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" 33 using namespace clang; 34 using namespace CodeGen; 35 36 /***/ 37 38 static 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? 54 static 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. 60 static 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. 69 static 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. 75 const CGFunctionInfo & 76 CodeGenTypes::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. 88 static 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. 102 static 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. 110 static 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. 121 static 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. 131 const CGFunctionInfo & 132 CodeGenTypes::arrangeFreeFunctionType(CanQual<FunctionProtoType> FTP) { 133 SmallVector<CanQualType, 16> argTypes; 134 return ::arrangeFreeFunctionType(*this, argTypes, FTP); 135 } 136 137 static 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. 167 const CGFunctionInfo & 168 CodeGenTypes::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. 183 const CGFunctionInfo & 184 CodeGenTypes::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. 200 const CGFunctionInfo & 201 CodeGenTypes::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. 225 const CGFunctionInfo & 226 CodeGenTypes::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. 246 const CGFunctionInfo & 247 CodeGenTypes::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. 272 const CGFunctionInfo & 273 CodeGenTypes::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. 285 const CGFunctionInfo & 286 CodeGenTypes::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 } 310 311 const CGFunctionInfo & 312 CodeGenTypes::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. 327 static const CGFunctionInfo & 328 arrangeFreeFunctionLikeCall(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. 360 const CGFunctionInfo & 361 CodeGenTypes::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. 368 const CGFunctionInfo & 369 CodeGenTypes::arrangeBlockFunctionCall(const CallArgList &args, 370 const FunctionType *fnType) { 371 return arrangeFreeFunctionLikeCall(*this, args, fnType, 1); 372 } 373 374 const CGFunctionInfo & 375 CodeGenTypes::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. 389 const CGFunctionInfo & 390 CodeGenTypes::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 } 404 405 const CGFunctionInfo & 406 CodeGenTypes::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 } 421 422 const 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. 430 const CGFunctionInfo & 431 CodeGenTypes::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 } 479 480 CGFunctionInfo *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 /***/ 504 505 void 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 } 549 550 llvm::Function::arg_iterator 551 CodeGenFunction::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. 617 static llvm::Value * 618 EnterStructPointerForCoercedAccess(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. 649 static 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. 684 static 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? 744 static 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. 770 static 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 /***/ 829 830 bool CodeGenModule::ReturnTypeUsesSRet(const CGFunctionInfo &FI) { 831 return FI.getReturnInfo().isIndirect(); 832 } 833 834 bool 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 getContext().getTargetInfo().useObjCFPRetForRealType(TargetInfo::Float); 841 case BuiltinType::Double: 842 return getContext().getTargetInfo().useObjCFPRetForRealType(TargetInfo::Double); 843 case BuiltinType::LongDouble: 844 return getContext().getTargetInfo().useObjCFPRetForRealType( 845 TargetInfo::LongDouble); 846 } 847 } 848 849 return false; 850 } 851 852 bool CodeGenModule::ReturnTypeUsesFP2Ret(QualType ResultType) { 853 if (const ComplexType *CT = ResultType->getAs<ComplexType>()) { 854 if (const BuiltinType *BT = CT->getElementType()->getAs<BuiltinType>()) { 855 if (BT->getKind() == BuiltinType::LongDouble) 856 return getContext().getTargetInfo().useObjCFP2RetForComplexLongDouble(); 857 } 858 } 859 860 return false; 861 } 862 863 llvm::FunctionType *CodeGenTypes::GetFunctionType(GlobalDecl GD) { 864 const CGFunctionInfo &FI = arrangeGlobalDeclaration(GD); 865 return GetFunctionType(FI); 866 } 867 868 llvm::FunctionType * 869 CodeGenTypes::GetFunctionType(const CGFunctionInfo &FI) { 870 871 bool Inserted = FunctionsBeingProcessed.insert(&FI); (void)Inserted; 872 assert(Inserted && "Recursively being processed?"); 873 874 SmallVector<llvm::Type*, 8> argTypes; 875 llvm::Type *resultType = 0; 876 877 const ABIArgInfo &retAI = FI.getReturnInfo(); 878 switch (retAI.getKind()) { 879 case ABIArgInfo::Expand: 880 llvm_unreachable("Invalid ABI kind for return argument"); 881 882 case ABIArgInfo::Extend: 883 case ABIArgInfo::Direct: 884 resultType = retAI.getCoerceToType(); 885 break; 886 887 case ABIArgInfo::Indirect: { 888 assert(!retAI.getIndirectAlign() && "Align unused on indirect return."); 889 resultType = llvm::Type::getVoidTy(getLLVMContext()); 890 891 QualType ret = FI.getReturnType(); 892 llvm::Type *ty = ConvertType(ret); 893 unsigned addressSpace = Context.getTargetAddressSpace(ret); 894 argTypes.push_back(llvm::PointerType::get(ty, addressSpace)); 895 break; 896 } 897 898 case ABIArgInfo::Ignore: 899 resultType = llvm::Type::getVoidTy(getLLVMContext()); 900 break; 901 } 902 903 // Add in all of the required arguments. 904 CGFunctionInfo::const_arg_iterator it = FI.arg_begin(), ie; 905 if (FI.isVariadic()) { 906 ie = it + FI.getRequiredArgs().getNumRequiredArgs(); 907 } else { 908 ie = FI.arg_end(); 909 } 910 for (; it != ie; ++it) { 911 const ABIArgInfo &argAI = it->info; 912 913 // Insert a padding type to ensure proper alignment. 914 if (llvm::Type *PaddingType = argAI.getPaddingType()) 915 argTypes.push_back(PaddingType); 916 917 switch (argAI.getKind()) { 918 case ABIArgInfo::Ignore: 919 break; 920 921 case ABIArgInfo::Indirect: { 922 // indirect arguments are always on the stack, which is addr space #0. 923 llvm::Type *LTy = ConvertTypeForMem(it->type); 924 argTypes.push_back(LTy->getPointerTo()); 925 break; 926 } 927 928 case ABIArgInfo::Extend: 929 case ABIArgInfo::Direct: { 930 // If the coerce-to type is a first class aggregate, flatten it. Either 931 // way is semantically identical, but fast-isel and the optimizer 932 // generally likes scalar values better than FCAs. 933 llvm::Type *argType = argAI.getCoerceToType(); 934 if (llvm::StructType *st = dyn_cast<llvm::StructType>(argType)) { 935 for (unsigned i = 0, e = st->getNumElements(); i != e; ++i) 936 argTypes.push_back(st->getElementType(i)); 937 } else { 938 argTypes.push_back(argType); 939 } 940 break; 941 } 942 943 case ABIArgInfo::Expand: 944 GetExpandedTypes(it->type, argTypes); 945 break; 946 } 947 } 948 949 bool Erased = FunctionsBeingProcessed.erase(&FI); (void)Erased; 950 assert(Erased && "Not in set?"); 951 952 return llvm::FunctionType::get(resultType, argTypes, FI.isVariadic()); 953 } 954 955 llvm::Type *CodeGenTypes::GetFunctionTypeForVTable(GlobalDecl GD) { 956 const CXXMethodDecl *MD = cast<CXXMethodDecl>(GD.getDecl()); 957 const FunctionProtoType *FPT = MD->getType()->getAs<FunctionProtoType>(); 958 959 if (!isFuncTypeConvertible(FPT)) 960 return llvm::StructType::get(getLLVMContext()); 961 962 const CGFunctionInfo *Info; 963 if (isa<CXXDestructorDecl>(MD)) 964 Info = &arrangeCXXDestructor(cast<CXXDestructorDecl>(MD), GD.getDtorType()); 965 else 966 Info = &arrangeCXXMethodDeclaration(MD); 967 return GetFunctionType(*Info); 968 } 969 970 void CodeGenModule::ConstructAttributeList(const CGFunctionInfo &FI, 971 const Decl *TargetDecl, 972 AttributeListType &PAL, 973 unsigned &CallingConv, 974 bool AttrOnCallSite) { 975 llvm::AttrBuilder FuncAttrs; 976 llvm::AttrBuilder RetAttrs; 977 978 CallingConv = FI.getEffectiveCallingConvention(); 979 980 if (FI.isNoReturn()) 981 FuncAttrs.addAttribute(llvm::Attribute::NoReturn); 982 983 // FIXME: handle sseregparm someday... 