CGExprAgg.cpp revision babcf9d04f4ed9d7ac96812e42c9e8fc0f1ae2c2
1//===--- CGExprAgg.cpp - Emit LLVM Code from Aggregate Expressions --------===// 2// 3// The LLVM Compiler Infrastructure 4// 5// This file is distributed under the University of Illinois Open Source 6// License. See LICENSE.TXT for details. 7// 8//===----------------------------------------------------------------------===// 9// 10// This contains code to emit Aggregate Expr nodes as LLVM code. 11// 12//===----------------------------------------------------------------------===// 13 14#include "CodeGenFunction.h" 15#include "CodeGenModule.h" 16#include "CGObjCRuntime.h" 17#include "clang/AST/ASTContext.h" 18#include "clang/AST/DeclCXX.h" 19#include "clang/AST/DeclTemplate.h" 20#include "clang/AST/StmtVisitor.h" 21#include "llvm/Constants.h" 22#include "llvm/Function.h" 23#include "llvm/GlobalVariable.h" 24#include "llvm/Intrinsics.h" 25using namespace clang; 26using namespace CodeGen; 27 28//===----------------------------------------------------------------------===// 29// Aggregate Expression Emitter 30//===----------------------------------------------------------------------===// 31 32namespace { 33class AggExprEmitter : public StmtVisitor<AggExprEmitter> { 34 CodeGenFunction &CGF; 35 CGBuilderTy &Builder; 36 AggValueSlot Dest; 37 bool IgnoreResult; 38 39 /// We want to use 'dest' as the return slot except under two 40 /// conditions: 41 /// - The destination slot requires garbage collection, so we 42 /// need to use the GC API. 43 /// - The destination slot is potentially aliased. 44 bool shouldUseDestForReturnSlot() const { 45 return !(Dest.requiresGCollection() || Dest.isPotentiallyAliased()); 46 } 47 48 ReturnValueSlot getReturnValueSlot() const { 49 if (!shouldUseDestForReturnSlot()) 50 return ReturnValueSlot(); 51 52 return ReturnValueSlot(Dest.getAddr(), Dest.isVolatile()); 53 } 54 55 AggValueSlot EnsureSlot(QualType T) { 56 if (!Dest.isIgnored()) return Dest; 57 return CGF.CreateAggTemp(T, "agg.tmp.ensured"); 58 } 59 60public: 61 AggExprEmitter(CodeGenFunction &cgf, AggValueSlot Dest, 62 bool ignore) 63 : CGF(cgf), Builder(CGF.Builder), Dest(Dest), 64 IgnoreResult(ignore) { 65 } 66 67 //===--------------------------------------------------------------------===// 68 // Utilities 69 //===--------------------------------------------------------------------===// 70 71 /// EmitAggLoadOfLValue - Given an expression with aggregate type that 72 /// represents a value lvalue, this method emits the address of the lvalue, 73 /// then loads the result into DestPtr. 74 void EmitAggLoadOfLValue(const Expr *E); 75 76 /// EmitFinalDestCopy - Perform the final copy to DestPtr, if desired. 77 void EmitFinalDestCopy(const Expr *E, LValue Src, bool Ignore = false); 78 void EmitFinalDestCopy(const Expr *E, RValue Src, bool Ignore = false, 79 unsigned Alignment = 0); 80 81 void EmitMoveFromReturnSlot(const Expr *E, RValue Src); 82 83 void EmitStdInitializerList(llvm::Value *DestPtr, InitListExpr *InitList); 84 void EmitArrayInit(llvm::Value *DestPtr, llvm::ArrayType *AType, 85 QualType elementType, InitListExpr *E); 86 87 AggValueSlot::NeedsGCBarriers_t needsGC(QualType T) { 88 if (CGF.getLangOptions().getGC() && TypeRequiresGCollection(T)) 89 return AggValueSlot::NeedsGCBarriers; 90 return AggValueSlot::DoesNotNeedGCBarriers; 91 } 92 93 bool TypeRequiresGCollection(QualType T); 94 95 //===--------------------------------------------------------------------===// 96 // Visitor Methods 97 //===--------------------------------------------------------------------===// 98 99 void VisitStmt(Stmt *S) { 100 CGF.ErrorUnsupported(S, "aggregate expression"); 101 } 102 void VisitParenExpr(ParenExpr *PE) { Visit(PE->getSubExpr()); } 103 void VisitGenericSelectionExpr(GenericSelectionExpr *GE) { 104 Visit(GE->getResultExpr()); 105 } 106 void VisitUnaryExtension(UnaryOperator *E) { Visit(E->getSubExpr()); } 107 void VisitSubstNonTypeTemplateParmExpr(SubstNonTypeTemplateParmExpr *E) { 108 return Visit(E->getReplacement()); 109 } 110 111 // l-values. 112 void VisitDeclRefExpr(DeclRefExpr *DRE) { EmitAggLoadOfLValue(DRE); } 113 void VisitMemberExpr(MemberExpr *ME) { EmitAggLoadOfLValue(ME); } 114 void VisitUnaryDeref(UnaryOperator *E) { EmitAggLoadOfLValue(E); } 115 void VisitStringLiteral(StringLiteral *E) { EmitAggLoadOfLValue(E); } 116 void VisitCompoundLiteralExpr(CompoundLiteralExpr *E); 117 void VisitArraySubscriptExpr(ArraySubscriptExpr *E) { 118 EmitAggLoadOfLValue(E); 119 } 120 void VisitBlockDeclRefExpr(const BlockDeclRefExpr *E) { 121 EmitAggLoadOfLValue(E); 122 } 123 void VisitPredefinedExpr(const PredefinedExpr *E) { 124 EmitAggLoadOfLValue(E); 125 } 126 127 // Operators. 128 void VisitCastExpr(CastExpr *E); 129 void VisitCallExpr(const CallExpr *E); 130 void VisitStmtExpr(const StmtExpr *E); 131 void VisitBinaryOperator(const BinaryOperator *BO); 132 void VisitPointerToDataMemberBinaryOperator(const BinaryOperator *BO); 133 void VisitBinAssign(const BinaryOperator *E); 134 void VisitBinComma(const BinaryOperator *E); 135 136 void VisitObjCMessageExpr(ObjCMessageExpr *E); 137 void VisitObjCIvarRefExpr(ObjCIvarRefExpr *E) { 138 EmitAggLoadOfLValue(E); 139 } 140 141 void VisitAbstractConditionalOperator(const AbstractConditionalOperator *CO); 142 void VisitChooseExpr(const ChooseExpr *CE); 143 void VisitInitListExpr(InitListExpr *E); 144 void VisitImplicitValueInitExpr(ImplicitValueInitExpr *E); 145 void VisitCXXDefaultArgExpr(CXXDefaultArgExpr *DAE) { 146 Visit(DAE->getExpr()); 147 } 148 void VisitCXXBindTemporaryExpr(CXXBindTemporaryExpr *E); 149 void VisitCXXConstructExpr(const CXXConstructExpr *E); 150 void VisitLambdaExpr(LambdaExpr *E); 151 void VisitExprWithCleanups(ExprWithCleanups *E); 152 void VisitCXXScalarValueInitExpr(CXXScalarValueInitExpr *E); 153 void VisitCXXTypeidExpr(CXXTypeidExpr *E) { EmitAggLoadOfLValue(E); } 154 void VisitMaterializeTemporaryExpr(MaterializeTemporaryExpr *E); 155 void VisitOpaqueValueExpr(OpaqueValueExpr *E); 156 157 void VisitPseudoObjectExpr(PseudoObjectExpr *E) { 158 if (E->isGLValue()) { 159 LValue LV = CGF.EmitPseudoObjectLValue(E); 160 return EmitFinalDestCopy(E, LV); 161 } 162 163 CGF.EmitPseudoObjectRValue(E, EnsureSlot(E->getType())); 164 } 165 166 void VisitVAArgExpr(VAArgExpr *E); 167 168 void EmitInitializationToLValue(Expr *E, LValue Address); 169 void EmitNullInitializationToLValue(LValue Address); 170 // case Expr::ChooseExprClass: 171 void VisitCXXThrowExpr(const CXXThrowExpr *E) { CGF.EmitCXXThrowExpr(E); } 172 void VisitAtomicExpr(AtomicExpr *E) { 173 CGF.EmitAtomicExpr(E, EnsureSlot(E->getType()).getAddr()); 174 } 175}; 176} // end anonymous namespace. 