1 //===---- CGBuiltin.cpp - Emit LLVM Code for builtins ---------------------===// 2 // 3 // The LLVM Compiler Infrastructure 4 // 5 // This file is distributed under the University of Illinois Open Source 6 // License. See LICENSE.TXT for details. 7 // 8 //===----------------------------------------------------------------------===// 9 // 10 // This contains code to emit Builtin calls as LLVM code. 11 // 12 //===----------------------------------------------------------------------===// 13 14 #include "TargetInfo.h" 15 #include "CodeGenFunction.h" 16 #include "CodeGenModule.h" 17 #include "CGObjCRuntime.h" 18 #include "clang/Basic/TargetInfo.h" 19 #include "clang/AST/APValue.h" 20 #include "clang/AST/ASTContext.h" 21 #include "clang/AST/Decl.h" 22 #include "clang/Basic/TargetBuiltins.h" 23 #include "llvm/Intrinsics.h" 24 #include "llvm/Target/TargetData.h" 25 26 using namespace clang; 27 using namespace CodeGen; 28 using namespace llvm; 29 30 /// getBuiltinLibFunction - Given a builtin id for a function like 31 /// "__builtin_fabsf", return a Function* for "fabsf". 32 llvm::Value *CodeGenModule::getBuiltinLibFunction(const FunctionDecl *FD, 33 unsigned BuiltinID) { 34 assert(Context.BuiltinInfo.isLibFunction(BuiltinID)); 35 36 // Get the name, skip over the __builtin_ prefix (if necessary). 37 StringRef Name; 38 GlobalDecl D(FD); 39 40 // If the builtin has been declared explicitly with an assembler label, 41 // use the mangled name. This differs from the plain label on platforms 42 // that prefix labels. 43 if (FD->hasAttr<AsmLabelAttr>()) 44 Name = getMangledName(D); 45 else 46 Name = Context.BuiltinInfo.GetName(BuiltinID) + 10; 47 48 llvm::FunctionType *Ty = 49 cast<llvm::FunctionType>(getTypes().ConvertType(FD->getType())); 50 51 return GetOrCreateLLVMFunction(Name, Ty, D, /*ForVTable=*/false); 52 } 53 54 /// Emit the conversions required to turn the given value into an 55 /// integer of the given size. 56 static Value *EmitToInt(CodeGenFunction &CGF, llvm::Value *V, 57 QualType T, llvm::IntegerType *IntType) { 58 V = CGF.EmitToMemory(V, T); 59 60 if (V->getType()->isPointerTy()) 61 return CGF.Builder.CreatePtrToInt(V, IntType); 62 63 assert(V->getType() == IntType); 64 return V; 65 } 66 67 static Value *EmitFromInt(CodeGenFunction &CGF, llvm::Value *V, 68 QualType T, llvm::Type *ResultType) { 69 V = CGF.EmitFromMemory(V, T); 70 71 if (ResultType->isPointerTy()) 72 return CGF.Builder.CreateIntToPtr(V, ResultType); 73 74 assert(V->getType() == ResultType); 75 return V; 76 } 77 78 /// Utility to insert an atomic instruction based on Instrinsic::ID 79 /// and the expression node. 80 static RValue EmitBinaryAtomic(CodeGenFunction &CGF, 81 llvm::AtomicRMWInst::BinOp Kind, 82 const CallExpr *E) { 83 QualType T = E->getType(); 84 assert(E->getArg(0)->getType()->isPointerType()); 85 assert(CGF.getContext().hasSameUnqualifiedType(T, 86 E->getArg(0)->getType()->getPointeeType())); 87 assert(CGF.getContext().hasSameUnqualifiedType(T, E->getArg(1)->getType())); 88 89 llvm::Value *DestPtr = CGF.EmitScalarExpr(E->getArg(0)); 90 unsigned AddrSpace = 91 cast<llvm::PointerType>(DestPtr->getType())->getAddressSpace(); 92 93 llvm::IntegerType *IntType = 94 llvm::IntegerType::get(CGF.getLLVMContext(), 95 CGF.getContext().getTypeSize(T)); 96 llvm::Type *IntPtrType = IntType->getPointerTo(AddrSpace); 97 98 llvm::Value *Args[2]; 99 Args[0] = CGF.Builder.CreateBitCast(DestPtr, IntPtrType); 100 Args[1] = CGF.EmitScalarExpr(E->getArg(1)); 101 llvm::Type *ValueType = Args[1]->getType(); 102 Args[1] = EmitToInt(CGF, Args[1], T, IntType); 103 104 llvm::Value *Result = 105 CGF.Builder.CreateAtomicRMW(Kind, Args[0], Args[1], 106 llvm::SequentiallyConsistent); 107 Result = EmitFromInt(CGF, Result, T, ValueType); 108 return RValue::get(Result); 109 } 110 111 /// Utility to insert an atomic instruction based Instrinsic::ID and 112 /// the expression node, where the return value is the result of the 113 /// operation. 114 static RValue EmitBinaryAtomicPost(CodeGenFunction &CGF, 115 llvm::AtomicRMWInst::BinOp Kind, 116 const CallExpr *E, 117 Instruction::BinaryOps Op) { 118 QualType T = E->getType(); 119 assert(E->getArg(0)->getType()->isPointerType()); 120 assert(CGF.getContext().hasSameUnqualifiedType(T, 121 E->getArg(0)->getType()->getPointeeType())); 122 assert(CGF.getContext().hasSameUnqualifiedType(T, E->getArg(1)->getType())); 123 124 llvm::Value *DestPtr = CGF.EmitScalarExpr(E->getArg(0)); 125 unsigned AddrSpace = 126 cast<llvm::PointerType>(DestPtr->getType())->getAddressSpace(); 127 128 llvm::IntegerType *IntType = 129 llvm::IntegerType::get(CGF.getLLVMContext(), 130 CGF.getContext().getTypeSize(T)); 131 llvm::Type *IntPtrType = IntType->getPointerTo(AddrSpace); 132 133 llvm::Value *Args[2]; 134 Args[1] = CGF.EmitScalarExpr(E->getArg(1)); 135 llvm::Type *ValueType = Args[1]->getType(); 136 Args[1] = EmitToInt(CGF, Args[1], T, IntType); 137 Args[0] = CGF.Builder.CreateBitCast(DestPtr, IntPtrType); 138 139 llvm::Value *Result = 140 CGF.Builder.CreateAtomicRMW(Kind, Args[0], Args[1], 141 llvm::SequentiallyConsistent); 142 Result = CGF.Builder.CreateBinOp(Op, Result, Args[1]); 143 Result = EmitFromInt(CGF, Result, T, ValueType); 144 return RValue::get(Result); 145 } 146 147 /// EmitFAbs - Emit a call to fabs/fabsf/fabsl, depending on the type of ValTy, 148 /// which must be a scalar floating point type. 149 static Value *EmitFAbs(CodeGenFunction &CGF, Value *V, QualType ValTy) { 150 const BuiltinType *ValTyP = ValTy->getAs<BuiltinType>(); 151 assert(ValTyP && "isn't scalar fp type!"); 152 153 StringRef FnName; 154 switch (ValTyP->getKind()) { 155 default: llvm_unreachable("Isn't a scalar fp type!"); 156 case BuiltinType::Float: FnName = "fabsf"; break; 157 case BuiltinType::Double: FnName = "fabs"; break; 158 case BuiltinType::LongDouble: FnName = "fabsl"; break; 159 } 160 161 // The prototype is something that takes and returns whatever V's type is. 162 llvm::FunctionType *FT = llvm::FunctionType::get(V->getType(), V->getType(), 163 false); 164 llvm::Value *Fn = CGF.CGM.CreateRuntimeFunction(FT, FnName); 165 166 return CGF.Builder.CreateCall(Fn, V, "abs"); 167 } 168 169 static RValue emitLibraryCall(CodeGenFunction &CGF, const FunctionDecl *Fn, 170 const CallExpr *E, llvm::Value *calleeValue) { 171 return CGF.EmitCall(E->getCallee()->getType(), calleeValue, 172 ReturnValueSlot(), E->arg_begin(), E->arg_end(), Fn); 173 } 174 175 RValue CodeGenFunction::EmitBuiltinExpr(const FunctionDecl *FD, 176 unsigned BuiltinID, const CallExpr *E) { 177 // See if we can constant fold this builtin. If so, don't emit it at all. 178 Expr::EvalResult Result; 179 if (E->Evaluate(Result, CGM.getContext()) && 180 !Result.hasSideEffects()) { 181 if (Result.Val.isInt()) 182 return RValue::get(llvm::ConstantInt::get(getLLVMContext(), 183 Result.Val.getInt())); 184 if (Result.Val.isFloat()) 185 return RValue::get(llvm::ConstantFP::get(getLLVMContext(), 186 Result.Val.getFloat())); 187 } 188 189 switch (BuiltinID) { 190 default: break; // Handle intrinsics and libm functions below. 191 case Builtin::BI__builtin___CFStringMakeConstantString: 192 case Builtin::BI__builtin___NSStringMakeConstantString: 193 return RValue::get(CGM.EmitConstantExpr(E, E->getType(), 0)); 194 case Builtin::BI__builtin_stdarg_start: 195 case Builtin::BI__builtin_va_start: 196 case Builtin::BI__builtin_va_end: { 197 Value *ArgValue = EmitVAListRef(E->getArg(0)); 198 llvm::Type *DestType = Int8PtrTy; 199 if (ArgValue->getType() != DestType) 200 ArgValue = Builder.CreateBitCast(ArgValue, DestType, 201 ArgValue->getName().data()); 202 203 Intrinsic::ID inst = (BuiltinID == Builtin::BI__builtin_va_end) ? 204 Intrinsic::vaend : Intrinsic::vastart; 205 return RValue::get(Builder.CreateCall(CGM.getIntrinsic(inst), ArgValue)); 206 } 207 case Builtin::BI__builtin_va_copy: { 208 Value *DstPtr = EmitVAListRef(E->getArg(0)); 209 Value *SrcPtr = EmitVAListRef(E->getArg(1)); 210 211 llvm::Type *Type = Int8PtrTy; 212 213 DstPtr = Builder.CreateBitCast(DstPtr, Type); 214 SrcPtr = Builder.CreateBitCast(SrcPtr, Type); 215 return RValue::get(Builder.CreateCall2(CGM.getIntrinsic(Intrinsic::vacopy), 216 DstPtr, SrcPtr)); 217 } 218 case Builtin::BI__builtin_abs: { 219 Value *ArgValue = EmitScalarExpr(E->getArg(0)); 220 221 Value *NegOp = Builder.CreateNeg(ArgValue, "neg"); 222 Value *CmpResult = 223 Builder.CreateICmpSGE(ArgValue, 224 llvm::Constant::getNullValue(ArgValue->getType()), 225 "abscond"); 226 Value *Result = 227 Builder.CreateSelect(CmpResult, ArgValue, NegOp, "abs"); 228 229 return RValue::get(Result); 230 } 231 case Builtin::BI__builtin_ctz: 232 case Builtin::BI__builtin_ctzl: 233 case Builtin::BI__builtin_ctzll: { 234 Value *ArgValue = EmitScalarExpr(E->getArg(0)); 235 236 llvm::Type *ArgType = ArgValue->getType(); 237 Value *F = CGM.getIntrinsic(Intrinsic::cttz, ArgType); 238 239 llvm::Type *ResultType = ConvertType(E->getType()); 240 Value *Result = Builder.CreateCall(F, ArgValue); 241 if (Result->getType() != ResultType) 242 Result = Builder.CreateIntCast(Result, ResultType, /*isSigned*/true, 243 "cast"); 244 return RValue::get(Result); 245 } 246 case Builtin::BI__builtin_clz: 247 case Builtin::BI__builtin_clzl: 248 case Builtin::BI__builtin_clzll: { 249 Value *ArgValue = EmitScalarExpr(E->getArg(0)); 250 251 llvm::Type *ArgType = ArgValue->getType(); 252 Value *F = CGM.getIntrinsic(Intrinsic::ctlz, ArgType); 253 254 llvm::Type *ResultType = ConvertType(E->getType()); 255 Value *Result = Builder.CreateCall(F, ArgValue); 256 if (Result->getType() != ResultType) 257 Result = Builder.CreateIntCast(Result, ResultType, /*isSigned*/true, 258 "cast"); 259 return RValue::get(Result); 260 } 261 case Builtin::BI__builtin_ffs: 262 case Builtin::BI__builtin_ffsl: 263 case Builtin::BI__builtin_ffsll: { 264 // ffs(x) -> x ? cttz(x) + 1 : 0 265 Value *ArgValue = EmitScalarExpr(E->getArg(0)); 266 267 llvm::Type *ArgType = ArgValue->getType(); 268 Value *F = CGM.getIntrinsic(Intrinsic::cttz, ArgType); 269 270 llvm::Type *ResultType = ConvertType(E->getType()); 271 Value *Tmp = Builder.CreateAdd(Builder.CreateCall(F, ArgValue), 272 llvm::ConstantInt::get(ArgType, 1)); 273 Value *Zero = llvm::Constant::getNullValue(ArgType); 274 Value *IsZero = Builder.CreateICmpEQ(ArgValue, Zero, "iszero"); 275 Value *Result = Builder.CreateSelect(IsZero, Zero, Tmp, "ffs"); 276 if (Result->getType() != ResultType) 277 Result = Builder.CreateIntCast(Result, ResultType, /*isSigned*/true, 278 "cast"); 279 return RValue::get(Result); 280 } 281 case Builtin::BI__builtin_parity: 282 case Builtin::BI__builtin_parityl: 283 case Builtin::BI__builtin_parityll: { 284 // parity(x) -> ctpop(x) & 1 285 Value *ArgValue = EmitScalarExpr(E->getArg(0)); 286 287 llvm::Type *ArgType = ArgValue->getType(); 288 Value *F = CGM.getIntrinsic(Intrinsic::ctpop, ArgType); 289 290 llvm::Type *ResultType = ConvertType(E->getType()); 291 Value *Tmp = Builder.CreateCall(F, ArgValue); 292 Value *Result = Builder.CreateAnd(Tmp, llvm::ConstantInt::get(ArgType, 1)); 293 if (Result->getType() != ResultType) 294 Result = Builder.CreateIntCast(Result, ResultType, /*isSigned*/true, 295 "cast"); 296 return RValue::get(Result); 297 } 298 case Builtin::BI__builtin_popcount: 299 case Builtin::BI__builtin_popcountl: 300 case Builtin::BI__builtin_popcountll: { 301 Value *ArgValue = EmitScalarExpr(E->getArg(0)); 302 303 llvm::Type *ArgType = ArgValue->getType(); 304 Value *F = CGM.getIntrinsic(Intrinsic::ctpop, ArgType); 305 306 llvm::Type *ResultType = ConvertType(E->getType()); 307 Value *Result = Builder.CreateCall(F, ArgValue); 308 if (Result->getType() != ResultType) 309 Result = Builder.CreateIntCast(Result, ResultType, /*isSigned*/true, 310 "cast"); 311 return RValue::get(Result); 312 } 313 case Builtin::BI__builtin_expect: { 314 Value *ArgValue = EmitScalarExpr(E->getArg(0)); 315 llvm::Type *ArgType = ArgValue->getType(); 316 317 Value *FnExpect = CGM.getIntrinsic(Intrinsic::expect, ArgType); 318 Value *ExpectedValue = EmitScalarExpr(E->getArg(1)); 319 320 Value *Result = Builder.CreateCall2(FnExpect, ArgValue, ExpectedValue, 321 "expval"); 322 return RValue::get(Result); 323 } 324 case Builtin::BI__builtin_bswap32: 325 case Builtin::BI__builtin_bswap64: { 326 Value *ArgValue = EmitScalarExpr(E->getArg(0)); 327 llvm::Type *ArgType = ArgValue->getType(); 328 Value *F = CGM.getIntrinsic(Intrinsic::bswap, ArgType); 329 return RValue::get(Builder.CreateCall(F, ArgValue)); 330 } 331 case Builtin::BI__builtin_object_size: { 332 // We pass this builtin onto the optimizer so that it can 333 // figure out the object size in more complex cases. 