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->EvaluateAsRValue(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 llvm::APSInt Size, DstSize; 552 if (!E->getArg(2)->EvaluateAsInt(Size, CGM.getContext()) || 553 !E->getArg(3)->EvaluateAsInt(DstSize, CGM.getContext())) 554 break; 555 if (Size.ugt(DstSize)) 556 break; 557 Value *Dest = EmitScalarExpr(E->getArg(0)); 558 Value *Src = EmitScalarExpr(E->getArg(1)); 559 Value *SizeVal = llvm::ConstantInt::get(Builder.getContext(), Size); 560 Builder.CreateMemCpy(Dest, Src, SizeVal, 1, false); 561 return RValue::get(Dest); 562 } 563 564 case Builtin::BI__builtin_objc_memmove_collectable: { 565 Value *Address = EmitScalarExpr(E->getArg(0)); 566 Value *SrcAddr = EmitScalarExpr(E->getArg(1)); 567 Value *SizeVal = EmitScalarExpr(E->getArg(2)); 568 CGM.getObjCRuntime().EmitGCMemmoveCollectable(*this, 569 Address, SrcAddr, SizeVal); 570 return RValue::get(Address); 571 } 572 573 case Builtin::BI__builtin___memmove_chk: { 574 // fold __builtin_memmove_chk(x, y, cst1, cst2) to memset iff cst1<=cst2. 575 llvm::APSInt Size, DstSize; 576 if (!E->getArg(2)->EvaluateAsInt(Size, CGM.getContext()) || 577 !E->getArg(3)->EvaluateAsInt(DstSize, CGM.getContext())) 578 break; 579 if (Size.ugt(DstSize)) 580 break; 581 Value *Dest = EmitScalarExpr(E->getArg(0)); 582 Value *Src = EmitScalarExpr(E->getArg(1)); 583 Value *SizeVal = llvm::ConstantInt::get(Builder.getContext(), Size); 584 Builder.CreateMemMove(Dest, Src, SizeVal, 1, false); 585 return RValue::get(Dest); 586 } 587 588 case Builtin::BImemmove: 589 case Builtin::BI__builtin_memmove: { 590 Value *Address = EmitScalarExpr(E->getArg(0)); 591 Value *SrcAddr = EmitScalarExpr(E->getArg(1)); 592 Value *SizeVal = EmitScalarExpr(E->getArg(2)); 593 Builder.CreateMemMove(Address, SrcAddr, SizeVal, 1, false); 594 return RValue::get(Address); 595 } 596 case Builtin::BImemset: 597 case Builtin::BI__builtin_memset: { 598 Value *Address = EmitScalarExpr(E->getArg(0)); 599 Value *ByteVal = Builder.CreateTrunc(EmitScalarExpr(E->getArg(1)), 600 Builder.getInt8Ty()); 601 Value *SizeVal = EmitScalarExpr(E->getArg(2)); 602 Builder.CreateMemSet(Address, ByteVal, SizeVal, 1, false); 603 return RValue::get(Address); 604 } 605 case Builtin::BI__builtin___memset_chk: { 606 // fold __builtin_memset_chk(x, y, cst1, cst2) to memset iff cst1<=cst2. 607 llvm::APSInt Size, DstSize; 608 if (!E->getArg(2)->EvaluateAsInt(Size, CGM.getContext()) || 609 !E->getArg(3)->EvaluateAsInt(DstSize, CGM.getContext())) 610 break; 611 if (Size.ugt(DstSize)) 612 break; 613 Value *Address = EmitScalarExpr(E->getArg(0)); 614 Value *ByteVal = Builder.CreateTrunc(EmitScalarExpr(E->getArg(1)), 615 Builder.getInt8Ty()); 616 Value *SizeVal = llvm::ConstantInt::get(Builder.getContext(), Size); 617 Builder.CreateMemSet(Address, ByteVal, SizeVal, 1, false); 618 619 return RValue::get(Address); 620 } 621 case Builtin::BI__builtin_dwarf_cfa: { 622 // The offset in bytes from the first argument to the CFA. 623 // 624 // Why on earth is this in the frontend? Is there any reason at 625 // all that the backend can't reasonably determine this while 626 // lowering llvm.eh.dwarf.cfa()? 627 // 628 // TODO: If there's a satisfactory reason, add a target hook for 629 // this instead of hard-coding 0, which is correct for most targets. 630 int32_t Offset = 0; 631 632 Value *F = CGM.getIntrinsic(Intrinsic::eh_dwarf_cfa); 633 return RValue::get(Builder.CreateCall(F, 634 llvm::ConstantInt::get(Int32Ty, Offset))); 635 } 636 case Builtin::BI__builtin_return_address: { 637 Value *Depth = EmitScalarExpr(E->getArg(0)); 638 Depth = Builder.CreateIntCast(Depth, Int32Ty, false); 639 Value *F = CGM.getIntrinsic(Intrinsic::returnaddress); 640 return RValue::get(Builder.CreateCall(F, Depth)); 641 } 642 case Builtin::BI__builtin_frame_address: { 643 Value *Depth = EmitScalarExpr(E->getArg(0)); 644 Depth = Builder.CreateIntCast(Depth, Int32Ty, false); 645 Value *F = CGM.getIntrinsic(Intrinsic::frameaddress); 646 return RValue::get(Builder.CreateCall(F, Depth)); 647 } 648 case Builtin::BI__builtin_extract_return_addr: { 649 Value *Address = EmitScalarExpr(E->getArg(0)); 650 Value *Result = getTargetHooks().decodeReturnAddress(*this, Address); 651 return RValue::get(Result); 652 } 653 case Builtin::BI__builtin_frob_return_addr: { 654 Value *Address = EmitScalarExpr(E->getArg(0)); 655 Value *Result = getTargetHooks().encodeReturnAddress(*this, Address); 656 return RValue::get(Result); 657 } 658 case Builtin::BI__builtin_dwarf_sp_column: { 659 llvm::IntegerType *Ty 660 = cast<llvm::IntegerType>(ConvertType(E->getType())); 661 int Column = getTargetHooks().getDwarfEHStackPointer(CGM); 662 if (Column == -1) { 663 CGM.ErrorUnsupported(E, "__builtin_dwarf_sp_column"); 664 return RValue::get(llvm::UndefValue::get(Ty)); 665 } 666 return RValue::get(llvm::ConstantInt::get(Ty, Column, true)); 667 } 668 case Builtin::BI__builtin_init_dwarf_reg_size_table: { 669 Value *Address = EmitScalarExpr(E->getArg(0)); 670 if (getTargetHooks().initDwarfEHRegSizeTable(*this, Address)) 671 CGM.ErrorUnsupported(E, "__builtin_init_dwarf_reg_size_table"); 672 return RValue::get(llvm::UndefValue::get(ConvertType(E->getType()))); 673 } 674 case Builtin::BI__builtin_eh_return: { 675 Value *Int = EmitScalarExpr(E->getArg(0)); 676 Value *Ptr = EmitScalarExpr(E->getArg(1)); 677 678 llvm::IntegerType *IntTy = cast<llvm::IntegerType>(Int->getType()); 679 assert((IntTy->getBitWidth() == 32 || IntTy->getBitWidth() == 64) && 680 "LLVM's __builtin_eh_return only supports 32- and 64-bit variants"); 681 Value *F = CGM.getIntrinsic(IntTy->getBitWidth() == 32 682 ? Intrinsic::eh_return_i32 683 : Intrinsic::eh_return_i64); 684 Builder.CreateCall2(F, Int, Ptr); 685 Builder.CreateUnreachable(); 686 687 // We do need to preserve an insertion point. 688 EmitBlock(createBasicBlock("builtin_eh_return.cont")); 689 690 return RValue::get(0); 691 } 692 case Builtin::BI__builtin_unwind_init: { 693 Value *F = CGM.getIntrinsic(Intrinsic::eh_unwind_init); 694 return RValue::get(Builder.CreateCall(F)); 695 } 696 case Builtin::BI__builtin_extend_pointer: { 697 // Extends a pointer to the size of an _Unwind_Word, which is 698 // uint64_t on all platforms. Generally this gets poked into a 699 // register and eventually used as an address, so if the 700 // addressing registers are wider than pointers and the platform 701 // doesn't implicitly ignore high-order bits when doing 702 // addressing, we need to make sure we zext / sext based on 703 // the platform's expectations. 704 // 705 // See: http://gcc.gnu.org/ml/gcc-bugs/2002-02/msg00237.html 706 707 // Cast the pointer to intptr_t. 708 Value *Ptr = EmitScalarExpr(E->getArg(0)); 709 Value *Result = Builder.CreatePtrToInt(Ptr, IntPtrTy, "extend.cast"); 710 711 // If that's 64 bits, we're done. 712 if (IntPtrTy->getBitWidth() == 64) 713 return RValue::get(Result); 714 715 // Otherwise, ask the codegen data what to do. 716 if (getTargetHooks().extendPointerWithSExt()) 717 return RValue::get(Builder.CreateSExt(Result, Int64Ty, "extend.sext")); 718 else 719 return RValue::get(Builder.CreateZExt(Result, Int64Ty, "extend.zext")); 720 } 721 case Builtin::BI__builtin_setjmp: { 722 // Buffer is a void**. 723 Value *Buf = EmitScalarExpr(E->getArg(0)); 724 725 // Store the frame pointer to the setjmp buffer. 726 Value *FrameAddr = 727 Builder.CreateCall(CGM.getIntrinsic(Intrinsic::frameaddress), 728 ConstantInt::get(Int32Ty, 0)); 729 Builder.CreateStore(FrameAddr, Buf); 730 731 // Store the stack pointer to the setjmp buffer. 732 Value *StackAddr = 733 Builder.CreateCall(CGM.getIntrinsic(Intrinsic::stacksave)); 734 Value *StackSaveSlot = 735 Builder.CreateGEP(Buf, ConstantInt::get(Int32Ty, 2)); 736 Builder.CreateStore(StackAddr, StackSaveSlot); 737 738 // Call LLVM's EH setjmp, which is lightweight. 739 Value *F = CGM.getIntrinsic(Intrinsic::eh_sjlj_setjmp); 740 Buf = Builder.CreateBitCast(Buf, Int8PtrTy); 741 return RValue::get(Builder.CreateCall(F, Buf)); 742 } 743 case Builtin::BI__builtin_longjmp: { 744 Value *Buf = EmitScalarExpr(E->getArg(0)); 745 Buf = Builder.CreateBitCast(Buf, Int8PtrTy); 746 747 // Call LLVM's EH longjmp, which is lightweight. 748 Builder.CreateCall(CGM.getIntrinsic(Intrinsic::eh_sjlj_longjmp), Buf); 749 750 // longjmp doesn't return; mark this as unreachable. 751 Builder.CreateUnreachable(); 752 753 // We do need to preserve an insertion point. 754 EmitBlock(createBasicBlock("longjmp.cont")); 755 756 return RValue::get(0); 757 } 758 case Builtin::BI__sync_fetch_and_add: 759 case Builtin::BI__sync_fetch_and_sub: 760 case Builtin::BI__sync_fetch_and_or: 761 case Builtin::BI__sync_fetch_and_and: 762 case Builtin::BI__sync_fetch_and_xor: 763 case Builtin::BI__sync_add_and_fetch: 764 case Builtin::BI__sync_sub_and_fetch: 765 case Builtin::BI__sync_and_and_fetch: 766 case Builtin::BI__sync_or_and_fetch: 767 case Builtin::BI__sync_xor_and_fetch: 768 case Builtin::BI__sync_val_compare_and_swap: 769 case Builtin::BI__sync_bool_compare_and_swap: 770 case Builtin::BI__sync_lock_test_and_set: 771 case Builtin::BI__sync_lock_release: 772 case Builtin::BI__sync_swap: 773 llvm_unreachable("Shouldn't make it through sema"); 774 case Builtin::BI__sync_fetch_and_add_1: 775 case Builtin::BI__sync_fetch_and_add_2: 776 case Builtin::BI__sync_fetch_and_add_4: 777 case Builtin::BI__sync_fetch_and_add_8: 778 case Builtin::BI__sync_fetch_and_add_16: 779 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Add, E); 780 case Builtin::BI__sync_fetch_and_sub_1: 781 case Builtin::BI__sync_fetch_and_sub_2: 782 case Builtin::BI__sync_fetch_and_sub_4: 783 case Builtin::BI__sync_fetch_and_sub_8: 784 case Builtin::BI__sync_fetch_and_sub_16: 785 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Sub, E); 786 case Builtin::BI__sync_fetch_and_or_1: 787 case Builtin::BI__sync_fetch_and_or_2: 788 case Builtin::BI__sync_fetch_and_or_4: 789 case Builtin::BI__sync_fetch_and_or_8: 790 case Builtin::BI__sync_fetch_and_or_16: 791 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Or, E); 792 case Builtin::BI__sync_fetch_and_and_1: 793 case Builtin::BI__sync_fetch_and_and_2: 794 case