177 178//===----------------------------------------------------------------------===// 179// Utilities 180//===----------------------------------------------------------------------===// 181 182/// EmitAggLoadOfLValue - Given an expression with aggregate type that 183/// represents a value lvalue, this method emits the address of the lvalue, 184/// then loads the result into DestPtr. 185void AggExprEmitter::EmitAggLoadOfLValue(const Expr *E) { 186 LValue LV = CGF.EmitLValue(E); 187 EmitFinalDestCopy(E, LV); 188} 189 190/// \brief True if the given aggregate type requires special GC API calls. 191bool AggExprEmitter::TypeRequiresGCollection(QualType T) { 192 // Only record types have members that might require garbage collection. 193 const RecordType *RecordTy = T->getAs<RecordType>(); 194 if (!RecordTy) return false; 195 196 // Don't mess with non-trivial C++ types. 197 RecordDecl *Record = RecordTy->getDecl(); 198 if (isa<CXXRecordDecl>(Record) && 199 (!cast<CXXRecordDecl>(Record)->hasTrivialCopyConstructor() || 200 !cast<CXXRecordDecl>(Record)->hasTrivialDestructor())) 201 return false; 202 203 // Check whether the type has an object member. 204 return Record->hasObjectMember(); 205} 206 207/// \brief Perform the final move to DestPtr if for some reason 208/// getReturnValueSlot() didn't use it directly. 209/// 210/// The idea is that you do something like this: 211/// RValue Result = EmitSomething(..., getReturnValueSlot()); 212/// EmitMoveFromReturnSlot(E, Result); 213/// 214/// If nothing interferes, this will cause the result to be emitted 215/// directly into the return value slot. Otherwise, a final move 216/// will be performed. 217void AggExprEmitter::EmitMoveFromReturnSlot(const Expr *E, RValue Src) { 218 if (shouldUseDestForReturnSlot()) { 219 // Logically, Dest.getAddr() should equal Src.getAggregateAddr(). 220 // The possibility of undef rvalues complicates that a lot, 221 // though, so we can't really assert. 222 return; 223 } 224 225 // Otherwise, do a final copy, 226 assert(Dest.getAddr() != Src.getAggregateAddr()); 227 EmitFinalDestCopy(E, Src, /*Ignore*/ true); 228} 229 230/// EmitFinalDestCopy - Perform the final copy to DestPtr, if desired. 231void AggExprEmitter::EmitFinalDestCopy(const Expr *E, RValue Src, bool Ignore, 232 unsigned Alignment) { 233 assert(Src.isAggregate() && "value must be aggregate value!"); 234 235 // If Dest is ignored, then we're evaluating an aggregate expression 236 // in a context (like an expression statement) that doesn't care 237 // about the result. C says that an lvalue-to-rvalue conversion is 238 // performed in these cases; C++ says that it is not. In either 239 // case, we don't actually need to do anything unless the value is 240 // volatile. 241 if (Dest.isIgnored()) { 242 if (!Src.isVolatileQualified() || 243 CGF.CGM.getLangOptions().CPlusPlus || 244 (IgnoreResult && Ignore)) 245 return; 246 247 // If the source is volatile, we must read from it; to do that, we need 248 // some place to put it. 249 Dest = CGF.CreateAggTemp(E->getType(), "agg.tmp"); 250 } 251 252 if (Dest.requiresGCollection()) { 253 CharUnits size = CGF.getContext().getTypeSizeInChars(E->getType()); 254 llvm::Type *SizeTy = CGF.ConvertType(CGF.getContext().getSizeType()); 255 llvm::Value *SizeVal = llvm::ConstantInt::get(SizeTy, size.getQuantity()); 256 CGF.CGM.getObjCRuntime().EmitGCMemmoveCollectable(CGF, 257 Dest.getAddr(), 258 Src.getAggregateAddr(), 259 SizeVal); 260 return; 261 } 262 // If the result of the assignment is used, copy the LHS there also. 263 // FIXME: Pass VolatileDest as well. I think we also need to merge volatile 264 // from the source as well, as we can't eliminate it if either operand 265 // is volatile, unless copy has volatile for both source and destination.. 266 CGF.EmitAggregateCopy(Dest.getAddr(), Src.getAggregateAddr(), E->getType(), 267 Dest.isVolatile()|Src.isVolatileQualified(), 268 Alignment); 269} 270 271/// EmitFinalDestCopy - Perform the final copy to DestPtr, if desired. 272void AggExprEmitter::EmitFinalDestCopy(const Expr *E, LValue Src, bool Ignore) { 273 assert(Src.isSimple() && "Can't have aggregate bitfield, vector, etc"); 274 275 CharUnits Alignment = std::min(Src.getAlignment(), Dest.getAlignment()); 276 EmitFinalDestCopy(E, Src.asAggregateRValue(), Ignore, Alignment.getQuantity()); 277} 278 279static QualType GetStdInitializerListElementType(QualType T) { 280 // Just assume that this is really std::initializer_list. 281 ClassTemplateSpecializationDecl *specialization = 282 cast<ClassTemplateSpecializationDecl>(T->castAs<RecordType>()->getDecl()); 283 return specialization->getTemplateArgs()[0].getAsType(); 284} 285 286/// \brief Prepare cleanup for the temporary array. 287static void EmitStdInitializerListCleanup(CodeGenFunction &CGF, 288 QualType arrayType, 289 llvm::Value *addr, 290 const InitListExpr *initList) { 291 QualType::DestructionKind dtorKind = arrayType.isDestructedType(); 292 if (!dtorKind) 293 return; // Type doesn't need destroying. 294 if (dtorKind != QualType::DK_cxx_destructor) { 295 CGF.ErrorUnsupported(initList, "ObjC ARC type in initializer_list"); 296 return; 297 } 298 299 CodeGenFunction::Destroyer *destroyer = CGF.getDestroyer(dtorKind); 300 CGF.pushDestroy(NormalAndEHCleanup, addr, arrayType, destroyer, 301 /*EHCleanup=*/true); 302} 303 304/// \brief Emit the initializer for a std::initializer_list initialized with a 305/// real initializer list. 306void AggExprEmitter::EmitStdInitializerList(llvm::Value *destPtr, 307 InitListExpr *initList) { 308 // We emit an array containing the elements, then have the init list point 309 // at the array. 310 ASTContext &ctx = CGF.getContext(); 311 unsigned numInits = initList->getNumInits(); 312 QualType element = GetStdInitializerListElementType(initList->getType()); 313 llvm::APInt size(ctx.getTypeSize(ctx.getSizeType()), numInits); 314 QualType array = ctx.getConstantArrayType(element, size, ArrayType::Normal,0); 315 llvm::Type *LTy = CGF.ConvertTypeForMem(array); 316 llvm::AllocaInst *alloc = CGF.CreateTempAlloca(LTy); 317 alloc->setAlignment(ctx.getTypeAlignInChars(array).getQuantity()); 318 alloc->setName(".initlist."); 319 320 EmitArrayInit(alloc, cast<llvm::ArrayType>(LTy), element, initList); 321 322 // FIXME: The diagnostics are somewhat out of place here. 323 RecordDecl *record = initList->getType()->castAs<RecordType>()->getDecl(); 324 RecordDecl::field_iterator field = record->field_begin(); 325 if (field == record->field_end()) { 326 CGF.ErrorUnsupported(initList, "weird std::initializer_list"); 327 return; 328 } 329 330 QualType elementPtr = ctx.getPointerType(element.withConst()); 331 332 // Start pointer. 