334 llvm::Type *ResType = ConvertType(E->getType()); 335 336 // LLVM only supports 0 and 2, make sure that we pass along that 337 // as a boolean. 338 Value *Ty = EmitScalarExpr(E->getArg(1)); 339 ConstantInt *CI = dyn_cast<ConstantInt>(Ty); 340 assert(CI); 341 uint64_t val = CI->getZExtValue(); 342 CI = ConstantInt::get(Builder.getInt1Ty(), (val & 0x2) >> 1); 343 344 Value *F = CGM.getIntrinsic(Intrinsic::objectsize, ResType); 345 return RValue::get(Builder.CreateCall2(F, 346 EmitScalarExpr(E->getArg(0)), 347 CI)); 348 } 349 case Builtin::BI__builtin_prefetch: { 350 Value *Locality, *RW, *Address = EmitScalarExpr(E->getArg(0)); 351 // FIXME: Technically these constants should of type 'int', yes? 352 RW = (E->getNumArgs() > 1) ? EmitScalarExpr(E->getArg(1)) : 353 llvm::ConstantInt::get(Int32Ty, 0); 354 Locality = (E->getNumArgs() > 2) ? EmitScalarExpr(E->getArg(2)) : 355 llvm::ConstantInt::get(Int32Ty, 3); 356 Value *Data = llvm::ConstantInt::get(Int32Ty, 1); 357 Value *F = CGM.getIntrinsic(Intrinsic::prefetch); 358 return RValue::get(Builder.CreateCall4(F, Address, RW, Locality, Data)); 359 } 360 case Builtin::BI__builtin_trap: { 361 Value *F = CGM.getIntrinsic(Intrinsic::trap); 362 return RValue::get(Builder.CreateCall(F)); 363 } 364 case Builtin::BI__builtin_unreachable: { 365 if (CatchUndefined) 366 EmitBranch(getTrapBB()); 367 else 368 Builder.CreateUnreachable(); 369 370 // We do need to preserve an insertion point. 371 EmitBlock(createBasicBlock("unreachable.cont")); 372 373 return RValue::get(0); 374 } 375 376 case Builtin::BI__builtin_powi: 377 case Builtin::BI__builtin_powif: 378 case Builtin::BI__builtin_powil: { 379 Value *Base = EmitScalarExpr(E->getArg(0)); 380 Value *Exponent = EmitScalarExpr(E->getArg(1)); 381 llvm::Type *ArgType = Base->getType(); 382 Value *F = CGM.getIntrinsic(Intrinsic::powi, ArgType); 383 return RValue::get(Builder.CreateCall2(F, Base, Exponent)); 384 } 385 386 case Builtin::BI__builtin_isgreater: 387 case Builtin::BI__builtin_isgreaterequal: 388 case Builtin::BI__builtin_isless: 389 case Builtin::BI__builtin_islessequal: 390 case Builtin::BI__builtin_islessgreater: 391 case Builtin::BI__builtin_isunordered: { 392 // Ordered comparisons: we know the arguments to these are matching scalar 393 // floating point values. 394 Value *LHS = EmitScalarExpr(E->getArg(0)); 395 Value *RHS = EmitScalarExpr(E->getArg(1)); 396 397 switch (BuiltinID) { 398 default: llvm_unreachable("Unknown ordered comparison"); 399 case Builtin::BI__builtin_isgreater: 400 LHS = Builder.CreateFCmpOGT(LHS, RHS, "cmp"); 401 break; 402 case Builtin::BI__builtin_isgreaterequal: 403 LHS = Builder.CreateFCmpOGE(LHS, RHS, "cmp"); 404 break; 405 case Builtin::BI__builtin_isless: 406 LHS = Builder.CreateFCmpOLT(LHS, RHS, "cmp"); 407 break; 408 case Builtin::BI__builtin_islessequal: 409 LHS = Builder.CreateFCmpOLE(LHS, RHS, "cmp"); 410 break; 411 case Builtin::BI__builtin_islessgreater: 412 LHS = Builder.CreateFCmpONE(LHS, RHS, "cmp"); 413 break; 414 case Builtin::BI__builtin_isunordered: 415 LHS = Builder.CreateFCmpUNO(LHS, RHS, "cmp"); 416 break; 417 } 418 // ZExt bool to int type. 419 return RValue::get(Builder.CreateZExt(LHS, ConvertType(E->getType()))); 420 } 421 case Builtin::BI__builtin_isnan: { 422 Value *V = EmitScalarExpr(E->getArg(0)); 423 V = Builder.CreateFCmpUNO(V, V, "cmp"); 424 return RValue::get(Builder.CreateZExt(V, ConvertType(E->getType()))); 425 } 426 427 case Builtin::BI__builtin_isinf: { 428 // isinf(x) --> fabs(x) == infinity 429 Value *V = EmitScalarExpr(E->getArg(0)); 430 V = EmitFAbs(*this, V, E->getArg(0)->getType()); 431 432 V = Builder.CreateFCmpOEQ(V, ConstantFP::getInfinity(V->getType()),"isinf"); 433 return RValue::get(Builder.CreateZExt(V, ConvertType(E->getType()))); 434 } 435 436 // TODO: BI__builtin_isinf_sign 437 // isinf_sign(x) -> isinf(x) ? (signbit(x) ? -1 : 1) : 0 438 439 case Builtin::BI__builtin_isnormal: { 440 // isnormal(x) --> x == x && fabsf(x) < infinity && fabsf(x) >= float_min 441 Value *V = EmitScalarExpr(E->getArg(0)); 442 Value *Eq = Builder.CreateFCmpOEQ(V, V, "iseq"); 443 444 Value *Abs = EmitFAbs(*this, V, E->getArg(0)->getType()); 445 Value *IsLessThanInf = 446 Builder.CreateFCmpULT(Abs, ConstantFP::getInfinity(V->getType()),"isinf"); 447 APFloat Smallest = APFloat::getSmallestNormalized( 448 getContext().getFloatTypeSemantics(E->getArg(0)->getType())); 449 Value *IsNormal = 450 Builder.CreateFCmpUGE(Abs, ConstantFP::get(V->getContext(), Smallest), 451 "isnormal"); 452 V = Builder.CreateAnd(Eq, IsLessThanInf, "and"); 453 V = Builder.CreateAnd(V, IsNormal, "and"); 454 return RValue::get(Builder.CreateZExt(V, ConvertType(E->getType()))); 455 } 456 457 case Builtin::BI__builtin_isfinite: { 458 // isfinite(x) --> x == x && fabs(x) != infinity; 459 Value *V = EmitScalarExpr(E->getArg(0)); 460 Value *Eq = Builder.CreateFCmpOEQ(V, V, "iseq"); 461 462 Value *Abs = EmitFAbs(*this, V, E->getArg(0)->getType()); 463 Value *IsNotInf = 464 Builder.CreateFCmpUNE(Abs, ConstantFP::getInfinity(V->getType()),"isinf"); 465 466 V = Builder.CreateAnd(Eq, IsNotInf, "and"); 467 return RValue::get(Builder.CreateZExt(V, ConvertType(E->getType()))); 468 } 469 470 case Builtin::BI__builtin_fpclassify: { 471 Value *V = EmitScalarExpr(E->getArg(5)); 472 llvm::Type *Ty = ConvertType(E->getArg(5)->getType()); 473 474 // Create Result 475 BasicBlock *Begin = Builder.GetInsertBlock(); 476 BasicBlock *End = createBasicBlock("fpclassify_end", this->CurFn); 477 Builder.SetInsertPoint(End); 478 PHINode *Result = 479 Builder.CreatePHI(ConvertType(E->getArg(0)->getType()), 4, 480 "fpclassify_result"); 481 482 // if (V==0) return FP_ZERO 483 Builder.SetInsertPoint(Begin); 484 Value *IsZero = Builder.CreateFCmpOEQ(V, Constant::getNullValue(Ty), 485 "iszero"); 486 Value *ZeroLiteral = EmitScalarExpr(E->getArg(4)); 487 BasicBlock *NotZero = createBasicBlock("fpclassify_not_zero", this->CurFn); 488 Builder.CreateCondBr(IsZero, End, NotZero); 489 Result->addIncoming(ZeroLiteral, Begin); 490 491 // if (V != V) return FP_NAN 492 Builder.SetInsertPoint(NotZero); 493 Value *IsNan = Builder.CreateFCmpUNO(V, V, "cmp"); 494 Value *NanLiteral = EmitScalarExpr(E->getArg(0)); 495 BasicBlock *NotNan = createBasicBlock("fpclassify_not_nan", this->CurFn); 496 Builder.CreateCondBr(IsNan, End, NotNan); 497 Result->addIncoming(NanLiteral, NotZero); 498 499 // if (fabs(V) == infinity) return FP_INFINITY 500 Builder.SetInsertPoint(NotNan); 501 Value *VAbs = EmitFAbs(*this, V, E->getArg(5)->getType()); 502 Value *IsInf = 503 Builder.CreateFCmpOEQ(VAbs, ConstantFP::getInfinity(V->getType()), 504 "isinf"); 505 Value *InfLiteral = EmitScalarExpr(E->getArg(1)); 506 BasicBlock *NotInf = createBasicBlock("fpclassify_not_inf", this->CurFn); 507 Builder.CreateCondBr(IsInf, End, NotInf); 508 Result->addIncoming(InfLiteral, NotNan); 509 510 // if (fabs(V) >= MIN_NORMAL) return FP_NORMAL else FP_SUBNORMAL 511 Builder.SetInsertPoint(NotInf); 512 APFloat Smallest = APFloat::getSmallestNormalized( 513 getContext().getFloatTypeSemantics(E->getArg(5)->getType())); 514 Value *IsNormal = 515 Builder.CreateFCmpUGE(VAbs, ConstantFP::get(V->getContext(), Smallest), 516 "isnormal"); 517 Value *NormalResult = 518 Builder.CreateSelect(IsNormal, EmitScalarExpr(E->getArg(2)), 519 EmitScalarExpr(E->getArg(3))); 520 Builder.CreateBr(End); 521 Result->addIncoming(NormalResult, NotInf); 522 523 // return Result 524 Builder.SetInsertPoint(End); 525 return RValue::get(Result); 526 } 527 528 case Builtin::BIalloca: 529 case Builtin::BI__builtin_alloca: { 530 Value *Size = EmitScalarExpr(E->getArg(0)); 531 return RValue::get(Builder.CreateAlloca(Builder.getInt8Ty(), Size)); 532 } 533 case Builtin::BIbzero: 534 case Builtin::BI__builtin_bzero: { 535 Value *Address = EmitScalarExpr(E->getArg(0)); 536 Value *SizeVal = EmitScalarExpr(E->getArg(1)); 537 Builder.CreateMemSet(Address, Builder.getInt8(0), SizeVal, 1, false); 538 return RValue::get(Address); 539 } 540 case Builtin::BImemcpy: 541 case Builtin::BI__builtin_memcpy: { 542 Value *Address = EmitScalarExpr(E->getArg(0)); 543 Value *SrcAddr = EmitScalarExpr(E->getArg(1)); 544 Value *SizeVal = EmitScalarExpr(E->getArg(2)); 545 Builder.CreateMemCpy(Address, SrcAddr, SizeVal, 1, false); 546 return RValue::get(Address); 547 } 548 549 case Builtin::BI__builtin___memcpy_chk: { 550 // fold __builtin_memcpy_chk(x, y, cst1, cst2) to memset iff cst1<=cst2. 551 if (!E->getArg(2)->isEvaluatable(CGM.getContext()) || 552 !E->getArg(3)->isEvaluatable(CGM.getContext())) 553 break; 554 llvm::APSInt Size = E->getArg(2)->EvaluateAsInt(CGM.getContext()); 555 llvm::APSInt DstSize = E->getArg(3)->EvaluateAsInt(CGM.getContext()); 556 if (Size.ugt(DstSize)) 557 break; 558 Value *Dest = EmitScalarExpr(E->getArg(0)); 559 Value *Src = EmitScalarExpr(E->getArg(1)); 560 Value *SizeVal = llvm::ConstantInt::get(Builder.getContext(), Size); 561 Builder.CreateMemCpy(Dest, Src, SizeVal, 1, false); 562 return RValue::get(Dest); 563 } 564 565 case Builtin::BI__builtin_objc_memmove_collectable: { 566 Value *Address = EmitScalarExpr(E->getArg(0)); 567 Value *SrcAddr = EmitScalarExpr(E->getArg(1)); 568 Value *SizeVal = EmitScalarExpr(E->getArg(2)); 569 CGM.getObjCRuntime().EmitGCMemmoveCollectable(*this, 570 Address, SrcAddr, SizeVal); 571 return RValue::get(Address); 572 } 573 574 case Builtin::BI__builtin___memmove_chk: { 575 // fold __builtin_memmove_chk(x, y, cst1, cst2) to memset iff cst1<=cst2. 576 if (!E->getArg(2)->isEvaluatable(CGM.getContext()) || 577 !E->getArg(3)->isEvaluatable(CGM.getContext())) 578 break; 579 llvm::APSInt Size = E->getArg(2)->EvaluateAsInt(CGM.getContext()); 580 llvm::APSInt DstSize = E->getArg(3)->EvaluateAsInt(CGM.getContext()); 581 if (Size.ugt(DstSize)) 582 break; 583 Value *Dest = EmitScalarExpr(E->getArg(0)); 584 Value *Src = EmitScalarExpr(E->getArg(1)); 585 Value *SizeVal = llvm::ConstantInt::get(Builder.getContext(), Size); 586 Builder.CreateMemMove(Dest, Src, SizeVal, 1, false); 587 return RValue::get(Dest); 588 } 589 590 case Builtin::BImemmove: 591 case Builtin::BI__builtin_memmove: { 592 Value *Address = EmitScalarExpr(E->getArg(0)); 593 Value *SrcAddr = EmitScalarExpr(E->getArg(1)); 594 Value *SizeVal = EmitScalarExpr(E->getArg(2)); 595 Builder.CreateMemMove(Address, SrcAddr, SizeVal, 1, false); 596 return RValue::get(Address); 597 } 598 case Builtin::BImemset: 599 case Builtin::BI__builtin_memset: { 600 Value *Address = EmitScalarExpr(E->getArg(0)); 601 Value *ByteVal = Builder.CreateTrunc(EmitScalarExpr(E->getArg(1)), 602 Builder.getInt8Ty()); 603 Value *SizeVal = EmitScalarExpr(E->getArg(2)); 604 Builder.CreateMemSet(Address, ByteVal, SizeVal, 1, false); 605 return RValue::get(Address); 606 } 607 case Builtin::BI__builtin___memset_chk: { 608 // fold __builtin_memset_chk(x, y, cst1, cst2) to memset iff cst1<=cst2. 609 if (!E->getArg(2)->isEvaluatable(CGM.getContext()) || 610 !E->getArg(3)->isEvaluatable(CGM.getContext())) 611 break; 612 llvm::APSInt Size = E->getArg(2)->EvaluateAsInt(CGM.getContext()); 613 llvm::APSInt DstSize = E->getArg(3)->EvaluateAsInt(CGM.getContext()); 614 if (Size.ugt(DstSize)) 615 break; 616 Value *Address = EmitScalarExpr(E->getArg(0)); 617 Value *ByteVal = Builder.CreateTrunc(EmitScalarExpr(E->getArg(1)), 618 Builder.getInt8Ty()); 619 Value *SizeVal = llvm::ConstantInt::get(Builder.getContext(), Size); 620 Builder.CreateMemSet(Address, ByteVal, SizeVal, 1, false); 621 622 return RValue::get(Address); 623 } 624 case Builtin::BI__builtin_dwarf_cfa: { 625 // The offset in bytes from the first argument to the CFA. 626 // 627 // Why on earth is this in the frontend? Is there any reason at 628 // all that the backend can't reasonably determine this while 629 // lowering llvm.eh.dwarf.cfa()? 630 // 631 // TODO: If there's a satisfactory reason, add a target hook for 632 // this instead of hard-coding 0, which is correct for most targets. 