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 "CGObjCRuntime.h" 16 #include "CodeGenModule.h" 17 #include "TargetInfo.h" 18 #include "clang/AST/ASTContext.h" 19 #include "clang/AST/Decl.h" 20 #include "clang/Basic/TargetBuiltins.h" 21 #include "clang/Basic/TargetInfo.h" 22 #include "clang/CodeGen/CGFunctionInfo.h" 23 #include "llvm/IR/DataLayout.h" 24 #include "llvm/IR/Intrinsics.h" 25 26 using namespace clang; 27 using namespace CodeGen; 28 using namespace llvm; 29 30 /// getBuiltinLibFunction - Given a builtin id for a function like 31 /// "__builtin_fabsf", return a Function* for "fabsf". 32 llvm::Value *CodeGenModule::getBuiltinLibFunction(const FunctionDecl *FD, 33 unsigned BuiltinID) { 34 assert(Context.BuiltinInfo.isLibFunction(BuiltinID)); 35 36 // Get the name, skip over the __builtin_ prefix (if necessary). 37 StringRef Name; 38 GlobalDecl D(FD); 39 40 // If the builtin has been declared explicitly with an assembler label, 41 // use the mangled name. This differs from the plain label on platforms 42 // that prefix labels. 43 if (FD->hasAttr<AsmLabelAttr>()) 44 Name = getMangledName(D); 45 else 46 Name = Context.BuiltinInfo.GetName(BuiltinID) + 10; 47 48 llvm::FunctionType *Ty = 49 cast<llvm::FunctionType>(getTypes().ConvertType(FD->getType())); 50 51 return GetOrCreateLLVMFunction(Name, Ty, D, /*ForVTable=*/false); 52 } 53 54 /// Emit the conversions required to turn the given value into an 55 /// integer of the given size. 56 static Value *EmitToInt(CodeGenFunction &CGF, llvm::Value *V, 57 QualType T, llvm::IntegerType *IntType) { 58 V = CGF.EmitToMemory(V, T); 59 60 if (V->getType()->isPointerTy()) 61 return CGF.Builder.CreatePtrToInt(V, IntType); 62 63 assert(V->getType() == IntType); 64 return V; 65 } 66 67 static Value *EmitFromInt(CodeGenFunction &CGF, llvm::Value *V, 68 QualType T, llvm::Type *ResultType) { 69 V = CGF.EmitFromMemory(V, T); 70 71 if (ResultType->isPointerTy()) 72 return CGF.Builder.CreateIntToPtr(V, ResultType); 73 74 assert(V->getType() == ResultType); 75 return V; 76 } 77 78 /// Utility to insert an atomic instruction based on Instrinsic::ID 79 /// and the expression node. 80 static RValue EmitBinaryAtomic(CodeGenFunction &CGF, 81 llvm::AtomicRMWInst::BinOp Kind, 82 const CallExpr *E) { 83 QualType T = E->getType(); 84 assert(E->getArg(0)->getType()->isPointerType()); 85 assert(CGF.getContext().hasSameUnqualifiedType(T, 86 E->getArg(0)->getType()->getPointeeType())); 87 assert(CGF.getContext().hasSameUnqualifiedType(T, E->getArg(1)->getType())); 88 89 llvm::Value *DestPtr = CGF.EmitScalarExpr(E->getArg(0)); 90 unsigned AddrSpace = DestPtr->getType()->getPointerAddressSpace(); 91 92 llvm::IntegerType *IntType = 93 llvm::IntegerType::get(CGF.getLLVMContext(), 94 CGF.getContext().getTypeSize(T)); 95 llvm::Type *IntPtrType = IntType->getPointerTo(AddrSpace); 96 97 llvm::Value *Args[2]; 98 Args[0] = CGF.Builder.CreateBitCast(DestPtr, IntPtrType); 99 Args[1] = CGF.EmitScalarExpr(E->getArg(1)); 100 llvm::Type *ValueType = Args[1]->getType(); 101 Args[1] = EmitToInt(CGF, Args[1], T, IntType); 102 103 llvm::Value *Result = 104 CGF.Builder.CreateAtomicRMW(Kind, Args[0], Args[1], 105 llvm::SequentiallyConsistent); 106 Result = EmitFromInt(CGF, Result, T, ValueType); 107 return RValue::get(Result); 108 } 109 110 /// Utility to insert an atomic instruction based Instrinsic::ID and 111 /// the expression node, where the return value is the result of the 112 /// operation. 113 static RValue EmitBinaryAtomicPost(CodeGenFunction &CGF, 114 llvm::AtomicRMWInst::BinOp Kind, 115 const CallExpr *E, 116 Instruction::BinaryOps Op) { 117 QualType T = E->getType(); 118 assert(E->getArg(0)->getType()->isPointerType()); 119 assert(CGF.getContext().hasSameUnqualifiedType(T, 120 E->getArg(0)->getType()->getPointeeType())); 121 assert(CGF.getContext().hasSameUnqualifiedType(T, E->getArg(1)->getType())); 122 123 llvm::Value *DestPtr = CGF.EmitScalarExpr(E->getArg(0)); 124 unsigned AddrSpace = DestPtr->getType()->getPointerAddressSpace(); 125 126 llvm::IntegerType *IntType = 127 llvm::IntegerType::get(CGF.getLLVMContext(), 128 CGF.getContext().getTypeSize(T)); 129 llvm::Type *IntPtrType = IntType->getPointerTo(AddrSpace); 130 131 llvm::Value *Args[2]; 132 Args[1] = CGF.EmitScalarExpr(E->getArg(1)); 133 llvm::Type *ValueType = Args[1]->getType(); 134 Args[1] = EmitToInt(CGF, Args[1], T, IntType); 135 Args[0] = CGF.Builder.CreateBitCast(DestPtr, IntPtrType); 136 137 llvm::Value *Result = 138 CGF.Builder.CreateAtomicRMW(Kind, Args[0], Args[1], 139 llvm::SequentiallyConsistent); 140 Result = CGF.Builder.CreateBinOp(Op, Result, Args[1]); 141 Result = EmitFromInt(CGF, Result, T, ValueType); 142 return RValue::get(Result); 143 } 144 145 /// EmitFAbs - Emit a call to fabs/fabsf/fabsl, depending on the type of ValTy, 146 /// which must be a scalar floating point type. 147 static Value *EmitFAbs(CodeGenFunction &CGF, Value *V, QualType ValTy) { 148 const BuiltinType *ValTyP = ValTy->getAs<BuiltinType>(); 149 assert(ValTyP && "isn't scalar fp type!"); 150 151 StringRef FnName; 152 switch (ValTyP->getKind()) { 153 default: llvm_unreachable("Isn't a scalar fp type!"); 154 case BuiltinType::Float: FnName = "fabsf"; break; 155 case BuiltinType::Double: FnName = "fabs"; break; 156 case BuiltinType::LongDouble: FnName = "fabsl"; break; 157 } 158 159 // The prototype is something that takes and returns whatever V's type is. 160 llvm::FunctionType *FT = llvm::FunctionType::get(V->getType(), V->getType(), 161 false); 162 llvm::Value *Fn = CGF.CGM.CreateRuntimeFunction(FT, FnName); 163 164 return CGF.EmitNounwindRuntimeCall(Fn, V, "abs"); 165 } 166 167 static RValue emitLibraryCall(CodeGenFunction &CGF, const FunctionDecl *Fn, 168 const CallExpr *E, llvm::Value *calleeValue) { 169 return CGF.EmitCall(E->getCallee()->getType(), calleeValue, E->getLocStart(), 170 ReturnValueSlot(), E->arg_begin(), E->arg_end(), Fn); 171 } 172 173 /// \brief Emit a call to llvm.{sadd,uadd,ssub,usub,smul,umul}.with.overflow.* 174 /// depending on IntrinsicID. 175 /// 176 /// \arg CGF The current codegen function. 177 /// \arg IntrinsicID The ID for the Intrinsic we wish to generate. 178 /// \arg X The first argument to the llvm.*.with.overflow.*. 179 /// \arg Y The second argument to the llvm.*.with.overflow.*. 180 /// \arg Carry The carry returned by the llvm.*.with.overflow.*. 181 /// \returns The result (i.e. sum/product) returned by the intrinsic. 182 static llvm::Value *EmitOverflowIntrinsic(CodeGenFunction &CGF, 183 const llvm::Intrinsic::ID IntrinsicID, 184 llvm::Value *X, llvm::Value *Y, 185 llvm::Value *&Carry) { 186 // Make sure we have integers of the same width. 187 assert(X->getType() == Y->getType() && 188 "Arguments must be the same type. (Did you forget to make sure both " 189 "arguments have the same integer width?)"); 190 191 llvm::Value *Callee = CGF.CGM.getIntrinsic(IntrinsicID, X->getType()); 192 llvm::Value *Tmp = CGF.Builder.CreateCall2(Callee, X, Y); 193 Carry = CGF.Builder.CreateExtractValue(Tmp, 1); 194 return CGF.Builder.CreateExtractValue(Tmp, 0); 195 } 196 197 RValue CodeGenFunction::EmitBuiltinExpr(const FunctionDecl *FD, 198 unsigned BuiltinID, const CallExpr *E) { 199 // See if we can constant fold this builtin. If so, don't emit it at all. 200 Expr::EvalResult Result; 201 if (E->EvaluateAsRValue(Result, CGM.getContext()) && 202 !Result.hasSideEffects()) { 203 if (Result.Val.isInt()) 204 return RValue::get(llvm::ConstantInt::get(getLLVMContext(), 205 Result.Val.getInt())); 206 if (Result.Val.isFloat()) 207 return RValue::get(llvm::ConstantFP::get(getLLVMContext(), 208 Result.Val.getFloat())); 209 } 210 211 switch (BuiltinID) { 212 default: break; // Handle intrinsics and libm functions below. 213 case Builtin::BI__builtin___CFStringMakeConstantString: 214 case Builtin::BI__builtin___NSStringMakeConstantString: 215 return RValue::get(CGM.EmitConstantExpr(E, E->getType(), 0)); 216 case Builtin::BI__builtin_stdarg_start: 217 case Builtin::BI__builtin_va_start: 218 case Builtin::BI__builtin_va_end: { 219 Value *ArgValue = EmitVAListRef(E->getArg(0)); 220 llvm::Type *DestType = Int8PtrTy; 221 if (ArgValue->getType() != DestType) 222 ArgValue = Builder.CreateBitCast(ArgValue, DestType, 223 ArgValue->getName().data()); 224 225 Intrinsic::ID inst = (BuiltinID == Builtin::BI__builtin_va_end) ? 226 Intrinsic::vaend : Intrinsic::vastart; 227 return RValue::get(Builder.CreateCall(CGM.getIntrinsic(inst), ArgValue)); 228 } 229 case Builtin::BI__builtin_va_copy: { 230 Value *DstPtr = EmitVAListRef(E->getArg(0)); 231 Value *SrcPtr = EmitVAListRef(E->getArg(1)); 232 233 llvm::Type *Type = Int8PtrTy; 234 235 DstPtr = Builder.CreateBitCast(DstPtr, Type); 236 SrcPtr = Builder.CreateBitCast(SrcPtr, Type); 237 return RValue::get(Builder.CreateCall2(CGM.getIntrinsic(Intrinsic::vacopy), 238 DstPtr, SrcPtr)); 239 } 240 case Builtin::BI__builtin_abs: 241 case Builtin::BI__builtin_labs: 242 case Builtin::BI__builtin_llabs: { 243 Value *ArgValue = EmitScalarExpr(E->getArg(0)); 244 245 Value *NegOp = Builder.CreateNeg(ArgValue, "neg"); 246 Value *CmpResult = 247 Builder.CreateICmpSGE(ArgValue, 248 llvm::Constant::getNullValue(ArgValue->getType()), 249 "abscond"); 250 Value *Result = 251 Builder.CreateSelect(CmpResult, ArgValue, NegOp, "abs"); 252 253 return RValue::get(Result); 254 } 255 256 case Builtin::BI__builtin_conj: 257 case Builtin::BI__builtin_conjf: 258 case Builtin::BI__builtin_conjl: { 259 ComplexPairTy ComplexVal = EmitComplexExpr(E->getArg(0)); 260 Value *Real = ComplexVal.first; 261 Value *Imag = ComplexVal.second; 262 Value *Zero = 263 Imag->getType()->isFPOrFPVectorTy() 264 ? llvm::ConstantFP::getZeroValueForNegation(Imag->getType()) 265 : llvm::Constant::getNullValue(Imag->getType()); 266 267 Imag = Builder.CreateFSub(Zero, Imag, "sub"); 268 return RValue::getComplex(std::make_pair(Real, Imag)); 269 } 270 case Builtin::BI__builtin_creal: 271 case Builtin::BI__builtin_crealf: 272 case Builtin::BI__builtin_creall: 273 case Builtin::BIcreal: 274 case Builtin::BIcrealf: 275 case Builtin::BIcreall: { 276 ComplexPairTy ComplexVal = EmitComplexExpr(E->getArg(0)); 277 return RValue::get(ComplexVal.first); 278 } 279 280 case Builtin::BI__builtin_cimag: 281 case Builtin::BI__builtin_cimagf: 282 case Builtin::BI__builtin_cimagl: 283 case Builtin::BIcimag: 284 case Builtin::BIcimagf: 285 case Builtin::BIcimagl: { 286 ComplexPairTy ComplexVal = EmitComplexExpr(E->getArg(0)); 287 return RValue::get(ComplexVal.second); 288 } 289 290 case Builtin::BI__builtin_ctzs: 291 case Builtin::BI__builtin_ctz: 292 case Builtin::BI__builtin_ctzl: 293 case Builtin::BI__builtin_ctzll: { 294 Value *ArgValue = EmitScalarExpr(E->getArg(0)); 295 296 llvm::Type *ArgType = ArgValue->getType(); 297 Value *F = CGM.getIntrinsic(Intrinsic::cttz, ArgType); 298 299 llvm::Type *ResultType = ConvertType(E->getType()); 300 Value *ZeroUndef = Builder.getInt1(getTarget().isCLZForZeroUndef()); 301 Value *Result = Builder.CreateCall2(F, ArgValue, ZeroUndef); 302 if (Result->getType() != ResultType) 303 Result = Builder.CreateIntCast(Result, ResultType, /*isSigned*/true, 304 "cast"); 305 return RValue::get(Result); 306 } 307 case Builtin::BI__builtin_clzs: 308 case Builtin::BI__builtin_clz: 309 case Builtin::BI__builtin_clzl: 310 case Builtin::BI__builtin_clzll: { 311 Value *ArgValue = EmitScalarExpr(E->getArg(0)); 312 313 llvm::Type *ArgType = ArgValue->getType(); 314 Value *F = CGM.getIntrinsic(Intrinsic::ctlz, ArgType); 315 316 llvm::Type *ResultType = ConvertType(E->getType()); 317 Value *ZeroUndef = Builder.getInt1(getTarget().isCLZForZeroUndef()); 318 Value *Result = Builder.CreateCall2(F, ArgValue, ZeroUndef); 319 if (Result->getType() != ResultType) 320 Result = Builder.CreateIntCast(Result, ResultType, /*isSigned*/true, 321 "cast"); 322 return RValue::get(Result); 323 } 324 case Builtin::BI__builtin_ffs: 325 case Builtin::BI__builtin_ffsl: 326 case Builtin::BI__builtin_ffsll: { 327 // ffs(x) -> x ? cttz(x) + 1 : 0 328 Value *ArgValue = EmitScalarExpr(E->getArg(0)); 329 330 llvm::Type *ArgType = ArgValue->getType(); 331 Value *F = CGM.getIntrinsic(Intrinsic::cttz, ArgType); 332 333 llvm::Type *ResultType = ConvertType(E->getType()); 334 Value *Tmp = Builder.CreateAdd(Builder.CreateCall2(F, ArgValue, 335 Builder.getTrue()), 336 llvm::ConstantInt::get(ArgType, 1)); 337 Value *Zero = llvm::Constant::getNullValue(ArgType); 338 Value *IsZero = Builder.CreateICmpEQ(ArgValue, Zero, "iszero"); 339 Value *Result = Builder.CreateSelect(IsZero, Zero, Tmp, "ffs"); 340 if (Result->getType() != ResultType) 341 Result = Builder.CreateIntCast(Result, ResultType, /*isSigned*/true, 342 "cast"); 343 return RValue::get(Result); 344 } 345 case Builtin::BI__builtin_parity: 346 case Builtin::BI__builtin_parityl: 347 case Builtin::BI__builtin_parityll: { 348 // parity(x) -> ctpop(x) & 1 349 Value *ArgValue = EmitScalarExpr(E->getArg(0)); 350 351 llvm::Type *ArgType = ArgValue->getType(); 352 Value *F = CGM.getIntrinsic(Intrinsic::ctpop, ArgType); 353 354 llvm::Type *ResultType = ConvertType(E->getType()); 355 Value *Tmp = Builder.CreateCall(F, ArgValue); 356 Value *Result = Builder.CreateAnd(Tmp, llvm::ConstantInt::get(ArgType, 1)); 357 if (Result->getType() != ResultType) 358 Result = Builder.CreateIntCast(Result, ResultType, /*isSigned*/true, 359 "cast"); 360 return RValue::get(Result); 361 } 362 case Builtin::BI__builtin_popcount: 363 case Builtin::BI__builtin_popcountl: 364 case Builtin::BI__builtin_popcountll: { 365 Value *ArgValue = EmitScalarExpr(E->getArg(0)); 366 367 llvm::Type *ArgType = ArgValue->getType(); 368 Value *F = CGM.getIntrinsic(Intrinsic::ctpop, ArgType); 369 370 llvm::Type *ResultType = ConvertType(E->getType()); 371 Value *Result = Builder.CreateCall(F, ArgValue); 372 if (Result->getType() != ResultType) 373 Result = Builder.CreateIntCast(Result, ResultType, /*isSigned*/true, 374 "cast"); 375 return RValue::get(Result); 376 } 377 case Builtin::BI__builtin_expect: { 378 Value *ArgValue = EmitScalarExpr(E->getArg(0)); 379 llvm::Type *ArgType = ArgValue->getType(); 380 381 Value *FnExpect = CGM.getIntrinsic(Intrinsic::expect, ArgType); 382 Value *ExpectedValue = EmitScalarExpr(E->getArg(1)); 383 384 Value *Result = Builder.CreateCall2(FnExpect, ArgValue, ExpectedValue, 385 "expval"); 386 return RValue::get(Result); 387 } 388 case Builtin::BI__builtin_bswap16: 389 case Builtin::BI__builtin_bswap32: 390 case Builtin::BI__builtin_bswap64: { 391 Value *ArgValue = EmitScalarExpr(E->getArg(0)); 392 llvm::Type *ArgType = ArgValue->getType(); 393 Value *F = CGM.getIntrinsic(Intrinsic::bswap, ArgType); 394 return RValue::get(Builder.CreateCall(F, ArgValue)); 395 } 396 case Builtin::BI__builtin_object_size: { 397 // We rely on constant folding to deal with expressions with side effects. 398 assert(!E->getArg(0)->HasSideEffects(getContext()) && 399 "should have been constant folded"); 400 401 // We pass this builtin onto the optimizer so that it can 402 // figure out the object size in more complex cases. 403 llvm::Type *ResType = ConvertType(E->getType()); 404 405 // LLVM only supports 0 and 2, make sure that we pass along that 406 // as a boolean. 407 Value *Ty = EmitScalarExpr(E->getArg(1)); 408 ConstantInt *CI = dyn_cast<ConstantInt>(Ty); 409 assert(CI); 410 uint64_t val = CI->getZExtValue(); 411 CI = ConstantInt::get(Builder.getInt1Ty(), (val & 0x2) >> 1); 412 // FIXME: Get right address space. 413 llvm::Type *Tys[] = { ResType, Builder.getInt8PtrTy(0) }; 414 Value *F = CGM.getIntrinsic(Intrinsic::objectsize, Tys); 415 return RValue::get(Builder.CreateCall2(F, EmitScalarExpr(E->getArg(0)),CI)); 416 } 417 case Builtin::BI__builtin_prefetch: { 418 Value *Locality, *RW, *Address = EmitScalarExpr(E->getArg(0)); 419 // FIXME: Technically these constants should of type 'int', yes? 420 RW = (E->getNumArgs() > 1) ? EmitScalarExpr(E->getArg(1)) : 421 llvm::ConstantInt::get(Int32Ty, 0); 422 Locality = (E->getNumArgs() > 2) ? EmitScalarExpr(E->getArg(2)) : 423 llvm::ConstantInt::get(Int32Ty, 3); 424 Value *Data = llvm::ConstantInt::get(Int32Ty, 1); 425 Value *F = CGM.getIntrinsic(Intrinsic::prefetch); 426 return RValue::get(Builder.CreateCall4(F, Address, RW, Locality, Data)); 427 } 428 case Builtin::BI__builtin_readcyclecounter: { 429 Value *F = CGM.getIntrinsic(Intrinsic::readcyclecounter); 430 return RValue::get(Builder.CreateCall(F)); 431 } 432 case Builtin::BI__builtin_trap: { 433 Value *F = CGM.getIntrinsic(Intrinsic::trap); 434 return RValue::get(Builder.CreateCall(F)); 435 } 436 case Builtin::BI__debugbreak: { 437 Value *F = CGM.getIntrinsic(Intrinsic::debugtrap); 438 return RValue::get(Builder.CreateCall(F)); 439 } 440 case Builtin::BI__builtin_unreachable: { 441 if (SanOpts->Unreachable) 442 EmitCheck(Builder.getFalse(), "builtin_unreachable", 443 EmitCheckSourceLocation(E->getExprLoc()), 444 ArrayRef<llvm::Value *>(), CRK_Unrecoverable); 445 else 446 Builder.CreateUnreachable(); 447 448 // We do need to preserve an insertion point. 449 EmitBlock(createBasicBlock("unreachable.cont")); 450 451 return RValue::get(0); 452 } 453 454 case Builtin::BI__builtin_powi: 455 case Builtin::BI__builtin_powif: 456 case Builtin::BI__builtin_powil: { 457 Value *Base = EmitScalarExpr(E->getArg(0)); 458 Value *Exponent = EmitScalarExpr(E->getArg(1)); 459 llvm::Type *ArgType = Base->getType(); 460 Value *F = CGM.getIntrinsic(Intrinsic::powi, ArgType); 461 return RValue::get(Builder.CreateCall2(F, Base, Exponent)); 462 } 463 464 case Builtin::BI__builtin_isgreater: 465 case Builtin::BI__builtin_isgreaterequal: 466 case Builtin::BI__builtin_isless: 467 case Builtin::BI__builtin_islessequal: 468 case Builtin::BI__builtin_islessgreater: 469 case Builtin::BI__builtin_isunordered: { 470 // Ordered comparisons: we know the arguments to these are matching scalar 471 // floating point values. 472 Value *LHS = EmitScalarExpr(E->getArg(0)); 473 Value *RHS = EmitScalarExpr(E->getArg(1)); 474 475 switch (BuiltinID) { 476 default: llvm_unreachable("Unknown ordered comparison"); 477 case Builtin::BI__builtin_isgreater: 478 LHS = Builder.CreateFCmpOGT(LHS, RHS, "cmp"); 479 break; 480 case Builtin::BI__builtin_isgreaterequal: 481 LHS = Builder.CreateFCmpOGE(LHS, RHS, "cmp"); 482 break; 483 case Builtin::BI__builtin_isless: 484 LHS = Builder.CreateFCmpOLT(LHS, RHS, "cmp"); 485 break; 486 case Builtin::BI__builtin_islessequal: 487 LHS = Builder.CreateFCmpOLE(LHS, RHS, "cmp"); 488 break; 489 case Builtin::BI__builtin_islessgreater: 490 LHS = Builder.CreateFCmpONE(LHS, RHS, "cmp"); 491 break; 492 case Builtin::BI__builtin_isunordered: 493 LHS = Builder.CreateFCmpUNO(LHS, RHS, "cmp"); 494 break; 495 } 496 // ZExt bool to int type. 497 return RValue::get(Builder.CreateZExt(LHS, ConvertType(E->getType()))); 498 } 499 case Builtin::BI__builtin_isnan: { 500 Value *V = EmitScalarExpr(E->getArg(0)); 501 V = Builder.CreateFCmpUNO(V, V, "cmp"); 502 return RValue::get(Builder.CreateZExt(V, ConvertType(E->getType()))); 503 } 504 505 case Builtin::BI__builtin_isinf: { 506 // isinf(x) --> fabs(x) == infinity 507 Value *V = EmitScalarExpr(E->getArg(0)); 508 V = EmitFAbs(*this, V, E->getArg(0)->getType()); 509 510 V = Builder.CreateFCmpOEQ(V, ConstantFP::getInfinity(V->getType()),"isinf"); 511 return RValue::get(Builder.CreateZExt(V, ConvertType(E->getType()))); 512 } 513 514 // TODO: BI__builtin_isinf_sign 515 // isinf_sign(x) -> isinf(x) ? (signbit(x) ? -1 : 1) : 0 516 517 case Builtin::BI__builtin_isnormal: { 518 // isnormal(x) --> x == x && fabsf(x) < infinity && fabsf(x) >= float_min 519 Value *V = EmitScalarExpr(E->getArg(0)); 520 Value *Eq = Builder.CreateFCmpOEQ(V, V, "iseq"); 521 522 Value *Abs = EmitFAbs(*this, V, E->getArg(0)->getType()); 523 Value *IsLessThanInf = 524 Builder.CreateFCmpULT(Abs, ConstantFP::getInfinity(V->getType()),"isinf"); 525 APFloat Smallest = APFloat::getSmallestNormalized( 526 getContext().getFloatTypeSemantics(E->getArg(0)->getType())); 527 Value *IsNormal = 528 Builder.CreateFCmpUGE(Abs, ConstantFP::get(V->getContext(), Smallest), 529 "isnormal"); 530 V = Builder.CreateAnd(Eq, IsLessThanInf, "and"); 531 V = Builder.CreateAnd(V, IsNormal, "and"); 532 return RValue::get(Builder.CreateZExt(V, ConvertType(E->getType()))); 533 } 534 535 case Builtin::BI__builtin_isfinite: { 536 // isfinite(x) --> x == x && fabs(x) != infinity; 537 Value *V = EmitScalarExpr(E->getArg(0)); 538 Value *Eq = Builder.CreateFCmpOEQ(V, V, "iseq"); 539 540 Value *Abs = EmitFAbs(*this, V, E->getArg(0)->getType()); 541 Value *IsNotInf = 542 Builder.CreateFCmpUNE(Abs, ConstantFP::getInfinity(V->getType()),"isinf"); 543 544 V = Builder.CreateAnd(Eq, IsNotInf, "and"); 545 return RValue::get(Builder.CreateZExt(V, ConvertType(E->getType()))); 546 } 547 548 case Builtin::BI__builtin_fpclassify: { 549 Value *V = EmitScalarExpr(E->getArg(5)); 550 llvm::Type *Ty = ConvertType(E->getArg(5)->getType()); 551 552 // Create Result 553 BasicBlock *Begin = Builder.GetInsertBlock(); 554 BasicBlock *End = createBasicBlock("fpclassify_end", this->CurFn); 555 Builder.SetInsertPoint(End); 556 PHINode *Result = 557 Builder.CreatePHI(ConvertType(E->getArg(0)->getType()), 4, 558 "fpclassify_result"); 559 560 // if (V==0) return FP_ZERO 561 Builder.SetInsertPoint(Begin); 562 Value *IsZero = Builder.CreateFCmpOEQ(V, Constant::getNullValue(Ty), 563 "iszero"); 564 Value *ZeroLiteral = EmitScalarExpr(E->getArg(4)); 565 BasicBlock *NotZero = createBasicBlock("fpclassify_not_zero", this->CurFn); 566 Builder.CreateCondBr(IsZero, End, NotZero); 567 Result->addIncoming(ZeroLiteral, Begin); 568 569 // if (V != V) return FP_NAN 570 Builder.SetInsertPoint(NotZero); 571 Value *IsNan = Builder.CreateFCmpUNO(V, V, "cmp"); 572 Value *NanLiteral = EmitScalarExpr(E->getArg(0)); 573 BasicBlock *NotNan = createBasicBlock("fpclassify_not_nan", this->CurFn); 574 Builder.CreateCondBr(IsNan, End, NotNan); 575 Result->addIncoming(NanLiteral, NotZero); 576 577 // if (fabs(V) == infinity) return FP_INFINITY 578 Builder.SetInsertPoint(NotNan); 579 Value *VAbs = EmitFAbs(*this, V, E->getArg(5)->getType()); 580 Value *IsInf = 581 Builder.CreateFCmpOEQ(VAbs, ConstantFP::getInfinity(V->getType()), 582 "isinf"); 583 Value *InfLiteral = EmitScalarExpr(E->getArg(1)); 584 BasicBlock *NotInf = createBasicBlock("fpclassify_not_inf", this->CurFn); 585 Builder.CreateCondBr(IsInf, End, NotInf); 586 Result->addIncoming(InfLiteral, NotNan); 587 588 // if (fabs(V) >= MIN_NORMAL) return FP_NORMAL else FP_SUBNORMAL 589 Builder.SetInsertPoint(NotInf); 590 APFloat Smallest = APFloat::getSmallestNormalized( 591 getContext().getFloatTypeSemantics(E->getArg(5)->getType())); 592 Value *IsNormal = 593 Builder.CreateFCmpUGE(VAbs, ConstantFP::get(V->getContext(), Smallest), 594 "isnormal"); 595 Value *NormalResult = 596 Builder.CreateSelect(IsNormal, EmitScalarExpr(E->getArg(2)), 597 EmitScalarExpr(E->getArg(3))); 598 Builder.CreateBr(End); 599 Result->addIncoming(NormalResult, NotInf); 600 601 // return Result 602 Builder.SetInsertPoint(End); 603 return RValue::get(Result); 604 } 605 606 case Builtin::BIalloca: 607 case Builtin::BI_alloca: 608 case Builtin::BI__builtin_alloca: { 609 Value *Size = EmitScalarExpr(E->getArg(0)); 610 return RValue::get(Builder.CreateAlloca(Builder.getInt8Ty(), Size)); 611 } 612 case Builtin::BIbzero: 613 case Builtin::BI__builtin_bzero: { 614 std::pair<llvm::Value*, unsigned> Dest = 615 EmitPointerWithAlignment(E->getArg(0)); 616 Value *SizeVal = EmitScalarExpr(E->getArg(1)); 617 Builder.CreateMemSet(Dest.first, Builder.getInt8(0), SizeVal, 618 Dest.second, false); 619 return RValue::get(Dest.first); 620 } 621 case Builtin::BImemcpy: 622 case Builtin::BI__builtin_memcpy: { 623 std::pair<llvm::Value*, unsigned> Dest = 624 EmitPointerWithAlignment(E->getArg(0)); 625 std::pair<llvm::Value*, unsigned> Src = 626 EmitPointerWithAlignment(E->getArg(1)); 627 Value *SizeVal = EmitScalarExpr(E->getArg(2)); 628 unsigned Align = std::min(Dest.second, Src.second); 629 Builder.CreateMemCpy(Dest.first, Src.first, SizeVal, Align, false); 630 return RValue::get(Dest.first); 631 } 632 633 case Builtin::BI__builtin___memcpy_chk: { 634 // fold __builtin_memcpy_chk(x, y, cst1, cst2) to memcpy iff cst1<=cst2. 635 llvm::APSInt Size, DstSize; 636 if (!E->getArg(2)->EvaluateAsInt(Size, CGM.getContext()) || 637 !E->getArg(3)->EvaluateAsInt(DstSize, CGM.getContext())) 638 break; 639 if (Size.ugt(DstSize)) 640 break; 641 std::pair<llvm::Value*, unsigned> Dest = 642 EmitPointerWithAlignment(E->getArg(0)); 643 std::pair<llvm::Value*, unsigned> Src = 644 EmitPointerWithAlignment(E->getArg(1)); 645 Value *SizeVal = llvm::ConstantInt::get(Builder.getContext(), Size); 646 unsigned Align = std::min(Dest.second, Src.second); 647 Builder.CreateMemCpy(Dest.first, Src.first, SizeVal, Align, false); 648 return RValue::get(Dest.first); 649 } 650 651 case Builtin::BI__builtin_objc_memmove_collectable: { 652 Value *Address = EmitScalarExpr(E->getArg(0)); 653 Value *SrcAddr = EmitScalarExpr(E->getArg(1)); 654 Value *SizeVal = EmitScalarExpr(E->getArg(2)); 655 CGM.getObjCRuntime().EmitGCMemmoveCollectable(*this, 656 Address, SrcAddr, SizeVal); 657 return RValue::get(Address); 658 } 659 660 case Builtin::BI__builtin___memmove_chk: { 661 // fold __builtin_memmove_chk(x, y, cst1, cst2) to memmove iff cst1<=cst2. 662 llvm::APSInt Size, DstSize; 663 if (!E->getArg(2)->EvaluateAsInt(Size, CGM.getContext()) || 664 !E->getArg(3)->EvaluateAsInt(DstSize, CGM.getContext())) 665 break; 666 if (Size.ugt(DstSize)) 667 break; 668 std::pair<llvm::Value*, unsigned> Dest = 669 EmitPointerWithAlignment(E->getArg(0)); 670 std::pair<llvm::Value*, unsigned> Src = 671 EmitPointerWithAlignment(E->getArg(1)); 672 Value *SizeVal = llvm::ConstantInt::get(Builder.getContext(), Size); 673 unsigned Align = std::min(Dest.second, Src.second); 674 Builder.CreateMemMove(Dest.first, Src.first, SizeVal, Align, false); 675 return RValue::get(Dest.first); 676 } 677 678 case Builtin::BImemmove: 679 case Builtin::BI__builtin_memmove: { 680 std::pair<llvm::Value*, unsigned> Dest = 681 EmitPointerWithAlignment(E->getArg(0)); 682 std::pair<llvm::Value*, unsigned> Src = 683 EmitPointerWithAlignment(E->getArg(1)); 684 Value *SizeVal = EmitScalarExpr(E->getArg(2)); 685 unsigned Align = std::min(Dest.second, Src.second); 686 Builder.CreateMemMove(Dest.first, Src.first, SizeVal, Align, false); 687 return RValue::get(Dest.first); 688 } 689 case Builtin::BImemset: 690 case Builtin::BI__builtin_memset: { 691 std::pair<llvm::Value*, unsigned> Dest = 692 EmitPointerWithAlignment(E->getArg(0)); 693 Value *ByteVal = Builder.CreateTrunc(EmitScalarExpr(E->getArg(1)), 694 Builder.getInt8Ty()); 695 Value *SizeVal = EmitScalarExpr(E->getArg(2)); 696 Builder.CreateMemSet(Dest.first, ByteVal, SizeVal, Dest.second, false); 697 return RValue::get(Dest.first); 698 } 699 case Builtin::BI__builtin___memset_chk: { 700 // fold __builtin_memset_chk(x, y, cst1, cst2) to memset iff cst1<=cst2. 