Builtin::BI__sync_fetch_and_and_4: 795 case Builtin::BI__sync_fetch_and_and_8: 796 case Builtin::BI__sync_fetch_and_and_16: 797 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::And, E); 798 case Builtin::BI__sync_fetch_and_xor_1: 799 case Builtin::BI__sync_fetch_and_xor_2: 800 case Builtin::BI__sync_fetch_and_xor_4: 801 case Builtin::BI__sync_fetch_and_xor_8: 802 case Builtin::BI__sync_fetch_and_xor_16: 803 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Xor, E); 804 805 // Clang extensions: not overloaded yet. 806 case Builtin::BI__sync_fetch_and_min: 807 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Min, E); 808 case Builtin::BI__sync_fetch_and_max: 809 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Max, E); 810 case Builtin::BI__sync_fetch_and_umin: 811 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::UMin, E); 812 case Builtin::BI__sync_fetch_and_umax: 813 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::UMax, E); 814 815 case Builtin::BI__sync_add_and_fetch_1: 816 case Builtin::BI__sync_add_and_fetch_2: 817 case Builtin::BI__sync_add_and_fetch_4: 818 case Builtin::BI__sync_add_and_fetch_8: 819 case Builtin::BI__sync_add_and_fetch_16: 820 return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::Add, E, 821 llvm::Instruction::Add); 822 case Builtin::BI__sync_sub_and_fetch_1: 823 case Builtin::BI__sync_sub_and_fetch_2: 824 case Builtin::BI__sync_sub_and_fetch_4: 825 case Builtin::BI__sync_sub_and_fetch_8: 826 case Builtin::BI__sync_sub_and_fetch_16: 827 return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::Sub, E, 828 llvm::Instruction::Sub); 829 case Builtin::BI__sync_and_and_fetch_1: 830 case Builtin::BI__sync_and_and_fetch_2: 831 case Builtin::BI__sync_and_and_fetch_4: 832 case Builtin::BI__sync_and_and_fetch_8: 833 case Builtin::BI__sync_and_and_fetch_16: 834 return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::And, E, 835 llvm::Instruction::And); 836 case Builtin::BI__sync_or_and_fetch_1: 837 case Builtin::BI__sync_or_and_fetch_2: 838 case Builtin::BI__sync_or_and_fetch_4: 839 case Builtin::BI__sync_or_and_fetch_8: 840 case Builtin::BI__sync_or_and_fetch_16: 841 return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::Or, E, 842 llvm::Instruction::Or); 843 case Builtin::BI__sync_xor_and_fetch_1: 844 case Builtin::BI__sync_xor_and_fetch_2: 845 case Builtin::BI__sync_xor_and_fetch_4: 846 case Builtin::BI__sync_xor_and_fetch_8: 847 case Builtin::BI__sync_xor_and_fetch_16: 848 return EmitBinaryAtomicPost(*this, llvm::AtomicRMWInst::Xor, E, 849 llvm::Instruction::Xor); 850 851 case Builtin::BI__sync_val_compare_and_swap_1: 852 case Builtin::BI__sync_val_compare_and_swap_2: 853 case Builtin::BI__sync_val_compare_and_swap_4: 854 case Builtin::BI__sync_val_compare_and_swap_8: 855 case Builtin::BI__sync_val_compare_and_swap_16: { 856 QualType T = E->getType(); 857 llvm::Value *DestPtr = EmitScalarExpr(E->getArg(0)); 858 unsigned AddrSpace = 859 cast<llvm::PointerType>(DestPtr->getType())->getAddressSpace(); 860 861 llvm::IntegerType *IntType = 862 llvm::IntegerType::get(getLLVMContext(), 863 getContext().getTypeSize(T)); 864 llvm::Type *IntPtrType = IntType->getPointerTo(AddrSpace); 865 866 Value *Args[3]; 867 Args[0] = Builder.CreateBitCast(DestPtr, IntPtrType); 868 Args[1] = EmitScalarExpr(E->getArg(1)); 869 llvm::Type *ValueType = Args[1]->getType(); 870 Args[1] = EmitToInt(*this, Args[1], T, IntType); 871 Args[2] = EmitToInt(*this, EmitScalarExpr(E->getArg(2)), T, IntType); 872 873 Value *Result = Builder.CreateAtomicCmpXchg(Args[0], Args[1], Args[2], 874 llvm::SequentiallyConsistent); 875 Result = EmitFromInt(*this, Result, T, ValueType); 876 return RValue::get(Result); 877 } 878 879 case Builtin::BI__sync_bool_compare_and_swap_1: 880 case Builtin::BI__sync_bool_compare_and_swap_2: 881 case Builtin::BI__sync_bool_compare_and_swap_4: 882 case Builtin::BI__sync_bool_compare_and_swap_8: 883 case Builtin::BI__sync_bool_compare_and_swap_16: { 884 QualType T = E->getArg(1)->getType(); 885 llvm::Value *DestPtr = EmitScalarExpr(E->getArg(0)); 886 unsigned AddrSpace = 887 cast<llvm::PointerType>(DestPtr->getType())->getAddressSpace(); 888 889 llvm::IntegerType *IntType = 890 llvm::IntegerType::get(getLLVMContext(), 891 getContext().getTypeSize(T)); 892 llvm::Type *IntPtrType = IntType->getPointerTo(AddrSpace); 893 894 Value *Args[3]; 895 Args[0] = Builder.CreateBitCast(DestPtr, IntPtrType); 896 Args[1] = EmitToInt(*this, EmitScalarExpr(E->getArg(1)), T, IntType); 897 Args[2] = EmitToInt(*this, EmitScalarExpr(E->getArg(2)), T, IntType); 898 899 Value *OldVal = Args[1]; 900 Value *PrevVal = Builder.CreateAtomicCmpXchg(Args[0], Args[1], Args[2], 901 llvm::SequentiallyConsistent); 902 Value *Result = Builder.CreateICmpEQ(PrevVal, OldVal); 903 // zext bool to int. 904 Result = Builder.CreateZExt(Result, ConvertType(E->getType())); 905 return RValue::get(Result); 906 } 907 908 case Builtin::BI__sync_swap_1: 909 case Builtin::BI__sync_swap_2: 910 case Builtin::BI__sync_swap_4: 911 case Builtin::BI__sync_swap_8: 912 case Builtin::BI__sync_swap_16: 913 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Xchg, E); 914 915 case Builtin::BI__sync_lock_test_and_set_1: 916 case Builtin::BI__sync_lock_test_and_set_2: 917 case Builtin::BI__sync_lock_test_and_set_4: 918 case Builtin::BI__sync_lock_test_and_set_8: 919 case Builtin::BI__sync_lock_test_and_set_16: 920 return EmitBinaryAtomic(*this, llvm::AtomicRMWInst::Xchg, E); 921 922 case Builtin::BI__sync_lock_release_1: 923 case Builtin::BI__sync_lock_release_2: 924 case Builtin::BI__sync_lock_release_4: 925 case Builtin::BI__sync_lock_release_8: 926 case Builtin::BI__sync_lock_release_16: { 927 Value *Ptr = EmitScalarExpr(E->getArg(0)); 928 llvm::Type *ElLLVMTy = 929 cast<llvm::PointerType>(Ptr->getType())->getElementType(); 930 llvm::StoreInst *Store = 931 Builder.CreateStore(llvm::Constant::getNullValue(ElLLVMTy), Ptr); 932 QualType ElTy = E->getArg(0)->getType()->getPointeeType(); 933 CharUnits StoreSize = getContext().getTypeSizeInChars(ElTy); 934 Store->setAlignment(StoreSize.getQuantity()); 935 Store->setAtomic(llvm::Release); 936 return RValue::get(0); 937 } 938 939 case Builtin::BI__sync_synchronize: { 940 // We assume this is supposed to correspond to a C++0x-style 941 // sequentially-consistent fence (i.e. this is only usable for 942 // synchonization, not device I/O or anything like that). This intrinsic 943 // is really badly designed in the sense that in theory, there isn't 944 // any way to safely use it... but in practice, it mostly works 945 // to use it with non-atomic loads and stores to get acquire/release 946 // semantics. 947 Builder.CreateFence(llvm::SequentiallyConsistent); 948 return RValue::get(0); 949 } 950 951 case Builtin::BI__atomic_thread_fence: 952 case Builtin::BI__atomic_signal_fence: { 953 llvm::SynchronizationScope Scope; 954 if (BuiltinID == Builtin::BI__atomic_signal_fence) 955 Scope = llvm::SingleThread; 956 else 957 Scope = llvm::CrossThread; 958 Value *Order = EmitScalarExpr(E->getArg(0)); 959 if (isa<llvm::ConstantInt>(Order)) { 960 int ord = cast<llvm::ConstantInt>(Order)->getZExtValue(); 961 switch (ord) { 962 case 0: // memory_order_relaxed 963 default: // invalid order 964 break; 965 case 1: // memory_order_consume 966 case 2: // memory_order_acquire 967 Builder.CreateFence(llvm::Acquire, Scope); 968 break; 969 case 3: // memory_order_release 970 Builder.CreateFence(llvm::Release, Scope); 971 break; 972 case 4: // memory_order_acq_rel 973 Builder.CreateFence(llvm::AcquireRelease, Scope); 974 break; 975 case 5: // memory_order_seq_cst 976 Builder.CreateFence(llvm::SequentiallyConsistent, Scope); 977 break; 978 } 979 return RValue::get(0); 980 } 981 982 llvm::BasicBlock *AcquireBB, *ReleaseBB, *AcqRelBB, *SeqCstBB; 983 AcquireBB = createBasicBlock("acquire", CurFn); 984 ReleaseBB = createBasicBlock("release", CurFn); 985 AcqRelBB = createBasicBlock("acqrel", CurFn); 986 SeqCstBB = createBasicBlock("seqcst", CurFn); 987 llvm::BasicBlock *ContBB = createBasicBlock("atomic.continue", CurFn); 988 989 Order = Builder.CreateIntCast(Order, Builder.getInt32Ty(), false); 990 llvm::SwitchInst *SI = Builder.CreateSwitch(Order, ContBB); 991 992 Builder.SetInsertPoint(AcquireBB); 993 Builder.CreateFence(llvm::Acquire, Scope); 994 Builder.CreateBr(ContBB); 995 SI->addCase(Builder.getInt32(1), AcquireBB); 996 SI->addCase(Builder.getInt32(2), AcquireBB); 997 998 Builder.SetInsertPoint(ReleaseBB); 999 Builder.CreateFence(llvm::Release, Scope); 1000 Builder.CreateBr(ContBB); 1001 SI->addCase(Builder.getInt32(3), ReleaseBB); 1002 1003 Builder.SetInsertPoint(AcqRelBB); 1004 Builder.CreateFence(llvm::AcquireRelease, Scope); 1005 Builder.CreateBr(ContBB); 1006 SI->addCase(Builder.getInt32(4), AcqRelBB); 1007 1008 Builder.SetInsertPoint(SeqCstBB); 1009 Builder.CreateFence(llvm::SequentiallyConsistent, Scope); 1010 Builder.CreateBr(ContBB); 1011 SI->addCase(Builder.getInt32(5), SeqCstBB); 1012 1013 Builder.SetInsertPoint(ContBB); 1014 return RValue::get(0); 1015 } 1016 1017 // Library functions with special handling. 1018 case Builtin::BIsqrt: 1019 case Builtin::BIsqrtf: 1020 case Builtin::BIsqrtl: { 1021 // TODO: there is currently no set of optimizer flags 1022 // sufficient for us to rewrite sqrt to @llvm.sqrt. 1023 // -fmath-errno=0 is not good enough; we need finiteness. 1024 // We could probably precondition the call with an ult 1025 // against 0, but is that worth the complexity? 1026 break; 1027 } 1028 1029 case Builtin::BIpow: 1030 case Builtin::BIpowf: 1031 case Builtin::BIpowl: { 1032 // Rewrite sqrt to intrinsic if allowed. 1033 if (!FD->hasAttr<ConstAttr>()) 1034 break; 1035 Value *Base = EmitScalarExpr(E->getArg(0)); 1036 Value *Exponent = EmitScalarExpr(E->getArg(1)); 1037 llvm::Type *ArgType = Base->getType(); 1038 Value *F = CGM.getIntrinsic(Intrinsic::pow, ArgType); 1039 return RValue::get(Builder.CreateCall2(F, Base, Exponent)); 1040 } 1041 1042 case Builtin::BIfma: 1043 case Builtin::BIfmaf: 1044 case Builtin::BIfmal: 1045 case Builtin::BI__builtin_fma: 1046 case Builtin::BI__builtin_fmaf: 1047 case Builtin::BI__builtin_fmal: { 1048 // Rewrite fma to intrinsic. 