1 //===--- CGCall.cpp - Encapsulate calling convention details ----*- C++ -*-===// 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 // These classes wrap the information about a call or function 11 // definition used to handle ABI compliancy. 12 // 13 //===----------------------------------------------------------------------===// 14 15 #include "CGCall.h" 16 #include "CGCXXABI.h" 17 #include "ABIInfo.h" 18 #include "CodeGenFunction.h" 19 #include "CodeGenModule.h" 20 #include "clang/Basic/TargetInfo.h" 21 #include "clang/AST/Decl.h" 22 #include "clang/AST/DeclCXX.h" 23 #include "clang/AST/DeclObjC.h" 24 #include "clang/Frontend/CodeGenOptions.h" 25 #include "llvm/Attributes.h" 26 #include "llvm/Support/CallSite.h" 27 #include "llvm/Target/TargetData.h" 28 #include "llvm/InlineAsm.h" 29 #include "llvm/Transforms/Utils/Local.h" 30 using namespace clang; 31 using namespace CodeGen; 32 33 /***/ 34 35 static unsigned ClangCallConvToLLVMCallConv(CallingConv CC) { 36 switch (CC) { 37 default: return llvm::CallingConv::C; 38 case CC_X86StdCall: return llvm::CallingConv::X86_StdCall; 39 case CC_X86FastCall: return llvm::CallingConv::X86_FastCall; 40 case CC_X86ThisCall: return llvm::CallingConv::X86_ThisCall; 41 case CC_AAPCS: return llvm::CallingConv::ARM_AAPCS; 42 case CC_AAPCS_VFP: return llvm::CallingConv::ARM_AAPCS_VFP; 43 // TODO: add support for CC_X86Pascal to llvm 44 } 45 } 46 47 /// Derives the 'this' type for codegen purposes, i.e. ignoring method 48 /// qualification. 49 /// FIXME: address space qualification? 50 static CanQualType GetThisType(ASTContext &Context, const CXXRecordDecl *RD) { 51 QualType RecTy = Context.getTagDeclType(RD)->getCanonicalTypeInternal(); 52 return Context.getPointerType(CanQualType::CreateUnsafe(RecTy)); 53 } 54 55 /// Returns the canonical formal type of the given C++ method. 56 static CanQual<FunctionProtoType> GetFormalType(const CXXMethodDecl *MD) { 57 return MD->getType()->getCanonicalTypeUnqualified() 58 .getAs<FunctionProtoType>(); 59 } 60 61 /// Returns the "extra-canonicalized" return type, which discards 62 /// qualifiers on the return type. Codegen doesn't care about them, 63 /// and it makes ABI code a little easier to be able to assume that 64 /// all parameter and return types are top-level unqualified. 65 static CanQualType GetReturnType(QualType RetTy) { 66 return RetTy->getCanonicalTypeUnqualified().getUnqualifiedType(); 67 } 68 69 const CGFunctionInfo & 70 CodeGenTypes::getFunctionInfo(CanQual<FunctionNoProtoType> FTNP) { 71 return getFunctionInfo(FTNP->getResultType().getUnqualifiedType(), 72 SmallVector<CanQualType, 16>(), 73 FTNP->getExtInfo()); 74 } 75 76 /// \param Args - contains any initial parameters besides those 77 /// in the formal type 78 static const CGFunctionInfo &getFunctionInfo(CodeGenTypes &CGT, 79 SmallVectorImpl<CanQualType> &ArgTys, 80 CanQual<FunctionProtoType> FTP) { 81 // FIXME: Kill copy. 82 for (unsigned i = 0, e = FTP->getNumArgs(); i != e; ++i) 83 ArgTys.push_back(FTP->getArgType(i)); 84 CanQualType ResTy = FTP->getResultType().getUnqualifiedType(); 85 return CGT.getFunctionInfo(ResTy, ArgTys, FTP->getExtInfo()); 86 } 87 88 const CGFunctionInfo & 89 CodeGenTypes::getFunctionInfo(CanQual<FunctionProtoType> FTP) { 90 SmallVector<CanQualType, 16> ArgTys; 91 return ::getFunctionInfo(*this, ArgTys, FTP); 92 } 93 94 static CallingConv getCallingConventionForDecl(const Decl *D) { 95 // Set the appropriate calling convention for the Function. 96 if (D->hasAttr<StdCallAttr>()) 97 return CC_X86StdCall; 98 99 if (D->hasAttr<FastCallAttr>()) 100 return CC_X86FastCall; 101 102 if (D->hasAttr<ThisCallAttr>()) 103 return CC_X86ThisCall; 104 105 if (D->hasAttr<PascalAttr>()) 106 return CC_X86Pascal; 107 108 if (PcsAttr *PCS = D->getAttr<PcsAttr>()) 109 return (PCS->getPCS() == PcsAttr::AAPCS ? CC_AAPCS : CC_AAPCS_VFP); 110 111 return CC_C; 112 } 113 114 const CGFunctionInfo &CodeGenTypes::getFunctionInfo(const CXXRecordDecl *RD, 115 const FunctionProtoType *FTP) { 116 SmallVector<CanQualType, 16> ArgTys; 117 118 // Add the 'this' pointer. 119 ArgTys.push_back(GetThisType(Context, RD)); 120 121 return ::getFunctionInfo(*this, ArgTys, 122 FTP->getCanonicalTypeUnqualified().getAs<FunctionProtoType>()); 123 } 124 125 const CGFunctionInfo &CodeGenTypes::getFunctionInfo(const CXXMethodDecl *MD) { 126 SmallVector<CanQualType, 16> ArgTys; 127 128 assert(!isa<CXXConstructorDecl>(MD) && "wrong method for contructors!"); 129 assert(!isa<CXXDestructorDecl>(MD) && "wrong method for destructors!"); 130 131 // Add the 'this' pointer unless this is a static method. 132 if (MD->isInstance()) 133 ArgTys.push_back(GetThisType(Context, MD->getParent())); 134 135 return ::getFunctionInfo(*this, ArgTys, GetFormalType(MD)); 136 } 137 138 const CGFunctionInfo &CodeGenTypes::getFunctionInfo(const CXXConstructorDecl *D, 139 CXXCtorType Type) { 140 SmallVector<CanQualType, 16> ArgTys; 141 ArgTys.push_back(GetThisType(Context, D->getParent())); 142 CanQualType ResTy = Context.VoidTy; 143 144 TheCXXABI.BuildConstructorSignature(D, Type, ResTy, ArgTys); 145 146 CanQual<FunctionProtoType> FTP = GetFormalType(D); 147 148 // Add the formal parameters. 149 for (unsigned i = 0, e = FTP->getNumArgs(); i != e; ++i) 150 ArgTys.push_back(FTP->getArgType(i)); 151 152 return getFunctionInfo(ResTy, ArgTys, FTP->getExtInfo()); 153 } 154 155 const CGFunctionInfo &CodeGenTypes::getFunctionInfo(const CXXDestructorDecl *D, 156 CXXDtorType Type) { 157 SmallVector<CanQualType, 2> ArgTys; 158 ArgTys.push_back(GetThisType(Context, D->getParent())); 159 CanQualType ResTy = Context.VoidTy; 160 161 TheCXXABI.BuildDestructorSignature(D, Type, ResTy, ArgTys); 162 163 CanQual<FunctionProtoType> FTP = GetFormalType(D); 164 assert(FTP->getNumArgs() == 0 && "dtor with formal parameters"); 165 166 return getFunctionInfo(ResTy, ArgTys, FTP->getExtInfo()); 167 } 168 169 const CGFunctionInfo &CodeGenTypes::getFunctionInfo(const FunctionDecl *FD) { 170 if (const CXXMethodDecl *MD = dyn_cast<CXXMethodDecl>(FD)) 171 if (MD->isInstance()) 172 return getFunctionInfo(MD); 173 174 CanQualType FTy = FD->getType()->getCanonicalTypeUnqualified(); 175 assert(isa<FunctionType>(FTy)); 176 if (isa<FunctionNoProtoType>(FTy)) 177 return getFunctionInfo(FTy.getAs<FunctionNoProtoType>()); 178 assert(isa<FunctionProtoType>(FTy)); 179 return getFunctionInfo(FTy.getAs<FunctionProtoType>()); 180 } 181 182 const CGFunctionInfo &CodeGenTypes::getFunctionInfo(const ObjCMethodDecl *MD) { 183 SmallVector<CanQualType, 16> ArgTys; 184 ArgTys.push_back(Context.getCanonicalParamType(MD->getSelfDecl()->getType())); 185 ArgTys.push_back(Context.getCanonicalParamType(Context.getObjCSelType())); 186 // FIXME: Kill copy? 187 for (ObjCMethodDecl::param_const_iterator i = MD->param_begin(), 188 e = MD->param_end(); i != e; ++i) { 189 ArgTys.push_back(Context.getCanonicalParamType((*i)->getType())); 190 } 191 192 FunctionType::ExtInfo einfo; 193 einfo = einfo.withCallingConv(getCallingConventionForDecl(MD)); 194 195 if (getContext().getLangOptions().ObjCAutoRefCount && 196 MD->hasAttr<NSReturnsRetainedAttr>()) 197 einfo = einfo.withProducesResult(true); 198 199 return getFunctionInfo(GetReturnType(MD->getResultType()), ArgTys, einfo); 200 } 201 202 const CGFunctionInfo &CodeGenTypes::getFunctionInfo(GlobalDecl GD) { 203 // FIXME: Do we need to handle ObjCMethodDecl? 204 const FunctionDecl *FD = cast<FunctionDecl>(GD.getDecl()); 205 206 if (const CXXConstructorDecl *CD = dyn_cast<CXXConstructorDecl>(FD)) 207 return getFunctionInfo(CD, GD.getCtorType()); 208 209 if (const CXXDestructorDecl *DD = dyn_cast<CXXDestructorDecl>(FD)) 210 return getFunctionInfo(DD, GD.getDtorType()); 211 212 return getFunctionInfo(FD); 213 } 214 215 const CGFunctionInfo &CodeGenTypes::getFunctionInfo(QualType ResTy, 216 const CallArgList &Args, 217 const FunctionType::ExtInfo &Info) { 218 // FIXME: Kill copy. 219 SmallVector<CanQualType, 16> ArgTys; 220 for (CallArgList::const_iterator i = Args.begin(), e = Args.end(); 221 i != e; ++i) 222 ArgTys.push_back(Context.getCanonicalParamType(i->Ty)); 223 return getFunctionInfo(GetReturnType(ResTy), ArgTys, Info); 224 } 225 226 const CGFunctionInfo &CodeGenTypes::getFunctionInfo(QualType ResTy, 227 const FunctionArgList &Args, 228 const FunctionType::ExtInfo &Info) { 229 // FIXME: Kill copy. 230 SmallVector<CanQualType, 16> ArgTys; 231 for (FunctionArgList::const_iterator i = Args.begin(), e = Args.end(); 232 i != e; ++i) 233 ArgTys.push_back(Context.getCanonicalParamType((*i)->getType())); 234 return getFunctionInfo(GetReturnType(ResTy), ArgTys, Info); 235 } 236 237 const CGFunctionInfo &CodeGenTypes::getNullaryFunctionInfo() { 238 SmallVector<CanQualType, 1> args; 239 return getFunctionInfo(getContext().VoidTy, args, FunctionType::ExtInfo()); 240 } 241 242 const CGFunctionInfo &CodeGenTypes::getFunctionInfo(CanQualType ResTy, 243 const SmallVectorImpl<CanQualType> &ArgTys, 244 const FunctionType::ExtInfo &Info) { 245 #ifndef NDEBUG 246 for (SmallVectorImpl<CanQualType>::const_iterator 247 I = ArgTys.begin(), E = ArgTys.end(); I != E; ++I) 248 assert(I->isCanonicalAsParam()); 249 #endif 250 251 unsigned CC = ClangCallConvToLLVMCallConv(Info.getCC()); 252 253 // Lookup or create unique function info. 254 llvm::FoldingSetNodeID ID; 255 CGFunctionInfo::Profile(ID, Info, ResTy, ArgTys.begin(), ArgTys.end()); 256 257 void *InsertPos = 0; 258 CGFunctionInfo *FI = FunctionInfos.FindNodeOrInsertPos(ID, InsertPos); 259 if (FI) 260 return *FI; 261 262 // Construct the function info. 263 FI = new CGFunctionInfo(CC, Info.getNoReturn(), Info.getProducesResult(), 264 Info.getHasRegParm(), Info.getRegParm(), ResTy, 265 ArgTys.data(), ArgTys.size()); 266 FunctionInfos.InsertNode(FI, InsertPos); 267 268 bool Inserted = FunctionsBeingProcessed.insert(FI); (void)Inserted; 269 assert(Inserted && "Recursively being processed?"); 270 271 // Compute ABI information. 