701 llvm::APSInt Size, DstSize; 702 if (!E->getArg(2)->EvaluateAsInt(Size, CGM.getContext()) || 703 !E->getArg(3)->EvaluateAsInt(DstSize, CGM.getContext())) 704 break; 705 if (Size.ugt(DstSize)) 706 break; 707 std::pair<llvm::Value*, unsigned> Dest = 708 EmitPointerWithAlignment(E->getArg(0)); 709 Value *ByteVal = Builder.CreateTrunc(EmitScalarExpr(E->getArg(1)), 710 Builder.getInt8Ty()); 711 Value *SizeVal = llvm::ConstantInt::get(Builder.getContext(), Size); 712 Builder.CreateMemSet(Dest.first, ByteVal, SizeVal, Dest.second, false); 713 return RValue::get(Dest.first); 714 } 715 case Builtin::BI__builtin_dwarf_cfa: { 716 // The offset in bytes from the first argument to the CFA. 717 // 718 // Why on earth is this in the frontend? Is there any reason at 719 // all that the backend can't reasonably determine this while 720 // lowering llvm.eh.dwarf.cfa()? 721 // 722 // TODO: If there's a satisfactory reason, add a target hook for 723 // this instead of hard-coding 0, which is correct for most targets. 724 int32_t Offset = 0; 725 726 Value *F = CGM.getIntrinsic(Intrinsic::eh_dwarf_cfa); 727 return RValue::get(Builder.CreateCall(F, 728 llvm::ConstantInt::get(Int32Ty, Offset))); 729 } 730 case Builtin::BI__builtin_return_address: { 731 Value *Depth = EmitScalarExpr(E->getArg(0)); 732 Depth = Builder.CreateIntCast(Depth, Int32Ty, false); 733 Value *F = CGM.getIntrinsic(Intrinsic::returnaddress); 734 return RValue::get(Builder.CreateCall(F, Depth)); 735 } 736 case Builtin::BI__builtin_frame_address: { 737 Value *Depth = EmitScalarExpr(E->getArg(0)); 738 Depth = Builder.CreateIntCast(Depth, Int32Ty, false); 739 Value *F = CGM.getIntrinsic(Intrinsic::frameaddress); 740 return RValue::get(Builder.CreateCall(F, Depth)); 741 } 742 case Builtin::BI__builtin_extract_return_addr: { 743 Value *Address = EmitScalarExpr(E->getArg(0)); 744 Value *Result = getTargetHooks().decodeReturnAddress(*this, Address); 745 return RValue::get(Result); 746 } 747 case Builtin::BI__builtin_frob_return_addr: { 748 Value *Address = EmitScalarExpr(E->getArg(0)); 749 Value *Result = getTargetHooks().encodeReturnAddress(*this, Address); 750 return RValue::get(Result); 751 } 752 case Builtin::BI__builtin_dwarf_sp_column: { 753 llvm::IntegerType *Ty 754 = cast<llvm::IntegerType>(ConvertType(E->getType())); 755 int Column = getTargetHooks().getDwarfEHStackPointer(CGM); 756 if (Column == -1) { 757 CGM.ErrorUnsupported(E, "__builtin_dwarf_sp_column"); 758 return RValue::get(llvm::UndefValue::get(Ty)); 759 } 760 return RValue::get(llvm::ConstantInt::get(Ty, Column, true)); 761 } 762 case Builtin::BI__builtin_init_dwarf_reg_size_table: { 763 Value *Address = EmitScalarExpr(E->getArg(0)); 764 if (getTargetHooks().initDwarfEHRegSizeTable(*this, Address)) 765 CGM.ErrorUnsupported(E, "__builtin_init_dwarf_reg_size_table"); 766 return RValue::get(llvm::UndefValue::get(ConvertType(E->getType()))); 767 } 768 case Builtin::BI__builtin_eh_return: { 769 Value *Int = EmitScalarExpr(E->getArg(0)); 770 Value *Ptr = EmitScalarExpr(E->getArg(1)); 771 772 llvm::IntegerType *IntTy = cast<llvm::IntegerType>(Int->getType()); 773 assert((IntTy->getBitWidth() == 32 || IntTy->getBitWidth() == 64) && 774 "LLVM's __builtin_eh_return only supports 32- and 64-bit variants"); 775 Value *F = CGM.getIntrinsic(IntTy->getBitWidth() == 32 776 ? Intrinsic::eh_return_i32 777 : Intrinsic::eh_return_i64); 778 Builder.CreateCall2(F, Int, Ptr); 779 Builder.CreateUnreachable(); 780 781 // We do need to preserve an insertion point. 782 EmitBlock(createBasicBlock("builtin_eh_return.cont")); 783 784 return RValue::get(0); 785 } 786 case Builtin::BI__builtin_unwind_init: { 787 Value *F = CGM.getIntrinsic(Intrinsic::eh_unwind_init); 788 return RValue::get(Builder.CreateCall(F)); 789 } 790 case Builtin::BI__builtin_extend_pointer: { 791 // Extends a pointer to the size of an _Unwind_Word, which is 792 // uint64_t on all platforms. Generally this gets poked into a 793 // register and eventually used as an address, so if the 794 // addressing registers are wider than pointers and the platform 795 // doesn't implicitly ignore high-order bits when doing 796 // addressing, we need to make sure we zext / sext based on 797 // the platform's expectations. 798 // 799 // See: http://gcc.gnu.org/ml/gcc-bugs/2002-02/msg00237.html 800 801 // Cast the pointer to intptr_t. 802 Value *Ptr = EmitScalarExpr(E->getArg(0)); 803 Value *Result = Builder.CreatePtrToInt(Ptr, IntPtrTy, "extend.cast"); 804 805 // If that's 64 bits, we're done. 806 if (IntPtrTy->getBitWidth() == 64) 807 return RValue::get(Result); 808 809 // Otherwise, ask the codegen data what to do. 810 if (getTargetHooks().extendPointerWithSExt()) 811 return RValue::get(Builder.CreateSExt(Result, Int64Ty, "extend.sext")); 812 else 813 return RValue::get(Builder.CreateZExt(Result, Int64Ty, "extend.zext")); 814 } 815 case Builtin::BI__builtin_setjmp: { 816 // Buffer is a void**. 817 Value *Buf = EmitScalarExpr(E->getArg(0)); 818 819 // Store the frame pointer to the setjmp buffer. 820 Value *FrameAddr = 821 Builder.CreateCall(CGM.getIntrinsic(Intrinsic::frameaddress), 822 ConstantInt::get(Int32Ty, 0)); 823 Builder.CreateStore(FrameAddr, Buf); 824 825 // Store the stack pointer to the setjmp buffer. 826 Value *StackAddr = 827 Builder.CreateCall(CGM.getIntrinsic(Intrinsic::stacksave)); 828 Value *StackSaveSlot = 829 Builder.CreateGEP(Buf, ConstantInt::get(Int32Ty, 2)); 830 Builder.CreateStore(StackAddr, StackSaveSlot); 831 832 // Call LLVM's EH setjmp, which is lightweight. 833 Value *F = CGM.getIntrinsic(Intrinsic::eh_sjlj_setjmp); 834 Buf = Builder.CreateBitCast(Buf, Int8PtrTy); 835 return RValue::get(Builder.CreateCall(F, Buf)); 836 } 837 case Builtin::BI__builtin_longjmp: { 838 Value *Buf = EmitScalarExpr(E->getArg(0)); 839 Buf = Builder.CreateBitCast(Buf, Int8PtrTy); 840 841 // Call LLVM's EH longjmp, which is lightweight. 842 Builder.CreateCall(CGM.getIntrinsic(Intrinsic::eh_sjlj_longjmp), Buf); 843 844 // longjmp doesn't return; mark this as unreachable. 845 Builder.CreateUnreachable(); 846 847 // We do need to preserve an insertion point. 848 EmitBlock(createBasicBlock("longjmp.cont")); 849 850 return RValue::get(0); 851 } 852 case Builtin::BI__sync_fetch_and_add: 853 case Builtin::BI__sync_fetch_and_sub: 854 case Builtin::BI__sync_fetch_and_or: 855 case Builtin::BI__sync_fetch_and_and: 856 case Builtin::BI__sync_fetch_and_xor: 857 case Builtin::BI__sync_add_and_fetch: 858 case Builtin::BI__sync_sub_and_fetch: 859 case Builtin::BI__sync_and_and_fetch: 860 case Builtin::BI__sync_or_and_fetch: 861 case Builtin::BI__sync_xor_and_fetch: 862 case Builtin::BI__sync_val_compare_and_swap: 863 case Builtin::BI__sync_bool_compare_and_swap: 864 case Builtin::BI__sync_lock_test_and_set: 865 case Builtin::BI__sync_lock_release: 866 case Builtin::BI__sync_swap: 867 llvm_unreachable("Shouldn't make it through sema"); 868 case Builtin::BI__sync_fetch_and_add_1: 869 case Builtin::BI__sync_fetch_and_add_2: 870 case Builtin::BI__sync_fetch_and_add_4: 871 case Builtin::BI__sync_fetch_and_add_8: 872 case Builtin::BI__sync_fetch_and_add_16: 873 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Add, E); 874 case Builtin::BI__sync_fetch_and_sub_1: 875 case Builtin::BI__sync_fetch_and_sub_2: 876 case Builtin::BI__sync_fetch_and_sub_4: 877 case Builtin::BI__sync_fetch_and_sub_8: 878 case Builtin::BI__sync_fetch_and_sub_16: 879 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Sub, E); 880 case Builtin::BI__sync_fetch_and_or_1: 881 case Builtin::BI__sync_fetch_and_or_2: 882 case Builtin::BI__sync_fetch_and_or_4: 883 case Builtin::BI__sync_fetch_and_or_8: 884 case Builtin::BI__sync_fetch_and_or_16: 885 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Or, E); 886 case Builtin::BI__sync_fetch_and_and_1: 887 case Builtin::BI__sync_fetch_and_and_2: 888 case Builtin::BI__sync_fetch_and_and_4: 889 case Builtin::BI__sync_fetch_and_and_8: 890 case Builtin::BI__sync_fetch_and_and_16: 891 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::And, E); 892 case Builtin::BI__sync_fetch_and_xor_1: 893 case Builtin::BI__sync_fetch_and_xor_2: 894 case Builtin::BI__sync_fetch_and_xor_4: 895 case Builtin::BI__sync_fetch_and_xor_8: 896 case Builtin::BI__sync_fetch_and_xor_16: 897 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Xor, E); 898 899 // Clang extensions: not overloaded yet. 900 case Builtin::BI__sync_fetch_and_min: 901 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Min, E); 902 case Builtin::BI__sync_fetch_and_max: 903 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Max, E); 904 case Builtin::BI__sync_fetch_and_umin: 905 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::UMin, E); 906 case Builtin::BI__sync_fetch_and_umax: 907 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::UMax, E); 908 909 case Builtin::BI__sync_add_and_fetch_1: 910 case Builtin::BI__sync_add_and_fetch_2: 911 case Builtin::BI__sync_add_and_fetch_4: 912 case Builtin::BI__sync_add_and_fetch_8: 913 case Builtin::BI__sync_add_and_fetch_16: 914 return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::Add, E, 915 llvm::Instruction::Add); 916 case Builtin::BI__sync_sub_and_fetch_1: 917 case Builtin::BI__sync_sub_and_fetch_2: 918 case Builtin::BI__sync_sub_and_fetch_4: 919 case Builtin::BI__sync_sub_and_fetch_8: 920 case Builtin::BI__sync_sub_and_fetch_16: 921 return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::Sub, E, 922 llvm::Instruction::Sub); 923 case Builtin::BI__sync_and_and_fetch_1: 924 case Builtin::BI__sync_and_and_fetch_2: 925 case Builtin::BI__sync_and_and_fetch_4: 926 case Builtin::BI__sync_and_and_fetch_8: 927 case Builtin::BI__sync_and_and_fetch_16: 928 return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::And, E, 929 llvm::Instruction::And); 930 case Builtin::BI__sync_or_and_fetch_1: 931 case Builtin::BI__sync_or_and_fetch_2: 932 case Builtin::BI__sync_or_and_fetch_4: 933 case Builtin::BI__sync_or_and_fetch_8: 934 case Builtin::BI__sync_or_and_fetch_16: 935 return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::Or, E, 936 llvm::Instruction::Or); 937 case Builtin::BI__sync_xor_and_fetch_1: 938 case Builtin::BI__sync_xor_and_fetch_2: 939 case Builtin::BI__sync_xor_and_fetch_4: 940 case Builtin::BI__sync_xor_and_fetch_8: 941 case Builtin::BI__sync_xor_and_fetch_16: 942 return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::Xor, E, 943 llvm::Instruction::Xor); 944 945 case Builtin::BI__sync_val_compare_and_swap_1: 946 case Builtin::BI__sync_val_compare_and_swap_2: 947 case Builtin::BI__sync_val_compare_and_swap_4: 948 case Builtin::BI__sync_val_compare_and_swap_8: 949 case Builtin::BI__sync_val_compare_and_swap_16: { 950 QualType T = E->getType(); 951 llvm::Value *DestPtr = EmitScalarExpr(E->getArg(0)); 952 unsigned AddrSpace = DestPtr->getType()->getPointerAddressSpace(); 953 954 llvm::IntegerType *IntType = 955 llvm::IntegerType::get(getLLVMContext(), 956 getContext().getTypeSize(T)); 957 llvm::Type *IntPtrType = IntType->getPointerTo(AddrSpace); 958 959 Value *Args[3]; 960 Args[0] = Builder.CreateBitCast(DestPtr, IntPtrType); 961 Args[1] = EmitScalarExpr(E->getArg(1)); 962 llvm::Type *ValueType = Args[1]->getType(); 963 Args[1] = EmitToInt(*this, Args[1], T, IntType); 964 Args[2] = EmitToInt(*this, EmitScalarExpr(E->getArg(2)), T, IntType); 965 966 Value *Result = Builder.CreateAtomicCmpXchg(Args[0], Args[1], Args[2], 967 llvm::SequentiallyConsistent); 968 Result = EmitFromInt(*this, Result, T, ValueType); 969 return RValue::get(Result); 970 } 971 972 case Builtin::BI__sync_bool_compare_and_swap_1: 973 case Builtin::BI__sync_bool_compare_and_swap_2: 974 case Builtin::BI__sync_bool_compare_and_swap_4: 975 case Builtin::BI__sync_bool_compare_and_swap_8: 976 case Builtin::BI__sync_bool_compare_and_swap_16: { 977 QualType T = E->getArg(1)->getType(); 978 llvm::Value *DestPtr = EmitScalarExpr(E->getArg(0)); 979 unsigned AddrSpace = DestPtr->getType()->getPointerAddressSpace(); 980 981 llvm::IntegerType *IntType = 982 llvm::IntegerType::get(getLLVMContext(), 983 getContext().getTypeSize(T)); 984 llvm::Type *IntPtrType = IntType->getPointerTo(AddrSpace); 985 986 Value *Args[3]; 987 Args[0] = Builder.CreateBitCast(DestPtr, IntPtrType); 988 Args[1] = EmitToInt(*this, EmitScalarExpr(E->getArg(1)), T, IntType); 989 Args[2] = EmitToInt(*this, EmitScalarExpr(E->getArg(2)), T, IntType); 990 991 Value *OldVal = Args[1]; 992 Value *PrevVal = Builder.CreateAtomicCmpXchg(Args[0], Args[1], Args[2], 993 llvm::SequentiallyConsistent); 994 Value *Result = Builder.CreateICmpEQ(PrevVal, OldVal); 995 // zext bool to int. 996 Result = Builder.CreateZExt(Result, ConvertType(E->getType())); 997 return RValue::get(Result); 998 } 999 1000 case Builtin::BI__sync_swap_1: 1001 case Builtin::BI__sync_swap_2: 1002 case Builtin::BI__sync_swap_4: 1003 case Builtin::BI__sync_swap_8: 1004 case Builtin::BI__sync_swap_16: 1005 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Xchg, E); 1006 1007 case Builtin::BI__sync_lock_test_and_set_1: 1008 case Builtin::BI__sync_lock_test_and_set_2: 1009 case Builtin::BI__sync_lock_test_and_set_4: 1010 case Builtin::BI__sync_lock_test_and_set_8: 1011 case Builtin::BI__sync_lock_test_and_set_16: 1012 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Xchg, E); 1013 1014 case Builtin::BI__sync_lock_release_1: 1015 case Builtin::BI__sync_lock_release_2: 1016 case Builtin::BI__sync_lock_release_4: 1017 case Builtin::BI__sync_lock_release_8: 1018 case Builtin::BI__sync_lock_release_16: { 1019 Value *Ptr = EmitScalarExpr(E->getArg(0)); 1020 QualType ElTy = E->getArg(0)->getType()->getPointeeType(); 1021 CharUnits StoreSize = getContext().getTypeSizeInChars(ElTy); 1022 llvm::Type *ITy = llvm::IntegerType::get(getLLVMContext(), 1023 StoreSize.getQuantity() * 8); 1024 Ptr = Builder.CreateBitCast(Ptr, ITy->getPointerTo()); 1025 llvm::StoreInst *Store = 1026 Builder.CreateStore(llvm::Constant::getNullValue(ITy), Ptr); 1027 Store->setAlignment(StoreSize.getQuantity()); 1028 Store->setAtomic(llvm::Release); 1029 return RValue::get(0); 1030 } 1031 1032 case Builtin::BI__sync_synchronize: { 1033 // We assume this is supposed to correspond to a C++0x-style 1034 // sequentially-consistent fence (i.e. this is only usable for 1035 // synchonization, not device I/O or anything like that). This intrinsic 1036 // is really badly designed in the sense that in theory, there isn't 1037 // any way to safely use it... but in practice, it mostly works 1038 // to use it with non-atomic loads and stores to get acquire/release 1039 // semantics. 1040 Builder.CreateFence(llvm::SequentiallyConsistent); 1041 return RValue::get(0); 1042 } 1043 1044 case Builtin::BI__c11_atomic_is_lock_free: 1045 case Builtin::BI__atomic_is_lock_free: { 1046 // Call "bool __atomic_is_lock_free(size_t size, void *ptr)". For the 1047 // __c11 builtin, ptr is 0 (indicating a properly-aligned object), since 1048 // _Atomic(T) is always properly-aligned. 1049 const char *LibCallName = "__atomic_is_lock_free"; 1050 CallArgList Args; 1051 Args.add(RValue::get(EmitScalarExpr(E->getArg(0))), 1052 getContext().getSizeType()); 1053 if (BuiltinID == Builtin::BI__atomic_is_lock_free) 1054 Args.add(RValue::get(EmitScalarExpr(E->getArg(1))), 1055 getContext().VoidPtrTy); 1056 else 1057 Args.add(RValue::get(llvm::Constant::getNullValue(VoidPtrTy)), 1058 getContext().VoidPtrTy); 1059 const CGFunctionInfo &FuncInfo = 1060 CGM.getTypes().arrangeFreeFunctionCall(E->getType(), Args, 1061 FunctionType::ExtInfo(), 1062 RequiredArgs::All); 1063 llvm::FunctionType *FTy = CGM.getTypes().GetFunctionType(FuncInfo); 1064 llvm::Constant *Func = CGM.CreateRuntimeFunction(FTy, LibCallName); 1065 return EmitCall(FuncInfo, Func, ReturnValueSlot(), Args); 1066 } 1067 1068 case Builtin::BI__atomic_test_and_set: { 1069 // Look at the argument type to determine whether this is a volatile 1070 // operation. The parameter type is always volatile. 1071 QualType PtrTy = E->getArg(0)->IgnoreImpCasts()->getType(); 1072 bool Volatile = 1073 PtrTy->castAs<PointerType>()->getPointeeType().isVolatileQualified(); 1074 1075 Value *Ptr = EmitScalarExpr(E->getArg(0)); 1076 unsigned AddrSpace = Ptr->getType()->getPointerAddressSpace(); 1077 Ptr = Builder.CreateBitCast(Ptr, Int8Ty->getPointerTo(AddrSpace)); 1078 Value *NewVal = Builder.getInt8(1); 1079 Value *Order = EmitScalarExpr(E->getArg(1)); 1080 if (isa<llvm::ConstantInt>(Order)) { 1081 int ord = cast<llvm::ConstantInt>(Order)->getZExtValue(); 1082 AtomicRMWInst *Result = 0; 1083 switch (ord) { 1084 case 0: // memory_order_relaxed 1085 default: // invalid order 1086 Result = Builder.CreateAtomicRMW(llvm::AtomicRMWInst::Xchg, 1087 Ptr, NewVal, 1088 llvm::Monotonic); 1089 break; 1090 case 1: // memory_order_consume 1091 case 2: // memory_order_acquire 1092 Result = Builder.CreateAtomicRMW(llvm::AtomicRMWInst::Xchg, 1093 Ptr, NewVal, 1094 llvm::Acquire); 1095 break; 1096 case 3: // memory_order_release 1097 Result = Builder.CreateAtomicRMW(llvm::AtomicRMWInst::Xchg, 1098 Ptr, NewVal, 1099 llvm::Release); 1100 break; 1101 case 4: // memory_order_acq_rel 1102 Result = Builder.CreateAtomicRMW(llvm::AtomicRMWInst::Xchg, 1103 Ptr, NewVal, 1104 llvm::AcquireRelease); 1105 break; 1106 case 5: // memory_order_seq_cst 1107 Result = Builder.CreateAtomicRMW(llvm::AtomicRMWInst::Xchg, 1108 Ptr, NewVal, 1109 llvm::SequentiallyConsistent); 1110 break; 1111 } 1112 Result->setVolatile(Volatile); 1113 return RValue::get(Builder.CreateIsNotNull(Result, "tobool")); 1114 } 1115 1116 llvm::BasicBlock *ContBB = createBasicBlock("atomic.continue", CurFn); 1117 1118 llvm::BasicBlock *BBs[5] = { 1119 createBasicBlock("monotonic", CurFn), 1120 createBasicBlock("acquire", CurFn), 1121 createBasicBlock("release", CurFn), 1122 createBasicBlock("acqrel", CurFn), 1123 createBasicBlock("seqcst", CurFn) 1124 }; 1125 llvm::AtomicOrdering Orders[5] = { 1126 llvm::Monotonic, llvm::Acquire, llvm::Release, 1127 llvm::AcquireRelease, llvm::SequentiallyConsistent 1128 }; 1129 1130 Order = Builder.CreateIntCast(Order, Builder.getInt32Ty(), false); 1131 llvm::SwitchInst *SI = Builder.CreateSwitch(Order, BBs[0]); 1132 1133 Builder.SetInsertPoint(ContBB); 1134 PHINode *Result = Builder.CreatePHI(Int8Ty, 5, "was_set"); 1135 1136 for (unsigned i = 0; i < 5; ++i) { 1137 Builder.SetInsertPoint(BBs[i]); 1138 AtomicRMWInst *RMW = Builder.CreateAtomicRMW(llvm::AtomicRMWInst::Xchg, 1139 Ptr, NewVal, Orders[i]); 1140 RMW->setVolatile(Volatile); 1141 Result->addIncoming(RMW, BBs[i]); 1142 Builder.CreateBr(ContBB); 1143 } 1144 1145 SI->addCase(Builder.getInt32(0), BBs[0]); 1146 SI->addCase(Builder.getInt32(1), BBs[1]); 1147 SI->addCase(Builder.getInt32(2), BBs[1]); 1148 SI->addCase(Builder.getInt32(3), BBs[2]); 1149 SI->addCase(Builder.getInt32(4), BBs[3]); 1150 SI->addCase(Builder.getInt32(5), BBs[4]); 1151 1152 Builder.SetInsertPoint(ContBB); 1153 return RValue::get(Builder.CreateIsNotNull(Result, "tobool")); 1154 } 1155 1156 case Builtin::BI__atomic_clear: { 1157 QualType PtrTy = E->getArg(0)->IgnoreImpCasts()->getType(); 1158 bool Volatile = 1159 PtrTy->castAs<PointerType>()->getPointeeType().isVolatileQualified(); 1160 1161 Value *Ptr = EmitScalarExpr(E->getArg(0)); 1162 unsigned AddrSpace = Ptr->getType()->getPointerAddressSpace(); 1163 Ptr = Builder.CreateBitCast(Ptr, Int8Ty->getPointerTo(AddrSpace)); 1164 Value *NewVal = Builder.getInt8(0); 1165 Value *Order = EmitScalarExpr(E->getArg(1)); 1166 if (isa<llvm::ConstantInt>(Order)) { 1167 int ord = cast<llvm::ConstantInt>(Order)->getZExtValue(); 1168 StoreInst *Store = Builder.CreateStore(NewVal, Ptr, Volatile); 1169 Store->setAlignment(1); 1170 switch (ord) { 1171 case 0: // memory_order_relaxed 1172 default: // invalid order 1173 Store->setOrdering(llvm::Monotonic); 1174 break; 1175 case 3: // memory_order_release 1176 Store->setOrdering(llvm::Release); 1177 break; 1178 case 5: // memory_order_seq_cst 1179 Store->setOrdering(llvm::SequentiallyConsistent); 1180 break; 1181 } 1182 return RValue::get(0); 1183 } 1184 1185 llvm::BasicBlock *ContBB = createBasicBlock("atomic.continue", CurFn); 1186 1187 llvm::BasicBlock *BBs[3] = { 1188 createBasicBlock("monotonic", CurFn), 1189 createBasicBlock("release", CurFn), 1190 createBasicBlock("seqcst", CurFn) 1191 }; 1192 llvm::AtomicOrdering Orders[3] = { 1193 llvm::Monotonic, llvm::Release, llvm::SequentiallyConsistent 1194 }; 1195 1196 Order = Builder.CreateIntCast(Order, Builder.getInt32Ty(), false); 1197 llvm::SwitchInst *SI = Builder.CreateSwitch(Order, BBs[0]); 1198 1199 for (unsigned i = 0; i < 3; ++i) { 1200 Builder.SetInsertPoint(BBs[i]); 1201 StoreInst *Store = Builder.CreateStore(NewVal, Ptr, Volatile); 1202 Store->setAlignment(1); 1203 Store->setOrdering(Orders[i]); 1204 Builder.CreateBr(ContBB); 1205 } 1206 1207 SI->addCase(Builder.getInt32(0), BBs[0]); 1208 SI->addCase(Builder.getInt32(3), BBs[1]); 1209 SI->addCase(Builder.getInt32(5), BBs[2]); 1210 1211 Builder.SetInsertPoint(ContBB); 1212 return RValue::get(0); 1213 } 1214 1215 case Builtin::BI__atomic_thread_fence: 1216 case Builtin::BI__atomic_signal_fence: 1217 case Builtin::BI__c11_atomic_thread_fence: 1218 case Builtin::BI__c11_atomic_signal_fence: { 1219 llvm::SynchronizationScope Scope; 1220 if (BuiltinID == Builtin::BI__atomic_signal_fence || 1221 BuiltinID == Builtin::BI__c11_atomic_signal_fence) 1222 Scope = llvm::SingleThread; 1223 else 1224 Scope = llvm::CrossThread; 1225 Value *Order = EmitScalarExpr(E->getArg(0)); 1226 if (isa<llvm::ConstantInt>(Order)) { 1227 int ord = cast<llvm::ConstantInt>(Order)->getZExtValue(); 1228 switch (ord) { 1229 case 0: // memory_order_relaxed 1230 default: // invalid order 1231 break; 1232 case 1: // memory_order_consume 1233 case 2: // memory_order_acquire 1234 Builder.CreateFence(llvm::Acquire, Scope); 1235 break; 1236 case 3: // memory_order_release 1237 Builder.CreateFence(llvm::Release, Scope); 1238 break; 1239 case 4: // memory_order_acq_rel 1240 Builder.CreateFence(llvm::AcquireRelease, Scope); 1241 break; 1242 case 5: // memory_order_seq_cst 1243 Builder.CreateFence(llvm::SequentiallyConsistent, Scope); 1244 break; 1245 } 1246 return RValue::get(0); 1247 } 1248 1249 llvm::BasicBlock *AcquireBB, *ReleaseBB, *AcqRelBB, *SeqCstBB; 1250 AcquireBB = createBasicBlock("acquire", CurFn); 1251 ReleaseBB = createBasicBlock("release", CurFn); 1252 AcqRelBB = createBasicBlock("acqrel", CurFn); 1253 SeqCstBB = createBasicBlock("seqcst", CurFn); 1254 llvm::BasicBlock *ContBB = createBasicBlock("atomic.continue", CurFn); 1255 1256 Order = Builder.CreateIntCast(Order, Builder.getInt32Ty(), false); 1257 llvm::SwitchInst *SI = Builder.CreateSwitch(Order, ContBB); 1258 1259 Builder.SetInsertPoint(AcquireBB); 1260 Builder.CreateFence(llvm::Acquire, Scope); 1261 Builder.CreateBr(ContBB); 1262 SI->addCase(Builder.getInt32(1), AcquireBB); 1263 SI->addCase(Builder.getInt32(2), AcquireBB); 1264 1265 Builder.SetInsertPoint(ReleaseBB); 1266 Builder.CreateFence(llvm::Release, Scope); 1267 Builder.CreateBr(ContBB); 1268 SI->addCase(Builder.getInt32(3), ReleaseBB); 1269 1270 Builder.SetInsertPoint(AcqRelBB); 1271 Builder.CreateFence(llvm::AcquireRelease, Scope); 1272 Builder.CreateBr(ContBB); 1273 SI->addCase(Builder.getInt32(4), AcqRelBB); 1274 1275 Builder.SetInsertPoint(SeqCstBB); 1276 Builder.CreateFence(llvm::SequentiallyConsistent, Scope); 1277 Builder.CreateBr(ContBB); 1278 SI->addCase(Builder.getInt32(5), SeqCstBB); 1279 1280 Builder.SetInsertPoint(ContBB); 1281 return RValue::get(0); 1282 } 1283 1284 // Library functions with special handling. 1285 case Builtin::BIsqrt: 1286 case Builtin::BIsqrtf: 1287 case Builtin::BIsqrtl: { 1288 // Transform a call to sqrt* into a @llvm.sqrt.* intrinsic call, but only 1289 // in finite- or unsafe-math mode (the intrinsic has different semantics 1290 // for handling negative numbers compared to the library function, so 1291 // -fmath-errno=0 is not enough). 1292 if (!FD->hasAttr<ConstAttr>()) 1293 break; 1294 if (!(CGM.getCodeGenOpts().UnsafeFPMath || 1295 CGM.getCodeGenOpts().NoNaNsFPMath)) 1296 break; 1297 Value *Arg0 = EmitScalarExpr(E->getArg(0)); 1298 llvm::Type *ArgType = Arg0->getType(); 1299 Value *F = CGM.getIntrinsic(Intrinsic::sqrt, ArgType); 1300 return RValue::get(Builder.CreateCall(F, Arg0)); 1301 } 1302 1303 case Builtin::BIpow: 1304 case Builtin::BIpowf: 1305 case Builtin::BIpowl: { 1306 // Transform a call to pow* into a @llvm.pow.* intrinsic call. 1307 if (!FD->hasAttr<ConstAttr>()) 1308 break; 1309 Value *Base = EmitScalarExpr(E->getArg(0)); 1310 Value *Exponent = EmitScalarExpr(E->getArg(1)); 1311 llvm::Type *ArgType = Base->getType(); 1312 Value *F = CGM.getIntrinsic(Intrinsic::pow, ArgType); 1313 return RValue::get(Builder.CreateCall2(F, Base, Exponent)); 1314 break; 1315 } 1316 1317 case Builtin::BIfma: 1318 case Builtin::BIfmaf: 1319 case Builtin::BIfmal: 1320 case Builtin::BI__builtin_fma: 1321 case Builtin::BI__builtin_fmaf: 1322 case Builtin::BI__builtin_fmal: { 1323 // Rewrite fma to intrinsic. 