CGBuiltin.cpp revision fe0af454943374758309a1066e3304b3a56af04e
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 25using namespace clang; 26using namespace CodeGen; 27using namespace llvm; 28 29/// getBuiltinLibFunction - Given a builtin id for a function like 30/// "__builtin_fabsf", return a Function* for "fabsf". 31llvm::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. 55static 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 66static 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. 79static 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. 112static 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. 146static 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.EmitNounwindRuntimeCall(Fn, V, "abs"); 164} 165 166static 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/// \brief Emit a call to llvm.{sadd,uadd,ssub,usub,smul,umul}.with.overflow.* 173/// depending on IntrinsicID. 174/// 175/// \arg CGF The current codegen function. 176/// \arg IntrinsicID The ID for the Intrinsic we wish to generate. 177/// \arg X The first argument to the llvm.*.with.overflow.*. 178/// \arg Y The second argument to the llvm.*.with.overflow.*. 179/// \arg Carry The carry returned by the llvm.*.with.overflow.*. 180/// \returns The result (i.e. sum/product) returned by the intrinsic. 181static llvm::Value *EmitOverflowIntrinsic(CodeGenFunction &CGF, 182 const llvm::Intrinsic::ID IntrinsicID, 183 llvm::Value *X, llvm::Value *Y, 184 llvm::Value *&Carry) { 185 // Make sure we have integers of the same width. 186 assert(X->getType() == Y->getType() && 187 "Arguments must be the same type. (Did you forget to make sure both " 188 "arguments have the same integer width?)"); 189 190 llvm::Value *Callee = CGF.CGM.getIntrinsic(IntrinsicID, X->getType()); 191 llvm::Value *Tmp = CGF.Builder.CreateCall2(Callee, X, Y); 192 Carry = CGF.Builder.CreateExtractValue(Tmp, 1); 193 return CGF.Builder.CreateExtractValue(Tmp, 0); 194} 195 196RValue CodeGenFunction::EmitBuiltinExpr(const FunctionDecl *FD, 197 unsigned BuiltinID, const CallExpr *E) { 198 // See if we can constant fold this builtin. If so, don't emit it at all. 199 Expr::EvalResult Result; 200 if (E->EvaluateAsRValue(Result, CGM.getContext()) && 201 !Result.hasSideEffects()) { 202 if (Result.Val.isInt()) 203 return RValue::get(llvm::ConstantInt::get(getLLVMContext(), 204 Result.Val.getInt())); 205 if (Result.Val.isFloat()) 206 return RValue::get(llvm::ConstantFP::get(getLLVMContext(), 207 Result.Val.getFloat())); 208 } 209 210 switch (BuiltinID) { 211 default: break; // Handle intrinsics and libm functions below. 212 case Builtin::BI__builtin___CFStringMakeConstantString: 213 case Builtin::BI__builtin___NSStringMakeConstantString: 214 return RValue::get(CGM.EmitConstantExpr(E, E->getType(), 0)); 215 case Builtin::BI__builtin_stdarg_start: 216 case Builtin::BI__builtin_va_start: 217 case Builtin::BI__builtin_va_end: { 218 Value *ArgValue = EmitVAListRef(E->getArg(0)); 219 llvm::Type *DestType = Int8PtrTy; 220 if (ArgValue->getType() != DestType) 221 ArgValue = Builder.CreateBitCast(ArgValue, DestType, 222 ArgValue->getName().data()); 223 224 Intrinsic::ID inst = (BuiltinID == Builtin::BI__builtin_va_end) ? 225 Intrinsic::vaend : Intrinsic::vastart; 226 return RValue::get(Builder.CreateCall(CGM.getIntrinsic(inst), ArgValue)); 227 } 228 case Builtin::BI__builtin_va_copy: { 229 Value *DstPtr = EmitVAListRef(E->getArg(0)); 230 Value *SrcPtr = EmitVAListRef(E->getArg(1)); 231 232 llvm::Type *Type = Int8PtrTy; 233 234 DstPtr = Builder.CreateBitCast(DstPtr, Type); 235 SrcPtr = Builder.CreateBitCast(SrcPtr, Type); 236 return RValue::get(Builder.CreateCall2(CGM.getIntrinsic(Intrinsic::vacopy), 237 DstPtr, SrcPtr)); 238 } 239 case Builtin::BI__builtin_abs: 240 case Builtin::BI__builtin_labs: 241 case Builtin::BI__builtin_llabs: { 242 Value *ArgValue = EmitScalarExpr(E->getArg(0)); 243 244 Value *NegOp = Builder.CreateNeg(ArgValue, "neg"); 245 Value *CmpResult = 246 Builder.CreateICmpSGE(ArgValue, 247 llvm::Constant::getNullValue(ArgValue->getType()), 248 "abscond"); 249 Value *Result = 250 Builder.CreateSelect(CmpResult, ArgValue, NegOp, "abs"); 251 252 return RValue::get(Result); 253 } 254 255 case Builtin::BI__builtin_conj: 256 case Builtin::BI__builtin_conjf: 257 case Builtin::BI__builtin_conjl: { 258 ComplexPairTy ComplexVal = EmitComplexExpr(E->getArg(0)); 259 Value *Real = ComplexVal.first; 260 Value *Imag = ComplexVal.second; 261 Value *Zero = 262 Imag->getType()->isFPOrFPVectorTy() 263 ? llvm::ConstantFP::getZeroValueForNegation(Imag->getType()) 264 : llvm::Constant::getNullValue(Imag->getType()); 265 266 Imag = Builder.CreateFSub(Zero, Imag, "sub"); 267 return RValue::getComplex(std::make_pair(Real, Imag)); 268 } 269 case Builtin::BI__builtin_creal: 270 case Builtin::BI__builtin_crealf: 271 case Builtin::BI__builtin_creall: 272 case Builtin::BIcreal: 273 case Builtin::BIcrealf: 274 case Builtin::BIcreall: { 275 ComplexPairTy ComplexVal = EmitComplexExpr(E->getArg(0)); 276 return RValue::get(ComplexVal.first); 277 } 278 279 case Builtin::BI__builtin_cimag: 280 case Builtin::BI__builtin_cimagf: 281 case Builtin::BI__builtin_cimagl: 282 case Builtin::BIcimag: 283 case Builtin::BIcimagf: 284 case Builtin::BIcimagl: { 285 ComplexPairTy ComplexVal = EmitComplexExpr(E->getArg(0)); 286 return RValue::get(ComplexVal.second); 287 } 288 289 case Builtin::BI__builtin_ctzs: 290 case Builtin::BI__builtin_ctz: 291 case Builtin::BI__builtin_ctzl: 292 case Builtin::BI__builtin_ctzll: { 293 Value *ArgValue = EmitScalarExpr(E->getArg(0)); 294 295 llvm::Type *ArgType = ArgValue->getType(); 296 Value *F = CGM.getIntrinsic(Intrinsic::cttz, ArgType); 297 298 llvm::Type *ResultType = ConvertType(E->getType()); 299 Value *ZeroUndef = Builder.getInt1(getTarget().isCLZForZeroUndef()); 300 Value *Result = Builder.CreateCall2(F, ArgValue, ZeroUndef); 301 if (Result->getType() != ResultType) 302 Result = Builder.CreateIntCast(Result, ResultType, /*isSigned*/true, 303 "cast"); 304 return RValue::get(Result); 305 } 306 case Builtin::BI__builtin_clzs: 307 case Builtin::BI__builtin_clz: 308 case Builtin::BI__builtin_clzl: 309 case Builtin::BI__builtin_clzll: { 310 Value *ArgValue = EmitScalarExpr(E->getArg(0)); 311 312 llvm::Type *ArgType = ArgValue->getType(); 313 Value *F = CGM.getIntrinsic(Intrinsic::ctlz, ArgType); 314 315 llvm::Type *ResultType = ConvertType(E->getType()); 316 Value *ZeroUndef = Builder.getInt1(getTarget().isCLZForZeroUndef()); 317 Value *Result = Builder.CreateCall2(F, ArgValue, ZeroUndef); 318 if (Result->getType() != ResultType) 319 Result = Builder.CreateIntCast(Result, ResultType, /*isSigned*/true, 320 "cast"); 321 return RValue::get(Result); 322 } 323 case Builtin::BI__builtin_ffs: 324 case Builtin::BI__builtin_ffsl: 325 case Builtin::BI__builtin_ffsll: { 326 // ffs(x) -> x ? cttz(x) + 1 : 0 327 Value *ArgValue = EmitScalarExpr(E->getArg(0)); 328 329 llvm::Type *ArgType = ArgValue->getType(); 330 Value *F = CGM.getIntrinsic(Intrinsic::cttz, ArgType); 331 332 llvm::Type *ResultType = ConvertType(E->getType()); 333 Value *Tmp = Builder.CreateAdd(Builder.CreateCall2(F, ArgValue, 334 Builder.getTrue()), 335 llvm::ConstantInt::get(ArgType, 1)); 336 Value *Zero = llvm::Constant::getNullValue(ArgType); 337 Value *IsZero = Builder.CreateICmpEQ(ArgValue, Zero, "iszero"); 338 Value *Result = Builder.CreateSelect(IsZero, Zero, Tmp, "ffs"); 339 if (Result->getType() != ResultType) 340 Result = Builder.CreateIntCast(Result, ResultType, /*isSigned*/true, 341 "cast"); 342 return RValue::get(Result); 343 } 344 case Builtin::BI__builtin_parity: 345 case Builtin::BI__builtin_parityl: 346 case Builtin::BI__builtin_parityll: { 347 // parity(x) -> ctpop(x) & 1 348 Value *ArgValue = EmitScalarExpr(E->getArg(0)); 349 350 llvm::Type *ArgType = ArgValue->getType(); 351 Value *F = CGM.getIntrinsic(Intrinsic::ctpop, ArgType); 352 353 llvm::Type *ResultType = ConvertType(E->getType()); 354 Value *Tmp = Builder.CreateCall(F, ArgValue); 355 Value *Result = Builder.CreateAnd(Tmp, llvm::ConstantInt::get(ArgType, 1)); 356 if (Result->getType() != ResultType) 357 Result = Builder.CreateIntCast(Result, ResultType, /*isSigned*/true, 358 "cast"); 359 return RValue::get(Result); 360 } 361 case Builtin::BI__builtin_popcount: 362 case Builtin::BI__builtin_popcountl: 363 case Builtin::BI__builtin_popcountll: { 364 Value *ArgValue = EmitScalarExpr(E->getArg(0)); 365 366 llvm::Type *ArgType = ArgValue->getType(); 367 Value *F = CGM.getIntrinsic(Intrinsic::ctpop, ArgType); 368 369 llvm::Type *ResultType = ConvertType(E->getType()); 370 Value *Result = Builder.CreateCall(F, ArgValue); 371 if (Result->getType() != ResultType) 372 Result = Builder.CreateIntCast(Result, ResultType, /*isSigned*/true, 373 "cast"); 374 return RValue::get(Result); 375 } 376 case Builtin::BI__builtin_expect: { 377 Value *ArgValue = EmitScalarExpr(E->getArg(0)); 378 llvm::Type *ArgType = ArgValue->getType(); 379 380 Value *FnExpect = CGM.getIntrinsic(Intrinsic::expect, ArgType); 381 Value *ExpectedValue = EmitScalarExpr(E->getArg(1)); 382 383 Value *Result = Builder.CreateCall2(FnExpect, ArgValue, ExpectedValue, 384 "expval"); 385 return RValue::get(Result); 386 } 387 case Builtin::BI__builtin_bswap16: 388 case Builtin::BI__builtin_bswap32: 389 case Builtin::BI__builtin_bswap64: { 390 Value *ArgValue = EmitScalarExpr(E->getArg(0)); 391 llvm::Type *ArgType = ArgValue->getType(); 392 Value *F = CGM.getIntrinsic(Intrinsic::bswap, ArgType); 393 return RValue::get(Builder.CreateCall(F, ArgValue)); 394 } 395 case Builtin::BI__builtin_object_size: { 396 // We rely on constant folding to deal with expressions with side effects. 397 assert(!E->getArg(0)->HasSideEffects(getContext()) && 398 "should have been constant folded"); 399 400 // We pass this builtin onto the optimizer so that it can 401 // figure out the object size in more complex cases. 402 llvm::Type *ResType = ConvertType(E->getType()); 403 404 // LLVM only supports 0 and 2, make sure that we pass along that 405 // as a boolean. 406 Value *Ty = EmitScalarExpr(E->getArg(1)); 407 ConstantInt *CI = dyn_cast<ConstantInt>(Ty); 408 assert(CI); 409 uint64_t val = CI->getZExtValue(); 410 CI = ConstantInt::get(Builder.getInt1Ty(), (val & 0x2) >> 1); 411 412 Value *F = CGM.getIntrinsic(Intrinsic::objectsize, ResType); 413 return RValue::get(Builder.CreateCall2(F, EmitScalarExpr(E->getArg(0)),CI)); 414 } 415 case Builtin::BI__builtin_prefetch: { 416 Value *Locality, *RW, *Address = EmitScalarExpr(E->getArg(0)); 417 // FIXME: Technically these constants should of type 'int', yes? 418 RW = (E->getNumArgs() > 1) ? EmitScalarExpr(E->getArg(1)) : 419 llvm::ConstantInt::get(Int32Ty, 0); 420 Locality = (E->getNumArgs() > 2) ? EmitScalarExpr(E->getArg(2)) : 421 llvm::ConstantInt::get(Int32Ty, 3); 422 Value *Data = llvm::ConstantInt::get(Int32Ty, 1); 423 Value *F = CGM.getIntrinsic(Intrinsic::prefetch); 424 return RValue::get(Builder.CreateCall4(F, Address, RW, Locality, Data)); 425 } 426 case Builtin::BI__builtin_readcyclecounter: { 427 Value *F = CGM.getIntrinsic(Intrinsic::readcyclecounter); 428 return RValue::get(Builder.CreateCall(F)); 429 } 430 case Builtin::BI__builtin_trap: { 431 Value *F = CGM.getIntrinsic(Intrinsic::trap); 432 return RValue::get(Builder.CreateCall(F)); 433 } 434 case Builtin::BI__debugbreak: { 435 Value *F = CGM.getIntrinsic(Intrinsic::debugtrap); 436 return RValue::get(Builder.CreateCall(F)); 437 } 438 case Builtin::BI__builtin_unreachable: { 439 if (SanOpts->Unreachable) 440 EmitCheck(Builder.getFalse(), "builtin_unreachable", 441 EmitCheckSourceLocation(E->getExprLoc()), 442 ArrayRef<llvm::Value *>(), CRK_Unrecoverable); 443 else 444 Builder.CreateUnreachable(); 445 446 // We do need to preserve an insertion point. 447 EmitBlock(createBasicBlock("unreachable.cont")); 448 449 return RValue::get(0); 450 } 451 452 case Builtin::BI__builtin_powi: 453 case Builtin::BI__builtin_powif: 454 case Builtin::BI__builtin_powil: { 455 Value *Base = EmitScalarExpr(E->getArg(0)); 456 Value *Exponent = EmitScalarExpr(E->getArg(1)); 457 llvm::Type *ArgType = Base->getType(); 458 Value *F = CGM.getIntrinsic(Intrinsic::powi, ArgType); 459 return RValue::get(Builder.CreateCall2(F, Base, Exponent)); 460 } 461 462 case Builtin::BI__builtin_isgreater: 463 case Builtin::BI__builtin_isgreaterequal: 464 case Builtin::BI__builtin_isless: 465 case Builtin::BI__builtin_islessequal: 466 case Builtin::BI__builtin_islessgreater: 467 case Builtin::BI__builtin_isunordered: { 468 // Ordered comparisons: we know the arguments to these are matching scalar 469 // floating point values. 470 Value *LHS = EmitScalarExpr(E->getArg(0)); 471 Value *RHS = EmitScalarExpr(E->getArg(1)); 472 473 switch (BuiltinID) { 474 default: llvm_unreachable("Unknown ordered comparison"); 475 case Builtin::BI__builtin_isgreater: 476 LHS = Builder.CreateFCmpOGT(LHS, RHS, "cmp"); 477 break; 478 case Builtin::BI__builtin_isgreaterequal: 479 LHS = Builder.CreateFCmpOGE(LHS, RHS, "cmp"); 480 break; 481 case Builtin::BI__builtin_isless: 482 LHS = Builder.CreateFCmpOLT(LHS, RHS, "cmp"); 483 break; 484 case Builtin::BI__builtin_islessequal: 485 LHS = Builder.CreateFCmpOLE(LHS, RHS, "cmp"); 486 break; 487 case Builtin::BI__builtin_islessgreater: 488 LHS = Builder.CreateFCmpONE(LHS, RHS, "cmp"); 489 break; 490 case Builtin::BI__builtin_isunordered: 491 LHS = Builder.CreateFCmpUNO(LHS, RHS, "cmp"); 492 break; 493 } 494 // ZExt bool to int type. 495 return RValue::get(Builder.CreateZExt(LHS, ConvertType(E->getType()))); 496 } 497 case Builtin::BI__builtin_isnan: { 498 Value *V = EmitScalarExpr(E->getArg(0)); 499 V = Builder.CreateFCmpUNO(V, V, "cmp"); 500 return RValue::get(Builder.CreateZExt(V, ConvertType(E->getType()))); 501 } 502 503 case Builtin::BI__builtin_isinf: { 504 // isinf(x) --> fabs(x) == infinity 505 Value *V = EmitScalarExpr(E->getArg(0)); 506 V = EmitFAbs(*this, V, E->getArg(0)->getType()); 507 508 V = Builder.CreateFCmpOEQ(V, ConstantFP::getInfinity(V->getType()),"isinf"); 509 return RValue::get(Builder.CreateZExt(V, ConvertType(E->getType()))); 510 } 511 512 // TODO: BI__builtin_isinf_sign 513 // isinf_sign(x) -> isinf(x) ? (signbit(x) ? -1 : 1) : 0 514 515 case Builtin::BI__builtin_isnormal: { 516 // isnormal(x) --> x == x && fabsf(x) < infinity && fabsf(x) >= float_min 517 Value *V = EmitScalarExpr(E->getArg(0)); 518 Value *Eq = Builder.CreateFCmpOEQ(V, V, "iseq"); 519 520 Value *Abs = EmitFAbs(*this, V, E->getArg(0)->getType()); 521 Value *IsLessThanInf = 522 Builder.CreateFCmpULT(Abs, ConstantFP::getInfinity(V->getType()),"isinf"); 523 APFloat Smallest = APFloat::getSmallestNormalized( 524 getContext().getFloatTypeSemantics(E->getArg(0)->getType())); 525 Value *IsNormal = 526 Builder.CreateFCmpUGE(Abs, ConstantFP::get(V->getContext(), Smallest), 527 "isnormal"); 528 V = Builder.CreateAnd(Eq, IsLessThanInf, "and"); 529 V = Builder.CreateAnd(V, IsNormal, "and"); 530 return RValue::get(Builder.CreateZExt(V, ConvertType(E->getType()))); 531 } 532 533 case Builtin::BI__builtin_isfinite: { 534 // isfinite(x) --> x == x && fabs(x) != infinity; 535 Value *V = EmitScalarExpr(E->getArg(0)); 536 Value *Eq = Builder.CreateFCmpOEQ(V, V, "iseq"); 537 538 Value *Abs = EmitFAbs(*this, V, E->getArg(0)->getType()); 539 Value *IsNotInf = 540 Builder.CreateFCmpUNE(Abs, ConstantFP::getInfinity(V->getType()),"isinf"); 541 542 V = Builder.CreateAnd(Eq, IsNotInf, "and"); 543 return RValue::get(Builder.CreateZExt(V, ConvertType(E->getType()))); 544 } 545 546 case Builtin::BI__builtin_fpclassify: { 547 Value *V = EmitScalarExpr(E->getArg(5)); 548 llvm::Type *Ty = ConvertType(E->getArg(5)->getType()); 549 550 // Create Result 551 BasicBlock *Begin = Builder.GetInsertBlock(); 552 BasicBlock *End = createBasicBlock("fpclassify_end", this->CurFn); 553 Builder.SetInsertPoint(End); 554 PHINode *Result = 555 Builder.CreatePHI(ConvertType(E->getArg(0)->getType()), 