CGCall.cpp revision 4ee7dc2369c1f0257a73b2e83a7d38fdebdd9176
1//===--- CGCall.cpp - Encapsulate calling convention details --------------===// 2// 3// The LLVM Compiler Infrastructure 4// 5// This file is distributed under the University of Illinois Open Source 6// License. See LICENSE.TXT for details. 7// 8//===----------------------------------------------------------------------===// 9// 10// These classes wrap the information about a call or function 11// definition used to handle ABI compliancy. 12// 13//===----------------------------------------------------------------------===// 14 15#include "CGCall.h" 16#include "ABIInfo.h" 17#include "CGCXXABI.h" 18#include "CodeGenFunction.h" 19#include "CodeGenModule.h" 20#include "TargetInfo.h" 21#include "clang/AST/Decl.h" 22#include "clang/AST/DeclCXX.h" 23#include "clang/AST/DeclObjC.h" 24#include "clang/Basic/TargetInfo.h" 25#include "clang/Frontend/CodeGenOptions.h" 26#include "llvm/ADT/StringExtras.h" 27#include "llvm/IR/Attributes.h" 28#include "llvm/IR/DataLayout.h" 29#include "llvm/IR/InlineAsm.h" 30#include "llvm/MC/SubtargetFeature.h" 31#include "llvm/Support/CallSite.h" 32#include "llvm/Transforms/Utils/Local.h" 33using namespace clang; 34using namespace CodeGen; 35 36/***/ 37 38static unsigned ClangCallConvToLLVMCallConv(CallingConv CC) { 39 switch (CC) { 40 default: return llvm::CallingConv::C; 41 case CC_X86StdCall: return llvm::CallingConv::X86_StdCall; 42 case CC_X86FastCall: return llvm::CallingConv::X86_FastCall; 43 case CC_X86ThisCall: return llvm::CallingConv::X86_ThisCall; 44 case CC_X86_64Win64: return llvm::CallingConv::X86_64_Win64; 45 case CC_X86_64SysV: return llvm::CallingConv::X86_64_SysV; 46 case CC_AAPCS: return llvm::CallingConv::ARM_AAPCS; 47 case CC_AAPCS_VFP: return llvm::CallingConv::ARM_AAPCS_VFP; 48 case CC_IntelOclBicc: return llvm::CallingConv::Intel_OCL_BI; 49 // TODO: add support for CC_X86Pascal to llvm 50 } 51} 52 53/// Derives the 'this' type for codegen purposes, i.e. ignoring method 54/// qualification. 55/// FIXME: address space qualification? 56static CanQualType GetThisType(ASTContext &Context, const CXXRecordDecl *RD) { 57 QualType RecTy = Context.getTagDeclType(RD)->getCanonicalTypeInternal(); 58 return Context.getPointerType(CanQualType::CreateUnsafe(RecTy)); 59} 60 61/// Returns the canonical formal type of the given C++ method. 62static CanQual<FunctionProtoType> GetFormalType(const CXXMethodDecl *MD) { 63 return MD->getType()->getCanonicalTypeUnqualified() 64 .getAs<FunctionProtoType>(); 65} 66 67/// Returns the "extra-canonicalized" return type, which discards 68/// qualifiers on the return type. Codegen doesn't care about them, 69/// and it makes ABI code a little easier to be able to assume that 70/// all parameter and return types are top-level unqualified. 71static CanQualType GetReturnType(QualType RetTy) { 72 return RetTy->getCanonicalTypeUnqualified().getUnqualifiedType(); 73} 74 75/// Arrange the argument and result information for a value of the given 76/// unprototyped freestanding function type. 77const CGFunctionInfo & 78CodeGenTypes::arrangeFreeFunctionType(CanQual<FunctionNoProtoType> FTNP) { 79 // When translating an unprototyped function type, always use a 80 // variadic type. 81 return arrangeLLVMFunctionInfo(FTNP->getResultType().getUnqualifiedType(), 82 None, FTNP->getExtInfo(), RequiredArgs(0)); 83} 84 85/// Arrange the LLVM function layout for a value of the given function 86/// type, on top of any implicit parameters already stored. Use the 87/// given ExtInfo instead of the ExtInfo from the function type. 88static const CGFunctionInfo &arrangeLLVMFunctionInfo(CodeGenTypes &CGT, 89 SmallVectorImpl<CanQualType> &prefix, 90 CanQual<FunctionProtoType> FTP, 91 FunctionType::ExtInfo extInfo) { 92 RequiredArgs required = RequiredArgs::forPrototypePlus(FTP, prefix.size()); 93 // FIXME: Kill copy. 94 for (unsigned i = 0, e = FTP->getNumArgs(); i != e; ++i) 95 prefix.push_back(FTP->getArgType(i)); 96 CanQualType resultType = FTP->getResultType().getUnqualifiedType(); 97 return CGT.arrangeLLVMFunctionInfo(resultType, prefix, extInfo, required); 98} 99 100/// Arrange the argument and result information for a free function (i.e. 101/// not a C++ or ObjC instance method) of the given type. 102static const CGFunctionInfo &arrangeFreeFunctionType(CodeGenTypes &CGT, 103 SmallVectorImpl<CanQualType> &prefix, 104 CanQual<FunctionProtoType> FTP) { 105 return arrangeLLVMFunctionInfo(CGT, prefix, FTP, FTP->getExtInfo()); 106} 107 108/// Arrange the argument and result information for a free function (i.e. 109/// not a C++ or ObjC instance method) of the given type. 110static const CGFunctionInfo &arrangeCXXMethodType(CodeGenTypes &CGT, 111 SmallVectorImpl<CanQualType> &prefix, 112 CanQual<FunctionProtoType> FTP) { 113 FunctionType::ExtInfo extInfo = FTP->getExtInfo(); 114 return arrangeLLVMFunctionInfo(CGT, prefix, FTP, extInfo); 115} 116 117/// Arrange the argument and result information for a value of the 118/// given freestanding function type. 119const CGFunctionInfo & 120CodeGenTypes::arrangeFreeFunctionType(CanQual<FunctionProtoType> FTP) { 121 SmallVector<CanQualType, 16> argTypes; 122 return ::arrangeFreeFunctionType(*this, argTypes, FTP); 123} 124 125static CallingConv getCallingConventionForDecl(const Decl *D) { 126 // Set the appropriate calling convention for the Function. 127 if (D->hasAttr<StdCallAttr>()) 128 return CC_X86StdCall; 129 130 if (D->hasAttr<FastCallAttr>()) 131 return CC_X86FastCall; 132 133 if (D->hasAttr<ThisCallAttr>()) 134 return CC_X86ThisCall; 135 136 if (D->hasAttr<PascalAttr>()) 137 return CC_X86Pascal; 138 139 if (PcsAttr *PCS = D->getAttr<PcsAttr>()) 140 return (PCS->getPCS() == PcsAttr::AAPCS ? CC_AAPCS : CC_AAPCS_VFP); 141 142 if (D->hasAttr<PnaclCallAttr>()) 143 return CC_PnaclCall; 144 145 if (D->hasAttr<IntelOclBiccAttr>()) 146 return CC_IntelOclBicc; 147 148 return CC_C; 149} 150 151/// Arrange the argument and result information for a call to an 152/// unknown C++ non-static member function of the given abstract type. 153/// (Zero value of RD means we don't have any meaningful "this" argument type, 154/// so fall back to a generic pointer type). 155/// The member function must be an ordinary function, i.e. not a 156/// constructor or destructor. 157const CGFunctionInfo & 158CodeGenTypes::arrangeCXXMethodType(const CXXRecordDecl *RD, 159 const FunctionProtoType *FTP) { 160 SmallVector<CanQualType, 16> argTypes; 161 162 // Add the 'this' pointer. 163 if (RD) 164 argTypes.push_back(GetThisType(Context, RD)); 165 else 166 argTypes.push_back(Context.VoidPtrTy); 167 168 return ::arrangeCXXMethodType(*this, argTypes, 169 FTP->getCanonicalTypeUnqualified().getAs<FunctionProtoType>()); 170} 171 172/// Arrange the argument and result information for a declaration or 173/// definition of the given C++ non-static member function. The 174/// member function must be an ordinary function, i.e. not a 175/// constructor or destructor. 176const CGFunctionInfo & 177CodeGenTypes::arrangeCXXMethodDeclaration(const CXXMethodDecl *MD) { 178 assert(!isa<CXXConstructorDecl>(MD) && "wrong method for constructors!"); 179 assert(!isa<CXXDestructorDecl>(MD) && "wrong method for destructors!"); 180 181 CanQual<FunctionProtoType> prototype = GetFormalType(MD); 182 183 if (MD->isInstance()) { 184 // The abstract case is perfectly fine. 185 const CXXRecordDecl *ThisType = 186 CGM.getCXXABI().getThisArgumentTypeForMethod(MD); 187 return arrangeCXXMethodType(ThisType, prototype.getTypePtr()); 188 } 189 190 return arrangeFreeFunctionType(prototype); 191} 192 193/// Arrange the argument and result information for a declaration 194/// or definition to the given constructor variant. 195const CGFunctionInfo & 196CodeGenTypes::arrangeCXXConstructorDeclaration(const CXXConstructorDecl *D, 197 CXXCtorType ctorKind) { 198 SmallVector<CanQualType, 16> argTypes; 199 argTypes.push_back(GetThisType(Context, D->getParent())); 200 201 GlobalDecl GD(D, ctorKind); 202 CanQualType resultType = 203 TheCXXABI.HasThisReturn(GD) ? argTypes.front() : Context.VoidTy; 204 205 TheCXXABI.BuildConstructorSignature(D, ctorKind, resultType, argTypes); 206 207 CanQual<FunctionProtoType> FTP = GetFormalType(D); 208 209 RequiredArgs required = RequiredArgs::forPrototypePlus(FTP, argTypes.size()); 210 211 // Add the formal parameters. 212 for (unsigned i = 0, e = FTP->getNumArgs(); i != e; ++i) 213 argTypes.push_back(FTP->getArgType(i)); 214 215 FunctionType::ExtInfo extInfo = FTP->getExtInfo(); 216 return arrangeLLVMFunctionInfo(resultType, argTypes, extInfo, required); 217} 218 219/// Arrange the argument and result information for a declaration, 220/// definition, or call to the given destructor variant. It so 221/// happens that all three cases produce the same information. 222const CGFunctionInfo & 223CodeGenTypes::arrangeCXXDestructor(const CXXDestructorDecl *D, 224 CXXDtorType dtorKind) { 225 SmallVector<CanQualType, 2> argTypes; 226 argTypes.push_back(GetThisType(Context, D->getParent())); 227 228 GlobalDecl GD(D, dtorKind); 229 CanQualType resultType = 230 TheCXXABI.HasThisReturn(GD) ? argTypes.front() : 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 return arrangeLLVMFunctionInfo(resultType, argTypes, extInfo, 240 RequiredArgs::All); 241} 242 243/// Arrange the argument and result information for the declaration or 244/// definition of the given function. 245const CGFunctionInfo & 246CodeGenTypes::arrangeFunctionDeclaration(const FunctionDecl *FD) { 247 if (const CXXMethodDecl *MD = dyn_cast<CXXMethodDecl>(FD)) 248 if (MD->isInstance()) 249 return arrangeCXXMethodDeclaration(MD); 250 251 CanQualType FTy = FD->getType()->getCanonicalTypeUnqualified(); 252 253 assert(isa<FunctionType>(FTy)); 254 255 // When declaring a function without a prototype, always use a 256 // non-variadic type. 257 if (isa<FunctionNoProtoType>(FTy)) { 258 CanQual<FunctionNoProtoType> noProto = FTy.getAs<FunctionNoProtoType>(); 259 return arrangeLLVMFunctionInfo(noProto->getResultType(), None, 260 noProto->getExtInfo(), RequiredArgs::All); 261 } 262 263 assert(isa<FunctionProtoType>(FTy)); 264 return arrangeFreeFunctionType(FTy.getAs<FunctionProtoType>()); 265} 266 267/// Arrange the argument and result information for the declaration or 268/// definition of an Objective-C method. 269const CGFunctionInfo & 270CodeGenTypes::arrangeObjCMethodDeclaration(const ObjCMethodDecl *MD) { 271 // It happens that this is the same as a call with no optional 272 // arguments, except also using the formal 'self' type. 273 return arrangeObjCMessageSendSignature(MD, MD->getSelfDecl()->getType()); 274} 275 276/// Arrange the argument and result information for the function type 277/// through which to perform a send to the given Objective-C method, 278/// using the given receiver type. The receiver type is not always 279/// the 'self' type of the method or even an Objective-C pointer type. 280/// This is *not* the right method for actually performing such a 281/// message send, due to the possibility of optional arguments. 282const CGFunctionInfo & 283CodeGenTypes::arrangeObjCMessageSendSignature(const ObjCMethodDecl *MD, 284 QualType receiverType) { 285 SmallVector<CanQualType, 16> argTys; 286 argTys.push_back(Context.getCanonicalParamType(receiverType)); 287 argTys.push_back(Context.getCanonicalParamType(Context.getObjCSelType())); 288 // FIXME: Kill copy? 289 for (ObjCMethodDecl::param_const_iterator i = MD->param_begin(), 290 e = MD->param_end(); i != e; ++i) { 291 argTys.push_back(Context.getCanonicalParamType((*i)->getType())); 292 } 293 294 FunctionType::ExtInfo einfo; 295 einfo = einfo.withCallingConv(getCallingConventionForDecl(MD)); 296 297 if (getContext().getLangOpts().ObjCAutoRefCount && 298 MD->hasAttr<NSReturnsRetainedAttr>()) 299 einfo = einfo.withProducesResult(true); 300 301 RequiredArgs required = 302 (MD->isVariadic() ? RequiredArgs(argTys.size()) : RequiredArgs::All); 303 304 return arrangeLLVMFunctionInfo(GetReturnType(MD->getResultType()), argTys, 305 einfo, required); 306} 307 308const CGFunctionInfo & 309CodeGenTypes::arrangeGlobalDeclaration(GlobalDecl GD) { 310 // FIXME: Do we need to handle ObjCMethodDecl? 