CGCall.cpp revision 64aa4b3ec7e62288e2e66c1935487ece995ca94b
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 "ABIInfo.h" 17#include "CGCXXABI.h" 18#include "CodeGenFunction.h" 19#include "CodeGenModule.h" 20#include "TargetInfo.h" 21#include "clang/AST/Decl.h" 22#include "clang/AST/DeclCXX.h" 23#include "clang/AST/DeclObjC.h" 24#include "clang/Basic/TargetInfo.h" 25#include "clang/Frontend/CodeGenOptions.h" 26#include "llvm/ADT/StringExtras.h" 27#include "llvm/IR/Attributes.h" 28#include "llvm/IR/DataLayout.h" 29#include "llvm/IR/InlineAsm.h" 30#include "llvm/MC/SubtargetFeature.h" 31#include "llvm/Support/CallSite.h" 32#include "llvm/Transforms/Utils/Local.h" 33using namespace clang; 34using namespace CodeGen; 35 36/***/ 37 38static unsigned ClangCallConvToLLVMCallConv(CallingConv CC) { 39 switch (CC) { 40 default: return llvm::CallingConv::C; 41 case CC_X86StdCall: return llvm::CallingConv::X86_StdCall; 42 case CC_X86FastCall: return llvm::CallingConv::X86_FastCall; 43 case CC_X86ThisCall: return llvm::CallingConv::X86_ThisCall; 44 case CC_AAPCS: return llvm::CallingConv::ARM_AAPCS; 45 case CC_AAPCS_VFP: return llvm::CallingConv::ARM_AAPCS_VFP; 46 case CC_IntelOclBicc: return llvm::CallingConv::Intel_OCL_BI; 47 // TODO: add support for CC_X86Pascal to llvm 48 } 49} 50 51/// Derives the 'this' type for codegen purposes, i.e. ignoring method 52/// qualification. 53/// FIXME: address space qualification? 54static CanQualType GetThisType(ASTContext &Context, const CXXRecordDecl *RD) { 55 QualType RecTy = Context.getTagDeclType(RD)->getCanonicalTypeInternal(); 56 return Context.getPointerType(CanQualType::CreateUnsafe(RecTy)); 57} 58 59/// Returns the canonical formal type of the given C++ method. 60static CanQual<FunctionProtoType> GetFormalType(const CXXMethodDecl *MD) { 61 return MD->getType()->getCanonicalTypeUnqualified() 62 .getAs<FunctionProtoType>(); 63} 64 65/// Returns the "extra-canonicalized" return type, which discards 66/// qualifiers on the return type. Codegen doesn't care about them, 67/// and it makes ABI code a little easier to be able to assume that 68/// all parameter and return types are top-level unqualified. 69static CanQualType GetReturnType(QualType RetTy) { 70 return RetTy->getCanonicalTypeUnqualified().getUnqualifiedType(); 71} 72 73/// Arrange the argument and result information for a value of the given 74/// unprototyped freestanding function type. 75const CGFunctionInfo & 76CodeGenTypes::arrangeFreeFunctionType(CanQual<FunctionNoProtoType> FTNP) { 77 // When translating an unprototyped function type, always use a 78 // variadic type. 79 return arrangeLLVMFunctionInfo(FTNP->getResultType().getUnqualifiedType(), 80 ArrayRef<CanQualType>(), 81 FTNP->getExtInfo(), 82 RequiredArgs(0)); 83} 84 85/// Arrange the LLVM function layout for a value of the given function 86/// type, on top of any implicit parameters already stored. Use the 87/// given ExtInfo instead of the ExtInfo from the function type. 88static const CGFunctionInfo &arrangeLLVMFunctionInfo(CodeGenTypes &CGT, 89 SmallVectorImpl<CanQualType> &prefix, 90 CanQual<FunctionProtoType> FTP, 91 FunctionType::ExtInfo extInfo) { 92 RequiredArgs required = RequiredArgs::forPrototypePlus(FTP, prefix.size()); 93 // FIXME: Kill copy. 94 for (unsigned i = 0, e = FTP->getNumArgs(); i != e; ++i) 95 prefix.push_back(FTP->getArgType(i)); 96 CanQualType resultType = FTP->getResultType().getUnqualifiedType(); 97 return CGT.arrangeLLVMFunctionInfo(resultType, prefix, extInfo, required); 98} 99 100/// Arrange the argument and result information for a free function (i.e. 101/// not a C++ or ObjC instance method) of the given type. 102static const CGFunctionInfo &arrangeFreeFunctionType(CodeGenTypes &CGT, 103 SmallVectorImpl<CanQualType> &prefix, 104 CanQual<FunctionProtoType> FTP) { 105 return arrangeLLVMFunctionInfo(CGT, prefix, FTP, FTP->getExtInfo()); 106} 107 108/// Given the formal ext-info of a C++ instance method, adjust it 109/// according to the C++ ABI in effect. 110static void adjustCXXMethodInfo(CodeGenTypes &CGT, 111 FunctionType::ExtInfo &extInfo, 112 bool isVariadic) { 113 if (extInfo.getCC() == CC_Default) { 114 CallingConv CC = CGT.getContext().getDefaultCXXMethodCallConv(isVariadic); 115 extInfo = extInfo.withCallingConv(CC); 116 } 117} 118 119/// Arrange the argument and result information for a free function (i.e. 120/// not a C++ or ObjC instance method) of the given type. 121static const CGFunctionInfo &arrangeCXXMethodType(CodeGenTypes &CGT, 122 SmallVectorImpl<CanQualType> &prefix, 123 CanQual<FunctionProtoType> FTP) { 124 FunctionType::ExtInfo extInfo = FTP->getExtInfo(); 125 adjustCXXMethodInfo(CGT, extInfo, FTP->isVariadic()); 126 return arrangeLLVMFunctionInfo(CGT, prefix, FTP, extInfo); 127} 128 129/// Arrange the argument and result information for a value of the 130/// given freestanding function type. 131const CGFunctionInfo & 132CodeGenTypes::arrangeFreeFunctionType(CanQual<FunctionProtoType> FTP) { 133 SmallVector<CanQualType, 16> argTypes; 134 return ::arrangeFreeFunctionType(*this, argTypes, FTP); 135} 136 137static CallingConv getCallingConventionForDecl(const Decl *D) { 138 // Set the appropriate calling convention for the Function. 139 if (D->hasAttr<StdCallAttr>()) 140 return CC_X86StdCall; 141 142 if (D->hasAttr<FastCallAttr>()) 143 return CC_X86FastCall; 144 145 if (D->hasAttr<ThisCallAttr>()) 146 return CC_X86ThisCall; 147 148 if (D->hasAttr<PascalAttr>()) 149 return CC_X86Pascal; 150 151 if (PcsAttr *PCS = D->getAttr<PcsAttr>()) 152 return (PCS->getPCS() == PcsAttr::AAPCS ? CC_AAPCS : CC_AAPCS_VFP); 153 154 if (D->hasAttr<PnaclCallAttr>()) 155 return CC_PnaclCall; 156 157 if (D->hasAttr<IntelOclBiccAttr>()) 158 return CC_IntelOclBicc; 159 160 return CC_C; 161} 162 163/// Arrange the argument and result information for a call to an 164/// unknown C++ non-static member function of the given abstract type. 165/// The member function must be an ordinary function, i.e. not a 166/// constructor or destructor. 167const CGFunctionInfo & 168CodeGenTypes::arrangeCXXMethodType(const CXXRecordDecl *RD, 169 const FunctionProtoType *FTP) { 170 SmallVector<CanQualType, 16> argTypes; 171 172 // Add the 'this' pointer. 173 argTypes.push_back(GetThisType(Context, RD)); 174 175 return ::arrangeCXXMethodType(*this, argTypes, 176 FTP->getCanonicalTypeUnqualified().getAs<FunctionProtoType>()); 177} 178 179/// Arrange the argument and result information for a declaration or 180/// definition of the given C++ non-static member function. The 181/// member function must be an ordinary function, i.e. not a 182/// constructor or destructor. 183const CGFunctionInfo & 184CodeGenTypes::arrangeCXXMethodDeclaration(const CXXMethodDecl *MD) { 185 assert(!isa<CXXConstructorDecl>(MD) && "wrong method for contructors!"); 186 assert(!isa<CXXDestructorDecl>(MD) && "wrong method for destructors!"); 187 188 CanQual<FunctionProtoType> prototype = GetFormalType(MD); 189 190 if (MD->isInstance()) { 191 // The abstract case is perfectly fine. 192 return arrangeCXXMethodType(MD->getParent(), prototype.getTypePtr()); 193 } 194 195 return arrangeFreeFunctionType(prototype); 196} 197 198/// Arrange the argument and result information for a declaration 199/// or definition to the given constructor variant. 200const CGFunctionInfo & 201CodeGenTypes::arrangeCXXConstructorDeclaration(const CXXConstructorDecl *D, 202 CXXCtorType ctorKind) { 203 SmallVector<CanQualType, 16> argTypes; 204 argTypes.push_back(GetThisType(Context, D->getParent())); 205 CanQualType resultType = Context.VoidTy; 206 207 TheCXXABI.BuildConstructorSignature(D, ctorKind, resultType, argTypes); 208 209 CanQual<FunctionProtoType> FTP = GetFormalType(D); 210 211 RequiredArgs required = RequiredArgs::forPrototypePlus(FTP, argTypes.size()); 212 213 // Add the formal parameters. 214 for (unsigned i = 0, e = FTP->getNumArgs(); i != e; ++i) 215 argTypes.push_back(FTP->getArgType(i)); 216 217 FunctionType::ExtInfo extInfo = FTP->getExtInfo(); 218 adjustCXXMethodInfo(*this, extInfo, FTP->isVariadic()); 219 return arrangeLLVMFunctionInfo(resultType, argTypes, extInfo, required); 220} 221 222/// Arrange the argument and result information for a declaration, 223/// definition, or call to the given destructor variant. It so 224/// happens that all three cases produce the same information. 225const CGFunctionInfo & 226CodeGenTypes::arrangeCXXDestructor(const CXXDestructorDecl *D, 227 CXXDtorType dtorKind) { 228 SmallVector<CanQualType, 2> argTypes; 229 argTypes.push_back(GetThisType(Context, D->getParent())); 230 CanQualType resultType = Context.VoidTy; 231 232 TheCXXABI.BuildDestructorSignature(D, dtorKind, resultType, argTypes); 233 234 CanQual<FunctionProtoType> FTP = GetFormalType(D); 235 assert(FTP->getNumArgs() == 0 && "dtor with formal parameters"); 236 assert(FTP->isVariadic() == 0 && "dtor with formal parameters"); 237 238 FunctionType::ExtInfo extInfo = FTP->getExtInfo(); 239 adjustCXXMethodInfo(*this, extInfo, false); 240 return arrangeLLVMFunctionInfo(resultType, argTypes, extInfo, 241 RequiredArgs::All); 242} 243 244/// Arrange the argument and result information for the declaration or 245/// definition of the given function. 246const CGFunctionInfo & 247CodeGenTypes::arrangeFunctionDeclaration(const FunctionDecl *FD) { 248 if (const CXXMethodDecl *MD = dyn_cast<CXXMethodDecl>(FD)) 249 if (MD->isInstance()) 250 return arrangeCXXMethodDeclaration(MD); 251 252 CanQualType FTy = FD->getType()->getCanonicalTypeUnqualified(); 253 254 assert(isa<FunctionType>(FTy)); 255 256 // When declaring a function without a prototype, always use a 257 // non-variadic type. 258 if (isa<FunctionNoProtoType>(FTy)) { 259 CanQual<FunctionNoProtoType> noProto = FTy.getAs<FunctionNoProtoType>(); 260 return arrangeLLVMFunctionInfo(noProto->getResultType(), 261 ArrayRef<CanQualType>(), 262 noProto->getExtInfo(), 263 RequiredArgs::All); 264 } 265 266 assert(isa<FunctionProtoType>(FTy)); 267 return arrangeFreeFunctionType(FTy.getAs<FunctionProtoType>()); 268} 269 270/// Arrange the argument and result information for the declaration or 271/// definition of an Objective-C method. 272const CGFunctionInfo & 273CodeGenTypes::arrangeObjCMethodDeclaration(const ObjCMethodDecl *MD) { 274 // It happens that this is the same as a call with no optional 275 // arguments, except also using the formal 'self' type. 276 return arrangeObjCMessageSendSignature(MD, MD->getSelfDecl()->getType()); 277} 278 279/// Arrange the argument and result information for the function type 280/// through which to perform a send to the given Objective-C method, 281/// using the given receiver type. The receiver type is not always 282/// the 'self' type of the method or even an Objective-C pointer type. 283/// This is *not* the right method for actually performing such a 284/// message send, due to the possibility of optional arguments. 285const CGFunctionInfo & 286CodeGenTypes::arrangeObjCMessageSendSignature(const ObjCMethodDecl *MD, 287 QualType receiverType) { 288 SmallVector<CanQualType, 16> argTys; 289 argTys.push_back(Context.getCanonicalParamType(receiverType)); 290 argTys.push_back(Context.getCanonicalParamType(Context.getObjCSelType())); 291 // FIXME: Kill copy? 292 for (ObjCMethodDecl::param_const_iterator i = MD->param_begin(), 293 e = MD->param_end(); i != e; ++i) { 294 argTys.push_back(Context.getCanonicalParamType((*i)->getType())); 295 } 296 297 FunctionType::ExtInfo einfo; 298 einfo = einfo.withCallingConv(getCallingConventionForDecl(MD)); 299 300 if (getContext().getLangOpts().ObjCAutoRefCount && 301 MD->hasAttr<NSReturnsRetainedAttr>()) 302 einfo = einfo.withProducesResult(true); 303 304 RequiredArgs required = 305 (MD->isVariadic() ? RequiredArgs(argTys.size()) : RequiredArgs::All); 306 307 return arrangeLLVMFunctionInfo(GetReturnType(MD->getResultType()), argTys, 308 einfo, required); 309} 310 311const CGFunctionInfo & 312CodeGenTypes::arrangeGlobalDeclaration(GlobalDecl GD) { 313 // FIXME: Do we need to handle ObjCMethodDecl? 