1//===--- CGExprAgg.cpp - Emit LLVM Code from Aggregate Expressions --------===//
2//
3// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4// See https://llvm.org/LICENSE.txt for license information.
5// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6//
7//===----------------------------------------------------------------------===//
8//
9// This contains code to emit Aggregate Expr nodes as LLVM code.
10//
11//===----------------------------------------------------------------------===//
12
13#include "CGCXXABI.h"
14#include "CGDebugInfo.h"
15#include "CGHLSLRuntime.h"
16#include "CGObjCRuntime.h"
17#include "CGRecordLayout.h"
18#include "CodeGenFunction.h"
19#include "CodeGenModule.h"
20#include "ConstantEmitter.h"
21#include "EHScopeStack.h"
22#include "TargetInfo.h"
23#include "clang/AST/ASTContext.h"
24#include "clang/AST/Attr.h"
25#include "clang/AST/DeclCXX.h"
26#include "clang/AST/DeclTemplate.h"
27#include "clang/AST/StmtVisitor.h"
28#include "llvm/IR/Constants.h"
29#include "llvm/IR/Function.h"
30#include "llvm/IR/GlobalVariable.h"
31#include "llvm/IR/Instruction.h"
32#include "llvm/IR/IntrinsicInst.h"
33#include "llvm/IR/Intrinsics.h"
34using namespace clang;
35using namespace CodeGen;
36
37//===----------------------------------------------------------------------===//
38// Aggregate Expression Emitter
39//===----------------------------------------------------------------------===//
40
41namespace {
42class AggExprEmitter : public StmtVisitor<AggExprEmitter> {
43 CodeGenFunction &CGF;
44 CGBuilderTy &Builder;
45 AggValueSlot Dest;
46 bool IsResultUnused;
47
48 AggValueSlot EnsureSlot(QualType T) {
49 if (!Dest.isIgnored())
50 return Dest;
51 return CGF.CreateAggTemp(T, Name: "agg.tmp.ensured");
52 }
53 void EnsureDest(QualType T) {
54 if (!Dest.isIgnored())
55 return;
56 Dest = CGF.CreateAggTemp(T, Name: "agg.tmp.ensured");
57 }
58
59 // Calls `Fn` with a valid return value slot, potentially creating a temporary
60 // to do so. If a temporary is created, an appropriate copy into `Dest` will
61 // be emitted, as will lifetime markers.
62 //
63 // The given function should take a ReturnValueSlot, and return an RValue that
64 // points to said slot.
65 void withReturnValueSlot(const Expr *E,
66 llvm::function_ref<RValue(ReturnValueSlot)> Fn);
67
68 void DoZeroInitPadding(uint64_t &PaddingStart, uint64_t PaddingEnd,
69 const FieldDecl *NextField);
70
71public:
72 AggExprEmitter(CodeGenFunction &cgf, AggValueSlot Dest, bool IsResultUnused)
73 : CGF(cgf), Builder(CGF.Builder), Dest(Dest),
74 IsResultUnused(IsResultUnused) {}
75
76 //===--------------------------------------------------------------------===//
77 // Utilities
78 //===--------------------------------------------------------------------===//
79
80 /// EmitAggLoadOfLValue - Given an expression with aggregate type that
81 /// represents a value lvalue, this method emits the address of the lvalue,
82 /// then loads the result into DestPtr.
83 void EmitAggLoadOfLValue(const Expr *E);
84
85 /// EmitFinalDestCopy - Perform the final copy to DestPtr, if desired.
86 /// SrcIsRValue is true if source comes from an RValue.
87 void EmitFinalDestCopy(QualType type, const LValue &src,
88 CodeGenFunction::ExprValueKind SrcValueKind =
89 CodeGenFunction::EVK_NonRValue);
90 void EmitFinalDestCopy(QualType type, RValue src);
91 void EmitCopy(QualType type, const AggValueSlot &dest,
92 const AggValueSlot &src);
93
94 void EmitArrayInit(Address DestPtr, llvm::ArrayType *AType, QualType ArrayQTy,
95 Expr *ExprToVisit, ArrayRef<Expr *> Args,
96 Expr *ArrayFiller);
97
98 AggValueSlot::NeedsGCBarriers_t needsGC(QualType T) {
99 if (CGF.getLangOpts().getGC() && TypeRequiresGCollection(T))
100 return AggValueSlot::NeedsGCBarriers;
101 return AggValueSlot::DoesNotNeedGCBarriers;
102 }
103
104 bool TypeRequiresGCollection(QualType T);
105
106 //===--------------------------------------------------------------------===//
107 // Visitor Methods
108 //===--------------------------------------------------------------------===//
109
110 void Visit(Expr *E) {
111 ApplyDebugLocation DL(CGF, E);
112 StmtVisitor<AggExprEmitter>::Visit(S: E);
113 }
114
115 void VisitStmt(Stmt *S) { CGF.ErrorUnsupported(S, Type: "aggregate expression"); }
116 void VisitParenExpr(ParenExpr *PE) { Visit(E: PE->getSubExpr()); }
117 void VisitGenericSelectionExpr(GenericSelectionExpr *GE) {
118 Visit(E: GE->getResultExpr());
119 }
120 void VisitCoawaitExpr(CoawaitExpr *E) {
121 CGF.EmitCoawaitExpr(E: *E, aggSlot: Dest, ignoreResult: IsResultUnused);
122 }
123 void VisitCoyieldExpr(CoyieldExpr *E) {
124 CGF.EmitCoyieldExpr(E: *E, aggSlot: Dest, ignoreResult: IsResultUnused);
125 }
126 void VisitUnaryCoawait(UnaryOperator *E) { Visit(E: E->getSubExpr()); }
127 void VisitUnaryExtension(UnaryOperator *E) { Visit(E: E->getSubExpr()); }
128 void VisitSubstNonTypeTemplateParmExpr(SubstNonTypeTemplateParmExpr *E) {
129 return Visit(E: E->getReplacement());
130 }
131
132 void VisitConstantExpr(ConstantExpr *E) {
133 EnsureDest(T: E->getType());
134
135 if (llvm::Value *Result = ConstantEmitter(CGF).tryEmitConstantExpr(CE: E)) {
136 CGF.CreateCoercedStore(
137 Src: Result, Dst: Dest.getAddress(),
138 DstSize: llvm::TypeSize::getFixed(
139 ExactSize: Dest.getPreferredSize(Ctx&: CGF.getContext(), Type: E->getType())
140 .getQuantity()),
141 DstIsVolatile: E->getType().isVolatileQualified());
142 return;
143 }
144 return Visit(E: E->getSubExpr());
145 }
146
147 // l-values.
148 void VisitDeclRefExpr(DeclRefExpr *E) { EmitAggLoadOfLValue(E); }
149 void VisitMemberExpr(MemberExpr *ME) { EmitAggLoadOfLValue(E: ME); }
150 void VisitUnaryDeref(UnaryOperator *E) { EmitAggLoadOfLValue(E); }
151 void VisitStringLiteral(StringLiteral *E) { EmitAggLoadOfLValue(E); }
152 void VisitCompoundLiteralExpr(CompoundLiteralExpr *E);
153 void VisitArraySubscriptExpr(ArraySubscriptExpr *E) {
154 EmitAggLoadOfLValue(E);
155 }
156 void VisitPredefinedExpr(const PredefinedExpr *E) { EmitAggLoadOfLValue(E); }
157
158 // Operators.
159 void VisitCastExpr(CastExpr *E);
160 void VisitCallExpr(const CallExpr *E);
161 void VisitStmtExpr(const StmtExpr *E);
162 void VisitBinaryOperator(const BinaryOperator *BO);
163 void VisitPointerToDataMemberBinaryOperator(const BinaryOperator *BO);
164 void VisitBinAssign(const BinaryOperator *E);
165 void VisitBinComma(const BinaryOperator *E);
166 void VisitBinCmp(const BinaryOperator *E);
167 void VisitCXXRewrittenBinaryOperator(CXXRewrittenBinaryOperator *E) {
168 Visit(E: E->getSemanticForm());
169 }
170
171 void VisitObjCMessageExpr(ObjCMessageExpr *E);
172 void VisitObjCIvarRefExpr(ObjCIvarRefExpr *E) { EmitAggLoadOfLValue(E); }
173
174 void VisitDesignatedInitUpdateExpr(DesignatedInitUpdateExpr *E);
175 void VisitAbstractConditionalOperator(const AbstractConditionalOperator *CO);
176 void VisitChooseExpr(const ChooseExpr *CE);
177 void VisitInitListExpr(InitListExpr *E);
178 void VisitCXXParenListOrInitListExpr(Expr *ExprToVisit, ArrayRef<Expr *> Args,
179 FieldDecl *InitializedFieldInUnion,
180 Expr *ArrayFiller);
181 void VisitArrayInitLoopExpr(const ArrayInitLoopExpr *E,
182 llvm::Value *outerBegin = nullptr);
183 void VisitImplicitValueInitExpr(ImplicitValueInitExpr *E);
184 void VisitNoInitExpr(NoInitExpr *E) {} // Do nothing.