311 const FunctionDecl *FD = cast<FunctionDecl>(GD.getDecl()); 312 313 if (const CXXConstructorDecl *CD = dyn_cast<CXXConstructorDecl>(FD)) 314 return arrangeCXXConstructorDeclaration(CD, GD.getCtorType()); 315 316 if (const CXXDestructorDecl *DD = dyn_cast<CXXDestructorDecl>(FD)) 317 return arrangeCXXDestructor(DD, GD.getDtorType()); 318 319 return arrangeFunctionDeclaration(FD); 320} 321 322/// Arrange a call as unto a free function, except possibly with an 323/// additional number of formal parameters considered required. 324static const CGFunctionInfo & 325arrangeFreeFunctionLikeCall(CodeGenTypes &CGT, 326 const CallArgList &args, 327 const FunctionType *fnType, 328 unsigned numExtraRequiredArgs) { 329 assert(args.size() >= numExtraRequiredArgs); 330 331 // In most cases, there are no optional arguments. 332 RequiredArgs required = RequiredArgs::All; 333 334 // If we have a variadic prototype, the required arguments are the 335 // extra prefix plus the arguments in the prototype. 336 if (const FunctionProtoType *proto = dyn_cast<FunctionProtoType>(fnType)) { 337 if (proto->isVariadic()) 338 required = RequiredArgs(proto->getNumArgs() + numExtraRequiredArgs); 339 340 // If we don't have a prototype at all, but we're supposed to 341 // explicitly use the variadic convention for unprototyped calls, 342 // treat all of the arguments as required but preserve the nominal 343 // possibility of variadics. 344 } else if (CGT.CGM.getTargetCodeGenInfo() 345 .isNoProtoCallVariadic(args, cast<FunctionNoProtoType>(fnType))) { 346 required = RequiredArgs(args.size()); 347 } 348 349 return CGT.arrangeFreeFunctionCall(fnType->getResultType(), args, 350 fnType->getExtInfo(), required); 351} 352 353/// Figure out the rules for calling a function with the given formal 354/// type using the given arguments. The arguments are necessary 355/// because the function might be unprototyped, in which case it's 356/// target-dependent in crazy ways. 357const CGFunctionInfo & 358CodeGenTypes::arrangeFreeFunctionCall(const CallArgList &args, 359 const FunctionType *fnType) { 360 return arrangeFreeFunctionLikeCall(*this, args, fnType, 0); 361} 362 363/// A block function call is essentially a free-function call with an 364/// extra implicit argument. 365const CGFunctionInfo & 366CodeGenTypes::arrangeBlockFunctionCall(const CallArgList &args, 367 const FunctionType *fnType) { 368 return arrangeFreeFunctionLikeCall(*this, args, fnType, 1); 369} 370 371const CGFunctionInfo & 372CodeGenTypes::arrangeFreeFunctionCall(QualType resultType, 373 const CallArgList &args, 374 FunctionType::ExtInfo info, 375 RequiredArgs required) { 376 // FIXME: Kill copy. 377 SmallVector<CanQualType, 16> argTypes; 378 for (CallArgList::const_iterator i = args.begin(), e = args.end(); 379 i != e; ++i) 380 argTypes.push_back(Context.getCanonicalParamType(i->Ty)); 381 return arrangeLLVMFunctionInfo(GetReturnType(resultType), argTypes, info, 382 required); 383} 384 385/// Arrange a call to a C++ method, passing the given arguments. 386const CGFunctionInfo & 387CodeGenTypes::arrangeCXXMethodCall(const CallArgList &args, 388 const FunctionProtoType *FPT, 389 RequiredArgs required) { 390 // FIXME: Kill copy. 391 SmallVector<CanQualType, 16> argTypes; 392 for (CallArgList::const_iterator i = args.begin(), e = args.end(); 393 i != e; ++i) 394 argTypes.push_back(Context.getCanonicalParamType(i->Ty)); 395 396 FunctionType::ExtInfo info = FPT->getExtInfo(); 397 return arrangeLLVMFunctionInfo(GetReturnType(FPT->getResultType()), 398 argTypes, info, required); 399} 400 401const CGFunctionInfo & 402CodeGenTypes::arrangeFunctionDeclaration(QualType resultType, 403 const FunctionArgList &args, 404 const FunctionType::ExtInfo &info, 405 bool isVariadic) { 406 // FIXME: Kill copy. 407 SmallVector<CanQualType, 16> argTypes; 408 for (FunctionArgList::const_iterator i = args.begin(), e = args.end(); 409 i != e; ++i) 410 argTypes.push_back(Context.getCanonicalParamType((*i)->getType())); 411 412 RequiredArgs required = 413 (isVariadic ? RequiredArgs(args.size()) : RequiredArgs::All); 414 return arrangeLLVMFunctionInfo(GetReturnType(resultType), argTypes, info, 415 required); 416} 417 418const CGFunctionInfo &CodeGenTypes::arrangeNullaryFunction() { 419 return arrangeLLVMFunctionInfo(getContext().VoidTy, None, 420 FunctionType::ExtInfo(), RequiredArgs::All); 421} 422 423/// Arrange the argument and result information for an abstract value 424/// of a given function type. This is the method which all of the 425/// above functions ultimately defer to. 426const CGFunctionInfo & 427CodeGenTypes::arrangeLLVMFunctionInfo(CanQualType resultType, 428 ArrayRef<CanQualType> argTypes, 429 FunctionType::ExtInfo info, 430 RequiredArgs required) { 431#ifndef NDEBUG 432 for (ArrayRef<CanQualType>::const_iterator 433 I = argTypes.begin(), E = argTypes.end(); I != E; ++I) 434 assert(I->isCanonicalAsParam()); 435#endif 436 437 unsigned CC = ClangCallConvToLLVMCallConv(info.getCC()); 438 439 // Lookup or create unique function info. 440 llvm::FoldingSetNodeID ID; 441 CGFunctionInfo::Profile(ID, info, required, resultType, argTypes); 442 443 void *insertPos = 0; 444 CGFunctionInfo *FI = FunctionInfos.FindNodeOrInsertPos(ID, insertPos); 445 if (FI) 446 return *FI; 447 448 // Construct the function info. We co-allocate the ArgInfos. 449 FI = CGFunctionInfo::create(CC, info, resultType, argTypes, required); 450 FunctionInfos.InsertNode(FI, insertPos); 451 452 bool inserted = FunctionsBeingProcessed.insert(FI); (void)inserted; 453 assert(inserted && "Recursively being processed?"); 454 455 // Compute ABI information. 456 getABIInfo().computeInfo(*FI); 457 458 // Loop over all of the computed argument and return value info. If any of 459 // them are direct or extend without a specified coerce type, specify the 460 // default now. 461 ABIArgInfo &retInfo = FI->getReturnInfo(); 462 if (retInfo.canHaveCoerceToType() && retInfo.getCoerceToType() == 0) 463 retInfo.setCoerceToType(ConvertType(FI->getReturnType())); 464 465 for (CGFunctionInfo::arg_iterator I = FI->arg_begin(), E = FI->arg_end(); 466 I != E; ++I) 467 if (I->info.canHaveCoerceToType() && I->info.getCoerceToType() == 0) 468 I->info.setCoerceToType(ConvertType(I->type)); 469 470 bool erased = FunctionsBeingProcessed.erase(FI); (void)erased; 471 assert(erased && "Not in set?"); 472 473 return *FI; 474} 475 476CGFunctionInfo *CGFunctionInfo::create(unsigned llvmCC, 477 const FunctionType::ExtInfo &info, 478 CanQualType resultType, 479 ArrayRef<CanQualType> argTypes, 480 RequiredArgs required) { 481 void *buffer = operator new(sizeof(CGFunctionInfo) + 482 sizeof(ArgInfo) * (argTypes.size() + 1)); 483 CGFunctionInfo *FI = new(buffer) CGFunctionInfo(); 484 FI->CallingConvention = llvmCC; 485 FI->EffectiveCallingConvention = llvmCC; 486 FI->ASTCallingConvention = info.getCC(); 487 FI->NoReturn = info.getNoReturn(); 488 FI->ReturnsRetained = info.getProducesResult(); 489 FI->Required = required; 490 FI->HasRegParm = info.getHasRegParm(); 491 FI->RegParm = info.getRegParm(); 492 FI->NumArgs = argTypes.size(); 493 FI->getArgsBuffer()[0].type = resultType; 494 for (unsigned i = 0, e = argTypes.size(); i != e; ++i) 495 FI->getArgsBuffer()[i + 1].type = argTypes[i]; 496 return FI; 497} 498 499/***/ 500 501void CodeGenTypes::GetExpandedTypes(QualType type, 502 SmallVectorImpl<llvm::Type*> &expandedTypes) { 503 if (const ConstantArrayType *AT = Context.getAsConstantArrayType(type)) { 504 uint64_t NumElts = AT->getSize().getZExtValue(); 505 for (uint64_t Elt = 0; Elt < NumElts; ++Elt) 506 GetExpandedTypes(AT->getElementType(), expandedTypes); 507 } else if (const RecordType *RT = type->getAs<RecordType>()) { 508 const RecordDecl *RD = RT->getDecl(); 509 assert(!RD->hasFlexibleArrayMember() && 510 "Cannot expand structure with flexible array."); 511 if (RD->isUnion()) { 512 // Unions can be here only in degenerative cases - all the fields are same 513 // after flattening. Thus we have to use the "largest" field. 514 const FieldDecl *LargestFD = 0; 515 CharUnits UnionSize = CharUnits::Zero(); 516 517 for (RecordDecl::field_iterator i = RD->field_begin(), e = RD->field_end(); 518 i != e; ++i) { 519 const FieldDecl *FD = *i; 520 assert(!FD->isBitField() && 521 "Cannot expand structure with bit-field members."); 522 CharUnits FieldSize = getContext().getTypeSizeInChars(FD->getType()); 523 if (UnionSize < FieldSize) { 524 UnionSize = FieldSize; 525 LargestFD = FD; 526 } 527 } 528 if (LargestFD) 529 GetExpandedTypes(LargestFD->getType(), expandedTypes); 530 } else { 531 for (RecordDecl::field_iterator i = RD->field_begin(), e = RD->field_end(); 532 i != e; ++i) { 533 assert(!i->isBitField() && 534 "Cannot expand structure with bit-field members."); 535 GetExpandedTypes(i->getType(), expandedTypes); 536 } 537 } 538 } else if (const ComplexType *CT = type->getAs<ComplexType>()) { 539 llvm::Type *EltTy = ConvertType(CT->getElementType()); 540 expandedTypes.push_back(EltTy); 541 expandedTypes.push_back(EltTy); 542 } else 543 expandedTypes.push_back(ConvertType(type)); 544} 545 546llvm::Function::arg_iterator 547CodeGenFunction::ExpandTypeFromArgs(QualType Ty, LValue LV, 548 llvm::Function::arg_iterator AI) { 549 assert(LV.isSimple() && 550 "Unexpected non-simple lvalue during struct expansion."); 551 552 if (const ConstantArrayType *AT = getContext().getAsConstantArrayType(Ty)) { 553 unsigned NumElts = AT->getSize().getZExtValue(); 554 QualType EltTy = AT->getElementType(); 555 for (unsigned Elt = 0; Elt < NumElts; ++Elt) { 556 llvm::Value *EltAddr = Builder.CreateConstGEP2_32(LV.getAddress(), 0, Elt); 557 LValue LV = MakeAddrLValue(EltAddr, EltTy); 558 AI = ExpandTypeFromArgs(EltTy, LV, AI); 559 } 560 } else if (const RecordType *RT = Ty->getAs<RecordType>()) { 561 RecordDecl *RD = RT->getDecl(); 562 if (RD->isUnion()) { 563 // Unions can be here only in degenerative cases - all the fields are same 564 // after flattening. Thus we have to use the "largest" field. 565 const FieldDecl *LargestFD = 0; 566 CharUnits UnionSize = CharUnits::Zero(); 567 568 for (RecordDecl::field_iterator i = RD->field_begin(), e = RD->field_end(); 569 i != e; ++i) { 570 const FieldDecl *FD = *i; 571 assert(!FD->isBitField() && 572 "Cannot expand structure with bit-field members."); 573 CharUnits FieldSize = getContext().getTypeSizeInChars(FD->getType()); 574 if (UnionSize < FieldSize) { 575 UnionSize = FieldSize; 576 LargestFD = FD; 577 } 578 } 579 if (LargestFD) { 580 // FIXME: What are the right qualifiers here? 581 LValue SubLV = EmitLValueForField(LV, LargestFD); 582 AI = ExpandTypeFromArgs(LargestFD->getType(), SubLV, AI); 583 } 584 } else { 585 for (RecordDecl::field_iterator i = RD->field_begin(), e = RD->field_end(); 586 i != e; ++i) { 587 FieldDecl *FD = *i; 588 QualType FT = FD->getType(); 589 590 // FIXME: What are the right qualifiers here? 591 LValue SubLV = EmitLValueForField(LV, FD); 592 AI = ExpandTypeFromArgs(FT, SubLV, AI); 593 } 594 } 595 } else if (const ComplexType *CT = Ty->getAs<ComplexType>()) { 596 QualType EltTy = CT->getElementType(); 597 llvm::Value *RealAddr = Builder.CreateStructGEP(LV.getAddress(), 0, "real"); 598 EmitStoreThroughLValue(RValue::get(AI++), MakeAddrLValue(RealAddr, EltTy)); 599 llvm::Value *ImagAddr = Builder.CreateStructGEP(LV.getAddress(), 1, "imag"); 600 EmitStoreThroughLValue(RValue::get(AI++), MakeAddrLValue(ImagAddr, EltTy)); 601 } else { 602 EmitStoreThroughLValue(RValue::get(AI), LV); 603 ++AI; 604 } 605 606 return AI; 607} 608 609/// EnterStructPointerForCoercedAccess - Given a struct pointer that we are 610/// accessing some number of bytes out of it, try to gep into the struct to get 611/// at its inner goodness. Dive as deep as possible without entering an element 612/// with an in-memory size smaller than DstSize. 613static llvm::Value * 614EnterStructPointerForCoercedAccess(llvm::Value *SrcPtr, 615 llvm::StructType *SrcSTy, 616 uint64_t DstSize, CodeGenFunction &CGF) { 617 // We can't dive into a zero-element struct. 618 if (SrcSTy->getNumElements() == 0) return SrcPtr; 619 620 llvm::Type *FirstElt = SrcSTy->getElementType(0); 621 622 // If the first elt is at least as large as what we're looking for, or if the 623 // first element is the same size as the whole struct, we can enter it. 624 uint64_t FirstEltSize = 625 CGF.CGM.getDataLayout().getTypeAllocSize(FirstElt); 626 if (FirstEltSize < DstSize && 627 FirstEltSize < CGF.CGM.getDataLayout().getTypeAllocSize(SrcSTy)) 628 return SrcPtr; 629 630 // GEP into the first element. 