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 "TargetInfo.h" 15 #include "CodeGenFunction.h" 16 #include "CodeGenModule.h" 17 #include "CGObjCRuntime.h" 18 #include "clang/Basic/TargetInfo.h" 19 #include "clang/AST/ASTContext.h" 20 #include "clang/AST/Decl.h" 21 #include "clang/Basic/TargetBuiltins.h" 22 #include "llvm/Intrinsics.h" 23 #include "llvm/Target/TargetData.h" 24 25 using namespace clang; 26 using namespace CodeGen; 27 using namespace llvm; 28 29 /// getBuiltinLibFunction - Given a builtin id for a function like 30 /// "__builtin_fabsf", return a Function* for "fabsf". 31 llvm::Value *CodeGenModule::getBuiltinLibFunction(const FunctionDecl *FD, 32 unsigned BuiltinID) { 33 assert(Context.BuiltinInfo.isLibFunction(BuiltinID)); 34 35 // Get the name, skip over the __builtin_ prefix (if necessary). 36 StringRef Name; 37 GlobalDecl D(FD); 38 39 // If the builtin has been declared explicitly with an assembler label, 40 // use the mangled name. This differs from the plain label on platforms 41 // that prefix labels. 42 if (FD->hasAttr<AsmLabelAttr>()) 43 Name = getMangledName(D); 44 else 45 Name = Context.BuiltinInfo.GetName(BuiltinID) + 10; 46 47 llvm::FunctionType *Ty = 48 cast<llvm::FunctionType>(getTypes().ConvertType(FD->getType())); 49 50 return GetOrCreateLLVMFunction(Name, Ty, D, /*ForVTable=*/false); 51 } 52 53 /// Emit the conversions required to turn the given value into an 54 /// integer of the given size. 55 static Value *EmitToInt(CodeGenFunction &CGF, llvm::Value *V, 56 QualType T, llvm::IntegerType *IntType) { 57 V = CGF.EmitToMemory(V, T); 58 59 if (V->getType()->isPointerTy()) 60 return CGF.Builder.CreatePtrToInt(V, IntType); 61 62 assert(V->getType() == IntType); 63 return V; 64 } 65 66 static Value *EmitFromInt(CodeGenFunction &CGF, llvm::Value *V, 67 QualType T, llvm::Type *ResultType) { 68 V = CGF.EmitFromMemory(V, T); 69 70 if (ResultType->isPointerTy()) 71 return CGF.Builder.CreateIntToPtr(V, ResultType); 72 73 assert(V->getType() == ResultType); 74 return V; 75 } 76 77 /// Utility to insert an atomic instruction based on Instrinsic::ID 78 /// and the expression node. 79 static RValue EmitBinaryAtomic(CodeGenFunction &CGF, 80 llvm::AtomicRMWInst::BinOp Kind, 81 const CallExpr *E) { 82 QualType T = E->getType(); 83 assert(E->getArg(0)->getType()->isPointerType()); 84 assert(CGF.getContext().hasSameUnqualifiedType(T, 85 E->getArg(0)->getType()->getPointeeType())); 86 assert(CGF.getContext().hasSameUnqualifiedType(T, E->getArg(1)->getType())); 87 88 llvm::Value *DestPtr = CGF.EmitScalarExpr(E->getArg(0)); 89 unsigned AddrSpace = 90 cast<llvm::PointerType>(DestPtr->getType())->getAddressSpace(); 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 = 125 cast<llvm::PointerType>(DestPtr->getType())->getAddressSpace(); 126 127 llvm::IntegerType *IntType = 128 llvm::IntegerType::get(CGF.getLLVMContext(), 129 CGF.getContext().getTypeSize(T)); 130 llvm::Type *IntPtrType = IntType->getPointerTo(AddrSpace); 131 132 llvm::Value *Args[2]; 133 Args[1] = CGF.EmitScalarExpr(E->getArg(1)); 134 llvm::Type *ValueType = Args[1]->getType(); 135 Args[1] = EmitToInt(CGF, Args[1], T, IntType); 136 Args[0] = CGF.Builder.CreateBitCast(DestPtr, IntPtrType); 137 138 llvm::Value *Result = 139 CGF.Builder.CreateAtomicRMW(Kind, Args[0], Args[1], 140 llvm::SequentiallyConsistent); 141 Result = CGF.Builder.CreateBinOp(Op, Result, Args[1]); 142 Result = EmitFromInt(CGF, Result, T, ValueType); 143 return RValue::get(Result); 144 } 145 146 /// EmitFAbs - Emit a call to fabs/fabsf/fabsl, depending on the type of ValTy, 147 /// which must be a scalar floating point type. 148 static Value *EmitFAbs(CodeGenFunction &CGF, Value *V, QualType ValTy) { 149 const BuiltinType *ValTyP = ValTy->getAs<BuiltinType>(); 150 assert(ValTyP && "isn't scalar fp type!"); 151 152 StringRef FnName; 153 switch (ValTyP->getKind()) { 154 default: llvm_unreachable("Isn't a scalar fp type!"); 155 case BuiltinType::Float: FnName = "fabsf"; break; 156 case BuiltinType::Double: FnName = "fabs"; break; 157 case BuiltinType::LongDouble: FnName = "fabsl"; break; 158 } 159 160 // The prototype is something that takes and returns whatever V's type is. 161 llvm::FunctionType *FT = llvm::FunctionType::get(V->getType(), V->getType(), 162 false); 163 llvm::Value *Fn = CGF.CGM.CreateRuntimeFunction(FT, FnName); 164 165 return CGF.Builder.CreateCall(Fn, V, "abs"); 166 } 167 168 static RValue emitLibraryCall(CodeGenFunction &CGF, const FunctionDecl *Fn, 169 const CallExpr *E, llvm::Value *calleeValue) { 170 return CGF.EmitCall(E->getCallee()->getType(), calleeValue, 171 ReturnValueSlot(), E->arg_begin(), E->arg_end(), Fn); 172 } 173 174 RValue CodeGenFunction::EmitBuiltinExpr(const FunctionDecl *FD, 175 unsigned BuiltinID, const CallExpr *E) { 176 // See if we can constant fold this builtin. If so, don't emit it at all. 177 Expr::EvalResult Result; 178 if (E->EvaluateAsRValue(Result, CGM.getContext()) && 179 !Result.hasSideEffects()) { 180 if (Result.Val.isInt()) 181 return RValue::get(llvm::ConstantInt::get(getLLVMContext(), 182 Result.Val.getInt())); 183 if (Result.Val.isFloat()) 184 return RValue::get(llvm::ConstantFP::get(getLLVMContext(), 185 Result.Val.getFloat())); 186 } 187 188 switch (BuiltinID) { 189 default: break; // Handle intrinsics and libm functions below. 190 case Builtin::BI__builtin___CFStringMakeConstantString: 191 case Builtin::BI__builtin___NSStringMakeConstantString: 192 return RValue::get(CGM.EmitConstantExpr(E, E->getType(), 0)); 193 case Builtin::BI__builtin_stdarg_start: 194 case Builtin::BI__builtin_va_start: 195 case Builtin::BI__builtin_va_end: { 196 Value *ArgValue = EmitVAListRef(E->getArg(0)); 197 llvm::Type *DestType = Int8PtrTy; 198 if (ArgValue->getType() != DestType) 199 ArgValue = Builder.CreateBitCast(ArgValue, DestType, 200 ArgValue->getName().data()); 201 202 Intrinsic::ID inst = (BuiltinID == Builtin::BI__builtin_va_end) ? 203 Intrinsic::vaend : Intrinsic::vastart; 204 return RValue::get(Builder.CreateCall(CGM.getIntrinsic(inst), ArgValue)); 205 } 206 case Builtin::BI__builtin_va_copy: { 207 Value *DstPtr = EmitVAListRef(E->getArg(0)); 208 Value *SrcPtr = EmitVAListRef(E->getArg(1)); 209 210 llvm::Type *Type = Int8PtrTy; 211 212 DstPtr = Builder.CreateBitCast(DstPtr, Type); 213 SrcPtr = Builder.CreateBitCast(SrcPtr, Type); 214 return RValue::get(Builder.CreateCall2(CGM.getIntrinsic(Intrinsic::vacopy), 215 DstPtr, SrcPtr)); 216 } 217 case Builtin::BI__builtin_abs: 218 case Builtin::BI__builtin_labs: 219 case Builtin::BI__builtin_llabs: { 220 Value *ArgValue = EmitScalarExpr(E->getArg(0)); 221 222 Value *NegOp = Builder.CreateNeg(ArgValue, "neg"); 223 Value *CmpResult = 224 Builder.CreateICmpSGE(ArgValue, 225 llvm::Constant::getNullValue(ArgValue->getType()), 226 "abscond"); 227 Value *Result = 228 Builder.CreateSelect(CmpResult, ArgValue, NegOp, "abs"); 229 230 return RValue::get(Result); 231 } 232 233 case Builtin::BI__builtin_conj: 234 case Builtin::BI__builtin_conjf: 235 case Builtin::BI__builtin_conjl: { 236 ComplexPairTy ComplexVal = EmitComplexExpr(E->getArg(0)); 237 Value *Real = ComplexVal.first; 238 Value *Imag = ComplexVal.second; 239 Value *Zero = 240 Imag->getType()->isFPOrFPVectorTy() 241 ? llvm::ConstantFP::getZeroValueForNegation(Imag->getType()) 242 : llvm::Constant::getNullValue(Imag->getType()); 243 244 Imag = Builder.CreateFSub(Zero, Imag, "sub"); 245 return RValue::getComplex(std::make_pair(Real, Imag)); 246 } 247 case Builtin::BI__builtin_creal: 248 case Builtin::BI__builtin_crealf: 249 case Builtin::BI__builtin_creall: { 250 ComplexPairTy ComplexVal = EmitComplexExpr(E->getArg(0)); 251 return RValue::get(ComplexVal.first); 252 } 253 254 case Builtin::BI__builtin_cimag: 255 case Builtin::BI__builtin_cimagf: 256 case Builtin::BI__builtin_cimagl: { 257 ComplexPairTy ComplexVal = EmitComplexExpr(E->getArg(0)); 258 return RValue::get(ComplexVal.second); 259 } 260 261 case Builtin::BI__builtin_ctzs: 262 case Builtin::BI__builtin_ctz: 263 case Builtin::BI__builtin_ctzl: 264 case Builtin::BI__builtin_ctzll: { 265 Value *ArgValue = EmitScalarExpr(E->getArg(0)); 266 267 llvm::Type *ArgType = ArgValue->getType(); 268 Value *F = CGM.getIntrinsic(Intrinsic::cttz, ArgType); 269 270 llvm::Type *ResultType = ConvertType(E->getType()); 271 Value *ZeroUndef = Builder.getInt1(Target.isCLZForZeroUndef()); 272 Value *Result = Builder.CreateCall2(F, ArgValue, ZeroUndef); 273 if (Result->getType() != ResultType) 274 Result = Builder.CreateIntCast(Result, ResultType, /*isSigned*/true, 275 "cast"); 276 return RValue::get(Result); 277 } 278 case Builtin::BI__builtin_clzs: 279 case Builtin::BI__builtin_clz: 280 case Builtin::BI__builtin_clzl: 281 case Builtin::BI__builtin_clzll: { 282 Value *ArgValue = EmitScalarExpr(E->getArg(0)); 283 284 llvm::Type *ArgType = ArgValue->getType(); 285 Value *F = CGM.getIntrinsic(Intrinsic::ctlz, ArgType); 286 287 llvm::Type *ResultType = ConvertType(E->getType()); 288 Value *ZeroUndef = Builder.getInt1(Target.isCLZForZeroUndef()); 289 Value *Result = Builder.CreateCall2(F, ArgValue, ZeroUndef); 290 if (Result->getType() != ResultType) 291 Result = Builder.CreateIntCast(Result, ResultType, /*isSigned*/true, 292 "cast"); 293 return RValue::get(Result); 294 } 295 case Builtin::BI__builtin_ffs: 296 case Builtin::BI__builtin_ffsl: 297 case Builtin::BI__builtin_ffsll: { 298 // ffs(x) -> x ? cttz(x) + 1 : 0 299 Value *ArgValue = EmitScalarExpr(E->getArg(0)); 300 301 llvm::Type *ArgType = ArgValue->getType(); 302 Value *F = CGM.getIntrinsic(Intrinsic::cttz, ArgType); 303 304 llvm::Type *ResultType = ConvertType(E->getType()); 305 Value *Tmp = Builder.CreateAdd(Builder.CreateCall2(F, ArgValue, 306 Builder.getTrue()), 307 llvm::ConstantInt::get(ArgType, 1)); 308 Value *Zero = llvm::Constant::getNullValue(ArgType); 309 Value *IsZero = Builder.CreateICmpEQ(ArgValue, Zero, "iszero"); 310 Value *Result = Builder.CreateSelect(IsZero, Zero, Tmp, "ffs"); 311 if (Result->getType() != ResultType) 312 Result = Builder.CreateIntCast(Result, ResultType, /*isSigned*/true, 313 "cast"); 314 return RValue::get(Result); 315 } 316 case Builtin::BI__builtin_parity: 317 case Builtin::BI__builtin_parityl: 318 case Builtin::BI__builtin_parityll: { 319 // parity(x) -> ctpop(x) & 1 320 Value *ArgValue = EmitScalarExpr(E->getArg(0)); 321 322 llvm::Type *ArgType = ArgValue->getType(); 323 Value *F = CGM.getIntrinsic(Intrinsic::ctpop, ArgType); 324 325 llvm::Type *ResultType = ConvertType(E->getType()); 326 Value *Tmp = Builder.CreateCall(F, ArgValue); 327 Value *Result = Builder.CreateAnd(Tmp, llvm::ConstantInt::get(ArgType, 1)); 328 if (Result->getType() != ResultType) 329 Result = Builder.CreateIntCast(Result, ResultType, /*isSigned*/true, 330 "cast"); 331 return RValue::get(Result); 332 } 333 case Builtin::BI__builtin_popcount: 334 case Builtin::BI__builtin_popcountl: 335 case Builtin::BI__builtin_popcountll: { 336 Value *ArgValue = EmitScalarExpr(E->getArg(0)); 337 338 llvm::Type *ArgType = ArgValue->getType(); 339 Value *F = CGM.getIntrinsic(Intrinsic::ctpop, ArgType); 340 341 llvm::Type *ResultType = ConvertType(E->getType()); 342 Value *Result = Builder.CreateCall(F, ArgValue); 343 if (Result->getType() != ResultType) 344 Result = Builder.CreateIntCast(Result, ResultType, /*isSigned*/true, 345 "cast"); 346 return RValue::get(Result); 347 } 348 case Builtin::BI__builtin_expect: { 349 Value *ArgValue = EmitScalarExpr(E->getArg(0)); 350 llvm::Type *ArgType = ArgValue->getType(); 351 352 Value *FnExpect = CGM.getIntrinsic(Intrinsic::expect, ArgType); 353 Value *ExpectedValue = EmitScalarExpr(E->getArg(1)); 354 355 Value *Result = Builder.CreateCall2(FnExpect, ArgValue, ExpectedValue, 356 "expval"); 357 return RValue::get(Result); 358 } 359 case Builtin::BI__builtin_bswap32: 360 case Builtin::BI__builtin_bswap64: { 361 Value *ArgValue = EmitScalarExpr(E->getArg(0)); 362 llvm::Type *ArgType = ArgValue->getType(); 363 Value *F = CGM.getIntrinsic(Intrinsic::bswap, ArgType); 364 return RValue::get(Builder.CreateCall(F, ArgValue)); 365 } 366 case Builtin::BI__builtin_object_size: { 367 // We rely on constant folding to deal with expressions with side effects. 368 assert(!E->getArg(0)->HasSideEffects(getContext()) && 369 "should have been constant folded"); 370 371 // We pass this builtin onto the optimizer so that it can 372 // figure out the object size in more complex cases. 373 llvm::Type *ResType = ConvertType(E->getType()); 374 375 // LLVM only supports 0 and 2, make sure that we pass along that 376 // as a boolean. 377 Value *Ty = EmitScalarExpr(E->getArg(1)); 378 ConstantInt *CI = dyn_cast<ConstantInt>(Ty); 379 assert(CI); 380 uint64_t val = CI->getZExtValue(); 381 CI = ConstantInt::get(Builder.getInt1Ty(), (val & 0x2) >> 1); 382 383 Value *F = CGM.getIntrinsic(Intrinsic::objectsize, ResType); 384 return RValue::get(Builder.CreateCall2(F, EmitScalarExpr(E->getArg(0)),CI)); 385 } 386 case Builtin::BI__builtin_prefetch: { 387 Value *Locality, *RW, *Address = EmitScalarExpr(E->getArg(0)); 388 // FIXME: Technically these constants should of type 'int', yes? 389 RW = (E->getNumArgs() > 1) ? EmitScalarExpr(E->getArg(1)) : 390 llvm::ConstantInt::get(Int32Ty, 0); 391 Locality = (E->getNumArgs() > 2) ? EmitScalarExpr(E->getArg(2)) : 392 llvm::ConstantInt::get(Int32Ty, 3); 393 Value *Data = llvm::ConstantInt::get(Int32Ty, 1); 394 Value *F = CGM.getIntrinsic(Intrinsic::prefetch); 395 return RValue::get(Builder.CreateCall4(F, Address, RW, Locality, Data)); 396 } 397 case Builtin::BI__builtin_readcyclecounter: { 398 Value *F = CGM.getIntrinsic(Intrinsic::readcyclecounter); 399 return RValue::get(Builder.CreateCall(F)); 400 } 401 case Builtin::BI__builtin_trap: { 402 Value *F = CGM.getIntrinsic(Intrinsic::trap); 403 return RValue::get(Builder.CreateCall(F)); 404 } 405 case Builtin::BI__builtin_unreachable: { 406 if (CatchUndefined) 407 EmitBranch(getTrapBB()); 408 else 409 Builder.CreateUnreachable(); 410 411 // We do need to preserve an insertion point. 