314 const FunctionDecl *FD = cast<FunctionDecl>(GD.getDecl()); 315 316 if (const CXXConstructorDecl *CD = dyn_cast<CXXConstructorDecl>(FD)) 317 return arrangeCXXConstructorDeclaration(CD, GD.getCtorType()); 318 319 if (const CXXDestructorDecl *DD = dyn_cast<CXXDestructorDecl>(FD)) 320 return arrangeCXXDestructor(DD, GD.getDtorType()); 321 322 return arrangeFunctionDeclaration(FD); 323} 324 325/// Arrange a call as unto a free function, except possibly with an 326/// additional number of formal parameters considered required. 327static const CGFunctionInfo & 328arrangeFreeFunctionLikeCall(CodeGenTypes &CGT, 329 const CallArgList &args, 330 const FunctionType *fnType, 331 unsigned numExtraRequiredArgs) { 332 assert(args.size() >= numExtraRequiredArgs); 333 334 // In most cases, there are no optional arguments. 335 RequiredArgs required = RequiredArgs::All; 336 337 // If we have a variadic prototype, the required arguments are the 338 // extra prefix plus the arguments in the prototype. 339 if (const FunctionProtoType *proto = dyn_cast<FunctionProtoType>(fnType)) { 340 if (proto->isVariadic()) 341 required = RequiredArgs(proto->getNumArgs() + numExtraRequiredArgs); 342 343 // If we don't have a prototype at all, but we're supposed to 344 // explicitly use the variadic convention for unprototyped calls, 345 // treat all of the arguments as required but preserve the nominal 346 // possibility of variadics. 347 } else if (CGT.CGM.getTargetCodeGenInfo() 348 .isNoProtoCallVariadic(args, cast<FunctionNoProtoType>(fnType))) { 349 required = RequiredArgs(args.size()); 350 } 351 352 return CGT.arrangeFreeFunctionCall(fnType->getResultType(), args, 353 fnType->getExtInfo(), required); 354} 355 356/// Figure out the rules for calling a function with the given formal 357/// type using the given arguments. The arguments are necessary 358/// because the function might be unprototyped, in which case it's 359/// target-dependent in crazy ways. 360const CGFunctionInfo & 361CodeGenTypes::arrangeFreeFunctionCall(const CallArgList &args, 362 const FunctionType *fnType) { 363 return arrangeFreeFunctionLikeCall(*this, args, fnType, 0); 364} 365 366/// A block function call is essentially a free-function call with an 367/// extra implicit argument. 368const CGFunctionInfo & 369CodeGenTypes::arrangeBlockFunctionCall(const CallArgList &args, 370 const FunctionType *fnType) { 371 return arrangeFreeFunctionLikeCall(*this, args, fnType, 1); 372} 373 374const CGFunctionInfo & 375CodeGenTypes::arrangeFreeFunctionCall(QualType resultType, 376 const CallArgList &args, 377 FunctionType::ExtInfo info, 378 RequiredArgs required) { 379 // FIXME: Kill copy. 380 SmallVector<CanQualType, 16> argTypes; 381 for (CallArgList::const_iterator i = args.begin(), e = args.end(); 382 i != e; ++i) 383 argTypes.push_back(Context.getCanonicalParamType(i->Ty)); 384 return arrangeLLVMFunctionInfo(GetReturnType(resultType), argTypes, info, 385 required); 386} 387 388/// Arrange a call to a C++ method, passing the given arguments. 389const CGFunctionInfo & 390CodeGenTypes::arrangeCXXMethodCall(const CallArgList &args, 391 const FunctionProtoType *FPT, 392 RequiredArgs required) { 393 // FIXME: Kill copy. 394 SmallVector<CanQualType, 16> argTypes; 395 for (CallArgList::const_iterator i = args.begin(), e = args.end(); 396 i != e; ++i) 397 argTypes.push_back(Context.getCanonicalParamType(i->Ty)); 398 399 FunctionType::ExtInfo info = FPT->getExtInfo(); 400 adjustCXXMethodInfo(*this, info, FPT->isVariadic()); 401 return arrangeLLVMFunctionInfo(GetReturnType(FPT->getResultType()), 402 argTypes, info, required); 403} 404 405const CGFunctionInfo & 406CodeGenTypes::arrangeFunctionDeclaration(QualType resultType, 407 const FunctionArgList &args, 408 const FunctionType::ExtInfo &info, 409 bool isVariadic) { 410 // FIXME: Kill copy. 411 SmallVector<CanQualType, 16> argTypes; 412 for (FunctionArgList::const_iterator i = args.begin(), e = args.end(); 413 i != e; ++i) 414 argTypes.push_back(Context.getCanonicalParamType((*i)->getType())); 415 416 RequiredArgs required = 417 (isVariadic ? RequiredArgs(args.size()) : RequiredArgs::All); 418 return arrangeLLVMFunctionInfo(GetReturnType(resultType), argTypes, info, 419 required); 420} 421 422const CGFunctionInfo &CodeGenTypes::arrangeNullaryFunction() { 423 return arrangeLLVMFunctionInfo(getContext().VoidTy, ArrayRef<CanQualType>(), 424 FunctionType::ExtInfo(), RequiredArgs::All); 425} 426 427/// Arrange the argument and result information for an abstract value 428/// of a given function type. This is the method which all of the 429/// above functions ultimately defer to. 430const CGFunctionInfo & 431CodeGenTypes::arrangeLLVMFunctionInfo(CanQualType resultType, 432 ArrayRef<CanQualType> argTypes, 433 FunctionType::ExtInfo info, 434 RequiredArgs required) { 435#ifndef NDEBUG 436 for (ArrayRef<CanQualType>::const_iterator 437 I = argTypes.begin(), E = argTypes.end(); I != E; ++I) 438 assert(I->isCanonicalAsParam()); 439#endif 440 441 unsigned CC = ClangCallConvToLLVMCallConv(info.getCC()); 442 443 // Lookup or create unique function info. 444 llvm::FoldingSetNodeID ID; 445 CGFunctionInfo::Profile(ID, info, required, resultType, argTypes); 446 447 void *insertPos = 0; 448 CGFunctionInfo *FI = FunctionInfos.FindNodeOrInsertPos(ID, insertPos); 449 if (FI) 450 return *FI; 451 452 // Construct the function info. We co-allocate the ArgInfos. 453 FI = CGFunctionInfo::create(CC, info, resultType, argTypes, required); 454 FunctionInfos.InsertNode(FI, insertPos); 455 456 bool inserted = FunctionsBeingProcessed.insert(FI); (void)inserted; 457 assert(inserted && "Recursively being processed?"); 458 459 // Compute ABI information. 460 getABIInfo().computeInfo(*FI); 461 462 // Loop over all of the computed argument and return value info. If any of 463 // them are direct or extend without a specified coerce type, specify the 464 // default now. 465 ABIArgInfo &retInfo = FI->getReturnInfo(); 466 if (retInfo.canHaveCoerceToType() && retInfo.getCoerceToType() == 0) 467 retInfo.setCoerceToType(ConvertType(FI->getReturnType())); 468 469 for (CGFunctionInfo::arg_iterator I = FI->arg_begin(), E = FI->arg_end(); 470 I != E; ++I) 471 if (I->info.canHaveCoerceToType() && I->info.getCoerceToType() == 0) 472 I->info.setCoerceToType(ConvertType(I->type)); 473 474 bool erased = FunctionsBeingProcessed.erase(FI); (void)erased; 475 assert(erased && "Not in set?"); 476 477 return *FI; 478} 479 480CGFunctionInfo *CGFunctionInfo::create(unsigned llvmCC, 481 const FunctionType::ExtInfo &info, 482 CanQualType resultType, 483 ArrayRef<CanQualType> argTypes, 484 RequiredArgs required) { 485 void *buffer = operator new(sizeof(CGFunctionInfo) + 486 sizeof(ArgInfo) * (argTypes.size() + 1)); 487 CGFunctionInfo *FI = new(buffer) CGFunctionInfo(); 488 FI->CallingConvention = llvmCC; 489 FI->EffectiveCallingConvention = llvmCC; 490 FI->ASTCallingConvention = info.getCC(); 491 FI->NoReturn = info.getNoReturn(); 492 FI->ReturnsRetained = info.getProducesResult(); 493 FI->Required = required; 494 FI->HasRegParm = info.getHasRegParm(); 495 FI->RegParm = info.getRegParm(); 496 FI->NumArgs = argTypes.size(); 497 FI->getArgsBuffer()[0].type = resultType; 498 for (unsigned i = 0, e = argTypes.size(); i != e; ++i) 499 FI->getArgsBuffer()[i + 1].type = argTypes[i]; 500 return FI; 501} 502 503/***/ 504 505void CodeGenTypes::GetExpandedTypes(QualType type, 506 SmallVectorImpl<llvm::Type*> &expandedTypes) { 507 if (const ConstantArrayType *AT = Context.getAsConstantArrayType(type)) { 508 uint64_t NumElts = AT->getSize().getZExtValue(); 509 for (uint64_t Elt = 0; Elt < NumElts; ++Elt) 510 GetExpandedTypes(AT->getElementType(), expandedTypes); 511 } else if (const RecordType *RT = type->getAs<RecordType>()) { 512 const RecordDecl *RD = RT->getDecl(); 513 assert(!RD->hasFlexibleArrayMember() && 514 "Cannot expand structure with flexible array."); 515 if (RD->isUnion()) { 516 // Unions can be here only in degenerative cases - all the fields are same 517 // after flattening. Thus we have to use the "largest" field. 518 const FieldDecl *LargestFD = 0; 519 CharUnits UnionSize = CharUnits::Zero(); 520 521 for (RecordDecl::field_iterator i = RD->field_begin(), e = RD->field_end(); 522 i != e; ++i) { 523 const FieldDecl *FD = *i; 524 assert(!FD->isBitField() && 525 "Cannot expand structure with bit-field members."); 526 CharUnits FieldSize = getContext().getTypeSizeInChars(FD->getType()); 527 if (UnionSize < FieldSize) { 528 UnionSize = FieldSize; 529 LargestFD = FD; 530 } 531 } 532 if (LargestFD) 533 GetExpandedTypes(LargestFD->getType(), expandedTypes); 534 } else { 535 for (RecordDecl::field_iterator i = RD->field_begin(), e = RD->field_end(); 536 i != e; ++i) { 537 assert(!i->isBitField() && 538 "Cannot expand structure with bit-field members."); 539 GetExpandedTypes(i->getType(), expandedTypes); 540 } 541 } 542 } else if (const ComplexType *CT = type->getAs<ComplexType>()) { 543 llvm::Type *EltTy = ConvertType(CT->getElementType()); 544 expandedTypes.push_back(EltTy); 545 expandedTypes.push_back(EltTy); 546 } else 547 expandedTypes.push_back(ConvertType(type)); 548} 549 550llvm::Function::arg_iterator 551CodeGenFunction::ExpandTypeFromArgs(QualType Ty, LValue LV, 552 llvm::Function::arg_iterator AI) { 553 assert(LV.isSimple() && 554 "Unexpected non-simple lvalue during struct expansion."); 555 556 if (const ConstantArrayType *AT = getContext().getAsConstantArrayType(Ty)) { 557 unsigned NumElts = AT->getSize().getZExtValue(); 558 QualType EltTy = AT->getElementType(); 559 for (unsigned Elt = 0; Elt < NumElts; ++Elt) { 560 llvm::Value *EltAddr = Builder.CreateConstGEP2_32(LV.getAddress(), 0, Elt); 561 LValue LV = MakeAddrLValue(EltAddr, EltTy); 562 AI = ExpandTypeFromArgs(EltTy, LV, AI); 563 } 564 } else if (const RecordType *RT = Ty->getAs<RecordType>()) { 565 RecordDecl *RD = RT->getDecl(); 566 if (RD->isUnion()) { 567 // Unions can be here only in degenerative cases - all the fields are same 568 // after flattening. Thus we have to use the "largest" field. 569 const FieldDecl *LargestFD = 0; 570 CharUnits UnionSize = CharUnits::Zero(); 571 572 for (RecordDecl::field_iterator i = RD->field_begin(), e = RD->field_end(); 573 i != e; ++i) { 574 const FieldDecl *FD = *i; 575 assert(!FD->isBitField() && 576 "Cannot expand structure with bit-field members."); 577 CharUnits FieldSize = getContext().getTypeSizeInChars(FD->getType()); 578 if (UnionSize < FieldSize) { 579 UnionSize = FieldSize; 580 LargestFD = FD; 581 } 582 } 583 if (LargestFD) { 584 // FIXME: What are the right qualifiers here? 585 LValue SubLV = EmitLValueForField(LV, LargestFD); 586 AI = ExpandTypeFromArgs(LargestFD->getType(), SubLV, AI); 587 } 588 } else { 589 for (RecordDecl::field_iterator i = RD->field_begin(), e = RD->field_end(); 590 i != e; ++i) { 591 FieldDecl *FD = *i; 592 QualType FT = FD->getType(); 593 594 // FIXME: What are the right qualifiers here? 595 LValue SubLV = EmitLValueForField(LV, FD); 596 AI = ExpandTypeFromArgs(FT, SubLV, AI); 597 } 598 } 599 } else if (const ComplexType *CT = Ty->getAs<ComplexType>()) { 600 QualType EltTy = CT->getElementType(); 601 llvm::Value *RealAddr = Builder.CreateStructGEP(LV.getAddress(), 0, "real"); 602 EmitStoreThroughLValue(RValue::get(AI++), MakeAddrLValue(RealAddr, EltTy)); 603 llvm::Value *ImagAddr = Builder.CreateStructGEP(LV.getAddress(), 1, "imag"); 604 EmitStoreThroughLValue(RValue::get(AI++), MakeAddrLValue(ImagAddr, EltTy)); 605 } else { 606 EmitStoreThroughLValue(RValue::get(AI), LV); 607 ++AI; 608 } 609 610 return AI; 611} 612 613/// EnterStructPointerForCoercedAccess - Given a struct pointer that we are 614/// accessing some number of bytes out of it, try to gep into the struct to get 615/// at its inner goodness. Dive as deep as possible without entering an element 616/// with an in-memory size smaller than DstSize. 617static llvm::Value * 618EnterStructPointerForCoercedAccess(llvm::Value *SrcPtr, 619 llvm::StructType *SrcSTy, 620 uint64_t DstSize, CodeGenFunction &CGF) { 621 // We can't dive into a zero-element struct. 