631 SrcPtr = CGF.Builder.CreateConstGEP2_32(SrcPtr, 0, 0, "coerce.dive"); 632 633 // If the first element is a struct, recurse. 634 llvm::Type *SrcTy = 635 cast<llvm::PointerType>(SrcPtr->getType())->getElementType(); 636 if (llvm::StructType *SrcSTy = dyn_cast<llvm::StructType>(SrcTy)) 637 return EnterStructPointerForCoercedAccess(SrcPtr, SrcSTy, DstSize, CGF); 638 639 return SrcPtr; 640} 641 642/// CoerceIntOrPtrToIntOrPtr - Convert a value Val to the specific Ty where both 643/// are either integers or pointers. This does a truncation of the value if it 644/// is too large or a zero extension if it is too small. 645/// 646/// This behaves as if the value were coerced through memory, so on big-endian 647/// targets the high bits are preserved in a truncation, while little-endian 648/// targets preserve the low bits. 649static llvm::Value *CoerceIntOrPtrToIntOrPtr(llvm::Value *Val, 650 llvm::Type *Ty, 651 CodeGenFunction &CGF) { 652 if (Val->getType() == Ty) 653 return Val; 654 655 if (isa<llvm::PointerType>(Val->getType())) { 656 // If this is Pointer->Pointer avoid conversion to and from int. 657 if (isa<llvm::PointerType>(Ty)) 658 return CGF.Builder.CreateBitCast(Val, Ty, "coerce.val"); 659 660 // Convert the pointer to an integer so we can play with its width. 661 Val = CGF.Builder.CreatePtrToInt(Val, CGF.IntPtrTy, "coerce.val.pi"); 662 } 663 664 llvm::Type *DestIntTy = Ty; 665 if (isa<llvm::PointerType>(DestIntTy)) 666 DestIntTy = CGF.IntPtrTy; 667 668 if (Val->getType() != DestIntTy) { 669 const llvm::DataLayout &DL = CGF.CGM.getDataLayout(); 670 if (DL.isBigEndian()) { 671 // Preserve the high bits on big-endian targets. 672 // That is what memory coercion does. 673 uint64_t SrcSize = DL.getTypeAllocSizeInBits(Val->getType()); 674 uint64_t DstSize = DL.getTypeAllocSizeInBits(DestIntTy); 675 if (SrcSize > DstSize) { 676 Val = CGF.Builder.CreateLShr(Val, SrcSize - DstSize, "coerce.highbits"); 677 Val = CGF.Builder.CreateTrunc(Val, DestIntTy, "coerce.val.ii"); 678 } else { 679 Val = CGF.Builder.CreateZExt(Val, DestIntTy, "coerce.val.ii"); 680 Val = CGF.Builder.CreateShl(Val, DstSize - SrcSize, "coerce.highbits"); 681 } 682 } else { 683 // Little-endian targets preserve the low bits. No shifts required. 684 Val = CGF.Builder.CreateIntCast(Val, DestIntTy, false, "coerce.val.ii"); 685 } 686 } 687 688 if (isa<llvm::PointerType>(Ty)) 689 Val = CGF.Builder.CreateIntToPtr(Val, Ty, "coerce.val.ip"); 690 return Val; 691} 692 693 694 695/// CreateCoercedLoad - Create a load from \arg SrcPtr interpreted as 696/// a pointer to an object of type \arg Ty. 697/// 698/// This safely handles the case when the src type is smaller than the 699/// destination type; in this situation the values of bits which not 700/// present in the src are undefined. 701static llvm::Value *CreateCoercedLoad(llvm::Value *SrcPtr, 702 llvm::Type *Ty, 703 CodeGenFunction &CGF) { 704 llvm::Type *SrcTy = 705 cast<llvm::PointerType>(SrcPtr->getType())->getElementType(); 706 707 // If SrcTy and Ty are the same, just do a load. 708 if (SrcTy == Ty) 709 return CGF.Builder.CreateLoad(SrcPtr); 710 711 uint64_t DstSize = CGF.CGM.getDataLayout().getTypeAllocSize(Ty); 712 713 if (llvm::StructType *SrcSTy = dyn_cast<llvm::StructType>(SrcTy)) { 714 SrcPtr = EnterStructPointerForCoercedAccess(SrcPtr, SrcSTy, DstSize, CGF); 715 SrcTy = cast<llvm::PointerType>(SrcPtr->getType())->getElementType(); 716 } 717 718 uint64_t SrcSize = CGF.CGM.getDataLayout().getTypeAllocSize(SrcTy); 719 720 // If the source and destination are integer or pointer types, just do an 721 // extension or truncation to the desired type. 722 if ((isa<llvm::IntegerType>(Ty) || isa<llvm::PointerType>(Ty)) && 723 (isa<llvm::IntegerType>(SrcTy) || isa<llvm::PointerType>(SrcTy))) { 724 llvm::LoadInst *Load = CGF.Builder.CreateLoad(SrcPtr); 725 return CoerceIntOrPtrToIntOrPtr(Load, Ty, CGF); 726 } 727 728 // If load is legal, just bitcast the src pointer. 729 if (SrcSize >= DstSize) { 730 // Generally SrcSize is never greater than DstSize, since this means we are 731 // losing bits. However, this can happen in cases where the structure has 732 // additional padding, for example due to a user specified alignment. 733 // 734 // FIXME: Assert that we aren't truncating non-padding bits when have access 735 // to that information. 736 llvm::Value *Casted = 737 CGF.Builder.CreateBitCast(SrcPtr, llvm::PointerType::getUnqual(Ty)); 738 llvm::LoadInst *Load = CGF.Builder.CreateLoad(Casted); 739 // FIXME: Use better alignment / avoid requiring aligned load. 740 Load->setAlignment(1); 741 return Load; 742 } 743 744 // Otherwise do coercion through memory. This is stupid, but 745 // simple. 746 llvm::Value *Tmp = CGF.CreateTempAlloca(Ty); 747 llvm::Type *I8PtrTy = CGF.Builder.getInt8PtrTy(); 748 llvm::Value *Casted = CGF.Builder.CreateBitCast(Tmp, I8PtrTy); 749 llvm::Value *SrcCasted = CGF.Builder.CreateBitCast(SrcPtr, I8PtrTy); 750 // FIXME: Use better alignment. 751 CGF.Builder.CreateMemCpy(Casted, SrcCasted, 752 llvm::ConstantInt::get(CGF.IntPtrTy, SrcSize), 753 1, false); 754 return CGF.Builder.CreateLoad(Tmp); 755} 756 757// Function to store a first-class aggregate into memory. We prefer to 758// store the elements rather than the aggregate to be more friendly to 759// fast-isel. 760// FIXME: Do we need to recurse here? 761static void BuildAggStore(CodeGenFunction &CGF, llvm::Value *Val, 762 llvm::Value *DestPtr, bool DestIsVolatile, 763 bool LowAlignment) { 764 // Prefer scalar stores to first-class aggregate stores. 765 if (llvm::StructType *STy = 766 dyn_cast<llvm::StructType>(Val->getType())) { 767 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 768 llvm::Value *EltPtr = CGF.Builder.CreateConstGEP2_32(DestPtr, 0, i); 769 llvm::Value *Elt = CGF.Builder.CreateExtractValue(Val, i); 770 llvm::StoreInst *SI = CGF.Builder.CreateStore(Elt, EltPtr, 771 DestIsVolatile); 772 if (LowAlignment) 773 SI->setAlignment(1); 774 } 775 } else { 776 llvm::StoreInst *SI = CGF.Builder.CreateStore(Val, DestPtr, DestIsVolatile); 777 if (LowAlignment) 778 SI->setAlignment(1); 779 } 780} 781 782/// CreateCoercedStore - Create a store to \arg DstPtr from \arg Src, 783/// where the source and destination may have different types. 784/// 785/// This safely handles the case when the src type is larger than the 786/// destination type; the upper bits of the src will be lost. 787static void CreateCoercedStore(llvm::Value *Src, 788 llvm::Value *DstPtr, 789 bool DstIsVolatile, 790 CodeGenFunction &CGF) { 791 llvm::Type *SrcTy = Src->getType(); 792 llvm::Type *DstTy = 793 cast<llvm::PointerType>(DstPtr->getType())->getElementType(); 794 if (SrcTy == DstTy) { 795 CGF.Builder.CreateStore(Src, DstPtr, DstIsVolatile); 796 return; 797 } 798 799 uint64_t SrcSize = CGF.CGM.getDataLayout().getTypeAllocSize(SrcTy); 800 801 if (llvm::StructType *DstSTy = dyn_cast<llvm::StructType>(DstTy)) { 802 DstPtr = EnterStructPointerForCoercedAccess(DstPtr, DstSTy, SrcSize, CGF); 803 DstTy = cast<llvm::PointerType>(DstPtr->getType())->getElementType(); 804 } 805 806 // If the source and destination are integer or pointer types, just do an 807 // extension or truncation to the desired type. 808 if ((isa<llvm::IntegerType>(SrcTy) || isa<llvm::PointerType>(SrcTy)) && 809 (isa<llvm::IntegerType>(DstTy) || isa<llvm::PointerType>(DstTy))) { 810 Src = CoerceIntOrPtrToIntOrPtr(Src, DstTy, CGF); 811 CGF.Builder.CreateStore(Src, DstPtr, DstIsVolatile); 812 return; 813 } 814 815 uint64_t DstSize = CGF.CGM.getDataLayout().getTypeAllocSize(DstTy); 816 817 // If store is legal, just bitcast the src pointer. 818 if (SrcSize <= DstSize) { 819 llvm::Value *Casted = 820 CGF.Builder.CreateBitCast(DstPtr, llvm::PointerType::getUnqual(SrcTy)); 821 // FIXME: Use better alignment / avoid requiring aligned store. 822 BuildAggStore(CGF, Src, Casted, DstIsVolatile, true); 823 } else { 824 // Otherwise do coercion through memory. This is stupid, but 825 // simple. 826 827 // Generally SrcSize is never greater than DstSize, since this means we are 828 // losing bits. However, this can happen in cases where the structure has 829 // additional padding, for example due to a user specified alignment. 830 // 831 // FIXME: Assert that we aren't truncating non-padding bits when have access 832 // to that information. 833 llvm::Value *Tmp = CGF.CreateTempAlloca(SrcTy); 834 CGF.Builder.CreateStore(Src, Tmp); 835 llvm::Type *I8PtrTy = CGF.Builder.getInt8PtrTy(); 836 llvm::Value *Casted = CGF.Builder.CreateBitCast(Tmp, I8PtrTy); 837 llvm::Value *DstCasted = CGF.Builder.CreateBitCast(DstPtr, I8PtrTy); 838 // FIXME: Use better alignment. 839 CGF.Builder.CreateMemCpy(DstCasted, Casted, 840 llvm::ConstantInt::get(CGF.IntPtrTy, DstSize), 841 1, false); 842 } 843} 844 845/***/ 846 847bool CodeGenModule::ReturnTypeUsesSRet(const CGFunctionInfo &FI) { 848 return FI.getReturnInfo().isIndirect(); 849} 850 851bool CodeGenModule::ReturnTypeUsesFPRet(QualType ResultType) { 852 if (const BuiltinType *BT = ResultType->getAs<BuiltinType>()) { 853 switch (BT->getKind()) { 854 default: 855 return false; 856 case BuiltinType::Float: 857 return getTarget().useObjCFPRetForRealType(TargetInfo::Float); 858 case BuiltinType::Double: 859 return getTarget().useObjCFPRetForRealType(TargetInfo::Double); 860 case BuiltinType::LongDouble: 861 return getTarget().useObjCFPRetForRealType(TargetInfo::LongDouble); 862 } 863 } 864 865 return false; 866} 867 868bool CodeGenModule::ReturnTypeUsesFP2Ret(QualType ResultType) { 869 if (const ComplexType *CT = ResultType->getAs<ComplexType>()) { 870 if (const BuiltinType *BT = CT->getElementType()->getAs<BuiltinType>()) { 871 if (BT->getKind() == BuiltinType::LongDouble) 872 return getTarget().useObjCFP2RetForComplexLongDouble(); 873 } 874 } 875 876 return false; 877} 878 879llvm::FunctionType *CodeGenTypes::GetFunctionType(GlobalDecl GD) { 880 const CGFunctionInfo &FI = arrangeGlobalDeclaration(GD); 881 return GetFunctionType(FI); 882} 883 884llvm::FunctionType * 885CodeGenTypes::GetFunctionType(const CGFunctionInfo &FI) { 886 887 bool Inserted = FunctionsBeingProcessed.insert(&FI); (void)Inserted; 888 assert(Inserted && "Recursively being processed?"); 889 890 SmallVector<llvm::Type*, 8> argTypes; 891 llvm::Type *resultType = 0; 892 893 const ABIArgInfo &retAI = FI.getReturnInfo(); 894 switch (retAI.getKind()) { 895 case ABIArgInfo::Expand: 896 llvm_unreachable("Invalid ABI kind for return argument"); 897 898 case ABIArgInfo::Extend: 899 case ABIArgInfo::Direct: 900 resultType = retAI.getCoerceToType(); 901 break; 902 903 case ABIArgInfo::Indirect: { 904 assert(!retAI.getIndirectAlign() && "Align unused on indirect return."); 905 resultType = llvm::Type::getVoidTy(getLLVMContext()); 906 907 QualType ret = FI.getReturnType(); 908 llvm::Type *ty = ConvertType(ret); 909 unsigned addressSpace = Context.getTargetAddressSpace(ret); 910 argTypes.push_back(llvm::PointerType::get(ty, addressSpace)); 911 break; 912 } 913 914 case ABIArgInfo::Ignore: 915 resultType = llvm::Type::getVoidTy(getLLVMContext()); 916 break; 917 } 918 919 // Add in all of the required arguments. 920 CGFunctionInfo::const_arg_iterator it = FI.arg_begin(), ie; 921 if (FI.isVariadic()) { 922 ie = it + FI.getRequiredArgs().getNumRequiredArgs(); 923 } else { 924 ie = FI.arg_end(); 925 } 926 for (; it != ie; ++it) { 927 const ABIArgInfo &argAI = it->info; 928 929 // Insert a padding type to ensure proper alignment. 930 if (llvm::Type *PaddingType = argAI.getPaddingType()) 931 argTypes.push_back(PaddingType); 932 933 switch (argAI.getKind()) { 934 case ABIArgInfo::Ignore: 935 break; 936 937 case ABIArgInfo::Indirect: { 938 // indirect arguments are always on the stack, which is addr space #0. 