185 void VisitCXXDefaultArgExpr(CXXDefaultArgExpr *DAE) {
186 CodeGenFunction::CXXDefaultArgExprScope Scope(CGF, DAE);
187 Visit(E: DAE->getExpr());
188 }
189 void VisitCXXDefaultInitExpr(CXXDefaultInitExpr *DIE) {
190 CodeGenFunction::CXXDefaultInitExprScope Scope(CGF, DIE);
191 Visit(E: DIE->getExpr());
192 }
193 void VisitCXXBindTemporaryExpr(CXXBindTemporaryExpr *E);
194 void VisitCXXConstructExpr(const CXXConstructExpr *E);
195 void VisitCXXInheritedCtorInitExpr(const CXXInheritedCtorInitExpr *E);
196 void VisitLambdaExpr(LambdaExpr *E);
197 void VisitCXXStdInitializerListExpr(CXXStdInitializerListExpr *E);
198 void VisitExprWithCleanups(ExprWithCleanups *E);
199 void VisitCXXScalarValueInitExpr(CXXScalarValueInitExpr *E);
200 void VisitCXXTypeidExpr(CXXTypeidExpr *E) { EmitAggLoadOfLValue(E); }
201 void VisitMaterializeTemporaryExpr(MaterializeTemporaryExpr *E);
202 void VisitOpaqueValueExpr(OpaqueValueExpr *E);
203
204 void VisitPseudoObjectExpr(PseudoObjectExpr *E) {
205 if (E->isGLValue()) {
206 LValue LV = CGF.EmitPseudoObjectLValue(e: E);
207 return EmitFinalDestCopy(type: E->getType(), src: LV);
208 }
209
210 AggValueSlot Slot = EnsureSlot(T: E->getType());
211 bool NeedsDestruction =
212 !Slot.isExternallyDestructed() &&
213 E->getType().isDestructedType() == QualType::DK_nontrivial_c_struct;
214 if (NeedsDestruction)
215 Slot.setExternallyDestructed();
216 CGF.EmitPseudoObjectRValue(e: E, slot: Slot);
217 if (NeedsDestruction)
218 CGF.pushDestroy(dtorKind: QualType::DK_nontrivial_c_struct, addr: Slot.getAddress(),
219 type: E->getType());
220 }
221
222 void VisitVAArgExpr(VAArgExpr *E);
223 void VisitCXXParenListInitExpr(CXXParenListInitExpr *E);
224 void VisitCXXParenListOrInitListExpr(Expr *ExprToVisit, ArrayRef<Expr *> Args,
225 Expr *ArrayFiller);
226
227 void EmitInitializationToLValue(Expr *E, LValue Address);
228 void EmitNullInitializationToLValue(LValue Address);
229 // case Expr::ChooseExprClass:
230 void VisitCXXThrowExpr(const CXXThrowExpr *E) { CGF.EmitCXXThrowExpr(E); }
231 void VisitAtomicExpr(AtomicExpr *E) {
232 RValue Res = CGF.EmitAtomicExpr(E);
233 EmitFinalDestCopy(type: E->getType(), src: Res);
234 }
235 void VisitPackIndexingExpr(PackIndexingExpr *E) {
236 Visit(E: E->getSelectedExpr());
237 }
238};
239} // end anonymous namespace.
240
241//===----------------------------------------------------------------------===//
242// Utilities
243//===----------------------------------------------------------------------===//
244
245/// EmitAggLoadOfLValue - Given an expression with aggregate type that
246/// represents a value lvalue, this method emits the address of the lvalue,
247/// then loads the result into DestPtr.
248void AggExprEmitter::EmitAggLoadOfLValue(const Expr *E) {
249 LValue LV = CGF.EmitLValue(E);
250
251 // If the type of the l-value is atomic, then do an atomic load.
252 if (LV.getType()->isAtomicType() || CGF.LValueIsSuitableForInlineAtomic(Src: LV)) {
253 CGF.EmitAtomicLoad(LV, SL: E->getExprLoc(), Slot: Dest);
254 return;
255 }
256
257 EmitFinalDestCopy(type: E->getType(), src: LV);
258}
259
260/// True if the given aggregate type requires special GC API calls.
261bool AggExprEmitter::TypeRequiresGCollection(QualType T) {
262 // Only record types have members that might require garbage collection.
263 const auto *Record = T->getAsRecordDecl();
264 if (!Record)
265 return false;
266
267 // Don't mess with non-trivial C++ types.
268 if (isa<CXXRecordDecl>(Val: Record) &&
269 (cast<CXXRecordDecl>(Val: Record)->hasNonTrivialCopyConstructor() ||
270 !cast<CXXRecordDecl>(Val: Record)->hasTrivialDestructor()))
271 return false;
272
273 // Check whether the type has an object member.
274 return Record->hasObjectMember();
275}
276
277void AggExprEmitter::withReturnValueSlot(
278 const Expr *E, llvm::function_ref<RValue(ReturnValueSlot)> EmitCall) {
279 QualType RetTy = E->getType();
280 bool RequiresDestruction =
281 !Dest.isExternallyDestructed() &&
282 RetTy.isDestructedType() == QualType::DK_nontrivial_c_struct;
283
284 // If it makes no observable difference, save a memcpy + temporary.
285 //
286 // We need to always provide our own temporary if destruction is required.
287 // Otherwise, EmitCall will emit its own, notice that it's "unused", and end
288 // its lifetime before we have the chance to emit a proper destructor call.
289 bool UseTemp = Dest.isPotentiallyAliased() || Dest.requiresGCollection() ||
290 (RequiresDestruction && Dest.isIgnored());
291
292 Address RetAddr = Address::invalid();
293
294 EHScopeStack::stable_iterator LifetimeEndBlock;
295 llvm::IntrinsicInst *LifetimeStartInst = nullptr;
296 if (!UseTemp) {
297 RetAddr = Dest.getAddress();
298 } else {
299 RetAddr = CGF.CreateMemTempWithoutCast(T: RetTy, Name: "tmp");
300 if (CGF.EmitLifetimeStart(Addr: RetAddr.getBasePointer())) {
301 LifetimeStartInst =
302 cast<llvm::IntrinsicInst>(Val: std::prev(x: Builder.GetInsertPoint()));
303 assert(LifetimeStartInst->getIntrinsicID() ==
304 llvm::Intrinsic::lifetime_start &&
305 "Last insertion wasn't a lifetime.start?");
306
307 CGF.pushFullExprCleanup<CodeGenFunction::CallLifetimeEnd>(
308 kind: NormalEHLifetimeMarker, A: RetAddr);
309 LifetimeEndBlock = CGF.EHStack.stable_begin();
310 }
311 }
312
313 RValue Src =
314 EmitCall(ReturnValueSlot(RetAddr, Dest.isVolatile(), IsResultUnused,
315 Dest.isExternallyDestructed()));
316
317 if (!UseTemp)
318 return;
319
320 assert(Dest.isIgnored() || Dest.emitRawPointer(CGF) !=
321 Src.getAggregatePointer(E->getType(), CGF));
322 EmitFinalDestCopy(type: E->getType(), src: Src);
323
324 if (!RequiresDestruction && LifetimeStartInst) {
325 // If there's no dtor to run, the copy was the last use of our temporary.
326 // Since we're not guaranteed to be in an ExprWithCleanups, clean up
327 // eagerly.
328 CGF.DeactivateCleanupBlock(Cleanup: LifetimeEndBlock, DominatingIP: LifetimeStartInst);
329 CGF.EmitLifetimeEnd(Addr: RetAddr.getBasePointer());
330 }
331}
332
333/// EmitFinalDestCopy - Perform the final copy to DestPtr, if desired.
334void AggExprEmitter::EmitFinalDestCopy(QualType type, RValue src) {
335 assert(src.isAggregate() && "value must be aggregate value!");
336 LValue srcLV = CGF.MakeAddrLValue(Addr: src.getAggregateAddress(), T: type);
337 EmitFinalDestCopy(type, src: srcLV, SrcValueKind: CodeGenFunction::EVK_RValue);
338}
339
340/// EmitFinalDestCopy - Perform the final copy to DestPtr, if desired.
341void AggExprEmitter::EmitFinalDestCopy(
342 QualType type, const LValue &src,
343 CodeGenFunction::ExprValueKind SrcValueKind) {
344 // If Dest is ignored, then we're evaluating an aggregate expression
345 // in a context that doesn't care about the result. Note that loads
346 // from volatile l-values force the existence of a non-ignored
347 // destination.
348 if (Dest.isIgnored())
349 return;
350
351 // Copy non-trivial C structs here.
352 LValue DstLV = CGF.MakeAddrLValue(
353 Addr: Dest.getAddress(), T: Dest.isVolatile() ? type.withVolatile() : type);
354
355 if (SrcValueKind == CodeGenFunction::EVK_RValue) {
356 if (type.isNonTrivialToPrimitiveDestructiveMove() == QualType::PCK_Struct) {
357 if (Dest.isPotentiallyAliased())
358 CGF.callCStructMoveAssignmentOperator(Dst: DstLV, Src: src);
359 else
360 CGF.callCStructMoveConstructor(Dst: DstLV, Src: src);
361 return;
362 }
363 } else {
364 if (type.isNonTrivialToPrimitiveCopy() == QualType::PCK_Struct) {
365 if (Dest.isPotentiallyAliased())
366 CGF.callCStructCopyAssignmentOperator(Dst: DstLV, Src: src);
367 else
368 CGF.callCStructCopyConstructor(Dst: DstLV, Src: src);
369 return;
370 }
371 }
372
373 AggValueSlot srcAgg = AggValueSlot::forLValue(
374 LV: src, isDestructed: AggValueSlot::IsDestructed, needsGC: needsGC(T: type), isAliased: AggValueSlot::IsAliased,
375 mayOverlap: AggValueSlot::MayOverlap);
376 EmitCopy(type, dest: Dest, src: srcAgg);
377}
378
379/// Perform a copy from the source into the destination.