1049 Value *FirstArg = EmitScalarExpr(E->getArg(0)); 1050 llvm::Type *ArgType = FirstArg->getType(); 1051 Value *F = CGM.getIntrinsic(Intrinsic::fma, ArgType); 1052 return RValue::get(Builder.CreateCall3(F, FirstArg, 1053 EmitScalarExpr(E->getArg(1)), 1054 EmitScalarExpr(E->getArg(2)))); 1055 } 1056 1057 case Builtin::BI__builtin_signbit: 1058 case Builtin::BI__builtin_signbitf: 1059 case Builtin::BI__builtin_signbitl: { 1060 LLVMContext &C = CGM.getLLVMContext(); 1061 1062 Value *Arg = EmitScalarExpr(E->getArg(0)); 1063 llvm::Type *ArgTy = Arg->getType(); 1064 if (ArgTy->isPPC_FP128Ty()) 1065 break; // FIXME: I'm not sure what the right implementation is here. 1066 int ArgWidth = ArgTy->getPrimitiveSizeInBits(); 1067 llvm::Type *ArgIntTy = llvm::IntegerType::get(C, ArgWidth); 1068 Value *BCArg = Builder.CreateBitCast(Arg, ArgIntTy); 1069 Value *ZeroCmp = llvm::Constant::getNullValue(ArgIntTy); 1070 Value *Result = Builder.CreateICmpSLT(BCArg, ZeroCmp); 1071 return RValue::get(Builder.CreateZExt(Result, ConvertType(E->getType()))); 1072 } 1073 case Builtin::BI__builtin_annotation: { 1074 llvm::Value *AnnVal = EmitScalarExpr(E->getArg(0)); 1075 llvm::Value *F = CGM.getIntrinsic(llvm::Intrinsic::annotation, 1076 AnnVal->getType()); 1077 1078 // Get the annotation string, go through casts. Sema requires this to be a 1079 // non-wide string literal, potentially casted, so the cast<> is safe. 1080 const Expr *AnnotationStrExpr = E->getArg(1)->IgnoreParenCasts(); 1081 llvm::StringRef Str = cast<StringLiteral>(AnnotationStrExpr)->getString(); 1082 return RValue::get(EmitAnnotationCall(F, AnnVal, Str, E->getExprLoc())); 1083 } 1084 } 1085 1086 // If this is an alias for a lib function (e.g. __builtin_sin), emit 1087 // the call using the normal call path, but using the unmangled 1088 // version of the function name. 1089 if (getContext().BuiltinInfo.isLibFunction(BuiltinID)) 1090 return emitLibraryCall(*this, FD, E, 1091 CGM.getBuiltinLibFunction(FD, BuiltinID)); 1092 1093 // If this is a predefined lib function (e.g. malloc), emit the call 1094 // using exactly the normal call path. 1095 if (getContext().BuiltinInfo.isPredefinedLibFunction(BuiltinID)) 1096 return emitLibraryCall(*this, FD, E, EmitScalarExpr(E->getCallee())); 1097 1098 // See if we have a target specific intrinsic. 1099 const char *Name = getContext().BuiltinInfo.GetName(BuiltinID); 1100 Intrinsic::ID IntrinsicID = Intrinsic::not_intrinsic; 1101 if (const char *Prefix = 1102 llvm::Triple::getArchTypePrefix(Target.getTriple().getArch())) 1103 IntrinsicID = Intrinsic::getIntrinsicForGCCBuiltin(Prefix, Name); 1104 1105 if (IntrinsicID != Intrinsic::not_intrinsic) { 1106 SmallVector<Value*, 16> Args; 1107 1108 // Find out if any arguments are required to be integer constant 1109 // expressions. 1110 unsigned ICEArguments = 0; 1111 ASTContext::GetBuiltinTypeError Error; 1112 getContext().GetBuiltinType(BuiltinID, Error, &ICEArguments); 1113 assert(Error == ASTContext::GE_None && "Should not codegen an error"); 1114 1115 Function *F = CGM.getIntrinsic(IntrinsicID); 1116 llvm::FunctionType *FTy = F->getFunctionType(); 1117 1118 for (unsigned i = 0, e = E->getNumArgs(); i != e; ++i) { 1119 Value *ArgValue; 1120 // If this is a normal argument, just emit it as a scalar. 1121 if ((ICEArguments & (1 << i)) == 0) { 1122 ArgValue = EmitScalarExpr(E->getArg(i)); 1123 } else { 1124 // If this is required to be a constant, constant fold it so that we 1125 // know that the generated intrinsic gets a ConstantInt. 1126 llvm::APSInt Result; 1127 bool IsConst = E->getArg(i)->isIntegerConstantExpr(Result,getContext()); 1128 assert(IsConst && "Constant arg isn't actually constant?"); 1129 (void)IsConst; 1130 ArgValue = llvm::ConstantInt::get(getLLVMContext(), Result); 1131 } 1132 1133 // If the intrinsic arg type is different from the builtin arg type 1134 // we need to do a bit cast. 1135 llvm::Type *PTy = FTy->getParamType(i); 1136 if (PTy != ArgValue->getType()) { 1137 assert(PTy->canLosslesslyBitCastTo(FTy->getParamType(i)) && 1138 "Must be able to losslessly bit cast to param"); 1139 ArgValue = Builder.CreateBitCast(ArgValue, PTy); 1140 } 1141 1142 Args.push_back(ArgValue); 1143 } 1144 1145 Value *V = Builder.CreateCall(F, Args); 1146 QualType BuiltinRetType = E->getType(); 1147 1148 llvm::Type *RetTy = llvm::Type::getVoidTy(getLLVMContext()); 1149 if (!BuiltinRetType->isVoidType()) RetTy = ConvertType(BuiltinRetType); 1150 1151 if (RetTy != V->getType()) { 1152 assert(V->getType()->canLosslesslyBitCastTo(RetTy) && 1153 "Must be able to losslessly bit cast result type"); 1154 V = Builder.CreateBitCast(V, RetTy); 1155 } 1156 1157 return RValue::get(V); 1158 } 1159 1160 // See if we have a target specific builtin that needs to be lowered. 1161 if (Value *V = EmitTargetBuiltinExpr(BuiltinID, E)) 1162 return RValue::get(V); 1163 1164 ErrorUnsupported(E, "builtin function"); 1165 1166 // Unknown builtin, for now just dump it out and return undef. 1167 if (hasAggregateLLVMType(E->getType())) 1168 return RValue::getAggregate(CreateMemTemp(E->getType())); 1169 return RValue::get(llvm::UndefValue::get(ConvertType(E->getType()))); 1170 } 1171 1172 Value *CodeGenFunction::EmitTargetBuiltinExpr(unsigned BuiltinID, 1173 const CallExpr *E) { 1174 switch (Target.getTriple().getArch()) { 1175 case llvm::Triple::arm: 1176 case llvm::Triple::thumb: 1177 return EmitARMBuiltinExpr(BuiltinID, E); 1178 case llvm::Triple::x86: 1179 case llvm::Triple::x86_64: 1180 return EmitX86BuiltinExpr(BuiltinID, E); 1181 case llvm::Triple::ppc: 1182 case llvm::Triple::ppc64: 1183 return EmitPPCBuiltinExpr(BuiltinID, E); 1184 default: 1185 return 0; 1186 } 1187 } 1188 1189 static llvm::VectorType *GetNeonType(LLVMContext &C, NeonTypeFlags TypeFlags) { 1190 int IsQuad = TypeFlags.isQuad(); 1191 switch (TypeFlags.getEltType()) { 1192 default: break; 1193 case NeonTypeFlags::Int8: 1194 case NeonTypeFlags::Poly8: 1195 return llvm::VectorType::get(llvm::Type::getInt8Ty(C), 8 << IsQuad); 1196 case NeonTypeFlags::Int16: 1197 case NeonTypeFlags::Poly16: 1198 case NeonTypeFlags::Float16: 1199 return llvm::VectorType::get(llvm::Type::getInt16Ty(C), 4 << IsQuad); 1200 case NeonTypeFlags::Int32: 1201 return llvm::VectorType::get(llvm::Type::getInt32Ty(C), 2 << IsQuad); 1202 case NeonTypeFlags::Int64: 1203 return llvm::VectorType::get(llvm::Type::getInt64Ty(C), 1 << IsQuad); 1204 case NeonTypeFlags::Float32: 1205 return llvm::VectorType::get(llvm::Type::getFloatTy(C), 2 << IsQuad); 1206 }; 1207 return 0; 1208 } 1209 1210 Value *CodeGenFunction::EmitNeonSplat(Value *V, Constant *C) { 1211 unsigned nElts = cast<llvm::VectorType>(V->getType())->getNumElements(); 1212 SmallVector<Constant*, 16> Indices(nElts, C); 1213 Value* SV = llvm::ConstantVector::get(Indices); 1214 return Builder.CreateShuffleVector(V, V, SV, "lane"); 1215 } 1216 1217 Value *CodeGenFunction::EmitNeonCall(Function *F, SmallVectorImpl<Value*> &Ops, 1218 const char *name, 1219 unsigned shift, bool rightshift) { 1220 unsigned j = 0; 1221 for (Function::const_arg_iterator ai = F->arg_begin(), ae = F->arg_end(); 1222 ai != ae; ++ai, ++j) 1223 if (shift > 0 && shift == j) 1224 Ops[j] = EmitNeonShiftVector(Ops[j], ai->getType(), rightshift); 1225 else 1226 Ops[j] = Builder.CreateBitCast(Ops[j], ai->getType(), name); 1227 1228 return Builder.CreateCall(F, Ops, name); 1229 } 1230 1231 Value *CodeGenFunction::EmitNeonShiftVector(Value *V, llvm::Type *Ty, 1232 bool neg) { 1233 ConstantInt *CI = cast<ConstantInt>(V); 1234 int SV = CI->getSExtValue(); 1235 1236 llvm::VectorType *VTy = cast<llvm::VectorType>(Ty); 1237 llvm::Constant *C = ConstantInt::get(VTy->getElementType(), neg ? -SV : SV); 1238 SmallVector<llvm::Constant*, 16> CV(VTy->getNumElements(), C); 1239 return llvm::ConstantVector::get(CV); 1240 } 1241 1242 /// GetPointeeAlignment - Given an expression with a pointer type, find the 1243 /// alignment of the type referenced by the pointer. Skip over implicit 1244 /// casts. 1245 static Value *GetPointeeAlignment(CodeGenFunction &CGF, const Expr *Addr) { 1246 unsigned Align = 1; 1247 // Check if the type is a pointer. The implicit cast operand might not be. 1248 while (Addr->getType()->isPointerType()) { 1249 QualType PtTy = Addr->getType()->getPointeeType(); 1250 unsigned NewA = CGF.getContext().getTypeAlignInChars(PtTy).getQuantity(); 1251 if (NewA > Align) 1252 Align = NewA; 1253 1254 // If the address is an implicit cast, repeat with the cast operand. 1255 if (const ImplicitCastExpr *CastAddr = dyn_cast<ImplicitCastExpr>(Addr)) { 1256 Addr = CastAddr->getSubExpr(); 1257 continue; 1258 } 1259 break; 1260 } 1261 return llvm::ConstantInt::get(CGF.Int32Ty, Align); 1262 } 1263 1264 Value *CodeGenFunction::EmitARMBuiltinExpr(unsigned BuiltinID, 1265 const CallExpr *E) { 1266 if (BuiltinID == ARM::BI__clear_cache) { 1267 const FunctionDecl *FD = E->getDirectCallee(); 1268 // Oddly people write this call without args on occasion and gcc accepts 1269 // it - it's also marked as varargs in the description file. 1270 SmallVector<Value*, 2> Ops; 1271 for (unsigned i = 0; i < E->getNumArgs(); i++) 1272 Ops.push_back(EmitScalarExpr(E->getArg(i))); 1273 llvm::Type *Ty = CGM.getTypes().ConvertType(FD->getType()); 1274 llvm::FunctionType *FTy = cast<llvm::FunctionType>(Ty); 1275 StringRef Name = FD->getName(); 1276 return Builder.CreateCall(CGM.CreateRuntimeFunction(FTy, Name), Ops); 1277 } 1278 1279 if (BuiltinID == ARM::BI__builtin_arm_ldrexd) { 1280 Function *F = CGM.getIntrinsic(Intrinsic::arm_ldrexd); 1281 1282 Value *LdPtr = EmitScalarExpr(E->getArg(0)); 1283 Value *Val = Builder.CreateCall(F, LdPtr, "ldrexd"); 1284 1285 Value *Val0 = Builder.CreateExtractValue(Val, 1); 1286 Value *Val1 = Builder.CreateExtractValue(Val, 0); 1287 Val0 = Builder.CreateZExt(Val0, Int64Ty); 1288 Val1 = Builder.CreateZExt(Val1, Int64Ty); 1289 1290 Value *ShiftCst = llvm::ConstantInt::get(Int64Ty, 32); 1291 Val = Builder.CreateShl(Val0, ShiftCst, "shl", true /* nuw */); 1292 return Builder.CreateOr(Val, Val1); 1293 } 1294 1295 if (BuiltinID == ARM::BI__builtin_arm_strexd) { 1296 Function *F = CGM.getIntrinsic(Intrinsic::arm_strexd); 1297 llvm::Type *STy = llvm::StructType::get(Int32Ty, Int32Ty, NULL); 1298 1299 Value *One = llvm::ConstantInt::get(Int32Ty, 1); 1300 Value *Tmp = Builder.CreateAlloca(Int64Ty, One); 1301 Value *Val = EmitScalarExpr(E->getArg(0)); 1302 Builder.CreateStore(Val, Tmp); 1303 1304 Value *LdPtr = Builder.CreateBitCast(Tmp,llvm::PointerType::getUnqual(STy)); 1305 Val = Builder.CreateLoad(LdPtr); 1306 1307 Value *Arg0 = Builder.CreateExtractValue(Val, 0); 1308 Value *Arg1 = Builder.CreateExtractValue(Val, 1); 1309 Value *StPtr = EmitScalarExpr(E->getArg(1)); 1310 return Builder.CreateCall3(F, Arg0, Arg1, StPtr, "strexd"); 1311 } 1312 1313 SmallVector<Value*, 4> Ops; 1314 for (unsigned i = 0, e = E->getNumArgs() - 1; i != e; i++) 1315 Ops.push_back(EmitScalarExpr(E->getArg(i))); 1316 1317 // vget_lane and vset_lane are not overloaded and do not have an extra 1318 // argument that specifies the vector type. 1319 switch (BuiltinID) { 1320 default: break; 1321 case ARM::BI__builtin_neon_vget_lane_i8: 1322 case ARM::BI__builtin_neon_vget_lane_i16: 1323 case ARM::BI__builtin_neon_vget_lane_i32: 1324 case ARM::BI__builtin_neon_vget_lane_i64: 1325 case ARM::BI__builtin_neon_vget_lane_f32: 1326 case ARM::BI__builtin_neon_vgetq_lane_i8: 1327 case ARM::BI__builtin_neon_vgetq_lane_i16: 1328 case ARM::BI__builtin_neon_vgetq_lane_i32: 1329 case ARM::BI__builtin_neon_vgetq_lane_i64: 1330 case ARM::BI__builtin_neon_vgetq_lane_f32: 1331 return Builder.CreateExtractElement(Ops[0], EmitScalarExpr(E->getArg(1)), 1332 "vget_lane"); 1333 case ARM::BI__builtin_neon_vset_lane_i8: 1334 case ARM::BI__builtin_neon_vset_lane_i16: 1335 case ARM::BI__builtin_neon_vset_lane_i32: 1336 case ARM::BI__builtin_neon_vset_lane_i64: 1337 case ARM::BI__builtin_neon_vset_lane_f32: 1338 case ARM::BI__builtin_neon_vsetq_lane_i8: 1339 case ARM::BI__builtin_neon_vsetq_lane_i16: 1340 case ARM::BI__builtin_neon_vsetq_lane_i32: 1341 case ARM::BI__builtin_neon_vsetq_lane_i64: 1342 case ARM::BI__builtin_neon_vsetq_lane_f32: 1343 Ops.push_back(EmitScalarExpr(E->getArg(2))); 1344 return Builder.CreateInsertElement(Ops[1], Ops[0], Ops[2], "vset_lane"); 1345 } 1346 1347 // Get the last argument, which specifies the vector type. 1348 llvm::APSInt Result; 1349 const Expr *Arg = E->getArg(E->getNumArgs()-1); 1350 if (!Arg->isIntegerConstantExpr(Result, getContext())) 1351 return 0; 1352 1353 if (BuiltinID == ARM::BI__builtin_arm_vcvtr_f || 1354 BuiltinID == ARM::BI__builtin_arm_vcvtr_d) { 1355 // Determine the overloaded type of this builtin. 1356 llvm::Type *Ty; 1357 if (BuiltinID == ARM::BI__builtin_arm_vcvtr_f) 1358 Ty = llvm::Type::getFloatTy(getLLVMContext()); 1359 else 1360 Ty = llvm::Type::getDoubleTy(getLLVMContext()); 1361 1362 // Determine whether this is an unsigned conversion or not. 1363 bool usgn = Result.getZExtValue() == 1; 1364 unsigned Int = usgn ? Intrinsic::arm_vcvtru : Intrinsic::arm_vcvtr; 1365 1366 // Call the appropriate intrinsic. 1367 Function *F = CGM.getIntrinsic(Int, Ty); 1368 return Builder.CreateCall(F, Ops, "vcvtr"); 1369 } 1370 1371 // Determine the type of this overloaded NEON intrinsic. 1372 NeonTypeFlags Type(Result.getZExtValue()); 1373 bool usgn = Type.isUnsigned(); 1374 bool quad = Type.isQuad(); 1375 bool rightShift = false; 1376 1377 llvm::VectorType *VTy = GetNeonType(getLLVMContext(), Type); 1378 llvm::Type *Ty = VTy; 1379 if (!Ty) 1380 return 0; 1381 1382 unsigned Int; 1383 switch (BuiltinID) { 1384 default: return 0; 1385 case ARM::BI__builtin_neon_vabd_v: 1386 case ARM::BI__builtin_neon_vabdq_v: 1387 Int = usgn ? Intrinsic::arm_neon_vabdu : Intrinsic::arm_neon_vabds; 1388 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vabd"); 1389 case ARM::BI__builtin_neon_vabs_v: 1390 case ARM::BI__builtin_neon_vabsq_v: 1391 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vabs, Ty), 1392 Ops, "vabs"); 1393 case ARM::BI__builtin_neon_vaddhn_v: 1394 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vaddhn, Ty), 1395 Ops, "vaddhn"); 1396 case ARM::BI__builtin_neon_vcale_v: 1397 std::swap(Ops[0], Ops[1]); 1398 case ARM::BI__builtin_neon_vcage_v: { 1399 Function *F = CGM.getIntrinsic(Intrinsic::arm_neon_vacged); 1400 return EmitNeonCall(F, Ops, "vcage"); 1401 } 1402 case ARM::BI__builtin_neon_vcaleq_v: 1403 std::swap(Ops[0], Ops[1]); 1404 case ARM::BI__builtin_neon_vcageq_v: { 1405 Function *F = CGM.getIntrinsic(Intrinsic::arm_neon_vacgeq); 1406 return EmitNeonCall(F, Ops, "vcage"); 1407 } 1408 case ARM::BI__builtin_neon_vcalt_v: 1409 std::swap(Ops[0], Ops[1]); 1410 case ARM::BI__builtin_neon_vcagt_v: { 1411 Function *F = CGM.getIntrinsic(Intrinsic::arm_neon_vacgtd); 1412 return EmitNeonCall(F, Ops, "vcagt"); 1413 } 1414 case ARM::BI__builtin_neon_vcaltq_v: 1415 std::swap(Ops[0], Ops[1]); 1416 case ARM::BI__builtin_neon_vcagtq_v: { 1417 Function *F = CGM.getIntrinsic(Intrinsic::arm_neon_vacgtq); 1418 return EmitNeonCall(F, Ops, "vcagt"); 1419 } 1420 case ARM::BI__builtin_neon_vcls_v: 1421 case ARM::BI__builtin_neon_vclsq_v: { 1422 Function *F = CGM.getIntrinsic(Intrinsic::arm_neon_vcls, Ty); 1423 return EmitNeonCall(F, Ops, "vcls"); 1424 } 1425 case ARM::BI__builtin_neon_vclz_v: 1426 case ARM::BI__builtin_neon_vclzq_v: { 1427 Function *F = CGM.getIntrinsic(Intrinsic::arm_neon_vclz, Ty); 1428 return EmitNeonCall(F, Ops, "vclz"); 1429 } 1430 case ARM::BI__builtin_neon_vcnt_v: 1431 case ARM::BI__builtin_neon_vcntq_v: { 1432 Function *F = CGM.getIntrinsic(Intrinsic::arm_neon_vcnt, Ty); 1433 return EmitNeonCall(F, Ops, "vcnt"); 1434 } 1435 case ARM::BI__builtin_neon_vcvt_f16_v: { 1436 assert(Type.getEltType() == NeonTypeFlags::Float16 && !quad && 1437 "unexpected vcvt_f16_v builtin"); 1438 Function *F = CGM.getIntrinsic(Intrinsic::arm_neon_vcvtfp2hf); 1439 return EmitNeonCall(F, Ops, "vcvt"); 1440 } 1441 case ARM::BI__builtin_neon_vcvt_f32_f16: { 1442 assert(Type.getEltType() == NeonTypeFlags::Float16 && !quad && 1443 "unexpected vcvt_f32_f16 builtin"); 1444 Function *F = CGM.getIntrinsic(Intrinsic::arm_neon_vcvthf2fp); 1445 return EmitNeonCall(F, Ops, "vcvt"); 1446 } 1447 case ARM::BI__builtin_neon_vcvt_f32_v: 1448 case ARM::BI__builtin_neon_vcvtq_f32_v: 1449 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 1450 Ty = GetNeonType(getLLVMContext(), 1451 NeonTypeFlags(NeonTypeFlags::Float32, false, quad)); 1452 return usgn ? Builder.CreateUIToFP(Ops[0], Ty, "vcvt") 1453 : Builder.CreateSIToFP(Ops[0], Ty, "vcvt"); 1454 case ARM::BI__builtin_neon_vcvt_s32_v: 1455 case ARM::BI__builtin_neon_vcvt_u32_v: 1456 case ARM::BI__builtin_neon_vcvtq_s32_v: 1457 case ARM::BI__builtin_neon_vcvtq_u32_v: { 1458 llvm::Type *FloatTy = 1459 GetNeonType(getLLVMContext(), 1460 NeonTypeFlags(NeonTypeFlags::Float32, false, quad)); 1461 Ops[0] = Builder.CreateBitCast(Ops[0], FloatTy); 1462 return usgn ? Builder.CreateFPToUI(Ops[0], Ty, "vcvt") 1463 : Builder.CreateFPToSI(Ops[0], Ty, "vcvt"); 1464 } 1465 case ARM::BI__builtin_neon_vcvt_n_f32_v: 1466 case ARM::BI__builtin_neon_vcvtq_n_f32_v: { 1467 llvm::Type *FloatTy = 1468 GetNeonType(getLLVMContext(), 1469 NeonTypeFlags(NeonTypeFlags::Float32, false, quad)); 1470 llvm::Type *Tys[2] = { FloatTy, Ty }; 1471 Int = usgn ? Intrinsic::arm_neon_vcvtfxu2fp 1472 : Intrinsic::arm_neon_vcvtfxs2fp; 1473 Function *F = CGM.getIntrinsic(Int, Tys); 1474 return EmitNeonCall(F, Ops, "vcvt_n"); 1475 } 1476 case ARM::BI__builtin_neon_vcvt_n_s32_v: 1477 case ARM::BI__builtin_neon_vcvt_n_u32_v: 1478 case ARM::BI__builtin_neon_vcvtq_n_s32_v: 1479 case ARM::BI__builtin_neon_vcvtq_n_u32_v: { 1480 llvm::Type *FloatTy = 1481 GetNeonType(getLLVMContext(), 1482 NeonTypeFlags(NeonTypeFlags::Float32, false, quad)); 1483 llvm::Type *Tys[2] = { Ty, FloatTy }; 1484 Int = usgn ? Intrinsic::arm_neon_vcvtfp2fxu 1485 : Intrinsic::arm_neon_vcvtfp2fxs; 1486 Function *F = CGM.getIntrinsic(Int, Tys); 1487 return EmitNeonCall(F, Ops, "vcvt_n"); 1488 } 1489 case ARM::BI__builtin_neon_vext_v: 1490 case ARM::BI__builtin_neon_vextq_v: { 1491 int CV = cast<ConstantInt>(Ops[2])->getSExtValue(); 1492 SmallVector<Constant*, 16> Indices; 1493 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) 1494 Indices.push_back(ConstantInt::get(Int32Ty, i+CV)); 1495 1496 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 1497 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 1498 Value *SV = llvm::ConstantVector::get(Indices); 1499 return Builder.CreateShuffleVector(Ops[0], Ops[1], SV, "vext"); 1500 } 1501 case ARM::BI__builtin_neon_vhadd_v: 1502 case ARM::BI__builtin_neon_vhaddq_v: 1503 Int = usgn ? Intrinsic::arm_neon_vhaddu : Intrinsic::arm_neon_vhadds; 1504 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vhadd"); 1505 case ARM::BI__builtin_neon_vhsub_v: 1506 case ARM::BI__builtin_neon_vhsubq_v: 1507 Int = usgn ? Intrinsic::arm_neon_vhsubu : Intrinsic::arm_neon_vhsubs; 1508 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vhsub"); 1509 case ARM::BI__builtin_neon_vld1_v: 1510 case ARM::BI__builtin_neon_vld1q_v: 1511 Ops.push_back(GetPointeeAlignment(*this, E->getArg(0))); 1512 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vld1, Ty), 1513 Ops, "vld1"); 1514 case ARM::BI__builtin_neon_vld1_lane_v: 1515 case ARM::BI__builtin_neon_vld1q_lane_v: 1516 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 1517 Ty = llvm::PointerType::getUnqual(VTy->getElementType()); 1518 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 1519 Ops[0] = Builder.CreateLoad(Ops[0]); 1520 return Builder.CreateInsertElement(Ops[1], Ops[0], Ops[2], "vld1_lane"); 1521 case ARM::BI__builtin_neon_vld1_dup_v: 1522 case ARM::BI__builtin_neon_vld1q_dup_v: { 1523 Value *V = UndefValue::get(Ty); 1524 Ty = llvm::PointerType::getUnqual(VTy->getElementType()); 1525 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 1526 Ops[0] = Builder.CreateLoad(Ops[0]); 1527 llvm::Constant *CI = ConstantInt::get(Int32Ty, 0); 1528 Ops[0] = Builder.CreateInsertElement(V, Ops[0], CI); 1529 return EmitNeonSplat(Ops[0], CI); 1530 } 1531 case ARM::BI__builtin_neon_vld2_v: 1532 case ARM::BI__builtin_neon_vld2q_v: { 1533 Function *F = CGM.getIntrinsic(Intrinsic::arm_neon_vld2, Ty); 1534 Value *Align = GetPointeeAlignment(*this, E->getArg(1)); 