622 if (SrcSTy->getNumElements() == 0) return SrcPtr; 623 624 llvm::Type *FirstElt = SrcSTy->getElementType(0); 625 626 // If the first elt is at least as large as what we're looking for, or if the 627 // first element is the same size as the whole struct, we can enter it. 628 uint64_t FirstEltSize = 629 CGF.CGM.getDataLayout().getTypeAllocSize(FirstElt); 630 if (FirstEltSize < DstSize && 631 FirstEltSize < CGF.CGM.getDataLayout().getTypeAllocSize(SrcSTy)) 632 return SrcPtr; 633 634 // GEP into the first element. 635 SrcPtr = CGF.Builder.CreateConstGEP2_32(SrcPtr, 0, 0, "coerce.dive"); 636 637 // If the first element is a struct, recurse. 638 llvm::Type *SrcTy = 639 cast<llvm::PointerType>(SrcPtr->getType())->getElementType(); 640 if (llvm::StructType *SrcSTy = dyn_cast<llvm::StructType>(SrcTy)) 641 return EnterStructPointerForCoercedAccess(SrcPtr, SrcSTy, DstSize, CGF); 642 643 return SrcPtr; 644} 645 646/// CoerceIntOrPtrToIntOrPtr - Convert a value Val to the specific Ty where both 647/// are either integers or pointers. This does a truncation of the value if it 648/// is too large or a zero extension if it is too small. 649static llvm::Value *CoerceIntOrPtrToIntOrPtr(llvm::Value *Val, 650 llvm::Type *Ty, 651 CodeGenFunction &CGF) { 652 if (Val->getType() == Ty) 653 return Val; 654 655 if (isa<llvm::PointerType>(Val->getType())) { 656 // If this is Pointer->Pointer avoid conversion to and from int. 657 if (isa<llvm::PointerType>(Ty)) 658 return CGF.Builder.CreateBitCast(Val, Ty, "coerce.val"); 659 660 // Convert the pointer to an integer so we can play with its width. 661 Val = CGF.Builder.CreatePtrToInt(Val, CGF.IntPtrTy, "coerce.val.pi"); 662 } 663 664 llvm::Type *DestIntTy = Ty; 665 if (isa<llvm::PointerType>(DestIntTy)) 666 DestIntTy = CGF.IntPtrTy; 667 668 if (Val->getType() != DestIntTy) 669 Val = CGF.Builder.CreateIntCast(Val, DestIntTy, false, "coerce.val.ii"); 670 671 if (isa<llvm::PointerType>(Ty)) 672 Val = CGF.Builder.CreateIntToPtr(Val, Ty, "coerce.val.ip"); 673 return Val; 674} 675 676 677 678/// CreateCoercedLoad - Create a load from \arg SrcPtr interpreted as 679/// a pointer to an object of type \arg Ty. 680/// 681/// This safely handles the case when the src type is smaller than the 682/// destination type; in this situation the values of bits which not 683/// present in the src are undefined. 684static llvm::Value *CreateCoercedLoad(llvm::Value *SrcPtr, 685 llvm::Type *Ty, 686 CodeGenFunction &CGF) { 687 llvm::Type *SrcTy = 688 cast<llvm::PointerType>(SrcPtr->getType())->getElementType(); 689 690 // If SrcTy and Ty are the same, just do a load. 691 if (SrcTy == Ty) 692 return CGF.Builder.CreateLoad(SrcPtr); 693 694 uint64_t DstSize = CGF.CGM.getDataLayout().getTypeAllocSize(Ty); 695 696 if (llvm::StructType *SrcSTy = dyn_cast<llvm::StructType>(SrcTy)) { 697 SrcPtr = EnterStructPointerForCoercedAccess(SrcPtr, SrcSTy, DstSize, CGF); 698 SrcTy = cast<llvm::PointerType>(SrcPtr->getType())->getElementType(); 699 } 700 701 uint64_t SrcSize = CGF.CGM.getDataLayout().getTypeAllocSize(SrcTy); 702 703 // If the source and destination are integer or pointer types, just do an 704 // extension or truncation to the desired type. 705 if ((isa<llvm::IntegerType>(Ty) || isa<llvm::PointerType>(Ty)) && 706 (isa<llvm::IntegerType>(SrcTy) || isa<llvm::PointerType>(SrcTy))) { 707 llvm::LoadInst *Load = CGF.Builder.CreateLoad(SrcPtr); 708 return CoerceIntOrPtrToIntOrPtr(Load, Ty, CGF); 709 } 710 711 // If load is legal, just bitcast the src pointer. 712 if (SrcSize >= DstSize) { 713 // Generally SrcSize is never greater than DstSize, since this means we are 714 // losing bits. However, this can happen in cases where the structure has 715 // additional padding, for example due to a user specified alignment. 716 // 717 // FIXME: Assert that we aren't truncating non-padding bits when have access 718 // to that information. 719 llvm::Value *Casted = 720 CGF.Builder.CreateBitCast(SrcPtr, llvm::PointerType::getUnqual(Ty)); 721 llvm::LoadInst *Load = CGF.Builder.CreateLoad(Casted); 722 // FIXME: Use better alignment / avoid requiring aligned load. 723 Load->setAlignment(1); 724 return Load; 725 } 726 727 // Otherwise do coercion through memory. This is stupid, but 728 // simple. 729 llvm::Value *Tmp = CGF.CreateTempAlloca(Ty); 730 llvm::Type *I8PtrTy = CGF.Builder.getInt8PtrTy(); 731 llvm::Value *Casted = CGF.Builder.CreateBitCast(Tmp, I8PtrTy); 732 llvm::Value *SrcCasted = CGF.Builder.CreateBitCast(SrcPtr, I8PtrTy); 733 // FIXME: Use better alignment. 734 CGF.Builder.CreateMemCpy(Casted, SrcCasted, 735 llvm::ConstantInt::get(CGF.IntPtrTy, SrcSize), 736 1, false); 737 return CGF.Builder.CreateLoad(Tmp); 738} 739 740// Function to store a first-class aggregate into memory. We prefer to 741// store the elements rather than the aggregate to be more friendly to 742// fast-isel. 743// FIXME: Do we need to recurse here? 744static void BuildAggStore(CodeGenFunction &CGF, llvm::Value *Val, 745 llvm::Value *DestPtr, bool DestIsVolatile, 746 bool LowAlignment) { 747 // Prefer scalar stores to first-class aggregate stores. 748 if (llvm::StructType *STy = 749 dyn_cast<llvm::StructType>(Val->getType())) { 750 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 751 llvm::Value *EltPtr = CGF.Builder.CreateConstGEP2_32(DestPtr, 0, i); 752 llvm::Value *Elt = CGF.Builder.CreateExtractValue(Val, i); 753 llvm::StoreInst *SI = CGF.Builder.CreateStore(Elt, EltPtr, 754 DestIsVolatile); 755 if (LowAlignment) 756 SI->setAlignment(1); 757 } 758 } else { 759 llvm::StoreInst *SI = CGF.Builder.CreateStore(Val, DestPtr, DestIsVolatile); 760 if (LowAlignment) 761 SI->setAlignment(1); 762 } 763} 764 765/// CreateCoercedStore - Create a store to \arg DstPtr from \arg Src, 766/// where the source and destination may have different types. 767/// 768/// This safely handles the case when the src type is larger than the 769/// destination type; the upper bits of the src will be lost. 770static void CreateCoercedStore(llvm::Value *Src, 771 llvm::Value *DstPtr, 772 bool DstIsVolatile, 773 CodeGenFunction &CGF) { 774 llvm::Type *SrcTy = Src->getType(); 775 llvm::Type *DstTy = 776 cast<llvm::PointerType>(DstPtr->getType())->getElementType(); 777 if (SrcTy == DstTy) { 778 CGF.Builder.CreateStore(Src, DstPtr, DstIsVolatile); 779 return; 780 } 781 782 uint64_t SrcSize = CGF.CGM.getDataLayout().getTypeAllocSize(SrcTy); 783 784 if (llvm::StructType *DstSTy = dyn_cast<llvm::StructType>(DstTy)) { 785 DstPtr = EnterStructPointerForCoercedAccess(DstPtr, DstSTy, SrcSize, CGF); 786 DstTy = cast<llvm::PointerType>(DstPtr->getType())->getElementType(); 787 } 788 789 // If the source and destination are integer or pointer types, just do an 790 // extension or truncation to the desired type. 791 if ((isa<llvm::IntegerType>(SrcTy) || isa<llvm::PointerType>(SrcTy)) && 792 (isa<llvm::IntegerType>(DstTy) || isa<llvm::PointerType>(DstTy))) { 793 Src = CoerceIntOrPtrToIntOrPtr(Src, DstTy, CGF); 794 CGF.Builder.CreateStore(Src, DstPtr, DstIsVolatile); 795 return; 796 } 797 798 uint64_t DstSize = CGF.CGM.getDataLayout().getTypeAllocSize(DstTy); 799 800 // If store is legal, just bitcast the src pointer. 801 if (SrcSize <= DstSize) { 802 llvm::Value *Casted = 803 CGF.Builder.CreateBitCast(DstPtr, llvm::PointerType::getUnqual(SrcTy)); 804 // FIXME: Use better alignment / avoid requiring aligned store. 805 BuildAggStore(CGF, Src, Casted, DstIsVolatile, true); 806 } else { 807 // Otherwise do coercion through memory. This is stupid, but 808 // simple. 809 810 // Generally SrcSize is never greater than DstSize, since this means we are 811 // losing bits. However, this can happen in cases where the structure has 812 // additional padding, for example due to a user specified alignment. 813 // 814 // FIXME: Assert that we aren't truncating non-padding bits when have access 815 // to that information. 816 llvm::Value *Tmp = CGF.CreateTempAlloca(SrcTy); 817 CGF.Builder.CreateStore(Src, Tmp); 818 llvm::Type *I8PtrTy = CGF.Builder.getInt8PtrTy(); 819 llvm::Value *Casted = CGF.Builder.CreateBitCast(Tmp, I8PtrTy); 820 llvm::Value *DstCasted = CGF.Builder.CreateBitCast(DstPtr, I8PtrTy); 821 // FIXME: Use better alignment. 822 CGF.Builder.CreateMemCpy(DstCasted, Casted, 823 llvm::ConstantInt::get(CGF.IntPtrTy, DstSize), 824 1, false); 825 } 826} 827 828/***/ 829 830bool CodeGenModule::ReturnTypeUsesSRet(const CGFunctionInfo &FI) { 831 return FI.getReturnInfo().isIndirect(); 832} 833 834bool CodeGenModule::ReturnTypeUsesFPRet(QualType ResultType) { 835 if (const BuiltinType *BT = ResultType->getAs<BuiltinType>()) { 836 switch (BT->getKind()) { 837 default: 838 return false; 839 case BuiltinType::Float: 840 return getTarget().useObjCFPRetForRealType(TargetInfo::Float); 841 case BuiltinType::Double: 842 return getTarget().useObjCFPRetForRealType(TargetInfo::Double); 843 case BuiltinType::LongDouble: 844 return getTarget().useObjCFPRetForRealType(TargetInfo::LongDouble); 845 } 846 } 847 848 return false; 849} 850 851bool CodeGenModule::ReturnTypeUsesFP2Ret(QualType ResultType) { 852 if (const ComplexType *CT = ResultType->getAs<ComplexType>()) { 853 if (const BuiltinType *BT = CT->getElementType()->getAs<BuiltinType>()) { 854 if (BT->getKind() == BuiltinType::LongDouble) 855 return getTarget().useObjCFP2RetForComplexLongDouble(); 856 } 857 } 858 859 return false; 860} 861 862llvm::FunctionType *CodeGenTypes::GetFunctionType(GlobalDecl GD) { 863 const CGFunctionInfo &FI = arrangeGlobalDeclaration(GD); 864 return GetFunctionType(FI); 865} 866 867llvm::FunctionType * 868CodeGenTypes::GetFunctionType(const CGFunctionInfo &FI) { 869 870 bool Inserted = FunctionsBeingProcessed.insert(&FI); (void)Inserted; 871 assert(Inserted && "Recursively being processed?"); 872 873 SmallVector<llvm::Type*, 8> argTypes; 874 llvm::Type *resultType = 0; 875 876 const ABIArgInfo &retAI = FI.getReturnInfo(); 877 switch (retAI.getKind()) { 878 case ABIArgInfo::Expand: 879 llvm_unreachable("Invalid ABI kind for return argument"); 880 881 case ABIArgInfo::Extend: 882 case ABIArgInfo::Direct: 883 resultType = retAI.getCoerceToType(); 884 break; 885 886 case ABIArgInfo::Indirect: { 887 assert(!retAI.getIndirectAlign() && "Align unused on indirect return."); 888 resultType = llvm::Type::getVoidTy(getLLVMContext()); 889 890 QualType ret = FI.getReturnType(); 891 llvm::Type *ty = ConvertType(ret); 892 unsigned addressSpace = Context.getTargetAddressSpace(ret); 893 argTypes.push_back(llvm::PointerType::get(ty, addressSpace)); 894 break; 895 } 896 897 case ABIArgInfo::Ignore: 898 resultType = llvm::Type::getVoidTy(getLLVMContext()); 899 break; 900 } 901 902 // Add in all of the required arguments. 903 CGFunctionInfo::const_arg_iterator it = FI.arg_begin(), ie; 904 if (FI.isVariadic()) { 905 ie = it + FI.getRequiredArgs().getNumRequiredArgs(); 906 } else { 907 ie = FI.arg_end(); 908 } 909 for (; it != ie; ++it) { 910 const ABIArgInfo &argAI = it->info; 911 912 // Insert a padding type to ensure proper alignment. 913 if (llvm::Type *PaddingType = argAI.getPaddingType()) 914 argTypes.push_back(PaddingType); 915 916 switch (argAI.getKind()) { 917 case ABIArgInfo::Ignore: 918 break; 919 920 case ABIArgInfo::Indirect: { 921 // indirect arguments are always on the stack, which is addr space #0. 