BI__builtin_ia32_vec_init_v8qi: 6094 case X86::BI__builtin_ia32_vec_init_v4hi: 6095 case X86::BI__builtin_ia32_vec_init_v2si: 6096 return Builder.CreateBitCast(BuildVector(Ops), 6097 llvm::Type::getX86_MMXTy(getLLVMContext())); 6098 case X86::BI__builtin_ia32_vec_ext_v2si: 6099 return Builder.CreateExtractElement(Ops[0], 6100 llvm::ConstantInt::get(Ops[1]->getType(), 0)); 6101 case X86::BI__builtin_ia32_ldmxcsr: { 6102 Value *Tmp = CreateMemTemp(E->getArg(0)->getType()); 6103 Builder.CreateStore(Ops[0], Tmp); 6104 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::x86_sse_ldmxcsr), 6105 Builder.CreateBitCast(Tmp, Int8PtrTy)); 6106 } 6107 case X86::BI__builtin_ia32_stmxcsr: { 6108 Value *Tmp = CreateMemTemp(E->getType()); 6109 Builder.CreateCall(CGM.getIntrinsic(Intrinsic::x86_sse_stmxcsr), 6110 Builder.CreateBitCast(Tmp, Int8PtrTy)); 6111 return Builder.CreateLoad(Tmp, "stmxcsr"); 6112 } 6113 case X86::BI__builtin_ia32_storehps: 6114 case X86::BI__builtin_ia32_storelps: { 6115 llvm::Type *PtrTy = llvm::PointerType::getUnqual(Int64Ty); 6116 llvm::Type *VecTy = llvm::VectorType::get(Int64Ty, 2); 6117 6118 // cast val v2i64 6119 Ops[1] = Builder.CreateBitCast(Ops[1], VecTy, "cast"); 6120 6121 // extract (0, 1) 6122 unsigned Index = BuiltinID == X86::BI__builtin_ia32_storelps ? 0 : 1; 6123 llvm::Value *Idx = llvm::ConstantInt::get(SizeTy, Index); 6124 Ops[1] = Builder.CreateExtractElement(Ops[1], Idx, "extract"); 6125 6126 // cast pointer to i64 & store 6127 Ops[0] = Builder.CreateBitCast(Ops[0], PtrTy); 6128 return Builder.CreateStore(Ops[1], Ops[0]); 6129 } 6130 case X86::BI__builtin_ia32_palignr128: 6131 case X86::BI__builtin_ia32_palignr256: { 6132 unsigned ShiftVal = cast<llvm::ConstantInt>(Ops[2])->getZExtValue(); 6133 6134 unsigned NumElts = 6135 cast<llvm::VectorType>(Ops[0]->getType())->getNumElements(); 6136 assert(NumElts % 16 == 0); 6137 unsigned NumLanes = NumElts / 16; 6138 unsigned NumLaneElts = NumElts / NumLanes; 6139 6140 // If palignr is shifting the pair of vectors more than the size of two 6141 // lanes, emit zero. 6142 if (ShiftVal >= (2 * NumLaneElts)) 6143 return llvm::Constant::getNullValue(ConvertType(E->getType())); 6144 6145 // If palignr is shifting the pair of input vectors more than one lane, 6146 // but less than two lanes, convert to shifting in zeroes. 6147 if (ShiftVal > NumLaneElts) { 6148 ShiftVal -= NumLaneElts; 6149 Ops[1] = Ops[0]; 6150 Ops[0] = llvm::Constant::getNullValue(Ops[0]->getType()); 6151 } 6152 6153 uint32_t Indices[32]; 6154 // 256-bit palignr operates on 128-bit lanes so we need to handle that 6155 for (unsigned l = 0; l != NumElts; l += NumLaneElts) { 6156 for (unsigned i = 0; i != NumLaneElts; ++i) { 6157 unsigned Idx = ShiftVal + i; 6158 if (Idx >= NumLaneElts) 6159 Idx += NumElts - NumLaneElts; // End of lane, switch operand. 6160 Indices[l + i] = Idx + l; 6161 } 6162 } 6163 6164 Value *SV = llvm::ConstantDataVector::get(getLLVMContext(), 6165 makeArrayRef(Indices, NumElts)); 6166 return Builder.CreateShuffleVector(Ops[1], Ops[0], SV, "palignr"); 6167 } 6168 case X86::BI__builtin_ia32_pslldqi256: { 6169 // Shift value is in bits so divide by 8. 6170 unsigned shiftVal = cast<llvm::ConstantInt>(Ops[1])->getZExtValue() >> 3; 6171 6172 // If pslldq is shifting the vector more than 15 bytes, emit zero. 6173 if (shiftVal >= 16) 6174 return llvm::Constant::getNullValue(ConvertType(E->getType())); 6175 6176 uint32_t Indices[32]; 6177 // 256-bit pslldq operates on 128-bit lanes so we need to handle that 6178 for (unsigned l = 0; l != 32; l += 16) { 6179 for (unsigned i = 0; i != 16; ++i) { 6180 unsigned Idx = 32 + i - shiftVal; 6181 if (Idx < 32) Idx -= 16; // end of lane, switch operand. 6182 Indices[l + i] = Idx + l; 6183 } 6184 } 6185 6186 llvm::Type *VecTy = llvm::VectorType::get(Int8Ty, 32); 6187 Ops[0] = Builder.CreateBitCast(Ops[0], VecTy, "cast"); 6188 Value *Zero = llvm::Constant::getNullValue(VecTy); 6189 6190 Value *SV = llvm::ConstantDataVector::get(getLLVMContext(), Indices); 6191 SV = Builder.CreateShuffleVector(Zero, Ops[0], SV, "pslldq"); 6192 llvm::Type *ResultType = ConvertType(E->getType()); 6193 return Builder.CreateBitCast(SV, ResultType, "cast"); 6194 } 6195 case X86::BI__builtin_ia32_psrldqi256: { 6196 // Shift value is in bits so divide by 8. 6197 unsigned shiftVal = cast<llvm::ConstantInt>(Ops[1])->getZExtValue() >> 3; 6198 6199 // If psrldq is shifting the vector more than 15 bytes, emit zero. 6200 if (shiftVal >= 16) 6201 return llvm::Constant::getNullValue(ConvertType(E->getType())); 6202 6203 uint32_t Indices[32]; 6204 // 256-bit psrldq operates on 128-bit lanes so we need to handle that 6205 for (unsigned l = 0; l != 32; l += 16) { 6206 for (unsigned i = 0; i != 16; ++i) { 6207 unsigned Idx = i + shiftVal; 6208 if (Idx >= 16) Idx += 16; // end of lane, switch operand. 