939 llvm::Type *LTy = ConvertTypeForMem(it->type); 940 argTypes.push_back(LTy->getPointerTo()); 941 break; 942 } 943 944 case ABIArgInfo::Extend: 945 case ABIArgInfo::Direct: { 946 // If the coerce-to type is a first class aggregate, flatten it. Either 947 // way is semantically identical, but fast-isel and the optimizer 948 // generally likes scalar values better than FCAs. 949 llvm::Type *argType = argAI.getCoerceToType(); 950 if (llvm::StructType *st = dyn_cast<llvm::StructType>(argType)) { 951 for (unsigned i = 0, e = st->getNumElements(); i != e; ++i) 952 argTypes.push_back(st->getElementType(i)); 953 } else { 954 argTypes.push_back(argType); 955 } 956 break; 957 } 958 959 case ABIArgInfo::Expand: 960 GetExpandedTypes(it->type, argTypes); 961 break; 962 } 963 } 964 965 bool Erased = FunctionsBeingProcessed.erase(&FI); (void)Erased; 966 assert(Erased && "Not in set?"); 967 968 return llvm::FunctionType::get(resultType, argTypes, FI.isVariadic()); 969} 970 971llvm::Type *CodeGenTypes::GetFunctionTypeForVTable(GlobalDecl GD) { 972 const CXXMethodDecl *MD = cast<CXXMethodDecl>(GD.getDecl()); 973 const FunctionProtoType *FPT = MD->getType()->getAs<FunctionProtoType>(); 974 975 if (!isFuncTypeConvertible(FPT)) 976 return llvm::StructType::get(getLLVMContext()); 977 978 const CGFunctionInfo *Info; 979 if (isa<CXXDestructorDecl>(MD)) 980 Info = &arrangeCXXDestructor(cast<CXXDestructorDecl>(MD), GD.getDtorType()); 981 else 982 Info = &arrangeCXXMethodDeclaration(MD); 983 return GetFunctionType(*Info); 984} 985 986void CodeGenModule::ConstructAttributeList(const CGFunctionInfo &FI, 987 const Decl *TargetDecl, 988 AttributeListType &PAL, 989 unsigned &CallingConv, 990 bool AttrOnCallSite) { 991 llvm::AttrBuilder FuncAttrs; 992 llvm::AttrBuilder RetAttrs; 993 994 CallingConv = FI.getEffectiveCallingConvention(); 995 996 if (FI.isNoReturn()) 997 FuncAttrs.addAttribute(llvm::Attribute::NoReturn); 998 999 // FIXME: handle sseregparm someday... 1000 if (TargetDecl) { 1001 if (TargetDecl->hasAttr<ReturnsTwiceAttr>()) 1002 FuncAttrs.addAttribute(llvm::Attribute::ReturnsTwice); 1003 if (TargetDecl->hasAttr<NoThrowAttr>()) 1004 FuncAttrs.addAttribute(llvm::Attribute::NoUnwind); 1005 if (TargetDecl->hasAttr<NoReturnAttr>()) 1006 FuncAttrs.addAttribute(llvm::Attribute::NoReturn); 1007 1008 if (const FunctionDecl *Fn = dyn_cast<FunctionDecl>(TargetDecl)) { 1009 const FunctionProtoType *FPT = Fn->getType()->getAs<FunctionProtoType>(); 1010 if (FPT && FPT->isNothrow(getContext())) 1011 FuncAttrs.addAttribute(llvm::Attribute::NoUnwind); 1012 // Don't use [[noreturn]] or _Noreturn for a call to a virtual function. 1013 // These attributes are not inherited by overloads. 1014 const CXXMethodDecl *MD = dyn_cast<CXXMethodDecl>(Fn); 1015 if (Fn->isNoReturn() && !(AttrOnCallSite && MD && MD->isVirtual())) 1016 FuncAttrs.addAttribute(llvm::Attribute::NoReturn); 1017 } 1018 1019 // 'const' and 'pure' attribute functions are also nounwind. 1020 if (TargetDecl->hasAttr<ConstAttr>()) { 1021 FuncAttrs.addAttribute(llvm::Attribute::ReadNone); 1022 FuncAttrs.addAttribute(llvm::Attribute::NoUnwind); 1023 } else if (TargetDecl->hasAttr<PureAttr>()) { 1024 FuncAttrs.addAttribute(llvm::Attribute::ReadOnly); 1025 FuncAttrs.addAttribute(llvm::Attribute::NoUnwind); 1026 } 1027 if (TargetDecl->hasAttr<MallocAttr>()) 1028 RetAttrs.addAttribute(llvm::Attribute::NoAlias); 1029 } 1030 1031 if (CodeGenOpts.OptimizeSize) 1032 FuncAttrs.addAttribute(llvm::Attribute::OptimizeForSize); 1033 if (CodeGenOpts.OptimizeSize == 2) 1034 FuncAttrs.addAttribute(llvm::Attribute::MinSize); 1035 if (CodeGenOpts.DisableRedZone) 1036 FuncAttrs.addAttribute(llvm::Attribute::NoRedZone); 1037 if (CodeGenOpts.NoImplicitFloat) 1038 FuncAttrs.addAttribute(llvm::Attribute::NoImplicitFloat); 1039 1040 if (AttrOnCallSite) { 1041 // Attributes that should go on the call site only. 1042 if (!CodeGenOpts.SimplifyLibCalls) 1043 FuncAttrs.addAttribute(llvm::Attribute::NoBuiltin); 1044 } else { 1045 // Attributes that should go on the function, but not the call site. 1046 if (!CodeGenOpts.DisableFPElim) { 1047 FuncAttrs.addAttribute("no-frame-pointer-elim", "false"); 1048 } else if (CodeGenOpts.OmitLeafFramePointer) { 1049 FuncAttrs.addAttribute("no-frame-pointer-elim", "false"); 1050 FuncAttrs.addAttribute("no-frame-pointer-elim-non-leaf"); 1051 } else { 1052 FuncAttrs.addAttribute("no-frame-pointer-elim", "true"); 1053 FuncAttrs.addAttribute("no-frame-pointer-elim-non-leaf"); 1054 } 1055 1056 FuncAttrs.addAttribute("less-precise-fpmad", 1057 llvm::toStringRef(CodeGenOpts.LessPreciseFPMAD)); 1058 FuncAttrs.addAttribute("no-infs-fp-math", 1059 llvm::toStringRef(CodeGenOpts.NoInfsFPMath)); 1060 FuncAttrs.addAttribute("no-nans-fp-math", 1061 llvm::toStringRef(CodeGenOpts.NoNaNsFPMath)); 1062 FuncAttrs.addAttribute("unsafe-fp-math", 1063 llvm::toStringRef(CodeGenOpts.UnsafeFPMath)); 1064 FuncAttrs.addAttribute("use-soft-float", 1065 llvm::toStringRef(CodeGenOpts.SoftFloat)); 1066 FuncAttrs.addAttribute("stack-protector-buffer-size", 1067 llvm::utostr(CodeGenOpts.SSPBufferSize)); 1068 1069 if (!CodeGenOpts.StackRealignment) 1070 FuncAttrs.addAttribute("no-realign-stack"); 1071 } 1072 1073 QualType RetTy = FI.getReturnType(); 1074 unsigned Index = 1; 1075 const ABIArgInfo &RetAI = FI.getReturnInfo(); 1076 switch (RetAI.getKind()) { 1077 case ABIArgInfo::Extend: 1078 if (RetTy->hasSignedIntegerRepresentation()) 1079 RetAttrs.addAttribute(llvm::Attribute::SExt); 1080 else if (RetTy->hasUnsignedIntegerRepresentation()) 1081 RetAttrs.addAttribute(llvm::Attribute::ZExt); 1082 // FALL THROUGH 1083 case ABIArgInfo::Direct: 1084 if (RetAI.getInReg()) 1085 RetAttrs.addAttribute(llvm::Attribute::InReg); 1086 break; 1087 case ABIArgInfo::Ignore: 1088 break; 1089 1090 case ABIArgInfo::Indirect: { 1091 llvm::AttrBuilder SRETAttrs; 1092 SRETAttrs.addAttribute(llvm::Attribute::StructRet); 1093 if (RetAI.getInReg()) 1094 SRETAttrs.addAttribute(llvm::Attribute::InReg); 1095 PAL.push_back(llvm:: 1096 AttributeSet::get(getLLVMContext(), Index, SRETAttrs)); 1097 1098 ++Index; 1099 // sret disables readnone and readonly 1100 FuncAttrs.removeAttribute(llvm::Attribute::ReadOnly) 1101 .removeAttribute(llvm::Attribute::ReadNone); 1102 break; 1103 } 1104 1105 case ABIArgInfo::Expand: 1106 llvm_unreachable("Invalid ABI kind for return argument"); 1107 } 1108 1109 if (RetAttrs.hasAttributes()) 1110 PAL.push_back(llvm:: 1111 AttributeSet::get(getLLVMContext(), 1112 llvm::AttributeSet::ReturnIndex, 1113 RetAttrs)); 1114 1115 for (CGFunctionInfo::const_arg_iterator it = FI.arg_begin(), 1116 ie = FI.arg_end(); it != ie; ++it) { 1117 QualType ParamType = it->type; 1118 const ABIArgInfo &AI = it->info; 1119 llvm::AttrBuilder Attrs; 1120 1121 if (AI.getPaddingType()) { 1122 if (AI.getPaddingInReg()) 1123 PAL.push_back(llvm::AttributeSet::get(getLLVMContext(), Index, 1124 llvm::Attribute::InReg)); 1125 // Increment Index if there is padding. 1126 ++Index; 1127 } 1128 1129 // 'restrict' -> 'noalias' is done in EmitFunctionProlog when we 1130 // have the corresponding parameter variable. It doesn't make 1131 // sense to do it here because parameters are so messed up. 1132 switch (AI.getKind()) { 1133 case ABIArgInfo::Extend: 1134 if (ParamType->isSignedIntegerOrEnumerationType()) 1135 Attrs.addAttribute(llvm::Attribute::SExt); 1136 else if (ParamType->isUnsignedIntegerOrEnumerationType()) 1137 Attrs.addAttribute(llvm::Attribute::ZExt); 1138 // FALL THROUGH 1139 case ABIArgInfo::Direct: 1140 if (AI.getInReg()) 1141 Attrs.addAttribute(llvm::Attribute::InReg); 1142 1143 // FIXME: handle sseregparm someday... 1144 1145 if (llvm::StructType *STy = 1146 dyn_cast<llvm::StructType>(AI.getCoerceToType())) { 1147 unsigned Extra = STy->getNumElements()-1; // 1 will be added below. 1148 if (Attrs.hasAttributes()) 1149 for (unsigned I = 0; I < Extra; ++I) 1150 PAL.push_back(llvm::AttributeSet::get(getLLVMContext(), Index + I, 1151 Attrs)); 1152 Index += Extra; 1153 } 1154 break; 1155 1156 case ABIArgInfo::Indirect: 1157 if (AI.getInReg()) 1158 Attrs.addAttribute(llvm::Attribute::InReg); 1159 1160 if (AI.getIndirectByVal()) 1161 Attrs.addAttribute(llvm::Attribute::ByVal); 1162 1163 Attrs.addAlignmentAttr(AI.getIndirectAlign()); 1164 1165 // byval disables readnone and readonly. 1166 FuncAttrs.removeAttribute(llvm::Attribute::ReadOnly) 1167 .removeAttribute(llvm::Attribute::ReadNone); 1168 break; 1169 1170 case ABIArgInfo::Ignore: 1171 // Skip increment, no matching LLVM parameter. 1172 continue; 1173 1174 case ABIArgInfo::Expand: { 1175 SmallVector<llvm::Type*, 8> types; 1176 // FIXME: This is rather inefficient. Do we ever actually need to do 1177 // anything here? The result should be just reconstructed on the other 1178 // side, so extension should be a non-issue. 1179 getTypes().GetExpandedTypes(ParamType, types); 1180 Index += types.size(); 1181 continue; 1182 } 1183 } 1184 1185 if (Attrs.hasAttributes()) 1186 PAL.push_back(llvm::AttributeSet::get(getLLVMContext(), Index, Attrs)); 1187 ++Index; 1188 } 1189 if (FuncAttrs.hasAttributes()) 1190 PAL.push_back(llvm:: 1191 AttributeSet::get(getLLVMContext(), 1192 llvm::AttributeSet::FunctionIndex, 1193 FuncAttrs)); 1194} 1195 1196/// An argument came in as a promoted argument; demote it back to its 1197/// declared type. 1198static llvm::Value *emitArgumentDemotion(CodeGenFunction &CGF, 1199 const VarDecl *var, 1200 llvm::Value *value) { 1201 llvm::Type *varType = CGF.ConvertType(var->getType()); 1202 1203 // This can happen with promotions that actually don't change the 1204 // underlying type, like the enum promotions. 1205 if (value->getType() == varType) return value; 1206 1207 assert((varType->isIntegerTy() || varType->isFloatingPointTy()) 1208 && "unexpected promotion type"); 1209 1210 if (isa<llvm::IntegerType>(varType)) 1211 return CGF.Builder.CreateTrunc(value, varType, "arg.unpromote"); 1212 1213 return CGF.Builder.CreateFPCast(value, varType, "arg.unpromote"); 1214} 1215 1216void CodeGenFunction::EmitFunctionProlog(const CGFunctionInfo &FI, 1217 llvm::Function *Fn, 1218 const FunctionArgList &Args) { 1219 // If this is an implicit-return-zero function, go ahead and 1220 // initialize the return value. TODO: it might be nice to have 1221 // a more general mechanism for this that didn't require synthesized 1222 // return statements. 1223 if (const FunctionDecl *FD = dyn_cast_or_null<FunctionDecl>(CurCodeDecl)) { 1224 if (FD->hasImplicitReturnZero()) { 1225 QualType RetTy = FD->getResultType().getUnqualifiedType(); 1226 llvm::Type* LLVMTy = CGM.getTypes().ConvertType(RetTy); 1227 llvm::Constant* Zero = llvm::Constant::getNullValue(LLVMTy); 1228 Builder.CreateStore(Zero, ReturnValue); 1229 } 1230 } 1231 1232 // FIXME: We no longer need the types from FunctionArgList; lift up and 1233 // simplify. 1234 1235 // Emit allocs for param decls. Give the LLVM Argument nodes names. 1236 llvm::Function::arg_iterator AI = Fn->arg_begin(); 1237 1238 // Name the struct return argument. 1239 if (CGM.ReturnTypeUsesSRet(FI)) { 1240 AI->setName("agg.result"); 1241 AI->addAttr(llvm::AttributeSet::get(getLLVMContext(), 1242 AI->getArgNo() + 1, 1243 llvm::Attribute::NoAlias)); 1244 ++AI; 1245 } 1246 1247 assert(FI.arg_size() == Args.size() && 1248 "Mismatch between function signature & arguments."); 1249 unsigned ArgNo = 1; 1250 CGFunctionInfo::const_arg_iterator info_it = FI.arg_begin(); 1251 for (FunctionArgList::const_iterator i = Args.begin(), e = Args.end(); 1252 i != e; ++i, ++info_it, ++ArgNo) { 1253 const VarDecl *Arg = *i; 1254 QualType Ty = info_it->type; 1255 const ABIArgInfo &ArgI = info_it->info; 1256 1257 bool isPromoted = 1258 isa<ParmVarDecl>(Arg) && cast<ParmVarDecl>(Arg)->isKNRPromoted(); 1259 1260 // Skip the dummy padding argument. 1261 if (ArgI.getPaddingType()) 1262 ++AI; 1263 1264 switch (ArgI.getKind()) { 1265 case ABIArgInfo::Indirect: { 1266 llvm::Value *V = AI; 1267 1268 if (!hasScalarEvaluationKind(Ty)) { 1269 // Aggregates and complex variables are accessed by reference. All we 1270 // need to do is realign the value, if requested 1271 if (ArgI.getIndirectRealign()) { 1272 llvm::Value *AlignedTemp = CreateMemTemp(Ty, "coerce"); 1273 1274 // Copy from the incoming argument pointer to the temporary with the 1275 // appropriate alignment. 1276 // 1277 // FIXME: We should have a common utility for generating an aggregate 1278 // copy. 1279 llvm::Type *I8PtrTy = Builder.getInt8PtrTy(); 1280 CharUnits Size = getContext().getTypeSizeInChars(Ty); 1281 llvm::Value *Dst = Builder.CreateBitCast(AlignedTemp, I8PtrTy); 1282 llvm::Value *Src = Builder.CreateBitCast(V, I8PtrTy); 1283 Builder.CreateMemCpy(Dst, 1284 Src, 1285 llvm::ConstantInt::get(IntPtrTy, 1286 Size.getQuantity()), 1287 ArgI.getIndirectAlign(), 1288 false); 1289 V = AlignedTemp; 1290 } 1291 } else { 1292 // Load scalar value from indirect argument. 