922 llvm::Type *LTy = ConvertTypeForMem(it->type); 923 argTypes.push_back(LTy->getPointerTo()); 924 break; 925 } 926 927 case ABIArgInfo::Extend: 928 case ABIArgInfo::Direct: { 929 // If the coerce-to type is a first class aggregate, flatten it. Either 930 // way is semantically identical, but fast-isel and the optimizer 931 // generally likes scalar values better than FCAs. 932 llvm::Type *argType = argAI.getCoerceToType(); 933 if (llvm::StructType *st = dyn_cast<llvm::StructType>(argType)) { 934 for (unsigned i = 0, e = st->getNumElements(); i != e; ++i) 935 argTypes.push_back(st->getElementType(i)); 936 } else { 937 argTypes.push_back(argType); 938 } 939 break; 940 } 941 942 case ABIArgInfo::Expand: 943 GetExpandedTypes(it->type, argTypes); 944 break; 945 } 946 } 947 948 bool Erased = FunctionsBeingProcessed.erase(&FI); (void)Erased; 949 assert(Erased && "Not in set?"); 950 951 return llvm::FunctionType::get(resultType, argTypes, FI.isVariadic()); 952} 953 954llvm::Type *CodeGenTypes::GetFunctionTypeForVTable(GlobalDecl GD) { 955 const CXXMethodDecl *MD = cast<CXXMethodDecl>(GD.getDecl()); 956 const FunctionProtoType *FPT = MD->getType()->getAs<FunctionProtoType>(); 957 958 if (!isFuncTypeConvertible(FPT)) 959 return llvm::StructType::get(getLLVMContext()); 960 961 const CGFunctionInfo *Info; 962 if (isa<CXXDestructorDecl>(MD)) 963 Info = &arrangeCXXDestructor(cast<CXXDestructorDecl>(MD), GD.getDtorType()); 964 else 965 Info = &arrangeCXXMethodDeclaration(MD); 966 return GetFunctionType(*Info); 967} 968 969void CodeGenModule::ConstructAttributeList(const CGFunctionInfo &FI, 970 const Decl *TargetDecl, 971 AttributeListType &PAL, 972 unsigned &CallingConv, 973 bool AttrOnCallSite) { 974 llvm::AttrBuilder FuncAttrs; 975 llvm::AttrBuilder RetAttrs; 976 977 CallingConv = FI.getEffectiveCallingConvention(); 978 979 if (FI.isNoReturn()) 980 FuncAttrs.addAttribute(llvm::Attribute::NoReturn); 981 982 // FIXME: handle sseregparm someday... 983 if (TargetDecl) { 984 if (TargetDecl->hasAttr<ReturnsTwiceAttr>()) 985 FuncAttrs.addAttribute(llvm::Attribute::ReturnsTwice); 986 if (TargetDecl->hasAttr<NoThrowAttr>()) 987 FuncAttrs.addAttribute(llvm::Attribute::NoUnwind); 988 if (TargetDecl->hasAttr<NoReturnAttr>()) 989 FuncAttrs.addAttribute(llvm::Attribute::NoReturn); 990 991 if (const FunctionDecl *Fn = dyn_cast<FunctionDecl>(TargetDecl)) { 992 const FunctionProtoType *FPT = Fn->getType()->getAs<FunctionProtoType>(); 993 if (FPT && FPT->isNothrow(getContext())) 994 FuncAttrs.addAttribute(llvm::Attribute::NoUnwind); 995 // Don't use [[noreturn]] or _Noreturn for a call to a virtual function. 996 // These attributes are not inherited by overloads. 997 const CXXMethodDecl *MD = dyn_cast<CXXMethodDecl>(Fn); 998 if (Fn->isNoReturn() && !(AttrOnCallSite && MD && MD->isVirtual())) 999 FuncAttrs.addAttribute(llvm::Attribute::NoReturn); 1000 } 1001 1002 // 'const' and 'pure' attribute functions are also nounwind. 1003 if (TargetDecl->hasAttr<ConstAttr>()) { 1004 FuncAttrs.addAttribute(llvm::Attribute::ReadNone); 1005 FuncAttrs.addAttribute(llvm::Attribute::NoUnwind); 1006 } else if (TargetDecl->hasAttr<PureAttr>()) { 1007 FuncAttrs.addAttribute(llvm::Attribute::ReadOnly); 1008 FuncAttrs.addAttribute(llvm::Attribute::NoUnwind); 1009 } 1010 if (TargetDecl->hasAttr<MallocAttr>()) 1011 RetAttrs.addAttribute(llvm::Attribute::NoAlias); 1012 } 1013 1014 if (CodeGenOpts.OptimizeSize) 1015 FuncAttrs.addAttribute(llvm::Attribute::OptimizeForSize); 1016 if (CodeGenOpts.OptimizeSize == 2) 1017 FuncAttrs.addAttribute(llvm::Attribute::MinSize); 1018 if (CodeGenOpts.DisableRedZone) 1019 FuncAttrs.addAttribute(llvm::Attribute::NoRedZone); 1020 if (CodeGenOpts.NoImplicitFloat) 1021 FuncAttrs.addAttribute(llvm::Attribute::NoImplicitFloat); 1022 1023 if (AttrOnCallSite) { 1024 // Attributes that should go on the call site only. 1025 if (!CodeGenOpts.SimplifyLibCalls) 1026 FuncAttrs.addAttribute(llvm::Attribute::NoBuiltin); 1027 } else { 1028 // Attributes that should go on the function, but not the call site. 1029 if (!CodeGenOpts.DisableFPElim) { 1030 FuncAttrs.addAttribute("no-frame-pointer-elim", "false"); 1031 FuncAttrs.addAttribute("no-frame-pointer-elim-non-leaf", "false"); 1032 } else if (CodeGenOpts.OmitLeafFramePointer) { 1033 FuncAttrs.addAttribute("no-frame-pointer-elim", "false"); 1034 FuncAttrs.addAttribute("no-frame-pointer-elim-non-leaf", "true"); 1035 } else { 1036 FuncAttrs.addAttribute("no-frame-pointer-elim", "true"); 1037 FuncAttrs.addAttribute("no-frame-pointer-elim-non-leaf", "true"); 1038 } 1039 1040 FuncAttrs.addAttribute("less-precise-fpmad", 1041 CodeGenOpts.LessPreciseFPMAD ? "true" : "false"); 1042 FuncAttrs.addAttribute("no-infs-fp-math", 1043 CodeGenOpts.NoInfsFPMath ? "true" : "false"); 1044 FuncAttrs.addAttribute("no-nans-fp-math", 1045 CodeGenOpts.NoNaNsFPMath ? "true" : "false"); 1046 FuncAttrs.addAttribute("unsafe-fp-math", 1047 CodeGenOpts.UnsafeFPMath ? "true" : "false"); 1048 FuncAttrs.addAttribute("use-soft-float", 1049 CodeGenOpts.SoftFloat ? "true" : "false"); 1050 } 1051 1052 QualType RetTy = FI.getReturnType(); 1053 unsigned Index = 1; 1054 const ABIArgInfo &RetAI = FI.getReturnInfo(); 1055 switch (RetAI.getKind()) { 1056 case ABIArgInfo::Extend: 1057 if (RetTy->hasSignedIntegerRepresentation()) 1058 RetAttrs.addAttribute(llvm::Attribute::SExt); 1059 else if (RetTy->hasUnsignedIntegerRepresentation()) 1060 RetAttrs.addAttribute(llvm::Attribute::ZExt); 1061 break; 1062 case ABIArgInfo::Direct: 1063 case ABIArgInfo::Ignore: 1064 break; 1065 1066 case ABIArgInfo::Indirect: { 1067 llvm::AttrBuilder SRETAttrs; 1068 SRETAttrs.addAttribute(llvm::Attribute::StructRet); 1069 if (RetAI.getInReg()) 1070 SRETAttrs.addAttribute(llvm::Attribute::InReg); 1071 PAL.push_back(llvm:: 1072 AttributeSet::get(getLLVMContext(), Index, SRETAttrs)); 1073 1074 ++Index; 1075 // sret disables readnone and readonly 1076 FuncAttrs.removeAttribute(llvm::Attribute::ReadOnly) 1077 .removeAttribute(llvm::Attribute::ReadNone); 1078 break; 1079 } 1080 1081 case ABIArgInfo::Expand: 1082 llvm_unreachable("Invalid ABI kind for return argument"); 1083 } 1084 1085 if (RetAttrs.hasAttributes()) 1086 PAL.push_back(llvm:: 1087 AttributeSet::get(getLLVMContext(), 1088 llvm::AttributeSet::ReturnIndex, 1089 RetAttrs)); 1090 1091 for (CGFunctionInfo::const_arg_iterator it = FI.arg_begin(), 1092 ie = FI.arg_end(); it != ie; ++it) { 1093 QualType ParamType = it->type; 1094 const ABIArgInfo &AI = it->info; 1095 llvm::AttrBuilder Attrs; 1096 1097 if (AI.getPaddingType()) { 1098 if (AI.getPaddingInReg()) 1099 PAL.push_back(llvm::AttributeSet::get(getLLVMContext(), Index, 1100 llvm::Attribute::InReg)); 1101 // Increment Index if there is padding. 1102 ++Index; 1103 } 1104 1105 // 'restrict' -> 'noalias' is done in EmitFunctionProlog when we 1106 // have the corresponding parameter variable. It doesn't make 1107 // sense to do it here because parameters are so messed up. 1108 switch (AI.getKind()) { 1109 case ABIArgInfo::Extend: 1110 if (ParamType->isSignedIntegerOrEnumerationType()) 1111 Attrs.addAttribute(llvm::Attribute::SExt); 1112 else if (ParamType->isUnsignedIntegerOrEnumerationType()) 1113 Attrs.addAttribute(llvm::Attribute::ZExt); 1114 // FALL THROUGH 1115 case ABIArgInfo::Direct: 1116 if (AI.getInReg()) 1117 Attrs.addAttribute(llvm::Attribute::InReg); 1118 1119 // FIXME: handle sseregparm someday... 1120 1121 if (llvm::StructType *STy = 1122 dyn_cast<llvm::StructType>(AI.getCoerceToType())) { 1123 unsigned Extra = STy->getNumElements()-1; // 1 will be added below. 1124 if (Attrs.hasAttributes()) 1125 for (unsigned I = 0; I < Extra; ++I) 1126 PAL.push_back(llvm::AttributeSet::get(getLLVMContext(), Index + I, 1127 Attrs)); 1128 Index += Extra; 1129 } 1130 break; 1131 1132 case ABIArgInfo::Indirect: 1133 if (AI.getInReg()) 1134 Attrs.addAttribute(llvm::Attribute::InReg); 1135 1136 if (AI.getIndirectByVal()) 1137 Attrs.addAttribute(llvm::Attribute::ByVal); 1138 1139 Attrs.addAlignmentAttr(AI.getIndirectAlign()); 1140 1141 // byval disables readnone and readonly. 1142 FuncAttrs.removeAttribute(llvm::Attribute::ReadOnly) 1143 .removeAttribute(llvm::Attribute::ReadNone); 1144 break; 1145 1146 case ABIArgInfo::Ignore: 1147 // Skip increment, no matching LLVM parameter. 1148 continue; 1149 1150 case ABIArgInfo::Expand: { 1151 SmallVector<llvm::Type*, 8> types; 1152 // FIXME: This is rather inefficient. Do we ever actually need to do 1153 // anything here? The result should be just reconstructed on the other 1154 // side, so extension should be a non-issue. 1155 getTypes().GetExpandedTypes(ParamType, types); 1156 Index += types.size(); 1157 continue; 1158 } 1159 } 1160 1161 if (Attrs.hasAttributes()) 1162 PAL.push_back(llvm::AttributeSet::get(getLLVMContext(), Index, Attrs)); 1163 ++Index; 1164 } 1165 if (FuncAttrs.hasAttributes()) 1166 PAL.push_back(llvm:: 1167 AttributeSet::get(getLLVMContext(), 1168 llvm::AttributeSet::FunctionIndex, 1169 FuncAttrs)); 1170} 1171 1172/// An argument came in as a promoted argument; demote it back to its 1173/// declared type. 1174static llvm::Value *emitArgumentDemotion(CodeGenFunction &CGF, 1175 const VarDecl *var, 1176 llvm::Value *value) { 1177 llvm::Type *varType = CGF.ConvertType(var->getType()); 1178 1179 // This can happen with promotions that actually don't change the 1180 // underlying type, like the enum promotions. 1181 if (value->getType() == varType) return value; 1182 1183 assert((varType->isIntegerTy() || varType->isFloatingPointTy()) 1184 && "unexpected promotion type"); 1185 1186 if (isa<llvm::IntegerType>(varType)) 1187 return CGF.Builder.CreateTrunc(value, varType, "arg.unpromote"); 1188 1189 return CGF.Builder.CreateFPCast(value, varType, "arg.unpromote"); 1190} 1191 1192void CodeGenFunction::EmitFunctionProlog(const CGFunctionInfo &FI, 1193 llvm::Function *Fn, 1194 const FunctionArgList &Args) { 1195 // If this is an implicit-return-zero function, go ahead and 1196 // initialize the return value. TODO: it might be nice to have 1197 // a more general mechanism for this that didn't require synthesized 1198 // return statements. 1199 if (const FunctionDecl *FD = dyn_cast_or_null<FunctionDecl>(CurFuncDecl)) { 1200 if (FD->hasImplicitReturnZero()) { 1201 QualType RetTy = FD->getResultType().getUnqualifiedType(); 1202 llvm::Type* LLVMTy = CGM.getTypes().ConvertType(RetTy); 1203 llvm::Constant* Zero = llvm::Constant::getNullValue(LLVMTy); 1204 Builder.CreateStore(Zero, ReturnValue); 1205 } 1206 } 1207 1208 // FIXME: We no longer need the types from FunctionArgList; lift up and 1209 // simplify. 1210 1211 // Emit allocs for param decls. Give the LLVM Argument nodes names. 1212 llvm::Function::arg_iterator AI = Fn->arg_begin(); 1213 1214 // Name the struct return argument. 1215 if (CGM.ReturnTypeUsesSRet(FI)) { 1216 AI->setName("agg.result"); 1217 AI->addAttr(llvm::AttributeSet::get(getLLVMContext(), 1218 AI->getArgNo() + 1, 1219 llvm::Attribute::NoAlias)); 1220 ++AI; 1221 } 1222 1223 assert(FI.arg_size() == Args.size() && 1224 "Mismatch between function signature & arguments."); 1225 unsigned ArgNo = 1; 1226 CGFunctionInfo::const_arg_iterator info_it = FI.arg_begin(); 1227 for (FunctionArgList::const_iterator i = Args.begin(), e = Args.end(); 1228 i != e; ++i, ++info_it, ++ArgNo) { 1229 const VarDecl *Arg = *i; 1230 QualType Ty = info_it->type; 1231 const ABIArgInfo &ArgI = info_it->info; 1232 1233 bool isPromoted = 1234 isa<ParmVarDecl>(Arg) && cast<ParmVarDecl>(Arg)->isKNRPromoted(); 1235 1236 // Skip the dummy padding argument. 1237 if (ArgI.getPaddingType()) 1238 ++AI; 1239 1240 switch (ArgI.getKind()) { 1241 case ABIArgInfo::Indirect: { 1242 llvm::Value *V = AI; 1243 1244 if (!hasScalarEvaluationKind(Ty)) { 1245 // Aggregates and complex variables are accessed by reference. All we 1246 // need to do is realign the value, if requested 1247 if (ArgI.getIndirectRealign()) { 1248 llvm::Value *AlignedTemp = CreateMemTemp(Ty, "coerce"); 1249 1250 // Copy from the incoming argument pointer to the temporary with the 1251 // appropriate alignment. 1252 // 1253 // FIXME: We should have a common utility for generating an aggregate 1254 // copy. 1255 llvm::Type *I8PtrTy = Builder.getInt8PtrTy(); 1256 CharUnits Size = getContext().getTypeSizeInChars(Ty); 1257 llvm::Value *Dst = Builder.CreateBitCast(AlignedTemp, I8PtrTy); 1258 llvm::Value *Src = Builder.CreateBitCast(V, I8PtrTy); 1259 Builder.CreateMemCpy(Dst, 1260 Src, 1261 llvm::ConstantInt::get(IntPtrTy, 1262 Size.getQuantity()), 1263 ArgI.getIndirectAlign(), 1264 false); 1265 V = AlignedTemp; 1266 } 1267 } else { 1268 // Load scalar value from indirect argument. 