6209 Indices[l + i] = Idx + l; 6210 } 6211 } 6212 6213 llvm::Type *VecTy = llvm::VectorType::get(Int8Ty, 32); 6214 Ops[0] = Builder.CreateBitCast(Ops[0], VecTy, "cast"); 6215 Value *Zero = llvm::Constant::getNullValue(VecTy); 6216 6217 Value *SV = llvm::ConstantDataVector::get(getLLVMContext(), Indices); 6218 SV = Builder.CreateShuffleVector(Ops[0], Zero, SV, "psrldq"); 6219 llvm::Type *ResultType = ConvertType(E->getType()); 6220 return Builder.CreateBitCast(SV, ResultType, "cast"); 6221 } 6222 case X86::BI__builtin_ia32_movntps: 6223 case X86::BI__builtin_ia32_movntps256: 6224 case X86::BI__builtin_ia32_movntpd: 6225 case X86::BI__builtin_ia32_movntpd256: 6226 case X86::BI__builtin_ia32_movntdq: 6227 case X86::BI__builtin_ia32_movntdq256: 6228 case X86::BI__builtin_ia32_movnti: 6229 case X86::BI__builtin_ia32_movnti64: { 6230 llvm::MDNode *Node = llvm::MDNode::get( 6231 getLLVMContext(), llvm::ConstantAsMetadata::get(Builder.getInt32(1))); 6232 6233 // Convert the type of the pointer to a pointer to the stored type. 6234 Value *BC = Builder.CreateBitCast(Ops[0], 6235 llvm::PointerType::getUnqual(Ops[1]->getType()), 6236 "cast"); 6237 StoreInst *SI = Builder.CreateStore(Ops[1], BC); 6238 SI->setMetadata(CGM.getModule().getMDKindID("nontemporal"), Node); 6239 6240 // If the operand is an integer, we can't assume alignment. Otherwise, 6241 // assume natural alignment. 6242 QualType ArgTy = E->getArg(1)->getType(); 6243 unsigned Align; 6244 if (ArgTy->isIntegerType()) 6245 Align = 1; 6246 else 6247 Align = getContext().getTypeSizeInChars(ArgTy).getQuantity(); 6248 SI->setAlignment(Align); 6249 return SI; 6250 } 6251 // 3DNow! 6252 case X86::BI__builtin_ia32_pswapdsf: 6253 case X86::BI__builtin_ia32_pswapdsi: { 6254 llvm::Type *MMXTy = llvm::Type::getX86_MMXTy(getLLVMContext()); 6255 Ops[0] = Builder.CreateBitCast(Ops[0], MMXTy, "cast"); 6256 llvm::Function *F = CGM.getIntrinsic(Intrinsic::x86_3dnowa_pswapd); 6257 return Builder.CreateCall(F, Ops, "pswapd"); 6258 } 6259 case X86::BI__builtin_ia32_rdrand16_step: 6260 case X86::BI__builtin_ia32_rdrand32_step: 6261 case X86::BI__builtin_ia32_rdrand64_step: 6262 case X86::BI__builtin_ia32_rdseed16_step: 6263 case X86::BI__builtin_ia32_rdseed32_step: 6264 case X86::BI__builtin_ia32_rdseed64_step: { 6265 Intrinsic::ID ID; 6266 switch (BuiltinID) { 6267 default: llvm_unreachable("Unsupported intrinsic!"); 6268 case X86::BI__builtin_ia32_rdrand16_step: 6269 ID = Intrinsic::x86_rdrand_16; 6270 break; 6271 case X86::BI__builtin_ia32_rdrand32_step: 6272 ID = Intrinsic::x86_rdrand_32; 6273 break; 6274 case X86::BI__builtin_ia32_rdrand64_step: 6275 ID = Intrinsic::x86_rdrand_64; 6276 break; 6277 case X86::BI__builtin_ia32_rdseed16_step: 6278 ID = Intrinsic::x86_rdseed_16; 6279 break; 6280 case X86::BI__builtin_ia32_rdseed32_step: 6281 ID = Intrinsic::x86_rdseed_32; 6282 break; 6283 case X86::BI__builtin_ia32_rdseed64_step: 6284 ID = Intrinsic::x86_rdseed_64; 6285 break; 6286 } 6287 6288 Value *Call = Builder.CreateCall(CGM.getIntrinsic(ID)); 6289 Builder.CreateStore(Builder.CreateExtractValue(Call, 0), Ops[0]); 6290 return Builder.CreateExtractValue(Call, 1); 6291 } 6292 // SSE comparison intrisics 6293 case X86::BI__builtin_ia32_cmpeqps: 6294 case X86::BI__builtin_ia32_cmpltps: 6295 case X86::BI__builtin_ia32_cmpleps: 6296 case X86::BI__builtin_ia32_cmpunordps: 6297 case X86::BI__builtin_ia32_cmpneqps: 6298 case X86::BI__builtin_ia32_cmpnltps: 6299 case X86::BI__builtin_ia32_cmpnleps: 6300 case X86::BI__builtin_ia32_cmpordps: 6301 case X86::BI__builtin_ia32_cmpeqss: 6302 case X86::BI__builtin_ia32_cmpltss: 6303 case X86::BI__builtin_ia32_cmpless: 6304 case X86::BI__builtin_ia32_cmpunordss: 6305 case X86::BI__builtin_ia32_cmpneqss: 6306 case X86::BI__builtin_ia32_cmpnltss: 6307 case X86::BI__builtin_ia32_cmpnless: 6308 case X86::BI__builtin_ia32_cmpordss: 6309 case X86::BI__builtin_ia32_cmpeqpd: 6310 case X86::BI__builtin_ia32_cmpltpd: 6311 case X86::BI__builtin_ia32_cmplepd: 6312 case X86::BI__builtin_ia32_cmpunordpd: 6313 case X86::BI__builtin_ia32_cmpneqpd: 6314 case X86::BI__builtin_ia32_cmpnltpd: 6315 case X86::BI__builtin_ia32_cmpnlepd: 6316 case X86::BI__builtin_ia32_cmpordpd: 6317 case X86::BI__builtin_ia32_cmpeqsd: 6318 case X86::BI__builtin_ia32_cmpltsd: 6319 case X86::BI__builtin_ia32_cmplesd: 6320 case X86::BI__builtin_ia32_cmpunordsd: 6321 case X86::BI__builtin_ia32_cmpneqsd: 6322 case X86::BI__builtin_ia32_cmpnltsd: 6323 case X86::BI__builtin_ia32_cmpnlesd: 6324 case X86::BI__builtin_ia32_cmpordsd: 6325 // These exist so that the builtin that takes an immediate can be bounds 6326 // checked by clang to avoid passing bad immediates to the backend. Since 6327 // AVX has a larger immediate than SSE we would need separate builtins to 6328 // do the different bounds checking. Rather than create a clang specific 6329 // SSE only builtin, this implements eight separate builtins to match gcc 6330 // implementation. 