4, 556 "fpclassify_result"); 557 558 // if (V==0) return FP_ZERO 559 Builder.SetInsertPoint(Begin); 560 Value *IsZero = Builder.CreateFCmpOEQ(V, Constant::getNullValue(Ty), 561 "iszero"); 562 Value *ZeroLiteral = EmitScalarExpr(E->getArg(4)); 563 BasicBlock *NotZero = createBasicBlock("fpclassify_not_zero", this->CurFn); 564 Builder.CreateCondBr(IsZero, End, NotZero); 565 Result->addIncoming(ZeroLiteral, Begin); 566 567 // if (V != V) return FP_NAN 568 Builder.SetInsertPoint(NotZero); 569 Value *IsNan = Builder.CreateFCmpUNO(V, V, "cmp"); 570 Value *NanLiteral = EmitScalarExpr(E->getArg(0)); 571 BasicBlock *NotNan = createBasicBlock("fpclassify_not_nan", this->CurFn); 572 Builder.CreateCondBr(IsNan, End, NotNan); 573 Result->addIncoming(NanLiteral, NotZero); 574 575 // if (fabs(V) == infinity) return FP_INFINITY 576 Builder.SetInsertPoint(NotNan); 577 Value *VAbs = EmitFAbs(*this, V, E->getArg(5)->getType()); 578 Value *IsInf = 579 Builder.CreateFCmpOEQ(VAbs, ConstantFP::getInfinity(V->getType()), 580 "isinf"); 581 Value *InfLiteral = EmitScalarExpr(E->getArg(1)); 582 BasicBlock *NotInf = createBasicBlock("fpclassify_not_inf", this->CurFn); 583 Builder.CreateCondBr(IsInf, End, NotInf); 584 Result->addIncoming(InfLiteral, NotNan); 585 586 // if (fabs(V) >= MIN_NORMAL) return FP_NORMAL else FP_SUBNORMAL 587 Builder.SetInsertPoint(NotInf); 588 APFloat Smallest = APFloat::getSmallestNormalized( 589 getContext().getFloatTypeSemantics(E->getArg(5)->getType())); 590 Value *IsNormal = 591 Builder.CreateFCmpUGE(VAbs, ConstantFP::get(V->getContext(), Smallest), 592 "isnormal"); 593 Value *NormalResult = 594 Builder.CreateSelect(IsNormal, EmitScalarExpr(E->getArg(2)), 595 EmitScalarExpr(E->getArg(3))); 596 Builder.CreateBr(End); 597 Result->addIncoming(NormalResult, NotInf); 598 599 // return Result 600 Builder.SetInsertPoint(End); 601 return RValue::get(Result); 602 } 603 604 case Builtin::BIalloca: 605 case Builtin::BI__builtin_alloca: { 606 Value *Size = EmitScalarExpr(E->getArg(0)); 607 return RValue::get(Builder.CreateAlloca(Builder.getInt8Ty(), Size)); 608 } 609 case Builtin::BIbzero: 610 case Builtin::BI__builtin_bzero: { 611 std::pair<llvm::Value*, unsigned> Dest = 612 EmitPointerWithAlignment(E->getArg(0)); 613 Value *SizeVal = EmitScalarExpr(E->getArg(1)); 614 Builder.CreateMemSet(Dest.first, Builder.getInt8(0), SizeVal, 615 Dest.second, false); 616 return RValue::get(Dest.first); 617 } 618 case Builtin::BImemcpy: 619 case Builtin::BI__builtin_memcpy: { 620 std::pair<llvm::Value*, unsigned> Dest = 621 EmitPointerWithAlignment(E->getArg(0)); 622 std::pair<llvm::Value*, unsigned> Src = 623 EmitPointerWithAlignment(E->getArg(1)); 624 Value *SizeVal = EmitScalarExpr(E->getArg(2)); 625 unsigned Align = std::min(Dest.second, Src.second); 626 Builder.CreateMemCpy(Dest.first, Src.first, SizeVal, Align, false); 627 return RValue::get(Dest.first); 628 } 629 630 case Builtin::BI__builtin___memcpy_chk: { 631 // fold __builtin_memcpy_chk(x, y, cst1, cst2) to memcpy iff cst1<=cst2. 632 llvm::APSInt Size, DstSize; 633 if (!E->getArg(2)->EvaluateAsInt(Size, CGM.getContext()) || 634 !E->getArg(3)->EvaluateAsInt(DstSize, CGM.getContext())) 635 break; 636 if (Size.ugt(DstSize)) 637 break; 638 std::pair<llvm::Value*, unsigned> Dest = 639 EmitPointerWithAlignment(E->getArg(0)); 640 std::pair<llvm::Value*, unsigned> Src = 641 EmitPointerWithAlignment(E->getArg(1)); 642 Value *SizeVal = llvm::ConstantInt::get(Builder.getContext(), Size); 643 unsigned Align = std::min(Dest.second, Src.second); 644 Builder.CreateMemCpy(Dest.first, Src.first, SizeVal, Align, false); 645 return RValue::get(Dest.first); 646 } 647 648 case Builtin::BI__builtin_objc_memmove_collectable: { 649 Value *Address = EmitScalarExpr(E->getArg(0)); 650 Value *SrcAddr = EmitScalarExpr(E->getArg(1)); 651 Value *SizeVal = EmitScalarExpr(E->getArg(2)); 652 CGM.getObjCRuntime().EmitGCMemmoveCollectable(*this, 653 Address, SrcAddr, SizeVal); 654 return RValue::get(Address); 655 } 656 657 case Builtin::BI__builtin___memmove_chk: { 658 // fold __builtin_memmove_chk(x, y, cst1, cst2) to memmove iff cst1<=cst2. 659 llvm::APSInt Size, DstSize; 660 if (!E->getArg(2)->EvaluateAsInt(Size, CGM.getContext()) || 661 !E->getArg(3)->EvaluateAsInt(DstSize, CGM.getContext())) 662 break; 663 if (Size.ugt(DstSize)) 664 break; 665 std::pair<llvm::Value*, unsigned> Dest = 666 EmitPointerWithAlignment(E->getArg(0)); 667 std::pair<llvm::Value*, unsigned> Src = 668 EmitPointerWithAlignment(E->getArg(1)); 669 Value *SizeVal = llvm::ConstantInt::get(Builder.getContext(), Size); 670 unsigned Align = std::min(Dest.second, Src.second); 671 Builder.CreateMemMove(Dest.first, Src.first, SizeVal, Align, false); 672 return RValue::get(Dest.first); 673 } 674 675 case Builtin::BImemmove: 676 case Builtin::BI__builtin_memmove: { 677 std::pair<llvm::Value*, unsigned> Dest = 678 EmitPointerWithAlignment(E->getArg(0)); 679 std::pair<llvm::Value*, unsigned> Src = 680 EmitPointerWithAlignment(E->getArg(1)); 681 Value *SizeVal = EmitScalarExpr(E->getArg(2)); 682 unsigned Align = std::min(Dest.second, Src.second); 683 Builder.CreateMemMove(Dest.first, Src.first, SizeVal, Align, false); 684 return RValue::get(Dest.first); 685 } 686 case Builtin::BImemset: 687 case Builtin::BI__builtin_memset: { 688 std::pair<llvm::Value*, unsigned> Dest = 689 EmitPointerWithAlignment(E->getArg(0)); 690 Value *ByteVal = Builder.CreateTrunc(EmitScalarExpr(E->getArg(1)), 691 Builder.getInt8Ty()); 692 Value *SizeVal = EmitScalarExpr(E->getArg(2)); 693 Builder.CreateMemSet(Dest.first, ByteVal, SizeVal, Dest.second, false); 694 return RValue::get(Dest.first); 695 } 696 case Builtin::BI__builtin___memset_chk: { 697 // fold __builtin_memset_chk(x, y, cst1, cst2) to memset iff cst1<=cst2. 698 llvm::APSInt Size, DstSize; 699 if (!E->getArg(2)->EvaluateAsInt(Size, CGM.getContext()) || 700 !E->getArg(3)->EvaluateAsInt(DstSize, CGM.getContext())) 701 break; 702 if (Size.ugt(DstSize)) 703 break; 704 std::pair<llvm::Value*, unsigned> Dest = 705 EmitPointerWithAlignment(E->getArg(0)); 706 Value *ByteVal = Builder.CreateTrunc(EmitScalarExpr(E->getArg(1)), 707 Builder.getInt8Ty()); 708 Value *SizeVal = llvm::ConstantInt::get(Builder.getContext(), Size); 709 Builder.CreateMemSet(Dest.first, ByteVal, SizeVal, Dest.second, false); 710 return RValue::get(Dest.first); 711 } 712 case Builtin::BI__builtin_dwarf_cfa: { 713 // The offset in bytes from the first argument to the CFA. 714 // 715 // Why on earth is this in the frontend? Is there any reason at 716 // all that the backend can't reasonably determine this while 717 // lowering llvm.eh.dwarf.cfa()? 718 // 719 // TODO: If there's a satisfactory reason, add a target hook for 720 // this instead of hard-coding 0, which is correct for most targets. 721 int32_t Offset = 0; 722 723 Value *F = CGM.getIntrinsic(Intrinsic::eh_dwarf_cfa); 724 return RValue::get(Builder.CreateCall(F, 725 llvm::ConstantInt::get(Int32Ty, Offset))); 726 } 727 case Builtin::BI__builtin_return_address: { 728 Value *Depth = EmitScalarExpr(E->getArg(0)); 729 Depth = Builder.CreateIntCast(Depth, Int32Ty, false); 730 Value *F = CGM.getIntrinsic(Intrinsic::returnaddress); 731 return RValue::get(Builder.CreateCall(F, Depth)); 732 } 733 case Builtin::BI__builtin_frame_address: { 734 Value *Depth = EmitScalarExpr(E->getArg(0)); 735 Depth = Builder.CreateIntCast(Depth, Int32Ty, false); 736 Value *F = CGM.getIntrinsic(Intrinsic::frameaddress); 737 return RValue::get(Builder.CreateCall(F, Depth)); 738 } 739 case Builtin::BI__builtin_extract_return_addr: { 740 Value *Address = EmitScalarExpr(E->getArg(0)); 741 Value *Result = getTargetHooks().decodeReturnAddress(*this, Address); 742 return RValue::get(Result); 743 } 744 case Builtin::BI__builtin_frob_return_addr: { 745 Value *Address = EmitScalarExpr(E->getArg(0)); 746 Value *Result = getTargetHooks().encodeReturnAddress(*this, Address); 747 return RValue::get(Result); 748 } 749 case Builtin::BI__builtin_dwarf_sp_column: { 750 llvm::IntegerType *Ty 751 = cast<llvm::IntegerType>(ConvertType(E->getType())); 752 int Column = getTargetHooks().getDwarfEHStackPointer(CGM); 753 if (Column == -1) { 754 CGM.ErrorUnsupported(E, "__builtin_dwarf_sp_column"); 755 return RValue::get(llvm::UndefValue::get(Ty)); 756 } 757 return RValue::get(llvm::ConstantInt::get(Ty, Column, true)); 758 } 759 case Builtin::BI__builtin_init_dwarf_reg_size_table: { 760 Value *Address = EmitScalarExpr(E->getArg(0)); 761 if (getTargetHooks().initDwarfEHRegSizeTable(*this, Address)) 762 CGM.ErrorUnsupported(E, "__builtin_init_dwarf_reg_size_table"); 763 return RValue::get(llvm::UndefValue::get(ConvertType(E->getType()))); 764 } 765 case Builtin::BI__builtin_eh_return: { 766 Value *Int = EmitScalarExpr(E->getArg(0)); 767 Value *Ptr = EmitScalarExpr(E->getArg(1)); 768 769 llvm::IntegerType *IntTy = cast<llvm::IntegerType>(Int->getType()); 770 assert((IntTy->getBitWidth() == 32 || IntTy->getBitWidth() == 64) && 771 "LLVM's __builtin_eh_return only supports 32- and 64-bit variants"); 772 Value *F = CGM.getIntrinsic(IntTy->getBitWidth() == 32 773 ? Intrinsic::eh_return_i32 774 : Intrinsic::eh_return_i64); 775 Builder.CreateCall2(F, Int, Ptr); 776 Builder.CreateUnreachable(); 777 778 // We do need to preserve an insertion point. 779 EmitBlock(createBasicBlock("builtin_eh_return.cont")); 780 781 return RValue::get(0); 782 } 783 case Builtin::BI__builtin_unwind_init: { 784 Value *F = CGM.getIntrinsic(Intrinsic::eh_unwind_init); 785 return RValue::get(Builder.CreateCall(F)); 786 } 787 case Builtin::BI__builtin_extend_pointer: { 788 // Extends a pointer to the size of an _Unwind_Word, which is 789 // uint64_t on all platforms. Generally this gets poked into a 790 // register and eventually used as an address, so if the 791 // addressing registers are wider than pointers and the platform 792 // doesn't implicitly ignore high-order bits when doing 793 // addressing, we need to make sure we zext / sext based on 794 // the platform's expectations. 795 // 796 // See: http://gcc.gnu.org/ml/gcc-bugs/2002-02/msg00237.html 797 798 // Cast the pointer to intptr_t. 799 Value *Ptr = EmitScalarExpr(E->getArg(0)); 800 Value *Result = Builder.CreatePtrToInt(Ptr, IntPtrTy, "extend.cast"); 801 802 // If that's 64 bits, we're done. 803 if (IntPtrTy->getBitWidth() == 64) 804 return RValue::get(Result); 805 806 // Otherwise, ask the codegen data what to do. 807 if (getTargetHooks().extendPointerWithSExt()) 808 return RValue::get(Builder.CreateSExt(Result, Int64Ty, "extend.sext")); 809 else 810 return RValue::get(Builder.CreateZExt(Result, Int64Ty, "extend.zext")); 811 } 812 case Builtin::BI__builtin_setjmp: { 813 // Buffer is a void**. 814 Value *Buf = EmitScalarExpr(E->getArg(0)); 815 816 // Store the frame pointer to the setjmp buffer. 817 Value *FrameAddr = 818 Builder.CreateCall(CGM.getIntrinsic(Intrinsic::frameaddress), 819 ConstantInt::get(Int32Ty, 0)); 820 Builder.CreateStore(FrameAddr, Buf); 821 822 // Store the stack pointer to the setjmp buffer. 823 Value *StackAddr = 824 Builder.CreateCall(CGM.getIntrinsic(Intrinsic::stacksave)); 825 Value *StackSaveSlot = 826 Builder.CreateGEP(Buf, ConstantInt::get(Int32Ty, 2)); 827 Builder.CreateStore(StackAddr, StackSaveSlot); 828 829 // Call LLVM's EH setjmp, which is lightweight. 830 Value *F = CGM.getIntrinsic(Intrinsic::eh_sjlj_setjmp); 831 Buf = Builder.CreateBitCast(Buf, Int8PtrTy); 832 return RValue::get(Builder.CreateCall(F, Buf)); 833 } 834 case Builtin::BI__builtin_longjmp: { 835 Value *Buf = EmitScalarExpr(E->getArg(0)); 836 Buf = Builder.CreateBitCast(Buf, Int8PtrTy); 837 838 // Call LLVM's EH longjmp, which is lightweight. 839 Builder.CreateCall(CGM.getIntrinsic(Intrinsic::eh_sjlj_longjmp), Buf); 840 841 // longjmp doesn't return; mark this as unreachable. 842 Builder.CreateUnreachable(); 843 844 // We do need to preserve an insertion point. 845 EmitBlock(createBasicBlock("longjmp.cont")); 846 847 return RValue::get(0); 848 } 849 case Builtin::BI__sync_fetch_and_add: 850 case Builtin::BI__sync_fetch_and_sub: 851 case Builtin::BI__sync_fetch_and_or: 852 case Builtin::BI__sync_fetch_and_and: 853 case Builtin::BI__sync_fetch_and_xor: 854 case Builtin::BI__sync_add_and_fetch: 855 case Builtin::BI__sync_sub_and_fetch: 856 case Builtin::BI__sync_and_and_fetch: 857 case Builtin::BI__sync_or_and_fetch: 858 case Builtin::BI__sync_xor_and_fetch: 859 case Builtin::BI__sync_val_compare_and_swap: 860 case Builtin::BI__sync_bool_compare_and_swap: 861 case Builtin::BI__sync_lock_test_and_set: 862 case Builtin::BI__sync_lock_release: 863 case Builtin::BI__sync_swap: 864 llvm_unreachable("Shouldn't make it through sema"); 865 case Builtin::BI__sync_fetch_and_add_1: 866 case Builtin::BI__sync_fetch_and_add_2: 867 case Builtin::BI__sync_fetch_and_add_4: 868 case Builtin::BI__sync_fetch_and_add_8: 869 case Builtin::BI__sync_fetch_and_add_16: 870 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Add, E); 871 case Builtin::BI__sync_fetch_and_sub_1: 872 case Builtin::BI__sync_fetch_and_sub_2: 873 case Builtin::BI__sync_fetch_and_sub_4: 874 case Builtin::BI__sync_fetch_and_sub_8: 875 case Builtin::BI__sync_fetch_and_sub_16: 876 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Sub, E); 877 case Builtin::BI__sync_fetch_and_or_1: 878 case Builtin::BI__sync_fetch_and_or_2: 879 case Builtin::BI__sync_fetch_and_or_4: 880 case Builtin::BI__sync_fetch_and_or_8: 881 case Builtin::BI__sync_fetch_and_or_16: 882 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Or, E); 883 case Builtin::BI__sync_fetch_and_and_1: 884 case Builtin::BI__sync_fetch_and_and_2: 885 case Builtin::BI__sync_fetch_and_and_4: 886 case Builtin::BI__sync_fetch_and_and_8: 887 case Builtin::BI__sync_fetch_and_and_16: 888 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::And, E); 889 case Builtin::BI__sync_fetch_and_xor_1: 890 case Builtin::BI__sync_fetch_and_xor_2: 891 case Builtin::BI__sync_fetch_and_xor_4: 892 case Builtin::BI__sync_fetch_and_xor_8: 893 case Builtin::BI__sync_fetch_and_xor_16: 894 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Xor, E); 895 896 // Clang extensions: not overloaded yet. 897 case Builtin::BI__sync_fetch_and_min: 898 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Min, E); 899 case Builtin::BI__sync_fetch_and_max: 900 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Max, E); 901 case Builtin::BI__sync_fetch_and_umin: 902 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::UMin, E); 903 case Builtin::BI__sync_fetch_and_umax: 904 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::UMax, E); 905 906 case Builtin::BI__sync_add_and_fetch_1: 907 case Builtin::BI__sync_add_and_fetch_2: 908 case Builtin::BI__sync_add_and_fetch_4: 909 case Builtin::BI__sync_add_and_fetch_8: 910 case Builtin::BI__sync_add_and_fetch_16: 911 return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::Add, E, 912 llvm::Instruction::Add); 913 case Builtin::BI__sync_sub_and_fetch_1: 914 case Builtin::BI__sync_sub_and_fetch_2: 915 case Builtin::BI__sync_sub_and_fetch_4: 916 case Builtin::BI__sync_sub_and_fetch_8: 917 case Builtin::BI__sync_sub_and_fetch_16: 918 return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::Sub, E, 919 llvm::Instruction::Sub); 920 case Builtin::BI__sync_and_and_fetch_1: 921 case Builtin::BI__sync_and_and_fetch_2: 922 case Builtin::BI__sync_and_and_fetch_4: 923 case Builtin::BI__sync_and_and_fetch_8: 924 case Builtin::BI__sync_and_and_fetch_16: 925 return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::And, E, 926 llvm::Instruction::And); 927 case Builtin::BI__sync_or_and_fetch_1: 928 case Builtin::BI__sync_or_and_fetch_2: 929 case Builtin::BI__sync_or_and_fetch_4: 930 case Builtin::BI__sync_or_and_fetch_8: 931 case Builtin::BI__sync_or_and_fetch_16: 932 return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::Or, E, 933 llvm::Instruction::Or); 934 case Builtin::BI__sync_xor_and_fetch_1: 935 case