1 //===---- CGBuiltin.cpp - Emit LLVM Code for builtins ---------------------===// 2 // 3 // The LLVM Compiler Infrastructure 4 // 5 // This file is distributed under the University of Illinois Open Source 6 // License. See LICENSE.TXT for details. 7 // 8 //===----------------------------------------------------------------------===// 9 // 10 // This contains code to emit Builtin calls as LLVM code. 11 // 12 //===----------------------------------------------------------------------===// 13 14 #include "CodeGenFunction.h" 15 #include "CGObjCRuntime.h" 16 #include "CodeGenModule.h" 17 #include "TargetInfo.h" 18 #include "clang/AST/ASTContext.h" 19 #include "clang/AST/Decl.h" 20 #include "clang/Basic/TargetBuiltins.h" 21 #include "clang/Basic/TargetInfo.h" 22 #include "llvm/IR/DataLayout.h" 23 #include "llvm/IR/Intrinsics.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 = DestPtr->getType()->getPointerAddressSpace(); 90 91 llvm::IntegerType *IntType = 92 llvm::IntegerType::get(CGF.getLLVMContext(), 93 CGF.getContext().getTypeSize(T)); 94 llvm::Type *IntPtrType = IntType->getPointerTo(AddrSpace); 95 96 llvm::Value *Args[2]; 97 Args[0] = CGF.Builder.CreateBitCast(DestPtr, IntPtrType); 98 Args[1] = CGF.EmitScalarExpr(E->getArg(1)); 99 llvm::Type *ValueType = Args[1]->getType(); 100 Args[1] = EmitToInt(CGF, Args[1], T, IntType); 101 102 llvm::Value *Result = 103 CGF.Builder.CreateAtomicRMW(Kind, Args[0], Args[1], 104 llvm::SequentiallyConsistent); 105 Result = EmitFromInt(CGF, Result, T, ValueType); 106 return RValue::get(Result); 107 } 108 109 /// Utility to insert an atomic instruction based Instrinsic::ID and 110 /// the expression node, where the return value is the result of the 111 /// operation. 112 static RValue EmitBinaryAtomicPost(CodeGenFunction &CGF, 113 llvm::AtomicRMWInst::BinOp Kind, 114 const CallExpr *E, 115 Instruction::BinaryOps Op) { 116 QualType T = E->getType(); 117 assert(E->getArg(0)->getType()->isPointerType()); 118 assert(CGF.getContext().hasSameUnqualifiedType(T, 119 E->getArg(0)->getType()->getPointeeType())); 120 assert(CGF.getContext().hasSameUnqualifiedType(T, E->getArg(1)->getType())); 121 122 llvm::Value *DestPtr = CGF.EmitScalarExpr(E->getArg(0)); 123 unsigned AddrSpace = DestPtr->getType()->getPointerAddressSpace(); 124 125 llvm::IntegerType *IntType = 126 llvm::IntegerType::get(CGF.getLLVMContext(), 127 CGF.getContext().getTypeSize(T)); 128 llvm::Type *IntPtrType = IntType->getPointerTo(AddrSpace); 129 130 llvm::Value *Args[2]; 131 Args[1] = CGF.EmitScalarExpr(E->getArg(1)); 132 llvm::Type *ValueType = Args[1]->getType(); 133 Args[1] = EmitToInt(CGF, Args[1], T, IntType); 134 Args[0] = CGF.Builder.CreateBitCast(DestPtr, IntPtrType); 135 136 llvm::Value *Result = 137 CGF.Builder.CreateAtomicRMW(Kind, Args[0], Args[1], 138 llvm::SequentiallyConsistent); 139 Result = CGF.Builder.CreateBinOp(Op, Result, Args[1]); 140 Result = EmitFromInt(CGF, Result, T, ValueType); 141 return RValue::get(Result); 142 } 143 144 /// EmitFAbs - Emit a call to fabs/fabsf/fabsl, depending on the type of ValTy, 145 /// which must be a scalar floating point type. 146 static Value *EmitFAbs(CodeGenFunction &CGF, Value *V, QualType ValTy) { 147 const BuiltinType *ValTyP = ValTy->getAs<BuiltinType>(); 148 assert(ValTyP && "isn't scalar fp type!"); 149 150 StringRef FnName; 151 switch (ValTyP->getKind()) { 152 default: llvm_unreachable("Isn't a scalar fp type!"); 153 case BuiltinType::Float: FnName = "fabsf"; break; 154 case BuiltinType::Double: FnName = "fabs"; break; 155 case BuiltinType::LongDouble: FnName = "fabsl"; break; 156 } 157 158 // The prototype is something that takes and returns whatever V's type is. 159 llvm::FunctionType *FT = llvm::FunctionType::get(V->getType(), V->getType(), 160 false); 161 llvm::Value *Fn = CGF.CGM.CreateRuntimeFunction(FT, FnName); 162 163 return CGF.Builder.CreateCall(Fn, V, "abs"); 164 } 165 166 static RValue emitLibraryCall(CodeGenFunction &CGF, const FunctionDecl *Fn, 167 const CallExpr *E, llvm::Value *calleeValue) { 168 return CGF.EmitCall(E->getCallee()->getType(), calleeValue, 169 ReturnValueSlot(), E->arg_begin(), E->arg_end(), Fn); 170 } 171 172 RValue CodeGenFunction::EmitBuiltinExpr(const FunctionDecl *FD, 173 unsigned BuiltinID, const CallExpr *E) { 174 // See if we can constant fold this builtin. If so, don't emit it at all. 175 Expr::EvalResult Result; 176 if (E->EvaluateAsRValue(Result, CGM.getContext()) && 177 !Result.hasSideEffects()) { 178 if (Result.Val.isInt()) 179 return RValue::get(llvm::ConstantInt::get(getLLVMContext(), 180 Result.Val.getInt())); 181 if (Result.Val.isFloat()) 182 return RValue::get(llvm::ConstantFP::get(getLLVMContext(), 183 Result.Val.getFloat())); 184 } 185 186 switch (BuiltinID) { 187 default: break; // Handle intrinsics and libm functions below. 188 case Builtin::BI__builtin___CFStringMakeConstantString: 189 case Builtin::BI__builtin___NSStringMakeConstantString: 190 return RValue::get(CGM.EmitConstantExpr(E, E->getType(), 0)); 191 case Builtin::BI__builtin_stdarg_start: 192 case Builtin::BI__builtin_va_start: 193 case Builtin::BI__builtin_va_end: { 194 Value *ArgValue = EmitVAListRef(E->getArg(0)); 195 llvm::Type *DestType = Int8PtrTy; 196 if (ArgValue->getType() != DestType) 197 ArgValue = Builder.CreateBitCast(ArgValue, DestType, 198 ArgValue->getName().data()); 199 200 Intrinsic::ID inst = (BuiltinID == Builtin::BI__builtin_va_end) ? 201 Intrinsic::vaend : Intrinsic::vastart; 202 return RValue::get(Builder.CreateCall(CGM.getIntrinsic(inst), ArgValue)); 203 } 204 case Builtin::BI__builtin_va_copy: { 205 Value *DstPtr = EmitVAListRef(E->getArg(0)); 206 Value *SrcPtr = EmitVAListRef(E->getArg(1)); 207 208 llvm::Type *Type = Int8PtrTy; 209 210 DstPtr = Builder.CreateBitCast(DstPtr, Type); 211 SrcPtr = Builder.CreateBitCast(SrcPtr, Type); 212 return RValue::get(Builder.CreateCall2(CGM.getIntrinsic(Intrinsic::vacopy), 213 DstPtr, SrcPtr)); 214 } 215 case Builtin::BI__builtin_abs: 216 case Builtin::BI__builtin_labs: 217 case Builtin::BI__builtin_llabs: { 218 Value *ArgValue = EmitScalarExpr(E->getArg(0)); 219 220 Value *NegOp = Builder.CreateNeg(ArgValue, "neg"); 221 Value *CmpResult = 222 Builder.CreateICmpSGE(ArgValue, 223 llvm::Constant::getNullValue(ArgValue->getType()), 224 "abscond"); 225 Value *Result = 226 Builder.CreateSelect(CmpResult, ArgValue, NegOp, "abs"); 227 228 return RValue::get(Result); 229 } 230 231 case Builtin::BI__builtin_conj: 232 case Builtin::BI__builtin_conjf: 233 case Builtin::BI__builtin_conjl: { 234 ComplexPairTy ComplexVal = EmitComplexExpr(E->getArg(0)); 235 Value *Real = ComplexVal.first; 236 Value *Imag = ComplexVal.second; 237 Value *Zero = 238 Imag->getType()->isFPOrFPVectorTy() 239 ? llvm::ConstantFP::getZeroValueForNegation(Imag->getType()) 240 : llvm::Constant::getNullValue(Imag->getType()); 241 242 Imag = Builder.CreateFSub(Zero, Imag, "sub"); 243 return RValue::getComplex(std::make_pair(Real, Imag)); 244 } 245 case Builtin::BI__builtin_creal: 246 case Builtin::BI__builtin_crealf: 247 case Builtin::BI__builtin_creall: 248 case Builtin::BIcreal: 249 case Builtin::BIcrealf: 250 case Builtin::BIcreall: { 251 ComplexPairTy ComplexVal = EmitComplexExpr(E->getArg(0)); 252 return RValue::get(ComplexVal.first); 253 } 254 255 case Builtin::BI__builtin_cimag: 256 case Builtin::BI__builtin_cimagf: 257 case Builtin::BI__builtin_cimagl: 258 case Builtin::BIcimag: 259 case Builtin::BIcimagf: 260 case Builtin::BIcimagl: { 261 ComplexPairTy ComplexVal = EmitComplexExpr(E->getArg(0)); 262 return RValue::get(ComplexVal.second); 263 } 264 265 case Builtin::BI__builtin_ctzs: 266 case Builtin::BI__builtin_ctz: 267 case Builtin::BI__builtin_ctzl: 268 case Builtin::BI__builtin_ctzll: { 269 Value *ArgValue = EmitScalarExpr(E->getArg(0)); 270 271 llvm::Type *ArgType = ArgValue->getType(); 272 Value *F = CGM.getIntrinsic(Intrinsic::cttz, ArgType); 273 274 llvm::Type *ResultType = ConvertType(E->getType()); 275 Value *ZeroUndef = Builder.getInt1(Target.isCLZForZeroUndef()); 276 Value *Result = Builder.CreateCall2(F, ArgValue, ZeroUndef); 277 if (Result->getType() != ResultType) 278 Result = Builder.CreateIntCast(Result, ResultType, /*isSigned*/true, 279 "cast"); 280 return RValue::get(Result); 281 } 282 case Builtin::BI__builtin_clzs: 283 case Builtin::BI__builtin_clz: 284 case Builtin::BI__builtin_clzl: 285 case Builtin::BI__builtin_clzll: { 286 Value *ArgValue = EmitScalarExpr(E->getArg(0)); 287 288 llvm::Type *ArgType = ArgValue->getType(); 289 Value *F = CGM.getIntrinsic(Intrinsic::ctlz, ArgType); 290 291 llvm::Type *ResultType = ConvertType(E->getType()); 292 Value *ZeroUndef = Builder.getInt1(Target.isCLZForZeroUndef()); 293 Value *Result = Builder.CreateCall2(F, ArgValue, ZeroUndef); 294 if (Result->getType() != ResultType) 295 Result = Builder.CreateIntCast(Result, ResultType, /*isSigned*/true, 296 "cast"); 297 return RValue::get(Result); 298 } 299 case Builtin::BI__builtin_ffs: 300 case Builtin::BI__builtin_ffsl: 301 case Builtin::BI__builtin_ffsll: { 302 // ffs(x) -> x ? cttz(x) + 1 : 0 303 Value *ArgValue = EmitScalarExpr(E->getArg(0)); 304 305 llvm::Type *ArgType = ArgValue->getType(); 306 Value *F = CGM.getIntrinsic(Intrinsic::cttz, ArgType); 307 308 llvm::Type *ResultType = ConvertType(E->getType()); 309 Value *Tmp = Builder.CreateAdd(Builder.CreateCall2(F, ArgValue, 310 Builder.getTrue()), 311 llvm::ConstantInt::get(ArgType, 1)); 312 Value *Zero = llvm::Constant::getNullValue(ArgType); 313 Value *IsZero = Builder.CreateICmpEQ(ArgValue, Zero, "iszero"); 314 Value *Result = Builder.CreateSelect(IsZero, Zero, Tmp, "ffs"); 315 if (Result->getType() != ResultType) 316 Result = Builder.CreateIntCast(Result, ResultType, /*isSigned*/true, 317 "cast"); 318 return RValue::get(Result); 319 } 320 case Builtin::BI__builtin_parity: 321 case Builtin::BI__builtin_parityl: 322 case Builtin::BI__builtin_parityll: { 323 // parity(x) -> ctpop(x) & 1 324 Value *ArgValue = EmitScalarExpr(E->getArg(0)); 325 326 llvm::Type *ArgType = ArgValue->getType(); 327 Value *F = CGM.getIntrinsic(Intrinsic::ctpop, ArgType); 328 329 llvm::Type *ResultType = ConvertType(E->getType()); 330 Value *Tmp = Builder.CreateCall(F, ArgValue); 331 Value *Result = Builder.CreateAnd(Tmp, llvm::ConstantInt::get(ArgType, 1)); 332 if (Result->getType() != ResultType) 333 Result = Builder.CreateIntCast(Result, ResultType, /*isSigned*/true, 334 "cast"); 335 return RValue::get(Result); 336 } 337 case Builtin::BI__builtin_popcount: 338 case Builtin::BI__builtin_popcountl: 339 case Builtin::BI__builtin_popcountll: { 340 Value *ArgValue = EmitScalarExpr(E->getArg(0)); 341 342 llvm::Type *ArgType = ArgValue->getType(); 343 Value *F = CGM.getIntrinsic(Intrinsic::ctpop, ArgType); 344 345 llvm::Type *ResultType = ConvertType(E->getType()); 346 Value *Result = Builder.CreateCall(F, ArgValue); 347 if (Result->getType() != ResultType) 348 Result = Builder.CreateIntCast(Result, ResultType, /*isSigned*/true, 349 "cast"); 350 return RValue::get(Result); 351 } 352 case Builtin::BI__builtin_expect: { 353 Value *ArgValue = EmitScalarExpr(E->getArg(0)); 354 llvm::Type *ArgType = ArgValue->getType(); 355 356 Value *FnExpect = CGM.getIntrinsic(Intrinsic::expect, ArgType); 357 Value *ExpectedValue = EmitScalarExpr(E->getArg(1)); 358 359 Value *Result = Builder.CreateCall2(FnExpect, ArgValue, ExpectedValue, 360 "expval"); 361 return RValue::get(Result); 362 } 363 case Builtin::BI__builtin_bswap16: 364 case Builtin::BI__builtin_bswap32: 365 case Builtin::BI__builtin_bswap64: { 366 Value *ArgValue = EmitScalarExpr(E->getArg(0)); 367 llvm::Type *ArgType = ArgValue->getType(); 368 Value *F = CGM.getIntrinsic(Intrinsic::bswap, ArgType); 369 return RValue::get(Builder.CreateCall(F, ArgValue)); 370 } 371 case Builtin::BI__builtin_object_size: { 372 // We rely on constant folding to deal with expressions with side effects. 373 assert(!E->getArg(0)->HasSideEffects(getContext()) && 374 "should have been constant folded"); 375 376 // We pass this builtin onto the optimizer so that it can 377 // figure out the object size in more complex cases. 378 llvm::Type *ResType = ConvertType(E->getType()); 379 380 // LLVM only supports 0 and 2, make sure that we pass along that 381 // as a boolean. 382 Value *Ty = EmitScalarExpr(E->getArg(1)); 383 ConstantInt *CI = dyn_cast<ConstantInt>(Ty); 384 assert(CI); 385 uint64_t val = CI->getZExtValue(); 386 CI = ConstantInt::get(Builder.getInt1Ty(), (val & 0x2) >> 1); 387 388 Value *F = CGM.getIntrinsic(Intrinsic::objectsize, ResType); 389 return RValue::get(Builder.CreateCall2(F, EmitScalarExpr(E->getArg(0)),CI)); 390 } 391 case Builtin::BI__builtin_prefetch: { 392 Value *Locality, *RW, *Address = EmitScalarExpr(E->getArg(0)); 393 // FIXME: Technically these constants should of type 'int', yes? 394 RW = (E->getNumArgs() > 1) ? EmitScalarExpr(E->getArg(1)) : 395 llvm::ConstantInt::get(Int32Ty, 0); 396 Locality = (E->getNumArgs() > 2) ? EmitScalarExpr(E->getArg(2)) : 397 llvm::ConstantInt::get(Int32Ty, 3); 398 Value *Data = llvm::ConstantInt::get(Int32Ty, 1); 399 Value *F = CGM.getIntrinsic(Intrinsic::prefetch); 400 return RValue::get(Builder.CreateCall4(F, Address, RW, Locality, Data)); 401 } 402 case Builtin::BI__builtin_readcyclecounter: { 403 Value *F = CGM.getIntrinsic(Intrinsic::readcyclecounter); 404 return RValue::get(Builder.CreateCall(F)); 405 } 406 case Builtin::BI__builtin_trap: { 407 Value *F = CGM.getIntrinsic(Intrinsic::trap); 408 return RValue::get(Builder.CreateCall(F)); 409 } 410 case Builtin::BI__debugbreak: { 411 Value *F = CGM.getIntrinsic(Intrinsic::debugtrap); 412 return RValue::get(Builder.CreateCall(F)); 413 } 414 case Builtin::BI__builtin_unreachable: { 415 if (getLangOpts().SanitizeUnreachable) 416 EmitCheck(Builder.getFalse(), "builtin_unreachable", 417 EmitCheckSourceLocation(E->getExprLoc()), 418 llvm::ArrayRef<llvm::Value *>(), CRK_Unrecoverable); 419 else 420 Builder.CreateUnreachable(); 421 422 // We do need to preserve an insertion point. 