412 EmitBlock(createBasicBlock("unreachable.cont")); 413 414 return RValue::get(0); 415 } 416 417 case Builtin::BI__builtin_powi: 418 case Builtin::BI__builtin_powif: 419 case Builtin::BI__builtin_powil: { 420 Value *Base = EmitScalarExpr(E->getArg(0)); 421 Value *Exponent = EmitScalarExpr(E->getArg(1)); 422 llvm::Type *ArgType = Base->getType(); 423 Value *F = CGM.getIntrinsic(Intrinsic::powi, ArgType); 424 return RValue::get(Builder.CreateCall2(F, Base, Exponent)); 425 } 426 427 case Builtin::BI__builtin_isgreater: 428 case Builtin::BI__builtin_isgreaterequal: 429 case Builtin::BI__builtin_isless: 430 case Builtin::BI__builtin_islessequal: 431 case Builtin::BI__builtin_islessgreater: 432 case Builtin::BI__builtin_isunordered: { 433 // Ordered comparisons: we know the arguments to these are matching scalar 434 // floating point values. 435 Value *LHS = EmitScalarExpr(E->getArg(0)); 436 Value *RHS = EmitScalarExpr(E->getArg(1)); 437 438 switch (BuiltinID) { 439 default: llvm_unreachable("Unknown ordered comparison"); 440 case Builtin::BI__builtin_isgreater: 441 LHS = Builder.CreateFCmpOGT(LHS, RHS, "cmp"); 442 break; 443 case Builtin::BI__builtin_isgreaterequal: 444 LHS = Builder.CreateFCmpOGE(LHS, RHS, "cmp"); 445 break; 446 case Builtin::BI__builtin_isless: 447 LHS = Builder.CreateFCmpOLT(LHS, RHS, "cmp"); 448 break; 449 case Builtin::BI__builtin_islessequal: 450 LHS = Builder.CreateFCmpOLE(LHS, RHS, "cmp"); 451 break; 452 case Builtin::BI__builtin_islessgreater: 453 LHS = Builder.CreateFCmpONE(LHS, RHS, "cmp"); 454 break; 455 case Builtin::BI__builtin_isunordered: 456 LHS = Builder.CreateFCmpUNO(LHS, RHS, "cmp"); 457 break; 458 } 459 // ZExt bool to int type. 460 return RValue::get(Builder.CreateZExt(LHS, ConvertType(E->getType()))); 461 } 462 case Builtin::BI__builtin_isnan: { 463 Value *V = EmitScalarExpr(E->getArg(0)); 464 V = Builder.CreateFCmpUNO(V, V, "cmp"); 465 return RValue::get(Builder.CreateZExt(V, ConvertType(E->getType()))); 466 } 467 468 case Builtin::BI__builtin_isinf: { 469 // isinf(x) --> fabs(x) == infinity 470 Value *V = EmitScalarExpr(E->getArg(0)); 471 V = EmitFAbs(*this, V, E->getArg(0)->getType()); 472 473 V = Builder.CreateFCmpOEQ(V, ConstantFP::getInfinity(V->getType()),"isinf"); 474 return RValue::get(Builder.CreateZExt(V, ConvertType(E->getType()))); 475 } 476 477 // TODO: BI__builtin_isinf_sign 478 // isinf_sign(x) -> isinf(x) ? (signbit(x) ? -1 : 1) : 0 479 480 case Builtin::BI__builtin_isnormal: { 481 // isnormal(x) --> x == x && fabsf(x) < infinity && fabsf(x) >= float_min 482 Value *V = EmitScalarExpr(E->getArg(0)); 483 Value *Eq = Builder.CreateFCmpOEQ(V, V, "iseq"); 484 485 Value *Abs = EmitFAbs(*this, V, E->getArg(0)->getType()); 486 Value *IsLessThanInf = 487 Builder.CreateFCmpULT(Abs, ConstantFP::getInfinity(V->getType()),"isinf"); 488 APFloat Smallest = APFloat::getSmallestNormalized( 489 getContext().getFloatTypeSemantics(E->getArg(0)->getType())); 490 Value *IsNormal = 491 Builder.CreateFCmpUGE(Abs, ConstantFP::get(V->getContext(), Smallest), 492 "isnormal"); 493 V = Builder.CreateAnd(Eq, IsLessThanInf, "and"); 494 V = Builder.CreateAnd(V, IsNormal, "and"); 495 return RValue::get(Builder.CreateZExt(V, ConvertType(E->getType()))); 496 } 497 498 case Builtin::BI__builtin_isfinite: { 499 // isfinite(x) --> x == x && fabs(x) != infinity; 500 Value *V = EmitScalarExpr(E->getArg(0)); 501 Value *Eq = Builder.CreateFCmpOEQ(V, V, "iseq"); 502 503 Value *Abs = EmitFAbs(*this, V, E->getArg(0)->getType()); 504 Value *IsNotInf = 505 Builder.CreateFCmpUNE(Abs, ConstantFP::getInfinity(V->getType()),"isinf"); 506 507 V = Builder.CreateAnd(Eq, IsNotInf, "and"); 508 return RValue::get(Builder.CreateZExt(V, ConvertType(E->getType()))); 509 } 510 511 case Builtin::BI__builtin_fpclassify: { 512 Value *V = EmitScalarExpr(E->getArg(5)); 513 llvm::Type *Ty = ConvertType(E->getArg(5)->getType()); 514 515 // Create Result 516 BasicBlock *Begin = Builder.GetInsertBlock(); 517 BasicBlock *End = createBasicBlock("fpclassify_end", this->CurFn); 518 Builder.SetInsertPoint(End); 519 PHINode *Result = 520 Builder.CreatePHI(ConvertType(E->getArg(0)->getType()), 4, 521 "fpclassify_result"); 522 523 // if (V==0) return FP_ZERO 524 Builder.SetInsertPoint(Begin); 525 Value *IsZero = Builder.CreateFCmpOEQ(V, Constant::getNullValue(Ty), 526 "iszero"); 527 Value *ZeroLiteral = EmitScalarExpr(E->getArg(4)); 528 BasicBlock *NotZero = createBasicBlock("fpclassify_not_zero", this->CurFn); 529 Builder.CreateCondBr(IsZero, End, NotZero); 530 Result->addIncoming(ZeroLiteral, Begin); 531 532 // if (V != V) return FP_NAN 533 Builder.SetInsertPoint(NotZero); 534 Value *IsNan = Builder.CreateFCmpUNO(V, V, "cmp"); 535 Value *NanLiteral = EmitScalarExpr(E->getArg(0)); 536 BasicBlock *NotNan = createBasicBlock("fpclassify_not_nan", this->CurFn); 537 Builder.CreateCondBr(IsNan, End, NotNan); 538 Result->addIncoming(NanLiteral, NotZero); 539 540 // if (fabs(V) == infinity) return FP_INFINITY 541 Builder.SetInsertPoint(NotNan); 542 Value *VAbs = EmitFAbs(*this, V, E->getArg(5)->getType()); 543 Value *IsInf = 544 Builder.CreateFCmpOEQ(VAbs, ConstantFP::getInfinity(V->getType()), 545 "isinf"); 546 Value *InfLiteral = EmitScalarExpr(E->getArg(1)); 547 BasicBlock *NotInf = createBasicBlock("fpclassify_not_inf", this->CurFn); 548 Builder.CreateCondBr(IsInf, End, NotInf); 549 Result->addIncoming(InfLiteral, NotNan); 550 551 // if (fabs(V) >= MIN_NORMAL) return FP_NORMAL else FP_SUBNORMAL 552 Builder.SetInsertPoint(NotInf); 553 APFloat Smallest = APFloat::getSmallestNormalized( 554 getContext().getFloatTypeSemantics(E->getArg(5)->getType())); 555 Value *IsNormal = 556 Builder.CreateFCmpUGE(VAbs, ConstantFP::get(V->getContext(), Smallest), 557 "isnormal"); 558 Value *NormalResult = 559 Builder.CreateSelect(IsNormal, EmitScalarExpr(E->getArg(2)), 560 EmitScalarExpr(E->getArg(3))); 561 Builder.CreateBr(End); 562 Result->addIncoming(NormalResult, NotInf); 563 564 // return Result 565 Builder.SetInsertPoint(End); 566 return RValue::get(Result); 567 } 568 569 case Builtin::BIalloca: 570 case Builtin::BI__builtin_alloca: { 571 Value *Size = EmitScalarExpr(E->getArg(0)); 572 return RValue::get(Builder.CreateAlloca(Builder.getInt8Ty(), Size)); 573 } 574 case Builtin::BIbzero: 575 case Builtin::BI__builtin_bzero: { 576 Value *Address = EmitScalarExpr(E->getArg(0)); 577 Value *SizeVal = EmitScalarExpr(E->getArg(1)); 578 unsigned Align = GetPointeeAlignment(E->getArg(0)); 579 Builder.CreateMemSet(Address, Builder.getInt8(0), SizeVal, Align, false); 580 return RValue::get(Address); 581 } 582 case Builtin::BImemcpy: 583 case Builtin::BI__builtin_memcpy: { 584 Value *Address = EmitScalarExpr(E->getArg(0)); 585 Value *SrcAddr = EmitScalarExpr(E->getArg(1)); 586 Value *SizeVal = EmitScalarExpr(E->getArg(2)); 587 unsigned Align = std::min(GetPointeeAlignment(E->getArg(0)), 588 GetPointeeAlignment(E->getArg(1))); 589 Builder.CreateMemCpy(Address, SrcAddr, SizeVal, Align, false); 590 return RValue::get(Address); 591 } 592 593 case Builtin::BI__builtin___memcpy_chk: { 594 // fold __builtin_memcpy_chk(x, y, cst1, cst2) to memset iff cst1<=cst2. 595 llvm::APSInt Size, DstSize; 596 if (!E->getArg(2)->EvaluateAsInt(Size, CGM.getContext()) || 597 !E->getArg(3)->EvaluateAsInt(DstSize, CGM.getContext())) 598 break; 599 if (Size.ugt(DstSize)) 600 break; 601 Value *Dest = EmitScalarExpr(E->getArg(0)); 602 Value *Src = EmitScalarExpr(E->getArg(1)); 603 Value *SizeVal = llvm::ConstantInt::get(Builder.getContext(), Size); 604 unsigned Align = std::min(GetPointeeAlignment(E->getArg(0)), 605 GetPointeeAlignment(E->getArg(1))); 606 Builder.CreateMemCpy(Dest, Src, SizeVal, Align, false); 607 return RValue::get(Dest); 608 } 609 610 case Builtin::BI__builtin_objc_memmove_collectable: { 611 Value *Address = EmitScalarExpr(E->getArg(0)); 612 Value *SrcAddr = EmitScalarExpr(E->getArg(1)); 613 Value *SizeVal = EmitScalarExpr(E->getArg(2)); 614 CGM.getObjCRuntime().EmitGCMemmoveCollectable(*this, 615 Address, SrcAddr, SizeVal); 616 return RValue::get(Address); 617 } 618 619 case Builtin::BI__builtin___memmove_chk: { 620 // fold __builtin_memmove_chk(x, y, cst1, cst2) to memset iff cst1<=cst2. 621 llvm::APSInt Size, DstSize; 622 if (!E->getArg(2)->EvaluateAsInt(Size, CGM.getContext()) || 623 !E->getArg(3)->EvaluateAsInt(DstSize, CGM.getContext())) 624 break; 625 if (Size.ugt(DstSize)) 626 break; 627 Value *Dest = EmitScalarExpr(E->getArg(0)); 628 Value *Src = EmitScalarExpr(E->getArg(1)); 629 Value *SizeVal = llvm::ConstantInt::get(Builder.getContext(), Size); 630 unsigned Align = std::min(GetPointeeAlignment(E->getArg(0)), 631 GetPointeeAlignment(E->getArg(1))); 632 Builder.CreateMemMove(Dest, Src, SizeVal, Align, false); 633 return RValue::get(Dest); 634 } 635 636 case Builtin::BImemmove: 637 case Builtin::BI__builtin_memmove: { 638 Value *Address = EmitScalarExpr(E->getArg(0)); 639 Value *SrcAddr = EmitScalarExpr(E->getArg(1)); 640 Value *SizeVal = EmitScalarExpr(E->getArg(2)); 641 unsigned Align = std::min(GetPointeeAlignment(E->getArg(0)), 642 GetPointeeAlignment(E->getArg(1))); 643 Builder.CreateMemMove(Address, SrcAddr, SizeVal, Align, false); 644 return RValue::get(Address); 645 } 646 case Builtin::BImemset: 647 case Builtin::BI__builtin_memset: { 648 Value *Address = EmitScalarExpr(E->getArg(0)); 649 Value *ByteVal = Builder.CreateTrunc(EmitScalarExpr(E->getArg(1)), 650 Builder.getInt8Ty()); 651 Value *SizeVal = EmitScalarExpr(E->getArg(2)); 652 unsigned Align = GetPointeeAlignment(E->getArg(0)); 653 Builder.CreateMemSet(Address, ByteVal, SizeVal, Align, false); 654 return RValue::get(Address); 655 } 656 case Builtin::BI__builtin___memset_chk: { 657 // fold __builtin_memset_chk(x, y, cst1, cst2) to memset iff cst1<=cst2. 658 llvm::APSInt Size, DstSize; 659 if (!E->getArg(2)->EvaluateAsInt(Size, CGM.getContext()) || 660 !E->getArg(3)->EvaluateAsInt(DstSize, CGM.getContext())) 661 break; 662 if (Size.ugt(DstSize)) 663 break; 664 Value *Address = EmitScalarExpr(E->getArg(0)); 665 Value *ByteVal = Builder.CreateTrunc(EmitScalarExpr(E->getArg(1)), 666 Builder.getInt8Ty()); 667 Value *SizeVal = llvm::ConstantInt::get(Builder.getContext(), Size); 668 unsigned Align = GetPointeeAlignment(E->getArg(0)); 669 Builder.CreateMemSet(Address, ByteVal, SizeVal, Align, false); 670 671 return RValue::get(Address); 672 } 673 case Builtin::BI__builtin_dwarf_cfa: { 674 // The offset in bytes from the first argument to the CFA. 675 // 676 // Why on earth is this in the frontend? Is there any reason at 677 // all that the backend can't reasonably determine this while 678 // lowering llvm.eh.dwarf.cfa()? 679 // 680 // TODO: If there's a satisfactory reason, add a target hook for 681 // this instead of hard-coding 0, which is correct for most targets. 682 int32_t Offset = 0; 683 684 Value *F = CGM.getIntrinsic(Intrinsic::eh_dwarf_cfa); 685 return RValue::get(Builder.CreateCall(F, 686 llvm::ConstantInt::get(Int32Ty, Offset))); 687 } 688 case Builtin::BI__builtin_return_address: { 689 Value *Depth = EmitScalarExpr(E->getArg(0)); 690 Depth = Builder.CreateIntCast(Depth, Int32Ty, false); 691 Value *F = CGM.getIntrinsic(Intrinsic::returnaddress); 692 return RValue::get(Builder.CreateCall(F, Depth)); 693 } 694 case Builtin::BI__builtin_frame_address: { 695 Value *Depth = EmitScalarExpr(E->getArg(0)); 696 Depth = Builder.CreateIntCast(Depth, Int32Ty, false); 697 Value *F = CGM.getIntrinsic(Intrinsic::frameaddress); 698 return RValue::get(Builder.CreateCall(F, Depth)); 699 } 700 case Builtin::BI__builtin_extract_return_addr: { 701 Value *Address = EmitScalarExpr(E->getArg(0)); 702 Value *Result = getTargetHooks().decodeReturnAddress(*this, Address); 703 return RValue::get(Result); 704 } 705 case Builtin::BI__builtin_frob_return_addr: { 706 Value *Address = EmitScalarExpr(E->getArg(0)); 707 Value *Result = getTargetHooks().encodeReturnAddress(*this, Address); 708 return RValue::get(Result); 709 } 710 case Builtin::BI__builtin_dwarf_sp_column: { 711 llvm::IntegerType *Ty 712 = cast<llvm::IntegerType>(ConvertType(E->getType())); 713 int Column = getTargetHooks().getDwarfEHStackPointer(CGM); 714 if (Column == -1) { 715 CGM.ErrorUnsupported(E, "__builtin_dwarf_sp_column"); 716 return RValue::get(llvm::UndefValue::get(Ty)); 717 } 718 return RValue::get(llvm::ConstantInt::get(Ty, Column, true)); 719 } 720 case Builtin::BI__builtin_init_dwarf_reg_size_table: { 721 Value *Address = EmitScalarExpr(E->getArg(0)); 722 if (getTargetHooks().initDwarfEHRegSizeTable(*this, Address)) 723 CGM.ErrorUnsupported(E, "__builtin_init_dwarf_reg_size_table"); 724 return RValue::get(llvm::UndefValue::get(ConvertType(E->getType()))); 725 } 726 case Builtin::BI__builtin_eh_return: { 727 Value *Int = EmitScalarExpr(E->getArg(0)); 728 Value *Ptr = EmitScalarExpr(E->getArg(1)); 729 730 llvm::IntegerType *IntTy = cast<llvm::IntegerType>(Int->getType()); 731 assert((IntTy->getBitWidth() == 32 || IntTy->getBitWidth() == 64) && 732 "LLVM's __builtin_eh_return only supports 32- and 64-bit variants"); 733 Value *F = CGM.getIntrinsic(IntTy->getBitWidth() == 32 734 ? Intrinsic::eh_return_i32 735 : Intrinsic::eh_return_i64); 736 Builder.CreateCall2(F, Int, Ptr); 737 Builder.CreateUnreachable(); 738 739 // We do need to preserve an insertion point. 