Builtin::BI__sync_xor_and_fetch_2: 936 case Builtin::BI__sync_xor_and_fetch_4: 937 case Builtin::BI__sync_xor_and_fetch_8: 938 case Builtin::BI__sync_xor_and_fetch_16: 939 return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::Xor, E, 940 llvm::Instruction::Xor); 941 942 case Builtin::BI__sync_val_compare_and_swap_1: 943 case Builtin::BI__sync_val_compare_and_swap_2: 944 case Builtin::BI__sync_val_compare_and_swap_4: 945 case Builtin::BI__sync_val_compare_and_swap_8: 946 case Builtin::BI__sync_val_compare_and_swap_16: { 947 QualType T = E->getType(); 948 llvm::Value *DestPtr = EmitScalarExpr(E->getArg(0)); 949 unsigned AddrSpace = DestPtr->getType()->getPointerAddressSpace(); 950 951 llvm::IntegerType *IntType = 952 llvm::IntegerType::get(getLLVMContext(), 953 getContext().getTypeSize(T)); 954 llvm::Type *IntPtrType = IntType->getPointerTo(AddrSpace); 955 956 Value *Args[3]; 957 Args[0] = Builder.CreateBitCast(DestPtr, IntPtrType); 958 Args[1] = EmitScalarExpr(E->getArg(1)); 959 llvm::Type *ValueType = Args[1]->getType(); 960 Args[1] = EmitToInt(*this, Args[1], T, IntType); 961 Args[2] = EmitToInt(*this, EmitScalarExpr(E->getArg(2)), T, IntType); 962 963 Value *Result = Builder.CreateAtomicCmpXchg(Args[0], Args[1], Args[2], 964 llvm::SequentiallyConsistent); 965 Result = EmitFromInt(*this, Result, T, ValueType); 966 return RValue::get(Result); 967 } 968 969 case Builtin::BI__sync_bool_compare_and_swap_1: 970 case Builtin::BI__sync_bool_compare_and_swap_2: 971 case Builtin::BI__sync_bool_compare_and_swap_4: 972 case Builtin::BI__sync_bool_compare_and_swap_8: 973 case Builtin::BI__sync_bool_compare_and_swap_16: { 974 QualType T = E->getArg(1)->getType(); 975 llvm::Value *DestPtr = EmitScalarExpr(E->getArg(0)); 976 unsigned AddrSpace = DestPtr->getType()->getPointerAddressSpace(); 977 978 llvm::IntegerType *IntType = 979 llvm::IntegerType::get(getLLVMContext(), 980 getContext().getTypeSize(T)); 981 llvm::Type *IntPtrType = IntType->getPointerTo(AddrSpace); 982 983 Value *Args[3]; 984 Args[0] = Builder.CreateBitCast(DestPtr, IntPtrType); 985 Args[1] = EmitToInt(*this, EmitScalarExpr(E->getArg(1)), T, IntType); 986 Args[2] = EmitToInt(*this, EmitScalarExpr(E->getArg(2)), T, IntType); 987 988 Value *OldVal = Args[1]; 989 Value *PrevVal = Builder.CreateAtomicCmpXchg(Args[0], Args[1], Args[2], 990 llvm::SequentiallyConsistent); 991 Value *Result = Builder.CreateICmpEQ(PrevVal, OldVal); 992 // zext bool to int. 993 Result = Builder.CreateZExt(Result, ConvertType(E->getType())); 994 return RValue::get(Result); 995 } 996 997 case Builtin::BI__sync_swap_1: 998 case Builtin::BI__sync_swap_2: 999 case Builtin::BI__sync_swap_4: 1000 case Builtin::BI__sync_swap_8: 1001 case Builtin::BI__sync_swap_16: 1002 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Xchg, E); 1003 1004 case Builtin::BI__sync_lock_test_and_set_1: 1005 case Builtin::BI__sync_lock_test_and_set_2: 1006 case Builtin::BI__sync_lock_test_and_set_4: 1007 case Builtin::BI__sync_lock_test_and_set_8: 1008 case Builtin::BI__sync_lock_test_and_set_16: 1009 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Xchg, E); 1010 1011 case Builtin::BI__sync_lock_release_1: 1012 case Builtin::BI__sync_lock_release_2: 1013 case Builtin::BI__sync_lock_release_4: 1014 case Builtin::BI__sync_lock_release_8: 1015 case Builtin::BI__sync_lock_release_16: { 1016 Value *Ptr = EmitScalarExpr(E->getArg(0)); 1017 QualType ElTy = E->getArg(0)->getType()->getPointeeType(); 1018 CharUnits StoreSize = getContext().getTypeSizeInChars(ElTy); 1019 llvm::Type *ITy = llvm::IntegerType::get(getLLVMContext(), 1020 StoreSize.getQuantity() * 8); 1021 Ptr = Builder.CreateBitCast(Ptr, ITy->getPointerTo()); 1022 llvm::StoreInst *Store = 1023 Builder.CreateStore(llvm::Constant::getNullValue(ITy), Ptr); 1024 Store->setAlignment(StoreSize.getQuantity()); 1025 Store->setAtomic(llvm::Release); 1026 return RValue::get(0); 1027 } 1028 1029 case Builtin::BI__sync_synchronize: { 1030 // We assume this is supposed to correspond to a C++0x-style 1031 // sequentially-consistent fence (i.e. this is only usable for 1032 // synchonization, not device I/O or anything like that). This intrinsic 1033 // is really badly designed in the sense that in theory, there isn't 1034 // any way to safely use it... but in practice, it mostly works 1035 // to use it with non-atomic loads and stores to get acquire/release 1036 // semantics. 1037 Builder.CreateFence(llvm::SequentiallyConsistent); 1038 return RValue::get(0); 1039 } 1040 1041 case Builtin::BI__c11_atomic_is_lock_free: 1042 case Builtin::BI__atomic_is_lock_free: { 1043 // Call "bool __atomic_is_lock_free(size_t size, void *ptr)". For the 1044 // __c11 builtin, ptr is 0 (indicating a properly-aligned object), since 1045 // _Atomic(T) is always properly-aligned. 1046 const char *LibCallName = "__atomic_is_lock_free"; 1047 CallArgList Args; 1048 Args.add(RValue::get(EmitScalarExpr(E->getArg(0))), 1049 getContext().getSizeType()); 1050 if (BuiltinID == Builtin::BI__atomic_is_lock_free) 1051 Args.add(RValue::get(EmitScalarExpr(E->getArg(1))), 1052 getContext().VoidPtrTy); 1053 else 1054 Args.add(RValue::get(llvm::Constant::getNullValue(VoidPtrTy)), 1055 getContext().VoidPtrTy); 1056 const CGFunctionInfo &FuncInfo = 1057 CGM.getTypes().arrangeFreeFunctionCall(E->getType(), Args, 1058 FunctionType::ExtInfo(), 1059 RequiredArgs::All); 1060 llvm::FunctionType *FTy = CGM.getTypes().GetFunctionType(FuncInfo); 1061 llvm::Constant *Func = CGM.CreateRuntimeFunction(FTy, LibCallName); 1062 return EmitCall(FuncInfo, Func, ReturnValueSlot(), Args); 1063 } 1064 1065 case Builtin::BI__atomic_test_and_set: { 1066 // Look at the argument type to determine whether this is a volatile 1067 // operation. The parameter type is always volatile. 1068 QualType PtrTy = E->getArg(0)->IgnoreImpCasts()->getType(); 1069 bool Volatile = 1070 PtrTy->castAs<PointerType>()->getPointeeType().isVolatileQualified(); 1071 1072 Value *Ptr = EmitScalarExpr(E->getArg(0)); 1073 unsigned AddrSpace = Ptr->getType()->getPointerAddressSpace(); 1074 Ptr = Builder.CreateBitCast(Ptr, Int8Ty->getPointerTo(AddrSpace)); 1075 Value *NewVal = Builder.getInt8(1); 1076 Value *Order = EmitScalarExpr(E->getArg(1)); 1077 if (isa<llvm::ConstantInt>(Order)) { 1078 int ord = cast<llvm::ConstantInt>(Order)->getZExtValue(); 1079 AtomicRMWInst *Result = 0; 1080 switch (ord) { 1081 case 0: // memory_order_relaxed 1082 default: // invalid order 1083 Result = Builder.CreateAtomicRMW(llvm::AtomicRMWInst::Xchg, 1084 Ptr, NewVal, 1085 llvm::Monotonic); 1086 break; 1087 case 1: // memory_order_consume 1088 case 2: // memory_order_acquire 1089 Result = Builder.CreateAtomicRMW(llvm::AtomicRMWInst::Xchg, 1090 Ptr, NewVal, 1091 llvm::Acquire); 1092 break; 1093 case 3: // memory_order_release 1094 Result = Builder.CreateAtomicRMW(llvm::AtomicRMWInst::Xchg, 1095 Ptr, NewVal, 1096 llvm::Release); 1097 break; 1098 case 4: // memory_order_acq_rel 1099 Result = Builder.CreateAtomicRMW(llvm::AtomicRMWInst::Xchg, 1100 Ptr, NewVal, 1101 llvm::AcquireRelease); 1102 break; 1103 case 5: // memory_order_seq_cst 1104 Result = Builder.CreateAtomicRMW(llvm::AtomicRMWInst::Xchg, 1105 Ptr, NewVal, 1106 llvm::SequentiallyConsistent); 1107 break; 1108 } 1109 Result->setVolatile(Volatile); 1110 return RValue::get(Builder.CreateIsNotNull(Result, "tobool")); 1111 } 1112 1113 llvm::BasicBlock *ContBB = createBasicBlock("atomic.continue", CurFn); 1114 1115 llvm::BasicBlock *BBs[5] = { 1116 createBasicBlock("monotonic", CurFn), 1117 createBasicBlock("acquire", CurFn), 1118 createBasicBlock("release", CurFn), 1119 createBasicBlock("acqrel", CurFn), 1120 createBasicBlock("seqcst", CurFn) 1121 }; 1122 llvm::AtomicOrdering Orders[5] = { 1123 llvm::Monotonic, llvm::Acquire, llvm::Release, 1124 llvm::AcquireRelease, llvm::SequentiallyConsistent 1125 }; 1126 1127 Order = Builder.CreateIntCast(Order, Builder.getInt32Ty(), false); 1128 llvm::SwitchInst *SI = Builder.CreateSwitch(Order, BBs[0]); 1129 1130 Builder.SetInsertPoint(ContBB); 1131 PHINode *Result = Builder.CreatePHI(Int8Ty, 5, "was_set"); 1132 1133 for (unsigned i = 0; i < 5; ++i) { 1134 Builder.SetInsertPoint(BBs[i]); 1135 AtomicRMWInst *RMW = Builder.CreateAtomicRMW(llvm::AtomicRMWInst::Xchg, 1136 Ptr, NewVal, Orders[i]); 1137 RMW->setVolatile(Volatile); 1138 Result->addIncoming(RMW, BBs[i]); 1139 Builder.CreateBr(ContBB); 1140 } 1141 1142 SI->addCase(Builder.getInt32(0), BBs[0]); 1143 SI->addCase(Builder.getInt32(1), BBs[1]); 1144 SI->addCase(Builder.getInt32(2), BBs[1]); 1145 SI->addCase(Builder.getInt32(3), BBs[2]); 1146 SI->addCase(Builder.getInt32(4), BBs[3]); 1147 SI->addCase(Builder.getInt32(5), BBs[4]); 1148 1149 Builder.SetInsertPoint(ContBB); 1150 return RValue::get(Builder.CreateIsNotNull(Result, "tobool")); 1151 } 1152 1153 case Builtin::BI__atomic_clear: { 1154 QualType PtrTy = E->getArg(0)->IgnoreImpCasts()->getType(); 1155 bool Volatile = 1156 PtrTy->castAs<PointerType>()->getPointeeType().isVolatileQualified(); 1157 1158 Value *Ptr = EmitScalarExpr(E->getArg(0)); 1159 unsigned AddrSpace = Ptr->getType()->getPointerAddressSpace(); 1160 Ptr = Builder.CreateBitCast(Ptr, Int8Ty->getPointerTo(AddrSpace)); 1161 Value *NewVal = Builder.getInt8(0); 1162 Value *Order = EmitScalarExpr(E->getArg(1)); 1163 if (isa<llvm::ConstantInt>(Order)) { 1164 int ord = cast<llvm::ConstantInt>(Order)->getZExtValue(); 1165 StoreInst *Store = Builder.CreateStore(NewVal, Ptr, Volatile); 1166 Store->setAlignment(1); 1167 switch (ord) { 1168 case 0: // memory_order_relaxed 1169 default: // invalid order 1170 Store->setOrdering(llvm::Monotonic); 1171 break; 1172 case 3: // memory_order_release 1173 Store->setOrdering(llvm::Release); 1174 break; 1175 case 5: // memory_order_seq_cst 1176 Store->setOrdering(llvm::SequentiallyConsistent); 1177 break; 1178 } 1179 return RValue::get(0); 1180 } 1181 1182 llvm::BasicBlock *ContBB = createBasicBlock("atomic.continue", CurFn); 1183 1184 llvm::BasicBlock *BBs[3] = { 1185 createBasicBlock("monotonic", CurFn), 1186 createBasicBlock("release", CurFn), 1187 createBasicBlock("seqcst", CurFn) 1188 }; 1189 llvm::AtomicOrdering Orders[3] = { 1190 llvm::Monotonic, llvm::Release, llvm::SequentiallyConsistent 1191 }; 1192 1193 Order = Builder.CreateIntCast(Order, Builder.getInt32Ty(), false); 1194 llvm::SwitchInst *SI = Builder.CreateSwitch(Order, BBs[0]); 1195 1196 for (unsigned i = 0; i < 3; ++i) { 1197 Builder.SetInsertPoint(BBs[i]); 1198 StoreInst *Store = Builder.CreateStore(NewVal, Ptr, Volatile); 1199 Store->setAlignment(1); 1200 Store->setOrdering(Orders[i]); 1201 Builder.CreateBr(ContBB); 1202 } 1203 1204 SI->addCase(Builder.getInt32(0), BBs[0]); 1205 SI->addCase(Builder.getInt32(3), BBs[1]); 1206 SI->addCase(Builder.getInt32(5), BBs[2]); 1207 1208 Builder.SetInsertPoint(ContBB); 1209 return RValue::get(0); 1210 } 1211 1212 case Builtin::BI__atomic_thread_fence: 1213 case Builtin::BI__atomic_signal_fence: 1214 case Builtin::BI__c11_atomic_thread_fence: 1215 case Builtin::BI__c11_atomic_signal_fence: { 1216 llvm::SynchronizationScope Scope; 1217 if (BuiltinID == Builtin::BI__atomic_signal_fence || 1218 BuiltinID == Builtin::BI__c11_atomic_signal_fence) 1219 Scope = llvm::SingleThread; 1220 else 1221 Scope = llvm::CrossThread; 1222 Value *Order = EmitScalarExpr(E->getArg(0)); 1223 if (isa<llvm::ConstantInt>(Order)) { 1224 int ord = cast<llvm::ConstantInt>(Order)->getZExtValue(); 1225 switch (ord) { 1226 case 0: // memory_order_relaxed 1227 default: // invalid order 1228 break; 1229 case 1: // memory_order_consume 1230 case 2: // memory_order_acquire 1231 Builder.CreateFence(llvm::Acquire, Scope); 1232 break; 1233 case 3: // memory_order_release 1234 Builder.CreateFence(llvm::Release, Scope); 1235 break; 1236 case 4: // memory_order_acq_rel 1237 Builder.CreateFence(llvm::AcquireRelease, Scope); 1238 break; 1239 case 5: // memory_order_seq_cst 1240 Builder.CreateFence(llvm::SequentiallyConsistent, Scope); 1241 break; 1242 } 1243 return RValue::get(0); 1244 } 1245 1246 llvm::BasicBlock *AcquireBB, *ReleaseBB, *AcqRelBB, *SeqCstBB; 1247 AcquireBB = createBasicBlock("acquire", CurFn); 1248 ReleaseBB = createBasicBlock("release", CurFn); 1249 AcqRelBB = createBasicBlock("acqrel", CurFn); 1250 SeqCstBB = createBasicBlock("seqcst", CurFn); 1251 llvm::BasicBlock *ContBB = createBasicBlock("atomic.continue", CurFn); 1252 1253 Order = Builder.CreateIntCast(Order, Builder.getInt32Ty(), false); 1254 llvm::SwitchInst *SI = Builder.CreateSwitch(Order, ContBB); 1255 1256 Builder.SetInsertPoint(AcquireBB); 1257 Builder.CreateFence(llvm::Acquire, Scope); 1258 Builder.CreateBr(ContBB); 1259 SI->addCase(Builder.getInt32(1), AcquireBB); 1260 SI->addCase(Builder.getInt32(2), AcquireBB); 1261 1262 Builder.SetInsertPoint(ReleaseBB); 1263 Builder.CreateFence(llvm::Release, Scope); 1264 Builder.CreateBr(ContBB); 1265 SI->addCase(Builder.getInt32(3), ReleaseBB); 1266 1267 Builder.SetInsertPoint(AcqRelBB); 1268 Builder.CreateFence(llvm::AcquireRelease, Scope); 1269 Builder.CreateBr(ContBB); 1270 SI->addCase(Builder.getInt32(4), AcqRelBB); 1271 1272 Builder.SetInsertPoint(SeqCstBB); 1273 Builder.CreateFence(llvm::SequentiallyConsistent, Scope); 1274 Builder.CreateBr(ContBB); 1275 SI->addCase(Builder.getInt32(5), SeqCstBB); 1276 1277 Builder.SetInsertPoint(ContBB); 1278 return RValue::get(0); 1279 } 1280 1281 // Library functions with special handling. 1282 case Builtin::BIsqrt: 1283 case Builtin::BIsqrtf: 1284 case Builtin::BIsqrtl: { 1285 // TODO: there is currently no set of optimizer flags 1286 // sufficient for us to rewrite sqrt to @llvm.sqrt. 1287 // -fmath-errno=0 is not good enough; we need finiteness. 1288 // We could probably precondition the call with an ult 1289 // against 0, but is that worth the complexity? 1290 break; 1291 } 1292 1293 case Builtin::BIpow: 1294 case Builtin::BIpowf: 1295 case Builtin::BIpowl: { 1296 // Transform a call to pow* into a @llvm.pow.* intrinsic call. 1297 if (!FD->hasAttr<ConstAttr>()) 1298 break; 1299 Value *Base = EmitScalarExpr(E->getArg(0)); 1300 Value *Exponent = EmitScalarExpr(E->getArg(1)); 1301 llvm::Type *ArgType = Base->getType(); 1302 Value *F = CGM.getIntrinsic(Intrinsic::pow, ArgType); 1303 return RValue::get(Builder.CreateCall2(F, Base, Exponent)); 1304 break; 1305 } 1306 1307 case Builtin::BIfma: 1308 case Builtin::BIfmaf: 1309 case Builtin::BIfmal: 1310 case Builtin::BI__builtin_fma: 1311 case Builtin::BI__builtin_fmaf: 1312 case Builtin::BI__builtin_fmal: { 1313 // Rewrite fma to intrinsic. 1314 Value *FirstArg = EmitScalarExpr(E->getArg(0)); 1315 llvm::Type *ArgType = FirstArg->getType(); 1316 Value *F = CGM.getIntrinsic(Intrinsic::fma, ArgType); 1317 return RValue::get(Builder.CreateCall3(F, FirstArg, 1318 EmitScalarExpr(E->getArg(1)), 1319 EmitScalarExpr(E->getArg(2)))); 1320 } 1321 1322 case Builtin::BI__builtin_signbit: 1323 case Builtin::BI__builtin_signbitf: 1324 case Builtin::BI__builtin_signbitl: { 1325 LLVMContext &C = CGM.getLLVMContext(); 1326 1327 Value *Arg = EmitScalarExpr(E->getArg(0)); 1328 llvm::Type *ArgTy = Arg->getType(); 1329 if (ArgTy->isPPC_FP128Ty()) 1330 break; // FIXME: I'm not sure what the right implementation is here. 1331 int ArgWidth = ArgTy->getPrimitiveSizeInBits(); 1332 llvm::Type *ArgIntTy = llvm::IntegerType::get(C, ArgWidth); 1333 Value *BCArg = Builder.CreateBitCast(Arg, ArgIntTy); 1334 Value *ZeroCmp = llvm::Constant::getNullValue(ArgIntTy); 1335 Value *Result = Builder.CreateICmpSLT(BCArg, ZeroCmp); 1336 return RValue::get(Builder.CreateZExt(Result, ConvertType(E->getType()))); 1337 } 1338 case Builtin::BI__builtin_annotation: { 1339 llvm::Value *AnnVal = EmitScalarExpr(E->getArg(0)); 1340 llvm::Value *F = CGM.getIntrinsic(llvm::Intrinsic::annotation, 1341 AnnVal->getType()); 1342 1343 // Get the annotation string, go through casts. Sema requires this to be a 1344 // non-wide string literal, potentially casted, so the cast<> is safe. 1345 const Expr *AnnotationStrExpr = E->getArg(1)->IgnoreParenCasts(); 1346 StringRef Str = cast<StringLiteral>(AnnotationStrExpr)->getString(); 1347 return RValue::get(EmitAnnotationCall(F, AnnVal, Str, E->getExprLoc())); 1348 } 1349 case Builtin::BI__builtin_addcb: 1350 case Builtin::BI__builtin_addcs: 1351 case Builtin::BI__builtin_addc: 1352 case Builtin::BI__builtin_addcl: 1353 case Builtin::BI__builtin_addcll: 1354 case Builtin::BI__builtin_subcb: 1355 case Builtin::BI__builtin_subcs: 1356 case Builtin::BI__builtin_subc: 1357 case Builtin::BI__builtin_subcl: 1358 case Builtin::BI__builtin_subcll: { 1359 1360 // We translate all of these builtins from expressions of the form: 1361 // int x = ..., y = ..., carryin = ..., carryout, result; 1362 // result = __builtin_addc(x, y, carryin, &carryout); 1363 // 1364 // to LLVM IR of the form: 1365 // 1366 // %tmp1 = call {i32, i1} @llvm.uadd.with.overflow.i32(i32 %x, i32 %y) 1367 // %tmpsum1 = extractvalue {i32, i1} %tmp1, 0 1368 // %carry1 = extractvalue {i32, i1} %tmp1, 1 1369 // %tmp2 = call {i32, i1} @llvm.uadd.with.overflow.i32(i32 %tmpsum1, 1370 // i32 %carryin) 1371 // %result = extractvalue {i32, i1} %tmp2, 0 1372 // %carry2 = extractvalue {i32, i1} %tmp2, 1 1373 // %tmp3 = or i1 %carry1, %carry2 1374 // %tmp4 = zext i1 %tmp3 to i32 1375 // store i32 %tmp4, i32* %carryout 1376 1377 // Scalarize our inputs. 