423 EmitBlock(createBasicBlock("unreachable.cont")); 424 425 return RValue::get(0); 426 } 427 428 case Builtin::BI__builtin_powi: 429 case Builtin::BI__builtin_powif: 430 case Builtin::BI__builtin_powil: { 431 Value *Base = EmitScalarExpr(E->getArg(0)); 432 Value *Exponent = EmitScalarExpr(E->getArg(1)); 433 llvm::Type *ArgType = Base->getType(); 434 Value *F = CGM.getIntrinsic(Intrinsic::powi, ArgType); 435 return RValue::get(Builder.CreateCall2(F, Base, Exponent)); 436 } 437 438 case Builtin::BI__builtin_isgreater: 439 case Builtin::BI__builtin_isgreaterequal: 440 case Builtin::BI__builtin_isless: 441 case Builtin::BI__builtin_islessequal: 442 case Builtin::BI__builtin_islessgreater: 443 case Builtin::BI__builtin_isunordered: { 444 // Ordered comparisons: we know the arguments to these are matching scalar 445 // floating point values. 446 Value *LHS = EmitScalarExpr(E->getArg(0)); 447 Value *RHS = EmitScalarExpr(E->getArg(1)); 448 449 switch (BuiltinID) { 450 default: llvm_unreachable("Unknown ordered comparison"); 451 case Builtin::BI__builtin_isgreater: 452 LHS = Builder.CreateFCmpOGT(LHS, RHS, "cmp"); 453 break; 454 case Builtin::BI__builtin_isgreaterequal: 455 LHS = Builder.CreateFCmpOGE(LHS, RHS, "cmp"); 456 break; 457 case Builtin::BI__builtin_isless: 458 LHS = Builder.CreateFCmpOLT(LHS, RHS, "cmp"); 459 break; 460 case Builtin::BI__builtin_islessequal: 461 LHS = Builder.CreateFCmpOLE(LHS, RHS, "cmp"); 462 break; 463 case Builtin::BI__builtin_islessgreater: 464 LHS = Builder.CreateFCmpONE(LHS, RHS, "cmp"); 465 break; 466 case Builtin::BI__builtin_isunordered: 467 LHS = Builder.CreateFCmpUNO(LHS, RHS, "cmp"); 468 break; 469 } 470 // ZExt bool to int type. 471 return RValue::get(Builder.CreateZExt(LHS, ConvertType(E->getType()))); 472 } 473 case Builtin::BI__builtin_isnan: { 474 Value *V = EmitScalarExpr(E->getArg(0)); 475 V = Builder.CreateFCmpUNO(V, V, "cmp"); 476 return RValue::get(Builder.CreateZExt(V, ConvertType(E->getType()))); 477 } 478 479 case Builtin::BI__builtin_isinf: { 480 // isinf(x) --> fabs(x) == infinity 481 Value *V = EmitScalarExpr(E->getArg(0)); 482 V = EmitFAbs(*this, V, E->getArg(0)->getType()); 483 484 V = Builder.CreateFCmpOEQ(V, ConstantFP::getInfinity(V->getType()),"isinf"); 485 return RValue::get(Builder.CreateZExt(V, ConvertType(E->getType()))); 486 } 487 488 // TODO: BI__builtin_isinf_sign 489 // isinf_sign(x) -> isinf(x) ? (signbit(x) ? -1 : 1) : 0 490 491 case Builtin::BI__builtin_isnormal: { 492 // isnormal(x) --> x == x && fabsf(x) < infinity && fabsf(x) >= float_min 493 Value *V = EmitScalarExpr(E->getArg(0)); 494 Value *Eq = Builder.CreateFCmpOEQ(V, V, "iseq"); 495 496 Value *Abs = EmitFAbs(*this, V, E->getArg(0)->getType()); 497 Value *IsLessThanInf = 498 Builder.CreateFCmpULT(Abs, ConstantFP::getInfinity(V->getType()),"isinf"); 499 APFloat Smallest = APFloat::getSmallestNormalized( 500 getContext().getFloatTypeSemantics(E->getArg(0)->getType())); 501 Value *IsNormal = 502 Builder.CreateFCmpUGE(Abs, ConstantFP::get(V->getContext(), Smallest), 503 "isnormal"); 504 V = Builder.CreateAnd(Eq, IsLessThanInf, "and"); 505 V = Builder.CreateAnd(V, IsNormal, "and"); 506 return RValue::get(Builder.CreateZExt(V, ConvertType(E->getType()))); 507 } 508 509 case Builtin::BI__builtin_isfinite: { 510 // isfinite(x) --> x == x && fabs(x) != infinity; 511 Value *V = EmitScalarExpr(E->getArg(0)); 512 Value *Eq = Builder.CreateFCmpOEQ(V, V, "iseq"); 513 514 Value *Abs = EmitFAbs(*this, V, E->getArg(0)->getType()); 515 Value *IsNotInf = 516 Builder.CreateFCmpUNE(Abs, ConstantFP::getInfinity(V->getType()),"isinf"); 517 518 V = Builder.CreateAnd(Eq, IsNotInf, "and"); 519 return RValue::get(Builder.CreateZExt(V, ConvertType(E->getType()))); 520 } 521 522 case Builtin::BI__builtin_fpclassify: { 523 Value *V = EmitScalarExpr(E->getArg(5)); 524 llvm::Type *Ty = ConvertType(E->getArg(5)->getType()); 525 526 // Create Result 527 BasicBlock *Begin = Builder.GetInsertBlock(); 528 BasicBlock *End = createBasicBlock("fpclassify_end", this->CurFn); 529 Builder.SetInsertPoint(End); 530 PHINode *Result = 531 Builder.CreatePHI(ConvertType(E->getArg(0)->getType()), 4, 532 "fpclassify_result"); 533 534 // if (V==0) return FP_ZERO 535 Builder.SetInsertPoint(Begin); 536 Value *IsZero = Builder.CreateFCmpOEQ(V, Constant::getNullValue(Ty), 537 "iszero"); 538 Value *ZeroLiteral = EmitScalarExpr(E->getArg(4)); 539 BasicBlock *NotZero = createBasicBlock("fpclassify_not_zero", this->CurFn); 540 Builder.CreateCondBr(IsZero, End, NotZero); 541 Result->addIncoming(ZeroLiteral, Begin); 542 543 // if (V != V) return FP_NAN 544 Builder.SetInsertPoint(NotZero); 545 Value *IsNan = Builder.CreateFCmpUNO(V, V, "cmp"); 546 Value *NanLiteral = EmitScalarExpr(E->getArg(0)); 547 BasicBlock *NotNan = createBasicBlock("fpclassify_not_nan", this->CurFn); 548 Builder.CreateCondBr(IsNan, End, NotNan); 549 Result->addIncoming(NanLiteral, NotZero); 550 551 // if (fabs(V) == infinity) return FP_INFINITY 552 Builder.SetInsertPoint(NotNan); 553 Value *VAbs = EmitFAbs(*this, V, E->getArg(5)->getType()); 554 Value *IsInf = 555 Builder.CreateFCmpOEQ(VAbs, ConstantFP::getInfinity(V->getType()), 556 "isinf"); 557 Value *InfLiteral = EmitScalarExpr(E->getArg(1)); 558 BasicBlock *NotInf = createBasicBlock("fpclassify_not_inf", this->CurFn); 559 Builder.CreateCondBr(IsInf, End, NotInf); 560 Result->addIncoming(InfLiteral, NotNan); 561 562 // if (fabs(V) >= MIN_NORMAL) return FP_NORMAL else FP_SUBNORMAL 563 Builder.SetInsertPoint(NotInf); 564 APFloat Smallest = APFloat::getSmallestNormalized( 565 getContext().getFloatTypeSemantics(E->getArg(5)->getType())); 566 Value *IsNormal = 567 Builder.CreateFCmpUGE(VAbs, ConstantFP::get(V->getContext(), Smallest), 568 "isnormal"); 569 Value *NormalResult = 570 Builder.CreateSelect(IsNormal, EmitScalarExpr(E->getArg(2)), 571 EmitScalarExpr(E->getArg(3))); 572 Builder.CreateBr(End); 573 Result->addIncoming(NormalResult, NotInf); 574 575 // return Result 576 Builder.SetInsertPoint(End); 577 return RValue::get(Result); 578 } 579 580 case Builtin::BIalloca: 581 case Builtin::BI__builtin_alloca: { 582 Value *Size = EmitScalarExpr(E->getArg(0)); 583 return RValue::get(Builder.CreateAlloca(Builder.getInt8Ty(), Size)); 584 } 585 case Builtin::BIbzero: 586 case Builtin::BI__builtin_bzero: { 587 std::pair<llvm::Value*, unsigned> Dest = 588 EmitPointerWithAlignment(E->getArg(0)); 589 Value *SizeVal = EmitScalarExpr(E->getArg(1)); 590 Builder.CreateMemSet(Dest.first, Builder.getInt8(0), SizeVal, 591 Dest.second, false); 592 return RValue::get(Dest.first); 593 } 594 case Builtin::BImemcpy: 595 case Builtin::BI__builtin_memcpy: { 596 std::pair<llvm::Value*, unsigned> Dest = 597 EmitPointerWithAlignment(E->getArg(0)); 598 std::pair<llvm::Value*, unsigned> Src = 599 EmitPointerWithAlignment(E->getArg(1)); 600 Value *SizeVal = EmitScalarExpr(E->getArg(2)); 601 unsigned Align = std::min(Dest.second, Src.second); 602 Builder.CreateMemCpy(Dest.first, Src.first, SizeVal, Align, false); 603 return RValue::get(Dest.first); 604 } 605 606 case Builtin::BI__builtin___memcpy_chk: { 607 // fold __builtin_memcpy_chk(x, y, cst1, cst2) to memcpy iff cst1<=cst2. 608 llvm::APSInt Size, DstSize; 609 if (!E->getArg(2)->EvaluateAsInt(Size, CGM.getContext()) || 610 !E->getArg(3)->EvaluateAsInt(DstSize, CGM.getContext())) 611 break; 612 if (Size.ugt(DstSize)) 613 break; 614 std::pair<llvm::Value*, unsigned> Dest = 615 EmitPointerWithAlignment(E->getArg(0)); 616 std::pair<llvm::Value*, unsigned> Src = 617 EmitPointerWithAlignment(E->getArg(1)); 618 Value *SizeVal = llvm::ConstantInt::get(Builder.getContext(), Size); 619 unsigned Align = std::min(Dest.second, Src.second); 620 Builder.CreateMemCpy(Dest.first, Src.first, SizeVal, Align, false); 621 return RValue::get(Dest.first); 622 } 623 624 case Builtin::BI__builtin_objc_memmove_collectable: { 625 Value *Address = EmitScalarExpr(E->getArg(0)); 626 Value *SrcAddr = EmitScalarExpr(E->getArg(1)); 627 Value *SizeVal = EmitScalarExpr(E->getArg(2)); 628 CGM.getObjCRuntime().EmitGCMemmoveCollectable(*this, 629 Address, SrcAddr, SizeVal); 630 return RValue::get(Address); 631 } 632 633 case Builtin::BI__builtin___memmove_chk: { 634 // fold __builtin_memmove_chk(x, y, cst1, cst2) to memmove iff cst1<=cst2. 635 llvm::APSInt Size, DstSize; 636 if (!E->getArg(2)->EvaluateAsInt(Size, CGM.getContext()) || 637 !E->getArg(3)->EvaluateAsInt(DstSize, CGM.getContext())) 638 break; 639 if (Size.ugt(DstSize)) 640 break; 641 std::pair<llvm::Value*, unsigned> Dest = 642 EmitPointerWithAlignment(E->getArg(0)); 643 std::pair<llvm::Value*, unsigned> Src = 644 EmitPointerWithAlignment(E->getArg(1)); 645 Value *SizeVal = llvm::ConstantInt::get(Builder.getContext(), Size); 646 unsigned Align = std::min(Dest.second, Src.second); 647 Builder.CreateMemMove(Dest.first, Src.first, SizeVal, Align, false); 648 return RValue::get(Dest.first); 649 } 650 651 case Builtin::BImemmove: 652 case Builtin::BI__builtin_memmove: { 653 std::pair<llvm::Value*, unsigned> Dest = 654 EmitPointerWithAlignment(E->getArg(0)); 655 std::pair<llvm::Value*, unsigned> Src = 656 EmitPointerWithAlignment(E->getArg(1)); 657 Value *SizeVal = EmitScalarExpr(E->getArg(2)); 658 unsigned Align = std::min(Dest.second, Src.second); 659 Builder.CreateMemMove(Dest.first, Src.first, SizeVal, Align, false); 660 return RValue::get(Dest.first); 661 } 662 case Builtin::BImemset: 663 case Builtin::BI__builtin_memset: { 664 std::pair<llvm::Value*, unsigned> Dest = 665 EmitPointerWithAlignment(E->getArg(0)); 666 Value *ByteVal = Builder.CreateTrunc(EmitScalarExpr(E->getArg(1)), 667 Builder.getInt8Ty()); 668 Value *SizeVal = EmitScalarExpr(E->getArg(2)); 669 Builder.CreateMemSet(Dest.first, ByteVal, SizeVal, Dest.second, false); 670 return RValue::get(Dest.first); 671 } 672 case Builtin::BI__builtin___memset_chk: { 673 // fold __builtin_memset_chk(x, y, cst1, cst2) to memset iff cst1<=cst2. 674 llvm::APSInt Size, DstSize; 675 if (!E->getArg(2)->EvaluateAsInt(Size, CGM.getContext()) || 676 !E->getArg(3)->EvaluateAsInt(DstSize, CGM.getContext())) 677 break; 678 if (Size.ugt(DstSize)) 679 break; 680 std::pair<llvm::Value*, unsigned> Dest = 681 EmitPointerWithAlignment(E->getArg(0)); 682 Value *ByteVal = Builder.CreateTrunc(EmitScalarExpr(E->getArg(1)), 683 Builder.getInt8Ty()); 684 Value *SizeVal = llvm::ConstantInt::get(Builder.getContext(), Size); 685 Builder.CreateMemSet(Dest.first, ByteVal, SizeVal, Dest.second, false); 686 return RValue::get(Dest.first); 687 } 688 case Builtin::BI__builtin_dwarf_cfa: { 689 // The offset in bytes from the first argument to the CFA. 690 // 691 // Why on earth is this in the frontend? Is there any reason at 692 // all that the backend can't reasonably determine this while 693 // lowering llvm.eh.dwarf.cfa()? 694 // 695 // TODO: If there's a satisfactory reason, add a target hook for 696 // this instead of hard-coding 0, which is correct for most targets. 697 int32_t Offset = 0; 698 699 Value *F = CGM.getIntrinsic(Intrinsic::eh_dwarf_cfa); 700 return RValue::get(Builder.CreateCall(F, 701 llvm::ConstantInt::get(Int32Ty, Offset))); 702 } 703 case Builtin::BI__builtin_return_address: { 704 Value *Depth = EmitScalarExpr(E->getArg(0)); 705 Depth = Builder.CreateIntCast(Depth, Int32Ty, false); 706 Value *F = CGM.getIntrinsic(Intrinsic::returnaddress); 707 return RValue::get(Builder.CreateCall(F, Depth)); 708 } 709 case Builtin::BI__builtin_frame_address: { 710 Value *Depth = EmitScalarExpr(E->getArg(0)); 711 Depth = Builder.CreateIntCast(Depth, Int32Ty, false); 712 Value *F = CGM.getIntrinsic(Intrinsic::frameaddress); 713 return RValue::get(Builder.CreateCall(F, Depth)); 714 } 715 case Builtin::BI__builtin_extract_return_addr: { 716 Value *Address = EmitScalarExpr(E->getArg(0)); 717 Value *Result = getTargetHooks().decodeReturnAddress(*this, Address); 718 return RValue::get(Result); 719 } 720 case Builtin::BI__builtin_frob_return_addr: { 721 Value *Address = EmitScalarExpr(E->getArg(0)); 722 Value *Result = getTargetHooks().encodeReturnAddress(*this, Address); 723 return RValue::get(Result); 724 } 725 case Builtin::BI__builtin_dwarf_sp_column: { 726 llvm::IntegerType *Ty 727 = cast<llvm::IntegerType>(ConvertType(E->getType())); 728 int Column = getTargetHooks().getDwarfEHStackPointer(CGM); 729 if (Column == -1) { 730 CGM.ErrorUnsupported(E, "__builtin_dwarf_sp_column"); 731 return RValue::get(llvm::UndefValue::get(Ty)); 732 } 733 return RValue::get(llvm::ConstantInt::get(Ty, Column, true)); 734 } 735 case Builtin::BI__builtin_init_dwarf_reg_size_table: { 736 Value *Address = EmitScalarExpr(E->getArg(0)); 737 if (getTargetHooks().initDwarfEHRegSizeTable(*this, Address)) 738 CGM.ErrorUnsupported(E, "__builtin_init_dwarf_reg_size_table"); 739 return RValue::get(llvm::UndefValue::get(ConvertType(E->getType()))); 740 } 741 case Builtin::BI__builtin_eh_return: { 742 Value *Int = EmitScalarExpr(E->getArg(0)); 743 Value *Ptr = EmitScalarExpr(E->getArg(1)); 744 745 llvm::IntegerType *IntTy = cast<llvm::IntegerType>(Int->getType()); 746 assert((IntTy->getBitWidth() == 32 || IntTy->getBitWidth() == 64) && 747 "LLVM's __builtin_eh_return only supports 32- and 64-bit variants"); 748 Value *F = CGM.getIntrinsic(IntTy->getBitWidth() == 32 749 ? Intrinsic::eh_return_i32 750 : Intrinsic::eh_return_i64); 751 Builder.CreateCall2(F, Int, Ptr); 752 Builder.CreateUnreachable(); 753 754 // We do need to preserve an insertion point. 