633 int32_t Offset = 0; 634 635 Value *F = CGM.getIntrinsic(Intrinsic::eh_dwarf_cfa); 636 return RValue::get(Builder.CreateCall(F, 637 llvm::ConstantInt::get(Int32Ty, Offset))); 638 } 639 case Builtin::BI__builtin_return_address: { 640 Value *Depth = EmitScalarExpr(E->getArg(0)); 641 Depth = Builder.CreateIntCast(Depth, Int32Ty, false); 642 Value *F = CGM.getIntrinsic(Intrinsic::returnaddress); 643 return RValue::get(Builder.CreateCall(F, Depth)); 644 } 645 case Builtin::BI__builtin_frame_address: { 646 Value *Depth = EmitScalarExpr(E->getArg(0)); 647 Depth = Builder.CreateIntCast(Depth, Int32Ty, false); 648 Value *F = CGM.getIntrinsic(Intrinsic::frameaddress); 649 return RValue::get(Builder.CreateCall(F, Depth)); 650 } 651 case Builtin::BI__builtin_extract_return_addr: { 652 Value *Address = EmitScalarExpr(E->getArg(0)); 653 Value *Result = getTargetHooks().decodeReturnAddress(*this, Address); 654 return RValue::get(Result); 655 } 656 case Builtin::BI__builtin_frob_return_addr: { 657 Value *Address = EmitScalarExpr(E->getArg(0)); 658 Value *Result = getTargetHooks().encodeReturnAddress(*this, Address); 659 return RValue::get(Result); 660 } 661 case Builtin::BI__builtin_dwarf_sp_column: { 662 llvm::IntegerType *Ty 663 = cast<llvm::IntegerType>(ConvertType(E->getType())); 664 int Column = getTargetHooks().getDwarfEHStackPointer(CGM); 665 if (Column == -1) { 666 CGM.ErrorUnsupported(E, "__builtin_dwarf_sp_column"); 667 return RValue::get(llvm::UndefValue::get(Ty)); 668 } 669 return RValue::get(llvm::ConstantInt::get(Ty, Column, true)); 670 } 671 case Builtin::BI__builtin_init_dwarf_reg_size_table: { 672 Value *Address = EmitScalarExpr(E->getArg(0)); 673 if (getTargetHooks().initDwarfEHRegSizeTable(*this, Address)) 674 CGM.ErrorUnsupported(E, "__builtin_init_dwarf_reg_size_table"); 675 return RValue::get(llvm::UndefValue::get(ConvertType(E->getType()))); 676 } 677 case Builtin::BI__builtin_eh_return: { 678 Value *Int = EmitScalarExpr(E->getArg(0)); 679 Value *Ptr = EmitScalarExpr(E->getArg(1)); 680 681 llvm::IntegerType *IntTy = cast<llvm::IntegerType>(Int->getType()); 682 assert((IntTy->getBitWidth() == 32 || IntTy->getBitWidth() == 64) && 683 "LLVM's __builtin_eh_return only supports 32- and 64-bit variants"); 684 Value *F = CGM.getIntrinsic(IntTy->getBitWidth() == 32 685 ? Intrinsic::eh_return_i32 686 : Intrinsic::eh_return_i64); 687 Builder.CreateCall2(F, Int, Ptr); 688 Builder.CreateUnreachable(); 689 690 // We do need to preserve an insertion point. 691 EmitBlock(createBasicBlock("builtin_eh_return.cont")); 692 693 return RValue::get(0); 694 } 695 case Builtin::BI__builtin_unwind_init: { 696 Value *F = CGM.getIntrinsic(Intrinsic::eh_unwind_init); 697 return RValue::get(Builder.CreateCall(F)); 698 } 699 case Builtin::BI__builtin_extend_pointer: { 700 // Extends a pointer to the size of an _Unwind_Word, which is 701 // uint64_t on all platforms. Generally this gets poked into a 702 // register and eventually used as an address, so if the 703 // addressing registers are wider than pointers and the platform 704 // doesn't implicitly ignore high-order bits when doing 705 // addressing, we need to make sure we zext / sext based on 706 // the platform's expectations. 707 // 708 // See: http://gcc.gnu.org/ml/gcc-bugs/2002-02/msg00237.html 709 710 // Cast the pointer to intptr_t. 711 Value *Ptr = EmitScalarExpr(E->getArg(0)); 712 Value *Result = Builder.CreatePtrToInt(Ptr, IntPtrTy, "extend.cast"); 713 714 // If that's 64 bits, we're done. 715 if (IntPtrTy->getBitWidth() == 64) 716 return RValue::get(Result); 717 718 // Otherwise, ask the codegen data what to do. 719 if (getTargetHooks().extendPointerWithSExt()) 720 return RValue::get(Builder.CreateSExt(Result, Int64Ty, "extend.sext")); 721 else 722 return RValue::get(Builder.CreateZExt(Result, Int64Ty, "extend.zext")); 723 } 724 case Builtin::BI__builtin_setjmp: { 725 // Buffer is a void**. 726 Value *Buf = EmitScalarExpr(E->getArg(0)); 727 728 // Store the frame pointer to the setjmp buffer. 729 Value *FrameAddr = 730 Builder.CreateCall(CGM.getIntrinsic(Intrinsic::frameaddress), 731 ConstantInt::get(Int32Ty, 0)); 732 Builder.CreateStore(FrameAddr, Buf); 733 734 // Store the stack pointer to the setjmp buffer. 735 Value *StackAddr = 736 Builder.CreateCall(CGM.getIntrinsic(Intrinsic::stacksave)); 737 Value *StackSaveSlot = 738 Builder.CreateGEP(Buf, ConstantInt::get(Int32Ty, 2)); 739 Builder.CreateStore(StackAddr, StackSaveSlot); 740 741 // Call LLVM's EH setjmp, which is lightweight. 742 Value *F = CGM.getIntrinsic(Intrinsic::eh_sjlj_setjmp); 743 Buf = Builder.CreateBitCast(Buf, Int8PtrTy); 744 return RValue::get(Builder.CreateCall(F, Buf)); 745 } 746 case Builtin::BI__builtin_longjmp: { 747 Value *Buf = EmitScalarExpr(E->getArg(0)); 748 Buf = Builder.CreateBitCast(Buf, Int8PtrTy); 749 750 // Call LLVM's EH longjmp, which is lightweight. 751 Builder.CreateCall(CGM.getIntrinsic(Intrinsic::eh_sjlj_longjmp), Buf); 752 753 // longjmp doesn't return; mark this as unreachable. 754 Builder.CreateUnreachable(); 755 756 // We do need to preserve an insertion point. 757 EmitBlock(createBasicBlock("longjmp.cont")); 758 759 return RValue::get(0); 760 } 761 case Builtin::BI__sync_fetch_and_add: 762 case Builtin::BI__sync_fetch_and_sub: 763 case Builtin::BI__sync_fetch_and_or: 764 case Builtin::BI__sync_fetch_and_and: 765 case Builtin::BI__sync_fetch_and_xor: 766 case Builtin::BI__sync_add_and_fetch: 767 case Builtin::BI__sync_sub_and_fetch: 768 case Builtin::BI__sync_and_and_fetch: 769 case Builtin::BI__sync_or_and_fetch: 770 case Builtin::BI__sync_xor_and_fetch: 771 case Builtin::BI__sync_val_compare_and_swap: 772 case Builtin::BI__sync_bool_compare_and_swap: 773 case Builtin::BI__sync_lock_test_and_set: 774 case Builtin::BI__sync_lock_release: 775 case Builtin::BI__sync_swap: 776 llvm_unreachable("Shouldn't make it through sema"); 777 case Builtin::BI__sync_fetch_and_add_1: 778 case Builtin::BI__sync_fetch_and_add_2: 779 case Builtin::BI__sync_fetch_and_add_4: 780 case Builtin::BI__sync_fetch_and_add_8: 781 case Builtin::BI__sync_fetch_and_add_16: 782 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Add, E); 783 case Builtin::BI__sync_fetch_and_sub_1: 784 case Builtin::BI__sync_fetch_and_sub_2: 785 case Builtin::BI__sync_fetch_and_sub_4: 786 case Builtin::BI__sync_fetch_and_sub_8: 787 case Builtin::BI__sync_fetch_and_sub_16: 788 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Sub, E); 789 case Builtin::BI__sync_fetch_and_or_1: 790 case Builtin::BI__sync_fetch_and_or_2: 791 case Builtin::BI__sync_fetch_and_or_4: 792 case Builtin::BI__sync_fetch_and_or_8: 793 case Builtin::BI__sync_fetch_and_or_16: 794 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Or, E); 795 case Builtin::BI__sync_fetch_and_and_1: 796 case Builtin::BI__sync_fetch_and_and_2: 797 case Builtin::BI__sync_fetch_and_and_4: 798 case Builtin::BI__sync_fetch_and_and_8: 799 case Builtin::BI__sync_fetch_and_and_16: 800 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::And, E); 801 case Builtin::BI__sync_fetch_and_xor_1: 802 case Builtin::BI__sync_fetch_and_xor_2: 803 case Builtin::BI__sync_fetch_and_xor_4: 804 case Builtin::BI__sync_fetch_and_xor_8: 805 case Builtin::BI__sync_fetch_and_xor_16: 806 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Xor, E); 807 808 // Clang extensions: not overloaded yet. 809 case Builtin::BI__sync_fetch_and_min: 810 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Min, E); 811 case Builtin::BI__sync_fetch_and_max: 812 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Max, E); 813 case Builtin::BI__sync_fetch_and_umin: 814 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::UMin, E); 815 case Builtin::BI__sync_fetch_and_umax: 816 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::UMax, E); 817 818 case Builtin::BI__sync_add_and_fetch_1: 819 case Builtin::BI__sync_add_and_fetch_2: 820 case Builtin::BI__sync_add_and_fetch_4: 821 case Builtin::BI__sync_add_and_fetch_8: 822 case Builtin::BI__sync_add_and_fetch_16: 823 return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::Add, E, 824 llvm::Instruction::Add); 825 case Builtin::BI__sync_sub_and_fetch_1: 826 case Builtin::BI__sync_sub_and_fetch_2: 827 case Builtin::BI__sync_sub_and_fetch_4: 828 case Builtin::BI__sync_sub_and_fetch_8: 829 case Builtin::BI__sync_sub_and_fetch_16: 830 return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::Sub, E, 831 llvm::Instruction::Sub); 832 case Builtin::BI__sync_and_and_fetch_1: 833 case Builtin::BI__sync_and_and_fetch_2: 834 case Builtin::BI__sync_and_and_fetch_4: 835 case Builtin::BI__sync_and_and_fetch_8: 836 case Builtin::BI__sync_and_and_fetch_16: 837 return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::And, E, 838 llvm::Instruction::And); 839 case Builtin::BI__sync_or_and_fetch_1: 840 case Builtin::BI__sync_or_and_fetch_2: 841 case Builtin::BI__sync_or_and_fetch_4: 842 case Builtin::BI__sync_or_and_fetch_8: 843 case Builtin::BI__sync_or_and_fetch_16: 844 return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::Or, E, 845 llvm::Instruction::Or); 846 case Builtin::BI__sync_xor_and_fetch_1: 847 case Builtin::BI__sync_xor_and_fetch_2: 848 case Builtin::BI__sync_xor_and_fetch_4: 849 case Builtin::BI__sync_xor_and_fetch_8: 850 case Builtin::BI__sync_xor_and_fetch_16: 851 return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::Xor, E, 852 llvm::Instruction::Xor); 853 854 case Builtin::BI__sync_val_compare_and_swap_1: 855 case Builtin::BI__sync_val_compare_and_swap_2: 856 case Builtin::BI__sync_val_compare_and_swap_4: 857 case Builtin::BI__sync_val_compare_and_swap_8: 858 case Builtin::BI__sync_val_compare_and_swap_16: { 859 QualType T = E->getType(); 860 llvm::Value *DestPtr = EmitScalarExpr(E->getArg(0)); 861 unsigned AddrSpace = 862 cast<llvm::PointerType>(DestPtr->getType())->getAddressSpace(); 863 864 llvm::IntegerType *IntType = 865 llvm::IntegerType::get(getLLVMContext(), 866 getContext().getTypeSize(T)); 867 llvm::Type *IntPtrType = IntType->getPointerTo(AddrSpace); 868 869 Value *Args[3]; 870 Args[0] = Builder.CreateBitCast(DestPtr, IntPtrType); 871 Args[1] = EmitScalarExpr(E->getArg(1)); 872 llvm::Type *ValueType = Args[1]->getType(); 873 Args[1] = EmitToInt(*this, Args[1], T, IntType); 874 Args[2] = EmitToInt(*this, EmitScalarExpr(E->getArg(2)), T, IntType); 875 876 Value *Result = Builder.CreateAtomicCmpXchg(Args[0], Args[1], Args[2], 877 llvm::SequentiallyConsistent); 878 Result = EmitFromInt(*this, Result, T, ValueType); 879 return RValue::get(Result); 880 } 881 882 case Builtin::BI__sync_bool_compare_and_swap_1: 883 case Builtin::BI__sync_bool_compare_and_swap_2: 884 case Builtin::BI__sync_bool_compare_and_swap_4: 885 case Builtin::BI__sync_bool_compare_and_swap_8: 886 case Builtin::BI__sync_bool_compare_and_swap_16: { 887 QualType T = E->getArg(1)->getType(); 888 llvm::Value *DestPtr = EmitScalarExpr(E->getArg(0)); 889 unsigned AddrSpace = 890 cast<llvm::PointerType>(DestPtr->getType())->getAddressSpace(); 891 892 llvm::IntegerType *IntType = 893 llvm::IntegerType::get(getLLVMContext(), 894 getContext().getTypeSize(T)); 895 llvm::Type *IntPtrType = IntType->getPointerTo(AddrSpace); 896 897 Value *Args[3]; 898 Args[0] = Builder.CreateBitCast(DestPtr, IntPtrType); 899 Args[1] = EmitToInt(*this, EmitScalarExpr(E->getArg(1)), T, IntType); 900 Args[2] = EmitToInt(*this, EmitScalarExpr(E->getArg(2)), T, IntType); 901 902 Value *OldVal = Args[1]; 903 Value *PrevVal = Builder.CreateAtomicCmpXchg(Args[0], Args[1], Args[2], 904 llvm::SequentiallyConsistent); 905 Value *Result = Builder.CreateICmpEQ(PrevVal, OldVal); 906 // zext bool to int. 907 Result = Builder.CreateZExt(Result, ConvertType(E->getType())); 908 return RValue::get(Result); 909 } 910 911 case Builtin::BI__sync_swap_1: 912 case Builtin::BI__sync_swap_2: 913 case Builtin::BI__sync_swap_4: 914 case Builtin::BI__sync_swap_8: 915 case Builtin::BI__sync_swap_16: 916 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Xchg, E); 917 918 case Builtin::BI__sync_lock_test_and_set_1: 919 case Builtin::BI__sync_lock_test_and_set_2: 920 case Builtin::BI__sync_lock_test_and_set_4: 921 case Builtin::BI__sync_lock_test_and_set_8: 922 case Builtin::BI__sync_lock_test_and_set_16: 923 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Xchg, E); 924 925 case Builtin::BI__sync_lock_release_1: 926 case Builtin::BI__sync_lock_release_2: 927 case Builtin::BI__sync_lock_release_4: 928 case Builtin::BI__sync_lock_release_8: 929 case Builtin::BI__sync_lock_release_16: { 930 Value *Ptr = EmitScalarExpr(E->getArg(0)); 931 llvm::Type *ElLLVMTy = 932 cast<llvm::PointerType>(Ptr->getType())->getElementType(); 933 llvm::StoreInst *Store = 934 Builder.CreateStore(llvm::Constant::getNullValue(ElLLVMTy), Ptr); 935 QualType ElTy = E->getArg(0)->getType()->getPointeeType(); 936 CharUnits StoreSize = getContext().getTypeSizeInChars(ElTy); 937 Store->setAlignment(StoreSize.getQuantity()); 938 Store->setAtomic(llvm::Release); 939 return RValue::get(0); 940 } 941 942 case Builtin::BI__sync_synchronize: { 943 // We assume this is supposed to correspond to a C++0x-style 944 // sequentially-consistent fence (i.e. this is only usable for 945 // synchonization, not device I/O or anything like that). This intrinsic 946 // is really badly designed in the sense that in theory, there isn't 947 // any way to safely use it... but in practice, it mostly works 948 // to use it with non-atomic loads and stores to get acquire/release 949 // semantics. 