1378 llvm::Value *X = EmitScalarExpr(E->getArg(0)); 1379 llvm::Value *Y = EmitScalarExpr(E->getArg(1)); 1380 llvm::Value *Carryin = EmitScalarExpr(E->getArg(2)); 1381 std::pair<llvm::Value*, unsigned> CarryOutPtr = 1382 EmitPointerWithAlignment(E->getArg(3)); 1383 1384 // Decide if we are lowering to a uadd.with.overflow or usub.with.overflow. 1385 llvm::Intrinsic::ID IntrinsicId; 1386 switch (BuiltinID) { 1387 default: llvm_unreachable("Unknown multiprecision builtin id."); 1388 case Builtin::BI__builtin_addcb: 1389 case Builtin::BI__builtin_addcs: 1390 case Builtin::BI__builtin_addc: 1391 case Builtin::BI__builtin_addcl: 1392 case Builtin::BI__builtin_addcll: 1393 IntrinsicId = llvm::Intrinsic::uadd_with_overflow; 1394 break; 1395 case Builtin::BI__builtin_subcb: 1396 case Builtin::BI__builtin_subcs: 1397 case Builtin::BI__builtin_subc: 1398 case Builtin::BI__builtin_subcl: 1399 case Builtin::BI__builtin_subcll: 1400 IntrinsicId = llvm::Intrinsic::usub_with_overflow; 1401 break; 1402 } 1403 1404 // Construct our resulting LLVM IR expression. 1405 llvm::Value *Carry1; 1406 llvm::Value *Sum1 = EmitOverflowIntrinsic(*this, IntrinsicId, 1407 X, Y, Carry1); 1408 llvm::Value *Carry2; 1409 llvm::Value *Sum2 = EmitOverflowIntrinsic(*this, IntrinsicId, 1410 Sum1, Carryin, Carry2); 1411 llvm::Value *CarryOut = Builder.CreateZExt(Builder.CreateOr(Carry1, Carry2), 1412 X->getType()); 1413 llvm::StoreInst *CarryOutStore = Builder.CreateStore(CarryOut, 1414 CarryOutPtr.first); 1415 CarryOutStore->setAlignment(CarryOutPtr.second); 1416 return RValue::get(Sum2); 1417 } 1418 case Builtin::BI__builtin_uadd_overflow: 1419 case Builtin::BI__builtin_uaddl_overflow: 1420 case Builtin::BI__builtin_uaddll_overflow: 1421 case Builtin::BI__builtin_usub_overflow: 1422 case Builtin::BI__builtin_usubl_overflow: 1423 case Builtin::BI__builtin_usubll_overflow: 1424 case Builtin::BI__builtin_umul_overflow: 1425 case Builtin::BI__builtin_umull_overflow: 1426 case Builtin::BI__builtin_umulll_overflow: 1427 case Builtin::BI__builtin_sadd_overflow: 1428 case Builtin::BI__builtin_saddl_overflow: 1429 case Builtin::BI__builtin_saddll_overflow: 1430 case Builtin::BI__builtin_ssub_overflow: 1431 case Builtin::BI__builtin_ssubl_overflow: 1432 case Builtin::BI__builtin_ssubll_overflow: 1433 case Builtin::BI__builtin_smul_overflow: 1434 case Builtin::BI__builtin_smull_overflow: 1435 case Builtin::BI__builtin_smulll_overflow: { 1436 1437 // We translate all of these builtins directly to the relevant llvm IR node. 1438 1439 // Scalarize our inputs. 1440 llvm::Value *X = EmitScalarExpr(E->getArg(0)); 1441 llvm::Value *Y = EmitScalarExpr(E->getArg(1)); 1442 std::pair<llvm::Value *, unsigned> SumOutPtr = 1443 EmitPointerWithAlignment(E->getArg(2)); 1444 1445 // Decide which of the overflow intrinsics we are lowering to: 1446 llvm::Intrinsic::ID IntrinsicId; 1447 switch (BuiltinID) { 1448 default: llvm_unreachable("Unknown security overflow builtin id."); 1449 case Builtin::BI__builtin_uadd_overflow: 1450 case Builtin::BI__builtin_uaddl_overflow: 1451 case Builtin::BI__builtin_uaddll_overflow: 1452 IntrinsicId = llvm::Intrinsic::uadd_with_overflow; 1453 break; 1454 case Builtin::BI__builtin_usub_overflow: 1455 case Builtin::BI__builtin_usubl_overflow: 1456 case Builtin::BI__builtin_usubll_overflow: 1457 IntrinsicId = llvm::Intrinsic::usub_with_overflow; 1458 break; 1459 case Builtin::BI__builtin_umul_overflow: 1460 case Builtin::BI__builtin_umull_overflow: 1461 case Builtin::BI__builtin_umulll_overflow: 1462 IntrinsicId = llvm::Intrinsic::umul_with_overflow; 1463 break; 1464 case Builtin::BI__builtin_sadd_overflow: 1465 case Builtin::BI__builtin_saddl_overflow: 1466 case Builtin::BI__builtin_saddll_overflow: 1467 IntrinsicId = llvm::Intrinsic::sadd_with_overflow; 1468 break; 1469 case Builtin::BI__builtin_ssub_overflow: 1470 case Builtin::BI__builtin_ssubl_overflow: 1471 case Builtin::BI__builtin_ssubll_overflow: 1472 IntrinsicId = llvm::Intrinsic::ssub_with_overflow; 1473 break; 1474 case Builtin::BI__builtin_smul_overflow: 1475 case Builtin::BI__builtin_smull_overflow: 1476 case Builtin::BI__builtin_smulll_overflow: 1477 IntrinsicId = llvm::Intrinsic::smul_with_overflow; 1478 break; 1479 } 1480 1481 1482 llvm::Value *Carry; 1483 llvm::Value *Sum = EmitOverflowIntrinsic(*this, IntrinsicId, X, Y, Carry); 1484 llvm::StoreInst *SumOutStore = Builder.CreateStore(Sum, SumOutPtr.first); 1485 SumOutStore->setAlignment(SumOutPtr.second); 1486 1487 return RValue::get(Carry); 1488 } 1489 case Builtin::BI__builtin_addressof: 1490 return RValue::get(EmitLValue(E->getArg(0)).getAddress()); 1491 case Builtin::BI__noop: 1492 return RValue::get(0); 1493 } 1494 1495 // If this is an alias for a lib function (e.g. __builtin_sin), emit 1496 // the call using the normal call path, but using the unmangled 1497 // version of the function name. 1498 if (getContext().BuiltinInfo.isLibFunction(BuiltinID)) 1499 return emitLibraryCall(*this, FD, E, 1500 CGM.getBuiltinLibFunction(FD, BuiltinID)); 1501 1502 // If this is a predefined lib function (e.g. malloc), emit the call 1503 // using exactly the normal call path. 1504 if (getContext().BuiltinInfo.isPredefinedLibFunction(BuiltinID)) 1505 return emitLibraryCall(*this, FD, E, EmitScalarExpr(E->getCallee())); 1506 1507 // See if we have a target specific intrinsic. 1508 const char *Name = getContext().BuiltinInfo.GetName(BuiltinID); 1509 Intrinsic::ID IntrinsicID = Intrinsic::not_intrinsic; 1510 if (const char *Prefix = 1511 llvm::Triple::getArchTypePrefix(getTarget().getTriple().getArch())) 1512 IntrinsicID = Intrinsic::getIntrinsicForGCCBuiltin(Prefix, Name); 1513 1514 if (IntrinsicID != Intrinsic::not_intrinsic) { 1515 SmallVector<Value*, 16> Args; 1516 1517 // Find out if any arguments are required to be integer constant 1518 // expressions. 1519 unsigned ICEArguments = 0; 1520 ASTContext::GetBuiltinTypeError Error; 1521 getContext().GetBuiltinType(BuiltinID, Error, &ICEArguments); 1522 assert(Error == ASTContext::GE_None && "Should not codegen an error"); 1523 1524 Function *F = CGM.getIntrinsic(IntrinsicID); 1525 llvm::FunctionType *FTy = F->getFunctionType(); 1526 1527 for (unsigned i = 0, e = E->getNumArgs(); i != e; ++i) { 1528 Value *ArgValue; 1529 // If this is a normal argument, just emit it as a scalar. 1530 if ((ICEArguments & (1 << i)) == 0) { 1531 ArgValue = EmitScalarExpr(E->getArg(i)); 1532 } else { 1533 // If this is required to be a constant, constant fold it so that we 1534 // know that the generated intrinsic gets a ConstantInt. 1535 llvm::APSInt Result; 1536 bool IsConst = E->getArg(i)->isIntegerConstantExpr(Result,getContext()); 1537 assert(IsConst && "Constant arg isn't actually constant?"); 1538 (void)IsConst; 1539 ArgValue = llvm::ConstantInt::get(getLLVMContext(), Result); 1540 } 1541 1542 // If the intrinsic arg type is different from the builtin arg type 1543 // we need to do a bit cast. 1544 llvm::Type *PTy = FTy->getParamType(i); 1545 if (PTy != ArgValue->getType()) { 1546 assert(PTy->canLosslesslyBitCastTo(FTy->getParamType(i)) && 1547 "Must be able to losslessly bit cast to param"); 1548 ArgValue = Builder.CreateBitCast(ArgValue, PTy); 1549 } 1550 1551 Args.push_back(ArgValue); 1552 } 1553 1554 Value *V = Builder.CreateCall(F, Args); 1555 QualType BuiltinRetType = E->getType(); 1556 1557 llvm::Type *RetTy = VoidTy; 1558 if (!BuiltinRetType->isVoidType()) 1559 RetTy = ConvertType(BuiltinRetType); 1560 1561 if (RetTy != V->getType()) { 1562 assert(V->getType()->canLosslesslyBitCastTo(RetTy) && 1563 "Must be able to losslessly bit cast result type"); 1564 V = Builder.CreateBitCast(V, RetTy); 1565 } 1566 1567 return RValue::get(V); 1568 } 1569 1570 // See if we have a target specific builtin that needs to be lowered. 1571 if (Value *V = EmitTargetBuiltinExpr(BuiltinID, E)) 1572 return RValue::get(V); 1573 1574 ErrorUnsupported(E, "builtin function"); 1575 1576 // Unknown builtin, for now just dump it out and return undef. 1577 return GetUndefRValue(E->getType()); 1578} 1579 1580Value *CodeGenFunction::EmitTargetBuiltinExpr(unsigned BuiltinID, 1581 const CallExpr *E) { 1582 switch (getTarget().getTriple().getArch()) { 1583 case llvm::Triple::aarch64: 1584 return EmitAArch64BuiltinExpr(BuiltinID, E); 1585 case llvm::Triple::arm: 1586 case llvm::Triple::thumb: 1587 return EmitARMBuiltinExpr(BuiltinID, E); 1588 case llvm::Triple::x86: 1589 case llvm::Triple::x86_64: 1590 return EmitX86BuiltinExpr(BuiltinID, E); 1591 case llvm::Triple::ppc: 1592 case llvm::Triple::ppc64: 1593 case llvm::Triple::ppc64le: 1594 return EmitPPCBuiltinExpr(BuiltinID, E); 1595 default: 1596 return 0; 1597 } 1598} 1599 1600static llvm::VectorType *GetNeonType(CodeGenFunction *CGF, 1601 NeonTypeFlags TypeFlags) { 1602 int IsQuad = TypeFlags.isQuad(); 1603 switch (TypeFlags.getEltType()) { 1604 case NeonTypeFlags::Int8: 1605 case NeonTypeFlags::Poly8: 1606 return llvm::VectorType::get(CGF->Int8Ty, 8 << IsQuad); 1607 case NeonTypeFlags::Int16: 1608 case NeonTypeFlags::Poly16: 1609 case NeonTypeFlags::Float16: 1610 return llvm::VectorType::get(CGF->Int16Ty, 4 << IsQuad); 1611 case NeonTypeFlags::Int32: 1612 return llvm::VectorType::get(CGF->Int32Ty, 2 << IsQuad); 1613 case NeonTypeFlags::Int64: 1614 return llvm::VectorType::get(CGF->Int64Ty, 1 << IsQuad); 1615 case NeonTypeFlags::Float32: 1616 return llvm::VectorType::get(CGF->FloatTy, 2 << IsQuad); 1617 case NeonTypeFlags::Float64: 1618 return llvm::VectorType::get(CGF->DoubleTy, 1 << IsQuad); 1619 } 1620 llvm_unreachable("Invalid NeonTypeFlags element type!"); 1621} 1622 1623static Value *EmitExtendedSHL(CodeGenFunction &CGF, 1624 SmallVectorImpl<Value*> &Ops, 1625 llvm::VectorType *VTy, bool usgn, bool isHigh) { 1626 CGBuilderTy Builder = CGF.Builder; 1627 if (isHigh){ 1628 unsigned NumElts = VTy->getNumElements(); 1629 unsigned EltBits = VTy->getElementType()->getPrimitiveSizeInBits(); 1630 llvm::Type *EltTy = 1631 llvm::IntegerType::get(VTy->getContext(), EltBits / 2); 1632 // The source operand type has twice as many elements of half the size. 1633 llvm::Type *SrcTy = llvm::VectorType::get(EltTy, NumElts * 2); 1634 SmallVector<Constant*, 8> Indices; 1635 for (unsigned i = 0; i != NumElts; i++) 1636 Indices.push_back(Builder.getInt32(i + NumElts)); 1637 Value *SV = llvm::ConstantVector::get(Indices); 1638 Value *Undef = llvm::UndefValue::get(SrcTy); 1639 Ops[0] = Builder.CreateBitCast(Ops[0], SrcTy); 1640 Ops[0] = Builder.CreateShuffleVector(Ops[0], Undef, SV); 1641 } else { 1642 llvm::Type *SrcTy = llvm::VectorType::getTruncatedElementVectorType(VTy); 1643 Ops[0] = Builder.CreateBitCast(Ops[0], SrcTy); 1644 } 1645 1646 if (usgn) 1647 Ops[0] = Builder.CreateZExt(Ops[0], VTy); 1648 else 1649 Ops[0] = Builder.CreateSExt(Ops[0], VTy); 1650 Ops[1] = CGF.EmitNeonShiftVector(Ops[1], VTy, false); 1651 return Builder.CreateShl(Ops[0], Ops[1], "vshl_n"); 1652} 1653 1654Value *CodeGenFunction::EmitNeonSplat(Value *V, Constant *C) { 1655 unsigned nElts = cast<llvm::VectorType>(V->getType())->getNumElements(); 1656 Value* SV = llvm::ConstantVector::getSplat(nElts, C); 1657 return Builder.CreateShuffleVector(V, V, SV, "lane"); 1658} 1659 1660Value *CodeGenFunction::EmitNeonCall(Function *F, SmallVectorImpl<Value*> &Ops, 1661 const char *name, 1662 unsigned shift, bool rightshift) { 1663 unsigned j = 0; 1664 for (Function::const_arg_iterator ai = F->arg_begin(), ae = F->arg_end(); 1665 ai != ae; ++ai, ++j) 1666 if (shift > 0 && shift == j) 1667 Ops[j] = EmitNeonShiftVector(Ops[j], ai->getType(), rightshift); 1668 else 1669 Ops[j] = Builder.CreateBitCast(Ops[j], ai->getType(), name); 1670 1671 return Builder.CreateCall(F, Ops, name); 1672} 1673 1674Value *CodeGenFunction::EmitNeonShiftVector(Value *V, llvm::Type *Ty, 1675 bool neg) { 1676 int SV = cast<ConstantInt>(V)->getSExtValue(); 1677 1678 llvm::VectorType *VTy = cast<llvm::VectorType>(Ty); 1679 llvm::Constant *C = ConstantInt::get(VTy->getElementType(), neg ? -SV : SV); 1680 return llvm::ConstantVector::getSplat(VTy->getNumElements(), C); 1681} 1682 1683/// GetPointeeAlignment - Given an expression with a pointer type, find the 1684/// alignment of the type referenced by the pointer. Skip over implicit 1685/// casts. 1686std::pair<llvm::Value*, unsigned> 1687CodeGenFunction::EmitPointerWithAlignment(const Expr *Addr) { 1688 assert(Addr->getType()->isPointerType()); 1689 Addr = Addr->IgnoreParens(); 1690 if (const ImplicitCastExpr *ICE = dyn_cast<ImplicitCastExpr>(Addr)) { 1691 if ((ICE->getCastKind() == CK_BitCast || ICE->getCastKind() == CK_NoOp) && 1692 ICE->getSubExpr()->getType()->isPointerType()) { 1693 std::pair<llvm::Value*, unsigned> Ptr = 1694 EmitPointerWithAlignment(ICE->getSubExpr()); 1695 Ptr.first = Builder.CreateBitCast(Ptr.first, 1696 ConvertType(Addr->getType())); 1697 return Ptr; 1698 } else if (ICE->getCastKind() == CK_ArrayToPointerDecay) { 1699 LValue LV = EmitLValue(ICE->getSubExpr()); 1700 unsigned Align = LV.getAlignment().getQuantity(); 1701 if (!Align) { 1702 // FIXME: Once LValues are fixed to always set alignment, 1703 // zap this code. 1704 QualType PtTy = ICE->getSubExpr()->getType(); 1705 if (!PtTy->isIncompleteType()) 1706 Align = getContext().getTypeAlignInChars(PtTy).getQuantity(); 1707 else 1708 Align = 1; 1709 } 1710 return std::make_pair(LV.getAddress(), Align); 1711 } 1712 } 1713 if (const UnaryOperator *UO = dyn_cast<UnaryOperator>(Addr)) { 1714 if (UO->getOpcode() == UO_AddrOf) { 1715 LValue LV = EmitLValue(UO->getSubExpr()); 1716 unsigned Align = LV.getAlignment().getQuantity(); 1717 if (!Align) { 1718 // FIXME: Once LValues are fixed to always set alignment, 1719 // zap this code. 1720 QualType PtTy = UO->getSubExpr()->getType(); 1721 if (!PtTy->isIncompleteType()) 1722 Align = getContext().getTypeAlignInChars(PtTy).getQuantity(); 1723 else 1724 Align = 1; 1725 } 1726 return std::make_pair(LV.getAddress(), Align); 1727 } 1728 } 1729 1730 unsigned Align = 1; 1731 QualType PtTy = Addr->getType()->getPointeeType(); 1732 if (!PtTy->isIncompleteType()) 1733 Align = getContext().getTypeAlignInChars(PtTy).getQuantity(); 1734 1735 return std::make_pair(EmitScalarExpr(Addr), Align); 1736} 1737 1738Value *CodeGenFunction::EmitAArch64BuiltinExpr(unsigned BuiltinID, 1739 const CallExpr *E) { 1740 if (BuiltinID == AArch64::BI__clear_cache) { 1741 assert(E->getNumArgs() == 2 && 1742 "Variadic __clear_cache slipped through on AArch64"); 1743 1744 const FunctionDecl *FD = E->getDirectCallee(); 1745 SmallVector<Value *, 2> Ops; 1746 for (unsigned i = 0; i < E->getNumArgs(); i++) 1747 Ops.push_back(EmitScalarExpr(E->getArg(i))); 1748 llvm::Type *Ty = CGM.getTypes().ConvertType(FD->getType()); 1749 llvm::FunctionType *FTy = cast<llvm::FunctionType>(Ty); 1750 StringRef Name = FD->getName(); 1751 return EmitNounwindRuntimeCall(CGM.CreateRuntimeFunction(FTy, Name), Ops); 1752 } 1753 1754 SmallVector<Value *, 4> Ops; 1755 for (unsigned i = 0, e = E->getNumArgs() - 1; i != e; i++) { 1756 Ops.push_back(EmitScalarExpr(E->getArg(i))); 1757 } 1758 1759 // Get the last argument, which specifies the vector type. 1760 llvm::APSInt Result; 1761 const Expr *Arg = E->getArg(E->getNumArgs() - 1); 1762 if (!Arg->isIntegerConstantExpr(Result, getContext())) 1763 return 0; 1764 1765 // Determine the type of this overloaded NEON intrinsic. 