755 EmitBlock(createBasicBlock("builtin_eh_return.cont")); 756 757 return RValue::get(0); 758 } 759 case Builtin::BI__builtin_unwind_init: { 760 Value *F = CGM.getIntrinsic(Intrinsic::eh_unwind_init); 761 return RValue::get(Builder.CreateCall(F)); 762 } 763 case Builtin::BI__builtin_extend_pointer: { 764 // Extends a pointer to the size of an _Unwind_Word, which is 765 // uint64_t on all platforms. Generally this gets poked into a 766 // register and eventually used as an address, so if the 767 // addressing registers are wider than pointers and the platform 768 // doesn't implicitly ignore high-order bits when doing 769 // addressing, we need to make sure we zext / sext based on 770 // the platform's expectations. 771 // 772 // See: http://gcc.gnu.org/ml/gcc-bugs/2002-02/msg00237.html 773 774 // Cast the pointer to intptr_t. 775 Value *Ptr = EmitScalarExpr(E->getArg(0)); 776 Value *Result = Builder.CreatePtrToInt(Ptr, IntPtrTy, "extend.cast"); 777 778 // If that's 64 bits, we're done. 779 if (IntPtrTy->getBitWidth() == 64) 780 return RValue::get(Result); 781 782 // Otherwise, ask the codegen data what to do. 783 if (getTargetHooks().extendPointerWithSExt()) 784 return RValue::get(Builder.CreateSExt(Result, Int64Ty, "extend.sext")); 785 else 786 return RValue::get(Builder.CreateZExt(Result, Int64Ty, "extend.zext")); 787 } 788 case Builtin::BI__builtin_setjmp: { 789 // Buffer is a void**. 790 Value *Buf = EmitScalarExpr(E->getArg(0)); 791 792 // Store the frame pointer to the setjmp buffer. 793 Value *FrameAddr = 794 Builder.CreateCall(CGM.getIntrinsic(Intrinsic::frameaddress), 795 ConstantInt::get(Int32Ty, 0)); 796 Builder.CreateStore(FrameAddr, Buf); 797 798 // Store the stack pointer to the setjmp buffer. 799 Value *StackAddr = 800 Builder.CreateCall(CGM.getIntrinsic(Intrinsic::stacksave)); 801 Value *StackSaveSlot = 802 Builder.CreateGEP(Buf, ConstantInt::get(Int32Ty, 2)); 803 Builder.CreateStore(StackAddr, StackSaveSlot); 804 805 // Call LLVM's EH setjmp, which is lightweight. 806 Value *F = CGM.getIntrinsic(Intrinsic::eh_sjlj_setjmp); 807 Buf = Builder.CreateBitCast(Buf, Int8PtrTy); 808 return RValue::get(Builder.CreateCall(F, Buf)); 809 } 810 case Builtin::BI__builtin_longjmp: { 811 Value *Buf = EmitScalarExpr(E->getArg(0)); 812 Buf = Builder.CreateBitCast(Buf, Int8PtrTy); 813 814 // Call LLVM's EH longjmp, which is lightweight. 815 Builder.CreateCall(CGM.getIntrinsic(Intrinsic::eh_sjlj_longjmp), Buf); 816 817 // longjmp doesn't return; mark this as unreachable. 818 Builder.CreateUnreachable(); 819 820 // We do need to preserve an insertion point. 821 EmitBlock(createBasicBlock("longjmp.cont")); 822 823 return RValue::get(0); 824 } 825 case Builtin::BI__sync_fetch_and_add: 826 case Builtin::BI__sync_fetch_and_sub: 827 case Builtin::BI__sync_fetch_and_or: 828 case Builtin::BI__sync_fetch_and_and: 829 case Builtin::BI__sync_fetch_and_xor: 830 case Builtin::BI__sync_add_and_fetch: 831 case Builtin::BI__sync_sub_and_fetch: 832 case Builtin::BI__sync_and_and_fetch: 833 case Builtin::BI__sync_or_and_fetch: 834 case Builtin::BI__sync_xor_and_fetch: 835 case Builtin::BI__sync_val_compare_and_swap: 836 case Builtin::BI__sync_bool_compare_and_swap: 837 case Builtin::BI__sync_lock_test_and_set: 838 case Builtin::BI__sync_lock_release: 839 case Builtin::BI__sync_swap: 840 llvm_unreachable("Shouldn't make it through sema"); 841 case Builtin::BI__sync_fetch_and_add_1: 842 case Builtin::BI__sync_fetch_and_add_2: 843 case Builtin::BI__sync_fetch_and_add_4: 844 case Builtin::BI__sync_fetch_and_add_8: 845 case Builtin::BI__sync_fetch_and_add_16: 846 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Add, E); 847 case Builtin::BI__sync_fetch_and_sub_1: 848 case Builtin::BI__sync_fetch_and_sub_2: 849 case Builtin::BI__sync_fetch_and_sub_4: 850 case Builtin::BI__sync_fetch_and_sub_8: 851 case Builtin::BI__sync_fetch_and_sub_16: 852 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Sub, E); 853 case Builtin::BI__sync_fetch_and_or_1: 854 case Builtin::BI__sync_fetch_and_or_2: 855 case Builtin::BI__sync_fetch_and_or_4: 856 case Builtin::BI__sync_fetch_and_or_8: 857 case Builtin::BI__sync_fetch_and_or_16: 858 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Or, E); 859 case Builtin::BI__sync_fetch_and_and_1: 860 case Builtin::BI__sync_fetch_and_and_2: 861 case Builtin::BI__sync_fetch_and_and_4: 862 case Builtin::BI__sync_fetch_and_and_8: 863 case Builtin::BI__sync_fetch_and_and_16: 864 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::And, E); 865 case Builtin::BI__sync_fetch_and_xor_1: 866 case Builtin::BI__sync_fetch_and_xor_2: 867 case Builtin::BI__sync_fetch_and_xor_4: 868 case Builtin::BI__sync_fetch_and_xor_8: 869 case Builtin::BI__sync_fetch_and_xor_16: 870 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Xor, E); 871 872 // Clang extensions: not overloaded yet. 873 case Builtin::BI__sync_fetch_and_min: 874 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Min, E); 875 case Builtin::BI__sync_fetch_and_max: 876 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Max, E); 877 case Builtin::BI__sync_fetch_and_umin: 878 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::UMin, E); 879 case Builtin::BI__sync_fetch_and_umax: 880 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::UMax, E); 881 882 case Builtin::BI__sync_add_and_fetch_1: 883 case Builtin::BI__sync_add_and_fetch_2: 884 case Builtin::BI__sync_add_and_fetch_4: 885 case Builtin::BI__sync_add_and_fetch_8: 886 case Builtin::BI__sync_add_and_fetch_16: 887 return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::Add, E, 888 llvm::Instruction::Add); 889 case Builtin::BI__sync_sub_and_fetch_1: 890 case Builtin::BI__sync_sub_and_fetch_2: 891 case Builtin::BI__sync_sub_and_fetch_4: 892 case Builtin::BI__sync_sub_and_fetch_8: 893 case Builtin::BI__sync_sub_and_fetch_16: 894 return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::Sub, E, 895 llvm::Instruction::Sub); 896 case Builtin::BI__sync_and_and_fetch_1: 897 case Builtin::BI__sync_and_and_fetch_2: 898 case Builtin::BI__sync_and_and_fetch_4: 899 case Builtin::BI__sync_and_and_fetch_8: 900 case Builtin::BI__sync_and_and_fetch_16: 901 return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::And, E, 902 llvm::Instruction::And); 903 case Builtin::BI__sync_or_and_fetch_1: 904 case Builtin::BI__sync_or_and_fetch_2: 905 case Builtin::BI__sync_or_and_fetch_4: 906 case Builtin::BI__sync_or_and_fetch_8: 907 case Builtin::BI__sync_or_and_fetch_16: 908 return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::Or, E, 909 llvm::Instruction::Or); 910 case Builtin::BI__sync_xor_and_fetch_1: 911 case Builtin::BI__sync_xor_and_fetch_2: 912 case Builtin::BI__sync_xor_and_fetch_4: 913 case Builtin::BI__sync_xor_and_fetch_8: 914 case Builtin::BI__sync_xor_and_fetch_16: 915 return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::Xor, E, 916 llvm::Instruction::Xor); 917 918 case Builtin::BI__sync_val_compare_and_swap_1: 919 case Builtin::BI__sync_val_compare_and_swap_2: 920 case Builtin::BI__sync_val_compare_and_swap_4: 921 case Builtin::BI__sync_val_compare_and_swap_8: 922 case Builtin::BI__sync_val_compare_and_swap_16: { 923 QualType T = E->getType(); 924 llvm::Value *DestPtr = EmitScalarExpr(E->getArg(0)); 925 unsigned AddrSpace = DestPtr->getType()->getPointerAddressSpace(); 926 927 llvm::IntegerType *IntType = 928 llvm::IntegerType::get(getLLVMContext(), 929 getContext().getTypeSize(T)); 930 llvm::Type *IntPtrType = IntType->getPointerTo(AddrSpace); 931 932 Value *Args[3]; 933 Args[0] = Builder.CreateBitCast(DestPtr, IntPtrType); 934 Args[1] = EmitScalarExpr(E->getArg(1)); 935 llvm::Type *ValueType = Args[1]->getType(); 936 Args[1] = EmitToInt(*this, Args[1], T, IntType); 937 Args[2] = EmitToInt(*this, EmitScalarExpr(E->getArg(2)), T, IntType); 938 939 Value *Result = Builder.CreateAtomicCmpXchg(Args[0], Args[1], Args[2], 940 llvm::SequentiallyConsistent); 941 Result = EmitFromInt(*this, Result, T, ValueType); 942 return RValue::get(Result); 943 } 944 945 case Builtin::BI__sync_bool_compare_and_swap_1: 946 case Builtin::BI__sync_bool_compare_and_swap_2: 947 case Builtin::BI__sync_bool_compare_and_swap_4: 948 case Builtin::BI__sync_bool_compare_and_swap_8: 949 case Builtin::BI__sync_bool_compare_and_swap_16: { 950 QualType T = E->getArg(1)->getType(); 951 llvm::Value *DestPtr = EmitScalarExpr(E->getArg(0)); 952 unsigned AddrSpace = DestPtr->getType()->getPointerAddressSpace(); 953 954 llvm::IntegerType *IntType = 955 llvm::IntegerType::get(getLLVMContext(), 956 getContext().getTypeSize(T)); 957 llvm::Type *IntPtrType = IntType->getPointerTo(AddrSpace); 958 959 Value *Args[3]; 960 Args[0] = Builder.CreateBitCast(DestPtr, IntPtrType); 961 Args[1] = EmitToInt(*this, EmitScalarExpr(E->getArg(1)), T, IntType); 962 Args[2] = EmitToInt(*this, EmitScalarExpr(E->getArg(2)), T, IntType); 963 964 Value *OldVal = Args[1]; 965 Value *PrevVal = Builder.CreateAtomicCmpXchg(Args[0], Args[1], Args[2], 966 llvm::SequentiallyConsistent); 967 Value *Result = Builder.CreateICmpEQ(PrevVal, OldVal); 968 // zext bool to int. 969 Result = Builder.CreateZExt(Result, ConvertType(E->getType())); 970 return RValue::get(Result); 971 } 972 973 case Builtin::BI__sync_swap_1: 974 case Builtin::BI__sync_swap_2: 975 case Builtin::BI__sync_swap_4: 976 case Builtin::BI__sync_swap_8: 977 case Builtin::BI__sync_swap_16: 978 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Xchg, E); 979 980 case Builtin::BI__sync_lock_test_and_set_1: 981 case Builtin::BI__sync_lock_test_and_set_2: 982 case Builtin::BI__sync_lock_test_and_set_4: 983 case Builtin::BI__sync_lock_test_and_set_8: 984 case Builtin::BI__sync_lock_test_and_set_16: 985 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Xchg, E); 986 987 case Builtin::BI__sync_lock_release_1: 988 case Builtin::BI__sync_lock_release_2: 989 case Builtin::BI__sync_lock_release_4: 990 case Builtin::BI__sync_lock_release_8: 991 case Builtin::BI__sync_lock_release_16: { 992 Value *Ptr = EmitScalarExpr(E->getArg(0)); 993 QualType ElTy = E->getArg(0)->getType()->getPointeeType(); 994 CharUnits StoreSize = getContext().getTypeSizeInChars(ElTy); 995 llvm::Type *ITy = llvm::IntegerType::get(getLLVMContext(), 996 StoreSize.getQuantity() * 8); 997 Ptr = Builder.CreateBitCast(Ptr, ITy->getPointerTo()); 998 llvm::StoreInst *Store = 999 Builder.CreateStore(llvm::Constant::getNullValue(ITy), Ptr); 1000 Store->setAlignment(StoreSize.getQuantity()); 1001 Store->setAtomic(llvm::Release); 1002 return RValue::get(0); 1003 } 1004 1005 case Builtin::BI__sync_synchronize: { 1006 // We assume this is supposed to correspond to a C++0x-style 1007 // sequentially-consistent fence (i.e. this is only usable for 1008 // synchonization, not device I/O or anything like that). This intrinsic 1009 // is really badly designed in the sense that in theory, there isn't 1010 // any way to safely use it... but in practice, it mostly works 1011 // to use it with non-atomic loads and stores to get acquire/release 1012 // semantics. 1013 Builder.CreateFence(llvm::SequentiallyConsistent); 1014 return RValue::get(0); 1015 } 1016 1017 case Builtin::BI__c11_atomic_is_lock_free: 1018 case Builtin::BI__atomic_is_lock_free: { 1019 // Call "bool __atomic_is_lock_free(size_t size, void *ptr)". For the 1020 // __c11 builtin, ptr is 0 (indicating a properly-aligned object), since 1021 // _Atomic(T) is always properly-aligned. 1022 const char *LibCallName = "__atomic_is_lock_free"; 1023 CallArgList Args; 1024 Args.add(RValue::get(EmitScalarExpr(E->getArg(0))), 1025 getContext().getSizeType()); 1026 if (BuiltinID == Builtin::BI__atomic_is_lock_free) 1027 Args.add(RValue::get(EmitScalarExpr(E->getArg(1))), 1028 getContext().VoidPtrTy); 1029 else 1030 Args.add(RValue::get(llvm::Constant::getNullValue(VoidPtrTy)), 1031 getContext().VoidPtrTy); 1032 const CGFunctionInfo &FuncInfo = 1033 CGM.getTypes().arrangeFreeFunctionCall(E->getType(), Args, 1034 FunctionType::ExtInfo(), 1035 RequiredArgs::All); 1036 llvm::FunctionType *FTy = CGM.getTypes().GetFunctionType(FuncInfo); 1037 llvm::Constant *Func = CGM.CreateRuntimeFunction(FTy, LibCallName); 1038 return EmitCall(FuncInfo, Func, ReturnValueSlot(), Args); 1039 } 1040 1041 case Builtin::BI__atomic_test_and_set: { 1042 // Look at the argument type to determine whether this is a volatile 1043 // operation. The parameter type is always volatile. 