740 EmitBlock(createBasicBlock("builtin_eh_return.cont")); 741 742 return RValue::get(0); 743 } 744 case Builtin::BI__builtin_unwind_init: { 745 Value *F = CGM.getIntrinsic(Intrinsic::eh_unwind_init); 746 return RValue::get(Builder.CreateCall(F)); 747 } 748 case Builtin::BI__builtin_extend_pointer: { 749 // Extends a pointer to the size of an _Unwind_Word, which is 750 // uint64_t on all platforms. Generally this gets poked into a 751 // register and eventually used as an address, so if the 752 // addressing registers are wider than pointers and the platform 753 // doesn't implicitly ignore high-order bits when doing 754 // addressing, we need to make sure we zext / sext based on 755 // the platform's expectations. 756 // 757 // See: http://gcc.gnu.org/ml/gcc-bugs/2002-02/msg00237.html 758 759 // Cast the pointer to intptr_t. 760 Value *Ptr = EmitScalarExpr(E->getArg(0)); 761 Value *Result = Builder.CreatePtrToInt(Ptr, IntPtrTy, "extend.cast"); 762 763 // If that's 64 bits, we're done. 764 if (IntPtrTy->getBitWidth() == 64) 765 return RValue::get(Result); 766 767 // Otherwise, ask the codegen data what to do. 768 if (getTargetHooks().extendPointerWithSExt()) 769 return RValue::get(Builder.CreateSExt(Result, Int64Ty, "extend.sext")); 770 else 771 return RValue::get(Builder.CreateZExt(Result, Int64Ty, "extend.zext")); 772 } 773 case Builtin::BI__builtin_setjmp: { 774 // Buffer is a void**. 775 Value *Buf = EmitScalarExpr(E->getArg(0)); 776 777 // Store the frame pointer to the setjmp buffer. 778 Value *FrameAddr = 779 Builder.CreateCall(CGM.getIntrinsic(Intrinsic::frameaddress), 780 ConstantInt::get(Int32Ty, 0)); 781 Builder.CreateStore(FrameAddr, Buf); 782 783 // Store the stack pointer to the setjmp buffer. 784 Value *StackAddr = 785 Builder.CreateCall(CGM.getIntrinsic(Intrinsic::stacksave)); 786 Value *StackSaveSlot = 787 Builder.CreateGEP(Buf, ConstantInt::get(Int32Ty, 2)); 788 Builder.CreateStore(StackAddr, StackSaveSlot); 789 790 // Call LLVM's EH setjmp, which is lightweight. 791 Value *F = CGM.getIntrinsic(Intrinsic::eh_sjlj_setjmp); 792 Buf = Builder.CreateBitCast(Buf, Int8PtrTy); 793 return RValue::get(Builder.CreateCall(F, Buf)); 794 } 795 case Builtin::BI__builtin_longjmp: { 796 Value *Buf = EmitScalarExpr(E->getArg(0)); 797 Buf = Builder.CreateBitCast(Buf, Int8PtrTy); 798 799 // Call LLVM's EH longjmp, which is lightweight. 800 Builder.CreateCall(CGM.getIntrinsic(Intrinsic::eh_sjlj_longjmp), Buf); 801 802 // longjmp doesn't return; mark this as unreachable. 803 Builder.CreateUnreachable(); 804 805 // We do need to preserve an insertion point. 806 EmitBlock(createBasicBlock("longjmp.cont")); 807 808 return RValue::get(0); 809 } 810 case Builtin::BI__sync_fetch_and_add: 811 case Builtin::BI__sync_fetch_and_sub: 812 case Builtin::BI__sync_fetch_and_or: 813 case Builtin::BI__sync_fetch_and_and: 814 case Builtin::BI__sync_fetch_and_xor: 815 case Builtin::BI__sync_add_and_fetch: 816 case Builtin::BI__sync_sub_and_fetch: 817 case Builtin::BI__sync_and_and_fetch: 818 case Builtin::BI__sync_or_and_fetch: 819 case Builtin::BI__sync_xor_and_fetch: 820 case Builtin::BI__sync_val_compare_and_swap: 821 case Builtin::BI__sync_bool_compare_and_swap: 822 case Builtin::BI__sync_lock_test_and_set: 823 case Builtin::BI__sync_lock_release: 824 case Builtin::BI__sync_swap: 825 llvm_unreachable("Shouldn't make it through sema"); 826 case Builtin::BI__sync_fetch_and_add_1: 827 case Builtin::BI__sync_fetch_and_add_2: 828 case Builtin::BI__sync_fetch_and_add_4: 829 case Builtin::BI__sync_fetch_and_add_8: 830 case Builtin::BI__sync_fetch_and_add_16: 831 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Add, E); 832 case Builtin::BI__sync_fetch_and_sub_1: 833 case Builtin::BI__sync_fetch_and_sub_2: 834 case Builtin::BI__sync_fetch_and_sub_4: 835 case Builtin::BI__sync_fetch_and_sub_8: 836 case Builtin::BI__sync_fetch_and_sub_16: 837 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Sub, E); 838 case Builtin::BI__sync_fetch_and_or_1: 839 case Builtin::BI__sync_fetch_and_or_2: 840 case Builtin::BI__sync_fetch_and_or_4: 841 case Builtin::BI__sync_fetch_and_or_8: 842 case Builtin::BI__sync_fetch_and_or_16: 843 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Or, E); 844 case Builtin::BI__sync_fetch_and_and_1: 845 case Builtin::BI__sync_fetch_and_and_2: 846 case Builtin::BI__sync_fetch_and_and_4: 847 case Builtin::BI__sync_fetch_and_and_8: 848 case Builtin::BI__sync_fetch_and_and_16: 849 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::And, E); 850 case Builtin::BI__sync_fetch_and_xor_1: 851 case Builtin::BI__sync_fetch_and_xor_2: 852 case Builtin::BI__sync_fetch_and_xor_4: 853 case Builtin::BI__sync_fetch_and_xor_8: 854 case Builtin::BI__sync_fetch_and_xor_16: 855 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Xor, E); 856 857 // Clang extensions: not overloaded yet. 858 case Builtin::BI__sync_fetch_and_min: 859 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Min, E); 860 case Builtin::BI__sync_fetch_and_max: 861 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Max, E); 862 case Builtin::BI__sync_fetch_and_umin: 863 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::UMin, E); 864 case Builtin::BI__sync_fetch_and_umax: 865 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::UMax, E); 866 867 case Builtin::BI__sync_add_and_fetch_1: 868 case Builtin::BI__sync_add_and_fetch_2: 869 case Builtin::BI__sync_add_and_fetch_4: 870 case Builtin::BI__sync_add_and_fetch_8: 871 case Builtin::BI__sync_add_and_fetch_16: 872 return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::Add, E, 873 llvm::Instruction::Add); 874 case Builtin::BI__sync_sub_and_fetch_1: 875 case Builtin::BI__sync_sub_and_fetch_2: 876 case Builtin::BI__sync_sub_and_fetch_4: 877 case Builtin::BI__sync_sub_and_fetch_8: 878 case Builtin::BI__sync_sub_and_fetch_16: 879 return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::Sub, E, 880 llvm::Instruction::Sub); 881 case Builtin::BI__sync_and_and_fetch_1: 882 case Builtin::BI__sync_and_and_fetch_2: 883 case Builtin::BI__sync_and_and_fetch_4: 884 case Builtin::BI__sync_and_and_fetch_8: 885 case Builtin::BI__sync_and_and_fetch_16: 886 return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::And, E, 887 llvm::Instruction::And); 888 case Builtin::BI__sync_or_and_fetch_1: 889 case Builtin::BI__sync_or_and_fetch_2: 890 case Builtin::BI__sync_or_and_fetch_4: 891 case Builtin::BI__sync_or_and_fetch_8: 892 case Builtin::BI__sync_or_and_fetch_16: 893 return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::Or, E, 894 llvm::Instruction::Or); 895 case Builtin::BI__sync_xor_and_fetch_1: 896 case Builtin::BI__sync_xor_and_fetch_2: 897 case Builtin::BI__sync_xor_and_fetch_4: 898 case Builtin::BI__sync_xor_and_fetch_8: 899 case Builtin::BI__sync_xor_and_fetch_16: 900 return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::Xor, E, 901 llvm::Instruction::Xor); 902 903 case Builtin::BI__sync_val_compare_and_swap_1: 904 case Builtin::BI__sync_val_compare_and_swap_2: 905 case Builtin::BI__sync_val_compare_and_swap_4: 906 case Builtin::BI__sync_val_compare_and_swap_8: 907 case Builtin::BI__sync_val_compare_and_swap_16: { 908 QualType T = E->getType(); 909 llvm::Value *DestPtr = EmitScalarExpr(E->getArg(0)); 910 unsigned AddrSpace = 911 cast<llvm::PointerType>(DestPtr->getType())->getAddressSpace(); 912 913 llvm::IntegerType *IntType = 914 llvm::IntegerType::get(getLLVMContext(), 915 getContext().getTypeSize(T)); 916 llvm::Type *IntPtrType = IntType->getPointerTo(AddrSpace); 917 918 Value *Args[3]; 919 Args[0] = Builder.CreateBitCast(DestPtr, IntPtrType); 920 Args[1] = EmitScalarExpr(E->getArg(1)); 921 llvm::Type *ValueType = Args[1]->getType(); 922 Args[1] = EmitToInt(*this, Args[1], T, IntType); 923 Args[2] = EmitToInt(*this, EmitScalarExpr(E->getArg(2)), T, IntType); 924 925 Value *Result = Builder.CreateAtomicCmpXchg(Args[0], Args[1], Args[2], 926 llvm::SequentiallyConsistent); 927 Result = EmitFromInt(*this, Result, T, ValueType); 928 return RValue::get(Result); 929 } 930 931 case Builtin::BI__sync_bool_compare_and_swap_1: 932 case Builtin::BI__sync_bool_compare_and_swap_2: 933 case Builtin::BI__sync_bool_compare_and_swap_4: 934 case Builtin::BI__sync_bool_compare_and_swap_8: 935 case Builtin::BI__sync_bool_compare_and_swap_16: { 936 QualType T = E->getArg(1)->getType(); 937 llvm::Value *DestPtr = EmitScalarExpr(E->getArg(0)); 938 unsigned AddrSpace = 939 cast<llvm::PointerType>(DestPtr->getType())->getAddressSpace(); 940 941 llvm::IntegerType *IntType = 942 llvm::IntegerType::get(getLLVMContext(), 943 getContext().getTypeSize(T)); 944 llvm::Type *IntPtrType = IntType->getPointerTo(AddrSpace); 945 946 Value *Args[3]; 947 Args[0] = Builder.CreateBitCast(DestPtr, IntPtrType); 948 Args[1] = EmitToInt(*this, EmitScalarExpr(E->getArg(1)), T, IntType); 949 Args[2] = EmitToInt(*this, EmitScalarExpr(E->getArg(2)), T, IntType); 950 951 Value *OldVal = Args[1]; 952 Value *PrevVal = Builder.CreateAtomicCmpXchg(Args[0], Args[1], Args[2], 953 llvm::SequentiallyConsistent); 954 Value *Result = Builder.CreateICmpEQ(PrevVal, OldVal); 955 // zext bool to int. 956 Result = Builder.CreateZExt(Result, ConvertType(E->getType())); 957 return RValue::get(Result); 958 } 959 960 case Builtin::BI__sync_swap_1: 961 case Builtin::BI__sync_swap_2: 962 case Builtin::BI__sync_swap_4: 963 case Builtin::BI__sync_swap_8: 964 case Builtin::BI__sync_swap_16: 965 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Xchg, E); 966 967 case Builtin::BI__sync_lock_test_and_set_1: 968 case Builtin::BI__sync_lock_test_and_set_2: 969 case Builtin::BI__sync_lock_test_and_set_4: 970 case Builtin::BI__sync_lock_test_and_set_8: 971 case Builtin::BI__sync_lock_test_and_set_16: 972 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Xchg, E); 973 974 case Builtin::BI__sync_lock_release_1: 975 case Builtin::BI__sync_lock_release_2: 976 case Builtin::BI__sync_lock_release_4: 977 case Builtin::BI__sync_lock_release_8: 978 case Builtin::BI__sync_lock_release_16: { 979 Value *Ptr = EmitScalarExpr(E->getArg(0)); 980 QualType ElTy = E->getArg(0)->getType()->getPointeeType(); 981 CharUnits StoreSize = getContext().getTypeSizeInChars(ElTy); 982 llvm::Type *ITy = llvm::IntegerType::get(getLLVMContext(), 983 StoreSize.getQuantity() * 8); 984 Ptr = Builder.CreateBitCast(Ptr, ITy->getPointerTo()); 985 llvm::StoreInst *Store = 986 Builder.CreateStore(llvm::Constant::getNullValue(ITy), Ptr); 987 Store->setAlignment(StoreSize.getQuantity()); 988 Store->setAtomic(llvm::Release); 989 return RValue::get(0); 990 } 991 992 case Builtin::BI__sync_synchronize: { 993 // We assume this is supposed to correspond to a C++0x-style 994 // sequentially-consistent fence (i.e. this is only usable for 995 // synchonization, not device I/O or anything like that). This intrinsic 996 // is really badly designed in the sense that in theory, there isn't 997 // any way to safely use it... but in practice, it mostly works 998 // to use it with non-atomic loads and stores to get acquire/release 999 // semantics. 1000 Builder.CreateFence(llvm::SequentiallyConsistent); 1001 return RValue::get(0); 1002 } 1003 1004 case Builtin::BI__c11_atomic_is_lock_free: 1005 case Builtin::BI__atomic_is_lock_free: { 1006 // Call "bool __atomic_is_lock_free(size_t size, void *ptr)". For the 1007 // __c11 builtin, ptr is 0 (indicating a properly-aligned object), since 1008 // _Atomic(T) is always properly-aligned. 1009 const char *LibCallName = "__atomic_is_lock_free"; 1010 CallArgList Args; 1011 Args.add(RValue::get(EmitScalarExpr(E->getArg(0))), 1012 getContext().getSizeType()); 1013 if (BuiltinID == Builtin::BI__atomic_is_lock_free) 1014 Args.add(RValue::get(EmitScalarExpr(E->getArg(1))), 1015 getContext().VoidPtrTy); 1016 else 1017 Args.add(RValue::get(llvm::Constant::getNullValue(VoidPtrTy)), 1018 getContext().VoidPtrTy); 1019 const CGFunctionInfo &FuncInfo = 1020 CGM.getTypes().arrangeFreeFunctionCall(E->getType(), Args, 1021 FunctionType::ExtInfo(), 1022 RequiredArgs::All); 1023 llvm::FunctionType *FTy = CGM.getTypes().GetFunctionType(FuncInfo); 1024 llvm::Constant *Func = CGM.CreateRuntimeFunction(FTy, LibCallName); 1025 return EmitCall(FuncInfo, Func, ReturnValueSlot(), Args); 1026 } 1027 1028 case Builtin::BI__atomic_test_and_set: { 1029 // Look at the argument type to determine whether this is a volatile 1030 // operation. The parameter type is always volatile. 