333 if (!ctx.hasSameType(field->getType(), elementPtr)) { 334 CGF.ErrorUnsupported(initList, "weird std::initializer_list"); 335 return; 336 } 337 LValue start = CGF.EmitLValueForFieldInitialization(destPtr, *field, 0); 338 llvm::Value *arrayStart = Builder.CreateStructGEP(alloc, 0, "arraystart"); 339 CGF.EmitStoreThroughLValue(RValue::get(arrayStart), start); 340 ++field; 341 342 if (field == record->field_end()) { 343 CGF.ErrorUnsupported(initList, "weird std::initializer_list"); 344 return; 345 } 346 LValue endOrLength = CGF.EmitLValueForFieldInitialization(destPtr, *field, 0); 347 if (ctx.hasSameType(field->getType(), elementPtr)) { 348 // End pointer. 349 llvm::Value *arrayEnd = Builder.CreateStructGEP(alloc,numInits, "arrayend"); 350 CGF.EmitStoreThroughLValue(RValue::get(arrayEnd), endOrLength); 351 } else if(ctx.hasSameType(field->getType(), ctx.getSizeType())) { 352 // Length. 353 CGF.EmitStoreThroughLValue(RValue::get(Builder.getInt(size)), endOrLength); 354 } else { 355 CGF.ErrorUnsupported(initList, "weird std::initializer_list"); 356 return; 357 } 358 359 if (!Dest.isExternallyDestructed()) 360 EmitStdInitializerListCleanup(CGF, array, alloc, initList); 361} 362 363/// \brief Emit initialization of an array from an initializer list. 364void AggExprEmitter::EmitArrayInit(llvm::Value *DestPtr, llvm::ArrayType *AType, 365 QualType elementType, InitListExpr *E) { 366 uint64_t NumInitElements = E->getNumInits(); 367 368 uint64_t NumArrayElements = AType->getNumElements(); 369 assert(NumInitElements <= NumArrayElements); 370 371 // DestPtr is an array*. Construct an elementType* by drilling 372 // down a level. 373 llvm::Value *zero = llvm::ConstantInt::get(CGF.SizeTy, 0); 374 llvm::Value *indices[] = { zero, zero }; 375 llvm::Value *begin = 376 Builder.CreateInBoundsGEP(DestPtr, indices, "arrayinit.begin"); 377 378 // Exception safety requires us to destroy all the 379 // already-constructed members if an initializer throws. 380 // For that, we'll need an EH cleanup. 381 QualType::DestructionKind dtorKind = elementType.isDestructedType(); 382 llvm::AllocaInst *endOfInit = 0; 383 EHScopeStack::stable_iterator cleanup; 384 llvm::Instruction *cleanupDominator = 0; 385 if (CGF.needsEHCleanup(dtorKind)) { 386 // In principle we could tell the cleanup where we are more 387 // directly, but the control flow can get so varied here that it 388 // would actually be quite complex. Therefore we go through an 389 // alloca. 390 endOfInit = CGF.CreateTempAlloca(begin->getType(), 391 "arrayinit.endOfInit"); 392 cleanupDominator = Builder.CreateStore(begin, endOfInit); 393 CGF.pushIrregularPartialArrayCleanup(begin, endOfInit, elementType, 394 CGF.getDestroyer(dtorKind)); 395 cleanup = CGF.EHStack.stable_begin(); 396 397 // Otherwise, remember that we didn't need a cleanup. 398 } else { 399 dtorKind = QualType::DK_none; 400 } 401 402 llvm::Value *one = llvm::ConstantInt::get(CGF.SizeTy, 1); 403 404 // The 'current element to initialize'. The invariants on this 405 // variable are complicated. Essentially, after each iteration of 406 // the loop, it points to the last initialized element, except 407 // that it points to the beginning of the array before any 408 // elements have been initialized. 409 llvm::Value *element = begin; 410 411 // Emit the explicit initializers. 412 for (uint64_t i = 0; i != NumInitElements; ++i) { 413 // Advance to the next element. 414 if (i > 0) { 415 element = Builder.CreateInBoundsGEP(element, one, "arrayinit.element"); 416 417 // Tell the cleanup that it needs to destroy up to this 418 // element. TODO: some of these stores can be trivially 419 // observed to be unnecessary. 420 if (endOfInit) Builder.CreateStore(element, endOfInit); 421 } 422 423 // If these are nested std::initializer_list inits, do them directly, 424 // because they are conceptually the same "location". 425 InitListExpr *initList = dyn_cast<InitListExpr>(E->getInit(i)); 426 if (initList && initList->initializesStdInitializerList()) { 427 EmitStdInitializerList(element, initList); 428 } else { 429 LValue elementLV = CGF.MakeAddrLValue(element, elementType); 430 EmitInitializationToLValue(E->getInit(i), elementLV); 431 } 432 } 433 434 // Check whether there's a non-trivial array-fill expression. 435 // Note that this will be a CXXConstructExpr even if the element 436 // type is an array (or array of array, etc.) of class type. 437 Expr *filler = E->getArrayFiller(); 438 bool hasTrivialFiller = true; 439 if (CXXConstructExpr *cons = dyn_cast_or_null<CXXConstructExpr>(filler)) { 440 assert(cons->getConstructor()->isDefaultConstructor()); 441 hasTrivialFiller = cons->getConstructor()->isTrivial(); 442 } 443 444 // Any remaining elements need to be zero-initialized, possibly 445 // using the filler expression. We can skip this if the we're 446 // emitting to zeroed memory. 447 if (NumInitElements != NumArrayElements && 448 !(Dest.isZeroed() && hasTrivialFiller && 449 CGF.getTypes().isZeroInitializable(elementType))) { 450 451 // Use an actual loop. This is basically 452 // do { *array++ = filler; } while (array != end); 453 454 // Advance to the start of the rest of the array. 455 if (NumInitElements) { 456 element = Builder.CreateInBoundsGEP(element, one, "arrayinit.start"); 457 if (endOfInit) Builder.CreateStore(element, endOfInit); 458 } 459 460 // Compute the end of the array. 461 llvm::Value *end = Builder.CreateInBoundsGEP(begin, 462 llvm::ConstantInt::get(CGF.SizeTy, NumArrayElements), 463 "arrayinit.end"); 464 465 llvm::BasicBlock *entryBB = Builder.GetInsertBlock(); 466 llvm::BasicBlock *bodyBB = CGF.createBasicBlock("arrayinit.body"); 467 468 // Jump into the body. 469 CGF.EmitBlock(bodyBB); 470 llvm::PHINode *currentElement = 471 Builder.CreatePHI(element->getType(), 2, "arrayinit.cur"); 472 currentElement->addIncoming(element, entryBB); 473 474 // Emit the actual filler expression. 475 LValue elementLV = CGF.MakeAddrLValue(currentElement, elementType); 476 if (filler) 477 EmitInitializationToLValue(filler, elementLV); 478 else 479 EmitNullInitializationToLValue(elementLV); 480 481 // Move on to the next element. 