1535 Ops[1] = Builder.CreateCall2(F, Ops[1], Align, "vld2"); 1536 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 1537 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 1538 return Builder.CreateStore(Ops[1], Ops[0]); 1539 } 1540 case ARM::BI__builtin_neon_vld3_v: 1541 case ARM::BI__builtin_neon_vld3q_v: { 1542 Function *F = CGM.getIntrinsic(Intrinsic::arm_neon_vld3, Ty); 1543 Value *Align = GetPointeeAlignment(*this, E->getArg(1)); 1544 Ops[1] = Builder.CreateCall2(F, Ops[1], Align, "vld3"); 1545 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 1546 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 1547 return Builder.CreateStore(Ops[1], Ops[0]); 1548 } 1549 case ARM::BI__builtin_neon_vld4_v: 1550 case ARM::BI__builtin_neon_vld4q_v: { 1551 Function *F = CGM.getIntrinsic(Intrinsic::arm_neon_vld4, Ty); 1552 Value *Align = GetPointeeAlignment(*this, E->getArg(1)); 1553 Ops[1] = Builder.CreateCall2(F, Ops[1], Align, "vld4"); 1554 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 1555 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 1556 return Builder.CreateStore(Ops[1], Ops[0]); 1557 } 1558 case ARM::BI__builtin_neon_vld2_lane_v: 1559 case ARM::BI__builtin_neon_vld2q_lane_v: { 1560 Function *F = CGM.getIntrinsic(Intrinsic::arm_neon_vld2lane, Ty); 1561 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 1562 Ops[3] = Builder.CreateBitCast(Ops[3], Ty); 1563 Ops.push_back(GetPointeeAlignment(*this, E->getArg(1))); 1564 Ops[1] = Builder.CreateCall(F, makeArrayRef(Ops).slice(1), "vld2_lane"); 1565 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 1566 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 1567 return Builder.CreateStore(Ops[1], Ops[0]); 1568 } 1569 case ARM::BI__builtin_neon_vld3_lane_v: 1570 case ARM::BI__builtin_neon_vld3q_lane_v: { 1571 Function *F = CGM.getIntrinsic(Intrinsic::arm_neon_vld3lane, Ty); 1572 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 1573 Ops[3] = Builder.CreateBitCast(Ops[3], Ty); 1574 Ops[4] = Builder.CreateBitCast(Ops[4], Ty); 1575 Ops.push_back(GetPointeeAlignment(*this, E->getArg(1))); 1576 Ops[1] = Builder.CreateCall(F, makeArrayRef(Ops).slice(1), "vld3_lane"); 1577 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 1578 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 1579 return Builder.CreateStore(Ops[1], Ops[0]); 1580 } 1581 case ARM::BI__builtin_neon_vld4_lane_v: 1582 case ARM::BI__builtin_neon_vld4q_lane_v: { 1583 Function *F = CGM.getIntrinsic(Intrinsic::arm_neon_vld4lane, Ty); 1584 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 1585 Ops[3] = Builder.CreateBitCast(Ops[3], Ty); 1586 Ops[4] = Builder.CreateBitCast(Ops[4], Ty); 1587 Ops[5] = Builder.CreateBitCast(Ops[5], Ty); 1588 Ops.push_back(GetPointeeAlignment(*this, E->getArg(1))); 1589 Ops[1] = Builder.CreateCall(F, makeArrayRef(Ops).slice(1), "vld3_lane"); 1590 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 1591 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 1592 return Builder.CreateStore(Ops[1], Ops[0]); 1593 } 1594 case ARM::BI__builtin_neon_vld2_dup_v: 1595 case ARM::BI__builtin_neon_vld3_dup_v: 1596 case ARM::BI__builtin_neon_vld4_dup_v: { 1597 // Handle 64-bit elements as a special-case. There is no "dup" needed. 1598 if (VTy->getElementType()->getPrimitiveSizeInBits() == 64) { 1599 switch (BuiltinID) { 1600 case ARM::BI__builtin_neon_vld2_dup_v: 1601 Int = Intrinsic::arm_neon_vld2; 1602 break; 1603 case ARM::BI__builtin_neon_vld3_dup_v: 1604 Int = Intrinsic::arm_neon_vld2; 1605 break; 1606 case ARM::BI__builtin_neon_vld4_dup_v: 1607 Int = Intrinsic::arm_neon_vld2; 1608 break; 1609 default: llvm_unreachable("unknown vld_dup intrinsic?"); 1610 } 1611 Function *F = CGM.getIntrinsic(Int, Ty); 1612 Value *Align = GetPointeeAlignment(*this, E->getArg(1)); 1613 Ops[1] = Builder.CreateCall2(F, Ops[1], Align, "vld_dup"); 1614 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 1615 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 1616 return Builder.CreateStore(Ops[1], Ops[0]); 1617 } 1618 switch (BuiltinID) { 1619 case ARM::BI__builtin_neon_vld2_dup_v: 1620 Int = Intrinsic::arm_neon_vld2lane; 1621 break; 1622 case ARM::BI__builtin_neon_vld3_dup_v: 1623 Int = Intrinsic::arm_neon_vld2lane; 1624 break; 1625 case ARM::BI__builtin_neon_vld4_dup_v: 1626 Int = Intrinsic::arm_neon_vld2lane; 1627 break; 1628 default: llvm_unreachable("unknown vld_dup intrinsic?"); 1629 } 1630 Function *F = CGM.getIntrinsic(Int, Ty); 1631 llvm::StructType *STy = cast<llvm::StructType>(F->getReturnType()); 1632 1633 SmallVector<Value*, 6> Args; 1634 Args.push_back(Ops[1]); 1635 Args.append(STy->getNumElements(), UndefValue::get(Ty)); 1636 1637 llvm::Constant *CI = ConstantInt::get(Int32Ty, 0); 1638 Args.push_back(CI); 1639 Args.push_back(GetPointeeAlignment(*this, E->getArg(1))); 1640 1641 Ops[1] = Builder.CreateCall(F, Args, "vld_dup"); 1642 // splat lane 0 to all elts in each vector of the result. 1643 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 1644 Value *Val = Builder.CreateExtractValue(Ops[1], i); 1645 Value *Elt = Builder.CreateBitCast(Val, Ty); 1646 Elt = EmitNeonSplat(Elt, CI); 1647 Elt = Builder.CreateBitCast(Elt, Val->getType()); 1648 Ops[1] = Builder.CreateInsertValue(Ops[1], Elt, i); 1649 } 1650 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 1651 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 1652 return Builder.CreateStore(Ops[1], Ops[0]); 1653 } 1654 case ARM::BI__builtin_neon_vmax_v: 1655 case ARM::BI__builtin_neon_vmaxq_v: 1656 Int = usgn ? Intrinsic::arm_neon_vmaxu : Intrinsic::arm_neon_vmaxs; 1657 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmax"); 1658 case ARM::BI__builtin_neon_vmin_v: 1659 case ARM::BI__builtin_neon_vminq_v: 1660 Int = usgn ? Intrinsic::arm_neon_vminu : Intrinsic::arm_neon_vmins; 1661 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmin"); 1662 case ARM::BI__builtin_neon_vmovl_v: { 1663 llvm::Type *DTy =llvm::VectorType::getTruncatedElementVectorType(VTy); 1664 Ops[0] = Builder.CreateBitCast(Ops[0], DTy); 1665 if (usgn) 1666 return Builder.CreateZExt(Ops[0], Ty, "vmovl"); 1667 return Builder.CreateSExt(Ops[0], Ty, "vmovl"); 1668 } 1669 case ARM::BI__builtin_neon_vmovn_v: { 1670 llvm::Type *QTy = llvm::VectorType::getExtendedElementVectorType(VTy); 1671 Ops[0] = Builder.CreateBitCast(Ops[0], QTy); 1672 return Builder.CreateTrunc(Ops[0], Ty, "vmovn"); 1673 } 1674 case ARM::BI__builtin_neon_vmul_v: 1675 case ARM::BI__builtin_neon_vmulq_v: 1676 assert(Type.isPoly() && "vmul builtin only supported for polynomial types"); 1677 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vmulp, Ty), 1678 Ops, "vmul"); 1679 case ARM::BI__builtin_neon_vmull_v: 1680 Int = usgn ? Intrinsic::arm_neon_vmullu : Intrinsic::arm_neon_vmulls; 1681 Int = Type.isPoly() ? (unsigned)Intrinsic::arm_neon_vmullp : Int; 1682 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vmull"); 1683 case ARM::BI__builtin_neon_vpadal_v: 1684 case ARM::BI__builtin_neon_vpadalq_v: { 1685 Int = usgn ? Intrinsic::arm_neon_vpadalu : Intrinsic::arm_neon_vpadals; 1686 // The source operand type has twice as many elements of half the size. 1687 unsigned EltBits = VTy->getElementType()->getPrimitiveSizeInBits(); 1688 llvm::Type *EltTy = 1689 llvm::IntegerType::get(getLLVMContext(), EltBits / 2); 1690 llvm::Type *NarrowTy = 1691 llvm::VectorType::get(EltTy, VTy->getNumElements() * 2); 1692 llvm::Type *Tys[2] = { Ty, NarrowTy }; 1693 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vpadal"); 1694 } 1695 case ARM::BI__builtin_neon_vpadd_v: 1696 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vpadd, Ty), 1697 Ops, "vpadd"); 1698 case ARM::BI__builtin_neon_vpaddl_v: 1699 case ARM::BI__builtin_neon_vpaddlq_v: { 1700 Int = usgn ? Intrinsic::arm_neon_vpaddlu : Intrinsic::arm_neon_vpaddls; 1701 // The source operand type has twice as many elements of half the size. 1702 unsigned EltBits = VTy->getElementType()->getPrimitiveSizeInBits(); 1703 llvm::Type *EltTy = llvm::IntegerType::get(getLLVMContext(), EltBits / 2); 1704 llvm::Type *NarrowTy = 1705 llvm::VectorType::get(EltTy, VTy->getNumElements() * 2); 1706 llvm::Type *Tys[2] = { Ty, NarrowTy }; 1707 return EmitNeonCall(CGM.getIntrinsic(Int, Tys), Ops, "vpaddl"); 1708 } 1709 case ARM::BI__builtin_neon_vpmax_v: 1710 Int = usgn ? Intrinsic::arm_neon_vpmaxu : Intrinsic::arm_neon_vpmaxs; 1711 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vpmax"); 1712 case ARM::BI__builtin_neon_vpmin_v: 1713 Int = usgn ? Intrinsic::arm_neon_vpminu : Intrinsic::arm_neon_vpmins; 1714 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vpmin"); 1715 case ARM::BI__builtin_neon_vqabs_v: 1716 case ARM::BI__builtin_neon_vqabsq_v: 1717 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vqabs, Ty), 1718 Ops, "vqabs"); 1719 case ARM::BI__builtin_neon_vqadd_v: 1720 case ARM::BI__builtin_neon_vqaddq_v: 1721 Int = usgn ? Intrinsic::arm_neon_vqaddu : Intrinsic::arm_neon_vqadds; 1722 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqadd"); 1723 case ARM::BI__builtin_neon_vqdmlal_v: 1724 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vqdmlal, Ty), 1725 Ops, "vqdmlal"); 1726 case ARM::BI__builtin_neon_vqdmlsl_v: 1727 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vqdmlsl, Ty), 1728 Ops, "vqdmlsl"); 1729 case ARM::BI__builtin_neon_vqdmulh_v: 1730 case ARM::BI__builtin_neon_vqdmulhq_v: 1731 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vqdmulh, Ty), 1732 Ops, "vqdmulh"); 1733 case ARM::BI__builtin_neon_vqdmull_v: 1734 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vqdmull, Ty), 1735 Ops, "vqdmull"); 1736 case ARM::BI__builtin_neon_vqmovn_v: 1737 Int = usgn ? Intrinsic::arm_neon_vqmovnu : Intrinsic::arm_neon_vqmovns; 1738 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqmovn"); 1739 case ARM::BI__builtin_neon_vqmovun_v: 1740 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vqmovnsu, Ty), 1741 Ops, "vqdmull"); 1742 case ARM::BI__builtin_neon_vqneg_v: 1743 case ARM::BI__builtin_neon_vqnegq_v: 1744 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vqneg, Ty), 1745 Ops, "vqneg"); 1746 case ARM::BI__builtin_neon_vqrdmulh_v: 