1031 QualType PtrTy = E->getArg(0)->IgnoreImpCasts()->getType(); 1032 bool Volatile = 1033 PtrTy->castAs<PointerType>()->getPointeeType().isVolatileQualified(); 1034 1035 Value *Ptr = EmitScalarExpr(E->getArg(0)); 1036 unsigned AddrSpace = 1037 cast<llvm::PointerType>(Ptr->getType())->getAddressSpace(); 1038 Ptr = Builder.CreateBitCast(Ptr, Int8Ty->getPointerTo(AddrSpace)); 1039 Value *NewVal = Builder.getInt8(1); 1040 Value *Order = EmitScalarExpr(E->getArg(1)); 1041 if (isa<llvm::ConstantInt>(Order)) { 1042 int ord = cast<llvm::ConstantInt>(Order)->getZExtValue(); 1043 AtomicRMWInst *Result = 0; 1044 switch (ord) { 1045 case 0: // memory_order_relaxed 1046 default: // invalid order 1047 Result = Builder.CreateAtomicRMW(llvm::AtomicRMWInst::Xchg, 1048 Ptr, NewVal, 1049 llvm::Monotonic); 1050 break; 1051 case 1: // memory_order_consume 1052 case 2: // memory_order_acquire 1053 Result = Builder.CreateAtomicRMW(llvm::AtomicRMWInst::Xchg, 1054 Ptr, NewVal, 1055 llvm::Acquire); 1056 break; 1057 case 3: // memory_order_release 1058 Result = Builder.CreateAtomicRMW(llvm::AtomicRMWInst::Xchg, 1059 Ptr, NewVal, 1060 llvm::Release); 1061 break; 1062 case 4: // memory_order_acq_rel 1063 Result = Builder.CreateAtomicRMW(llvm::AtomicRMWInst::Xchg, 1064 Ptr, NewVal, 1065 llvm::AcquireRelease); 1066 break; 1067 case 5: // memory_order_seq_cst 1068 Result = Builder.CreateAtomicRMW(llvm::AtomicRMWInst::Xchg, 1069 Ptr, NewVal, 1070 llvm::SequentiallyConsistent); 1071 break; 1072 } 1073 Result->setVolatile(Volatile); 1074 return RValue::get(Builder.CreateIsNotNull(Result, "tobool")); 1075 } 1076 1077 llvm::BasicBlock *ContBB = createBasicBlock("atomic.continue", CurFn); 1078 1079 llvm::BasicBlock *BBs[5] = { 1080 createBasicBlock("monotonic", CurFn), 1081 createBasicBlock("acquire", CurFn), 1082 createBasicBlock("release", CurFn), 1083 createBasicBlock("acqrel", CurFn), 1084 createBasicBlock("seqcst", CurFn) 1085 }; 1086 llvm::AtomicOrdering Orders[5] = { 1087 llvm::Monotonic, llvm::Acquire, llvm::Release, 1088 llvm::AcquireRelease, llvm::SequentiallyConsistent 1089 }; 1090 1091 Order = Builder.CreateIntCast(Order, Builder.getInt32Ty(), false); 1092 llvm::SwitchInst *SI = Builder.CreateSwitch(Order, BBs[0]); 1093 1094 Builder.SetInsertPoint(ContBB); 1095 PHINode *Result = Builder.CreatePHI(Int8Ty, 5, "was_set"); 1096 1097 for (unsigned i = 0; i < 5; ++i) { 1098 Builder.SetInsertPoint(BBs[i]); 1099 AtomicRMWInst *RMW = Builder.CreateAtomicRMW(llvm::AtomicRMWInst::Xchg, 1100 Ptr, NewVal, Orders[i]); 1101 RMW->setVolatile(Volatile); 1102 Result->addIncoming(RMW, BBs[i]); 1103 Builder.CreateBr(ContBB); 1104 } 1105 1106 SI->addCase(Builder.getInt32(0), BBs[0]); 1107 SI->addCase(Builder.getInt32(1), BBs[1]); 1108 SI->addCase(Builder.getInt32(2), BBs[1]); 1109 SI->addCase(Builder.getInt32(3), BBs[2]); 1110 SI->addCase(Builder.getInt32(4), BBs[3]); 1111 SI->addCase(Builder.getInt32(5), BBs[4]); 1112 1113 Builder.SetInsertPoint(ContBB); 1114 return RValue::get(Builder.CreateIsNotNull(Result, "tobool")); 1115 } 1116 1117 case Builtin::BI__atomic_clear: { 1118 QualType PtrTy = E->getArg(0)->IgnoreImpCasts()->getType(); 1119 bool Volatile = 1120 PtrTy->castAs<PointerType>()->getPointeeType().isVolatileQualified(); 1121 1122 Value *Ptr = EmitScalarExpr(E->getArg(0)); 1123 unsigned AddrSpace = 1124 cast<llvm::PointerType>(Ptr->getType())->getAddressSpace(); 1125 Ptr = Builder.CreateBitCast(Ptr, Int8Ty->getPointerTo(AddrSpace)); 1126 Value *NewVal = Builder.getInt8(0); 1127 Value *Order = EmitScalarExpr(E->getArg(1)); 1128 if (isa<llvm::ConstantInt>(Order)) { 1129 int ord = cast<llvm::ConstantInt>(Order)->getZExtValue(); 1130 StoreInst *Store = Builder.CreateStore(NewVal, Ptr, Volatile); 1131 Store->setAlignment(1); 1132 switch (ord) { 1133 case 0: // memory_order_relaxed 1134 default: // invalid order 1135 Store->setOrdering(llvm::Monotonic); 1136 break; 1137 case 3: // memory_order_release 1138 Store->setOrdering(llvm::Release); 1139 break; 1140 case 5: // memory_order_seq_cst 1141 Store->setOrdering(llvm::SequentiallyConsistent); 1142 break; 1143 } 1144 return RValue::get(0); 1145 } 1146 1147 llvm::BasicBlock *ContBB = createBasicBlock("atomic.continue", CurFn); 1148 1149 llvm::BasicBlock *BBs[3] = { 1150 createBasicBlock("monotonic", CurFn), 1151 createBasicBlock("release", CurFn), 1152 createBasicBlock("seqcst", CurFn) 1153 }; 1154 llvm::AtomicOrdering Orders[3] = { 1155 llvm::Monotonic, llvm::Release, llvm::SequentiallyConsistent 1156 }; 1157 1158 Order = Builder.CreateIntCast(Order, Builder.getInt32Ty(), false); 1159 llvm::SwitchInst *SI = Builder.CreateSwitch(Order, BBs[0]); 1160 1161 for (unsigned i = 0; i < 3; ++i) { 1162 Builder.SetInsertPoint(BBs[i]); 1163 StoreInst *Store = Builder.CreateStore(NewVal, Ptr, Volatile); 1164 Store->setAlignment(1); 1165 Store->setOrdering(Orders[i]); 1166 Builder.CreateBr(ContBB); 1167 } 1168 1169 SI->addCase(Builder.getInt32(0), BBs[0]); 1170 SI->addCase(Builder.getInt32(3), BBs[1]); 1171 SI->addCase(Builder.getInt32(5), BBs[2]); 1172 1173 Builder.SetInsertPoint(ContBB); 1174 return RValue::get(0); 1175 } 1176 1177 case Builtin::BI__atomic_thread_fence: 1178 case Builtin::BI__atomic_signal_fence: 1179 case Builtin::BI__c11_atomic_thread_fence: 1180 case Builtin::BI__c11_atomic_signal_fence: { 1181 llvm::SynchronizationScope Scope; 1182 if (BuiltinID == Builtin::BI__atomic_signal_fence || 1183 BuiltinID == Builtin::BI__c11_atomic_signal_fence) 1184 Scope = llvm::SingleThread; 1185 else 1186 Scope = llvm::CrossThread; 1187 Value *Order = EmitScalarExpr(E->getArg(0)); 1188 if (isa<llvm::ConstantInt>(Order)) { 1189 int ord = cast<llvm::ConstantInt>(Order)->getZExtValue(); 1190 switch (ord) { 1191 case 0: // memory_order_relaxed 1192 default: // invalid order 1193 break; 1194 case 1: // memory_order_consume 1195 case 2: // memory_order_acquire 1196 Builder.CreateFence(llvm::Acquire, Scope); 1197 break; 1198 case 3: // memory_order_release 1199 Builder.CreateFence(llvm::Release, Scope); 1200 break; 1201 case 4: // memory_order_acq_rel 1202 Builder.CreateFence(llvm::AcquireRelease, Scope); 1203 break; 1204 case 5: // memory_order_seq_cst 1205 Builder.CreateFence(llvm::SequentiallyConsistent, Scope); 1206 break; 1207 } 1208 return RValue::get(0); 1209 } 1210 1211 llvm::BasicBlock *AcquireBB, *ReleaseBB, *AcqRelBB, *SeqCstBB; 1212 AcquireBB = createBasicBlock("acquire", CurFn); 1213 ReleaseBB = createBasicBlock("release", CurFn); 1214 AcqRelBB = createBasicBlock("acqrel", CurFn); 1215 SeqCstBB = createBasicBlock("seqcst", CurFn); 1216 llvm::BasicBlock *ContBB = createBasicBlock("atomic.continue", CurFn); 1217 1218 Order = Builder.CreateIntCast(Order, Builder.getInt32Ty(), false); 1219 llvm::SwitchInst *SI = Builder.CreateSwitch(Order, ContBB); 1220 1221 Builder.SetInsertPoint(AcquireBB); 1222 Builder.CreateFence(llvm::Acquire, Scope); 1223 Builder.CreateBr(ContBB); 1224 SI->addCase(Builder.getInt32(1), AcquireBB); 1225 SI->addCase(Builder.getInt32(2), AcquireBB); 1226 1227 Builder.SetInsertPoint(ReleaseBB); 1228 Builder.CreateFence(llvm::Release, Scope); 1229 Builder.CreateBr(ContBB); 1230 SI->addCase(Builder.getInt32(3), ReleaseBB); 1231 1232 Builder.SetInsertPoint(AcqRelBB); 1233 Builder.CreateFence(llvm::AcquireRelease, Scope); 1234 Builder.CreateBr(ContBB); 1235 SI->addCase(Builder.getInt32(4), AcqRelBB); 1236 1237 Builder.SetInsertPoint(SeqCstBB); 1238 Builder.CreateFence(llvm::SequentiallyConsistent, Scope); 1239 Builder.CreateBr(ContBB); 1240 SI->addCase(Builder.getInt32(5), SeqCstBB); 1241 1242 Builder.SetInsertPoint(ContBB); 1243 return RValue::get(0); 1244 } 1245 1246 // Library functions with special handling. 1247 case Builtin::BIsqrt: 1248 case Builtin::BIsqrtf: 1249 case Builtin::BIsqrtl: { 1250 // TODO: there is currently no set of optimizer flags 1251 // sufficient for us to rewrite sqrt to @llvm.sqrt. 1252 // -fmath-errno=0 is not good enough; we need finiteness. 1253 // We could probably precondition the call with an ult 1254 // against 0, but is that worth the complexity? 1255 break; 1256 } 1257 1258 case Builtin::BIpow: 1259 case Builtin::BIpowf: 1260 case Builtin::BIpowl: { 1261 // Rewrite sqrt to intrinsic if allowed. 1262 if (!FD->hasAttr<ConstAttr>()) 1263 break; 1264 Value *Base = EmitScalarExpr(E->getArg(0)); 1265 Value *Exponent = EmitScalarExpr(E->getArg(1)); 1266 llvm::Type *ArgType = Base->getType(); 1267 Value *F = CGM.getIntrinsic(Intrinsic::pow, ArgType); 1268 return RValue::get(Builder.CreateCall2(F, Base, Exponent)); 1269 } 1270 1271 case Builtin::BIfma: 1272 case Builtin::BIfmaf: 1273 case Builtin::BIfmal: 1274 case Builtin::BI__builtin_fma: 1275 case Builtin::BI__builtin_fmaf: 1276 case Builtin::BI__builtin_fmal: { 1277 // Rewrite fma to intrinsic. 1278 Value *FirstArg = EmitScalarExpr(E->getArg(0)); 1279 llvm::Type *ArgType = FirstArg->getType(); 1280 Value *F = CGM.getIntrinsic(Intrinsic::fma, ArgType); 1281 return RValue::get(Builder.CreateCall3(F, FirstArg, 1282 EmitScalarExpr(E->getArg(1)), 1283 EmitScalarExpr(E->getArg(2)))); 1284 } 1285 1286 case Builtin::BI__builtin_signbit: 1287 case Builtin::BI__builtin_signbitf: 1288 case Builtin::BI__builtin_signbitl: { 1289 LLVMContext &C = CGM.getLLVMContext(); 1290 1291 Value *Arg = EmitScalarExpr(E->getArg(0)); 1292 llvm::Type *ArgTy = Arg->getType(); 1293 if (ArgTy->isPPC_FP128Ty()) 1294 break; // FIXME: I'm not sure what the right implementation is here. 1295 int ArgWidth = ArgTy->getPrimitiveSizeInBits(); 1296 llvm::Type *ArgIntTy = llvm::IntegerType::get(C, ArgWidth); 1297 Value *BCArg = Builder.CreateBitCast(Arg, ArgIntTy); 1298 Value *ZeroCmp = llvm::Constant::getNullValue(ArgIntTy); 1299 Value *Result = Builder.CreateICmpSLT(BCArg, ZeroCmp); 1300 return RValue::get(Builder.CreateZExt(Result, ConvertType(E->getType()))); 1301 } 1302 case Builtin::BI__builtin_annotation: { 1303 llvm::Value *AnnVal = EmitScalarExpr(E->getArg(0)); 1304 llvm::Value *F = CGM.getIntrinsic(llvm::Intrinsic::annotation, 1305 AnnVal->getType()); 1306 1307 // Get the annotation string, go through casts. Sema requires this to be a 1308 // non-wide string literal, potentially casted, so the cast<> is safe. 1309 const Expr *AnnotationStrExpr = E->getArg(1)->IgnoreParenCasts(); 1310 llvm::StringRef Str = cast<StringLiteral>(AnnotationStrExpr)->getString(); 1311 return RValue::get(EmitAnnotationCall(F, AnnVal, Str, E->getExprLoc())); 1312 } 1313 } 1314 1315 // If this is an alias for a lib function (e.g. __builtin_sin), emit 1316 // the call using the normal call path, but using the unmangled 1317 // version of the function name. 1318 if (getContext().BuiltinInfo.isLibFunction(BuiltinID)) 1319 return emitLibraryCall(*this, FD, E, 1320 CGM.getBuiltinLibFunction(FD, BuiltinID)); 1321 1322 // If this is a predefined lib function (e.g. malloc), emit the call 1323 // using exactly the normal call path. 1324 if (getContext().BuiltinInfo.isPredefinedLibFunction(BuiltinID)) 1325 return emitLibraryCall(*this, FD, E, EmitScalarExpr(E->getCallee())); 1326 1327 // See if we have a target specific intrinsic. 1328 const char *Name = getContext().BuiltinInfo.GetName(BuiltinID); 1329 Intrinsic::ID IntrinsicID = Intrinsic::not_intrinsic; 1330 if (const char *Prefix = 1331 llvm::Triple::getArchTypePrefix(Target.getTriple().getArch())) 1332 IntrinsicID = Intrinsic::getIntrinsicForGCCBuiltin(Prefix, Name); 1333 1334 if (IntrinsicID != Intrinsic::not_intrinsic) { 1335 SmallVector<Value*, 16> Args; 1336 1337 // Find out if any arguments are required to be integer constant 1338 // expressions. 1339 unsigned ICEArguments = 0; 1340 ASTContext::GetBuiltinTypeError Error; 1341 getContext().GetBuiltinType(BuiltinID, Error, &ICEArguments); 1342 assert(Error == ASTContext::GE_None && "Should not codegen an error"); 1343 1344 Function *F = CGM.getIntrinsic(IntrinsicID); 1345 llvm::FunctionType *FTy = F->getFunctionType(); 1346 1347 for (unsigned i = 0, e = E->getNumArgs(); i != e; ++i) { 1348 Value *ArgValue; 1349 // If this is a normal argument, just emit it as a scalar. 1350 if ((ICEArguments & (1 << i)) == 0) { 1351 ArgValue = EmitScalarExpr(E->getArg(i)); 1352 } else { 1353 // If this is required to be a constant, constant fold it so that we 1354 // know that the generated intrinsic gets a ConstantInt. 1355 llvm::APSInt Result; 1356 bool IsConst = E->getArg(i)->isIntegerConstantExpr(Result,getContext()); 1357 assert(IsConst && "Constant arg isn't actually constant?"); 1358 (void)IsConst; 1359 ArgValue = llvm::ConstantInt::get(getLLVMContext(), Result); 1360 } 1361 1362 // If the intrinsic arg type is different from the builtin arg type 1363 // we need to do a bit cast. 1364 llvm::Type *PTy = FTy->getParamType(i); 1365 if (PTy != ArgValue->getType()) { 1366 assert(PTy->canLosslesslyBitCastTo(FTy->getParamType(i)) && 1367 "Must be able to losslessly bit cast to param"); 1368 ArgValue = Builder.CreateBitCast(ArgValue, PTy); 1369 } 1370 1371 Args.push_back(ArgValue); 1372 } 1373 1374 Value *V = Builder.CreateCall(F, Args); 1375 QualType BuiltinRetType = E->getType(); 1376 1377 llvm::Type *RetTy = VoidTy; 1378 if (!BuiltinRetType->isVoidType()) 1379 RetTy = ConvertType(BuiltinRetType); 1380 1381 if (RetTy != V->getType()) { 1382 assert(V->getType()->canLosslesslyBitCastTo(RetTy) && 1383 "Must be able to losslessly bit cast result type"); 1384 V = Builder.CreateBitCast(V, RetTy); 1385 } 1386 1387 return RValue::get(V); 1388 } 1389 1390 // See if we have a target specific builtin that needs to be lowered. 