1293 CharUnits Alignment = getContext().getTypeAlignInChars(Ty); 1294 V = EmitLoadOfScalar(V, false, Alignment.getQuantity(), Ty, 1295 Arg->getLocStart()); 1296 1297 if (isPromoted) 1298 V = emitArgumentDemotion(*this, Arg, V); 1299 } 1300 EmitParmDecl(*Arg, V, ArgNo); 1301 break; 1302 } 1303 1304 case ABIArgInfo::Extend: 1305 case ABIArgInfo::Direct: { 1306 1307 // If we have the trivial case, handle it with no muss and fuss. 1308 if (!isa<llvm::StructType>(ArgI.getCoerceToType()) && 1309 ArgI.getCoerceToType() == ConvertType(Ty) && 1310 ArgI.getDirectOffset() == 0) { 1311 assert(AI != Fn->arg_end() && "Argument mismatch!"); 1312 llvm::Value *V = AI; 1313 1314 if (Arg->getType().isRestrictQualified()) 1315 AI->addAttr(llvm::AttributeSet::get(getLLVMContext(), 1316 AI->getArgNo() + 1, 1317 llvm::Attribute::NoAlias)); 1318 1319 // Ensure the argument is the correct type. 1320 if (V->getType() != ArgI.getCoerceToType()) 1321 V = Builder.CreateBitCast(V, ArgI.getCoerceToType()); 1322 1323 if (isPromoted) 1324 V = emitArgumentDemotion(*this, Arg, V); 1325 1326 if (const CXXMethodDecl *MD = 1327 dyn_cast_or_null<CXXMethodDecl>(CurCodeDecl)) { 1328 if (MD->isVirtual() && Arg == CXXABIThisDecl) 1329 V = CGM.getCXXABI(). 1330 adjustThisParameterInVirtualFunctionPrologue(*this, CurGD, V); 1331 } 1332 1333 // Because of merging of function types from multiple decls it is 1334 // possible for the type of an argument to not match the corresponding 1335 // type in the function type. Since we are codegening the callee 1336 // in here, add a cast to the argument type. 1337 llvm::Type *LTy = ConvertType(Arg->getType()); 1338 if (V->getType() != LTy) 1339 V = Builder.CreateBitCast(V, LTy); 1340 1341 EmitParmDecl(*Arg, V, ArgNo); 1342 break; 1343 } 1344 1345 llvm::AllocaInst *Alloca = CreateMemTemp(Ty, Arg->getName()); 1346 1347 // The alignment we need to use is the max of the requested alignment for 1348 // the argument plus the alignment required by our access code below. 1349 unsigned AlignmentToUse = 1350 CGM.getDataLayout().getABITypeAlignment(ArgI.getCoerceToType()); 1351 AlignmentToUse = std::max(AlignmentToUse, 1352 (unsigned)getContext().getDeclAlign(Arg).getQuantity()); 1353 1354 Alloca->setAlignment(AlignmentToUse); 1355 llvm::Value *V = Alloca; 1356 llvm::Value *Ptr = V; // Pointer to store into. 1357 1358 // If the value is offset in memory, apply the offset now. 1359 if (unsigned Offs = ArgI.getDirectOffset()) { 1360 Ptr = Builder.CreateBitCast(Ptr, Builder.getInt8PtrTy()); 1361 Ptr = Builder.CreateConstGEP1_32(Ptr, Offs); 1362 Ptr = Builder.CreateBitCast(Ptr, 1363 llvm::PointerType::getUnqual(ArgI.getCoerceToType())); 1364 } 1365 1366 // If the coerce-to type is a first class aggregate, we flatten it and 1367 // pass the elements. Either way is semantically identical, but fast-isel 1368 // and the optimizer generally likes scalar values better than FCAs. 1369 llvm::StructType *STy = dyn_cast<llvm::StructType>(ArgI.getCoerceToType()); 1370 if (STy && STy->getNumElements() > 1) { 1371 uint64_t SrcSize = CGM.getDataLayout().getTypeAllocSize(STy); 1372 llvm::Type *DstTy = 1373 cast<llvm::PointerType>(Ptr->getType())->getElementType(); 1374 uint64_t DstSize = CGM.getDataLayout().getTypeAllocSize(DstTy); 1375 1376 if (SrcSize <= DstSize) { 1377 Ptr = Builder.CreateBitCast(Ptr, llvm::PointerType::getUnqual(STy)); 1378 1379 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 1380 assert(AI != Fn->arg_end() && "Argument mismatch!"); 1381 AI->setName(Arg->getName() + ".coerce" + Twine(i)); 1382 llvm::Value *EltPtr = Builder.CreateConstGEP2_32(Ptr, 0, i); 1383 Builder.CreateStore(AI++, EltPtr); 1384 } 1385 } else { 1386 llvm::AllocaInst *TempAlloca = 1387 CreateTempAlloca(ArgI.getCoerceToType(), "coerce"); 1388 TempAlloca->setAlignment(AlignmentToUse); 1389 llvm::Value *TempV = TempAlloca; 1390 1391 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 1392 assert(AI != Fn->arg_end() && "Argument mismatch!"); 1393 AI->setName(Arg->getName() + ".coerce" + Twine(i)); 1394 llvm::Value *EltPtr = Builder.CreateConstGEP2_32(TempV, 0, i); 1395 Builder.CreateStore(AI++, EltPtr); 1396 } 1397 1398 Builder.CreateMemCpy(Ptr, TempV, DstSize, AlignmentToUse); 1399 } 1400 } else { 1401 // Simple case, just do a coerced store of the argument into the alloca. 1402 assert(AI != Fn->arg_end() && "Argument mismatch!"); 1403 AI->setName(Arg->getName() + ".coerce"); 1404 CreateCoercedStore(AI++, Ptr, /*DestIsVolatile=*/false, *this); 1405 } 1406 1407 1408 // Match to what EmitParmDecl is expecting for this type. 1409 if (CodeGenFunction::hasScalarEvaluationKind(Ty)) { 1410 V = EmitLoadOfScalar(V, false, AlignmentToUse, Ty, Arg->getLocStart()); 1411 if (isPromoted) 1412 V = emitArgumentDemotion(*this, Arg, V); 1413 } 1414 EmitParmDecl(*Arg, V, ArgNo); 1415 continue; // Skip ++AI increment, already done. 1416 } 1417 1418 case ABIArgInfo::Expand: { 1419 // If this structure was expanded into multiple arguments then 1420 // we need to create a temporary and reconstruct it from the 1421 // arguments. 1422 llvm::AllocaInst *Alloca = CreateMemTemp(Ty); 1423 CharUnits Align = getContext().getDeclAlign(Arg); 1424 Alloca->setAlignment(Align.getQuantity()); 1425 LValue LV = MakeAddrLValue(Alloca, Ty, Align); 1426 llvm::Function::arg_iterator End = ExpandTypeFromArgs(Ty, LV, AI); 1427 EmitParmDecl(*Arg, Alloca, ArgNo); 1428 1429 // Name the arguments used in expansion and increment AI. 1430 unsigned Index = 0; 1431 for (; AI != End; ++AI, ++Index) 1432 AI->setName(Arg->getName() + "." + Twine(Index)); 1433 continue; 1434 } 1435 1436 case ABIArgInfo::Ignore: 1437 // Initialize the local variable appropriately. 1438 if (!hasScalarEvaluationKind(Ty)) 1439 EmitParmDecl(*Arg, CreateMemTemp(Ty), ArgNo); 1440 else 1441 EmitParmDecl(*Arg, llvm::UndefValue::get(ConvertType(Arg->getType())), 1442 ArgNo); 1443 1444 // Skip increment, no matching LLVM parameter. 1445 continue; 1446 } 1447 1448 ++AI; 1449 } 1450 assert(AI == Fn->arg_end() && "Argument mismatch!"); 1451} 1452 1453static void eraseUnusedBitCasts(llvm::Instruction *insn) { 1454 while (insn->use_empty()) { 1455 llvm::BitCastInst *bitcast = dyn_cast<llvm::BitCastInst>(insn); 1456 if (!bitcast) return; 1457 1458 // This is "safe" because we would have used a ConstantExpr otherwise. 1459 insn = cast<llvm::Instruction>(bitcast->getOperand(0)); 1460 bitcast->eraseFromParent(); 1461 } 1462} 1463 1464/// Try to emit a fused autorelease of a return result. 1465static llvm::Value *tryEmitFusedAutoreleaseOfResult(CodeGenFunction &CGF, 1466 llvm::Value *result) { 1467 // We must be immediately followed the cast. 1468 llvm::BasicBlock *BB = CGF.Builder.GetInsertBlock(); 1469 if (BB->empty()) return 0; 1470 if (&BB->back() != result) return 0; 1471 1472 llvm::Type *resultType = result->getType(); 1473 1474 // result is in a BasicBlock and is therefore an Instruction. 1475 llvm::Instruction *generator = cast<llvm::Instruction>(result); 1476 1477 SmallVector<llvm::Instruction*,4> insnsToKill; 1478 1479 // Look for: 1480 // %generator = bitcast %type1* %generator2 to %type2* 1481 while (llvm::BitCastInst *bitcast = dyn_cast<llvm::BitCastInst>(generator)) { 1482 // We would have emitted this as a constant if the operand weren't 1483 // an Instruction. 1484 generator = cast<llvm::Instruction>(bitcast->getOperand(0)); 1485 1486 // Require the generator to be immediately followed by the cast. 1487 if (generator->getNextNode() != bitcast) 1488 return 0; 1489 1490 insnsToKill.push_back(bitcast); 1491 } 1492 1493 // Look for: 1494 // %generator = call i8* @objc_retain(i8* %originalResult) 1495 // or 1496 // %generator = call i8* @objc_retainAutoreleasedReturnValue(i8* %originalResult) 1497 llvm::CallInst *call = dyn_cast<llvm::CallInst>(generator); 1498 if (!call) return 0; 1499 1500 bool doRetainAutorelease; 1501 1502 if (call->getCalledValue() == CGF.CGM.getARCEntrypoints().objc_retain) { 1503 doRetainAutorelease = true; 1504 } else if (call->getCalledValue() == CGF.CGM.getARCEntrypoints() 1505 .objc_retainAutoreleasedReturnValue) { 1506 doRetainAutorelease = false; 1507 1508 // If we emitted an assembly marker for this call (and the 1509 // ARCEntrypoints field should have been set if so), go looking 1510 // for that call. If we can't find it, we can't do this 1511 // optimization. But it should always be the immediately previous 1512 // instruction, unless we needed bitcasts around the call. 1513 if (CGF.CGM.getARCEntrypoints().retainAutoreleasedReturnValueMarker) { 1514 llvm::Instruction *prev = call->getPrevNode(); 1515 assert(prev); 1516 if (isa<llvm::BitCastInst>(prev)) { 1517 prev = prev->getPrevNode(); 1518 assert(prev); 1519 } 1520 assert(isa<llvm::CallInst>(prev)); 1521 assert(cast<llvm::CallInst>(prev)->getCalledValue() == 1522 CGF.CGM.getARCEntrypoints().retainAutoreleasedReturnValueMarker); 1523 insnsToKill.push_back(prev); 1524 } 1525 } else { 1526 return 0; 1527 } 1528 1529 result = call->getArgOperand(0); 1530 insnsToKill.push_back(call); 1531 1532 // Keep killing bitcasts, for sanity. Note that we no longer care 1533 // about precise ordering as long as there's exactly one use. 1534 while (llvm::BitCastInst *bitcast = dyn_cast<llvm::BitCastInst>(result)) { 1535 if (!bitcast->hasOneUse()) break; 1536 insnsToKill.push_back(bitcast); 1537 result = bitcast->getOperand(0); 1538 } 1539 1540 // Delete all the unnecessary instructions, from latest to earliest. 1541 for (SmallVectorImpl<llvm::Instruction*>::iterator 1542 i = insnsToKill.begin(), e = insnsToKill.end(); i != e; ++i) 1543 (*i)->eraseFromParent(); 1544 1545 // Do the fused retain/autorelease if we were asked to. 1546 if (doRetainAutorelease) 1547 result = CGF.EmitARCRetainAutoreleaseReturnValue(result); 1548 1549 // Cast back to the result type. 1550 return CGF.Builder.CreateBitCast(result, resultType); 1551} 1552 1553/// If this is a +1 of the value of an immutable 'self', remove it. 1554static llvm::Value *tryRemoveRetainOfSelf(CodeGenFunction &CGF, 1555 llvm::Value *result) { 1556 // This is only applicable to a method with an immutable 'self'. 1557 const ObjCMethodDecl *method = 1558 dyn_cast_or_null<ObjCMethodDecl>(CGF.CurCodeDecl); 1559 if (!method) return 0; 1560 const VarDecl *self = method->getSelfDecl(); 1561 if (!self->getType().isConstQualified()) return 0; 1562 1563 // Look for a retain call. 1564 llvm::CallInst *retainCall = 1565 dyn_cast<llvm::CallInst>(result->stripPointerCasts()); 1566 if (!retainCall || 1567 retainCall->getCalledValue() != CGF.CGM.getARCEntrypoints().objc_retain) 1568 return 0; 1569 1570 // Look for an ordinary load of 'self'. 1571 llvm::Value *retainedValue = retainCall->getArgOperand(0); 1572 llvm::LoadInst *load = 1573 dyn_cast<llvm::LoadInst>(retainedValue->stripPointerCasts()); 1574 if (!load || load->isAtomic() || load->isVolatile() || 1575 load->getPointerOperand() != CGF.GetAddrOfLocalVar(self)) 1576 return 0; 1577 1578 // Okay! Burn it all down. This relies for correctness on the 1579 // assumption that the retain is emitted as part of the return and 1580 // that thereafter everything is used "linearly". 1581 llvm::Type *resultType = result->getType(); 1582 eraseUnusedBitCasts(cast<llvm::Instruction>(result)); 1583 assert(retainCall->use_empty()); 1584 retainCall->eraseFromParent(); 1585 eraseUnusedBitCasts(cast<llvm::Instruction>(retainedValue)); 1586 1587 return CGF.Builder.CreateBitCast(load, resultType); 1588} 1589 1590/// Emit an ARC autorelease of the result of a function. 1591/// 1592/// \return the value to actually return from the function 1593static llvm::Value *emitAutoreleaseOfResult(CodeGenFunction &CGF, 1594 llvm::Value *result) { 1595 // If we're returning 'self', kill the initial retain. This is a 1596 // heuristic attempt to "encourage correctness" in the really unfortunate 1597 // case where we have a return of self during a dealloc and we desperately 1598 // need to avoid the possible autorelease. 1599 if (llvm::Value *self = tryRemoveRetainOfSelf(CGF, result)) 1600 return self; 1601 1602 // At -O0, try to emit a fused retain/autorelease. 1603 if (CGF.shouldUseFusedARCCalls()) 1604 if (llvm::Value *fused = tryEmitFusedAutoreleaseOfResult(CGF, result)) 1605 return fused; 1606 1607 return CGF.EmitARCAutoreleaseReturnValue(result); 1608} 1609 1610/// Heuristically search for a dominating store to the return-value slot. 1611static llvm::StoreInst *findDominatingStoreToReturnValue(CodeGenFunction &CGF) { 1612 // If there are multiple uses of the return-value slot, just check 1613 // for something immediately preceding the IP. Sometimes this can 1614 // happen with how we generate implicit-returns; it can also happen 1615 // with noreturn cleanups. 