1747 case ARM::BI__builtin_neon_vqrdmulhq_v: 1748 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vqrdmulh, Ty), 1749 Ops, "vqrdmulh"); 1750 case ARM::BI__builtin_neon_vqrshl_v: 1751 case ARM::BI__builtin_neon_vqrshlq_v: 1752 Int = usgn ? Intrinsic::arm_neon_vqrshiftu : Intrinsic::arm_neon_vqrshifts; 1753 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqrshl"); 1754 case ARM::BI__builtin_neon_vqrshrn_n_v: 1755 Int = usgn ? Intrinsic::arm_neon_vqrshiftnu : Intrinsic::arm_neon_vqrshiftns; 1756 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqrshrn_n", 1757 1, true); 1758 case ARM::BI__builtin_neon_vqrshrun_n_v: 1759 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vqrshiftnsu, Ty), 1760 Ops, "vqrshrun_n", 1, true); 1761 case ARM::BI__builtin_neon_vqshl_v: 1762 case ARM::BI__builtin_neon_vqshlq_v: 1763 Int = usgn ? Intrinsic::arm_neon_vqshiftu : Intrinsic::arm_neon_vqshifts; 1764 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqshl"); 1765 case ARM::BI__builtin_neon_vqshl_n_v: 1766 case ARM::BI__builtin_neon_vqshlq_n_v: 1767 Int = usgn ? Intrinsic::arm_neon_vqshiftu : Intrinsic::arm_neon_vqshifts; 1768 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqshl_n", 1769 1, false); 1770 case ARM::BI__builtin_neon_vqshlu_n_v: 1771 case ARM::BI__builtin_neon_vqshluq_n_v: 1772 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vqshiftsu, Ty), 1773 Ops, "vqshlu", 1, false); 1774 case ARM::BI__builtin_neon_vqshrn_n_v: 1775 Int = usgn ? Intrinsic::arm_neon_vqshiftnu : Intrinsic::arm_neon_vqshiftns; 1776 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqshrn_n", 1777 1, true); 1778 case ARM::BI__builtin_neon_vqshrun_n_v: 1779 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vqshiftnsu, Ty), 1780 Ops, "vqshrun_n", 1, true); 1781 case ARM::BI__builtin_neon_vqsub_v: 1782 case ARM::BI__builtin_neon_vqsubq_v: 1783 Int = usgn ? Intrinsic::arm_neon_vqsubu : Intrinsic::arm_neon_vqsubs; 1784 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vqsub"); 1785 case ARM::BI__builtin_neon_vraddhn_v: 1786 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vraddhn, Ty), 1787 Ops, "vraddhn"); 1788 case ARM::BI__builtin_neon_vrecpe_v: 1789 case ARM::BI__builtin_neon_vrecpeq_v: 1790 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vrecpe, Ty), 1791 Ops, "vrecpe"); 1792 case ARM::BI__builtin_neon_vrecps_v: 1793 case ARM::BI__builtin_neon_vrecpsq_v: 1794 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vrecps, Ty), 1795 Ops, "vrecps"); 1796 case ARM::BI__builtin_neon_vrhadd_v: 1797 case ARM::BI__builtin_neon_vrhaddq_v: 1798 Int = usgn ? Intrinsic::arm_neon_vrhaddu : Intrinsic::arm_neon_vrhadds; 1799 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrhadd"); 1800 case ARM::BI__builtin_neon_vrshl_v: 1801 case ARM::BI__builtin_neon_vrshlq_v: 1802 Int = usgn ? Intrinsic::arm_neon_vrshiftu : Intrinsic::arm_neon_vrshifts; 1803 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrshl"); 1804 case ARM::BI__builtin_neon_vrshrn_n_v: 1805 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vrshiftn, Ty), 1806 Ops, "vrshrn_n", 1, true); 1807 case ARM::BI__builtin_neon_vrshr_n_v: 1808 case ARM::BI__builtin_neon_vrshrq_n_v: 1809 Int = usgn ? Intrinsic::arm_neon_vrshiftu : Intrinsic::arm_neon_vrshifts; 1810 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vrshr_n", 1, true); 1811 case ARM::BI__builtin_neon_vrsqrte_v: 1812 case ARM::BI__builtin_neon_vrsqrteq_v: 1813 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vrsqrte, Ty), 1814 Ops, "vrsqrte"); 1815 case ARM::BI__builtin_neon_vrsqrts_v: 1816 case ARM::BI__builtin_neon_vrsqrtsq_v: 1817 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vrsqrts, Ty), 1818 Ops, "vrsqrts"); 1819 case ARM::BI__builtin_neon_vrsra_n_v: 1820 case ARM::BI__builtin_neon_vrsraq_n_v: 1821 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 1822 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 1823 Ops[2] = EmitNeonShiftVector(Ops[2], Ty, true); 1824 Int = usgn ? Intrinsic::arm_neon_vrshiftu : Intrinsic::arm_neon_vrshifts; 1825 Ops[1] = Builder.CreateCall2(CGM.getIntrinsic(Int, Ty), Ops[1], Ops[2]); 1826 return Builder.CreateAdd(Ops[0], Ops[1], "vrsra_n"); 1827 case ARM::BI__builtin_neon_vrsubhn_v: 1828 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vrsubhn, Ty), 1829 Ops, "vrsubhn"); 1830 case ARM::BI__builtin_neon_vshl_v: 1831 case ARM::BI__builtin_neon_vshlq_v: 1832 Int = usgn ? Intrinsic::arm_neon_vshiftu : Intrinsic::arm_neon_vshifts; 1833 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vshl"); 1834 case ARM::BI__builtin_neon_vshll_n_v: 1835 Int = usgn ? Intrinsic::arm_neon_vshiftlu : Intrinsic::arm_neon_vshiftls; 1836 return EmitNeonCall(CGM.getIntrinsic(Int, Ty), Ops, "vshll", 1); 1837 case ARM::BI__builtin_neon_vshl_n_v: 1838 case ARM::BI__builtin_neon_vshlq_n_v: 1839 Ops[1] = EmitNeonShiftVector(Ops[1], Ty, false); 1840 return Builder.CreateShl(Builder.CreateBitCast(Ops[0],Ty), Ops[1], "vshl_n"); 1841 case ARM::BI__builtin_neon_vshrn_n_v: 1842 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vshiftn, Ty), 1843 Ops, "vshrn_n", 1, true); 1844 case ARM::BI__builtin_neon_vshr_n_v: 1845 case ARM::BI__builtin_neon_vshrq_n_v: 1846 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 1847 Ops[1] = EmitNeonShiftVector(Ops[1], Ty, false); 1848 if (usgn) 1849 return Builder.CreateLShr(Ops[0], Ops[1], "vshr_n"); 1850 else 1851 return Builder.CreateAShr(Ops[0], Ops[1], "vshr_n"); 1852 case ARM::BI__builtin_neon_vsri_n_v: 1853 case ARM::BI__builtin_neon_vsriq_n_v: 1854 rightShift = true; 1855 case ARM::BI__builtin_neon_vsli_n_v: 1856 case ARM::BI__builtin_neon_vsliq_n_v: 1857 Ops[2] = EmitNeonShiftVector(Ops[2], Ty, rightShift); 1858 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vshiftins, Ty), 1859 Ops, "vsli_n"); 1860 case ARM::BI__builtin_neon_vsra_n_v: 1861 case ARM::BI__builtin_neon_vsraq_n_v: 1862 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 1863 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 1864 Ops[2] = EmitNeonShiftVector(Ops[2], Ty, false); 1865 if (usgn) 1866 Ops[1] = Builder.CreateLShr(Ops[1], Ops[2], "vsra_n"); 1867 else 1868 Ops[1] = Builder.CreateAShr(Ops[1], Ops[2], "vsra_n"); 1869 return Builder.CreateAdd(Ops[0], Ops[1]); 1870 case ARM::BI__builtin_neon_vst1_v: 1871 case ARM::BI__builtin_neon_vst1q_v: 1872 Ops.push_back(GetPointeeAlignment(*this, E->getArg(0))); 1873 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vst1, Ty), 1874 Ops, ""); 1875 case ARM::BI__builtin_neon_vst1_lane_v: 1876 case ARM::BI__builtin_neon_vst1q_lane_v: 1877 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 1878 Ops[1] = Builder.CreateExtractElement(Ops[1], Ops[2]); 1879 Ty = llvm::PointerType::getUnqual(Ops[1]->getType()); 1880 return Builder.CreateStore(Ops[1], Builder.CreateBitCast(Ops[0], Ty)); 1881 case ARM::BI__builtin_neon_vst2_v: 1882 case ARM::BI__builtin_neon_vst2q_v: 1883 Ops.push_back(GetPointeeAlignment(*this, E->getArg(0))); 1884 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vst2, Ty), 1885 Ops, ""); 1886 case ARM::BI__builtin_neon_vst2_lane_v: 1887 case ARM::BI__builtin_neon_vst2q_lane_v: 1888 Ops.push_back(GetPointeeAlignment(*this, E->getArg(0))); 1889 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vst2lane, Ty), 1890 Ops, ""); 1891 case ARM::BI__builtin_neon_vst3_v: 1892 case ARM::BI__builtin_neon_vst3q_v: 1893 Ops.push_back(GetPointeeAlignment(*this, E->getArg(0))); 1894 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vst3, Ty), 1895 Ops, ""); 1896 case ARM::BI__builtin_neon_vst3_lane_v: 1897 case ARM::BI__builtin_neon_vst3q_lane_v: 1898 Ops.push_back(GetPointeeAlignment(*this, E->getArg(0))); 1899 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vst3lane, Ty), 1900 Ops, ""); 1901 case ARM::BI__builtin_neon_vst4_v: 1902 case ARM::BI__builtin_neon_vst4q_v: 1903 Ops.push_back(GetPointeeAlignment(*this, E->getArg(0))); 1904 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vst4, Ty), 1905 Ops, ""); 1906 case ARM::BI__builtin_neon_vst4_lane_v: 1907 case ARM::BI__builtin_neon_vst4q_lane_v: 1908 Ops.push_back(GetPointeeAlignment(*this, E->getArg(0))); 1909 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vst4lane, Ty), 1910 Ops, ""); 1911 case ARM::BI__builtin_neon_vsubhn_v: 1912 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vsubhn, Ty), 1913 Ops, "vsubhn"); 1914 case ARM::BI__builtin_neon_vtbl1_v: 1915 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbl1), 1916 Ops, "vtbl1"); 1917 case ARM::BI__builtin_neon_vtbl2_v: 1918 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbl2), 1919 Ops, "vtbl2"); 1920 case ARM::BI__builtin_neon_vtbl3_v: 1921 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbl3), 1922 Ops, "vtbl3"); 1923 case ARM::BI__builtin_neon_vtbl4_v: 1924 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbl4), 1925 Ops, "vtbl4"); 1926 case ARM::BI__builtin_neon_vtbx1_v: 1927 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbx1), 1928 Ops, "vtbx1"); 1929 case ARM::BI__builtin_neon_vtbx2_v: 1930 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbx2), 1931 Ops, "vtbx2"); 1932 case ARM::BI__builtin_neon_vtbx3_v: 1933 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbx3), 1934 Ops, "vtbx3"); 1935 case ARM::BI__builtin_neon_vtbx4_v: 1936 return EmitNeonCall(CGM.getIntrinsic(Intrinsic::arm_neon_vtbx4), 1937 Ops, "vtbx4"); 1938 case ARM::BI__builtin_neon_vtst_v: 1939 case ARM::BI__builtin_neon_vtstq_v: { 1940 Ops[0] = Builder.CreateBitCast(Ops[0], Ty); 1941 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 1942 Ops[0] = Builder.CreateAnd(Ops[0], Ops[1]); 1943 Ops[0] = Builder.CreateICmp(ICmpInst::ICMP_NE, Ops[0], 1944 ConstantAggregateZero::get(Ty)); 1945 return Builder.CreateSExt(Ops[0], Ty, "vtst"); 1946 } 1947 case ARM::BI__builtin_neon_vtrn_v: 1948 case ARM::BI__builtin_neon_vtrnq_v: { 1949 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty)); 1950 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 1951 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 1952 Value *SV = 0; 1953 1954 for (unsigned vi = 0; vi != 2; ++vi) { 1955 SmallVector<Constant*, 16> Indices; 1956 for (unsigned i = 