380///
381/// \param type - the type of the aggregate being copied; qualifiers are
382/// ignored
383void AggExprEmitter::EmitCopy(QualType type, const AggValueSlot &dest,
384 const AggValueSlot &src) {
385 if (dest.requiresGCollection()) {
386 CharUnits sz = dest.getPreferredSize(Ctx&: CGF.getContext(), Type: type);
387 llvm::Value *size = llvm::ConstantInt::get(Ty: CGF.SizeTy, V: sz.getQuantity());
388 CGF.CGM.getObjCRuntime().EmitGCMemmoveCollectable(CGF, DestPtr: dest.getAddress(),
389 SrcPtr: src.getAddress(), Size: size);
390 return;
391 }
392
393 // If the result of the assignment is used, copy the LHS there also.
394 // It's volatile if either side is. Use the minimum alignment of
395 // the two sides.
396 LValue DestLV = CGF.MakeAddrLValue(Addr: dest.getAddress(), T: type);
397 LValue SrcLV = CGF.MakeAddrLValue(Addr: src.getAddress(), T: type);
398 CGF.EmitAggregateCopy(Dest: DestLV, Src: SrcLV, EltTy: type, MayOverlap: dest.mayOverlap(),
399 isVolatile: dest.isVolatile() || src.isVolatile());
400}
401
402/// Emit the initializer for a std::initializer_list initialized with a
403/// real initializer list.
404void AggExprEmitter::VisitCXXStdInitializerListExpr(
405 CXXStdInitializerListExpr *E) {
406 // Emit an array containing the elements. The array is externally destructed
407 // if the std::initializer_list object is.
408 ASTContext &Ctx = CGF.getContext();
409 LValue Array = CGF.EmitLValue(E: E->getSubExpr());
410 assert(Array.isSimple() && "initializer_list array not a simple lvalue");
411 Address ArrayPtr = Array.getAddress();
412
413 const ConstantArrayType *ArrayType =
414 Ctx.getAsConstantArrayType(T: E->getSubExpr()->getType());
415 assert(ArrayType && "std::initializer_list constructed from non-array");
416
417 auto *Record = E->getType()->castAsRecordDecl();
418 RecordDecl::field_iterator Field = Record->field_begin();
419 assert(Field != Record->field_end() &&
420 Ctx.hasSameType(Field->getType()->getPointeeType(),
421 ArrayType->getElementType()) &&
422 "Expected std::initializer_list first field to be const E *");
423
424 // Start pointer.
425 AggValueSlot Dest = EnsureSlot(T: E->getType());
426 LValue DestLV = CGF.MakeAddrLValue(Addr: Dest.getAddress(), T: E->getType());
427 LValue Start = CGF.EmitLValueForFieldInitialization(Base: DestLV, Field: *Field);
428 llvm::Value *ArrayStart = ArrayPtr.emitRawPointer(CGF);
429 CGF.EmitStoreThroughLValue(Src: RValue::get(V: ArrayStart), Dst: Start);
430 ++Field;
431 assert(Field != Record->field_end() &&
432 "Expected std::initializer_list to have two fields");
433
434 llvm::Value *Size = Builder.getInt(AI: ArrayType->getSize());
435 LValue EndOrLength = CGF.EmitLValueForFieldInitialization(Base: DestLV, Field: *Field);
436 if (Ctx.hasSameType(T1: Field->getType(), T2: Ctx.getSizeType())) {
437 // Length.
438 CGF.EmitStoreThroughLValue(Src: RValue::get(V: Size), Dst: EndOrLength);
439
440 } else {
441 // End pointer.
442 assert(Field->getType()->isPointerType() &&
443 Ctx.hasSameType(Field->getType()->getPointeeType(),
444 ArrayType->getElementType()) &&
445 "Expected std::initializer_list second field to be const E *");
446 llvm::Value *Zero = llvm::ConstantInt::get(Ty: CGF.PtrDiffTy, V: 0);
447 llvm::Value *IdxEnd[] = {Zero, Size};
448 llvm::Value *ArrayEnd = Builder.CreateInBoundsGEP(
449 Ty: ArrayPtr.getElementType(), Ptr: ArrayPtr.emitRawPointer(CGF), IdxList: IdxEnd,
450 Name: "arrayend");
451 CGF.EmitStoreThroughLValue(Src: RValue::get(V: ArrayEnd), Dst: EndOrLength);
452 }
453
454 assert(++Field == Record->field_end() &&
455 "Expected std::initializer_list to only have two fields");
456}
457
458/// Determine if E is a trivial array filler, that is, one that is
459/// equivalent to zero-initialization.
460static bool isTrivialFiller(Expr *E) {
461 if (!E)
462 return true;
463
464 if (isa<ImplicitValueInitExpr>(Val: E))
465 return true;
466
467 if (auto *ILE = dyn_cast<InitListExpr>(Val: E)) {
468 if (ILE->getNumInits())
469 return false;
470 return isTrivialFiller(E: ILE->getArrayFiller());
471 }
472
473 if (auto *Cons = dyn_cast_or_null<CXXConstructExpr>(Val: E))
474 return Cons->getConstructor()->isDefaultConstructor() &&
475 Cons->getConstructor()->isTrivial();
476
477 // FIXME: Are there other cases where we can avoid emitting an initializer?
478 return false;
479}
480
481// emit an elementwise cast where the RHS is a scalar or vector
482// or emit an aggregate splat cast
483static void EmitHLSLScalarElementwiseAndSplatCasts(CodeGenFunction &CGF,
484 LValue DestVal,
485 llvm::Value *SrcVal,
486 QualType SrcTy,
487 SourceLocation Loc) {
488 // Flatten our destination
489 SmallVector<LValue, 16> StoreList;
490 CGF.FlattenAccessAndTypeLValue(LVal: DestVal, AccessList&: StoreList);
491
492 bool isVector = false;
493 if (auto *VT = SrcTy->getAs<VectorType>()) {
494 isVector = true;
495 SrcTy = VT->getElementType();
496 assert(StoreList.size() <= VT->getNumElements() &&
497 "Cannot perform HLSL flat cast when vector source \
498 object has less elements than flattened destination \
499 object.");
500 }
501
502 for (unsigned I = 0, Size = StoreList.size(); I < Size; I++) {
503 LValue DestLVal = StoreList[I];
504 llvm::Value *Load =
505 isVector ? CGF.Builder.CreateExtractElement(Vec: SrcVal, Idx: I, Name: "vec.load")
506 : SrcVal;
507 llvm::Value *Cast =
508 CGF.EmitScalarConversion(Src: Load, SrcTy, DstTy: DestLVal.getType(), Loc);
509 CGF.EmitStoreThroughLValue(Src: RValue::get(V: Cast), Dst: DestLVal);
510 }
511}
512
513// emit a flat cast where the RHS is an aggregate
514static void EmitHLSLElementwiseCast(CodeGenFunction &CGF, LValue DestVal,
515 LValue SrcVal, SourceLocation Loc) {
516 // Flatten our destination
517 SmallVector<LValue, 16> StoreList;
518 CGF.FlattenAccessAndTypeLValue(LVal: DestVal, AccessList&: StoreList);
519 // Flatten our src
520 SmallVector<LValue, 16> LoadList;
521 CGF.FlattenAccessAndTypeLValue(LVal: SrcVal, AccessList&: LoadList);
522
523 assert(StoreList.size() <= LoadList.size() &&
524 "Cannot perform HLSL elementwise cast when flattened source object \
525 has less elements than flattened destination object.");
526 // apply casts to what we load from LoadList
527 // and store result in Dest
528 for (unsigned I = 0, E = StoreList.size(); I < E; I++) {
529 LValue DestLVal = StoreList[I];
530 LValue SrcLVal = LoadList[I];
531 RValue RVal = CGF.EmitLoadOfLValue(V: SrcLVal, Loc);
532 assert(RVal.isScalar() && "All flattened source values should be scalars");
533 llvm::Value *Val = RVal.getScalarVal();
534 llvm::Value *Cast = CGF.EmitScalarConversion(Src: Val, SrcTy: SrcLVal.getType(),
535 DstTy: DestLVal.getType(), Loc);
536 CGF.EmitStoreThroughLValue(Src: RValue::get(V: Cast), Dst: DestLVal);
537 }
538}
539
540/// Emit initialization of an array from an initializer list. ExprToVisit must
541/// be either an InitListEpxr a CXXParenInitListExpr.
542void AggExprEmitter::EmitArrayInit(Address DestPtr, llvm::ArrayType *AType,
543 QualType ArrayQTy, Expr *ExprToVisit,
544 ArrayRef<Expr *> Args, Expr *ArrayFiller) {
545 uint64_t NumInitElements = Args.size();
546
547 uint64_t NumArrayElements = AType->getNumElements();
548 for (const auto *Init : Args) {
549 if (const auto *Embed = dyn_cast<EmbedExpr>(Val: Init->IgnoreParenImpCasts())) {
550 NumInitElements += Embed->getDataElementCount() - 1;
551 if (NumInitElements > NumArrayElements) {
552 NumInitElements = NumArrayElements;
553 break;
554 }
555 }
556 }
557
558 assert(NumInitElements <= NumArrayElements);
559
560 QualType elementType =
561 CGF.getContext().getAsArrayType(T: ArrayQTy)->getElementType();
562 CharUnits elementSize = CGF.getContext().getTypeSizeInChars(T: elementType);
563 CharUnits elementAlign =
564 DestPtr.getAlignment().alignmentOfArrayElement(elementSize);
565 llvm::Type *llvmElementType = CGF.ConvertTypeForMem(T: elementType);
566
567 // Consider initializing the array by copying from a global. For this to be
568 // more efficient than per-element initialization, the size of the elements
569 // with explicit initializers should be large enough.