272 getABIInfo().computeInfo(*FI); 273 274 // Loop over all of the computed argument and return value info. If any of 275 // them are direct or extend without a specified coerce type, specify the 276 // default now. 277 ABIArgInfo &RetInfo = FI->getReturnInfo(); 278 if (RetInfo.canHaveCoerceToType() && RetInfo.getCoerceToType() == 0) 279 RetInfo.setCoerceToType(ConvertType(FI->getReturnType())); 280 281 for (CGFunctionInfo::arg_iterator I = FI->arg_begin(), E = FI->arg_end(); 282 I != E; ++I) 283 if (I->info.canHaveCoerceToType() && I->info.getCoerceToType() == 0) 284 I->info.setCoerceToType(ConvertType(I->type)); 285 286 bool Erased = FunctionsBeingProcessed.erase(FI); (void)Erased; 287 assert(Erased && "Not in set?"); 288 289 return *FI; 290 } 291 292 CGFunctionInfo::CGFunctionInfo(unsigned _CallingConvention, 293 bool _NoReturn, bool returnsRetained, 294 bool _HasRegParm, unsigned _RegParm, 295 CanQualType ResTy, 296 const CanQualType *ArgTys, 297 unsigned NumArgTys) 298 : CallingConvention(_CallingConvention), 299 EffectiveCallingConvention(_CallingConvention), 300 NoReturn(_NoReturn), ReturnsRetained(returnsRetained), 301 HasRegParm(_HasRegParm), RegParm(_RegParm) 302 { 303 NumArgs = NumArgTys; 304 305 // FIXME: Coallocate with the CGFunctionInfo object. 306 Args = new ArgInfo[1 + NumArgTys]; 307 Args[0].type = ResTy; 308 for (unsigned i = 0; i != NumArgTys; ++i) 309 Args[1 + i].type = ArgTys[i]; 310 } 311 312 /***/ 313 314 void CodeGenTypes::GetExpandedTypes(QualType type, 315 SmallVectorImpl<llvm::Type*> &expandedTypes) { 316 if (const ConstantArrayType *AT = Context.getAsConstantArrayType(type)) { 317 uint64_t NumElts = AT->getSize().getZExtValue(); 318 for (uint64_t Elt = 0; Elt < NumElts; ++Elt) 319 GetExpandedTypes(AT->getElementType(), expandedTypes); 320 } else if (const RecordType *RT = type->getAsStructureType()) { 321 const RecordDecl *RD = RT->getDecl(); 322 assert(!RD->hasFlexibleArrayMember() && 323 "Cannot expand structure with flexible array."); 324 for (RecordDecl::field_iterator i = RD->field_begin(), e = RD->field_end(); 325 i != e; ++i) { 326 const FieldDecl *FD = *i; 327 assert(!FD->isBitField() && 328 "Cannot expand structure with bit-field members."); 329 GetExpandedTypes(FD->getType(), expandedTypes); 330 } 331 } else if (const ComplexType *CT = type->getAs<ComplexType>()) { 332 llvm::Type *EltTy = ConvertType(CT->getElementType()); 333 expandedTypes.push_back(EltTy); 334 expandedTypes.push_back(EltTy); 335 } else 336 expandedTypes.push_back(ConvertType(type)); 337 } 338 339 llvm::Function::arg_iterator 340 CodeGenFunction::ExpandTypeFromArgs(QualType Ty, LValue LV, 341 llvm::Function::arg_iterator AI) { 342 assert(LV.isSimple() && 343 "Unexpected non-simple lvalue during struct expansion."); 344 llvm::Value *Addr = LV.getAddress(); 345 346 if (const ConstantArrayType *AT = getContext().getAsConstantArrayType(Ty)) { 347 unsigned NumElts = AT->getSize().getZExtValue(); 348 QualType EltTy = AT->getElementType(); 349 for (unsigned Elt = 0; Elt < NumElts; ++Elt) { 350 llvm::Value *EltAddr = Builder.CreateConstGEP2_32(Addr, 0, Elt); 351 LValue LV = MakeAddrLValue(EltAddr, EltTy); 352 AI = ExpandTypeFromArgs(EltTy, LV, AI); 353 } 354 } else if (const RecordType *RT = Ty->getAsStructureType()) { 355 RecordDecl *RD = RT->getDecl(); 356 for (RecordDecl::field_iterator i = RD->field_begin(), e = RD->field_end(); 357 i != e; ++i) { 358 FieldDecl *FD = *i; 359 QualType FT = FD->getType(); 360 361 // FIXME: What are the right qualifiers here? 362 LValue LV = EmitLValueForField(Addr, FD, 0); 363 AI = ExpandTypeFromArgs(FT, LV, AI); 364 } 365 } else if (const ComplexType *CT = Ty->getAs<ComplexType>()) { 366 QualType EltTy = CT->getElementType(); 367 llvm::Value *RealAddr = Builder.CreateStructGEP(Addr, 0, "real"); 368 EmitStoreThroughLValue(RValue::get(AI++), MakeAddrLValue(RealAddr, EltTy)); 369 llvm::Value *ImagAddr = Builder.CreateStructGEP(Addr, 0, "imag"); 370 EmitStoreThroughLValue(RValue::get(AI++), MakeAddrLValue(ImagAddr, EltTy)); 371 } else { 372 EmitStoreThroughLValue(RValue::get(AI), LV); 373 ++AI; 374 } 375 376 return AI; 377 } 378 379 /// EnterStructPointerForCoercedAccess - Given a struct pointer that we are 380 /// accessing some number of bytes out of it, try to gep into the struct to get 381 /// at its inner goodness. Dive as deep as possible without entering an element 382 /// with an in-memory size smaller than DstSize. 383 static llvm::Value * 384 EnterStructPointerForCoercedAccess(llvm::Value *SrcPtr, 385 llvm::StructType *SrcSTy, 386 uint64_t DstSize, CodeGenFunction &CGF) { 387 // We can't dive into a zero-element struct. 388 if (SrcSTy->getNumElements() == 0) return SrcPtr; 389 390 llvm::Type *FirstElt = SrcSTy->getElementType(0); 391 392 // If the first elt is at least as large as what we're looking for, or if the 393 // first element is the same size as the whole struct, we can enter it. 394 uint64_t FirstEltSize = 395 CGF.CGM.getTargetData().getTypeAllocSize(FirstElt); 396 if (FirstEltSize < DstSize && 397 FirstEltSize < CGF.CGM.getTargetData().getTypeAllocSize(SrcSTy)) 398 return SrcPtr; 399 400 // GEP into the first element. 401 SrcPtr = CGF.Builder.CreateConstGEP2_32(SrcPtr, 0, 0, "coerce.dive"); 402 403 // If the first element is a struct, recurse. 404 llvm::Type *SrcTy = 405 cast<llvm::PointerType>(SrcPtr->getType())->getElementType(); 406 if (llvm::StructType *SrcSTy = dyn_cast<llvm::StructType>(SrcTy)) 407 return EnterStructPointerForCoercedAccess(SrcPtr, SrcSTy, DstSize, CGF); 408 409 return SrcPtr; 410 } 411 412 /// CoerceIntOrPtrToIntOrPtr - Convert a value Val to the specific Ty where both 413 /// are either integers or pointers. This does a truncation of the value if it 414 /// is too large or a zero extension if it is too small. 415 static llvm::Value *CoerceIntOrPtrToIntOrPtr(llvm::Value *Val, 416 llvm::Type *Ty, 417 CodeGenFunction &CGF) { 418 if (Val->getType() == Ty) 419 return Val; 420 421 if (isa<llvm::PointerType>(Val->getType())) { 422 // If this is Pointer->Pointer avoid conversion to and from int. 423 if (isa<llvm::PointerType>(Ty)) 424 return CGF.Builder.CreateBitCast(Val, Ty, "coerce.val"); 425 426 // Convert the pointer to an integer so we can play with its width. 427 Val = CGF.Builder.CreatePtrToInt(Val, CGF.IntPtrTy, "coerce.val.pi"); 428 } 429 430 llvm::Type *DestIntTy = Ty; 431 if (isa<llvm::PointerType>(DestIntTy)) 432 DestIntTy = CGF.IntPtrTy; 433 434 if (Val->getType() != DestIntTy) 435 Val = CGF.Builder.CreateIntCast(Val, DestIntTy, false, "coerce.val.ii"); 436 437 if (isa<llvm::PointerType>(Ty)) 438 Val = CGF.Builder.CreateIntToPtr(Val, Ty, "coerce.val.ip"); 439 return Val; 440 } 441 442 443 444 /// CreateCoercedLoad - Create a load from \arg SrcPtr interpreted as 445 /// a pointer to an object of type \arg Ty. 446 /// 447 /// This safely handles the case when the src type is smaller than the 448 /// destination type; in this situation the values of bits which not 449 /// present in the src are undefined. 