6331 6332 // Choose the immediate. 6333 unsigned Imm; 6334 switch (BuiltinID) { 6335 default: llvm_unreachable("Unsupported intrinsic!"); 6336 case X86::BI__builtin_ia32_cmpeqps: 6337 case X86::BI__builtin_ia32_cmpeqss: 6338 case X86::BI__builtin_ia32_cmpeqpd: 6339 case X86::BI__builtin_ia32_cmpeqsd: 6340 Imm = 0; 6341 break; 6342 case X86::BI__builtin_ia32_cmpltps: 6343 case X86::BI__builtin_ia32_cmpltss: 6344 case X86::BI__builtin_ia32_cmpltpd: 6345 case X86::BI__builtin_ia32_cmpltsd: 6346 Imm = 1; 6347 break; 6348 case X86::BI__builtin_ia32_cmpleps: 6349 case X86::BI__builtin_ia32_cmpless: 6350 case X86::BI__builtin_ia32_cmplepd: 6351 case X86::BI__builtin_ia32_cmplesd: 6352 Imm = 2; 6353 break; 6354 case X86::BI__builtin_ia32_cmpunordps: 6355 case X86::BI__builtin_ia32_cmpunordss: 6356 case X86::BI__builtin_ia32_cmpunordpd: 6357 case X86::BI__builtin_ia32_cmpunordsd: 6358 Imm = 3; 6359 break; 6360 case X86::BI__builtin_ia32_cmpneqps: 6361 case X86::BI__builtin_ia32_cmpneqss: 6362 case X86::BI__builtin_ia32_cmpneqpd: 6363 case X86::BI__builtin_ia32_cmpneqsd: 6364 Imm = 4; 6365 break; 6366 case X86::BI__builtin_ia32_cmpnltps: 6367 case X86::BI__builtin_ia32_cmpnltss: 6368 case X86::BI__builtin_ia32_cmpnltpd: 6369 case X86::BI__builtin_ia32_cmpnltsd: 6370 Imm = 5; 6371 break; 6372 case X86::BI__builtin_ia32_cmpnleps: 6373 case X86::BI__builtin_ia32_cmpnless: 6374 case X86::BI__builtin_ia32_cmpnlepd: 6375 case X86::BI__builtin_ia32_cmpnlesd: 6376 Imm = 6; 6377 break; 6378 case X86::BI__builtin_ia32_cmpordps: 6379 case X86::BI__builtin_ia32_cmpordss: 6380 case X86::BI__builtin_ia32_cmpordpd: 6381 case X86::BI__builtin_ia32_cmpordsd: 6382 Imm = 7; 6383 break; 6384 } 6385 6386 // Choose the intrinsic ID. 6387 const char *name; 6388 Intrinsic::ID ID; 6389 switch (BuiltinID) { 6390 default: llvm_unreachable("Unsupported intrinsic!"); 6391 case X86::BI__builtin_ia32_cmpeqps: 6392 case X86::BI__builtin_ia32_cmpltps: 6393 case X86::BI__builtin_ia32_cmpleps: 6394 case X86::BI__builtin_ia32_cmpunordps: 6395 case X86::BI__builtin_ia32_cmpneqps: 6396 case X86::BI__builtin_ia32_cmpnltps: 6397 case X86::BI__builtin_ia32_cmpnleps: 6398 case X86::BI__builtin_ia32_cmpordps: 6399 name = "cmpps"; 6400 ID = Intrinsic::x86_sse_cmp_ps; 6401 break; 6402 case X86::BI__builtin_ia32_cmpeqss: 6403 case X86::BI__builtin_ia32_cmpltss: 6404 case X86::BI__builtin_ia32_cmpless: 6405 case X86::BI__builtin_ia32_cmpunordss: 6406 case X86::BI__builtin_ia32_cmpneqss: 6407 case X86::BI__builtin_ia32_cmpnltss: 6408 case X86::BI__builtin_ia32_cmpnless: 6409 case X86::BI__builtin_ia32_cmpordss: 6410 name = "cmpss"; 6411 ID = Intrinsic::x86_sse_cmp_ss; 6412 break; 6413 case X86::BI__builtin_ia32_cmpeqpd: 6414 case X86::BI__builtin_ia32_cmpltpd: 6415 case X86::BI__builtin_ia32_cmplepd: 6416 case X86::BI__builtin_ia32_cmpunordpd: 6417 case X86::BI__builtin_ia32_cmpneqpd: 6418 case X86::BI__builtin_ia32_cmpnltpd: 6419 case X86::BI__builtin_ia32_cmpnlepd: 6420 case X86::BI__builtin_ia32_cmpordpd: 6421 name = "cmppd"; 6422 ID = Intrinsic::x86_sse2_cmp_pd; 6423 break; 6424 case X86::BI__builtin_ia32_cmpeqsd: 6425 case X86::BI__builtin_ia32_cmpltsd: 6426 case X86::BI__builtin_ia32_cmplesd: 6427 case X86::BI__builtin_ia32_cmpunordsd: 6428 case X86::BI__builtin_ia32_cmpneqsd: 6429 case X86::BI__builtin_ia32_cmpnltsd: 6430 case X86::BI__builtin_ia32_cmpnlesd: 6431 case X86::BI__builtin_ia32_cmpordsd: 6432 name = "cmpsd"; 6433 ID = Intrinsic::x86_sse2_cmp_sd; 6434 break; 6435 } 6436 6437 Ops.push_back(llvm::ConstantInt::get(Int8Ty, Imm)); 6438 llvm::Function *F = CGM.getIntrinsic(ID); 6439 return Builder.CreateCall(F, Ops, name); 6440 } 6441 } 6442 6443 6444 Value *CodeGenFunction::EmitPPCBuiltinExpr(unsigned BuiltinID, 6445 const CallExpr *E) { 6446 SmallVector<Value*, 4> Ops; 6447 6448 for (unsigned i = 0, e = E->getNumArgs(); i != e; i++) 6449 Ops.push_back(EmitScalarExpr(E->getArg(i))); 6450 6451 Intrinsic::ID ID = Intrinsic::not_intrinsic; 6452 6453 switch (BuiltinID) { 6454 default: return nullptr; 6455 6456 // vec_ld, vec_lvsl, vec_lvsr 6457 case PPC::BI__builtin_altivec_lvx: 6458 case PPC::BI__builtin_altivec_lvxl: 6459 case PPC::BI__builtin_altivec_lvebx: 