984 if (TargetDecl) { 985 if (TargetDecl->hasAttr<ReturnsTwiceAttr>()) 986 FuncAttrs.addAttribute(llvm::Attribute::ReturnsTwice); 987 if (TargetDecl->hasAttr<NoThrowAttr>()) 988 FuncAttrs.addAttribute(llvm::Attribute::NoUnwind); 989 if (TargetDecl->hasAttr<NoReturnAttr>()) 990 FuncAttrs.addAttribute(llvm::Attribute::NoReturn); 991 992 if (const FunctionDecl *Fn = dyn_cast<FunctionDecl>(TargetDecl)) { 993 const FunctionProtoType *FPT = Fn->getType()->getAs<FunctionProtoType>(); 994 if (FPT && FPT->isNothrow(getContext())) 995 FuncAttrs.addAttribute(llvm::Attribute::NoUnwind); 996 // Don't use [[noreturn]] or _Noreturn for a call to a virtual function. 997 // These attributes are not inherited by overloads. 998 const CXXMethodDecl *MD = dyn_cast<CXXMethodDecl>(Fn); 999 if (Fn->isNoReturn() && !(AttrOnCallSite && MD && MD->isVirtual())) 1000 FuncAttrs.addAttribute(llvm::Attribute::NoReturn); 1001 } 1002 1003 // 'const' and 'pure' attribute functions are also nounwind. 1004 if (TargetDecl->hasAttr<ConstAttr>()) { 1005 FuncAttrs.addAttribute(llvm::Attribute::ReadNone); 1006 FuncAttrs.addAttribute(llvm::Attribute::NoUnwind); 1007 } else if (TargetDecl->hasAttr<PureAttr>()) { 1008 FuncAttrs.addAttribute(llvm::Attribute::ReadOnly); 1009 FuncAttrs.addAttribute(llvm::Attribute::NoUnwind); 1010 } 1011 if (TargetDecl->hasAttr<MallocAttr>()) 1012 RetAttrs.addAttribute(llvm::Attribute::NoAlias); 1013 } 1014 1015 if (CodeGenOpts.OptimizeSize) 1016 FuncAttrs.addAttribute(llvm::Attribute::OptimizeForSize); 1017 if (CodeGenOpts.OptimizeSize == 2) 1018 FuncAttrs.addAttribute(llvm::Attribute::MinSize); 1019 if (CodeGenOpts.DisableRedZone) 1020 FuncAttrs.addAttribute(llvm::Attribute::NoRedZone); 1021 if (CodeGenOpts.NoImplicitFloat) 1022 FuncAttrs.addAttribute(llvm::Attribute::NoImplicitFloat); 1023 1024 if (AttrOnCallSite) { 1025 // Attributes that should go on the call site only. 1026 if (!CodeGenOpts.SimplifyLibCalls) 1027 FuncAttrs.addAttribute(llvm::Attribute::NoBuiltin); 1028 } else { 1029 // Attributes that should go on the function, but not the call site. 1030 if (!CodeGenOpts.DisableFPElim) { 1031 FuncAttrs.addAttribute("no-frame-pointer-elim", "false"); 1032 FuncAttrs.addAttribute("no-frame-pointer-elim-non-leaf", "false"); 1033 } else if (CodeGenOpts.OmitLeafFramePointer) { 1034 FuncAttrs.addAttribute("no-frame-pointer-elim", "false"); 1035 FuncAttrs.addAttribute("no-frame-pointer-elim-non-leaf", "true"); 1036 } else { 1037 FuncAttrs.addAttribute("no-frame-pointer-elim", "true"); 1038 FuncAttrs.addAttribute("no-frame-pointer-elim-non-leaf", "true"); 1039 } 1040 1041 FuncAttrs.addAttribute("less-precise-fpmad", 1042 CodeGenOpts.LessPreciseFPMAD ? "true" : "false"); 1043 FuncAttrs.addAttribute("no-infs-fp-math", 1044 CodeGenOpts.NoInfsFPMath ? "true" : "false"); 1045 FuncAttrs.addAttribute("no-nans-fp-math", 1046 CodeGenOpts.NoNaNsFPMath ? "true" : "false"); 1047 FuncAttrs.addAttribute("unsafe-fp-math", 1048 CodeGenOpts.UnsafeFPMath ? "true" : "false"); 1049 FuncAttrs.addAttribute("use-soft-float", 1050 CodeGenOpts.SoftFloat ? "true" : "false"); 1051 } 1052 1053 QualType RetTy = FI.getReturnType(); 1054 unsigned Index = 1; 1055 const ABIArgInfo &RetAI = FI.getReturnInfo(); 1056 switch (RetAI.getKind()) { 1057 case ABIArgInfo::Extend: 1058 if (RetTy->hasSignedIntegerRepresentation()) 1059 RetAttrs.addAttribute(llvm::Attribute::SExt); 1060 else if (RetTy->hasUnsignedIntegerRepresentation()) 1061 RetAttrs.addAttribute(llvm::Attribute::ZExt); 1062 break; 1063 case ABIArgInfo::Direct: 1064 case ABIArgInfo::Ignore: 1065 break; 1066 1067 case ABIArgInfo::Indirect: { 1068 llvm::AttrBuilder SRETAttrs; 1069 SRETAttrs.addAttribute(llvm::Attribute::StructRet); 1070 if (RetAI.getInReg()) 1071 SRETAttrs.addAttribute(llvm::Attribute::InReg); 1072 PAL.push_back(llvm:: 1073 AttributeSet::get(getLLVMContext(), Index, SRETAttrs)); 1074 1075 ++Index; 1076 // sret disables readnone and readonly 1077 FuncAttrs.removeAttribute(llvm::Attribute::ReadOnly) 1078 .removeAttribute(llvm::Attribute::ReadNone); 1079 break; 1080 } 1081 1082 case ABIArgInfo::Expand: 1083 llvm_unreachable("Invalid ABI kind for return argument"); 1084 } 1085 1086 if (RetAttrs.hasAttributes()) 1087 PAL.push_back(llvm:: 1088 AttributeSet::get(getLLVMContext(), 1089 llvm::AttributeSet::ReturnIndex, 1090 RetAttrs)); 1091 1092 for (CGFunctionInfo::const_arg_iterator it = FI.arg_begin(), 1093 ie = FI.arg_end(); it != ie; ++it) { 1094 QualType ParamType = it->type; 1095 const ABIArgInfo &AI = it->info; 1096 llvm::AttrBuilder Attrs; 1097 1098 if (AI.getPaddingType()) { 1099 if (AI.getPaddingInReg()) 1100 PAL.push_back(llvm::AttributeSet::get(getLLVMContext(), Index, 1101 llvm::Attribute::InReg)); 1102 // Increment Index if there is padding. 1103 ++Index; 1104 } 1105 1106 // 'restrict' -> 'noalias' is done in EmitFunctionProlog when we 1107 // have the corresponding parameter variable. It doesn't make 1108 // sense to do it here because parameters are so messed up. 1109 switch (AI.getKind()) { 1110 case ABIArgInfo::Extend: 1111 if (ParamType->isSignedIntegerOrEnumerationType()) 1112 Attrs.addAttribute(llvm::Attribute::SExt); 1113 else if (ParamType->isUnsignedIntegerOrEnumerationType()) 1114 Attrs.addAttribute(llvm::Attribute::ZExt); 1115 // FALL THROUGH 1116 case ABIArgInfo::Direct: 1117 if (AI.getInReg()) 1118 Attrs.addAttribute(llvm::Attribute::InReg); 1119 1120 // FIXME: handle sseregparm someday... 1121 1122 if (llvm::StructType *STy = 1123 dyn_cast<llvm::StructType>(AI.getCoerceToType())) { 1124 unsigned Extra = STy->getNumElements()-1; // 1 will be added below. 1125 if (Attrs.hasAttributes()) 1126 for (unsigned I = 0; I < Extra; ++I) 1127 PAL.push_back(llvm::AttributeSet::get(getLLVMContext(), Index + I, 1128 Attrs)); 1129 Index += Extra; 1130 } 1131 break; 1132 1133 case ABIArgInfo::Indirect: 1134 if (AI.getInReg()) 1135 Attrs.addAttribute(llvm::Attribute::InReg); 1136 1137 if (AI.getIndirectByVal()) 1138 Attrs.addAttribute(llvm::Attribute::ByVal); 1139 1140 Attrs.addAlignmentAttr(AI.getIndirectAlign()); 1141 1142 // byval disables readnone and readonly. 1143 FuncAttrs.removeAttribute(llvm::Attribute::ReadOnly) 1144 .removeAttribute(llvm::Attribute::ReadNone); 1145 break; 1146 1147 case ABIArgInfo::Ignore: 1148 // Skip increment, no matching LLVM parameter. 1149 continue; 1150 1151 case ABIArgInfo::Expand: { 1152 SmallVector<llvm::Type*, 8> types; 1153 // FIXME: This is rather inefficient. Do we ever actually need to do 1154 // anything here? The result should be just reconstructed on the other 1155 // side, so extension should be a non-issue. 1156 getTypes().GetExpandedTypes(ParamType, types); 1157 Index += types.size(); 1158 continue; 1159 } 1160 } 1161 1162 if (Attrs.hasAttributes()) 1163 PAL.push_back(llvm::AttributeSet::get(getLLVMContext(), Index, Attrs)); 1164 ++Index; 1165 } 1166 if (FuncAttrs.hasAttributes()) 1167 PAL.push_back(llvm:: 1168 AttributeSet::get(getLLVMContext(), 1169 llvm::AttributeSet::FunctionIndex, 1170 FuncAttrs)); 1171 } 1172 1173 /// An argument came in as a promoted argument; demote it back to its 1174 /// declared type. 1175 static llvm::Value *emitArgumentDemotion(CodeGenFunction &CGF, 1176 const VarDecl *var, 1177 llvm::Value *value) { 1178 llvm::Type *varType = CGF.ConvertType(var->getType()); 1179 1180 // This can happen with promotions that actually don't change the 1181 // underlying type, like the enum promotions. 