1391 if (Value *V = EmitTargetBuiltinExpr(BuiltinID, E)) 1392 return RValue::get(V); 1393 1394 ErrorUnsupported(E, "builtin function"); 1395 1396 // Unknown builtin, for now just dump it out and return undef. 1397 if (hasAggregateLLVMType(E->getType())) 1398 return RValue::getAggregate(CreateMemTemp(E->getType())); 1399 return RValue::get(llvm::UndefValue::get(ConvertType(E->getType()))); 1400 } 1401 1402 Value *CodeGenFunction::EmitTargetBuiltinExpr(unsigned BuiltinID, 1403 const CallExpr *E) { 1404 switch (Target.getTriple().getArch()) { 1405 case llvm::Triple::arm: 1406 case llvm::Triple::thumb: 1407 return EmitARMBuiltinExpr(BuiltinID, E); 1408 case llvm::Triple::x86: 1409 case llvm::Triple::x86_64: 1410 return EmitX86BuiltinExpr(BuiltinID, E); 1411 case llvm::Triple::ppc: 1412 case llvm::Triple::ppc64: 1413 return EmitPPCBuiltinExpr(BuiltinID, E); 1414 default: 1415 return 0; 1416 } 1417 } 1418 1419 static llvm::VectorType *GetNeonType(CodeGenFunction *CGF, 1420 NeonTypeFlags TypeFlags) { 1421 int IsQuad = TypeFlags.isQuad(); 1422 switch (TypeFlags.getEltType()) { 1423 case NeonTypeFlags::Int8: 1424 case NeonTypeFlags::Poly8: 1425 return llvm::VectorType::get(CGF->Int8Ty, 8 << IsQuad); 1426 case NeonTypeFlags::Int16: 1427 case NeonTypeFlags::Poly16: 1428 case NeonTypeFlags::Float16: 1429 return llvm::VectorType::get(CGF->Int16Ty, 4 << IsQuad); 1430 case NeonTypeFlags::Int32: 1431 return llvm::VectorType::get(CGF->Int32Ty, 2 << IsQuad); 1432 case NeonTypeFlags::Int64: 1433 return llvm::VectorType::get(CGF->Int64Ty, 1 << IsQuad); 1434 case NeonTypeFlags::Float32: 1435 return llvm::VectorType::get(CGF->FloatTy, 2 << IsQuad); 1436 } 1437 llvm_unreachable("Invalid NeonTypeFlags element type!"); 1438 } 1439 1440 Value *CodeGenFunction::EmitNeonSplat(Value *V, Constant *C) { 1441 unsigned nElts = cast<llvm::VectorType>(V->getType())->getNumElements(); 1442 Value* SV = llvm::ConstantVector::getSplat(nElts, C); 1443 return Builder.CreateShuffleVector(V, V, SV, "lane"); 1444 } 1445 1446 Value *CodeGenFunction::EmitNeonCall(Function *F, SmallVectorImpl<Value*> &Ops, 1447 const char *name, 1448 unsigned shift, bool rightshift) { 1449 unsigned j = 0; 1450 for (Function::const_arg_iterator ai = F->arg_begin(), ae = F->arg_end(); 1451 ai != ae; ++ai, ++j) 1452 if (shift > 0 && shift == j) 1453 Ops[j] = EmitNeonShiftVector(Ops[j], ai->getType(), rightshift); 1454 else 1455 Ops[j] = Builder.CreateBitCast(Ops[j], ai->getType(), name); 1456 1457 return Builder.CreateCall(F, Ops, name); 1458 } 1459 1460 Value *CodeGenFunction::EmitNeonShiftVector(Value *V, llvm::Type *Ty, 1461 bool neg) { 1462 int SV = cast<ConstantInt>(V)->getSExtValue(); 1463 1464 llvm::VectorType *VTy = cast<llvm::VectorType>(Ty); 1465 llvm::Constant *C = ConstantInt::get(VTy->getElementType(), neg ? -SV : SV); 1466 return llvm::ConstantVector::getSplat(VTy->getNumElements(), C); 1467 } 1468 1469 /// GetPointeeAlignment - Given an expression with a pointer type, find the 1470 /// alignment of the type referenced by the pointer. Skip over implicit 1471 /// casts. 1472 unsigned CodeGenFunction::GetPointeeAlignment(const Expr *Addr) { 1473 unsigned Align = 1; 1474 // Check if the type is a pointer. The implicit cast operand might not be. 1475 while (Addr->getType()->isPointerType()) { 1476 QualType PtTy = Addr->getType()->getPointeeType(); 1477 1478 // Can't get alignment of incomplete types. 1479 if (!PtTy->isIncompleteType()) { 1480 unsigned NewA = getContext().getTypeAlignInChars(PtTy).getQuantity(); 1481 if (NewA > Align) 1482 Align = NewA; 1483 } 1484 1485 // If the address is an implicit cast, repeat with the cast operand. 1486 if (const ImplicitCastExpr *CastAddr = dyn_cast<ImplicitCastExpr>(Addr)) { 1487 Addr = CastAddr->getSubExpr(); 1488 continue; 1489 } 1490 break; 1491 } 1492 return Align; 1493 } 1494 1495 /// GetPointeeAlignmentValue - Given an expression with a pointer type, find 1496 /// the alignment of the type referenced by the pointer. Skip over implicit 1497 /// casts. Return the alignment as an llvm::Value. 1498 Value *CodeGenFunction::GetPointeeAlignmentValue(const Expr *Addr) { 1499 return llvm::ConstantInt::get(Int32Ty, GetPointeeAlignment(Addr)); 1500 } 1501 1502 Value *CodeGenFunction::EmitARMBuiltinExpr(unsigned BuiltinID, 1503 const CallExpr *E) { 1504 if (BuiltinID == ARM::BI__clear_cache) { 1505 const FunctionDecl *FD = E->getDirectCallee(); 1506 // Oddly people write this call without args on occasion and gcc accepts 1507 // it - it's also marked as varargs in the description file. 1508 SmallVector<Value*, 2> Ops; 1509 for (unsigned i = 0; i < E->getNumArgs(); i++) 1510 Ops.push_back(EmitScalarExpr(E->getArg(i))); 1511 llvm::Type *Ty = CGM.getTypes().ConvertType(FD->getType()); 1512 llvm::FunctionType *FTy = cast<llvm::FunctionType>(Ty); 1513 StringRef Name = FD->getName(); 1514 return Builder.CreateCall(CGM.CreateRuntimeFunction(FTy, Name), Ops); 1515 } 1516 1517 if (BuiltinID == ARM::BI__builtin_arm_ldrexd) { 1518 Function *F = CGM.getIntrinsic(Intrinsic::arm_ldrexd); 1519 1520 Value *LdPtr = EmitScalarExpr(E->getArg(0)); 1521 Value *Val = Builder.CreateCall(F, LdPtr, "ldrexd"); 1522 1523 Value *Val0 = Builder.CreateExtractValue(Val, 1); 1524 Value *Val1 = Builder.CreateExtractValue(Val, 0); 1525 Val0 = Builder.CreateZExt(Val0, Int64Ty); 1526 Val1 = Builder.CreateZExt(Val1, Int64Ty); 1527 1528 Value *ShiftCst = llvm::ConstantInt::get(Int64Ty, 32); 1529 Val = Builder.CreateShl(Val0, ShiftCst, "shl", true /* nuw */); 1530 return Builder.CreateOr(Val, Val1); 1531 } 1532 1533 if (BuiltinID == ARM::BI__builtin_arm_strexd) { 1534 Function *F = CGM.getIntrinsic(Intrinsic::arm_strexd); 1535 llvm::Type *STy = llvm::StructType::get(Int32Ty, Int32Ty, NULL); 1536 1537 Value *One = llvm::ConstantInt::get(Int32Ty, 1); 1538 Value *Tmp = Builder.CreateAlloca(Int64Ty, One); 1539 Value *Val = EmitScalarExpr(E->getArg(0)); 1540 Builder.CreateStore(Val, Tmp); 1541 1542 Value *LdPtr = Builder.CreateBitCast(Tmp,llvm::PointerType::getUnqual(STy)); 1543 Val = Builder.CreateLoad(LdPtr); 1544 1545 Value *Arg0 = Builder.CreateExtractValue(Val, 0); 1546 Value *Arg1 = Builder.CreateExtractValue(Val, 1); 1547 Value *StPtr = EmitScalarExpr(E->getArg(1)); 1548 return Builder.CreateCall3(F, Arg0, Arg1, StPtr, "strexd"); 1549 } 1550 1551 SmallVector<Value*, 4> Ops; 1552 for (unsigned i = 0, e = E->getNumArgs() - 1; i != e; i++) 1553 Ops.push_back(EmitScalarExpr(E->getArg(i))); 1554 1555 // vget_lane and vset_lane are not overloaded and do not have an extra 1556 // argument that specifies the vector type. 1557 switch (BuiltinID) { 1558 default: break; 1559 case ARM::BI__builtin_neon_vget_lane_i8: 1560 case ARM::BI__builtin_neon_vget_lane_i16: 1561 case ARM::BI__builtin_neon_vget_lane_i32: 1562 case ARM::BI__builtin_neon_vget_lane_i64: 1563 case ARM::BI__builtin_neon_vget_lane_f32: 1564 case ARM::BI__builtin_neon_vgetq_lane_i8: 1565 case ARM::BI__builtin_neon_vgetq_lane_i16: 1566 case ARM::BI__builtin_neon_vgetq_lane_i32: 1567 case ARM::BI__builtin_neon_vgetq_lane_i64: 1568 case ARM::BI__builtin_neon_vgetq_lane_f32: 1569 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 1570 "vget_lane"); 1571 case ARM::BI__builtin_neon_vset_lane_i8: 1572 case ARM::BI__builtin_neon_vset_lane_i16: 1573 case ARM::BI__builtin_neon_vset_lane_i32: 1574 case ARM::BI__builtin_neon_vset_lane_i64: 1575 case ARM::BI__builtin_neon_vset_lane_f32: 1576 case ARM::BI__builtin_neon_vsetq_lane_i8: 1577 case ARM::BI__builtin_neon_vsetq_lane_i16: 1578 case ARM::BI__builtin_neon_vsetq_lane_i32: 1579 case ARM::BI__builtin_neon_vsetq_lane_i64: 1580 case ARM::BI__builtin_neon_vsetq_lane_f32: 1581 Ops.push_back(EmitScalarExpr(E->getArg(2))); 1582 return Builder.CreateInsertElement(Ops[1], Ops[0], Ops[2], "vset_lane"); 1583 } 1584 1585 // Get the last argument, which specifies the vector type. 1586 llvm::APSInt Result; 1587 const Expr *Arg = E->getArg(E->getNumArgs()-1); 1588 if (!Arg->isIntegerConstantExpr(Result, getContext())) 1589 return 0; 1590 1591 if (BuiltinID == ARM::BI__builtin_arm_vcvtr_f || 1592 BuiltinID == ARM::BI__builtin_arm_vcvtr_d) { 1593 // Determine the overloaded type of this builtin. 1594 llvm::Type *Ty; 1595 if (BuiltinID == ARM::BI__builtin_arm_vcvtr_f) 1596 Ty = FloatTy; 1597 else 1598 Ty = DoubleTy; 1599 1600 // Determine whether this is an unsigned conversion or not. 1601 bool usgn = Result.getZExtValue() == 1; 1602 unsigned Int = usgn ? Intrinsic::arm_vcvtru : Intrinsic::arm_vcvtr; 1603 1604 // Call the appropriate intrinsic. 1605 Function *F = CGM.getIntrinsic(Int, Ty); 1606 return Builder.CreateCall(F, Ops, "vcvtr"); 1607 } 1608 1609 // Determine the type of this overloaded NEON intrinsic. 1610 NeonTypeFlags Type(Result.getZExtValue()); 1611 bool usgn = Type.isUnsigned(); 1612 bool quad = Type.isQuad(); 1613 bool rightShift = false; 1614 1615 llvm::VectorType *VTy = GetNeonType(this, Type); 1616 llvm::Type *Ty = VTy; 1617 if (!Ty) 1618 return 0; 1619 1620 unsigned Int; 1621 switch (BuiltinID) { 1622 default: return 0; 1623 case ARM::BI__builtin_neon_vabd_v: 1624 case ARM::BI__builtin_neon_vabdq_v: 1625 Int = usgn ? Intrinsic::arm_neon_vabdu : Intrinsic::arm_neon_vabds; 1626 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vabd"); 1627 case ARM::BI__builtin_neon_vabs_v: 1628 case ARM::BI__builtin_neon_vabsq_v: 1629 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vabs, Ty), 1630 Ops, "vabs"); 1631 case ARM::BI__builtin_neon_vaddhn_v: 1632 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vaddhn, Ty), 1633 Ops, "vaddhn"); 1634 case ARM::BI__builtin_neon_vcale_v: 1635 std::swap(Ops[0], Ops[1]); 1636 case ARM::BI__builtin_neon_vcage_v: { 1637 Function *F = CGM.getIntrinsic(Intrinsic::arm_neon_vacged); 1638 return EmitNeonCall(F, Ops, "vcage"); 1639 } 1640 case ARM::BI__builtin_neon_vcaleq_v: 1641 std::swap(Ops[0], Ops[1]); 1642 case ARM::BI__builtin_neon_vcageq_v: { 1643 Function *F = CGM.getIntrinsic(Intrinsic::arm_neon_vacgeq); 1644 return EmitNeonCall(F, Ops, "vcage"); 1645 } 1646 case ARM::BI__builtin_neon_vcalt_v: 1647 std::swap(Ops[0], Ops[1]); 1648 case ARM::BI__builtin_neon_vcagt_v: { 1649 Function *F = CGM.getIntrinsic(Intrinsic::arm_neon_vacgtd); 1650 return EmitNeonCall(F, Ops, "vcagt"); 1651 } 1652 case ARM::BI__builtin_neon_vcaltq_v: 1653 std::swap(Ops[0], Ops[1]); 1654 case ARM::BI__builtin_neon_vcagtq_v: { 1655 Function *F = CGM.getIntrinsic(Intrinsic::arm_neon_vacgtq); 1656 return EmitNeonCall(F, Ops, "vcagt"); 1657 } 1658 case ARM::BI__builtin_neon_vcls_v: 1659 case ARM::BI__builtin_neon_vclsq_v: { 1660 Function *F = CGM.getIntrinsic(Intrinsic::arm_neon_vcls, Ty); 1661 return EmitNeonCall(F, Ops, "vcls"); 1662 } 1663 case ARM::BI__builtin_neon_vclz_v: 1664 case ARM::BI__builtin_neon_vclzq_v: { 1665 // Generate target-independent intrinsic; also need to add second argument 1666 // for whether or not clz of zero is undefined; on ARM it isn't. 1667 Function *F = CGM.getIntrinsic(Intrinsic::ctlz, Ty); 1668 Ops.push_back(Builder.getInt1(Target.isCLZForZeroUndef())); 1669 return EmitNeonCall(F, Ops, "vclz"); 1670 } 1671 case ARM::BI__builtin_neon_vcnt_v: 1672 case ARM::BI__builtin_neon_vcntq_v: { 1673 // generate target-independent intrinsic 1674 Function *F = CGM.getIntrinsic(Intrinsic::ctpop, Ty); 1675 return EmitNeonCall(F, Ops, "vctpop"); 1676 } 1677 case ARM::BI__builtin_neon_vcvt_f16_v: { 1678 assert(Type.getEltType() == NeonTypeFlags::Float16 && !quad && 1679 "unexpected vcvt_f16_v builtin"); 1680 Function *F = CGM.getIntrinsic(Intrinsic::arm_neon_vcvtfp2hf); 1681 return EmitNeonCall(F, Ops, "vcvt"); 1682 } 1683 case ARM::BI__builtin_neon_vcvt_f32_f16: { 1684 assert(Type.getEltType() == NeonTypeFlags::Float16 && !quad && 1685 "unexpected vcvt_f32_f16 builtin"); 1686 Function *F = CGM.getIntrinsic(Intrinsic::arm_neon_vcvthf2fp); 1687 return EmitNeonCall(F, Ops, "vcvt"); 1688 } 1689 case ARM::BI__builtin_neon_vcvt_f32_v: 1690 case ARM::BI__builtin_neon_vcvtq_f32_v: 1691 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 1692 Ty = GetNeonType(this, NeonTypeFlags(NeonTypeFlags::Float32, false, quad)); 1693 return usgn ? Builder.CreateUIToFP(Ops[0], Ty, "vcvt") 1694 : Builder.CreateSIToFP(Ops[0], Ty, "vcvt"); 1695 case ARM::BI__builtin_neon_vcvt_s32_v: 1696 case ARM::BI__builtin_neon_vcvt_u32_v: 1697 case ARM::BI__builtin_neon_vcvtq_s32_v: 1698 case ARM::BI__builtin_neon_vcvtq_u32_v: { 1699 llvm::Type *FloatTy = 1700 GetNeonType(this, NeonTypeFlags(NeonTypeFlags::Float32, false, quad)); 1701 Ops[0] = Builder.CreateBitCast(Ops[0], FloatTy); 1702 return usgn ? Builder.CreateFPToUI(Ops[0], Ty, "vcvt") 1703 : Builder.CreateFPToSI(Ops[0], Ty, "vcvt"); 1704 } 1705 case ARM::BI__builtin_neon_vcvt_n_f32_v: 1706 case ARM::BI__builtin_neon_vcvtq_n_f32_v: { 1707 llvm::Type *FloatTy = 1708 GetNeonType(this, NeonTypeFlags(NeonTypeFlags::Float32, false, quad)); 1709 llvm::Type *Tys[2] = { FloatTy, Ty }; 1710 Int = usgn ? Intrinsic::arm_neon_vcvtfxu2fp 1711 : Intrinsic::arm_neon_vcvtfxs2fp; 1712 Function *F = CGM.getIntrinsic(Int, Tys); 1713 return EmitNeonCall(F, Ops, "vcvt_n"); 1714 } 1715 case ARM::BI__builtin_neon_vcvt_n_s32_v: 1716 case ARM::BI__builtin_neon_vcvt_n_u32_v: 1717 case ARM::BI__builtin_neon_vcvtq_n_s32_v: 1718 case ARM::BI__builtin_neon_vcvtq_n_u32_v: { 1719 llvm::Type *FloatTy = 1720 GetNeonType(this, NeonTypeFlags(NeonTypeFlags::Float32, false, quad)); 1721 llvm::Type *Tys[2] = { Ty, FloatTy }; 1722 Int = usgn ? Intrinsic::arm_neon_vcvtfp2fxu 1723 : Intrinsic::arm_neon_vcvtfp2fxs; 1724 Function *F = CGM.getIntrinsic(Int, Tys); 1725 return EmitNeonCall(F, Ops, "vcvt_n"); 1726 } 1727 case ARM::BI__builtin_neon_vext_v: 1728 case ARM::BI__builtin_neon_vextq_v: { 1729 int CV = cast<ConstantInt>(Ops[2])->getSExtValue(); 1730 SmallVector<Constant*, 16> Indices; 1731 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) 1732 Indices.push_back(ConstantInt::get(Int32Ty, i+CV)); 1733 1734 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 1735 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 1736 Value *SV = llvm::ConstantVector::get(Indices); 1737 return Builder.CreateShuffleVector(Ops[0], Ops[1], SV, "vext"); 1738 } 1739 case ARM::BI__builtin_neon_vhadd_v: 1740 case ARM::BI__builtin_neon_vhaddq_v: 1741 Int = usgn ? Intrinsic::arm_neon_vhaddu : Intrinsic::arm_neon_vhadds; 1742 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vhadd"); 1743 case ARM::BI__builtin_neon_vhsub_v: 1744 case ARM::BI__builtin_neon_vhsubq_v: 1745 Int = usgn ? Intrinsic::arm_neon_vhsubu : Intrinsic::arm_neon_vhsubs; 1746 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vhsub"); 1747 case ARM::BI__builtin_neon_vld1_v: 1748 case ARM::BI__builtin_neon_vld1q_v: 1749 Ops.push_back(GetPointeeAlignmentValue(E->getArg(0))); 1750 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vld1, Ty), 1751 Ops, "vld1"); 1752 case ARM::BI__builtin_neon_vld1q_lane_v: 1753 // Handle 64-bit integer elements as a special case. Use shuffles of 1754 // one-element vectors to avoid poor code for i64 in the backend. 