482 llvm::Value *nextElement = 483 Builder.CreateInBoundsGEP(currentElement, one, "arrayinit.next"); 484 485 // Tell the EH cleanup that we finished with the last element. 486 if (endOfInit) Builder.CreateStore(nextElement, endOfInit); 487 488 // Leave the loop if we're done. 489 llvm::Value *done = Builder.CreateICmpEQ(nextElement, end, 490 "arrayinit.done"); 491 llvm::BasicBlock *endBB = CGF.createBasicBlock("arrayinit.end"); 492 Builder.CreateCondBr(done, endBB, bodyBB); 493 currentElement->addIncoming(nextElement, Builder.GetInsertBlock()); 494 495 CGF.EmitBlock(endBB); 496 } 497 498 // Leave the partial-array cleanup if we entered one. 499 if (dtorKind) CGF.DeactivateCleanupBlock(cleanup, cleanupDominator); 500} 501 502//===----------------------------------------------------------------------===// 503// Visitor Methods 504//===----------------------------------------------------------------------===// 505 506void AggExprEmitter::VisitMaterializeTemporaryExpr(MaterializeTemporaryExpr *E){ 507 Visit(E->GetTemporaryExpr()); 508} 509 510void AggExprEmitter::VisitOpaqueValueExpr(OpaqueValueExpr *e) { 511 EmitFinalDestCopy(e, CGF.getOpaqueLValueMapping(e)); 512} 513 514void 515AggExprEmitter::VisitCompoundLiteralExpr(CompoundLiteralExpr *E) { 516 if (E->getType().isPODType(CGF.getContext())) { 517 // For a POD type, just emit a load of the lvalue + a copy, because our 518 // compound literal might alias the destination. 519 // FIXME: This is a band-aid; the real problem appears to be in our handling 520 // of assignments, where we store directly into the LHS without checking 521 // whether anything in the RHS aliases. 522 EmitAggLoadOfLValue(E); 523 return; 524 } 525 526 AggValueSlot Slot = EnsureSlot(E->getType()); 527 CGF.EmitAggExpr(E->getInitializer(), Slot); 528} 529 530 531void AggExprEmitter::VisitCastExpr(CastExpr *E) { 532 switch (E->getCastKind()) { 533 case CK_Dynamic: { 534 assert(isa<CXXDynamicCastExpr>(E) && "CK_Dynamic without a dynamic_cast?"); 535 LValue LV = CGF.EmitCheckedLValue(E->getSubExpr()); 536 // FIXME: Do we also need to handle property references here? 537 if (LV.isSimple()) 538 CGF.EmitDynamicCast(LV.getAddress(), cast<CXXDynamicCastExpr>(E)); 539 else 540 CGF.CGM.ErrorUnsupported(E, "non-simple lvalue dynamic_cast"); 541 542 if (!Dest.isIgnored()) 543 CGF.CGM.ErrorUnsupported(E, "lvalue dynamic_cast with a destination"); 544 break; 545 } 546 547 case CK_ToUnion: { 548 if (Dest.isIgnored()) break; 549 550 // GCC union extension 551 QualType Ty = E->getSubExpr()->getType(); 552 QualType PtrTy = CGF.getContext().getPointerType(Ty); 553 llvm::Value *CastPtr = Builder.CreateBitCast(Dest.getAddr(), 554 CGF.ConvertType(PtrTy)); 555 EmitInitializationToLValue(E->getSubExpr(), 556 CGF.MakeAddrLValue(CastPtr, Ty)); 557 break; 558 } 559 560 case CK_DerivedToBase: 561 case CK_BaseToDerived: 562 case CK_UncheckedDerivedToBase: { 563 llvm_unreachable("cannot perform hierarchy conversion in EmitAggExpr: " 564 "should have been unpacked before we got here"); 565 } 566 567 case CK_LValueToRValue: // hope for downstream optimization 568 case CK_NoOp: 569 case CK_AtomicToNonAtomic: 570 case CK_NonAtomicToAtomic: 571 case CK_UserDefinedConversion: 572 case CK_ConstructorConversion: 573 assert(CGF.getContext().hasSameUnqualifiedType(E->getSubExpr()->getType(), 574 E->getType()) && 575 "Implicit cast types must be compatible"); 576 Visit(E->getSubExpr()); 577 break; 578 579 case CK_LValueBitCast: 580 llvm_unreachable("should not be emitting lvalue bitcast as rvalue"); 581 582 case CK_Dependent: 583 case CK_BitCast: 584 case CK_ArrayToPointerDecay: 585 case CK_FunctionToPointerDecay: 586 case CK_NullToPointer: 587 case CK_NullToMemberPointer: 588 case CK_BaseToDerivedMemberPointer: 589 case CK_DerivedToBaseMemberPointer: 590 case CK_MemberPointerToBoolean: 591 case CK_ReinterpretMemberPointer: 592 case CK_IntegralToPointer: 593 case CK_PointerToIntegral: 594 case CK_PointerToBoolean: 595 case CK_ToVoid: 596 case CK_VectorSplat: 597 case CK_IntegralCast: 598 case CK_IntegralToBoolean: 599 case CK_IntegralToFloating: 600 case CK_FloatingToIntegral: 601 case CK_FloatingToBoolean: 602 case CK_FloatingCast: 603 case CK_CPointerToObjCPointerCast: 604 case CK_BlockPointerToObjCPointerCast: 605 case CK_AnyPointerToBlockPointerCast: 606 case CK_ObjCObjectLValueCast: 607 case CK_FloatingRealToComplex: 608 case CK_FloatingComplexToReal: 609 case CK_FloatingComplexToBoolean: 610 case CK_FloatingComplexCast: 611 case CK_FloatingComplexToIntegralComplex: 612 case CK_IntegralRealToComplex: 613 case CK_IntegralComplexToReal: 614 case CK_IntegralComplexToBoolean: 615 case CK_IntegralComplexCast: 616 case CK_IntegralComplexToFloatingComplex: 617 case CK_ARCProduceObject: 618 case CK_ARCConsumeObject: 619 case CK_ARCReclaimReturnedObject: 620 case CK_ARCExtendBlockObject: 621 case CK_CopyAndAutoreleaseBlockObject: 622 llvm_unreachable("cast kind invalid for aggregate types"); 623 } 624} 625 626void AggExprEmitter::VisitCallExpr(const CallExpr *E) { 627 if (E->getCallReturnType()->isReferenceType()) { 628 EmitAggLoadOfLValue(E); 629 return; 630 } 631 632 RValue RV = CGF.EmitCallExpr(E, getReturnValueSlot()); 633 EmitMoveFromReturnSlot(E, RV); 634} 635 636void AggExprEmitter::VisitObjCMessageExpr(ObjCMessageExpr *E) { 637 RValue RV = CGF.EmitObjCMessageExpr(E, getReturnValueSlot()); 638 EmitMoveFromReturnSlot(E, RV); 639} 640 641void AggExprEmitter::VisitBinComma(const BinaryOperator *E) { 642 CGF.EmitIgnoredExpr(E->getLHS()); 643 Visit(E->getRHS()); 644} 645 646void AggExprEmitter::VisitStmtExpr(const StmtExpr *E) { 647 CodeGenFunction::StmtExprEvaluation eval(CGF); 648 CGF.EmitCompoundStmt(*E->getSubStmt(), true, Dest); 649} 650 651void AggExprEmitter::VisitBinaryOperator(const BinaryOperator *E) { 652 if (E->getOpcode() == BO_PtrMemD || E->getOpcode() == BO_PtrMemI) 653 VisitPointerToDataMemberBinaryOperator(E); 654 else 655 CGF.ErrorUnsupported(E, "aggregate binary expression"); 656} 657 658void AggExprEmitter::VisitPointerToDataMemberBinaryOperator( 659 const BinaryOperator *E) { 660 LValue LV = CGF.EmitPointerToDataMemberBinaryExpr(E); 661 EmitFinalDestCopy(E, LV); 662} 663 664void AggExprEmitter::VisitBinAssign(const BinaryOperator *E) { 665 // For an assignment to work, the value on the right has 666 // to be compatible with the value on the left. 667 assert(CGF.getContext().hasSameUnqualifiedType(E->getLHS()->getType(), 