1182 if (value->getType() == varType) return value; 1183 1184 assert((varType->isIntegerTy() || varType->isFloatingPointTy()) 1185 && "unexpected promotion type"); 1186 1187 if (isa<llvm::IntegerType>(varType)) 1188 return CGF.Builder.CreateTrunc(value, varType, "arg.unpromote"); 1189 1190 return CGF.Builder.CreateFPCast(value, varType, "arg.unpromote"); 1191 } 1192 1193 void CodeGenFunction::EmitFunctionProlog(const CGFunctionInfo &FI, 1194 llvm::Function *Fn, 1195 const FunctionArgList &Args) { 1196 // If this is an implicit-return-zero function, go ahead and 1197 // initialize the return value. TODO: it might be nice to have 1198 // a more general mechanism for this that didn't require synthesized 1199 // return statements. 1200 if (const FunctionDecl *FD = dyn_cast_or_null<FunctionDecl>(CurFuncDecl)) { 1201 if (FD->hasImplicitReturnZero()) { 1202 QualType RetTy = FD->getResultType().getUnqualifiedType(); 1203 llvm::Type* LLVMTy = CGM.getTypes().ConvertType(RetTy); 1204 llvm::Constant* Zero = llvm::Constant::getNullValue(LLVMTy); 1205 Builder.CreateStore(Zero, ReturnValue); 1206 } 1207 } 1208 1209 // FIXME: We no longer need the types from FunctionArgList; lift up and 1210 // simplify. 1211 1212 // Emit allocs for param decls. Give the LLVM Argument nodes names. 1213 llvm::Function::arg_iterator AI = Fn->arg_begin(); 1214 1215 // Name the struct return argument. 1216 if (CGM.ReturnTypeUsesSRet(FI)) { 1217 AI->setName("agg.result"); 1218 AI->addAttr(llvm::AttributeSet::get(getLLVMContext(), 1219 AI->getArgNo() + 1, 1220 llvm::Attribute::NoAlias)); 1221 ++AI; 1222 } 1223 1224 assert(FI.arg_size() == Args.size() && 1225 "Mismatch between function signature & arguments."); 1226 unsigned ArgNo = 1; 1227 CGFunctionInfo::const_arg_iterator info_it = FI.arg_begin(); 1228 for (FunctionArgList::const_iterator i = Args.begin(), e = Args.end(); 1229 i != e; ++i, ++info_it, ++ArgNo) { 1230 const VarDecl *Arg = *i; 1231 QualType Ty = info_it->type; 1232 const ABIArgInfo &ArgI = info_it->info; 1233 1234 bool isPromoted = 1235 isa<ParmVarDecl>(Arg) && cast<ParmVarDecl>(Arg)->isKNRPromoted(); 1236 1237 // Skip the dummy padding argument. 1238 if (ArgI.getPaddingType()) 1239 ++AI; 1240 1241 switch (ArgI.getKind()) { 1242 case ABIArgInfo::Indirect: { 1243 llvm::Value *V = AI; 1244 1245 if (!hasScalarEvaluationKind(Ty)) { 1246 // Aggregates and complex variables are accessed by reference. All we 1247 // need to do is realign the value, if requested 1248 if (ArgI.getIndirectRealign()) { 1249 llvm::Value *AlignedTemp = CreateMemTemp(Ty, "coerce"); 1250 1251 // Copy from the incoming argument pointer to the temporary with the 1252 // appropriate alignment. 1253 // 1254 // FIXME: We should have a common utility for generating an aggregate 1255 // copy. 1256 llvm::Type *I8PtrTy = Builder.getInt8PtrTy(); 1257 CharUnits Size = getContext().getTypeSizeInChars(Ty); 1258 llvm::Value *Dst = Builder.CreateBitCast(AlignedTemp, I8PtrTy); 1259 llvm::Value *Src = Builder.CreateBitCast(V, I8PtrTy); 1260 Builder.CreateMemCpy(Dst, 1261 Src, 1262 llvm::ConstantInt::get(IntPtrTy, 1263 Size.getQuantity()), 1264 ArgI.getIndirectAlign(), 1265 false); 1266 V = AlignedTemp; 1267 } 1268 } else { 1269 // Load scalar value from indirect argument. 1270 CharUnits Alignment = getContext().getTypeAlignInChars(Ty); 1271 V = EmitLoadOfScalar(V, false, Alignment.getQuantity(), Ty); 1272 1273 if (isPromoted) 1274 V = emitArgumentDemotion(*this, Arg, V); 1275 } 1276 EmitParmDecl(*Arg, V, ArgNo); 1277 break; 1278 } 1279 1280 case ABIArgInfo::Extend: 1281 case ABIArgInfo::Direct: { 1282 1283 // If we have the trivial case, handle it with no muss and fuss. 1284 if (!isa<llvm::StructType>(ArgI.getCoerceToType()) && 1285 ArgI.getCoerceToType() == ConvertType(Ty) && 1286 ArgI.getDirectOffset() == 0) { 1287 assert(AI != Fn->arg_end() && "Argument mismatch!"); 1288 llvm::Value *V = AI; 1289 1290 if (Arg->getType().isRestrictQualified()) 1291 AI->addAttr(llvm::AttributeSet::get(getLLVMContext(), 1292 AI->getArgNo() + 1, 1293 llvm::Attribute::NoAlias)); 1294 1295 // Ensure the argument is the correct type. 1296 if (V->getType() != ArgI.getCoerceToType()) 1297 V = Builder.CreateBitCast(V, ArgI.getCoerceToType()); 1298 1299 if (isPromoted) 1300 V = emitArgumentDemotion(*this, Arg, V); 1301 1302 // Because of merging of function types from multiple decls it is 1303 // possible for the type of an argument to not match the corresponding 1304 // type in the function type. Since we are codegening the callee 1305 // in here, add a cast to the argument type. 1306 llvm::Type *LTy = ConvertType(Arg->getType()); 1307 if (V->getType() != LTy) 1308 V = Builder.CreateBitCast(V, LTy); 1309 1310 EmitParmDecl(*Arg, V, ArgNo); 1311 break; 1312 } 1313 1314 llvm::AllocaInst *Alloca = CreateMemTemp(Ty, Arg->getName()); 1315 1316 // The alignment we need to use is the max of the requested alignment for 1317 // the argument plus the alignment required by our access code below. 1318 unsigned AlignmentToUse = 1319 CGM.getDataLayout().getABITypeAlignment(ArgI.getCoerceToType()); 1320 AlignmentToUse = std::max(AlignmentToUse, 1321 (unsigned)getContext().getDeclAlign(Arg).getQuantity()); 1322 1323 Alloca->setAlignment(AlignmentToUse); 1324 llvm::Value *V = Alloca; 1325 llvm::Value *Ptr = V; // Pointer to store into. 1326 1327 // If the value is offset in memory, apply the offset now. 1328 if (unsigned Offs = ArgI.getDirectOffset()) { 1329 Ptr = Builder.CreateBitCast(Ptr, Builder.getInt8PtrTy()); 1330 Ptr = Builder.CreateConstGEP1_32(Ptr, Offs); 1331 Ptr = Builder.CreateBitCast(Ptr, 1332 llvm::PointerType::getUnqual(ArgI.getCoerceToType())); 1333 } 1334 1335 // If the coerce-to type is a first class aggregate, we flatten it and 1336 // pass the elements. Either way is semantically identical, but fast-isel 1337 // and the optimizer generally likes scalar values better than FCAs. 1338 llvm::StructType *STy = dyn_cast<llvm::StructType>(ArgI.getCoerceToType()); 1339 if (STy && STy->getNumElements() > 1) { 1340 uint64_t SrcSize = CGM.getDataLayout().getTypeAllocSize(STy); 1341 llvm::Type *DstTy = 1342 cast<llvm::PointerType>(Ptr->getType())->getElementType(); 1343 uint64_t DstSize = CGM.getDataLayout().getTypeAllocSize(DstTy); 1344 1345 if (SrcSize <= DstSize) { 1346 Ptr = Builder.CreateBitCast(Ptr, llvm::PointerType::getUnqual(STy)); 1347 1348 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 1349 assert(AI != Fn->arg_end() && "Argument mismatch!"); 1350 AI->setName(Arg->getName() + ".coerce" + Twine(i)); 1351 llvm::Value *EltPtr = Builder.CreateConstGEP2_32(Ptr, 0, i); 1352 Builder.CreateStore(AI++, EltPtr); 1353 } 1354 } else { 1355 llvm::AllocaInst *TempAlloca = 1356 CreateTempAlloca(ArgI.getCoerceToType(), "coerce"); 1357 TempAlloca->setAlignment(AlignmentToUse); 1358 llvm::Value *TempV = TempAlloca; 1359 1360 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 1361 assert(AI != Fn->arg_end() && "Argument mismatch!"); 1362 AI->setName(Arg->getName() + ".coerce" + Twine(i)); 1363 llvm::Value *EltPtr = Builder.CreateConstGEP2_32(TempV, 0, i); 1364 Builder.CreateStore(AI++, EltPtr); 1365 } 1366 1367 Builder.CreateMemCpy(Ptr, TempV, DstSize, AlignmentToUse); 1368 } 1369 } else { 1370 // Simple case, just do a coerced store of the argument into the alloca. 1371 assert(AI != Fn->arg_end() && "Argument mismatch!"); 1372 AI->setName(Arg->getName() + ".coerce"); 1373 CreateCoercedStore(AI++, Ptr, /*DestIsVolatile=*/false, *this); 1374 } 1375 1376 1377 // Match to what EmitParmDecl is expecting for this type. 1378 if (CodeGenFunction::hasScalarEvaluationKind(Ty)) { 1379 V = EmitLoadOfScalar(V, false, AlignmentToUse, Ty); 1380 if (isPromoted) 1381 V = emitArgumentDemotion(*this, Arg, V); 1382 } 1383 EmitParmDecl(*Arg, V, ArgNo); 1384 continue; // Skip ++AI increment, already done. 1385 } 1386 1387 case ABIArgInfo::Expand: { 1388 // If this structure was expanded into multiple arguments then 1389 // we need to create a temporary and reconstruct it from the 1390 // arguments. 