450 static llvm::Value *CreateCoercedLoad(llvm::Value *SrcPtr, 451 llvm::Type *Ty, 452 CodeGenFunction &CGF) { 453 llvm::Type *SrcTy = 454 cast<llvm::PointerType>(SrcPtr->getType())->getElementType(); 455 456 // If SrcTy and Ty are the same, just do a load. 457 if (SrcTy == Ty) 458 return CGF.Builder.CreateLoad(SrcPtr); 459 460 uint64_t DstSize = CGF.CGM.getTargetData().getTypeAllocSize(Ty); 461 462 if (llvm::StructType *SrcSTy = dyn_cast<llvm::StructType>(SrcTy)) { 463 SrcPtr = EnterStructPointerForCoercedAccess(SrcPtr, SrcSTy, DstSize, CGF); 464 SrcTy = cast<llvm::PointerType>(SrcPtr->getType())->getElementType(); 465 } 466 467 uint64_t SrcSize = CGF.CGM.getTargetData().getTypeAllocSize(SrcTy); 468 469 // If the source and destination are integer or pointer types, just do an 470 // extension or truncation to the desired type. 471 if ((isa<llvm::IntegerType>(Ty) || isa<llvm::PointerType>(Ty)) && 472 (isa<llvm::IntegerType>(SrcTy) || isa<llvm::PointerType>(SrcTy))) { 473 llvm::LoadInst *Load = CGF.Builder.CreateLoad(SrcPtr); 474 return CoerceIntOrPtrToIntOrPtr(Load, Ty, CGF); 475 } 476 477 // If load is legal, just bitcast the src pointer. 478 if (SrcSize >= DstSize) { 479 // Generally SrcSize is never greater than DstSize, since this means we are 480 // losing bits. However, this can happen in cases where the structure has 481 // additional padding, for example due to a user specified alignment. 482 // 483 // FIXME: Assert that we aren't truncating non-padding bits when have access 484 // to that information. 485 llvm::Value *Casted = 486 CGF.Builder.CreateBitCast(SrcPtr, llvm::PointerType::getUnqual(Ty)); 487 llvm::LoadInst *Load = CGF.Builder.CreateLoad(Casted); 488 // FIXME: Use better alignment / avoid requiring aligned load. 489 Load->setAlignment(1); 490 return Load; 491 } 492 493 // Otherwise do coercion through memory. This is stupid, but 494 // simple. 495 llvm::Value *Tmp = CGF.CreateTempAlloca(Ty); 496 llvm::Value *Casted = 497 CGF.Builder.CreateBitCast(Tmp, llvm::PointerType::getUnqual(SrcTy)); 498 llvm::StoreInst *Store = 499 CGF.Builder.CreateStore(CGF.Builder.CreateLoad(SrcPtr), Casted); 500 // FIXME: Use better alignment / avoid requiring aligned store. 501 Store->setAlignment(1); 502 return CGF.Builder.CreateLoad(Tmp); 503 } 504 505 // Function to store a first-class aggregate into memory. We prefer to 506 // store the elements rather than the aggregate to be more friendly to 507 // fast-isel. 508 // FIXME: Do we need to recurse here? 509 static void BuildAggStore(CodeGenFunction &CGF, llvm::Value *Val, 510 llvm::Value *DestPtr, bool DestIsVolatile, 511 bool LowAlignment) { 512 // Prefer scalar stores to first-class aggregate stores. 513 if (llvm::StructType *STy = 514 dyn_cast<llvm::StructType>(Val->getType())) { 515 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 516 llvm::Value *EltPtr = CGF.Builder.CreateConstGEP2_32(DestPtr, 0, i); 517 llvm::Value *Elt = CGF.Builder.CreateExtractValue(Val, i); 518 llvm::StoreInst *SI = CGF.Builder.CreateStore(Elt, EltPtr, 519 DestIsVolatile); 520 if (LowAlignment) 521 SI->setAlignment(1); 522 } 523 } else { 524 CGF.Builder.CreateStore(Val, DestPtr, DestIsVolatile); 525 } 526 } 527 528 /// CreateCoercedStore - Create a store to \arg DstPtr from \arg Src, 529 /// where the source and destination may have different types. 530 /// 531 /// This safely handles the case when the src type is larger than the 532 /// destination type; the upper bits of the src will be lost. 533 static void CreateCoercedStore(llvm::Value *Src, 534 llvm::Value *DstPtr, 535 bool DstIsVolatile, 536 CodeGenFunction &CGF) { 537 llvm::Type *SrcTy = Src->getType(); 538 llvm::Type *DstTy = 539 cast<llvm::PointerType>(DstPtr->getType())->getElementType(); 540 if (SrcTy == DstTy) { 541 CGF.Builder.CreateStore(Src, DstPtr, DstIsVolatile); 542 return; 543 } 544 545 uint64_t SrcSize = CGF.CGM.getTargetData().getTypeAllocSize(SrcTy); 546 547 if (llvm::StructType *DstSTy = dyn_cast<llvm::StructType>(DstTy)) { 548 DstPtr = EnterStructPointerForCoercedAccess(DstPtr, DstSTy, SrcSize, CGF); 549 DstTy = cast<llvm::PointerType>(DstPtr->getType())->getElementType(); 550 } 551 552 // If the source and destination are integer or pointer types, just do an 553 // extension or truncation to the desired type. 554 if ((isa<llvm::IntegerType>(SrcTy) || isa<llvm::PointerType>(SrcTy)) && 555 (isa<llvm::IntegerType>(DstTy) || isa<llvm::PointerType>(DstTy))) { 556 Src = CoerceIntOrPtrToIntOrPtr(Src, DstTy, CGF); 557 CGF.Builder.CreateStore(Src, DstPtr, DstIsVolatile); 558 return; 559 } 560 561 uint64_t DstSize = CGF.CGM.getTargetData().getTypeAllocSize(DstTy); 562 563 // If store is legal, just bitcast the src pointer. 564 if (SrcSize <= DstSize) { 565 llvm::Value *Casted = 566 CGF.Builder.CreateBitCast(DstPtr, llvm::PointerType::getUnqual(SrcTy)); 567 // FIXME: Use better alignment / avoid requiring aligned store. 568 BuildAggStore(CGF, Src, Casted, DstIsVolatile, true); 569 } else { 570 // Otherwise do coercion through memory. This is stupid, but 571 // simple. 572 573 // Generally SrcSize is never greater than DstSize, since this means we are 574 // losing bits. However, this can happen in cases where the structure has 575 // additional padding, for example due to a user specified alignment. 576 // 577 // FIXME: Assert that we aren't truncating non-padding bits when have access 578 // to that information. 579 llvm::Value *Tmp = CGF.CreateTempAlloca(SrcTy); 580 CGF.Builder.CreateStore(Src, Tmp); 581 llvm::Value *Casted = 582 CGF.Builder.CreateBitCast(Tmp, llvm::PointerType::getUnqual(DstTy)); 583 llvm::LoadInst *Load = CGF.Builder.CreateLoad(Casted); 584 // FIXME: Use better alignment / avoid requiring aligned load. 585 Load->setAlignment(1); 586 CGF.Builder.CreateStore(Load, DstPtr, DstIsVolatile); 587 } 588 } 589 590 /***/ 591 592 bool CodeGenModule::ReturnTypeUsesSRet(const CGFunctionInfo &FI) { 593 return FI.getReturnInfo().isIndirect(); 594 } 595 596 bool CodeGenModule::ReturnTypeUsesFPRet(QualType ResultType) { 597 if (const BuiltinType *BT = ResultType->getAs<BuiltinType>()) { 598 switch (BT->getKind()) { 599 default: 600 return false; 601 case BuiltinType::Float: 602 return getContext().getTargetInfo().useObjCFPRetForRealType(TargetInfo::Float); 603 case BuiltinType::Double: 604 return getContext().getTargetInfo().useObjCFPRetForRealType(TargetInfo::Double); 605 case BuiltinType::LongDouble: 606 return getContext().getTargetInfo().useObjCFPRetForRealType( 607 TargetInfo::LongDouble); 608 } 609 } 610 611 return false; 612 } 613 614 llvm::FunctionType *CodeGenTypes::GetFunctionType(GlobalDecl GD) { 615 const CGFunctionInfo &FI = getFunctionInfo(GD); 616 617 // For definition purposes, don't consider a K&R function variadic. 618 bool Variadic = false; 619 if (const FunctionProtoType *FPT = 620 cast<FunctionDecl>(GD.getDecl())->getType()->getAs<FunctionProtoType>()) 621 Variadic = FPT->isVariadic(); 622 623 return GetFunctionType(FI, Variadic); 624 } 625 626 llvm::FunctionType * 627 CodeGenTypes::GetFunctionType(const CGFunctionInfo &FI, bool isVariadic) { 628 629 bool Inserted = FunctionsBeingProcessed.insert(&FI); (void)Inserted; 630 assert(Inserted && "Recursively being processed?"); 631 632 SmallVector<llvm::Type*, 8> argTypes; 633 llvm::Type *resultType = 0; 634 635 const ABIArgInfo &retAI = FI.getReturnInfo(); 636 switch (retAI.getKind()) { 637 case ABIArgInfo::Expand: 638 llvm_unreachable("Invalid ABI kind for return argument"); 639 640 case ABIArgInfo::Extend: 641 case ABIArgInfo::Direct: 642 resultType = retAI.getCoerceToType(); 643 break; 644 645 case ABIArgInfo::Indirect: { 646 assert(!retAI.getIndirectAlign() && "Align unused on indirect return."); 647 resultType = llvm::Type::getVoidTy(getLLVMContext()); 648 649 QualType ret = FI.getReturnType(); 650 llvm::Type *ty = ConvertType(ret); 651 unsigned addressSpace = Context.getTargetAddressSpace(ret); 652 argTypes.push_back(llvm::PointerType::get(ty, addressSpace)); 653 break; 654 } 655 656 case ABIArgInfo::Ignore: 657 resultType = llvm::Type::getVoidTy(getLLVMContext()); 658 break; 659 } 660 661 for (CGFunctionInfo::const_arg_iterator it = FI.arg_begin(), 662 ie = FI.arg_end(); it != ie; ++it) { 663 const ABIArgInfo &argAI = it->info; 664 665 switch (argAI.getKind()) { 666 case ABIArgInfo::Ignore: 667 break; 668 669 case ABIArgInfo::Indirect: { 670 // indirect arguments are always on the stack, which is addr space #0. 