0, e = VTy->getNumElements(); i != e; i += 2) { 1957 Indices.push_back(ConstantInt::get(Int32Ty, i+vi)); 1958 Indices.push_back(ConstantInt::get(Int32Ty, i+e+vi)); 1959 } 1960 Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ops[0], vi); 1961 SV = llvm::ConstantVector::get(Indices); 1962 SV = Builder.CreateShuffleVector(Ops[1], Ops[2], SV, "vtrn"); 1963 SV = Builder.CreateStore(SV, Addr); 1964 } 1965 return SV; 1966 } 1967 case ARM::BI__builtin_neon_vuzp_v: 1968 case ARM::BI__builtin_neon_vuzpq_v: { 1969 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty)); 1970 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 1971 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 1972 Value *SV = 0; 1973 1974 for (unsigned vi = 0; vi != 2; ++vi) { 1975 SmallVector<Constant*, 16> Indices; 1976 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) 1977 Indices.push_back(ConstantInt::get(Int32Ty, 2*i+vi)); 1978 1979 Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ops[0], vi); 1980 SV = llvm::ConstantVector::get(Indices); 1981 SV = Builder.CreateShuffleVector(Ops[1], Ops[2], SV, "vuzp"); 1982 SV = Builder.CreateStore(SV, Addr); 1983 } 1984 return SV; 1985 } 1986 case ARM::BI__builtin_neon_vzip_v: 1987 case ARM::BI__builtin_neon_vzipq_v: { 1988 Ops[0] = Builder.CreateBitCast(Ops[0], llvm::PointerType::getUnqual(Ty)); 1989 Ops[1] = Builder.CreateBitCast(Ops[1], Ty); 1990 Ops[2] = Builder.CreateBitCast(Ops[2], Ty); 1991 Value *SV = 0; 1992 1993 for (unsigned vi = 0; vi != 2; ++vi) { 1994 SmallVector<Constant*, 16> Indices; 1995 for (unsigned i = 0, e = VTy->getNumElements(); i != e; i += 2) { 1996 Indices.push_back(ConstantInt::get(Int32Ty, (i + vi*e) >> 1)); 1997 Indices.push_back(ConstantInt::get(Int32Ty, ((i + vi*e) >> 1)+e)); 1998 } 1999 Value *Addr = Builder.CreateConstInBoundsGEP1_32(Ops[0], vi); 2000 SV = llvm::ConstantVector::get(Indices); 2001 SV = Builder.CreateShuffleVector(Ops[1], Ops[2], SV, "vzip"); 2002 SV = Builder.CreateStore(SV, Addr); 2003 } 2004 return SV; 2005 } 2006 } 2007 } 2008 2009 llvm::Value *CodeGenFunction:: 2010 BuildVector(const SmallVectorImpl<llvm::Value*> &Ops) { 2011 assert((Ops.size() & (Ops.size() - 1)) == 0 && 2012 "Not a power-of-two sized vector!"); 2013 bool AllConstants = true; 2014 for (unsigned i = 0, e = Ops.size(); i != e && AllConstants; ++i) 2015 AllConstants &= isa<Constant>(Ops[i]); 2016 2017 // If this is a constant vector, create a ConstantVector. 2018 if (AllConstants) { 2019 std::vector<llvm::Constant*> CstOps; 2020 for (unsigned i = 0, e = Ops.size(); i != e; ++i) 2021 CstOps.push_back(cast<Constant>(Ops[i])); 2022 return llvm::ConstantVector::get(CstOps); 2023 } 2024 2025 // Otherwise, insertelement the values to build the vector. 2026 Value *Result = 2027 llvm::UndefValue::get(llvm::VectorType::get(Ops[0]->getType(), Ops.size())); 2028 2029 for (unsigned i = 0, e = Ops.size(); i != e; ++i) 2030 Result = Builder.CreateInsertElement(Result, Ops[i], 2031 llvm::ConstantInt::get(llvm::Type::getInt32Ty(getLLVMContext()), i)); 2032 2033 return Result; 2034 } 2035 2036 Value *CodeGenFunction::EmitX86BuiltinExpr(unsigned BuiltinID, 2037 const CallExpr *E) { 2038 SmallVector<Value*, 4> Ops; 2039 2040 // Find out if any arguments are required to be integer constant expressions. 2041 unsigned ICEArguments = 0; 2042 ASTContext::GetBuiltinTypeError Error; 2043 getContext().GetBuiltinType(BuiltinID, Error, &ICEArguments); 2044 assert(Error == ASTContext::GE_None && "Should not codegen an error"); 2045 2046 for (unsigned i = 0, e = E->getNumArgs(); i != e; i++) { 2047 // If this is a normal argument, just emit it as a scalar. 2048 if ((ICEArguments & (1 << i)) == 0) { 2049 Ops.push_back(EmitScalarExpr(E->getArg(i))); 2050 continue; 2051 } 2052 2053 // If this is required to be a constant, constant fold it so that we know 2054 // that the generated intrinsic gets a ConstantInt. 2055 llvm::APSInt Result; 2056 bool IsConst = E->getArg(i)->isIntegerConstantExpr(Result, getContext()); 2057 assert(IsConst && "Constant arg isn't actually constant?"); (void)IsConst; 2058 Ops.push_back(llvm::ConstantInt::get(getLLVMContext(), Result)); 2059 } 2060 2061 switch (BuiltinID) { 2062 default: return 0; 2063 case X86::BI__builtin_ia32_pslldi128: 2064 case X86::BI__builtin_ia32_psllqi128: 2065 case X86::BI__builtin_ia32_psllwi128: 2066 case X86::BI__builtin_ia32_psradi128: 2067 case X86::BI__builtin_ia32_psrawi128: 2068 case X86::BI__builtin_ia32_psrldi128: 2069 case X86::BI__builtin_ia32_psrlqi128: 2070 case X86::BI__builtin_ia32_psrlwi128: { 2071 Ops[1] = Builder.CreateZExt(Ops[1], Int64Ty, "zext"); 2072 llvm::Type *Ty = llvm::VectorType::get(Int64Ty, 2); 2073 llvm::Value *Zero = llvm::ConstantInt::get(Int32Ty, 0); 2074 Ops[1] = Builder.CreateInsertElement(llvm::UndefValue::get(Ty), 2075 Ops[1], Zero, "insert"); 2076 Ops[1] = Builder.CreateBitCast(Ops[1], Ops[0]->getType(), "bitcast"); 2077 const char *name = 0; 2078 Intrinsic::ID ID = Intrinsic::not_intrinsic; 2079 2080 switch (BuiltinID) { 2081 default: llvm_unreachable("Unsupported shift intrinsic!"); 2082 case X86::BI__builtin_ia32_pslldi128: 2083 name = "pslldi"; 2084 ID = Intrinsic::x86_sse2_psll_d; 2085 break; 2086 case X86::BI__builtin_ia32_psllqi128: 2087 name = "psllqi"; 2088 ID = Intrinsic::x86_sse2_psll_q; 2089 break; 2090 case X86::BI__builtin_ia32_psllwi128: 2091 name = "psllwi"; 2092 ID = Intrinsic::x86_sse2_psll_w; 2093 break; 2094 case X86::BI__builtin_ia32_psradi128: 2095 name = "psradi"; 2096 ID = Intrinsic::x86_sse2_psra_d; 2097 break; 2098 case X86::BI__builtin_ia32_psrawi128: 2099 name = "psrawi"; 2100 ID = Intrinsic::x86_sse2_psra_w; 2101 break; 2102 case X86::BI__builtin_ia32_psrldi128: 2103 name = "psrldi"; 2104 ID = Intrinsic::x86_sse2_psrl_d; 2105 break; 2106 case X86::BI__builtin_ia32_psrlqi128: 2107 name = "psrlqi"; 2108 ID = Intrinsic::x86_sse2_psrl_q; 2109 break; 2110 case X86::BI__builtin_ia32_psrlwi128: 2111 name = "psrlwi"; 2112 ID = Intrinsic::x86_sse2_psrl_w; 2113 break; 2114 } 2115 llvm::Function *F = CGM.getIntrinsic(ID); 2116 return Builder.CreateCall(F, Ops, name); 2117 } 2118 case X86::BI__builtin_ia32_vec_init_v8qi: 2119 case X86::BI__builtin_ia32_vec_init_v4hi: 2120 case X86::BI__builtin_ia32_vec_init_v2si: 2121 return Builder.CreateBitCast(BuildVector(Ops), 2122 llvm::Type::getX86_MMXTy(getLLVMContext())); 2123 case X86::BI__builtin_ia32_vec_ext_v2si: 2124 return Builder.CreateExtractElement(Ops[0], 2125 llvm::ConstantInt::get(Ops[1]->getType(), 0)); 2126 case X86::BI__builtin_ia32_pslldi: 2127 case X86::BI__builtin_ia32_psllqi: 2128 case X86::BI__builtin_ia32_psllwi: 2129 case X86::BI__builtin_ia32_psradi: 2130 case X86::BI__builtin_ia32_psrawi: 2131 case X86::BI__builtin_ia32_psrldi: 2132 case X86::BI__builtin_ia32_psrlqi: 2133 case X86::BI__builtin_ia32_psrlwi: { 2134 Ops[1] = Builder.CreateZExt(Ops[1], Int64Ty, "zext"); 2135 llvm::Type *Ty = llvm::VectorType::get(Int64Ty, 1); 2136 Ops[1] = Builder.CreateBitCast(Ops[1], Ty, "bitcast"); 2137 const char *name = 0; 2138 Intrinsic::ID ID = Intrinsic::not_intrinsic; 2139 2140 switch (BuiltinID) { 2141 default: llvm_unreachable("Unsupported shift intrinsic!"); 2142 case X86::BI__builtin_ia32_pslldi: 2143 name = "pslldi"; 2144 ID = Intrinsic::x86_mmx_psll_d; 2145 break; 2146 case X86::BI__builtin_ia32_psllqi: 2147 name = "psllqi"; 2148 ID = Intrinsic::x86_mmx_psll_q; 2149 break; 2150 case X86::BI__builtin_ia32_psllwi: 2151 name = "psllwi"; 2152 ID = Intrinsic::x86_mmx_psll_w; 2153 break; 2154 case X86::BI__builtin_ia32_psradi: 2155 name = "psradi"; 2156 ID = Intrinsic::x86_mmx_psra_d; 2157 break; 2158 case X86::BI__builtin_ia32_psrawi: 2159 name = "psrawi"; 2160 ID = Intrinsic::x86_mmx_psra_w; 2161 break; 2162 case X86::BI__builtin_ia32_psrldi: 2163 name = "psrldi"; 2164 ID = Intrinsic::x86_mmx_psrl_d; 2165 break; 2166 case X86::BI__builtin_ia32_psrlqi: 2167 name = "psrlqi"; 2168 ID = Intrinsic::x86_mmx_psrl_q; 2169 break; 2170 case X86::BI__builtin_ia32_psrlwi: 2171 name = "psrlwi"; 2172 ID = Intrinsic::x86_mmx_psrl_w; 2173 break; 2174 } 2175 llvm::Function *F = CGM.getIntrinsic(ID); 2176 return Builder.CreateCall(F, Ops, name); 2177 } 2178 case X86::BI__builtin_ia32_cmpps: { 2179 llvm::Function *F = CGM.getIntrinsic(Intrinsic::x86_sse_cmp_ps); 2180 return Builder.CreateCall(F, Ops, "cmpps"); 2181 } 2182 case X86::BI__builtin_ia32_cmpss: { 2183 llvm::Function *F = CGM.getIntrinsic(Intrinsic::x86_sse_cmp_ss); 2184 return Builder.CreateCall(F, Ops, "cmpss"); 2185 } 2186 case X86::BI__builtin_ia32_ldmxcsr: { 2187 llvm::Type *PtrTy = Int8PtrTy; 2188 Value *One = llvm::ConstantInt::get(Int32Ty, 1); 2189 Value *Tmp = Builder.CreateAlloca(Int32Ty, One); 2190 Builder.CreateStore(Ops[0], Tmp); 2191 return Builder.CreateCall(CGM.getIntrinsic(Intrinsic::x86_sse_ldmxcsr), 2192 Builder.CreateBitCast(Tmp, PtrTy)); 2193 } 2194 case X86::BI__builtin_ia32_stmxcsr: { 2195 llvm::Type *PtrTy = Int8PtrTy; 2196 Value *One = llvm::ConstantInt::get(Int32Ty, 1); 2197 Value *Tmp = Builder.CreateAlloca(Int32Ty, One); 2198 Builder.CreateCall(CGM.getIntrinsic(Intrinsic::x86_sse_stmxcsr), 2199 Builder.CreateBitCast(Tmp, PtrTy)); 2200 return Builder.CreateLoad(Tmp, "stmxcsr"); 2201 } 2202 case X86::BI__builtin_ia32_cmppd: { 2203 llvm::Function *F = CGM.getIntrinsic(Intrinsic::x86_sse2_cmp_pd); 2204 return Builder.CreateCall(F, Ops, "cmppd"); 2205 } 2206 case X86::BI__builtin_ia32_cmpsd: { 2207 llvm::Function *F = CGM.getIntrinsic(Intrinsic::x86_sse2_cmp_sd); 2208 return Builder.CreateCall(F, Ops, "cmpsd"); 2209 } 2210 case X86::BI__builtin_ia32_storehps: 2211 case X86::BI__builtin_ia32_storelps: { 2212 llvm::Type *PtrTy = llvm::PointerType::getUnqual(Int64Ty); 2213 llvm::Type *VecTy = llvm::VectorType::get(Int64Ty, 2); 2214 2215 // cast val v2i64 2216 Ops[1] = Builder.CreateBitCast(Ops[1], VecTy, "cast"); 2217 2218 // extract (0, 1) 2219 unsigned Index = BuiltinID == X86::BI__builtin_ia32_storelps ? 