668 E->getRHS()->getType()) 669 && "Invalid assignment"); 670 671 if (const DeclRefExpr *DRE = dyn_cast<DeclRefExpr>(E->getLHS())) 672 if (const VarDecl *VD = dyn_cast<VarDecl>(DRE->getDecl())) 673 if (VD->hasAttr<BlocksAttr>() && 674 E->getRHS()->HasSideEffects(CGF.getContext())) { 675 // When __block variable on LHS, the RHS must be evaluated first 676 // as it may change the 'forwarding' field via call to Block_copy. 677 LValue RHS = CGF.EmitLValue(E->getRHS()); 678 LValue LHS = CGF.EmitLValue(E->getLHS()); 679 Dest = AggValueSlot::forLValue(LHS, AggValueSlot::IsDestructed, 680 needsGC(E->getLHS()->getType()), 681 AggValueSlot::IsAliased); 682 EmitFinalDestCopy(E, RHS, true); 683 return; 684 } 685 686 LValue LHS = CGF.EmitLValue(E->getLHS()); 687 688 // Codegen the RHS so that it stores directly into the LHS. 689 AggValueSlot LHSSlot = 690 AggValueSlot::forLValue(LHS, AggValueSlot::IsDestructed, 691 needsGC(E->getLHS()->getType()), 692 AggValueSlot::IsAliased); 693 CGF.EmitAggExpr(E->getRHS(), LHSSlot, false); 694 EmitFinalDestCopy(E, LHS, true); 695} 696 697void AggExprEmitter:: 698VisitAbstractConditionalOperator(const AbstractConditionalOperator *E) { 699 llvm::BasicBlock *LHSBlock = CGF.createBasicBlock("cond.true"); 700 llvm::BasicBlock *RHSBlock = CGF.createBasicBlock("cond.false"); 701 llvm::BasicBlock *ContBlock = CGF.createBasicBlock("cond.end"); 702 703 // Bind the common expression if necessary. 704 CodeGenFunction::OpaqueValueMapping binding(CGF, E); 705 706 CodeGenFunction::ConditionalEvaluation eval(CGF); 707 CGF.EmitBranchOnBoolExpr(E->getCond(), LHSBlock, RHSBlock); 708 709 // Save whether the destination's lifetime is externally managed. 710 bool isExternallyDestructed = Dest.isExternallyDestructed(); 711 712 eval.begin(CGF); 713 CGF.EmitBlock(LHSBlock); 714 Visit(E->getTrueExpr()); 715 eval.end(CGF); 716 717 assert(CGF.HaveInsertPoint() && "expression evaluation ended with no IP!"); 718 CGF.Builder.CreateBr(ContBlock); 719 720 // If the result of an agg expression is unused, then the emission 721 // of the LHS might need to create a destination slot. That's fine 722 // with us, and we can safely emit the RHS into the same slot, but 723 // we shouldn't claim that it's already being destructed. 724 Dest.setExternallyDestructed(isExternallyDestructed); 725 726 eval.begin(CGF); 727 CGF.EmitBlock(RHSBlock); 728 Visit(E->getFalseExpr()); 729 eval.end(CGF); 730 731 CGF.EmitBlock(ContBlock); 732} 733 734void AggExprEmitter::VisitChooseExpr(const ChooseExpr *CE) { 735 Visit(CE->getChosenSubExpr(CGF.getContext())); 736} 737 738void AggExprEmitter::VisitVAArgExpr(VAArgExpr *VE) { 739 llvm::Value *ArgValue = CGF.EmitVAListRef(VE->getSubExpr()); 740 llvm::Value *ArgPtr = CGF.EmitVAArg(ArgValue, VE->getType()); 741 742 if (!ArgPtr) { 743 CGF.ErrorUnsupported(VE, "aggregate va_arg expression"); 744 return; 745 } 746 747 EmitFinalDestCopy(VE, CGF.MakeAddrLValue(ArgPtr, VE->getType())); 748} 749 750void AggExprEmitter::VisitCXXBindTemporaryExpr(CXXBindTemporaryExpr *E) { 751 // Ensure that we have a slot, but if we already do, remember 752 // whether it was externally destructed. 753 bool wasExternallyDestructed = Dest.isExternallyDestructed(); 754 Dest = EnsureSlot(E->getType()); 755 756 // We're going to push a destructor if there isn't already one. 757 Dest.setExternallyDestructed(); 758 759 Visit(E->getSubExpr()); 760 761 // Push that destructor we promised. 762 if (!wasExternallyDestructed) 763 CGF.EmitCXXTemporary(E->getTemporary(), E->getType(), Dest.getAddr()); 764} 765 766void 767AggExprEmitter::VisitCXXConstructExpr(const CXXConstructExpr *E) { 768 AggValueSlot Slot = EnsureSlot(E->getType()); 769 CGF.EmitCXXConstructExpr(E, Slot); 770} 771 772void 773AggExprEmitter::VisitLambdaExpr(LambdaExpr *E) { 774 AggValueSlot Slot = EnsureSlot(E->getType()); 775 CGF.EmitLambdaExpr(E, Slot); 776} 777 778void AggExprEmitter::VisitExprWithCleanups(ExprWithCleanups *E) { 779 CGF.enterFullExpression(E); 780 CodeGenFunction::RunCleanupsScope cleanups(CGF); 781 Visit(E->getSubExpr()); 782} 783 784void AggExprEmitter::VisitCXXScalarValueInitExpr(CXXScalarValueInitExpr *E) { 785 QualType T = E->getType(); 786 AggValueSlot Slot = EnsureSlot(T); 787 EmitNullInitializationToLValue(CGF.MakeAddrLValue(Slot.getAddr(), T)); 788} 789 790void AggExprEmitter::VisitImplicitValueInitExpr(ImplicitValueInitExpr *E) { 791 QualType T = E->getType(); 792 AggValueSlot Slot = EnsureSlot(T); 793 EmitNullInitializationToLValue(CGF.MakeAddrLValue(Slot.getAddr(), T)); 794} 795 796/// isSimpleZero - If emitting this value will obviously just cause a store of 797/// zero to memory, return true. This can return false if uncertain, so it just 798/// handles simple cases. 799static bool isSimpleZero(const Expr *E, CodeGenFunction &CGF) { 800 E = E->IgnoreParens(); 801 802 // 0 803 if (const IntegerLiteral *IL = dyn_cast<IntegerLiteral>(E)) 804 return IL->getValue() == 0; 805 // +0.0 806 if (const FloatingLiteral *FL = dyn_cast<FloatingLiteral>(E)) 807 return FL->getValue().isPosZero(); 808 // int() 809 if ((isa<ImplicitValueInitExpr>(E) || isa<CXXScalarValueInitExpr>(E)) && 810 CGF.getTypes().isZeroInitializable(E->getType())) 811 return true; 812 // (int*)0 - Null pointer expressions. 813 if (const CastExpr *ICE = dyn_cast<CastExpr>(E)) 814 return ICE->getCastKind() == CK_NullToPointer; 815 // '\0' 816 if (const CharacterLiteral *CL = dyn_cast<CharacterLiteral>(E)) 817 return CL->getValue() == 0; 818 819 // Otherwise, hard case: conservatively return false. 820 return false; 821} 822 823 824void 825AggExprEmitter::EmitInitializationToLValue(Expr* E, LValue LV) { 826 QualType type = LV.getType(); 827 // FIXME: Ignore result? 828 // FIXME: Are initializers affected by volatile? 