950 Builder.CreateFence(llvm::SequentiallyConsistent); 951 return RValue::get(0); 952 } 953 954 case Builtin::BI__builtin_llvm_memory_barrier: { 955 Value *C[5] = { 956 EmitScalarExpr(E->getArg(0)), 957 EmitScalarExpr(E->getArg(1)), 958 EmitScalarExpr(E->getArg(2)), 959 EmitScalarExpr(E->getArg(3)), 960 EmitScalarExpr(E->getArg(4)) 961 }; 962 Builder.CreateCall(CGM.getIntrinsic(Intrinsic::memory_barrier), C); 963 return RValue::get(0); 964 } 965 966 // Library functions with special handling. 967 case Builtin::BIsqrt: 968 case Builtin::BIsqrtf: 969 case Builtin::BIsqrtl: { 970 // TODO: there is currently no set of optimizer flags 971 // sufficient for us to rewrite sqrt to @llvm.sqrt. 972 // -fmath-errno=0 is not good enough; we need finiteness. 973 // We could probably precondition the call with an ult 974 // against 0, but is that worth the complexity? 975 break; 976 } 977 978 case Builtin::BIpow: 979 case Builtin::BIpowf: 980 case Builtin::BIpowl: { 981 // Rewrite sqrt to intrinsic if allowed. 982 if (!FD->hasAttr<ConstAttr>()) 983 break; 984 Value *Base = EmitScalarExpr(E->getArg(0)); 985 Value *Exponent = EmitScalarExpr(E->getArg(1)); 986 llvm::Type *ArgType = Base->getType(); 987 Value *F = CGM.getIntrinsic(Intrinsic::pow, ArgType); 988 return RValue::get(Builder.CreateCall2(F, Base, Exponent)); 989 } 990 991 case Builtin::BIfma: 992 case Builtin::BIfmaf: 993 case Builtin::BIfmal: 994 case Builtin::BI__builtin_fma: 995 case Builtin::BI__builtin_fmaf: 996 case Builtin::BI__builtin_fmal: { 997 // Rewrite fma to intrinsic. 998 Value *FirstArg = EmitScalarExpr(E->getArg(0)); 999 llvm::Type *ArgType = FirstArg->getType(); 1000 Value *F = CGM.getIntrinsic(Intrinsic::fma, ArgType); 1001 return RValue::get(Builder.CreateCall3(F, FirstArg, 1002 EmitScalarExpr(E->getArg(1)), 1003 EmitScalarExpr(E->getArg(2)))); 1004 } 1005 1006 case Builtin::BI__builtin_signbit: 1007 case Builtin::BI__builtin_signbitf: 1008 case Builtin::BI__builtin_signbitl: { 1009 LLVMContext &C = CGM.getLLVMContext(); 1010 1011 Value *Arg = EmitScalarExpr(E->getArg(0)); 1012 llvm::Type *ArgTy = Arg->getType(); 1013 if (ArgTy->isPPC_FP128Ty()) 1014 break; // FIXME: I'm not sure what the right implementation is here. 1015 int ArgWidth = ArgTy->getPrimitiveSizeInBits(); 1016 llvm::Type *ArgIntTy = llvm::IntegerType::get(C, ArgWidth); 1017 Value *BCArg = Builder.CreateBitCast(Arg, ArgIntTy); 1018 Value *ZeroCmp = llvm::Constant::getNullValue(ArgIntTy); 1019 Value *Result = Builder.CreateICmpSLT(BCArg, ZeroCmp); 1020 return RValue::get(Builder.CreateZExt(Result, ConvertType(E->getType()))); 1021 } 1022 case Builtin::BI__builtin_annotation: { 1023 llvm::Value *AnnVal = EmitScalarExpr(E->getArg(0)); 1024 llvm::Value *F = CGM.getIntrinsic(llvm::Intrinsic::annotation, 1025 AnnVal->getType()); 1026 1027 // Get the annotation string, go through casts. Sema requires this to be a 1028 // non-wide string literal, potentially casted, so the cast<> is safe. 1029 const Expr *AnnotationStrExpr = E->getArg(1)->IgnoreParenCasts(); 1030 llvm::StringRef Str = cast<StringLiteral>(AnnotationStrExpr)->getString(); 1031 return RValue::get(EmitAnnotationCall(F, AnnVal, Str, E->getExprLoc())); 1032 } 1033 } 1034 1035 // If this is an alias for a lib function (e.g. __builtin_sin), emit 1036 // the call using the normal call path, but using the unmangled 1037 // version of the function name. 1038 if (getContext().BuiltinInfo.isLibFunction(BuiltinID)) 1039 return emitLibraryCall(*this, FD, E, 1040 CGM.getBuiltinLibFunction(FD, BuiltinID)); 1041 1042 // If this is a predefined lib function (e.g. malloc), emit the call 1043 // using exactly the normal call path. 1044 if (getContext().BuiltinInfo.isPredefinedLibFunction(BuiltinID)) 1045 return emitLibraryCall(*this, FD, E, EmitScalarExpr(E->getCallee())); 1046 1047 // See if we have a target specific intrinsic. 1048 const char *Name = getContext().BuiltinInfo.GetName(BuiltinID); 1049 Intrinsic::ID IntrinsicID = Intrinsic::not_intrinsic; 1050 if (const char *Prefix = 1051 llvm::Triple::getArchTypePrefix(Target.getTriple().getArch())) 1052 IntrinsicID = Intrinsic::getIntrinsicForGCCBuiltin(Prefix, Name); 1053 1054 if (IntrinsicID != Intrinsic::not_intrinsic) { 1055 SmallVector<Value*, 16> Args; 1056 1057 // Find out if any arguments are required to be integer constant 1058 // expressions. 1059 unsigned ICEArguments = 0; 1060 ASTContext::GetBuiltinTypeError Error; 1061 getContext().GetBuiltinType(BuiltinID, Error, &ICEArguments); 1062 assert(Error == ASTContext::GE_None && "Should not codegen an error"); 1063 1064 Function *F = CGM.getIntrinsic(IntrinsicID); 1065 llvm::FunctionType *FTy = F->getFunctionType(); 1066 1067 for (unsigned i = 0, e = E->getNumArgs(); i != e; ++i) { 1068 Value *ArgValue; 1069 // If this is a normal argument, just emit it as a scalar. 1070 if ((ICEArguments & (1 << i)) == 0) { 1071 ArgValue = EmitScalarExpr(E->getArg(i)); 1072 } else { 1073 // If this is required to be a constant, constant fold it so that we 1074 // know that the generated intrinsic gets a ConstantInt. 1075 llvm::APSInt Result; 1076 bool IsConst = E->getArg(i)->isIntegerConstantExpr(Result,getContext()); 1077 assert(IsConst && "Constant arg isn't actually constant?"); 1078 (void)IsConst; 1079 ArgValue = llvm::ConstantInt::get(getLLVMContext(), Result); 1080 } 1081 1082 // If the intrinsic arg type is different from the builtin arg type 1083 // we need to do a bit cast. 1084 llvm::Type *PTy = FTy->getParamType(i); 1085 if (PTy != ArgValue->getType()) { 1086 assert(PTy->canLosslesslyBitCastTo(FTy->getParamType(i)) && 1087 "Must be able to losslessly bit cast to param"); 1088 ArgValue = Builder.CreateBitCast(ArgValue, PTy); 1089 } 1090 1091 Args.push_back(ArgValue); 1092 } 1093 1094 Value *V = Builder.CreateCall(F, Args); 1095 QualType BuiltinRetType = E->getType(); 1096 1097 llvm::Type *RetTy = llvm::Type::getVoidTy(getLLVMContext()); 1098 if (!BuiltinRetType->isVoidType()) RetTy = ConvertType(BuiltinRetType); 1099 1100 if (RetTy != V->getType()) { 1101 assert(V->getType()->canLosslesslyBitCastTo(RetTy) && 1102 "Must be able to losslessly bit cast result type"); 1103 V = Builder.CreateBitCast(V, RetTy); 1104 } 1105 1106 return RValue::get(V); 1107 } 1108 1109 // See if we have a target specific builtin that needs to be lowered. 1110 if (Value *V = EmitTargetBuiltinExpr(BuiltinID, E)) 1111 return RValue::get(V); 1112 1113 ErrorUnsupported(E, "builtin function"); 1114 1115 // Unknown builtin, for now just dump it out and return undef. 1116 if (hasAggregateLLVMType(E->getType())) 1117 return RValue::getAggregate(CreateMemTemp(E->getType())); 1118 return RValue::get(llvm::UndefValue::get(ConvertType(E->getType()))); 1119 } 1120 1121 Value *CodeGenFunction::EmitTargetBuiltinExpr(unsigned BuiltinID, 1122 const CallExpr *E) { 1123 switch (Target.getTriple().getArch()) { 1124 case llvm::Triple::arm: 1125 case llvm::Triple::thumb: 1126 return EmitARMBuiltinExpr(BuiltinID, E); 1127 case llvm::Triple::x86: 1128 case llvm::Triple::x86_64: 1129 return EmitX86BuiltinExpr(BuiltinID, E); 1130 case llvm::Triple::ppc: 1131 case llvm::Triple::ppc64: 1132 return EmitPPCBuiltinExpr(BuiltinID, E); 1133 default: 1134 return 0; 1135 } 1136 } 1137 1138 static llvm::VectorType *GetNeonType(LLVMContext &C, unsigned type, bool q) { 1139 switch (type) { 1140 default: break; 1141 case 0: 1142 case 5: return llvm::VectorType::get(llvm::Type::getInt8Ty(C), 8 << (int)q); 1143 case 6: 1144 case 7: 1145 case 1: return llvm::VectorType::get(llvm::Type::getInt16Ty(C),4 << (int)q); 1146 case 2: return llvm::VectorType::get(llvm::Type::getInt32Ty(C),2 << (int)q); 1147 case 3: return llvm::VectorType::get(llvm::Type::getInt64Ty(C),1 << (int)q); 1148 case 4: return llvm::VectorType::get(llvm::Type::getFloatTy(C),2 << (int)q); 1149 }; 1150 return 0; 1151 } 1152 1153 Value *CodeGenFunction::EmitNeonSplat(Value *V, Constant *C) { 1154 unsigned nElts = cast<llvm::VectorType>(V->getType())->getNumElements(); 1155 SmallVector<Constant*, 16> Indices(nElts, C); 1156 Value* SV = llvm::ConstantVector::get(Indices); 1157 return Builder.CreateShuffleVector(V, V, SV, "lane"); 1158 } 1159 1160 Value *CodeGenFunction::EmitNeonCall(Function *F, SmallVectorImpl<Value*> &Ops, 1161 const char *name, 1162 unsigned shift, bool rightshift) { 1163 unsigned j = 0; 1164 for (Function::const_arg_iterator ai = F->arg_begin(), ae = F->arg_end(); 1165 ai != ae; ++ai, ++j) 1166 if (shift > 0 && shift == j) 1167 Ops[j] = EmitNeonShiftVector(Ops[j], ai->getType(), rightshift); 1168 else 1169 Ops[j] = Builder.CreateBitCast(Ops[j], ai->getType(), name); 1170 1171 return Builder.CreateCall(F, Ops, name); 1172 } 1173 1174 Value *CodeGenFunction::EmitNeonShiftVector(Value *V, llvm::Type *Ty, 1175 bool neg) { 1176 ConstantInt *CI = cast<ConstantInt>(V); 1177 int SV = CI->getSExtValue(); 1178 1179 llvm::VectorType *VTy = cast<llvm::VectorType>(Ty); 1180 llvm::Constant *C = ConstantInt::get(VTy->getElementType(), neg ? -SV : SV); 1181 SmallVector<llvm::Constant*, 16> CV(VTy->getNumElements(), C); 1182 return llvm::ConstantVector::get(CV); 1183 } 1184 1185 /// GetPointeeAlignment - Given an expression with a pointer type, find the 1186 /// alignment of the type referenced by the pointer. Skip over implicit 1187 /// casts. 1188 static Value *GetPointeeAlignment(CodeGenFunction &CGF, const Expr *Addr) { 1189 unsigned Align = 1; 1190 // Check if the type is a pointer. The implicit cast operand might not be. 1191 while (Addr->getType()->isPointerType()) { 1192 QualType PtTy = Addr->getType()->getPointeeType(); 1193 unsigned NewA = CGF.getContext().getTypeAlignInChars(PtTy).getQuantity(); 1194 if (NewA > Align) 1195 Align = NewA; 1196 1197 // If the address is an implicit cast, repeat with the cast operand. 1198 if (const ImplicitCastExpr *CastAddr = dyn_cast<ImplicitCastExpr>(Addr)) { 1199 Addr = CastAddr->getSubExpr(); 1200 continue; 1201 } 1202 break; 1203 } 1204 return llvm::ConstantInt::get(CGF.Int32Ty, Align); 1205 } 1206 1207 Value *CodeGenFunction::EmitARMBuiltinExpr(unsigned BuiltinID, 1208 const CallExpr *E) { 1209 if (BuiltinID == ARM::BI__clear_cache) { 1210 const FunctionDecl *FD = E->getDirectCallee(); 1211 // Oddly people write this call without args on occasion and gcc accepts 1212 // it - it's also marked as varargs in the description file. 1213 SmallVector<Value*, 2> Ops; 1214 for (unsigned i = 0; i < E->getNumArgs(); i++) 1215 Ops.push_back(EmitScalarExpr(E->getArg(i))); 1216 llvm::Type *Ty = CGM.getTypes().ConvertType(FD->getType()); 1217 llvm::FunctionType *FTy = cast<llvm::FunctionType>(Ty); 1218 StringRef Name = FD->getName(); 1219 return Builder.CreateCall(CGM.CreateRuntimeFunction(FTy, Name), Ops); 1220 } 1221 1222 if (BuiltinID == ARM::BI__builtin_arm_ldrexd) { 1223 Function *F = CGM.getIntrinsic(Intrinsic::arm_ldrexd); 1224 1225 Value *LdPtr = EmitScalarExpr(E->getArg(0)); 1226 Value *Val = Builder.CreateCall(F, LdPtr, "ldrexd"); 1227 1228 Value *Val0 = Builder.CreateExtractValue(Val, 1); 1229 Value *Val1 = Builder.CreateExtractValue(Val, 0); 1230 Val0 = Builder.CreateZExt(Val0, Int64Ty); 1231 Val1 = Builder.CreateZExt(Val1, Int64Ty); 1232 1233 Value *ShiftCst = llvm::ConstantInt::get(Int64Ty, 32); 1234 Val = Builder.CreateShl(Val0, ShiftCst, "shl", true /* nuw */); 1235 return Builder.CreateOr(Val, Val1); 1236 } 1237 1238 if (BuiltinID == ARM::BI__builtin_arm_strexd) { 1239 Function *F = CGM.getIntrinsic(Intrinsic::arm_strexd); 1240 llvm::Type *STy = llvm::StructType::get(Int32Ty, Int32Ty, NULL); 1241 1242 Value *One = llvm::ConstantInt::get(Int32Ty, 1); 1243 Value *Tmp = Builder.CreateAlloca(Int64Ty, One); 1244 Value *Val = EmitScalarExpr(E->getArg(0)); 1245 Builder.CreateStore(Val, Tmp); 1246 1247 Value *LdPtr = Builder.CreateBitCast(Tmp,llvm::PointerType::getUnqual(STy)); 1248 Val = Builder.CreateLoad(LdPtr); 1249 1250 Value *Arg0 = Builder.CreateExtractValue(Val, 0); 1251 Value *Arg1 = Builder.CreateExtractValue(Val, 1); 1252 Value *StPtr = EmitScalarExpr(E->getArg(1)); 1253 return Builder.CreateCall3(F, Arg0, Arg1, StPtr, "strexd"); 1254 } 1255 1256 SmallVector<Value*, 4> Ops; 1257 for (unsigned i = 0, e = E->getNumArgs() - 1; i != e; i++) 1258 Ops.push_back(EmitScalarExpr(E->getArg(i))); 1259 1260 // vget_lane and vset_lane are not overloaded and do not have an extra 1261 // argument that specifies the vector type. 1262 switch (BuiltinID) { 1263 default: break; 1264 case ARM::BI__builtin_neon_vget_lane_i8: 1265 case ARM::BI__builtin_neon_vget_lane_i16: 1266 case ARM::BI__builtin_neon_vget_lane_i32: 1267 case ARM::BI__builtin_neon_vget_lane_i64: 1268 case ARM::BI__builtin_neon_vget_lane_f32: 1269 case ARM::BI__builtin_neon_vgetq_lane_i8: 1270 case ARM::BI__builtin_neon_vgetq_lane_i16: 1271 case ARM::BI__builtin_neon_vgetq_lane_i32: 1272 case ARM::BI__builtin_neon_vgetq_lane_i64: 1273 case ARM::BI__builtin_neon_vgetq_lane_f32: 1274 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 1275 "vget_lane"); 1276 case ARM::BI__builtin_neon_vset_lane_i8: 1277 case ARM::BI__builtin_neon_vset_lane_i16: 1278 case ARM::BI__builtin_neon_vset_lane_i32: 1279 case ARM::BI__builtin_neon_vset_lane_i64: 1280 case ARM::BI__builtin_neon_vset_lane_f32: 1281 case ARM::BI__builtin_neon_vsetq_lane_i8: 1282 case ARM::BI__builtin_neon_vsetq_lane_i16: 1283 case ARM::BI__builtin_neon_vsetq_lane_i32: 1284 case ARM::BI__builtin_neon_vsetq_lane_i64: 1285 case ARM::BI__builtin_neon_vsetq_lane_f32: 1286 Ops.push_back(EmitScalarExpr(E->getArg(2))); 1287 return Builder.CreateInsertElement(Ops[1], Ops[0], Ops[2], "vset_lane"); 1288 } 1289 1290 // Get the last argument, which specifies the vector type. 1291 llvm::APSInt Result; 1292 const Expr *Arg = E->getArg(E->getNumArgs()-1); 1293 if (!Arg->isIntegerConstantExpr(Result, getContext())) 1294 return 0; 1295 1296 if (BuiltinID == ARM::BI__builtin_arm_vcvtr_f || 1297 BuiltinID == ARM::BI__builtin_arm_vcvtr_d) { 1298 // Determine the overloaded type of this builtin. 1299 llvm::Type *Ty; 1300 if (BuiltinID == ARM::BI__builtin_arm_vcvtr_f) 1301 Ty = llvm::Type::getFloatTy(getLLVMContext()); 1302 else 1303 Ty = llvm::Type::getDoubleTy(getLLVMContext()); 1304 1305 // Determine whether this is an unsigned conversion or not. 1306 bool usgn = Result.getZExtValue() == 1; 1307 unsigned Int = usgn ? Intrinsic::arm_vcvtru : Intrinsic::arm_vcvtr; 1308 1309 // Call the appropriate intrinsic. 1310 Function *F = CGM.getIntrinsic(Int, Ty); 1311 return Builder.CreateCall(F, Ops, "vcvtr"); 1312 } 1313 1314 // Determine the type of this overloaded NEON intrinsic. 1315 unsigned type = Result.getZExtValue(); 1316 bool usgn = type & 0x08; 1317 bool quad = type & 0x10; 1318 bool poly = (type & 0x7) == 5 || (type & 0x7) == 6; 1319 (void)poly; // Only used in assert()s. 1320 bool rightShift = false; 1321 1322 llvm::VectorType *VTy = GetNeonType(getLLVMContext(), type & 0x7, quad); 1323 llvm::Type *Ty = VTy; 1324 if (!Ty) 1325 return 0; 1326 1327 unsigned Int; 1328 switch (BuiltinID) { 1329 default: return 0; 1330 case ARM::BI__builtin_neon_vabd_v: 1331 case ARM::BI__builtin_neon_vabdq_v: 1332 Int = usgn ? Intrinsic::arm_neon_vabdu : Intrinsic::arm_neon_vabds; 1333 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vabd"); 1334 case ARM::BI__builtin_neon_vabs_v: 1335 case ARM::BI__builtin_neon_vabsq_v: 1336 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vabs, Ty), 1337 Ops, "vabs"); 1338 case ARM::BI__builtin_neon_vaddhn_v: 1339 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vaddhn, Ty), 1340 Ops, "vaddhn"); 1341 case ARM::BI__builtin_neon_vcale_v: 1342 std::swap(Ops[0], Ops[1]); 1343 case ARM::BI__builtin_neon_vcage_v: { 1344 Function *F = CGM.getIntrinsic(Intrinsic::arm_neon_vacged); 1345 return EmitNeonCall(F, Ops, "vcage"); 1346 } 1347 case ARM::BI__builtin_neon_vcaleq_v: 1348 std::swap(Ops[0], Ops[1]); 1349 case ARM::BI__builtin_neon_vcageq_v: { 1350 Function *F = CGM.getIntrinsic(Intrinsic::arm_neon_vacgeq); 1351 return EmitNeonCall(F, Ops, "vcage"); 1352 } 1353 case ARM::BI__builtin_neon_vcalt_v: 1354 std::swap(Ops[0], Ops[1]); 1355 case ARM::BI__builtin_neon_vcagt_v: { 1356 Function *F = CGM.getIntrinsic(Intrinsic::arm_neon_vacgtd); 1357 return EmitNeonCall(F, Ops, "vcagt"); 1358 } 1359 case ARM::BI__builtin_neon_vcaltq_v: 1360 std::swap(Ops[0], Ops[1]); 1361 case ARM::BI__builtin_neon_vcagtq_v: { 1362 Function *F = CGM.getIntrinsic(Intrinsic::arm_neon_vacgtq); 1363 return EmitNeonCall(F, Ops, "vcagt"); 1364 } 1365 case ARM::BI__builtin_neon_vcls_v: 1366 case ARM::BI__builtin_neon_vclsq_v: { 1367 Function *F = CGM.getIntrinsic(Intrinsic::arm_neon_vcls, Ty); 1368 return EmitNeonCall(F, Ops, "vcls"); 1369 } 1370 case ARM::BI__builtin_neon_vclz_v: 1371 case ARM::BI__builtin_neon_vclzq_v: { 1372 Function *F = CGM.getIntrinsic(Intrinsic::arm_neon_vclz, Ty); 1373 return EmitNeonCall(F, Ops, "vclz"); 1374 } 1375 case ARM::BI__builtin_neon_vcnt_v: 1376 case ARM::BI__builtin_neon_vcntq_v: { 1377 Function *F = CGM.getIntrinsic(Intrinsic::arm_neon_vcnt, Ty); 1378 return EmitNeonCall(F, Ops, "vcnt"); 1379 } 1380 case ARM::BI__builtin_neon_vcvt_f16_v: { 1381 assert((type & 0x7) == 7 && !quad && "unexpected vcvt_f16_v builtin"); 1382 Function *F = CGM.getIntrinsic(Intrinsic::arm_neon_vcvtfp2hf); 1383 return EmitNeonCall(F, Ops, "vcvt"); 1384 } 1385 case ARM::BI__builtin_neon_vcvt_f32_f16: { 1386 assert((type & 0x7) == 7 && !quad && "unexpected vcvt_f32_f16 builtin"); 1387 Function *F = CGM.getIntrinsic(Intrinsic::arm_neon_vcvthf2fp); 1388 return EmitNeonCall(F, Ops, "vcvt"); 1389 } 1390 case ARM::BI__builtin_neon_vcvt_f32_v: 1391 case ARM::BI__builtin_neon_vcvtq_f32_v: { 1392 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 1393 Ty = GetNeonType(getLLVMContext(), 4, quad); 1394 return usgn ? Builder.CreateUIToFP(Ops[0], Ty, "vcvt") 1395 : Builder.CreateSIToFP(Ops[0], Ty, "vcvt"); 1396 } 1397 case ARM::BI__builtin_neon_vcvt_s32_v: 1398 case ARM::BI__builtin_neon_vcvt_u32_v: 1399 case ARM::BI__builtin_neon_vcvtq_s32_v: 1400 case ARM::BI__builtin_neon_vcvtq_u32_v: { 1401 Ops[0] = Builder.CreateBitCast(Ops[0], GetNeonType(getLLVMContext(), 4, quad)); 1402 return usgn ? Builder.CreateFPToUI(Ops[0], Ty, "vcvt") 1403 : Builder.CreateFPToSI(Ops[0], Ty, "vcvt"); 1404 } 1405 case ARM::BI__builtin_neon_vcvt_n_f32_v: 1406 case ARM::BI__builtin_neon_vcvtq_n_f32_v: { 1407 llvm::Type *Tys[2] = { GetNeonType(getLLVMContext(), 4, quad), Ty }; 1408 Int = usgn ? Intrinsic::arm_neon_vcvtfxu2fp : Intrinsic::arm_neon_vcvtfxs2fp; 1409 Function *F = CGM.getIntrinsic(Int, Tys); 1410 return EmitNeonCall(F, Ops, "vcvt_n"); 1411 } 1412 case ARM::BI__builtin_neon_vcvt_n_s32_v: 1413 case ARM::BI__builtin_neon_vcvt_n_u32_v: 1414 case ARM::BI__builtin_neon_vcvtq_n_s32_v: 1415 case ARM::BI__builtin_neon_vcvtq_n_u32_v: { 1416 llvm::Type *Tys[2] = { Ty, GetNeonType(getLLVMContext(), 4, quad) }; 1417 Int = usgn ? Intrinsic::arm_neon_vcvtfp2fxu : Intrinsic::arm_neon_vcvtfp2fxs; 1418 Function *F = CGM.getIntrinsic(Int, Tys); 1419 return EmitNeonCall(F, Ops, "vcvt_n"); 1420 } 1421 case ARM::BI__builtin_neon_vext_v: 1422 case ARM::BI__builtin_neon_vextq_v: { 1423 int CV = cast<ConstantInt>(Ops[2])->getSExtValue(); 1424 SmallVector<Constant*, 16> Indices; 1425 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) 1426 Indices.push_back(ConstantInt::get(Int32Ty, i+CV)); 1427 1428 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 1429 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 1430 Value *SV = llvm::ConstantVector::get(Indices); 1431 return Builder.CreateShuffleVector(Ops[0], Ops[1], SV, "vext"); 1432 } 1433 case ARM::BI__builtin_neon_vhadd_v: 1434 case ARM::BI__builtin_neon_vhaddq_v: 1435 Int = usgn ? Intrinsic::arm_neon_vhaddu : Intrinsic::arm_neon_vhadds; 1436 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vhadd"); 1437 case ARM::BI__builtin_neon_vhsub_v: 1438 case ARM::BI__builtin_neon_vhsubq_v: 1439 Int = usgn ? Intrinsic::arm_neon_vhsubu : Intrinsic::arm_neon_vhsubs; 1440 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vhsub"); 1441 case ARM::BI__builtin_neon_vld1_v: 1442 case ARM::BI__builtin_neon_vld1q_v: 1443 Ops.push_back(GetPointeeAlignment(*this, E->getArg(0))); 1444 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vld1, Ty), 1445 Ops, "vld1"); 1446 case ARM::BI__builtin_neon_vld1_lane_v: 1447 case ARM::BI__builtin_neon_vld1q_lane_v: 1448 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 1449 Ty = llvm::PointerType::getUnqual(VTy->getElementType()); 1450 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 1451 Ops[0] = Builder.CreateLoad(Ops[0]); 1452 return Builder.CreateInsertElement(Ops[1], Ops[0], Ops[2], "vld1_lane"); 1453 case ARM::BI__builtin_neon_vld1_dup_v: 1454 case ARM::BI__builtin_neon_vld1q_dup_v: { 1455 Value *V = UndefValue::get(Ty); 1456 Ty = llvm::PointerType::getUnqual(VTy->getElementType()); 1457 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 1458 Ops[0] = Builder.CreateLoad(Ops[0]); 1459 llvm::Constant *CI = ConstantInt::get(Int32Ty, 0); 1460 Ops[0] = Builder.CreateInsertElement(V, Ops[0], CI); 1461 return EmitNeonSplat(Ops[0], CI); 1462 } 1463 case ARM::BI__builtin_neon_vld2_v: 1464 case ARM::BI__builtin_neon_vld2q_v: { 1465 Function *F = CGM.getIntrinsic(Intrinsic::arm_neon_vld2, Ty); 1466 Value *Align = GetPointeeAlignment(*this, E->getArg(1)); 1467 Ops[1] = Builder.CreateCall2(F, Ops[1], Align, "vld2"); 1468 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 1469 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 1470 return Builder.CreateStore(Ops[1], Ops[0]); 1471 } 1472 case ARM::BI__builtin_neon_vld3_v: 1473 case ARM::BI__builtin_neon_vld3q_v: { 1474 Function *F = CGM.getIntrinsic(Intrinsic::arm_neon_vld3, Ty); 1475 Value *Align = GetPointeeAlignment(*this, E->getArg(1)); 1476 Ops[1] = Builder.CreateCall2(F, Ops[1], Align, "vld3"); 1477 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 1478 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 1479 return Builder.CreateStore(Ops[1], Ops[0]); 1480 } 1481 case ARM::BI__builtin_neon_vld4_v: 1482 case ARM::BI__builtin_neon_vld4q_v: { 1483 Function *F = CGM.getIntrinsic(Intrinsic::arm_neon_vld4, Ty); 1484 Value *Align = GetPointeeAlignment(*this, E->getArg(1)); 1485 Ops[1] = Builder.CreateCall2(F, Ops[1], Align, "vld4"); 1486 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 1487 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 1488 return Builder.CreateStore(Ops[1], Ops[0]); 1489 } 1490 case ARM::BI__builtin_neon_vld2_lane_v: 1491 case ARM::BI__builtin_neon_vld2q_lane_v: { 1492 Function *F = CGM.getIntrinsic(Intrinsic::arm_neon_vld2lane, Ty); 1493 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 1494 Ops[3] = Builder.CreateBitCast(Ops[3], Ty); 1495 Ops.push_back(GetPointeeAlignment(*this, E->getArg(1))); 1496 Ops[1] = Builder.CreateCall(F, makeArrayRef(Ops).slice(1), "vld2_lane"); 1497 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 1498 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 1499 return Builder.CreateStore(Ops[1], Ops[0]); 1500 } 1501 case ARM::BI__builtin_neon_vld3_lane_v: 1502 case ARM::BI__builtin_neon_vld3q_lane_v: { 1503 Function *F = CGM.getIntrinsic(Intrinsic::arm_neon_vld3lane, Ty); 1504 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 1505 Ops[3] = Builder.CreateBitCast(Ops[3], Ty); 1506 Ops[4] = Builder.CreateBitCast(Ops[4], Ty); 1507 Ops.push_back(GetPointeeAlignment(*this, E->getArg(1))); 1508 Ops[1] = Builder.CreateCall(F, makeArrayRef(Ops).slice(1), "vld3_lane"); 1509 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 1510 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 1511 return Builder.CreateStore(Ops[1], Ops[0]); 1512 } 1513 case ARM::BI__builtin_neon_vld4_lane_v: 1514 case ARM::BI__builtin_neon_vld4q_lane_v: { 1515 Function *F = CGM.getIntrinsic(Intrinsic::arm_neon_vld4lane, Ty); 1516 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 1517 Ops[3] = Builder.CreateBitCast(Ops[3], Ty); 1518 Ops[4] = Builder.CreateBitCast(Ops[4], Ty); 1519 Ops[5] = Builder.CreateBitCast(Ops[5], Ty); 1520 Ops.push_back(GetPointeeAlignment(*this, E->getArg(1))); 1521 Ops[1] = Builder.CreateCall(F, makeArrayRef(Ops).slice(1), "vld3_lane"); 1522 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 1523 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 1524 return Builder.CreateStore(Ops[1], Ops[0]); 1525 } 1526 case ARM::BI__builtin_neon_vld2_dup_v: 1527 case ARM::BI__builtin_neon_vld3_dup_v: 1528 case ARM::BI__builtin_neon_vld4_dup_v: { 1529 // Handle 64-bit elements as a special-case. There is no "dup" needed. 