671 llvm::Type *LTy = ConvertTypeForMem(it->type); 672 argTypes.push_back(LTy->getPointerTo()); 673 break; 674 } 675 676 case ABIArgInfo::Extend: 677 case ABIArgInfo::Direct: { 678 // If the coerce-to type is a first class aggregate, flatten it. Either 679 // way is semantically identical, but fast-isel and the optimizer 680 // generally likes scalar values better than FCAs. 681 llvm::Type *argType = argAI.getCoerceToType(); 682 if (llvm::StructType *st = dyn_cast<llvm::StructType>(argType)) { 683 for (unsigned i = 0, e = st->getNumElements(); i != e; ++i) 684 argTypes.push_back(st->getElementType(i)); 685 } else { 686 argTypes.push_back(argType); 687 } 688 break; 689 } 690 691 case ABIArgInfo::Expand: 692 GetExpandedTypes(it->type, argTypes); 693 break; 694 } 695 } 696 697 bool Erased = FunctionsBeingProcessed.erase(&FI); (void)Erased; 698 assert(Erased && "Not in set?"); 699 700 return llvm::FunctionType::get(resultType, argTypes, isVariadic); 701 } 702 703 llvm::Type *CodeGenTypes::GetFunctionTypeForVTable(GlobalDecl GD) { 704 const CXXMethodDecl *MD = cast<CXXMethodDecl>(GD.getDecl()); 705 const FunctionProtoType *FPT = MD->getType()->getAs<FunctionProtoType>(); 706 707 if (!isFuncTypeConvertible(FPT)) 708 return llvm::StructType::get(getLLVMContext()); 709 710 const CGFunctionInfo *Info; 711 if (isa<CXXDestructorDecl>(MD)) 712 Info = &getFunctionInfo(cast<CXXDestructorDecl>(MD), GD.getDtorType()); 713 else 714 Info = &getFunctionInfo(MD); 715 return GetFunctionType(*Info, FPT->isVariadic()); 716 } 717 718 void CodeGenModule::ConstructAttributeList(const CGFunctionInfo &FI, 719 const Decl *TargetDecl, 720 AttributeListType &PAL, 721 unsigned &CallingConv) { 722 unsigned FuncAttrs = 0; 723 unsigned RetAttrs = 0; 724 725 CallingConv = FI.getEffectiveCallingConvention(); 726 727 if (FI.isNoReturn()) 728 FuncAttrs |= llvm::Attribute::NoReturn; 729 730 // FIXME: handle sseregparm someday... 731 if (TargetDecl) { 732 if (TargetDecl->hasAttr<ReturnsTwiceAttr>()) 733 FuncAttrs |= llvm::Attribute::ReturnsTwice; 734 if (TargetDecl->hasAttr<NoThrowAttr>()) 735 FuncAttrs |= llvm::Attribute::NoUnwind; 736 else if (const FunctionDecl *Fn = dyn_cast<FunctionDecl>(TargetDecl)) { 737 const FunctionProtoType *FPT = Fn->getType()->getAs<FunctionProtoType>(); 738 if (FPT && FPT->isNothrow(getContext())) 739 FuncAttrs |= llvm::Attribute::NoUnwind; 740 } 741 742 if (TargetDecl->hasAttr<NoReturnAttr>()) 743 FuncAttrs |= llvm::Attribute::NoReturn; 744 745 if (TargetDecl->hasAttr<ReturnsTwiceAttr>()) 746 FuncAttrs |= llvm::Attribute::ReturnsTwice; 747 748 // 'const' and 'pure' attribute functions are also nounwind. 749 if (TargetDecl->hasAttr<ConstAttr>()) { 750 FuncAttrs |= llvm::Attribute::ReadNone; 751 FuncAttrs |= llvm::Attribute::NoUnwind; 752 } else if (TargetDecl->hasAttr<PureAttr>()) { 753 FuncAttrs |= llvm::Attribute::ReadOnly; 754 FuncAttrs |= llvm::Attribute::NoUnwind; 755 } 756 if (TargetDecl->hasAttr<MallocAttr>()) 757 RetAttrs |= llvm::Attribute::NoAlias; 758 } 759 760 if (CodeGenOpts.OptimizeSize) 761 FuncAttrs |= llvm::Attribute::OptimizeForSize; 762 if (CodeGenOpts.DisableRedZone) 763 FuncAttrs |= llvm::Attribute::NoRedZone; 764 if (CodeGenOpts.NoImplicitFloat) 765 FuncAttrs |= llvm::Attribute::NoImplicitFloat; 766 767 QualType RetTy = FI.getReturnType(); 768 unsigned Index = 1; 769 const ABIArgInfo &RetAI = FI.getReturnInfo(); 770 switch (RetAI.getKind()) { 771 case ABIArgInfo::Extend: 772 if (RetTy->hasSignedIntegerRepresentation()) 773 RetAttrs |= llvm::Attribute::SExt; 774 else if (RetTy->hasUnsignedIntegerRepresentation()) 775 RetAttrs |= llvm::Attribute::ZExt; 776 break; 777 case ABIArgInfo::Direct: 778 case ABIArgInfo::Ignore: 779 break; 780 781 case ABIArgInfo::Indirect: 782 PAL.push_back(llvm::AttributeWithIndex::get(Index, 783 llvm::Attribute::StructRet)); 784 ++Index; 785 // sret disables readnone and readonly 786 FuncAttrs &= ~(llvm::Attribute::ReadOnly | 787 llvm::Attribute::ReadNone); 788 break; 789 790 case ABIArgInfo::Expand: 791 llvm_unreachable("Invalid ABI kind for return argument"); 792 } 793 794 if (RetAttrs) 795 PAL.push_back(llvm::AttributeWithIndex::get(0, RetAttrs)); 796 797 // FIXME: RegParm should be reduced in case of global register variable. 798 signed RegParm; 799 if (FI.getHasRegParm()) 800 RegParm = FI.getRegParm(); 801 else 802 RegParm = CodeGenOpts.NumRegisterParameters; 803 804 unsigned PointerWidth = getContext().getTargetInfo().getPointerWidth(0); 805 for (CGFunctionInfo::const_arg_iterator it = FI.arg_begin(), 806 ie = FI.arg_end(); it != ie; ++it) { 807 QualType ParamType = it->type; 808 const ABIArgInfo &AI = it->info; 809 unsigned Attributes = 0; 810 811 // 'restrict' -> 'noalias' is done in EmitFunctionProlog when we 812 // have the corresponding parameter variable. It doesn't make 813 // sense to do it here because parameters are so messed up. 814 switch (AI.getKind()) { 815 case ABIArgInfo::Extend: 816 if (ParamType->isSignedIntegerOrEnumerationType()) 817 Attributes |= llvm::Attribute::SExt; 818 else if (ParamType->isUnsignedIntegerOrEnumerationType()) 819 Attributes |= llvm::Attribute::ZExt; 820 // FALL THROUGH 821 case ABIArgInfo::Direct: 822 if (RegParm > 0 && 823 (ParamType->isIntegerType() || ParamType->isPointerType())) { 824 RegParm -= 825 (Context.getTypeSize(ParamType) + PointerWidth - 1) / PointerWidth; 826 if (RegParm >= 0) 827 Attributes |= llvm::Attribute::InReg; 828 } 829 // FIXME: handle sseregparm someday... 830 831 if (llvm::StructType *STy = 832 dyn_cast<llvm::StructType>(AI.getCoerceToType())) 833 Index += STy->getNumElements()-1; // 1 will be added below. 834 break; 835 836 case ABIArgInfo::Indirect: 837 if (AI.getIndirectByVal()) 838 Attributes |= llvm::Attribute::ByVal; 839 840 Attributes |= 841 llvm::Attribute::constructAlignmentFromInt(AI.getIndirectAlign()); 842 // byval disables readnone and readonly. 843 FuncAttrs &= ~(llvm::Attribute::ReadOnly | 844 llvm::Attribute::ReadNone); 845 break; 846 847 case ABIArgInfo::Ignore: 848 // Skip increment, no matching LLVM parameter. 849 continue; 850 851 case ABIArgInfo::Expand: { 852 SmallVector<llvm::Type*, 8> types; 853 // FIXME: This is rather inefficient. Do we ever actually need to do 854 // anything here? The result should be just reconstructed on the other 855 // side, so extension should be a non-issue. 856 getTypes().GetExpandedTypes(ParamType, types); 857 Index += types.size(); 858 continue; 859 } 860 } 861 862 if (Attributes) 863 PAL.push_back(llvm::AttributeWithIndex::get(Index, Attributes)); 864 ++Index; 865 } 866 if (FuncAttrs) 867 PAL.push_back(llvm::AttributeWithIndex::get(~0, FuncAttrs)); 868 } 869 870 /// An argument came in as a promoted argument; demote it back to its 871 /// declared type. 872 static llvm::Value *emitArgumentDemotion(CodeGenFunction &CGF, 873 const VarDecl *var, 874 llvm::Value *value) { 875 llvm::Type *varType = CGF.ConvertType(var->getType()); 876 877 // This can happen with promotions that actually don't change the 878 // underlying type, like the enum promotions. 879 if (value->getType() == varType) return value; 880 881 assert((varType->isIntegerTy() || varType->isFloatingPointTy()) 882 && "unexpected promotion type"); 883 884 if (isa<llvm::IntegerType>(varType)) 885 return CGF.Builder.CreateTrunc(value, varType, "arg.unpromote"); 886 887 return CGF.Builder.CreateFPCast(value, varType, "arg.unpromote"); 888 } 889 890 void CodeGenFunction::EmitFunctionProlog(const CGFunctionInfo &FI, 891 llvm::Function *Fn, 892 const FunctionArgList &Args) { 893 // If this is an implicit-return-zero function, go ahead and 894 // initialize the return value. TODO: it might be nice to have 895 // a more general mechanism for this that didn't require synthesized 896 // return statements. 897 if (const FunctionDecl *FD = dyn_cast_or_null<FunctionDecl>(CurFuncDecl)) { 898 if (FD->hasImplicitReturnZero()) { 899 QualType RetTy = FD->getResultType().getUnqualifiedType(); 900 llvm::Type* LLVMTy = CGM.getTypes().ConvertType(RetTy); 901 llvm::Constant* Zero = llvm::Constant::getNullValue(LLVMTy); 902 Builder.CreateStore(Zero, ReturnValue); 903 } 904 } 905 906 // FIXME: We no longer need the types from FunctionArgList; lift up and 907 // simplify. 908 909 // Emit allocs for param decls. Give the LLVM Argument nodes names. 910 llvm::Function::arg_iterator AI = Fn->arg_begin(); 911 912 // Name the struct return argument. 913 if (CGM.ReturnTypeUsesSRet(FI)) { 914 AI->setName("agg.result"); 915 AI->addAttr(llvm::Attribute::NoAlias); 916 ++AI; 917 } 918 919 assert(FI.arg_size() == Args.size() && 920 "Mismatch between function signature & arguments."); 921 unsigned ArgNo = 1; 922 CGFunctionInfo::const_arg_iterator info_it = FI.arg_begin(); 923 for (FunctionArgList::const_iterator i = Args.begin(), e = Args.end(); 924 i != e; ++i, ++info_it, ++ArgNo) { 925 const VarDecl *Arg = *i; 926 QualType Ty = info_it->type; 927 const ABIArgInfo &ArgI = info_it->info; 928 929 bool isPromoted = 930 isa<ParmVarDecl>(Arg) && cast<ParmVarDecl>(Arg)->isKNRPromoted(); 931 932 switch (ArgI.getKind()) { 933 case ABIArgInfo::Indirect: { 934 llvm::Value *V = AI; 935 936 if (hasAggregateLLVMType(Ty)) { 937 // Aggregates and complex variables are accessed by reference. All we 938 // need to do is realign the value, if requested 939 if (ArgI.getIndirectRealign()) { 940 llvm::Value *AlignedTemp = CreateMemTemp(Ty, "coerce"); 941 942 // Copy from the incoming argument pointer to the temporary with the 943 // appropriate alignment. 