1766 NeonTypeFlags Type(Result.getZExtValue()); 1767 bool usgn = Type.isUnsigned(); 1768 1769 llvm::VectorType *VTy = GetNeonType(this, Type); 1770 llvm::Type *Ty = VTy; 1771 if (!Ty) 1772 return 0; 1773 1774 unsigned Int; 1775 switch (BuiltinID) { 1776 default: 1777 return 0; 1778 1779 // AArch64 builtins mapping to legacy ARM v7 builtins. 1780 // FIXME: the mapped builtins listed correspond to what has been tested 1781 // in aarch64-neon-intrinsics.c so far. 1782 case AArch64::BI__builtin_neon_vmul_v: 1783 return EmitARMBuiltinExpr(ARM::BI__builtin_neon_vmul_v, E); 1784 case AArch64::BI__builtin_neon_vmulq_v: 1785 return EmitARMBuiltinExpr(ARM::BI__builtin_neon_vmulq_v, E); 1786 case AArch64::BI__builtin_neon_vabd_v: 1787 return EmitARMBuiltinExpr(ARM::BI__builtin_neon_vabd_v, E); 1788 case AArch64::BI__builtin_neon_vabdq_v: 1789 return EmitARMBuiltinExpr(ARM::BI__builtin_neon_vabdq_v, E); 1790 case AArch64::BI__builtin_neon_vfma_v: 1791 return EmitARMBuiltinExpr(ARM::BI__builtin_neon_vfma_v, E); 1792 case AArch64::BI__builtin_neon_vfmaq_v: 1793 return EmitARMBuiltinExpr(ARM::BI__builtin_neon_vfmaq_v, E); 1794 case AArch64::BI__builtin_neon_vbsl_v: 1795 return EmitARMBuiltinExpr(ARM::BI__builtin_neon_vbsl_v, E); 1796 case AArch64::BI__builtin_neon_vbslq_v: 1797 return EmitARMBuiltinExpr(ARM::BI__builtin_neon_vbslq_v, E); 1798 case AArch64::BI__builtin_neon_vrsqrts_v: 1799 return EmitARMBuiltinExpr(ARM::BI__builtin_neon_vrsqrts_v, E); 1800 case AArch64::BI__builtin_neon_vrsqrtsq_v: 1801 return EmitARMBuiltinExpr(ARM::BI__builtin_neon_vrsqrtsq_v, E); 1802 case AArch64::BI__builtin_neon_vrecps_v: 1803 return EmitARMBuiltinExpr(ARM::BI__builtin_neon_vrecps_v, E); 1804 case AArch64::BI__builtin_neon_vrecpsq_v: 1805 return EmitARMBuiltinExpr(ARM::BI__builtin_neon_vrecpsq_v, E); 1806 case AArch64::BI__builtin_neon_vcage_v: 1807 return EmitARMBuiltinExpr(ARM::BI__builtin_neon_vcage_v, E); 1808 case AArch64::BI__builtin_neon_vcale_v: 1809 return EmitARMBuiltinExpr(ARM::BI__builtin_neon_vcale_v, E); 1810 case AArch64::BI__builtin_neon_vcaleq_v: 1811 std::swap(Ops[0], Ops[1]); 1812 case AArch64::BI__builtin_neon_vcageq_v: { 1813 Function *F; 1814 if (VTy->getElementType()->isIntegerTy(64)) 1815 F = CGM.getIntrinsic(Intrinsic::aarch64_neon_vacgeq); 1816 else 1817 F = CGM.getIntrinsic(Intrinsic::arm_neon_vacgeq); 1818 return EmitNeonCall(F, Ops, "vcage"); 1819 } 1820 case AArch64::BI__builtin_neon_vcalt_v: 1821 return EmitARMBuiltinExpr(ARM::BI__builtin_neon_vcalt_v, E); 1822 case AArch64::BI__builtin_neon_vcagt_v: 1823 return EmitARMBuiltinExpr(ARM::BI__builtin_neon_vcagt_v, E); 1824 case AArch64::BI__builtin_neon_vcaltq_v: 1825 std::swap(Ops[0], Ops[1]); 1826 case AArch64::BI__builtin_neon_vcagtq_v: { 1827 Function *F; 1828 if (VTy->getElementType()->isIntegerTy(64)) 1829 F = CGM.getIntrinsic(Intrinsic::aarch64_neon_vacgtq); 1830 else 1831 F = CGM.getIntrinsic(Intrinsic::arm_neon_vacgtq); 1832 return EmitNeonCall(F, Ops, "vcagt"); 1833 } 1834 case AArch64::BI__builtin_neon_vtst_v: 1835 return EmitARMBuiltinExpr(ARM::BI__builtin_neon_vtst_v, E); 1836 case AArch64::BI__builtin_neon_vtstq_v: 1837 return EmitARMBuiltinExpr(ARM::BI__builtin_neon_vtstq_v, E); 1838 case AArch64::BI__builtin_neon_vhadd_v: 1839 return EmitARMBuiltinExpr(ARM::BI__builtin_neon_vhadd_v, E); 1840 case AArch64::BI__builtin_neon_vhaddq_v: 1841 return EmitARMBuiltinExpr(ARM::BI__builtin_neon_vhaddq_v, E); 1842 case AArch64::BI__builtin_neon_vhsub_v: 1843 return EmitARMBuiltinExpr(ARM::BI__builtin_neon_vhsub_v, E); 1844 case AArch64::BI__builtin_neon_vhsubq_v: 1845 return EmitARMBuiltinExpr(ARM::BI__builtin_neon_vhsubq_v, E); 1846 case AArch64::BI__builtin_neon_vrhadd_v: 1847 return EmitARMBuiltinExpr(ARM::BI__builtin_neon_vrhadd_v, E); 1848 case AArch64::BI__builtin_neon_vrhaddq_v: 1849 return EmitARMBuiltinExpr(ARM::BI__builtin_neon_vrhaddq_v, E); 1850 case AArch64::BI__builtin_neon_vqadd_v: 1851 return EmitARMBuiltinExpr(ARM::BI__builtin_neon_vqadd_v, E); 1852 case AArch64::BI__builtin_neon_vqaddq_v: 1853 return EmitARMBuiltinExpr(ARM::BI__builtin_neon_vqaddq_v, E); 1854 case AArch64::BI__builtin_neon_vqsub_v: 1855 return EmitARMBuiltinExpr(ARM::BI__builtin_neon_vqsub_v, E); 1856 case AArch64::BI__builtin_neon_vqsubq_v: 1857 return EmitARMBuiltinExpr(ARM::BI__builtin_neon_vqsubq_v, E); 1858 case AArch64::BI__builtin_neon_vshl_v: 1859 return EmitARMBuiltinExpr(ARM::BI__builtin_neon_vshl_v, E); 1860 case AArch64::BI__builtin_neon_vshlq_v: 1861 return EmitARMBuiltinExpr(ARM::BI__builtin_neon_vshlq_v, E); 1862 case AArch64::BI__builtin_neon_vqshl_v: 1863 return EmitARMBuiltinExpr(ARM::BI__builtin_neon_vqshl_v, E); 1864 case AArch64::BI__builtin_neon_vqshlq_v: 1865 return EmitARMBuiltinExpr(ARM::BI__builtin_neon_vqshlq_v, E); 1866 case AArch64::BI__builtin_neon_vrshl_v: 1867 return EmitARMBuiltinExpr(ARM::BI__builtin_neon_vrshl_v, E); 1868 case AArch64::BI__builtin_neon_vrshlq_v: 1869 return EmitARMBuiltinExpr(ARM::BI__builtin_neon_vrshlq_v, E); 1870 case AArch64::BI__builtin_neon_vqrshl_v: 1871 return EmitARMBuiltinExpr(ARM::BI__builtin_neon_vqrshl_v, E); 1872 case AArch64::BI__builtin_neon_vqrshlq_v: 1873 return EmitARMBuiltinExpr(ARM::BI__builtin_neon_vqrshlq_v, E); 1874 case AArch64::BI__builtin_neon_vmax_v: 1875 return EmitARMBuiltinExpr(ARM::BI__builtin_neon_vmax_v, E); 1876 case AArch64::BI__builtin_neon_vmaxq_v: 1877 return EmitARMBuiltinExpr(ARM::BI__builtin_neon_vmaxq_v, E); 1878 case AArch64::BI__builtin_neon_vmin_v: 1879 return EmitARMBuiltinExpr(ARM::BI__builtin_neon_vmin_v, E); 1880 case AArch64::BI__builtin_neon_vminq_v: 1881 return EmitARMBuiltinExpr(ARM::BI__builtin_neon_vminq_v, E); 1882 case AArch64::BI__builtin_neon_vpmax_v: 1883 return EmitARMBuiltinExpr(ARM::BI__builtin_neon_vpmax_v, E); 1884 case AArch64::BI__builtin_neon_vpmin_v: 1885 return EmitARMBuiltinExpr(ARM::BI__builtin_neon_vpmin_v, E); 1886 case AArch64::BI__builtin_neon_vpadd_v: 1887 return EmitARMBuiltinExpr(ARM::BI__builtin_neon_vpadd_v, E); 1888 case AArch64::BI__builtin_neon_vqdmulh_v: 1889 return EmitARMBuiltinExpr(ARM::BI__builtin_neon_vqdmulh_v, E); 1890 case AArch64::BI__builtin_neon_vqdmulhq_v: 1891 return EmitARMBuiltinExpr(ARM::BI__builtin_neon_vqdmulhq_v, E); 1892 case AArch64::BI__builtin_neon_vqrdmulh_v: 1893 return EmitARMBuiltinExpr(ARM::BI__builtin_neon_vqrdmulh_v, E); 1894 case AArch64::BI__builtin_neon_vqrdmulhq_v: 1895 return EmitARMBuiltinExpr(ARM::BI__builtin_neon_vqrdmulhq_v, E); 1896 case AArch64::BI__builtin_neon_vshl_n_v: 1897 return EmitARMBuiltinExpr(ARM::BI__builtin_neon_vshl_n_v, E); 1898 case AArch64::BI__builtin_neon_vshlq_n_v: 1899 return EmitARMBuiltinExpr(ARM::BI__builtin_neon_vshlq_n_v, E); 1900 case AArch64::BI__builtin_neon_vmovl_v: 1901 return EmitARMBuiltinExpr(ARM::BI__builtin_neon_vmovl_v, E); 1902 case AArch64::BI__builtin_neon_vshll_n_v: 1903 return EmitExtendedSHL(*this, Ops, VTy, usgn, false); 1904 case AArch64::BI__builtin_neon_vmovl_high_v: 1905 Ops.push_back(ConstantInt::get(Int32Ty, 0)); 1906 case AArch64::BI__builtin_neon_vshll_high_n_v: 1907 return EmitExtendedSHL(*this, Ops, VTy, usgn, true); 1908 1909 // AArch64-only builtins 1910 case AArch64::BI__builtin_neon_vfms_v: 1911 case AArch64::BI__builtin_neon_vfmsq_v: { 1912 Value *F = CGM.getIntrinsic(Intrinsic::fma, Ty); 1913 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 1914 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 1915 Ops[1] = Builder.CreateFNeg(Ops[1]); 1916 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 1917 1918 // LLVM's fma intrinsic puts the accumulator in the last position, but the 1919 // AArch64 intrinsic has it first. 1920 return Builder.CreateCall3(F, Ops[1], Ops[2], Ops[0]); 1921 } 1922 case AArch64::BI__builtin_neon_vmaxnm_v: 1923 case AArch64::BI__builtin_neon_vmaxnmq_v: { 1924 Int = Intrinsic::aarch64_neon_vmaxnm; 1925 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmaxnm"); 1926 } 1927 case AArch64::BI__builtin_neon_vminnm_v: 1928 case AArch64::BI__builtin_neon_vminnmq_v: { 1929 Int = Intrinsic::aarch64_neon_vminnm; 1930 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vminnm"); 1931 } 1932 case AArch64::BI__builtin_neon_vpmaxnm_v: 1933 case AArch64::BI__builtin_neon_vpmaxnmq_v: { 1934 Int = Intrinsic::aarch64_neon_vpmaxnm; 1935 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vpmaxnm"); 1936 } 1937 case AArch64::BI__builtin_neon_vpminnm_v: 1938 case AArch64::BI__builtin_neon_vpminnmq_v: { 1939 Int = Intrinsic::aarch64_neon_vpminnm; 1940 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vpminnm"); 1941 } 1942 case AArch64::BI__builtin_neon_vpmaxq_v: { 1943 Int = usgn ? Intrinsic::arm_neon_vpmaxu : Intrinsic::arm_neon_vpmaxs; 1944 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vpmax"); 1945 } 1946 case AArch64::BI__builtin_neon_vpminq_v: { 1947 Int = usgn ? Intrinsic::arm_neon_vpminu : Intrinsic::arm_neon_vpmins; 1948 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vpmin"); 1949 } 1950 case AArch64::BI__builtin_neon_vpaddq_v: { 1951 Int = Intrinsic::arm_neon_vpadd; 1952 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vpadd"); 1953 } 1954 case AArch64::BI__builtin_neon_vmulx_v: 1955 case AArch64::BI__builtin_neon_vmulxq_v: { 1956 Int = Intrinsic::aarch64_neon_vmulx; 1957 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmulx"); 1958 } 1959 } 1960} 1961 1962Value *CodeGenFunction::EmitARMBuiltinExpr(unsigned BuiltinID, 1963 const CallExpr *E) { 1964 if (BuiltinID == ARM::BI__clear_cache) { 1965 assert(E->getNumArgs() == 2 && "__clear_cache takes 2 arguments"); 1966 const FunctionDecl *FD = E->getDirectCallee(); 1967 SmallVector<Value*, 2> Ops; 1968 for (unsigned i = 0; i < 2; i++) 1969 Ops.push_back(EmitScalarExpr(E->getArg(i))); 1970 llvm::Type *Ty = CGM.getTypes().ConvertType(FD->getType()); 1971 llvm::FunctionType *FTy = cast<llvm::FunctionType>(Ty); 1972 StringRef Name = FD->getName(); 1973 return EmitNounwindRuntimeCall(CGM.CreateRuntimeFunction(FTy, Name), Ops); 1974 } 1975 1976 if (BuiltinID == ARM::BI__builtin_arm_ldrexd || 1977 (BuiltinID == ARM::BI__builtin_arm_ldrex && 1978 getContext().getTypeSize(E->getType()) == 64)) { 1979 Function *F = CGM.getIntrinsic(Intrinsic::arm_ldrexd); 1980 1981 Value *LdPtr = EmitScalarExpr(E->getArg(0)); 1982 Value *Val = Builder.CreateCall(F, Builder.CreateBitCast(LdPtr, Int8PtrTy), 1983 "ldrexd"); 1984 1985 Value *Val0 = Builder.CreateExtractValue(Val, 1); 1986 Value *Val1 = Builder.CreateExtractValue(Val, 0); 1987 Val0 = Builder.CreateZExt(Val0, Int64Ty); 1988 Val1 = Builder.CreateZExt(Val1, Int64Ty); 1989 1990 Value *ShiftCst = llvm::ConstantInt::get(Int64Ty, 32); 1991 Val = Builder.CreateShl(Val0, ShiftCst, "shl", true /* nuw */); 1992 Val = Builder.CreateOr(Val, Val1); 1993 return Builder.CreateBitCast(Val, ConvertType(E->getType())); 1994 } 1995 1996 if (BuiltinID == ARM::BI__builtin_arm_ldrex) { 1997 Value *LoadAddr = EmitScalarExpr(E->getArg(0)); 1998 1999 QualType Ty = E->getType(); 2000 llvm::Type *RealResTy = ConvertType(Ty); 2001 llvm::Type *IntResTy = llvm::IntegerType::get(getLLVMContext(), 2002 getContext().getTypeSize(Ty)); 2003 LoadAddr = Builder.CreateBitCast(LoadAddr, IntResTy->getPointerTo()); 2004 2005 Function *F = CGM.getIntrinsic(Intrinsic::arm_ldrex, LoadAddr->getType()); 2006 Value *Val = Builder.CreateCall(F, LoadAddr, "ldrex"); 2007 2008 if (RealResTy->isPointerTy()) 2009 return Builder.CreateIntToPtr(Val, RealResTy); 2010 else { 2011 Val = Builder.CreateTruncOrBitCast(Val, IntResTy); 2012 return Builder.CreateBitCast(Val, RealResTy); 2013 } 2014 } 2015 2016 if (BuiltinID == ARM::BI__builtin_arm_strexd || 2017 (BuiltinID == ARM::BI__builtin_arm_strex && 2018 getContext().getTypeSize(E->getArg(0)->getType()) == 64)) { 2019 Function *F = CGM.getIntrinsic(Intrinsic::arm_strexd); 2020 llvm::Type *STy = llvm::StructType::get(Int32Ty, Int32Ty, NULL); 2021 2022 Value *Tmp = CreateMemTemp(E->getArg(0)->getType()); 2023 Value *Val = EmitScalarExpr(E->getArg(0)); 2024 Builder.CreateStore(Val, Tmp); 2025 2026 Value *LdPtr = Builder.CreateBitCast(Tmp,llvm::PointerType::getUnqual(STy)); 2027 Val = Builder.CreateLoad(LdPtr); 2028 2029 Value *Arg0 = Builder.CreateExtractValue(Val, 0); 2030 Value *Arg1 = Builder.CreateExtractValue(Val, 1); 2031 Value *StPtr = Builder.CreateBitCast(EmitScalarExpr(E->getArg(1)), Int8PtrTy); 2032 return Builder.CreateCall3(F, Arg0, Arg1, StPtr, "strexd"); 2033 } 2034 2035 if (BuiltinID == ARM::BI__builtin_arm_strex) { 2036 Value *StoreVal = EmitScalarExpr(E->getArg(0)); 2037 Value *StoreAddr = EmitScalarExpr(E->getArg(1)); 2038 2039 QualType Ty = E->getArg(0)->getType(); 2040 llvm::Type *StoreTy = llvm::IntegerType::get(getLLVMContext(), 2041 getContext().getTypeSize(Ty)); 2042 StoreAddr = Builder.CreateBitCast(StoreAddr, StoreTy->getPointerTo()); 2043 2044 if (StoreVal->getType()->isPointerTy()) 2045 StoreVal = Builder.CreatePtrToInt(StoreVal, Int32Ty); 2046 else { 2047 StoreVal = Builder.CreateBitCast(StoreVal, StoreTy); 2048 StoreVal = Builder.CreateZExtOrBitCast(StoreVal, Int32Ty); 2049 } 2050 2051 Function *F = CGM.getIntrinsic(Intrinsic::arm_strex, StoreAddr->getType()); 2052 return Builder.CreateCall2(F, StoreVal, StoreAddr, "strex"); 2053 } 2054 2055 if (BuiltinID == ARM::BI__builtin_arm_clrex) { 2056 Function *F = CGM.getIntrinsic(Intrinsic::arm_clrex); 2057 return Builder.CreateCall(F); 2058 } 2059 2060 SmallVector<Value*, 4> Ops; 2061 llvm::Value *Align = 0; 2062 for (unsigned i = 0, e = E->getNumArgs() - 1; i != e; i++) { 2063 if (i == 0) { 2064 switch (BuiltinID) { 2065 case ARM::BI__builtin_neon_vld1_v: 2066 case ARM::BI__builtin_neon_vld1q_v: 2067 case ARM::BI__builtin_neon_vld1q_lane_v: 2068 case ARM::BI__builtin_neon_vld1_lane_v: 2069 case ARM::BI__builtin_neon_vld1_dup_v: 2070 case ARM::BI__builtin_neon_vld1q_dup_v: 2071 case ARM::BI__builtin_neon_vst1_v: 2072 case ARM::BI__builtin_neon_vst1q_v: 2073 case ARM::BI__builtin_neon_vst1q_lane_v: 2074 case ARM::BI__builtin_neon_vst1_lane_v: 2075 case ARM::BI__builtin_neon_vst2_v: 2076 case ARM::BI__builtin_neon_vst2q_v: 2077 case ARM::BI__builtin_neon_vst2_lane_v: 2078 case ARM::BI__builtin_neon_vst2q_lane_v: 2079 case ARM::BI__builtin_neon_vst3_v: 2080 case ARM::BI__builtin_neon_vst3q_v: 2081 case ARM::BI__builtin_neon_vst3_lane_v: 2082 case ARM::BI__builtin_neon_vst3q_lane_v: 2083 case ARM::BI__builtin_neon_vst4_v: 2084 case ARM::BI__builtin_neon_vst4q_v: 2085 case ARM::BI__builtin_neon_vst4_lane_v: 2086 case ARM::BI__builtin_neon_vst4q_lane_v: 2087 // Get the alignment for the argument in addition to the value; 2088 // we'll use it later. 2089 std::pair<llvm::Value*, unsigned> Src = 2090 EmitPointerWithAlignment(E->getArg(0)); 2091 Ops.push_back(Src.first); 2092 Align = Builder.getInt32(Src.second); 2093 continue; 2094 } 2095 } 2096 if (i == 1) { 2097 switch (BuiltinID) { 2098 case ARM::BI__builtin_neon_vld2_v: 2099 case ARM::BI__builtin_neon_vld2q_v: 2100 case ARM::BI__builtin_neon_vld3_v: 2101 case ARM::BI__builtin_neon_vld3q_v: 2102 case ARM::BI__builtin_neon_vld4_v: 2103 case ARM::BI__builtin_neon_vld4q_v: 2104 case ARM::BI__builtin_neon_vld2_lane_v: 2105 case ARM::BI__builtin_neon_vld2q_lane_v: 2106 case ARM::BI__builtin_neon_vld3_lane_v: 2107 case ARM::BI__builtin_neon_vld3q_lane_v: 2108 case ARM::BI__builtin_neon_vld4_lane_v: 2109 case ARM::BI__builtin_neon_vld4q_lane_v: 2110 case ARM::BI__builtin_neon_vld2_dup_v: 2111 case ARM::BI__builtin_neon_vld3_dup_v: 2112 case ARM::BI__builtin_neon_vld4_dup_v: 2113 // Get the alignment for the argument in addition to the value; 2114 // we'll use it later. 