6460 case PPC::BI__builtin_altivec_lvehx: 6461 case PPC::BI__builtin_altivec_lvewx: 6462 case PPC::BI__builtin_altivec_lvsl: 6463 case PPC::BI__builtin_altivec_lvsr: 6464 case PPC::BI__builtin_vsx_lxvd2x: 6465 case PPC::BI__builtin_vsx_lxvw4x: 6466 { 6467 Ops[1] = Builder.CreateBitCast(Ops[1], Int8PtrTy); 6468 6469 Ops[0] = Builder.CreateGEP(Ops[1], Ops[0]); 6470 Ops.pop_back(); 6471 6472 switch (BuiltinID) { 6473 default: llvm_unreachable("Unsupported ld/lvsl/lvsr intrinsic!"); 6474 case PPC::BI__builtin_altivec_lvx: 6475 ID = Intrinsic::ppc_altivec_lvx; 6476 break; 6477 case PPC::BI__builtin_altivec_lvxl: 6478 ID = Intrinsic::ppc_altivec_lvxl; 6479 break; 6480 case PPC::BI__builtin_altivec_lvebx: 6481 ID = Intrinsic::ppc_altivec_lvebx; 6482 break; 6483 case PPC::BI__builtin_altivec_lvehx: 6484 ID = Intrinsic::ppc_altivec_lvehx; 6485 break; 6486 case PPC::BI__builtin_altivec_lvewx: 6487 ID = Intrinsic::ppc_altivec_lvewx; 6488 break; 6489 case PPC::BI__builtin_altivec_lvsl: 6490 ID = Intrinsic::ppc_altivec_lvsl; 6491 break; 6492 case PPC::BI__builtin_altivec_lvsr: 6493 ID = Intrinsic::ppc_altivec_lvsr; 6494 break; 6495 case PPC::BI__builtin_vsx_lxvd2x: 6496 ID = Intrinsic::ppc_vsx_lxvd2x; 6497 break; 6498 case PPC::BI__builtin_vsx_lxvw4x: 6499 ID = Intrinsic::ppc_vsx_lxvw4x; 6500 break; 6501 } 6502 llvm::Function *F = CGM.getIntrinsic(ID); 6503 return Builder.CreateCall(F, Ops, ""); 6504 } 6505 6506 // vec_st 6507 case PPC::BI__builtin_altivec_stvx: 6508 case PPC::BI__builtin_altivec_stvxl: 6509 case PPC::BI__builtin_altivec_stvebx: 6510 case PPC::BI__builtin_altivec_stvehx: 6511 case PPC::BI__builtin_altivec_stvewx: 6512 case PPC::BI__builtin_vsx_stxvd2x: 6513 case PPC::BI__builtin_vsx_stxvw4x: 6514 { 6515 Ops[2] = Builder.CreateBitCast(Ops[2], Int8PtrTy); 6516 Ops[1] = Builder.CreateGEP(Ops[2], Ops[1]); 6517 Ops.pop_back(); 6518 6519 switch (BuiltinID) { 6520 default: llvm_unreachable("Unsupported st intrinsic!"); 6521 case PPC::BI__builtin_altivec_stvx: 6522 ID = Intrinsic::ppc_altivec_stvx; 6523 break; 6524 case PPC::BI__builtin_altivec_stvxl: 6525 ID = Intrinsic::ppc_altivec_stvxl; 6526 break; 6527 case PPC::BI__builtin_altivec_stvebx: 6528 ID = Intrinsic::ppc_altivec_stvebx; 6529 break; 6530 case PPC::BI__builtin_altivec_stvehx: 6531 ID = Intrinsic::ppc_altivec_stvehx; 6532 break; 6533 case PPC::BI__builtin_altivec_stvewx: 6534 ID = Intrinsic::ppc_altivec_stvewx; 6535 break; 6536 case PPC::BI__builtin_vsx_stxvd2x: 6537 ID = Intrinsic::ppc_vsx_stxvd2x; 6538 break; 6539 case PPC::BI__builtin_vsx_stxvw4x: 6540 ID = Intrinsic::ppc_vsx_stxvw4x; 6541 break; 6542 } 6543 llvm::Function *F = CGM.getIntrinsic(ID); 6544 return Builder.CreateCall(F, Ops, ""); 6545 } 6546 // Square root 6547 case PPC::BI__builtin_vsx_xvsqrtsp: 6548 case PPC::BI__builtin_vsx_xvsqrtdp: { 6549 llvm::Type *ResultType = ConvertType(E->getType()); 6550 Value *X = EmitScalarExpr(E->getArg(0)); 6551 ID = Intrinsic::sqrt; 6552 llvm::Function *F = CGM.getIntrinsic(ID, ResultType); 6553 return Builder.CreateCall(F, X); 6554 } 6555 // Count leading zeros 6556 case PPC::BI__builtin_altivec_vclzb: 6557 case PPC::BI__builtin_altivec_vclzh: 6558 case PPC::BI__builtin_altivec_vclzw: 6559 case PPC::BI__builtin_altivec_vclzd: { 6560 llvm::Type *ResultType = ConvertType(E->getType()); 6561 Value *X = EmitScalarExpr(E->getArg(0)); 6562 Value *Undef = ConstantInt::get(Builder.getInt1Ty(), false); 6563 Function *F = CGM.getIntrinsic(Intrinsic::ctlz, ResultType); 6564 return Builder.CreateCall(F, {X, Undef}); 6565 } 6566 // Copy sign 6567 case PPC::BI__builtin_vsx_xvcpsgnsp: 6568 case PPC::BI__builtin_vsx_xvcpsgndp: { 6569 llvm::Type *ResultType = ConvertType(E->getType()); 6570 Value *X = EmitScalarExpr(E->getArg(0)); 6571 Value *Y = EmitScalarExpr(E->getArg(1)); 6572 ID = Intrinsic::copysign; 6573 llvm::Function *F = CGM.getIntrinsic(ID, ResultType); 6574 return Builder.CreateCall(F, {X, Y}); 6575 } 6576 // Rounding/truncation 6577 case PPC::BI__builtin_vsx_xvrspip: 6578 case PPC::BI__builtin_vsx_xvrdpip: 6579 case PPC::BI__builtin_vsx_xvrdpim: 6580 case PPC::BI__builtin_vsx_xvrspim: 6581 case PPC::BI__builtin_vsx_xvrdpi: 6582 case PPC::BI__builtin_vsx_xvrspi: 6583 case PPC::BI__builtin_vsx_xvrdpic: 6584 case PPC::BI__builtin_vsx_xvrspic: 6585 case PPC::BI__builtin_vsx_xvrdpiz: 6586 case PPC::BI__builtin_vsx_xvrspiz: { 6587 llvm::Type *ResultType = ConvertType(E->getType()); 6588 Value *X = EmitScalarExpr(E->getArg(0)); 6589 if (BuiltinID == PPC::BI__builtin_vsx_xvrdpim || 6590 BuiltinID == PPC::BI__builtin_vsx_xvrspim) 6591 ID = Intrinsic::floor; 6592 else if (BuiltinID == PPC::BI__builtin_vsx_xvrdpi || 