1616 if (!CGF.ReturnValue->hasOneUse()) { 1617 llvm::BasicBlock *IP = CGF.Builder.GetInsertBlock(); 1618 if (IP->empty()) return 0; 1619 llvm::StoreInst *store = dyn_cast<llvm::StoreInst>(&IP->back()); 1620 if (!store) return 0; 1621 if (store->getPointerOperand() != CGF.ReturnValue) return 0; 1622 assert(!store->isAtomic() && !store->isVolatile()); // see below 1623 return store; 1624 } 1625 1626 llvm::StoreInst *store = 1627 dyn_cast<llvm::StoreInst>(CGF.ReturnValue->use_back()); 1628 if (!store) return 0; 1629 1630 // These aren't actually possible for non-coerced returns, and we 1631 // only care about non-coerced returns on this code path. 1632 assert(!store->isAtomic() && !store->isVolatile()); 1633 1634 // Now do a first-and-dirty dominance check: just walk up the 1635 // single-predecessors chain from the current insertion point. 1636 llvm::BasicBlock *StoreBB = store->getParent(); 1637 llvm::BasicBlock *IP = CGF.Builder.GetInsertBlock(); 1638 while (IP != StoreBB) { 1639 if (!(IP = IP->getSinglePredecessor())) 1640 return 0; 1641 } 1642 1643 // Okay, the store's basic block dominates the insertion point; we 1644 // can do our thing. 1645 return store; 1646} 1647 1648void CodeGenFunction::EmitFunctionEpilog(const CGFunctionInfo &FI, 1649 bool EmitRetDbgLoc, 1650 SourceLocation EndLoc) { 1651 // Functions with no result always return void. 1652 if (ReturnValue == 0) { 1653 Builder.CreateRetVoid(); 1654 return; 1655 } 1656 1657 llvm::DebugLoc RetDbgLoc; 1658 llvm::Value *RV = 0; 1659 QualType RetTy = FI.getReturnType(); 1660 const ABIArgInfo &RetAI = FI.getReturnInfo(); 1661 1662 switch (RetAI.getKind()) { 1663 case ABIArgInfo::Indirect: { 1664 switch (getEvaluationKind(RetTy)) { 1665 case TEK_Complex: { 1666 ComplexPairTy RT = 1667 EmitLoadOfComplex(MakeNaturalAlignAddrLValue(ReturnValue, RetTy), 1668 EndLoc); 1669 EmitStoreOfComplex(RT, 1670 MakeNaturalAlignAddrLValue(CurFn->arg_begin(), RetTy), 1671 /*isInit*/ true); 1672 break; 1673 } 1674 case TEK_Aggregate: 1675 // Do nothing; aggregrates get evaluated directly into the destination. 1676 break; 1677 case TEK_Scalar: 1678 EmitStoreOfScalar(Builder.CreateLoad(ReturnValue), 1679 MakeNaturalAlignAddrLValue(CurFn->arg_begin(), RetTy), 1680 /*isInit*/ true); 1681 break; 1682 } 1683 break; 1684 } 1685 1686 case ABIArgInfo::Extend: 1687 case ABIArgInfo::Direct: 1688 if (RetAI.getCoerceToType() == ConvertType(RetTy) && 1689 RetAI.getDirectOffset() == 0) { 1690 // The internal return value temp always will have pointer-to-return-type 1691 // type, just do a load. 1692 1693 // If there is a dominating store to ReturnValue, we can elide 1694 // the load, zap the store, and usually zap the alloca. 1695 if (llvm::StoreInst *SI = findDominatingStoreToReturnValue(*this)) { 1696 // Reuse the debug location from the store unless there is 1697 // cleanup code to be emitted between the store and return 1698 // instruction. 1699 if (EmitRetDbgLoc && !AutoreleaseResult) 1700 RetDbgLoc = SI->getDebugLoc(); 1701 // Get the stored value and nuke the now-dead store. 1702 RV = SI->getValueOperand(); 1703 SI->eraseFromParent(); 1704 1705 // If that was the only use of the return value, nuke it as well now. 1706 if (ReturnValue->use_empty() && isa<llvm::AllocaInst>(ReturnValue)) { 1707 cast<llvm::AllocaInst>(ReturnValue)->eraseFromParent(); 1708 ReturnValue = 0; 1709 } 1710 1711 // Otherwise, we have to do a simple load. 1712 } else { 1713 RV = Builder.CreateLoad(ReturnValue); 1714 } 1715 } else { 1716 llvm::Value *V = ReturnValue; 1717 // If the value is offset in memory, apply the offset now. 1718 if (unsigned Offs = RetAI.getDirectOffset()) { 1719 V = Builder.CreateBitCast(V, Builder.getInt8PtrTy()); 1720 V = Builder.CreateConstGEP1_32(V, Offs); 1721 V = Builder.CreateBitCast(V, 1722 llvm::PointerType::getUnqual(RetAI.getCoerceToType())); 1723 } 1724 1725 RV = CreateCoercedLoad(V, RetAI.getCoerceToType(), *this); 1726 } 1727 1728 // In ARC, end functions that return a retainable type with a call 1729 // to objc_autoreleaseReturnValue. 1730 if (AutoreleaseResult) { 1731 assert(getLangOpts().ObjCAutoRefCount && 1732 !FI.isReturnsRetained() && 1733 RetTy->isObjCRetainableType()); 1734 RV = emitAutoreleaseOfResult(*this, RV); 1735 } 1736 1737 break; 1738 1739 case ABIArgInfo::Ignore: 1740 break; 1741 1742 case ABIArgInfo::Expand: 1743 llvm_unreachable("Invalid ABI kind for return argument"); 1744 } 1745 1746 llvm::Instruction *Ret = RV ? Builder.CreateRet(RV) : Builder.CreateRetVoid(); 1747 if (!RetDbgLoc.isUnknown()) 1748 Ret->setDebugLoc(RetDbgLoc); 1749} 1750 1751void CodeGenFunction::EmitDelegateCallArg(CallArgList &args, 1752 const VarDecl *param, 1753 SourceLocation loc) { 1754 // StartFunction converted the ABI-lowered parameter(s) into a 1755 // local alloca. We need to turn that into an r-value suitable 1756 // for EmitCall. 1757 llvm::Value *local = GetAddrOfLocalVar(param); 1758 1759 QualType type = param->getType(); 1760 1761 // For the most part, we just need to load the alloca, except: 1762 // 1) aggregate r-values are actually pointers to temporaries, and 1763 // 2) references to non-scalars are pointers directly to the aggregate. 1764 // I don't know why references to scalars are different here. 1765 if (const ReferenceType *ref = type->getAs<ReferenceType>()) { 1766 if (!hasScalarEvaluationKind(ref->getPointeeType())) 1767 return args.add(RValue::getAggregate(local), type); 1768 1769 // Locals which are references to scalars are represented 1770 // with allocas holding the pointer. 1771 return args.add(RValue::get(Builder.CreateLoad(local)), type); 1772 } 1773 1774 args.add(convertTempToRValue(local, type, loc), type); 1775} 1776 1777static bool isProvablyNull(llvm::Value *addr) { 1778 return isa<llvm::ConstantPointerNull>(addr); 1779} 1780 1781static bool isProvablyNonNull(llvm::Value *addr) { 1782 return isa<llvm::AllocaInst>(addr); 1783} 1784 1785/// Emit the actual writing-back of a writeback. 1786static void emitWriteback(CodeGenFunction &CGF, 1787 const CallArgList::Writeback &writeback) { 1788 const LValue &srcLV = writeback.Source; 1789 llvm::Value *srcAddr = srcLV.getAddress(); 1790 assert(!isProvablyNull(srcAddr) && 1791 "shouldn't have writeback for provably null argument"); 1792 1793 llvm::BasicBlock *contBB = 0; 1794 1795 // If the argument wasn't provably non-null, we need to null check 1796 // before doing the store. 1797 bool provablyNonNull = isProvablyNonNull(srcAddr); 1798 if (!provablyNonNull) { 1799 llvm::BasicBlock *writebackBB = CGF.createBasicBlock("icr.writeback"); 1800 contBB = CGF.createBasicBlock("icr.done"); 1801 1802 llvm::Value *isNull = CGF.Builder.CreateIsNull(srcAddr, "icr.isnull"); 1803 CGF.Builder.CreateCondBr(isNull, contBB, writebackBB); 1804 CGF.EmitBlock(writebackBB); 1805 } 1806 1807 // Load the value to writeback. 1808 llvm::Value *value = CGF.Builder.CreateLoad(writeback.Temporary); 1809 1810 // Cast it back, in case we're writing an id to a Foo* or something. 1811 value = CGF.Builder.CreateBitCast(value, 1812 cast<llvm::PointerType>(srcAddr->getType())->getElementType(), 1813 "icr.writeback-cast"); 1814 1815 // Perform the writeback. 1816 1817 // If we have a "to use" value, it's something we need to emit a use 1818 // of. This has to be carefully threaded in: if it's done after the 1819 // release it's potentially undefined behavior (and the optimizer 1820 // will ignore it), and if it happens before the retain then the 1821 // optimizer could move the release there. 1822 if (writeback.ToUse) { 1823 assert(srcLV.getObjCLifetime() == Qualifiers::OCL_Strong); 1824 1825 // Retain the new value. No need to block-copy here: the block's 1826 // being passed up the stack. 1827 value = CGF.EmitARCRetainNonBlock(value); 1828 1829 // Emit the intrinsic use here. 1830 CGF.EmitARCIntrinsicUse(writeback.ToUse); 1831 1832 // Load the old value (primitively). 1833 llvm::Value *oldValue = CGF.EmitLoadOfScalar(srcLV, SourceLocation()); 1834 1835 // Put the new value in place (primitively). 1836 CGF.EmitStoreOfScalar(value, srcLV, /*init*/ false); 1837 1838 // Release the old value. 1839 CGF.EmitARCRelease(oldValue, srcLV.isARCPreciseLifetime()); 1840 1841 // Otherwise, we can just do a normal lvalue store. 1842 } else { 1843 CGF.EmitStoreThroughLValue(RValue::get(value), srcLV); 1844 } 1845 1846 // Jump to the continuation block. 1847 if (!provablyNonNull) 1848 CGF.EmitBlock(contBB); 1849} 1850 1851static void emitWritebacks(CodeGenFunction &CGF, 1852 const CallArgList &args) { 1853 for (CallArgList::writeback_iterator 1854 i = args.writeback_begin(), e = args.writeback_end(); i != e; ++i) 1855 emitWriteback(CGF, *i); 1856} 1857 1858static void deactivateArgCleanupsBeforeCall(CodeGenFunction &CGF, 1859 const CallArgList &CallArgs) { 1860 assert(CGF.getTarget().getCXXABI().isArgumentDestroyedByCallee()); 1861 ArrayRef<CallArgList::CallArgCleanup> Cleanups = 1862 CallArgs.getCleanupsToDeactivate(); 1863 // Iterate in reverse to increase the likelihood of popping the cleanup. 1864 for (ArrayRef<CallArgList::CallArgCleanup>::reverse_iterator 1865 I = Cleanups.rbegin(), E = Cleanups.rend(); I != E; ++I) { 1866 CGF.DeactivateCleanupBlock(I->Cleanup, I->IsActiveIP); 1867 I->IsActiveIP->eraseFromParent(); 1868 } 1869} 1870 1871static const Expr *maybeGetUnaryAddrOfOperand(const Expr *E) { 1872 if (const UnaryOperator *uop = dyn_cast<UnaryOperator>(E->IgnoreParens())) 1873 if (uop->getOpcode() == UO_AddrOf) 1874 return uop->getSubExpr(); 1875 return 0; 1876} 1877 1878/// Emit an argument that's being passed call-by-writeback. That is, 1879/// we are passing the address of 1880static void emitWritebackArg(CodeGenFunction &CGF, CallArgList &args, 1881 const ObjCIndirectCopyRestoreExpr *CRE) { 1882 LValue srcLV; 1883 1884 // Make an optimistic effort to emit the address as an l-value. 1885 // This can fail if the the argument expression is more complicated. 1886 if (const Expr *lvExpr = maybeGetUnaryAddrOfOperand(CRE->getSubExpr())) { 1887 srcLV = CGF.EmitLValue(lvExpr); 1888 1889 // Otherwise, just emit it as a scalar. 1890 } else { 1891 llvm::Value *srcAddr = CGF.EmitScalarExpr(CRE->getSubExpr()); 1892 1893 QualType srcAddrType = 1894 CRE->getSubExpr()->getType()->castAs<PointerType>()->getPointeeType(); 1895 srcLV = CGF.MakeNaturalAlignAddrLValue(srcAddr, srcAddrType); 1896 } 1897 llvm::Value *srcAddr = srcLV.getAddress(); 1898 1899 // The dest and src types don't necessarily match in LLVM terms 1900 // because of the crazy ObjC compatibility rules. 1901 1902 llvm::PointerType *destType = 1903 cast<llvm::PointerType>(CGF.ConvertType(CRE->getType())); 1904 1905 // If the address is a constant null, just pass the appropriate null. 1906 if (isProvablyNull(srcAddr)) { 1907 args.add(RValue::get(llvm::ConstantPointerNull::get(destType)), 1908 CRE->getType()); 1909 return; 1910 } 1911 1912 // Create the temporary. 1913 llvm::Value *temp = CGF.CreateTempAlloca(destType->getElementType(), 1914 "icr.temp"); 1915 // Loading an l-value can introduce a cleanup if the l-value is __weak, 1916 // and that cleanup will be conditional if we can't prove that the l-value 1917 // isn't null, so we need to register a dominating point so that the cleanups 1918 // system will make valid IR. 1919 CodeGenFunction::ConditionalEvaluation condEval(CGF); 1920 1921 // Zero-initialize it if we're not doing a copy-initialization. 1922 bool shouldCopy = CRE->shouldCopy(); 1923 if (!shouldCopy) { 1924 llvm::Value *null = 1925 llvm::ConstantPointerNull::get( 1926 cast<llvm::PointerType>(destType->getElementType())); 1927 CGF.Builder.CreateStore(null, temp); 1928 } 1929 1930 llvm::BasicBlock *contBB = 0; 1931 llvm::BasicBlock *originBB = 0; 1932 1933 // If the address is *not* known to be non-null, we need to switch. 1934 llvm::Value *finalArgument; 1935 1936 bool provablyNonNull = isProvablyNonNull(srcAddr); 1937 if (provablyNonNull) { 1938 finalArgument = temp; 1939 } else { 1940 llvm::Value *isNull = CGF.Builder.CreateIsNull(srcAddr, "icr.isnull"); 1941 1942 finalArgument = CGF.Builder.CreateSelect(isNull, 1943 llvm::ConstantPointerNull::get(destType), 1944 temp, "icr.argument"); 1945 1946 // If we need to copy, then the load has to be conditional, which 1947 // means we need control flow. 1948 if (shouldCopy) { 1949 originBB = CGF.Builder.GetInsertBlock(); 1950 contBB = CGF.createBasicBlock("icr.cont"); 1951 llvm::BasicBlock *copyBB = CGF.createBasicBlock("icr.copy"); 1952 CGF.Builder.CreateCondBr(isNull, contBB, copyBB); 1953 CGF.EmitBlock(copyBB); 1954 condEval.begin(CGF); 1955 } 1956 } 1957 1958 llvm::Value *valueToUse = 0; 1959 1960 // Perform a copy if necessary. 