1755 if (VTy->getElementType()->isIntegerTy(64)) { 1756 // Extract the other lane. 1757 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 1758 int Lane = cast<ConstantInt>(Ops[2])->getZExtValue(); 1759 Value *SV = llvm::ConstantVector::get(ConstantInt::get(Int32Ty, 1-Lane)); 1760 Ops[1] = Builder.CreateShuffleVector(Ops[1], Ops[1], SV); 1761 // Load the value as a one-element vector. 1762 Ty = llvm::VectorType::get(VTy->getElementType(), 1); 1763 Function *F = CGM.getIntrinsic(Intrinsic::arm_neon_vld1, Ty); 1764 Value *Ld = Builder.CreateCall2(F, Ops[0], 1765 GetPointeeAlignmentValue(E->getArg(0))); 1766 // Combine them. 1767 SmallVector<Constant*, 2> Indices; 1768 Indices.push_back(ConstantInt::get(Int32Ty, 1-Lane)); 1769 Indices.push_back(ConstantInt::get(Int32Ty, Lane)); 1770 SV = llvm::ConstantVector::get(Indices); 1771 return Builder.CreateShuffleVector(Ops[1], Ld, SV, "vld1q_lane"); 1772 } 1773 // fall through 1774 case ARM::BI__builtin_neon_vld1_lane_v: { 1775 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 1776 Ty = llvm::PointerType::getUnqual(VTy->getElementType()); 1777 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 1778 LoadInst *Ld = Builder.CreateLoad(Ops[0]); 1779 Value *Align = GetPointeeAlignmentValue(E->getArg(0)); 1780 Ld->setAlignment(cast<ConstantInt>(Align)->getZExtValue()); 1781 return Builder.CreateInsertElement(Ops[1], Ld, Ops[2], "vld1_lane"); 1782 } 1783 case ARM::BI__builtin_neon_vld1_dup_v: 1784 case ARM::BI__builtin_neon_vld1q_dup_v: { 1785 Value *V = UndefValue::get(Ty); 1786 Ty = llvm::PointerType::getUnqual(VTy->getElementType()); 1787 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 1788 LoadInst *Ld = Builder.CreateLoad(Ops[0]); 1789 Value *Align = GetPointeeAlignmentValue(E->getArg(0)); 1790 Ld->setAlignment(cast<ConstantInt>(Align)->getZExtValue()); 1791 llvm::Constant *CI = ConstantInt::get(Int32Ty, 0); 1792 Ops[0] = Builder.CreateInsertElement(V, Ld, CI); 1793 return EmitNeonSplat(Ops[0], CI); 1794 } 1795 case ARM::BI__builtin_neon_vld2_v: 1796 case ARM::BI__builtin_neon_vld2q_v: { 1797 Function *F = CGM.getIntrinsic(Intrinsic::arm_neon_vld2, Ty); 1798 Value *Align = GetPointeeAlignmentValue(E->getArg(1)); 1799 Ops[1] = Builder.CreateCall2(F, Ops[1], Align, "vld2"); 1800 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 1801 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 1802 return Builder.CreateStore(Ops[1], Ops[0]); 1803 } 1804 case ARM::BI__builtin_neon_vld3_v: 1805 case ARM::BI__builtin_neon_vld3q_v: { 1806 Function *F = CGM.getIntrinsic(Intrinsic::arm_neon_vld3, Ty); 1807 Value *Align = GetPointeeAlignmentValue(E->getArg(1)); 1808 Ops[1] = Builder.CreateCall2(F, Ops[1], Align, "vld3"); 1809 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 1810 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 1811 return Builder.CreateStore(Ops[1], Ops[0]); 1812 } 1813 case ARM::BI__builtin_neon_vld4_v: 1814 case ARM::BI__builtin_neon_vld4q_v: { 1815 Function *F = CGM.getIntrinsic(Intrinsic::arm_neon_vld4, Ty); 1816 Value *Align = GetPointeeAlignmentValue(E->getArg(1)); 1817 Ops[1] = Builder.CreateCall2(F, Ops[1], Align, "vld4"); 1818 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 1819 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 1820 return Builder.CreateStore(Ops[1], Ops[0]); 1821 } 1822 case ARM::BI__builtin_neon_vld2_lane_v: 1823 case ARM::BI__builtin_neon_vld2q_lane_v: { 1824 Function *F = CGM.getIntrinsic(Intrinsic::arm_neon_vld2lane, Ty); 1825 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 1826 Ops[3] = Builder.CreateBitCast(Ops[3], Ty); 1827 Ops.push_back(GetPointeeAlignmentValue(E->getArg(1))); 1828 Ops[1] = Builder.CreateCall(F, makeArrayRef(Ops).slice(1), "vld2_lane"); 1829 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 1830 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 1831 return Builder.CreateStore(Ops[1], Ops[0]); 1832 } 1833 case ARM::BI__builtin_neon_vld3_lane_v: 1834 case ARM::BI__builtin_neon_vld3q_lane_v: { 1835 Function *F = CGM.getIntrinsic(Intrinsic::arm_neon_vld3lane, Ty); 1836 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 1837 Ops[3] = Builder.CreateBitCast(Ops[3], Ty); 1838 Ops[4] = Builder.CreateBitCast(Ops[4], Ty); 1839 Ops.push_back(GetPointeeAlignmentValue(E->getArg(1))); 1840 Ops[1] = Builder.CreateCall(F, makeArrayRef(Ops).slice(1), "vld3_lane"); 1841 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 1842 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 1843 return Builder.CreateStore(Ops[1], Ops[0]); 1844 } 1845 case ARM::BI__builtin_neon_vld4_lane_v: 1846 case ARM::BI__builtin_neon_vld4q_lane_v: { 1847 Function *F = CGM.getIntrinsic(Intrinsic::arm_neon_vld4lane, Ty); 1848 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 1849 Ops[3] = Builder.CreateBitCast(Ops[3], Ty); 1850 Ops[4] = Builder.CreateBitCast(Ops[4], Ty); 1851 Ops[5] = Builder.CreateBitCast(Ops[5], Ty); 1852 Ops.push_back(GetPointeeAlignmentValue(E->getArg(1))); 1853 Ops[1] = Builder.CreateCall(F, makeArrayRef(Ops).slice(1), "vld3_lane"); 1854 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 1855 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 1856 return Builder.CreateStore(Ops[1], Ops[0]); 1857 } 1858 case ARM::BI__builtin_neon_vld2_dup_v: 1859 case ARM::BI__builtin_neon_vld3_dup_v: 1860 case ARM::BI__builtin_neon_vld4_dup_v: { 1861 // Handle 64-bit elements as a special-case. There is no "dup" needed. 1862 if (VTy->getElementType()->getPrimitiveSizeInBits() == 64) { 1863 switch (BuiltinID) { 1864 case ARM::BI__builtin_neon_vld2_dup_v: 1865 Int = Intrinsic::arm_neon_vld2; 1866 break; 1867 case ARM::BI__builtin_neon_vld3_dup_v: 1868 Int = Intrinsic::arm_neon_vld3; 1869 break; 1870 case ARM::BI__builtin_neon_vld4_dup_v: 1871 Int = Intrinsic::arm_neon_vld4; 1872 break; 1873 default: llvm_unreachable("unknown vld_dup intrinsic?"); 1874 } 1875 Function *F = CGM.getIntrinsic(Int, Ty); 1876 Value *Align = GetPointeeAlignmentValue(E->getArg(1)); 1877 Ops[1] = Builder.CreateCall2(F, Ops[1], Align, "vld_dup"); 1878 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 1879 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 1880 return Builder.CreateStore(Ops[1], Ops[0]); 1881 } 1882 switch (BuiltinID) { 1883 case ARM::BI__builtin_neon_vld2_dup_v: 1884 Int = Intrinsic::arm_neon_vld2lane; 1885 break; 1886 case ARM::BI__builtin_neon_vld3_dup_v: 1887 Int = Intrinsic::arm_neon_vld3lane; 1888 break; 1889 case ARM::BI__builtin_neon_vld4_dup_v: 1890 Int = Intrinsic::arm_neon_vld4lane; 1891 break; 1892 default: llvm_unreachable("unknown vld_dup intrinsic?"); 1893 } 1894 Function *F = CGM.getIntrinsic(Int, Ty); 1895 llvm::StructType *STy = cast<llvm::StructType>(F->getReturnType()); 1896 1897 SmallVector<Value*, 6> Args; 1898 Args.push_back(Ops[1]); 1899 Args.append(STy->getNumElements(), UndefValue::get(Ty)); 1900 1901 llvm::Constant *CI = ConstantInt::get(Int32Ty, 0); 1902 Args.push_back(CI); 1903 Args.push_back(GetPointeeAlignmentValue(E->getArg(1))); 1904 1905 Ops[1] = Builder.CreateCall(F, Args, "vld_dup"); 1906 // splat lane 0 to all elts in each vector of the result. 1907 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 1908 Value *Val = Builder.CreateExtractValue(Ops[1], i); 1909 Value *Elt = Builder.CreateBitCast(Val, Ty); 1910 Elt = EmitNeonSplat(Elt, CI); 1911 Elt = Builder.CreateBitCast(Elt, Val->getType()); 1912 Ops[1] = Builder.CreateInsertValue(Ops[1], Elt, i); 1913 } 1914 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 1915 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 1916 return Builder.CreateStore(Ops[1], Ops[0]); 1917 } 1918 case ARM::BI__builtin_neon_vmax_v: 1919 case ARM::BI__builtin_neon_vmaxq_v: 1920 Int = usgn ? Intrinsic::arm_neon_vmaxu : Intrinsic::arm_neon_vmaxs; 1921 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmax"); 1922 case ARM::BI__builtin_neon_vmin_v: 1923 case ARM::BI__builtin_neon_vminq_v: 1924 Int = usgn ? Intrinsic::arm_neon_vminu : Intrinsic::arm_neon_vmins; 1925 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmin"); 1926 case ARM::BI__builtin_neon_vmovl_v: { 1927 llvm::Type *DTy =llvm::VectorType::getTruncatedElementVectorType(VTy); 1928 Ops[0] = Builder.CreateBitCast(Ops[0], DTy); 1929 if (usgn) 1930 return Builder.CreateZExt(Ops[0], Ty, "vmovl"); 1931 return Builder.CreateSExt(Ops[0], Ty, "vmovl"); 1932 } 1933 case ARM::BI__builtin_neon_vmovn_v: { 1934 llvm::Type *QTy = llvm::VectorType::getExtendedElementVectorType(VTy); 1935 Ops[0] = Builder.CreateBitCast(Ops[0], QTy); 1936 return Builder.CreateTrunc(Ops[0], Ty, "vmovn"); 1937 } 1938 case ARM::BI__builtin_neon_vmul_v: 1939 case ARM::BI__builtin_neon_vmulq_v: 1940 assert(Type.isPoly() && "vmul builtin only supported for polynomial types"); 1941 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vmulp, Ty), 1942 Ops, "vmul"); 1943 case ARM::BI__builtin_neon_vmull_v: 1944 Int = usgn ? Intrinsic::arm_neon_vmullu : Intrinsic::arm_neon_vmulls; 1945 Int = Type.isPoly() ? (unsigned)Intrinsic::arm_neon_vmullp : Int; 1946 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmull"); 1947 case ARM::BI__builtin_neon_vpadal_v: 1948 case ARM::BI__builtin_neon_vpadalq_v: { 1949 Int = usgn ? Intrinsic::arm_neon_vpadalu : Intrinsic::arm_neon_vpadals; 1950 // The source operand type has twice as many elements of half the size. 1951 unsigned EltBits = VTy->getElementType()->getPrimitiveSizeInBits(); 1952 llvm::Type *EltTy = 1953 llvm::IntegerType::get(getLLVMContext(), EltBits / 2); 1954 llvm::Type *NarrowTy = 1955 llvm::VectorType::get(EltTy, VTy->getNumElements() * 2); 1956 llvm::Type *Tys[2] = { Ty, NarrowTy }; 1957 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vpadal"); 1958 } 1959 case ARM::BI__builtin_neon_vpadd_v: 1960 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vpadd, Ty), 1961 Ops, "vpadd"); 1962 case ARM::BI__builtin_neon_vpaddl_v: 1963 case ARM::BI__builtin_neon_vpaddlq_v: { 1964 Int = usgn ? Intrinsic::arm_neon_vpaddlu : Intrinsic::arm_neon_vpaddls; 1965 // The source operand type has twice as many elements of half the size. 1966 unsigned EltBits = VTy->getElementType()->getPrimitiveSizeInBits(); 1967 llvm::Type *EltTy = llvm::IntegerType::get(getLLVMContext(), EltBits / 2); 1968 llvm::Type *NarrowTy = 1969 llvm::VectorType::get(EltTy, VTy->getNumElements() * 2); 1970 llvm::Type *Tys[2] = { Ty, NarrowTy }; 1971 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vpaddl"); 1972 } 1973 case ARM::BI__builtin_neon_vpmax_v: 1974 Int = usgn ? Intrinsic::arm_neon_vpmaxu : Intrinsic::arm_neon_vpmaxs; 1975 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vpmax"); 1976 case ARM::BI__builtin_neon_vpmin_v: 1977 Int = usgn ? Intrinsic::arm_neon_vpminu : Intrinsic::arm_neon_vpmins; 1978 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vpmin"); 1979 case ARM::BI__builtin_neon_vqabs_v: 1980 case ARM::BI__builtin_neon_vqabsq_v: 1981 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vqabs, Ty), 1982 Ops, "vqabs"); 1983 case ARM::BI__builtin_neon_vqadd_v: 1984 case ARM::BI__builtin_neon_vqaddq_v: 1985 Int = usgn ? Intrinsic::arm_neon_vqaddu : Intrinsic::arm_neon_vqadds; 1986 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqadd"); 1987 case ARM::BI__builtin_neon_vqdmlal_v: 1988 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vqdmlal, Ty), 1989 Ops, "vqdmlal"); 1990 case ARM::BI__builtin_neon_vqdmlsl_v: 1991 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vqdmlsl, Ty), 1992 Ops, "vqdmlsl"); 1993 case ARM::BI__builtin_neon_vqdmulh_v: 1994 case ARM::BI__builtin_neon_vqdmulhq_v: 1995 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vqdmulh, Ty), 