6593 BuiltinID == PPC::BI__builtin_vsx_xvrspi) 6594 ID = Intrinsic::round; 6595 else if (BuiltinID == PPC::BI__builtin_vsx_xvrdpic || 6596 BuiltinID == PPC::BI__builtin_vsx_xvrspic) 6597 ID = Intrinsic::nearbyint; 6598 else if (BuiltinID == PPC::BI__builtin_vsx_xvrdpip || 6599 BuiltinID == PPC::BI__builtin_vsx_xvrspip) 6600 ID = Intrinsic::ceil; 6601 else if (BuiltinID == PPC::BI__builtin_vsx_xvrdpiz || 6602 BuiltinID == PPC::BI__builtin_vsx_xvrspiz) 6603 ID = Intrinsic::trunc; 6604 llvm::Function *F = CGM.getIntrinsic(ID, ResultType); 6605 return Builder.CreateCall(F, X); 6606 } 6607 // FMA variations 6608 case PPC::BI__builtin_vsx_xvmaddadp: 6609 case PPC::BI__builtin_vsx_xvmaddasp: 6610 case PPC::BI__builtin_vsx_xvnmaddadp: 6611 case PPC::BI__builtin_vsx_xvnmaddasp: 6612 case PPC::BI__builtin_vsx_xvmsubadp: 6613 case PPC::BI__builtin_vsx_xvmsubasp: 6614 case PPC::BI__builtin_vsx_xvnmsubadp: 6615 case PPC::BI__builtin_vsx_xvnmsubasp: { 6616 llvm::Type *ResultType = ConvertType(E->getType()); 6617 Value *X = EmitScalarExpr(E->getArg(0)); 6618 Value *Y = EmitScalarExpr(E->getArg(1)); 6619 Value *Z = EmitScalarExpr(E->getArg(2)); 6620 Value *Zero = llvm::ConstantFP::getZeroValueForNegation(ResultType); 6621 llvm::Function *F = CGM.getIntrinsic(Intrinsic::fma, ResultType); 6622 switch (BuiltinID) { 6623 case PPC::BI__builtin_vsx_xvmaddadp: 6624 case PPC::BI__builtin_vsx_xvmaddasp: 6625 return Builder.CreateCall(F, {X, Y, Z}); 6626 case PPC::BI__builtin_vsx_xvnmaddadp: 6627 case PPC::BI__builtin_vsx_xvnmaddasp: 6628 return Builder.CreateFSub(Zero, 6629 Builder.CreateCall(F, {X, Y, Z}), "sub"); 6630 case PPC::BI__builtin_vsx_xvmsubadp: 6631 case PPC::BI__builtin_vsx_xvmsubasp: 6632 return Builder.CreateCall(F, 6633 {X, Y, Builder.CreateFSub(Zero, Z, "sub")}); 6634 case PPC::BI__builtin_vsx_xvnmsubadp: 6635 case PPC::BI__builtin_vsx_xvnmsubasp: 6636 Value *FsubRes = 6637 Builder.CreateCall(F, {X, Y, Builder.CreateFSub(Zero, Z, "sub")}); 6638 return Builder.CreateFSub(Zero, FsubRes, "sub"); 6639 } 6640 llvm_unreachable("Unknown FMA operation"); 6641 return nullptr; // Suppress no-return warning 6642 } 6643 } 6644 } 6645 6646 // Emit an intrinsic that has 1 float or double. 6647 static Value *emitUnaryFPBuiltin(CodeGenFunction &CGF, 6648 const CallExpr *E, 6649 unsigned IntrinsicID) { 6650 llvm::Value *Src0 = CGF.EmitScalarExpr(E->getArg(0)); 6651 6652 Value *F = CGF.CGM.getIntrinsic(IntrinsicID, Src0->getType()); 6653 return CGF.Builder.CreateCall(F, Src0); 6654 } 6655 6656 // Emit an intrinsic that has 3 float or double operands. 6657 static Value *emitTernaryFPBuiltin(CodeGenFunction &CGF, 6658 const CallExpr *E, 6659 unsigned IntrinsicID) { 6660 llvm::Value *Src0 = CGF.EmitScalarExpr(E->getArg(0)); 6661 llvm::Value *Src1 = CGF.EmitScalarExpr(E->getArg(1)); 6662 llvm::Value *Src2 = CGF.EmitScalarExpr(E->getArg(2)); 6663 6664 Value *F = CGF.CGM.getIntrinsic(IntrinsicID, Src0->getType()); 6665 return CGF.Builder.CreateCall(F, {Src0, Src1, Src2}); 6666 } 6667 6668 // Emit an intrinsic that has 1 float or double operand, and 1 integer. 6669 static Value *emitFPIntBuiltin(CodeGenFunction &CGF, 6670 const CallExpr *E, 6671 unsigned IntrinsicID) { 6672 llvm::Value *Src0 = CGF.EmitScalarExpr(E->getArg(0)); 6673 llvm::Value *Src1 = CGF.EmitScalarExpr(E->getArg(1)); 6674 6675 Value *F = CGF.CGM.getIntrinsic(IntrinsicID, Src0->getType()); 6676 return CGF.Builder.CreateCall(F, {Src0, Src1}); 6677 } 6678 6679 Value *CodeGenFunction::EmitAMDGPUBuiltinExpr(unsigned BuiltinID, 6680 const CallExpr *E) { 6681 switch (BuiltinID) { 6682 case AMDGPU::BI__builtin_amdgpu_div_scale: 6683 case AMDGPU::BI__builtin_amdgpu_div_scalef: { 6684 // Translate from the intrinsics's struct return to the builtin's out 6685 // argument. 6686 6687 std::pair<llvm::Value *, unsigned> FlagOutPtr 6688 = EmitPointerWithAlignment(E->getArg(3)); 6689 6690 llvm::Value *X = EmitScalarExpr(E->getArg(0)); 6691 llvm::Value *Y = EmitScalarExpr(E->getArg(1)); 6692 llvm::Value *Z = EmitScalarExpr(E->getArg(2)); 6693 6694 llvm::Value *Callee = CGM.getIntrinsic(Intrinsic::AMDGPU_div_scale, 6695 X->getType()); 6696 6697 llvm::Value *Tmp = Builder.CreateCall(Callee, {X, Y, Z}); 6698 6699 llvm::Value *Result = Builder.CreateExtractValue(Tmp, 0); 6700 llvm::Value *Flag = Builder.CreateExtractValue(Tmp, 1); 6701 6702 llvm::Type *RealFlagType 6703 = FlagOutPtr.first->getType()->getPointerElementType(); 6704 6705 llvm::Value *FlagExt = Builder.CreateZExt(Flag, RealFlagType); 6706 llvm::StoreInst *FlagStore = Builder.CreateStore(FlagExt, FlagOutPtr.first); 6707 FlagStore->setAlignment(FlagOutPtr.second); 6708 return Result; 6709 } 6710 