2115 std::pair<llvm::Value*, unsigned> Src = 2116 EmitPointerWithAlignment(E->getArg(1)); 2117 Ops.push_back(Src.first); 2118 Align = Builder.getInt32(Src.second); 2119 continue; 2120 } 2121 } 2122 Ops.push_back(EmitScalarExpr(E->getArg(i))); 2123 } 2124 2125 // vget_lane and vset_lane are not overloaded and do not have an extra 2126 // argument that specifies the vector type. 2127 switch (BuiltinID) { 2128 default: break; 2129 case ARM::BI__builtin_neon_vget_lane_i8: 2130 case ARM::BI__builtin_neon_vget_lane_i16: 2131 case ARM::BI__builtin_neon_vget_lane_i32: 2132 case ARM::BI__builtin_neon_vget_lane_i64: 2133 case ARM::BI__builtin_neon_vget_lane_f32: 2134 case ARM::BI__builtin_neon_vgetq_lane_i8: 2135 case ARM::BI__builtin_neon_vgetq_lane_i16: 2136 case ARM::BI__builtin_neon_vgetq_lane_i32: 2137 case ARM::BI__builtin_neon_vgetq_lane_i64: 2138 case ARM::BI__builtin_neon_vgetq_lane_f32: 2139 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 2140 "vget_lane"); 2141 case ARM::BI__builtin_neon_vset_lane_i8: 2142 case ARM::BI__builtin_neon_vset_lane_i16: 2143 case ARM::BI__builtin_neon_vset_lane_i32: 2144 case ARM::BI__builtin_neon_vset_lane_i64: 2145 case ARM::BI__builtin_neon_vset_lane_f32: 2146 case ARM::BI__builtin_neon_vsetq_lane_i8: 2147 case ARM::BI__builtin_neon_vsetq_lane_i16: 2148 case ARM::BI__builtin_neon_vsetq_lane_i32: 2149 case ARM::BI__builtin_neon_vsetq_lane_i64: 2150 case ARM::BI__builtin_neon_vsetq_lane_f32: 2151 Ops.push_back(EmitScalarExpr(E->getArg(2))); 2152 return Builder.CreateInsertElement(Ops[1], Ops[0], Ops[2], "vset_lane"); 2153 } 2154 2155 // Get the last argument, which specifies the vector type. 2156 llvm::APSInt Result; 2157 const Expr *Arg = E->getArg(E->getNumArgs()-1); 2158 if (!Arg->isIntegerConstantExpr(Result, getContext())) 2159 return 0; 2160 2161 if (BuiltinID == ARM::BI__builtin_arm_vcvtr_f || 2162 BuiltinID == ARM::BI__builtin_arm_vcvtr_d) { 2163 // Determine the overloaded type of this builtin. 2164 llvm::Type *Ty; 2165 if (BuiltinID == ARM::BI__builtin_arm_vcvtr_f) 2166 Ty = FloatTy; 2167 else 2168 Ty = DoubleTy; 2169 2170 // Determine whether this is an unsigned conversion or not. 2171 bool usgn = Result.getZExtValue() == 1; 2172 unsigned Int = usgn ? Intrinsic::arm_vcvtru : Intrinsic::arm_vcvtr; 2173 2174 // Call the appropriate intrinsic. 2175 Function *F = CGM.getIntrinsic(Int, Ty); 2176 return Builder.CreateCall(F, Ops, "vcvtr"); 2177 } 2178 2179 // Determine the type of this overloaded NEON intrinsic. 2180 NeonTypeFlags Type(Result.getZExtValue()); 2181 bool usgn = Type.isUnsigned(); 2182 bool quad = Type.isQuad(); 2183 bool rightShift = false; 2184 2185 llvm::VectorType *VTy = GetNeonType(this, Type); 2186 llvm::Type *Ty = VTy; 2187 if (!Ty) 2188 return 0; 2189 2190 unsigned Int; 2191 switch (BuiltinID) { 2192 default: return 0; 2193 case ARM::BI__builtin_neon_vbsl_v: 2194 case ARM::BI__builtin_neon_vbslq_v: 2195 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vbsl, Ty), 2196 Ops, "vbsl"); 2197 case ARM::BI__builtin_neon_vabd_v: 2198 case ARM::BI__builtin_neon_vabdq_v: 2199 Int = usgn ? Intrinsic::arm_neon_vabdu : Intrinsic::arm_neon_vabds; 2200 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vabd"); 2201 case ARM::BI__builtin_neon_vabs_v: 2202 case ARM::BI__builtin_neon_vabsq_v: 2203 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vabs, Ty), 2204 Ops, "vabs"); 2205 case ARM::BI__builtin_neon_vaddhn_v: 2206 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vaddhn, Ty), 2207 Ops, "vaddhn"); 2208 case ARM::BI__builtin_neon_vcale_v: 2209 std::swap(Ops[0], Ops[1]); 2210 case ARM::BI__builtin_neon_vcage_v: { 2211 Function *F = CGM.getIntrinsic(Intrinsic::arm_neon_vacged); 2212 return EmitNeonCall(F, Ops, "vcage"); 2213 } 2214 case ARM::BI__builtin_neon_vcaleq_v: 2215 std::swap(Ops[0], Ops[1]); 2216 case ARM::BI__builtin_neon_vcageq_v: { 2217 Function *F = CGM.getIntrinsic(Intrinsic::arm_neon_vacgeq); 2218 return EmitNeonCall(F, Ops, "vcage"); 2219 } 2220 case ARM::BI__builtin_neon_vcalt_v: 2221 std::swap(Ops[0], Ops[1]); 2222 case ARM::BI__builtin_neon_vcagt_v: { 2223 Function *F = CGM.getIntrinsic(Intrinsic::arm_neon_vacgtd); 2224 return EmitNeonCall(F, Ops, "vcagt"); 2225 } 2226 case ARM::BI__builtin_neon_vcaltq_v: 2227 std::swap(Ops[0], Ops[1]); 2228 case ARM::BI__builtin_neon_vcagtq_v: { 2229 Function *F = CGM.getIntrinsic(Intrinsic::arm_neon_vacgtq); 2230 return EmitNeonCall(F, Ops, "vcagt"); 2231 } 2232 case ARM::BI__builtin_neon_vcls_v: 2233 case ARM::BI__builtin_neon_vclsq_v: { 2234 Function *F = CGM.getIntrinsic(Intrinsic::arm_neon_vcls, Ty); 2235 return EmitNeonCall(F, Ops, "vcls"); 2236 } 2237 case ARM::BI__builtin_neon_vclz_v: 2238 case ARM::BI__builtin_neon_vclzq_v: { 2239 // Generate target-independent intrinsic; also need to add second argument 2240 // for whether or not clz of zero is undefined; on ARM it isn't. 2241 Function *F = CGM.getIntrinsic(Intrinsic::ctlz, Ty); 2242 Ops.push_back(Builder.getInt1(getTarget().isCLZForZeroUndef())); 2243 return EmitNeonCall(F, Ops, "vclz"); 2244 } 2245 case ARM::BI__builtin_neon_vcnt_v: 2246 case ARM::BI__builtin_neon_vcntq_v: { 2247 // generate target-independent intrinsic 2248 Function *F = CGM.getIntrinsic(Intrinsic::ctpop, Ty); 2249 return EmitNeonCall(F, Ops, "vctpop"); 2250 } 2251 case ARM::BI__builtin_neon_vcvt_f16_v: { 2252 assert(Type.getEltType() == NeonTypeFlags::Float16 && !quad && 2253 "unexpected vcvt_f16_v builtin"); 2254 Function *F = CGM.getIntrinsic(Intrinsic::arm_neon_vcvtfp2hf); 2255 return EmitNeonCall(F, Ops, "vcvt"); 2256 } 2257 case ARM::BI__builtin_neon_vcvt_f32_f16: { 2258 assert(Type.getEltType() == NeonTypeFlags::Float16 && !quad && 2259 "unexpected vcvt_f32_f16 builtin"); 2260 Function *F = CGM.getIntrinsic(Intrinsic::arm_neon_vcvthf2fp); 2261 return EmitNeonCall(F, Ops, "vcvt"); 2262 } 2263 case ARM::BI__builtin_neon_vcvt_f32_v: 2264 case ARM::BI__builtin_neon_vcvtq_f32_v: 2265 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 2266 Ty = GetNeonType(this, NeonTypeFlags(NeonTypeFlags::Float32, false, quad)); 2267 return usgn ? Builder.CreateUIToFP(Ops[0], Ty, "vcvt") 2268 : Builder.CreateSIToFP(Ops[0], Ty, "vcvt"); 2269 case ARM::BI__builtin_neon_vcvt_s32_v: 2270 case ARM::BI__builtin_neon_vcvt_u32_v: 2271 case ARM::BI__builtin_neon_vcvtq_s32_v: 2272 case ARM::BI__builtin_neon_vcvtq_u32_v: { 2273 llvm::Type *FloatTy = 2274 GetNeonType(this, NeonTypeFlags(NeonTypeFlags::Float32, false, quad)); 2275 Ops[0] = Builder.CreateBitCast(Ops[0], FloatTy); 2276 return usgn ? Builder.CreateFPToUI(Ops[0], Ty, "vcvt") 2277 : Builder.CreateFPToSI(Ops[0], Ty, "vcvt"); 2278 } 2279 case ARM::BI__builtin_neon_vcvt_n_f32_v: 2280 case ARM::BI__builtin_neon_vcvtq_n_f32_v: { 2281 llvm::Type *FloatTy = 2282 GetNeonType(this, NeonTypeFlags(NeonTypeFlags::Float32, false, quad)); 2283 llvm::Type *Tys[2] = { FloatTy, Ty }; 2284 Int = usgn ? Intrinsic::arm_neon_vcvtfxu2fp 2285 : Intrinsic::arm_neon_vcvtfxs2fp; 2286 Function *F = CGM.getIntrinsic(Int, Tys); 2287 return EmitNeonCall(F, Ops, "vcvt_n"); 2288 } 2289 case ARM::BI__builtin_neon_vcvt_n_s32_v: 2290 case ARM::BI__builtin_neon_vcvt_n_u32_v: 2291 case ARM::BI__builtin_neon_vcvtq_n_s32_v: 2292 case ARM::BI__builtin_neon_vcvtq_n_u32_v: { 2293 llvm::Type *FloatTy = 2294 GetNeonType(this, NeonTypeFlags(NeonTypeFlags::Float32, false, quad)); 2295 llvm::Type *Tys[2] = { Ty, FloatTy }; 2296 Int = usgn ? Intrinsic::arm_neon_vcvtfp2fxu 2297 : Intrinsic::arm_neon_vcvtfp2fxs; 2298 Function *F = CGM.getIntrinsic(Int, Tys); 2299 return EmitNeonCall(F, Ops, "vcvt_n"); 2300 } 2301 case ARM::BI__builtin_neon_vext_v: 2302 case ARM::BI__builtin_neon_vextq_v: { 2303 int CV = cast<ConstantInt>(Ops[2])->getSExtValue(); 2304 SmallVector<Constant*, 16> Indices; 2305 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) 2306 Indices.push_back(ConstantInt::get(Int32Ty, i+CV)); 2307 2308 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 2309 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 2310 Value *SV = llvm::ConstantVector::get(Indices); 2311 return Builder.CreateShuffleVector(Ops[0], Ops[1], SV, "vext"); 2312 } 2313 case ARM::BI__builtin_neon_vhadd_v: 2314 case ARM::BI__builtin_neon_vhaddq_v: 2315 Int = usgn ? Intrinsic::arm_neon_vhaddu : Intrinsic::arm_neon_vhadds; 2316 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vhadd"); 2317 case ARM::BI__builtin_neon_vhsub_v: 2318 case ARM::BI__builtin_neon_vhsubq_v: 2319 Int = usgn ? Intrinsic::arm_neon_vhsubu : Intrinsic::arm_neon_vhsubs; 2320 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vhsub"); 2321 case ARM::BI__builtin_neon_vld1_v: 2322 case ARM::BI__builtin_neon_vld1q_v: 2323 Ops.push_back(Align); 2324 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vld1, Ty), 2325 Ops, "vld1"); 2326 case ARM::BI__builtin_neon_vld1q_lane_v: 2327 // Handle 64-bit integer elements as a special case. Use shuffles of 2328 // one-element vectors to avoid poor code for i64 in the backend. 2329 if (VTy->getElementType()->isIntegerTy(64)) { 2330 // Extract the other lane. 2331 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 2332 int Lane = cast<ConstantInt>(Ops[2])->getZExtValue(); 2333 Value *SV = llvm::ConstantVector::get(ConstantInt::get(Int32Ty, 1-Lane)); 2334 Ops[1] = Builder.CreateShuffleVector(Ops[1], Ops[1], SV); 2335 // Load the value as a one-element vector. 2336 Ty = llvm::VectorType::get(VTy->getElementType(), 1); 2337 Function *F = CGM.getIntrinsic(Intrinsic::arm_neon_vld1, Ty); 2338 Value *Ld = Builder.CreateCall2(F, Ops[0], Align); 2339 // Combine them. 2340 SmallVector<Constant*, 2> Indices; 2341 Indices.push_back(ConstantInt::get(Int32Ty, 1-Lane)); 2342 Indices.push_back(ConstantInt::get(Int32Ty, Lane)); 2343 SV = llvm::ConstantVector::get(Indices); 2344 return Builder.CreateShuffleVector(Ops[1], Ld, SV, "vld1q_lane"); 2345 } 2346 // fall through 2347 case ARM::BI__builtin_neon_vld1_lane_v: { 2348 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 2349 Ty = llvm::PointerType::getUnqual(VTy->getElementType()); 2350 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 2351 LoadInst *Ld = Builder.CreateLoad(Ops[0]); 2352 Ld->setAlignment(cast<ConstantInt>(Align)->getZExtValue()); 2353 return Builder.CreateInsertElement(Ops[1], Ld, Ops[2], "vld1_lane"); 2354 } 2355 case ARM::BI__builtin_neon_vld1_dup_v: 2356 case ARM::BI__builtin_neon_vld1q_dup_v: { 2357 Value *V = UndefValue::get(Ty); 2358 Ty = llvm::PointerType::getUnqual(VTy->getElementType()); 2359 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 2360 LoadInst *Ld = Builder.CreateLoad(Ops[0]); 2361 Ld->setAlignment(cast<ConstantInt>(Align)->getZExtValue()); 2362 llvm::Constant *CI = ConstantInt::get(Int32Ty, 0); 2363 Ops[0] = Builder.CreateInsertElement(V, Ld, CI); 2364 return EmitNeonSplat(Ops[0], CI); 2365 } 2366 case ARM::BI__builtin_neon_vld2_v: 2367 case ARM::BI__builtin_neon_vld2q_v: { 2368 Function *F = CGM.getIntrinsic(Intrinsic::arm_neon_vld2, Ty); 2369 Ops[1] = Builder.CreateCall2(F, Ops[1], Align, "vld2"); 2370 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 2371 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 2372 return Builder.CreateStore(Ops[1], Ops[0]); 2373 } 2374 case ARM::BI__builtin_neon_vld3_v: 2375 case ARM::BI__builtin_neon_vld3q_v: { 2376 Function *F = CGM.getIntrinsic(Intrinsic::arm_neon_vld3, Ty); 2377 Ops[1] = Builder.CreateCall2(F, Ops[1], Align, "vld3"); 2378 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 2379 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 2380 return Builder.CreateStore(Ops[1], Ops[0]); 2381 } 2382 case ARM::BI__builtin_neon_vld4_v: 2383 case ARM::BI__builtin_neon_vld4q_v: { 2384 Function *F = CGM.getIntrinsic(Intrinsic::arm_neon_vld4, Ty); 2385 Ops[1] = Builder.CreateCall2(F, Ops[1], Align, "vld4"); 2386 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 2387 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 2388 return Builder.CreateStore(Ops[1], Ops[0]); 2389 } 2390 case ARM::BI__builtin_neon_vld2_lane_v: 2391 case ARM::BI__builtin_neon_vld2q_lane_v: { 2392 Function *F = CGM.getIntrinsic(Intrinsic::arm_neon_vld2lane, Ty); 2393 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 2394 Ops[3] = Builder.CreateBitCast(Ops[3], Ty); 2395 Ops.push_back(Align); 2396 Ops[1] = Builder.CreateCall(F, makeArrayRef(Ops).slice(1), "vld2_lane"); 2397 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 2398 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 2399 return Builder.CreateStore(Ops[1], Ops[0]); 2400 } 2401 case ARM::BI__builtin_neon_vld3_lane_v: 2402 case ARM::BI__builtin_neon_vld3q_lane_v: { 2403 Function *F = CGM.getIntrinsic(Intrinsic::arm_neon_vld3lane, Ty); 2404 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 2405 Ops[3] = Builder.CreateBitCast(Ops[3], Ty); 2406 Ops[4] = Builder.CreateBitCast(Ops[4], Ty); 2407 Ops.push_back(Align); 2408 Ops[1] = Builder.CreateCall(F, makeArrayRef(Ops).slice(1), "vld3_lane"); 2409 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 2410 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 2411 return Builder.CreateStore(Ops[1], Ops[0]); 2412 } 2413 case ARM::BI__builtin_neon_vld4_lane_v: 2414 case ARM::BI__builtin_neon_vld4q_lane_v: { 2415 Function *F = CGM.getIntrinsic(Intrinsic::arm_neon_vld4lane, Ty); 2416 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 2417 Ops[3] = Builder.CreateBitCast(Ops[3], Ty); 2418 Ops[4] = Builder.CreateBitCast(Ops[4], Ty); 2419 Ops[5] = Builder.CreateBitCast(Ops[5], Ty); 2420 Ops.push_back(Align); 2421 Ops[1] = Builder.CreateCall(F, makeArrayRef(Ops).slice(1), "vld3_lane"); 2422 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 2423 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 2424 return Builder.CreateStore(Ops[1], Ops[0]); 2425 } 2426 case ARM::BI__builtin_neon_vld2_dup_v: 2427 case ARM::BI__builtin_neon_vld3_dup_v: 2428 case ARM::BI__builtin_neon_vld4_dup_v: { 2429 // Handle 64-bit elements as a special-case. There is no "dup" needed. 2430 if (VTy->getElementType()->getPrimitiveSizeInBits() == 64) { 2431 switch (BuiltinID) { 2432 case ARM::BI__builtin_neon_vld2_dup_v: 2433 Int = Intrinsic::arm_neon_vld2; 2434 break; 2435 case ARM::BI__builtin_neon_vld3_dup_v: 2436 Int = Intrinsic::arm_neon_vld3; 2437 break; 2438 case ARM::BI__builtin_neon_vld4_dup_v: 2439 Int = Intrinsic::arm_neon_vld4; 2440 break; 2441 default: llvm_unreachable("unknown vld_dup intrinsic?"); 2442 } 2443 Function *F = CGM.getIntrinsic(Int, Ty); 2444 Ops[1] = Builder.CreateCall2(F, Ops[1], Align, "vld_dup"); 2445 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 2446 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 2447 return Builder.CreateStore(Ops[1], Ops[0]); 2448 } 2449 switch (BuiltinID) { 2450 case ARM::BI__builtin_neon_vld2_dup_v: 2451 Int = Intrinsic::arm_neon_vld2lane; 2452 break; 2453 case ARM::BI__builtin_neon_vld3_dup_v: 2454 Int = Intrinsic::arm_neon_vld3lane; 2455 break; 2456 case ARM::BI__builtin_neon_vld4_dup_v: 2457 Int = Intrinsic::arm_neon_vld4lane; 2458 break; 2459 default: llvm_unreachable("unknown vld_dup intrinsic?"); 2460 } 2461 Function *F = CGM.getIntrinsic(Int, Ty); 2462 llvm::StructType *STy = cast<llvm::StructType>(F->getReturnType()); 2463 2464 SmallVector<Value*, 6> Args; 2465 Args.push_back(Ops[1]); 2466 Args.append(STy->getNumElements(), UndefValue::get(Ty)); 2467 2468 llvm::Constant *CI = ConstantInt::get(Int32Ty, 0); 2469 Args.push_back(CI); 2470 Args.push_back(Align); 2471 2472 Ops[1] = Builder.CreateCall(F, Args, "vld_dup"); 2473 // splat lane 0 to all elts in each vector of the result. 