1324 Value *FirstArg = EmitScalarExpr(E->getArg(0)); 1325 llvm::Type *ArgType = FirstArg->getType(); 1326 Value *F = CGM.getIntrinsic(Intrinsic::fma, ArgType); 1327 return RValue::get(Builder.CreateCall3(F, FirstArg, 1328 EmitScalarExpr(E->getArg(1)), 1329 EmitScalarExpr(E->getArg(2)))); 1330 } 1331 1332 case Builtin::BI__builtin_signbit: 1333 case Builtin::BI__builtin_signbitf: 1334 case Builtin::BI__builtin_signbitl: { 1335 LLVMContext &C = CGM.getLLVMContext(); 1336 1337 Value *Arg = EmitScalarExpr(E->getArg(0)); 1338 llvm::Type *ArgTy = Arg->getType(); 1339 if (ArgTy->isPPC_FP128Ty()) 1340 break; // FIXME: I'm not sure what the right implementation is here. 1341 int ArgWidth = ArgTy->getPrimitiveSizeInBits(); 1342 llvm::Type *ArgIntTy = llvm::IntegerType::get(C, ArgWidth); 1343 Value *BCArg = Builder.CreateBitCast(Arg, ArgIntTy); 1344 Value *ZeroCmp = llvm::Constant::getNullValue(ArgIntTy); 1345 Value *Result = Builder.CreateICmpSLT(BCArg, ZeroCmp); 1346 return RValue::get(Builder.CreateZExt(Result, ConvertType(E->getType()))); 1347 } 1348 case Builtin::BI__builtin_annotation: { 1349 llvm::Value *AnnVal = EmitScalarExpr(E->getArg(0)); 1350 llvm::Value *F = CGM.getIntrinsic(llvm::Intrinsic::annotation, 1351 AnnVal->getType()); 1352 1353 // Get the annotation string, go through casts. Sema requires this to be a 1354 // non-wide string literal, potentially casted, so the cast<> is safe. 1355 const Expr *AnnotationStrExpr = E->getArg(1)->IgnoreParenCasts(); 1356 StringRef Str = cast<StringLiteral>(AnnotationStrExpr)->getString(); 1357 return RValue::get(EmitAnnotationCall(F, AnnVal, Str, E->getExprLoc())); 1358 } 1359 case Builtin::BI__builtin_addcb: 1360 case Builtin::BI__builtin_addcs: 1361 case Builtin::BI__builtin_addc: 1362 case Builtin::BI__builtin_addcl: 1363 case Builtin::BI__builtin_addcll: 1364 case Builtin::BI__builtin_subcb: 1365 case Builtin::BI__builtin_subcs: 1366 case Builtin::BI__builtin_subc: 1367 case Builtin::BI__builtin_subcl: 1368 case Builtin::BI__builtin_subcll: { 1369 1370 // We translate all of these builtins from expressions of the form: 1371 // int x = ..., y = ..., carryin = ..., carryout, result; 1372 // result = __builtin_addc(x, y, carryin, &carryout); 1373 // 1374 // to LLVM IR of the form: 1375 // 1376 // %tmp1 = call {i32, i1} @llvm.uadd.with.overflow.i32(i32 %x, i32 %y) 1377 // %tmpsum1 = extractvalue {i32, i1} %tmp1, 0 1378 // %carry1 = extractvalue {i32, i1} %tmp1, 1 1379 // %tmp2 = call {i32, i1} @llvm.uadd.with.overflow.i32(i32 %tmpsum1, 1380 // i32 %carryin) 1381 // %result = extractvalue {i32, i1} %tmp2, 0 1382 // %carry2 = extractvalue {i32, i1} %tmp2, 1 1383 // %tmp3 = or i1 %carry1, %carry2 1384 // %tmp4 = zext i1 %tmp3 to i32 1385 // store i32 %tmp4, i32* %carryout 1386 1387 // Scalarize our inputs. 1388 llvm::Value *X = EmitScalarExpr(E->getArg(0)); 1389 llvm::Value *Y = EmitScalarExpr(E->getArg(1)); 1390 llvm::Value *Carryin = EmitScalarExpr(E->getArg(2)); 1391 std::pair<llvm::Value*, unsigned> CarryOutPtr = 1392 EmitPointerWithAlignment(E->getArg(3)); 1393 1394 // Decide if we are lowering to a uadd.with.overflow or usub.with.overflow. 1395 llvm::Intrinsic::ID IntrinsicId; 1396 switch (BuiltinID) { 1397 default: llvm_unreachable("Unknown multiprecision builtin id."); 1398 case Builtin::BI__builtin_addcb: 1399 case Builtin::BI__builtin_addcs: 1400 case Builtin::BI__builtin_addc: 1401 case Builtin::BI__builtin_addcl: 1402 case Builtin::BI__builtin_addcll: 1403 IntrinsicId = llvm::Intrinsic::uadd_with_overflow; 1404 break; 1405 case Builtin::BI__builtin_subcb: 1406 case Builtin::BI__builtin_subcs: 1407 case Builtin::BI__builtin_subc: 1408 case Builtin::BI__builtin_subcl: 1409 case Builtin::BI__builtin_subcll: 1410 IntrinsicId = llvm::Intrinsic::usub_with_overflow; 1411 break; 1412 } 1413 1414 // Construct our resulting LLVM IR expression. 1415 llvm::Value *Carry1; 1416 llvm::Value *Sum1 = EmitOverflowIntrinsic(*this, IntrinsicId, 1417 X, Y, Carry1); 1418 llvm::Value *Carry2; 1419 llvm::Value *Sum2 = EmitOverflowIntrinsic(*this, IntrinsicId, 1420 Sum1, Carryin, Carry2); 1421 llvm::Value *CarryOut = Builder.CreateZExt(Builder.CreateOr(Carry1, Carry2), 1422 X->getType()); 1423 llvm::StoreInst *CarryOutStore = Builder.CreateStore(CarryOut, 1424 CarryOutPtr.first); 1425 CarryOutStore->setAlignment(CarryOutPtr.second); 1426 return RValue::get(Sum2); 1427 } 1428 case Builtin::BI__builtin_uadd_overflow: 1429 case Builtin::BI__builtin_uaddl_overflow: 1430 case Builtin::BI__builtin_uaddll_overflow: 1431 case Builtin::BI__builtin_usub_overflow: 1432 case Builtin::BI__builtin_usubl_overflow: 1433 case Builtin::BI__builtin_usubll_overflow: 1434 case Builtin::BI__builtin_umul_overflow: 1435 case Builtin::BI__builtin_umull_overflow: 1436 case Builtin::BI__builtin_umulll_overflow: 1437 case Builtin::BI__builtin_sadd_overflow: 1438 case Builtin::BI__builtin_saddl_overflow: 1439 case Builtin::BI__builtin_saddll_overflow: 1440 case Builtin::BI__builtin_ssub_overflow: 1441 case Builtin::BI__builtin_ssubl_overflow: 1442 case Builtin::BI__builtin_ssubll_overflow: 1443 case Builtin::BI__builtin_smul_overflow: 1444 case Builtin::BI__builtin_smull_overflow: 1445 case Builtin::BI__builtin_smulll_overflow: { 1446 1447 // We translate all of these builtins directly to the relevant llvm IR node. 1448 1449 // Scalarize our inputs. 1450 llvm::Value *X = EmitScalarExpr(E->getArg(0)); 1451 llvm::Value *Y = EmitScalarExpr(E->getArg(1)); 1452 std::pair<llvm::Value *, unsigned> SumOutPtr = 1453 EmitPointerWithAlignment(E->getArg(2)); 1454 1455 // Decide which of the overflow intrinsics we are lowering to: 1456 llvm::Intrinsic::ID IntrinsicId; 1457 switch (BuiltinID) { 1458 default: llvm_unreachable("Unknown security overflow builtin id."); 1459 case Builtin::BI__builtin_uadd_overflow: 1460 case Builtin::BI__builtin_uaddl_overflow: 1461 case Builtin::BI__builtin_uaddll_overflow: 1462 IntrinsicId = llvm::Intrinsic::uadd_with_overflow; 1463 break; 1464 case Builtin::BI__builtin_usub_overflow: 1465 case Builtin::BI__builtin_usubl_overflow: 1466 case Builtin::BI__builtin_usubll_overflow: 1467 IntrinsicId = llvm::Intrinsic::usub_with_overflow; 1468 break; 1469 case Builtin::BI__builtin_umul_overflow: 1470 case Builtin::BI__builtin_umull_overflow: 1471 case Builtin::BI__builtin_umulll_overflow: 1472 IntrinsicId = llvm::Intrinsic::umul_with_overflow; 1473 break; 1474 case Builtin::BI__builtin_sadd_overflow: 1475 case Builtin::BI__builtin_saddl_overflow: 1476 case Builtin::BI__builtin_saddll_overflow: 1477 IntrinsicId = llvm::Intrinsic::sadd_with_overflow; 1478 break; 1479 case Builtin::BI__builtin_ssub_overflow: 1480 case Builtin::BI__builtin_ssubl_overflow: 1481 case Builtin::BI__builtin_ssubll_overflow: 1482 IntrinsicId = llvm::Intrinsic::ssub_with_overflow; 1483 break; 1484 case Builtin::BI__builtin_smul_overflow: 1485 case Builtin::BI__builtin_smull_overflow: 1486 case Builtin::BI__builtin_smulll_overflow: 1487 IntrinsicId = llvm::Intrinsic::smul_with_overflow; 1488 break; 1489 } 1490 1491 1492 llvm::Value *Carry; 1493 llvm::Value *Sum = EmitOverflowIntrinsic(*this, IntrinsicId, X, Y, Carry); 1494 llvm::StoreInst *SumOutStore = Builder.CreateStore(Sum, SumOutPtr.first); 1495 SumOutStore->setAlignment(SumOutPtr.second); 1496 1497 return RValue::get(Carry); 1498 } 1499 case Builtin::BI__builtin_addressof: 1500 return RValue::get(EmitLValue(E->getArg(0)).getAddress()); 1501 case Builtin::BI__noop: 1502 return RValue::get(0); 1503 } 1504 1505 // If this is an alias for a lib function (e.g. __builtin_sin), emit 1506 // the call using the normal call path, but using the unmangled 1507 // version of the function name. 1508 if (getContext().BuiltinInfo.isLibFunction(BuiltinID)) 1509 return emitLibraryCall(*this, FD, E, 1510 CGM.getBuiltinLibFunction(FD, BuiltinID)); 1511 1512 // If this is a predefined lib function (e.g. malloc), emit the call 1513 // using exactly the normal call path. 1514 if (getContext().BuiltinInfo.isPredefinedLibFunction(BuiltinID)) 1515 return emitLibraryCall(*this, FD, E, EmitScalarExpr(E->getCallee())); 1516 1517 // See if we have a target specific intrinsic. 1518 const char *Name = getContext().BuiltinInfo.GetName(BuiltinID); 1519 Intrinsic::ID IntrinsicID = Intrinsic::not_intrinsic; 1520 if (const char *Prefix = 1521 llvm::Triple::getArchTypePrefix(getTarget().getTriple().getArch())) 1522 IntrinsicID = Intrinsic::getIntrinsicForGCCBuiltin(Prefix, Name); 1523 1524 if (IntrinsicID != Intrinsic::not_intrinsic) { 1525 SmallVector<Value*, 16> Args; 1526 1527 // Find out if any arguments are required to be integer constant 1528 // expressions. 1529 unsigned ICEArguments = 0; 1530 ASTContext::GetBuiltinTypeError Error; 1531 getContext().GetBuiltinType(BuiltinID, Error, &ICEArguments); 1532 assert(Error == ASTContext::GE_None && "Should not codegen an error"); 1533 1534 Function *F = CGM.getIntrinsic(IntrinsicID); 1535 llvm::FunctionType *FTy = F->getFunctionType(); 1536 1537 for (unsigned i = 0, e = E->getNumArgs(); i != e; ++i) { 1538 Value *ArgValue; 1539 // If this is a normal argument, just emit it as a scalar. 1540 if ((ICEArguments & (1 << i)) == 0) { 1541 ArgValue = EmitScalarExpr(E->getArg(i)); 1542 } else { 1543 // If this is required to be a constant, constant fold it so that we 1544 // know that the generated intrinsic gets a ConstantInt. 1545 llvm::APSInt Result; 1546 bool IsConst = E->getArg(i)->isIntegerConstantExpr(Result,getContext()); 1547 assert(IsConst && "Constant arg isn't actually constant?"); 1548 (void)IsConst; 1549 ArgValue = llvm::ConstantInt::get(getLLVMContext(), Result); 1550 } 1551 1552 // If the intrinsic arg type is different from the builtin arg type 1553 // we need to do a bit cast. 1554 llvm::Type *PTy = FTy->getParamType(i); 1555 if (PTy != ArgValue->getType()) { 1556 assert(PTy->canLosslesslyBitCastTo(FTy->getParamType(i)) && 1557 "Must be able to losslessly bit cast to param"); 1558 ArgValue = Builder.CreateBitCast(ArgValue, PTy); 1559 } 1560 1561 Args.push_back(ArgValue); 1562 } 1563 1564 Value *V = Builder.CreateCall(F, Args); 1565 QualType BuiltinRetType = E->getType(); 1566 1567 llvm::Type *RetTy = VoidTy; 1568 if (!BuiltinRetType->isVoidType()) 1569 RetTy = ConvertType(BuiltinRetType); 1570 1571 if (RetTy != V->getType()) { 1572 assert(V->getType()->canLosslesslyBitCastTo(RetTy) && 1573 "Must be able to losslessly bit cast result type"); 1574 V = Builder.CreateBitCast(V, RetTy); 1575 } 1576 1577 return RValue::get(V); 1578 } 1579 1580 // See if we have a target specific builtin that needs to be lowered. 1581 if (Value *V = EmitTargetBuiltinExpr(BuiltinID, E)) 1582 return RValue::get(V); 1583 1584 ErrorUnsupported(E, "builtin function"); 1585 1586 // Unknown builtin, for now just dump it out and return undef. 1587 return GetUndefRValue(E->getType()); 1588 } 1589 1590 Value *CodeGenFunction::EmitTargetBuiltinExpr(unsigned BuiltinID, 1591 const CallExpr *E) { 1592 switch (getTarget().getTriple().getArch()) { 1593 case llvm::Triple::aarch64: 1594 return EmitAArch64BuiltinExpr(BuiltinID, E); 1595 case llvm::Triple::arm: 1596 case llvm::Triple::thumb: 1597 return EmitARMBuiltinExpr(BuiltinID, E); 1598 case llvm::Triple::x86: 1599 case llvm::Triple::x86_64: 1600 return EmitX86BuiltinExpr(BuiltinID, E); 1601 case llvm::Triple::ppc: 1602 case llvm::Triple::ppc64: 1603 case llvm::Triple::ppc64le: 1604 return EmitPPCBuiltinExpr(BuiltinID, E); 1605 default: 1606 return 0; 1607 } 1608 } 1609 1610 static llvm::VectorType *GetNeonType(CodeGenFunction *CGF, 1611 NeonTypeFlags TypeFlags, 1612 bool V1Ty=false) { 1613 int IsQuad = TypeFlags.isQuad(); 1614 switch (TypeFlags.getEltType()) { 1615 case NeonTypeFlags::Int8: 1616 case NeonTypeFlags::Poly8: 1617 return llvm::VectorType::get(CGF->Int8Ty, V1Ty ? 1 : (8 << IsQuad)); 1618 case NeonTypeFlags::Int16: 1619 case NeonTypeFlags::Poly16: 1620 case NeonTypeFlags::Float16: 1621 return llvm::VectorType::get(CGF->Int16Ty, V1Ty ? 1 : (4 << IsQuad)); 1622 case NeonTypeFlags::Int32: 1623 return llvm::VectorType::get(CGF->Int32Ty, V1Ty ? 1 : (2 << IsQuad)); 1624 case NeonTypeFlags::Int64: 1625 case NeonTypeFlags::Poly64: 1626 return llvm::VectorType::get(CGF->Int64Ty, V1Ty ? 1 : (1 << IsQuad)); 1627 case NeonTypeFlags::Poly128: 1628 // FIXME: i128 and f128 doesn't get fully support in Clang and llvm. 1629 // There is a lot of i128 and f128 API missing. 1630 // so we use v16i8 to represent poly128 and get pattern matched. 1631 return llvm::VectorType::get(CGF->Int8Ty, 16); 1632 case NeonTypeFlags::Float32: 1633 return llvm::VectorType::get(CGF->FloatTy, V1Ty ? 1 : (2 << IsQuad)); 1634 case NeonTypeFlags::Float64: 1635 return llvm::VectorType::get(CGF->DoubleTy, V1Ty ? 1 : (1 << IsQuad)); 1636 } 1637 llvm_unreachable("Unknown vector element type!"); 1638 } 1639 1640 Value *CodeGenFunction::EmitNeonSplat(Value *V, Constant *C) { 1641 unsigned nElts = cast<llvm::VectorType>(V->getType())->getNumElements(); 1642 Value* SV = llvm::ConstantVector::getSplat(nElts, C); 1643 return Builder.CreateShuffleVector(V, V, SV, "lane"); 1644 } 1645 1646 Value *CodeGenFunction::EmitNeonCall(Function *F, SmallVectorImpl<Value*> &Ops, 1647 const char *name, 1648 unsigned shift, bool rightshift) { 1649 unsigned j = 0; 1650 for (Function::const_arg_iterator ai = F->arg_begin(), ae = F->arg_end(); 1651 ai != ae; ++ai, ++j) 1652 if (shift > 0 && shift == j) 1653 Ops[j] = EmitNeonShiftVector(Ops[j], ai->getType(), rightshift); 1654 else 1655 Ops[j] = Builder.CreateBitCast(Ops[j], ai->getType(), name); 1656 1657 return Builder.CreateCall(F, Ops, name); 1658 } 1659 1660 Value *CodeGenFunction::EmitNeonShiftVector(Value *V, llvm::Type *Ty, 1661 bool neg) { 1662 int SV = cast<ConstantInt>(V)->getSExtValue(); 1663 1664 llvm::VectorType *VTy = cast<llvm::VectorType>(Ty); 1665 llvm::Constant *C = ConstantInt::get(VTy->getElementType(), neg ? -SV : SV); 1666 return llvm::ConstantVector::getSplat(VTy->getNumElements(), C); 1667 } 1668 1669 // \brief Right-shift a vector by a constant. 1670 Value *CodeGenFunction::EmitNeonRShiftImm(Value *Vec, Value *Shift, 1671 llvm::Type *Ty, bool usgn, 1672 const char *name) { 1673 llvm::VectorType *VTy = cast<llvm::VectorType>(Ty); 1674 1675 int ShiftAmt = cast<ConstantInt>(Shift)->getSExtValue(); 1676 int EltSize = VTy->getScalarSizeInBits(); 1677 1678 Vec = Builder.CreateBitCast(Vec, Ty); 1679 1680 // lshr/ashr are undefined when the shift amount is equal to the vector 1681 // element size. 1682 if (ShiftAmt == EltSize) { 1683 if (usgn) { 1684 // Right-shifting an unsigned value by its size yields 0. 1685 llvm::Constant *Zero = ConstantInt::get(VTy->getElementType(), 0); 1686 return llvm::ConstantVector::getSplat(VTy->getNumElements(), Zero); 1687 } else { 1688 // Right-shifting a signed value by its size is equivalent 1689 // to a shift of size-1. 1690 --ShiftAmt; 1691 Shift = ConstantInt::get(VTy->getElementType(), ShiftAmt); 1692 } 1693 } 1694 1695 Shift = EmitNeonShiftVector(Shift, Ty, false); 1696 if (usgn) 1697 return Builder.CreateLShr(Vec, Shift, name); 1698 else 1699 return Builder.CreateAShr(Vec, Shift, name); 1700 } 1701 1702 /// GetPointeeAlignment - Given an expression with a pointer type, find the 1703 /// alignment of the type referenced by the pointer. Skip over implicit 1704 /// casts. 1705 std::pair<llvm::Value*, unsigned> 1706 CodeGenFunction::EmitPointerWithAlignment(const Expr *Addr) { 1707 assert(Addr->getType()->isPointerType()); 1708 Addr = Addr->IgnoreParens(); 1709 if (const ImplicitCastExpr *ICE = dyn_cast<ImplicitCastExpr>(Addr)) { 1710 if ((ICE->getCastKind() == CK_BitCast || ICE->getCastKind() == CK_NoOp) && 1711 ICE->getSubExpr()->getType()->isPointerType()) { 1712 std::pair<llvm::Value*, unsigned> Ptr = 1713 EmitPointerWithAlignment(ICE->getSubExpr()); 1714 Ptr.first = Builder.CreateBitCast(Ptr.first, 1715 ConvertType(Addr->getType())); 1716 return Ptr; 1717 } else if (ICE->getCastKind() == CK_ArrayToPointerDecay) { 1718 LValue LV = EmitLValue(ICE->getSubExpr()); 1719 unsigned Align = LV.getAlignment().getQuantity(); 1720 if (!Align) { 1721 // FIXME: Once LValues are fixed to always set alignment, 1722 // zap this code. 1723 QualType PtTy = ICE->getSubExpr()->getType(); 1724 if (!PtTy->isIncompleteType()) 1725 Align = getContext().getTypeAlignInChars(PtTy).getQuantity(); 1726 else 1727 Align = 1; 1728 } 1729 return std::make_pair(LV.getAddress(), Align); 1730 } 1731 } 1732 if (const UnaryOperator *UO = dyn_cast<UnaryOperator>(Addr)) { 1733 if (UO->getOpcode() == UO_AddrOf) { 1734 LValue LV = EmitLValue(UO->getSubExpr()); 1735 unsigned Align = LV.getAlignment().getQuantity(); 1736 if (!Align) { 1737 // FIXME: Once LValues are fixed to always set alignment, 1738 // zap this code. 1739 QualType PtTy = UO->getSubExpr()->getType(); 1740 if (!PtTy->isIncompleteType()) 1741 Align = getContext().getTypeAlignInChars(PtTy).getQuantity(); 1742 else 1743 Align = 1; 1744 } 1745 return std::make_pair(LV.getAddress(), Align); 1746 } 1747 } 1748 1749 unsigned Align = 1; 1750 QualType PtTy = Addr->getType()->getPointeeType(); 1751 if (!PtTy->isIncompleteType()) 1752 Align = getContext().getTypeAlignInChars(PtTy).getQuantity(); 1753 1754 return std::make_pair(EmitScalarExpr(Addr), Align); 1755 } 1756 1757 Value *CodeGenFunction::EmitCommonNeonBuiltinExpr(unsigned BuiltinID, 1758 const CallExpr *E, 1759 SmallVectorImpl<Value *> &Ops, 1760 llvm::Value *Align) { 1761 // Get the last argument, which specifies the vector type. 1762 llvm::APSInt Result; 1763 const Expr *Arg = E->getArg(E->getNumArgs() - 1); 1764 if (!Arg->isIntegerConstantExpr(Result, getContext())) 1765 return 0; 1766 1767 // Determine the type of this overloaded NEON intrinsic. 1768 NeonTypeFlags Type(Result.getZExtValue()); 1769 bool Usgn = Type.isUnsigned(); 1770 bool Quad = Type.isQuad(); 1771 1772 llvm::VectorType *VTy = GetNeonType(this, Type); 1773 llvm::Type *Ty = VTy; 1774 if (!Ty) 1775 return 0; 1776 1777 unsigned Int; 1778 switch (BuiltinID) { 1779 default: break; 1780 case NEON::BI__builtin_neon_vabs_v: 1781 case NEON::BI__builtin_neon_vabsq_v: 1782 if (VTy->getElementType()->isFloatingPointTy()) 1783 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::fabs, Ty), Ops, "vabs"); 1784 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vabs, Ty), Ops, 1785 "vabs"); 1786 case NEON::BI__builtin_neon_vaeseq_v: 1787 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_aese), 1788 Ops, "aese"); 1789 case NEON::BI__builtin_neon_vaesdq_v: 1790 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_aesd), 1791 Ops, "aesd"); 1792 case NEON::BI__builtin_neon_vaesmcq_v: 1793 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_aesmc), 1794 Ops, "aesmc"); 1795 case NEON::BI__builtin_neon_vaesimcq_v: 1796 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_aesimc), 1797 Ops, "aesimc"); 1798 case NEON::BI__builtin_neon_vabd_v: 1799 case NEON::BI__builtin_neon_vabdq_v: 1800 Int = Usgn ? Intrinsic::arm_neon_vabdu : Intrinsic::arm_neon_vabds; 1801 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vabd"); 1802 case NEON::BI__builtin_neon_vaddhn_v: { 1803 llvm::VectorType *SrcTy = 1804 llvm::VectorType::getExtendedElementVectorType(VTy); 1805 1806 // %sum = add <4 x i32> %lhs, %rhs 1807 Ops[0] = Builder.CreateBitCast(Ops[0], SrcTy); 1808 Ops[1] = Builder.CreateBitCast(Ops[1], SrcTy); 1809 Ops[0] = Builder.CreateAdd(Ops[0], Ops[1], "vaddhn"); 1810 1811 // %high = lshr <4 x i32> %sum, <i32 16, i32 16, i32 16, i32 16> 1812 Constant *ShiftAmt = ConstantInt::get(SrcTy->getElementType(), 1813 SrcTy->getScalarSizeInBits() / 2); 1814 ShiftAmt = ConstantVector::getSplat(VTy->getNumElements(), ShiftAmt); 1815 Ops[0] = Builder.CreateLShr(Ops[0], ShiftAmt, "vaddhn"); 1816 1817 // %res = trunc <4 x i32> %high to <4 x i16> 1818 return Builder.CreateTrunc(Ops[0], VTy, "vaddhn"); 1819 } 1820 case NEON::BI__builtin_neon_vbsl_v: 1821 case NEON::BI__builtin_neon_vbslq_v: 1822 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vbsl, Ty), 1823 Ops, "vbsl"); 1824 case NEON::BI__builtin_neon_vcale_v: 1825 case NEON::BI__builtin_neon_vcaleq_v: 1826 std::swap(Ops[0], Ops[1]); 1827 case NEON::BI__builtin_neon_vcage_v: 1828 case NEON::BI__builtin_neon_vcageq_v: { 1829 llvm::Type *VecFlt = llvm::VectorType::get( 1830 VTy->getScalarSizeInBits() == 32 ? FloatTy : DoubleTy, 1831 VTy->getNumElements()); 1832 llvm::Type *Tys[] = { VTy, VecFlt }; 1833 Function *F = CGM.getIntrinsic(Intrinsic::arm_neon_vacge, Tys); 1834 return EmitNeonCall(F, Ops, "vcage"); 1835 } 1836 case NEON::BI__builtin_neon_vcalt_v: 1837 case NEON::BI__builtin_neon_vcaltq_v: 1838 std::swap(Ops[0], Ops[1]); 1839 case NEON::BI__builtin_neon_vcagt_v: 1840 case NEON::BI__builtin_neon_vcagtq_v: { 1841 llvm::Type *VecFlt = llvm::VectorType::get( 1842 VTy->getScalarSizeInBits() == 32 ? FloatTy : DoubleTy, 1843 VTy->getNumElements()); 1844 llvm::Type *Tys[] = { VTy, VecFlt }; 1845 Function *F = CGM.getIntrinsic(Intrinsic::arm_neon_vacgt, Tys); 1846 return EmitNeonCall(F, Ops, "vcagt"); 1847 } 1848 case NEON::BI__builtin_neon_vcls_v: 1849 case NEON::BI__builtin_neon_vclsq_v: { 1850 Function *F = CGM.getIntrinsic(Intrinsic::arm_neon_vcls, Ty); 1851 return EmitNeonCall(F, Ops, "vcls"); 1852 } 1853 case NEON::BI__builtin_neon_vclz_v: 1854 case NEON::BI__builtin_neon_vclzq_v: { 1855 // Generate target-independent intrinsic; also need to add second argument 1856 // for whether or not clz of zero is undefined; on ARM it isn't. 1857 Function *F = CGM.getIntrinsic(Intrinsic::ctlz, Ty); 1858 Ops.push_back(Builder.getInt1(getTarget().isCLZForZeroUndef())); 1859 return EmitNeonCall(F, Ops, "vclz"); 1860 } 1861 case NEON::BI__builtin_neon_vcnt_v: 1862 case NEON::BI__builtin_neon_vcntq_v: { 1863 // generate target-independent intrinsic 1864 Function *F = CGM.getIntrinsic(Intrinsic::ctpop, Ty); 1865 return EmitNeonCall(F, Ops, "vctpop"); 1866 } 1867 case NEON::BI__builtin_neon_vcvt_f16_v: { 1868 assert(Type.getEltType() == NeonTypeFlags::Float16 && !Quad && 1869 "unexpected vcvt_f16_v builtin"); 1870 Function *F = CGM.getIntrinsic(Intrinsic::arm_neon_vcvtfp2hf); 1871 return EmitNeonCall(F, Ops, "vcvt"); 1872 } 1873 case NEON::BI__builtin_neon_vcvt_f32_f16: { 1874 assert(Type.getEltType() == NeonTypeFlags::Float16 && !Quad && 1875 "unexpected vcvt_f32_f16 builtin"); 1876 Function *F = CGM.getIntrinsic(Intrinsic::arm_neon_vcvthf2fp); 1877 return EmitNeonCall(F, Ops, "vcvt"); 1878 } 1879 case NEON::BI__builtin_neon_vcvt_f32_v: 1880 case NEON::BI__builtin_neon_vcvtq_f32_v: 1881 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 1882 Ty = GetNeonType(this, NeonTypeFlags(NeonTypeFlags::Float32, false, Quad)); 1883 return Usgn ? Builder.CreateUIToFP(Ops[0], Ty, "vcvt") 1884 : Builder.CreateSIToFP(Ops[0], Ty, "vcvt"); 1885 case NEON::BI__builtin_neon_vcvt_n_f32_v: 1886 case NEON::BI__builtin_neon_vcvtq_n_f32_v: { 1887 llvm::Type *FloatTy = 1888 GetNeonType(this, NeonTypeFlags(NeonTypeFlags::Float32, false, Quad)); 1889 llvm::Type *Tys[2] = { FloatTy, Ty }; 1890 Int = Usgn ? Intrinsic::arm_neon_vcvtfxu2fp 1891 : Intrinsic::arm_neon_vcvtfxs2fp; 1892 Function *F = CGM.getIntrinsic(Int, Tys); 1893 return EmitNeonCall(F, Ops, "vcvt_n"); 1894 } 1895 case NEON::BI__builtin_neon_vcvt_n_s32_v: 1896 case NEON::BI__builtin_neon_vcvt_n_u32_v: 1897 case NEON::BI__builtin_neon_vcvt_n_s64_v: 1898 case NEON::BI__builtin_neon_vcvt_n_u64_v: 1899 case NEON::BI__builtin_neon_vcvtq_n_s32_v: 1900 case NEON::BI__builtin_neon_vcvtq_n_u32_v: 1901 case NEON::BI__builtin_neon_vcvtq_n_s64_v: 1902 case NEON::BI__builtin_neon_vcvtq_n_u64_v: { 1903 bool Double = 1904 (cast<llvm::IntegerType>(VTy->getElementType())->getBitWidth() == 64); 1905 llvm::Type *FloatTy = 1906 GetNeonType(this, NeonTypeFlags(Double ? NeonTypeFlags::Float64 1907 : NeonTypeFlags::Float32, 1908 false, Quad)); 1909 llvm::Type *Tys[2] = { Ty, FloatTy }; 1910 Int = Usgn ? Intrinsic::arm_neon_vcvtfp2fxu 1911 : Intrinsic::arm_neon_vcvtfp2fxs; 1912 Function *F = CGM.getIntrinsic(Int, Tys); 1913 return EmitNeonCall(F, Ops, "vcvt_n"); 1914 } 1915 case NEON::BI__builtin_neon_vcvt_s32_v: 1916 case NEON::BI__builtin_neon_vcvt_u32_v: 1917 case NEON::BI__builtin_neon_vcvtq_s32_v: 1918 case NEON::BI__builtin_neon_vcvtq_u32_v: { 1919 llvm::Type *FloatTy = 1920 GetNeonType(this, NeonTypeFlags(NeonTypeFlags::Float32, false, Quad)); 1921 Ops[0] = Builder.CreateBitCast(Ops[0], FloatTy); 1922 return Usgn ? Builder.CreateFPToUI(Ops[0], Ty, "vcvt") 1923 : Builder.CreateFPToSI(Ops[0], Ty, "vcvt"); 1924 } 1925 case NEON::BI__builtin_neon_vext_v: 1926 case NEON::BI__builtin_neon_vextq_v: { 1927 int CV = cast<ConstantInt>(Ops[2])->getSExtValue(); 1928 SmallVector<Constant*, 16> Indices; 1929 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) 1930 Indices.push_back(ConstantInt::get(Int32Ty, i+CV)); 1931 1932 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 1933 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 1934 Value *SV = llvm::ConstantVector::get(Indices); 1935 return Builder.CreateShuffleVector(Ops[0], Ops[1], SV, "vext"); 1936 } 1937 case NEON::BI__builtin_neon_vfma_v: 1938 case NEON::BI__builtin_neon_vfmaq_v: { 1939 Value *F = CGM.getIntrinsic(Intrinsic::fma, Ty); 1940 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 1941 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 1942 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 1943 1944 // NEON intrinsic puts accumulator first, unlike the LLVM fma. 1945 return Builder.CreateCall3(F, Ops[1], Ops[2], Ops[0]); 1946 } 1947 case NEON::BI__builtin_neon_vhadd_v: 1948 case NEON::BI__builtin_neon_vhaddq_v: 1949 Int = Usgn ? Intrinsic::arm_neon_vhaddu : Intrinsic::arm_neon_vhadds; 1950 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vhadd"); 1951 case NEON::BI__builtin_neon_vhsub_v: 1952 case NEON::BI__builtin_neon_vhsubq_v: 1953 Int = Usgn ? Intrinsic::arm_neon_vhsubu : Intrinsic::arm_neon_vhsubs; 1954 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vhsub"); 1955 case NEON::BI__builtin_neon_vld1_v: 1956 case NEON::BI__builtin_neon_vld1q_v: 1957 Ops.push_back(Align); 1958 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vld1, Ty), 1959 Ops, "vld1"); 1960 case NEON::BI__builtin_neon_vld2_v: 1961 case NEON::BI__builtin_neon_vld2q_v: { 1962 Function *F = CGM.getIntrinsic(Intrinsic::arm_neon_vld2, Ty); 1963 Ops[1] = Builder.CreateCall2(F, Ops[1], Align, "vld2"); 1964 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 1965 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 1966 return Builder.CreateStore(Ops[1], Ops[0]); 1967 } 1968 case NEON::BI__builtin_neon_vld3_v: 1969 case NEON::BI__builtin_neon_vld3q_v: { 1970 Function *F = CGM.getIntrinsic(Intrinsic::arm_neon_vld3, Ty); 1971 Ops[1] = Builder.CreateCall2(F, Ops[1], Align, "vld3"); 1972 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 1973 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 1974 return Builder.CreateStore(Ops[1], Ops[0]); 1975 } 1976 case NEON::BI__builtin_neon_vld4_v: 1977 case NEON::BI__builtin_neon_vld4q_v: { 1978 Function *F = CGM.getIntrinsic(Intrinsic::arm_neon_vld4, Ty); 1979 Ops[1] = Builder.CreateCall2(F, Ops[1], Align, "vld4"); 1980 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 1981 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 1982 return Builder.CreateStore(Ops[1], Ops[0]); 1983 } 1984 case NEON::BI__builtin_neon_vld1_dup_v: 1985 case NEON::BI__builtin_neon_vld1q_dup_v: { 1986 Value *V = UndefValue::get(Ty); 1987 Ty = llvm::PointerType::getUnqual(VTy->getElementType()); 1988 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 1989 LoadInst *Ld = Builder.CreateLoad(Ops[0]); 1990 Ld->setAlignment(cast<ConstantInt>(Align)->getZExtValue()); 