829 if (Dest.isZeroed() && isSimpleZero(E, CGF)) { 830 // Storing "i32 0" to a zero'd memory location is a noop. 831 } else if (isa<ImplicitValueInitExpr>(E)) { 832 EmitNullInitializationToLValue(LV); 833 } else if (type->isReferenceType()) { 834 RValue RV = CGF.EmitReferenceBindingToExpr(E, /*InitializedDecl=*/0); 835 CGF.EmitStoreThroughLValue(RV, LV); 836 } else if (type->isAnyComplexType()) { 837 CGF.EmitComplexExprIntoAddr(E, LV.getAddress(), false); 838 } else if (CGF.hasAggregateLLVMType(type)) { 839 CGF.EmitAggExpr(E, AggValueSlot::forLValue(LV, 840 AggValueSlot::IsDestructed, 841 AggValueSlot::DoesNotNeedGCBarriers, 842 AggValueSlot::IsNotAliased, 843 Dest.isZeroed())); 844 } else if (LV.isSimple()) { 845 CGF.EmitScalarInit(E, /*D=*/0, LV, /*Captured=*/false); 846 } else { 847 CGF.EmitStoreThroughLValue(RValue::get(CGF.EmitScalarExpr(E)), LV); 848 } 849} 850 851void AggExprEmitter::EmitNullInitializationToLValue(LValue lv) { 852 QualType type = lv.getType(); 853 854 // If the destination slot is already zeroed out before the aggregate is 855 // copied into it, we don't have to emit any zeros here. 856 if (Dest.isZeroed() && CGF.getTypes().isZeroInitializable(type)) 857 return; 858 859 if (!CGF.hasAggregateLLVMType(type)) { 860 // For non-aggregates, we can store zero. 861 llvm::Value *null = llvm::Constant::getNullValue(CGF.ConvertType(type)); 862 // Note that the following is not equivalent to 863 // EmitStoreThroughBitfieldLValue for ARC types. 864 if (lv.isBitField()) { 865 CGF.EmitStoreThroughBitfieldLValue(RValue::get(null), lv); 866 } else { 867 assert(lv.isSimple()); 868 CGF.EmitStoreOfScalar(null, lv, /* isInitialization */ true); 869 } 870 } else { 871 // There's a potential optimization opportunity in combining 872 // memsets; that would be easy for arrays, but relatively 873 // difficult for structures with the current code. 874 CGF.EmitNullInitialization(lv.getAddress(), lv.getType()); 875 } 876} 877 878void AggExprEmitter::VisitInitListExpr(InitListExpr *E) { 879#if 0 880 // FIXME: Assess perf here? Figure out what cases are worth optimizing here 881 // (Length of globals? Chunks of zeroed-out space?). 882 // 883 // If we can, prefer a copy from a global; this is a lot less code for long 884 // globals, and it's easier for the current optimizers to analyze. 885 if (llvm::Constant* C = CGF.CGM.EmitConstantExpr(E, E->getType(), &CGF)) { 886 llvm::GlobalVariable* GV = 887 new llvm::GlobalVariable(CGF.CGM.getModule(), C->getType(), true, 888 llvm::GlobalValue::InternalLinkage, C, ""); 889 EmitFinalDestCopy(E, CGF.MakeAddrLValue(GV, E->getType())); 890 return; 891 } 892#endif 893 if (E->hadArrayRangeDesignator()) 894 CGF.ErrorUnsupported(E, "GNU array range designator extension"); 895 896 if (E->initializesStdInitializerList()) { 897 EmitStdInitializerList(Dest.getAddr(), E); 898 return; 899 } 900 901 llvm::Value *DestPtr = Dest.getAddr(); 902 903 // Handle initialization of an array. 904 if (E->getType()->isArrayType()) { 905 if (E->getNumInits() > 0) { 906 QualType T1 = E->getType(); 907 QualType T2 = E->getInit(0)->getType(); 908 if (CGF.getContext().hasSameUnqualifiedType(T1, T2)) { 909 EmitAggLoadOfLValue(E->getInit(0)); 910 return; 911 } 912 } 913 914 QualType elementType = 915 CGF.getContext().getAsArrayType(E->getType())->getElementType(); 916 917 llvm::PointerType *APType = 918 cast<llvm::PointerType>(DestPtr->getType()); 919 llvm::ArrayType *AType = 920 cast<llvm::ArrayType>(APType->getElementType()); 921 922 EmitArrayInit(DestPtr, AType, elementType, E); 923 return; 924 } 925 926 assert(E->getType()->isRecordType() && "Only support structs/unions here!"); 927 928 // Do struct initialization; this code just sets each individual member 929 // to the approprate value. This makes bitfield support automatic; 930 // the disadvantage is that the generated code is more difficult for 931 // the optimizer, especially with bitfields. 932 unsigned NumInitElements = E->getNumInits(); 933 RecordDecl *record = E->getType()->castAs<RecordType>()->getDecl(); 934 935 if (record->isUnion()) { 936 // Only initialize one field of a union. The field itself is 937 // specified by the initializer list. 938 if (!E->getInitializedFieldInUnion()) { 939 // Empty union; we have nothing to do. 940 941#ifndef NDEBUG 942 // Make sure that it's really an empty and not a failure of 943 // semantic analysis. 944 for (RecordDecl::field_iterator Field = record->field_begin(), 945 FieldEnd = record->field_end(); 946 Field != FieldEnd; ++Field) 947 assert(Field->isUnnamedBitfield() && "Only unnamed bitfields allowed"); 948#endif 949 return; 950 } 951 952 // FIXME: volatility 953 FieldDecl *Field = E->getInitializedFieldInUnion(); 954 955 LValue FieldLoc = CGF.EmitLValueForFieldInitialization(DestPtr, Field, 0); 956 if (NumInitElements) { 957 // Store the initializer into the field 958 EmitInitializationToLValue(E->getInit(0), FieldLoc); 959 } else { 960 // Default-initialize to null. 961 EmitNullInitializationToLValue(FieldLoc); 962 } 963 964 return; 965 } 966 967 // We'll need to enter cleanup scopes in case any of the member 968 // initializers throw an exception. 969 SmallVector<EHScopeStack::stable_iterator, 16> cleanups; 970 llvm::Instruction *cleanupDominator = 0; 971 972 // Here we iterate over the fields; this makes it simpler to both 973 // default-initialize fields and skip over unnamed fields. 974 unsigned curInitIndex = 0; 975 for (RecordDecl::field_iterator field = record->field_begin(), 976 fieldEnd = record->field_end(); 977 field != fieldEnd; ++field) { 978 // We're done once we hit the flexible array member. 979 if (field->getType()->isIncompleteArrayType()) 980 break; 981 982 // Always skip anonymous bitfields. 983 if (field->isUnnamedBitfield()) 984 continue; 985 986 // We're done if we reach the end of the explicit initializers, we 987 // have a zeroed object, and the rest of the fields are 988 // zero-initializable. 989 if (curInitIndex == NumInitElements && Dest.isZeroed() && 990 CGF.getTypes().isZeroInitializable(E->getType())) 991 break; 992 993 // FIXME: volatility 994 LValue LV = CGF.EmitLValueForFieldInitialization(DestPtr, *field, 0); 995 // We never generate write-barries for initialized fields. 996 LV.setNonGC(true); 997 998 if (curInitIndex < NumInitElements) { 999 // Store the initializer into the field. 1000 EmitInitializationToLValue(E->getInit(curInitIndex++), LV); 1001 } else { 1002 // We're out of initalizers; default-initialize to null 1003 EmitNullInitializationToLValue(LV); 1004 } 1005 1006 // Push a destructor if necessary. 1007 // FIXME: if we have an array of structures, all explicitly 1008 // initialized, we can end up pushing a linear number of cleanups. 1009 bool pushedCleanup = false; 1010 if (QualType::DestructionKind dtorKind 1011 = field->getType().isDestructedType()) { 1012 assert(LV.isSimple()); 1013 if (CGF.needsEHCleanup(dtorKind)) { 1014 if (!cleanupDominator) 1015 cleanupDominator = CGF.Builder.CreateUnreachable(); // placeholder 1016 1017 CGF.pushDestroy(EHCleanup, LV.getAddress(), field->getType(), 1018 CGF.getDestroyer(dtorKind), false); 1019 cleanups.push_back(CGF.EHStack.stable_begin()); 1020 pushedCleanup = true; 1021 } 1022 } 1023 1024 // If the GEP didn't get used because of a dead zero init or something 1025 // else, clean it up for -O0 builds and general tidiness. 