1269 CharUnits Alignment = getContext().getTypeAlignInChars(Ty); 1270 V = EmitLoadOfScalar(V, false, Alignment.getQuantity(), Ty); 1271 1272 if (isPromoted) 1273 V = emitArgumentDemotion(*this, Arg, V); 1274 } 1275 EmitParmDecl(*Arg, V, ArgNo); 1276 break; 1277 } 1278 1279 case ABIArgInfo::Extend: 1280 case ABIArgInfo::Direct: { 1281 1282 // If we have the trivial case, handle it with no muss and fuss. 1283 if (!isa<llvm::StructType>(ArgI.getCoerceToType()) && 1284 ArgI.getCoerceToType() == ConvertType(Ty) && 1285 ArgI.getDirectOffset() == 0) { 1286 assert(AI != Fn->arg_end() && "Argument mismatch!"); 1287 llvm::Value *V = AI; 1288 1289 if (Arg->getType().isRestrictQualified()) 1290 AI->addAttr(llvm::AttributeSet::get(getLLVMContext(), 1291 AI->getArgNo() + 1, 1292 llvm::Attribute::NoAlias)); 1293 1294 // Ensure the argument is the correct type. 1295 if (V->getType() != ArgI.getCoerceToType()) 1296 V = Builder.CreateBitCast(V, ArgI.getCoerceToType()); 1297 1298 if (isPromoted) 1299 V = emitArgumentDemotion(*this, Arg, V); 1300 1301 // Because of merging of function types from multiple decls it is 1302 // possible for the type of an argument to not match the corresponding 1303 // type in the function type. Since we are codegening the callee 1304 // in here, add a cast to the argument type. 1305 llvm::Type *LTy = ConvertType(Arg->getType()); 1306 if (V->getType() != LTy) 1307 V = Builder.CreateBitCast(V, LTy); 1308 1309 EmitParmDecl(*Arg, V, ArgNo); 1310 break; 1311 } 1312 1313 llvm::AllocaInst *Alloca = CreateMemTemp(Ty, Arg->getName()); 1314 1315 // The alignment we need to use is the max of the requested alignment for 1316 // the argument plus the alignment required by our access code below. 1317 unsigned AlignmentToUse = 1318 CGM.getDataLayout().getABITypeAlignment(ArgI.getCoerceToType()); 1319 AlignmentToUse = std::max(AlignmentToUse, 1320 (unsigned)getContext().getDeclAlign(Arg).getQuantity()); 1321 1322 Alloca->setAlignment(AlignmentToUse); 1323 llvm::Value *V = Alloca; 1324 llvm::Value *Ptr = V; // Pointer to store into. 1325 1326 // If the value is offset in memory, apply the offset now. 1327 if (unsigned Offs = ArgI.getDirectOffset()) { 1328 Ptr = Builder.CreateBitCast(Ptr, Builder.getInt8PtrTy()); 1329 Ptr = Builder.CreateConstGEP1_32(Ptr, Offs); 1330 Ptr = Builder.CreateBitCast(Ptr, 1331 llvm::PointerType::getUnqual(ArgI.getCoerceToType())); 1332 } 1333 1334 // If the coerce-to type is a first class aggregate, we flatten it and 1335 // pass the elements. Either way is semantically identical, but fast-isel 1336 // and the optimizer generally likes scalar values better than FCAs. 1337 llvm::StructType *STy = dyn_cast<llvm::StructType>(ArgI.getCoerceToType()); 1338 if (STy && STy->getNumElements() > 1) { 1339 uint64_t SrcSize = CGM.getDataLayout().getTypeAllocSize(STy); 1340 llvm::Type *DstTy = 1341 cast<llvm::PointerType>(Ptr->getType())->getElementType(); 1342 uint64_t DstSize = CGM.getDataLayout().getTypeAllocSize(DstTy); 1343 1344 if (SrcSize <= DstSize) { 1345 Ptr = Builder.CreateBitCast(Ptr, llvm::PointerType::getUnqual(STy)); 1346 1347 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 1348 assert(AI != Fn->arg_end() && "Argument mismatch!"); 1349 AI->setName(Arg->getName() + ".coerce" + Twine(i)); 1350 llvm::Value *EltPtr = Builder.CreateConstGEP2_32(Ptr, 0, i); 1351 Builder.CreateStore(AI++, EltPtr); 1352 } 1353 } else { 1354 llvm::AllocaInst *TempAlloca = 1355 CreateTempAlloca(ArgI.getCoerceToType(), "coerce"); 1356 TempAlloca->setAlignment(AlignmentToUse); 1357 llvm::Value *TempV = TempAlloca; 1358 1359 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 1360 assert(AI != Fn->arg_end() && "Argument mismatch!"); 1361 AI->setName(Arg->getName() + ".coerce" + Twine(i)); 1362 llvm::Value *EltPtr = Builder.CreateConstGEP2_32(TempV, 0, i); 1363 Builder.CreateStore(AI++, EltPtr); 1364 } 1365 1366 Builder.CreateMemCpy(Ptr, TempV, DstSize, AlignmentToUse); 1367 } 1368 } else { 1369 // Simple case, just do a coerced store of the argument into the alloca. 1370 assert(AI != Fn->arg_end() && "Argument mismatch!"); 1371 AI->setName(Arg->getName() + ".coerce"); 1372 CreateCoercedStore(AI++, Ptr, /*DestIsVolatile=*/false, *this); 1373 } 1374 1375 1376 // Match to what EmitParmDecl is expecting for this type. 1377 if (CodeGenFunction::hasScalarEvaluationKind(Ty)) { 1378 V = EmitLoadOfScalar(V, false, AlignmentToUse, Ty); 1379 if (isPromoted) 1380 V = emitArgumentDemotion(*this, Arg, V); 1381 } 1382 EmitParmDecl(*Arg, V, ArgNo); 1383 continue; // Skip ++AI increment, already done. 1384 } 1385 1386 case ABIArgInfo::Expand: { 1387 // If this structure was expanded into multiple arguments then 1388 // we need to create a temporary and reconstruct it from the 1389 // arguments. 1390 llvm::AllocaInst *Alloca = CreateMemTemp(Ty); 1391 CharUnits Align = getContext().getDeclAlign(Arg); 1392 Alloca->setAlignment(Align.getQuantity()); 1393 LValue LV = MakeAddrLValue(Alloca, Ty, Align); 1394 llvm::Function::arg_iterator End = ExpandTypeFromArgs(Ty, LV, AI); 1395 EmitParmDecl(*Arg, Alloca, ArgNo); 1396 1397 // Name the arguments used in expansion and increment AI. 1398 unsigned Index = 0; 1399 for (; AI != End; ++AI, ++Index) 1400 AI->setName(Arg->getName() + "." + Twine(Index)); 1401 continue; 1402 } 1403 1404 case ABIArgInfo::Ignore: 1405 // Initialize the local variable appropriately. 1406 if (!hasScalarEvaluationKind(Ty)) 1407 EmitParmDecl(*Arg, CreateMemTemp(Ty), ArgNo); 1408 else 1409 EmitParmDecl(*Arg, llvm::UndefValue::get(ConvertType(Arg->getType())), 1410 ArgNo); 1411 1412 // Skip increment, no matching LLVM parameter. 1413 continue; 1414 } 1415 1416 ++AI; 1417 } 1418 assert(AI == Fn->arg_end() && "Argument mismatch!"); 1419} 1420 1421static void eraseUnusedBitCasts(llvm::Instruction *insn) { 1422 while (insn->use_empty()) { 1423 llvm::BitCastInst *bitcast = dyn_cast<llvm::BitCastInst>(insn); 1424 if (!bitcast) return; 1425 1426 // This is "safe" because we would have used a ConstantExpr otherwise. 1427 insn = cast<llvm::Instruction>(bitcast->getOperand(0)); 1428 bitcast->eraseFromParent(); 1429 } 1430} 1431 1432/// Try to emit a fused autorelease of a return result. 1433static llvm::Value *tryEmitFusedAutoreleaseOfResult(CodeGenFunction &CGF, 1434 llvm::Value *result) { 1435 // We must be immediately followed the cast. 1436 llvm::BasicBlock *BB = CGF.Builder.GetInsertBlock(); 1437 if (BB->empty()) return 0; 1438 if (&BB->back() != result) return 0; 1439 1440 llvm::Type *resultType = result->getType(); 1441 1442 // result is in a BasicBlock and is therefore an Instruction. 1443 llvm::Instruction *generator = cast<llvm::Instruction>(result); 1444 1445 SmallVector<llvm::Instruction*,4> insnsToKill; 1446 1447 // Look for: 1448 // %generator = bitcast %type1* %generator2 to %type2* 1449 while (llvm::BitCastInst *bitcast = dyn_cast<llvm::BitCastInst>(generator)) { 1450 // We would have emitted this as a constant if the operand weren't 1451 // an Instruction. 1452 generator = cast<llvm::Instruction>(bitcast->getOperand(0)); 1453 1454 // Require the generator to be immediately followed by the cast. 1455 if (generator->getNextNode() != bitcast) 1456 return 0; 1457 1458 insnsToKill.push_back(bitcast); 1459 } 1460 1461 // Look for: 1462 // %generator = call i8* @objc_retain(i8* %originalResult) 1463 // or 1464 // %generator = call i8* @objc_retainAutoreleasedReturnValue(i8* %originalResult) 1465 llvm::CallInst *call = dyn_cast<llvm::CallInst>(generator); 1466 if (!call) return 0; 1467 1468 bool doRetainAutorelease; 1469 1470 if (call->getCalledValue() == CGF.CGM.getARCEntrypoints().objc_retain) { 1471 doRetainAutorelease = true; 1472 } else if (call->getCalledValue() == CGF.CGM.getARCEntrypoints() 1473 .objc_retainAutoreleasedReturnValue) { 1474 doRetainAutorelease = false; 1475 1476 // If we emitted an assembly marker for this call (and the 1477 // ARCEntrypoints field should have been set if so), go looking 1478 // for that call. If we can't find it, we can't do this 1479 // optimization. But it should always be the immediately previous 1480 // instruction, unless we needed bitcasts around the call. 1481 if (CGF.CGM.getARCEntrypoints().retainAutoreleasedReturnValueMarker) { 1482 llvm::Instruction *prev = call->getPrevNode(); 1483 assert(prev); 1484 if (isa<llvm::BitCastInst>(prev)) { 1485 prev = prev->getPrevNode(); 1486 assert(prev); 1487 } 1488 assert(isa<llvm::CallInst>(prev)); 1489 assert(cast<llvm::CallInst>(prev)->getCalledValue() == 1490 CGF.CGM.getARCEntrypoints().retainAutoreleasedReturnValueMarker); 1491 insnsToKill.push_back(prev); 1492 } 1493 } else { 1494 return 0; 1495 } 1496 1497 result = call->getArgOperand(0); 1498 insnsToKill.push_back(call); 1499 1500 // Keep killing bitcasts, for sanity. Note that we no longer care 1501 // about precise ordering as long as there's exactly one use. 1502 while (llvm::BitCastInst *bitcast = dyn_cast<llvm::BitCastInst>(result)) { 1503 if (!bitcast->hasOneUse()) break; 1504 insnsToKill.push_back(bitcast); 1505 result = bitcast->getOperand(0); 1506 } 1507 1508 // Delete all the unnecessary instructions, from latest to earliest. 1509 for (SmallVectorImpl<llvm::Instruction*>::iterator 1510 i = insnsToKill.begin(), e = insnsToKill.end(); i != e; ++i) 1511 (*i)->eraseFromParent(); 1512 1513 // Do the fused retain/autorelease if we were asked to. 1514 if (doRetainAutorelease) 1515 result = CGF.EmitARCRetainAutoreleaseReturnValue(result); 1516 1517 // Cast back to the result type. 1518 return CGF.Builder.CreateBitCast(result, resultType); 1519} 1520 1521/// If this is a +1 of the value of an immutable 'self', remove it. 1522static llvm::Value *tryRemoveRetainOfSelf(CodeGenFunction &CGF, 1523 llvm::Value *result) { 1524 // This is only applicable to a method with an immutable 'self'. 1525 const ObjCMethodDecl *method = 1526 dyn_cast_or_null<ObjCMethodDecl>(CGF.CurCodeDecl); 1527 if (!method) return 0; 1528 const VarDecl *self = method->getSelfDecl(); 1529 if (!self->getType().isConstQualified()) return 0; 1530 1531 // Look for a retain call. 1532 llvm::CallInst *retainCall = 1533 dyn_cast<llvm::CallInst>(result->stripPointerCasts()); 1534 if (!retainCall || 1535 retainCall->getCalledValue() != CGF.CGM.getARCEntrypoints().objc_retain) 1536 return 0; 1537 1538 // Look for an ordinary load of 'self'. 1539 llvm::Value *retainedValue = retainCall->getArgOperand(0); 1540 llvm::LoadInst *load = 1541 dyn_cast<llvm::LoadInst>(retainedValue->stripPointerCasts()); 1542 if (!load || load->isAtomic() || load->isVolatile() || 1543 load->getPointerOperand() != CGF.GetAddrOfLocalVar(self)) 1544 return 0; 1545 1546 // Okay! Burn it all down. This relies for correctness on the 1547 // assumption that the retain is emitted as part of the return and 1548 // that thereafter everything is used "linearly". 1549 llvm::Type *resultType = result->getType(); 1550 eraseUnusedBitCasts(cast<llvm::Instruction>(result)); 1551 assert(retainCall->use_empty()); 1552 retainCall->eraseFromParent(); 1553 eraseUnusedBitCasts(cast<llvm::Instruction>(retainedValue)); 1554 1555 return CGF.Builder.CreateBitCast(load, resultType); 1556} 1557 1558/// Emit an ARC autorelease of the result of a function. 1559/// 1560/// \return the value to actually return from the function 1561static llvm::Value *emitAutoreleaseOfResult(CodeGenFunction &CGF, 1562 llvm::Value *result) { 1563 // If we're returning 'self', kill the initial retain. This is a 1564 // heuristic attempt to "encourage correctness" in the really unfortunate 1565 // case where we have a return of self during a dealloc and we desperately 1566 // need to avoid the possible autorelease. 1567 if (llvm::Value *self = tryRemoveRetainOfSelf(CGF, result)) 1568 return self; 1569 1570 // At -O0, try to emit a fused retain/autorelease. 1571 if (CGF.shouldUseFusedARCCalls()) 1572 if (llvm::Value *fused = tryEmitFusedAutoreleaseOfResult(CGF, result)) 1573 return fused; 1574 1575 return CGF.EmitARCAutoreleaseReturnValue(result); 1576} 1577 1578/// Heuristically search for a dominating store to the return-value slot. 1579static llvm::StoreInst *findDominatingStoreToReturnValue(CodeGenFunction &CGF) { 1580 // If there are multiple uses of the return-value slot, just check 1581 // for something immediately preceding the IP. Sometimes this can 1582 // happen with how we generate implicit-returns; it can also happen 1583 // with noreturn cleanups. 