1996 Ops, "vqdmulh"); 1997 case ARM::BI__builtin_neon_vqdmull_v: 1998 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vqdmull, Ty), 1999 Ops, "vqdmull"); 2000 case ARM::BI__builtin_neon_vqmovn_v: 2001 Int = usgn ? Intrinsic::arm_neon_vqmovnu : Intrinsic::arm_neon_vqmovns; 2002 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqmovn"); 2003 case ARM::BI__builtin_neon_vqmovun_v: 2004 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vqmovnsu, Ty), 2005 Ops, "vqdmull"); 2006 case ARM::BI__builtin_neon_vqneg_v: 2007 case ARM::BI__builtin_neon_vqnegq_v: 2008 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vqneg, Ty), 2009 Ops, "vqneg"); 2010 case ARM::BI__builtin_neon_vqrdmulh_v: 2011 case ARM::BI__builtin_neon_vqrdmulhq_v: 2012 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vqrdmulh, Ty), 2013 Ops, "vqrdmulh"); 2014 case ARM::BI__builtin_neon_vqrshl_v: 2015 case ARM::BI__builtin_neon_vqrshlq_v: 2016 Int = usgn ? Intrinsic::arm_neon_vqrshiftu : Intrinsic::arm_neon_vqrshifts; 2017 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqrshl"); 2018 case ARM::BI__builtin_neon_vqrshrn_n_v: 2019 Int = usgn ? Intrinsic::arm_neon_vqrshiftnu : Intrinsic::arm_neon_vqrshiftns; 2020 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqrshrn_n", 2021 1, true); 2022 case ARM::BI__builtin_neon_vqrshrun_n_v: 2023 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vqrshiftnsu, Ty), 2024 Ops, "vqrshrun_n", 1, true); 2025 case ARM::BI__builtin_neon_vqshl_v: 2026 case ARM::BI__builtin_neon_vqshlq_v: 2027 Int = usgn ? Intrinsic::arm_neon_vqshiftu : Intrinsic::arm_neon_vqshifts; 2028 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqshl"); 2029 case ARM::BI__builtin_neon_vqshl_n_v: 2030 case ARM::BI__builtin_neon_vqshlq_n_v: 2031 Int = usgn ? Intrinsic::arm_neon_vqshiftu : Intrinsic::arm_neon_vqshifts; 2032 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqshl_n", 2033 1, false); 2034 case ARM::BI__builtin_neon_vqshlu_n_v: 2035 case ARM::BI__builtin_neon_vqshluq_n_v: 2036 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vqshiftsu, Ty), 2037 Ops, "vqshlu", 1, false); 2038 case ARM::BI__builtin_neon_vqshrn_n_v: 2039 Int = usgn ? Intrinsic::arm_neon_vqshiftnu : Intrinsic::arm_neon_vqshiftns; 2040 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqshrn_n", 2041 1, true); 2042 case ARM::BI__builtin_neon_vqshrun_n_v: 2043 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vqshiftnsu, Ty), 2044 Ops, "vqshrun_n", 1, true); 2045 case ARM::BI__builtin_neon_vqsub_v: 2046 case ARM::BI__builtin_neon_vqsubq_v: 2047 Int = usgn ? Intrinsic::arm_neon_vqsubu : Intrinsic::arm_neon_vqsubs; 2048 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqsub"); 2049 case ARM::BI__builtin_neon_vraddhn_v: 2050 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vraddhn, Ty), 2051 Ops, "vraddhn"); 2052 case ARM::BI__builtin_neon_vrecpe_v: 2053 case ARM::BI__builtin_neon_vrecpeq_v: 2054 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vrecpe, Ty), 2055 Ops, "vrecpe"); 2056 case ARM::BI__builtin_neon_vrecps_v: 2057 case ARM::BI__builtin_neon_vrecpsq_v: 2058 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vrecps, Ty), 2059 Ops, "vrecps"); 2060 case ARM::BI__builtin_neon_vrhadd_v: 2061 case ARM::BI__builtin_neon_vrhaddq_v: 2062 Int = usgn ? Intrinsic::arm_neon_vrhaddu : Intrinsic::arm_neon_vrhadds; 2063 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrhadd"); 2064 case ARM::BI__builtin_neon_vrshl_v: 2065 case ARM::BI__builtin_neon_vrshlq_v: 2066 Int = usgn ? Intrinsic::arm_neon_vrshiftu : Intrinsic::arm_neon_vrshifts; 2067 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrshl"); 2068 case ARM::BI__builtin_neon_vrshrn_n_v: 2069 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vrshiftn, Ty), 2070 Ops, "vrshrn_n", 1, true); 2071 case ARM::BI__builtin_neon_vrshr_n_v: 2072 case ARM::BI__builtin_neon_vrshrq_n_v: 2073 Int = usgn ? Intrinsic::arm_neon_vrshiftu : Intrinsic::arm_neon_vrshifts; 2074 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrshr_n", 1, true); 2075 case ARM::BI__builtin_neon_vrsqrte_v: 2076 case ARM::BI__builtin_neon_vrsqrteq_v: 2077 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vrsqrte, Ty), 2078 Ops, "vrsqrte"); 2079 case ARM::BI__builtin_neon_vrsqrts_v: 2080 case ARM::BI__builtin_neon_vrsqrtsq_v: 2081 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vrsqrts, Ty), 2082 Ops, "vrsqrts"); 2083 case ARM::BI__builtin_neon_vrsra_n_v: 2084 case ARM::BI__builtin_neon_vrsraq_n_v: 2085 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 2086 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 2087 Ops[2] = EmitNeonShiftVector(Ops[2], Ty, true); 2088 Int = usgn ? Intrinsic::arm_neon_vrshiftu : Intrinsic::arm_neon_vrshifts; 2089 Ops[1] = Builder.CreateCall2(CGM.getIntrinsic(Int, Ty), Ops[1], Ops[2]); 2090 return Builder.CreateAdd(Ops[0], Ops[1], "vrsra_n"); 2091 case ARM::BI__builtin_neon_vrsubhn_v: 2092 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vrsubhn, Ty), 2093 Ops, "vrsubhn"); 2094 case ARM::BI__builtin_neon_vshl_v: 2095 case ARM::BI__builtin_neon_vshlq_v: 2096 Int = usgn ? Intrinsic::arm_neon_vshiftu : Intrinsic::arm_neon_vshifts; 2097 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vshl"); 2098 case ARM::BI__builtin_neon_vshll_n_v: 2099 Int = usgn ? Intrinsic::arm_neon_vshiftlu : Intrinsic::arm_neon_vshiftls; 2100 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vshll", 1); 2101 case ARM::BI__builtin_neon_vshl_n_v: 2102 case ARM::BI__builtin_neon_vshlq_n_v: 2103 Ops[1] = EmitNeonShiftVector(Ops[1], Ty, false); 2104 return Builder.CreateShl(Builder.CreateBitCast(Ops[0],Ty), Ops[1], "vshl_n"); 2105 case ARM::BI__builtin_neon_vshrn_n_v: 2106 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vshiftn, Ty), 2107 Ops, "vshrn_n", 1, true); 2108 case ARM::BI__builtin_neon_vshr_n_v: 2109 case ARM::BI__builtin_neon_vshrq_n_v: 2110 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 2111 Ops[1] = EmitNeonShiftVector(Ops[1], Ty, false); 2112 if (usgn) 2113 return Builder.CreateLShr(Ops[0], Ops[1], "vshr_n"); 2114 else 2115 return Builder.CreateAShr(Ops[0], Ops[1], "vshr_n"); 2116 case ARM::BI__builtin_neon_vsri_n_v: 2117 case ARM::BI__builtin_neon_vsriq_n_v: 2118 rightShift = true; 2119 case ARM::BI__builtin_neon_vsli_n_v: 2120 case ARM::BI__builtin_neon_vsliq_n_v: 2121 Ops[2] = EmitNeonShiftVector(Ops[2], Ty, rightShift); 2122 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vshiftins, Ty), 2123 Ops, "vsli_n"); 2124 case ARM::BI__builtin_neon_vsra_n_v: 2125 case ARM::BI__builtin_neon_vsraq_n_v: 2126 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 2127 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 2128 Ops[2] = EmitNeonShiftVector(Ops[2], Ty, false); 2129 if (usgn) 2130 Ops[1] = Builder.CreateLShr(Ops[1], Ops[2], "vsra_n"); 2131 else 2132 Ops[1] = Builder.CreateAShr(Ops[1], Ops[2], "vsra_n"); 2133 return Builder.CreateAdd(Ops[0], Ops[1]); 2134 case ARM::BI__builtin_neon_vst1_v: 2135 case ARM::BI__builtin_neon_vst1q_v: 2136 Ops.push_back(GetPointeeAlignmentValue(E->getArg(0))); 2137 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vst1, Ty), 2138 Ops, ""); 2139 case ARM::BI__builtin_neon_vst1q_lane_v: 2140 // Handle 64-bit integer elements as a special case. Use a shuffle to get 2141 // a one-element vector and avoid poor code for i64 in the backend. 2142 if (VTy->getElementType()->isIntegerTy(64)) { 2143 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 2144 Value *SV = llvm::ConstantVector::get(cast<llvm::Constant>(Ops[2])); 2145 Ops[1] = Builder.CreateShuffleVector(Ops[1], Ops[1], SV); 2146 Ops[2] = GetPointeeAlignmentValue(E->getArg(0)); 2147 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::arm_neon_vst1, 2148 Ops[1]->getType()), Ops); 2149 } 2150 // fall through 2151 case ARM::BI__builtin_neon_vst1_lane_v: { 2152 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 2153 Ops[1] = Builder.CreateExtractElement(Ops[1], Ops[2]); 2154 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 2155 StoreInst *St = Builder.CreateStore(Ops[1], 2156 Builder.CreateBitCast(Ops[0], Ty)); 2157 Value *Align = GetPointeeAlignmentValue(E->getArg(0)); 2158 St->setAlignment(cast<ConstantInt>(Align)->getZExtValue()); 2159 return St; 2160 } 2161 case ARM::BI__builtin_neon_vst2_v: 2162 case ARM::BI__builtin_neon_vst2q_v: 2163 Ops.push_back(GetPointeeAlignmentValue(E->getArg(0))); 2164 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vst2, Ty), 2165 Ops, ""); 2166 case ARM::BI__builtin_neon_vst2_lane_v: 2167 case ARM::BI__builtin_neon_vst2q_lane_v: 2168 Ops.push_back(GetPointeeAlignmentValue(E->getArg(0))); 2169 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vst2lane, Ty), 2170 Ops, ""); 2171 case ARM::BI__builtin_neon_vst3_v: 2172 case ARM::BI__builtin_neon_vst3q_v: 2173 Ops.push_back(GetPointeeAlignmentValue(E->getArg(0))); 2174 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vst3, Ty), 2175 Ops, ""); 2176 case ARM::BI__builtin_neon_vst3_lane_v: 2177 case ARM::BI__builtin_neon_vst3q_lane_v: 2178 Ops.push_back(GetPointeeAlignmentValue(E->getArg(0))); 2179 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vst3lane, Ty), 2180 Ops, ""); 2181 case ARM::BI__builtin_neon_vst4_v: 2182 case ARM::BI__builtin_neon_vst4q_v: 2183 Ops.push_back(GetPointeeAlignmentValue(E->getArg(0))); 2184 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vst4, Ty), 2185 Ops, ""); 2186 case ARM::BI__builtin_neon_vst4_lane_v: 2187 case ARM::BI__builtin_neon_vst4q_lane_v: 2188 Ops.push_back(GetPointeeAlignmentValue(E->getArg(0))); 2189 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vst4lane, Ty), 2190 Ops, ""); 2191 case ARM::BI__builtin_neon_vsubhn_v: 2192 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vsubhn, Ty), 2193 Ops, "vsubhn"); 2194 case ARM::BI__builtin_neon_vtbl1_v: 2195 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbl1), 2196 Ops, "vtbl1"); 2197 case ARM::BI__builtin_neon_vtbl2_v: 2198 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbl2), 2199 Ops, "vtbl2"); 2200 case ARM::BI__builtin_neon_vtbl3_v: 2201 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbl3), 2202 Ops, "vtbl3"); 2203 case ARM::BI__builtin_neon_vtbl4_v: 2204 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbl4), 2205 Ops, "vtbl4"); 2206 case ARM::BI__builtin_neon_vtbx1_v: 2207 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbx1), 2208 Ops, "vtbx1"); 2209 case ARM::BI__builtin_neon_vtbx2_v: 2210 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbx2), 2211 Ops, "vtbx2"); 2212 case ARM::BI__builtin_neon_vtbx3_v: 2213 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbx3), 2214 Ops, "vtbx3"); 2215 case ARM::BI__builtin_neon_vtbx4_v: 2216 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbx4), 2217 Ops, "vtbx4"); 2218 case ARM::BI__builtin_neon_vtst_v: 2219 case ARM::BI__builtin_neon_vtstq_v: { 2220 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 2221 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 2222 Ops[0] = Builder.CreateAnd(Ops[0], Ops[1]); 2223 Ops[0] = Builder.CreateICmp(ICmpInst::ICMP_NE, Ops[0], 2224 ConstantAggregateZero::get(Ty)); 2225 return Builder.CreateSExt(Ops[0], Ty, "vtst"); 2226 } 2227 case ARM::BI__builtin_neon_vtrn_v: 2228 case ARM::BI__builtin_neon_vtrnq_v: { 2229 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty)); 2230 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 2231 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 2232 Value *SV = 0; 2233 2234 for (unsigned vi = 0; vi != 2; ++vi) { 2235 SmallVector<Constant*, 16> Indices; 2236 for (unsigned i = 0, e = VTy->getNumElements(); i != e; i += 2) { 2237 Indices.push_back(Builder.getInt32(i+vi)); 2238 Indices.push_back(Builder.getInt32(i+e+vi)); 2239 } 2240 Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ops[0], vi); 2241 SV = llvm::ConstantVector::get(Indices); 2242 SV = Builder.CreateShuffleVector(Ops[1], Ops[2], SV, "vtrn"); 2243 SV = Builder.CreateStore(SV, Addr); 2244 } 2245 return SV; 2246 } 2247 case ARM::BI__builtin_neon_vuzp_v: 2248 case ARM::BI__builtin_neon_vuzpq_v: { 2249 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty)); 2250 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 2251 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 2252 Value *SV = 0; 2253 2254 for (unsigned vi = 0; vi != 2; ++vi) { 2255 SmallVector<Constant*, 16> Indices; 2256 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) 2257 Indices.push_back(ConstantInt::get(Int32Ty, 2*i+vi)); 2258 2259 Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ops[0], vi); 2260 SV = llvm::ConstantVector::get(Indices); 2261 SV = Builder.CreateShuffleVector(Ops[1], Ops[2], SV, "vuzp"); 2262 SV = Builder.CreateStore(SV, Addr); 2263 } 2264 return SV; 2265 } 2266 case ARM::BI__builtin_neon_vzip_v: 2267 case ARM::BI__builtin_neon_vzipq_v: { 2268 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty)); 2269 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 2270 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 2271 Value *SV = 0; 2272 2273 for (unsigned vi = 0; vi != 2; ++vi) { 2274 SmallVector<Constant*, 16> Indices; 2275 for (unsigned i = 0, e = VTy->getNumElements(); i != e; i += 2) { 2276 Indices.push_back(ConstantInt::get(Int32Ty, (i + vi*e) >> 1)); 2277 Indices.push_back(ConstantInt::get(Int32Ty, ((i + vi*e) >> 1)+e)); 2278 } 2279 Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ops[0], vi); 2280 SV = llvm::ConstantVector::get(Indices); 2281 SV = Builder.CreateShuffleVector(Ops[1], Ops[2], SV, "vzip"); 2282 SV = Builder.CreateStore(SV, Addr); 2283 } 2284 return SV; 2285 } 2286 } 2287 } 2288 2289 llvm::Value *CodeGenFunction:: 2290 BuildVector(ArrayRef<llvm::Value*> Ops) { 2291 assert((Ops.size() & (Ops.size() - 1)) == 0 && 2292 "Not a power-of-two sized vector!"); 2293 bool AllConstants = true; 2294 for (unsigned i = 0, e = Ops.size(); i != e && AllConstants; ++i) 2295 AllConstants &= isa<Constant>(Ops[i]); 2296 2297 // If this is a constant vector, create a ConstantVector. 