1961 if (shouldCopy) { 1962 RValue srcRV = CGF.EmitLoadOfLValue(srcLV, SourceLocation()); 1963 assert(srcRV.isScalar()); 1964 1965 llvm::Value *src = srcRV.getScalarVal(); 1966 src = CGF.Builder.CreateBitCast(src, destType->getElementType(), 1967 "icr.cast"); 1968 1969 // Use an ordinary store, not a store-to-lvalue. 1970 CGF.Builder.CreateStore(src, temp); 1971 1972 // If optimization is enabled, and the value was held in a 1973 // __strong variable, we need to tell the optimizer that this 1974 // value has to stay alive until we're doing the store back. 1975 // This is because the temporary is effectively unretained, 1976 // and so otherwise we can violate the high-level semantics. 1977 if (CGF.CGM.getCodeGenOpts().OptimizationLevel != 0 && 1978 srcLV.getObjCLifetime() == Qualifiers::OCL_Strong) { 1979 valueToUse = src; 1980 } 1981 } 1982 1983 // Finish the control flow if we needed it. 1984 if (shouldCopy && !provablyNonNull) { 1985 llvm::BasicBlock *copyBB = CGF.Builder.GetInsertBlock(); 1986 CGF.EmitBlock(contBB); 1987 1988 // Make a phi for the value to intrinsically use. 1989 if (valueToUse) { 1990 llvm::PHINode *phiToUse = CGF.Builder.CreatePHI(valueToUse->getType(), 2, 1991 "icr.to-use"); 1992 phiToUse->addIncoming(valueToUse, copyBB); 1993 phiToUse->addIncoming(llvm::UndefValue::get(valueToUse->getType()), 1994 originBB); 1995 valueToUse = phiToUse; 1996 } 1997 1998 condEval.end(CGF); 1999 } 2000 2001 args.addWriteback(srcLV, temp, valueToUse); 2002 args.add(RValue::get(finalArgument), CRE->getType()); 2003} 2004 2005void CodeGenFunction::EmitCallArg(CallArgList &args, const Expr *E, 2006 QualType type) { 2007 if (const ObjCIndirectCopyRestoreExpr *CRE 2008 = dyn_cast<ObjCIndirectCopyRestoreExpr>(E)) { 2009 assert(getLangOpts().ObjCAutoRefCount); 2010 assert(getContext().hasSameType(E->getType(), type)); 2011 return emitWritebackArg(*this, args, CRE); 2012 } 2013 2014 assert(type->isReferenceType() == E->isGLValue() && 2015 "reference binding to unmaterialized r-value!"); 2016 2017 if (E->isGLValue()) { 2018 assert(E->getObjectKind() == OK_Ordinary); 2019 return args.add(EmitReferenceBindingToExpr(E), type); 2020 } 2021 2022 bool HasAggregateEvalKind = hasAggregateEvaluationKind(type); 2023 2024 // In the Microsoft C++ ABI, aggregate arguments are destructed by the callee. 2025 // However, we still have to push an EH-only cleanup in case we unwind before 2026 // we make it to the call. 2027 if (HasAggregateEvalKind && 2028 CGM.getTarget().getCXXABI().isArgumentDestroyedByCallee()) { 2029 const CXXRecordDecl *RD = type->getAsCXXRecordDecl(); 2030 if (RD && RD->hasNonTrivialDestructor()) { 2031 AggValueSlot Slot = CreateAggTemp(type, "agg.arg.tmp"); 2032 Slot.setExternallyDestructed(); 2033 EmitAggExpr(E, Slot); 2034 RValue RV = Slot.asRValue(); 2035 args.add(RV, type); 2036 2037 pushDestroy(EHCleanup, RV.getAggregateAddr(), type, destroyCXXObject, 2038 /*useEHCleanupForArray*/ true); 2039 // This unreachable is a temporary marker which will be removed later. 2040 llvm::Instruction *IsActive = Builder.CreateUnreachable(); 2041 args.addArgCleanupDeactivation(EHStack.getInnermostEHScope(), IsActive); 2042 return; 2043 } 2044 } 2045 2046 if (HasAggregateEvalKind && isa<ImplicitCastExpr>(E) && 2047 cast<CastExpr>(E)->getCastKind() == CK_LValueToRValue) { 2048 LValue L = EmitLValue(cast<CastExpr>(E)->getSubExpr()); 2049 assert(L.isSimple()); 2050 if (L.getAlignment() >= getContext().getTypeAlignInChars(type)) { 2051 args.add(L.asAggregateRValue(), type, /*NeedsCopy*/true); 2052 } else { 2053 // We can't represent a misaligned lvalue in the CallArgList, so copy 2054 // to an aligned temporary now. 2055 llvm::Value *tmp = CreateMemTemp(type); 2056 EmitAggregateCopy(tmp, L.getAddress(), type, L.isVolatile(), 2057 L.getAlignment()); 2058 args.add(RValue::getAggregate(tmp), type); 2059 } 2060 return; 2061 } 2062 2063 args.add(EmitAnyExprToTemp(E), type); 2064} 2065 2066// In ObjC ARC mode with no ObjC ARC exception safety, tell the ARC 2067// optimizer it can aggressively ignore unwind edges. 2068void 2069CodeGenFunction::AddObjCARCExceptionMetadata(llvm::Instruction *Inst) { 2070 if (CGM.getCodeGenOpts().OptimizationLevel != 0 && 2071 !CGM.getCodeGenOpts().ObjCAutoRefCountExceptions) 2072 Inst->setMetadata("clang.arc.no_objc_arc_exceptions", 2073 CGM.getNoObjCARCExceptionsMetadata()); 2074} 2075 2076/// Emits a call to the given no-arguments nounwind runtime function. 2077llvm::CallInst * 2078CodeGenFunction::EmitNounwindRuntimeCall(llvm::Value *callee, 2079 const llvm::Twine &name) { 2080 return EmitNounwindRuntimeCall(callee, ArrayRef<llvm::Value*>(), name); 2081} 2082 2083/// Emits a call to the given nounwind runtime function. 2084llvm::CallInst * 2085CodeGenFunction::EmitNounwindRuntimeCall(llvm::Value *callee, 2086 ArrayRef<llvm::Value*> args, 2087 const llvm::Twine &name) { 2088 llvm::CallInst *call = EmitRuntimeCall(callee, args, name); 2089 call->setDoesNotThrow(); 2090 return call; 2091} 2092 2093/// Emits a simple call (never an invoke) to the given no-arguments 2094/// runtime function. 2095llvm::CallInst * 2096CodeGenFunction::EmitRuntimeCall(llvm::Value *callee, 2097 const llvm::Twine &name) { 2098 return EmitRuntimeCall(callee, ArrayRef<llvm::Value*>(), name); 2099} 2100 2101/// Emits a simple call (never an invoke) to the given runtime 2102/// function. 2103llvm::CallInst * 2104CodeGenFunction::EmitRuntimeCall(llvm::Value *callee, 2105 ArrayRef<llvm::Value*> args, 2106 const llvm::Twine &name) { 2107 llvm::CallInst *call = Builder.CreateCall(callee, args, name); 2108 call->setCallingConv(getRuntimeCC()); 2109 return call; 2110} 2111 2112/// Emits a call or invoke to the given noreturn runtime function. 2113void CodeGenFunction::EmitNoreturnRuntimeCallOrInvoke(llvm::Value *callee, 2114 ArrayRef<llvm::Value*> args) { 2115 if (getInvokeDest()) { 2116 llvm::InvokeInst *invoke = 2117 Builder.CreateInvoke(callee, 2118 getUnreachableBlock(), 2119 getInvokeDest(), 2120 args); 2121 invoke->setDoesNotReturn(); 2122 invoke->setCallingConv(getRuntimeCC()); 2123 } else { 2124 llvm::CallInst *call = Builder.CreateCall(callee, args); 2125 call->setDoesNotReturn(); 2126 call->setCallingConv(getRuntimeCC()); 2127 Builder.CreateUnreachable(); 2128 } 2129} 2130 2131/// Emits a call or invoke instruction to the given nullary runtime 2132/// function. 2133llvm::CallSite 2134CodeGenFunction::EmitRuntimeCallOrInvoke(llvm::Value *callee, 2135 const Twine &name) { 2136 return EmitRuntimeCallOrInvoke(callee, ArrayRef<llvm::Value*>(), name); 2137} 2138 2139/// Emits a call or invoke instruction to the given runtime function. 2140llvm::CallSite 2141CodeGenFunction::EmitRuntimeCallOrInvoke(llvm::Value *callee, 2142 ArrayRef<llvm::Value*> args, 2143 const Twine &name) { 2144 llvm::CallSite callSite = EmitCallOrInvoke(callee, args, name); 2145 callSite.setCallingConv(getRuntimeCC()); 2146 return callSite; 2147} 2148 2149llvm::CallSite 2150CodeGenFunction::EmitCallOrInvoke(llvm::Value *Callee, 2151 const Twine &Name) { 2152 return EmitCallOrInvoke(Callee, ArrayRef<llvm::Value *>(), Name); 2153} 2154 2155/// Emits a call or invoke instruction to the given function, depending 2156/// on the current state of the EH stack. 2157llvm::CallSite 2158CodeGenFunction::EmitCallOrInvoke(llvm::Value *Callee, 2159 ArrayRef<llvm::Value *> Args, 2160 const Twine &Name) { 2161 llvm::BasicBlock *InvokeDest = getInvokeDest(); 2162 2163 llvm::Instruction *Inst; 2164 if (!InvokeDest) 2165 Inst = Builder.CreateCall(Callee, Args, Name); 2166 else { 2167 llvm::BasicBlock *ContBB = createBasicBlock("invoke.cont"); 2168 Inst = Builder.CreateInvoke(Callee, ContBB, InvokeDest, Args, Name); 2169 EmitBlock(ContBB); 2170 } 2171 2172 // In ObjC ARC mode with no ObjC ARC exception safety, tell the ARC 2173 // optimizer it can aggressively ignore unwind edges. 2174 if (CGM.getLangOpts().ObjCAutoRefCount) 2175 AddObjCARCExceptionMetadata(Inst); 2176 2177 return Inst; 2178} 2179 2180static void checkArgMatches(llvm::Value *Elt, unsigned &ArgNo, 2181 llvm::FunctionType *FTy) { 2182 if (ArgNo < FTy->getNumParams()) 2183 assert(Elt->getType() == FTy->getParamType(ArgNo)); 2184 else 2185 assert(FTy->isVarArg()); 2186 ++ArgNo; 2187} 2188 2189void CodeGenFunction::ExpandTypeToArgs(QualType Ty, RValue RV, 2190 SmallVectorImpl<llvm::Value *> &Args, 2191 llvm::FunctionType *IRFuncTy) { 2192 if (const ConstantArrayType *AT = getContext().getAsConstantArrayType(Ty)) { 2193 unsigned NumElts = AT->getSize().getZExtValue(); 2194 QualType EltTy = AT->getElementType(); 2195 llvm::Value *Addr = RV.getAggregateAddr(); 2196 for (unsigned Elt = 0; Elt < NumElts; ++Elt) { 2197 llvm::Value *EltAddr = Builder.CreateConstGEP2_32(Addr, 0, Elt); 2198 RValue EltRV = convertTempToRValue(EltAddr, EltTy, SourceLocation()); 2199 ExpandTypeToArgs(EltTy, EltRV, Args, IRFuncTy); 2200 } 2201 } else if (const RecordType *RT = Ty->getAs<RecordType>()) { 2202 RecordDecl *RD = RT->getDecl(); 2203 assert(RV.isAggregate() && "Unexpected rvalue during struct expansion"); 2204 LValue LV = MakeAddrLValue(RV.getAggregateAddr(), Ty); 2205 2206 if (RD->isUnion()) { 2207 const FieldDecl *LargestFD = 0; 2208 CharUnits UnionSize = CharUnits::Zero(); 2209 2210 for (RecordDecl::field_iterator i = RD->field_begin(), e = RD->field_end(); 2211 i != e; ++i) { 2212 const FieldDecl *FD = *i; 2213 assert(!FD->isBitField() && 2214 "Cannot expand structure with bit-field members."); 2215 CharUnits FieldSize = getContext().getTypeSizeInChars(FD->getType()); 2216 if (UnionSize < FieldSize) { 2217 UnionSize = FieldSize; 2218 LargestFD = FD; 2219 } 2220 } 2221 if (LargestFD) { 2222 RValue FldRV = EmitRValueForField(LV, LargestFD, SourceLocation()); 2223 ExpandTypeToArgs(LargestFD->getType(), FldRV, Args, IRFuncTy); 2224 } 2225 } else { 2226 for (RecordDecl::field_iterator i = RD->field_begin(), e = RD->field_end(); 2227 i != e; ++i) { 2228 FieldDecl *FD = *i; 2229 2230 RValue FldRV = EmitRValueForField(LV, FD, SourceLocation()); 2231 ExpandTypeToArgs(FD->getType(), FldRV, Args, IRFuncTy); 2232 } 2233 } 2234 } else if (Ty->isAnyComplexType()) { 2235 ComplexPairTy CV = RV.getComplexVal(); 2236 Args.push_back(CV.first); 2237 Args.push_back(CV.second); 2238 } else { 2239 assert(RV.isScalar() && 2240 "Unexpected non-scalar rvalue during struct expansion."); 2241 2242 // Insert a bitcast as needed. 2243 llvm::Value *V = RV.getScalarVal(); 2244 if (Args.size() < IRFuncTy->getNumParams() && 2245 V->getType() != IRFuncTy->getParamType(Args.size())) 2246 V = Builder.CreateBitCast(V, IRFuncTy->getParamType(Args.size())); 2247 2248 Args.push_back(V); 2249 } 2250} 2251 2252 2253RValue CodeGenFunction::EmitCall(const CGFunctionInfo &CallInfo, 2254 llvm::Value *Callee, 2255 ReturnValueSlot ReturnValue, 2256 const CallArgList &CallArgs, 2257 const Decl *TargetDecl, 2258 llvm::Instruction **callOrInvoke) { 2259 // FIXME: We no longer need the types from CallArgs; lift up and simplify. 2260 SmallVector<llvm::Value*, 16> Args; 2261 2262 // Handle struct-return functions by passing a pointer to the 2263 // location that we would like to return into. 2264 QualType RetTy = CallInfo.getReturnType(); 2265 const ABIArgInfo &RetAI = CallInfo.getReturnInfo(); 2266 2267 // IRArgNo - Keep track of the argument number in the callee we're looking at. 2268 unsigned IRArgNo = 0; 2269 llvm::FunctionType *IRFuncTy = 2270 cast<llvm::FunctionType>( 2271 cast<llvm::PointerType>(Callee->getType())->getElementType()); 2272 2273 // If the call returns a temporary with struct return, create a temporary 2274 // alloca to hold the result, unless one is given to us. 2275 if (CGM.ReturnTypeUsesSRet(CallInfo)) { 2276 llvm::Value *Value = ReturnValue.getValue(); 2277 if (!Value) 2278 Value = CreateMemTemp(RetTy); 2279 Args.push_back(Value); 2280 checkArgMatches(Value, IRArgNo, IRFuncTy); 2281 } 2282 2283 assert(CallInfo.arg_size() == CallArgs.size() && 2284 "Mismatch between function signature & arguments."); 2285 CGFunctionInfo::const_arg_iterator info_it = CallInfo.arg_begin(); 2286 for (CallArgList::const_iterator I = CallArgs.begin(), E = CallArgs.end(); 2287 I != E; ++I, ++info_it) { 2288 const ABIArgInfo &ArgInfo = info_it->info; 2289 RValue RV = I->RV; 2290 2291 CharUnits TypeAlign = getContext().getTypeAlignInChars(I->Ty); 2292 2293 // Insert a padding argument to ensure proper alignment. 