1044 QualType PtrTy = E->getArg(0)->IgnoreImpCasts()->getType(); 1045 bool Volatile = 1046 PtrTy->castAs<PointerType>()->getPointeeType().isVolatileQualified(); 1047 1048 Value *Ptr = EmitScalarExpr(E->getArg(0)); 1049 unsigned AddrSpace = Ptr->getType()->getPointerAddressSpace(); 1050 Ptr = Builder.CreateBitCast(Ptr, Int8Ty->getPointerTo(AddrSpace)); 1051 Value *NewVal = Builder.getInt8(1); 1052 Value *Order = EmitScalarExpr(E->getArg(1)); 1053 if (isa<llvm::ConstantInt>(Order)) { 1054 int ord = cast<llvm::ConstantInt>(Order)->getZExtValue(); 1055 AtomicRMWInst *Result = 0; 1056 switch (ord) { 1057 case 0: // memory_order_relaxed 1058 default: // invalid order 1059 Result = Builder.CreateAtomicRMW(llvm::AtomicRMWInst::Xchg, 1060 Ptr, NewVal, 1061 llvm::Monotonic); 1062 break; 1063 case 1: // memory_order_consume 1064 case 2: // memory_order_acquire 1065 Result = Builder.CreateAtomicRMW(llvm::AtomicRMWInst::Xchg, 1066 Ptr, NewVal, 1067 llvm::Acquire); 1068 break; 1069 case 3: // memory_order_release 1070 Result = Builder.CreateAtomicRMW(llvm::AtomicRMWInst::Xchg, 1071 Ptr, NewVal, 1072 llvm::Release); 1073 break; 1074 case 4: // memory_order_acq_rel 1075 Result = Builder.CreateAtomicRMW(llvm::AtomicRMWInst::Xchg, 1076 Ptr, NewVal, 1077 llvm::AcquireRelease); 1078 break; 1079 case 5: // memory_order_seq_cst 1080 Result = Builder.CreateAtomicRMW(llvm::AtomicRMWInst::Xchg, 1081 Ptr, NewVal, 1082 llvm::SequentiallyConsistent); 1083 break; 1084 } 1085 Result->setVolatile(Volatile); 1086 return RValue::get(Builder.CreateIsNotNull(Result, "tobool")); 1087 } 1088 1089 llvm::BasicBlock *ContBB = createBasicBlock("atomic.continue", CurFn); 1090 1091 llvm::BasicBlock *BBs[5] = { 1092 createBasicBlock("monotonic", CurFn), 1093 createBasicBlock("acquire", CurFn), 1094 createBasicBlock("release", CurFn), 1095 createBasicBlock("acqrel", CurFn), 1096 createBasicBlock("seqcst", CurFn) 1097 }; 1098 llvm::AtomicOrdering Orders[5] = { 1099 llvm::Monotonic, llvm::Acquire, llvm::Release, 1100 llvm::AcquireRelease, llvm::SequentiallyConsistent 1101 }; 1102 1103 Order = Builder.CreateIntCast(Order, Builder.getInt32Ty(), false); 1104 llvm::SwitchInst *SI = Builder.CreateSwitch(Order, BBs[0]); 1105 1106 Builder.SetInsertPoint(ContBB); 1107 PHINode *Result = Builder.CreatePHI(Int8Ty, 5, "was_set"); 1108 1109 for (unsigned i = 0; i < 5; ++i) { 1110 Builder.SetInsertPoint(BBs[i]); 1111 AtomicRMWInst *RMW = Builder.CreateAtomicRMW(llvm::AtomicRMWInst::Xchg, 1112 Ptr, NewVal, Orders[i]); 1113 RMW->setVolatile(Volatile); 1114 Result->addIncoming(RMW, BBs[i]); 1115 Builder.CreateBr(ContBB); 1116 } 1117 1118 SI->addCase(Builder.getInt32(0), BBs[0]); 1119 SI->addCase(Builder.getInt32(1), BBs[1]); 1120 SI->addCase(Builder.getInt32(2), BBs[1]); 1121 SI->addCase(Builder.getInt32(3), BBs[2]); 1122 SI->addCase(Builder.getInt32(4), BBs[3]); 1123 SI->addCase(Builder.getInt32(5), BBs[4]); 1124 1125 Builder.SetInsertPoint(ContBB); 1126 return RValue::get(Builder.CreateIsNotNull(Result, "tobool")); 1127 } 1128 1129 case Builtin::BI__atomic_clear: { 1130 QualType PtrTy = E->getArg(0)->IgnoreImpCasts()->getType(); 1131 bool Volatile = 1132 PtrTy->castAs<PointerType>()->getPointeeType().isVolatileQualified(); 1133 1134 Value *Ptr = EmitScalarExpr(E->getArg(0)); 1135 unsigned AddrSpace = Ptr->getType()->getPointerAddressSpace(); 1136 Ptr = Builder.CreateBitCast(Ptr, Int8Ty->getPointerTo(AddrSpace)); 1137 Value *NewVal = Builder.getInt8(0); 1138 Value *Order = EmitScalarExpr(E->getArg(1)); 1139 if (isa<llvm::ConstantInt>(Order)) { 1140 int ord = cast<llvm::ConstantInt>(Order)->getZExtValue(); 1141 StoreInst *Store = Builder.CreateStore(NewVal, Ptr, Volatile); 1142 Store->setAlignment(1); 1143 switch (ord) { 1144 case 0: // memory_order_relaxed 1145 default: // invalid order 1146 Store->setOrdering(llvm::Monotonic); 1147 break; 1148 case 3: // memory_order_release 1149 Store->setOrdering(llvm::Release); 1150 break; 1151 case 5: // memory_order_seq_cst 1152 Store->setOrdering(llvm::SequentiallyConsistent); 1153 break; 1154 } 1155 return RValue::get(0); 1156 } 1157 1158 llvm::BasicBlock *ContBB = createBasicBlock("atomic.continue", CurFn); 1159 1160 llvm::BasicBlock *BBs[3] = { 1161 createBasicBlock("monotonic", CurFn), 1162 createBasicBlock("release", CurFn), 1163 createBasicBlock("seqcst", CurFn) 1164 }; 1165 llvm::AtomicOrdering Orders[3] = { 1166 llvm::Monotonic, llvm::Release, llvm::SequentiallyConsistent 1167 }; 1168 1169 Order = Builder.CreateIntCast(Order, Builder.getInt32Ty(), false); 1170 llvm::SwitchInst *SI = Builder.CreateSwitch(Order, BBs[0]); 1171 1172 for (unsigned i = 0; i < 3; ++i) { 1173 Builder.SetInsertPoint(BBs[i]); 1174 StoreInst *Store = Builder.CreateStore(NewVal, Ptr, Volatile); 1175 Store->setAlignment(1); 1176 Store->setOrdering(Orders[i]); 1177 Builder.CreateBr(ContBB); 1178 } 1179 1180 SI->addCase(Builder.getInt32(0), BBs[0]); 1181 SI->addCase(Builder.getInt32(3), BBs[1]); 1182 SI->addCase(Builder.getInt32(5), BBs[2]); 1183 1184 Builder.SetInsertPoint(ContBB); 1185 return RValue::get(0); 1186 } 1187 1188 case Builtin::BI__atomic_thread_fence: 1189 case Builtin::BI__atomic_signal_fence: 1190 case Builtin::BI__c11_atomic_thread_fence: 1191 case Builtin::BI__c11_atomic_signal_fence: { 1192 llvm::SynchronizationScope Scope; 1193 if (BuiltinID == Builtin::BI__atomic_signal_fence || 1194 BuiltinID == Builtin::BI__c11_atomic_signal_fence) 1195 Scope = llvm::SingleThread; 1196 else 1197 Scope = llvm::CrossThread; 1198 Value *Order = EmitScalarExpr(E->getArg(0)); 1199 if (isa<llvm::ConstantInt>(Order)) { 1200 int ord = cast<llvm::ConstantInt>(Order)->getZExtValue(); 1201 switch (ord) { 1202 case 0: // memory_order_relaxed 1203 default: // invalid order 1204 break; 1205 case 1: // memory_order_consume 1206 case 2: // memory_order_acquire 1207 Builder.CreateFence(llvm::Acquire, Scope); 1208 break; 1209 case 3: // memory_order_release 1210 Builder.CreateFence(llvm::Release, Scope); 1211 break; 1212 case 4: // memory_order_acq_rel 1213 Builder.CreateFence(llvm::AcquireRelease, Scope); 1214 break; 1215 case 5: // memory_order_seq_cst 1216 Builder.CreateFence(llvm::SequentiallyConsistent, Scope); 1217 break; 1218 } 1219 return RValue::get(0); 1220 } 1221 1222 llvm::BasicBlock *AcquireBB, *ReleaseBB, *AcqRelBB, *SeqCstBB; 1223 AcquireBB = createBasicBlock("acquire", CurFn); 1224 ReleaseBB = createBasicBlock("release", CurFn); 1225 AcqRelBB = createBasicBlock("acqrel", CurFn); 1226 SeqCstBB = createBasicBlock("seqcst", CurFn); 1227 llvm::BasicBlock *ContBB = createBasicBlock("atomic.continue", CurFn); 1228 1229 Order = Builder.CreateIntCast(Order, Builder.getInt32Ty(), false); 1230 llvm::SwitchInst *SI = Builder.CreateSwitch(Order, ContBB); 1231 1232 Builder.SetInsertPoint(AcquireBB); 1233 Builder.CreateFence(llvm::Acquire, Scope); 1234 Builder.CreateBr(ContBB); 1235 SI->addCase(Builder.getInt32(1), AcquireBB); 1236 SI->addCase(Builder.getInt32(2), AcquireBB); 1237 1238 Builder.SetInsertPoint(ReleaseBB); 1239 Builder.CreateFence(llvm::Release, Scope); 1240 Builder.CreateBr(ContBB); 1241 SI->addCase(Builder.getInt32(3), ReleaseBB); 1242 1243 Builder.SetInsertPoint(AcqRelBB); 1244 Builder.CreateFence(llvm::AcquireRelease, Scope); 1245 Builder.CreateBr(ContBB); 1246 SI->addCase(Builder.getInt32(4), AcqRelBB); 1247 1248 Builder.SetInsertPoint(SeqCstBB); 1249 Builder.CreateFence(llvm::SequentiallyConsistent, Scope); 1250 Builder.CreateBr(ContBB); 1251 SI->addCase(Builder.getInt32(5), SeqCstBB); 1252 1253 Builder.SetInsertPoint(ContBB); 1254 return RValue::get(0); 1255 } 1256 1257 // Library functions with special handling. 1258 case Builtin::BIsqrt: 1259 case Builtin::BIsqrtf: 1260 case Builtin::BIsqrtl: { 1261 // TODO: there is currently no set of optimizer flags 1262 // sufficient for us to rewrite sqrt to @llvm.sqrt. 1263 // -fmath-errno=0 is not good enough; we need finiteness. 1264 // We could probably precondition the call with an ult 1265 // against 0, but is that worth the complexity? 1266 break; 1267 } 1268 1269 case Builtin::BIpow: 1270 case Builtin::BIpowf: 1271 case Builtin::BIpowl: { 1272 // Rewrite sqrt to intrinsic if allowed. 1273 if (!FD->hasAttr<ConstAttr>()) 1274 break; 1275 Value *Base = EmitScalarExpr(E->getArg(0)); 1276 Value *Exponent = EmitScalarExpr(E->getArg(1)); 1277 llvm::Type *ArgType = Base->getType(); 1278 Value *F = CGM.getIntrinsic(Intrinsic::pow, ArgType); 1279 return RValue::get(Builder.CreateCall2(F, Base, Exponent)); 1280 } 1281 1282 case Builtin::BIfma: 1283 case Builtin::BIfmaf: 1284 case Builtin::BIfmal: 1285 case Builtin::BI__builtin_fma: 1286 case Builtin::BI__builtin_fmaf: 1287 case Builtin::BI__builtin_fmal: { 1288 // Rewrite fma to intrinsic. 1289 Value *FirstArg = EmitScalarExpr(E->getArg(0)); 1290 llvm::Type *ArgType = FirstArg->getType(); 1291 Value *F = CGM.getIntrinsic(Intrinsic::fma, ArgType); 1292 return RValue::get(Builder.CreateCall3(F, FirstArg, 1293 EmitScalarExpr(E->getArg(1)), 1294 EmitScalarExpr(E->getArg(2)))); 1295 } 1296 1297 case Builtin::BI__builtin_signbit: 1298 case Builtin::BI__builtin_signbitf: 1299 case Builtin::BI__builtin_signbitl: { 1300 LLVMContext &C = CGM.getLLVMContext(); 1301 1302 Value *Arg = EmitScalarExpr(E->getArg(0)); 1303 llvm::Type *ArgTy = Arg->getType(); 1304 if (ArgTy->isPPC_FP128Ty()) 1305 break; // FIXME: I'm not sure what the right implementation is here. 1306 int ArgWidth = ArgTy->getPrimitiveSizeInBits(); 1307 llvm::Type *ArgIntTy = llvm::IntegerType::get(C, ArgWidth); 1308 Value *BCArg = Builder.CreateBitCast(Arg, ArgIntTy); 1309 Value *ZeroCmp = llvm::Constant::getNullValue(ArgIntTy); 1310 Value *Result = Builder.CreateICmpSLT(BCArg, ZeroCmp); 1311 return RValue::get(Builder.CreateZExt(Result, ConvertType(E->getType()))); 1312 } 1313 case Builtin::BI__builtin_annotation: { 1314 llvm::Value *AnnVal = EmitScalarExpr(E->getArg(0)); 1315 llvm::Value *F = CGM.getIntrinsic(llvm::Intrinsic::annotation, 1316 AnnVal->getType()); 1317 1318 // Get the annotation string, go through casts. Sema requires this to be a 1319 // non-wide string literal, potentially casted, so the cast<> is safe. 1320 const Expr *AnnotationStrExpr = E->getArg(1)->IgnoreParenCasts(); 1321 llvm::StringRef Str = cast<StringLiteral>(AnnotationStrExpr)->getString(); 1322 return RValue::get(EmitAnnotationCall(F, AnnVal, Str, E->getExprLoc())); 1323 } 1324 case Builtin::BI__noop: 1325 return RValue::get(0); 1326 } 1327 1328 // If this is an alias for a lib function (e.g. __builtin_sin), emit 1329 // the call using the normal call path, but using the unmangled 1330 // version of the function name. 1331 if (getContext().BuiltinInfo.isLibFunction(BuiltinID)) 1332 return emitLibraryCall(*this, FD, E, 1333 CGM.getBuiltinLibFunction(FD, BuiltinID)); 1334 1335 // If this is a predefined lib function (e.g. malloc), emit the call 1336 // using exactly the normal call path. 1337 if (getContext().BuiltinInfo.isPredefinedLibFunction(BuiltinID)) 1338 return emitLibraryCall(*this, FD, E, EmitScalarExpr(E->getCallee())); 1339 1340 // See if we have a target specific intrinsic. 1341 const char *Name = getContext().BuiltinInfo.GetName(BuiltinID); 1342 Intrinsic::ID IntrinsicID = Intrinsic::not_intrinsic; 1343 if (const char *Prefix = 1344 llvm::Triple::getArchTypePrefix(Target.getTriple().getArch())) 1345 IntrinsicID = Intrinsic::getIntrinsicForGCCBuiltin(Prefix, Name); 1346 1347 if (IntrinsicID != Intrinsic::not_intrinsic) { 1348 SmallVector<Value*, 16> Args; 1349 1350 // Find out if any arguments are required to be integer constant 1351 // expressions. 1352 unsigned ICEArguments = 0; 1353 ASTContext::GetBuiltinTypeError Error; 1354 getContext().GetBuiltinType(BuiltinID, Error, &ICEArguments); 1355 assert(Error == ASTContext::GE_None && "Should not codegen an error"); 1356 1357 Function *F = CGM.getIntrinsic(IntrinsicID); 1358 llvm::FunctionType *FTy = F->getFunctionType(); 1359 1360 for (unsigned i = 0, e = E->getNumArgs(); i != e; ++i) { 1361 Value *ArgValue; 1362 // If this is a normal argument, just emit it as a scalar. 1363 if ((ICEArguments & (1 << i)) == 0) { 1364 ArgValue = EmitScalarExpr(E->getArg(i)); 1365 } else { 1366 // If this is required to be a constant, constant fold it so that we 1367 // know that the generated intrinsic gets a ConstantInt. 1368 llvm::APSInt Result; 1369 bool IsConst = E->getArg(i)->isIntegerConstantExpr(Result,getContext()); 1370 assert(IsConst && "Constant arg isn't actually constant?"); 1371 (void)IsConst; 1372 ArgValue = llvm::ConstantInt::get(getLLVMContext(), Result); 1373 } 1374 1375 // If the intrinsic arg type is different from the builtin arg type 1376 // we need to do a bit cast. 1377 llvm::Type *PTy = FTy->getParamType(i); 1378 if (PTy != ArgValue->getType()) { 1379 assert(PTy->canLosslesslyBitCastTo(FTy->getParamType(i)) && 1380 "Must be able to losslessly bit cast to param"); 1381 ArgValue = Builder.CreateBitCast(ArgValue, PTy); 1382 } 1383 1384 Args.push_back(ArgValue); 1385 } 1386 1387 Value *V = Builder.CreateCall(F, Args); 1388 QualType BuiltinRetType = E->getType(); 1389 1390 llvm::Type *RetTy = VoidTy; 1391 if (!BuiltinRetType->isVoidType()) 1392 RetTy = ConvertType(BuiltinRetType); 1393 1394 if (RetTy != V->getType()) { 1395 assert(V->getType()->canLosslesslyBitCastTo(RetTy) && 1396 "Must be able to losslessly bit cast result type"); 1397 V = Builder.CreateBitCast(V, RetTy); 1398 } 1399 1400 return RValue::get(V); 1401 } 1402 1403 // See if we have a target specific builtin that needs to be lowered. 1404 if (Value *V = EmitTargetBuiltinExpr(BuiltinID, E)) 1405 return RValue::get(V); 1406 1407 ErrorUnsupported(E, "builtin function"); 1408 1409 // Unknown builtin, for now just dump it out and return undef. 