1584 if (!CGF.ReturnValue->hasOneUse()) { 1585 llvm::BasicBlock *IP = CGF.Builder.GetInsertBlock(); 1586 if (IP->empty()) return 0; 1587 llvm::StoreInst *store = dyn_cast<llvm::StoreInst>(&IP->back()); 1588 if (!store) return 0; 1589 if (store->getPointerOperand() != CGF.ReturnValue) return 0; 1590 assert(!store->isAtomic() && !store->isVolatile()); // see below 1591 return store; 1592 } 1593 1594 llvm::StoreInst *store = 1595 dyn_cast<llvm::StoreInst>(CGF.ReturnValue->use_back()); 1596 if (!store) return 0; 1597 1598 // These aren't actually possible for non-coerced returns, and we 1599 // only care about non-coerced returns on this code path. 1600 assert(!store->isAtomic() && !store->isVolatile()); 1601 1602 // Now do a first-and-dirty dominance check: just walk up the 1603 // single-predecessors chain from the current insertion point. 1604 llvm::BasicBlock *StoreBB = store->getParent(); 1605 llvm::BasicBlock *IP = CGF.Builder.GetInsertBlock(); 1606 while (IP != StoreBB) { 1607 if (!(IP = IP->getSinglePredecessor())) 1608 return 0; 1609 } 1610 1611 // Okay, the store's basic block dominates the insertion point; we 1612 // can do our thing. 1613 return store; 1614} 1615 1616/// Check whether 'this' argument of a callsite matches 'this' of the caller. 1617static bool checkThisPointer(llvm::Value *ThisArg, llvm::Value *This) { 1618 if (ThisArg == This) 1619 return true; 1620 // Check whether ThisArg is a bitcast of This. 1621 llvm::BitCastInst *Bitcast; 1622 if ((Bitcast = dyn_cast<llvm::BitCastInst>(ThisArg)) && 1623 Bitcast->getOperand(0) == This) 1624 return true; 1625 return false; 1626} 1627 1628void CodeGenFunction::EmitFunctionEpilog(const CGFunctionInfo &FI) { 1629 // Functions with no result always return void. 1630 if (ReturnValue == 0) { 1631 Builder.CreateRetVoid(); 1632 return; 1633 } 1634 1635 llvm::DebugLoc RetDbgLoc; 1636 llvm::Value *RV = 0; 1637 QualType RetTy = FI.getReturnType(); 1638 const ABIArgInfo &RetAI = FI.getReturnInfo(); 1639 1640 switch (RetAI.getKind()) { 1641 case ABIArgInfo::Indirect: { 1642 switch (getEvaluationKind(RetTy)) { 1643 case TEK_Complex: { 1644 ComplexPairTy RT = 1645 EmitLoadOfComplex(MakeNaturalAlignAddrLValue(ReturnValue, RetTy)); 1646 EmitStoreOfComplex(RT, 1647 MakeNaturalAlignAddrLValue(CurFn->arg_begin(), RetTy), 1648 /*isInit*/ true); 1649 break; 1650 } 1651 case TEK_Aggregate: 1652 // Do nothing; aggregrates get evaluated directly into the destination. 1653 break; 1654 case TEK_Scalar: 1655 EmitStoreOfScalar(Builder.CreateLoad(ReturnValue), 1656 MakeNaturalAlignAddrLValue(CurFn->arg_begin(), RetTy), 1657 /*isInit*/ true); 1658 break; 1659 } 1660 break; 1661 } 1662 1663 case ABIArgInfo::Extend: 1664 case ABIArgInfo::Direct: 1665 if (RetAI.getCoerceToType() == ConvertType(RetTy) && 1666 RetAI.getDirectOffset() == 0) { 1667 // The internal return value temp always will have pointer-to-return-type 1668 // type, just do a load. 1669 1670 // If there is a dominating store to ReturnValue, we can elide 1671 // the load, zap the store, and usually zap the alloca. 1672 if (llvm::StoreInst *SI = findDominatingStoreToReturnValue(*this)) { 1673 // Get the stored value and nuke the now-dead store. 1674 RetDbgLoc = SI->getDebugLoc(); 1675 RV = SI->getValueOperand(); 1676 SI->eraseFromParent(); 1677 1678 // If that was the only use of the return value, nuke it as well now. 1679 if (ReturnValue->use_empty() && isa<llvm::AllocaInst>(ReturnValue)) { 1680 cast<llvm::AllocaInst>(ReturnValue)->eraseFromParent(); 1681 ReturnValue = 0; 1682 } 1683 1684 // Otherwise, we have to do a simple load. 1685 } else { 1686 RV = Builder.CreateLoad(ReturnValue); 1687 } 1688 } else { 1689 llvm::Value *V = ReturnValue; 1690 // If the value is offset in memory, apply the offset now. 1691 if (unsigned Offs = RetAI.getDirectOffset()) { 1692 V = Builder.CreateBitCast(V, Builder.getInt8PtrTy()); 1693 V = Builder.CreateConstGEP1_32(V, Offs); 1694 V = Builder.CreateBitCast(V, 1695 llvm::PointerType::getUnqual(RetAI.getCoerceToType())); 1696 } 1697 1698 RV = CreateCoercedLoad(V, RetAI.getCoerceToType(), *this); 1699 } 1700 1701 // In ARC, end functions that return a retainable type with a call 1702 // to objc_autoreleaseReturnValue. 1703 if (AutoreleaseResult) { 1704 assert(getLangOpts().ObjCAutoRefCount && 1705 !FI.isReturnsRetained() && 1706 RetTy->isObjCRetainableType()); 1707 RV = emitAutoreleaseOfResult(*this, RV); 1708 } 1709 1710 break; 1711 1712 case ABIArgInfo::Ignore: 1713 break; 1714 1715 case ABIArgInfo::Expand: 1716 llvm_unreachable("Invalid ABI kind for return argument"); 1717 } 1718 1719 // If this function returns 'this', the last instruction is a CallInst 1720 // that returns 'this', and 'this' argument of the CallInst points to 1721 // the same object as CXXThisValue, use the return value from the CallInst. 1722 // We will not need to keep 'this' alive through the callsite. It also enables 1723 // optimizations in the backend, such as tail call optimization. 1724 if (CalleeWithThisReturn && CGM.getCXXABI().HasThisReturn(CurGD)) { 1725 llvm::BasicBlock *IP = Builder.GetInsertBlock(); 1726 llvm::CallInst *Callsite; 1727 if (!IP->empty() && (Callsite = dyn_cast<llvm::CallInst>(&IP->back())) && 1728 Callsite->getCalledFunction() == CalleeWithThisReturn && 1729 checkThisPointer(Callsite->getOperand(0), CXXThisValue)) 1730 RV = Builder.CreateBitCast(Callsite, RetAI.getCoerceToType()); 1731 } 1732 llvm::Instruction *Ret = RV ? Builder.CreateRet(RV) : Builder.CreateRetVoid(); 1733 if (!RetDbgLoc.isUnknown()) 1734 Ret->setDebugLoc(RetDbgLoc); 1735} 1736 1737void CodeGenFunction::EmitDelegateCallArg(CallArgList &args, 1738 const VarDecl *param) { 1739 // StartFunction converted the ABI-lowered parameter(s) into a 1740 // local alloca. We need to turn that into an r-value suitable 1741 // for EmitCall. 1742 llvm::Value *local = GetAddrOfLocalVar(param); 1743 1744 QualType type = param->getType(); 1745 1746 // For the most part, we just need to load the alloca, except: 1747 // 1) aggregate r-values are actually pointers to temporaries, and 1748 // 2) references to non-scalars are pointers directly to the aggregate. 1749 // I don't know why references to scalars are different here. 1750 if (const ReferenceType *ref = type->getAs<ReferenceType>()) { 1751 if (!hasScalarEvaluationKind(ref->getPointeeType())) 1752 return args.add(RValue::getAggregate(local), type); 1753 1754 // Locals which are references to scalars are represented 1755 // with allocas holding the pointer. 1756 return args.add(RValue::get(Builder.CreateLoad(local)), type); 1757 } 1758 1759 args.add(convertTempToRValue(local, type), type); 1760} 1761 1762static bool isProvablyNull(llvm::Value *addr) { 1763 return isa<llvm::ConstantPointerNull>(addr); 1764} 1765 1766static bool isProvablyNonNull(llvm::Value *addr) { 1767 return isa<llvm::AllocaInst>(addr); 1768} 1769 1770/// Emit the actual writing-back of a writeback. 1771static void emitWriteback(CodeGenFunction &CGF, 1772 const CallArgList::Writeback &writeback) { 1773 const LValue &srcLV = writeback.Source; 1774 llvm::Value *srcAddr = srcLV.getAddress(); 1775 assert(!isProvablyNull(srcAddr) && 1776 "shouldn't have writeback for provably null argument"); 1777 1778 llvm::BasicBlock *contBB = 0; 1779 1780 // If the argument wasn't provably non-null, we need to null check 1781 // before doing the store. 1782 bool provablyNonNull = isProvablyNonNull(srcAddr); 1783 if (!provablyNonNull) { 1784 llvm::BasicBlock *writebackBB = CGF.createBasicBlock("icr.writeback"); 1785 contBB = CGF.createBasicBlock("icr.done"); 1786 1787 llvm::Value *isNull = CGF.Builder.CreateIsNull(srcAddr, "icr.isnull"); 1788 CGF.Builder.CreateCondBr(isNull, contBB, writebackBB); 1789 CGF.EmitBlock(writebackBB); 1790 } 1791 1792 // Load the value to writeback. 1793 llvm::Value *value = CGF.Builder.CreateLoad(writeback.Temporary); 1794 1795 // Cast it back, in case we're writing an id to a Foo* or something. 1796 value = CGF.Builder.CreateBitCast(value, 1797 cast<llvm::PointerType>(srcAddr->getType())->getElementType(), 1798 "icr.writeback-cast"); 1799 1800 // Perform the writeback. 1801 1802 // If we have a "to use" value, it's something we need to emit a use 1803 // of. This has to be carefully threaded in: if it's done after the 1804 // release it's potentially undefined behavior (and the optimizer 1805 // will ignore it), and if it happens before the retain then the 1806 // optimizer could move the release there. 1807 if (writeback.ToUse) { 1808 assert(srcLV.getObjCLifetime() == Qualifiers::OCL_Strong); 1809 1810 // Retain the new value. No need to block-copy here: the block's 1811 // being passed up the stack. 1812 value = CGF.EmitARCRetainNonBlock(value); 1813 1814 // Emit the intrinsic use here. 1815 CGF.EmitARCIntrinsicUse(writeback.ToUse); 1816 1817 // Load the old value (primitively). 1818 llvm::Value *oldValue = CGF.EmitLoadOfScalar(srcLV); 1819 1820 // Put the new value in place (primitively). 1821 CGF.EmitStoreOfScalar(value, srcLV, /*init*/ false); 1822 1823 // Release the old value. 1824 CGF.EmitARCRelease(oldValue, srcLV.isARCPreciseLifetime()); 1825 1826 // Otherwise, we can just do a normal lvalue store. 1827 } else { 1828 CGF.EmitStoreThroughLValue(RValue::get(value), srcLV); 1829 } 1830 1831 // Jump to the continuation block. 1832 if (!provablyNonNull) 1833 CGF.EmitBlock(contBB); 1834} 1835 1836static void emitWritebacks(CodeGenFunction &CGF, 1837 const CallArgList &args) { 1838 for (CallArgList::writeback_iterator 1839 i = args.writeback_begin(), e = args.writeback_end(); i != e; ++i) 1840 emitWriteback(CGF, *i); 1841} 1842 1843static const Expr *maybeGetUnaryAddrOfOperand(const Expr *E) { 1844 if (const UnaryOperator *uop = dyn_cast<UnaryOperator>(E->IgnoreParens())) 1845 if (uop->getOpcode() == UO_AddrOf) 1846 return uop->getSubExpr(); 1847 return 0; 1848} 1849 1850/// Emit an argument that's being passed call-by-writeback. That is, 1851/// we are passing the address of 1852static void emitWritebackArg(CodeGenFunction &CGF, CallArgList &args, 1853 const ObjCIndirectCopyRestoreExpr *CRE) { 1854 LValue srcLV; 1855 1856 // Make an optimistic effort to emit the address as an l-value. 1857 // This can fail if the the argument expression is more complicated. 1858 if (const Expr *lvExpr = maybeGetUnaryAddrOfOperand(CRE->getSubExpr())) { 1859 srcLV = CGF.EmitLValue(lvExpr); 1860 1861 // Otherwise, just emit it as a scalar. 1862 } else { 1863 llvm::Value *srcAddr = CGF.EmitScalarExpr(CRE->getSubExpr()); 1864 1865 QualType srcAddrType = 1866 CRE->getSubExpr()->getType()->castAs<PointerType>()->getPointeeType(); 1867 srcLV = CGF.MakeNaturalAlignAddrLValue(srcAddr, srcAddrType); 1868 } 1869 llvm::Value *srcAddr = srcLV.getAddress(); 1870 1871 // The dest and src types don't necessarily match in LLVM terms 1872 // because of the crazy ObjC compatibility rules. 1873 1874 llvm::PointerType *destType = 1875 cast<llvm::PointerType>(CGF.ConvertType(CRE->getType())); 1876 1877 // If the address is a constant null, just pass the appropriate null. 1878 if (isProvablyNull(srcAddr)) { 1879 args.add(RValue::get(llvm::ConstantPointerNull::get(destType)), 1880 CRE->getType()); 1881 return; 1882 } 1883 1884 // Create the temporary. 1885 llvm::Value *temp = CGF.CreateTempAlloca(destType->getElementType(), 1886 "icr.temp"); 1887 // Loading an l-value can introduce a cleanup if the l-value is __weak, 1888 // and that cleanup will be conditional if we can't prove that the l-value 1889 // isn't null, so we need to register a dominating point so that the cleanups 1890 // system will make valid IR. 1891 CodeGenFunction::ConditionalEvaluation condEval(CGF); 1892 1893 // Zero-initialize it if we're not doing a copy-initialization. 