1391 llvm::AllocaInst *Alloca = CreateMemTemp(Ty); 1392 CharUnits Align = getContext().getDeclAlign(Arg); 1393 Alloca->setAlignment(Align.getQuantity()); 1394 LValue LV = MakeAddrLValue(Alloca, Ty, Align); 1395 llvm::Function::arg_iterator End = ExpandTypeFromArgs(Ty, LV, AI); 1396 EmitParmDecl(*Arg, Alloca, ArgNo); 1397 1398 // Name the arguments used in expansion and increment AI. 1399 unsigned Index = 0; 1400 for (; AI != End; ++AI, ++Index) 1401 AI->setName(Arg->getName() + "." + Twine(Index)); 1402 continue; 1403 } 1404 1405 case ABIArgInfo::Ignore: 1406 // Initialize the local variable appropriately. 1407 if (!hasScalarEvaluationKind(Ty)) 1408 EmitParmDecl(*Arg, CreateMemTemp(Ty), ArgNo); 1409 else 1410 EmitParmDecl(*Arg, llvm::UndefValue::get(ConvertType(Arg->getType())), 1411 ArgNo); 1412 1413 // Skip increment, no matching LLVM parameter. 1414 continue; 1415 } 1416 1417 ++AI; 1418 } 1419 assert(AI == Fn->arg_end() && "Argument mismatch!"); 1420 } 1421 1422 static void eraseUnusedBitCasts(llvm::Instruction *insn) { 1423 while (insn->use_empty()) { 1424 llvm::BitCastInst *bitcast = dyn_cast<llvm::BitCastInst>(insn); 1425 if (!bitcast) return; 1426 1427 // This is "safe" because we would have used a ConstantExpr otherwise. 1428 insn = cast<llvm::Instruction>(bitcast->getOperand(0)); 1429 bitcast->eraseFromParent(); 1430 } 1431 } 1432 1433 /// Try to emit a fused autorelease of a return result. 1434 static llvm::Value *tryEmitFusedAutoreleaseOfResult(CodeGenFunction &CGF, 1435 llvm::Value *result) { 1436 // We must be immediately followed the cast. 1437 llvm::BasicBlock *BB = CGF.Builder.GetInsertBlock(); 1438 if (BB->empty()) return 0; 1439 if (&BB->back() != result) return 0; 1440 1441 llvm::Type *resultType = result->getType(); 1442 1443 // result is in a BasicBlock and is therefore an Instruction. 1444 llvm::Instruction *generator = cast<llvm::Instruction>(result); 1445 1446 SmallVector<llvm::Instruction*,4> insnsToKill; 1447 1448 // Look for: 1449 // %generator = bitcast %type1* %generator2 to %type2* 1450 while (llvm::BitCastInst *bitcast = dyn_cast<llvm::BitCastInst>(generator)) { 1451 // We would have emitted this as a constant if the operand weren't 1452 // an Instruction. 1453 generator = cast<llvm::Instruction>(bitcast->getOperand(0)); 1454 1455 // Require the generator to be immediately followed by the cast. 1456 if (generator->getNextNode() != bitcast) 1457 return 0; 1458 1459 insnsToKill.push_back(bitcast); 1460 } 1461 1462 // Look for: 1463 // %generator = call i8* @objc_retain(i8* %originalResult) 1464 // or 1465 // %generator = call i8* @objc_retainAutoreleasedReturnValue(i8* %originalResult) 1466 llvm::CallInst *call = dyn_cast<llvm::CallInst>(generator); 1467 if (!call) return 0; 1468 1469 bool doRetainAutorelease; 1470 1471 if (call->getCalledValue() == CGF.CGM.getARCEntrypoints().objc_retain) { 1472 doRetainAutorelease = true; 1473 } else if (call->getCalledValue() == CGF.CGM.getARCEntrypoints() 1474 .objc_retainAutoreleasedReturnValue) { 1475 doRetainAutorelease = false; 1476 1477 // If we emitted an assembly marker for this call (and the 1478 // ARCEntrypoints field should have been set if so), go looking 1479 // for that call. If we can't find it, we can't do this 1480 // optimization. But it should always be the immediately previous 1481 // instruction, unless we needed bitcasts around the call. 1482 if (CGF.CGM.getARCEntrypoints().retainAutoreleasedReturnValueMarker) { 1483 llvm::Instruction *prev = call->getPrevNode(); 1484 assert(prev); 1485 if (isa<llvm::BitCastInst>(prev)) { 1486 prev = prev->getPrevNode(); 1487 assert(prev); 1488 } 1489 assert(isa<llvm::CallInst>(prev)); 1490 assert(cast<llvm::CallInst>(prev)->getCalledValue() == 1491 CGF.CGM.getARCEntrypoints().retainAutoreleasedReturnValueMarker); 1492 insnsToKill.push_back(prev); 1493 } 1494 } else { 1495 return 0; 1496 } 1497 1498 result = call->getArgOperand(0); 1499 insnsToKill.push_back(call); 1500 1501 // Keep killing bitcasts, for sanity. Note that we no longer care 1502 // about precise ordering as long as there's exactly one use. 1503 while (llvm::BitCastInst *bitcast = dyn_cast<llvm::BitCastInst>(result)) { 1504 if (!bitcast->hasOneUse()) break; 1505 insnsToKill.push_back(bitcast); 1506 result = bitcast->getOperand(0); 1507 } 1508 1509 // Delete all the unnecessary instructions, from latest to earliest. 1510 for (SmallVectorImpl<llvm::Instruction*>::iterator 1511 i = insnsToKill.begin(), e = insnsToKill.end(); i != e; ++i) 1512 (*i)->eraseFromParent(); 1513 1514 // Do the fused retain/autorelease if we were asked to. 1515 if (doRetainAutorelease) 1516 result = CGF.EmitARCRetainAutoreleaseReturnValue(result); 1517 1518 // Cast back to the result type. 1519 return CGF.Builder.CreateBitCast(result, resultType); 1520 } 1521 1522 /// If this is a +1 of the value of an immutable 'self', remove it. 1523 static llvm::Value *tryRemoveRetainOfSelf(CodeGenFunction &CGF, 1524 llvm::Value *result) { 1525 // This is only applicable to a method with an immutable 'self'. 1526 const ObjCMethodDecl *method = 1527 dyn_cast_or_null<ObjCMethodDecl>(CGF.CurCodeDecl); 1528 if (!method) return 0; 1529 const VarDecl *self = method->getSelfDecl(); 1530 if (!self->getType().isConstQualified()) return 0; 1531 1532 // Look for a retain call. 1533 llvm::CallInst *retainCall = 1534 dyn_cast<llvm::CallInst>(result->stripPointerCasts()); 1535 if (!retainCall || 1536 retainCall->getCalledValue() != CGF.CGM.getARCEntrypoints().objc_retain) 1537 return 0; 1538 1539 // Look for an ordinary load of 'self'. 1540 llvm::Value *retainedValue = retainCall->getArgOperand(0); 1541 llvm::LoadInst *load = 1542 dyn_cast<llvm::LoadInst>(retainedValue->stripPointerCasts()); 1543 if (!load || load->isAtomic() || load->isVolatile() || 1544 load->getPointerOperand() != CGF.GetAddrOfLocalVar(self)) 1545 return 0; 1546 1547 // Okay! Burn it all down. This relies for correctness on the 1548 // assumption that the retain is emitted as part of the return and 1549 // that thereafter everything is used "linearly". 1550 llvm::Type *resultType = result->getType(); 1551 eraseUnusedBitCasts(cast<llvm::Instruction>(result)); 1552 assert(retainCall->use_empty()); 1553 retainCall->eraseFromParent(); 1554 eraseUnusedBitCasts(cast<llvm::Instruction>(retainedValue)); 1555 1556 return CGF.Builder.CreateBitCast(load, resultType); 1557 } 1558 1559 /// Emit an ARC autorelease of the result of a function. 1560 /// 1561 /// \return the value to actually return from the function 1562 static llvm::Value *emitAutoreleaseOfResult(CodeGenFunction &CGF, 1563 llvm::Value *result) { 1564 // If we're returning 'self', kill the initial retain. This is a 1565 // heuristic attempt to "encourage correctness" in the really unfortunate 1566 // case where we have a return of self during a dealloc and we desperately 1567 // need to avoid the possible autorelease. 1568 if (llvm::Value *self = tryRemoveRetainOfSelf(CGF, result)) 1569 return self; 1570 1571 // At -O0, try to emit a fused retain/autorelease. 1572 if (CGF.shouldUseFusedARCCalls()) 1573 if (llvm::Value *fused = tryEmitFusedAutoreleaseOfResult(CGF, result)) 1574 return fused; 1575 1576 return CGF.EmitARCAutoreleaseReturnValue(result); 1577 } 1578 1579 /// Heuristically search for a dominating store to the return-value slot. 1580 static llvm::StoreInst *findDominatingStoreToReturnValue(CodeGenFunction &CGF) { 1581 // If there are multiple uses of the return-value slot, just check 1582 // for something immediately preceding the IP. Sometimes this can 1583 // happen with how we generate implicit-returns; it can also happen 1584 // with noreturn cleanups. 