2294 if (llvm::Type *PaddingType = ArgInfo.getPaddingType()) { 2295 Args.push_back(llvm::UndefValue::get(PaddingType)); 2296 ++IRArgNo; 2297 } 2298 2299 switch (ArgInfo.getKind()) { 2300 case ABIArgInfo::Indirect: { 2301 if (RV.isScalar() || RV.isComplex()) { 2302 // Make a temporary alloca to pass the argument. 2303 llvm::AllocaInst *AI = CreateMemTemp(I->Ty); 2304 if (ArgInfo.getIndirectAlign() > AI->getAlignment()) 2305 AI->setAlignment(ArgInfo.getIndirectAlign()); 2306 Args.push_back(AI); 2307 2308 LValue argLV = 2309 MakeAddrLValue(Args.back(), I->Ty, TypeAlign); 2310 2311 if (RV.isScalar()) 2312 EmitStoreOfScalar(RV.getScalarVal(), argLV, /*init*/ true); 2313 else 2314 EmitStoreOfComplex(RV.getComplexVal(), argLV, /*init*/ true); 2315 2316 // Validate argument match. 2317 checkArgMatches(AI, IRArgNo, IRFuncTy); 2318 } else { 2319 // We want to avoid creating an unnecessary temporary+copy here; 2320 // however, we need one in three cases: 2321 // 1. If the argument is not byval, and we are required to copy the 2322 // source. (This case doesn't occur on any common architecture.) 2323 // 2. If the argument is byval, RV is not sufficiently aligned, and 2324 // we cannot force it to be sufficiently aligned. 2325 // 3. If the argument is byval, but RV is located in an address space 2326 // different than that of the argument (0). 2327 llvm::Value *Addr = RV.getAggregateAddr(); 2328 unsigned Align = ArgInfo.getIndirectAlign(); 2329 const llvm::DataLayout *TD = &CGM.getDataLayout(); 2330 const unsigned RVAddrSpace = Addr->getType()->getPointerAddressSpace(); 2331 const unsigned ArgAddrSpace = (IRArgNo < IRFuncTy->getNumParams() ? 2332 IRFuncTy->getParamType(IRArgNo)->getPointerAddressSpace() : 0); 2333 if ((!ArgInfo.getIndirectByVal() && I->NeedsCopy) || 2334 (ArgInfo.getIndirectByVal() && TypeAlign.getQuantity() < Align && 2335 llvm::getOrEnforceKnownAlignment(Addr, Align, TD) < Align) || 2336 (ArgInfo.getIndirectByVal() && (RVAddrSpace != ArgAddrSpace))) { 2337 // Create an aligned temporary, and copy to it. 2338 llvm::AllocaInst *AI = CreateMemTemp(I->Ty); 2339 if (Align > AI->getAlignment()) 2340 AI->setAlignment(Align); 2341 Args.push_back(AI); 2342 EmitAggregateCopy(AI, Addr, I->Ty, RV.isVolatileQualified()); 2343 2344 // Validate argument match. 2345 checkArgMatches(AI, IRArgNo, IRFuncTy); 2346 } else { 2347 // Skip the extra memcpy call. 2348 Args.push_back(Addr); 2349 2350 // Validate argument match. 2351 checkArgMatches(Addr, IRArgNo, IRFuncTy); 2352 } 2353 } 2354 break; 2355 } 2356 2357 case ABIArgInfo::Ignore: 2358 break; 2359 2360 case ABIArgInfo::Extend: 2361 case ABIArgInfo::Direct: { 2362 if (!isa<llvm::StructType>(ArgInfo.getCoerceToType()) && 2363 ArgInfo.getCoerceToType() == ConvertType(info_it->type) && 2364 ArgInfo.getDirectOffset() == 0) { 2365 llvm::Value *V; 2366 if (RV.isScalar()) 2367 V = RV.getScalarVal(); 2368 else 2369 V = Builder.CreateLoad(RV.getAggregateAddr()); 2370 2371 // If the argument doesn't match, perform a bitcast to coerce it. This 2372 // can happen due to trivial type mismatches. 2373 if (IRArgNo < IRFuncTy->getNumParams() && 2374 V->getType() != IRFuncTy->getParamType(IRArgNo)) 2375 V = Builder.CreateBitCast(V, IRFuncTy->getParamType(IRArgNo)); 2376 Args.push_back(V); 2377 2378 checkArgMatches(V, IRArgNo, IRFuncTy); 2379 break; 2380 } 2381 2382 // FIXME: Avoid the conversion through memory if possible. 2383 llvm::Value *SrcPtr; 2384 if (RV.isScalar() || RV.isComplex()) { 2385 SrcPtr = CreateMemTemp(I->Ty, "coerce"); 2386 LValue SrcLV = MakeAddrLValue(SrcPtr, I->Ty, TypeAlign); 2387 if (RV.isScalar()) { 2388 EmitStoreOfScalar(RV.getScalarVal(), SrcLV, /*init*/ true); 2389 } else { 2390 EmitStoreOfComplex(RV.getComplexVal(), SrcLV, /*init*/ true); 2391 } 2392 } else 2393 SrcPtr = RV.getAggregateAddr(); 2394 2395 // If the value is offset in memory, apply the offset now. 2396 if (unsigned Offs = ArgInfo.getDirectOffset()) { 2397 SrcPtr = Builder.CreateBitCast(SrcPtr, Builder.getInt8PtrTy()); 2398 SrcPtr = Builder.CreateConstGEP1_32(SrcPtr, Offs); 2399 SrcPtr = Builder.CreateBitCast(SrcPtr, 2400 llvm::PointerType::getUnqual(ArgInfo.getCoerceToType())); 2401 2402 } 2403 2404 // If the coerce-to type is a first class aggregate, we flatten it and 2405 // pass the elements. Either way is semantically identical, but fast-isel 2406 // and the optimizer generally likes scalar values better than FCAs. 2407 if (llvm::StructType *STy = 2408 dyn_cast<llvm::StructType>(ArgInfo.getCoerceToType())) { 2409 llvm::Type *SrcTy = 2410 cast<llvm::PointerType>(SrcPtr->getType())->getElementType(); 2411 uint64_t SrcSize = CGM.getDataLayout().getTypeAllocSize(SrcTy); 2412 uint64_t DstSize = CGM.getDataLayout().getTypeAllocSize(STy); 2413 2414 // If the source type is smaller than the destination type of the 2415 // coerce-to logic, copy the source value into a temp alloca the size 2416 // of the destination type to allow loading all of it. The bits past 2417 // the source value are left undef. 2418 if (SrcSize < DstSize) { 2419 llvm::AllocaInst *TempAlloca 2420 = CreateTempAlloca(STy, SrcPtr->getName() + ".coerce"); 2421 Builder.CreateMemCpy(TempAlloca, SrcPtr, SrcSize, 0); 2422 SrcPtr = TempAlloca; 2423 } else { 2424 SrcPtr = Builder.CreateBitCast(SrcPtr, 2425 llvm::PointerType::getUnqual(STy)); 2426 } 2427 2428 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { 2429 llvm::Value *EltPtr = Builder.CreateConstGEP2_32(SrcPtr, 0, i); 2430 llvm::LoadInst *LI = Builder.CreateLoad(EltPtr); 2431 // We don't know what we're loading from. 2432 LI->setAlignment(1); 2433 Args.push_back(LI); 2434 2435 // Validate argument match. 2436 checkArgMatches(LI, IRArgNo, IRFuncTy); 2437 } 2438 } else { 2439 // In the simple case, just pass the coerced loaded value. 2440 Args.push_back(CreateCoercedLoad(SrcPtr, ArgInfo.getCoerceToType(), 2441 *this)); 2442 2443 // Validate argument match. 2444 checkArgMatches(Args.back(), IRArgNo, IRFuncTy); 2445 } 2446 2447 break; 2448 } 2449 2450 case ABIArgInfo::Expand: 2451 ExpandTypeToArgs(I->Ty, RV, Args, IRFuncTy); 2452 IRArgNo = Args.size(); 2453 break; 2454 } 2455 } 2456 2457 if (!CallArgs.getCleanupsToDeactivate().empty()) 2458 deactivateArgCleanupsBeforeCall(*this, CallArgs); 2459 2460 // If the callee is a bitcast of a function to a varargs pointer to function 2461 // type, check to see if we can remove the bitcast. This handles some cases 2462 // with unprototyped functions. 2463 if (llvm::ConstantExpr *CE = dyn_cast<llvm::ConstantExpr>(Callee)) 2464 if (llvm::Function *CalleeF = dyn_cast<llvm::Function>(CE->getOperand(0))) { 2465 llvm::PointerType *CurPT=cast<llvm::PointerType>(Callee->getType()); 2466 llvm::FunctionType *CurFT = 2467 cast<llvm::FunctionType>(CurPT->getElementType()); 2468 llvm::FunctionType *ActualFT = CalleeF->getFunctionType(); 2469 2470 if (CE->getOpcode() == llvm::Instruction::BitCast && 2471 ActualFT->getReturnType() == CurFT->getReturnType() && 2472 ActualFT->getNumParams() == CurFT->getNumParams() && 2473 ActualFT->getNumParams() == Args.size() && 2474 (CurFT->isVarArg() || !ActualFT->isVarArg())) { 2475 bool ArgsMatch = true; 2476 for (unsigned i = 0, e = ActualFT->getNumParams(); i != e; ++i) 2477 if (ActualFT->getParamType(i) != CurFT->getParamType(i)) { 2478 ArgsMatch = false; 2479 break; 2480 } 2481 2482 // Strip the cast if we can get away with it. This is a nice cleanup, 2483 // but also allows us to inline the function at -O0 if it is marked 2484 // always_inline. 2485 if (ArgsMatch) 2486 Callee = CalleeF; 2487 } 2488 } 2489 2490 unsigned CallingConv; 2491 CodeGen::AttributeListType AttributeList; 2492 CGM.ConstructAttributeList(CallInfo, TargetDecl, AttributeList, 2493 CallingConv, true); 2494 llvm::AttributeSet Attrs = llvm::AttributeSet::get(getLLVMContext(), 2495 AttributeList); 2496 2497 llvm::BasicBlock *InvokeDest = 0; 2498 if (!Attrs.hasAttribute(llvm::AttributeSet::FunctionIndex, 2499 llvm::Attribute::NoUnwind)) 2500 InvokeDest = getInvokeDest(); 2501 2502 llvm::CallSite CS; 2503 if (!InvokeDest) { 2504 CS = Builder.CreateCall(Callee, Args); 2505 } else { 2506 llvm::BasicBlock *Cont = createBasicBlock("invoke.cont"); 2507 CS = Builder.CreateInvoke(Callee, Cont, InvokeDest, Args); 2508 EmitBlock(Cont); 2509 } 2510 if (callOrInvoke) 2511 *callOrInvoke = CS.getInstruction(); 2512 2513 CS.setAttributes(Attrs); 2514 CS.setCallingConv(static_cast<llvm::CallingConv::ID>(CallingConv)); 2515 2516 // In ObjC ARC mode with no ObjC ARC exception safety, tell the ARC 2517 // optimizer it can aggressively ignore unwind edges. 2518 if (CGM.getLangOpts().ObjCAutoRefCount) 2519 AddObjCARCExceptionMetadata(CS.getInstruction()); 2520 2521 // If the call doesn't return, finish the basic block and clear the 2522 // insertion point; this allows the rest of IRgen to discard 2523 // unreachable code. 2524 if (CS.doesNotReturn()) { 2525 Builder.CreateUnreachable(); 2526 Builder.ClearInsertionPoint(); 2527 2528 // FIXME: For now, emit a dummy basic block because expr emitters in 2529 // generally are not ready to handle emitting expressions at unreachable 2530 // points. 2531 EnsureInsertPoint(); 2532 2533 // Return a reasonable RValue. 2534 return GetUndefRValue(RetTy); 2535 } 2536 2537 llvm::Instruction *CI = CS.getInstruction(); 2538 if (Builder.isNamePreserving() && !CI->getType()->isVoidTy()) 2539 CI->setName("call"); 2540 2541 // Emit any writebacks immediately. Arguably this should happen 2542 // after any return-value munging. 2543 if (CallArgs.hasWritebacks()) 2544 emitWritebacks(*this, CallArgs); 2545 2546 switch (RetAI.getKind()) { 2547 case ABIArgInfo::Indirect: 2548 return convertTempToRValue(Args[0], RetTy, SourceLocation()); 2549 2550 case ABIArgInfo::Ignore: 2551 // If we are ignoring an argument that had a result, make sure to 2552 // construct the appropriate return value for our caller. 2553 return GetUndefRValue(RetTy); 2554 2555 case ABIArgInfo::Extend: 2556 case ABIArgInfo::Direct: { 2557 llvm::Type *RetIRTy = ConvertType(RetTy); 2558 if (RetAI.getCoerceToType() == RetIRTy && RetAI.getDirectOffset() == 0) { 2559 switch (getEvaluationKind(RetTy)) { 2560 case TEK_Complex: { 2561 llvm::Value *Real = Builder.CreateExtractValue(CI, 0); 2562 llvm::Value *Imag = Builder.CreateExtractValue(CI, 1); 2563 return RValue::getComplex(std::make_pair(Real, Imag)); 2564 } 2565 case TEK_Aggregate: { 2566 llvm::Value *DestPtr = ReturnValue.getValue(); 2567 bool DestIsVolatile = ReturnValue.isVolatile(); 2568 2569 if (!DestPtr) { 2570 DestPtr = CreateMemTemp(RetTy, "agg.tmp"); 2571 DestIsVolatile = false; 2572 } 2573 BuildAggStore(*this, CI, DestPtr, DestIsVolatile, false); 2574 return RValue::getAggregate(DestPtr); 2575 } 2576 case TEK_Scalar: { 2577 // If the argument doesn't match, perform a bitcast to coerce it. This 2578 // can happen due to trivial type mismatches. 2579 llvm::Value *V = CI; 2580 if (V->getType() != RetIRTy) 2581 V = Builder.CreateBitCast(V, RetIRTy); 2582 return RValue::get(V); 2583 } 2584 } 2585 llvm_unreachable("bad evaluation kind"); 2586 } 2587 2588 llvm::Value *DestPtr = ReturnValue.getValue(); 2589 bool DestIsVolatile = ReturnValue.isVolatile(); 2590 2591 if (!DestPtr) { 2592 DestPtr = CreateMemTemp(RetTy, "coerce"); 2593 DestIsVolatile = false; 2594 } 2595 2596 // If the value is offset in memory, apply the offset now. 2597 llvm::Value *StorePtr = DestPtr; 2598 if (unsigned Offs = RetAI.getDirectOffset()) { 2599 StorePtr = Builder.CreateBitCast(StorePtr, Builder.getInt8PtrTy()); 2600 StorePtr = Builder.CreateConstGEP1_32(StorePtr, Offs); 2601 StorePtr = Builder.CreateBitCast(StorePtr, 2602 llvm::PointerType::getUnqual(RetAI.getCoerceToType())); 2603 } 2604 CreateCoercedStore(CI, StorePtr, DestIsVolatile, *this); 2605 2606 return convertTempToRValue(DestPtr, RetTy, SourceLocation()); 2607 } 2608 2609 case ABIArgInfo::Expand: 2610 llvm_unreachable("Invalid ABI kind for return argument"); 2611 } 2612 2613 llvm_unreachable("Unhandled ABIArgInfo::Kind"); 2614} 2615 2616/* VarArg handling */ 2617 2618llvm::Value *CodeGenFunction::EmitVAArg(llvm::Value *VAListAddr, QualType Ty) { 2619 return CGM.getTypes().getABIInfo().EmitVAArg(VAListAddr, Ty, *this); 2620} 2621