570 if (NumInitElements * elementSize.getQuantity() > 16 &&
571 elementType.isTriviallyCopyableType(Context: CGF.getContext())) {
572 CodeGen::CodeGenModule &CGM = CGF.CGM;
573 ConstantEmitter Emitter(CGF);
574 QualType GVArrayQTy = CGM.getContext().getAddrSpaceQualType(
575 T: CGM.getContext().removeAddrSpaceQualType(T: ArrayQTy),
576 AddressSpace: CGM.GetGlobalConstantAddressSpace());
577 LangAS AS = GVArrayQTy.getAddressSpace();
578 if (llvm::Constant *C =
579 Emitter.tryEmitForInitializer(E: ExprToVisit, destAddrSpace: AS, destType: GVArrayQTy)) {
580 auto GV = new llvm::GlobalVariable(
581 CGM.getModule(), C->getType(),
582 /* isConstant= */ true, llvm::GlobalValue::PrivateLinkage, C,
583 "constinit",
584 /* InsertBefore= */ nullptr, llvm::GlobalVariable::NotThreadLocal,
585 CGM.getContext().getTargetAddressSpace(AS));
586 Emitter.finalize(global: GV);
587 CharUnits Align = CGM.getContext().getTypeAlignInChars(T: GVArrayQTy);
588 GV->setAlignment(Align.getAsAlign());
589 Address GVAddr(GV, GV->getValueType(), Align);
590 EmitFinalDestCopy(type: ArrayQTy, src: CGF.MakeAddrLValue(Addr: GVAddr, T: GVArrayQTy));
591 return;
592 }
593 }
594
595 // Exception safety requires us to destroy all the
596 // already-constructed members if an initializer throws.
597 // For that, we'll need an EH cleanup.
598 QualType::DestructionKind dtorKind = elementType.isDestructedType();
599 Address endOfInit = Address::invalid();
600 CodeGenFunction::CleanupDeactivationScope deactivation(CGF);
601
602 llvm::Value *begin = DestPtr.emitRawPointer(CGF);
603 if (dtorKind) {
604 CodeGenFunction::AllocaTrackerRAII allocaTracker(CGF);
605 // In principle we could tell the cleanup where we are more
606 // directly, but the control flow can get so varied here that it
607 // would actually be quite complex. Therefore we go through an
608 // alloca.
609 llvm::Instruction *dominatingIP =
610 Builder.CreateFlagLoad(Addr: llvm::ConstantInt::getNullValue(Ty: CGF.Int8PtrTy));
611 endOfInit = CGF.CreateTempAlloca(Ty: begin->getType(), align: CGF.getPointerAlign(),
612 Name: "arrayinit.endOfInit");
613 Builder.CreateStore(Val: begin, Addr: endOfInit);
614 CGF.pushIrregularPartialArrayCleanup(arrayBegin: begin, arrayEndPointer: endOfInit, elementType,
615 elementAlignment: elementAlign,
616 destroyer: CGF.getDestroyer(destructionKind: dtorKind));
617 cast<EHCleanupScope>(Val&: *CGF.EHStack.find(sp: CGF.EHStack.stable_begin()))
618 .AddAuxAllocas(Allocas: allocaTracker.Take());
619
620 CGF.DeferredDeactivationCleanupStack.push_back(
621 Elt: {.Cleanup: CGF.EHStack.stable_begin(), .DominatingIP: dominatingIP});
622 }
623
624 llvm::Value *one = llvm::ConstantInt::get(Ty: CGF.SizeTy, V: 1);
625
626 auto Emit = [&](Expr *Init, uint64_t ArrayIndex) {
627 llvm::Value *element = begin;
628 if (ArrayIndex > 0) {
629 element = Builder.CreateInBoundsGEP(
630 Ty: llvmElementType, Ptr: begin,
631 IdxList: llvm::ConstantInt::get(Ty: CGF.SizeTy, V: ArrayIndex), Name: "arrayinit.element");
632
633 // Tell the cleanup that it needs to destroy up to this
634 // element. TODO: some of these stores can be trivially
635 // observed to be unnecessary.
636 if (endOfInit.isValid())
637 Builder.CreateStore(Val: element, Addr: endOfInit);
638 }
639
640 LValue elementLV = CGF.MakeAddrLValue(
641 Addr: Address(element, llvmElementType, elementAlign), T: elementType);
642 EmitInitializationToLValue(E: Init, Address: elementLV);
643 return true;
644 };
645
646 unsigned ArrayIndex = 0;
647 // Emit the explicit initializers.
648 for (uint64_t i = 0; i != NumInitElements; ++i) {
649 if (ArrayIndex >= NumInitElements)
650 break;
651 if (auto *EmbedS = dyn_cast<EmbedExpr>(Val: Args[i]->IgnoreParenImpCasts())) {
652 EmbedS->doForEachDataElement(C&: Emit, StartingIndexInArray&: ArrayIndex);
653 } else {
654 Emit(Args[i], ArrayIndex);
655 ArrayIndex++;
656 }
657 }
658
659 // Check whether there's a non-trivial array-fill expression.
660 bool hasTrivialFiller = isTrivialFiller(E: ArrayFiller);
661
662 // Any remaining elements need to be zero-initialized, possibly
663 // using the filler expression. We can skip this if the we're
664 // emitting to zeroed memory.
665 if (NumInitElements != NumArrayElements &&
666 !(Dest.isZeroed() && hasTrivialFiller &&
667 CGF.getTypes().isZeroInitializable(T: elementType))) {
668
669 // Use an actual loop. This is basically
670 // do { *array++ = filler; } while (array != end);
671
672 // Advance to the start of the rest of the array.
673 llvm::Value *element = begin;
674 if (NumInitElements) {
675 element = Builder.CreateInBoundsGEP(
676 Ty: llvmElementType, Ptr: element,
677 IdxList: llvm::ConstantInt::get(Ty: CGF.SizeTy, V: NumInitElements),
678 Name: "arrayinit.start");
679 if (endOfInit.isValid())
680 Builder.CreateStore(Val: element, Addr: endOfInit);
681 }
682
683 // Compute the end of the array.
684 llvm::Value *end = Builder.CreateInBoundsGEP(
685 Ty: llvmElementType, Ptr: begin,
686 IdxList: llvm::ConstantInt::get(Ty: CGF.SizeTy, V: NumArrayElements), Name: "arrayinit.end");
687
688 llvm::BasicBlock *entryBB = Builder.GetInsertBlock();
689 llvm::BasicBlock *bodyBB = CGF.createBasicBlock(name: "arrayinit.body");
690
691 // Jump into the body.
692 CGF.EmitBlock(BB: bodyBB);
693 llvm::PHINode *currentElement =
694 Builder.CreatePHI(Ty: element->getType(), NumReservedValues: 2, Name: "arrayinit.cur");
695 currentElement->addIncoming(V: element, BB: entryBB);
696
697 // Emit the actual filler expression.
698 {
699 // C++1z [class.temporary]p5:
700 // when a default constructor is called to initialize an element of
701 // an array with no corresponding initializer [...] the destruction of
702 // every temporary created in a default argument is sequenced before
703 // the construction of the next array element, if any
704 CodeGenFunction::RunCleanupsScope CleanupsScope(CGF);
705 LValue elementLV = CGF.MakeAddrLValue(
706 Addr: Address(currentElement, llvmElementType, elementAlign), T: elementType);
707 if (ArrayFiller)
708 EmitInitializationToLValue(E: ArrayFiller, Address: elementLV);
709 else
710 EmitNullInitializationToLValue(Address: elementLV);
711 }
712
713 // Move on to the next element.
714 llvm::Value *nextElement = Builder.CreateInBoundsGEP(
715 Ty: llvmElementType, Ptr: currentElement, IdxList: one, Name: "arrayinit.next");
716
717 // Tell the EH cleanup that we finished with the last element.
718 if (endOfInit.isValid())
719 Builder.CreateStore(Val: nextElement, Addr: endOfInit);
720
721 // Leave the loop if we're done.
722 llvm::Value *done =
723 Builder.CreateICmpEQ(LHS: nextElement, RHS: end, Name: "arrayinit.done");
724 llvm::BasicBlock *endBB = CGF.createBasicBlock(name: "arrayinit.end");
725 Builder.CreateCondBr(Cond: done, True: endBB, False: bodyBB);
726 currentElement->addIncoming(V: nextElement, BB: Builder.GetInsertBlock());
727
728 CGF.EmitBlock(BB: endBB);
729 }
730}
731
732//===----------------------------------------------------------------------===//
733// Visitor Methods
734//===----------------------------------------------------------------------===//
735
736void AggExprEmitter::VisitMaterializeTemporaryExpr(
737 MaterializeTemporaryExpr *E) {
738 Visit(E: E->getSubExpr());
739}
740
741void AggExprEmitter::VisitOpaqueValueExpr(OpaqueValueExpr *e) {
742 // If this is a unique OVE, just visit its source expression.
743 if (e->isUnique())
744 Visit(E: e->getSourceExpr());
745 else
746 EmitFinalDestCopy(type: e->getType(), src: CGF.getOrCreateOpaqueLValueMapping(e));
747}
748
749void AggExprEmitter::VisitCompoundLiteralExpr(CompoundLiteralExpr *E) {
750 if (Dest.isPotentiallyAliased()) {
751 // Just emit a load of the lvalue + a copy, because our compound literal
752 // might alias the destination.
753 EmitAggLoadOfLValue(E);
754 return;
755 }
756
757 AggValueSlot Slot = EnsureSlot(T: E->getType());
758
759 // Block-scope compound literals are destroyed at the end of the enclosing
760 // scope in C.