1585 if (!CGF.ReturnValue->hasOneUse()) { 1586 llvm::BasicBlock *IP = CGF.Builder.GetInsertBlock(); 1587 if (IP->empty()) return 0; 1588 llvm::StoreInst *store = dyn_cast<llvm::StoreInst>(&IP->back()); 1589 if (!store) return 0; 1590 if (store->getPointerOperand() != CGF.ReturnValue) return 0; 1591 assert(!store->isAtomic() && !store->isVolatile()); // see below 1592 return store; 1593 } 1594 1595 llvm::StoreInst *store = 1596 dyn_cast<llvm::StoreInst>(CGF.ReturnValue->use_back()); 1597 if (!store) return 0; 1598 1599 // These aren't actually possible for non-coerced returns, and we 1600 // only care about non-coerced returns on this code path. 1601 assert(!store->isAtomic() && !store->isVolatile()); 1602 1603 // Now do a first-and-dirty dominance check: just walk up the 1604 // single-predecessors chain from the current insertion point. 1605 llvm::BasicBlock *StoreBB = store->getParent(); 1606 llvm::BasicBlock *IP = CGF.Builder.GetInsertBlock(); 1607 while (IP != StoreBB) { 1608 if (!(IP = IP->getSinglePredecessor())) 1609 return 0; 1610 } 1611 1612 // Okay, the store's basic block dominates the insertion point; we 1613 // can do our thing. 1614 return store; 1615 } 1616 1617 void CodeGenFunction::EmitFunctionEpilog(const CGFunctionInfo &FI) { 1618 // Functions with no result always return void. 1619 if (ReturnValue == 0) { 1620 Builder.CreateRetVoid(); 1621 return; 1622 } 1623 1624 llvm::DebugLoc RetDbgLoc; 1625 llvm::Value *RV = 0; 1626 QualType RetTy = FI.getReturnType(); 1627 const ABIArgInfo &RetAI = FI.getReturnInfo(); 1628 1629 switch (RetAI.getKind()) { 1630 case ABIArgInfo::Indirect: { 1631 switch (getEvaluationKind(RetTy)) { 1632 case TEK_Complex: { 1633 ComplexPairTy RT = 1634 EmitLoadOfComplex(MakeNaturalAlignAddrLValue(ReturnValue, RetTy)); 1635 EmitStoreOfComplex(RT, 1636 MakeNaturalAlignAddrLValue(CurFn->arg_begin(), RetTy), 1637 /*isInit*/ true); 1638 break; 1639 } 1640 case TEK_Aggregate: 1641 // Do nothing; aggregrates get evaluated directly into the destination. 1642 break; 1643 case TEK_Scalar: 1644 EmitStoreOfScalar(Builder.CreateLoad(ReturnValue), 1645 MakeNaturalAlignAddrLValue(CurFn->arg_begin(), RetTy), 1646 /*isInit*/ true); 1647 break; 1648 } 1649 break; 1650 } 1651 1652 case ABIArgInfo::Extend: 1653 case ABIArgInfo::Direct: 1654 if (RetAI.getCoerceToType() == ConvertType(RetTy) && 1655 RetAI.getDirectOffset() == 0) { 1656 // The internal return value temp always will have pointer-to-return-type 1657 // type, just do a load. 1658 1659 // If there is a dominating store to ReturnValue, we can elide 1660 // the load, zap the store, and usually zap the alloca. 1661 if (llvm::StoreInst *SI = findDominatingStoreToReturnValue(*this)) { 1662 // Get the stored value and nuke the now-dead store. 1663 RetDbgLoc = SI->getDebugLoc(); 1664 RV = SI->getValueOperand(); 1665 SI->eraseFromParent(); 1666 1667 // If that was the only use of the return value, nuke it as well now. 1668 if (ReturnValue->use_empty() && isa<llvm::AllocaInst>(ReturnValue)) { 1669 cast<llvm::AllocaInst>(ReturnValue)->eraseFromParent(); 1670 ReturnValue = 0; 1671 } 1672 1673 // Otherwise, we have to do a simple load. 1674 } else { 1675 RV = Builder.CreateLoad(ReturnValue); 1676 } 1677 } else { 1678 llvm::Value *V = ReturnValue; 1679 // If the value is offset in memory, apply the offset now. 1680 if (unsigned Offs = RetAI.getDirectOffset()) { 1681 V = Builder.CreateBitCast(V, Builder.getInt8PtrTy()); 1682 V = Builder.CreateConstGEP1_32(V, Offs); 1683 V = Builder.CreateBitCast(V, 1684 llvm::PointerType::getUnqual(RetAI.getCoerceToType())); 1685 } 1686 1687 RV = CreateCoercedLoad(V, RetAI.getCoerceToType(), *this); 1688 } 1689 1690 // In ARC, end functions that return a retainable type with a call 1691 // to objc_autoreleaseReturnValue. 1692 if (AutoreleaseResult) { 1693 assert(getLangOpts().ObjCAutoRefCount && 1694 !FI.isReturnsRetained() && 1695 RetTy->isObjCRetainableType()); 1696 RV = emitAutoreleaseOfResult(*this, RV); 1697 } 1698 1699 break; 1700 1701 case ABIArgInfo::Ignore: 1702 break; 1703 1704 case ABIArgInfo::Expand: 1705 llvm_unreachable("Invalid ABI kind for return argument"); 1706 } 1707 1708 llvm::Instruction *Ret = RV ? Builder.CreateRet(RV) : Builder.CreateRetVoid(); 1709 if (!RetDbgLoc.isUnknown()) 1710 Ret->setDebugLoc(RetDbgLoc); 1711 } 1712 1713 void CodeGenFunction::EmitDelegateCallArg(CallArgList &args, 1714 const VarDecl *param) { 1715 // StartFunction converted the ABI-lowered parameter(s) into a 1716 // local alloca. We need to turn that into an r-value suitable 1717 // for EmitCall. 1718 llvm::Value *local = GetAddrOfLocalVar(param); 1719 1720 QualType type = param->getType(); 1721 1722 // For the most part, we just need to load the alloca, except: 1723 // 1) aggregate r-values are actually pointers to temporaries, and 1724 // 2) references to non-scalars are pointers directly to the aggregate. 1725 // I don't know why references to scalars are different here. 1726 if (const ReferenceType *ref = type->getAs<ReferenceType>()) { 1727 if (!hasScalarEvaluationKind(ref->getPointeeType())) 1728 return args.add(RValue::getAggregate(local), type); 1729 1730 // Locals which are references to scalars are represented 1731 // with allocas holding the pointer. 1732 return args.add(RValue::get(Builder.CreateLoad(local)), type); 1733 } 1734 1735 args.add(convertTempToRValue(local, type), type); 1736 } 1737 1738 static bool isProvablyNull(llvm::Value *addr) { 1739 return isa<llvm::ConstantPointerNull>(addr); 1740 } 1741 1742 static bool isProvablyNonNull(llvm::Value *addr) { 1743 return isa<llvm::AllocaInst>(addr); 1744 } 1745 1746 /// Emit the actual writing-back of a writeback. 1747 static void emitWriteback(CodeGenFunction &CGF, 1748 const CallArgList::Writeback &writeback) { 1749 llvm::Value *srcAddr = writeback.Address; 1750 assert(!isProvablyNull(srcAddr) && 1751 "shouldn't have writeback for provably null argument"); 1752 1753 llvm::BasicBlock *contBB = 0; 1754 1755 // If the argument wasn't provably non-null, we need to null check 1756 // before doing the store. 1757 bool provablyNonNull = isProvablyNonNull(srcAddr); 1758 if (!provablyNonNull) { 1759 llvm::BasicBlock *writebackBB = CGF.createBasicBlock("icr.writeback"); 1760 contBB = CGF.createBasicBlock("icr.done"); 1761 1762 llvm::Value *isNull = CGF.Builder.CreateIsNull(srcAddr, "icr.isnull"); 1763 CGF.Builder.CreateCondBr(isNull, contBB, writebackBB); 1764 CGF.EmitBlock(writebackBB); 1765 } 1766 1767 // Load the value to writeback. 1768 llvm::Value *value = CGF.Builder.CreateLoad(writeback.Temporary); 1769 1770 // Cast it back, in case we're writing an id to a Foo* or something. 1771 value = CGF.Builder.CreateBitCast(value, 1772 cast<llvm::PointerType>(srcAddr->getType())->getElementType(), 1773 "icr.writeback-cast"); 1774 1775 // Perform the writeback. 1776 QualType srcAddrType = writeback.AddressType; 1777 CGF.EmitStoreThroughLValue(RValue::get(value), 1778 CGF.MakeAddrLValue(srcAddr, srcAddrType)); 1779 1780 // Jump to the continuation block. 1781 if (!provablyNonNull) 1782 CGF.EmitBlock(contBB); 1783 } 1784 1785 static void emitWritebacks(CodeGenFunction &CGF, 1786 const CallArgList &args) { 1787 for (CallArgList::writeback_iterator 1788 i = args.writeback_begin(), e = args.writeback_end(); i != e; ++i) 1789 emitWriteback(CGF, *i); 1790 } 1791 1792 /// Emit an argument that's being passed call-by-writeback. That is, 1793 /// we are passing the address of 1794 static void emitWritebackArg(CodeGenFunction &CGF, CallArgList &args, 1795 const ObjCIndirectCopyRestoreExpr *CRE) { 1796 llvm::Value *srcAddr = CGF.EmitScalarExpr(CRE->getSubExpr()); 1797 1798 // The dest and src types don't necessarily match in LLVM terms 1799 // because of the crazy ObjC compatibility rules. 1800 1801 llvm::PointerType *destType = 1802 cast<llvm::PointerType>(CGF.ConvertType(CRE->getType())); 1803 1804 // If the address is a constant null, just pass the appropriate null. 1805 if (isProvablyNull(srcAddr)) { 1806 args.add(RValue::get(llvm::ConstantPointerNull::get(destType)), 1807 CRE->getType()); 1808 return; 1809 } 1810 1811 QualType srcAddrType = 1812 CRE->getSubExpr()->getType()->castAs<PointerType>()->getPointeeType(); 1813 1814 // Create the temporary. 