944 // 945 // FIXME: We should have a common utility for generating an aggregate 946 // copy. 947 llvm::Type *I8PtrTy = Builder.getInt8PtrTy(); 948 CharUnits Size = getContext().getTypeSizeInChars(Ty); 949 llvm::Value *Dst = Builder.CreateBitCast(AlignedTemp, I8PtrTy); 950 llvm::Value *Src = Builder.CreateBitCast(V, I8PtrTy); 951 Builder.CreateMemCpy(Dst, 952 Src, 953 llvm::ConstantInt::get(IntPtrTy, 954 Size.getQuantity()), 955 ArgI.getIndirectAlign(), 956 false); 957 V = AlignedTemp; 958 } 959 } else { 960 // Load scalar value from indirect argument. 961 CharUnits Alignment = getContext().getTypeAlignInChars(Ty); 962 V = EmitLoadOfScalar(V, false, Alignment.getQuantity(), Ty); 963 964 if (isPromoted) 965 V = emitArgumentDemotion(*this, Arg, V); 966 } 967 EmitParmDecl(*Arg, V, ArgNo); 968 break; 969 } 970 971 case ABIArgInfo::Extend: 972 case ABIArgInfo::Direct: { 973 // If we have the trivial case, handle it with no muss and fuss. 974 if (!isa<llvm::StructType>(ArgI.getCoerceToType()) && 975 ArgI.getCoerceToType() == ConvertType(Ty) && 976 ArgI.getDirectOffset() == 0) { 977 assert(AI != Fn->arg_end() && "Argument mismatch!"); 978 llvm::Value *V = AI; 979 980 if (Arg->getType().isRestrictQualified()) 981 AI->addAttr(llvm::Attribute::NoAlias); 982 983 // Ensure the argument is the correct type. 984 if (V->getType() != ArgI.getCoerceToType()) 985 V = Builder.CreateBitCast(V, ArgI.getCoerceToType()); 986 987 if (isPromoted) 988 V = emitArgumentDemotion(*this, Arg, V); 989 990 EmitParmDecl(*Arg, V, ArgNo); 991 break; 992 } 993 994 llvm::AllocaInst *Alloca = CreateMemTemp(Ty, "coerce"); 995 996 // The alignment we need to use is the max of the requested alignment for 997 // the argument plus the alignment required by our access code below. 998 unsigned AlignmentToUse = 999 CGM.getTargetData().getABITypeAlignment(ArgI.getCoerceToType()); 1000 AlignmentToUse = std::max(AlignmentToUse, 1001 (unsigned)getContext().getDeclAlign(Arg).getQuantity()); 1002 1003 Alloca->setAlignment(AlignmentToUse); 1004 llvm::Value *V = Alloca; 1005 llvm::Value *Ptr = V; // Pointer to store into. 1006 1007 // If the value is offset in memory, apply the offset now. 1008 if (unsigned Offs = ArgI.getDirectOffset()) { 1009 Ptr = Builder.CreateBitCast(Ptr, Builder.getInt8PtrTy()); 1010 Ptr = Builder.CreateConstGEP1_32(Ptr, Offs); 1011 Ptr = Builder.CreateBitCast(Ptr, 1012 llvm::PointerType::getUnqual(ArgI.getCoerceToType())); 1013 } 1014 1015 // If the coerce-to type is a first class aggregate, we flatten it and 1016 // pass the elements. Either way is semantically identical, but fast-isel 1017 // and the optimizer generally likes scalar values better than FCAs. 1018 if (llvm::StructType *STy = 1019 dyn_cast<llvm::StructType>(ArgI.getCoerceToType())) { 1020 Ptr = Builder.CreateBitCast(Ptr, llvm::PointerType::getUnqual(STy)); 1021 1022 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 1023 assert(AI != Fn->arg_end() && "Argument mismatch!"); 1024 AI->setName(Arg->getName() + ".coerce" + Twine(i)); 1025 llvm::Value *EltPtr = Builder.CreateConstGEP2_32(Ptr, 0, i); 1026 Builder.CreateStore(AI++, EltPtr); 1027 } 1028 } else { 1029 // Simple case, just do a coerced store of the argument into the alloca. 1030 assert(AI != Fn->arg_end() && "Argument mismatch!"); 1031 AI->setName(Arg->getName() + ".coerce"); 1032 CreateCoercedStore(AI++, Ptr, /*DestIsVolatile=*/false, *this); 1033 } 1034 1035 1036 // Match to what EmitParmDecl is expecting for this type. 1037 if (!CodeGenFunction::hasAggregateLLVMType(Ty)) { 1038 V = EmitLoadOfScalar(V, false, AlignmentToUse, Ty); 1039 if (isPromoted) 1040 V = emitArgumentDemotion(*this, Arg, V); 1041 } 1042 EmitParmDecl(*Arg, V, ArgNo); 1043 continue; // Skip ++AI increment, already done. 1044 } 1045 1046 case ABIArgInfo::Expand: { 1047 // If this structure was expanded into multiple arguments then 1048 // we need to create a temporary and reconstruct it from the 1049 // arguments. 1050 llvm::Value *Temp = CreateMemTemp(Ty, Arg->getName() + ".addr"); 1051 llvm::Function::arg_iterator End = 1052 ExpandTypeFromArgs(Ty, MakeAddrLValue(Temp, Ty), AI); 1053 EmitParmDecl(*Arg, Temp, ArgNo); 1054 1055 // Name the arguments used in expansion and increment AI. 1056 unsigned Index = 0; 1057 for (; AI != End; ++AI, ++Index) 1058 AI->setName(Arg->getName() + "." + Twine(Index)); 1059 continue; 1060 } 1061 1062 case ABIArgInfo::Ignore: 1063 // Initialize the local variable appropriately. 1064 if (hasAggregateLLVMType(Ty)) 1065 EmitParmDecl(*Arg, CreateMemTemp(Ty), ArgNo); 1066 else 1067 EmitParmDecl(*Arg, llvm::UndefValue::get(ConvertType(Arg->getType())), 1068 ArgNo); 1069 1070 // Skip increment, no matching LLVM parameter. 1071 continue; 1072 } 1073 1074 ++AI; 1075 } 1076 assert(AI == Fn->arg_end() && "Argument mismatch!"); 1077 } 1078 1079 /// Try to emit a fused autorelease of a return result. 1080 static llvm::Value *tryEmitFusedAutoreleaseOfResult(CodeGenFunction &CGF, 1081 llvm::Value *result) { 1082 // We must be immediately followed the cast. 1083 llvm::BasicBlock *BB = CGF.Builder.GetInsertBlock(); 1084 if (BB->empty()) return 0; 1085 if (&BB->back() != result) return 0; 1086 1087 llvm::Type *resultType = result->getType(); 1088 1089 // result is in a BasicBlock and is therefore an Instruction. 1090 llvm::Instruction *generator = cast<llvm::Instruction>(result); 1091 1092 SmallVector<llvm::Instruction*,4> insnsToKill; 1093 1094 // Look for: 1095 // %generator = bitcast %type1* %generator2 to %type2* 1096 while (llvm::BitCastInst *bitcast = dyn_cast<llvm::BitCastInst>(generator)) { 1097 // We would have emitted this as a constant if the operand weren't 1098 // an Instruction. 1099 generator = cast<llvm::Instruction>(bitcast->getOperand(0)); 1100 1101 // Require the generator to be immediately followed by the cast. 1102 if (generator->getNextNode() != bitcast) 1103 return 0; 1104 1105 insnsToKill.push_back(bitcast); 1106 } 1107 1108 // Look for: 1109 // %generator = call i8* @objc_retain(i8* %originalResult) 1110 // or 1111 // %generator = call i8* @objc_retainAutoreleasedReturnValue(i8* %originalResult) 1112 llvm::CallInst *call = dyn_cast<llvm::CallInst>(generator); 1113 if (!call) return 0; 1114 1115 bool doRetainAutorelease; 1116 1117 if (call->getCalledValue() == CGF.CGM.getARCEntrypoints().objc_retain) { 1118 doRetainAutorelease = true; 1119 } else if (call->getCalledValue() == CGF.CGM.getARCEntrypoints() 1120 .objc_retainAutoreleasedReturnValue) { 1121 doRetainAutorelease = false; 1122 1123 // Look for an inline asm immediately preceding the call and kill it, too. 1124 llvm::Instruction *prev = call->getPrevNode(); 1125 if (llvm::CallInst *asmCall = dyn_cast_or_null<llvm::CallInst>(prev)) 1126 if (asmCall->getCalledValue() 1127 == CGF.CGM.getARCEntrypoints().retainAutoreleasedReturnValueMarker) 1128 insnsToKill.push_back(prev); 1129 } else { 1130 return 0; 1131 } 1132 1133 result = call->getArgOperand(0); 1134 insnsToKill.push_back(call); 1135 1136 // Keep killing bitcasts, for sanity. Note that we no longer care 1137 // about precise ordering as long as there's exactly one use. 1138 while (llvm::BitCastInst *bitcast = dyn_cast<llvm::BitCastInst>(result)) { 1139 if (!bitcast->hasOneUse()) break; 1140 insnsToKill.push_back(bitcast); 1141 result = bitcast->getOperand(0); 1142 } 1143 1144 // Delete all the unnecessary instructions, from latest to earliest. 1145 for (SmallVectorImpl<llvm::Instruction*>::iterator 1146 i = insnsToKill.begin(), e = insnsToKill.end(); i != e; ++i) 1147 (*i)->eraseFromParent(); 1148 1149 // Do the fused retain/autorelease if we were asked to. 1150 if (doRetainAutorelease) 1151 result = CGF.EmitARCRetainAutoreleaseReturnValue(result); 1152 1153 // Cast back to the result type. 1154 return CGF.Builder.CreateBitCast(result, resultType); 1155 } 1156 1157 /// Emit an ARC autorelease of the result of a function. 1158 static llvm::Value *emitAutoreleaseOfResult(CodeGenFunction &CGF, 1159 llvm::Value *result) { 1160 // At -O0, try to emit a fused retain/autorelease. 