1026 if (!pushedCleanup && LV.isSimple()) 1027 if (llvm::GetElementPtrInst *GEP = 1028 dyn_cast<llvm::GetElementPtrInst>(LV.getAddress())) 1029 if (GEP->use_empty()) 1030 GEP->eraseFromParent(); 1031 } 1032 1033 // Deactivate all the partial cleanups in reverse order, which 1034 // generally means popping them. 1035 for (unsigned i = cleanups.size(); i != 0; --i) 1036 CGF.DeactivateCleanupBlock(cleanups[i-1], cleanupDominator); 1037 1038 // Destroy the placeholder if we made one. 1039 if (cleanupDominator) 1040 cleanupDominator->eraseFromParent(); 1041} 1042 1043//===----------------------------------------------------------------------===// 1044// Entry Points into this File 1045//===----------------------------------------------------------------------===// 1046 1047/// GetNumNonZeroBytesInInit - Get an approximate count of the number of 1048/// non-zero bytes that will be stored when outputting the initializer for the 1049/// specified initializer expression. 1050static CharUnits GetNumNonZeroBytesInInit(const Expr *E, CodeGenFunction &CGF) { 1051 E = E->IgnoreParens(); 1052 1053 // 0 and 0.0 won't require any non-zero stores! 1054 if (isSimpleZero(E, CGF)) return CharUnits::Zero(); 1055 1056 // If this is an initlist expr, sum up the size of sizes of the (present) 1057 // elements. If this is something weird, assume the whole thing is non-zero. 1058 const InitListExpr *ILE = dyn_cast<InitListExpr>(E); 1059 if (ILE == 0 || !CGF.getTypes().isZeroInitializable(ILE->getType())) 1060 return CGF.getContext().getTypeSizeInChars(E->getType()); 1061 1062 // InitListExprs for structs have to be handled carefully. If there are 1063 // reference members, we need to consider the size of the reference, not the 1064 // referencee. InitListExprs for unions and arrays can't have references. 1065 if (const RecordType *RT = E->getType()->getAs<RecordType>()) { 1066 if (!RT->isUnionType()) { 1067 RecordDecl *SD = E->getType()->getAs<RecordType>()->getDecl(); 1068 CharUnits NumNonZeroBytes = CharUnits::Zero(); 1069 1070 unsigned ILEElement = 0; 1071 for (RecordDecl::field_iterator Field = SD->field_begin(), 1072 FieldEnd = SD->field_end(); Field != FieldEnd; ++Field) { 1073 // We're done once we hit the flexible array member or run out of 1074 // InitListExpr elements. 1075 if (Field->getType()->isIncompleteArrayType() || 1076 ILEElement == ILE->getNumInits()) 1077 break; 1078 if (Field->isUnnamedBitfield()) 1079 continue; 1080 1081 const Expr *E = ILE->getInit(ILEElement++); 1082 1083 // Reference values are always non-null and have the width of a pointer. 1084 if (Field->getType()->isReferenceType()) 1085 NumNonZeroBytes += CGF.getContext().toCharUnitsFromBits( 1086 CGF.getContext().getTargetInfo().getPointerWidth(0)); 1087 else 1088 NumNonZeroBytes += GetNumNonZeroBytesInInit(E, CGF); 1089 } 1090 1091 return NumNonZeroBytes; 1092 } 1093 } 1094 1095 1096 CharUnits NumNonZeroBytes = CharUnits::Zero(); 1097 for (unsigned i = 0, e = ILE->getNumInits(); i != e; ++i) 1098 NumNonZeroBytes += GetNumNonZeroBytesInInit(ILE->getInit(i), CGF); 1099 return NumNonZeroBytes; 1100} 1101 1102/// CheckAggExprForMemSetUse - If the initializer is large and has a lot of 1103/// zeros in it, emit a memset and avoid storing the individual zeros. 1104/// 1105static void CheckAggExprForMemSetUse(AggValueSlot &Slot, const Expr *E, 1106 CodeGenFunction &CGF) { 1107 // If the slot is already known to be zeroed, nothing to do. Don't mess with 1108 // volatile stores. 1109 if (Slot.isZeroed() || Slot.isVolatile() || Slot.getAddr() == 0) return; 1110 1111 // C++ objects with a user-declared constructor don't need zero'ing. 1112 if (CGF.getContext().getLangOptions().CPlusPlus) 1113 if (const RecordType *RT = CGF.getContext() 1114 .getBaseElementType(E->getType())->getAs<RecordType>()) { 1115 const CXXRecordDecl *RD = cast<CXXRecordDecl>(RT->getDecl()); 1116 if (RD->hasUserDeclaredConstructor()) 1117 return; 1118 } 1119 1120 // If the type is 16-bytes or smaller, prefer individual stores over memset. 1121 std::pair<CharUnits, CharUnits> TypeInfo = 1122 CGF.getContext().getTypeInfoInChars(E->getType()); 1123 if (TypeInfo.first <= CharUnits::fromQuantity(16)) 1124 return; 1125 1126 // Check to see if over 3/4 of the initializer are known to be zero. If so, 1127 // we prefer to emit memset + individual stores for the rest. 1128 CharUnits NumNonZeroBytes = GetNumNonZeroBytesInInit(E, CGF); 1129 if (NumNonZeroBytes*4 > TypeInfo.first) 1130 return; 1131 1132 // Okay, it seems like a good idea to use an initial memset, emit the call. 1133 llvm::Constant *SizeVal = CGF.Builder.getInt64(TypeInfo.first.getQuantity()); 1134 CharUnits Align = TypeInfo.second; 1135 1136 llvm::Value *Loc = Slot.getAddr(); 1137 1138 Loc = CGF.Builder.CreateBitCast(Loc, CGF.Int8PtrTy); 1139 CGF.Builder.CreateMemSet(Loc, CGF.Builder.getInt8(0), SizeVal, 1140 Align.getQuantity(), false); 1141 1142 // Tell the AggExprEmitter that the slot is known zero. 1143 Slot.setZeroed(); 1144} 1145 1146 1147 1148 1149/// EmitAggExpr - Emit the computation of the specified expression of aggregate 1150/// type. The result is computed into DestPtr. Note that if DestPtr is null, 1151/// the value of the aggregate expression is not needed. If VolatileDest is 1152/// true, DestPtr cannot be 0. 1153/// 1154/// \param IsInitializer - true if this evaluation is initializing an 1155/// object whose lifetime is already being managed. 1156void CodeGenFunction::EmitAggExpr(const Expr *E, AggValueSlot Slot, 1157 bool IgnoreResult) { 1158 assert(E && hasAggregateLLVMType(E->getType()) && 1159 "Invalid aggregate expression to emit"); 1160 assert((Slot.getAddr() != 0 || Slot.isIgnored()) && 1161 "slot has bits but no address"); 1162 1163 // Optimize the slot if possible. 