1410 if (hasAggregateLLVMType(E->getType())) 1411 return RValue::getAggregate(CreateMemTemp(E->getType())); 1412 return RValue::get(llvm::UndefValue::get(ConvertType(E->getType()))); 1413 } 1414 1415 Value *CodeGenFunction::EmitTargetBuiltinExpr(unsigned BuiltinID, 1416 const CallExpr *E) { 1417 switch (Target.getTriple().getArch()) { 1418 case llvm::Triple::arm: 1419 case llvm::Triple::thumb: 1420 return EmitARMBuiltinExpr(BuiltinID, E); 1421 case llvm::Triple::x86: 1422 case llvm::Triple::x86_64: 1423 return EmitX86BuiltinExpr(BuiltinID, E); 1424 case llvm::Triple::ppc: 1425 case llvm::Triple::ppc64: 1426 return EmitPPCBuiltinExpr(BuiltinID, E); 1427 default: 1428 return 0; 1429 } 1430 } 1431 1432 static llvm::VectorType *GetNeonType(CodeGenFunction *CGF, 1433 NeonTypeFlags TypeFlags) { 1434 int IsQuad = TypeFlags.isQuad(); 1435 switch (TypeFlags.getEltType()) { 1436 case NeonTypeFlags::Int8: 1437 case NeonTypeFlags::Poly8: 1438 return llvm::VectorType::get(CGF->Int8Ty, 8 << IsQuad); 1439 case NeonTypeFlags::Int16: 1440 case NeonTypeFlags::Poly16: 1441 case NeonTypeFlags::Float16: 1442 return llvm::VectorType::get(CGF->Int16Ty, 4 << IsQuad); 1443 case NeonTypeFlags::Int32: 1444 return llvm::VectorType::get(CGF->Int32Ty, 2 << IsQuad); 1445 case NeonTypeFlags::Int64: 1446 return llvm::VectorType::get(CGF->Int64Ty, 1 << IsQuad); 1447 case NeonTypeFlags::Float32: 1448 return llvm::VectorType::get(CGF->FloatTy, 2 << IsQuad); 1449 } 1450 llvm_unreachable("Invalid NeonTypeFlags element type!"); 1451 } 1452 1453 Value *CodeGenFunction::EmitNeonSplat(Value *V, Constant *C) { 1454 unsigned nElts = cast<llvm::VectorType>(V->getType())->getNumElements(); 1455 Value* SV = llvm::ConstantVector::getSplat(nElts, C); 1456 return Builder.CreateShuffleVector(V, V, SV, "lane"); 1457 } 1458 1459 Value *CodeGenFunction::EmitNeonCall(Function *F, SmallVectorImpl<Value*> &Ops, 1460 const char *name, 1461 unsigned shift, bool rightshift) { 1462 unsigned j = 0; 1463 for (Function::const_arg_iterator ai = F->arg_begin(), ae = F->arg_end(); 1464 ai != ae; ++ai, ++j) 1465 if (shift > 0 && shift == j) 1466 Ops[j] = EmitNeonShiftVector(Ops[j], ai->getType(), rightshift); 1467 else 1468 Ops[j] = Builder.CreateBitCast(Ops[j], ai->getType(), name); 1469 1470 return Builder.CreateCall(F, Ops, name); 1471 } 1472 1473 Value *CodeGenFunction::EmitNeonShiftVector(Value *V, llvm::Type *Ty, 1474 bool neg) { 1475 int SV = cast<ConstantInt>(V)->getSExtValue(); 1476 1477 llvm::VectorType *VTy = cast<llvm::VectorType>(Ty); 1478 llvm::Constant *C = ConstantInt::get(VTy->getElementType(), neg ? -SV : SV); 1479 return llvm::ConstantVector::getSplat(VTy->getNumElements(), C); 1480 } 1481 1482 /// GetPointeeAlignment - Given an expression with a pointer type, find the 1483 /// alignment of the type referenced by the pointer. Skip over implicit 1484 /// casts. 1485 std::pair<llvm::Value*, unsigned> 1486 CodeGenFunction::EmitPointerWithAlignment(const Expr *Addr) { 1487 assert(Addr->getType()->isPointerType()); 1488 Addr = Addr->IgnoreParens(); 1489 if (const ImplicitCastExpr *ICE = dyn_cast<ImplicitCastExpr>(Addr)) { 1490 if ((ICE->getCastKind() == CK_BitCast || ICE->getCastKind() == CK_NoOp) && 1491 ICE->getSubExpr()->getType()->isPointerType()) { 1492 std::pair<llvm::Value*, unsigned> Ptr = 1493 EmitPointerWithAlignment(ICE->getSubExpr()); 1494 Ptr.first = Builder.CreateBitCast(Ptr.first, 1495 ConvertType(Addr->getType())); 1496 return Ptr; 1497 } else if (ICE->getCastKind() == CK_ArrayToPointerDecay) { 1498 LValue LV = EmitLValue(ICE->getSubExpr()); 1499 unsigned Align = LV.getAlignment().getQuantity(); 1500 if (!Align) { 1501 // FIXME: Once LValues are fixed to always set alignment, 1502 // zap this code. 1503 QualType PtTy = ICE->getSubExpr()->getType(); 1504 if (!PtTy->isIncompleteType()) 1505 Align = getContext().getTypeAlignInChars(PtTy).getQuantity(); 1506 else 1507 Align = 1; 1508 } 1509 return std::make_pair(LV.getAddress(), Align); 1510 } 1511 } 1512 if (const UnaryOperator *UO = dyn_cast<UnaryOperator>(Addr)) { 1513 if (UO->getOpcode() == UO_AddrOf) { 1514 LValue LV = EmitLValue(UO->getSubExpr()); 1515 unsigned Align = LV.getAlignment().getQuantity(); 1516 if (!Align) { 1517 // FIXME: Once LValues are fixed to always set alignment, 1518 // zap this code. 1519 QualType PtTy = UO->getSubExpr()->getType(); 1520 if (!PtTy->isIncompleteType()) 1521 Align = getContext().getTypeAlignInChars(PtTy).getQuantity(); 1522 else 1523 Align = 1; 1524 } 1525 return std::make_pair(LV.getAddress(), Align); 1526 } 1527 } 1528 1529 unsigned Align = 1; 1530 QualType PtTy = Addr->getType()->getPointeeType(); 1531 if (!PtTy->isIncompleteType()) 1532 Align = getContext().getTypeAlignInChars(PtTy).getQuantity(); 1533 1534 return std::make_pair(EmitScalarExpr(Addr), Align); 1535 } 1536 1537 Value *CodeGenFunction::EmitARMBuiltinExpr(unsigned BuiltinID, 1538 const CallExpr *E) { 1539 if (BuiltinID == ARM::BI__clear_cache) { 1540 const FunctionDecl *FD = E->getDirectCallee(); 1541 // Oddly people write this call without args on occasion and gcc accepts 1542 // it - it's also marked as varargs in the description file. 1543 SmallVector<Value*, 2> Ops; 1544 for (unsigned i = 0; i < E->getNumArgs(); i++) 1545 Ops.push_back(EmitScalarExpr(E->getArg(i))); 1546 llvm::Type *Ty = CGM.getTypes().ConvertType(FD->getType()); 1547 llvm::FunctionType *FTy = cast<llvm::FunctionType>(Ty); 1548 StringRef Name = FD->getName(); 1549 return Builder.CreateCall(CGM.CreateRuntimeFunction(FTy, Name), Ops); 1550 } 1551 1552 if (BuiltinID == ARM::BI__builtin_arm_ldrexd) { 1553 Function *F = CGM.getIntrinsic(Intrinsic::arm_ldrexd); 1554 1555 Value *LdPtr = EmitScalarExpr(E->getArg(0)); 1556 Value *Val = Builder.CreateCall(F, LdPtr, "ldrexd"); 1557 1558 Value *Val0 = Builder.CreateExtractValue(Val, 1); 1559 Value *Val1 = Builder.CreateExtractValue(Val, 0); 1560 Val0 = Builder.CreateZExt(Val0, Int64Ty); 1561 Val1 = Builder.CreateZExt(Val1, Int64Ty); 1562 1563 Value *ShiftCst = llvm::ConstantInt::get(Int64Ty, 32); 1564 Val = Builder.CreateShl(Val0, ShiftCst, "shl", true /* nuw */); 1565 return Builder.CreateOr(Val, Val1); 1566 } 1567 1568 if (BuiltinID == ARM::BI__builtin_arm_strexd) { 1569 Function *F = CGM.getIntrinsic(Intrinsic::arm_strexd); 1570 llvm::Type *STy = llvm::StructType::get(Int32Ty, Int32Ty, NULL); 1571 1572 Value *One = llvm::ConstantInt::get(Int32Ty, 1); 1573 Value *Tmp = Builder.CreateAlloca(Int64Ty, One); 1574 Value *Val = EmitScalarExpr(E->getArg(0)); 1575 Builder.CreateStore(Val, Tmp); 1576 1577 Value *LdPtr = Builder.CreateBitCast(Tmp,llvm::PointerType::getUnqual(STy)); 1578 Val = Builder.CreateLoad(LdPtr); 1579 1580 Value *Arg0 = Builder.CreateExtractValue(Val, 0); 1581 Value *Arg1 = Builder.CreateExtractValue(Val, 1); 1582 Value *StPtr = EmitScalarExpr(E->getArg(1)); 1583 return Builder.CreateCall3(F, Arg0, Arg1, StPtr, "strexd"); 1584 } 1585 1586 SmallVector<Value*, 4> Ops; 1587 llvm::Value *Align = 0; 1588 for (unsigned i = 0, e = E->getNumArgs() - 1; i != e; i++) { 1589 if (i == 0) { 1590 switch (BuiltinID) { 1591 case ARM::BI__builtin_neon_vld1_v: 1592 case ARM::BI__builtin_neon_vld1q_v: 1593 case ARM::BI__builtin_neon_vld1q_lane_v: 1594 case ARM::BI__builtin_neon_vld1_lane_v: 1595 case ARM::BI__builtin_neon_vld1_dup_v: 1596 case ARM::BI__builtin_neon_vld1q_dup_v: 1597 case ARM::BI__builtin_neon_vst1_v: 1598 case ARM::BI__builtin_neon_vst1q_v: 1599 case ARM::BI__builtin_neon_vst1q_lane_v: 1600 case ARM::BI__builtin_neon_vst1_lane_v: 1601 case ARM::BI__builtin_neon_vst2_v: 1602 case ARM::BI__builtin_neon_vst2q_v: 1603 case ARM::BI__builtin_neon_vst2_lane_v: 1604 case ARM::BI__builtin_neon_vst2q_lane_v: 1605 case ARM::BI__builtin_neon_vst3_v: 1606 case ARM::BI__builtin_neon_vst3q_v: 1607 case ARM::BI__builtin_neon_vst3_lane_v: 1608 case ARM::BI__builtin_neon_vst3q_lane_v: 1609 case ARM::BI__builtin_neon_vst4_v: 1610 case ARM::BI__builtin_neon_vst4q_v: 1611 case ARM::BI__builtin_neon_vst4_lane_v: 1612 case ARM::BI__builtin_neon_vst4q_lane_v: 1613 // Get the alignment for the argument in addition to the value; 1614 // we'll use it later. 1615 std::pair<llvm::Value*, unsigned> Src = 1616 EmitPointerWithAlignment(E->getArg(0)); 1617 Ops.push_back(Src.first); 1618 Align = Builder.getInt32(Src.second); 1619 continue; 1620 } 1621 } 1622 if (i == 1) { 1623 switch (BuiltinID) { 1624 case ARM::BI__builtin_neon_vld2_v: 1625 case ARM::BI__builtin_neon_vld2q_v: 1626 case ARM::BI__builtin_neon_vld3_v: 1627 case ARM::BI__builtin_neon_vld3q_v: 1628 case ARM::BI__builtin_neon_vld4_v: 1629 case ARM::BI__builtin_neon_vld4q_v: 1630 case ARM::BI__builtin_neon_vld2_lane_v: 1631 case ARM::BI__builtin_neon_vld2q_lane_v: 1632 case ARM::BI__builtin_neon_vld3_lane_v: 1633 case ARM::BI__builtin_neon_vld3q_lane_v: 1634 case ARM::BI__builtin_neon_vld4_lane_v: 1635 case ARM::BI__builtin_neon_vld4q_lane_v: 1636 case ARM::BI__builtin_neon_vld2_dup_v: 1637 case ARM::BI__builtin_neon_vld3_dup_v: 1638 case ARM::BI__builtin_neon_vld4_dup_v: 1639 // Get the alignment for the argument in addition to the value; 1640 // we'll use it later. 1641 std::pair<llvm::Value*, unsigned> Src = 1642 EmitPointerWithAlignment(E->getArg(1)); 1643 Ops.push_back(Src.first); 1644 Align = Builder.getInt32(Src.second); 1645 continue; 1646 } 1647 } 1648 Ops.push_back(EmitScalarExpr(E->getArg(i))); 1649 } 1650 1651 // vget_lane and vset_lane are not overloaded and do not have an extra 1652 // argument that specifies the vector type. 1653 switch (BuiltinID) { 1654 default: break; 1655 case ARM::BI__builtin_neon_vget_lane_i8: 1656 case ARM::BI__builtin_neon_vget_lane_i16: 1657 case ARM::BI__builtin_neon_vget_lane_i32: 1658 case ARM::BI__builtin_neon_vget_lane_i64: 1659 case ARM::BI__builtin_neon_vget_lane_f32: 1660 case ARM::BI__builtin_neon_vgetq_lane_i8: 1661 case ARM::BI__builtin_neon_vgetq_lane_i16: 1662 case ARM::BI__builtin_neon_vgetq_lane_i32: 1663 case ARM::BI__builtin_neon_vgetq_lane_i64: 1664 case ARM::BI__builtin_neon_vgetq_lane_f32: 1665 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 1666 "vget_lane"); 1667 case ARM::BI__builtin_neon_vset_lane_i8: 1668 case ARM::BI__builtin_neon_vset_lane_i16: 1669 case ARM::BI__builtin_neon_vset_lane_i32: 1670 case ARM::BI__builtin_neon_vset_lane_i64: 1671 case ARM::BI__builtin_neon_vset_lane_f32: 1672 case ARM::BI__builtin_neon_vsetq_lane_i8: 1673 case ARM::BI__builtin_neon_vsetq_lane_i16: 1674 case ARM::BI__builtin_neon_vsetq_lane_i32: 1675 case ARM::BI__builtin_neon_vsetq_lane_i64: 1676 case ARM::BI__builtin_neon_vsetq_lane_f32: 1677 Ops.push_back(EmitScalarExpr(E->getArg(2))); 1678 return Builder.CreateInsertElement(Ops[1], Ops[0], Ops[2], "vset_lane"); 1679 } 1680 1681 // Get the last argument, which specifies the vector type. 1682 llvm::APSInt Result; 1683 const Expr *Arg = E->getArg(E->getNumArgs()-1); 1684 if (!Arg->isIntegerConstantExpr(Result, getContext())) 1685 return 0; 1686 1687 if (BuiltinID == ARM::BI__builtin_arm_vcvtr_f || 1688 BuiltinID == ARM::BI__builtin_arm_vcvtr_d) { 1689 // Determine the overloaded type of this builtin. 1690 llvm::Type *Ty; 1691 if (BuiltinID == ARM::BI__builtin_arm_vcvtr_f) 1692 Ty = FloatTy; 1693 else 1694 Ty = DoubleTy; 1695 1696 // Determine whether this is an unsigned conversion or not. 1697 bool usgn = Result.getZExtValue() == 1; 1698 unsigned Int = usgn ? Intrinsic::arm_vcvtru : Intrinsic::arm_vcvtr; 1699 1700 // Call the appropriate intrinsic. 1701 Function *F = CGM.getIntrinsic(Int, Ty); 1702 return Builder.CreateCall(F, Ops, "vcvtr"); 1703 } 1704 1705 // Determine the type of this overloaded NEON intrinsic. 1706 NeonTypeFlags Type(Result.getZExtValue()); 1707 bool usgn = Type.isUnsigned(); 1708 bool quad = Type.isQuad(); 1709 bool rightShift = false; 1710 1711 llvm::VectorType *VTy = GetNeonType(this, Type); 1712 llvm::Type *Ty = VTy; 1713 if (!Ty) 1714 return 0; 1715 1716 unsigned Int; 1717 switch (BuiltinID) { 1718 default: return 0; 1719 case ARM::BI__builtin_neon_vbsl_v: 1720 case ARM::BI__builtin_neon_vbslq_v: 1721 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vbsl, Ty), 1722 Ops, "vbsl"); 1723 case ARM::BI__builtin_neon_vabd_v: 1724 case ARM::BI__builtin_neon_vabdq_v: 1725 Int = usgn ? Intrinsic::arm_neon_vabdu : Intrinsic::arm_neon_vabds; 1726 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vabd"); 1727 case ARM::BI__builtin_neon_vabs_v: 1728 case ARM::BI__builtin_neon_vabsq_v: 1729 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vabs, Ty), 1730 Ops, "vabs"); 1731 case ARM::BI__builtin_neon_vaddhn_v: 1732 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vaddhn, Ty), 1733 Ops, "vaddhn"); 1734 case ARM::BI__builtin_neon_vcale_v: 1735 std::swap(Ops[0], Ops[1]); 1736 case ARM::BI__builtin_neon_vcage_v: { 1737 Function *F = CGM.getIntrinsic(Intrinsic::arm_neon_vacged); 1738 return EmitNeonCall(F, Ops, "vcage"); 1739 } 1740 case ARM::BI__builtin_neon_vcaleq_v: 1741 std::swap(Ops[0], Ops[1]); 1742 case ARM::BI__builtin_neon_vcageq_v: { 1743 Function *F = CGM.getIntrinsic(Intrinsic::arm_neon_vacgeq); 1744 return EmitNeonCall(F, Ops, "vcage"); 1745 } 1746 case ARM::BI__builtin_neon_vcalt_v: 1747 std::swap(Ops[0], Ops[1]); 1748 case ARM::BI__builtin_neon_vcagt_v: { 1749 Function *F = CGM.getIntrinsic(Intrinsic::arm_neon_vacgtd); 1750 return EmitNeonCall(F, Ops, "vcagt"); 1751 } 1752 case ARM::BI__builtin_neon_vcaltq_v: 1753 std::swap(Ops[0], Ops[1]); 1754 case ARM::BI__builtin_neon_vcagtq_v: { 1755 Function *F = CGM.getIntrinsic(Intrinsic::arm_neon_vacgtq); 1756 return EmitNeonCall(F, Ops, "vcagt"); 1757 } 1758 case ARM::BI__builtin_neon_vcls_v: 1759 case ARM::BI__builtin_neon_vclsq_v: { 1760 Function *F = CGM.getIntrinsic(Intrinsic::arm_neon_vcls, Ty); 1761 return EmitNeonCall(F, Ops, "vcls"); 1762 } 1763 case ARM::BI__builtin_neon_vclz_v: 1764 case ARM::BI__builtin_neon_vclzq_v: { 1765 // Generate target-independent intrinsic; also need to add second argument 1766 // for whether or not clz of zero is undefined; on ARM it isn't. 1767 Function *F = CGM.getIntrinsic(Intrinsic::ctlz, Ty); 1768 Ops.push_back(Builder.getInt1(Target.isCLZForZeroUndef())); 1769 return EmitNeonCall(F, Ops, "vclz"); 1770 } 1771 case ARM::BI__builtin_neon_vcnt_v: 1772 case ARM::BI__builtin_neon_vcntq_v: { 1773 // generate target-independent intrinsic 1774 Function *F = CGM.getIntrinsic(Intrinsic::ctpop, Ty); 1775 return EmitNeonCall(F, Ops, "vctpop"); 1776 } 1777 case ARM::BI__builtin_neon_vcvt_f16_v: { 1778 assert(Type.getEltType() == NeonTypeFlags::Float16 && !quad && 1779 "unexpected vcvt_f16_v builtin"); 1780 Function *F = CGM.getIntrinsic(Intrinsic::arm_neon_vcvtfp2hf); 1781 return EmitNeonCall(F, Ops, "vcvt"); 1782 } 1783 case ARM::BI__builtin_neon_vcvt_f32_f16: { 1784 assert(Type.getEltType() == NeonTypeFlags::Float16 && !quad && 1785 "unexpected vcvt_f32_f16 