761 bool Destruct =
762 !CGF.getLangOpts().CPlusPlus && !Slot.isExternallyDestructed();
763 if (Destruct)
764 Slot.setExternallyDestructed();
765
766 CGF.EmitAggExpr(E: E->getInitializer(), AS: Slot);
767
768 if (Destruct)
769 if (QualType::DestructionKind DtorKind = E->getType().isDestructedType())
770 CGF.pushLifetimeExtendedDestroy(
771 kind: CGF.getCleanupKind(kind: DtorKind), addr: Slot.getAddress(), type: E->getType(),
772 destroyer: CGF.getDestroyer(destructionKind: DtorKind), useEHCleanupForArray: DtorKind & EHCleanup);
773}
774
775/// Attempt to look through various unimportant expressions to find a
776/// cast of the given kind.
777static Expr *findPeephole(Expr *op, CastKind kind, const ASTContext &ctx) {
778 op = op->IgnoreParenNoopCasts(Ctx: ctx);
779 if (auto castE = dyn_cast<CastExpr>(Val: op)) {
780 if (castE->getCastKind() == kind)
781 return castE->getSubExpr();
782 }
783 return nullptr;
784}
785
786void AggExprEmitter::VisitCastExpr(CastExpr *E) {
787 if (const auto *ECE = dyn_cast<ExplicitCastExpr>(Val: E))
788 CGF.CGM.EmitExplicitCastExprType(E: ECE, CGF: &CGF);
789 switch (E->getCastKind()) {
790 case CK_Dynamic: {
791 // FIXME: Can this actually happen? We have no test coverage for it.
792 assert(isa<CXXDynamicCastExpr>(E) && "CK_Dynamic without a dynamic_cast?");
793 LValue LV =
794 CGF.EmitCheckedLValue(E: E->getSubExpr(), TCK: CodeGenFunction::TCK_Load);
795 // FIXME: Do we also need to handle property references here?
796 if (LV.isSimple())
797 CGF.EmitDynamicCast(V: LV.getAddress(), DCE: cast<CXXDynamicCastExpr>(Val: E));
798 else
799 CGF.CGM.ErrorUnsupported(S: E, Type: "non-simple lvalue dynamic_cast");
800
801 if (!Dest.isIgnored())
802 CGF.CGM.ErrorUnsupported(S: E, Type: "lvalue dynamic_cast with a destination");
803 break;
804 }
805
806 case CK_ToUnion: {
807 // Evaluate even if the destination is ignored.
808 if (Dest.isIgnored()) {
809 CGF.EmitAnyExpr(E: E->getSubExpr(), aggSlot: AggValueSlot::ignored(),
810 /*ignoreResult=*/true);
811 break;
812 }
813
814 // GCC union extension
815 QualType Ty = E->getSubExpr()->getType();
816 Address CastPtr = Dest.getAddress().withElementType(ElemTy: CGF.ConvertType(T: Ty));
817 EmitInitializationToLValue(E: E->getSubExpr(),
818 Address: CGF.MakeAddrLValue(Addr: CastPtr, T: Ty));
819 break;
820 }
821
822 case CK_LValueToRValueBitCast: {
823 if (Dest.isIgnored()) {
824 CGF.EmitAnyExpr(E: E->getSubExpr(), aggSlot: AggValueSlot::ignored(),
825 /*ignoreResult=*/true);
826 break;
827 }
828
829 LValue SourceLV = CGF.EmitLValue(E: E->getSubExpr());
830 Address SourceAddress = SourceLV.getAddress().withElementType(ElemTy: CGF.Int8Ty);
831 Address DestAddress = Dest.getAddress().withElementType(ElemTy: CGF.Int8Ty);
832 llvm::Value *SizeVal = llvm::ConstantInt::get(
833 Ty: CGF.SizeTy,
834 V: CGF.getContext().getTypeSizeInChars(T: E->getType()).getQuantity());
835 Builder.CreateMemCpy(Dest: DestAddress, Src: SourceAddress, Size: SizeVal);
836 break;
837 }
838
839 case CK_DerivedToBase:
840 case CK_BaseToDerived:
841 case CK_UncheckedDerivedToBase: {
842 llvm_unreachable("cannot perform hierarchy conversion in EmitAggExpr: "
843 "should have been unpacked before we got here");
844 }
845
846 case CK_NonAtomicToAtomic:
847 case CK_AtomicToNonAtomic: {
848 bool isToAtomic = (E->getCastKind() == CK_NonAtomicToAtomic);
849
850 // Determine the atomic and value types.
851 QualType atomicType = E->getSubExpr()->getType();
852 QualType valueType = E->getType();
853 if (isToAtomic)
854 std::swap(a&: atomicType, b&: valueType);
855
856 assert(atomicType->isAtomicType());
857 assert(CGF.getContext().hasSameUnqualifiedType(
858 valueType, atomicType->castAs<AtomicType>()->getValueType()));
859
860 // Just recurse normally if we're ignoring the result or the
861 // atomic type doesn't change representation.
862 if (Dest.isIgnored() || !CGF.CGM.isPaddedAtomicType(type: atomicType)) {
863 return Visit(E: E->getSubExpr());
864 }
865
866 CastKind peepholeTarget =
867 (isToAtomic ? CK_AtomicToNonAtomic : CK_NonAtomicToAtomic);
868
869 // These two cases are reverses of each other; try to peephole them.
870 if (Expr *op =
871 findPeephole(op: E->getSubExpr(), kind: peepholeTarget, ctx: CGF.getContext())) {
872 assert(CGF.getContext().hasSameUnqualifiedType(op->getType(),
873 E->getType()) &&
874 "peephole significantly changed types?");
875 return Visit(E: op);
876 }
877
878 // If we're converting an r-value of non-atomic type to an r-value
879 // of atomic type, just emit directly into the relevant sub-object.
880 if (isToAtomic) {
881 AggValueSlot valueDest = Dest;
882 if (!valueDest.isIgnored() && CGF.CGM.isPaddedAtomicType(type: atomicType)) {
883 // Zero-initialize. (Strictly speaking, we only need to initialize
884 // the padding at the end, but this is simpler.)
885 if (!Dest.isZeroed())
886 CGF.EmitNullInitialization(DestPtr: Dest.getAddress(), Ty: atomicType);
887
888 // Build a GEP to refer to the subobject.
889 Address valueAddr =
890 CGF.Builder.CreateStructGEP(Addr: valueDest.getAddress(), Index: 0);
891 valueDest = AggValueSlot::forAddr(
892 addr: valueAddr, quals: valueDest.getQualifiers(),
893 isDestructed: valueDest.isExternallyDestructed(), needsGC: valueDest.requiresGCollection(),
894 isAliased: valueDest.isPotentiallyAliased(), mayOverlap: AggValueSlot::DoesNotOverlap,
895 isZeroed: AggValueSlot::IsZeroed);
896 }
897
898 CGF.EmitAggExpr(E: E->getSubExpr(), AS: valueDest);
899 return;
900 }
901
902 // Otherwise, we're converting an atomic type to a non-atomic type.
903 // Make an atomic temporary, emit into that, and then copy the value out.
904 AggValueSlot atomicSlot =
905 CGF.CreateAggTemp(T: atomicType, Name: "atomic-to-nonatomic.temp");
906 CGF.EmitAggExpr(E: E->getSubExpr(), AS: atomicSlot);
907
908 Address valueAddr = Builder.CreateStructGEP(Addr: atomicSlot.getAddress(), Index: 0);
909 RValue rvalue = RValue::getAggregate(addr: valueAddr, isVolatile: atomicSlot.isVolatile());
910 return EmitFinalDestCopy(type: valueType, src: rvalue);
911 }
912 case CK_AddressSpaceConversion:
913 return Visit(E: E->getSubExpr());
914
915 case CK_LValueToRValue:
916 // If we're loading from a volatile type, force the destination
917 // into existence.