1161 if (CGF.shouldUseFusedARCCalls()) 1162 if (llvm::Value *fused = tryEmitFusedAutoreleaseOfResult(CGF, result)) 1163 return fused; 1164 1165 return CGF.EmitARCAutoreleaseReturnValue(result); 1166 } 1167 1168 void CodeGenFunction::EmitFunctionEpilog(const CGFunctionInfo &FI) { 1169 // Functions with no result always return void. 1170 if (ReturnValue == 0) { 1171 Builder.CreateRetVoid(); 1172 return; 1173 } 1174 1175 llvm::DebugLoc RetDbgLoc; 1176 llvm::Value *RV = 0; 1177 QualType RetTy = FI.getReturnType(); 1178 const ABIArgInfo &RetAI = FI.getReturnInfo(); 1179 1180 switch (RetAI.getKind()) { 1181 case ABIArgInfo::Indirect: { 1182 unsigned Alignment = getContext().getTypeAlignInChars(RetTy).getQuantity(); 1183 if (RetTy->isAnyComplexType()) { 1184 ComplexPairTy RT = LoadComplexFromAddr(ReturnValue, false); 1185 StoreComplexToAddr(RT, CurFn->arg_begin(), false); 1186 } else if (CodeGenFunction::hasAggregateLLVMType(RetTy)) { 1187 // Do nothing; aggregrates get evaluated directly into the destination. 1188 } else { 1189 EmitStoreOfScalar(Builder.CreateLoad(ReturnValue), CurFn->arg_begin(), 1190 false, Alignment, RetTy); 1191 } 1192 break; 1193 } 1194 1195 case ABIArgInfo::Extend: 1196 case ABIArgInfo::Direct: 1197 if (RetAI.getCoerceToType() == ConvertType(RetTy) && 1198 RetAI.getDirectOffset() == 0) { 1199 // The internal return value temp always will have pointer-to-return-type 1200 // type, just do a load. 1201 1202 // If the instruction right before the insertion point is a store to the 1203 // return value, we can elide the load, zap the store, and usually zap the 1204 // alloca. 1205 llvm::BasicBlock *InsertBB = Builder.GetInsertBlock(); 1206 llvm::StoreInst *SI = 0; 1207 if (InsertBB->empty() || 1208 !(SI = dyn_cast<llvm::StoreInst>(&InsertBB->back())) || 1209 SI->getPointerOperand() != ReturnValue || SI->isVolatile()) { 1210 RV = Builder.CreateLoad(ReturnValue); 1211 } else { 1212 // Get the stored value and nuke the now-dead store. 1213 RetDbgLoc = SI->getDebugLoc(); 1214 RV = SI->getValueOperand(); 1215 SI->eraseFromParent(); 1216 1217 // If that was the only use of the return value, nuke it as well now. 1218 if (ReturnValue->use_empty() && isa<llvm::AllocaInst>(ReturnValue)) { 1219 cast<llvm::AllocaInst>(ReturnValue)->eraseFromParent(); 1220 ReturnValue = 0; 1221 } 1222 } 1223 } else { 1224 llvm::Value *V = ReturnValue; 1225 // If the value is offset in memory, apply the offset now. 1226 if (unsigned Offs = RetAI.getDirectOffset()) { 1227 V = Builder.CreateBitCast(V, Builder.getInt8PtrTy()); 1228 V = Builder.CreateConstGEP1_32(V, Offs); 1229 V = Builder.CreateBitCast(V, 1230 llvm::PointerType::getUnqual(RetAI.getCoerceToType())); 1231 } 1232 1233 RV = CreateCoercedLoad(V, RetAI.getCoerceToType(), *this); 1234 } 1235 1236 // In ARC, end functions that return a retainable type with a call 1237 // to objc_autoreleaseReturnValue. 1238 if (AutoreleaseResult) { 1239 assert(getLangOptions().ObjCAutoRefCount && 1240 !FI.isReturnsRetained() && 1241 RetTy->isObjCRetainableType()); 1242 RV = emitAutoreleaseOfResult(*this, RV); 1243 } 1244 1245 break; 1246 1247 case ABIArgInfo::Ignore: 1248 break; 1249 1250 case ABIArgInfo::Expand: 1251 llvm_unreachable("Invalid ABI kind for return argument"); 1252 } 1253 1254 llvm::Instruction *Ret = RV ? Builder.CreateRet(RV) : Builder.CreateRetVoid(); 1255 if (!RetDbgLoc.isUnknown()) 1256 Ret->setDebugLoc(RetDbgLoc); 1257 } 1258 1259 void CodeGenFunction::EmitDelegateCallArg(CallArgList &args, 1260 const VarDecl *param) { 1261 // StartFunction converted the ABI-lowered parameter(s) into a 1262 // local alloca. We need to turn that into an r-value suitable 1263 // for EmitCall. 1264 llvm::Value *local = GetAddrOfLocalVar(param); 1265 1266 QualType type = param->getType(); 1267 1268 // For the most part, we just need to load the alloca, except: 1269 // 1) aggregate r-values are actually pointers to temporaries, and 1270 // 2) references to aggregates are pointers directly to the aggregate. 1271 // I don't know why references to non-aggregates are different here. 1272 if (const ReferenceType *ref = type->getAs<ReferenceType>()) { 1273 if (hasAggregateLLVMType(ref->getPointeeType())) 1274 return args.add(RValue::getAggregate(local), type); 1275 1276 // Locals which are references to scalars are represented 1277 // with allocas holding the pointer. 1278 return args.add(RValue::get(Builder.CreateLoad(local)), type); 1279 } 1280 1281 if (type->isAnyComplexType()) { 1282 ComplexPairTy complex = LoadComplexFromAddr(local, /*volatile*/ false); 1283 return args.add(RValue::getComplex(complex), type); 1284 } 1285 1286 if (hasAggregateLLVMType(type)) 1287 return args.add(RValue::getAggregate(local), type); 1288 1289 unsigned alignment = getContext().getDeclAlign(param).getQuantity(); 1290 llvm::Value *value = EmitLoadOfScalar(local, false, alignment, type); 1291 return args.add(RValue::get(value), type); 1292 } 1293 1294 static bool isProvablyNull(llvm::Value *addr) { 1295 return isa<llvm::ConstantPointerNull>(addr); 1296 } 1297 1298 static bool isProvablyNonNull(llvm::Value *addr) { 1299 return isa<llvm::AllocaInst>(addr); 1300 } 1301 1302 /// Emit the actual writing-back of a writeback. 1303 static void emitWriteback(CodeGenFunction &CGF, 1304 const CallArgList::Writeback &writeback) { 1305 llvm::Value *srcAddr = writeback.Address; 1306 assert(!isProvablyNull(srcAddr) && 1307 "shouldn't have writeback for provably null argument"); 1308 1309 llvm::BasicBlock *contBB = 0; 1310 1311 // If the argument wasn't provably non-null, we need to null check 1312 // before doing the store. 1313 bool provablyNonNull = isProvablyNonNull(srcAddr); 1314 if (!provablyNonNull) { 1315 llvm::BasicBlock *writebackBB = CGF.createBasicBlock("icr.writeback"); 1316 contBB = CGF.createBasicBlock("icr.done"); 1317 1318 llvm::Value *isNull = CGF.Builder.CreateIsNull(srcAddr, "icr.isnull"); 1319 CGF.Builder.CreateCondBr(isNull, contBB, writebackBB); 1320 CGF.EmitBlock(writebackBB); 1321 } 1322 1323 // Load the value to writeback. 1324 llvm::Value *value = CGF.Builder.CreateLoad(writeback.Temporary); 1325 1326 // Cast it back, in case we're writing an id to a Foo* or something. 1327 value = CGF.Builder.CreateBitCast(value, 1328 cast<llvm::PointerType>(srcAddr->getType())->getElementType(), 1329 "icr.writeback-cast"); 1330 1331 // Perform the writeback. 1332 QualType srcAddrType = writeback.AddressType; 1333 CGF.EmitStoreThroughLValue(RValue::get(value), 1334 CGF.MakeAddrLValue(srcAddr, srcAddrType)); 1335 1336 // Jump to the continuation block. 1337 if (!provablyNonNull) 1338 CGF.EmitBlock(contBB); 1339 } 1340 1341 static void emitWritebacks(CodeGenFunction &CGF, 1342 const CallArgList &args) { 1343 for (CallArgList::writeback_iterator 1344 i = args.writeback_begin(), e = args.writeback_end(); i != e; ++i) 1345 emitWriteback(CGF, *i); 1346 } 1347 1348 /// Emit an argument that's being passed call-by-writeback. That is, 1349 /// we are passing the address of 1350 static void emitWritebackArg(CodeGenFunction &CGF, CallArgList &args, 1351 const ObjCIndirectCopyRestoreExpr *CRE) { 1352 llvm::Value *srcAddr = CGF.EmitScalarExpr(CRE->getSubExpr()); 1353 1354 // The dest and src types don't necessarily match in LLVM terms 1355 // because of the crazy ObjC compatibility rules. 1356 1357 llvm::PointerType *destType = 1358 cast<llvm::PointerType>(CGF.ConvertType(CRE->getType())); 1359 1360 // If the address is a constant null, just pass the appropriate null. 1361 if (isProvablyNull(srcAddr)) { 1362 args.add(RValue::get(llvm::ConstantPointerNull::get(destType)), 1363 CRE->getType()); 1364 return; 1365 } 1366 1367 QualType srcAddrType = 1368 CRE->getSubExpr()->getType()->castAs<PointerType>()->getPointeeType(); 1369 1370 // Create the temporary. 1371 llvm::Value *temp = CGF.CreateTempAlloca(destType->getElementType(), 1372 "icr.temp"); 1373 1374 // Zero-initialize it if we're not doing a copy-initialization. 1375 bool shouldCopy = CRE->shouldCopy(); 1376 if (!shouldCopy) { 1377 llvm::Value *null = 1378 llvm::ConstantPointerNull::get( 1379 cast<llvm::PointerType>(destType->getElementType())); 1380 CGF.Builder.CreateStore(null, temp); 1381 } 1382 1383 llvm::BasicBlock *contBB = 0; 1384 1385 // If the address is *not* known to be non-null, we need to switch. 1386 llvm::Value *finalArgument; 1387 1388 bool provablyNonNull = isProvablyNonNull(srcAddr); 1389 if (provablyNonNull) { 1390 finalArgument = temp; 1391 } else { 1392 llvm::Value *isNull = CGF.Builder.CreateIsNull(srcAddr, "icr.isnull"); 1393 1394 finalArgument = CGF.Builder.CreateSelect(isNull, 1395 llvm::ConstantPointerNull::get(destType), 1396 temp, "icr.argument"); 1397 1398 // If we need to copy, then the load has to be conditional, which 1399 // means we need control flow. 1400 if (shouldCopy) { 1401 contBB = CGF.createBasicBlock("icr.cont"); 1402 llvm::BasicBlock *copyBB = CGF.createBasicBlock("icr.copy"); 1403 CGF.Builder.CreateCondBr(isNull, contBB, copyBB); 1404 CGF.EmitBlock(copyBB); 1405 } 1406 } 1407 1408 // Perform a copy if necessary. 