1894 bool shouldCopy = CRE->shouldCopy(); 1895 if (!shouldCopy) { 1896 llvm::Value *null = 1897 llvm::ConstantPointerNull::get( 1898 cast<llvm::PointerType>(destType->getElementType())); 1899 CGF.Builder.CreateStore(null, temp); 1900 } 1901 1902 llvm::BasicBlock *contBB = 0; 1903 llvm::BasicBlock *originBB = 0; 1904 1905 // If the address is *not* known to be non-null, we need to switch. 1906 llvm::Value *finalArgument; 1907 1908 bool provablyNonNull = isProvablyNonNull(srcAddr); 1909 if (provablyNonNull) { 1910 finalArgument = temp; 1911 } else { 1912 llvm::Value *isNull = CGF.Builder.CreateIsNull(srcAddr, "icr.isnull"); 1913 1914 finalArgument = CGF.Builder.CreateSelect(isNull, 1915 llvm::ConstantPointerNull::get(destType), 1916 temp, "icr.argument"); 1917 1918 // If we need to copy, then the load has to be conditional, which 1919 // means we need control flow. 1920 if (shouldCopy) { 1921 originBB = CGF.Builder.GetInsertBlock(); 1922 contBB = CGF.createBasicBlock("icr.cont"); 1923 llvm::BasicBlock *copyBB = CGF.createBasicBlock("icr.copy"); 1924 CGF.Builder.CreateCondBr(isNull, contBB, copyBB); 1925 CGF.EmitBlock(copyBB); 1926 condEval.begin(CGF); 1927 } 1928 } 1929 1930 llvm::Value *valueToUse = 0; 1931 1932 // Perform a copy if necessary. 1933 if (shouldCopy) { 1934 RValue srcRV = CGF.EmitLoadOfLValue(srcLV); 1935 assert(srcRV.isScalar()); 1936 1937 llvm::Value *src = srcRV.getScalarVal(); 1938 src = CGF.Builder.CreateBitCast(src, destType->getElementType(), 1939 "icr.cast"); 1940 1941 // Use an ordinary store, not a store-to-lvalue. 1942 CGF.Builder.CreateStore(src, temp); 1943 1944 // If optimization is enabled, and the value was held in a 1945 // __strong variable, we need to tell the optimizer that this 1946 // value has to stay alive until we're doing the store back. 1947 // This is because the temporary is effectively unretained, 1948 // and so otherwise we can violate the high-level semantics. 1949 if (CGF.CGM.getCodeGenOpts().OptimizationLevel != 0 && 1950 srcLV.getObjCLifetime() == Qualifiers::OCL_Strong) { 1951 valueToUse = src; 1952 } 1953 } 1954 1955 // Finish the control flow if we needed it. 1956 if (shouldCopy && !provablyNonNull) { 1957 llvm::BasicBlock *copyBB = CGF.Builder.GetInsertBlock(); 1958 CGF.EmitBlock(contBB); 1959 1960 // Make a phi for the value to intrinsically use. 1961 if (valueToUse) { 1962 llvm::PHINode *phiToUse = CGF.Builder.CreatePHI(valueToUse->getType(), 2, 1963 "icr.to-use"); 1964 phiToUse->addIncoming(valueToUse, copyBB); 1965 phiToUse->addIncoming(llvm::UndefValue::get(valueToUse->getType()), 1966 originBB); 1967 valueToUse = phiToUse; 1968 } 1969 1970 condEval.end(CGF); 1971 } 1972 1973 args.addWriteback(srcLV, temp, valueToUse); 1974 args.add(RValue::get(finalArgument), CRE->getType()); 1975} 1976 1977void CodeGenFunction::EmitCallArg(CallArgList &args, const Expr *E, 1978 QualType type) { 1979 if (const ObjCIndirectCopyRestoreExpr *CRE 1980 = dyn_cast<ObjCIndirectCopyRestoreExpr>(E)) { 1981 assert(getLangOpts().ObjCAutoRefCount); 1982 assert(getContext().hasSameType(E->getType(), type)); 1983 return emitWritebackArg(*this, args, CRE); 1984 } 1985 1986 assert(type->isReferenceType() == E->isGLValue() && 1987 "reference binding to unmaterialized r-value!"); 1988 1989 if (E->isGLValue()) { 1990 assert(E->getObjectKind() == OK_Ordinary); 1991 return args.add(EmitReferenceBindingToExpr(E, /*InitializedDecl=*/0), 1992 type); 1993 } 1994 1995 if (hasAggregateEvaluationKind(type) && 1996 isa<ImplicitCastExpr>(E) && 1997 cast<CastExpr>(E)->getCastKind() == CK_LValueToRValue) { 1998 LValue L = EmitLValue(cast<CastExpr>(E)->getSubExpr()); 1999 assert(L.isSimple()); 2000 args.add(L.asAggregateRValue(), type, /*NeedsCopy*/true); 2001 return; 2002 } 2003 2004 args.add(EmitAnyExprToTemp(E), type); 2005} 2006 2007// In ObjC ARC mode with no ObjC ARC exception safety, tell the ARC 2008// optimizer it can aggressively ignore unwind edges. 2009void 2010CodeGenFunction::AddObjCARCExceptionMetadata(llvm::Instruction *Inst) { 2011 if (CGM.getCodeGenOpts().OptimizationLevel != 0 && 2012 !CGM.getCodeGenOpts().ObjCAutoRefCountExceptions) 2013 Inst->setMetadata("clang.arc.no_objc_arc_exceptions", 2014 CGM.getNoObjCARCExceptionsMetadata()); 2015} 2016 2017/// Emits a call to the given no-arguments nounwind runtime function. 2018llvm::CallInst * 2019CodeGenFunction::EmitNounwindRuntimeCall(llvm::Value *callee, 2020 const llvm::Twine &name) { 2021 return EmitNounwindRuntimeCall(callee, ArrayRef<llvm::Value*>(), name); 2022} 2023 2024/// Emits a call to the given nounwind runtime function. 2025llvm::CallInst * 2026CodeGenFunction::EmitNounwindRuntimeCall(llvm::Value *callee, 2027 ArrayRef<llvm::Value*> args, 2028 const llvm::Twine &name) { 2029 llvm::CallInst *call = EmitRuntimeCall(callee, args, name); 2030 call->setDoesNotThrow(); 2031 return call; 2032} 2033 2034/// Emits a simple call (never an invoke) to the given no-arguments 2035/// runtime function. 2036llvm::CallInst * 2037CodeGenFunction::EmitRuntimeCall(llvm::Value *callee, 2038 const llvm::Twine &name) { 2039 return EmitRuntimeCall(callee, ArrayRef<llvm::Value*>(), name); 2040} 2041 2042/// Emits a simple call (never an invoke) to the given runtime 2043/// function. 2044llvm::CallInst * 2045CodeGenFunction::EmitRuntimeCall(llvm::Value *callee, 2046 ArrayRef<llvm::Value*> args, 2047 const llvm::Twine &name) { 2048 llvm::CallInst *call = Builder.CreateCall(callee, args, name); 2049 call->setCallingConv(getRuntimeCC()); 2050 return call; 2051} 2052 2053/// Emits a call or invoke to the given noreturn runtime function. 2054void CodeGenFunction::EmitNoreturnRuntimeCallOrInvoke(llvm::Value *callee, 2055 ArrayRef<llvm::Value*> args) { 2056 if (getInvokeDest()) { 2057 llvm::InvokeInst *invoke = 2058 Builder.CreateInvoke(callee, 2059 getUnreachableBlock(), 2060 getInvokeDest(), 2061 args); 2062 invoke->setDoesNotReturn(); 2063 invoke->setCallingConv(getRuntimeCC()); 2064 } else { 2065 llvm::CallInst *call = Builder.CreateCall(callee, args); 2066 call->setDoesNotReturn(); 2067 call->setCallingConv(getRuntimeCC()); 2068 Builder.CreateUnreachable(); 2069 } 2070} 2071 2072/// Emits a call or invoke instruction to the given nullary runtime 2073/// function. 2074llvm::CallSite 2075CodeGenFunction::EmitRuntimeCallOrInvoke(llvm::Value *callee, 2076 const Twine &name) { 2077 return EmitRuntimeCallOrInvoke(callee, ArrayRef<llvm::Value*>(), name); 2078} 2079 2080/// Emits a call or invoke instruction to the given runtime function. 2081llvm::CallSite 2082CodeGenFunction::EmitRuntimeCallOrInvoke(llvm::Value *callee, 2083 ArrayRef<llvm::Value*> args, 2084 const Twine &name) { 2085 llvm::CallSite callSite = EmitCallOrInvoke(callee, args, name); 2086 callSite.setCallingConv(getRuntimeCC()); 2087 return callSite; 2088} 2089 2090llvm::CallSite 2091CodeGenFunction::EmitCallOrInvoke(llvm::Value *Callee, 2092 const Twine &Name) { 2093 return EmitCallOrInvoke(Callee, ArrayRef<llvm::Value *>(), Name); 2094} 2095 2096/// Emits a call or invoke instruction to the given function, depending 2097/// on the current state of the EH stack. 2098llvm::CallSite 2099CodeGenFunction::EmitCallOrInvoke(llvm::Value *Callee, 2100 ArrayRef<llvm::Value *> Args, 2101 const Twine &Name) { 2102 llvm::BasicBlock *InvokeDest = getInvokeDest(); 2103 2104 llvm::Instruction *Inst; 2105 if (!InvokeDest) 2106 Inst = Builder.CreateCall(Callee, Args, Name); 2107 else { 2108 llvm::BasicBlock *ContBB = createBasicBlock("invoke.cont"); 2109 Inst = Builder.CreateInvoke(Callee, ContBB, InvokeDest, Args, Name); 2110 EmitBlock(ContBB); 2111 } 2112 2113 // In ObjC ARC mode with no ObjC ARC exception safety, tell the ARC 2114 // optimizer it can aggressively ignore unwind edges. 2115 if (CGM.getLangOpts().ObjCAutoRefCount) 2116 AddObjCARCExceptionMetadata(Inst); 2117 2118 return Inst; 2119} 2120 2121static void checkArgMatches(llvm::Value *Elt, unsigned &ArgNo, 2122 llvm::FunctionType *FTy) { 2123 if (ArgNo < FTy->getNumParams()) 2124 assert(Elt->getType() == FTy->getParamType(ArgNo)); 2125 else 2126 assert(FTy->isVarArg()); 2127 ++ArgNo; 2128} 2129 2130void CodeGenFunction::ExpandTypeToArgs(QualType Ty, RValue RV, 2131 SmallVector<llvm::Value*,16> &Args, 2132 llvm::FunctionType *IRFuncTy) { 2133 if (const ConstantArrayType *AT = getContext().getAsConstantArrayType(Ty)) { 2134 unsigned NumElts = AT->getSize().getZExtValue(); 2135 QualType EltTy = AT->getElementType(); 2136 llvm::Value *Addr = RV.getAggregateAddr(); 2137 for (unsigned Elt = 0; Elt < NumElts; ++Elt) { 2138 llvm::Value *EltAddr = Builder.CreateConstGEP2_32(Addr, 0, Elt); 2139 RValue EltRV = convertTempToRValue(EltAddr, EltTy); 2140 ExpandTypeToArgs(EltTy, EltRV, Args, IRFuncTy); 2141 } 2142 } else if (const RecordType *RT = Ty->getAs<RecordType>()) { 2143 RecordDecl *RD = RT->getDecl(); 2144 assert(RV.isAggregate() && "Unexpected rvalue during struct expansion"); 2145 LValue LV = MakeAddrLValue(RV.getAggregateAddr(), Ty); 2146 2147 if (RD->isUnion()) { 2148 const FieldDecl *LargestFD = 0; 2149 CharUnits UnionSize = CharUnits::Zero(); 2150 2151 for (RecordDecl::field_iterator i = RD->field_begin(), e = RD->field_end(); 2152 i != e; ++i) { 2153 const FieldDecl *FD = *i; 2154 assert(!FD->isBitField() && 2155 "Cannot expand structure with bit-field members."); 2156 CharUnits FieldSize = getContext().getTypeSizeInChars(FD->getType()); 2157 if (UnionSize < FieldSize) { 2158 UnionSize = FieldSize; 2159 LargestFD = FD; 2160 } 2161 } 2162 if (LargestFD) { 2163 RValue FldRV = EmitRValueForField(LV, LargestFD); 2164 ExpandTypeToArgs(LargestFD->getType(), FldRV, Args, IRFuncTy); 2165 } 2166 } else { 2167 for (RecordDecl::field_iterator i = RD->field_begin(), e = RD->field_end(); 2168 i != e; ++i) { 2169 FieldDecl *FD = *i; 2170 2171 RValue FldRV = EmitRValueForField(LV, FD); 2172 ExpandTypeToArgs(FD->getType(), FldRV, Args, IRFuncTy); 2173 } 2174 } 2175 } else if (Ty->isAnyComplexType()) { 2176 ComplexPairTy CV = RV.getComplexVal(); 2177 Args.push_back(CV.first); 2178 Args.push_back(CV.second); 2179 } else { 2180 assert(RV.isScalar() && 2181 "Unexpected non-scalar rvalue during struct expansion."); 2182 2183 // Insert a bitcast as needed. 2184 llvm::Value *V = RV.getScalarVal(); 2185 if (Args.size() < IRFuncTy->getNumParams() && 2186 V->getType() != IRFuncTy->getParamType(Args.size())) 2187 V = Builder.CreateBitCast(V, IRFuncTy->getParamType(Args.size())); 2188 2189 Args.push_back(V); 2190 } 2191} 2192 2193 2194RValue CodeGenFunction::EmitCall(const CGFunctionInfo &CallInfo, 2195 llvm::Value *Callee, 2196 ReturnValueSlot ReturnValue, 2197 const CallArgList &CallArgs, 2198 const Decl *TargetDecl, 2199 llvm::Instruction **callOrInvoke) { 2200 // FIXME: We no longer need the types from CallArgs; lift up and simplify. 2201 SmallVector<llvm::Value*, 16> Args; 2202 2203 // Handle struct-return functions by passing a pointer to the 2204 // location that we would like to return into. 2205 QualType RetTy = CallInfo.getReturnType(); 2206 const ABIArgInfo &RetAI = CallInfo.getReturnInfo(); 2207 2208 // IRArgNo - Keep track of the argument number in the callee we're looking at. 2209 unsigned IRArgNo = 0; 2210 llvm::FunctionType *IRFuncTy = 2211 cast<llvm::FunctionType>( 2212 cast<llvm::PointerType>(Callee->getType())->getElementType()); 2213 2214 // If the call returns a temporary with struct return, create a temporary 2215 // alloca to hold the result, unless one is given to us. 2216 if (CGM.ReturnTypeUsesSRet(CallInfo)) { 2217 llvm::Value *Value = ReturnValue.getValue(); 2218 if (!Value) 2219 Value = CreateMemTemp(RetTy); 2220 Args.push_back(Value); 2221 checkArgMatches(Value, IRArgNo, IRFuncTy); 2222 } 2223 2224 assert(CallInfo.arg_size() == CallArgs.size() && 2225 "Mismatch between function signature & arguments."); 2226 CGFunctionInfo::const_arg_iterator info_it = CallInfo.arg_begin(); 2227 for (CallArgList::const_iterator I = CallArgs.begin(), E = CallArgs.end(); 2228 I != E; ++I, ++info_it) { 2229 const ABIArgInfo &ArgInfo = info_it->info; 2230 RValue RV = I->RV; 2231 2232 CharUnits TypeAlign = getContext().getTypeAlignInChars(I->Ty); 2233 2234 // Insert a padding argument to ensure proper alignment. 