1530 if (VTy->getElementType()->getPrimitiveSizeInBits() == 64) { 1531 switch (BuiltinID) { 1532 case ARM::BI__builtin_neon_vld2_dup_v: 1533 Int = Intrinsic::arm_neon_vld2; 1534 break; 1535 case ARM::BI__builtin_neon_vld3_dup_v: 1536 Int = Intrinsic::arm_neon_vld2; 1537 break; 1538 case ARM::BI__builtin_neon_vld4_dup_v: 1539 Int = Intrinsic::arm_neon_vld2; 1540 break; 1541 default: llvm_unreachable("unknown vld_dup intrinsic?"); 1542 } 1543 Function *F = CGM.getIntrinsic(Int, Ty); 1544 Value *Align = GetPointeeAlignment(*this, E->getArg(1)); 1545 Ops[1] = Builder.CreateCall2(F, Ops[1], Align, "vld_dup"); 1546 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 1547 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 1548 return Builder.CreateStore(Ops[1], Ops[0]); 1549 } 1550 switch (BuiltinID) { 1551 case ARM::BI__builtin_neon_vld2_dup_v: 1552 Int = Intrinsic::arm_neon_vld2lane; 1553 break; 1554 case ARM::BI__builtin_neon_vld3_dup_v: 1555 Int = Intrinsic::arm_neon_vld2lane; 1556 break; 1557 case ARM::BI__builtin_neon_vld4_dup_v: 1558 Int = Intrinsic::arm_neon_vld2lane; 1559 break; 1560 default: llvm_unreachable("unknown vld_dup intrinsic?"); 1561 } 1562 Function *F = CGM.getIntrinsic(Int, Ty); 1563 llvm::StructType *STy = cast<llvm::StructType>(F->getReturnType()); 1564 1565 SmallVector<Value*, 6> Args; 1566 Args.push_back(Ops[1]); 1567 Args.append(STy->getNumElements(), UndefValue::get(Ty)); 1568 1569 llvm::Constant *CI = ConstantInt::get(Int32Ty, 0); 1570 Args.push_back(CI); 1571 Args.push_back(GetPointeeAlignment(*this, E->getArg(1))); 1572 1573 Ops[1] = Builder.CreateCall(F, Args, "vld_dup"); 1574 // splat lane 0 to all elts in each vector of the result. 1575 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 1576 Value *Val = Builder.CreateExtractValue(Ops[1], i); 1577 Value *Elt = Builder.CreateBitCast(Val, Ty); 1578 Elt = EmitNeonSplat(Elt, CI); 1579 Elt = Builder.CreateBitCast(Elt, Val->getType()); 1580 Ops[1] = Builder.CreateInsertValue(Ops[1], Elt, i); 1581 } 1582 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 1583 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 1584 return Builder.CreateStore(Ops[1], Ops[0]); 1585 } 1586 case ARM::BI__builtin_neon_vmax_v: 1587 case ARM::BI__builtin_neon_vmaxq_v: 1588 Int = usgn ? Intrinsic::arm_neon_vmaxu : Intrinsic::arm_neon_vmaxs; 1589 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmax"); 1590 case ARM::BI__builtin_neon_vmin_v: 1591 case ARM::BI__builtin_neon_vminq_v: 1592 Int = usgn ? Intrinsic::arm_neon_vminu : Intrinsic::arm_neon_vmins; 1593 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmin"); 1594 case ARM::BI__builtin_neon_vmovl_v: { 1595 llvm::Type *DTy =llvm::VectorType::getTruncatedElementVectorType(VTy); 1596 Ops[0] = Builder.CreateBitCast(Ops[0], DTy); 1597 if (usgn) 1598 return Builder.CreateZExt(Ops[0], Ty, "vmovl"); 1599 return Builder.CreateSExt(Ops[0], Ty, "vmovl"); 1600 } 1601 case ARM::BI__builtin_neon_vmovn_v: { 1602 llvm::Type *QTy = llvm::VectorType::getExtendedElementVectorType(VTy); 1603 Ops[0] = Builder.CreateBitCast(Ops[0], QTy); 1604 return Builder.CreateTrunc(Ops[0], Ty, "vmovn"); 1605 } 1606 case ARM::BI__builtin_neon_vmul_v: 1607 case ARM::BI__builtin_neon_vmulq_v: 1608 assert(poly && "vmul builtin only supported for polynomial types"); 1609 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vmulp, Ty), 1610 Ops, "vmul"); 1611 case ARM::BI__builtin_neon_vmull_v: 1612 Int = usgn ? Intrinsic::arm_neon_vmullu : Intrinsic::arm_neon_vmulls; 1613 Int = poly ? (unsigned)Intrinsic::arm_neon_vmullp : Int; 1614 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmull"); 1615 case ARM::BI__builtin_neon_vpadal_v: 1616 case ARM::BI__builtin_neon_vpadalq_v: { 1617 Int = usgn ? Intrinsic::arm_neon_vpadalu : Intrinsic::arm_neon_vpadals; 1618 // The source operand type has twice as many elements of half the size. 1619 unsigned EltBits = VTy->getElementType()->getPrimitiveSizeInBits(); 1620 llvm::Type *EltTy = 1621 llvm::IntegerType::get(getLLVMContext(), EltBits / 2); 1622 llvm::Type *NarrowTy = 1623 llvm::VectorType::get(EltTy, VTy->getNumElements() * 2); 1624 llvm::Type *Tys[2] = { Ty, NarrowTy }; 1625 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vpadal"); 1626 } 1627 case ARM::BI__builtin_neon_vpadd_v: 1628 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vpadd, Ty), 1629 Ops, "vpadd"); 1630 case ARM::BI__builtin_neon_vpaddl_v: 1631 case ARM::BI__builtin_neon_vpaddlq_v: { 1632 Int = usgn ? Intrinsic::arm_neon_vpaddlu : Intrinsic::arm_neon_vpaddls; 1633 // The source operand type has twice as many elements of half the size. 1634 unsigned EltBits = VTy->getElementType()->getPrimitiveSizeInBits(); 1635 llvm::Type *EltTy = llvm::IntegerType::get(getLLVMContext(), EltBits / 2); 1636 llvm::Type *NarrowTy = 1637 llvm::VectorType::get(EltTy, VTy->getNumElements() * 2); 1638 llvm::Type *Tys[2] = { Ty, NarrowTy }; 1639 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vpaddl"); 1640 } 1641 case ARM::BI__builtin_neon_vpmax_v: 1642 Int = usgn ? Intrinsic::arm_neon_vpmaxu : Intrinsic::arm_neon_vpmaxs; 1643 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vpmax"); 1644 case ARM::BI__builtin_neon_vpmin_v: 1645 Int = usgn ? Intrinsic::arm_neon_vpminu : Intrinsic::arm_neon_vpmins; 1646 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vpmin"); 1647 case ARM::BI__builtin_neon_vqabs_v: 1648 case ARM::BI__builtin_neon_vqabsq_v: 1649 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vqabs, Ty), 1650 Ops, "vqabs"); 1651 case ARM::BI__builtin_neon_vqadd_v: 1652 case ARM::BI__builtin_neon_vqaddq_v: 1653 Int = usgn ? Intrinsic::arm_neon_vqaddu : Intrinsic::arm_neon_vqadds; 1654 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqadd"); 1655 case ARM::BI__builtin_neon_vqdmlal_v: 1656 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vqdmlal, Ty), 1657 Ops, "vqdmlal"); 1658 case ARM::BI__builtin_neon_vqdmlsl_v: 1659 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vqdmlsl, Ty), 1660 Ops, "vqdmlsl"); 1661 case ARM::BI__builtin_neon_vqdmulh_v: 1662 case ARM::BI__builtin_neon_vqdmulhq_v: 1663 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vqdmulh, Ty), 1664 Ops, "vqdmulh"); 1665 case ARM::BI__builtin_neon_vqdmull_v: 1666 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vqdmull, Ty), 1667 Ops, "vqdmull"); 1668 case ARM::BI__builtin_neon_vqmovn_v: 1669 Int = usgn ? Intrinsic::arm_neon_vqmovnu : Intrinsic::arm_neon_vqmovns; 1670 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqmovn"); 1671 case ARM::BI__builtin_neon_vqmovun_v: 1672 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vqmovnsu, Ty), 1673 Ops, "vqdmull"); 1674 case ARM::BI__builtin_neon_vqneg_v: 1675 case ARM::BI__builtin_neon_vqnegq_v: 1676 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vqneg, Ty), 1677 Ops, "vqneg"); 1678 case ARM::BI__builtin_neon_vqrdmulh_v: 1679 case ARM::BI__builtin_neon_vqrdmulhq_v: 1680 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vqrdmulh, Ty), 1681 Ops, "vqrdmulh"); 1682 case ARM::BI__builtin_neon_vqrshl_v: 1683 case ARM::BI__builtin_neon_vqrshlq_v: 1684 Int = usgn ? Intrinsic::arm_neon_vqrshiftu : Intrinsic::arm_neon_vqrshifts; 1685 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqrshl"); 1686 case ARM::BI__builtin_neon_vqrshrn_n_v: 1687 Int = usgn ? Intrinsic::arm_neon_vqrshiftnu : Intrinsic::arm_neon_vqrshiftns; 1688 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqrshrn_n", 1689 1, true); 1690 case ARM::BI__builtin_neon_vqrshrun_n_v: 1691 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vqrshiftnsu, Ty), 1692 Ops, "vqrshrun_n", 1, true); 1693 case ARM::BI__builtin_neon_vqshl_v: 1694 case ARM::BI__builtin_neon_vqshlq_v: 1695 Int = usgn ? Intrinsic::arm_neon_vqshiftu : Intrinsic::arm_neon_vqshifts; 1696 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqshl"); 1697 case ARM::BI__builtin_neon_vqshl_n_v: 1698 case ARM::BI__builtin_neon_vqshlq_n_v: 1699 Int = usgn ? Intrinsic::arm_neon_vqshiftu : Intrinsic::arm_neon_vqshifts; 1700 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqshl_n", 1701 1, false); 1702 case ARM::BI__builtin_neon_vqshlu_n_v: 1703 case ARM::BI__builtin_neon_vqshluq_n_v: 1704 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vqshiftsu, Ty), 1705 Ops, "vqshlu", 1, false); 1706 case ARM::BI__builtin_neon_vqshrn_n_v: 1707 Int = usgn ? Intrinsic::arm_neon_vqshiftnu : Intrinsic::arm_neon_vqshiftns; 1708 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqshrn_n", 1709 1, true); 1710 case ARM::BI__builtin_neon_vqshrun_n_v: 1711 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vqshiftnsu, Ty), 1712 Ops, "vqshrun_n", 1, true); 1713 case ARM::BI__builtin_neon_vqsub_v: 1714 case ARM::BI__builtin_neon_vqsubq_v: 1715 Int = usgn ? Intrinsic::arm_neon_vqsubu : Intrinsic::arm_neon_vqsubs; 1716 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqsub"); 1717 case ARM::BI__builtin_neon_vraddhn_v: 1718 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vraddhn, Ty), 1719 Ops, "vraddhn"); 1720 case ARM::BI__builtin_neon_vrecpe_v: 1721 case ARM::BI__builtin_neon_vrecpeq_v: 1722 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vrecpe, Ty), 1723 Ops, "vrecpe"); 1724 case ARM::BI__builtin_neon_vrecps_v: 1725 case ARM::BI__builtin_neon_vrecpsq_v: 1726 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vrecps, Ty), 1727 Ops, "vrecps"); 1728 case ARM::BI__builtin_neon_vrhadd_v: 1729 case ARM::BI__builtin_neon_vrhaddq_v: 1730 Int = usgn ? Intrinsic::arm_neon_vrhaddu : Intrinsic::arm_neon_vrhadds; 1731 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrhadd"); 1732 case ARM::BI__builtin_neon_vrshl_v: 1733 case ARM::BI__builtin_neon_vrshlq_v: 1734 Int = usgn ? Intrinsic::arm_neon_vrshiftu : Intrinsic::arm_neon_vrshifts; 1735 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrshl"); 1736 case ARM::BI__builtin_neon_vrshrn_n_v: 1737 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vrshiftn, Ty), 1738 Ops, "vrshrn_n", 1, true); 1739 case ARM::BI__builtin_neon_vrshr_n_v: 1740 case ARM::BI__builtin_neon_vrshrq_n_v: 1741 Int = usgn ? Intrinsic::arm_neon_vrshiftu : Intrinsic::arm_neon_vrshifts; 1742 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrshr_n", 1, true); 1743 case ARM::BI__builtin_neon_vrsqrte_v: 1744 case ARM::BI__builtin_neon_vrsqrteq_v: 1745 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vrsqrte, Ty), 1746 Ops, "vrsqrte"); 1747 case ARM::BI__builtin_neon_vrsqrts_v: 1748 case ARM::BI__builtin_neon_vrsqrtsq_v: 1749 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vrsqrts, Ty), 1750 Ops, "vrsqrts"); 1751 case ARM::BI__builtin_neon_vrsra_n_v: 1752 case ARM::BI__builtin_neon_vrsraq_n_v: 1753 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 1754 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 1755 Ops[2] = EmitNeonShiftVector(Ops[2], Ty, true); 1756 Int = usgn ? Intrinsic::arm_neon_vrshiftu : Intrinsic::arm_neon_vrshifts; 1757 Ops[1] = Builder.CreateCall2(CGM.getIntrinsic(Int, Ty), Ops[1], Ops[2]); 1758 return Builder.CreateAdd(Ops[0], Ops[1], "vrsra_n"); 1759 case ARM::BI__builtin_neon_vrsubhn_v: 1760 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vrsubhn, Ty), 1761 Ops, "vrsubhn"); 1762 case ARM::BI__builtin_neon_vshl_v: 1763 case ARM::BI__builtin_neon_vshlq_v: 1764 Int = usgn ? Intrinsic::arm_neon_vshiftu : Intrinsic::arm_neon_vshifts; 1765 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vshl"); 1766 case ARM::BI__builtin_neon_vshll_n_v: 1767 Int = usgn ? Intrinsic::arm_neon_vshiftlu : Intrinsic::arm_neon_vshiftls; 1768 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vshll", 1); 1769 case ARM::BI__builtin_neon_vshl_n_v: 1770 case ARM::BI__builtin_neon_vshlq_n_v: 1771 Ops[1] = EmitNeonShiftVector(Ops[1], Ty, false); 1772 return Builder.CreateShl(Builder.CreateBitCast(Ops[0],Ty), Ops[1], "vshl_n"); 1773 case ARM::BI__builtin_neon_vshrn_n_v: 1774 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vshiftn, Ty), 1775 Ops, "vshrn_n", 1, true); 1776 case ARM::BI__builtin_neon_vshr_n_v: 1777 case ARM::BI__builtin_neon_vshrq_n_v: 1778 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 1779 Ops[1] = EmitNeonShiftVector(Ops[1], Ty, false); 1780 if (usgn) 1781 return Builder.CreateLShr(Ops[0], Ops[1], "vshr_n"); 1782 else 1783 return Builder.CreateAShr(Ops[0], Ops[1], "vshr_n"); 1784 case ARM::BI__builtin_neon_vsri_n_v: 1785 case