2474 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 2475 Value *Val = Builder.CreateExtractValue(Ops[1], i); 2476 Value *Elt = Builder.CreateBitCast(Val, Ty); 2477 Elt = EmitNeonSplat(Elt, CI); 2478 Elt = Builder.CreateBitCast(Elt, Val->getType()); 2479 Ops[1] = Builder.CreateInsertValue(Ops[1], Elt, i); 2480 } 2481 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 2482 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 2483 return Builder.CreateStore(Ops[1], Ops[0]); 2484 } 2485 case ARM::BI__builtin_neon_vmax_v: 2486 case ARM::BI__builtin_neon_vmaxq_v: 2487 Int = usgn ? Intrinsic::arm_neon_vmaxu : Intrinsic::arm_neon_vmaxs; 2488 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmax"); 2489 case ARM::BI__builtin_neon_vmin_v: 2490 case ARM::BI__builtin_neon_vminq_v: 2491 Int = usgn ? Intrinsic::arm_neon_vminu : Intrinsic::arm_neon_vmins; 2492 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmin"); 2493 case ARM::BI__builtin_neon_vmovl_v: { 2494 llvm::Type *DTy =llvm::VectorType::getTruncatedElementVectorType(VTy); 2495 Ops[0] = Builder.CreateBitCast(Ops[0], DTy); 2496 if (usgn) 2497 return Builder.CreateZExt(Ops[0], Ty, "vmovl"); 2498 return Builder.CreateSExt(Ops[0], Ty, "vmovl"); 2499 } 2500 case ARM::BI__builtin_neon_vmovn_v: { 2501 llvm::Type *QTy = llvm::VectorType::getExtendedElementVectorType(VTy); 2502 Ops[0] = Builder.CreateBitCast(Ops[0], QTy); 2503 return Builder.CreateTrunc(Ops[0], Ty, "vmovn"); 2504 } 2505 case ARM::BI__builtin_neon_vmul_v: 2506 case ARM::BI__builtin_neon_vmulq_v: 2507 assert(Type.isPoly() && "vmul builtin only supported for polynomial types"); 2508 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vmulp, Ty), 2509 Ops, "vmul"); 2510 case ARM::BI__builtin_neon_vmull_v: 2511 Int = usgn ? Intrinsic::arm_neon_vmullu : Intrinsic::arm_neon_vmulls; 2512 Int = Type.isPoly() ? (unsigned)Intrinsic::arm_neon_vmullp : Int; 2513 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmull"); 2514 case ARM::BI__builtin_neon_vfma_v: 2515 case ARM::BI__builtin_neon_vfmaq_v: { 2516 Value *F = CGM.getIntrinsic(Intrinsic::fma, Ty); 2517 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 2518 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 2519 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 2520 2521 // NEON intrinsic puts accumulator first, unlike the LLVM fma. 2522 return Builder.CreateCall3(F, Ops[1], Ops[2], Ops[0]); 2523 } 2524 case ARM::BI__builtin_neon_vpadal_v: 2525 case ARM::BI__builtin_neon_vpadalq_v: { 2526 Int = usgn ? Intrinsic::arm_neon_vpadalu : Intrinsic::arm_neon_vpadals; 2527 // The source operand type has twice as many elements of half the size. 2528 unsigned EltBits = VTy->getElementType()->getPrimitiveSizeInBits(); 2529 llvm::Type *EltTy = 2530 llvm::IntegerType::get(getLLVMContext(), EltBits / 2); 2531 llvm::Type *NarrowTy = 2532 llvm::VectorType::get(EltTy, VTy->getNumElements() * 2); 2533 llvm::Type *Tys[2] = { Ty, NarrowTy }; 2534 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vpadal"); 2535 } 2536 case ARM::BI__builtin_neon_vpadd_v: 2537 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vpadd, Ty), 2538 Ops, "vpadd"); 2539 case ARM::BI__builtin_neon_vpaddl_v: 2540 case ARM::BI__builtin_neon_vpaddlq_v: { 2541 Int = usgn ? Intrinsic::arm_neon_vpaddlu : Intrinsic::arm_neon_vpaddls; 2542 // The source operand type has twice as many elements of half the size. 2543 unsigned EltBits = VTy->getElementType()->getPrimitiveSizeInBits(); 2544 llvm::Type *EltTy = llvm::IntegerType::get(getLLVMContext(), EltBits / 2); 2545 llvm::Type *NarrowTy = 2546 llvm::VectorType::get(EltTy, VTy->getNumElements() * 2); 2547 llvm::Type *Tys[2] = { Ty, NarrowTy }; 2548 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vpaddl"); 2549 } 2550 case ARM::BI__builtin_neon_vpmax_v: 2551 Int = usgn ? Intrinsic::arm_neon_vpmaxu : Intrinsic::arm_neon_vpmaxs; 2552 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vpmax"); 2553 case ARM::BI__builtin_neon_vpmin_v: 2554 Int = usgn ? Intrinsic::arm_neon_vpminu : Intrinsic::arm_neon_vpmins; 2555 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vpmin"); 2556 case ARM::BI__builtin_neon_vqabs_v: 2557 case ARM::BI__builtin_neon_vqabsq_v: 2558 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vqabs, Ty), 2559 Ops, "vqabs"); 2560 case ARM::BI__builtin_neon_vqadd_v: 2561 case ARM::BI__builtin_neon_vqaddq_v: 2562 Int = usgn ? Intrinsic::arm_neon_vqaddu : Intrinsic::arm_neon_vqadds; 2563 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqadd"); 2564 case ARM::BI__builtin_neon_vqdmlal_v: 2565 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vqdmlal, Ty), 2566 Ops, "vqdmlal"); 2567 case ARM::BI__builtin_neon_vqdmlsl_v: 2568 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vqdmlsl, Ty), 2569 Ops, "vqdmlsl"); 2570 case ARM::BI__builtin_neon_vqdmulh_v: 2571 case ARM::BI__builtin_neon_vqdmulhq_v: 2572 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vqdmulh, Ty), 2573 Ops, "vqdmulh"); 2574 case ARM::BI__builtin_neon_vqdmull_v: 2575 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vqdmull, Ty), 2576 Ops, "vqdmull"); 2577 case ARM::BI__builtin_neon_vqmovn_v: 2578 Int = usgn ? Intrinsic::arm_neon_vqmovnu : Intrinsic::arm_neon_vqmovns; 2579 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqmovn"); 2580 case ARM::BI__builtin_neon_vqmovun_v: 2581 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vqmovnsu, Ty), 2582 Ops, "vqdmull"); 2583 case ARM::BI__builtin_neon_vqneg_v: 2584 case ARM::BI__builtin_neon_vqnegq_v: 2585 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vqneg, Ty), 2586 Ops, "vqneg"); 2587 case ARM::BI__builtin_neon_vqrdmulh_v: 2588 case ARM::BI__builtin_neon_vqrdmulhq_v: 2589 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vqrdmulh, Ty), 2590 Ops, "vqrdmulh"); 2591 case ARM::BI__builtin_neon_vqrshl_v: 2592 case ARM::BI__builtin_neon_vqrshlq_v: 2593 Int = usgn ? Intrinsic::arm_neon_vqrshiftu : Intrinsic::arm_neon_vqrshifts; 2594 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqrshl"); 2595 case ARM::BI__builtin_neon_vqrshrn_n_v: 2596 Int = 2597 usgn ? Intrinsic::arm_neon_vqrshiftnu : Intrinsic::arm_neon_vqrshiftns; 2598 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqrshrn_n", 2599 1, true); 2600 case ARM::BI__builtin_neon_vqrshrun_n_v: 2601 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vqrshiftnsu, Ty), 2602 Ops, "vqrshrun_n", 1, true); 2603 case ARM::BI__builtin_neon_vqshl_v: 2604 case ARM::BI__builtin_neon_vqshlq_v: 2605 Int = usgn ? Intrinsic::arm_neon_vqshiftu : Intrinsic::arm_neon_vqshifts; 2606 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqshl"); 2607 case ARM::BI__builtin_neon_vqshl_n_v: 2608 case ARM::BI__builtin_neon_vqshlq_n_v: 2609 Int = usgn ? Intrinsic::arm_neon_vqshiftu : Intrinsic::arm_neon_vqshifts; 2610 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqshl_n", 2611 1, false); 2612 case ARM::BI__builtin_neon_vqshlu_n_v: 2613 case ARM::BI__builtin_neon_vqshluq_n_v: 2614 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vqshiftsu, Ty), 2615 Ops, "vqshlu", 1, false); 2616 case ARM::BI__builtin_neon_vqshrn_n_v: 2617 Int = usgn ? Intrinsic::arm_neon_vqshiftnu : Intrinsic::arm_neon_vqshiftns; 2618 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqshrn_n", 2619 1, true); 2620 case ARM::BI__builtin_neon_vqshrun_n_v: 2621 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vqshiftnsu, Ty), 2622 Ops, "vqshrun_n", 1, true); 2623 case ARM::BI__builtin_neon_vqsub_v: 2624 case ARM::BI__builtin_neon_vqsubq_v: 2625 Int = usgn ? Intrinsic::arm_neon_vqsubu : Intrinsic::arm_neon_vqsubs; 2626 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqsub"); 2627 case ARM::BI__builtin_neon_vraddhn_v: 2628 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vraddhn, Ty), 2629 Ops, "vraddhn"); 2630 case ARM::BI__builtin_neon_vrecpe_v: 2631 case ARM::BI__builtin_neon_vrecpeq_v: 2632 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vrecpe, Ty), 2633 Ops, "vrecpe"); 2634 case ARM::BI__builtin_neon_vrecps_v: 2635 case ARM::BI__builtin_neon_vrecpsq_v: 2636 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vrecps, Ty), 2637 Ops, "vrecps"); 2638 case ARM::BI__builtin_neon_vrhadd_v: 2639 case ARM::BI__builtin_neon_vrhaddq_v: 2640 Int = usgn ? Intrinsic::arm_neon_vrhaddu : Intrinsic::arm_neon_vrhadds; 2641 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrhadd"); 2642 case ARM::BI__builtin_neon_vrshl_v: 2643 case ARM::BI__builtin_neon_vrshlq_v: 2644 Int = usgn ? Intrinsic::arm_neon_vrshiftu : Intrinsic::arm_neon_vrshifts; 2645 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrshl"); 2646 case ARM::BI__builtin_neon_vrshrn_n_v: 2647 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vrshiftn, Ty), 2648 Ops, "vrshrn_n", 1, true); 2649 case ARM::BI__builtin_neon_vrshr_n_v: 2650 case ARM::BI__builtin_neon_vrshrq_n_v: 2651 Int = usgn ? Intrinsic::arm_neon_vrshiftu : Intrinsic::arm_neon_vrshifts; 2652 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrshr_n", 1, true); 2653 case ARM::BI__builtin_neon_vrsqrte_v: 2654 case ARM::BI__builtin_neon_vrsqrteq_v: 2655 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vrsqrte, Ty), 2656 Ops, "vrsqrte"); 2657 case ARM::BI__builtin_neon_vrsqrts_v: 2658 case ARM::BI__builtin_neon_vrsqrtsq_v: 2659 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vrsqrts, Ty), 2660 Ops, "vrsqrts"); 2661 case ARM::BI__builtin_neon_vrsra_n_v: 2662 case ARM::BI__builtin_neon_vrsraq_n_v: 2663 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 2664 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 2665 Ops[2] = EmitNeonShiftVector(Ops[2], Ty, true); 2666 Int = usgn ? Intrinsic::arm_neon_vrshiftu : Intrinsic::arm_neon_vrshifts; 2667 Ops[1] = Builder.CreateCall2(CGM.getIntrinsic(Int, Ty), Ops[1], Ops[2]); 2668 return Builder.CreateAdd(Ops[0], Ops[1], "vrsra_n"); 2669 case ARM::BI__builtin_neon_vrsubhn_v: 2670 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vrsubhn, Ty), 2671 Ops, "vrsubhn"); 2672 case ARM::BI__builtin_neon_vshl_v: 2673 case ARM::BI__builtin_neon_vshlq_v: 2674 Int = usgn ? Intrinsic::arm_neon_vshiftu : Intrinsic::arm_neon_vshifts; 2675 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vshl"); 2676 case ARM::BI__builtin_neon_vshll_n_v: 2677 Int = usgn ? Intrinsic::arm_neon_vshiftlu : Intrinsic::arm_neon_vshiftls; 2678 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vshll", 1); 2679 case ARM::BI__builtin_neon_vshl_n_v: 2680 case ARM::BI__builtin_neon_vshlq_n_v: 2681 Ops[1] = EmitNeonShiftVector(Ops[1], Ty, false); 2682 return Builder.CreateShl(Builder.CreateBitCast(Ops[0],Ty), Ops[1], 2683 "vshl_n"); 2684 case ARM::BI__builtin_neon_vshrn_n_v: 2685 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vshiftn, Ty), 2686 Ops, "vshrn_n", 1, true); 2687 case ARM::BI__builtin_neon_vshr_n_v: 2688 case ARM::BI__builtin_neon_vshrq_n_v: 2689 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 2690 Ops[1] = EmitNeonShiftVector(Ops[1], Ty, false); 2691 if (usgn) 2692 return Builder.CreateLShr(Ops[0], Ops[1], "vshr_n"); 2693 else 2694 return Builder.CreateAShr(Ops[0], Ops[1], "vshr_n"); 2695 case ARM::BI__builtin_neon_vsri_n_v: 2696 case ARM::BI__builtin_neon_vsriq_n_v: 2697 rightShift = true; 2698 case ARM::BI__builtin_neon_vsli_n_v: 2699 case ARM::BI__builtin_neon_vsliq_n_v: 2700 Ops[2] = EmitNeonShiftVector(Ops[2], Ty, rightShift); 2701 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vshiftins, Ty), 2702 Ops, "vsli_n"); 2703 case ARM::BI__builtin_neon_vsra_n_v: 2704 case ARM::BI__builtin_neon_vsraq_n_v: 2705 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 2706 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 2707 Ops[2] = EmitNeonShiftVector(Ops[2], Ty, false); 2708 if (usgn) 2709 Ops[1] = Builder.CreateLShr(Ops[1], Ops[2], "vsra_n"); 2710 else 2711 Ops[1] = Builder.CreateAShr(Ops[1], Ops[2], "vsra_n"); 2712 return Builder.CreateAdd(Ops[0], Ops[1]); 2713 case ARM::BI__builtin_neon_vst1_v: 2714 case ARM::BI__builtin_neon_vst1q_v: 2715 Ops.push_back(Align); 2716 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vst1, Ty), 2717 Ops, ""); 2718 case ARM::BI__builtin_neon_vst1q_lane_v: 2719 // Handle 64-bit integer elements as a special case. Use a shuffle to get 2720 // a one-element vector and avoid poor code for i64 in the backend. 2721 if (VTy->getElementType()->isIntegerTy(64)) { 2722 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 2723 Value *SV = llvm::ConstantVector::get(cast<llvm::Constant>(Ops[2])); 2724 Ops[1] = Builder.CreateShuffleVector(Ops[1], Ops[1], SV); 2725 Ops[2] = Align; 2726 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::arm_neon_vst1, 2727 Ops[1]->getType()), Ops); 2728 } 2729 // fall through 2730 case ARM::BI__builtin_neon_vst1_lane_v: { 2731 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 2732 Ops[1] = Builder.CreateExtractElement(Ops[1], Ops[2]); 2733 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 2734 StoreInst *St = Builder.CreateStore(Ops[1], 2735 Builder.CreateBitCast(Ops[0], Ty)); 2736 St->setAlignment(cast<ConstantInt>(Align)->getZExtValue()); 2737 return St; 2738 } 2739 case ARM::BI__builtin_neon_vst2_v: 2740 case ARM::BI__builtin_neon_vst2q_v: 2741 Ops.push_back(Align); 2742 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vst2, Ty), 2743 Ops, ""); 2744 case ARM::BI__builtin_neon_vst2_lane_v: 2745 case ARM::BI__builtin_neon_vst2q_lane_v: 2746 Ops.push_back(Align); 2747 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vst2lane, Ty), 2748 Ops, ""); 2749 case ARM::BI__builtin_neon_vst3_v: 2750 case ARM::BI__builtin_neon_vst3q_v: 2751 Ops.push_back(Align); 2752 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vst3, Ty), 2753 Ops, ""); 2754 case ARM::BI__builtin_neon_vst3_lane_v: 2755 case ARM::BI__builtin_neon_vst3q_lane_v: 2756 Ops.push_back(Align); 2757 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vst3lane, Ty), 2758 Ops, ""); 2759 case ARM::BI__builtin_neon_vst4_v: 2760 case ARM::BI__builtin_neon_vst4q_v: 2761 Ops.push_back(Align); 2762 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vst4, Ty), 2763 Ops, ""); 2764 case ARM::BI__builtin_neon_vst4_lane_v: 2765 case ARM::BI__builtin_neon_vst4q_lane_v: 2766 Ops.push_back(Align); 2767 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vst4lane, Ty), 2768 Ops, ""); 2769 case ARM::BI__builtin_neon_vsubhn_v: 2770 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vsubhn, Ty), 2771 Ops, "vsubhn"); 2772 case ARM::BI__builtin_neon_vtbl1_v: 2773 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbl1), 2774 Ops, "vtbl1"); 2775 case ARM::BI__builtin_neon_vtbl2_v: 2776 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbl2), 2777 Ops, "vtbl2"); 2778 case ARM::BI__builtin_neon_vtbl3_v: 2779 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbl3), 2780 Ops, "vtbl3"); 2781 case ARM::BI__builtin_neon_vtbl4_v: 2782 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbl4), 2783 Ops, "vtbl4"); 2784 case ARM::BI__builtin_neon_vtbx1_v: 2785 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbx1), 2786 Ops, "vtbx1"); 2787 case ARM::BI__builtin_neon_vtbx2_v: 2788 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbx2), 2789 Ops, "vtbx2"); 2790 case ARM::BI__builtin_neon_vtbx3_v: 2791 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbx3), 2792 Ops, "vtbx3"); 2793 case ARM::BI__builtin_neon_vtbx4_v: 2794 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbx4), 2795 Ops, "vtbx4"); 2796 case ARM::BI__builtin_neon_vtst_v: 2797 case ARM::BI__builtin_neon_vtstq_v: { 2798 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 2799 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 2800 Ops[0] = Builder.CreateAnd(Ops[0], Ops[1]); 2801 Ops[0] = Builder.CreateICmp(ICmpInst::ICMP_NE, Ops[0], 2802 ConstantAggregateZero::get(Ty)); 2803 return Builder.CreateSExt(Ops[0], Ty, "vtst"); 2804 } 2805 case ARM::BI__builtin_neon_vtrn_v: 2806 case