1991 llvm::Constant *CI = ConstantInt::get(Int32Ty, 0); 1992 Ops[0] = Builder.CreateInsertElement(V, Ld, CI); 1993 return EmitNeonSplat(Ops[0], CI); 1994 } 1995 case NEON::BI__builtin_neon_vld2_lane_v: 1996 case NEON::BI__builtin_neon_vld2q_lane_v: { 1997 Function *F = CGM.getIntrinsic(Intrinsic::arm_neon_vld2lane, Ty); 1998 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 1999 Ops[3] = Builder.CreateBitCast(Ops[3], Ty); 2000 Ops.push_back(Align); 2001 Ops[1] = Builder.CreateCall(F, makeArrayRef(Ops).slice(1), "vld2_lane"); 2002 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 2003 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 2004 return Builder.CreateStore(Ops[1], Ops[0]); 2005 } 2006 case NEON::BI__builtin_neon_vld3_lane_v: 2007 case NEON::BI__builtin_neon_vld3q_lane_v: { 2008 Function *F = CGM.getIntrinsic(Intrinsic::arm_neon_vld3lane, Ty); 2009 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 2010 Ops[3] = Builder.CreateBitCast(Ops[3], Ty); 2011 Ops[4] = Builder.CreateBitCast(Ops[4], Ty); 2012 Ops.push_back(Align); 2013 Ops[1] = Builder.CreateCall(F, makeArrayRef(Ops).slice(1), "vld3_lane"); 2014 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 2015 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 2016 return Builder.CreateStore(Ops[1], Ops[0]); 2017 } 2018 case NEON::BI__builtin_neon_vld4_lane_v: 2019 case NEON::BI__builtin_neon_vld4q_lane_v: { 2020 Function *F = CGM.getIntrinsic(Intrinsic::arm_neon_vld4lane, Ty); 2021 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 2022 Ops[3] = Builder.CreateBitCast(Ops[3], Ty); 2023 Ops[4] = Builder.CreateBitCast(Ops[4], Ty); 2024 Ops[5] = Builder.CreateBitCast(Ops[5], Ty); 2025 Ops.push_back(Align); 2026 Ops[1] = Builder.CreateCall(F, makeArrayRef(Ops).slice(1), "vld3_lane"); 2027 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 2028 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 2029 return Builder.CreateStore(Ops[1], Ops[0]); 2030 } 2031 case NEON::BI__builtin_neon_vmax_v: 2032 case NEON::BI__builtin_neon_vmaxq_v: 2033 Int = Usgn ? Intrinsic::arm_neon_vmaxu : Intrinsic::arm_neon_vmaxs; 2034 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmax"); 2035 case NEON::BI__builtin_neon_vmin_v: 2036 case NEON::BI__builtin_neon_vminq_v: 2037 Int = Usgn ? Intrinsic::arm_neon_vminu : Intrinsic::arm_neon_vmins; 2038 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmin"); 2039 case NEON::BI__builtin_neon_vmovl_v: { 2040 llvm::Type *DTy =llvm::VectorType::getTruncatedElementVectorType(VTy); 2041 Ops[0] = Builder.CreateBitCast(Ops[0], DTy); 2042 if (Usgn) 2043 return Builder.CreateZExt(Ops[0], Ty, "vmovl"); 2044 return Builder.CreateSExt(Ops[0], Ty, "vmovl"); 2045 } 2046 case NEON::BI__builtin_neon_vmovn_v: { 2047 llvm::Type *QTy = llvm::VectorType::getExtendedElementVectorType(VTy); 2048 Ops[0] = Builder.CreateBitCast(Ops[0], QTy); 2049 return Builder.CreateTrunc(Ops[0], Ty, "vmovn"); 2050 } 2051 case NEON::BI__builtin_neon_vmul_v: 2052 case NEON::BI__builtin_neon_vmulq_v: 2053 assert(Type.isPoly() && "vmul builtin only supported for polynomial types"); 2054 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vmulp, Ty), 2055 Ops, "vmul"); 2056 case NEON::BI__builtin_neon_vmull_v: 2057 // FIXME: the integer vmull operations could be emitted in terms of pure 2058 // LLVM IR (2 exts followed by a mul). Unfortunately LLVM has a habit of 2059 // hoisting the exts outside loops. Until global ISel comes along that can 2060 // see through such movement this leads to bad CodeGen. So we need an 2061 // intrinsic for now. 2062 Int = Usgn ? Intrinsic::arm_neon_vmullu : Intrinsic::arm_neon_vmulls; 2063 Int = Type.isPoly() ? (unsigned)Intrinsic::arm_neon_vmullp : Int; 2064 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmull"); 2065 case NEON::BI__builtin_neon_vpadal_v: 2066 case NEON::BI__builtin_neon_vpadalq_v: { 2067 Int = Usgn ? Intrinsic::arm_neon_vpadalu : Intrinsic::arm_neon_vpadals; 2068 // The source operand type has twice as many elements of half the size. 2069 unsigned EltBits = VTy->getElementType()->getPrimitiveSizeInBits(); 2070 llvm::Type *EltTy = 2071 llvm::IntegerType::get(getLLVMContext(), EltBits / 2); 2072 llvm::Type *NarrowTy = 2073 llvm::VectorType::get(EltTy, VTy->getNumElements() * 2); 2074 llvm::Type *Tys[2] = { Ty, NarrowTy }; 2075 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vpadal"); 2076 } 2077 case NEON::BI__builtin_neon_vpadd_v: 2078 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vpadd, Ty), 2079 Ops, "vpadd"); 2080 case NEON::BI__builtin_neon_vpaddl_v: 2081 case NEON::BI__builtin_neon_vpaddlq_v: { 2082 Int = Usgn ? Intrinsic::arm_neon_vpaddlu : Intrinsic::arm_neon_vpaddls; 2083 // The source operand type has twice as many elements of half the size. 2084 unsigned EltBits = VTy->getElementType()->getPrimitiveSizeInBits(); 2085 llvm::Type *EltTy = llvm::IntegerType::get(getLLVMContext(), EltBits / 2); 2086 llvm::Type *NarrowTy = 2087 llvm::VectorType::get(EltTy, VTy->getNumElements() * 2); 2088 llvm::Type *Tys[2] = { Ty, NarrowTy }; 2089 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vpaddl"); 2090 } 2091 case NEON::BI__builtin_neon_vpmax_v: 2092 Int = Usgn ? Intrinsic::arm_neon_vpmaxu : Intrinsic::arm_neon_vpmaxs; 2093 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vpmax"); 2094 case NEON::BI__builtin_neon_vpmin_v: 2095 Int = Usgn ? Intrinsic::arm_neon_vpminu : Intrinsic::arm_neon_vpmins; 2096 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vpmin"); 2097 case NEON::BI__builtin_neon_vqabs_v: 2098 case NEON::BI__builtin_neon_vqabsq_v: 2099 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vqabs, Ty), 2100 Ops, "vqabs"); 2101 case NEON::BI__builtin_neon_vqadd_v: 2102 case NEON::BI__builtin_neon_vqaddq_v: 2103 Int = Usgn ? Intrinsic::arm_neon_vqaddu : Intrinsic::arm_neon_vqadds; 2104 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqadd"); 2105 case NEON::BI__builtin_neon_vqmovn_v: 2106 Int = Usgn ? Intrinsic::arm_neon_vqmovnu : Intrinsic::arm_neon_vqmovns; 2107 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqmovn"); 2108 case NEON::BI__builtin_neon_vqmovun_v: 2109 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vqmovnsu, Ty), 2110 Ops, "vqdmull"); 2111 case NEON::BI__builtin_neon_vqneg_v: 2112 case NEON::BI__builtin_neon_vqnegq_v: 2113 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vqneg, Ty), 2114 Ops, "vqneg"); 2115 case NEON::BI__builtin_neon_vqsub_v: 2116 case NEON::BI__builtin_neon_vqsubq_v: 2117 Int = Usgn ? Intrinsic::arm_neon_vqsubu : Intrinsic::arm_neon_vqsubs; 2118 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqsub"); 2119 case NEON::BI__builtin_neon_vqdmlal_v: { 2120 SmallVector<Value *, 2> MulOps(Ops.begin() + 1, Ops.end()); 2121 Value *Mul = EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vqdmull, Ty), 2122 MulOps, "vqdmlal"); 2123 2124 SmallVector<Value *, 2> AddOps; 2125 AddOps.push_back(Ops[0]); 2126 AddOps.push_back(Mul); 2127 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vqadds, Ty), 2128 AddOps, "vqdmlal"); 2129 } 2130 case NEON::BI__builtin_neon_vqdmlsl_v: { 2131 SmallVector<Value *, 2> MulOps(Ops.begin() + 1, Ops.end()); 2132 Value *Mul = EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vqdmull, Ty), 2133 MulOps, "vqdmlsl"); 2134 2135 SmallVector<Value *, 2> SubOps; 2136 SubOps.push_back(Ops[0]); 2137 SubOps.push_back(Mul); 2138 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vqsubs, Ty), 2139 SubOps, "vqdmlsl"); 2140 } 2141 case NEON::BI__builtin_neon_vqdmulh_v: 2142 case NEON::BI__builtin_neon_vqdmulhq_v: 2143 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vqdmulh, Ty), 2144 Ops, "vqdmulh"); 2145 case NEON::BI__builtin_neon_vqdmull_v: 2146 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vqdmull, Ty), 2147 Ops, "vqdmull"); 2148 case NEON::BI__builtin_neon_vqshl_n_v: 2149 case NEON::BI__builtin_neon_vqshlq_n_v: 2150 Int = Usgn ? Intrinsic::arm_neon_vqshiftu : Intrinsic::arm_neon_vqshifts; 2151 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqshl_n", 2152 1, false); 2153 case NEON::BI__builtin_neon_vqrdmulh_v: 2154 case NEON::BI__builtin_neon_vqrdmulhq_v: 2155 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vqrdmulh, Ty), 2156 Ops, "vqrdmulh"); 2157 case NEON::BI__builtin_neon_vqrshl_v: 2158 case NEON::BI__builtin_neon_vqrshlq_v: 2159 Int = Usgn ? Intrinsic::arm_neon_vqrshiftu : Intrinsic::arm_neon_vqrshifts; 2160 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqrshl"); 2161 case NEON::BI__builtin_neon_vqshl_v: 2162 case NEON::BI__builtin_neon_vqshlq_v: 2163 Int = Usgn ? Intrinsic::arm_neon_vqshiftu : Intrinsic::arm_neon_vqshifts; 2164 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqshl"); 2165 case NEON::BI__builtin_neon_vraddhn_v: 2166 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vraddhn, Ty), 2167 Ops, "vraddhn"); 2168 case NEON::BI__builtin_neon_vrecpe_v: 2169 case NEON::BI__builtin_neon_vrecpeq_v: 2170 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vrecpe, Ty), 2171 Ops, "vrecpe"); 2172 case NEON::BI__builtin_neon_vrecps_v: 2173 case NEON::BI__builtin_neon_vrecpsq_v: 2174 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vrecps, Ty), 2175 Ops, "vrecps"); 2176 case NEON::BI__builtin_neon_vrhadd_v: 2177 case NEON::BI__builtin_neon_vrhaddq_v: 2178 Int = Usgn ? Intrinsic::arm_neon_vrhaddu : Intrinsic::arm_neon_vrhadds; 2179 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrhadd"); 2180 case NEON::BI__builtin_neon_vrshl_v: 2181 case NEON::BI__builtin_neon_vrshlq_v: 2182 Int = Usgn ? Intrinsic::arm_neon_vrshiftu : Intrinsic::arm_neon_vrshifts; 2183 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrshl"); 2184 case NEON::BI__builtin_neon_vrsqrte_v: 2185 case NEON::BI__builtin_neon_vrsqrteq_v: 2186 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vrsqrte, Ty), 2187 Ops, "vrsqrte"); 2188 case NEON::BI__builtin_neon_vrsqrts_v: 2189 case NEON::BI__builtin_neon_vrsqrtsq_v: 2190 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vrsqrts, Ty), 2191 Ops, "vrsqrts"); 2192 case NEON::BI__builtin_neon_vrsubhn_v: 2193 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vrsubhn, Ty), 2194 Ops, "vrsubhn"); 2195 case NEON::BI__builtin_neon_vsha1su1q_v: 2196 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_sha1su1), 2197 Ops, "sha1su1"); 2198 case NEON::BI__builtin_neon_vsha256su0q_v: 2199 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_sha256su0), 2200 Ops, "sha256su0"); 2201 case NEON::BI__builtin_neon_vsha1su0q_v: 2202 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_sha1su0), 2203 Ops, "sha1su0"); 2204 case NEON::BI__builtin_neon_vsha256hq_v: 2205 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_sha256h), 2206 Ops, "sha256h"); 2207 case NEON::BI__builtin_neon_vsha256h2q_v: 2208 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_sha256h2), 2209 Ops, "sha256h2"); 2210 case NEON::BI__builtin_neon_vsha256su1q_v: 2211 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_sha256su1), 2212 Ops, "sha256su1"); 2213 case NEON::BI__builtin_neon_vshl_n_v: 2214 case NEON::BI__builtin_neon_vshlq_n_v: 2215 Ops[1] = EmitNeonShiftVector(Ops[1], Ty, false); 2216 return Builder.CreateShl(Builder.CreateBitCast(Ops[0],Ty), Ops[1], 2217 "vshl_n"); 2218 case NEON::BI__builtin_neon_vshll_n_v: { 2219 llvm::Type *SrcTy = llvm::VectorType::getTruncatedElementVectorType(VTy); 2220 Ops[0] = Builder.CreateBitCast(Ops[0], SrcTy); 2221 if (Usgn) 2222 Ops[0] = Builder.CreateZExt(Ops[0], VTy); 2223 else 2224 Ops[0] = Builder.CreateSExt(Ops[0], VTy); 2225 Ops[1] = EmitNeonShiftVector(Ops[1], VTy, false); 2226 return Builder.CreateShl(Ops[0], Ops[1], "vshll_n"); 2227 } 2228 case NEON::BI__builtin_neon_vshl_v: 2229 case NEON::BI__builtin_neon_vshlq_v: 2230 Int = Usgn ? Intrinsic::arm_neon_vshiftu : Intrinsic::arm_neon_vshifts; 2231 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vshl"); 2232 case NEON::BI__builtin_neon_vshrn_n_v: { 2233 llvm::Type *SrcTy = llvm::VectorType::getExtendedElementVectorType(VTy); 2234 Ops[0] = Builder.CreateBitCast(Ops[0], SrcTy); 2235 Ops[1] = EmitNeonShiftVector(Ops[1], SrcTy, false); 2236 if (Usgn) 2237 Ops[0] = Builder.CreateLShr(Ops[0], Ops[1]); 2238 else 2239 Ops[0] = Builder.CreateAShr(Ops[0], Ops[1]); 2240 return Builder.CreateTrunc(Ops[0], Ty, "vshrn_n"); 2241 } 2242 case NEON::BI__builtin_neon_vshr_n_v: 2243 case NEON::BI__builtin_neon_vshrq_n_v: 2244 return EmitNeonRShiftImm(Ops[0], Ops[1], Ty, Usgn, "vshr_n"); 2245 case NEON::BI__builtin_neon_vst1_v: 2246 case NEON::BI__builtin_neon_vst1q_v: 2247 Ops.push_back(Align); 2248 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vst1, Ty), 2249 Ops, ""); 2250 case NEON::BI__builtin_neon_vst2_v: 2251 case NEON::BI__builtin_neon_vst2q_v: 2252 Ops.push_back(Align); 2253 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vst2, Ty), 2254 Ops, ""); 2255 case NEON::BI__builtin_neon_vst3_v: 2256 case NEON::BI__builtin_neon_vst3q_v: 2257 Ops.push_back(Align); 2258 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vst3, Ty), 2259 Ops, ""); 2260 case NEON::BI__builtin_neon_vst4_v: 2261 case NEON::BI__builtin_neon_vst4q_v: 2262 Ops.push_back(Align); 2263 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vst4, Ty), 2264 Ops, ""); 2265 case NEON::BI__builtin_neon_vst2_lane_v: 2266 case NEON::BI__builtin_neon_vst2q_lane_v: 2267 Ops.push_back(Align); 2268 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vst2lane, Ty), 2269 Ops, ""); 2270 case NEON::BI__builtin_neon_vst3_lane_v: 2271 case NEON::BI__builtin_neon_vst3q_lane_v: 2272 Ops.push_back(Align); 2273 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vst3lane, Ty), 2274 Ops, ""); 2275 case NEON::BI__builtin_neon_vst4_lane_v: 2276 case NEON::BI__builtin_neon_vst4q_lane_v: 2277 Ops.push_back(Align); 2278 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vst4lane, Ty), 2279 Ops, ""); 2280 case NEON::BI__builtin_neon_vsubhn_v: { 2281 llvm::VectorType *SrcTy = 2282 llvm::VectorType::getExtendedElementVectorType(VTy); 2283 2284 // %sum = add <4 x i32> %lhs, %rhs 2285 Ops[0] = Builder.CreateBitCast(Ops[0], SrcTy); 2286 Ops[1] = Builder.CreateBitCast(Ops[1], SrcTy); 2287 Ops[0] = Builder.CreateSub(Ops[0], Ops[1], "vsubhn"); 2288 2289 // %high = lshr <4 x i32> %sum, <i32 16, i32 16, i32 16, i32 16> 2290 Constant *ShiftAmt = ConstantInt::get(SrcTy->getElementType(), 2291 SrcTy->getScalarSizeInBits() / 2); 2292 ShiftAmt = ConstantVector::getSplat(VTy->getNumElements(), ShiftAmt); 2293 Ops[0] = Builder.CreateLShr(Ops[0], ShiftAmt, "vsubhn"); 2294 2295 // %res = trunc <4 x i32> %high to <4 x i16> 2296 return Builder.CreateTrunc(Ops[0], VTy, "vsubhn"); 2297 } 2298 case NEON::BI__builtin_neon_vtrn_v: 2299 case NEON::BI__builtin_neon_vtrnq_v: { 2300 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty)); 2301 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 2302 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 2303 Value *SV = 0; 2304 2305 for (unsigned vi = 0; vi != 2; ++vi) { 2306 SmallVector<Constant*, 16> Indices; 2307 for (unsigned i = 0, e = VTy->getNumElements(); i != e; i += 2) { 2308 Indices.push_back(Builder.getInt32(i+vi)); 2309 Indices.push_back(Builder.getInt32(i+e+vi)); 2310 } 2311 Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ops[0], vi); 2312 SV = llvm::ConstantVector::get(Indices); 2313 SV = Builder.CreateShuffleVector(Ops[1], Ops[2], SV, "vtrn"); 2314 SV = Builder.CreateStore(SV, Addr); 2315 } 2316 return SV; 2317 } 2318 case NEON::BI__builtin_neon_vtst_v: 2319 case NEON::BI__builtin_neon_vtstq_v: { 2320 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 2321 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 2322 Ops[0] = Builder.CreateAnd(Ops[0], Ops[1]); 2323 Ops[0] = Builder.CreateICmp(ICmpInst::ICMP_NE, Ops[0], 2324 ConstantAggregateZero::get(Ty)); 2325 return Builder.CreateSExt(Ops[0], Ty, "vtst"); 2326 } 2327 case NEON::BI__builtin_neon_vuzp_v: 2328 case NEON::BI__builtin_neon_vuzpq_v: { 2329 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty)); 2330 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 2331 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 2332 Value *SV = 0; 2333 2334 for (unsigned vi = 0; vi != 2; ++vi) { 2335 SmallVector<Constant*, 16> Indices; 2336 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) 2337 Indices.push_back(ConstantInt::get(Int32Ty, 2*i+vi)); 2338 2339 Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ops[0], vi); 2340 SV = llvm::ConstantVector::get(Indices); 2341 SV = Builder.CreateShuffleVector(Ops[1], Ops[2], SV, "vuzp"); 2342 SV = Builder.CreateStore(SV, Addr); 2343 } 2344 return SV; 2345 } 2346 case NEON::BI__builtin_neon_vzip_v: 2347 case NEON::BI__builtin_neon_vzipq_v: { 2348 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty)); 2349 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 2350 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 2351 Value *SV = 0; 2352 2353 for (unsigned vi = 0; vi != 2; ++vi) { 2354 SmallVector<Constant*, 16> Indices; 2355 for (unsigned i = 0, e = VTy->getNumElements(); i != e; i += 2) { 2356 Indices.push_back(ConstantInt::get(Int32Ty, (i + vi*e) >> 1)); 2357 Indices.push_back(ConstantInt::get(Int32Ty, ((i + vi*e) >> 1)+e)); 2358 } 2359 Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ops[0], vi); 2360 SV = llvm::ConstantVector::get(Indices); 2361 SV = Builder.CreateShuffleVector(Ops[1], Ops[2], SV, "vzip"); 2362 SV = Builder.CreateStore(SV, Addr); 2363 } 2364 return SV; 2365 } 2366 } 2367 2368 return 0; 2369 } 2370 2371 static Value *EmitAArch64ScalarBuiltinExpr(CodeGenFunction &CGF, 2372 unsigned BuiltinID, 2373 const CallExpr *E) { 2374 unsigned int Int = 0; 2375 unsigned IntTypes = 0; 2376 enum { 2377 ScalarRet = (1 << 0), 2378 VectorRet = (1 << 1), 2379 ScalarArg0 = (1 << 2), 2380 VectorGetArg0 = (1 << 3), 2381 VectorCastArg0 = (1 << 4), 2382 ScalarArg1 = (1 << 5), 2383 VectorGetArg1 = (1 << 6), 2384 VectorCastArg1 = (1 << 7), 2385 ScalarFpCmpzArg1 = (1 << 8) 2386 }; 2387 const char *s = NULL; 2388 2389 SmallVector<Value *, 4> Ops; 2390 for (unsigned i = 0, e = E->getNumArgs(); i != e; i++) { 2391 Ops.push_back(CGF.EmitScalarExpr(E->getArg(i))); 2392 } 2393 2394 // AArch64 scalar builtins are not overloaded, they do not have an extra 2395 // argument that specifies the vector type, need to handle each case. 2396 switch (BuiltinID) { 2397 default: break; 2398 case NEON::BI__builtin_neon_vdups_lane_f32: 2399 case NEON::BI__builtin_neon_vdupd_lane_f64: 2400 case NEON::BI__builtin_neon_vdups_laneq_f32: 2401 case NEON::BI__builtin_neon_vdupd_laneq_f64: { 2402 return CGF.Builder.CreateExtractElement(Ops[0], Ops[1], "vdup_lane"); 2403 } 2404 case NEON::BI__builtin_neon_vdupb_lane_i8: 2405 case NEON::BI__builtin_neon_vduph_lane_i16: 2406 case NEON::BI__builtin_neon_vdups_lane_i32: 2407 case NEON::BI__builtin_neon_vdupd_lane_i64: 2408 case NEON::BI__builtin_neon_vdupb_laneq_i8: 2409 case NEON::BI__builtin_neon_vduph_laneq_i16: 2410 case NEON::BI__builtin_neon_vdups_laneq_i32: 2411 case NEON::BI__builtin_neon_vdupd_laneq_i64: { 2412 // The backend treats Neon scalar types as v1ix types 2413 // So we want to dup lane from any vector to v1ix vector 2414 // with shufflevector 2415 s = "vdup_lane"; 2416 Value* SV = llvm::ConstantVector::getSplat(1, cast<ConstantInt>(Ops[1])); 2417 Value *Result = CGF.Builder.CreateShuffleVector(Ops[0], Ops[0], SV, s); 2418 llvm::Type *Ty = CGF.ConvertType(E->getCallReturnType()); 2419 // AArch64 intrinsic one-element vector type cast to 2420 // scalar type expected by the builtin 2421 return CGF.Builder.CreateBitCast(Result, Ty, s); 2422 } 2423 case NEON::BI__builtin_neon_vqdmlalh_lane_s16 : 2424 case NEON::BI__builtin_neon_vqdmlalh_laneq_s16 : 2425 case NEON::BI__builtin_neon_vqdmlals_lane_s32 : 2426 case NEON::BI__builtin_neon_vqdmlals_laneq_s32 : 2427 case NEON::BI__builtin_neon_vqdmlslh_lane_s16 : 2428 case NEON::BI__builtin_neon_vqdmlslh_laneq_s16 : 2429 case NEON::BI__builtin_neon_vqdmlsls_lane_s32 : 2430 case NEON::BI__builtin_neon_vqdmlsls_laneq_s32 : { 2431 Int = Intrinsic::arm_neon_vqadds; 2432 if (BuiltinID == NEON::BI__builtin_neon_vqdmlslh_lane_s16 || 2433 BuiltinID == NEON::BI__builtin_neon_vqdmlslh_laneq_s16 || 2434 BuiltinID == NEON::BI__builtin_neon_vqdmlsls_lane_s32 || 2435 BuiltinID == NEON::BI__builtin_neon_vqdmlsls_laneq_s32) { 2436 Int = Intrinsic::arm_neon_vqsubs; 2437 } 2438 // create vqdmull call with b * c[i] 2439 llvm::Type *Ty = CGF.ConvertType(E->getArg(1)->getType()); 2440 llvm::VectorType *OpVTy = llvm::VectorType::get(Ty, 1); 2441 Ty = CGF.ConvertType(E->getArg(0)->getType()); 2442 llvm::VectorType *ResVTy = llvm::VectorType::get(Ty, 1); 2443 Value *F = CGF.CGM.getIntrinsic(Intrinsic::arm_neon_vqdmull, ResVTy); 2444 Value *V = UndefValue::get(OpVTy); 2445 llvm::Constant *CI = ConstantInt::get(CGF.Int32Ty, 0); 2446 SmallVector<Value *, 2> MulOps; 2447 MulOps.push_back(Ops[1]); 2448 MulOps.push_back(Ops[2]); 2449 MulOps[0] = CGF.Builder.CreateInsertElement(V, MulOps[0], CI); 2450 MulOps[1] = CGF.Builder.CreateExtractElement(MulOps[1], Ops[3], "extract"); 2451 MulOps[1] = CGF.Builder.CreateInsertElement(V, MulOps[1], CI); 2452 Value *MulRes = CGF.Builder.CreateCall2(F, MulOps[0], MulOps[1]); 2453 // create vqadds call with a +/- vqdmull result 2454 F = CGF.CGM.getIntrinsic(Int, ResVTy); 2455 SmallVector<Value *, 2> AddOps; 2456 AddOps.push_back(Ops[0]); 2457 AddOps.push_back(MulRes); 2458 V = UndefValue::get(ResVTy); 2459 AddOps[0] = CGF.Builder.CreateInsertElement(V, AddOps[0], CI); 2460 Value *AddRes = CGF.Builder.CreateCall2(F, AddOps[0], AddOps[1]); 2461 return CGF.Builder.CreateBitCast(AddRes, Ty); 2462 } 2463 case NEON::BI__builtin_neon_vfmas_lane_f32: 2464 case NEON::BI__builtin_neon_vfmas_laneq_f32: 2465 case NEON::BI__builtin_neon_vfmad_lane_f64: 2466 case NEON::BI__builtin_neon_vfmad_laneq_f64: { 2467 llvm::Type *Ty = CGF.ConvertType(E->getCallReturnType()); 2468 Value *F = CGF.CGM.getIntrinsic(Intrinsic::fma, Ty); 2469 Ops[2] = CGF.Builder.CreateExtractElement(Ops[2], Ops[3], "extract"); 2470 return CGF.Builder.CreateCall3(F, Ops[1], Ops[2], Ops[0]); 2471 } 2472 // Scalar Floating-point Multiply Extended 2473 case NEON::BI__builtin_neon_vmulxs_f32: 2474 case NEON::BI__builtin_neon_vmulxd_f64: { 2475 Int = Intrinsic::aarch64_neon_vmulx; 2476 llvm::Type *Ty = CGF.ConvertType(E->getCallReturnType()); 2477 return CGF.EmitNeonCall(CGF.CGM.getIntrinsic(Int, Ty), Ops, "vmulx"); 2478 } 2479 case NEON::BI__builtin_neon_vmul_n_f64: { 2480 // v1f64 vmul_n_f64 should be mapped to Neon scalar mul lane 2481 llvm::Type *VTy = GetNeonType(&CGF, 2482 NeonTypeFlags(NeonTypeFlags::Float64, false, false)); 2483 Ops[0] = CGF.Builder.CreateBitCast(Ops[0], VTy); 2484 llvm::Value *Idx = llvm::ConstantInt::get(CGF.Int32Ty, 0); 2485 Ops[0] = CGF.Builder.CreateExtractElement(Ops[0], Idx, "extract"); 2486 Value *Result = CGF.Builder.CreateFMul(Ops[0], Ops[1]); 2487 return CGF.Builder.CreateBitCast(Result, VTy); 2488 } 2489 case NEON::BI__builtin_neon_vget_lane_i8: 2490 case NEON::BI__builtin_neon_vget_lane_i16: 2491 case NEON::BI__builtin_neon_vget_lane_i32: 2492 case NEON::BI__builtin_neon_vget_lane_i64: 2493 case NEON::BI__builtin_neon_vget_lane_f32: 2494 case NEON::BI__builtin_neon_vget_lane_f64: 2495 case NEON::BI__builtin_neon_vgetq_lane_i8: 2496 case NEON::BI__builtin_neon_vgetq_lane_i16: 2497 case NEON::BI__builtin_neon_vgetq_lane_i32: 2498 case NEON::BI__builtin_neon_vgetq_lane_i64: 2499 case NEON::BI__builtin_neon_vgetq_lane_f32: 2500 case NEON::BI__builtin_neon_vgetq_lane_f64: 2501 return CGF.EmitARMBuiltinExpr(NEON::BI__builtin_neon_vget_lane_i8, E); 2502 case NEON::BI__builtin_neon_vset_lane_i8: 2503 case NEON::BI__builtin_neon_vset_lane_i16: 2504 case NEON::BI__builtin_neon_vset_lane_i32: 2505 case NEON::BI__builtin_neon_vset_lane_i64: 2506 case NEON::BI__builtin_neon_vset_lane_f32: 2507 case NEON::BI__builtin_neon_vset_lane_f64: 2508 case NEON::BI__builtin_neon_vsetq_lane_i8: 2509 case NEON::BI__builtin_neon_vsetq_lane_i16: 2510 case NEON::BI__builtin_neon_vsetq_lane_i32: 2511 case NEON::BI__builtin_neon_vsetq_lane_i64: 2512 case NEON::BI__builtin_neon_vsetq_lane_f32: 2513 case NEON::BI__builtin_neon_vsetq_lane_f64: 2514 return CGF.EmitARMBuiltinExpr(NEON::BI__builtin_neon_vset_lane_i8, E); 2515 // Crypto 2516 case NEON::BI__builtin_neon_vsha1h_u32: 2517 Int = Intrinsic::arm_neon_sha1h; 2518 s = "sha1h"; break; 2519 case NEON::BI__builtin_neon_vsha1cq_u32: 2520 Int = Intrinsic::arm_neon_sha1c; 2521 s = "sha1c"; break; 2522 case NEON::BI__builtin_neon_vsha1pq_u32: 2523 Int = Intrinsic::arm_neon_sha1p; 2524 s = "sha1p"; break; 2525 case NEON::BI__builtin_neon_vsha1mq_u32: 2526 Int = Intrinsic::arm_neon_sha1m; 2527 s = "sha1m"; break; 2528 // Scalar Add 2529 case NEON::BI__builtin_neon_vaddd_s64: 2530 Int = Intrinsic::aarch64_neon_vaddds; 2531 s = "vaddds"; break; 2532 case NEON::BI__builtin_neon_vaddd_u64: 2533 Int = Intrinsic::aarch64_neon_vadddu; 2534 s = "vadddu"; break; 2535 // Scalar Sub 2536 case NEON::BI__builtin_neon_vsubd_s64: 2537 Int = Intrinsic::aarch64_neon_vsubds; 2538 s = "vsubds"; break; 2539 case NEON::BI__builtin_neon_vsubd_u64: 2540 Int = Intrinsic::aarch64_neon_vsubdu; 2541 s = "vsubdu"; break; 2542 // Scalar Saturating Add 2543 case NEON::BI__builtin_neon_vqaddb_s8: 2544 case NEON::BI__builtin_neon_vqaddh_s16: 2545 case NEON::BI__builtin_neon_vqadds_s32: 2546 case NEON::BI__builtin_neon_vqaddd_s64: 2547 Int = Intrinsic::arm_neon_vqadds; 2548 s = "vqadds"; IntTypes = VectorRet; break; 2549 case NEON::BI__builtin_neon_vqaddb_u8: 2550 case NEON::BI__builtin_neon_vqaddh_u16: 2551 case NEON::BI__builtin_neon_vqadds_u32: 2552 case NEON::BI__builtin_neon_vqaddd_u64: 2553 Int = Intrinsic::arm_neon_vqaddu; 2554 s = "vqaddu"; IntTypes = VectorRet; break; 2555 // Scalar Saturating Sub 2556 case NEON::BI__builtin_neon_vqsubb_s8: 2557 case NEON::BI__builtin_neon_vqsubh_s16: 2558 case NEON::BI__builtin_neon_vqsubs_s32: 2559 case NEON::BI__builtin_neon_vqsubd_s64: 2560 Int = Intrinsic::arm_neon_vqsubs; 2561 s = "vqsubs"; IntTypes = VectorRet; break; 2562 case NEON::BI__builtin_neon_vqsubb_u8: 2563 case NEON::BI__builtin_neon_vqsubh_u16: 2564 case NEON::BI__builtin_neon_vqsubs_u32: 2565 case NEON::BI__builtin_neon_vqsubd_u64: 2566 Int = Intrinsic::arm_neon_vqsubu; 2567 s = "vqsubu"; IntTypes = VectorRet; break; 2568 // Scalar Shift Left 2569 case NEON::BI__builtin_neon_vshld_s64: 2570 Int = Intrinsic::aarch64_neon_vshlds; 2571 s = "vshlds"; break; 2572 case NEON::BI__builtin_neon_vshld_u64: 2573 Int = Intrinsic::aarch64_neon_vshldu; 2574 s = "vshldu"; break; 2575 // Scalar Saturating Shift Left 2576 case NEON::BI__builtin_neon_vqshlb_s8: 2577 case NEON::BI__builtin_neon_vqshlh_s16: 2578 case NEON::BI__builtin_neon_vqshls_s32: 2579 case NEON::BI__builtin_neon_vqshld_s64: 2580 Int = Intrinsic::aarch64_neon_vqshls; 2581 s = "vqshls"; IntTypes = VectorRet; break; 2582 case NEON::BI__builtin_neon_vqshlb_u8: 2583 case NEON::BI__builtin_neon_vqshlh_u16: 2584 case NEON::BI__builtin_neon_vqshls_u32: 2585 case NEON::BI__builtin_neon_vqshld_u64: 2586 Int = Intrinsic::aarch64_neon_vqshlu; 2587 s = "vqshlu"; IntTypes = VectorRet; break; 2588 // Scalar Rouding Shift Left 2589 case NEON::BI__builtin_neon_vrshld_s64: 2590 Int = Intrinsic::aarch64_neon_vrshlds; 2591 s = "vrshlds"; break; 2592 case NEON::BI__builtin_neon_vrshld_u64: 2593 Int = Intrinsic::aarch64_neon_vrshldu; 2594 s = "vrshldu"; break; 2595 // Scalar Saturating Rouding Shift Left 2596 case NEON::BI__builtin_neon_vqrshlb_s8: 2597 case NEON::BI__builtin_neon_vqrshlh_s16: 2598 case NEON::BI__builtin_neon_vqrshls_s32: 2599 case NEON::BI__builtin_neon_vqrshld_s64: 2600 Int = Intrinsic::aarch64_neon_vqrshls; 2601 s = "vqrshls"; IntTypes = VectorRet; break; 2602 case NEON::BI__builtin_neon_vqrshlb_u8: 2603 case NEON::BI__builtin_neon_vqrshlh_u16: 2604 case NEON::BI__builtin_neon_vqrshls_u32: 2605 case NEON::BI__builtin_neon_vqrshld_u64: 2606 Int = Intrinsic::aarch64_neon_vqrshlu; 2607 s = "vqrshlu"; IntTypes = VectorRet; break; 2608 // Scalar Reduce Pairwise Add 2609 case NEON::BI__builtin_neon_vpaddd_s64: 2610 case NEON::BI__builtin_neon_vpaddd_u64: 2611 Int = Intrinsic::aarch64_neon_vpadd; 2612 s = "vpadd"; break; 2613 case NEON::BI__builtin_neon_vaddv_f32: 2614 case NEON::BI__builtin_neon_vaddvq_f32: 2615 case NEON::BI__builtin_neon_vaddvq_f64: 2616 case NEON::BI__builtin_neon_vpadds_f32: 2617 case NEON::BI__builtin_neon_vpaddd_f64: 2618 Int = Intrinsic::aarch64_neon_vpfadd; 2619 s = "vpfadd"; IntTypes = ScalarRet | VectorCastArg0; break; 2620 // Scalar Reduce Pairwise Floating Point Max 2621 case NEON::BI__builtin_neon_vmaxv_f32: 2622 case NEON::BI__builtin_neon_vpmaxs_f32: 2623 case NEON::BI__builtin_neon_vmaxvq_f64: 2624 case NEON::BI__builtin_neon_vpmaxqd_f64: 2625 Int = Intrinsic::aarch64_neon_vpmax; 2626 s = "vpmax"; IntTypes = ScalarRet | VectorCastArg0; break; 2627 // Scalar Reduce Pairwise Floating Point Min 2628 case NEON::BI__builtin_neon_vminv_f32: 2629 case NEON::BI__builtin_neon_vpmins_f32: 2630 case NEON::BI__builtin_neon_vminvq_f64: 2631 case NEON::BI__builtin_neon_vpminqd_f64: 2632 Int = Intrinsic::aarch64_neon_vpmin; 2633 s = "vpmin"; IntTypes = ScalarRet | VectorCastArg0; break; 2634 // Scalar Reduce Pairwise Floating Point Maxnm 2635 case NEON::BI__builtin_neon_vmaxnmv_f32: 2636 case NEON::BI__builtin_neon_vpmaxnms_f32: 2637 case NEON::BI__builtin_neon_vmaxnmvq_f64: 2638 case NEON::BI__builtin_neon_vpmaxnmqd_f64: 2639 Int = Intrinsic::aarch64_neon_vpfmaxnm; 2640 s = "vpfmaxnm"; IntTypes = ScalarRet | VectorCastArg0; break; 2641 // Scalar Reduce Pairwise Floating Point Minnm 2642 case NEON::BI__builtin_neon_vminnmv_f32: 2643 case NEON::BI__builtin_neon_vpminnms_f32: 2644 case NEON::BI__builtin_neon_vminnmvq_f64: 2645 case NEON::BI__builtin_neon_vpminnmqd_f64: 2646 Int = Intrinsic::aarch64_neon_vpfminnm; 2647 s = "vpfminnm"; IntTypes = ScalarRet | VectorCastArg0; break; 2648 // The followings are intrinsics with scalar results generated AcrossVec vectors 2649 case NEON::BI__builtin_neon_vaddlv_s8: 2650 case NEON::BI__builtin_neon_vaddlv_s16: 2651 case NEON::BI__builtin_neon_vaddlv_s32: 2652 case NEON::BI__builtin_neon_vaddlvq_s8: 2653 case NEON::BI__builtin_neon_vaddlvq_s16: 2654 case NEON::BI__builtin_neon_vaddlvq_s32: 2655 Int = Intrinsic::aarch64_neon_saddlv; 2656 s = "saddlv"; IntTypes = VectorRet | VectorCastArg1; break; 2657 case NEON::BI__builtin_neon_vaddlv_u8: 2658 case NEON::BI__builtin_neon_vaddlv_u16: 2659 case NEON::BI__builtin_neon_vaddlv_u32: 2660 case NEON::BI__builtin_neon_vaddlvq_u8: 2661 case NEON::BI__builtin_neon_vaddlvq_u16: 2662 case NEON::BI__builtin_neon_vaddlvq_u32: 2663 Int = Intrinsic::aarch64_neon_uaddlv; 2664 s = "uaddlv"; IntTypes = VectorRet | VectorCastArg1; break; 2665 case NEON::BI__builtin_neon_vmaxv_s8: 2666 case NEON::BI__builtin_neon_vmaxv_s16: 2667 case NEON::BI__builtin_neon_vmaxv_s32: 2668 case NEON::BI__builtin_neon_vmaxvq_s8: 2669 case NEON::BI__builtin_neon_vmaxvq_s16: 2670 case NEON::BI__builtin_neon_vmaxvq_s32: 2671 Int = Intrinsic::aarch64_neon_smaxv; 2672 s = "smaxv"; IntTypes = VectorRet | VectorCastArg1; break; 2673 case NEON::BI__builtin_neon_vmaxv_u8: 2674 case NEON::BI__builtin_neon_vmaxv_u16: 2675 case NEON::BI__builtin_neon_vmaxv_u32: 2676 case NEON::BI__builtin_neon_vmaxvq_u8: 2677 case NEON::BI__builtin_neon_vmaxvq_u16: 2678 case NEON::BI__builtin_neon_vmaxvq_u32: 2679 Int = Intrinsic::aarch64_neon_umaxv; 2680 s = "umaxv"; IntTypes = VectorRet | VectorCastArg1; break; 2681 case NEON::BI__builtin_neon_vminv_s8: 2682 case NEON::BI__builtin_neon_vminv_s16: 2683 case NEON::BI__builtin_neon_vminv_s32: 2684 case NEON::BI__builtin_neon_vminvq_s8: 2685 case NEON::BI__builtin_neon_vminvq_s16: 2686 case NEON::BI__builtin_neon_vminvq_s32: 2687 Int = Intrinsic::aarch64_neon_sminv; 2688 s = "sminv"; IntTypes = VectorRet | VectorCastArg1; break; 2689 case NEON::BI__builtin_neon_vminv_u8: 2690 case NEON::BI__builtin_neon_vminv_u16: 2691 case NEON::BI__builtin_neon_vminv_u32: 2692 case NEON::BI__builtin_neon_vminvq_u8: 2693 case NEON::BI__builtin_neon_vminvq_u16: 2694 case NEON::BI__builtin_neon_vminvq_u32: 2695 Int = Intrinsic::aarch64_neon_uminv; 2696 s = "uminv"; IntTypes = VectorRet | VectorCastArg1; break; 2697 case NEON::BI__builtin_neon_vaddv_s8: 2698 case NEON::BI__builtin_neon_vaddv_s16: 2699 case NEON::BI__builtin_neon_vaddv_s32: 2700 case NEON::BI__builtin_neon_vaddvq_s8: 2701 case NEON::BI__builtin_neon_vaddvq_s16: 2702 case NEON::BI__builtin_neon_vaddvq_s32: 2703 case NEON::BI__builtin_neon_vaddvq_s64: 2704 case NEON::BI__builtin_neon_vaddv_u8: 2705 case NEON::BI__builtin_neon_vaddv_u16: 2706 case NEON::BI__builtin_neon_vaddv_u32: 2707 case NEON::BI__builtin_neon_vaddvq_u8: 2708 case NEON::BI__builtin_neon_vaddvq_u16: 2709 case NEON::BI__builtin_neon_vaddvq_u32: 2710 case NEON::BI__builtin_neon_vaddvq_u64: 2711 Int = Intrinsic::aarch64_neon_vaddv; 2712 s = "vaddv"; IntTypes = VectorRet | VectorCastArg1; break; 2713 case NEON::BI__builtin_neon_vmaxvq_f32: 2714 Int = Intrinsic::aarch64_neon_vmaxv; 2715 s = "vmaxv"; break; 2716 case NEON::BI__builtin_neon_vminvq_f32: 2717 Int = Intrinsic::aarch64_neon_vminv; 2718 s = "vminv"; break; 2719 case NEON::BI__builtin_neon_vmaxnmvq_f32: 2720 Int = Intrinsic::aarch64_neon_vmaxnmv; 2721 s = "vmaxnmv"; break; 2722 case NEON::BI__builtin_neon_vminnmvq_f32: 2723 Int = Intrinsic::aarch64_neon_vminnmv; 2724 s = "vminnmv"; break; 2725 // Scalar Integer Saturating Doubling Multiply Half High 2726 case NEON::BI__builtin_neon_vqdmulhh_s16: 2727 case NEON::BI__builtin_neon_vqdmulhs_s32: 2728 Int = Intrinsic::arm_neon_vqdmulh; 2729 s = "vqdmulh"; IntTypes = VectorRet; break; 2730 // Scalar Integer Saturating Rounding Doubling Multiply Half High 2731 case NEON::BI__builtin_neon_vqrdmulhh_s16: 2732 case NEON::BI__builtin_neon_vqrdmulhs_s32: 2733 Int = Intrinsic::arm_neon_vqrdmulh; 2734 s = "vqrdmulh"; IntTypes = VectorRet; break; 2735 // Scalar Floating-point Reciprocal Step 2736 case NEON::BI__builtin_neon_vrecpss_f32: 2737 case NEON::BI__builtin_neon_vrecpsd_f64: 2738 Int = Intrinsic::aarch64_neon_vrecps; 2739 s = "vrecps"; IntTypes = ScalarRet; break; 2740 // Scalar Floating-point Reciprocal Square Root Step 2741 case NEON::BI__builtin_neon_vrsqrtss_f32: 2742 case NEON::BI__builtin_neon_vrsqrtsd_f64: 2743 Int = Intrinsic::aarch64_neon_vrsqrts; 2744 s = "vrsqrts"; IntTypes = ScalarRet; break; 2745 // Scalar Signed Integer Convert To Floating-point 2746 case NEON::BI__builtin_neon_vcvts_f32_s32: 2747 case NEON::BI__builtin_neon_vcvtd_f64_s64: 2748 Int = Intrinsic::aarch64_neon_vcvtint2fps; 2749 s = "vcvtf"; IntTypes = ScalarRet | VectorGetArg0; break; 2750 // Scalar Unsigned Integer Convert To Floating-point 2751 case NEON::BI__builtin_neon_vcvts_f32_u32: 2752 case NEON::BI__builtin_neon_vcvtd_f64_u64: 2753 Int = Intrinsic::aarch64_neon_vcvtint2fpu; 2754 s = "vcvtf"; IntTypes = ScalarRet | VectorGetArg0; break; 2755 // Scalar Floating-point Converts 2756 case NEON::BI__builtin_neon_vcvtxd_f32_f64: 2757 Int = Intrinsic::aarch64_neon_fcvtxn; 2758 s = "vcvtxn"; break; 2759 case NEON::BI__builtin_neon_vcvtas_s32_f32: 2760 case NEON::BI__builtin_neon_vcvtad_s64_f64: 2761 Int = Intrinsic::aarch64_neon_fcvtas; 2762 s = "vcvtas"; IntTypes = VectorRet | ScalarArg1; break; 2763 case NEON::BI__builtin_neon_vcvtas_u32_f32: 2764 case NEON::BI__builtin_neon_vcvtad_u64_f64: 2765 Int = Intrinsic::aarch64_neon_fcvtau; 2766 s = "vcvtau"; IntTypes = VectorRet | ScalarArg1; break; 2767 case NEON::BI__builtin_neon_vcvtms_s32_f32: 2768 case NEON::BI__builtin_neon_vcvtmd_s64_f64: 2769 Int = Intrinsic::aarch64_neon_fcvtms; 2770 s = "vcvtms"; IntTypes = VectorRet | ScalarArg1; break; 2771 case NEON::BI__builtin_neon_vcvtms_u32_f32: 2772 case NEON::BI__builtin_neon_vcvtmd_u64_f64: 2773 Int = Intrinsic::aarch64_neon_fcvtmu; 2774 s = "vcvtmu"; IntTypes = VectorRet | ScalarArg1; break; 2775 case NEON::BI__builtin_neon_vcvtns_s32_f32: 2776 case NEON::BI__builtin_neon_vcvtnd_s64_f64: 2777 Int = Intrinsic::aarch64_neon_fcvtns; 2778 s = "vcvtns"; IntTypes = VectorRet | ScalarArg1; break; 2779 case NEON::BI__builtin_neon_vcvtns_u32_f32: 2780 case NEON::BI__builtin_neon_vcvtnd_u64_f64: 2781 Int = Intrinsic::aarch64_neon_fcvtnu; 2782 s = "vcvtnu"; IntTypes = VectorRet | ScalarArg1; break; 2783 case NEON::BI__builtin_neon_vcvtps_s32_f32: 2784 case NEON::BI__builtin_neon_vcvtpd_s64_f64: 2785 Int = Intrinsic::aarch64_neon_fcvtps; 2786 s = "vcvtps"; IntTypes = VectorRet | ScalarArg1; break; 2787 case NEON::BI__builtin_neon_vcvtps_u32_f32: 2788 case NEON::BI__builtin_neon_vcvtpd_u64_f64: 2789 Int = Intrinsic::aarch64_neon_fcvtpu; 2790 s = "vcvtpu"; IntTypes = VectorRet | ScalarArg1; break; 2791 case NEON::BI__builtin_neon_vcvts_s32_f32: 2792 case NEON::BI__builtin_neon_vcvtd_s64_f64: 2793 Int = Intrinsic::aarch64_neon_fcvtzs; 2794 s = "vcvtzs"; IntTypes = VectorRet | ScalarArg1; break; 2795 case NEON::BI__builtin_neon_vcvts_u32_f32: 2796 case NEON::BI__builtin_neon_vcvtd_u64_f64: 2797 Int = Intrinsic::aarch64_neon_fcvtzu; 2798 s = "vcvtzu"; IntTypes = VectorRet | ScalarArg1; break; 2799 // Scalar Floating-point Reciprocal Estimate 2800 case NEON::BI__builtin_neon_vrecpes_f32: 2801 case NEON::BI__builtin_neon_vrecped_f64: 2802 Int = Intrinsic::aarch64_neon_vrecpe; 2803 s = "vrecpe"; IntTypes = ScalarRet; break; 2804 // Scalar Floating-point Reciprocal Exponent 2805 case NEON::BI__builtin_neon_vrecpxs_f32: 2806 case NEON::BI__builtin_neon_vrecpxd_f64: 2807 Int = Intrinsic::aarch64_neon_vrecpx; 2808 s = "vrecpx"; IntTypes = ScalarRet; break; 2809 // Scalar Floating-point Reciprocal Square Root Estimate 2810 case NEON::BI__builtin_neon_vrsqrtes_f32: 2811 case NEON::BI__builtin_neon_vrsqrted_f64: 2812 Int = Intrinsic::aarch64_neon_vrsqrte; 2813 s = "vrsqrte"; IntTypes = ScalarRet; break; 2814 // Scalar Compare Equal 2815 case NEON::BI__builtin_neon_vceqd_s64: 2816 case NEON::BI__builtin_neon_vceqd_u64: 2817 Int = Intrinsic::aarch64_neon_vceq; s = "vceq"; 2818 IntTypes = VectorRet | VectorGetArg0 | VectorGetArg1; break; 2819 // Scalar Compare Equal To Zero 2820 case NEON::BI__builtin_neon_vceqzd_s64: 2821 case NEON::BI__builtin_neon_vceqzd_u64: 2822 Int = Intrinsic::aarch64_neon_vceq; s = "vceq"; 2823 // Add implicit zero operand. 2824 Ops.push_back(llvm::Constant::getNullValue(Ops[0]->getType())); 2825 IntTypes = VectorRet | VectorGetArg0 | VectorGetArg1; break; 2826 // Scalar Compare Greater Than or Equal 2827 case NEON::BI__builtin_neon_vcged_s64: 2828 Int = Intrinsic::aarch64_neon_vcge; s = "vcge"; 2829 IntTypes = VectorRet | VectorGetArg0 | VectorGetArg1; break; 2830 case NEON::BI__builtin_neon_vcged_u64: 2831 Int = Intrinsic::aarch64_neon_vchs; s = "vcge"; 2832 IntTypes = VectorRet | VectorGetArg0 | VectorGetArg1; break; 2833 // Scalar Compare Greater Than or Equal To Zero 2834 case NEON::BI__builtin_neon_vcgezd_s64: 2835 Int = Intrinsic::aarch64_neon_vcge; s = "vcge"; 2836 // Add implicit zero operand. 2837 Ops.push_back(llvm::Constant::getNullValue(Ops[0]->getType())); 2838 IntTypes = VectorRet | VectorGetArg0 | VectorGetArg1; break; 2839 // Scalar Compare Greater Than 2840 case NEON::BI__builtin_neon_vcgtd_s64: 2841 Int = Intrinsic::aarch64_neon_vcgt; s = "vcgt"; 2842 IntTypes = VectorRet | VectorGetArg0 | VectorGetArg1; break; 2843 case NEON::BI__builtin_neon_vcgtd_u64: 2844 Int = Intrinsic::aarch64_neon_vchi; s = "vcgt"; 2845 IntTypes = VectorRet | VectorGetArg0 | VectorGetArg1; break; 2846 // Scalar Compare Greater Than Zero 2847 case NEON::BI__builtin_neon_vcgtzd_s64: 2848 Int = Intrinsic::aarch64_neon_vcgt; s = "vcgt"; 2849 // Add implicit zero operand. 2850 Ops.push_back(llvm::Constant::getNullValue(Ops[0]->getType())); 2851 IntTypes = VectorRet | VectorGetArg0 | VectorGetArg1; break; 2852 // Scalar Compare Less Than or Equal 2853 case NEON::BI__builtin_neon_vcled_s64: 2854 Int = Intrinsic::aarch64_neon_vcge; s = "vcge"; 2855 std::swap(Ops[0], Ops[1]); 2856 IntTypes = VectorRet | VectorGetArg0 | VectorGetArg1; break; 2857 case NEON::BI__builtin_neon_vcled_u64: 2858 Int = Intrinsic::aarch64_neon_vchs; s = "vchs"; 2859 std::swap(Ops[0], Ops[1]); 2860 IntTypes = VectorRet | VectorGetArg0 | VectorGetArg1; break; 2861 // Scalar Compare Less Than or Equal To Zero 2862 case NEON::BI__builtin_neon_vclezd_s64: 2863 Int = Intrinsic::aarch64_neon_vclez; s = "vcle"; 2864 // Add implicit zero operand. 2865 Ops.push_back(llvm::Constant::getNullValue(Ops[0]->getType())); 2866 IntTypes = VectorRet | VectorGetArg0 | VectorGetArg1; break; 2867 // Scalar Compare Less Than 2868 case NEON::BI__builtin_neon_vcltd_s64: 2869 Int = Intrinsic::aarch64_neon_vcgt; s = "vcgt"; 2870 std::swap(Ops[0], Ops[1]); 2871 IntTypes = VectorRet | VectorGetArg0 | VectorGetArg1; break; 2872 case NEON::BI__builtin_neon_vcltd_u64: 2873 Int = Intrinsic::aarch64_neon_vchi; s = "vchi"; 2874 std::swap(Ops[0], Ops[1]); 2875 IntTypes = VectorRet | VectorGetArg0 | VectorGetArg1; break; 2876 // Scalar Compare Less Than Zero 2877 case NEON::BI__builtin_neon_vcltzd_s64: 2878 Int = Intrinsic::aarch64_neon_vcltz; s = "vclt"; 2879 // Add implicit zero operand. 2880 Ops.push_back(llvm::Constant::getNullValue(Ops[0]->getType())); 2881 IntTypes = VectorRet | VectorGetArg0 | VectorGetArg1; break; 2882 // Scalar Floating-point Compare Equal 2883 case NEON::BI__builtin_neon_vceqs_f32: 2884 case NEON::BI__builtin_neon_vceqd_f64: 2885 Int = Intrinsic::aarch64_neon_fceq; s = "vceq"; 2886 IntTypes = VectorRet | ScalarArg0 | ScalarArg1; break; 2887 // Scalar Floating-point Compare Equal To Zero 2888 case NEON::BI__builtin_neon_vceqzs_f32: 2889 case NEON::BI__builtin_neon_vceqzd_f64: 2890 Int = Intrinsic::aarch64_neon_fceq; s = "vceq"; 2891 // Add implicit zero operand. 2892 Ops.push_back(llvm::Constant::getNullValue(CGF.FloatTy)); 2893 IntTypes = VectorRet | ScalarArg0 | ScalarFpCmpzArg1; break; 2894 // Scalar Floating-point Compare Greater Than Or Equal 2895 case NEON::BI__builtin_neon_vcges_f32: 2896 case NEON::BI__builtin_neon_vcged_f64: 2897 Int = Intrinsic::aarch64_neon_fcge; s = "vcge"; 2898 IntTypes = VectorRet | ScalarArg0 | ScalarArg1; break; 2899 // Scalar Floating-point Compare Greater Than Or Equal To Zero 2900 case NEON::BI__builtin_neon_vcgezs_f32: 2901 case NEON::BI__builtin_neon_vcgezd_f64: 2902 Int = Intrinsic::aarch64_neon_fcge; s = "vcge"; 2903 // Add implicit zero operand. 2904 Ops.push_back(llvm::Constant::getNullValue(CGF.FloatTy)); 2905 IntTypes = VectorRet | ScalarArg0 | ScalarFpCmpzArg1; break; 2906 // Scalar Floating-point Compare Greather Than 2907 case NEON::BI__builtin_neon_vcgts_f32: 2908 case NEON::BI__builtin_neon_vcgtd_f64: 2909 Int = Intrinsic::aarch64_neon_fcgt; s = "vcgt"; 2910 IntTypes = VectorRet | ScalarArg0 | ScalarArg1; break; 2911 // Scalar Floating-point Compare Greather Than Zero 2912 case NEON::BI__builtin_neon_vcgtzs_f32: 2913 case NEON::BI__builtin_neon_vcgtzd_f64: 2914 Int = Intrinsic::aarch64_neon_fcgt; s = "vcgt"; 2915 // Add implicit zero operand. 2916 Ops.push_back(llvm::Constant::getNullValue(CGF.FloatTy)); 2917 IntTypes = VectorRet | ScalarArg0 | ScalarFpCmpzArg1; break; 2918 // Scalar Floating-point Compare Less Than or Equal 2919 case NEON::BI__builtin_neon_vcles_f32: 2920 case NEON::BI__builtin_neon_vcled_f64: 2921 Int = Intrinsic::aarch64_neon_fcge; s = "vcge"; 2922 std::swap(Ops[0], Ops[1]); 2923 IntTypes = VectorRet | ScalarArg0 | ScalarArg1; break; 2924 // Scalar Floating-point Compare Less Than Or Equal To Zero 2925 case NEON::BI__builtin_neon_vclezs_f32: 2926 case NEON::BI__builtin_neon_vclezd_f64: 2927 Int = Intrinsic::aarch64_neon_fclez; s = "vcle"; 2928 // Add implicit zero operand. 2929 Ops.push_back(llvm::Constant::getNullValue(CGF.FloatTy)); 2930 IntTypes = VectorRet | ScalarArg0 | ScalarFpCmpzArg1; break; 2931 // Scalar Floating-point Compare Less Than Zero 2932 case NEON::BI__builtin_neon_vclts_f32: 2933 case NEON::BI__builtin_neon_vcltd_f64: 2934 Int = Intrinsic::aarch64_neon_fcgt; s = "vcgt"; 2935 std::swap(Ops[0], Ops[1]); 2936 IntTypes = VectorRet | ScalarArg0 | ScalarArg1; break; 2937 // Scalar Floating-point Compare Less Than Zero 2938 case NEON::BI__builtin_neon_vcltzs_f32: 2939 case NEON::BI__builtin_neon_vcltzd_f64: 2940 Int = Intrinsic::aarch64_neon_fcltz; s = "vclt"; 2941 // Add implicit zero operand. 2942 Ops.push_back(llvm::Constant::getNullValue(CGF.FloatTy)); 2943 IntTypes = VectorRet | ScalarArg0 | ScalarFpCmpzArg1; break; 2944 // Scalar Floating-point Absolute Compare Greater Than Or Equal 2945 case NEON::BI__builtin_neon_vcages_f32: 2946 case NEON::BI__builtin_neon_vcaged_f64: 2947 Int = Intrinsic::aarch64_neon_fcage; s = "vcage"; 2948 IntTypes = VectorRet | ScalarArg0 | ScalarArg1; break; 2949 // Scalar Floating-point Absolute Compare Greater Than 2950 case NEON::BI__builtin_neon_vcagts_f32: 2951 case NEON::BI__builtin_neon_vcagtd_f64: 2952 Int = Intrinsic::aarch64_neon_fcagt; s = "vcagt"; 2953 IntTypes = VectorRet | ScalarArg0 | ScalarArg1; break; 2954 // Scalar Floating-point Absolute Compare Less Than Or Equal 2955 case NEON::BI__builtin_neon_vcales_f32: 2956 case NEON::BI__builtin_neon_vcaled_f64: 2957 Int = Intrinsic::aarch64_neon_fcage; s = "vcage"; 2958 std::swap(Ops[0], Ops[1]); 2959 IntTypes = VectorRet | ScalarArg0 | ScalarArg1; break; 2960 // Scalar Floating-point Absolute Compare Less Than 2961 case NEON::BI__builtin_neon_vcalts_f32: 2962 case NEON::BI__builtin_neon_vcaltd_f64: 2963 Int = Intrinsic::aarch64_neon_fcagt; s = "vcalt"; 2964 std::swap(Ops[0], Ops[1]); 2965 IntTypes = VectorRet | ScalarArg0 | ScalarArg1; break; 2966 // Scalar Compare Bitwise Test Bits 2967 case NEON::BI__builtin_neon_vtstd_s64: 2968 case NEON::BI__builtin_neon_vtstd_u64: 2969 Int = Intrinsic::aarch64_neon_vtstd; s = "vtst"; 2970 IntTypes = VectorRet | VectorGetArg0 | VectorGetArg1; break; 2971 // Scalar Absolute Value 2972 case NEON::BI__builtin_neon_vabsd_s64: 2973 Int = Intrinsic::aarch64_neon_vabs; 2974 s = "vabs"; break; 2975 // Scalar Absolute Difference 2976 case NEON::BI__builtin_neon_vabds_f32: 2977 case NEON::BI__builtin_neon_vabdd_f64: 2978 Int = Intrinsic::aarch64_neon_vabd; 2979 s = "vabd"; IntTypes = ScalarRet; break; 2980 // Scalar Signed Saturating Absolute Value 2981 case NEON::BI__builtin_neon_vqabsb_s8: 2982 case NEON::BI__builtin_neon_vqabsh_s16: 2983 case NEON::BI__builtin_neon_vqabss_s32: 2984 case NEON::BI__builtin_neon_vqabsd_s64: 2985 Int = Intrinsic::arm_neon_vqabs; 2986 s = "vqabs"; IntTypes = VectorRet; break; 2987 // Scalar Negate 2988 case NEON::BI__builtin_neon_vnegd_s64: 2989 Int = Intrinsic::aarch64_neon_vneg; 2990 s = "vneg"; break; 2991 // Scalar Signed Saturating Negate 2992 case NEON::BI__builtin_neon_vqnegb_s8: 2993 case NEON::BI__builtin_neon_vqnegh_s16: 2994 case NEON::BI__builtin_neon_vqnegs_s32: 2995 case NEON::BI__builtin_neon_vqnegd_s64: 2996 Int = Intrinsic::arm_neon_vqneg; 2997 s = "vqneg"; IntTypes = VectorRet; break; 2998 // Scalar Signed Saturating Accumulated of Unsigned Value 2999 case NEON::BI__builtin_neon_vuqaddb_s8: 3000 case NEON::BI__builtin_neon_vuqaddh_s16: 3001 case NEON::BI__builtin_neon_vuqadds_s32: 3002 case NEON::BI__builtin_neon_vuqaddd_s64: 3003 Int = Intrinsic::aarch64_neon_vuqadd; 3004 s = "vuqadd"; IntTypes = VectorRet; break; 3005 // Scalar Unsigned Saturating Accumulated of Signed Value 3006 case NEON::BI__builtin_neon_vsqaddb_u8: 3007 case NEON::BI__builtin_neon_vsqaddh_u16: 3008 case NEON::BI__builtin_neon_vsqadds_u32: 3009 case NEON::BI__builtin_neon_vsqaddd_u64: 3010 Int = Intrinsic::aarch64_neon_vsqadd; 3011 s = "vsqadd"; IntTypes = VectorRet; break; 3012 // Signed Saturating Doubling Multiply-Add Long 3013 case NEON::BI__builtin_neon_vqdmlalh_s16: 3014 case NEON::BI__builtin_neon_vqdmlals_s32: 3015 Int = Intrinsic::aarch64_neon_vqdmlal; 3016 s = "vqdmlal"; IntTypes = VectorRet; break; 3017 // Signed Saturating Doubling Multiply-Subtract Long 3018 case NEON::BI__builtin_neon_vqdmlslh_s16: 3019 case NEON::BI__builtin_neon_vqdmlsls_s32: 3020 Int = Intrinsic::aarch64_neon_vqdmlsl; 3021 s = "vqdmlsl"; IntTypes = VectorRet; break; 3022 // Signed Saturating Doubling Multiply Long 3023 case NEON::BI__builtin_neon_vqdmullh_s16: 3024 case NEON::BI__builtin_neon_vqdmulls_s32: 3025 Int = Intrinsic::arm_neon_vqdmull; 3026 s = "vqdmull"; IntTypes = VectorRet; break; 3027 // Scalar Signed Saturating Extract Unsigned Narrow 3028 case NEON::BI__builtin_neon_vqmovunh_s16: 3029 case NEON::BI__builtin_neon_vqmovuns_s32: 3030 case NEON::BI__builtin_neon_vqmovund_s64: 3031 Int = Intrinsic::arm_neon_vqmovnsu; 3032 s = "vqmovun"; IntTypes = VectorRet; break; 3033 // Scalar Signed Saturating Extract Narrow 3034 case NEON::BI__builtin_neon_vqmovnh_s16: 3035 case NEON::BI__builtin_neon_vqmovns_s32: 3036 case NEON::BI__builtin_neon_vqmovnd_s64: 3037 Int = Intrinsic::arm_neon_vqmovns; 3038 s = "vqmovn"; IntTypes = VectorRet; break; 3039 // Scalar Unsigned Saturating Extract Narrow 3040 case NEON::BI__builtin_neon_vqmovnh_u16: 3041 case NEON::BI__builtin_neon_vqmovns_u32: 3042 case NEON::BI__builtin_neon_vqmovnd_u64: 3043 Int = Intrinsic::arm_neon_vqmovnu; 3044 s = "vqmovn"; IntTypes = VectorRet; break; 3045 // Scalar Signed Shift Right (Immediate) 3046 case NEON::BI__builtin_neon_vshrd_n_s64: 3047 Int = Intrinsic::aarch64_neon_vshrds_n; 3048 s = "vsshr"; break; 3049 // Scalar Unsigned Shift Right (Immediate) 3050 case NEON::BI__builtin_neon_vshrd_n_u64: 3051 Int = Intrinsic::aarch64_neon_vshrdu_n; 3052 s = "vushr"; break; 3053 // Scalar Signed Rounding Shift Right (Immediate) 3054 case NEON::BI__builtin_neon_vrshrd_n_s64: 3055 Int = Intrinsic::aarch64_neon_vsrshr; 3056 s = "vsrshr"; IntTypes = VectorRet; break; 3057 // Scalar Unsigned Rounding Shift Right (Immediate) 3058 case NEON::BI__builtin_neon_vrshrd_n_u64: 3059 Int = Intrinsic::aarch64_neon_vurshr; 3060 s = "vurshr"; IntTypes = VectorRet; break; 3061 // Scalar Signed Shift Right and Accumulate (Immediate) 3062 case NEON::BI__builtin_neon_vsrad_n_s64: 3063 Int = Intrinsic::aarch64_neon_vsrads_n; 3064 s = "vssra"; break; 3065 // Scalar Unsigned Shift Right and Accumulate (Immediate) 3066 case NEON::BI__builtin_neon_vsrad_n_u64: 3067 Int = Intrinsic::aarch64_neon_vsradu_n; 3068 s = "vusra"; break; 3069 // Scalar Signed Rounding Shift Right and Accumulate (Immediate) 3070 case NEON::BI__builtin_neon_vrsrad_n_s64: 3071 Int = Intrinsic::aarch64_neon_vrsrads_n; 3072 s = "vsrsra"; break; 3073 // Scalar Unsigned Rounding Shift Right and Accumulate (Immediate) 3074 case NEON::BI__builtin_neon_vrsrad_n_u64: 3075 Int = Intrinsic::aarch64_neon_vrsradu_n; 3076 s = "vursra"; break; 3077 // Scalar Signed/Unsigned Shift Left (Immediate) 3078 case NEON::BI__builtin_neon_vshld_n_s64: 3079 case NEON::BI__builtin_neon_vshld_n_u64: 3080 Int = Intrinsic::aarch64_neon_vshld_n; 3081 s = "vshl"; break; 3082 // Signed Saturating Shift Left (Immediate) 3083 case NEON::BI__builtin_neon_vqshlb_n_s8: 3084 case NEON::BI__builtin_neon_vqshlh_n_s16: 3085 case NEON::BI__builtin_neon_vqshls_n_s32: 3086 case NEON::BI__builtin_neon_vqshld_n_s64: 3087 Int = Intrinsic::aarch64_neon_vqshls_n; 3088 s = "vsqshl"; IntTypes = VectorRet; break; 3089 // Unsigned Saturating Shift Left (Immediate) 3090 case NEON::BI__builtin_neon_vqshlb_n_u8: 3091 case NEON::BI__builtin_neon_vqshlh_n_u16: 3092 case NEON::BI__builtin_neon_vqshls_n_u32: 3093 case NEON::BI__builtin_neon_vqshld_n_u64: 3094 Int = Intrinsic::aarch64_neon_vqshlu_n; 3095 s = "vuqshl"; IntTypes = VectorRet; break; 3096 // Signed Saturating Shift Left Unsigned (Immediate) 3097 case NEON::BI__builtin_neon_vqshlub_n_s8: 3098 case NEON::BI__builtin_neon_vqshluh_n_s16: 3099 case NEON::BI__builtin_neon_vqshlus_n_s32: 3100 case NEON::BI__builtin_neon_vqshlud_n_s64: 3101 Int = Intrinsic::aarch64_neon_vsqshlu; 3102 s = "vsqshlu"; IntTypes = VectorRet; break; 3103 // Shift Right And Insert (Immediate) 3104 case NEON::BI__builtin_neon_vsrid_n_s64: 3105 case NEON::BI__builtin_neon_vsrid_n_u64: 3106 Int = Intrinsic::aarch64_neon_vsri; 3107 s = "vsri"; IntTypes = VectorRet; break; 3108 // Shift Left And Insert (Immediate) 3109 case NEON::BI__builtin_neon_vslid_n_s64: 3110 case NEON::BI__builtin_neon_vslid_n_u64: 3111 Int = Intrinsic::aarch64_neon_vsli; 3112 s = "vsli"; IntTypes = VectorRet; break; 3113 // Signed Saturating Shift Right Narrow (Immediate) 3114 case NEON::BI__builtin_neon_vqshrnh_n_s16: 3115 case NEON::BI__builtin_neon_vqshrns_n_s32: 3116 case NEON::BI__builtin_neon_vqshrnd_n_s64: 3117 Int = Intrinsic::aarch64_neon_vsqshrn; 3118 s = "vsqshrn"; IntTypes = VectorRet; break; 3119 // Unsigned Saturating Shift Right Narrow (Immediate) 3120 case NEON::BI__builtin_neon_vqshrnh_n_u16: 3121 case NEON::BI__builtin_neon_vqshrns_n_u32: 3122 case NEON::BI__builtin_neon_vqshrnd_n_u64: 3123 Int = Intrinsic::aarch64_neon_vuqshrn; 3124 s = "vuqshrn"; IntTypes = VectorRet; break; 3125 // Signed Saturating Rounded Shift Right Narrow (Immediate) 3126 case NEON::BI__builtin_neon_vqrshrnh_n_s16: 3127 case NEON::BI__builtin_neon_vqrshrns_n_s32: 3128 case NEON::BI__builtin_neon_vqrshrnd_n_s64: 3129 Int = Intrinsic::aarch64_neon_vsqrshrn; 3130 s = "vsqrshrn"; IntTypes = VectorRet; break; 3131 // Unsigned Saturating Rounded Shift Right Narrow (Immediate) 3132 case NEON::BI__builtin_neon_vqrshrnh_n_u16: 3133 case NEON::BI__builtin_neon_vqrshrns_n_u32: 3134 case NEON::BI__builtin_neon_vqrshrnd_n_u64: 3135 Int = Intrinsic::aarch64_neon_vuqrshrn; 3136 s = "vuqrshrn"; IntTypes = VectorRet; break; 3137 // Signed Saturating Shift Right Unsigned Narrow (Immediate) 3138 case NEON::BI__builtin_neon_vqshrunh_n_s16: 3139 case NEON::BI__builtin_neon_vqshruns_n_s32: 3140 case NEON::BI__builtin_neon_vqshrund_n_s64: 3141 Int = Intrinsic::aarch64_neon_vsqshrun; 3142 s = "vsqshrun"; IntTypes = VectorRet; break; 3143 // Signed Saturating Rounded Shift Right Unsigned Narrow (Immediate) 3144 case NEON::BI__builtin_neon_vqrshrunh_n_s16: 3145 case NEON::BI__builtin_neon_vqrshruns_n_s32: 3146 case NEON::BI__builtin_neon_vqrshrund_n_s64: 3147 Int = Intrinsic::aarch64_neon_vsqrshrun; 3148 s = "vsqrshrun"; IntTypes = VectorRet; break; 3149 // Scalar Signed Fixed-point Convert To Floating-Point (Immediate) 3150 case NEON::BI__builtin_neon_vcvts_n_f32_s32: 3151 case NEON::BI__builtin_neon_vcvtd_n_f64_s64: 3152 Int = Intrinsic::aarch64_neon_vcvtfxs2fp_n; 3153 s = "vcvtf"; IntTypes = ScalarRet | VectorGetArg0; break; 3154 // Scalar Unsigned Fixed-point Convert To Floating-Point (Immediate) 3155 case NEON::BI__builtin_neon_vcvts_n_f32_u32: 3156 case NEON::BI__builtin_neon_vcvtd_n_f64_u64: 3157 Int = Intrinsic::aarch64_neon_vcvtfxu2fp_n; 3158 s = "vcvtf"; IntTypes = ScalarRet | VectorGetArg0; break; 3159 // Scalar Floating-point Convert To Signed Fixed-point (Immediate) 3160 case NEON::BI__builtin_neon_vcvts_n_s32_f32: 3161 case NEON::BI__builtin_neon_vcvtd_n_s64_f64: 3162 Int = Intrinsic::aarch64_neon_vcvtfp2fxs_n; 3163 s = "fcvtzs"; IntTypes = VectorRet | ScalarArg0; break; 3164 // Scalar Floating-point Convert To Unsigned Fixed-point (Immediate) 3165 case NEON::BI__builtin_neon_vcvts_n_u32_f32: 3166 case NEON::BI__builtin_neon_vcvtd_n_u64_f64: 3167 Int = Intrinsic::aarch64_neon_vcvtfp2fxu_n; 3168 s = "fcvtzu"; IntTypes = VectorRet | ScalarArg0; break; 3169 case NEON::BI__builtin_neon_vmull_p64: 3170 Int = Intrinsic::aarch64_neon_vmull_p64; 3171 s = "vmull"; break; 3172 } 3173 3174 if (!Int) 3175 return 0; 3176 3177 // Determine the type(s) of this overloaded AArch64 intrinsic. 