918 if (E->getSubExpr()->getType().isVolatileQualified()) {
919 bool Destruct =
920 !Dest.isExternallyDestructed() &&
921 E->getType().isDestructedType() == QualType::DK_nontrivial_c_struct;
922 if (Destruct)
923 Dest.setExternallyDestructed();
924 EnsureDest(T: E->getType());
925 Visit(E: E->getSubExpr());
926
927 if (Destruct)
928 CGF.pushDestroy(dtorKind: QualType::DK_nontrivial_c_struct, addr: Dest.getAddress(),
929 type: E->getType());
930
931 return;
932 }
933
934 [[fallthrough]];
935
936 case CK_HLSLArrayRValue:
937 Visit(E: E->getSubExpr());
938 break;
939 case CK_HLSLAggregateSplatCast: {
940 Expr *Src = E->getSubExpr();
941 QualType SrcTy = Src->getType();
942 RValue RV = CGF.EmitAnyExpr(E: Src);
943 LValue DestLVal = CGF.MakeAddrLValue(Addr: Dest.getAddress(), T: E->getType());
944 SourceLocation Loc = E->getExprLoc();
945
946 assert(RV.isScalar() && SrcTy->isScalarType() &&
947 "RHS of HLSL splat cast must be a scalar.");
948 llvm::Value *SrcVal = RV.getScalarVal();
949 EmitHLSLScalarElementwiseAndSplatCasts(CGF, DestVal: DestLVal, SrcVal, SrcTy, Loc);
950 break;
951 }
952 case CK_HLSLElementwiseCast: {
953 Expr *Src = E->getSubExpr();
954 QualType SrcTy = Src->getType();
955 RValue RV = CGF.EmitAnyExpr(E: Src);
956 LValue DestLVal = CGF.MakeAddrLValue(Addr: Dest.getAddress(), T: E->getType());
957 SourceLocation Loc = E->getExprLoc();
958
959 if (RV.isScalar()) {
960 llvm::Value *SrcVal = RV.getScalarVal();
961 assert(SrcTy->isVectorType() &&
962 "HLSL Elementwise cast doesn't handle splatting.");
963 EmitHLSLScalarElementwiseAndSplatCasts(CGF, DestVal: DestLVal, SrcVal, SrcTy, Loc);
964 } else {
965 assert(RV.isAggregate() &&
966 "Can't perform HLSL Aggregate cast on a complex type.");
967 Address SrcVal = RV.getAggregateAddress();
968 EmitHLSLElementwiseCast(CGF, DestVal: DestLVal, SrcVal: CGF.MakeAddrLValue(Addr: SrcVal, T: SrcTy),
969 Loc);
970 }
971 break;
972 }
973 case CK_NoOp:
974 case CK_UserDefinedConversion:
975 case CK_ConstructorConversion:
976 assert(CGF.getContext().hasSameUnqualifiedType(E->getSubExpr()->getType(),
977 E->getType()) &&
978 "Implicit cast types must be compatible");
979 Visit(E: E->getSubExpr());
980 break;
981
982 case CK_LValueBitCast:
983 llvm_unreachable("should not be emitting lvalue bitcast as rvalue");
984
985 case CK_Dependent:
986 case CK_BitCast:
987 case CK_ArrayToPointerDecay:
988 case CK_FunctionToPointerDecay:
989 case CK_NullToPointer:
990 case CK_NullToMemberPointer:
991 case CK_BaseToDerivedMemberPointer:
992 case CK_DerivedToBaseMemberPointer:
993 case CK_MemberPointerToBoolean:
994 case CK_ReinterpretMemberPointer:
995 case CK_IntegralToPointer:
996 case CK_PointerToIntegral:
997 case CK_PointerToBoolean:
998 case CK_ToVoid:
999 case CK_VectorSplat:
1000 case CK_IntegralCast:
1001 case CK_BooleanToSignedIntegral:
1002 case CK_IntegralToBoolean:
1003 case CK_IntegralToFloating:
1004 case CK_FloatingToIntegral:
1005 case CK_FloatingToBoolean:
1006 case CK_FloatingCast:
1007 case CK_CPointerToObjCPointerCast:
1008 case CK_BlockPointerToObjCPointerCast:
1009 case CK_AnyPointerToBlockPointerCast:
1010 case CK_ObjCObjectLValueCast:
1011 case CK_FloatingRealToComplex:
1012 case CK_FloatingComplexToReal:
1013 case CK_FloatingComplexToBoolean:
1014 case CK_FloatingComplexCast:
1015 case CK_FloatingComplexToIntegralComplex:
1016 case CK_IntegralRealToComplex:
1017 case CK_IntegralComplexToReal:
1018 case CK_IntegralComplexToBoolean:
1019 case CK_IntegralComplexCast:
1020 case CK_IntegralComplexToFloatingComplex:
1021 case CK_ARCProduceObject:
1022 case CK_ARCConsumeObject:
1023 case CK_ARCReclaimReturnedObject:
1024 case CK_ARCExtendBlockObject:
1025 case CK_CopyAndAutoreleaseBlockObject:
1026 case CK_BuiltinFnToFnPtr:
1027 case CK_ZeroToOCLOpaqueType:
1028 case CK_MatrixCast:
1029 case CK_HLSLVectorTruncation:
1030 case CK_HLSLMatrixTruncation:
1031 case CK_IntToOCLSampler:
1032 case CK_FloatingToFixedPoint:
1033 case CK_FixedPointToFloating:
1034 case CK_FixedPointCast:
1035 case CK_FixedPointToBoolean:
1036 case CK_FixedPointToIntegral:
1037 case CK_IntegralToFixedPoint:
1038 llvm_unreachable("cast kind invalid for aggregate types");
1039 }
1040}
1041
1042void AggExprEmitter::VisitCallExpr(const CallExpr *E) {
1043 if (E->getCallReturnType(Ctx: CGF.getContext())->isReferenceType()) {
1044 EmitAggLoadOfLValue(E);
1045 return;
1046 }
1047
1048 withReturnValueSlot(
1049 E, EmitCall: [&](ReturnValueSlot Slot) { return CGF.EmitCallExpr(E, ReturnValue: Slot); });
1050}
1051
1052void AggExprEmitter::VisitObjCMessageExpr(ObjCMessageExpr *E) {
1053 withReturnValueSlot(E, EmitCall: [&](ReturnValueSlot Slot) {
1054 return CGF.EmitObjCMessageExpr(E, Return: Slot);
1055 });
1056}
1057
1058void AggExprEmitter::VisitBinComma(const BinaryOperator *E) {
1059 CGF.EmitIgnoredExpr(E: E->getLHS());
1060 Visit(E: E->getRHS());
1061}
1062
1063void AggExprEmitter::VisitStmtExpr(const StmtExpr *E) {
1064 CodeGenFunction::StmtExprEvaluation eval(CGF);
1065 CGF.EmitCompoundStmt(S: *E->getSubStmt(), GetLast: true, AVS: Dest);
1066}
1067
1068enum CompareKind {
1069 CK_Less,
1070 CK_Greater,
1071 CK_Equal,
1072};
1073
1074static llvm::Value *EmitCompare(CGBuilderTy &Builder, CodeGenFunction &CGF,
1075 const BinaryOperator *E, llvm::Value *LHS,
1076 llvm::Value *RHS, CompareKind Kind,
1077 const char *NameSuffix = "") {
1078 QualType ArgTy = E->getLHS()->getType();
1079 if (const ComplexType *CT = ArgTy->getAs<ComplexType>())
1080 ArgTy = CT->getElementType();
1081
1082 if (const auto *MPT = ArgTy->getAs<MemberPointerType>()) {
1083 assert(Kind == CK_Equal &&
1084 "member pointers may only be compared for equality");
1085 return CGF.CGM.getCXXABI().EmitMemberPointerComparison(
1086 CGF, L: LHS, R: RHS, MPT, /*IsInequality*/ Inequality: false);
1087 }
1088
1089 // Compute the comparison instructions for the specified comparison kind.
1090 struct CmpInstInfo {
1091 const char *Name;
1092 llvm::CmpInst::Predicate FCmp;
1093 llvm::CmpInst::Predicate SCmp;
1094 llvm::CmpInst::Predicate UCmp;
1095 };
1096 CmpInstInfo InstInfo = [&]() -> CmpInstInfo {
1097 using FI = llvm::FCmpInst;
1098 using II = llvm::ICmpInst;
1099 switch (Kind) {
1100 case CK_Less:
1101 return {.Name: "cmp.lt", .FCmp: FI::FCMP_OLT, .SCmp: II::ICMP_SLT, .UCmp: II::ICMP_ULT};
1102 case CK_Greater:
1103 return {.Name: "cmp.gt", .FCmp: FI::FCMP_OGT, .SCmp: II::ICMP_SGT, .UCmp: II::ICMP_UGT};
1104 case CK_Equal:
1105 return {.Name: "cmp.eq", .FCmp: FI::FCMP_OEQ, .SCmp: II::ICMP_EQ, .UCmp: II::ICMP_EQ};
1106 }
1107 llvm_unreachable("Unrecognised CompareKind enum");
1108 }();
1109
1110 if (ArgTy->hasFloatingRepresentation())
1111 return Builder.CreateFCmp(P: InstInfo.FCmp, LHS, RHS,
1112 Name: llvm::Twine(InstInfo.Name) + NameSuffix);
1113 if (ArgTy->isIntegralOrEnumerationType() || ArgTy->isPointerType()) {
1114 auto Inst =
1115 ArgTy->hasSignedIntegerRepresentation() ? InstInfo.SCmp : InstInfo.UCmp;
1116 return Builder.CreateICmp(P: Inst, LHS, RHS,
1117 Name: llvm::Twine(InstInfo.Name) + NameSuffix);
1118 }
1119
1120 llvm_unreachable("unsupported aggregate binary expression should have "
1121 "already been handled");
1122}
1123
1124void AggExprEmitter::VisitBinCmp(const BinaryOperator *E) {
1125 using llvm::BasicBlock;
1126 using llvm::PHINode;
1127 using llvm::Value;
1128 assert(CGF.getContext().hasSameType(E->getLHS()->getType(),
1129 E->getRHS()->getType()));
1130 const ComparisonCategoryInfo &CmpInfo =
1131 CGF.getContext().CompCategories.getInfoForType(Ty: E->getType());
1132 assert(CmpInfo.Record->isTriviallyCopyable() &&
1133 "cannot copy non-trivially copyable aggregate");
1134
1135 QualType ArgTy = E->getLHS()->getType();
1136
1137 if (!ArgTy->isIntegralOrEnumerationType() && !ArgTy->isRealFloatingType() &&
1138 !ArgTy->isNullPtrType() && !ArgTy->isPointerType() &&
1139 !ArgTy->isMemberPointerType() && !ArgTy->isAnyComplexType()) {
1140 return CGF.ErrorUnsupported(S: E, Type: "aggregate three-way comparison");
1141 }
1142 bool IsComplex = ArgTy->isAnyComplexType();
1143
1144 // Evaluate the operands to the expression and extract their values.