2298 if (AllConstants) { 2299 SmallVector<llvm::Constant*, 16> CstOps; 2300 for (unsigned i = 0, e = Ops.size(); i != e; ++i) 2301 CstOps.push_back(cast<Constant>(Ops[i])); 2302 return llvm::ConstantVector::get(CstOps); 2303 } 2304 2305 // Otherwise, insertelement the values to build the vector. 2306 Value *Result = 2307 llvm::UndefValue::get(llvm::VectorType::get(Ops[0]->getType(), Ops.size())); 2308 2309 for (unsigned i = 0, e = Ops.size(); i != e; ++i) 2310 Result = Builder.CreateInsertElement(Result, Ops[i], Builder.getInt32(i)); 2311 2312 return Result; 2313 } 2314 2315 Value *CodeGenFunction::EmitX86BuiltinExpr(unsigned BuiltinID, 2316 const CallExpr *E) { 2317 SmallVector<Value*, 4> Ops; 2318 2319 // Find out if any arguments are required to be integer constant expressions. 2320 unsigned ICEArguments = 0; 2321 ASTContext::GetBuiltinTypeError Error; 2322 getContext().GetBuiltinType(BuiltinID, Error, &ICEArguments); 2323 assert(Error == ASTContext::GE_None && "Should not codegen an error"); 2324 2325 for (unsigned i = 0, e = E->getNumArgs(); i != e; i++) { 2326 // If this is a normal argument, just emit it as a scalar. 2327 if ((ICEArguments & (1 << i)) == 0) { 2328 Ops.push_back(EmitScalarExpr(E->getArg(i))); 2329 continue; 2330 } 2331 2332 // If this is required to be a constant, constant fold it so that we know 2333 // that the generated intrinsic gets a ConstantInt. 2334 llvm::APSInt Result; 2335 bool IsConst = E->getArg(i)->isIntegerConstantExpr(Result, getContext()); 2336 assert(IsConst && "Constant arg isn't actually constant?"); (void)IsConst; 2337 Ops.push_back(llvm::ConstantInt::get(getLLVMContext(), Result)); 2338 } 2339 2340 switch (BuiltinID) { 2341 default: return 0; 2342 case X86::BI__builtin_ia32_vec_init_v8qi: 2343 case X86::BI__builtin_ia32_vec_init_v4hi: 2344 case X86::BI__builtin_ia32_vec_init_v2si: 2345 return Builder.CreateBitCast(BuildVector(Ops), 2346 llvm::Type::getX86_MMXTy(getLLVMContext())); 2347 case X86::BI__builtin_ia32_vec_ext_v2si: 2348 return Builder.CreateExtractElement(Ops[0], 2349 llvm::ConstantInt::get(Ops[1]->getType(), 0)); 2350 case X86::BI__builtin_ia32_ldmxcsr: { 2351 llvm::Type *PtrTy = Int8PtrTy; 2352 Value *One = llvm::ConstantInt::get(Int32Ty, 1); 2353 Value *Tmp = Builder.CreateAlloca(Int32Ty, One); 2354 Builder.CreateStore(Ops[0], Tmp); 2355 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::x86_sse_ldmxcsr), 2356 Builder.CreateBitCast(Tmp, PtrTy)); 2357 } 2358 case X86::BI__builtin_ia32_stmxcsr: { 2359 llvm::Type *PtrTy = Int8PtrTy; 2360 Value *One = llvm::ConstantInt::get(Int32Ty, 1); 2361 Value *Tmp = Builder.CreateAlloca(Int32Ty, One); 2362 Builder.CreateCall(CGM.getIntrinsic(Intrinsic::x86_sse_stmxcsr), 2363 Builder.CreateBitCast(Tmp, PtrTy)); 2364 return Builder.CreateLoad(Tmp, "stmxcsr"); 2365 } 2366 case X86::BI__builtin_ia32_storehps: 2367 case X86::BI__builtin_ia32_storelps: { 2368 llvm::Type *PtrTy = llvm::PointerType::getUnqual(Int64Ty); 2369 llvm::Type *VecTy = llvm::VectorType::get(Int64Ty, 2); 2370 2371 // cast val v2i64 2372 Ops[1] = Builder.CreateBitCast(Ops[1], VecTy, "cast"); 2373 2374 // extract (0, 1) 2375 unsigned Index = BuiltinID == X86::BI__builtin_ia32_storelps ? 0 : 1; 2376 llvm::Value *Idx = llvm::ConstantInt::get(Int32Ty, Index); 2377 Ops[1] = Builder.CreateExtractElement(Ops[1], Idx, "extract"); 2378 2379 // cast pointer to i64 & store 2380 Ops[0] = Builder.CreateBitCast(Ops[0], PtrTy); 2381 return Builder.CreateStore(Ops[1], Ops[0]); 2382 } 2383 case X86::BI__builtin_ia32_palignr: { 2384 unsigned shiftVal = cast<llvm::ConstantInt>(Ops[2])->getZExtValue(); 2385 2386 // If palignr is shifting the pair of input vectors less than 9 bytes, 2387 // emit a shuffle instruction. 2388 if (shiftVal <= 8) { 2389 SmallVector<llvm::Constant*, 8> Indices; 2390 for (unsigned i = 0; i != 8; ++i) 2391 Indices.push_back(llvm::ConstantInt::get(Int32Ty, shiftVal + i)); 2392 2393 Value* SV = llvm::ConstantVector::get(Indices); 2394 return Builder.CreateShuffleVector(Ops[1], Ops[0], SV, "palignr"); 2395 } 2396 2397 // If palignr is shifting the pair of input vectors more than 8 but less 2398 // than 16 bytes, emit a logical right shift of the destination. 2399 if (shiftVal < 16) { 2400 // MMX has these as 1 x i64 vectors for some odd optimization reasons. 2401 llvm::Type *VecTy = llvm::VectorType::get(Int64Ty, 1); 2402 2403 Ops[0] = Builder.CreateBitCast(Ops[0], VecTy, "cast"); 2404 Ops[1] = llvm::ConstantInt::get(VecTy, (shiftVal-8) * 8); 2405 2406 // create i32 constant 2407 llvm::Function *F = CGM.getIntrinsic(Intrinsic::x86_mmx_psrl_q); 2408 return Builder.CreateCall(F, makeArrayRef(&Ops[0], 2), "palignr"); 2409 } 2410 2411 // If palignr is shifting the pair of vectors more than 16 bytes, emit zero. 2412 return llvm::Constant::getNullValue(ConvertType(E->getType())); 2413 } 2414 case X86::BI__builtin_ia32_palignr128: { 2415 unsigned shiftVal = cast<llvm::ConstantInt>(Ops[2])->getZExtValue(); 2416 2417 // If palignr is shifting the pair of input vectors less than 17 bytes, 2418 // emit a shuffle instruction. 2419 if (shiftVal <= 16) { 2420 SmallVector<llvm::Constant*, 16> Indices; 2421 for (unsigned i = 0; i != 16; ++i) 2422 Indices.push_back(llvm::ConstantInt::get(Int32Ty, shiftVal + i)); 2423 2424 Value* SV = llvm::ConstantVector::get(Indices); 2425 return Builder.CreateShuffleVector(Ops[1], Ops[0], SV, "palignr"); 2426 } 2427 2428 // If palignr is shifting the pair of input vectors more than 16 but less 2429 // than 32 bytes, emit a logical right shift of the destination. 2430 if (shiftVal < 32) { 2431 llvm::Type *VecTy = llvm::VectorType::get(Int64Ty, 2); 2432 2433 Ops[0] = Builder.CreateBitCast(Ops[0], VecTy, "cast"); 2434 Ops[1] = llvm::ConstantInt::get(Int32Ty, (shiftVal-16) * 8); 2435 2436 // create i32 constant 2437 llvm::Function *F = CGM.getIntrinsic(Intrinsic::x86_sse2_psrl_dq); 2438 return Builder.CreateCall(F, makeArrayRef(&Ops[0], 2), "palignr"); 2439 } 2440 2441 // If palignr is shifting the pair of vectors more than 32 bytes, emit zero. 2442 return llvm::Constant::getNullValue(ConvertType(E->getType())); 2443 } 2444 case X86::BI__builtin_ia32_palignr256: { 2445 unsigned shiftVal = cast<llvm::ConstantInt>(Ops[2])->getZExtValue(); 2446 2447 // If palignr is shifting the pair of input vectors less than 17 bytes, 2448 // emit a shuffle instruction. 2449 if (shiftVal <= 16) { 2450 SmallVector<llvm::Constant*, 32> Indices; 2451 // 256-bit palignr operates on 128-bit lanes so we need to handle that 2452 for (unsigned l = 0; l != 2; ++l) { 2453 unsigned LaneStart = l * 16; 2454 unsigned LaneEnd = (l+1) * 16; 2455 for (unsigned i = 0; i != 16; ++i) { 2456 unsigned Idx = shiftVal + i + LaneStart; 2457 if (Idx >= LaneEnd) Idx += 16; // end of lane, switch operand 2458 Indices.push_back(llvm::ConstantInt::get(Int32Ty, Idx)); 2459 } 2460 } 2461 2462 Value* SV = llvm::ConstantVector::get(Indices); 2463 return Builder.CreateShuffleVector(Ops[1], Ops[0], SV, "palignr"); 2464 } 2465 2466 // If palignr is shifting the pair of input vectors more than 16 but less 2467 // than 32 bytes, emit a logical right shift of the destination. 2468 if (shiftVal < 32) { 2469 llvm::Type *VecTy = llvm::VectorType::get(Int64Ty, 4); 2470 2471 Ops[0] = Builder.CreateBitCast(Ops[0], VecTy, "cast"); 2472 Ops[1] = llvm::ConstantInt::get(Int32Ty, (shiftVal-16) * 8); 2473 2474 // create i32 constant 2475 llvm::Function *F = CGM.getIntrinsic(Intrinsic::x86_avx2_psrl_dq); 2476 return Builder.CreateCall(F, makeArrayRef(&Ops[0], 2), "palignr"); 2477 } 2478 2479 // If palignr is shifting the pair of vectors more than 32 bytes, emit zero. 2480 return llvm::Constant::getNullValue(ConvertType(E->getType())); 2481 } 2482 case X86::BI__builtin_ia32_movntps: 2483 case X86::BI__builtin_ia32_movntps256: 2484 case X86::BI__builtin_ia32_movntpd: 2485 case X86::BI__builtin_ia32_movntpd256: 2486 case X86::BI__builtin_ia32_movntdq: 2487 case X86::BI__builtin_ia32_movntdq256: 2488 case X86::BI__builtin_ia32_movnti: { 2489 llvm::MDNode *Node = llvm::MDNode::get(getLLVMContext(), 2490 Builder.getInt32(1)); 2491 2492 // Convert the type of the pointer to a pointer to the stored type. 2493 Value *BC = Builder.CreateBitCast(Ops[0], 2494 llvm::PointerType::getUnqual(Ops[1]->getType()), 2495 "cast"); 2496 StoreInst *SI = Builder.CreateStore(Ops[1], BC); 2497 SI->setMetadata(CGM.getModule().getMDKindID("nontemporal"), Node); 2498 SI->setAlignment(16); 2499 return SI; 2500 } 2501 // 3DNow! 2502 case X86::BI__builtin_ia32_pswapdsf: 2503 case X86::BI__builtin_ia32_pswapdsi: { 2504 const char *name = 0; 2505 Intrinsic::ID ID = Intrinsic::not_intrinsic; 2506 switch(BuiltinID) { 2507 default: llvm_unreachable("Unsupported intrinsic!"); 2508 case X86::BI__builtin_ia32_pswapdsf: 2509 case X86::BI__builtin_ia32_pswapdsi: 2510 name = "pswapd"; 2511 ID = Intrinsic::x86_3dnowa_pswapd; 2512 break; 2513 } 2514 llvm::Type *MMXTy = llvm::Type::getX86_MMXTy(getLLVMContext()); 2515 Ops[0] = Builder.CreateBitCast(Ops[0], MMXTy, "cast"); 2516 llvm::Function *F = CGM.getIntrinsic(ID); 2517 return Builder.CreateCall(F, Ops, name); 2518 } 2519 case X86::BI__builtin_ia32_rdrand16_step: 2520 case X86::BI__builtin_ia32_rdrand32_step: 2521 case X86::BI__builtin_ia32_rdrand64_step: { 2522 Intrinsic::ID ID; 2523 switch (BuiltinID) { 2524 default: llvm_unreachable("Unsupported intrinsic!"); 2525 case X86::BI__builtin_ia32_rdrand16_step: 2526 ID = Intrinsic::x86_rdrand_16; 2527 break; 2528 case X86::BI__builtin_ia32_rdrand32_step: 2529 ID = Intrinsic::x86_rdrand_32; 2530 break; 2531 case X86::BI__builtin_ia32_rdrand64_step: 2532 ID = Intrinsic::x86_rdrand_64; 2533 break; 2534 } 2535 2536 Value *Call = Builder.CreateCall(CGM.getIntrinsic(ID)); 2537 Builder.CreateStore(Builder.CreateExtractValue(Call, 0), Ops[0]); 2538 return Builder.CreateExtractValue(Call, 1); 2539 } 2540 } 2541 } 2542 2543 2544 Value *CodeGenFunction::EmitPPCBuiltinExpr(unsigned BuiltinID, 2545 const CallExpr *E) { 2546 SmallVector<Value*, 4> Ops; 2547 2548 for (unsigned i = 0, e = E->getNumArgs(); i != e; i++) 2549 Ops.push_back(EmitScalarExpr(E->getArg(i))); 2550 2551 Intrinsic::ID ID = Intrinsic::not_intrinsic; 2552 2553 switch (BuiltinID) { 2554 default: return 0; 2555 2556 // vec_ld, vec_lvsl, vec_lvsr 2557 case PPC::BI__builtin_altivec_lvx: 2558 case PPC::BI__builtin_altivec_lvxl: 2559 case PPC::BI__builtin_altivec_lvebx: 2560 case PPC::BI__builtin_altivec_lvehx: 2561 case PPC::BI__builtin_altivec_lvewx: 2562 case PPC::BI__builtin_altivec_lvsl: 2563 case PPC::BI__builtin_altivec_lvsr: 2564 { 2565 Ops[1] = Builder.CreateBitCast(Ops[1], Int8PtrTy); 2566 2567 Ops[0] = Builder.CreateGEP(Ops[1], Ops[0]); 2568 Ops.pop_back(); 2569 2570 switch (BuiltinID) { 2571 default: llvm_unreachable("Unsupported ld/lvsl/lvsr intrinsic!"); 2572 case PPC::BI__builtin_altivec_lvx: 2573 ID = Intrinsic::ppc_altivec_lvx; 2574 break; 2575 case PPC::BI__builtin_altivec_lvxl: 2576 ID = Intrinsic::ppc_altivec_lvxl; 2577 break; 2578 case PPC::BI__builtin_altivec_lvebx: 2579 ID = Intrinsic::ppc_altivec_lvebx; 2580 break; 2581 case PPC::BI__builtin_altivec_lvehx: 2582 ID = Intrinsic::ppc_altivec_lvehx; 2583 break; 2584 case PPC::BI__builtin_altivec_lvewx: 2585 ID = Intrinsic::ppc_altivec_lvewx; 2586 break; 2587 case PPC::BI__builtin_altivec_lvsl: 2588 ID = Intrinsic::ppc_altivec_lvsl; 2589 break; 2590 case PPC::BI__builtin_altivec_lvsr: 2591 ID = Intrinsic::ppc_altivec_lvsr; 2592 break; 2593 } 2594 llvm::Function *F = CGM.getIntrinsic(ID); 2595 return Builder.CreateCall(F, Ops, ""); 2596 } 2597 2598 // vec_st 2599 case PPC::BI__builtin_altivec_stvx: 2600 case PPC::BI__builtin_altivec_stvxl: 2601 case PPC::BI__builtin_altivec_stvebx: 2602 case PPC::BI__builtin_altivec_stvehx: 2603 case PPC::BI__builtin_altivec_stvewx: 2604 { 2605 Ops[2] = Builder.CreateBitCast(Ops[2], Int8PtrTy); 2606 Ops[1] = Builder.CreateGEP(Ops[2], Ops[1]); 2607 Ops.pop_back(); 2608 2609 switch (BuiltinID) { 2610 default: llvm_unreachable("Unsupported st intrinsic!"); 2611 case PPC::BI__builtin_altivec_stvx: 2612 ID = Intrinsic::ppc_altivec_stvx; 2613 break; 2614 case PPC::BI__builtin_altivec_stvxl: 2615 ID = Intrinsic::ppc_altivec_stvxl; 2616 break; 2617 case PPC::BI__builtin_altivec_stvebx: 2618 ID = Intrinsic::ppc_altivec_stvebx; 2619 break; 2620 case PPC::BI__builtin_altivec_stvehx: 2621 ID = Intrinsic::ppc_altivec_stvehx; 2622 break; 2623 case PPC::BI__builtin_altivec_stvewx: 2624 ID = Intrinsic::ppc_altivec_stvewx; 2625 break; 2626 } 2627 llvm::Function *F = CGM.getIntrinsic(ID); 2628 return Builder.CreateCall(F, Ops, ""); 2629 } 2630 } 2631 } 2632