0 : 1; 2220 llvm::Value *Idx = llvm::ConstantInt::get(Int32Ty, Index); 2221 Ops[1] = Builder.CreateExtractElement(Ops[1], Idx, "extract"); 2222 2223 // cast pointer to i64 & store 2224 Ops[0] = Builder.CreateBitCast(Ops[0], PtrTy); 2225 return Builder.CreateStore(Ops[1], Ops[0]); 2226 } 2227 case X86::BI__builtin_ia32_palignr: { 2228 unsigned shiftVal = cast<llvm::ConstantInt>(Ops[2])->getZExtValue(); 2229 2230 // If palignr is shifting the pair of input vectors less than 9 bytes, 2231 // emit a shuffle instruction. 2232 if (shiftVal <= 8) { 2233 SmallVector<llvm::Constant*, 8> Indices; 2234 for (unsigned i = 0; i != 8; ++i) 2235 Indices.push_back(llvm::ConstantInt::get(Int32Ty, shiftVal + i)); 2236 2237 Value* SV = llvm::ConstantVector::get(Indices); 2238 return Builder.CreateShuffleVector(Ops[1], Ops[0], SV, "palignr"); 2239 } 2240 2241 // If palignr is shifting the pair of input vectors more than 8 but less 2242 // than 16 bytes, emit a logical right shift of the destination. 2243 if (shiftVal < 16) { 2244 // MMX has these as 1 x i64 vectors for some odd optimization reasons. 2245 llvm::Type *VecTy = llvm::VectorType::get(Int64Ty, 1); 2246 2247 Ops[0] = Builder.CreateBitCast(Ops[0], VecTy, "cast"); 2248 Ops[1] = llvm::ConstantInt::get(VecTy, (shiftVal-8) * 8); 2249 2250 // create i32 constant 2251 llvm::Function *F = CGM.getIntrinsic(Intrinsic::x86_mmx_psrl_q); 2252 return Builder.CreateCall(F, makeArrayRef(&Ops[0], 2), "palignr"); 2253 } 2254 2255 // If palignr is shifting the pair of vectors more than 16 bytes, emit zero. 2256 return llvm::Constant::getNullValue(ConvertType(E->getType())); 2257 } 2258 case X86::BI__builtin_ia32_palignr128: { 2259 unsigned shiftVal = cast<llvm::ConstantInt>(Ops[2])->getZExtValue(); 2260 2261 // If palignr is shifting the pair of input vectors less than 17 bytes, 2262 // emit a shuffle instruction. 2263 if (shiftVal <= 16) { 2264 SmallVector<llvm::Constant*, 16> Indices; 2265 for (unsigned i = 0; i != 16; ++i) 2266 Indices.push_back(llvm::ConstantInt::get(Int32Ty, shiftVal + i)); 2267 2268 Value* SV = llvm::ConstantVector::get(Indices); 2269 return Builder.CreateShuffleVector(Ops[1], Ops[0], SV, "palignr"); 2270 } 2271 2272 // If palignr is shifting the pair of input vectors more than 16 but less 2273 // than 32 bytes, emit a logical right shift of the destination. 2274 if (shiftVal < 32) { 2275 llvm::Type *VecTy = llvm::VectorType::get(Int64Ty, 2); 2276 2277 Ops[0] = Builder.CreateBitCast(Ops[0], VecTy, "cast"); 2278 Ops[1] = llvm::ConstantInt::get(Int32Ty, (shiftVal-16) * 8); 2279 2280 // create i32 constant 2281 llvm::Function *F = CGM.getIntrinsic(Intrinsic::x86_sse2_psrl_dq); 2282 return Builder.CreateCall(F, makeArrayRef(&Ops[0], 2), "palignr"); 2283 } 2284 2285 // If palignr is shifting the pair of vectors more than 32 bytes, emit zero. 2286 return llvm::Constant::getNullValue(ConvertType(E->getType())); 2287 } 2288 case X86::BI__builtin_ia32_movntps: 2289 case X86::BI__builtin_ia32_movntpd: 2290 case X86::BI__builtin_ia32_movntdq: 2291 case X86::BI__builtin_ia32_movnti: { 2292 llvm::MDNode *Node = llvm::MDNode::get(getLLVMContext(), 2293 Builder.getInt32(1)); 2294 2295 // Convert the type of the pointer to a pointer to the stored type. 2296 Value *BC = Builder.CreateBitCast(Ops[0], 2297 llvm::PointerType::getUnqual(Ops[1]->getType()), 2298 "cast"); 2299 StoreInst *SI = Builder.CreateStore(Ops[1], BC); 2300 SI->setMetadata(CGM.getModule().getMDKindID("nontemporal"), Node); 2301 SI->setAlignment(16); 2302 return SI; 2303 } 2304 // 3DNow! 2305 case X86::BI__builtin_ia32_pavgusb: 2306 case X86::BI__builtin_ia32_pf2id: 2307 case X86::BI__builtin_ia32_pfacc: 2308 case X86::BI__builtin_ia32_pfadd: 2309 case X86::BI__builtin_ia32_pfcmpeq: 2310 case X86::BI__builtin_ia32_pfcmpge: 2311 case X86::BI__builtin_ia32_pfcmpgt: 2312 case X86::BI__builtin_ia32_pfmax: 2313 case X86::BI__builtin_ia32_pfmin: 2314 case X86::BI__builtin_ia32_pfmul: 2315 case X86::BI__builtin_ia32_pfrcp: 2316 case X86::BI__builtin_ia32_pfrcpit1: 2317 case X86::BI__builtin_ia32_pfrcpit2: 2318 case X86::BI__builtin_ia32_pfrsqrt: 2319 case X86::BI__builtin_ia32_pfrsqit1: 2320 case X86::BI__builtin_ia32_pfrsqrtit1: 2321 case X86::BI__builtin_ia32_pfsub: 2322 case X86::BI__builtin_ia32_pfsubr: 2323 case X86::BI__builtin_ia32_pi2fd: 2324 case X86::BI__builtin_ia32_pmulhrw: 2325 case X86::BI__builtin_ia32_pf2iw: 2326 case X86::BI__builtin_ia32_pfnacc: 2327 case X86::BI__builtin_ia32_pfpnacc: 2328 case X86::BI__builtin_ia32_pi2fw: 2329 case X86::BI__builtin_ia32_pswapdsf: 2330 case X86::BI__builtin_ia32_pswapdsi: { 2331 const char *name = 0; 2332 Intrinsic::ID ID = Intrinsic::not_intrinsic; 2333 switch(BuiltinID) { 2334 case X86::BI__builtin_ia32_pavgusb: 2335 name = "pavgusb"; 2336 ID = Intrinsic::x86_3dnow_pavgusb; 2337 break; 2338 case X86::BI__builtin_ia32_pf2id: 2339 name = "pf2id"; 2340 ID = Intrinsic::x86_3dnow_pf2id; 2341 break; 2342 case X86::BI__builtin_ia32_pfacc: 2343 name = "pfacc"; 2344 ID = Intrinsic::x86_3dnow_pfacc; 2345 break; 2346 case X86::BI__builtin_ia32_pfadd: 2347 name = "pfadd"; 2348 ID = Intrinsic::x86_3dnow_pfadd; 2349 break; 2350 case X86::BI__builtin_ia32_pfcmpeq: 2351 name = "pfcmpeq"; 2352 ID = Intrinsic::x86_3dnow_pfcmpeq; 2353 break; 2354 case X86::BI__builtin_ia32_pfcmpge: 2355 name = "pfcmpge"; 2356 ID = Intrinsic::x86_3dnow_pfcmpge; 2357 break; 2358 case X86::BI__builtin_ia32_pfcmpgt: 2359 name = "pfcmpgt"; 2360 ID = Intrinsic::x86_3dnow_pfcmpgt; 2361 break; 2362 case X86::BI__builtin_ia32_pfmax: 2363 name = "pfmax"; 2364 ID = Intrinsic::x86_3dnow_pfmax; 2365 break; 2366 case X86::BI__builtin_ia32_pfmin: 2367 name = "pfmin"; 2368 ID = Intrinsic::x86_3dnow_pfmin; 2369 break; 2370 case X86::BI__builtin_ia32_pfmul: 2371 name = "pfmul"; 2372 ID = Intrinsic::x86_3dnow_pfmul; 2373 break; 2374 case X86::BI__builtin_ia32_pfrcp: 2375 name = "pfrcp"; 2376 ID = Intrinsic::x86_3dnow_pfrcp; 2377 break; 2378 case X86::BI__builtin_ia32_pfrcpit1: 2379 name = "pfrcpit1"; 2380 ID = Intrinsic::x86_3dnow_pfrcpit1; 2381 break; 2382 case X86::BI__builtin_ia32_pfrcpit2: 2383 name = "pfrcpit2"; 2384 ID = Intrinsic::x86_3dnow_pfrcpit2; 2385 break; 2386 case X86::BI__builtin_ia32_pfrsqrt: 2387 name = "pfrsqrt"; 2388 ID = Intrinsic::x86_3dnow_pfrsqrt; 2389 break; 2390 case X86::BI__builtin_ia32_pfrsqit1: 2391 case X86::BI__builtin_ia32_pfrsqrtit1: 2392 name = "pfrsqit1"; 2393 ID = Intrinsic::x86_3dnow_pfrsqit1; 2394 break; 2395 case X86::BI__builtin_ia32_pfsub: 2396 name = "pfsub"; 2397 ID = Intrinsic::x86_3dnow_pfsub; 2398 break; 2399 case X86::BI__builtin_ia32_pfsubr: 2400 name = "pfsubr"; 2401 ID = Intrinsic::x86_3dnow_pfsubr; 2402 break; 2403 case X86::BI__builtin_ia32_pi2fd: 2404 name = "pi2fd"; 2405 ID = Intrinsic::x86_3dnow_pi2fd; 2406 break; 2407 case X86::BI__builtin_ia32_pmulhrw: 2408 name = "pmulhrw"; 2409 ID = Intrinsic::x86_3dnow_pmulhrw; 2410 break; 2411 case X86::BI__builtin_ia32_pf2iw: 2412 name = "pf2iw"; 2413 ID = Intrinsic::x86_3dnowa_pf2iw; 2414 break; 2415 case X86::BI__builtin_ia32_pfnacc: 2416 name = "pfnacc"; 2417 ID = Intrinsic::x86_3dnowa_pfnacc; 2418 break; 2419 case X86::BI__builtin_ia32_pfpnacc: 2420 name = "pfpnacc"; 2421 ID = Intrinsic::x86_3dnowa_pfpnacc; 2422 break; 2423 case X86::BI__builtin_ia32_pi2fw: 2424 name = "pi2fw"; 2425 ID = Intrinsic::x86_3dnowa_pi2fw; 2426 break; 2427 case X86::BI__builtin_ia32_pswapdsf: 2428 case X86::BI__builtin_ia32_pswapdsi: 2429 name = "pswapd"; 2430 ID = Intrinsic::x86_3dnowa_pswapd; 2431 break; 2432 } 2433 llvm::Function *F = CGM.getIntrinsic(ID); 2434 return Builder.CreateCall(F, Ops, name); 2435 } 2436 } 2437 } 2438 2439 Value *CodeGenFunction::EmitPPCBuiltinExpr(unsigned BuiltinID, 2440 const CallExpr *E) { 2441 SmallVector<Value*, 4> Ops; 2442 2443 for (unsigned i = 0, e = E->getNumArgs(); i != e; i++) 2444 Ops.push_back(EmitScalarExpr(E->getArg(i))); 2445 2446 Intrinsic::ID ID = Intrinsic::not_intrinsic; 2447 2448 switch (BuiltinID) { 2449 default: return 0; 2450 2451 // vec_ld, vec_lvsl, vec_lvsr 2452 case PPC::BI__builtin_altivec_lvx: 2453 case PPC::BI__builtin_altivec_lvxl: 2454 case PPC::BI__builtin_altivec_lvebx: 2455 case PPC::BI__builtin_altivec_lvehx: 2456 case PPC::BI__builtin_altivec_lvewx: 2457 case PPC::BI__builtin_altivec_lvsl: 2458 case PPC::BI__builtin_altivec_lvsr: 2459 { 2460 Ops[1] = Builder.CreateBitCast(Ops[1], Int8PtrTy); 2461 2462 Ops[0] = Builder.CreateGEP(Ops[1], Ops[0]); 2463 Ops.pop_back(); 2464 2465 switch (BuiltinID) { 2466 default: llvm_unreachable("Unsupported ld/lvsl/lvsr intrinsic!"); 2467 case PPC::BI__builtin_altivec_lvx: 2468 ID = Intrinsic::ppc_altivec_lvx; 2469 break; 2470 case PPC::BI__builtin_altivec_lvxl: 2471 ID = Intrinsic::ppc_altivec_lvxl; 2472 break; 2473 case PPC::BI__builtin_altivec_lvebx: 2474 ID = Intrinsic::ppc_altivec_lvebx; 2475 break; 2476 case PPC::BI__builtin_altivec_lvehx: 2477 ID = Intrinsic::ppc_altivec_lvehx; 2478 break; 2479 case PPC::BI__builtin_altivec_lvewx: 2480 ID = Intrinsic::ppc_altivec_lvewx; 2481 break; 2482 case PPC::BI__builtin_altivec_lvsl: 2483 ID = Intrinsic::ppc_altivec_lvsl; 2484 break; 2485 case PPC::BI__builtin_altivec_lvsr: 2486 ID = Intrinsic::ppc_altivec_lvsr; 2487 break; 2488 } 2489 llvm::Function *F = CGM.getIntrinsic(ID); 2490 return Builder.CreateCall(F, Ops, ""); 2491 } 2492 2493 // vec_st 2494 case PPC::BI__builtin_altivec_stvx: 2495 case PPC::BI__builtin_altivec_stvxl: 2496 case PPC::BI__builtin_altivec_stvebx: 2497 case PPC::BI__builtin_altivec_stvehx: 2498 case PPC::BI__builtin_altivec_stvewx: 2499 { 2500 Ops[2] = Builder.CreateBitCast(Ops[2], Int8PtrTy); 2501 Ops[1] = Builder.CreateGEP(Ops[2], Ops[1]); 2502 Ops.pop_back(); 2503 2504 switch (BuiltinID) { 2505 default: llvm_unreachable("Unsupported st intrinsic!"); 2506 case PPC::BI__builtin_altivec_stvx: 2507 ID = Intrinsic::ppc_altivec_stvx; 2508 break; 2509 case PPC::BI__builtin_altivec_stvxl: 2510 ID = Intrinsic::ppc_altivec_stvxl; 2511 break; 2512 case PPC::BI__builtin_altivec_stvebx: 2513 ID = Intrinsic::ppc_altivec_stvebx; 2514 break; 2515 case PPC::BI__builtin_altivec_stvehx: 2516 ID = Intrinsic::ppc_altivec_stvehx; 2517 break; 2518 case PPC::BI__builtin_altivec_stvewx: 2519 ID = Intrinsic::ppc_altivec_stvewx; 2520 break; 2521 } 2522 llvm::Function *F = CGM.getIntrinsic(ID); 2523 return Builder.CreateCall(F, Ops, ""); 2524 } 2525 } 2526 return 0; 2527 } 2528