1815 llvm::Value *temp = CGF.CreateTempAlloca(destType->getElementType(), 1816 "icr.temp"); 1817 // Loading an l-value can introduce a cleanup if the l-value is __weak, 1818 // and that cleanup will be conditional if we can't prove that the l-value 1819 // isn't null, so we need to register a dominating point so that the cleanups 1820 // system will make valid IR. 1821 CodeGenFunction::ConditionalEvaluation condEval(CGF); 1822 1823 // Zero-initialize it if we're not doing a copy-initialization. 1824 bool shouldCopy = CRE->shouldCopy(); 1825 if (!shouldCopy) { 1826 llvm::Value *null = 1827 llvm::ConstantPointerNull::get( 1828 cast<llvm::PointerType>(destType->getElementType())); 1829 CGF.Builder.CreateStore(null, temp); 1830 } 1831 1832 llvm::BasicBlock *contBB = 0; 1833 1834 // If the address is *not* known to be non-null, we need to switch. 1835 llvm::Value *finalArgument; 1836 1837 bool provablyNonNull = isProvablyNonNull(srcAddr); 1838 if (provablyNonNull) { 1839 finalArgument = temp; 1840 } else { 1841 llvm::Value *isNull = CGF.Builder.CreateIsNull(srcAddr, "icr.isnull"); 1842 1843 finalArgument = CGF.Builder.CreateSelect(isNull, 1844 llvm::ConstantPointerNull::get(destType), 1845 temp, "icr.argument"); 1846 1847 // If we need to copy, then the load has to be conditional, which 1848 // means we need control flow. 1849 if (shouldCopy) { 1850 contBB = CGF.createBasicBlock("icr.cont"); 1851 llvm::BasicBlock *copyBB = CGF.createBasicBlock("icr.copy"); 1852 CGF.Builder.CreateCondBr(isNull, contBB, copyBB); 1853 CGF.EmitBlock(copyBB); 1854 condEval.begin(CGF); 1855 } 1856 } 1857 1858 // Perform a copy if necessary. 1859 if (shouldCopy) { 1860 LValue srcLV = CGF.MakeAddrLValue(srcAddr, srcAddrType); 1861 RValue srcRV = CGF.EmitLoadOfLValue(srcLV); 1862 assert(srcRV.isScalar()); 1863 1864 llvm::Value *src = srcRV.getScalarVal(); 1865 src = CGF.Builder.CreateBitCast(src, destType->getElementType(), 1866 "icr.cast"); 1867 1868 // Use an ordinary store, not a store-to-lvalue. 1869 CGF.Builder.CreateStore(src, temp); 1870 } 1871 1872 // Finish the control flow if we needed it. 1873 if (shouldCopy && !provablyNonNull) { 1874 CGF.EmitBlock(contBB); 1875 condEval.end(CGF); 1876 } 1877 1878 args.addWriteback(srcAddr, srcAddrType, temp); 1879 args.add(RValue::get(finalArgument), CRE->getType()); 1880 } 1881 1882 void CodeGenFunction::EmitCallArg(CallArgList &args, const Expr *E, 1883 QualType type) { 1884 if (const ObjCIndirectCopyRestoreExpr *CRE 1885 = dyn_cast<ObjCIndirectCopyRestoreExpr>(E)) { 1886 assert(getLangOpts().ObjCAutoRefCount); 1887 assert(getContext().hasSameType(E->getType(), type)); 1888 return emitWritebackArg(*this, args, CRE); 1889 } 1890 1891 assert(type->isReferenceType() == E->isGLValue() && 1892 "reference binding to unmaterialized r-value!"); 1893 1894 if (E->isGLValue()) { 1895 assert(E->getObjectKind() == OK_Ordinary); 1896 return args.add(EmitReferenceBindingToExpr(E, /*InitializedDecl=*/0), 1897 type); 1898 } 1899 1900 if (hasAggregateEvaluationKind(type) && 1901 isa<ImplicitCastExpr>(E) && 1902 cast<CastExpr>(E)->getCastKind() == CK_LValueToRValue) { 1903 LValue L = EmitLValue(cast<CastExpr>(E)->getSubExpr()); 1904 assert(L.isSimple()); 1905 args.add(L.asAggregateRValue(), type, /*NeedsCopy*/true); 1906 return; 1907 } 1908 1909 args.add(EmitAnyExprToTemp(E), type); 1910 } 1911 1912 // In ObjC ARC mode with no ObjC ARC exception safety, tell the ARC 1913 // optimizer it can aggressively ignore unwind edges. 1914 void 1915 CodeGenFunction::AddObjCARCExceptionMetadata(llvm::Instruction *Inst) { 1916 if (CGM.getCodeGenOpts().OptimizationLevel != 0 && 1917 !CGM.getCodeGenOpts().ObjCAutoRefCountExceptions) 1918 Inst->setMetadata("clang.arc.no_objc_arc_exceptions", 1919 CGM.getNoObjCARCExceptionsMetadata()); 1920 } 1921 1922 /// Emits a call to the given no-arguments nounwind runtime function. 1923 llvm::CallInst * 1924 CodeGenFunction::EmitNounwindRuntimeCall(llvm::Value *callee, 1925 const llvm::Twine &name) { 1926 return EmitNounwindRuntimeCall(callee, ArrayRef<llvm::Value*>(), name); 1927 } 1928 1929 /// Emits a call to the given nounwind runtime function. 1930 llvm::CallInst * 1931 CodeGenFunction::EmitNounwindRuntimeCall(llvm::Value *callee, 1932 ArrayRef<llvm::Value*> args, 1933 const llvm::Twine &name) { 1934 llvm::CallInst *call = EmitRuntimeCall(callee, args, name); 1935 call->setDoesNotThrow(); 1936 return call; 1937 } 1938 1939 /// Emits a simple call (never an invoke) to the given no-arguments 1940 /// runtime function. 1941 llvm::CallInst * 1942 CodeGenFunction::EmitRuntimeCall(llvm::Value *callee, 1943 const llvm::Twine &name) { 1944 return EmitRuntimeCall(callee, ArrayRef<llvm::Value*>(), name); 1945 } 1946 1947 /// Emits a simple call (never an invoke) to the given runtime 1948 /// function. 1949 llvm::CallInst * 1950 CodeGenFunction::EmitRuntimeCall(llvm::Value *callee, 1951 ArrayRef<llvm::Value*> args, 1952 const llvm::Twine &name) { 1953 llvm::CallInst *call = Builder.CreateCall(callee, args, name); 1954 call->setCallingConv(getRuntimeCC()); 1955 return call; 1956 } 1957 1958 /// Emits a call or invoke to the given noreturn runtime function. 1959 void CodeGenFunction::EmitNoreturnRuntimeCallOrInvoke(llvm::Value *callee, 1960 ArrayRef<llvm::Value*> args) { 1961 if (getInvokeDest()) { 1962 llvm::InvokeInst *invoke = 1963 Builder.CreateInvoke(callee, 1964 getUnreachableBlock(), 1965 getInvokeDest(), 1966 args); 1967 invoke->setDoesNotReturn(); 1968 invoke->setCallingConv(getRuntimeCC()); 1969 } else { 1970 llvm::CallInst *call = Builder.CreateCall(callee, args); 1971 call->setDoesNotReturn(); 1972 call->setCallingConv(getRuntimeCC()); 1973 Builder.CreateUnreachable(); 1974 } 1975 } 1976 1977 /// Emits a call or invoke instruction to the given nullary runtime 1978 /// function. 1979 llvm::CallSite 1980 CodeGenFunction::EmitRuntimeCallOrInvoke(llvm::Value *callee, 1981 const Twine &name) { 1982 return EmitRuntimeCallOrInvoke(callee, ArrayRef<llvm::Value*>(), name); 1983 } 1984 1985 /// Emits a call or invoke instruction to the given runtime function. 1986 llvm::CallSite 1987 CodeGenFunction::EmitRuntimeCallOrInvoke(llvm::Value *callee, 1988 ArrayRef<llvm::Value*> args, 1989 const Twine &name) { 1990 llvm::CallSite callSite = EmitCallOrInvoke(callee, args, name); 1991 callSite.setCallingConv(getRuntimeCC()); 1992 return callSite; 1993 } 1994 1995 llvm::CallSite 1996 CodeGenFunction::EmitCallOrInvoke(llvm::Value *Callee, 1997 const Twine &Name) { 1998 return EmitCallOrInvoke(Callee, ArrayRef<llvm::Value *>(), Name); 1999 } 2000 2001 /// Emits a call or invoke instruction to the given function, depending 2002 /// on the current state of the EH stack. 2003 llvm::CallSite 2004 CodeGenFunction::EmitCallOrInvoke(llvm::Value *Callee, 2005 ArrayRef<llvm::Value *> Args, 2006 const Twine &Name) { 2007 llvm::BasicBlock *InvokeDest = getInvokeDest(); 2008 2009 llvm::Instruction *Inst; 2010 if (!InvokeDest) 2011 Inst = Builder.CreateCall(Callee, Args, Name); 2012 else { 2013 llvm::BasicBlock *ContBB = createBasicBlock("invoke.cont"); 2014 Inst = Builder.CreateInvoke(Callee, ContBB, InvokeDest, Args, Name); 2015 EmitBlock(ContBB); 2016 } 2017 2018 // In ObjC ARC mode with no ObjC ARC exception safety, tell the ARC 2019 // optimizer it can aggressively ignore unwind edges. 