1164 CheckAggExprForMemSetUse(Slot, E, *this); 1165 1166 AggExprEmitter(*this, Slot, IgnoreResult).Visit(const_cast<Expr*>(E)); 1167} 1168 1169LValue CodeGenFunction::EmitAggExprToLValue(const Expr *E) { 1170 assert(hasAggregateLLVMType(E->getType()) && "Invalid argument!"); 1171 llvm::Value *Temp = CreateMemTemp(E->getType()); 1172 LValue LV = MakeAddrLValue(Temp, E->getType()); 1173 EmitAggExpr(E, AggValueSlot::forLValue(LV, AggValueSlot::IsNotDestructed, 1174 AggValueSlot::DoesNotNeedGCBarriers, 1175 AggValueSlot::IsNotAliased)); 1176 return LV; 1177} 1178 1179void CodeGenFunction::EmitAggregateCopy(llvm::Value *DestPtr, 1180 llvm::Value *SrcPtr, QualType Ty, 1181 bool isVolatile, unsigned Alignment) { 1182 assert(!Ty->isAnyComplexType() && "Shouldn't happen for complex"); 1183 1184 if (getContext().getLangOptions().CPlusPlus) { 1185 if (const RecordType *RT = Ty->getAs<RecordType>()) { 1186 CXXRecordDecl *Record = cast<CXXRecordDecl>(RT->getDecl()); 1187 assert((Record->hasTrivialCopyConstructor() || 1188 Record->hasTrivialCopyAssignment() || 1189 Record->hasTrivialMoveConstructor() || 1190 Record->hasTrivialMoveAssignment()) && 1191 "Trying to aggregate-copy a type without a trivial copy " 1192 "constructor or assignment operator"); 1193 // Ignore empty classes in C++. 1194 if (Record->isEmpty()) 1195 return; 1196 } 1197 } 1198 1199 // Aggregate assignment turns into llvm.memcpy. This is almost valid per 1200 // C99 6.5.16.1p3, which states "If the value being stored in an object is 1201 // read from another object that overlaps in anyway the storage of the first 1202 // object, then the overlap shall be exact and the two objects shall have 1203 // qualified or unqualified versions of a compatible type." 1204 // 1205 // memcpy is not defined if the source and destination pointers are exactly 1206 // equal, but other compilers do this optimization, and almost every memcpy 1207 // implementation handles this case safely. If there is a libc that does not 1208 // safely handle this, we can add a target hook. 1209 1210 // Get size and alignment info for this aggregate. 1211 std::pair<CharUnits, CharUnits> TypeInfo = 1212 getContext().getTypeInfoInChars(Ty); 1213 1214 if (!Alignment) 1215 Alignment = TypeInfo.second.getQuantity(); 1216 1217 // FIXME: Handle variable sized types. 1218 1219 // FIXME: If we have a volatile struct, the optimizer can remove what might 1220 // appear to be `extra' memory ops: 1221 // 1222 // volatile struct { int i; } a, b; 1223 // 1224 // int main() { 1225 // a = b; 1226 // a = b; 1227 // } 1228 // 1229 // we need to use a different call here. We use isVolatile to indicate when 1230 // either the source or the destination is volatile. 1231 1232 llvm::PointerType *DPT = cast<llvm::PointerType>(DestPtr->getType()); 1233 llvm::Type *DBP = 1234 llvm::Type::getInt8PtrTy(getLLVMContext(), DPT->getAddressSpace()); 1235 DestPtr = Builder.CreateBitCast(DestPtr, DBP); 1236 1237 llvm::PointerType *SPT = cast<llvm::PointerType>(SrcPtr->getType()); 1238 llvm::Type *SBP = 1239 llvm::Type::getInt8PtrTy(getLLVMContext(), SPT->getAddressSpace()); 1240 SrcPtr = Builder.CreateBitCast(SrcPtr, SBP); 1241 1242 // Don't do any of the memmove_collectable tests if GC isn't set. 1243 if (CGM.getLangOptions().getGC() == LangOptions::NonGC) { 1244 // fall through 1245 } else if (const RecordType *RecordTy = Ty->getAs<RecordType>()) { 1246 RecordDecl *Record = RecordTy->getDecl(); 1247 if (Record->hasObjectMember()) { 1248 CharUnits size = TypeInfo.first; 1249 llvm::Type *SizeTy = ConvertType(getContext().getSizeType()); 1250 llvm::Value *SizeVal = llvm::ConstantInt::get(SizeTy, size.getQuantity()); 1251 CGM.getObjCRuntime().EmitGCMemmoveCollectable(*this, DestPtr, SrcPtr, 1252 SizeVal); 1253 return; 1254 } 1255 } else if (Ty->isArrayType()) { 1256 QualType BaseType = getContext().getBaseElementType(Ty); 1257 if (const RecordType *RecordTy = BaseType->getAs<RecordType>()) { 1258 if (RecordTy->getDecl()->hasObjectMember()) { 1259 CharUnits size = TypeInfo.first; 1260 llvm::Type *SizeTy = ConvertType(getContext().getSizeType()); 1261 llvm::Value *SizeVal = 1262 llvm::ConstantInt::get(SizeTy, size.getQuantity()); 1263 CGM.getObjCRuntime().EmitGCMemmoveCollectable(*this, DestPtr, SrcPtr, 1264 SizeVal); 1265 return; 1266 } 1267 } 1268 } 1269 1270 Builder.CreateMemCpy(DestPtr, SrcPtr, 1271 llvm::ConstantInt::get(IntPtrTy, 1272 TypeInfo.first.getQuantity()), 1273 Alignment, isVolatile); 1274} 1275 1276void CodeGenFunction::MaybeEmitStdInitializerListCleanup(llvm::Value *loc, 1277 const Expr *init) { 1278 const ExprWithCleanups *cleanups = dyn_cast<ExprWithCleanups>(init); 1279 if (cleanups) 1280 init = cleanups->getSubExpr(); 1281 1282 if (isa<InitListExpr>(init) && 1283 cast<InitListExpr>(init)->initializesStdInitializerList()) { 1284 // We initialized this std::initializer_list with an initializer list. 1285 // A backing array was created. Push a cleanup for it. 1286 EmitStdInitializerListCleanup(loc, cast<InitListExpr>(init)); 1287 } 1288} 1289 1290static void EmitRecursiveStdInitializerListCleanup(CodeGenFunction &CGF, 1291 llvm::Value *arrayStart, 1292 const InitListExpr *init) { 1293 // Check if there are any recursive cleanups to do, i.e. if we have 1294 // std::initializer_list<std::initializer_list<obj>> list = {{obj()}}; 1295 // then we need to destroy the inner array as well. 1296 for (unsigned i = 0, e = init->getNumInits(); i != e; ++i) { 1297 const InitListExpr *subInit = dyn_cast<InitListExpr>(init->getInit(i)); 1298 if (!subInit || !subInit->initializesStdInitializerList()) 1299 continue; 1300 1301 // This one needs to be destroyed. Get the address of the std::init_list. 1302 llvm::Value *offset = llvm::ConstantInt::get(CGF.SizeTy, i); 1303 llvm::Value *loc = CGF.Builder.CreateInBoundsGEP(arrayStart, offset, 1304 "std.initlist"); 1305 CGF.EmitStdInitializerListCleanup(loc, subInit); 1306 } 1307} 1308 1309void CodeGenFunction::EmitStdInitializerListCleanup(llvm::Value *loc, 1310 const InitListExpr *init) { 1311 ASTContext &ctx = getContext(); 1312 QualType element = GetStdInitializerListElementType(init->getType()); 1313 unsigned numInits = init->getNumInits(); 1314 llvm::APInt size(ctx.getTypeSize(ctx.getSizeType()), numInits); 1315 QualType array =ctx.getConstantArrayType(element, size, ArrayType::Normal, 0); 1316 QualType arrayPtr = ctx.getPointerType(array); 1317 llvm::Type *arrayPtrType = ConvertType(arrayPtr); 1318 1319 // lvalue is the location of a std::initializer_list, which as its first 1320 // element has a pointer to the array we want to destroy. 1321 llvm::Value *startPointer = Builder.CreateStructGEP(loc, 0, "startPointer"); 1322 llvm::Value *startAddress = Builder.CreateLoad(startPointer, "startAddress"); 1323 1324 ::EmitRecursiveStdInitializerListCleanup(*this, startAddress, init); 1325 1326 llvm::Value *arrayAddress = 1327 Builder.CreateBitCast(startAddress, arrayPtrType, "arrayAddress"); 1328 ::EmitStdInitializerListCleanup(*this, array, arrayAddress, init); 1329} 1330