ARM::BI__builtin_neon_vtrnq_v: { 2807 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty)); 2808 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 2809 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 2810 Value *SV = 0; 2811 2812 for (unsigned vi = 0; vi != 2; ++vi) { 2813 SmallVector<Constant*, 16> Indices; 2814 for (unsigned i = 0, e = VTy->getNumElements(); i != e; i += 2) { 2815 Indices.push_back(Builder.getInt32(i+vi)); 2816 Indices.push_back(Builder.getInt32(i+e+vi)); 2817 } 2818 Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ops[0], vi); 2819 SV = llvm::ConstantVector::get(Indices); 2820 SV = Builder.CreateShuffleVector(Ops[1], Ops[2], SV, "vtrn"); 2821 SV = Builder.CreateStore(SV, Addr); 2822 } 2823 return SV; 2824 } 2825 case ARM::BI__builtin_neon_vuzp_v: 2826 case ARM::BI__builtin_neon_vuzpq_v: { 2827 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty)); 2828 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 2829 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 2830 Value *SV = 0; 2831 2832 for (unsigned vi = 0; vi != 2; ++vi) { 2833 SmallVector<Constant*, 16> Indices; 2834 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) 2835 Indices.push_back(ConstantInt::get(Int32Ty, 2*i+vi)); 2836 2837 Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ops[0], vi); 2838 SV = llvm::ConstantVector::get(Indices); 2839 SV = Builder.CreateShuffleVector(Ops[1], Ops[2], SV, "vuzp"); 2840 SV = Builder.CreateStore(SV, Addr); 2841 } 2842 return SV; 2843 } 2844 case ARM::BI__builtin_neon_vzip_v: 2845 case ARM::BI__builtin_neon_vzipq_v: { 2846 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty)); 2847 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 2848 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 2849 Value *SV = 0; 2850 2851 for (unsigned vi = 0; vi != 2; ++vi) { 2852 SmallVector<Constant*, 16> Indices; 2853 for (unsigned i = 0, e = VTy->getNumElements(); i != e; i += 2) { 2854 Indices.push_back(ConstantInt::get(Int32Ty, (i + vi*e) >> 1)); 2855 Indices.push_back(ConstantInt::get(Int32Ty, ((i + vi*e) >> 1)+e)); 2856 } 2857 Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ops[0], vi); 2858 SV = llvm::ConstantVector::get(Indices); 2859 SV = Builder.CreateShuffleVector(Ops[1], Ops[2], SV, "vzip"); 2860 SV = Builder.CreateStore(SV, Addr); 2861 } 2862 return SV; 2863 } 2864 } 2865} 2866 2867llvm::Value *CodeGenFunction:: 2868BuildVector(ArrayRef<llvm::Value*> Ops) { 2869 assert((Ops.size() & (Ops.size() - 1)) == 0 && 2870 "Not a power-of-two sized vector!"); 2871 bool AllConstants = true; 2872 for (unsigned i = 0, e = Ops.size(); i != e && AllConstants; ++i) 2873 AllConstants &= isa<Constant>(Ops[i]); 2874 2875 // If this is a constant vector, create a ConstantVector. 2876 if (AllConstants) { 2877 SmallVector<llvm::Constant*, 16> CstOps; 2878 for (unsigned i = 0, e = Ops.size(); i != e; ++i) 2879 CstOps.push_back(cast<Constant>(Ops[i])); 2880 return llvm::ConstantVector::get(CstOps); 2881 } 2882 2883 // Otherwise, insertelement the values to build the vector. 2884 Value *Result = 2885 llvm::UndefValue::get(llvm::VectorType::get(Ops[0]->getType(), Ops.size())); 2886 2887 for (unsigned i = 0, e = Ops.size(); i != e; ++i) 2888 Result = Builder.CreateInsertElement(Result, Ops[i], Builder.getInt32(i)); 2889 2890 return Result; 2891} 2892 2893Value *CodeGenFunction::EmitX86BuiltinExpr(unsigned BuiltinID, 2894 const CallExpr *E) { 2895 SmallVector<Value*, 4> Ops; 2896 2897 // Find out if any arguments are required to be integer constant expressions. 2898 unsigned ICEArguments = 0; 2899 ASTContext::GetBuiltinTypeError Error; 2900 getContext().GetBuiltinType(BuiltinID, Error, &ICEArguments); 2901 assert(Error == ASTContext::GE_None && "Should not codegen an error"); 2902 2903 for (unsigned i = 0, e = E->getNumArgs(); i != e; i++) { 2904 // If this is a normal argument, just emit it as a scalar. 2905 if ((ICEArguments & (1 << i)) == 0) { 2906 Ops.push_back(EmitScalarExpr(E->getArg(i))); 2907 continue; 2908 } 2909 2910 // If this is required to be a constant, constant fold it so that we know 2911 // that the generated intrinsic gets a ConstantInt. 2912 llvm::APSInt Result; 2913 bool IsConst = E->getArg(i)->isIntegerConstantExpr(Result, getContext()); 2914 assert(IsConst && "Constant arg isn't actually constant?"); (void)IsConst; 2915 Ops.push_back(llvm::ConstantInt::get(getLLVMContext(), Result)); 2916 } 2917 2918 switch (BuiltinID) { 2919 default: return 0; 2920 case X86::BI__builtin_ia32_vec_init_v8qi: 2921 case X86::BI__builtin_ia32_vec_init_v4hi: 2922 case X86::BI__builtin_ia32_vec_init_v2si: 2923 return Builder.CreateBitCast(BuildVector(Ops), 2924 llvm::Type::getX86_MMXTy(getLLVMContext())); 2925 case X86::BI__builtin_ia32_vec_ext_v2si: 2926 return Builder.CreateExtractElement(Ops[0], 2927 llvm::ConstantInt::get(Ops[1]->getType(), 0)); 2928 case X86::BI__builtin_ia32_ldmxcsr: { 2929 Value *Tmp = CreateMemTemp(E->getArg(0)->getType()); 2930 Builder.CreateStore(Ops[0], Tmp); 2931 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::x86_sse_ldmxcsr), 2932 Builder.CreateBitCast(Tmp, Int8PtrTy)); 2933 } 2934 case X86::BI__builtin_ia32_stmxcsr: { 2935 Value *Tmp = CreateMemTemp(E->getType()); 2936 Builder.CreateCall(CGM.getIntrinsic(Intrinsic::x86_sse_stmxcsr), 2937 Builder.CreateBitCast(Tmp, Int8PtrTy)); 2938 return Builder.CreateLoad(Tmp, "stmxcsr"); 2939 } 2940 case X86::BI__builtin_ia32_storehps: 2941 case X86::BI__builtin_ia32_storelps: { 2942 llvm::Type *PtrTy = llvm::PointerType::getUnqual(Int64Ty); 2943 llvm::Type *VecTy = llvm::VectorType::get(Int64Ty, 2); 2944 2945 // cast val v2i64 2946 Ops[1] = Builder.CreateBitCast(Ops[1], VecTy, "cast"); 2947 2948 // extract (0, 1) 2949 unsigned Index = BuiltinID == X86::BI__builtin_ia32_storelps ? 0 : 1; 2950 llvm::Value *Idx = llvm::ConstantInt::get(Int32Ty, Index); 2951 Ops[1] = Builder.CreateExtractElement(Ops[1], Idx, "extract"); 2952 2953 // cast pointer to i64 & store 2954 Ops[0] = Builder.CreateBitCast(Ops[0], PtrTy); 2955 return Builder.CreateStore(Ops[1], Ops[0]); 2956 } 2957 case X86::BI__builtin_ia32_palignr: { 2958 unsigned shiftVal = cast<llvm::ConstantInt>(Ops[2])->getZExtValue(); 2959 2960 // If palignr is shifting the pair of input vectors less than 9 bytes, 2961 // emit a shuffle instruction. 2962 if (shiftVal <= 8) { 2963 SmallVector<llvm::Constant*, 8> Indices; 2964 for (unsigned i = 0; i != 8; ++i) 2965 Indices.push_back(llvm::ConstantInt::get(Int32Ty, shiftVal + i)); 2966 2967 Value* SV = llvm::ConstantVector::get(Indices); 2968 return Builder.CreateShuffleVector(Ops[1], Ops[0], SV, "palignr"); 2969 } 2970 2971 // If palignr is shifting the pair of input vectors more than 8 but less 2972 // than 16 bytes, emit a logical right shift of the destination. 2973 if (shiftVal < 16) { 2974 // MMX has these as 1 x i64 vectors for some odd optimization reasons. 2975 llvm::Type *VecTy = llvm::VectorType::get(Int64Ty, 1); 2976 2977 Ops[0] = Builder.CreateBitCast(Ops[0], VecTy, "cast"); 2978 Ops[1] = llvm::ConstantInt::get(VecTy, (shiftVal-8) * 8); 2979 2980 // create i32 constant 2981 llvm::Function *F = CGM.getIntrinsic(Intrinsic::x86_mmx_psrl_q); 2982 return Builder.CreateCall(F, makeArrayRef(&Ops[0], 2), "palignr"); 2983 } 2984 2985 // If palignr is shifting the pair of vectors more than 16 bytes, emit zero. 2986 return llvm::Constant::getNullValue(ConvertType(E->getType())); 2987 } 2988 case X86::BI__builtin_ia32_palignr128: { 2989 unsigned shiftVal = cast<llvm::ConstantInt>(Ops[2])->getZExtValue(); 2990 2991 // If palignr is shifting the pair of input vectors less than 17 bytes, 2992 // emit a shuffle instruction. 2993 if (shiftVal <= 16) { 2994 SmallVector<llvm::Constant*, 16> Indices; 2995 for (unsigned i = 0; i != 16; ++i) 2996 Indices.push_back(llvm::ConstantInt::get(Int32Ty, shiftVal + i)); 2997 2998 Value* SV = llvm::ConstantVector::get(Indices); 2999 return Builder.CreateShuffleVector(Ops[1], Ops[0], SV, "palignr"); 3000 } 3001 3002 // If palignr is shifting the pair of input vectors more than 16 but less 3003 // than 32 bytes, emit a logical right shift of the destination. 3004 if (shiftVal < 32) { 3005 llvm::Type *VecTy = llvm::VectorType::get(Int64Ty, 2); 3006 3007 Ops[0] = Builder.CreateBitCast(Ops[0], VecTy, "cast"); 3008 Ops[1] = llvm::ConstantInt::get(Int32Ty, (shiftVal-16) * 8); 3009 3010 // create i32 constant 3011 llvm::Function *F = CGM.getIntrinsic(Intrinsic::x86_sse2_psrl_dq); 3012 return Builder.CreateCall(F, makeArrayRef(&Ops[0], 2), "palignr"); 3013 } 3014 3015 // If palignr is shifting the pair of vectors more than 32 bytes, emit zero. 3016 return llvm::Constant::getNullValue(ConvertType(E->getType())); 3017 } 3018 case X86::BI__builtin_ia32_palignr256: { 3019 unsigned shiftVal = cast<llvm::ConstantInt>(Ops[2])->getZExtValue(); 3020 3021 // If palignr is shifting the pair of input vectors less than 17 bytes, 3022 // emit a shuffle instruction. 3023 if (shiftVal <= 16) { 3024 SmallVector<llvm::Constant*, 32> Indices; 3025 // 256-bit palignr operates on 128-bit lanes so we need to handle that 3026 for (unsigned l = 0; l != 2; ++l) { 3027 unsigned LaneStart = l * 16; 3028 unsigned LaneEnd = (l+1) * 16; 3029 for (unsigned i = 0; i != 16; ++i) { 3030 unsigned Idx = shiftVal + i + LaneStart; 3031 if (Idx >= LaneEnd) Idx += 16; // end of lane, switch operand 3032 Indices.push_back(llvm::ConstantInt::get(Int32Ty, Idx)); 3033 } 3034 } 3035 3036 Value* SV = llvm::ConstantVector::get(Indices); 3037 return Builder.CreateShuffleVector(Ops[1], Ops[0], SV, "palignr"); 3038 } 3039 3040 // If palignr is shifting the pair of input vectors more than 16 but less 3041 // than 32 bytes, emit a logical right shift of the destination. 3042 if (shiftVal < 32) { 3043 llvm::Type *VecTy = llvm::VectorType::get(Int64Ty, 4); 3044 3045 Ops[0] = Builder.CreateBitCast(Ops[0], VecTy, "cast"); 3046 Ops[1] = llvm::ConstantInt::get(Int32Ty, (shiftVal-16) * 8); 3047 3048 // create i32 constant 3049 llvm::Function *F = CGM.getIntrinsic(Intrinsic::x86_avx2_psrl_dq); 3050 return Builder.CreateCall(F, makeArrayRef(&Ops[0], 2), "palignr"); 3051 } 3052 3053 // If palignr is shifting the pair of vectors more than 32 bytes, emit zero. 3054 return llvm::Constant::getNullValue(ConvertType(E->getType())); 3055 } 3056 case X86::BI__builtin_ia32_movntps: 3057 case X86::BI__builtin_ia32_movntps256: 3058 case X86::BI__builtin_ia32_movntpd: 3059 case X86::BI__builtin_ia32_movntpd256: 3060 case X86::BI__builtin_ia32_movntdq: 3061 case X86::BI__builtin_ia32_movntdq256: 3062 case X86::BI__builtin_ia32_movnti: { 3063 llvm::MDNode *Node = llvm::MDNode::get(getLLVMContext(), 3064 Builder.getInt32(1)); 3065 3066 // Convert the type of the pointer to a pointer to the stored type. 3067 Value *BC = Builder.CreateBitCast(Ops[0], 3068 llvm::PointerType::getUnqual(Ops[1]->getType()), 3069 "cast"); 3070 StoreInst *SI = Builder.CreateStore(Ops[1], BC); 3071 SI->setMetadata(CGM.getModule().getMDKindID("nontemporal"), Node); 3072 SI->setAlignment(16); 3073 return SI; 3074 } 3075 // 3DNow! 3076 case X86::BI__builtin_ia32_pswapdsf: 3077 case X86::BI__builtin_ia32_pswapdsi: { 3078 const char *name = 0; 3079 Intrinsic::ID ID = Intrinsic::not_intrinsic; 3080 switch(BuiltinID) { 3081 default: llvm_unreachable("Unsupported intrinsic!"); 3082 case X86::BI__builtin_ia32_pswapdsf: 3083 case X86::BI__builtin_ia32_pswapdsi: 3084 name = "pswapd"; 3085 ID = Intrinsic::x86_3dnowa_pswapd; 3086 break; 3087 } 3088 llvm::Type *MMXTy = llvm::Type::getX86_MMXTy(getLLVMContext()); 3089 Ops[0] = Builder.CreateBitCast(Ops[0], MMXTy, "cast"); 3090 llvm::Function *F = CGM.getIntrinsic(ID); 3091 return Builder.CreateCall(F, Ops, name); 3092 } 3093 case X86::BI__builtin_ia32_rdrand16_step: 3094 case X86::BI__builtin_ia32_rdrand32_step: 3095 case X86::BI__builtin_ia32_rdrand64_step: 3096 case X86::BI__builtin_ia32_rdseed16_step: 3097 case X86::BI__builtin_ia32_rdseed32_step: 3098 case X86::BI__builtin_ia32_rdseed64_step: { 3099 Intrinsic::ID ID; 3100 switch (BuiltinID) { 3101 default: llvm_unreachable("Unsupported intrinsic!"); 3102 case X86::BI__builtin_ia32_rdrand16_step: 3103 ID = Intrinsic::x86_rdrand_16; 3104 break; 3105 case X86::BI__builtin_ia32_rdrand32_step: 3106 ID = Intrinsic::x86_rdrand_32; 3107 break; 3108 case X86::BI__builtin_ia32_rdrand64_step: 3109 ID = Intrinsic::x86_rdrand_64; 3110 break; 3111 case X86::BI__builtin_ia32_rdseed16_step: 3112 ID = Intrinsic::x86_rdseed_16; 3113 break; 3114 case X86::BI__builtin_ia32_rdseed32_step: 3115 ID = Intrinsic::x86_rdseed_32; 3116 break; 3117 case X86::BI__builtin_ia32_rdseed64_step: 3118 ID = Intrinsic::x86_rdseed_64; 3119 break; 3120 } 3121 3122 Value *Call = Builder.CreateCall(CGM.getIntrinsic(ID)); 3123 Builder.CreateStore(Builder.CreateExtractValue(Call, 0), Ops[0]); 3124 return Builder.CreateExtractValue(Call, 1); 3125 } 3126 // AVX2 broadcast 3127 case X86::BI__builtin_ia32_vbroadcastsi256: { 3128 Value *VecTmp = CreateMemTemp(E->getArg(0)->getType()); 3129 Builder.CreateStore(Ops[0], VecTmp); 3130 Value *F = CGM.getIntrinsic(Intrinsic::x86_avx2_vbroadcasti128); 3131 return Builder.CreateCall(F, Builder.CreateBitCast(VecTmp, Int8PtrTy)); 3132 } 3133 } 3134} 3135 3136 3137Value *CodeGenFunction::EmitPPCBuiltinExpr(unsigned BuiltinID, 3138 const CallExpr *E) { 3139 SmallVector<Value*, 4> Ops; 3140 3141 for (unsigned i = 0, e = E->getNumArgs(); i != e; i++) 3142 Ops.push_back(EmitScalarExpr(E->getArg(i))); 3143 3144 Intrinsic::ID ID = Intrinsic::not_intrinsic; 3145 3146 switch (BuiltinID) { 3147 default: return 0; 3148 3149 // vec_ld, vec_lvsl, vec_lvsr 3150 case PPC::BI__builtin_altivec_lvx: 3151 case PPC::BI__builtin_altivec_lvxl: 3152 case PPC::BI__builtin_altivec_lvebx: 3153 case PPC::BI__builtin_altivec_lvehx: 3154 case PPC::BI__builtin_altivec_lvewx: 3155 case PPC::BI__builtin_altivec_lvsl: 3156 case PPC::BI__builtin_altivec_lvsr: 3157 { 3158 Ops[1] = Builder.CreateBitCast(Ops[1], Int8PtrTy); 3159 3160 Ops[0] = Builder.CreateGEP(Ops[1], Ops[0]); 3161 Ops.pop_back(); 3162 3163 switch (BuiltinID) { 3164 default: llvm_unreachable("Unsupported ld/lvsl/lvsr intrinsic!"); 3165 case PPC::BI__builtin_altivec_lvx: 3166 ID = Intrinsic::ppc_altivec_lvx; 3167 break; 3168 case PPC::BI__builtin_altivec_lvxl: 3169 ID = Intrinsic::ppc_altivec_lvxl; 3170 break; 3171 case PPC::BI__builtin_altivec_lvebx: 3172 ID = Intrinsic::ppc_altivec_lvebx; 3173 break; 3174 case PPC::BI__builtin_altivec_lvehx: 3175 ID = Intrinsic::ppc_altivec_lvehx; 3176 break; 3177 case PPC::BI__builtin_altivec_lvewx: 3178 ID = Intrinsic::ppc_altivec_lvewx; 3179 break; 3180 case PPC::BI__builtin_altivec_lvsl: 3181 ID = Intrinsic::ppc_altivec_lvsl; 3182 break; 3183 case PPC::BI__builtin_altivec_lvsr: 3184 ID = Intrinsic::ppc_altivec_lvsr; 3185 break; 3186 } 3187 llvm::Function *F = CGM.getIntrinsic(ID); 3188 return Builder.CreateCall(F, Ops, ""); 3189 } 3190 3191 // vec_st 3192 case PPC::BI__builtin_altivec_stvx: 3193 case PPC::BI__builtin_altivec_stvxl: 3194 case PPC::BI__builtin_altivec_stvebx: 3195 case PPC::BI__builtin_altivec_stvehx: 3196 case PPC::BI__builtin_altivec_stvewx: 3197 { 3198 Ops[2] = Builder.CreateBitCast(Ops[2], Int8PtrTy); 3199 Ops[1] = Builder.CreateGEP(Ops[2], Ops[1]); 3200 Ops.pop_back(); 3201 3202 switch (BuiltinID) { 3203 default: llvm_unreachable("Unsupported st intrinsic!"); 3204 case PPC::BI__builtin_altivec_stvx: 3205 ID = Intrinsic::ppc_altivec_stvx; 3206 break; 3207 case PPC::BI__builtin_altivec_stvxl: 3208 ID = Intrinsic::ppc_altivec_stvxl; 3209 break; 3210 case PPC::BI__builtin_altivec_stvebx: 3211 ID = Intrinsic::ppc_altivec_stvebx; 3212 break; 3213 case PPC::BI__builtin_altivec_stvehx: 3214 ID = Intrinsic::ppc_altivec_stvehx; 3215 break; 3216 case PPC::BI__builtin_altivec_stvewx: 3217 ID = Intrinsic::ppc_altivec_stvewx; 3218 break; 3219 } 3220 llvm::Function *F = CGM.getIntrinsic(ID); 3221 return Builder.CreateCall(F, Ops, ""); 3222 } 3223 } 3224} 3225