builtin"); 1786 Function *F = CGM.getIntrinsic(Intrinsic::arm_neon_vcvthf2fp); 1787 return EmitNeonCall(F, Ops, "vcvt"); 1788 } 1789 case ARM::BI__builtin_neon_vcvt_f32_v: 1790 case ARM::BI__builtin_neon_vcvtq_f32_v: 1791 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 1792 Ty = GetNeonType(this, NeonTypeFlags(NeonTypeFlags::Float32, false, quad)); 1793 return usgn ? Builder.CreateUIToFP(Ops[0], Ty, "vcvt") 1794 : Builder.CreateSIToFP(Ops[0], Ty, "vcvt"); 1795 case ARM::BI__builtin_neon_vcvt_s32_v: 1796 case ARM::BI__builtin_neon_vcvt_u32_v: 1797 case ARM::BI__builtin_neon_vcvtq_s32_v: 1798 case ARM::BI__builtin_neon_vcvtq_u32_v: { 1799 llvm::Type *FloatTy = 1800 GetNeonType(this, NeonTypeFlags(NeonTypeFlags::Float32, false, quad)); 1801 Ops[0] = Builder.CreateBitCast(Ops[0], FloatTy); 1802 return usgn ? Builder.CreateFPToUI(Ops[0], Ty, "vcvt") 1803 : Builder.CreateFPToSI(Ops[0], Ty, "vcvt"); 1804 } 1805 case ARM::BI__builtin_neon_vcvt_n_f32_v: 1806 case ARM::BI__builtin_neon_vcvtq_n_f32_v: { 1807 llvm::Type *FloatTy = 1808 GetNeonType(this, NeonTypeFlags(NeonTypeFlags::Float32, false, quad)); 1809 llvm::Type *Tys[2] = { FloatTy, Ty }; 1810 Int = usgn ? Intrinsic::arm_neon_vcvtfxu2fp 1811 : Intrinsic::arm_neon_vcvtfxs2fp; 1812 Function *F = CGM.getIntrinsic(Int, Tys); 1813 return EmitNeonCall(F, Ops, "vcvt_n"); 1814 } 1815 case ARM::BI__builtin_neon_vcvt_n_s32_v: 1816 case ARM::BI__builtin_neon_vcvt_n_u32_v: 1817 case ARM::BI__builtin_neon_vcvtq_n_s32_v: 1818 case ARM::BI__builtin_neon_vcvtq_n_u32_v: { 1819 llvm::Type *FloatTy = 1820 GetNeonType(this, NeonTypeFlags(NeonTypeFlags::Float32, false, quad)); 1821 llvm::Type *Tys[2] = { Ty, FloatTy }; 1822 Int = usgn ? Intrinsic::arm_neon_vcvtfp2fxu 1823 : Intrinsic::arm_neon_vcvtfp2fxs; 1824 Function *F = CGM.getIntrinsic(Int, Tys); 1825 return EmitNeonCall(F, Ops, "vcvt_n"); 1826 } 1827 case ARM::BI__builtin_neon_vext_v: 1828 case ARM::BI__builtin_neon_vextq_v: { 1829 int CV = cast<ConstantInt>(Ops[2])->getSExtValue(); 1830 SmallVector<Constant*, 16> Indices; 1831 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) 1832 Indices.push_back(ConstantInt::get(Int32Ty, i+CV)); 1833 1834 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 1835 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 1836 Value *SV = llvm::ConstantVector::get(Indices); 1837 return Builder.CreateShuffleVector(Ops[0], Ops[1], SV, "vext"); 1838 } 1839 case ARM::BI__builtin_neon_vhadd_v: 1840 case ARM::BI__builtin_neon_vhaddq_v: 1841 Int = usgn ? Intrinsic::arm_neon_vhaddu : Intrinsic::arm_neon_vhadds; 1842 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vhadd"); 1843 case ARM::BI__builtin_neon_vhsub_v: 1844 case ARM::BI__builtin_neon_vhsubq_v: 1845 Int = usgn ? Intrinsic::arm_neon_vhsubu : Intrinsic::arm_neon_vhsubs; 1846 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vhsub"); 1847 case ARM::BI__builtin_neon_vld1_v: 1848 case ARM::BI__builtin_neon_vld1q_v: 1849 Ops.push_back(Align); 1850 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vld1, Ty), 1851 Ops, "vld1"); 1852 case ARM::BI__builtin_neon_vld1q_lane_v: 1853 // Handle 64-bit integer elements as a special case. Use shuffles of 1854 // one-element vectors to avoid poor code for i64 in the backend. 1855 if (VTy->getElementType()->isIntegerTy(64)) { 1856 // Extract the other lane. 1857 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 1858 int Lane = cast<ConstantInt>(Ops[2])->getZExtValue(); 1859 Value *SV = llvm::ConstantVector::get(ConstantInt::get(Int32Ty, 1-Lane)); 1860 Ops[1] = Builder.CreateShuffleVector(Ops[1], Ops[1], SV); 1861 // Load the value as a one-element vector. 1862 Ty = llvm::VectorType::get(VTy->getElementType(), 1); 1863 Function *F = CGM.getIntrinsic(Intrinsic::arm_neon_vld1, Ty); 1864 Value *Ld = Builder.CreateCall2(F, Ops[0], Align); 1865 // Combine them. 1866 SmallVector<Constant*, 2> Indices; 1867 Indices.push_back(ConstantInt::get(Int32Ty, 1-Lane)); 1868 Indices.push_back(ConstantInt::get(Int32Ty, Lane)); 1869 SV = llvm::ConstantVector::get(Indices); 1870 return Builder.CreateShuffleVector(Ops[1], Ld, SV, "vld1q_lane"); 1871 } 1872 // fall through 1873 case ARM::BI__builtin_neon_vld1_lane_v: { 1874 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 1875 Ty = llvm::PointerType::getUnqual(VTy->getElementType()); 1876 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 1877 LoadInst *Ld = Builder.CreateLoad(Ops[0]); 1878 Ld->setAlignment(cast<ConstantInt>(Align)->getZExtValue()); 1879 return Builder.CreateInsertElement(Ops[1], Ld, Ops[2], "vld1_lane"); 1880 } 1881 case ARM::BI__builtin_neon_vld1_dup_v: 1882 case ARM::BI__builtin_neon_vld1q_dup_v: { 1883 Value *V = UndefValue::get(Ty); 1884 Ty = llvm::PointerType::getUnqual(VTy->getElementType()); 1885 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 1886 LoadInst *Ld = Builder.CreateLoad(Ops[0]); 1887 Ld->setAlignment(cast<ConstantInt>(Align)->getZExtValue()); 1888 llvm::Constant *CI = ConstantInt::get(Int32Ty, 0); 1889 Ops[0] = Builder.CreateInsertElement(V, Ld, CI); 1890 return EmitNeonSplat(Ops[0], CI); 1891 } 1892 case ARM::BI__builtin_neon_vld2_v: 1893 case ARM::BI__builtin_neon_vld2q_v: { 1894 Function *F = CGM.getIntrinsic(Intrinsic::arm_neon_vld2, Ty); 1895 Ops[1] = Builder.CreateCall2(F, Ops[1], Align, "vld2"); 1896 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 1897 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 1898 return Builder.CreateStore(Ops[1], Ops[0]); 1899 } 1900 case ARM::BI__builtin_neon_vld3_v: 1901 case ARM::BI__builtin_neon_vld3q_v: { 1902 Function *F = CGM.getIntrinsic(Intrinsic::arm_neon_vld3, Ty); 1903 Ops[1] = Builder.CreateCall2(F, Ops[1], Align, "vld3"); 1904 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 1905 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 1906 return Builder.CreateStore(Ops[1], Ops[0]); 1907 } 1908 case ARM::BI__builtin_neon_vld4_v: 1909 case ARM::BI__builtin_neon_vld4q_v: { 1910 Function *F = CGM.getIntrinsic(Intrinsic::arm_neon_vld4, Ty); 1911 Ops[1] = Builder.CreateCall2(F, Ops[1], Align, "vld4"); 1912 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 1913 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 1914 return Builder.CreateStore(Ops[1], Ops[0]); 1915 } 1916 case ARM::BI__builtin_neon_vld2_lane_v: 1917 case ARM::BI__builtin_neon_vld2q_lane_v: { 1918 Function *F = CGM.getIntrinsic(Intrinsic::arm_neon_vld2lane, Ty); 1919 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 1920 Ops[3] = Builder.CreateBitCast(Ops[3], Ty); 1921 Ops.push_back(Align); 1922 Ops[1] = Builder.CreateCall(F, makeArrayRef(Ops).slice(1), "vld2_lane"); 1923 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 1924 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 1925 return Builder.CreateStore(Ops[1], Ops[0]); 1926 } 1927 case ARM::BI__builtin_neon_vld3_lane_v: 1928 case ARM::BI__builtin_neon_vld3q_lane_v: { 1929 Function *F = CGM.getIntrinsic(Intrinsic::arm_neon_vld3lane, Ty); 1930 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 1931 Ops[3] = Builder.CreateBitCast(Ops[3], Ty); 1932 Ops[4] = Builder.CreateBitCast(Ops[4], Ty); 1933 Ops.push_back(Align); 1934 Ops[1] = Builder.CreateCall(F, makeArrayRef(Ops).slice(1), "vld3_lane"); 1935 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 1936 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 1937 return Builder.CreateStore(Ops[1], Ops[0]); 1938 } 1939 case ARM::BI__builtin_neon_vld4_lane_v: 1940 case ARM::BI__builtin_neon_vld4q_lane_v: { 1941 Function *F = CGM.getIntrinsic(Intrinsic::arm_neon_vld4lane, Ty); 1942 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 1943 Ops[3] = Builder.CreateBitCast(Ops[3], Ty); 1944 Ops[4] = Builder.CreateBitCast(Ops[4], Ty); 1945 Ops[5] = Builder.CreateBitCast(Ops[5], Ty); 1946 Ops.push_back(Align); 1947 Ops[1] = Builder.CreateCall(F, makeArrayRef(Ops).slice(1), "vld3_lane"); 1948 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 1949 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 1950 return Builder.CreateStore(Ops[1], Ops[0]); 1951 } 1952 case ARM::BI__builtin_neon_vld2_dup_v: 1953 case ARM::BI__builtin_neon_vld3_dup_v: 1954 case ARM::BI__builtin_neon_vld4_dup_v: { 1955 // Handle 64-bit elements as a special-case. There is no "dup" needed. 1956 if (VTy->getElementType()->getPrimitiveSizeInBits() == 64) { 1957 switch (BuiltinID) { 1958 case ARM::BI__builtin_neon_vld2_dup_v: 1959 Int = Intrinsic::arm_neon_vld2; 1960 break; 1961 case ARM::BI__builtin_neon_vld3_dup_v: 1962 Int = Intrinsic::arm_neon_vld3; 1963 break; 1964 case ARM::BI__builtin_neon_vld4_dup_v: 1965 Int = Intrinsic::arm_neon_vld4; 1966 break; 1967 default: llvm_unreachable("unknown vld_dup intrinsic?"); 1968 } 1969 Function *F = CGM.getIntrinsic(Int, Ty); 1970 Ops[1] = Builder.CreateCall2(F, Ops[1], Align, "vld_dup"); 1971 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 1972 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 1973 return Builder.CreateStore(Ops[1], Ops[0]); 1974 } 1975 switch (BuiltinID) { 1976 case ARM::BI__builtin_neon_vld2_dup_v: 1977 Int = Intrinsic::arm_neon_vld2lane; 1978 break; 1979 case ARM::BI__builtin_neon_vld3_dup_v: 1980 Int = Intrinsic::arm_neon_vld3lane; 1981 break; 1982 case ARM::BI__builtin_neon_vld4_dup_v: 1983 Int = Intrinsic::arm_neon_vld4lane; 1984 break; 1985 default: llvm_unreachable("unknown vld_dup intrinsic?"); 1986 } 1987 Function *F = CGM.getIntrinsic(Int, Ty); 1988 llvm::StructType *STy = cast<llvm::StructType>(F->getReturnType()); 1989 1990 SmallVector<Value*, 6> Args; 1991 Args.push_back(Ops[1]); 1992 Args.append(STy->getNumElements(), UndefValue::get(Ty)); 1993 1994 llvm::Constant *CI = ConstantInt::get(Int32Ty, 0); 1995 Args.push_back(CI); 1996 Args.push_back(Align); 1997 1998 Ops[1] = Builder.CreateCall(F, Args, "vld_dup"); 1999 // splat lane 0 to all elts in each vector of the result. 2000 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 2001 Value *Val = Builder.CreateExtractValue(Ops[1], i); 2002 Value *Elt = Builder.CreateBitCast(Val, Ty); 2003 Elt = EmitNeonSplat(Elt, CI); 2004 Elt = Builder.CreateBitCast(Elt, Val->getType()); 2005 Ops[1] = Builder.CreateInsertValue(Ops[1], Elt, i); 2006 } 2007 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 2008 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 2009 return Builder.CreateStore(Ops[1], Ops[0]); 2010 } 2011 case ARM::BI__builtin_neon_vmax_v: 2012 case ARM::BI__builtin_neon_vmaxq_v: 2013 Int = usgn ? Intrinsic::arm_neon_vmaxu : Intrinsic::arm_neon_vmaxs; 2014 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmax"); 2015 case ARM::BI__builtin_neon_vmin_v: 2016 case ARM::BI__builtin_neon_vminq_v: 2017 Int = usgn ? Intrinsic::arm_neon_vminu : Intrinsic::arm_neon_vmins; 2018 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmin"); 2019 case ARM::BI__builtin_neon_vmovl_v: { 2020 llvm::Type *DTy =llvm::VectorType::getTruncatedElementVectorType(VTy); 2021 Ops[0] = Builder.CreateBitCast(Ops[0], DTy); 2022 if (usgn) 2023 return Builder.CreateZExt(Ops[0], Ty, "vmovl"); 2024 return Builder.CreateSExt(Ops[0], Ty, "vmovl"); 2025 } 2026 case ARM::BI__builtin_neon_vmovn_v: { 2027 llvm::Type *QTy = llvm::VectorType::getExtendedElementVectorType(VTy); 2028 Ops[0] = Builder.CreateBitCast(Ops[0], QTy); 2029 return Builder.CreateTrunc(Ops[0], Ty, "vmovn"); 2030 } 2031 case ARM::BI__builtin_neon_vmul_v: 2032 case ARM::BI__builtin_neon_vmulq_v: 2033 assert(Type.isPoly() && "vmul builtin only supported for polynomial types"); 2034 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vmulp, Ty), 2035 Ops, "vmul"); 2036 case ARM::BI__builtin_neon_vmull_v: 2037 Int = usgn ? Intrinsic::arm_neon_vmullu : Intrinsic::arm_neon_vmulls; 2038 Int = Type.isPoly() ? (unsigned)Intrinsic::arm_neon_vmullp : Int; 2039 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmull"); 2040 case ARM::BI__builtin_neon_vfma_v: 2041 case ARM::BI__builtin_neon_vfmaq_v: { 2042 Value *F = CGM.getIntrinsic(Intrinsic::fma, Ty); 2043 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 2044 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 2045 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 2046 return Builder.CreateCall3(F, Ops[0], Ops[1], Ops[2]); 2047 } 2048 case ARM::BI__builtin_neon_vpadal_v: 2049 case ARM::BI__builtin_neon_vpadalq_v: { 2050 Int = usgn ? Intrinsic::arm_neon_vpadalu : Intrinsic::arm_neon_vpadals; 2051 // The source operand type has twice as many elements of half the size. 2052 unsigned EltBits = VTy->getElementType()->getPrimitiveSizeInBits(); 2053 llvm::Type *EltTy = 2054 llvm::IntegerType::get(getLLVMContext(), EltBits / 2); 2055 llvm::Type *NarrowTy = 2056 llvm::VectorType::get(EltTy, VTy->getNumElements() * 2); 2057 llvm::Type *Tys[2] = { Ty, NarrowTy }; 2058 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vpadal"); 2059 } 2060 case ARM::BI__builtin_neon_vpadd_v: 2061 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vpadd, Ty), 2062 Ops, "vpadd"); 2063 case ARM::BI__builtin_neon_vpaddl_v: 2064 case ARM::BI__builtin_neon_vpaddlq_v: { 2065 Int = usgn ? Intrinsic::arm_neon_vpaddlu : Intrinsic::arm_neon_vpaddls; 2066 // The source operand type has twice as many elements of half the size. 2067 unsigned EltBits = VTy->getElementType()->getPrimitiveSizeInBits(); 2068 llvm::Type *EltTy = llvm::IntegerType::get(getLLVMContext(), EltBits / 2); 2069 llvm::Type *NarrowTy = 2070 llvm::VectorType::get(EltTy, VTy->getNumElements() * 2); 2071 llvm::Type *Tys[2] = { Ty, NarrowTy }; 2072 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vpaddl"); 2073 } 2074 case ARM::BI__builtin_neon_vpmax_v: 2075 Int = usgn ? Intrinsic::arm_neon_vpmaxu : Intrinsic::arm_neon_vpmaxs; 2076 