1145 auto EmitOperand = [&](Expr *E) -> std::pair<Value *, Value *> {
1146 RValue RV = CGF.EmitAnyExpr(E);
1147 if (RV.isScalar())
1148 return {RV.getScalarVal(), nullptr};
1149 if (RV.isAggregate())
1150 return {RV.getAggregatePointer(PointeeType: E->getType(), CGF), nullptr};
1151 assert(RV.isComplex());
1152 return RV.getComplexVal();
1153 };
1154 auto LHSValues = EmitOperand(E->getLHS()),
1155 RHSValues = EmitOperand(E->getRHS());
1156
1157 auto EmitCmp = [&](CompareKind K) {
1158 Value *Cmp = EmitCompare(Builder, CGF, E, LHS: LHSValues.first, RHS: RHSValues.first,
1159 Kind: K, NameSuffix: IsComplex ? ".r" : "");
1160 if (!IsComplex)
1161 return Cmp;
1162 assert(K == CompareKind::CK_Equal);
1163 Value *CmpImag = EmitCompare(Builder, CGF, E, LHS: LHSValues.second,
1164 RHS: RHSValues.second, Kind: K, NameSuffix: ".i");
1165 return Builder.CreateAnd(LHS: Cmp, RHS: CmpImag, Name: "and.eq");
1166 };
1167 auto EmitCmpRes = [&](const ComparisonCategoryInfo::ValueInfo *VInfo) {
1168 return Builder.getInt(AI: VInfo->getIntValue());
1169 };
1170
1171 Value *Select;
1172 if (ArgTy->isNullPtrType()) {
1173 Select = EmitCmpRes(CmpInfo.getEqualOrEquiv());
1174 } else if (!CmpInfo.isPartial()) {
1175 Value *SelectOne =
1176 Builder.CreateSelect(C: EmitCmp(CK_Less), True: EmitCmpRes(CmpInfo.getLess()),
1177 False: EmitCmpRes(CmpInfo.getGreater()), Name: "sel.lt");
1178 Select = Builder.CreateSelect(C: EmitCmp(CK_Equal),
1179 True: EmitCmpRes(CmpInfo.getEqualOrEquiv()),
1180 False: SelectOne, Name: "sel.eq");
1181 } else {
1182 Value *SelectEq = Builder.CreateSelect(
1183 C: EmitCmp(CK_Equal), True: EmitCmpRes(CmpInfo.getEqualOrEquiv()),
1184 False: EmitCmpRes(CmpInfo.getUnordered()), Name: "sel.eq");
1185 Value *SelectGT = Builder.CreateSelect(C: EmitCmp(CK_Greater),
1186 True: EmitCmpRes(CmpInfo.getGreater()),
1187 False: SelectEq, Name: "sel.gt");
1188 Select = Builder.CreateSelect(
1189 C: EmitCmp(CK_Less), True: EmitCmpRes(CmpInfo.getLess()), False: SelectGT, Name: "sel.lt");
1190 }
1191 // Create the return value in the destination slot.
1192 EnsureDest(T: E->getType());
1193 LValue DestLV = CGF.MakeAddrLValue(Addr: Dest.getAddress(), T: E->getType());
1194
1195 // Emit the address of the first (and only) field in the comparison category
1196 // type, and initialize it from the constant integer value selected above.
1197 LValue FieldLV = CGF.EmitLValueForFieldInitialization(
1198 Base: DestLV, Field: *CmpInfo.Record->field_begin());
1199 CGF.EmitStoreThroughLValue(Src: RValue::get(V: Select), Dst: FieldLV, /*IsInit*/ isInit: true);
1200
1201 // All done! The result is in the Dest slot.
1202}
1203
1204void AggExprEmitter::VisitBinaryOperator(const BinaryOperator *E) {
1205 if (E->getOpcode() == BO_PtrMemD || E->getOpcode() == BO_PtrMemI)
1206 VisitPointerToDataMemberBinaryOperator(BO: E);
1207 else
1208 CGF.ErrorUnsupported(S: E, Type: "aggregate binary expression");
1209}
1210
1211void AggExprEmitter::VisitPointerToDataMemberBinaryOperator(
1212 const BinaryOperator *E) {
1213 LValue LV = CGF.EmitPointerToDataMemberBinaryExpr(E);
1214 EmitFinalDestCopy(type: E->getType(), src: LV);
1215}
1216
1217/// Is the value of the given expression possibly a reference to or
1218/// into a __block variable?
1219static bool isBlockVarRef(const Expr *E) {
1220 // Make sure we look through parens.
1221 E = E->IgnoreParens();
1222
1223 // Check for a direct reference to a __block variable.
1224 if (const DeclRefExpr *DRE = dyn_cast<DeclRefExpr>(Val: E)) {
1225 const VarDecl *var = dyn_cast<VarDecl>(Val: DRE->getDecl());
1226 return (var && var->hasAttr<BlocksAttr>());
1227 }
1228
1229 // More complicated stuff.
1230
1231 // Binary operators.
1232 if (const BinaryOperator *op = dyn_cast<BinaryOperator>(Val: E)) {
1233 // For an assignment or pointer-to-member operation, just care
1234 // about the LHS.
1235 if (op->isAssignmentOp() || op->isPtrMemOp())
1236 return isBlockVarRef(E: op->getLHS());
1237
1238 // For a comma, just care about the RHS.
1239 if (op->getOpcode() == BO_Comma)
1240 return isBlockVarRef(E: op->getRHS());
1241
1242 // FIXME: pointer arithmetic?
1243 return false;
1244
1245 // Check both sides of a conditional operator.
1246 } else if (const AbstractConditionalOperator *op =
1247 dyn_cast<AbstractConditionalOperator>(Val: E)) {
1248 return isBlockVarRef(E: op->getTrueExpr()) ||
1249 isBlockVarRef(E: op->getFalseExpr());
1250
1251 // OVEs are required to support BinaryConditionalOperators.
1252 } else if (const OpaqueValueExpr *op = dyn_cast<OpaqueValueExpr>(Val: E)) {
1253 if (const Expr *src = op->getSourceExpr())
1254 return isBlockVarRef(E: src);
1255
1256 // Casts are necessary to get things like (*(int*)&var) = foo().
1257 // We don't really care about the kind of cast here, except
1258 // we don't want to look through l2r casts, because it's okay
1259 // to get the *value* in a __block variable.
1260 } else if (const CastExpr *cast = dyn_cast<CastExpr>(Val: E)) {
1261 if (cast->getCastKind() == CK_LValueToRValue)
1262 return false;
1263 return isBlockVarRef(E: cast->getSubExpr());
1264
1265 // Handle unary operators. Again, just aggressively look through
1266 // it, ignoring the operation.
1267 } else if (const UnaryOperator *uop = dyn_cast<UnaryOperator>(Val: E)) {
1268 return isBlockVarRef(E: uop->getSubExpr());
1269
1270 // Look into the base of a field access.
1271 } else if (const MemberExpr *mem = dyn_cast<MemberExpr>(Val: E)) {
1272 return isBlockVarRef(E: mem->getBase());
1273
1274 // Look into the base of a subscript.
1275 } else if (const ArraySubscriptExpr *sub = dyn_cast<ArraySubscriptExpr>(Val: E)) {
1276 return isBlockVarRef(E: sub->getBase());
1277 }
1278
1279 return false;
1280}
1281
1282void AggExprEmitter::VisitBinAssign(const BinaryOperator *E) {
1283 ApplyAtomGroup Grp(CGF.getDebugInfo());
1284 // For an assignment to work, the value on the right has
1285 // to be compatible with the value on the left.
1286 assert(CGF.getContext().hasSameUnqualifiedType(E->getLHS()->getType(),
1287 E->getRHS()->getType()) &&
1288 "Invalid assignment");
1289
1290 // If the LHS might be a __block variable, and the RHS can
1291 // potentially cause a block copy, we need to evaluate the RHS first
1292 // so that the assignment goes the right place.
1293 // This is pretty semantically fragile.
1294 if (isBlockVarRef(E: E->getLHS()) &&
1295 E->getRHS()->HasSideEffects(Ctx: CGF.getContext())) {
1296 // Ensure that we have a destination, and evaluate the RHS into that.
1297 EnsureDest(T: E->getRHS()->getType());
1298 Visit(E: E->getRHS());
1299
1300 // Now emit the LHS and copy into it.
1301 LValue LHS = CGF.EmitCheckedLValue(E: E->getLHS(), TCK: CodeGenFunction::TCK_Store);
1302
1303 // That copy is an atomic copy if the LHS is atomic.
1304 if (LHS.getType()->isAtomicType() ||
1305 CGF.LValueIsSuitableForInlineAtomic(Src: LHS)) {
1306 CGF.EmitAtomicStore(rvalue: Dest.asRValue(), lvalue: LHS, /*isInit*/ false);
1307 return;
1308 }
1309
1310 EmitCopy(type: E->getLHS()->getType(),
1311 dest: AggValueSlot::forLValue(LV: LHS, isDestructed: AggValueSlot::IsDestructed,
1312 needsGC: needsGC(T: E->getLHS()->getType()),
1313 isAliased: AggValueSlot::IsAliased,
1314 mayOverlap: AggValueSlot::MayOverlap),
1315 src: Dest);
1316 return;
1317 }
1318
1319 LValue LHS = CGF.EmitLValue(E: E->getLHS());
1320
1321 // If we have an atomic type, evaluate into the destination and then
1322 // do an atomic copy.
1323 if (LHS.getType()->isAtomicType() ||
1324 CGF.LValueIsSuitableForInlineAtomic(Src: LHS)) {
1325 EnsureDest(T: E->getRHS()->getType());
1326 Visit(E: E->getRHS());
1327 CGF.EmitAtomicStore(rvalue: Dest.asRValue(), lvalue: LHS, /*isInit*/ false);
1328 return;
1329 }
1330
1331 // Codegen the RHS so that it stores directly into the LHS.
1332 AggValueSlot LHSSlot = AggValueSlot::forLValue(
1333 LV: LHS, isDestructed: AggValueSlot::IsDestructed, needsGC: needsGC(T: E->getLHS()->getType()),
1334 isAliased: AggValueSlot::IsAliased, mayOverlap: AggValueSlot::MayOverlap);
1335 // A non-volatile aggregate destination might have volatile member.