case AMDGPU::BI__builtin_amdgpu_div_fmas: 6711 case AMDGPU::BI__builtin_amdgpu_div_fmasf: { 6712 llvm::Value *Src0 = EmitScalarExpr(E->getArg(0)); 6713 llvm::Value *Src1 = EmitScalarExpr(E->getArg(1)); 6714 llvm::Value *Src2 = EmitScalarExpr(E->getArg(2)); 6715 llvm::Value *Src3 = EmitScalarExpr(E->getArg(3)); 6716 6717 llvm::Value *F = CGM.getIntrinsic(Intrinsic::AMDGPU_div_fmas, 6718 Src0->getType()); 6719 llvm::Value *Src3ToBool = Builder.CreateIsNotNull(Src3); 6720 return Builder.CreateCall(F, {Src0, Src1, Src2, Src3ToBool}); 6721 } 6722 case AMDGPU::BI__builtin_amdgpu_div_fixup: 6723 case AMDGPU::BI__builtin_amdgpu_div_fixupf: 6724 return emitTernaryFPBuiltin(*this, E, Intrinsic::AMDGPU_div_fixup); 6725 case AMDGPU::BI__builtin_amdgpu_trig_preop: 6726 case AMDGPU::BI__builtin_amdgpu_trig_preopf: 6727 return emitFPIntBuiltin(*this, E, Intrinsic::AMDGPU_trig_preop); 6728 case AMDGPU::BI__builtin_amdgpu_rcp: 6729 case AMDGPU::BI__builtin_amdgpu_rcpf: 6730 return emitUnaryFPBuiltin(*this, E, Intrinsic::AMDGPU_rcp); 6731 case AMDGPU::BI__builtin_amdgpu_rsq: 6732 case AMDGPU::BI__builtin_amdgpu_rsqf: 6733 return emitUnaryFPBuiltin(*this, E, Intrinsic::AMDGPU_rsq); 6734 case AMDGPU::BI__builtin_amdgpu_rsq_clamped: 6735 case AMDGPU::BI__builtin_amdgpu_rsq_clampedf: 6736 return emitUnaryFPBuiltin(*this, E, Intrinsic::AMDGPU_rsq_clamped); 6737 case AMDGPU::BI__builtin_amdgpu_ldexp: 6738 case AMDGPU::BI__builtin_amdgpu_ldexpf: 6739 return emitFPIntBuiltin(*this, E, Intrinsic::AMDGPU_ldexp); 6740 case AMDGPU::BI__builtin_amdgpu_class: 6741 case AMDGPU::BI__builtin_amdgpu_classf: 6742 return emitFPIntBuiltin(*this, E, Intrinsic::AMDGPU_class); 6743 default: 6744 return nullptr; 6745 } 6746 } 6747 6748 /// Handle a SystemZ function in which the final argument is a pointer 6749 /// to an int that receives the post-instruction CC value. At the LLVM level 6750 /// this is represented as a function that returns a {result, cc} pair. 6751 static Value *EmitSystemZIntrinsicWithCC(CodeGenFunction &CGF, 6752 unsigned IntrinsicID, 6753 const CallExpr *E) { 6754 unsigned NumArgs = E->getNumArgs() - 1; 6755 SmallVector<Value *, 8> Args(NumArgs); 6756 for (unsigned I = 0; I < NumArgs; ++I) 6757 Args[I] = CGF.EmitScalarExpr(E->getArg(I)); 6758 Value *CCPtr = CGF.EmitScalarExpr(E->getArg(NumArgs)); 6759 Value *F = CGF.CGM.getIntrinsic(IntrinsicID); 6760 Value *Call = CGF.Builder.CreateCall(F, Args); 6761 Value *CC = CGF.Builder.CreateExtractValue(Call, 1); 6762 CGF.Builder.CreateStore(CC, CCPtr); 6763 return CGF.Builder.CreateExtractValue(Call, 0); 6764 } 6765 6766 Value *CodeGenFunction::EmitSystemZBuiltinExpr(unsigned BuiltinID, 6767 const CallExpr *E) { 6768 switch (BuiltinID) { 6769 case SystemZ::BI__builtin_tbegin: { 6770 Value *TDB = EmitScalarExpr(E->getArg(0)); 6771 Value *Control = llvm::ConstantInt::get(Int32Ty, 0xff0c); 6772 Value *F = CGM.getIntrinsic(Intrinsic::s390_tbegin); 6773 return Builder.CreateCall(F, {TDB, Control}); 6774 } 6775 case SystemZ::BI__builtin_tbegin_nofloat: { 6776 Value *TDB = EmitScalarExpr(E->getArg(0)); 6777 Value *Control = llvm::ConstantInt::get(Int32Ty, 0xff0c); 6778 Value *F = CGM.getIntrinsic(Intrinsic::s390_tbegin_nofloat); 6779 return Builder.CreateCall(F, {TDB, Control}); 6780 } 6781 case SystemZ::BI__builtin_tbeginc: { 6782 Value *TDB = llvm::ConstantPointerNull::get(Int8PtrTy); 6783 Value *Control = llvm::ConstantInt::get(Int32Ty, 0xff08); 6784 Value *F = CGM.getIntrinsic(Intrinsic::s390_tbeginc); 6785 return Builder.CreateCall(F, {TDB, Control}); 6786 } 6787 case SystemZ::BI__builtin_tabort: { 6788 Value *Data = EmitScalarExpr(E->getArg(0)); 6789 Value *F = CGM.getIntrinsic(Intrinsic::s390_tabort); 6790 return Builder.CreateCall(F, Builder.CreateSExt(Data, Int64Ty, "tabort")); 6791 } 6792 case SystemZ::BI__builtin_non_tx_store: { 6793 Value *Address = EmitScalarExpr(E->getArg(0)); 6794 Value *Data = EmitScalarExpr(E->getArg(1)); 6795 Value *F = CGM.getIntrinsic(Intrinsic::s390_ntstg); 6796 return Builder.CreateCall(F, {Data, Address}); 6797 } 6798 6799 // Vector builtins. Note that most vector builtins are mapped automatically 6800 // to target-specific LLVM intrinsics. The ones handled specially here can 6801 // be represented via standard LLVM IR, which is preferable to enable common 6802 // LLVM optimizations. 