2235 if (llvm::Type *PaddingType = ArgInfo.getPaddingType()) { 2236 Args.push_back(llvm::UndefValue::get(PaddingType)); 2237 ++IRArgNo; 2238 } 2239 2240 switch (ArgInfo.getKind()) { 2241 case ABIArgInfo::Indirect: { 2242 if (RV.isScalar() || RV.isComplex()) { 2243 // Make a temporary alloca to pass the argument. 2244 llvm::AllocaInst *AI = CreateMemTemp(I->Ty); 2245 if (ArgInfo.getIndirectAlign() > AI->getAlignment()) 2246 AI->setAlignment(ArgInfo.getIndirectAlign()); 2247 Args.push_back(AI); 2248 2249 LValue argLV = 2250 MakeAddrLValue(Args.back(), I->Ty, TypeAlign); 2251 2252 if (RV.isScalar()) 2253 EmitStoreOfScalar(RV.getScalarVal(), argLV, /*init*/ true); 2254 else 2255 EmitStoreOfComplex(RV.getComplexVal(), argLV, /*init*/ true); 2256 2257 // Validate argument match. 2258 checkArgMatches(AI, IRArgNo, IRFuncTy); 2259 } else { 2260 // We want to avoid creating an unnecessary temporary+copy here; 2261 // however, we need one in three cases: 2262 // 1. If the argument is not byval, and we are required to copy the 2263 // source. (This case doesn't occur on any common architecture.) 2264 // 2. If the argument is byval, RV is not sufficiently aligned, and 2265 // we cannot force it to be sufficiently aligned. 2266 // 3. If the argument is byval, but RV is located in an address space 2267 // different than that of the argument (0). 2268 llvm::Value *Addr = RV.getAggregateAddr(); 2269 unsigned Align = ArgInfo.getIndirectAlign(); 2270 const llvm::DataLayout *TD = &CGM.getDataLayout(); 2271 const unsigned RVAddrSpace = Addr->getType()->getPointerAddressSpace(); 2272 const unsigned ArgAddrSpace = (IRArgNo < IRFuncTy->getNumParams() ? 2273 IRFuncTy->getParamType(IRArgNo)->getPointerAddressSpace() : 0); 2274 if ((!ArgInfo.getIndirectByVal() && I->NeedsCopy) || 2275 (ArgInfo.getIndirectByVal() && TypeAlign.getQuantity() < Align && 2276 llvm::getOrEnforceKnownAlignment(Addr, Align, TD) < Align) || 2277 (ArgInfo.getIndirectByVal() && (RVAddrSpace != ArgAddrSpace))) { 2278 // Create an aligned temporary, and copy to it. 2279 llvm::AllocaInst *AI = CreateMemTemp(I->Ty); 2280 if (Align > AI->getAlignment()) 2281 AI->setAlignment(Align); 2282 Args.push_back(AI); 2283 EmitAggregateCopy(AI, Addr, I->Ty, RV.isVolatileQualified()); 2284 2285 // Validate argument match. 2286 checkArgMatches(AI, IRArgNo, IRFuncTy); 2287 } else { 2288 // Skip the extra memcpy call. 2289 Args.push_back(Addr); 2290 2291 // Validate argument match. 2292 checkArgMatches(Addr, IRArgNo, IRFuncTy); 2293 } 2294 } 2295 break; 2296 } 2297 2298 case ABIArgInfo::Ignore: 2299 break; 2300 2301 case ABIArgInfo::Extend: 2302 case ABIArgInfo::Direct: { 2303 if (!isa<llvm::StructType>(ArgInfo.getCoerceToType()) && 2304 ArgInfo.getCoerceToType() == ConvertType(info_it->type) && 2305 ArgInfo.getDirectOffset() == 0) { 2306 llvm::Value *V; 2307 if (RV.isScalar()) 2308 V = RV.getScalarVal(); 2309 else 2310 V = Builder.CreateLoad(RV.getAggregateAddr()); 2311 2312 // If the argument doesn't match, perform a bitcast to coerce it. This 2313 // can happen due to trivial type mismatches. 2314 if (IRArgNo < IRFuncTy->getNumParams() && 2315 V->getType() != IRFuncTy->getParamType(IRArgNo)) 2316 V = Builder.CreateBitCast(V, IRFuncTy->getParamType(IRArgNo)); 2317 Args.push_back(V); 2318 2319 checkArgMatches(V, IRArgNo, IRFuncTy); 2320 break; 2321 } 2322 2323 // FIXME: Avoid the conversion through memory if possible. 2324 llvm::Value *SrcPtr; 2325 if (RV.isScalar() || RV.isComplex()) { 2326 SrcPtr = CreateMemTemp(I->Ty, "coerce"); 2327 LValue SrcLV = MakeAddrLValue(SrcPtr, I->Ty, TypeAlign); 2328 if (RV.isScalar()) { 2329 EmitStoreOfScalar(RV.getScalarVal(), SrcLV, /*init*/ true); 2330 } else { 2331 EmitStoreOfComplex(RV.getComplexVal(), SrcLV, /*init*/ true); 2332 } 2333 } else 2334 SrcPtr = RV.getAggregateAddr(); 2335 2336 // If the value is offset in memory, apply the offset now. 2337 if (unsigned Offs = ArgInfo.getDirectOffset()) { 2338 SrcPtr = Builder.CreateBitCast(SrcPtr, Builder.getInt8PtrTy()); 2339 SrcPtr = Builder.CreateConstGEP1_32(SrcPtr, Offs); 2340 SrcPtr = Builder.CreateBitCast(SrcPtr, 2341 llvm::PointerType::getUnqual(ArgInfo.getCoerceToType())); 2342 2343 } 2344 2345 // If the coerce-to type is a first class aggregate, we flatten it and 2346 // pass the elements. Either way is semantically identical, but fast-isel 2347 // and the optimizer generally likes scalar values better than FCAs. 2348 if (llvm::StructType *STy = 2349 dyn_cast<llvm::StructType>(ArgInfo.getCoerceToType())) { 2350 llvm::Type *SrcTy = 2351 cast<llvm::PointerType>(SrcPtr->getType())->getElementType(); 2352 uint64_t SrcSize = CGM.getDataLayout().getTypeAllocSize(SrcTy); 2353 uint64_t DstSize = CGM.getDataLayout().getTypeAllocSize(STy); 2354 2355 // If the source type is smaller than the destination type of the 2356 // coerce-to logic, copy the source value into a temp alloca the size 2357 // of the destination type to allow loading all of it. The bits past 2358 // the source value are left undef. 2359 if (SrcSize < DstSize) { 2360 llvm::AllocaInst *TempAlloca 2361 = CreateTempAlloca(STy, SrcPtr->getName() + ".coerce"); 2362 Builder.CreateMemCpy(TempAlloca, SrcPtr, SrcSize, 0); 2363 SrcPtr = TempAlloca; 2364 } else { 2365 SrcPtr = Builder.CreateBitCast(SrcPtr, 2366 llvm::PointerType::getUnqual(STy)); 2367 } 2368 2369 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 2370 llvm::Value *EltPtr = Builder.CreateConstGEP2_32(SrcPtr, 0, i); 2371 llvm::LoadInst *LI = Builder.CreateLoad(EltPtr); 2372 // We don't know what we're loading from. 2373 LI->setAlignment(1); 2374 Args.push_back(LI); 2375 2376 // Validate argument match. 2377 checkArgMatches(LI, IRArgNo, IRFuncTy); 2378 } 2379 } else { 2380 // In the simple case, just pass the coerced loaded value. 2381 Args.push_back(CreateCoercedLoad(SrcPtr, ArgInfo.getCoerceToType(), 2382 *this)); 2383 2384 // Validate argument match. 2385 checkArgMatches(Args.back(), IRArgNo, IRFuncTy); 2386 } 2387 2388 break; 2389 } 2390 2391 case ABIArgInfo::Expand: 2392 ExpandTypeToArgs(I->Ty, RV, Args, IRFuncTy); 2393 IRArgNo = Args.size(); 2394 break; 2395 } 2396 } 2397 2398 // If the callee is a bitcast of a function to a varargs pointer to function 2399 // type, check to see if we can remove the bitcast. This handles some cases 2400 // with unprototyped functions. 2401 if (llvm::ConstantExpr *CE = dyn_cast<llvm::ConstantExpr>(Callee)) 2402 if (llvm::Function *CalleeF = dyn_cast<llvm::Function>(CE->getOperand(0))) { 2403 llvm::PointerType *CurPT=cast<llvm::PointerType>(Callee->getType()); 2404 llvm::FunctionType *CurFT = 2405 cast<llvm::FunctionType>(CurPT->getElementType()); 2406 llvm::FunctionType *ActualFT = CalleeF->getFunctionType(); 2407 2408 if (CE->getOpcode() == llvm::Instruction::BitCast && 2409 ActualFT->getReturnType() == CurFT->getReturnType() && 2410 ActualFT->getNumParams() == CurFT->getNumParams() && 2411 ActualFT->getNumParams() == Args.size() && 2412 (CurFT->isVarArg() || !ActualFT->isVarArg())) { 2413 bool ArgsMatch = true; 2414 for (unsigned i = 0, e = ActualFT->getNumParams(); i != e; ++i) 2415 if (ActualFT->getParamType(i) != CurFT->getParamType(i)) { 2416 ArgsMatch = false; 2417 break; 2418 } 2419 2420 // Strip the cast if we can get away with it. This is a nice cleanup, 2421 // but also allows us to inline the function at -O0 if it is marked 2422 // always_inline. 2423 if (ArgsMatch) 2424 Callee = CalleeF; 2425 } 2426 } 2427 2428 unsigned CallingConv; 2429 CodeGen::AttributeListType AttributeList; 2430 CGM.ConstructAttributeList(CallInfo, TargetDecl, AttributeList, 2431 CallingConv, true); 2432 llvm::AttributeSet Attrs = llvm::AttributeSet::get(getLLVMContext(), 2433 AttributeList); 2434 2435 llvm::BasicBlock *InvokeDest = 0; 2436 if (!Attrs.hasAttribute(llvm::AttributeSet::FunctionIndex, 2437 llvm::Attribute::NoUnwind)) 2438 InvokeDest = getInvokeDest(); 2439 2440 llvm::CallSite CS; 2441 if (!InvokeDest) { 2442 CS = Builder.CreateCall(Callee, Args); 2443 } else { 2444 llvm::BasicBlock *Cont = createBasicBlock("invoke.cont"); 2445 CS = Builder.CreateInvoke(Callee, Cont, InvokeDest, Args); 2446 EmitBlock(Cont); 2447 } 2448 if (callOrInvoke) 2449 *callOrInvoke = CS.getInstruction(); 2450 2451 CS.setAttributes(Attrs); 2452 CS.setCallingConv(static_cast<llvm::CallingConv::ID>(CallingConv)); 2453 2454 // In ObjC ARC mode with no ObjC ARC exception safety, tell the ARC 2455 // optimizer it can aggressively ignore unwind edges. 2456 if (CGM.getLangOpts().ObjCAutoRefCount) 2457 AddObjCARCExceptionMetadata(CS.getInstruction()); 2458 2459 // If the call doesn't return, finish the basic block and clear the 2460 // insertion point; this allows the rest of IRgen to discard 2461 // unreachable code. 2462 if (CS.doesNotReturn()) { 2463 Builder.CreateUnreachable(); 2464 Builder.ClearInsertionPoint(); 2465 2466 // FIXME: For now, emit a dummy basic block because expr emitters in 2467 // generally are not ready to handle emitting expressions at unreachable 2468 // points. 2469 EnsureInsertPoint(); 2470 2471 // Return a reasonable RValue. 2472 return GetUndefRValue(RetTy); 2473 } 2474 2475 llvm::Instruction *CI = CS.getInstruction(); 2476 if (Builder.isNamePreserving() && !CI->getType()->isVoidTy()) 2477 CI->setName("call"); 2478 2479 // Emit any writebacks immediately. Arguably this should happen 2480 // after any return-value munging. 2481 if (CallArgs.hasWritebacks()) 2482 emitWritebacks(*this, CallArgs); 2483 2484 switch (RetAI.getKind()) { 2485 case ABIArgInfo::Indirect: 2486 return convertTempToRValue(Args[0], RetTy); 2487 2488 case ABIArgInfo::Ignore: 2489 // If we are ignoring an argument that had a result, make sure to 2490 // construct the appropriate return value for our caller. 2491 return GetUndefRValue(RetTy); 2492 2493 case ABIArgInfo::Extend: 2494 case ABIArgInfo::Direct: { 2495 llvm::Type *RetIRTy = ConvertType(RetTy); 2496 if (RetAI.getCoerceToType() == RetIRTy && RetAI.getDirectOffset() == 0) { 2497 switch (getEvaluationKind(RetTy)) { 2498 case TEK_Complex: { 2499 llvm::Value *Real = Builder.CreateExtractValue(CI, 0); 2500 llvm::Value *Imag = Builder.CreateExtractValue(CI, 1); 2501 return RValue::getComplex(std::make_pair(Real, Imag)); 2502 } 2503 case TEK_Aggregate: { 2504 llvm::Value *DestPtr = ReturnValue.getValue(); 2505 bool DestIsVolatile = ReturnValue.isVolatile(); 2506 2507 if (!DestPtr) { 2508 DestPtr = CreateMemTemp(RetTy, "agg.tmp"); 2509 DestIsVolatile = false; 2510 } 2511 BuildAggStore(*this, CI, DestPtr, DestIsVolatile, false); 2512 return RValue::getAggregate(DestPtr); 2513 } 2514 case TEK_Scalar: { 2515 // If the argument doesn't match, perform a bitcast to coerce it. This 2516 // can happen due to trivial type mismatches. 2517 llvm::Value *V = CI; 2518 if (V->getType() != RetIRTy) 2519 V = Builder.CreateBitCast(V, RetIRTy); 2520 return RValue::get(V); 2521 } 2522 } 2523 llvm_unreachable("bad evaluation kind"); 2524 } 2525 2526 llvm::Value *DestPtr = ReturnValue.getValue(); 2527 bool DestIsVolatile = ReturnValue.isVolatile(); 2528 2529 if (!DestPtr) { 2530 DestPtr = CreateMemTemp(RetTy, "coerce"); 2531 DestIsVolatile = false; 2532 } 2533 2534 // If the value is offset in memory, apply the offset now. 2535 llvm::Value *StorePtr = DestPtr; 2536 if (unsigned Offs = RetAI.getDirectOffset()) { 2537 StorePtr = Builder.CreateBitCast(StorePtr, Builder.getInt8PtrTy()); 2538 StorePtr = Builder.CreateConstGEP1_32(StorePtr, Offs); 2539 StorePtr = Builder.CreateBitCast(StorePtr, 2540 llvm::PointerType::getUnqual(RetAI.getCoerceToType())); 2541 } 2542 CreateCoercedStore(CI, StorePtr, DestIsVolatile, *this); 2543 2544 return convertTempToRValue(DestPtr, RetTy); 2545 } 2546 2547 case ABIArgInfo::Expand: 2548 llvm_unreachable("Invalid ABI kind for return argument"); 2549 } 2550 2551 llvm_unreachable("Unhandled ABIArgInfo::Kind"); 2552} 2553 2554/* VarArg handling */ 2555 2556llvm::Value *CodeGenFunction::EmitVAArg(llvm::Value *VAListAddr, QualType Ty) { 2557 return CGM.getTypes().getABIInfo().EmitVAArg(VAListAddr, Ty, *this); 2558} 2559