3178 Function *F = 0; 3179 SmallVector<llvm::Type *, 3> Tys; 3180 3181 // Return type. 3182 if (IntTypes & (ScalarRet | VectorRet)) { 3183 llvm::Type *Ty = CGF.ConvertType(E->getCallReturnType()); 3184 if (IntTypes & ScalarRet) { 3185 // Scalar return value. 3186 Tys.push_back(Ty); 3187 } else if (IntTypes & VectorRet) { 3188 // Convert the scalar return type to one-vector element type. 3189 Tys.push_back(llvm::VectorType::get(Ty, 1)); 3190 } 3191 } 3192 3193 // Arguments. 3194 if (IntTypes & (ScalarArg0 | VectorGetArg0 | VectorCastArg0)) { 3195 const Expr *Arg = E->getArg(0); 3196 llvm::Type *Ty = CGF.ConvertType(Arg->getType()); 3197 if (IntTypes & ScalarArg0) { 3198 // Scalar argument. 3199 Tys.push_back(Ty); 3200 } else if (IntTypes & VectorGetArg0) { 3201 // Convert the scalar argument to one-vector element type. 3202 Tys.push_back(llvm::VectorType::get(Ty, 1)); 3203 } else if (IntTypes & VectorCastArg0) { 3204 // Cast the argument to vector type. 3205 Tys.push_back(cast<llvm::VectorType>(Ty)); 3206 } 3207 } 3208 3209 // The only intrinsics that require a 2nd argument are the compare intrinsics. 3210 // However, the builtins don't always have a 2nd argument (e.g., 3211 // floating-point compare to zero), so we inspect the first argument to 3212 // determine the type. 3213 if (IntTypes & (ScalarArg1 | VectorGetArg1 | VectorCastArg1)) { 3214 const Expr *Arg = E->getArg(0); 3215 llvm::Type *Ty = CGF.ConvertType(Arg->getType()); 3216 if (IntTypes & ScalarArg1) { 3217 // Scalar argument. 3218 Tys.push_back(Ty); 3219 } else if (IntTypes & VectorGetArg1) { 3220 // Convert the scalar argument to one-vector element type. 3221 Tys.push_back(llvm::VectorType::get(Ty, 1)); 3222 } else if (IntTypes & VectorCastArg1) { 3223 // Cast the argument to a vector type. 3224 Tys.push_back(cast<llvm::VectorType>(Ty)); 3225 } 3226 } else if (IntTypes & ScalarFpCmpzArg1) { 3227 // Floating-point zero argument. 3228 Tys.push_back(CGF.FloatTy); 3229 } 3230 3231 if (IntTypes) 3232 F = CGF.CGM.getIntrinsic(Int, Tys); 3233 else 3234 F = CGF.CGM.getIntrinsic(Int); 3235 3236 Value *Result = CGF.EmitNeonCall(F, Ops, s); 3237 llvm::Type *ResultType = CGF.ConvertType(E->getType()); 3238 // AArch64 intrinsic one-element vector type cast to 3239 // scalar type expected by the builtin 3240 return CGF.Builder.CreateBitCast(Result, ResultType, s); 3241 } 3242 3243 Value *CodeGenFunction::EmitAArch64CompareBuiltinExpr( 3244 Value *Op, llvm::Type *Ty, const CmpInst::Predicate Fp, 3245 const CmpInst::Predicate Ip, const Twine &Name) { 3246 llvm::Type *OTy = ((llvm::User *)Op)->getOperand(0)->getType(); 3247 if (OTy->isPointerTy()) 3248 OTy = Ty; 3249 Op = Builder.CreateBitCast(Op, OTy); 3250 if (((llvm::VectorType *)OTy)->getElementType()->isFloatingPointTy()) { 3251 Op = Builder.CreateFCmp(Fp, Op, ConstantAggregateZero::get(OTy)); 3252 } else { 3253 Op = Builder.CreateICmp(Ip, Op, ConstantAggregateZero::get(OTy)); 3254 } 3255 return Builder.CreateSExt(Op, Ty, Name); 3256 } 3257 3258 static Value *packTBLDVectorList(CodeGenFunction &CGF, ArrayRef<Value *> Ops, 3259 Value *ExtOp, Value *IndexOp, 3260 llvm::Type *ResTy, unsigned IntID, 3261 const char *Name) { 3262 SmallVector<Value *, 2> TblOps; 3263 if (ExtOp) 3264 TblOps.push_back(ExtOp); 3265 3266 // Build a vector containing sequential number like (0, 1, 2, ..., 15) 3267 SmallVector<Constant*, 16> Indices; 3268 llvm::VectorType *TblTy = cast<llvm::VectorType>(Ops[0]->getType()); 3269 for (unsigned i = 0, e = TblTy->getNumElements(); i != e; ++i) { 3270 Indices.push_back(ConstantInt::get(CGF.Int32Ty, 2*i)); 3271 Indices.push_back(ConstantInt::get(CGF.Int32Ty, 2*i+1)); 3272 } 3273 Value *SV = llvm::ConstantVector::get(Indices); 3274 3275 int PairPos = 0, End = Ops.size() - 1; 3276 while (PairPos < End) { 3277 TblOps.push_back(CGF.Builder.CreateShuffleVector(Ops[PairPos], 3278 Ops[PairPos+1], SV, Name)); 3279 PairPos += 2; 3280 } 3281 3282 // If there's an odd number of 64-bit lookup table, fill the high 64-bit 3283 // of the 128-bit lookup table with zero. 3284 if (PairPos == End) { 3285 Value *ZeroTbl = ConstantAggregateZero::get(TblTy); 3286 TblOps.push_back(CGF.Builder.CreateShuffleVector(Ops[PairPos], 3287 ZeroTbl, SV, Name)); 3288 } 3289 3290 TblTy = llvm::VectorType::get(TblTy->getElementType(), 3291 2*TblTy->getNumElements()); 3292 llvm::Type *Tys[2] = { ResTy, TblTy }; 3293 3294 Function *TblF; 3295 TblOps.push_back(IndexOp); 3296 TblF = CGF.CGM.getIntrinsic(IntID, Tys); 3297 3298 return CGF.EmitNeonCall(TblF, TblOps, Name); 3299 } 3300 3301 static Value *EmitAArch64TblBuiltinExpr(CodeGenFunction &CGF, 3302 unsigned BuiltinID, 3303 const CallExpr *E) { 3304 unsigned int Int = 0; 3305 const char *s = NULL; 3306 3307 unsigned TblPos; 3308 switch (BuiltinID) { 3309 default: 3310 return 0; 3311 case NEON::BI__builtin_neon_vtbl1_v: 3312 case NEON::BI__builtin_neon_vqtbl1_v: 3313 case NEON::BI__builtin_neon_vqtbl1q_v: 3314 case NEON::BI__builtin_neon_vtbl2_v: 3315 case NEON::BI__builtin_neon_vqtbl2_v: 3316 case NEON::BI__builtin_neon_vqtbl2q_v: 3317 case NEON::BI__builtin_neon_vtbl3_v: 3318 case NEON::BI__builtin_neon_vqtbl3_v: 3319 case NEON::BI__builtin_neon_vqtbl3q_v: 3320 case NEON::BI__builtin_neon_vtbl4_v: 3321 case NEON::BI__builtin_neon_vqtbl4_v: 3322 case NEON::BI__builtin_neon_vqtbl4q_v: 3323 TblPos = 0; 3324 break; 3325 case NEON::BI__builtin_neon_vtbx1_v: 3326 case NEON::BI__builtin_neon_vqtbx1_v: 3327 case NEON::BI__builtin_neon_vqtbx1q_v: 3328 case NEON::BI__builtin_neon_vtbx2_v: 3329 case NEON::BI__builtin_neon_vqtbx2_v: 3330 case NEON::BI__builtin_neon_vqtbx2q_v: 3331 case NEON::BI__builtin_neon_vtbx3_v: 3332 case NEON::BI__builtin_neon_vqtbx3_v: 3333 case NEON::BI__builtin_neon_vqtbx3q_v: 3334 case NEON::BI__builtin_neon_vtbx4_v: 3335 case NEON::BI__builtin_neon_vqtbx4_v: 3336 case NEON::BI__builtin_neon_vqtbx4q_v: 3337 TblPos = 1; 3338 break; 3339 } 3340 3341 assert(E->getNumArgs() >= 3); 3342 3343 // Get the last argument, which specifies the vector type. 3344 llvm::APSInt Result; 3345 const Expr *Arg = E->getArg(E->getNumArgs() - 1); 3346 if (!Arg->isIntegerConstantExpr(Result, CGF.getContext())) 3347 return 0; 3348 3349 // Determine the type of this overloaded NEON intrinsic. 3350 NeonTypeFlags Type(Result.getZExtValue()); 3351 llvm::VectorType *VTy = GetNeonType(&CGF, Type); 3352 llvm::Type *Ty = VTy; 3353 if (!Ty) 3354 return 0; 3355 3356 SmallVector<Value *, 4> Ops; 3357 for (unsigned i = 0, e = E->getNumArgs() - 1; i != e; i++) { 3358 Ops.push_back(CGF.EmitScalarExpr(E->getArg(i))); 3359 } 3360 3361 Arg = E->getArg(TblPos); 3362 llvm::Type *TblTy = CGF.ConvertType(Arg->getType()); 3363 llvm::VectorType *VTblTy = cast<llvm::VectorType>(TblTy); 3364 llvm::Type *Tys[2] = { Ty, VTblTy }; 3365 unsigned nElts = VTy->getNumElements(); 3366 3367 // AArch64 scalar builtins are not overloaded, they do not have an extra 3368 // argument that specifies the vector type, need to handle each case. 3369 SmallVector<Value *, 2> TblOps; 3370 switch (BuiltinID) { 3371 case NEON::BI__builtin_neon_vtbl1_v: { 3372 TblOps.push_back(Ops[0]); 3373 return packTBLDVectorList(CGF, TblOps, 0, Ops[1], Ty, 3374 Intrinsic::aarch64_neon_vtbl1, "vtbl1"); 3375 } 3376 case NEON::BI__builtin_neon_vtbl2_v: { 3377 TblOps.push_back(Ops[0]); 3378 TblOps.push_back(Ops[1]); 3379 return packTBLDVectorList(CGF, TblOps, 0, Ops[2], Ty, 3380 Intrinsic::aarch64_neon_vtbl1, "vtbl1"); 3381 } 3382 case NEON::BI__builtin_neon_vtbl3_v: { 3383 TblOps.push_back(Ops[0]); 3384 TblOps.push_back(Ops[1]); 3385 TblOps.push_back(Ops[2]); 3386 return packTBLDVectorList(CGF, TblOps, 0, Ops[3], Ty, 3387 Intrinsic::aarch64_neon_vtbl2, "vtbl2"); 3388 } 3389 case NEON::BI__builtin_neon_vtbl4_v: { 3390 TblOps.push_back(Ops[0]); 3391 TblOps.push_back(Ops[1]); 3392 TblOps.push_back(Ops[2]); 3393 TblOps.push_back(Ops[3]); 3394 return packTBLDVectorList(CGF, TblOps, 0, Ops[4], Ty, 3395 Intrinsic::aarch64_neon_vtbl2, "vtbl2"); 3396 } 3397 case NEON::BI__builtin_neon_vtbx1_v: { 3398 TblOps.push_back(Ops[1]); 3399 Value *TblRes = packTBLDVectorList(CGF, TblOps, 0, Ops[2], Ty, 3400 Intrinsic::aarch64_neon_vtbl1, "vtbl1"); 3401 3402 llvm::Constant *Eight = ConstantInt::get(VTy->getElementType(), 8); 3403 Value* EightV = llvm::ConstantVector::getSplat(nElts, Eight); 3404 Value *CmpRes = CGF.Builder.CreateICmp(ICmpInst::ICMP_UGE, Ops[2], EightV); 3405 CmpRes = CGF.Builder.CreateSExt(CmpRes, Ty); 3406 3407 SmallVector<Value *, 4> BslOps; 3408 BslOps.push_back(CmpRes); 3409 BslOps.push_back(Ops[0]); 3410 BslOps.push_back(TblRes); 3411 Function *BslF = CGF.CGM.getIntrinsic(Intrinsic::arm_neon_vbsl, Ty); 3412 return CGF.EmitNeonCall(BslF, BslOps, "vbsl"); 3413 } 3414 case NEON::BI__builtin_neon_vtbx2_v: { 3415 TblOps.push_back(Ops[1]); 3416 TblOps.push_back(Ops[2]); 3417 return packTBLDVectorList(CGF, TblOps, Ops[0], Ops[3], Ty, 3418 Intrinsic::aarch64_neon_vtbx1, "vtbx1"); 3419 } 3420 case NEON::BI__builtin_neon_vtbx3_v: { 3421 TblOps.push_back(Ops[1]); 3422 TblOps.push_back(Ops[2]); 3423 TblOps.push_back(Ops[3]); 3424 Value *TblRes = packTBLDVectorList(CGF, TblOps, 0, Ops[4], Ty, 3425 Intrinsic::aarch64_neon_vtbl2, "vtbl2"); 3426 3427 llvm::Constant *TwentyFour = ConstantInt::get(VTy->getElementType(), 24); 3428 Value* TwentyFourV = llvm::ConstantVector::getSplat(nElts, TwentyFour); 3429 Value *CmpRes = CGF.Builder.CreateICmp(ICmpInst::ICMP_UGE, Ops[4], 3430 TwentyFourV); 3431 CmpRes = CGF.Builder.CreateSExt(CmpRes, Ty); 3432 3433 SmallVector<Value *, 4> BslOps; 3434 BslOps.push_back(CmpRes); 3435 BslOps.push_back(Ops[0]); 3436 BslOps.push_back(TblRes); 3437 Function *BslF = CGF.CGM.getIntrinsic(Intrinsic::arm_neon_vbsl, Ty); 3438 return CGF.EmitNeonCall(BslF, BslOps, "vbsl"); 3439 } 3440 case NEON::BI__builtin_neon_vtbx4_v: { 3441 TblOps.push_back(Ops[1]); 3442 TblOps.push_back(Ops[2]); 3443 TblOps.push_back(Ops[3]); 3444 TblOps.push_back(Ops[4]); 3445 return packTBLDVectorList(CGF, TblOps, Ops[0], Ops[5], Ty, 3446 Intrinsic::aarch64_neon_vtbx2, "vtbx2"); 3447 } 3448 case NEON::BI__builtin_neon_vqtbl1_v: 3449 case NEON::BI__builtin_neon_vqtbl1q_v: 3450 Int = Intrinsic::aarch64_neon_vtbl1; s = "vtbl1"; break; 3451 case NEON::BI__builtin_neon_vqtbl2_v: 3452 case NEON::BI__builtin_neon_vqtbl2q_v: { 3453 Int = Intrinsic::aarch64_neon_vtbl2; s = "vtbl2"; break; 3454 case NEON::BI__builtin_neon_vqtbl3_v: 3455 case NEON::BI__builtin_neon_vqtbl3q_v: 3456 Int = Intrinsic::aarch64_neon_vtbl3; s = "vtbl3"; break; 3457 case NEON::BI__builtin_neon_vqtbl4_v: 3458 case NEON::BI__builtin_neon_vqtbl4q_v: 3459 Int = Intrinsic::aarch64_neon_vtbl4; s = "vtbl4"; break; 3460 case NEON::BI__builtin_neon_vqtbx1_v: 3461 case NEON::BI__builtin_neon_vqtbx1q_v: 3462 Int = Intrinsic::aarch64_neon_vtbx1; s = "vtbx1"; break; 3463 case NEON::BI__builtin_neon_vqtbx2_v: 3464 case NEON::BI__builtin_neon_vqtbx2q_v: 3465 Int = Intrinsic::aarch64_neon_vtbx2; s = "vtbx2"; break; 3466 case NEON::BI__builtin_neon_vqtbx3_v: 3467 case NEON::BI__builtin_neon_vqtbx3q_v: 3468 Int = Intrinsic::aarch64_neon_vtbx3; s = "vtbx3"; break; 3469 case NEON::BI__builtin_neon_vqtbx4_v: 3470 case NEON::BI__builtin_neon_vqtbx4q_v: 3471 Int = Intrinsic::aarch64_neon_vtbx4; s = "vtbx4"; break; 3472 } 3473 } 3474 3475 if (!Int) 3476 return 0; 3477 3478 Function *F = CGF.CGM.getIntrinsic(Int, Tys); 3479 return CGF.EmitNeonCall(F, Ops, s); 3480 } 3481 3482 Value *CodeGenFunction::EmitAArch64BuiltinExpr(unsigned BuiltinID, 3483 const CallExpr *E) { 3484 // Process AArch64 scalar builtins 3485 if (Value *Result = EmitAArch64ScalarBuiltinExpr(*this, BuiltinID, E)) 3486 return Result; 3487 3488 // Process AArch64 table lookup builtins 3489 if (Value *Result = EmitAArch64TblBuiltinExpr(*this, BuiltinID, E)) 3490 return Result; 3491 3492 if (BuiltinID == AArch64::BI__clear_cache) { 3493 assert(E->getNumArgs() == 2 && 3494 "Variadic __clear_cache slipped through on AArch64"); 3495 3496 const FunctionDecl *FD = E->getDirectCallee(); 3497 SmallVector<Value *, 2> Ops; 3498 for (unsigned i = 0; i < E->getNumArgs(); i++) 3499 Ops.push_back(EmitScalarExpr(E->getArg(i))); 3500 llvm::Type *Ty = CGM.getTypes().ConvertType(FD->getType()); 3501 llvm::FunctionType *FTy = cast<llvm::FunctionType>(Ty); 3502 StringRef Name = FD->getName(); 3503 return EmitNounwindRuntimeCall(CGM.CreateRuntimeFunction(FTy, Name), Ops); 3504 } 3505 3506 SmallVector<Value *, 4> Ops; 3507 llvm::Value *Align = 0; // Alignment for load/store 3508 3509 if (BuiltinID == NEON::BI__builtin_neon_vldrq_p128) { 3510 Value *Op = EmitScalarExpr(E->getArg(0)); 3511 unsigned addressSpace = 3512 cast<llvm::PointerType>(Op->getType())->getAddressSpace(); 3513 llvm::Type *Ty = llvm::Type::getFP128PtrTy(getLLVMContext(), addressSpace); 3514 Op = Builder.CreateBitCast(Op, Ty); 3515 Op = Builder.CreateLoad(Op); 3516 Ty = llvm::Type::getIntNTy(getLLVMContext(), 128); 3517 return Builder.CreateBitCast(Op, Ty); 3518 } 3519 if (BuiltinID == NEON::BI__builtin_neon_vstrq_p128) { 3520 Value *Op0 = EmitScalarExpr(E->getArg(0)); 3521 unsigned addressSpace = 3522 cast<llvm::PointerType>(Op0->getType())->getAddressSpace(); 3523 llvm::Type *PTy = llvm::Type::getFP128PtrTy(getLLVMContext(), addressSpace); 3524 Op0 = Builder.CreateBitCast(Op0, PTy); 3525 Value *Op1 = EmitScalarExpr(E->getArg(1)); 3526 llvm::Type *Ty = llvm::Type::getFP128Ty(getLLVMContext()); 3527 Op1 = Builder.CreateBitCast(Op1, Ty); 3528 return Builder.CreateStore(Op1, Op0); 3529 } 3530 for (unsigned i = 0, e = E->getNumArgs() - 1; i != e; i++) { 3531 if (i == 0) { 3532 switch (BuiltinID) { 3533 case NEON::BI__builtin_neon_vld1_v: 3534 case NEON::BI__builtin_neon_vld1q_v: 3535 case NEON::BI__builtin_neon_vst1_v: 3536 case NEON::BI__builtin_neon_vst1q_v: 3537 case NEON::BI__builtin_neon_vst2_v: 3538 case NEON::BI__builtin_neon_vst2q_v: 3539 case NEON::BI__builtin_neon_vst3_v: 3540 case NEON::BI__builtin_neon_vst3q_v: 3541 case NEON::BI__builtin_neon_vst4_v: 3542 case NEON::BI__builtin_neon_vst4q_v: 3543 case NEON::BI__builtin_neon_vst1_x2_v: 3544 case NEON::BI__builtin_neon_vst1q_x2_v: 3545 case NEON::BI__builtin_neon_vst1_x3_v: 3546 case NEON::BI__builtin_neon_vst1q_x3_v: 3547 case NEON::BI__builtin_neon_vst1_x4_v: 3548 case NEON::BI__builtin_neon_vst1q_x4_v: 3549 // Handle ld1/st1 lane in this function a little different from ARM. 3550 case NEON::BI__builtin_neon_vld1_lane_v: 3551 case NEON::BI__builtin_neon_vld1q_lane_v: 3552 case NEON::BI__builtin_neon_vst1_lane_v: 3553 case NEON::BI__builtin_neon_vst1q_lane_v: 3554 case NEON::BI__builtin_neon_vst2_lane_v: 3555 case NEON::BI__builtin_neon_vst2q_lane_v: 3556 case NEON::BI__builtin_neon_vst3_lane_v: 3557 case NEON::BI__builtin_neon_vst3q_lane_v: 3558 case NEON::BI__builtin_neon_vst4_lane_v: 3559 case NEON::BI__builtin_neon_vst4q_lane_v: 3560 case NEON::BI__builtin_neon_vld1_dup_v: 3561 case NEON::BI__builtin_neon_vld1q_dup_v: 3562 // Get the alignment for the argument in addition to the value; 3563 // we'll use it later. 3564 std::pair<llvm::Value *, unsigned> Src = 3565 EmitPointerWithAlignment(E->getArg(0)); 3566 Ops.push_back(Src.first); 3567 Align = Builder.getInt32(Src.second); 3568 continue; 3569 } 3570 } 3571 if (i == 1) { 3572 switch (BuiltinID) { 3573 case NEON::BI__builtin_neon_vld2_v: 3574 case NEON::BI__builtin_neon_vld2q_v: 3575 case NEON::BI__builtin_neon_vld3_v: 3576 case NEON::BI__builtin_neon_vld3q_v: 3577 case NEON::BI__builtin_neon_vld4_v: 3578 case NEON::BI__builtin_neon_vld4q_v: 3579 case NEON::BI__builtin_neon_vld1_x2_v: 3580 case NEON::BI__builtin_neon_vld1q_x2_v: 3581 case NEON::BI__builtin_neon_vld1_x3_v: 3582 case NEON::BI__builtin_neon_vld1q_x3_v: 3583 case NEON::BI__builtin_neon_vld1_x4_v: 3584 case NEON::BI__builtin_neon_vld1q_x4_v: 3585 // Handle ld1/st1 dup lane in this function a little different from ARM. 3586 case NEON::BI__builtin_neon_vld2_dup_v: 3587 case NEON::BI__builtin_neon_vld2q_dup_v: 3588 case NEON::BI__builtin_neon_vld3_dup_v: 3589 case NEON::BI__builtin_neon_vld3q_dup_v: 3590 case NEON::BI__builtin_neon_vld4_dup_v: 3591 case NEON::BI__builtin_neon_vld4q_dup_v: 3592 case NEON::BI__builtin_neon_vld2_lane_v: 3593 case NEON::BI__builtin_neon_vld2q_lane_v: 3594 case NEON::BI__builtin_neon_vld3_lane_v: 3595 case NEON::BI__builtin_neon_vld3q_lane_v: 3596 case NEON::BI__builtin_neon_vld4_lane_v: 3597 case NEON::BI__builtin_neon_vld4q_lane_v: 3598 // Get the alignment for the argument in addition to the value; 3599 // we'll use it later. 3600 std::pair<llvm::Value *, unsigned> Src = 3601 EmitPointerWithAlignment(E->getArg(1)); 3602 Ops.push_back(Src.first); 3603 Align = Builder.getInt32(Src.second); 3604 continue; 3605 } 3606 } 3607 Ops.push_back(EmitScalarExpr(E->getArg(i))); 3608 } 3609 3610 // Get the last argument, which specifies the vector type. 3611 llvm::APSInt Result; 3612 const Expr *Arg = E->getArg(E->getNumArgs() - 1); 3613 if (!Arg->isIntegerConstantExpr(Result, getContext())) 3614 return 0; 3615 3616 // Determine the type of this overloaded NEON intrinsic. 3617 NeonTypeFlags Type(Result.getZExtValue()); 3618 bool usgn = Type.isUnsigned(); 3619 bool quad = Type.isQuad(); 3620 3621 llvm::VectorType *VTy = GetNeonType(this, Type); 3622 llvm::Type *Ty = VTy; 3623 if (!Ty) 3624 return 0; 3625 3626 // Many NEON builtins have identical semantics and uses in ARM and 3627 // AArch64. Emit these in a single function. 3628 if (Value *Result = EmitCommonNeonBuiltinExpr(BuiltinID, E, Ops, Align)) 3629 return Result; 3630 3631 unsigned Int; 3632 switch (BuiltinID) { 3633 default: 3634 return 0; 3635 3636 // AArch64 builtins mapping to legacy ARM v7 builtins. 3637 // FIXME: the mapped builtins listed correspond to what has been tested 3638 // in aarch64-neon-intrinsics.c so far. 3639 3640 // Shift by immediate 3641 case NEON::BI__builtin_neon_vrshr_n_v: 3642 case NEON::BI__builtin_neon_vrshrq_n_v: 3643 Int = usgn ? Intrinsic::aarch64_neon_vurshr 3644 : Intrinsic::aarch64_neon_vsrshr; 3645 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrshr_n"); 3646 case NEON::BI__builtin_neon_vsra_n_v: 3647 if (VTy->getElementType()->isIntegerTy(64)) { 3648 Int = usgn ? Intrinsic::aarch64_neon_vsradu_n 3649 : Intrinsic::aarch64_neon_vsrads_n; 3650 return EmitNeonCall(CGM.getIntrinsic(Int), Ops, "vsra_n"); 3651 } 3652 return EmitARMBuiltinExpr(NEON::BI__builtin_neon_vsra_n_v, E); 3653 case NEON::BI__builtin_neon_vsraq_n_v: 3654 return EmitARMBuiltinExpr(NEON::BI__builtin_neon_vsraq_n_v, E); 3655 case NEON::BI__builtin_neon_vrsra_n_v: 3656 if (VTy->getElementType()->isIntegerTy(64)) { 3657 Int = usgn ? Intrinsic::aarch64_neon_vrsradu_n 3658 : Intrinsic::aarch64_neon_vrsrads_n; 3659 return EmitNeonCall(CGM.getIntrinsic(Int), Ops, "vrsra_n"); 3660 } 3661 // fall through 3662 case NEON::BI__builtin_neon_vrsraq_n_v: { 3663 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 3664 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 3665 Int = usgn ? Intrinsic::aarch64_neon_vurshr 3666 : Intrinsic::aarch64_neon_vsrshr; 3667 Ops[1] = Builder.CreateCall2(CGM.getIntrinsic(Int, Ty), Ops[1], Ops[2]); 3668 return Builder.CreateAdd(Ops[0], Ops[1], "vrsra_n"); 3669 } 3670 case NEON::BI__builtin_neon_vqshlu_n_v: 3671 case NEON::BI__builtin_neon_vqshluq_n_v: 3672 Int = Intrinsic::aarch64_neon_vsqshlu; 3673 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqshlu_n"); 3674 case NEON::BI__builtin_neon_vsri_n_v: 3675 case NEON::BI__builtin_neon_vsriq_n_v: 3676 Int = Intrinsic::aarch64_neon_vsri; 3677 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vsri_n"); 3678 case NEON::BI__builtin_neon_vsli_n_v: 3679 case NEON::BI__builtin_neon_vsliq_n_v: 3680 Int = Intrinsic::aarch64_neon_vsli; 3681 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vsli_n"); 3682 case NEON::BI__builtin_neon_vqshrun_n_v: 3683 Int = Intrinsic::aarch64_neon_vsqshrun; 3684 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqshrun_n"); 3685 case NEON::BI__builtin_neon_vrshrn_n_v: 3686 Int = Intrinsic::aarch64_neon_vrshrn; 3687 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrshrn_n"); 3688 case NEON::BI__builtin_neon_vqrshrun_n_v: 3689 Int = Intrinsic::aarch64_neon_vsqrshrun; 3690 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqrshrun_n"); 3691 case NEON::BI__builtin_neon_vqshrn_n_v: 3692 Int = usgn ? Intrinsic::aarch64_neon_vuqshrn 3693 : Intrinsic::aarch64_neon_vsqshrn; 3694 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqshrn_n"); 3695 case NEON::BI__builtin_neon_vqrshrn_n_v: 3696 Int = usgn ? Intrinsic::aarch64_neon_vuqrshrn 3697 : Intrinsic::aarch64_neon_vsqrshrn; 3698 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqrshrn_n"); 3699 3700 // Convert 3701 case NEON::BI__builtin_neon_vcvt_n_f64_v: 3702 case NEON::BI__builtin_neon_vcvtq_n_f64_v: { 3703 llvm::Type *FloatTy = 3704 GetNeonType(this, NeonTypeFlags(NeonTypeFlags::Float64, false, quad)); 3705 llvm::Type *Tys[2] = { FloatTy, Ty }; 3706 Int = usgn ? Intrinsic::arm_neon_vcvtfxu2fp 3707 : Intrinsic::arm_neon_vcvtfxs2fp; 3708 Function *F = CGM.getIntrinsic(Int, Tys); 3709 return EmitNeonCall(F, Ops, "vcvt_n"); 3710 } 3711 3712 // Load/Store 3713 case NEON::BI__builtin_neon_vld1_x2_v: 3714 case NEON::BI__builtin_neon_vld1q_x2_v: 3715 case NEON::BI__builtin_neon_vld1_x3_v: 3716 case NEON::BI__builtin_neon_vld1q_x3_v: 3717 case NEON::BI__builtin_neon_vld1_x4_v: 3718 case NEON::BI__builtin_neon_vld1q_x4_v: { 3719 unsigned Int; 3720 switch (BuiltinID) { 3721 case NEON::BI__builtin_neon_vld1_x2_v: 3722 case NEON::BI__builtin_neon_vld1q_x2_v: 3723 Int = Intrinsic::aarch64_neon_vld1x2; 3724 break; 3725 case NEON::BI__builtin_neon_vld1_x3_v: 3726 case NEON::BI__builtin_neon_vld1q_x3_v: 3727 Int = Intrinsic::aarch64_neon_vld1x3; 3728 break; 3729 case NEON::BI__builtin_neon_vld1_x4_v: 3730 case NEON::BI__builtin_neon_vld1q_x4_v: 3731 Int = Intrinsic::aarch64_neon_vld1x4; 3732 break; 3733 } 3734 Function *F = CGM.getIntrinsic(Int, Ty); 3735 Ops[1] = Builder.CreateCall2(F, Ops[1], Align, "vld1xN"); 3736 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 3737 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 3738 return Builder.CreateStore(Ops[1], Ops[0]); 3739 } 3740 case NEON::BI__builtin_neon_vst1_x2_v: 3741 case NEON::BI__builtin_neon_vst1q_x2_v: 3742 case NEON::BI__builtin_neon_vst1_x3_v: 3743 case NEON::BI__builtin_neon_vst1q_x3_v: 3744 case NEON::BI__builtin_neon_vst1_x4_v: 3745 case NEON::BI__builtin_neon_vst1q_x4_v: { 3746 Ops.push_back(Align); 3747 unsigned Int; 3748 switch (BuiltinID) { 3749 case NEON::BI__builtin_neon_vst1_x2_v: 3750 case NEON::BI__builtin_neon_vst1q_x2_v: 3751 Int = Intrinsic::aarch64_neon_vst1x2; 3752 break; 3753 case NEON::BI__builtin_neon_vst1_x3_v: 3754 case NEON::BI__builtin_neon_vst1q_x3_v: 3755 Int = Intrinsic::aarch64_neon_vst1x3; 3756 break; 3757 case NEON::BI__builtin_neon_vst1_x4_v: 3758 case NEON::BI__builtin_neon_vst1q_x4_v: 3759 Int = Intrinsic::aarch64_neon_vst1x4; 3760 break; 3761 } 3762 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, ""); 3763 } 3764 case NEON::BI__builtin_neon_vld1_lane_v: 3765 case NEON::BI__builtin_neon_vld1q_lane_v: { 3766 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 3767 Ty = llvm::PointerType::getUnqual(VTy->getElementType()); 3768 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 3769 LoadInst *Ld = Builder.CreateLoad(Ops[0]); 3770 Ld->setAlignment(cast<ConstantInt>(Align)->getZExtValue()); 3771 return Builder.CreateInsertElement(Ops[1], Ld, Ops[2], "vld1_lane"); 3772 } 3773 case NEON::BI__builtin_neon_vst1_lane_v: 3774 case NEON::BI__builtin_neon_vst1q_lane_v: { 3775 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 3776 Ops[1] = Builder.CreateExtractElement(Ops[1], Ops[2]); 3777 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 3778 StoreInst *St = 3779 Builder.CreateStore(Ops[1], Builder.CreateBitCast(Ops[0], Ty)); 3780 St->setAlignment(cast<ConstantInt>(Align)->getZExtValue()); 3781 return St; 3782 } 3783 case NEON::BI__builtin_neon_vld2_dup_v: 3784 case NEON::BI__builtin_neon_vld2q_dup_v: 3785 case NEON::BI__builtin_neon_vld3_dup_v: 3786 case NEON::BI__builtin_neon_vld3q_dup_v: 3787 case NEON::BI__builtin_neon_vld4_dup_v: 3788 case NEON::BI__builtin_neon_vld4q_dup_v: { 3789 // Handle 64-bit x 1 elements as a special-case. There is no "dup" needed. 