6803 6804 case SystemZ::BI__builtin_s390_vpopctb: 6805 case SystemZ::BI__builtin_s390_vpopcth: 6806 case SystemZ::BI__builtin_s390_vpopctf: 6807 case SystemZ::BI__builtin_s390_vpopctg: { 6808 llvm::Type *ResultType = ConvertType(E->getType()); 6809 Value *X = EmitScalarExpr(E->getArg(0)); 6810 Function *F = CGM.getIntrinsic(Intrinsic::ctpop, ResultType); 6811 return Builder.CreateCall(F, X); 6812 } 6813 6814 case SystemZ::BI__builtin_s390_vclzb: 6815 case SystemZ::BI__builtin_s390_vclzh: 6816 case SystemZ::BI__builtin_s390_vclzf: 6817 case SystemZ::BI__builtin_s390_vclzg: { 6818 llvm::Type *ResultType = ConvertType(E->getType()); 6819 Value *X = EmitScalarExpr(E->getArg(0)); 6820 Value *Undef = ConstantInt::get(Builder.getInt1Ty(), false); 6821 Function *F = CGM.getIntrinsic(Intrinsic::ctlz, ResultType); 6822 return Builder.CreateCall(F, {X, Undef}); 6823 } 6824 6825 case SystemZ::BI__builtin_s390_vctzb: 6826 case SystemZ::BI__builtin_s390_vctzh: 6827 case SystemZ::BI__builtin_s390_vctzf: 6828 case SystemZ::BI__builtin_s390_vctzg: { 6829 llvm::Type *ResultType = ConvertType(E->getType()); 6830 Value *X = EmitScalarExpr(E->getArg(0)); 6831 Value *Undef = ConstantInt::get(Builder.getInt1Ty(), false); 6832 Function *F = CGM.getIntrinsic(Intrinsic::cttz, ResultType); 6833 return Builder.CreateCall(F, {X, Undef}); 6834 } 6835 6836 case SystemZ::BI__builtin_s390_vfsqdb: { 6837 llvm::Type *ResultType = ConvertType(E->getType()); 6838 Value *X = EmitScalarExpr(E->getArg(0)); 6839 Function *F = CGM.getIntrinsic(Intrinsic::sqrt, ResultType); 6840 return Builder.CreateCall(F, X); 6841 } 6842 case SystemZ::BI__builtin_s390_vfmadb: { 6843 llvm::Type *ResultType = ConvertType(E->getType()); 6844 Value *X = EmitScalarExpr(E->getArg(0)); 6845 Value *Y = EmitScalarExpr(E->getArg(1)); 6846 Value *Z = EmitScalarExpr(E->getArg(2)); 6847 Function *F = CGM.getIntrinsic(Intrinsic::fma, ResultType); 6848 return Builder.CreateCall(F, {X, Y, Z}); 6849 } 6850 case SystemZ::BI__builtin_s390_vfmsdb: { 6851 llvm::Type *ResultType = ConvertType(E->getType()); 6852 Value *X = EmitScalarExpr(E->getArg(0)); 6853 Value *Y = EmitScalarExpr(E->getArg(1)); 6854 Value *Z = EmitScalarExpr(E->getArg(2)); 6855 Value *Zero = llvm::ConstantFP::getZeroValueForNegation(ResultType); 6856 Function *F = CGM.getIntrinsic(Intrinsic::fma, ResultType); 6857 return Builder.CreateCall(F, {X, Y, Builder.CreateFSub(Zero, Z, "sub")}); 6858 } 6859 case SystemZ::BI__builtin_s390_vflpdb: { 6860 llvm::Type *ResultType = ConvertType(E->getType()); 6861 Value *X = EmitScalarExpr(E->getArg(0)); 6862 Function *F = CGM.getIntrinsic(Intrinsic::fabs, ResultType); 6863 return Builder.CreateCall(F, X); 6864 } 6865 case SystemZ::BI__builtin_s390_vflndb: { 6866 llvm::Type *ResultType = ConvertType(E->getType()); 6867 Value *X = EmitScalarExpr(E->getArg(0)); 6868 Value *Zero = llvm::ConstantFP::getZeroValueForNegation(ResultType); 6869 Function *F = CGM.getIntrinsic(Intrinsic::fabs, ResultType); 6870 return Builder.CreateFSub(Zero, Builder.CreateCall(F, X), "sub"); 6871 } 6872 case SystemZ::BI__builtin_s390_vfidb: { 6873 llvm::Type *ResultType = ConvertType(E->getType()); 6874 Value *X = EmitScalarExpr(E->getArg(0)); 6875 // Constant-fold the M4 and M5 mask arguments. 6876 llvm::APSInt M4, M5; 6877 bool IsConstM4 = E->getArg(1)->isIntegerConstantExpr(M4, getContext()); 6878 bool IsConstM5 = E->getArg(2)->isIntegerConstantExpr(M5, getContext()); 6879 assert(IsConstM4 && IsConstM5 && "Constant arg isn't actually constant?"); 6880 (void)IsConstM4; (void)IsConstM5; 6881 // Check whether this instance of vfidb can be represented via a LLVM 6882 // standard intrinsic. We only support some combinations of M4 and M5. 6883 Intrinsic::ID ID = Intrinsic::not_intrinsic; 6884 switch (M4.getZExtValue()) { 6885 default: break; 6886 case 0: // IEEE-inexact exception allowed 6887 switch (M5.getZExtValue()) { 6888 default: break; 6889 case 0: ID = Intrinsic::rint; break; 6890 } 6891 break; 6892 case 4: // IEEE-inexact exception suppressed 6893 switch (M5.getZExtValue()) { 6894 default: break; 6895 case 0: ID = Intrinsic::nearbyint; break; 6896 case 1: ID = Intrinsic::round; break; 6897 case 5: ID = Intrinsic::trunc; break; 6898 case 6: ID = Intrinsic::ceil; break; 6899 case 7: ID = Intrinsic::floor; break; 6900 } 6901 break; 6902 } 6903 if (ID != Intrinsic::not_intrinsic) { 6904 Function *F = CGM.getIntrinsic(ID, ResultType); 6905 return Builder.CreateCall(F, X); 6906 } 6907 Function *F = CGM.getIntrinsic(Intrinsic::s390_vfidb); 6908 Value *M4Value = llvm::ConstantInt::get(getLLVMContext(), M4); 6909 Value *M5Value = llvm::ConstantInt::get(getLLVMContext(), M5); 6910 return Builder.CreateCall(F, {X, M4Value, M5Value}); 6911 } 6912 6913 // Vector intrisincs that output the post-instruction CC value. 