1409 if (shouldCopy) { 1410 LValue srcLV = CGF.MakeAddrLValue(srcAddr, srcAddrType); 1411 RValue srcRV = CGF.EmitLoadOfLValue(srcLV); 1412 assert(srcRV.isScalar()); 1413 1414 llvm::Value *src = srcRV.getScalarVal(); 1415 src = CGF.Builder.CreateBitCast(src, destType->getElementType(), 1416 "icr.cast"); 1417 1418 // Use an ordinary store, not a store-to-lvalue. 1419 CGF.Builder.CreateStore(src, temp); 1420 } 1421 1422 // Finish the control flow if we needed it. 1423 if (shouldCopy && !provablyNonNull) 1424 CGF.EmitBlock(contBB); 1425 1426 args.addWriteback(srcAddr, srcAddrType, temp); 1427 args.add(RValue::get(finalArgument), CRE->getType()); 1428 } 1429 1430 void CodeGenFunction::EmitCallArg(CallArgList &args, const Expr *E, 1431 QualType type) { 1432 if (const ObjCIndirectCopyRestoreExpr *CRE 1433 = dyn_cast<ObjCIndirectCopyRestoreExpr>(E)) { 1434 assert(getContext().getLangOptions().ObjCAutoRefCount); 1435 assert(getContext().hasSameType(E->getType(), type)); 1436 return emitWritebackArg(*this, args, CRE); 1437 } 1438 1439 assert(type->isReferenceType() == E->isGLValue() && 1440 "reference binding to unmaterialized r-value!"); 1441 1442 if (E->isGLValue()) { 1443 assert(E->getObjectKind() == OK_Ordinary); 1444 return args.add(EmitReferenceBindingToExpr(E, /*InitializedDecl=*/0), 1445 type); 1446 } 1447 1448 if (hasAggregateLLVMType(type) && !E->getType()->isAnyComplexType() && 1449 isa<ImplicitCastExpr>(E) && 1450 cast<CastExpr>(E)->getCastKind() == CK_LValueToRValue) { 1451 LValue L = EmitLValue(cast<CastExpr>(E)->getSubExpr()); 1452 assert(L.isSimple()); 1453 args.add(RValue::getAggregate(L.getAddress(), L.isVolatileQualified()), 1454 type, /*NeedsCopy*/true); 1455 return; 1456 } 1457 1458 args.add(EmitAnyExprToTemp(E), type); 1459 } 1460 1461 /// Emits a call or invoke instruction to the given function, depending 1462 /// on the current state of the EH stack. 1463 llvm::CallSite 1464 CodeGenFunction::EmitCallOrInvoke(llvm::Value *Callee, 1465 ArrayRef<llvm::Value *> Args, 1466 const Twine &Name) { 1467 llvm::BasicBlock *InvokeDest = getInvokeDest(); 1468 if (!InvokeDest) 1469 return Builder.CreateCall(Callee, Args, Name); 1470 1471 llvm::BasicBlock *ContBB = createBasicBlock("invoke.cont"); 1472 llvm::InvokeInst *Invoke = Builder.CreateInvoke(Callee, ContBB, InvokeDest, 1473 Args, Name); 1474 EmitBlock(ContBB); 1475 return Invoke; 1476 } 1477 1478 llvm::CallSite 1479 CodeGenFunction::EmitCallOrInvoke(llvm::Value *Callee, 1480 const Twine &Name) { 1481 return EmitCallOrInvoke(Callee, ArrayRef<llvm::Value *>(), Name); 1482 } 1483 1484 static void checkArgMatches(llvm::Value *Elt, unsigned &ArgNo, 1485 llvm::FunctionType *FTy) { 1486 if (ArgNo < FTy->getNumParams()) 1487 assert(Elt->getType() == FTy->getParamType(ArgNo)); 1488 else 1489 assert(FTy->isVarArg()); 1490 ++ArgNo; 1491 } 1492 1493 void CodeGenFunction::ExpandTypeToArgs(QualType Ty, RValue RV, 1494 SmallVector<llvm::Value*,16> &Args, 1495 llvm::FunctionType *IRFuncTy) { 1496 if (const ConstantArrayType *AT = getContext().getAsConstantArrayType(Ty)) { 1497 unsigned NumElts = AT->getSize().getZExtValue(); 1498 QualType EltTy = AT->getElementType(); 1499 llvm::Value *Addr = RV.getAggregateAddr(); 1500 for (unsigned Elt = 0; Elt < NumElts; ++Elt) { 1501 llvm::Value *EltAddr = Builder.CreateConstGEP2_32(Addr, 0, Elt); 1502 LValue LV = MakeAddrLValue(EltAddr, EltTy); 1503 RValue EltRV; 1504 if (CodeGenFunction::hasAggregateLLVMType(EltTy)) 1505 EltRV = RValue::getAggregate(LV.getAddress()); 1506 else 1507 EltRV = EmitLoadOfLValue(LV); 1508 ExpandTypeToArgs(EltTy, EltRV, Args, IRFuncTy); 1509 } 1510 } else if (const RecordType *RT = Ty->getAsStructureType()) { 1511 RecordDecl *RD = RT->getDecl(); 1512 assert(RV.isAggregate() && "Unexpected rvalue during struct expansion"); 1513 llvm::Value *Addr = RV.getAggregateAddr(); 1514 for (RecordDecl::field_iterator i = RD->field_begin(), e = RD->field_end(); 1515 i != e; ++i) { 1516 FieldDecl *FD = *i; 1517 QualType FT = FD->getType(); 1518 1519 // FIXME: What are the right qualifiers here? 1520 LValue LV = EmitLValueForField(Addr, FD, 0); 1521 RValue FldRV; 1522 if (CodeGenFunction::hasAggregateLLVMType(FT)) 1523 FldRV = RValue::getAggregate(LV.getAddress()); 1524 else 1525 FldRV = EmitLoadOfLValue(LV); 1526 ExpandTypeToArgs(FT, FldRV, Args, IRFuncTy); 1527 } 1528 } else if (isa<ComplexType>(Ty)) { 1529 ComplexPairTy CV = RV.getComplexVal(); 1530 Args.push_back(CV.first); 1531 Args.push_back(CV.second); 1532 } else { 1533 assert(RV.isScalar() && 1534 "Unexpected non-scalar rvalue during struct expansion."); 1535 1536 // Insert a bitcast as needed. 1537 llvm::Value *V = RV.getScalarVal(); 1538 if (Args.size() < IRFuncTy->getNumParams() && 1539 V->getType() != IRFuncTy->getParamType(Args.size())) 1540 V = Builder.CreateBitCast(V, IRFuncTy->getParamType(Args.size())); 1541 1542 Args.push_back(V); 1543 } 1544 } 1545 1546 1547 RValue CodeGenFunction::EmitCall(const CGFunctionInfo &CallInfo, 1548 llvm::Value *Callee, 1549 ReturnValueSlot ReturnValue, 1550 const CallArgList &CallArgs, 1551 const Decl *TargetDecl, 1552 llvm::Instruction **callOrInvoke) { 1553 // FIXME: We no longer need the types from CallArgs; lift up and simplify. 1554 SmallVector<llvm::Value*, 16> Args; 1555 1556 // Handle struct-return functions by passing a pointer to the 1557 // location that we would like to return into. 1558 QualType RetTy = CallInfo.getReturnType(); 1559 const ABIArgInfo &RetAI = CallInfo.getReturnInfo(); 1560 1561 // IRArgNo - Keep track of the argument number in the callee we're looking at. 1562 unsigned IRArgNo = 0; 1563 llvm::FunctionType *IRFuncTy = 1564 cast<llvm::FunctionType>( 1565 cast<llvm::PointerType>(Callee->getType())->getElementType()); 1566 1567 // If the call returns a temporary with struct return, create a temporary 1568 // alloca to hold the result, unless one is given to us. 1569 if (CGM.ReturnTypeUsesSRet(CallInfo)) { 1570 llvm::Value *Value = ReturnValue.getValue(); 1571 if (!Value) 1572 Value = CreateMemTemp(RetTy); 1573 Args.push_back(Value); 1574 checkArgMatches(Value, IRArgNo, IRFuncTy); 1575 } 1576 1577 assert(CallInfo.arg_size() == CallArgs.size() && 1578 "Mismatch between function signature & arguments."); 1579 CGFunctionInfo::const_arg_iterator info_it = CallInfo.arg_begin(); 1580 for (CallArgList::const_iterator I = CallArgs.begin(), E = CallArgs.end(); 1581 I != E; ++I, ++info_it) { 1582 const ABIArgInfo &ArgInfo = info_it->info; 1583 RValue RV = I->RV; 1584 1585 unsigned TypeAlign = 1586 getContext().getTypeAlignInChars(I->Ty).getQuantity(); 1587 switch (ArgInfo.getKind()) { 1588 case ABIArgInfo::Indirect: { 1589 if (RV.isScalar() || RV.isComplex()) { 1590 // Make a temporary alloca to pass the argument. 1591 llvm::AllocaInst *AI = CreateMemTemp(I->Ty); 1592 if (ArgInfo.getIndirectAlign() > AI->getAlignment()) 1593 AI->setAlignment(ArgInfo.getIndirectAlign()); 1594 Args.push_back(AI); 1595 1596 if (RV.isScalar()) 1597 EmitStoreOfScalar(RV.getScalarVal(), Args.back(), false, 1598 TypeAlign, I->Ty); 1599 else 1600 StoreComplexToAddr(RV.getComplexVal(), Args.back(), false); 1601 1602 // Validate argument match. 1603 checkArgMatches(AI, IRArgNo, IRFuncTy); 1604 } else { 1605 // We want to avoid creating an unnecessary temporary+copy here; 1606 // however, we need one in two cases: 1607 // 1. If the argument is not byval, and we are required to copy the 1608 // source. (This case doesn't occur on any common architecture.) 1609 // 2. If the argument is byval, RV is not sufficiently aligned, and 1610 // we cannot force it to be sufficiently aligned. 1611 llvm::Value *Addr = RV.getAggregateAddr(); 1612 unsigned Align = ArgInfo.getIndirectAlign(); 1613 const llvm::TargetData *TD = &CGM.getTargetData(); 1614 if ((!ArgInfo.getIndirectByVal() && I->NeedsCopy) || 1615 (ArgInfo.getIndirectByVal() && TypeAlign < Align && 1616 llvm::getOrEnforceKnownAlignment(Addr, Align, TD) < Align)) { 1617 // Create an aligned temporary, and copy to it. 1618 llvm::AllocaInst *AI = CreateMemTemp(I->Ty); 1619 if (Align > AI->getAlignment()) 1620 AI->setAlignment(Align); 1621 Args.push_back(AI); 1622 EmitAggregateCopy(AI, Addr, I->Ty, RV.isVolatileQualified()); 1623 1624 // Validate argument match. 1625 checkArgMatches(AI, IRArgNo, IRFuncTy); 1626 } else { 1627 // Skip the extra memcpy call. 1628 Args.push_back(Addr); 1629 1630 // Validate argument match. 1631 checkArgMatches(Addr, IRArgNo, IRFuncTy); 1632 } 1633 } 1634 break; 1635 } 1636 1637 case ABIArgInfo::Ignore: 1638 break; 1639 1640 case ABIArgInfo::Extend: 1641 case ABIArgInfo::Direct: { 1642 if (!isa<llvm::StructType>(ArgInfo.getCoerceToType()) && 1643 ArgInfo.getCoerceToType() == ConvertType(info_it->type) && 1644 ArgInfo.getDirectOffset() == 0) { 1645 llvm::Value *V; 1646 if (RV.isScalar()) 1647 V = RV.getScalarVal(); 1648 else 1649 V = Builder.CreateLoad(RV.getAggregateAddr()); 1650 1651 // If the argument doesn't match, perform a bitcast to coerce it. This 1652 // can happen due to trivial type mismatches. 