3790 if (VTy->getElementType()->getPrimitiveSizeInBits() == 64 && 3791 VTy->getNumElements() == 1) { 3792 switch (BuiltinID) { 3793 case NEON::BI__builtin_neon_vld2_dup_v: 3794 Int = Intrinsic::arm_neon_vld2; 3795 break; 3796 case NEON::BI__builtin_neon_vld3_dup_v: 3797 Int = Intrinsic::arm_neon_vld3; 3798 break; 3799 case NEON::BI__builtin_neon_vld4_dup_v: 3800 Int = Intrinsic::arm_neon_vld4; 3801 break; 3802 default: 3803 llvm_unreachable("unknown vld_dup intrinsic?"); 3804 } 3805 Function *F = CGM.getIntrinsic(Int, Ty); 3806 Ops[1] = Builder.CreateCall2(F, Ops[1], Align, "vld_dup"); 3807 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 3808 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 3809 return Builder.CreateStore(Ops[1], Ops[0]); 3810 } 3811 switch (BuiltinID) { 3812 case NEON::BI__builtin_neon_vld2_dup_v: 3813 case NEON::BI__builtin_neon_vld2q_dup_v: 3814 Int = Intrinsic::arm_neon_vld2lane; 3815 break; 3816 case NEON::BI__builtin_neon_vld3_dup_v: 3817 case NEON::BI__builtin_neon_vld3q_dup_v: 3818 Int = Intrinsic::arm_neon_vld3lane; 3819 break; 3820 case NEON::BI__builtin_neon_vld4_dup_v: 3821 case NEON::BI__builtin_neon_vld4q_dup_v: 3822 Int = Intrinsic::arm_neon_vld4lane; 3823 break; 3824 } 3825 Function *F = CGM.getIntrinsic(Int, Ty); 3826 llvm::StructType *STy = cast<llvm::StructType>(F->getReturnType()); 3827 3828 SmallVector<Value *, 6> Args; 3829 Args.push_back(Ops[1]); 3830 Args.append(STy->getNumElements(), UndefValue::get(Ty)); 3831 3832 llvm::Constant *CI = ConstantInt::get(Int32Ty, 0); 3833 Args.push_back(CI); 3834 Args.push_back(Align); 3835 3836 Ops[1] = Builder.CreateCall(F, Args, "vld_dup"); 3837 // splat lane 0 to all elts in each vector of the result. 3838 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 3839 Value *Val = Builder.CreateExtractValue(Ops[1], i); 3840 Value *Elt = Builder.CreateBitCast(Val, Ty); 3841 Elt = EmitNeonSplat(Elt, CI); 3842 Elt = Builder.CreateBitCast(Elt, Val->getType()); 3843 Ops[1] = Builder.CreateInsertValue(Ops[1], Elt, i); 3844 } 3845 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 3846 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 3847 return Builder.CreateStore(Ops[1], Ops[0]); 3848 } 3849 3850 case NEON::BI__builtin_neon_vmul_lane_v: 3851 case NEON::BI__builtin_neon_vmul_laneq_v: { 3852 // v1f64 vmul_lane should be mapped to Neon scalar mul lane 3853 bool Quad = false; 3854 if (BuiltinID == NEON::BI__builtin_neon_vmul_laneq_v) 3855 Quad = true; 3856 Ops[0] = Builder.CreateBitCast(Ops[0], DoubleTy); 3857 llvm::Type *VTy = GetNeonType(this, 3858 NeonTypeFlags(NeonTypeFlags::Float64, false, Quad)); 3859 Ops[1] = Builder.CreateBitCast(Ops[1], VTy); 3860 Ops[1] = Builder.CreateExtractElement(Ops[1], Ops[2], "extract"); 3861 Value *Result = Builder.CreateFMul(Ops[0], Ops[1]); 3862 return Builder.CreateBitCast(Result, Ty); 3863 } 3864 3865 // AArch64-only builtins 3866 case NEON::BI__builtin_neon_vfmaq_laneq_v: { 3867 Value *F = CGM.getIntrinsic(Intrinsic::fma, Ty); 3868 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 3869 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 3870 3871 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 3872 Ops[2] = EmitNeonSplat(Ops[2], cast<ConstantInt>(Ops[3])); 3873 return Builder.CreateCall3(F, Ops[2], Ops[1], Ops[0]); 3874 } 3875 case NEON::BI__builtin_neon_vfmaq_lane_v: { 3876 Value *F = CGM.getIntrinsic(Intrinsic::fma, Ty); 3877 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 3878 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 3879 3880 llvm::VectorType *VTy = cast<llvm::VectorType>(Ty); 3881 llvm::Type *STy = llvm::VectorType::get(VTy->getElementType(), 3882 VTy->getNumElements() / 2); 3883 Ops[2] = Builder.CreateBitCast(Ops[2], STy); 3884 Value* SV = llvm::ConstantVector::getSplat(VTy->getNumElements(), 3885 cast<ConstantInt>(Ops[3])); 3886 Ops[2] = Builder.CreateShuffleVector(Ops[2], Ops[2], SV, "lane"); 3887 3888 return Builder.CreateCall3(F, Ops[2], Ops[1], Ops[0]); 3889 } 3890 case NEON::BI__builtin_neon_vfma_lane_v: { 3891 llvm::VectorType *VTy = cast<llvm::VectorType>(Ty); 3892 // v1f64 fma should be mapped to Neon scalar f64 fma 3893 if (VTy && VTy->getElementType() == DoubleTy) { 3894 Ops[0] = Builder.CreateBitCast(Ops[0], DoubleTy); 3895 Ops[1] = Builder.CreateBitCast(Ops[1], DoubleTy); 3896 llvm::Type *VTy = GetNeonType(this, 3897 NeonTypeFlags(NeonTypeFlags::Float64, false, false)); 3898 Ops[2] = Builder.CreateBitCast(Ops[2], VTy); 3899 Ops[2] = Builder.CreateExtractElement(Ops[2], Ops[3], "extract"); 3900 Value *F = CGM.getIntrinsic(Intrinsic::fma, DoubleTy); 3901 Value *Result = Builder.CreateCall3(F, Ops[1], Ops[2], Ops[0]); 3902 return Builder.CreateBitCast(Result, Ty); 3903 } 3904 Value *F = CGM.getIntrinsic(Intrinsic::fma, Ty); 3905 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 3906 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 3907 3908 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 3909 Ops[2] = EmitNeonSplat(Ops[2], cast<ConstantInt>(Ops[3])); 3910 return Builder.CreateCall3(F, Ops[2], Ops[1], Ops[0]); 3911 } 3912 case NEON::BI__builtin_neon_vfma_laneq_v: { 3913 llvm::VectorType *VTy = cast<llvm::VectorType>(Ty); 3914 // v1f64 fma should be mapped to Neon scalar f64 fma 3915 if (VTy && VTy->getElementType() == DoubleTy) { 3916 Ops[0] = Builder.CreateBitCast(Ops[0], DoubleTy); 3917 Ops[1] = Builder.CreateBitCast(Ops[1], DoubleTy); 3918 llvm::Type *VTy = GetNeonType(this, 3919 NeonTypeFlags(NeonTypeFlags::Float64, false, true)); 3920 Ops[2] = Builder.CreateBitCast(Ops[2], VTy); 3921 Ops[2] = Builder.CreateExtractElement(Ops[2], Ops[3], "extract"); 3922 Value *F = CGM.getIntrinsic(Intrinsic::fma, DoubleTy); 3923 Value *Result = Builder.CreateCall3(F, Ops[1], Ops[2], Ops[0]); 3924 return Builder.CreateBitCast(Result, Ty); 3925 } 3926 Value *F = CGM.getIntrinsic(Intrinsic::fma, Ty); 3927 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 3928 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 3929 3930 llvm::Type *STy = llvm::VectorType::get(VTy->getElementType(), 3931 VTy->getNumElements() * 2); 3932 Ops[2] = Builder.CreateBitCast(Ops[2], STy); 3933 Value* SV = llvm::ConstantVector::getSplat(VTy->getNumElements(), 3934 cast<ConstantInt>(Ops[3])); 3935 Ops[2] = Builder.CreateShuffleVector(Ops[2], Ops[2], SV, "lane"); 3936 3937 return Builder.CreateCall3(F, Ops[2], Ops[1], Ops[0]); 3938 } 3939 case NEON::BI__builtin_neon_vfms_v: 3940 case NEON::BI__builtin_neon_vfmsq_v: { 3941 Value *F = CGM.getIntrinsic(Intrinsic::fma, Ty); 3942 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 3943 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 3944 Ops[1] = Builder.CreateFNeg(Ops[1]); 3945 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 3946 3947 // LLVM's fma intrinsic puts the accumulator in the last position, but the 3948 // AArch64 intrinsic has it first. 3949 return Builder.CreateCall3(F, Ops[1], Ops[2], Ops[0]); 3950 } 3951 case NEON::BI__builtin_neon_vmaxnm_v: 3952 case NEON::BI__builtin_neon_vmaxnmq_v: { 3953 Int = Intrinsic::aarch64_neon_vmaxnm; 3954 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmaxnm"); 3955 } 3956 case NEON::BI__builtin_neon_vminnm_v: 3957 case NEON::BI__builtin_neon_vminnmq_v: { 3958 Int = Intrinsic::aarch64_neon_vminnm; 3959 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vminnm"); 3960 } 3961 case NEON::BI__builtin_neon_vpmaxnm_v: 3962 case NEON::BI__builtin_neon_vpmaxnmq_v: { 3963 Int = Intrinsic::aarch64_neon_vpmaxnm; 3964 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vpmaxnm"); 3965 } 3966 case NEON::BI__builtin_neon_vpminnm_v: 3967 case NEON::BI__builtin_neon_vpminnmq_v: { 3968 Int = Intrinsic::aarch64_neon_vpminnm; 3969 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vpminnm"); 3970 } 3971 case NEON::BI__builtin_neon_vpmaxq_v: { 3972 Int = usgn ? Intrinsic::arm_neon_vpmaxu : Intrinsic::arm_neon_vpmaxs; 3973 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vpmax"); 3974 } 3975 case NEON::BI__builtin_neon_vpminq_v: { 3976 Int = usgn ? Intrinsic::arm_neon_vpminu : Intrinsic::arm_neon_vpmins; 3977 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vpmin"); 3978 } 3979 case NEON::BI__builtin_neon_vpaddq_v: { 3980 Int = Intrinsic::arm_neon_vpadd; 3981 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vpadd"); 3982 } 3983 case NEON::BI__builtin_neon_vmulx_v: 3984 case NEON::BI__builtin_neon_vmulxq_v: { 3985 Int = Intrinsic::aarch64_neon_vmulx; 3986 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmulx"); 3987 } 3988 case NEON::BI__builtin_neon_vsqadd_v: 3989 case NEON::BI__builtin_neon_vsqaddq_v: { 3990 Int = Intrinsic::aarch64_neon_usqadd; 3991 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vsqadd"); 3992 } 3993 case NEON::BI__builtin_neon_vuqadd_v: 3994 case NEON::BI__builtin_neon_vuqaddq_v: { 3995 Int = Intrinsic::aarch64_neon_suqadd; 3996 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vuqadd"); 3997 } 3998 case NEON::BI__builtin_neon_vrbit_v: 3999 case NEON::BI__builtin_neon_vrbitq_v: 4000 Int = Intrinsic::aarch64_neon_rbit; 4001 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrbit"); 4002 case NEON::BI__builtin_neon_vcvt_f32_f64: { 4003 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 4004 Ty = GetNeonType(this, NeonTypeFlags(NeonTypeFlags::Float32, false, false)); 4005 return Builder.CreateFPTrunc(Ops[0], Ty, "vcvt"); 4006 } 4007 case NEON::BI__builtin_neon_vcvtx_f32_v: { 4008 llvm::Type *EltTy = FloatTy; 4009 llvm::Type *ResTy = llvm::VectorType::get(EltTy, 2); 4010 llvm::Type *Tys[2] = { ResTy, Ty }; 4011 Int = Intrinsic::aarch64_neon_vcvtxn; 4012 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vcvtx_f32_f64"); 4013 } 4014 case NEON::BI__builtin_neon_vcvt_f64_f32: { 4015 llvm::Type *OpTy = 4016 GetNeonType(this, NeonTypeFlags(NeonTypeFlags::Float32, false, false)); 4017 Ops[0] = Builder.CreateBitCast(Ops[0], OpTy); 4018 return Builder.CreateFPExt(Ops[0], Ty, "vcvt"); 4019 } 4020 case NEON::BI__builtin_neon_vcvt_f64_v: 4021 case NEON::BI__builtin_neon_vcvtq_f64_v: { 4022 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 4023 Ty = GetNeonType(this, NeonTypeFlags(NeonTypeFlags::Float64, false, quad)); 4024 return usgn ? Builder.CreateUIToFP(Ops[0], Ty, "vcvt") 4025 : Builder.CreateSIToFP(Ops[0], Ty, "vcvt"); 4026 } 4027 case NEON::BI__builtin_neon_vrndn_v: 4028 case NEON::BI__builtin_neon_vrndnq_v: { 4029 Int = Intrinsic::aarch64_neon_frintn; 4030 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrndn"); 4031 } 4032 case NEON::BI__builtin_neon_vrnda_v: 4033 case NEON::BI__builtin_neon_vrndaq_v: { 4034 Int = Intrinsic::round; 4035 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrnda"); 4036 } 4037 case NEON::BI__builtin_neon_vrndp_v: 4038 case NEON::BI__builtin_neon_vrndpq_v: { 4039 Int = Intrinsic::ceil; 4040 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrndp"); 4041 } 4042 case NEON::BI__builtin_neon_vrndm_v: 4043 case NEON::BI__builtin_neon_vrndmq_v: { 4044 Int = Intrinsic::floor; 4045 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrndm"); 4046 } 4047 case NEON::BI__builtin_neon_vrndx_v: 4048 case NEON::BI__builtin_neon_vrndxq_v: { 4049 Int = Intrinsic::rint; 4050 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrndx"); 4051 } 4052 case NEON::BI__builtin_neon_vrnd_v: 4053 case NEON::BI__builtin_neon_vrndq_v: { 4054 Int = Intrinsic::trunc; 4055 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrnd"); 4056 } 4057 case NEON::BI__builtin_neon_vrndi_v: 4058 case NEON::BI__builtin_neon_vrndiq_v: { 4059 Int = Intrinsic::nearbyint; 4060 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrndi"); 4061 } 4062 case NEON::BI__builtin_neon_vcvt_s64_v: 4063 case NEON::BI__builtin_neon_vcvt_u64_v: 4064 case NEON::BI__builtin_neon_vcvtq_s64_v: 4065 case NEON::BI__builtin_neon_vcvtq_u64_v: { 4066 llvm::Type *DoubleTy = 4067 GetNeonType(this, NeonTypeFlags(NeonTypeFlags::Float64, false, quad)); 4068 Ops[0] = Builder.CreateBitCast(Ops[0], DoubleTy); 4069 return usgn ? Builder.CreateFPToUI(Ops[0], Ty, "vcvt") 4070 : Builder.CreateFPToSI(Ops[0], Ty, "vcvt"); 4071 } 4072 case NEON::BI__builtin_neon_vcvtn_s32_v: 4073 case NEON::BI__builtin_neon_vcvtnq_s32_v: { 4074 llvm::Type *OpTy = llvm::VectorType::get(FloatTy, VTy->getNumElements()); 4075 llvm::Type *Tys[2] = { Ty, OpTy }; 4076 Int = Intrinsic::arm_neon_vcvtns; 4077 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vcvtns_f32"); 4078 } 4079 case NEON::BI__builtin_neon_vcvtn_s64_v: 4080 case NEON::BI__builtin_neon_vcvtnq_s64_v: { 4081 llvm::Type *OpTy = llvm::VectorType::get(DoubleTy, VTy->getNumElements()); 4082 llvm::Type *Tys[2] = { Ty, OpTy }; 4083 Int = Intrinsic::arm_neon_vcvtns; 4084 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vcvtns_f64"); 4085 } 4086 case NEON::BI__builtin_neon_vcvtn_u32_v: 4087 case NEON::BI__builtin_neon_vcvtnq_u32_v: { 4088 llvm::Type *OpTy = llvm::VectorType::get(FloatTy, VTy->getNumElements()); 4089 llvm::Type *Tys[2] = { Ty, OpTy }; 4090 Int = Intrinsic::arm_neon_vcvtnu; 4091 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vcvtnu_f32"); 4092 } 4093 case NEON::BI__builtin_neon_vcvtn_u64_v: 4094 case NEON::BI__builtin_neon_vcvtnq_u64_v: { 4095 llvm::Type *OpTy = llvm::VectorType::get(DoubleTy, VTy->getNumElements()); 4096 llvm::Type *Tys[2] = { Ty, OpTy }; 4097 Int = Intrinsic::arm_neon_vcvtnu; 4098 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vcvtnu_f64"); 4099 } 4100 case NEON::BI__builtin_neon_vcvtp_s32_v: 4101 case NEON::BI__builtin_neon_vcvtpq_s32_v: { 4102 llvm::Type *OpTy = llvm::VectorType::get(FloatTy, VTy->getNumElements()); 4103 llvm::Type *Tys[2] = { Ty, OpTy }; 4104 Int = Intrinsic::arm_neon_vcvtps; 4105 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vcvtps_f32"); 4106 } 4107 case NEON::BI__builtin_neon_vcvtp_s64_v: 4108 case NEON::BI__builtin_neon_vcvtpq_s64_v: { 4109 llvm::Type *OpTy = llvm::VectorType::get(DoubleTy, VTy->getNumElements()); 4110 llvm::Type *Tys[2] = { Ty, OpTy }; 4111 Int = Intrinsic::arm_neon_vcvtps; 4112 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vcvtps_f64"); 4113 } 4114 case NEON::BI__builtin_neon_vcvtp_u32_v: 4115 case NEON::BI__builtin_neon_vcvtpq_u32_v: { 4116 llvm::Type *OpTy = llvm::VectorType::get(FloatTy, VTy->getNumElements()); 4117 llvm::Type *Tys[2] = { Ty, OpTy }; 4118 Int = Intrinsic::arm_neon_vcvtpu; 4119 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vcvtpu_f32"); 4120 } 4121 case NEON::BI__builtin_neon_vcvtp_u64_v: 4122 case NEON::BI__builtin_neon_vcvtpq_u64_v: { 4123 llvm::Type *OpTy = llvm::VectorType::get(DoubleTy, VTy->getNumElements()); 4124 llvm::Type *Tys[2] = { Ty, OpTy }; 4125 Int = Intrinsic::arm_neon_vcvtpu; 4126 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vcvtpu_f64"); 4127 } 4128 case NEON::BI__builtin_neon_vcvtm_s32_v: 4129 case NEON::BI__builtin_neon_vcvtmq_s32_v: { 4130 llvm::Type *OpTy = llvm::VectorType::get(FloatTy, VTy->getNumElements()); 4131 llvm::Type *Tys[2] = { Ty, OpTy }; 4132 Int = Intrinsic::arm_neon_vcvtms; 4133 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vcvtms_f32"); 4134 } 4135 case NEON::BI__builtin_neon_vcvtm_s64_v: 4136 case NEON::BI__builtin_neon_vcvtmq_s64_v: { 4137 llvm::Type *OpTy = llvm::VectorType::get(DoubleTy, VTy->getNumElements()); 4138 llvm::Type *Tys[2] = { Ty, OpTy }; 4139 Int = Intrinsic::arm_neon_vcvtms; 4140 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vcvtms_f64"); 4141 } 4142 case NEON::BI__builtin_neon_vcvtm_u32_v: 4143 case NEON::BI__builtin_neon_vcvtmq_u32_v: { 4144 llvm::Type *OpTy = llvm::VectorType::get(FloatTy, VTy->getNumElements()); 4145 llvm::Type *Tys[2] = { Ty, OpTy }; 4146 Int = Intrinsic::arm_neon_vcvtmu; 4147 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vcvtmu_f32"); 4148 } 4149 case NEON::BI__builtin_neon_vcvtm_u64_v: 4150 case NEON::BI__builtin_neon_vcvtmq_u64_v: { 4151 llvm::Type *OpTy = llvm::VectorType::get(DoubleTy, VTy->getNumElements()); 4152 llvm::Type *Tys[2] = { Ty, OpTy }; 4153 Int = Intrinsic::arm_neon_vcvtmu; 4154 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vcvtmu_f64"); 4155 } 4156 case NEON::BI__builtin_neon_vcvta_s32_v: 4157 case NEON::BI__builtin_neon_vcvtaq_s32_v: { 4158 llvm::Type *OpTy = llvm::VectorType::get(FloatTy, VTy->getNumElements()); 4159 llvm::Type *Tys[2] = { Ty, OpTy }; 4160 Int = Intrinsic::arm_neon_vcvtas; 4161 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vcvtas_f32"); 4162 } 4163 case NEON::BI__builtin_neon_vcvta_s64_v: 4164 case NEON::BI__builtin_neon_vcvtaq_s64_v: { 4165 llvm::Type *OpTy = llvm::VectorType::get(DoubleTy, VTy->getNumElements()); 4166 llvm::Type *Tys[2] = { Ty, OpTy }; 4167 Int = Intrinsic::arm_neon_vcvtas; 4168 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vcvtas_f64"); 4169 } 4170 case NEON::BI__builtin_neon_vcvta_u32_v: 4171 case NEON::BI__builtin_neon_vcvtaq_u32_v: { 4172 llvm::Type *OpTy = llvm::VectorType::get(FloatTy, VTy->getNumElements()); 4173 llvm::Type *Tys[2] = { Ty, OpTy }; 4174 Int = Intrinsic::arm_neon_vcvtau; 4175 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vcvtau_f32"); 4176 } 4177 case NEON::BI__builtin_neon_vcvta_u64_v: 4178 case NEON::BI__builtin_neon_vcvtaq_u64_v: { 4179 llvm::Type *OpTy = llvm::VectorType::get(DoubleTy, VTy->getNumElements()); 4180 llvm::Type *Tys[2] = { Ty, OpTy }; 4181 Int = Intrinsic::arm_neon_vcvtau; 4182 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vcvtau_f64"); 4183 } 4184 case NEON::BI__builtin_neon_vsqrt_v: 4185 case NEON::BI__builtin_neon_vsqrtq_v: { 4186 Int = Intrinsic::sqrt; 4187 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vsqrt"); 4188 } 4189 case NEON::BI__builtin_neon_vceqz_v: 4190 case NEON::BI__builtin_neon_vceqzq_v: 4191 return EmitAArch64CompareBuiltinExpr(Ops[0], Ty, ICmpInst::FCMP_OEQ, 4192 ICmpInst::ICMP_EQ, "vceqz"); 4193 case NEON::BI__builtin_neon_vcgez_v: 4194 case NEON::BI__builtin_neon_vcgezq_v: 4195 return EmitAArch64CompareBuiltinExpr(Ops[0], Ty, ICmpInst::FCMP_OGE, 4196 ICmpInst::ICMP_SGE, "vcgez"); 4197 case NEON::BI__builtin_neon_vclez_v: 4198 case NEON::BI__builtin_neon_vclezq_v: 4199 return EmitAArch64CompareBuiltinExpr(Ops[0], Ty, ICmpInst::FCMP_OLE, 4200 ICmpInst::ICMP_SLE, "vclez"); 4201 case NEON::BI__builtin_neon_vcgtz_v: 4202 case NEON::BI__builtin_neon_vcgtzq_v: 4203 return EmitAArch64CompareBuiltinExpr(Ops[0], Ty, ICmpInst::FCMP_OGT, 4204 ICmpInst::ICMP_SGT, "vcgtz"); 4205 case NEON::BI__builtin_neon_vcltz_v: 4206 case NEON::BI__builtin_neon_vcltzq_v: 4207 return EmitAArch64CompareBuiltinExpr(Ops[0], Ty, ICmpInst::FCMP_OLT, 4208 ICmpInst::ICMP_SLT, "vcltz"); 4209 } 4210 } 4211 4212 Value *CodeGenFunction::EmitARMBuiltinExpr(unsigned BuiltinID, 4213 const CallExpr *E) { 4214 if (BuiltinID == ARM::BI__clear_cache) { 4215 assert(E->getNumArgs() == 2 && "__clear_cache takes 2 arguments"); 4216 const FunctionDecl *FD = E->getDirectCallee(); 4217 SmallVector<Value*, 2> Ops; 4218 for (unsigned i = 0; i < 2; i++) 4219 Ops.push_back(EmitScalarExpr(E->getArg(i))); 4220 llvm::Type *Ty = CGM.getTypes().ConvertType(FD->getType()); 4221 llvm::FunctionType *FTy = cast<llvm::FunctionType>(Ty); 4222 StringRef Name = FD->getName(); 4223 return EmitNounwindRuntimeCall(CGM.CreateRuntimeFunction(FTy, Name), Ops); 4224 } 4225 4226 if (BuiltinID == ARM::BI__builtin_arm_ldrexd || 4227 (BuiltinID == ARM::BI__builtin_arm_ldrex && 4228 getContext().getTypeSize(E->getType()) == 64)) { 4229 Function *F = CGM.getIntrinsic(Intrinsic::arm_ldrexd); 4230 4231 Value *LdPtr = EmitScalarExpr(E->getArg(0)); 4232 Value *Val = Builder.CreateCall(F, Builder.CreateBitCast(LdPtr, Int8PtrTy), 4233 "ldrexd"); 4234 4235 Value *Val0 = Builder.CreateExtractValue(Val, 1); 4236 Value *Val1 = Builder.CreateExtractValue(Val, 0); 4237 Val0 = Builder.CreateZExt(Val0, Int64Ty); 4238 Val1 = Builder.CreateZExt(Val1, Int64Ty); 4239 4240 Value *ShiftCst = llvm::ConstantInt::get(Int64Ty, 32); 4241 Val = Builder.CreateShl(Val0, ShiftCst, "shl", true /* nuw */); 4242 Val = Builder.CreateOr(Val, Val1); 4243 return Builder.CreateBitCast(Val, ConvertType(E->getType())); 4244 } 4245 4246 if (BuiltinID == ARM::BI__builtin_arm_ldrex) { 4247 Value *LoadAddr = EmitScalarExpr(E->getArg(0)); 4248 4249 QualType Ty = E->getType(); 4250 llvm::Type *RealResTy = ConvertType(Ty); 4251 llvm::Type *IntResTy = llvm::IntegerType::get(getLLVMContext(), 4252 getContext().getTypeSize(Ty)); 4253 LoadAddr = Builder.CreateBitCast(LoadAddr, IntResTy->getPointerTo()); 4254 4255 Function *F = CGM.getIntrinsic(Intrinsic::arm_ldrex, LoadAddr->getType()); 4256 Value *Val = Builder.CreateCall(F, LoadAddr, "ldrex"); 4257 4258 if (RealResTy->isPointerTy()) 4259 return Builder.CreateIntToPtr(Val, RealResTy); 4260 else { 4261 Val = Builder.CreateTruncOrBitCast(Val, IntResTy); 4262 return Builder.CreateBitCast(Val, RealResTy); 4263 } 4264 } 4265 4266 if (BuiltinID == ARM::BI__builtin_arm_strexd || 4267 (BuiltinID == ARM::BI__builtin_arm_strex && 4268 getContext().getTypeSize(E->getArg(0)->getType()) == 64)) { 4269 Function *F = CGM.getIntrinsic(Intrinsic::arm_strexd); 4270 llvm::Type *STy = llvm::StructType::get(Int32Ty, Int32Ty, NULL); 4271 4272 Value *Tmp = CreateMemTemp(E->getArg(0)->getType()); 4273 Value *Val = EmitScalarExpr(E->getArg(0)); 4274 Builder.CreateStore(Val, Tmp); 4275 4276 Value *LdPtr = Builder.CreateBitCast(Tmp,llvm::PointerType::getUnqual(STy)); 4277 Val = Builder.CreateLoad(LdPtr); 4278 4279 Value *Arg0 = Builder.CreateExtractValue(Val, 0); 4280 Value *Arg1 = Builder.CreateExtractValue(Val, 1); 4281 Value *StPtr = Builder.CreateBitCast(EmitScalarExpr(E->getArg(1)), Int8PtrTy); 4282 return Builder.CreateCall3(F, Arg0, Arg1, StPtr, "strexd"); 4283 } 4284 4285 if (BuiltinID == ARM::BI__builtin_arm_strex) { 4286 Value *StoreVal = EmitScalarExpr(E->getArg(0)); 4287 Value *StoreAddr = EmitScalarExpr(E->getArg(1)); 4288 4289 QualType Ty = E->getArg(0)->getType(); 4290 llvm::Type *StoreTy = llvm::IntegerType::get(getLLVMContext(), 4291 getContext().getTypeSize(Ty)); 4292 StoreAddr = Builder.CreateBitCast(StoreAddr, StoreTy->getPointerTo()); 4293 4294 if (StoreVal->getType()->isPointerTy()) 4295 StoreVal = Builder.CreatePtrToInt(StoreVal, Int32Ty); 4296 else { 4297 StoreVal = Builder.CreateBitCast(StoreVal, StoreTy); 4298 StoreVal = Builder.CreateZExtOrBitCast(StoreVal, Int32Ty); 4299 } 4300 4301 Function *F = CGM.getIntrinsic(Intrinsic::arm_strex, StoreAddr->getType()); 4302 return Builder.CreateCall2(F, StoreVal, StoreAddr, "strex"); 4303 } 4304 4305 if (BuiltinID == ARM::BI__builtin_arm_clrex) { 4306 Function *F = CGM.getIntrinsic(Intrinsic::arm_clrex); 4307 return Builder.CreateCall(F); 4308 } 4309 4310 if (BuiltinID == ARM::BI__builtin_arm_sevl) { 4311 Function *F = CGM.getIntrinsic(Intrinsic::arm_sevl); 4312 return Builder.CreateCall(F); 4313 } 4314 4315 // CRC32 4316 Intrinsic::ID CRCIntrinsicID = Intrinsic::not_intrinsic; 4317 switch (BuiltinID) { 4318 case ARM::BI__builtin_arm_crc32b: 4319 CRCIntrinsicID = Intrinsic::arm_crc32b; break; 4320 case ARM::BI__builtin_arm_crc32cb: 4321 CRCIntrinsicID = Intrinsic::arm_crc32cb; break; 4322 case ARM::BI__builtin_arm_crc32h: 4323 CRCIntrinsicID = Intrinsic::arm_crc32h; break; 4324 case ARM::BI__builtin_arm_crc32ch: 4325 CRCIntrinsicID = Intrinsic::arm_crc32ch; break; 4326 case ARM::BI__builtin_arm_crc32w: 4327 case ARM::BI__builtin_arm_crc32d: 4328 CRCIntrinsicID = Intrinsic::arm_crc32w; break; 4329 case ARM::BI__builtin_arm_crc32cw: 4330 case ARM::BI__builtin_arm_crc32cd: 4331 CRCIntrinsicID = Intrinsic::arm_crc32cw; break; 4332 } 4333 4334 if (CRCIntrinsicID != Intrinsic::not_intrinsic) { 4335 Value *Arg0 = EmitScalarExpr(E->getArg(0)); 4336 Value *Arg1 = EmitScalarExpr(E->getArg(1)); 4337 4338 // crc32{c,}d intrinsics are implemnted as two calls to crc32{c,}w 4339 // intrinsics, hence we need different codegen for these cases. 4340 if (BuiltinID == ARM::BI__builtin_arm_crc32d || 4341 BuiltinID == ARM::BI__builtin_arm_crc32cd) { 4342 Value *C1 = llvm::ConstantInt::get(Int64Ty, 32); 4343 Value *Arg1a = Builder.CreateTruncOrBitCast(Arg1, Int32Ty); 4344 Value *Arg1b = Builder.CreateLShr(Arg1, C1); 4345 Arg1b = Builder.CreateTruncOrBitCast(Arg1b, Int32Ty); 4346 4347 Function *F = CGM.getIntrinsic(CRCIntrinsicID); 4348 Value *Res = Builder.CreateCall2(F, Arg0, Arg1a); 4349 return Builder.CreateCall2(F, Res, Arg1b); 4350 } else { 4351 Arg1 = Builder.CreateZExtOrBitCast(Arg1, Int32Ty); 4352 4353 Function *F = CGM.getIntrinsic(CRCIntrinsicID); 4354 return Builder.CreateCall2(F, Arg0, Arg1); 4355 } 4356 } 4357 4358 SmallVector<Value*, 4> Ops; 4359 llvm::Value *Align = 0; 4360 for (unsigned i = 0, e = E->getNumArgs() - 1; i != e; i++) { 4361 if (i == 0) { 4362 switch (BuiltinID) { 4363 case NEON::BI__builtin_neon_vld1_v: 4364 case NEON::BI__builtin_neon_vld1q_v: 4365 case NEON::BI__builtin_neon_vld1q_lane_v: 4366 case NEON::BI__builtin_neon_vld1_lane_v: 4367 case NEON::BI__builtin_neon_vld1_dup_v: 4368 case NEON::BI__builtin_neon_vld1q_dup_v: 4369 case NEON::BI__builtin_neon_vst1_v: 4370 case NEON::BI__builtin_neon_vst1q_v: 4371 case NEON::BI__builtin_neon_vst1q_lane_v: 4372 case NEON::BI__builtin_neon_vst1_lane_v: 4373 case NEON::BI__builtin_neon_vst2_v: 4374 case NEON::BI__builtin_neon_vst2q_v: 4375 case NEON::BI__builtin_neon_vst2_lane_v: 4376 case NEON::BI__builtin_neon_vst2q_lane_v: 4377 case NEON::BI__builtin_neon_vst3_v: 4378 case NEON::BI__builtin_neon_vst3q_v: 4379 case NEON::BI__builtin_neon_vst3_lane_v: 4380 case NEON::BI__builtin_neon_vst3q_lane_v: 4381 case NEON::BI__builtin_neon_vst4_v: 4382 case NEON::BI__builtin_neon_vst4q_v: 4383 case NEON::BI__builtin_neon_vst4_lane_v: 4384 case NEON::BI__builtin_neon_vst4q_lane_v: 4385 // Get the alignment for the argument in addition to the value; 4386 // we'll use it later. 4387 std::pair<llvm::Value*, unsigned> Src = 4388 EmitPointerWithAlignment(E->getArg(0)); 4389 Ops.push_back(Src.first); 4390 Align = Builder.getInt32(Src.second); 4391 continue; 4392 } 4393 } 4394 if (i == 1) { 4395 switch (BuiltinID) { 4396 case NEON::BI__builtin_neon_vld2_v: 4397 case NEON::BI__builtin_neon_vld2q_v: 4398 case NEON::BI__builtin_neon_vld3_v: 4399 case NEON::BI__builtin_neon_vld3q_v: 4400 case NEON::BI__builtin_neon_vld4_v: 4401 case NEON::BI__builtin_neon_vld4q_v: 4402 case NEON::BI__builtin_neon_vld2_lane_v: 4403 case NEON::BI__builtin_neon_vld2q_lane_v: 4404 case NEON::BI__builtin_neon_vld3_lane_v: 4405 case NEON::BI__builtin_neon_vld3q_lane_v: 4406 case NEON::BI__builtin_neon_vld4_lane_v: 4407 case NEON::BI__builtin_neon_vld4q_lane_v: 4408 case NEON::BI__builtin_neon_vld2_dup_v: 4409 case NEON::BI__builtin_neon_vld3_dup_v: 4410 case NEON::BI__builtin_neon_vld4_dup_v: 4411 // Get the alignment for the argument in addition to the value; 4412 // we'll use it later. 