6914 6915 #define INTRINSIC_WITH_CC(NAME) \ 6916 case SystemZ::BI__builtin_##NAME: \ 6917 return EmitSystemZIntrinsicWithCC(*this, Intrinsic::NAME, E) 6918 6919 INTRINSIC_WITH_CC(s390_vpkshs); 6920 INTRINSIC_WITH_CC(s390_vpksfs); 6921 INTRINSIC_WITH_CC(s390_vpksgs); 6922 6923 INTRINSIC_WITH_CC(s390_vpklshs); 6924 INTRINSIC_WITH_CC(s390_vpklsfs); 6925 INTRINSIC_WITH_CC(s390_vpklsgs); 6926 6927 INTRINSIC_WITH_CC(s390_vceqbs); 6928 INTRINSIC_WITH_CC(s390_vceqhs); 6929 INTRINSIC_WITH_CC(s390_vceqfs); 6930 INTRINSIC_WITH_CC(s390_vceqgs); 6931 6932 INTRINSIC_WITH_CC(s390_vchbs); 6933 INTRINSIC_WITH_CC(s390_vchhs); 6934 INTRINSIC_WITH_CC(s390_vchfs); 6935 INTRINSIC_WITH_CC(s390_vchgs); 6936 6937 INTRINSIC_WITH_CC(s390_vchlbs); 6938 INTRINSIC_WITH_CC(s390_vchlhs); 6939 INTRINSIC_WITH_CC(s390_vchlfs); 6940 INTRINSIC_WITH_CC(s390_vchlgs); 6941 6942 INTRINSIC_WITH_CC(s390_vfaebs); 6943 INTRINSIC_WITH_CC(s390_vfaehs); 6944 INTRINSIC_WITH_CC(s390_vfaefs); 6945 6946 INTRINSIC_WITH_CC(s390_vfaezbs); 6947 INTRINSIC_WITH_CC(s390_vfaezhs); 6948 INTRINSIC_WITH_CC(s390_vfaezfs); 6949 6950 INTRINSIC_WITH_CC(s390_vfeebs); 6951 INTRINSIC_WITH_CC(s390_vfeehs); 6952 INTRINSIC_WITH_CC(s390_vfeefs); 6953 6954 INTRINSIC_WITH_CC(s390_vfeezbs); 6955 INTRINSIC_WITH_CC(s390_vfeezhs); 6956 INTRINSIC_WITH_CC(s390_vfeezfs); 6957 6958 INTRINSIC_WITH_CC(s390_vfenebs); 6959 INTRINSIC_WITH_CC(s390_vfenehs); 6960 INTRINSIC_WITH_CC(s390_vfenefs); 6961 6962 INTRINSIC_WITH_CC(s390_vfenezbs); 6963 INTRINSIC_WITH_CC(s390_vfenezhs); 6964 INTRINSIC_WITH_CC(s390_vfenezfs); 6965 6966 INTRINSIC_WITH_CC(s390_vistrbs); 6967 INTRINSIC_WITH_CC(s390_vistrhs); 6968 INTRINSIC_WITH_CC(s390_vistrfs); 6969 6970 INTRINSIC_WITH_CC(s390_vstrcbs); 6971 INTRINSIC_WITH_CC(s390_vstrchs); 6972 INTRINSIC_WITH_CC(s390_vstrcfs); 6973 6974 INTRINSIC_WITH_CC(s390_vstrczbs); 6975 INTRINSIC_WITH_CC(s390_vstrczhs); 6976 INTRINSIC_WITH_CC(s390_vstrczfs); 6977 6978 INTRINSIC_WITH_CC(s390_vfcedbs); 6979 INTRINSIC_WITH_CC(s390_vfchdbs); 6980 INTRINSIC_WITH_CC(s390_vfchedbs); 6981 6982 INTRINSIC_WITH_CC(s390_vftcidb); 6983 6984 #undef INTRINSIC_WITH_CC 6985 6986 default: 6987 return nullptr; 6988 } 6989 } 6990 6991 Value *CodeGenFunction::EmitNVPTXBuiltinExpr(unsigned BuiltinID, 6992 const CallExpr *E) { 6993 switch (BuiltinID) { 6994 case NVPTX::BI__nvvm_atom_add_gen_i: 6995 case NVPTX::BI__nvvm_atom_add_gen_l: 6996 case NVPTX::BI__nvvm_atom_add_gen_ll: 6997 return MakeBinaryAtomicValue(*this, llvm::AtomicRMWInst::Add, E); 6998 6999 case NVPTX::BI__nvvm_atom_sub_gen_i: 7000 case NVPTX::BI__nvvm_atom_sub_gen_l: 7001 case NVPTX::BI__nvvm_atom_sub_gen_ll: 7002 return MakeBinaryAtomicValue(*this, llvm::AtomicRMWInst::Sub, E); 7003 7004 case NVPTX::BI__nvvm_atom_and_gen_i: 7005 case NVPTX::BI__nvvm_atom_and_gen_l: 7006 case NVPTX::BI__nvvm_atom_and_gen_ll: 7007 return MakeBinaryAtomicValue(*this, llvm::AtomicRMWInst::And, E); 7008 7009 case NVPTX::BI__nvvm_atom_or_gen_i: 7010 case NVPTX::BI__nvvm_atom_or_gen_l: 7011 case NVPTX::BI__nvvm_atom_or_gen_ll: 7012 return MakeBinaryAtomicValue(*this, llvm::AtomicRMWInst::Or, E); 7013 7014 case NVPTX::BI__nvvm_atom_xor_gen_i: 7015 case NVPTX::BI__nvvm_atom_xor_gen_l: 7016 case NVPTX::BI__nvvm_atom_xor_gen_ll: 7017 return MakeBinaryAtomicValue(*this, llvm::AtomicRMWInst::Xor, E); 7018 7019 case NVPTX::BI__nvvm_atom_xchg_gen_i: 7020 case NVPTX::BI__nvvm_atom_xchg_gen_l: 7021 case NVPTX::BI__nvvm_atom_xchg_gen_ll: 7022 return MakeBinaryAtomicValue(*this, llvm::AtomicRMWInst::Xchg, E); 7023 7024 case NVPTX::BI__nvvm_atom_max_gen_i: 7025 case NVPTX::BI__nvvm_atom_max_gen_l: 7026 case NVPTX::BI__nvvm_atom_max_gen_ll: 7027 case NVPTX::BI__nvvm_atom_max_gen_ui: 7028 case NVPTX::BI__nvvm_atom_max_gen_ul: 7029 case NVPTX::BI__nvvm_atom_max_gen_ull: 7030 return MakeBinaryAtomicValue(*this, llvm::AtomicRMWInst::Max, E); 7031 7032 case NVPTX::BI__nvvm_atom_min_gen_i: 7033 case NVPTX::BI__nvvm_atom_min_gen_l: 7034 case NVPTX::BI__nvvm_atom_min_gen_ll: 7035 case NVPTX::BI__nvvm_atom_min_gen_ui: 7036 case NVPTX::BI__nvvm_atom_min_gen_ul: 7037 case NVPTX::BI__nvvm_atom_min_gen_ull: 7038 return MakeBinaryAtomicValue(*this, llvm::AtomicRMWInst::Min, E); 7039 7040 case NVPTX::BI__nvvm_atom_cas_gen_i: 7041 case NVPTX::BI__nvvm_atom_cas_gen_l: 7042 case NVPTX::BI__nvvm_atom_cas_gen_ll: 7043 return MakeAtomicCmpXchgValue(*this, E, true); 7044 7045 case NVPTX::BI__nvvm_atom_add_gen_f: { 7046 Value *Ptr = EmitScalarExpr(E->getArg(0)); 7047 Value *Val = EmitScalarExpr(E->getArg(1)); 7048 // atomicrmw only deals with integer arguments so we need to use 7049 // LLVM's nvvm_atomic_load_add_f32 intrinsic for that. 7050 Value *FnALAF32 = 7051 CGM.getIntrinsic(Intrinsic::nvvm_atomic_load_add_f32, Ptr->getType()); 7052 return Builder.CreateCall(FnALAF32, {Ptr, Val}); 7053 } 7054 7055 default: 7056 return nullptr; 7057 } 7058 } 7059