2020 if (CGM.getLangOpts().ObjCAutoRefCount) 2021 AddObjCARCExceptionMetadata(Inst); 2022 2023 return Inst; 2024 } 2025 2026 static void checkArgMatches(llvm::Value *Elt, unsigned &ArgNo, 2027 llvm::FunctionType *FTy) { 2028 if (ArgNo < FTy->getNumParams()) 2029 assert(Elt->getType() == FTy->getParamType(ArgNo)); 2030 else 2031 assert(FTy->isVarArg()); 2032 ++ArgNo; 2033 } 2034 2035 void CodeGenFunction::ExpandTypeToArgs(QualType Ty, RValue RV, 2036 SmallVector<llvm::Value*,16> &Args, 2037 llvm::FunctionType *IRFuncTy) { 2038 if (const ConstantArrayType *AT = getContext().getAsConstantArrayType(Ty)) { 2039 unsigned NumElts = AT->getSize().getZExtValue(); 2040 QualType EltTy = AT->getElementType(); 2041 llvm::Value *Addr = RV.getAggregateAddr(); 2042 for (unsigned Elt = 0; Elt < NumElts; ++Elt) { 2043 llvm::Value *EltAddr = Builder.CreateConstGEP2_32(Addr, 0, Elt); 2044 RValue EltRV = convertTempToRValue(EltAddr, EltTy); 2045 ExpandTypeToArgs(EltTy, EltRV, Args, IRFuncTy); 2046 } 2047 } else if (const RecordType *RT = Ty->getAs<RecordType>()) { 2048 RecordDecl *RD = RT->getDecl(); 2049 assert(RV.isAggregate() && "Unexpected rvalue during struct expansion"); 2050 LValue LV = MakeAddrLValue(RV.getAggregateAddr(), Ty); 2051 2052 if (RD->isUnion()) { 2053 const FieldDecl *LargestFD = 0; 2054 CharUnits UnionSize = CharUnits::Zero(); 2055 2056 for (RecordDecl::field_iterator i = RD->field_begin(), e = RD->field_end(); 2057 i != e; ++i) { 2058 const FieldDecl *FD = *i; 2059 assert(!FD->isBitField() && 2060 "Cannot expand structure with bit-field members."); 2061 CharUnits FieldSize = getContext().getTypeSizeInChars(FD->getType()); 2062 if (UnionSize < FieldSize) { 2063 UnionSize = FieldSize; 2064 LargestFD = FD; 2065 } 2066 } 2067 if (LargestFD) { 2068 RValue FldRV = EmitRValueForField(LV, LargestFD); 2069 ExpandTypeToArgs(LargestFD->getType(), FldRV, Args, IRFuncTy); 2070 } 2071 } else { 2072 for (RecordDecl::field_iterator i = RD->field_begin(), e = RD->field_end(); 2073 i != e; ++i) { 2074 FieldDecl *FD = *i; 2075 2076 RValue FldRV = EmitRValueForField(LV, FD); 2077 ExpandTypeToArgs(FD->getType(), FldRV, Args, IRFuncTy); 2078 } 2079 } 2080 } else if (Ty->isAnyComplexType()) { 2081 ComplexPairTy CV = RV.getComplexVal(); 2082 Args.push_back(CV.first); 2083 Args.push_back(CV.second); 2084 } else { 2085 assert(RV.isScalar() && 2086 "Unexpected non-scalar rvalue during struct expansion."); 2087 2088 // Insert a bitcast as needed. 2089 llvm::Value *V = RV.getScalarVal(); 2090 if (Args.size() < IRFuncTy->getNumParams() && 2091 V->getType() != IRFuncTy->getParamType(Args.size())) 2092 V = Builder.CreateBitCast(V, IRFuncTy->getParamType(Args.size())); 2093 2094 Args.push_back(V); 2095 } 2096 } 2097 2098 2099 RValue CodeGenFunction::EmitCall(const CGFunctionInfo &CallInfo, 2100 llvm::Value *Callee, 2101 ReturnValueSlot ReturnValue, 2102 const CallArgList &CallArgs, 2103 const Decl *TargetDecl, 2104 llvm::Instruction **callOrInvoke) { 2105 // FIXME: We no longer need the types from CallArgs; lift up and simplify. 2106 SmallVector<llvm::Value*, 16> Args; 2107 2108 // Handle struct-return functions by passing a pointer to the 2109 // location that we would like to return into. 2110 QualType RetTy = CallInfo.getReturnType(); 2111 const ABIArgInfo &RetAI = CallInfo.getReturnInfo(); 2112 2113 // IRArgNo - Keep track of the argument number in the callee we're looking at. 2114 unsigned IRArgNo = 0; 2115 llvm::FunctionType *IRFuncTy = 2116 cast<llvm::FunctionType>( 2117 cast<llvm::PointerType>(Callee->getType())->getElementType()); 2118 2119 // If the call returns a temporary with struct return, create a temporary 2120 // alloca to hold the result, unless one is given to us. 2121 if (CGM.ReturnTypeUsesSRet(CallInfo)) { 2122 llvm::Value *Value = ReturnValue.getValue(); 2123 if (!Value) 2124 Value = CreateMemTemp(RetTy); 2125 Args.push_back(Value); 2126 checkArgMatches(Value, IRArgNo, IRFuncTy); 2127 } 2128 2129 assert(CallInfo.arg_size() == CallArgs.size() && 2130 "Mismatch between function signature & arguments."); 2131 CGFunctionInfo::const_arg_iterator info_it = CallInfo.arg_begin(); 2132 for (CallArgList::const_iterator I = CallArgs.begin(), E = CallArgs.end(); 2133 I != E; ++I, ++info_it) { 2134 const ABIArgInfo &ArgInfo = info_it->info; 2135 RValue RV = I->RV; 2136 2137 CharUnits TypeAlign = getContext().getTypeAlignInChars(I->Ty); 2138 2139 // Insert a padding argument to ensure proper alignment. 2140 if (llvm::Type *PaddingType = ArgInfo.getPaddingType()) { 2141 Args.push_back(llvm::UndefValue::get(PaddingType)); 2142 ++IRArgNo; 2143 } 2144 2145 switch (ArgInfo.getKind()) { 2146 case ABIArgInfo::Indirect: { 2147 if (RV.isScalar() || RV.isComplex()) { 2148 // Make a temporary alloca to pass the argument. 2149 llvm::AllocaInst *AI = CreateMemTemp(I->Ty); 2150 if (ArgInfo.getIndirectAlign() > AI->getAlignment()) 2151 AI->setAlignment(ArgInfo.getIndirectAlign()); 2152 Args.push_back(AI); 2153 2154 LValue argLV = 2155 MakeAddrLValue(Args.back(), I->Ty, TypeAlign); 2156 2157 if (RV.isScalar()) 2158 EmitStoreOfScalar(RV.getScalarVal(), argLV, /*init*/ true); 2159 else 2160 EmitStoreOfComplex(RV.getComplexVal(), argLV, /*init*/ true); 2161 2162 // Validate argument match. 2163 checkArgMatches(AI, IRArgNo, IRFuncTy); 2164 } else { 2165 // We want to avoid creating an unnecessary temporary+copy here; 2166 // however, we need one in three cases: 2167 // 1. If the argument is not byval, and we are required to copy the 2168 // source. (This case doesn't occur on any common architecture.) 2169 // 2. If the argument is byval, RV is not sufficiently aligned, and 2170 // we cannot force it to be sufficiently aligned. 2171 // 3. If the argument is byval, but RV is located in an address space 2172 // different than that of the argument (0). 2173 llvm::Value *Addr = RV.getAggregateAddr(); 2174 unsigned Align = ArgInfo.getIndirectAlign(); 2175 const llvm::DataLayout *TD = &CGM.getDataLayout(); 2176 const unsigned RVAddrSpace = Addr->getType()->getPointerAddressSpace(); 2177 const unsigned ArgAddrSpace = (IRArgNo < IRFuncTy->getNumParams() ? 2178 IRFuncTy->getParamType(IRArgNo)->getPointerAddressSpace() : 0); 2179 if ((!ArgInfo.getIndirectByVal() && I->NeedsCopy) || 2180 (ArgInfo.getIndirectByVal() && TypeAlign.getQuantity() < Align && 2181 llvm::getOrEnforceKnownAlignment(Addr, Align, TD) < Align) || 2182 (ArgInfo.getIndirectByVal() && (RVAddrSpace != ArgAddrSpace))) { 2183 // Create an aligned temporary, and copy to it. 2184 llvm::AllocaInst *AI = CreateMemTemp(I->Ty); 2185 if (Align > AI->getAlignment()) 2186 AI->setAlignment(Align); 2187 Args.push_back(AI); 2188 EmitAggregateCopy(AI, Addr, I->Ty, RV.isVolatileQualified()); 2189 2190 // Validate argument match. 2191 checkArgMatches(AI, IRArgNo, IRFuncTy); 2192 } else { 2193 // Skip the extra memcpy call. 2194 Args.push_back(Addr); 2195 2196 // Validate argument match. 2197 checkArgMatches(Addr, IRArgNo, IRFuncTy); 2198 } 2199 } 2200 break; 2201 } 2202 2203 case ABIArgInfo::Ignore: 2204 break; 2205 2206 case ABIArgInfo::Extend: 2207 case ABIArgInfo::Direct: { 2208 if (!isa<llvm::StructType>(ArgInfo.getCoerceToType()) && 2209 ArgInfo.getCoerceToType() == ConvertType(info_it->type) && 2210 ArgInfo.getDirectOffset() == 0) { 2211 llvm::Value *V; 2212 if (RV.isScalar()) 2213 V = RV.getScalarVal(); 2214 else 2215 V = Builder.CreateLoad(RV.getAggregateAddr()); 2216 2217 // If the argument doesn't match, perform a bitcast to coerce it. This 2218 // can happen due to trivial type mismatches. 2219 if (IRArgNo < IRFuncTy->getNumParams() && 2220 V->getType() != IRFuncTy->getParamType(IRArgNo)) 2221 V = Builder.CreateBitCast(V, IRFuncTy->getParamType(IRArgNo)); 2222 Args.push_back(V); 2223 2224 checkArgMatches(V, IRArgNo, IRFuncTy); 2225 break; 2226 } 2227 2228 // FIXME: Avoid the conversion through memory if possible. 