ARM::BI__builtin_neon_vsriq_n_v: 1786 rightShift = true; 1787 case ARM::BI__builtin_neon_vsli_n_v: 1788 case ARM::BI__builtin_neon_vsliq_n_v: 1789 Ops[2] = EmitNeonShiftVector(Ops[2], Ty, rightShift); 1790 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vshiftins, Ty), 1791 Ops, "vsli_n"); 1792 case ARM::BI__builtin_neon_vsra_n_v: 1793 case ARM::BI__builtin_neon_vsraq_n_v: 1794 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 1795 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 1796 Ops[2] = EmitNeonShiftVector(Ops[2], Ty, false); 1797 if (usgn) 1798 Ops[1] = Builder.CreateLShr(Ops[1], Ops[2], "vsra_n"); 1799 else 1800 Ops[1] = Builder.CreateAShr(Ops[1], Ops[2], "vsra_n"); 1801 return Builder.CreateAdd(Ops[0], Ops[1]); 1802 case ARM::BI__builtin_neon_vst1_v: 1803 case ARM::BI__builtin_neon_vst1q_v: 1804 Ops.push_back(GetPointeeAlignment(*this, E->getArg(0))); 1805 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vst1, Ty), 1806 Ops, ""); 1807 case ARM::BI__builtin_neon_vst1_lane_v: 1808 case ARM::BI__builtin_neon_vst1q_lane_v: 1809 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 1810 Ops[1] = Builder.CreateExtractElement(Ops[1], Ops[2]); 1811 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 1812 return Builder.CreateStore(Ops[1], Builder.CreateBitCast(Ops[0], Ty)); 1813 case ARM::BI__builtin_neon_vst2_v: 1814 case ARM::BI__builtin_neon_vst2q_v: 1815 Ops.push_back(GetPointeeAlignment(*this, E->getArg(0))); 1816 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vst2, Ty), 1817 Ops, ""); 1818 case ARM::BI__builtin_neon_vst2_lane_v: 1819 case ARM::BI__builtin_neon_vst2q_lane_v: 1820 Ops.push_back(GetPointeeAlignment(*this, E->getArg(0))); 1821 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vst2lane, Ty), 1822 Ops, ""); 1823 case ARM::BI__builtin_neon_vst3_v: 1824 case ARM::BI__builtin_neon_vst3q_v: 1825 Ops.push_back(GetPointeeAlignment(*this, E->getArg(0))); 1826 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vst3, Ty), 1827 Ops, ""); 1828 case ARM::BI__builtin_neon_vst3_lane_v: 1829 case ARM::BI__builtin_neon_vst3q_lane_v: 1830 Ops.push_back(GetPointeeAlignment(*this, E->getArg(0))); 1831 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vst3lane, Ty), 1832 Ops, ""); 1833 case ARM::BI__builtin_neon_vst4_v: 1834 case ARM::BI__builtin_neon_vst4q_v: 1835 Ops.push_back(GetPointeeAlignment(*this, E->getArg(0))); 1836 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vst4, Ty), 1837 Ops, ""); 1838 case ARM::BI__builtin_neon_vst4_lane_v: 1839 case ARM::BI__builtin_neon_vst4q_lane_v: 1840 Ops.push_back(GetPointeeAlignment(*this, E->getArg(0))); 1841 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vst4lane, Ty), 1842 Ops, ""); 1843 case ARM::BI__builtin_neon_vsubhn_v: 1844 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vsubhn, Ty), 1845 Ops, "vsubhn"); 1846 case ARM::BI__builtin_neon_vtbl1_v: 1847 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbl1), 1848 Ops, "vtbl1"); 1849 case ARM::BI__builtin_neon_vtbl2_v: 1850 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbl2), 1851 Ops, "vtbl2"); 1852 case ARM::BI__builtin_neon_vtbl3_v: 1853 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbl3), 1854 Ops, "vtbl3"); 1855 case ARM::BI__builtin_neon_vtbl4_v: 1856 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbl4), 1857 Ops, "vtbl4"); 1858 case ARM::BI__builtin_neon_vtbx1_v: 1859 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbx1), 1860 Ops, "vtbx1"); 1861 case ARM::BI__builtin_neon_vtbx2_v: 1862 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbx2), 1863 Ops, "vtbx2"); 1864 case ARM::BI__builtin_neon_vtbx3_v: 1865 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbx3), 1866 Ops, "vtbx3"); 1867 case ARM::BI__builtin_neon_vtbx4_v: 1868 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbx4), 1869 Ops, "vtbx4"); 1870 case ARM::BI__builtin_neon_vtst_v: 1871 case ARM::BI__builtin_neon_vtstq_v: { 1872 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 1873 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 1874 Ops[0] = Builder.CreateAnd(Ops[0], Ops[1]); 1875 Ops[0] = Builder.CreateICmp(ICmpInst::ICMP_NE, Ops[0], 1876 ConstantAggregateZero::get(Ty)); 1877 return Builder.CreateSExt(Ops[0], Ty, "vtst"); 1878 } 1879 case ARM::BI__builtin_neon_vtrn_v: 1880 case ARM::BI__builtin_neon_vtrnq_v: { 1881 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty)); 1882 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 1883 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 1884 Value *SV = 0; 1885 1886 for (unsigned vi = 0; vi != 2; ++vi) { 1887 SmallVector<Constant*, 16> Indices; 1888 for (unsigned i = 0, e = VTy->getNumElements(); i != e; i += 2) { 1889 Indices.push_back(ConstantInt::get(Int32Ty, i+vi)); 1890 Indices.push_back(ConstantInt::get(Int32Ty, i+e+vi)); 1891 } 1892 Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ops[0], vi); 1893 SV = llvm::ConstantVector::get(Indices); 1894 SV = Builder.CreateShuffleVector(Ops[1], Ops[2], SV, "vtrn"); 1895 SV = Builder.CreateStore(SV, Addr); 1896 } 1897 return SV; 1898 } 1899 case ARM::BI__builtin_neon_vuzp_v: 1900 case ARM::BI__builtin_neon_vuzpq_v: { 1901 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty)); 1902 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 1903 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 1904 Value *SV = 0; 1905 1906 for (unsigned vi = 0; vi != 2; ++vi) { 1907 SmallVector<Constant*, 16> Indices; 1908 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) 1909 Indices.push_back(ConstantInt::get(Int32Ty, 2*i+vi)); 1910 1911 Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ops[0], vi); 1912 SV = llvm::ConstantVector::get(Indices); 1913 SV = Builder.CreateShuffleVector(Ops[1], Ops[2], SV, "vuzp"); 1914 SV = Builder.CreateStore(SV, Addr); 1915 } 1916 return SV; 1917 } 1918 case ARM::BI__builtin_neon_vzip_v: 1919 case ARM::BI__builtin_neon_vzipq_v: { 1920 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty)); 1921 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 1922 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 1923 Value *SV = 0; 1924 1925 for (unsigned vi = 0; vi != 2; ++vi) { 1926 SmallVector<Constant*, 16> Indices; 1927 for (unsigned i = 0, e = VTy->getNumElements(); i != e; i += 2) { 1928 Indices.push_back(ConstantInt::get(Int32Ty, (i + vi*e) >> 1)); 1929 Indices.push_back(ConstantInt::get(Int32Ty, ((i + vi*e) >> 1)+e)); 1930 } 1931 Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ops[0], vi); 1932 SV = llvm::ConstantVector::get(Indices); 1933 SV = Builder.CreateShuffleVector(Ops[1], Ops[2], SV, "vzip"); 1934 SV = Builder.CreateStore(SV, Addr); 1935 } 1936 return SV; 1937 } 1938 } 1939 } 1940 1941 llvm::Value *CodeGenFunction:: 1942 BuildVector(const SmallVectorImpl<llvm::Value*> &Ops) { 1943 assert((Ops.size() & (Ops.size() - 1)) == 0 && 1944 "Not a power-of-two sized vector!"); 1945 bool AllConstants = true; 1946 for (unsigned i = 0, e = Ops.size(); i != e && AllConstants; ++i) 1947 AllConstants &= isa<Constant>(Ops[i]); 1948 1949 // If this is a constant vector, create a ConstantVector. 1950 if (AllConstants) { 1951 std::vector<llvm::Constant*> CstOps; 1952 for (unsigned i = 0, e = Ops.size(); i != e; ++i) 1953 CstOps.push_back(cast<Constant>(Ops[i])); 1954 return llvm::ConstantVector::get(CstOps); 1955 } 1956 1957 // Otherwise, insertelement the values to build the vector. 1958 Value *Result = 1959 llvm::UndefValue::get(llvm::VectorType::get(Ops[0]->getType(), Ops.size())); 1960 1961 for (unsigned i = 0, e = Ops.size(); i != e; ++i) 1962 Result = Builder.CreateInsertElement(Result, Ops[i], 1963 llvm::ConstantInt::get(llvm::Type::getInt32Ty(getLLVMContext()), i)); 1964 1965 return Result; 1966 } 1967 1968 Value *CodeGenFunction::EmitX86BuiltinExpr(unsigned BuiltinID, 1969 const CallExpr *E) { 1970 SmallVector<Value*, 4> Ops; 1971 1972 // Find out if any arguments are required to be integer constant expressions. 1973 unsigned ICEArguments = 0; 1974 ASTContext::GetBuiltinTypeError Error; 1975 getContext().GetBuiltinType(BuiltinID, Error, &ICEArguments); 1976 assert(Error == ASTContext::GE_None && "Should not codegen an error"); 1977 1978 for (unsigned i = 0, e = E->getNumArgs(); i != e; i++) { 1979 // If this is a normal argument, just emit it as a scalar. 1980 if ((ICEArguments & (1 << i)) == 0) { 1981 Ops.push_back(EmitScalarExpr(E->getArg(i))); 1982 continue; 1983 } 1984 1985 // If this is required to be a constant, constant fold it so that we know 1986 // that the generated intrinsic gets a ConstantInt. 1987 llvm::APSInt Result; 1988 bool IsConst = E->getArg(i)->isIntegerConstantExpr(Result, getContext()); 1989 assert(IsConst && "Constant arg isn't actually constant?"); (void)IsConst; 1990 Ops.push_back(llvm::ConstantInt::get(getLLVMContext(), Result)); 1991 } 1992 1993 switch (BuiltinID) { 1994 default: return 0; 1995 case X86::BI__builtin_ia32_pslldi128: 1996 case X86::BI__builtin_ia32_psllqi128: 1997 case X86::BI__builtin_ia32_psllwi128: 1998 case X86::BI__builtin_ia32_psradi128: 1999 case X86::BI__builtin_ia32_psrawi128: 2000 case X86::BI__builtin_ia32_psrldi128: 2001 case X86::BI__builtin_ia32_psrlqi128: 2002 case X86::BI__builtin_ia32_psrlwi128: { 2003 Ops[1] = Builder.CreateZExt(Ops[1], Int64Ty, "zext"); 2004 llvm::Type *Ty = llvm::VectorType::get(Int64Ty, 2); 2005 llvm::Value *Zero = llvm::ConstantInt::get(Int32Ty, 0); 2006 Ops[1] = Builder.CreateInsertElement(llvm::UndefValue::get(Ty), 2007 Ops[1], Zero, "insert"); 2008 Ops[1] = Builder.CreateBitCast(Ops[1], Ops[0]->getType(), "bitcast"); 2009 const char *name = 0; 2010 Intrinsic::ID ID = Intrinsic::not_intrinsic; 2011 2012 switch (BuiltinID) { 2013 default: llvm_unreachable("Unsupported shift intrinsic!"); 2014 case X86::BI__builtin_ia32_pslldi128: 2015 name = "pslldi"; 2016 ID = Intrinsic::x86_sse2_psll_d; 2017 break; 2018 case X86::BI__builtin_ia32_psllqi128: 2019 name = "psllqi"; 2020 ID = Intrinsic::x86_sse2_psll_q; 2021 break; 2022 case X86::BI__builtin_ia32_psllwi128: 2023 name = "psllwi"; 2024 ID = Intrinsic::x86_sse2_psll_w; 2025 break; 2026 case X86::BI__builtin_ia32_psradi128: 2027 name = "psradi"; 2028 ID = Intrinsic::x86_sse2_psra_d; 2029 break; 2030 case X86::BI__builtin_ia32_psrawi128: 2031 name = "psrawi"; 2032 ID = Intrinsic::x86_sse2_psra_w; 2033 break; 2034 case X86::BI__builtin_ia32_psrldi128: 2035 name = "psrldi"; 2036 ID = Intrinsic::x86_sse2_psrl_d; 2037 break; 2038 case X86::BI__builtin_ia32_psrlqi128: 2039 name = "psrlqi"; 2040 ID = Intrinsic::x86_sse2_psrl_q; 2041 break; 2042 case X86::BI__builtin_ia32_psrlwi128: 2043 name = "psrlwi"; 2044 ID = Intrinsic::x86_sse2_psrl_w; 2045 break; 2046 } 2047 llvm::Function *F = CGM.getIntrinsic(ID); 2048 return Builder.CreateCall(F, Ops, name); 2049 } 2050 case X86::BI__builtin_ia32_vec_init_v8qi: 2051 case X86::BI__builtin_ia32_vec_init_v4hi: 2052 case X86::BI__builtin_ia32_vec_init_v2si: 2053 return Builder.CreateBitCast(BuildVector(Ops), 2054 llvm::Type::getX86_MMXTy(getLLVMContext())); 2055 case X86::BI__builtin_ia32_vec_ext_v2si: 2056 return Builder.CreateExtractElement(Ops[0], 2057 llvm::ConstantInt::get(Ops[1]->getType(), 0)); 2058 case X86::BI__builtin_ia32_pslldi: 2059 case X86::BI__builtin_ia32_psllqi: 2060 case X86::BI__builtin_ia32_psllwi: 2061 case X86::BI__builtin_ia32_psradi: 2062 case X86::BI__builtin_ia32_psrawi: 2063 case X86::BI__builtin_ia32_psrldi: 2064 case X86::BI__builtin_ia32_psrlqi: 2065 case X86::BI__builtin_ia32_psrlwi: { 2066 Ops[1] = Builder.CreateZExt(Ops[1], Int64Ty, "zext"); 2067 llvm::Type *Ty = llvm::VectorType::get(Int64Ty, 1); 2068 Ops[1] = Builder.CreateBitCast(Ops[1], Ty, "bitcast"); 2069 const char *name = 0; 2070 Intrinsic::ID ID = Intrinsic::not_intrinsic; 2071 2072 switch (BuiltinID) { 2073 default: llvm_unreachable("Unsupported shift intrinsic!"); 2074 case X86::BI__builtin_ia32_pslldi: 2075 name = "pslldi"; 2076 ID = Intrinsic::x86_mmx_psll_d; 2077 break; 2078 case X86::BI__builtin_ia32_psllqi: 2079 name = "psllqi"; 2080 ID = Intrinsic::x86_mmx_psll_q; 2081 break; 2082 case X86::BI__builtin_ia32_psllwi: 2083 name = "psllwi"; 2084 ID = Intrinsic::x86_mmx_psll_w; 2085 break; 2086 case X86::BI__builtin_ia32_psradi: 2087 name = "psradi"; 2088 ID = Intrinsic::x86_mmx_psra_d; 2089 break; 2090 case X86::BI__builtin_ia32_psrawi: 2091 name = "psrawi"; 2092 ID = Intrinsic::x86_mmx_psra_w; 2093 break; 2094 case X86::BI__builtin_ia32_psrldi: 2095 name = "psrldi"; 2096 ID = Intrinsic::x86_mmx_psrl_d; 