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vpmax"); 2077 case ARM::BI__builtin_neon_vpmin_v: 2078 Int = usgn ? Intrinsic::arm_neon_vpminu : Intrinsic::arm_neon_vpmins; 2079 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vpmin"); 2080 case ARM::BI__builtin_neon_vqabs_v: 2081 case ARM::BI__builtin_neon_vqabsq_v: 2082 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vqabs, Ty), 2083 Ops, "vqabs"); 2084 case ARM::BI__builtin_neon_vqadd_v: 2085 case ARM::BI__builtin_neon_vqaddq_v: 2086 Int = usgn ? Intrinsic::arm_neon_vqaddu : Intrinsic::arm_neon_vqadds; 2087 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqadd"); 2088 case ARM::BI__builtin_neon_vqdmlal_v: 2089 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vqdmlal, Ty), 2090 Ops, "vqdmlal"); 2091 case ARM::BI__builtin_neon_vqdmlsl_v: 2092 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vqdmlsl, Ty), 2093 Ops, "vqdmlsl"); 2094 case ARM::BI__builtin_neon_vqdmulh_v: 2095 case ARM::BI__builtin_neon_vqdmulhq_v: 2096 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vqdmulh, Ty), 2097 Ops, "vqdmulh"); 2098 case ARM::BI__builtin_neon_vqdmull_v: 2099 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vqdmull, Ty), 2100 Ops, "vqdmull"); 2101 case ARM::BI__builtin_neon_vqmovn_v: 2102 Int = usgn ? Intrinsic::arm_neon_vqmovnu : Intrinsic::arm_neon_vqmovns; 2103 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqmovn"); 2104 case ARM::BI__builtin_neon_vqmovun_v: 2105 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vqmovnsu, Ty), 2106 Ops, "vqdmull"); 2107 case ARM::BI__builtin_neon_vqneg_v: 2108 case ARM::BI__builtin_neon_vqnegq_v: 2109 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vqneg, Ty), 2110 Ops, "vqneg"); 2111 case ARM::BI__builtin_neon_vqrdmulh_v: 2112 case ARM::BI__builtin_neon_vqrdmulhq_v: 2113 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vqrdmulh, Ty), 2114 Ops, "vqrdmulh"); 2115 case ARM::BI__builtin_neon_vqrshl_v: 2116 case ARM::BI__builtin_neon_vqrshlq_v: 2117 Int = usgn ? Intrinsic::arm_neon_vqrshiftu : Intrinsic::arm_neon_vqrshifts; 2118 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqrshl"); 2119 case ARM::BI__builtin_neon_vqrshrn_n_v: 2120 Int = 2121 usgn ? Intrinsic::arm_neon_vqrshiftnu : Intrinsic::arm_neon_vqrshiftns; 2122 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqrshrn_n", 2123 1, true); 2124 case ARM::BI__builtin_neon_vqrshrun_n_v: 2125 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vqrshiftnsu, Ty), 2126 Ops, "vqrshrun_n", 1, true); 2127 case ARM::BI__builtin_neon_vqshl_v: 2128 case ARM::BI__builtin_neon_vqshlq_v: 2129 Int = usgn ? Intrinsic::arm_neon_vqshiftu : Intrinsic::arm_neon_vqshifts; 2130 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqshl"); 2131 case ARM::BI__builtin_neon_vqshl_n_v: 2132 case ARM::BI__builtin_neon_vqshlq_n_v: 2133 Int = usgn ? Intrinsic::arm_neon_vqshiftu : Intrinsic::arm_neon_vqshifts; 2134 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqshl_n", 2135 1, false); 2136 case ARM::BI__builtin_neon_vqshlu_n_v: 2137 case ARM::BI__builtin_neon_vqshluq_n_v: 2138 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vqshiftsu, Ty), 2139 Ops, "vqshlu", 1, false); 2140 case ARM::BI__builtin_neon_vqshrn_n_v: 2141 Int = usgn ? Intrinsic::arm_neon_vqshiftnu : Intrinsic::arm_neon_vqshiftns; 2142 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqshrn_n", 2143 1, true); 2144 case ARM::BI__builtin_neon_vqshrun_n_v: 2145 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vqshiftnsu, Ty), 2146 Ops, "vqshrun_n", 1, true); 2147 case ARM::BI__builtin_neon_vqsub_v: 2148 case ARM::BI__builtin_neon_vqsubq_v: 2149 Int = usgn ? Intrinsic::arm_neon_vqsubu : Intrinsic::arm_neon_vqsubs; 2150 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqsub"); 2151 case ARM::BI__builtin_neon_vraddhn_v: 2152 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vraddhn, Ty), 2153 Ops, "vraddhn"); 2154 case ARM::BI__builtin_neon_vrecpe_v: 2155 case ARM::BI__builtin_neon_vrecpeq_v: 2156 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vrecpe, Ty), 2157 Ops, "vrecpe"); 2158 case ARM::BI__builtin_neon_vrecps_v: 2159 case ARM::BI__builtin_neon_vrecpsq_v: 2160 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vrecps, Ty), 2161 Ops, "vrecps"); 2162 case ARM::BI__builtin_neon_vrhadd_v: 2163 case ARM::BI__builtin_neon_vrhaddq_v: 2164 Int = usgn ? Intrinsic::arm_neon_vrhaddu : Intrinsic::arm_neon_vrhadds; 2165 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrhadd"); 2166 case ARM::BI__builtin_neon_vrshl_v: 2167 case ARM::BI__builtin_neon_vrshlq_v: 2168 Int = usgn ? Intrinsic::arm_neon_vrshiftu : Intrinsic::arm_neon_vrshifts; 2169 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrshl"); 2170 case ARM::BI__builtin_neon_vrshrn_n_v: 2171 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vrshiftn, Ty), 2172 Ops, "vrshrn_n", 1, true); 2173 case ARM::BI__builtin_neon_vrshr_n_v: 2174 case ARM::BI__builtin_neon_vrshrq_n_v: 2175 Int = usgn ? Intrinsic::arm_neon_vrshiftu : Intrinsic::arm_neon_vrshifts; 2176 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrshr_n", 1, true); 2177 case ARM::BI__builtin_neon_vrsqrte_v: 2178 case ARM::BI__builtin_neon_vrsqrteq_v: 2179 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vrsqrte, Ty), 2180 Ops, "vrsqrte"); 2181 case ARM::BI__builtin_neon_vrsqrts_v: 2182 case ARM::BI__builtin_neon_vrsqrtsq_v: 2183 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vrsqrts, Ty), 2184 Ops, "vrsqrts"); 2185 case ARM::BI__builtin_neon_vrsra_n_v: 2186 case ARM::BI__builtin_neon_vrsraq_n_v: 2187 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 2188 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 2189 Ops[2] = EmitNeonShiftVector(Ops[2], Ty, true); 2190 Int = usgn ? Intrinsic::arm_neon_vrshiftu : Intrinsic::arm_neon_vrshifts; 2191 Ops[1] = Builder.CreateCall2(CGM.getIntrinsic(Int, Ty), Ops[1], Ops[2]); 2192 return Builder.CreateAdd(Ops[0], Ops[1], "vrsra_n"); 2193 case ARM::BI__builtin_neon_vrsubhn_v: 2194 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vrsubhn, Ty), 2195 Ops, "vrsubhn"); 2196 case ARM::BI__builtin_neon_vshl_v: 2197 case ARM::BI__builtin_neon_vshlq_v: 2198 Int = usgn ? Intrinsic::arm_neon_vshiftu : Intrinsic::arm_neon_vshifts; 2199 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vshl"); 2200 case ARM::BI__builtin_neon_vshll_n_v: 2201 Int = usgn ? Intrinsic::arm_neon_vshiftlu : Intrinsic::arm_neon_vshiftls; 2202 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vshll", 1); 2203 case ARM::BI__builtin_neon_vshl_n_v: 2204 case ARM::BI__builtin_neon_vshlq_n_v: 2205 Ops[1] = EmitNeonShiftVector(Ops[1], Ty, false); 2206 return Builder.CreateShl(Builder.CreateBitCast(Ops[0],Ty), Ops[1], 2207 "vshl_n"); 2208 case ARM::BI__builtin_neon_vshrn_n_v: 2209 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vshiftn, Ty), 2210 Ops, "vshrn_n", 1, true); 2211 case ARM::BI__builtin_neon_vshr_n_v: 2212 case ARM::BI__builtin_neon_vshrq_n_v: 2213 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 2214 Ops[1] = EmitNeonShiftVector(Ops[1], Ty, false); 2215 if (usgn) 2216 return Builder.CreateLShr(Ops[0], Ops[1], "vshr_n"); 2217 else 2218 return Builder.CreateAShr(Ops[0], Ops[1], "vshr_n"); 2219 case ARM::BI__builtin_neon_vsri_n_v: 2220 case ARM::BI__builtin_neon_vsriq_n_v: 2221 rightShift = true; 2222 case ARM::BI__builtin_neon_vsli_n_v: 2223 case ARM::BI__builtin_neon_vsliq_n_v: 2224 Ops[2] = EmitNeonShiftVector(Ops[2], Ty, rightShift); 2225 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vshiftins, Ty), 2226 Ops, "vsli_n"); 2227 case ARM::BI__builtin_neon_vsra_n_v: 2228 case ARM::BI__builtin_neon_vsraq_n_v: 2229 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 2230 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 2231 Ops[2] = EmitNeonShiftVector(Ops[2], Ty, false); 2232 if (usgn) 2233 Ops[1] = Builder.CreateLShr(Ops[1], Ops[2], "vsra_n"); 2234 else 2235 Ops[1] = Builder.CreateAShr(Ops[1], Ops[2], "vsra_n"); 2236 return Builder.CreateAdd(Ops[0], Ops[1]); 2237 case ARM::BI__builtin_neon_vst1_v: 2238 case ARM::BI__builtin_neon_vst1q_v: 2239 Ops.push_back(Align); 2240 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vst1, Ty), 2241 Ops, ""); 2242 case ARM::BI__builtin_neon_vst1q_lane_v: 2243 // Handle 64-bit integer elements as a special case. Use a shuffle to get 2244 // a one-element vector and avoid poor code for i64 in the backend. 2245 if (VTy->getElementType()->isIntegerTy(64)) { 2246 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 2247 Value *SV = llvm::ConstantVector::get(cast<llvm::Constant>(Ops[2])); 2248 Ops[1] = Builder.CreateShuffleVector(Ops[1], Ops[1], SV); 2249 Ops[2] = Align; 2250 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::arm_neon_vst1, 2251 Ops[1]->getType()), Ops); 2252 } 2253 // fall through 2254 case ARM::BI__builtin_neon_vst1_lane_v: { 2255 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 2256 Ops[1] = Builder.CreateExtractElement(Ops[1], Ops[2]); 2257 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 2258 StoreInst *St = Builder.CreateStore(Ops[1], 2259 Builder.CreateBitCast(Ops[0], Ty)); 2260 St->setAlignment(cast<ConstantInt>(Align)->getZExtValue()); 2261 return St; 2262 } 2263 case ARM::BI__builtin_neon_vst2_v: 2264 case ARM::BI__builtin_neon_vst2q_v: 2265 Ops.push_back(Align); 2266 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vst2, Ty), 2267 Ops, ""); 2268 case ARM::BI__builtin_neon_vst2_lane_v: 2269 case ARM::BI__builtin_neon_vst2q_lane_v: 2270 Ops.push_back(Align); 2271 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vst2lane, Ty), 2272 Ops, ""); 2273 case ARM::BI__builtin_neon_vst3_v: 2274 case ARM::BI__builtin_neon_vst3q_v: 2275 Ops.push_back(Align); 2276 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vst3, Ty), 2277 Ops, ""); 2278 case ARM::BI__builtin_neon_vst3_lane_v: 2279 case ARM::BI__builtin_neon_vst3q_lane_v: 2280 Ops.push_back(Align); 2281 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vst3lane, Ty), 2282 Ops, ""); 2283 case ARM::BI__builtin_neon_vst4_v: 2284 case ARM::BI__builtin_neon_vst4q_v: 2285 Ops.push_back(Align); 2286 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vst4, Ty), 2287 Ops, ""); 2288 case ARM::BI__builtin_neon_vst4_lane_v: 2289 case ARM::BI__builtin_neon_vst4q_lane_v: 2290 Ops.push_back(Align); 2291 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vst4lane, Ty), 2292 Ops, ""); 2293 case ARM::BI__builtin_neon_vsubhn_v: 2294 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vsubhn, Ty), 2295 Ops, "vsubhn"); 2296 case ARM::BI__builtin_neon_vtbl1_v: 2297 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbl1), 2298 Ops, "vtbl1"); 2299 case ARM::BI__builtin_neon_vtbl2_v: 2300 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbl2), 2301 Ops, "vtbl2"); 2302 case ARM::BI__builtin_neon_vtbl3_v: 2303 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbl3), 2304 Ops, "vtbl3"); 2305 case ARM::BI__builtin_neon_vtbl4_v: 2306 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbl4), 2307 Ops, "vtbl4"); 2308 case ARM::BI__builtin_neon_vtbx1_v: 2309 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbx1), 2310 Ops, "vtbx1"); 2311 case ARM::BI__builtin_neon_vtbx2_v: 2312 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbx2), 2313 Ops, "vtbx2"); 2314 case ARM::BI__builtin_neon_vtbx3_v: 2315 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbx3), 2316 Ops, "vtbx3"); 2317 case ARM::BI__builtin_neon_vtbx4_v: 2318 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbx4), 2319 Ops, "vtbx4"); 2320 case ARM::BI__builtin_neon_vtst_v: 2321 case ARM::BI__builtin_neon_vtstq_v: { 2322 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 2323 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 2324 Ops[0] = Builder.CreateAnd(Ops[0], Ops[1]); 2325 Ops[0] = Builder.CreateICmp(ICmpInst::ICMP_NE, Ops[0], 2326 ConstantAggregateZero::get(Ty)); 2327 return Builder.CreateSExt(Ops[0], Ty, "vtst"); 2328 } 2329 case ARM::BI__builtin_neon_vtrn_v: 2330 case ARM::BI__builtin_neon_vtrnq_v: { 2331 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty)); 2332 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 2333 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 2334 Value *SV = 0; 2335 2336 for (unsigned vi = 0; vi != 2; ++vi) { 2337 SmallVector<Constant*, 16> Indices; 2338 for (unsigned i = 0, e = VTy->getNumElements(); i != e; i += 2) { 2339 Indices.push_back(Builder.getInt32(i+vi)); 2340 Indices.push_back(Builder.getInt32(i+e+vi)); 2341 } 2342 Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ops[0], vi); 2343 SV = llvm::ConstantVector::get(Indices); 2344 SV = Builder.CreateShuffleVector(Ops[1], Ops[2], SV, "vtrn"); 2345 SV = Builder.CreateStore(SV, Addr); 2346 } 2347 return SV; 2348 } 2349 case ARM::BI__builtin_neon_vuzp_v: 2350 case ARM::BI__builtin_neon_vuzpq_v: { 2351 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty)); 2352 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 2353 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 2354 Value *SV = 0; 2355 2356 for (unsigned vi = 0; vi != 2; ++vi) { 2357 SmallVector<Constant*, 16> Indices; 2358 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) 2359 Indices.push_back(ConstantInt::get(Int32Ty, 2*i+vi)); 2360 2361 Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ops[0], vi); 2362 SV = llvm::ConstantVector::get(Indices); 2363 SV = Builder.CreateShuffleVector(Ops[1], Ops[2], SV, "vuzp"); 2364 SV = Builder.CreateStore(SV, Addr); 2365 } 2366 return SV; 2367 } 2368 case ARM::BI__builtin_neon_vzip_v: 2369 case ARM::BI__builtin_neon_vzipq_v: { 2370 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty)); 2371 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 2372 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 2373 Value *SV = 0; 2374 2375 for (unsigned vi = 0; vi != 2; ++vi) { 2376 SmallVector<Constant*, 16> Indices; 2377 for (unsigned i = 0, e = VTy->getNumElements(); i != e; i += 2) { 2378 Indices.push_back(ConstantInt::get(Int32Ty, (i + vi*e) >> 1)); 2379 Indices.push_back(ConstantInt::get(Int32Ty, ((i + vi*e) >> 1)+e)); 2380 } 2381 Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ops[0], vi); 2382 SV = llvm::ConstantVector::get(Indices); 2383 SV = Builder.CreateShuffleVector(Ops[1], Ops[2], SV, "vzip"); 2384 SV = Builder.CreateStore(SV, Addr); 2385 } 2386 return SV; 2387 } 2388 } 2389 } 2390 2391 llvm::Value *CodeGenFunction:: 2392 BuildVector(ArrayRef<llvm::Value*> Ops) { 2393 assert((Ops.size() & (Ops.size() - 1)) == 0 && 2394 "Not a power-of-two sized vector!"); 2395 bool AllConstants = true; 2396 for (unsigned i = 0, e = Ops.size(); i != e && AllConstants; ++i) 2397 AllConstants &= isa<Constant>(Ops[i]); 2398 2399 // If this is a constant vector, create a ConstantVector. 