1653 if (IRArgNo < IRFuncTy->getNumParams() && 1654 V->getType() != IRFuncTy->getParamType(IRArgNo)) 1655 V = Builder.CreateBitCast(V, IRFuncTy->getParamType(IRArgNo)); 1656 Args.push_back(V); 1657 1658 checkArgMatches(V, IRArgNo, IRFuncTy); 1659 break; 1660 } 1661 1662 // FIXME: Avoid the conversion through memory if possible. 1663 llvm::Value *SrcPtr; 1664 if (RV.isScalar()) { 1665 SrcPtr = CreateMemTemp(I->Ty, "coerce"); 1666 EmitStoreOfScalar(RV.getScalarVal(), SrcPtr, false, TypeAlign, I->Ty); 1667 } else if (RV.isComplex()) { 1668 SrcPtr = CreateMemTemp(I->Ty, "coerce"); 1669 StoreComplexToAddr(RV.getComplexVal(), SrcPtr, false); 1670 } else 1671 SrcPtr = RV.getAggregateAddr(); 1672 1673 // If the value is offset in memory, apply the offset now. 1674 if (unsigned Offs = ArgInfo.getDirectOffset()) { 1675 SrcPtr = Builder.CreateBitCast(SrcPtr, Builder.getInt8PtrTy()); 1676 SrcPtr = Builder.CreateConstGEP1_32(SrcPtr, Offs); 1677 SrcPtr = Builder.CreateBitCast(SrcPtr, 1678 llvm::PointerType::getUnqual(ArgInfo.getCoerceToType())); 1679 1680 } 1681 1682 // If the coerce-to type is a first class aggregate, we flatten it and 1683 // pass the elements. Either way is semantically identical, but fast-isel 1684 // and the optimizer generally likes scalar values better than FCAs. 1685 if (llvm::StructType *STy = 1686 dyn_cast<llvm::StructType>(ArgInfo.getCoerceToType())) { 1687 SrcPtr = Builder.CreateBitCast(SrcPtr, 1688 llvm::PointerType::getUnqual(STy)); 1689 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 1690 llvm::Value *EltPtr = Builder.CreateConstGEP2_32(SrcPtr, 0, i); 1691 llvm::LoadInst *LI = Builder.CreateLoad(EltPtr); 1692 // We don't know what we're loading from. 1693 LI->setAlignment(1); 1694 Args.push_back(LI); 1695 1696 // Validate argument match. 1697 checkArgMatches(LI, IRArgNo, IRFuncTy); 1698 } 1699 } else { 1700 // In the simple case, just pass the coerced loaded value. 1701 Args.push_back(CreateCoercedLoad(SrcPtr, ArgInfo.getCoerceToType(), 1702 *this)); 1703 1704 // Validate argument match. 1705 checkArgMatches(Args.back(), IRArgNo, IRFuncTy); 1706 } 1707 1708 break; 1709 } 1710 1711 case ABIArgInfo::Expand: 1712 ExpandTypeToArgs(I->Ty, RV, Args, IRFuncTy); 1713 IRArgNo = Args.size(); 1714 break; 1715 } 1716 } 1717 1718 // If the callee is a bitcast of a function to a varargs pointer to function 1719 // type, check to see if we can remove the bitcast. This handles some cases 1720 // with unprototyped functions. 1721 if (llvm::ConstantExpr *CE = dyn_cast<llvm::ConstantExpr>(Callee)) 1722 if (llvm::Function *CalleeF = dyn_cast<llvm::Function>(CE->getOperand(0))) { 1723 llvm::PointerType *CurPT=cast<llvm::PointerType>(Callee->getType()); 1724 llvm::FunctionType *CurFT = 1725 cast<llvm::FunctionType>(CurPT->getElementType()); 1726 llvm::FunctionType *ActualFT = CalleeF->getFunctionType(); 1727 1728 if (CE->getOpcode() == llvm::Instruction::BitCast && 1729 ActualFT->getReturnType() == CurFT->getReturnType() && 1730 ActualFT->getNumParams() == CurFT->getNumParams() && 1731 ActualFT->getNumParams() == Args.size() && 1732 (CurFT->isVarArg() || !ActualFT->isVarArg())) { 1733 bool ArgsMatch = true; 1734 for (unsigned i = 0, e = ActualFT->getNumParams(); i != e; ++i) 1735 if (ActualFT->getParamType(i) != CurFT->getParamType(i)) { 1736 ArgsMatch = false; 1737 break; 1738 } 1739 1740 // Strip the cast if we can get away with it. This is a nice cleanup, 1741 // but also allows us to inline the function at -O0 if it is marked 1742 // always_inline. 1743 if (ArgsMatch) 1744 Callee = CalleeF; 1745 } 1746 } 1747 1748 unsigned CallingConv; 1749 CodeGen::AttributeListType AttributeList; 1750 CGM.ConstructAttributeList(CallInfo, TargetDecl, AttributeList, CallingConv); 1751 llvm::AttrListPtr Attrs = llvm::AttrListPtr::get(AttributeList.begin(), 1752 AttributeList.end()); 1753 1754 llvm::BasicBlock *InvokeDest = 0; 1755 if (!(Attrs.getFnAttributes() & llvm::Attribute::NoUnwind)) 1756 InvokeDest = getInvokeDest(); 1757 1758 llvm::CallSite CS; 1759 if (!InvokeDest) { 1760 CS = Builder.CreateCall(Callee, Args); 1761 } else { 1762 llvm::BasicBlock *Cont = createBasicBlock("invoke.cont"); 1763 CS = Builder.CreateInvoke(Callee, Cont, InvokeDest, Args); 1764 EmitBlock(Cont); 1765 } 1766 if (callOrInvoke) 1767 *callOrInvoke = CS.getInstruction(); 1768 1769 CS.setAttributes(Attrs); 1770 CS.setCallingConv(static_cast<llvm::CallingConv::ID>(CallingConv)); 1771 1772 // If the call doesn't return, finish the basic block and clear the 1773 // insertion point; this allows the rest of IRgen to discard 1774 // unreachable code. 1775 if (CS.doesNotReturn()) { 1776 Builder.CreateUnreachable(); 1777 Builder.ClearInsertionPoint(); 1778 1779 // FIXME: For now, emit a dummy basic block because expr emitters in 1780 // generally are not ready to handle emitting expressions at unreachable 1781 // points. 1782 EnsureInsertPoint(); 1783 1784 // Return a reasonable RValue. 1785 return GetUndefRValue(RetTy); 1786 } 1787 1788 llvm::Instruction *CI = CS.getInstruction(); 1789 if (Builder.isNamePreserving() && !CI->getType()->isVoidTy()) 1790 CI->setName("call"); 1791 1792 // Emit any writebacks immediately. Arguably this should happen 1793 // after any return-value munging. 1794 if (CallArgs.hasWritebacks()) 1795 emitWritebacks(*this, CallArgs); 1796 1797 switch (RetAI.getKind()) { 1798 case ABIArgInfo::Indirect: { 1799 unsigned Alignment = getContext().getTypeAlignInChars(RetTy).getQuantity(); 1800 if (RetTy->isAnyComplexType()) 1801 return RValue::getComplex(LoadComplexFromAddr(Args[0], false)); 1802 if (CodeGenFunction::hasAggregateLLVMType(RetTy)) 1803 return RValue::getAggregate(Args[0]); 1804 return RValue::get(EmitLoadOfScalar(Args[0], false, Alignment, RetTy)); 1805 } 1806 1807 case ABIArgInfo::Ignore: 1808 // If we are ignoring an argument that had a result, make sure to 1809 // construct the appropriate return value for our caller. 1810 return GetUndefRValue(RetTy); 1811 1812 case ABIArgInfo::Extend: 1813 case ABIArgInfo::Direct: { 1814 llvm::Type *RetIRTy = ConvertType(RetTy); 1815 if (RetAI.getCoerceToType() == RetIRTy && RetAI.getDirectOffset() == 0) { 1816 if (RetTy->isAnyComplexType()) { 1817 llvm::Value *Real = Builder.CreateExtractValue(CI, 0); 1818 llvm::Value *Imag = Builder.CreateExtractValue(CI, 1); 1819 return RValue::getComplex(std::make_pair(Real, Imag)); 1820 } 1821 if (CodeGenFunction::hasAggregateLLVMType(RetTy)) { 1822 llvm::Value *DestPtr = ReturnValue.getValue(); 1823 bool DestIsVolatile = ReturnValue.isVolatile(); 1824 1825 if (!DestPtr) { 1826 DestPtr = CreateMemTemp(RetTy, "agg.tmp"); 1827 DestIsVolatile = false; 1828 } 1829 BuildAggStore(*this, CI, DestPtr, DestIsVolatile, false); 1830 return RValue::getAggregate(DestPtr); 1831 } 1832 1833 // If the argument doesn't match, perform a bitcast to coerce it. This 1834 // can happen due to trivial type mismatches. 1835 llvm::Value *V = CI; 1836 if (V->getType() != RetIRTy) 1837 V = Builder.CreateBitCast(V, RetIRTy); 1838 return RValue::get(V); 1839 } 1840 1841 llvm::Value *DestPtr = ReturnValue.getValue(); 1842 bool DestIsVolatile = ReturnValue.isVolatile(); 1843 1844 if (!DestPtr) { 1845 DestPtr = CreateMemTemp(RetTy, "coerce"); 1846 DestIsVolatile = false; 1847 } 1848 1849 // If the value is offset in memory, apply the offset now. 1850 llvm::Value *StorePtr = DestPtr; 1851 if (unsigned Offs = RetAI.getDirectOffset()) { 1852 StorePtr = Builder.CreateBitCast(StorePtr, Builder.getInt8PtrTy()); 1853 StorePtr = Builder.CreateConstGEP1_32(StorePtr, Offs); 1854 StorePtr = Builder.CreateBitCast(StorePtr, 1855 llvm::PointerType::getUnqual(RetAI.getCoerceToType())); 1856 } 1857 CreateCoercedStore(CI, StorePtr, DestIsVolatile, *this); 1858 1859 unsigned Alignment = getContext().getTypeAlignInChars(RetTy).getQuantity(); 1860 if (RetTy->isAnyComplexType()) 1861 return RValue::getComplex(LoadComplexFromAddr(DestPtr, false)); 1862 if (CodeGenFunction::hasAggregateLLVMType(RetTy)) 1863 return RValue::getAggregate(DestPtr); 1864 return RValue::get(EmitLoadOfScalar(DestPtr, false, Alignment, RetTy)); 1865 } 1866 1867 case ABIArgInfo::Expand: 1868 llvm_unreachable("Invalid ABI kind for return argument"); 1869 } 1870 1871 llvm_unreachable("Unhandled ABIArgInfo::Kind"); 1872 } 1873 1874 /* VarArg handling */ 1875 1876 llvm::Value *CodeGenFunction::EmitVAArg(llvm::Value *VAListAddr, QualType Ty) { 1877 return CGM.getTypes().getABIInfo().EmitVAArg(VAListAddr, Ty, *this); 1878 } 1879