2097 break; 2098 case X86::BI__builtin_ia32_psrlqi: 2099 name = "psrlqi"; 2100 ID = Intrinsic::x86_mmx_psrl_q; 2101 break; 2102 case X86::BI__builtin_ia32_psrlwi: 2103 name = "psrlwi"; 2104 ID = Intrinsic::x86_mmx_psrl_w; 2105 break; 2106 } 2107 llvm::Function *F = CGM.getIntrinsic(ID); 2108 return Builder.CreateCall(F, Ops, name); 2109 } 2110 case X86::BI__builtin_ia32_cmpps: { 2111 llvm::Function *F = CGM.getIntrinsic(Intrinsic::x86_sse_cmp_ps); 2112 return Builder.CreateCall(F, Ops, "cmpps"); 2113 } 2114 case X86::BI__builtin_ia32_cmpss: { 2115 llvm::Function *F = CGM.getIntrinsic(Intrinsic::x86_sse_cmp_ss); 2116 return Builder.CreateCall(F, Ops, "cmpss"); 2117 } 2118 case X86::BI__builtin_ia32_ldmxcsr: { 2119 llvm::Type *PtrTy = Int8PtrTy; 2120 Value *One = llvm::ConstantInt::get(Int32Ty, 1); 2121 Value *Tmp = Builder.CreateAlloca(Int32Ty, One); 2122 Builder.CreateStore(Ops[0], Tmp); 2123 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::x86_sse_ldmxcsr), 2124 Builder.CreateBitCast(Tmp, PtrTy)); 2125 } 2126 case X86::BI__builtin_ia32_stmxcsr: { 2127 llvm::Type *PtrTy = Int8PtrTy; 2128 Value *One = llvm::ConstantInt::get(Int32Ty, 1); 2129 Value *Tmp = Builder.CreateAlloca(Int32Ty, One); 2130 Builder.CreateCall(CGM.getIntrinsic(Intrinsic::x86_sse_stmxcsr), 2131 Builder.CreateBitCast(Tmp, PtrTy)); 2132 return Builder.CreateLoad(Tmp, "stmxcsr"); 2133 } 2134 case X86::BI__builtin_ia32_cmppd: { 2135 llvm::Function *F = CGM.getIntrinsic(Intrinsic::x86_sse2_cmp_pd); 2136 return Builder.CreateCall(F, Ops, "cmppd"); 2137 } 2138 case X86::BI__builtin_ia32_cmpsd: { 2139 llvm::Function *F = CGM.getIntrinsic(Intrinsic::x86_sse2_cmp_sd); 2140 return Builder.CreateCall(F, Ops, "cmpsd"); 2141 } 2142 case X86::BI__builtin_ia32_storehps: 2143 case X86::BI__builtin_ia32_storelps: { 2144 llvm::Type *PtrTy = llvm::PointerType::getUnqual(Int64Ty); 2145 llvm::Type *VecTy = llvm::VectorType::get(Int64Ty, 2); 2146 2147 // cast val v2i64 2148 Ops[1] = Builder.CreateBitCast(Ops[1], VecTy, "cast"); 2149 2150 // extract (0, 1) 2151 unsigned Index = BuiltinID == X86::BI__builtin_ia32_storelps ? 0 : 1; 2152 llvm::Value *Idx = llvm::ConstantInt::get(Int32Ty, Index); 2153 Ops[1] = Builder.CreateExtractElement(Ops[1], Idx, "extract"); 2154 2155 // cast pointer to i64 & store 2156 Ops[0] = Builder.CreateBitCast(Ops[0], PtrTy); 2157 return Builder.CreateStore(Ops[1], Ops[0]); 2158 } 2159 case X86::BI__builtin_ia32_palignr: { 2160 unsigned shiftVal = cast<llvm::ConstantInt>(Ops[2])->getZExtValue(); 2161 2162 // If palignr is shifting the pair of input vectors less than 9 bytes, 2163 // emit a shuffle instruction. 2164 if (shiftVal <= 8) { 2165 SmallVector<llvm::Constant*, 8> Indices; 2166 for (unsigned i = 0; i != 8; ++i) 2167 Indices.push_back(llvm::ConstantInt::get(Int32Ty, shiftVal + i)); 2168 2169 Value* SV = llvm::ConstantVector::get(Indices); 2170 return Builder.CreateShuffleVector(Ops[1], Ops[0], SV, "palignr"); 2171 } 2172 2173 // If palignr is shifting the pair of input vectors more than 8 but less 2174 // than 16 bytes, emit a logical right shift of the destination. 2175 if (shiftVal < 16) { 2176 // MMX has these as 1 x i64 vectors for some odd optimization reasons. 2177 llvm::Type *VecTy = llvm::VectorType::get(Int64Ty, 1); 2178 2179 Ops[0] = Builder.CreateBitCast(Ops[0], VecTy, "cast"); 2180 Ops[1] = llvm::ConstantInt::get(VecTy, (shiftVal-8) * 8); 2181 2182 // create i32 constant 2183 llvm::Function *F = CGM.getIntrinsic(Intrinsic::x86_mmx_psrl_q); 2184 return Builder.CreateCall(F, makeArrayRef(&Ops[0], 2), "palignr"); 2185 } 2186 2187 // If palignr is shifting the pair of vectors more than 16 bytes, emit zero. 2188 return llvm::Constant::getNullValue(ConvertType(E->getType())); 2189 } 2190 case X86::BI__builtin_ia32_palignr128: { 2191 unsigned shiftVal = cast<llvm::ConstantInt>(Ops[2])->getZExtValue(); 2192 2193 // If palignr is shifting the pair of input vectors less than 17 bytes, 2194 // emit a shuffle instruction. 2195 if (shiftVal <= 16) { 2196 SmallVector<llvm::Constant*, 16> Indices; 2197 for (unsigned i = 0; i != 16; ++i) 2198 Indices.push_back(llvm::ConstantInt::get(Int32Ty, shiftVal + i)); 2199 2200 Value* SV = llvm::ConstantVector::get(Indices); 2201 return Builder.CreateShuffleVector(Ops[1], Ops[0], SV, "palignr"); 2202 } 2203 2204 // If palignr is shifting the pair of input vectors more than 16 but less 2205 // than 32 bytes, emit a logical right shift of the destination. 2206 if (shiftVal < 32) { 2207 llvm::Type *VecTy = llvm::VectorType::get(Int64Ty, 2); 2208 2209 Ops[0] = Builder.CreateBitCast(Ops[0], VecTy, "cast"); 2210 Ops[1] = llvm::ConstantInt::get(Int32Ty, (shiftVal-16) * 8); 2211 2212 // create i32 constant 2213 llvm::Function *F = CGM.getIntrinsic(Intrinsic::x86_sse2_psrl_dq); 2214 return Builder.CreateCall(F, makeArrayRef(&Ops[0], 2), "palignr"); 2215 } 2216 2217 // If palignr is shifting the pair of vectors more than 32 bytes, emit zero. 2218 return llvm::Constant::getNullValue(ConvertType(E->getType())); 2219 } 2220 case X86::BI__builtin_ia32_movntps: 2221 case X86::BI__builtin_ia32_movntpd: 2222 case X86::BI__builtin_ia32_movntdq: 2223 case X86::BI__builtin_ia32_movnti: { 2224 llvm::MDNode *Node = llvm::MDNode::get(getLLVMContext(), 2225 Builder.getInt32(1)); 2226 2227 // Convert the type of the pointer to a pointer to the stored type. 2228 Value *BC = Builder.CreateBitCast(Ops[0], 2229 llvm::PointerType::getUnqual(Ops[1]->getType()), 2230 "cast"); 2231 StoreInst *SI = Builder.CreateStore(Ops[1], BC); 2232 SI->setMetadata(CGM.getModule().getMDKindID("nontemporal"), Node); 2233 SI->setAlignment(16); 2234 return SI; 2235 } 2236 // 3DNow! 2237 case X86::BI__builtin_ia32_pavgusb: 2238 case X86::BI__builtin_ia32_pf2id: 2239 case X86::BI__builtin_ia32_pfacc: 2240 case X86::BI__builtin_ia32_pfadd: 2241 case X86::BI__builtin_ia32_pfcmpeq: 2242 case X86::BI__builtin_ia32_pfcmpge: 2243 case X86::BI__builtin_ia32_pfcmpgt: 2244 case X86::BI__builtin_ia32_pfmax: 2245 case X86::BI__builtin_ia32_pfmin: 2246 case X86::BI__builtin_ia32_pfmul: 2247 case X86::BI__builtin_ia32_pfrcp: 2248 case X86::BI__builtin_ia32_pfrcpit1: 2249 case X86::BI__builtin_ia32_pfrcpit2: 2250 case X86::BI__builtin_ia32_pfrsqrt: 2251 case X86::BI__builtin_ia32_pfrsqit1: 2252 case X86::BI__builtin_ia32_pfrsqrtit1: 2253 case X86::BI__builtin_ia32_pfsub: 2254 case X86::BI__builtin_ia32_pfsubr: 2255 case X86::BI__builtin_ia32_pi2fd: 2256 case X86::BI__builtin_ia32_pmulhrw: 2257 case X86::BI__builtin_ia32_pf2iw: 2258 case X86::BI__builtin_ia32_pfnacc: 2259 case X86::BI__builtin_ia32_pfpnacc: 2260 case X86::BI__builtin_ia32_pi2fw: 2261 case X86::BI__builtin_ia32_pswapdsf: 2262 case X86::BI__builtin_ia32_pswapdsi: { 2263 const char *name = 0; 2264 Intrinsic::ID ID = Intrinsic::not_intrinsic; 2265 switch(BuiltinID) { 2266 case X86::BI__builtin_ia32_pavgusb: 2267 name = "pavgusb"; 2268 ID = Intrinsic::x86_3dnow_pavgusb; 2269 break; 2270 case X86::BI__builtin_ia32_pf2id: 2271 name = "pf2id"; 2272 ID = Intrinsic::x86_3dnow_pf2id; 2273 break; 2274 case X86::BI__builtin_ia32_pfacc: 2275 name = "pfacc"; 2276 ID = Intrinsic::x86_3dnow_pfacc; 2277 break; 2278 case X86::BI__builtin_ia32_pfadd: 2279 name = "pfadd"; 2280 ID = Intrinsic::x86_3dnow_pfadd; 2281 break; 2282 case X86::BI__builtin_ia32_pfcmpeq: 2283 name = "pfcmpeq"; 2284 ID = Intrinsic::x86_3dnow_pfcmpeq; 2285 break; 2286 case X86::BI__builtin_ia32_pfcmpge: 2287 name = "pfcmpge"; 2288 ID = Intrinsic::x86_3dnow_pfcmpge; 2289 break; 2290 case X86::BI__builtin_ia32_pfcmpgt: 2291 name = "pfcmpgt"; 2292 ID = Intrinsic::x86_3dnow_pfcmpgt; 2293 break; 2294 case X86::BI__builtin_ia32_pfmax: 2295 name = "pfmax"; 2296 ID = Intrinsic::x86_3dnow_pfmax; 2297 break; 2298 case X86::BI__builtin_ia32_pfmin: 2299 name = "pfmin"; 2300 ID = Intrinsic::x86_3dnow_pfmin; 2301 break; 2302 case X86::BI__builtin_ia32_pfmul: 2303 name = "pfmul"; 2304 ID = Intrinsic::x86_3dnow_pfmul; 2305 break; 2306 case X86::BI__builtin_ia32_pfrcp: 2307 name = "pfrcp"; 2308 ID = Intrinsic::x86_3dnow_pfrcp; 2309 break; 2310 case X86::BI__builtin_ia32_pfrcpit1: 2311 name = "pfrcpit1"; 2312 ID = Intrinsic::x86_3dnow_pfrcpit1; 2313 break; 2314 case X86::BI__builtin_ia32_pfrcpit2: 2315 name = "pfrcpit2"; 2316 ID = Intrinsic::x86_3dnow_pfrcpit2; 2317 break; 2318 case X86::BI__builtin_ia32_pfrsqrt: 2319 name = "pfrsqrt"; 2320 ID = Intrinsic::x86_3dnow_pfrsqrt; 2321 break; 2322 case X86::BI__builtin_ia32_pfrsqit1: 2323 case X86::BI__builtin_ia32_pfrsqrtit1: 2324 name = "pfrsqit1"; 2325 ID = Intrinsic::x86_3dnow_pfrsqit1; 2326 break; 2327 case X86::BI__builtin_ia32_pfsub: 2328 name = "pfsub"; 2329 ID = Intrinsic::x86_3dnow_pfsub; 2330 break; 2331 case X86::BI__builtin_ia32_pfsubr: 2332 name = "pfsubr"; 2333 ID = Intrinsic::x86_3dnow_pfsubr; 2334 break; 2335 case X86::BI__builtin_ia32_pi2fd: 2336 name = "pi2fd"; 2337 ID = Intrinsic::x86_3dnow_pi2fd; 2338 break; 2339 case X86::BI__builtin_ia32_pmulhrw: 2340 name = "pmulhrw"; 2341 ID = Intrinsic::x86_3dnow_pmulhrw; 2342 break; 2343 case X86::BI__builtin_ia32_pf2iw: 2344 name = "pf2iw"; 2345 ID = Intrinsic::x86_3dnowa_pf2iw; 2346 break; 2347 case X86::BI__builtin_ia32_pfnacc: 2348 name = "pfnacc"; 2349 ID = Intrinsic::x86_3dnowa_pfnacc; 2350 break; 2351 case X86::BI__builtin_ia32_pfpnacc: 2352 name = "pfpnacc"; 2353 ID = Intrinsic::x86_3dnowa_pfpnacc; 2354 break; 2355 case X86::BI__builtin_ia32_pi2fw: 2356 name = "pi2fw"; 2357 ID = Intrinsic::x86_3dnowa_pi2fw; 2358 break; 2359 case X86::BI__builtin_ia32_pswapdsf: 2360 case X86::BI__builtin_ia32_pswapdsi: 2361 name = "pswapd"; 2362 ID = Intrinsic::x86_3dnowa_pswapd; 2363 break; 2364 } 2365 llvm::Function *F = CGM.getIntrinsic(ID); 2366 return Builder.CreateCall(F, Ops, name); 2367 } 2368 } 2369 } 2370 2371 Value *CodeGenFunction::EmitPPCBuiltinExpr(unsigned BuiltinID, 2372 const CallExpr *E) { 2373 SmallVector<Value*, 4> Ops; 2374 2375 for (unsigned i = 0, e = E->getNumArgs(); i != e; i++) 2376 Ops.push_back(EmitScalarExpr(E->getArg(i))); 2377 2378 Intrinsic::ID ID = Intrinsic::not_intrinsic; 2379 2380 switch (BuiltinID) { 2381 default: return 0; 2382 2383 // vec_ld, vec_lvsl, vec_lvsr 2384 case PPC::BI__builtin_altivec_lvx: 2385 case PPC::BI__builtin_altivec_lvxl: 2386 case PPC::BI__builtin_altivec_lvebx: 2387 case PPC::BI__builtin_altivec_lvehx: 2388 case PPC::BI__builtin_altivec_lvewx: 2389 case PPC::BI__builtin_altivec_lvsl: 2390 case PPC::BI__builtin_altivec_lvsr: 2391 { 2392 Ops[1] = Builder.CreateBitCast(Ops[1], Int8PtrTy); 2393 2394 Ops[0] = Builder.CreateGEP(Ops[1], Ops[0]); 2395 Ops.pop_back(); 2396 2397 switch (BuiltinID) { 2398 default: llvm_unreachable("Unsupported ld/lvsl/lvsr intrinsic!"); 2399 case PPC::BI__builtin_altivec_lvx: 2400 ID = Intrinsic::ppc_altivec_lvx; 2401 break; 2402 case PPC::BI__builtin_altivec_lvxl: 2403 ID = Intrinsic::ppc_altivec_lvxl; 2404 break; 2405 case PPC::BI__builtin_altivec_lvebx: 2406 ID = Intrinsic::ppc_altivec_lvebx; 2407 break; 2408 case PPC::BI__builtin_altivec_lvehx: 2409 ID = Intrinsic::ppc_altivec_lvehx; 2410 break; 2411 case PPC::BI__builtin_altivec_lvewx: 2412 ID = Intrinsic::ppc_altivec_lvewx; 2413 break; 2414 case PPC::BI__builtin_altivec_lvsl: 2415 ID = Intrinsic::ppc_altivec_lvsl; 2416 break; 2417 case PPC::BI__builtin_altivec_lvsr: 2418 ID = Intrinsic::ppc_altivec_lvsr; 2419 break; 2420 } 2421 llvm::Function *F = CGM.getIntrinsic(ID); 2422 return Builder.CreateCall(F, Ops, ""); 2423 } 2424 2425 // vec_st 2426 case PPC::BI__builtin_altivec_stvx: 2427 case PPC::BI__builtin_altivec_stvxl: 2428 case PPC::BI__builtin_altivec_stvebx: 2429 case PPC::BI__builtin_altivec_stvehx: 2430 case PPC::BI__builtin_altivec_stvewx: 2431 { 2432 Ops[2] = Builder.CreateBitCast(Ops[2], Int8PtrTy); 2433 Ops[1] = Builder.CreateGEP(Ops[2], Ops[1]); 2434 Ops.pop_back(); 2435 2436 switch (BuiltinID) { 2437 default: llvm_unreachable("Unsupported st intrinsic!"); 2438 case PPC::BI__builtin_altivec_stvx: 2439 ID = Intrinsic::ppc_altivec_stvx; 2440 break; 2441 case PPC::BI__builtin_altivec_stvxl: 2442 ID = Intrinsic::ppc_altivec_stvxl; 2443 break; 2444 case PPC::BI__builtin_altivec_stvebx: 2445 ID = Intrinsic::ppc_altivec_stvebx; 2446 break; 2447 case PPC::BI__builtin_altivec_stvehx: 2448 ID = Intrinsic::ppc_altivec_stvehx; 2449 break; 2450 case PPC::BI__builtin_altivec_stvewx: 2451 ID = Intrinsic::ppc_altivec_stvewx; 2452 break; 2453 } 2454 llvm::Function *F = CGM.getIntrinsic(ID); 2455 return Builder.CreateCall(F, Ops, ""); 2456 } 2457 } 2458 return 0; 2459 } 2460