2400 if (AllConstants) { 2401 SmallVector<llvm::Constant*, 16> CstOps; 2402 for (unsigned i = 0, e = Ops.size(); i != e; ++i) 2403 CstOps.push_back(cast<Constant>(Ops[i])); 2404 return llvm::ConstantVector::get(CstOps); 2405 } 2406 2407 // Otherwise, insertelement the values to build the vector. 2408 Value *Result = 2409 llvm::UndefValue::get(llvm::VectorType::get(Ops[0]->getType(), Ops.size())); 2410 2411 for (unsigned i = 0, e = Ops.size(); i != e; ++i) 2412 Result = Builder.CreateInsertElement(Result, Ops[i], Builder.getInt32(i)); 2413 2414 return Result; 2415 } 2416 2417 Value *CodeGenFunction::EmitX86BuiltinExpr(unsigned BuiltinID, 2418 const CallExpr *E) { 2419 SmallVector<Value*, 4> Ops; 2420 2421 // Find out if any arguments are required to be integer constant expressions. 2422 unsigned ICEArguments = 0; 2423 ASTContext::GetBuiltinTypeError Error; 2424 getContext().GetBuiltinType(BuiltinID, Error, &ICEArguments); 2425 assert(Error == ASTContext::GE_None && "Should not codegen an error"); 2426 2427 for (unsigned i = 0, e = E->getNumArgs(); i != e; i++) { 2428 // If this is a normal argument, just emit it as a scalar. 2429 if ((ICEArguments & (1 << i)) == 0) { 2430 Ops.push_back(EmitScalarExpr(E->getArg(i))); 2431 continue; 2432 } 2433 2434 // If this is required to be a constant, constant fold it so that we know 2435 // that the generated intrinsic gets a ConstantInt. 2436 llvm::APSInt Result; 2437 bool IsConst = E->getArg(i)->isIntegerConstantExpr(Result, getContext()); 2438 assert(IsConst && "Constant arg isn't actually constant?"); (void)IsConst; 2439 Ops.push_back(llvm::ConstantInt::get(getLLVMContext(), Result)); 2440 } 2441 2442 switch (BuiltinID) { 2443 default: return 0; 2444 case X86::BI__builtin_ia32_vec_init_v8qi: 2445 case X86::BI__builtin_ia32_vec_init_v4hi: 2446 case X86::BI__builtin_ia32_vec_init_v2si: 2447 return Builder.CreateBitCast(BuildVector(Ops), 2448 llvm::Type::getX86_MMXTy(getLLVMContext())); 2449 case X86::BI__builtin_ia32_vec_ext_v2si: 2450 return Builder.CreateExtractElement(Ops[0], 2451 llvm::ConstantInt::get(Ops[1]->getType(), 0)); 2452 case X86::BI__builtin_ia32_ldmxcsr: { 2453 llvm::Type *PtrTy = Int8PtrTy; 2454 Value *One = llvm::ConstantInt::get(Int32Ty, 1); 2455 Value *Tmp = Builder.CreateAlloca(Int32Ty, One); 2456 Builder.CreateStore(Ops[0], Tmp); 2457 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::x86_sse_ldmxcsr), 2458 Builder.CreateBitCast(Tmp, PtrTy)); 2459 } 2460 case X86::BI__builtin_ia32_stmxcsr: { 2461 llvm::Type *PtrTy = Int8PtrTy; 2462 Value *One = llvm::ConstantInt::get(Int32Ty, 1); 2463 Value *Tmp = Builder.CreateAlloca(Int32Ty, One); 2464 Builder.CreateCall(CGM.getIntrinsic(Intrinsic::x86_sse_stmxcsr), 2465 Builder.CreateBitCast(Tmp, PtrTy)); 2466 return Builder.CreateLoad(Tmp, "stmxcsr"); 2467 } 2468 case X86::BI__builtin_ia32_storehps: 2469 case X86::BI__builtin_ia32_storelps: { 2470 llvm::Type *PtrTy = llvm::PointerType::getUnqual(Int64Ty); 2471 llvm::Type *VecTy = llvm::VectorType::get(Int64Ty, 2); 2472 2473 // cast val v2i64 2474 Ops[1] = Builder.CreateBitCast(Ops[1], VecTy, "cast"); 2475 2476 // extract (0, 1) 2477 unsigned Index = BuiltinID == X86::BI__builtin_ia32_storelps ? 0 : 1; 2478 llvm::Value *Idx = llvm::ConstantInt::get(Int32Ty, Index); 2479 Ops[1] = Builder.CreateExtractElement(Ops[1], Idx, "extract"); 2480 2481 // cast pointer to i64 & store 2482 Ops[0] = Builder.CreateBitCast(Ops[0], PtrTy); 2483 return Builder.CreateStore(Ops[1], Ops[0]); 2484 } 2485 case X86::BI__builtin_ia32_palignr: { 2486 unsigned shiftVal = cast<llvm::ConstantInt>(Ops[2])->getZExtValue(); 2487 2488 // If palignr is shifting the pair of input vectors less than 9 bytes, 2489 // emit a shuffle instruction. 2490 if (shiftVal <= 8) { 2491 SmallVector<llvm::Constant*, 8> Indices; 2492 for (unsigned i = 0; i != 8; ++i) 2493 Indices.push_back(llvm::ConstantInt::get(Int32Ty, shiftVal + i)); 2494 2495 Value* SV = llvm::ConstantVector::get(Indices); 2496 return Builder.CreateShuffleVector(Ops[1], Ops[0], SV, "palignr"); 2497 } 2498 2499 // If palignr is shifting the pair of input vectors more than 8 but less 2500 // than 16 bytes, emit a logical right shift of the destination. 2501 if (shiftVal < 16) { 2502 // MMX has these as 1 x i64 vectors for some odd optimization reasons. 2503 llvm::Type *VecTy = llvm::VectorType::get(Int64Ty, 1); 2504 2505 Ops[0] = Builder.CreateBitCast(Ops[0], VecTy, "cast"); 2506 Ops[1] = llvm::ConstantInt::get(VecTy, (shiftVal-8) * 8); 2507 2508 // create i32 constant 2509 llvm::Function *F = CGM.getIntrinsic(Intrinsic::x86_mmx_psrl_q); 2510 return Builder.CreateCall(F, makeArrayRef(&Ops[0], 2), "palignr"); 2511 } 2512 2513 // If palignr is shifting the pair of vectors more than 16 bytes, emit zero. 2514 return llvm::Constant::getNullValue(ConvertType(E->getType())); 2515 } 2516 case X86::BI__builtin_ia32_palignr128: { 2517 unsigned shiftVal = cast<llvm::ConstantInt>(Ops[2])->getZExtValue(); 2518 2519 // If palignr is shifting the pair of input vectors less than 17 bytes, 2520 // emit a shuffle instruction. 2521 if (shiftVal <= 16) { 2522 SmallVector<llvm::Constant*, 16> Indices; 2523 for (unsigned i = 0; i != 16; ++i) 2524 Indices.push_back(llvm::ConstantInt::get(Int32Ty, shiftVal + i)); 2525 2526 Value* SV = llvm::ConstantVector::get(Indices); 2527 return Builder.CreateShuffleVector(Ops[1], Ops[0], SV, "palignr"); 2528 } 2529 2530 // If palignr is shifting the pair of input vectors more than 16 but less 2531 // than 32 bytes, emit a logical right shift of the destination. 2532 if (shiftVal < 32) { 2533 llvm::Type *VecTy = llvm::VectorType::get(Int64Ty, 2); 2534 2535 Ops[0] = Builder.CreateBitCast(Ops[0], VecTy, "cast"); 2536 Ops[1] = llvm::ConstantInt::get(Int32Ty, (shiftVal-16) * 8); 2537 2538 // create i32 constant 2539 llvm::Function *F = CGM.getIntrinsic(Intrinsic::x86_sse2_psrl_dq); 2540 return Builder.CreateCall(F, makeArrayRef(&Ops[0], 2), "palignr"); 2541 } 2542 2543 // If palignr is shifting the pair of vectors more than 32 bytes, emit zero. 2544 return llvm::Constant::getNullValue(ConvertType(E->getType())); 2545 } 2546 case X86::BI__builtin_ia32_palignr256: { 2547 unsigned shiftVal = cast<llvm::ConstantInt>(Ops[2])->getZExtValue(); 2548 2549 // If palignr is shifting the pair of input vectors less than 17 bytes, 2550 // emit a shuffle instruction. 2551 if (shiftVal <= 16) { 2552 SmallVector<llvm::Constant*, 32> Indices; 2553 // 256-bit palignr operates on 128-bit lanes so we need to handle that 2554 for (unsigned l = 0; l != 2; ++l) { 2555 unsigned LaneStart = l * 16; 2556 unsigned LaneEnd = (l+1) * 16; 2557 for (unsigned i = 0; i != 16; ++i) { 2558 unsigned Idx = shiftVal + i + LaneStart; 2559 if (Idx >= LaneEnd) Idx += 16; // end of lane, switch operand 2560 Indices.push_back(llvm::ConstantInt::get(Int32Ty, Idx)); 2561 } 2562 } 2563 2564 Value* SV = llvm::ConstantVector::get(Indices); 2565 return Builder.CreateShuffleVector(Ops[1], Ops[0], SV, "palignr"); 2566 } 2567 2568 // If palignr is shifting the pair of input vectors more than 16 but less 2569 // than 32 bytes, emit a logical right shift of the destination. 2570 if (shiftVal < 32) { 2571 llvm::Type *VecTy = llvm::VectorType::get(Int64Ty, 4); 2572 2573 Ops[0] = Builder.CreateBitCast(Ops[0], VecTy, "cast"); 2574 Ops[1] = llvm::ConstantInt::get(Int32Ty, (shiftVal-16) * 8); 2575 2576 // create i32 constant 2577 llvm::Function *F = CGM.getIntrinsic(Intrinsic::x86_avx2_psrl_dq); 2578 return Builder.CreateCall(F, makeArrayRef(&Ops[0], 2), "palignr"); 2579 } 2580 2581 // If palignr is shifting the pair of vectors more than 32 bytes, emit zero. 2582 return llvm::Constant::getNullValue(ConvertType(E->getType())); 2583 } 2584 case X86::BI__builtin_ia32_movntps: 2585 case X86::BI__builtin_ia32_movntps256: 2586 case X86::BI__builtin_ia32_movntpd: 2587 case X86::BI__builtin_ia32_movntpd256: 2588 case X86::BI__builtin_ia32_movntdq: 2589 case X86::BI__builtin_ia32_movntdq256: 2590 case X86::BI__builtin_ia32_movnti: { 2591 llvm::MDNode *Node = llvm::MDNode::get(getLLVMContext(), 2592 Builder.getInt32(1)); 2593 2594 // Convert the type of the pointer to a pointer to the stored type. 2595 Value *BC = Builder.CreateBitCast(Ops[0], 2596 llvm::PointerType::getUnqual(Ops[1]->getType()), 2597 "cast"); 2598 StoreInst *SI = Builder.CreateStore(Ops[1], BC); 2599 SI->setMetadata(CGM.getModule().getMDKindID("nontemporal"), Node); 2600 SI->setAlignment(16); 2601 return SI; 2602 } 2603 // 3DNow! 2604 case X86::BI__builtin_ia32_pswapdsf: 2605 case X86::BI__builtin_ia32_pswapdsi: { 2606 const char *name = 0; 2607 Intrinsic::ID ID = Intrinsic::not_intrinsic; 2608 switch(BuiltinID) { 2609 default: llvm_unreachable("Unsupported intrinsic!"); 2610 case X86::BI__builtin_ia32_pswapdsf: 2611 case X86::BI__builtin_ia32_pswapdsi: 2612 name = "pswapd"; 2613 ID = Intrinsic::x86_3dnowa_pswapd; 2614 break; 2615 } 2616 llvm::Type *MMXTy = llvm::Type::getX86_MMXTy(getLLVMContext()); 2617 Ops[0] = Builder.CreateBitCast(Ops[0], MMXTy, "cast"); 2618 llvm::Function *F = CGM.getIntrinsic(ID); 2619 return Builder.CreateCall(F, Ops, name); 2620 } 2621 case X86::BI__builtin_ia32_rdrand16_step: 2622 case X86::BI__builtin_ia32_rdrand32_step: 2623 case X86::BI__builtin_ia32_rdrand64_step: { 2624 Intrinsic::ID ID; 2625 switch (BuiltinID) { 2626 default: llvm_unreachable("Unsupported intrinsic!"); 2627 case X86::BI__builtin_ia32_rdrand16_step: 2628 ID = Intrinsic::x86_rdrand_16; 2629 break; 2630 case X86::BI__builtin_ia32_rdrand32_step: 2631 ID = Intrinsic::x86_rdrand_32; 2632 break; 2633 case X86::BI__builtin_ia32_rdrand64_step: 2634 ID = Intrinsic::x86_rdrand_64; 2635 break; 2636 } 2637 2638 Value *Call = Builder.CreateCall(CGM.getIntrinsic(ID)); 2639 Builder.CreateStore(Builder.CreateExtractValue(Call, 0), Ops[0]); 2640 return Builder.CreateExtractValue(Call, 1); 2641 } 2642 } 2643 } 2644 2645 2646 Value *CodeGenFunction::EmitPPCBuiltinExpr(unsigned BuiltinID, 2647 const CallExpr *E) { 2648 SmallVector<Value*, 4> Ops; 2649 2650 for (unsigned i = 0, e = E->getNumArgs(); i != e; i++) 2651 Ops.push_back(EmitScalarExpr(E->getArg(i))); 2652 2653 Intrinsic::ID ID = Intrinsic::not_intrinsic; 2654 2655 switch (BuiltinID) { 2656 default: return 0; 2657 2658 // vec_ld, vec_lvsl, vec_lvsr 2659 case PPC::BI__builtin_altivec_lvx: 2660 case PPC::BI__builtin_altivec_lvxl: 2661 case PPC::BI__builtin_altivec_lvebx: 2662 case PPC::BI__builtin_altivec_lvehx: 2663 case PPC::BI__builtin_altivec_lvewx: 2664 case PPC::BI__builtin_altivec_lvsl: 2665 case PPC::BI__builtin_altivec_lvsr: 2666 { 2667 Ops[1] = Builder.CreateBitCast(Ops[1], Int8PtrTy); 2668 2669 Ops[0] = Builder.CreateGEP(Ops[1], Ops[0]); 2670 Ops.pop_back(); 2671 2672 switch (BuiltinID) { 2673 default: llvm_unreachable("Unsupported ld/lvsl/lvsr intrinsic!"); 2674 case PPC::BI__builtin_altivec_lvx: 2675 ID = Intrinsic::ppc_altivec_lvx; 2676 break; 2677 case PPC::BI__builtin_altivec_lvxl: 2678 ID = Intrinsic::ppc_altivec_lvxl; 2679 break; 2680 case PPC::BI__builtin_altivec_lvebx: 2681 ID = Intrinsic::ppc_altivec_lvebx; 2682 break; 2683 case PPC::BI__builtin_altivec_lvehx: 2684 ID = Intrinsic::ppc_altivec_lvehx; 2685 break; 2686 case PPC::BI__builtin_altivec_lvewx: 2687 ID = Intrinsic::ppc_altivec_lvewx; 2688 break; 2689 case PPC::BI__builtin_altivec_lvsl: 2690 ID = Intrinsic::ppc_altivec_lvsl; 2691 break; 2692 case PPC::BI__builtin_altivec_lvsr: 2693 ID = Intrinsic::ppc_altivec_lvsr; 2694 break; 2695 } 2696 llvm::Function *F = CGM.getIntrinsic(ID); 2697 return Builder.CreateCall(F, Ops, ""); 2698 } 2699 2700 // vec_st 2701 case PPC::BI__builtin_altivec_stvx: 2702 case PPC::BI__builtin_altivec_stvxl: 2703 case PPC::BI__builtin_altivec_stvebx: 2704 case PPC::BI__builtin_altivec_stvehx: 2705 case PPC::BI__builtin_altivec_stvewx: 2706 { 2707 Ops[2] = Builder.CreateBitCast(Ops[2], Int8PtrTy); 2708 Ops[1] = Builder.CreateGEP(Ops[2], Ops[1]); 2709 Ops.pop_back(); 2710 2711 switch (BuiltinID) { 2712 default: llvm_unreachable("Unsupported st intrinsic!"); 2713 case PPC::BI__builtin_altivec_stvx: 2714 ID = Intrinsic::ppc_altivec_stvx; 2715 break; 2716 case PPC::BI__builtin_altivec_stvxl: 2717 ID = Intrinsic::ppc_altivec_stvxl; 2718 break; 2719 case PPC::BI__builtin_altivec_stvebx: 2720 ID = Intrinsic::ppc_altivec_stvebx; 2721 break; 2722 case PPC::BI__builtin_altivec_stvehx: 2723 ID = Intrinsic::ppc_altivec_stvehx; 2724 break; 2725 case PPC::BI__builtin_altivec_stvewx: 2726 ID = Intrinsic::ppc_altivec_stvewx; 2727 break; 2728 } 2729 llvm::Function *F = CGM.getIntrinsic(ID); 2730 return Builder.CreateCall(F, Ops, ""); 2731 } 2732 } 2733 } 2734