2229 llvm::Value *SrcPtr; 2230 if (RV.isScalar() || RV.isComplex()) { 2231 SrcPtr = CreateMemTemp(I->Ty, "coerce"); 2232 LValue SrcLV = MakeAddrLValue(SrcPtr, I->Ty, TypeAlign); 2233 if (RV.isScalar()) { 2234 EmitStoreOfScalar(RV.getScalarVal(), SrcLV, /*init*/ true); 2235 } else { 2236 EmitStoreOfComplex(RV.getComplexVal(), SrcLV, /*init*/ true); 2237 } 2238 } else 2239 SrcPtr = RV.getAggregateAddr(); 2240 2241 // If the value is offset in memory, apply the offset now. 2242 if (unsigned Offs = ArgInfo.getDirectOffset()) { 2243 SrcPtr = Builder.CreateBitCast(SrcPtr, Builder.getInt8PtrTy()); 2244 SrcPtr = Builder.CreateConstGEP1_32(SrcPtr, Offs); 2245 SrcPtr = Builder.CreateBitCast(SrcPtr, 2246 llvm::PointerType::getUnqual(ArgInfo.getCoerceToType())); 2247 2248 } 2249 2250 // If the coerce-to type is a first class aggregate, we flatten it and 2251 // pass the elements. Either way is semantically identical, but fast-isel 2252 // and the optimizer generally likes scalar values better than FCAs. 2253 if (llvm::StructType *STy = 2254 dyn_cast<llvm::StructType>(ArgInfo.getCoerceToType())) { 2255 llvm::Type *SrcTy = 2256 cast<llvm::PointerType>(SrcPtr->getType())->getElementType(); 2257 uint64_t SrcSize = CGM.getDataLayout().getTypeAllocSize(SrcTy); 2258 uint64_t DstSize = CGM.getDataLayout().getTypeAllocSize(STy); 2259 2260 // If the source type is smaller than the destination type of the 2261 // coerce-to logic, copy the source value into a temp alloca the size 2262 // of the destination type to allow loading all of it. The bits past 2263 // the source value are left undef. 2264 if (SrcSize < DstSize) { 2265 llvm::AllocaInst *TempAlloca 2266 = CreateTempAlloca(STy, SrcPtr->getName() + ".coerce"); 2267 Builder.CreateMemCpy(TempAlloca, SrcPtr, SrcSize, 0); 2268 SrcPtr = TempAlloca; 2269 } else { 2270 SrcPtr = Builder.CreateBitCast(SrcPtr, 2271 llvm::PointerType::getUnqual(STy)); 2272 } 2273 2274 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 2275 llvm::Value *EltPtr = Builder.CreateConstGEP2_32(SrcPtr, 0, i); 2276 llvm::LoadInst *LI = Builder.CreateLoad(EltPtr); 2277 // We don't know what we're loading from. 2278 LI->setAlignment(1); 2279 Args.push_back(LI); 2280 2281 // Validate argument match. 2282 checkArgMatches(LI, IRArgNo, IRFuncTy); 2283 } 2284 } else { 2285 // In the simple case, just pass the coerced loaded value. 2286 Args.push_back(CreateCoercedLoad(SrcPtr, ArgInfo.getCoerceToType(), 2287 *this)); 2288 2289 // Validate argument match. 2290 checkArgMatches(Args.back(), IRArgNo, IRFuncTy); 2291 } 2292 2293 break; 2294 } 2295 2296 case ABIArgInfo::Expand: 2297 ExpandTypeToArgs(I->Ty, RV, Args, IRFuncTy); 2298 IRArgNo = Args.size(); 2299 break; 2300 } 2301 } 2302 2303 // If the callee is a bitcast of a function to a varargs pointer to function 2304 // type, check to see if we can remove the bitcast. This handles some cases 2305 // with unprototyped functions. 2306 if (llvm::ConstantExpr *CE = dyn_cast<llvm::ConstantExpr>(Callee)) 2307 if (llvm::Function *CalleeF = dyn_cast<llvm::Function>(CE->getOperand(0))) { 2308 llvm::PointerType *CurPT=cast<llvm::PointerType>(Callee->getType()); 2309 llvm::FunctionType *CurFT = 2310 cast<llvm::FunctionType>(CurPT->getElementType()); 2311 llvm::FunctionType *ActualFT = CalleeF->getFunctionType(); 2312 2313 if (CE->getOpcode() == llvm::Instruction::BitCast && 2314 ActualFT->getReturnType() == CurFT->getReturnType() && 2315 ActualFT->getNumParams() == CurFT->getNumParams() && 2316 ActualFT->getNumParams() == Args.size() && 2317 (CurFT->isVarArg() || !ActualFT->isVarArg())) { 2318 bool ArgsMatch = true; 2319 for (unsigned i = 0, e = ActualFT->getNumParams(); i != e; ++i) 2320 if (ActualFT->getParamType(i) != CurFT->getParamType(i)) { 2321 ArgsMatch = false; 2322 break; 2323 } 2324 2325 // Strip the cast if we can get away with it. This is a nice cleanup, 2326 // but also allows us to inline the function at -O0 if it is marked 2327 // always_inline. 2328 if (ArgsMatch) 2329 Callee = CalleeF; 2330 } 2331 } 2332 2333 unsigned CallingConv; 2334 CodeGen::AttributeListType AttributeList; 2335 CGM.ConstructAttributeList(CallInfo, TargetDecl, AttributeList, 2336 CallingConv, true); 2337 llvm::AttributeSet Attrs = llvm::AttributeSet::get(getLLVMContext(), 2338 AttributeList); 2339 2340 llvm::BasicBlock *InvokeDest = 0; 2341 if (!Attrs.hasAttribute(llvm::AttributeSet::FunctionIndex, 2342 llvm::Attribute::NoUnwind)) 2343 InvokeDest = getInvokeDest(); 2344 2345 llvm::CallSite CS; 2346 if (!InvokeDest) { 2347 CS = Builder.CreateCall(Callee, Args); 2348 } else { 2349 llvm::BasicBlock *Cont = createBasicBlock("invoke.cont"); 2350 CS = Builder.CreateInvoke(Callee, Cont, InvokeDest, Args); 2351 EmitBlock(Cont); 2352 } 2353 if (callOrInvoke) 2354 *callOrInvoke = CS.getInstruction(); 2355 2356 CS.setAttributes(Attrs); 2357 CS.setCallingConv(static_cast<llvm::CallingConv::ID>(CallingConv)); 2358 2359 // In ObjC ARC mode with no ObjC ARC exception safety, tell the ARC 2360 // optimizer it can aggressively ignore unwind edges. 2361 if (CGM.getLangOpts().ObjCAutoRefCount) 2362 AddObjCARCExceptionMetadata(CS.getInstruction()); 2363 2364 // If the call doesn't return, finish the basic block and clear the 2365 // insertion point; this allows the rest of IRgen to discard 2366 // unreachable code. 2367 if (CS.doesNotReturn()) { 2368 Builder.CreateUnreachable(); 2369 Builder.ClearInsertionPoint(); 2370 2371 // FIXME: For now, emit a dummy basic block because expr emitters in 2372 // generally are not ready to handle emitting expressions at unreachable 2373 // points. 2374 EnsureInsertPoint(); 2375 2376 // Return a reasonable RValue. 2377 return GetUndefRValue(RetTy); 2378 } 2379 2380 llvm::Instruction *CI = CS.getInstruction(); 2381 if (Builder.isNamePreserving() && !CI->getType()->isVoidTy()) 2382 CI->setName("call"); 2383 2384 // Emit any writebacks immediately. Arguably this should happen 2385 // after any return-value munging. 2386 if (CallArgs.hasWritebacks()) 2387 emitWritebacks(*this, CallArgs); 2388 2389 switch (RetAI.getKind()) { 2390 case ABIArgInfo::Indirect: 2391 return convertTempToRValue(Args[0], RetTy); 2392 2393 case ABIArgInfo::Ignore: 2394 // If we are ignoring an argument that had a result, make sure to 2395 // construct the appropriate return value for our caller. 2396 return GetUndefRValue(RetTy); 2397 2398 case ABIArgInfo::Extend: 2399 case ABIArgInfo::Direct: { 2400 llvm::Type *RetIRTy = ConvertType(RetTy); 2401 if (RetAI.getCoerceToType() == RetIRTy && RetAI.getDirectOffset() == 0) { 2402 switch (getEvaluationKind(RetTy)) { 2403 case TEK_Complex: { 2404 llvm::Value *Real = Builder.CreateExtractValue(CI, 0); 2405 llvm::Value *Imag = Builder.CreateExtractValue(CI, 1); 2406 return RValue::getComplex(std::make_pair(Real, Imag)); 2407 } 2408 case TEK_Aggregate: { 2409 llvm::Value *DestPtr = ReturnValue.getValue(); 2410 bool DestIsVolatile = ReturnValue.isVolatile(); 2411 2412 if (!DestPtr) { 2413 DestPtr = CreateMemTemp(RetTy, "agg.tmp"); 2414 DestIsVolatile = false; 2415 } 2416 BuildAggStore(*this, CI, DestPtr, DestIsVolatile, false); 2417 return RValue::getAggregate(DestPtr); 2418 } 2419 case TEK_Scalar: { 2420 // If the argument doesn't match, perform a bitcast to coerce it. This 2421 // can happen due to trivial type mismatches. 2422 llvm::Value *V = CI; 2423 if (V->getType() != RetIRTy) 2424 V = Builder.CreateBitCast(V, RetIRTy); 2425 return RValue::get(V); 2426 } 2427 } 2428 llvm_unreachable("bad evaluation kind"); 2429 } 2430 2431 llvm::Value *DestPtr = ReturnValue.getValue(); 2432 bool DestIsVolatile = ReturnValue.isVolatile(); 2433 2434 if (!DestPtr) { 2435 DestPtr = CreateMemTemp(RetTy, "coerce"); 2436 DestIsVolatile = false; 2437 } 2438 2439 // If the value is offset in memory, apply the offset now. 2440 llvm::Value *StorePtr = DestPtr; 2441 if (unsigned Offs = RetAI.getDirectOffset()) { 2442 StorePtr = Builder.CreateBitCast(StorePtr, Builder.getInt8PtrTy()); 2443 StorePtr = Builder.CreateConstGEP1_32(StorePtr, Offs); 2444 StorePtr = Builder.CreateBitCast(StorePtr, 2445 llvm::PointerType::getUnqual(RetAI.getCoerceToType())); 2446 } 2447 CreateCoercedStore(CI, StorePtr, DestIsVolatile, *this); 2448 2449 return convertTempToRValue(DestPtr, RetTy); 2450 } 2451 2452 case ABIArgInfo::Expand: 2453 llvm_unreachable("Invalid ABI kind for return argument"); 2454 } 2455 2456 llvm_unreachable("Unhandled ABIArgInfo::Kind"); 2457 } 2458 2459 /* VarArg handling */ 2460 2461 llvm::Value *CodeGenFunction::EmitVAArg(llvm::Value *VAListAddr, QualType Ty) { 2462 return CGM.getTypes().getABIInfo().EmitVAArg(VAListAddr, Ty, *this); 2463 } 2464