4413 std::pair<llvm::Value*, unsigned> Src = 4414 EmitPointerWithAlignment(E->getArg(1)); 4415 Ops.push_back(Src.first); 4416 Align = Builder.getInt32(Src.second); 4417 continue; 4418 } 4419 } 4420 Ops.push_back(EmitScalarExpr(E->getArg(i))); 4421 } 4422 4423 switch (BuiltinID) { 4424 default: break; 4425 // vget_lane and vset_lane are not overloaded and do not have an extra 4426 // argument that specifies the vector type. 4427 case NEON::BI__builtin_neon_vget_lane_i8: 4428 case NEON::BI__builtin_neon_vget_lane_i16: 4429 case NEON::BI__builtin_neon_vget_lane_i32: 4430 case NEON::BI__builtin_neon_vget_lane_i64: 4431 case NEON::BI__builtin_neon_vget_lane_f32: 4432 case NEON::BI__builtin_neon_vgetq_lane_i8: 4433 case NEON::BI__builtin_neon_vgetq_lane_i16: 4434 case NEON::BI__builtin_neon_vgetq_lane_i32: 4435 case NEON::BI__builtin_neon_vgetq_lane_i64: 4436 case NEON::BI__builtin_neon_vgetq_lane_f32: 4437 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 4438 "vget_lane"); 4439 case NEON::BI__builtin_neon_vset_lane_i8: 4440 case NEON::BI__builtin_neon_vset_lane_i16: 4441 case NEON::BI__builtin_neon_vset_lane_i32: 4442 case NEON::BI__builtin_neon_vset_lane_i64: 4443 case NEON::BI__builtin_neon_vset_lane_f32: 4444 case NEON::BI__builtin_neon_vsetq_lane_i8: 4445 case NEON::BI__builtin_neon_vsetq_lane_i16: 4446 case NEON::BI__builtin_neon_vsetq_lane_i32: 4447 case NEON::BI__builtin_neon_vsetq_lane_i64: 4448 case NEON::BI__builtin_neon_vsetq_lane_f32: 4449 Ops.push_back(EmitScalarExpr(E->getArg(2))); 4450 return Builder.CreateInsertElement(Ops[1], Ops[0], Ops[2], "vset_lane"); 4451 4452 // Non-polymorphic crypto instructions also not overloaded 4453 case NEON::BI__builtin_neon_vsha1h_u32: 4454 Ops.push_back(EmitScalarExpr(E->getArg(0))); 4455 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_sha1h), Ops, 4456 "vsha1h"); 4457 case NEON::BI__builtin_neon_vsha1cq_u32: 4458 Ops.push_back(EmitScalarExpr(E->getArg(2))); 4459 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_sha1c), Ops, 4460 "vsha1h"); 4461 case NEON::BI__builtin_neon_vsha1pq_u32: 4462 Ops.push_back(EmitScalarExpr(E->getArg(2))); 4463 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_sha1p), Ops, 4464 "vsha1h"); 4465 case NEON::BI__builtin_neon_vsha1mq_u32: 4466 Ops.push_back(EmitScalarExpr(E->getArg(2))); 4467 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_sha1m), Ops, 4468 "vsha1h"); 4469 } 4470 4471 // Get the last argument, which specifies the vector type. 4472 llvm::APSInt Result; 4473 const Expr *Arg = E->getArg(E->getNumArgs()-1); 4474 if (!Arg->isIntegerConstantExpr(Result, getContext())) 4475 return 0; 4476 4477 if (BuiltinID == ARM::BI__builtin_arm_vcvtr_f || 4478 BuiltinID == ARM::BI__builtin_arm_vcvtr_d) { 4479 // Determine the overloaded type of this builtin. 4480 llvm::Type *Ty; 4481 if (BuiltinID == ARM::BI__builtin_arm_vcvtr_f) 4482 Ty = FloatTy; 4483 else 4484 Ty = DoubleTy; 4485 4486 // Determine whether this is an unsigned conversion or not. 4487 bool usgn = Result.getZExtValue() == 1; 4488 unsigned Int = usgn ? Intrinsic::arm_vcvtru : Intrinsic::arm_vcvtr; 4489 4490 // Call the appropriate intrinsic. 4491 Function *F = CGM.getIntrinsic(Int, Ty); 4492 return Builder.CreateCall(F, Ops, "vcvtr"); 4493 } 4494 4495 // Determine the type of this overloaded NEON intrinsic. 4496 NeonTypeFlags Type(Result.getZExtValue()); 4497 bool usgn = Type.isUnsigned(); 4498 bool rightShift = false; 4499 4500 llvm::VectorType *VTy = GetNeonType(this, Type); 4501 llvm::Type *Ty = VTy; 4502 if (!Ty) 4503 return 0; 4504 4505 // Many NEON builtins have identical semantics and uses in ARM and 4506 // AArch64. Emit these in a single function. 4507 if (Value *Result = EmitCommonNeonBuiltinExpr(BuiltinID, E, Ops, Align)) 4508 return Result; 4509 4510 unsigned Int; 4511 switch (BuiltinID) { 4512 default: return 0; 4513 case NEON::BI__builtin_neon_vld1q_lane_v: 4514 // Handle 64-bit integer elements as a special case. Use shuffles of 4515 // one-element vectors to avoid poor code for i64 in the backend. 4516 if (VTy->getElementType()->isIntegerTy(64)) { 4517 // Extract the other lane. 4518 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 4519 int Lane = cast<ConstantInt>(Ops[2])->getZExtValue(); 4520 Value *SV = llvm::ConstantVector::get(ConstantInt::get(Int32Ty, 1-Lane)); 4521 Ops[1] = Builder.CreateShuffleVector(Ops[1], Ops[1], SV); 4522 // Load the value as a one-element vector. 4523 Ty = llvm::VectorType::get(VTy->getElementType(), 1); 4524 Function *F = CGM.getIntrinsic(Intrinsic::arm_neon_vld1, Ty); 4525 Value *Ld = Builder.CreateCall2(F, Ops[0], Align); 4526 // Combine them. 4527 SmallVector<Constant*, 2> Indices; 4528 Indices.push_back(ConstantInt::get(Int32Ty, 1-Lane)); 4529 Indices.push_back(ConstantInt::get(Int32Ty, Lane)); 4530 SV = llvm::ConstantVector::get(Indices); 4531 return Builder.CreateShuffleVector(Ops[1], Ld, SV, "vld1q_lane"); 4532 } 4533 // fall through 4534 case NEON::BI__builtin_neon_vld1_lane_v: { 4535 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 4536 Ty = llvm::PointerType::getUnqual(VTy->getElementType()); 4537 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 4538 LoadInst *Ld = Builder.CreateLoad(Ops[0]); 4539 Ld->setAlignment(cast<ConstantInt>(Align)->getZExtValue()); 4540 return Builder.CreateInsertElement(Ops[1], Ld, Ops[2], "vld1_lane"); 4541 } 4542 case NEON::BI__builtin_neon_vld2_dup_v: 4543 case NEON::BI__builtin_neon_vld3_dup_v: 4544 case NEON::BI__builtin_neon_vld4_dup_v: { 4545 // Handle 64-bit elements as a special-case. There is no "dup" needed. 4546 if (VTy->getElementType()->getPrimitiveSizeInBits() == 64) { 4547 switch (BuiltinID) { 4548 case NEON::BI__builtin_neon_vld2_dup_v: 4549 Int = Intrinsic::arm_neon_vld2; 4550 break; 4551 case NEON::BI__builtin_neon_vld3_dup_v: 4552 Int = Intrinsic::arm_neon_vld3; 4553 break; 4554 case NEON::BI__builtin_neon_vld4_dup_v: 4555 Int = Intrinsic::arm_neon_vld4; 4556 break; 4557 default: llvm_unreachable("unknown vld_dup intrinsic?"); 4558 } 4559 Function *F = CGM.getIntrinsic(Int, Ty); 4560 Ops[1] = Builder.CreateCall2(F, Ops[1], Align, "vld_dup"); 4561 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 4562 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 4563 return Builder.CreateStore(Ops[1], Ops[0]); 4564 } 4565 switch (BuiltinID) { 4566 case NEON::BI__builtin_neon_vld2_dup_v: 4567 Int = Intrinsic::arm_neon_vld2lane; 4568 break; 4569 case NEON::BI__builtin_neon_vld3_dup_v: 4570 Int = Intrinsic::arm_neon_vld3lane; 4571 break; 4572 case NEON::BI__builtin_neon_vld4_dup_v: 4573 Int = Intrinsic::arm_neon_vld4lane; 4574 break; 4575 default: llvm_unreachable("unknown vld_dup intrinsic?"); 4576 } 4577 Function *F = CGM.getIntrinsic(Int, Ty); 4578 llvm::StructType *STy = cast<llvm::StructType>(F->getReturnType()); 4579 4580 SmallVector<Value*, 6> Args; 4581 Args.push_back(Ops[1]); 4582 Args.append(STy->getNumElements(), UndefValue::get(Ty)); 4583 4584 llvm::Constant *CI = ConstantInt::get(Int32Ty, 0); 4585 Args.push_back(CI); 4586 Args.push_back(Align); 4587 4588 Ops[1] = Builder.CreateCall(F, Args, "vld_dup"); 4589 // splat lane 0 to all elts in each vector of the result. 4590 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 4591 Value *Val = Builder.CreateExtractValue(Ops[1], i); 4592 Value *Elt = Builder.CreateBitCast(Val, Ty); 4593 Elt = EmitNeonSplat(Elt, CI); 4594 Elt = Builder.CreateBitCast(Elt, Val->getType()); 4595 Ops[1] = Builder.CreateInsertValue(Ops[1], Elt, i); 4596 } 4597 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 4598 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 4599 return Builder.CreateStore(Ops[1], Ops[0]); 4600 } 4601 case NEON::BI__builtin_neon_vqrshrn_n_v: 4602 Int = 4603 usgn ? Intrinsic::arm_neon_vqrshiftnu : Intrinsic::arm_neon_vqrshiftns; 4604 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqrshrn_n", 4605 1, true); 4606 case NEON::BI__builtin_neon_vqrshrun_n_v: 4607 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vqrshiftnsu, Ty), 4608 Ops, "vqrshrun_n", 1, true); 4609 case NEON::BI__builtin_neon_vqshlu_n_v: 4610 case NEON::BI__builtin_neon_vqshluq_n_v: 4611 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vqshiftsu, Ty), 4612 Ops, "vqshlu", 1, false); 4613 case NEON::BI__builtin_neon_vqshrn_n_v: 4614 Int = usgn ? Intrinsic::arm_neon_vqshiftnu : Intrinsic::arm_neon_vqshiftns; 4615 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqshrn_n", 4616 1, true); 4617 case NEON::BI__builtin_neon_vqshrun_n_v: 4618 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vqshiftnsu, Ty), 4619 Ops, "vqshrun_n", 1, true); 4620 case NEON::BI__builtin_neon_vrecpe_v: 4621 case NEON::BI__builtin_neon_vrecpeq_v: 4622 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vrecpe, Ty), 4623 Ops, "vrecpe"); 4624 case NEON::BI__builtin_neon_vrshrn_n_v: 4625 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vrshiftn, Ty), 4626 Ops, "vrshrn_n", 1, true); 4627 case NEON::BI__builtin_neon_vrshr_n_v: 4628 case NEON::BI__builtin_neon_vrshrq_n_v: 4629 Int = usgn ? Intrinsic::arm_neon_vrshiftu : Intrinsic::arm_neon_vrshifts; 4630 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrshr_n", 1, true); 4631 case NEON::BI__builtin_neon_vrsra_n_v: 4632 case NEON::BI__builtin_neon_vrsraq_n_v: 4633 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 4634 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 4635 Ops[2] = EmitNeonShiftVector(Ops[2], Ty, true); 4636 Int = usgn ? Intrinsic::arm_neon_vrshiftu : Intrinsic::arm_neon_vrshifts; 4637 Ops[1] = Builder.CreateCall2(CGM.getIntrinsic(Int, Ty), Ops[1], Ops[2]); 4638 return Builder.CreateAdd(Ops[0], Ops[1], "vrsra_n"); 4639 case NEON::BI__builtin_neon_vsri_n_v: 4640 case NEON::BI__builtin_neon_vsriq_n_v: 4641 rightShift = true; 4642 case NEON::BI__builtin_neon_vsli_n_v: 4643 case NEON::BI__builtin_neon_vsliq_n_v: 4644 Ops[2] = EmitNeonShiftVector(Ops[2], Ty, rightShift); 4645 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vshiftins, Ty), 4646 Ops, "vsli_n"); 4647 case NEON::BI__builtin_neon_vsra_n_v: 4648 case NEON::BI__builtin_neon_vsraq_n_v: 4649 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 4650 Ops[1] = EmitNeonRShiftImm(Ops[1], Ops[2], Ty, usgn, "vsra_n"); 4651 return Builder.CreateAdd(Ops[0], Ops[1]); 4652 case NEON::BI__builtin_neon_vst1q_lane_v: 4653 // Handle 64-bit integer elements as a special case. Use a shuffle to get 4654 // a one-element vector and avoid poor code for i64 in the backend. 4655 if (VTy->getElementType()->isIntegerTy(64)) { 4656 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 4657 Value *SV = llvm::ConstantVector::get(cast<llvm::Constant>(Ops[2])); 4658 Ops[1] = Builder.CreateShuffleVector(Ops[1], Ops[1], SV); 4659 Ops[2] = Align; 4660 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::arm_neon_vst1, 4661 Ops[1]->getType()), Ops); 4662 } 4663 // fall through 4664 case NEON::BI__builtin_neon_vst1_lane_v: { 4665 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 4666 Ops[1] = Builder.CreateExtractElement(Ops[1], Ops[2]); 4667 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 4668 StoreInst *St = Builder.CreateStore(Ops[1], 4669 Builder.CreateBitCast(Ops[0], Ty)); 4670 St->setAlignment(cast<ConstantInt>(Align)->getZExtValue()); 4671 return St; 4672 } 4673 case NEON::BI__builtin_neon_vtbl1_v: 4674 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbl1), 4675 Ops, "vtbl1"); 4676 case NEON::BI__builtin_neon_vtbl2_v: 4677 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbl2), 4678 Ops, "vtbl2"); 4679 case NEON::BI__builtin_neon_vtbl3_v: 4680 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbl3), 4681 Ops, "vtbl3"); 4682 case NEON::BI__builtin_neon_vtbl4_v: 4683 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbl4), 4684 Ops, "vtbl4"); 4685 case NEON::BI__builtin_neon_vtbx1_v: 4686 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbx1), 4687 Ops, "vtbx1"); 4688 case NEON::BI__builtin_neon_vtbx2_v: 4689 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbx2), 4690 Ops, "vtbx2"); 4691 case NEON::BI__builtin_neon_vtbx3_v: 4692 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbx3), 4693 Ops, "vtbx3"); 4694 case NEON::BI__builtin_neon_vtbx4_v: 4695 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbx4), 4696 Ops, "vtbx4"); 4697 } 4698 } 4699 4700 llvm::Value *CodeGenFunction:: 4701 BuildVector(ArrayRef<llvm::Value*> Ops) { 4702 assert((Ops.size() & (Ops.size() - 1)) == 0 && 4703 "Not a power-of-two sized vector!"); 4704 bool AllConstants = true; 4705 for (unsigned i = 0, e = Ops.size(); i != e && AllConstants; ++i) 4706 AllConstants &= isa<Constant>(Ops[i]); 4707 4708 // If this is a constant vector, create a ConstantVector. 4709 if (AllConstants) { 4710 SmallVector<llvm::Constant*, 16> CstOps; 4711 for (unsigned i = 0, e = Ops.size(); i != e; ++i) 4712 CstOps.push_back(cast<Constant>(Ops[i])); 4713 return llvm::ConstantVector::get(CstOps); 4714 } 4715 4716 // Otherwise, insertelement the values to build the vector. 4717 Value *Result = 4718 llvm::UndefValue::get(llvm::VectorType::get(Ops[0]->getType(), Ops.size())); 4719 4720 for (unsigned i = 0, e = Ops.size(); i != e; ++i) 4721 Result = Builder.CreateInsertElement(Result, Ops[i], Builder.getInt32(i)); 4722 4723 return Result; 4724 } 4725 4726 Value *CodeGenFunction::EmitX86BuiltinExpr(unsigned BuiltinID, 4727 const CallExpr *E) { 4728 SmallVector<Value*, 4> Ops; 4729 4730 // Find out if any arguments are required to be integer constant expressions. 4731 unsigned ICEArguments = 0; 4732 ASTContext::GetBuiltinTypeError Error; 4733 getContext().GetBuiltinType(BuiltinID, Error, &ICEArguments); 4734 assert(Error == ASTContext::GE_None && "Should not codegen an error"); 4735 4736 for (unsigned i = 0, e = E->getNumArgs(); i != e; i++) { 4737 // If this is a normal argument, just emit it as a scalar. 4738 if ((ICEArguments & (1 << i)) == 0) { 4739 Ops.push_back(EmitScalarExpr(E->getArg(i))); 4740 continue; 4741 } 4742 4743 // If this is required to be a constant, constant fold it so that we know 4744 // that the generated intrinsic gets a ConstantInt. 4745 llvm::APSInt Result; 4746 bool IsConst = E->getArg(i)->isIntegerConstantExpr(Result, getContext()); 4747 assert(IsConst && "Constant arg isn't actually constant?"); (void)IsConst; 4748 Ops.push_back(llvm::ConstantInt::get(getLLVMContext(), Result)); 4749 } 4750 4751 switch (BuiltinID) { 4752 default: return 0; 4753 case X86::BI__builtin_ia32_vec_init_v8qi: 4754 case X86::BI__builtin_ia32_vec_init_v4hi: 4755 case X86::BI__builtin_ia32_vec_init_v2si: 4756 return Builder.CreateBitCast(BuildVector(Ops), 4757 llvm::Type::getX86_MMXTy(getLLVMContext())); 4758 case X86::BI__builtin_ia32_vec_ext_v2si: 4759 return Builder.CreateExtractElement(Ops[0], 4760 llvm::ConstantInt::get(Ops[1]->getType(), 0)); 4761 case X86::BI__builtin_ia32_ldmxcsr: { 4762 Value *Tmp = CreateMemTemp(E->getArg(0)->getType()); 4763 Builder.CreateStore(Ops[0], Tmp); 4764 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::x86_sse_ldmxcsr), 4765 Builder.CreateBitCast(Tmp, Int8PtrTy)); 4766 } 4767 case X86::BI__builtin_ia32_stmxcsr: { 4768 Value *Tmp = CreateMemTemp(E->getType()); 4769 Builder.CreateCall(CGM.getIntrinsic(Intrinsic::x86_sse_stmxcsr), 4770 Builder.CreateBitCast(Tmp, Int8PtrTy)); 4771 return Builder.CreateLoad(Tmp, "stmxcsr"); 4772 } 4773 case X86::BI__builtin_ia32_storehps: 4774 case X86::BI__builtin_ia32_storelps: { 4775 llvm::Type *PtrTy = llvm::PointerType::getUnqual(Int64Ty); 4776 llvm::Type *VecTy = llvm::VectorType::get(Int64Ty, 2); 4777 4778 // cast val v2i64 4779 Ops[1] = Builder.CreateBitCast(Ops[1], VecTy, "cast"); 4780 4781 // extract (0, 1) 4782 unsigned Index = BuiltinID == X86::BI__builtin_ia32_storelps ? 0 : 1; 4783 llvm::Value *Idx = llvm::ConstantInt::get(Int32Ty, Index); 4784 Ops[1] = Builder.CreateExtractElement(Ops[1], Idx, "extract"); 4785 4786 // cast pointer to i64 & store 4787 Ops[0] = Builder.CreateBitCast(Ops[0], PtrTy); 4788 return Builder.CreateStore(Ops[1], Ops[0]); 4789 } 4790 case X86::BI__builtin_ia32_palignr: { 4791 unsigned shiftVal = cast<llvm::ConstantInt>(Ops[2])->getZExtValue(); 4792 4793 // If palignr is shifting the pair of input vectors less than 9 bytes, 4794 // emit a shuffle instruction. 4795 if (shiftVal <= 8) { 4796 SmallVector<llvm::Constant*, 8> Indices; 4797 for (unsigned i = 0; i != 8; ++i) 4798 Indices.push_back(llvm::ConstantInt::get(Int32Ty, shiftVal + i)); 4799 4800 Value* SV = llvm::ConstantVector::get(Indices); 4801 return Builder.CreateShuffleVector(Ops[1], Ops[0], SV, "palignr"); 4802 } 4803 4804 // If palignr is shifting the pair of input vectors more than 8 but less 4805 // than 16 bytes, emit a logical right shift of the destination. 4806 if (shiftVal < 16) { 4807 // MMX has these as 1 x i64 vectors for some odd optimization reasons. 4808 llvm::Type *VecTy = llvm::VectorType::get(Int64Ty, 1); 4809 4810 Ops[0] = Builder.CreateBitCast(Ops[0], VecTy, "cast"); 4811 Ops[1] = llvm::ConstantInt::get(VecTy, (shiftVal-8) * 8); 4812 4813 // create i32 constant 4814 llvm::Function *F = CGM.getIntrinsic(Intrinsic::x86_mmx_psrl_q); 4815 return Builder.CreateCall(F, makeArrayRef(&Ops[0], 2), "palignr"); 4816 } 4817 4818 // If palignr is shifting the pair of vectors more than 16 bytes, emit zero. 4819 return llvm::Constant::getNullValue(ConvertType(E->getType())); 4820 } 4821 case X86::BI__builtin_ia32_palignr128: { 4822 unsigned shiftVal = cast<llvm::ConstantInt>(Ops[2])->getZExtValue(); 4823 4824 // If palignr is shifting the pair of input vectors less than 17 bytes, 4825 // emit a shuffle instruction. 4826 if (shiftVal <= 16) { 4827 SmallVector<llvm::Constant*, 16> Indices; 4828 for (unsigned i = 0; i != 16; ++i) 4829 Indices.push_back(llvm::ConstantInt::get(Int32Ty, shiftVal + i)); 4830 4831 Value* SV = llvm::ConstantVector::get(Indices); 4832 return Builder.CreateShuffleVector(Ops[1], Ops[0], SV, "palignr"); 4833 } 4834 4835 // If palignr is shifting the pair of input vectors more than 16 but less 4836 // than 32 bytes, emit a logical right shift of the destination. 4837 if (shiftVal < 32) { 4838 llvm::Type *VecTy = llvm::VectorType::get(Int64Ty, 2); 4839 4840 Ops[0] = Builder.CreateBitCast(Ops[0], VecTy, "cast"); 4841 Ops[1] = llvm::ConstantInt::get(Int32Ty, (shiftVal-16) * 8); 4842 4843 // create i32 constant 4844 llvm::Function *F = CGM.getIntrinsic(Intrinsic::x86_sse2_psrl_dq); 4845 return Builder.CreateCall(F, makeArrayRef(&Ops[0], 2), "palignr"); 4846 } 4847 4848 // If palignr is shifting the pair of vectors more than 32 bytes, emit zero. 4849 return llvm::Constant::getNullValue(ConvertType(E->getType())); 4850 } 4851 case X86::BI__builtin_ia32_palignr256: { 4852 unsigned shiftVal = cast<llvm::ConstantInt>(Ops[2])->getZExtValue(); 4853 4854 // If palignr is shifting the pair of input vectors less than 17 bytes, 4855 // emit a shuffle instruction. 4856 if (shiftVal <= 16) { 4857 SmallVector<llvm::Constant*, 32> Indices; 4858 // 256-bit palignr operates on 128-bit lanes so we need to handle that 4859 for (unsigned l = 0; l != 2; ++l) { 4860 unsigned LaneStart = l * 16; 4861 unsigned LaneEnd = (l+1) * 16; 4862 for (unsigned i = 0; i != 16; ++i) { 4863 unsigned Idx = shiftVal + i + LaneStart; 4864 if (Idx >= LaneEnd) Idx += 16; // end of lane, switch operand 4865 Indices.push_back(llvm::ConstantInt::get(Int32Ty, Idx)); 4866 } 4867 } 4868 4869 Value* SV = llvm::ConstantVector::get(Indices); 4870 return Builder.CreateShuffleVector(Ops[1], Ops[0], SV, "palignr"); 4871 } 4872 4873 // If palignr is shifting the pair of input vectors more than 16 but less 4874 // than 32 bytes, emit a logical right shift of the destination. 4875 if (shiftVal < 32) { 4876 llvm::Type *VecTy = llvm::VectorType::get(Int64Ty, 4); 4877 4878 Ops[0] = Builder.CreateBitCast(Ops[0], VecTy, "cast"); 4879 Ops[1] = llvm::ConstantInt::get(Int32Ty, (shiftVal-16) * 8); 4880 4881 // create i32 constant 4882 llvm::Function *F = CGM.getIntrinsic(Intrinsic::x86_avx2_psrl_dq); 4883 return Builder.CreateCall(F, makeArrayRef(&Ops[0], 2), "palignr"); 4884 } 4885 4886 // If palignr is shifting the pair of vectors more than 32 bytes, emit zero. 4887 return llvm::Constant::getNullValue(ConvertType(E->getType())); 4888 } 4889 case X86::BI__builtin_ia32_movntps: 4890 case X86::BI__builtin_ia32_movntps256: 4891 case X86::BI__builtin_ia32_movntpd: 4892 case X86::BI__builtin_ia32_movntpd256: 4893 case X86::BI__builtin_ia32_movntdq: 4894 case X86::BI__builtin_ia32_movntdq256: 4895 case X86::BI__builtin_ia32_movnti: 4896 case X86::BI__builtin_ia32_movnti64: { 4897 llvm::MDNode *Node = llvm::MDNode::get(getLLVMContext(), 4898 Builder.getInt32(1)); 4899 4900 // Convert the type of the pointer to a pointer to the stored type. 4901 Value *BC = Builder.CreateBitCast(Ops[0], 4902 llvm::PointerType::getUnqual(Ops[1]->getType()), 4903 "cast"); 4904 StoreInst *SI = Builder.CreateStore(Ops[1], BC); 4905 SI->setMetadata(CGM.getModule().getMDKindID("nontemporal"), Node); 4906 4907 // If the operand is an integer, we can't assume alignment. Otherwise, 4908 // assume natural alignment. 4909 QualType ArgTy = E->getArg(1)->getType(); 4910 unsigned Align; 4911 if (ArgTy->isIntegerType()) 4912 Align = 1; 4913 else 4914 Align = getContext().getTypeSizeInChars(ArgTy).getQuantity(); 4915 SI->setAlignment(Align); 4916 return SI; 4917 } 4918 // 3DNow! 4919 case X86::BI__builtin_ia32_pswapdsf: 4920 case X86::BI__builtin_ia32_pswapdsi: { 4921 const char *name = 0; 4922 Intrinsic::ID ID = Intrinsic::not_intrinsic; 4923 switch(BuiltinID) { 4924 default: llvm_unreachable("Unsupported intrinsic!"); 4925 case X86::BI__builtin_ia32_pswapdsf: 4926 case X86::BI__builtin_ia32_pswapdsi: 4927 name = "pswapd"; 4928 ID = Intrinsic::x86_3dnowa_pswapd; 4929 break; 4930 } 4931 llvm::Type *MMXTy = llvm::Type::getX86_MMXTy(getLLVMContext()); 4932 Ops[0] = Builder.CreateBitCast(Ops[0], MMXTy, "cast"); 4933 llvm::Function *F = CGM.getIntrinsic(ID); 4934 return Builder.CreateCall(F, Ops, name); 4935 } 4936 case X86::BI__builtin_ia32_rdrand16_step: 4937 case X86::BI__builtin_ia32_rdrand32_step: 4938 case X86::BI__builtin_ia32_rdrand64_step: 4939 case X86::BI__builtin_ia32_rdseed16_step: 4940 case X86::BI__builtin_ia32_rdseed32_step: 4941 case X86::BI__builtin_ia32_rdseed64_step: { 4942 Intrinsic::ID ID; 4943 switch (BuiltinID) { 4944 default: llvm_unreachable("Unsupported intrinsic!"); 4945 case X86::BI__builtin_ia32_rdrand16_step: 4946 ID = Intrinsic::x86_rdrand_16; 4947 break; 4948 case X86::BI__builtin_ia32_rdrand32_step: 4949 ID = Intrinsic::x86_rdrand_32; 4950 break; 4951 case X86::BI__builtin_ia32_rdrand64_step: 4952 ID = Intrinsic::x86_rdrand_64; 4953 break; 4954 case X86::BI__builtin_ia32_rdseed16_step: 4955 ID = Intrinsic::x86_rdseed_16; 4956 break; 4957 case X86::BI__builtin_ia32_rdseed32_step: 4958 ID = Intrinsic::x86_rdseed_32; 4959 break; 4960 case X86::BI__builtin_ia32_rdseed64_step: 4961 ID = Intrinsic::x86_rdseed_64; 4962 break; 4963 } 4964 4965 Value *Call = Builder.CreateCall(CGM.getIntrinsic(ID)); 4966 Builder.CreateStore(Builder.CreateExtractValue(Call, 0), Ops[0]); 4967 return Builder.CreateExtractValue(Call, 1); 4968 } 4969 // AVX2 broadcast 4970 case X86::BI__builtin_ia32_vbroadcastsi256: { 4971 Value *VecTmp = CreateMemTemp(E->getArg(0)->getType()); 4972 Builder.CreateStore(Ops[0], VecTmp); 4973 Value *F = CGM.getIntrinsic(Intrinsic::x86_avx2_vbroadcasti128); 4974 return Builder.CreateCall(F, Builder.CreateBitCast(VecTmp, Int8PtrTy)); 4975 } 4976 } 4977 } 4978 4979 4980 Value *CodeGenFunction::EmitPPCBuiltinExpr(unsigned BuiltinID, 4981 const CallExpr *E) { 4982 SmallVector<Value*, 4> Ops; 4983 4984 for (unsigned i = 0, e = E->getNumArgs(); i != e; i++) 4985 Ops.push_back(EmitScalarExpr(E->getArg(i))); 4986 4987 Intrinsic::ID ID = Intrinsic::not_intrinsic; 4988 4989 switch (BuiltinID) { 4990 default: return 0; 4991 4992 // vec_ld, vec_lvsl, vec_lvsr 4993 case PPC::BI__builtin_altivec_lvx: 4994 case PPC::BI__builtin_altivec_lvxl: 4995 case PPC::BI__builtin_altivec_lvebx: 4996 case PPC::BI__builtin_altivec_lvehx: 4997 case PPC::BI__builtin_altivec_lvewx: 4998 case PPC::BI__builtin_altivec_lvsl: 4999 case PPC::BI__builtin_altivec_lvsr: 5000 { 5001 Ops[1] = Builder.CreateBitCast(Ops[1], Int8PtrTy); 5002 5003 Ops[0] = Builder.CreateGEP(Ops[1], Ops[0]); 5004 Ops.pop_back(); 5005 5006 switch (BuiltinID) { 5007 default: llvm_unreachable("Unsupported ld/lvsl/lvsr intrinsic!"); 5008 case PPC::BI__builtin_altivec_lvx: 5009 ID = Intrinsic::ppc_altivec_lvx; 5010 break; 5011 case PPC::BI__builtin_altivec_lvxl: 5012 ID = Intrinsic::ppc_altivec_lvxl; 5013 break; 5014 case PPC::BI__builtin_altivec_lvebx: 5015 ID = Intrinsic::ppc_altivec_lvebx; 5016 break; 5017 case PPC::BI__builtin_altivec_lvehx: 5018 ID = Intrinsic::ppc_altivec_lvehx; 5019 break; 5020 case PPC::BI__builtin_altivec_lvewx: 5021 ID = Intrinsic::ppc_altivec_lvewx; 5022 break; 5023 case PPC::BI__builtin_altivec_lvsl: 5024 ID = Intrinsic::ppc_altivec_lvsl; 5025 break; 5026 case PPC::BI__builtin_altivec_lvsr: 5027 ID = Intrinsic::ppc_altivec_lvsr; 5028 break; 5029 } 5030 llvm::Function *F = CGM.getIntrinsic(ID); 5031 return Builder.CreateCall(F, Ops, ""); 5032 } 5033 5034 // vec_st 5035 case PPC::BI__builtin_altivec_stvx: 5036 case PPC::BI__builtin_altivec_stvxl: 5037 case PPC::BI__builtin_altivec_stvebx: 5038 case PPC::BI__builtin_altivec_stvehx: 5039 case PPC::BI__builtin_altivec_stvewx: 5040 { 5041 Ops[2] = Builder.CreateBitCast(Ops[2], Int8PtrTy); 5042 Ops[1] = Builder.CreateGEP(Ops[2], Ops[1]); 5043 Ops.pop_back(); 5044 5045 switch (BuiltinID) { 5046 default: llvm_unreachable("Unsupported st intrinsic!"); 5047 case PPC::BI__builtin_altivec_stvx: 5048 ID = Intrinsic::ppc_altivec_stvx; 5049 break; 5050 case PPC::BI__builtin_altivec_stvxl: 5051 ID = Intrinsic::ppc_altivec_stvxl; 5052 break; 5053 case PPC::BI__builtin_altivec_stvebx: 5054 ID = Intrinsic::ppc_altivec_stvebx; 5055 break; 5056 case PPC::BI__builtin_altivec_stvehx: 5057 ID = Intrinsic::ppc_altivec_stvehx; 5058 break; 5059 case PPC::BI__builtin_altivec_stvewx: 5060 ID = Intrinsic::ppc_altivec_stvewx; 5061 break; 5062 } 5063 llvm::Function *F = CGM.getIntrinsic(ID); 5064 return Builder.CreateCall(F, Ops, ""); 5065 } 5066 } 5067 } 5068