1336 if (!LHSSlot.isVolatile() && CGF.hasVolatileMember(T: E->getLHS()->getType()))
1337 LHSSlot.setVolatile(true);
1338
1339 CGF.EmitAggExpr(E: E->getRHS(), AS: LHSSlot);
1340
1341 // Copy into the destination if the assignment isn't ignored.
1342 EmitFinalDestCopy(type: E->getType(), src: LHS);
1343
1344 if (!Dest.isIgnored() && !Dest.isExternallyDestructed() &&
1345 E->getType().isDestructedType() == QualType::DK_nontrivial_c_struct)
1346 CGF.pushDestroy(dtorKind: QualType::DK_nontrivial_c_struct, addr: Dest.getAddress(),
1347 type: E->getType());
1348}
1349
1350void AggExprEmitter::VisitAbstractConditionalOperator(
1351 const AbstractConditionalOperator *E) {
1352 llvm::BasicBlock *LHSBlock = CGF.createBasicBlock(name: "cond.true");
1353 llvm::BasicBlock *RHSBlock = CGF.createBasicBlock(name: "cond.false");
1354 llvm::BasicBlock *ContBlock = CGF.createBasicBlock(name: "cond.end");
1355
1356 // Bind the common expression if necessary.
1357 CodeGenFunction::OpaqueValueMapping binding(CGF, E);
1358
1359 CodeGenFunction::ConditionalEvaluation eval(CGF);
1360 CGF.EmitBranchOnBoolExpr(Cond: E->getCond(), TrueBlock: LHSBlock, FalseBlock: RHSBlock,
1361 TrueCount: CGF.getProfileCount(S: E));
1362
1363 // Save whether the destination's lifetime is externally managed.
1364 bool isExternallyDestructed = Dest.isExternallyDestructed();
1365 bool destructNonTrivialCStruct =
1366 !isExternallyDestructed &&
1367 E->getType().isDestructedType() == QualType::DK_nontrivial_c_struct;
1368 isExternallyDestructed |= destructNonTrivialCStruct;
1369 Dest.setExternallyDestructed(isExternallyDestructed);
1370
1371 eval.begin(CGF);
1372 CGF.EmitBlock(BB: LHSBlock);
1373 CGF.incrementProfileCounter(ExecSkip: CGF.UseExecPath, S: E);
1374 Visit(E: E->getTrueExpr());
1375 eval.end(CGF);
1376
1377 assert(CGF.HaveInsertPoint() && "expression evaluation ended with no IP!");
1378 CGF.Builder.CreateBr(Dest: ContBlock);
1379
1380 // If the result of an agg expression is unused, then the emission
1381 // of the LHS might need to create a destination slot. That's fine
1382 // with us, and we can safely emit the RHS into the same slot, but
1383 // we shouldn't claim that it's already being destructed.
1384 Dest.setExternallyDestructed(isExternallyDestructed);
1385
1386 eval.begin(CGF);
1387 CGF.EmitBlock(BB: RHSBlock);
1388 CGF.incrementProfileCounter(ExecSkip: CGF.UseSkipPath, S: E);
1389 Visit(E: E->getFalseExpr());
1390 eval.end(CGF);
1391
1392 if (destructNonTrivialCStruct)
1393 CGF.pushDestroy(dtorKind: QualType::DK_nontrivial_c_struct, addr: Dest.getAddress(),
1394 type: E->getType());
1395
1396 CGF.EmitBlock(BB: ContBlock);
1397}
1398
1399void AggExprEmitter::VisitChooseExpr(const ChooseExpr *CE) {
1400 Visit(E: CE->getChosenSubExpr());
1401}
1402
1403void AggExprEmitter::VisitVAArgExpr(VAArgExpr *VE) {
1404 Address ArgValue = Address::invalid();
1405 CGF.EmitVAArg(VE, VAListAddr&: ArgValue, Slot: Dest);
1406
1407 // If EmitVAArg fails, emit an error.
1408 if (!ArgValue.isValid()) {
1409 CGF.ErrorUnsupported(S: VE, Type: "aggregate va_arg expression");
1410 return;
1411 }
1412}
1413
1414void AggExprEmitter::VisitCXXBindTemporaryExpr(CXXBindTemporaryExpr *E) {
1415 // Ensure that we have a slot, but if we already do, remember
1416 // whether it was externally destructed.
1417 bool wasExternallyDestructed = Dest.isExternallyDestructed();
1418 EnsureDest(T: E->getType());
1419
1420 // We're going to push a destructor if there isn't already one.
1421 Dest.setExternallyDestructed();
1422
1423 Visit(E: E->getSubExpr());
1424
1425 // Push that destructor we promised.
1426 if (!wasExternallyDestructed)
1427 CGF.EmitCXXTemporary(Temporary: E->getTemporary(), TempType: E->getType(), Ptr: Dest.getAddress());
1428}
1429
1430void AggExprEmitter::VisitCXXConstructExpr(const CXXConstructExpr *E) {
1431 AggValueSlot Slot = EnsureSlot(T: E->getType());
1432 CGF.EmitCXXConstructExpr(E, Dest: Slot);
1433}
1434
1435void AggExprEmitter::VisitCXXInheritedCtorInitExpr(
1436 const CXXInheritedCtorInitExpr *E) {
1437 AggValueSlot Slot = EnsureSlot(T: E->getType());
1438 CGF.EmitInheritedCXXConstructorCall(D: E->getConstructor(), ForVirtualBase: E->constructsVBase(),
1439 This: Slot.getAddress(),
1440 InheritedFromVBase: E->inheritedFromVBase(), E);
1441}
1442
1443void AggExprEmitter::VisitLambdaExpr(LambdaExpr *E) {
1444 AggValueSlot Slot = EnsureSlot(T: E->getType());
1445 LValue SlotLV = CGF.MakeAddrLValue(Addr: Slot.getAddress(), T: E->getType());
1446
1447 // We'll need to enter cleanup scopes in case any of the element
1448 // initializers throws an exception or contains branch out of the expressions.
1449 CodeGenFunction::CleanupDeactivationScope scope(CGF);
1450
1451 CXXRecordDecl::field_iterator CurField = E->getLambdaClass()->field_begin();
1452 for (LambdaExpr::const_capture_init_iterator i = E->capture_init_begin(),
1453 e = E->capture_init_end();
1454 i != e; ++i, ++CurField) {
1455 // Emit initialization
1456 LValue LV = CGF.EmitLValueForFieldInitialization(Base: SlotLV, Field: *CurField);
1457 if (CurField->hasCapturedVLAType()) {
1458 CGF.EmitLambdaVLACapture(VAT: CurField->getCapturedVLAType(), LV);
1459 continue;
1460 }
1461
1462 EmitInitializationToLValue(E: *i, Address: LV);
1463
1464 // Push a destructor if necessary.
1465 if (QualType::DestructionKind DtorKind =
1466 CurField->getType().isDestructedType()) {
1467 assert(LV.isSimple());
1468 if (DtorKind)
1469 CGF.pushDestroyAndDeferDeactivation(cleanupKind: NormalAndEHCleanup, addr: LV.getAddress(),
1470 type: CurField->getType(),
1471 destroyer: CGF.getDestroyer(destructionKind: DtorKind), useEHCleanupForArray: false);
1472 }
1473 }
1474}
1475
1476void AggExprEmitter::VisitExprWithCleanups(ExprWithCleanups *E) {
1477 CodeGenFunction::RunCleanupsScope cleanups(CGF);
1478 Visit(E: E->getSubExpr());
1479}
1480
1481void AggExprEmitter::VisitCXXScalarValueInitExpr(CXXScalarValueInitExpr *E) {
1482 QualType T = E->getType();
1483 AggValueSlot Slot = EnsureSlot(T);
1484 EmitNullInitializationToLValue(Address: CGF.MakeAddrLValue(Addr: Slot.getAddress(), T));
1485}
1486
1487void AggExprEmitter::VisitImplicitValueInitExpr(ImplicitValueInitExpr *E) {
1488 QualType T = E->getType();
1489 AggValueSlot Slot = EnsureSlot(T);
1490 EmitNullInitializationToLValue(Address: CGF.MakeAddrLValue(Addr: Slot.getAddress(), T));
1491}
1492
1493/// Determine whether the given cast kind is known to always convert values
1494/// with all zero bits in their value representation to values with all zero
1495/// bits in their value representation.
1496static bool castPreservesZero(const CastExpr *CE) {
1497 switch (CE->getCastKind()) {
1498 // No-ops.
1499 case CK_NoOp:
1500 case CK_UserDefinedConversion:
1501 case CK_ConstructorConversion:
1502 case CK_BitCast:
1503 case CK_ToUnion:
1504 case CK_ToVoid:
1505 // Conversions between (possibly-complex) integral, (possibly-complex)
1506 // floating-point, and bool.
1507 case CK_BooleanToSignedIntegral:
1508 case CK_FloatingCast:
1509 case CK_FloatingComplexCast:
1510 case CK_FloatingComplexToBoolean:
1511 case CK_FloatingComplexToIntegralComplex:
1512 case CK_FloatingComplexToReal:
1513 case CK_FloatingRealToComplex:
1514 case CK_FloatingToBoolean:
1515 case CK_FloatingToIntegral:
1516 case CK_IntegralCast:
1517 case CK_IntegralComplexCast:
1518 case CK_IntegralComplexToBoolean:
1519 case CK_IntegralComplexToFloatingComplex:
1520 case CK_IntegralComplexToReal:
1521 case CK_IntegralRealToComplex:
1522 case CK_IntegralToBoolean:
1523 case CK_IntegralToFloating:
1524 // Reinterpreting integers as pointers and vice versa.
1525 case CK_IntegralToPointer:
1526 case CK_PointerToIntegral:
1527 // Language extensions.
1528 case CK_VectorSplat:
1529 case CK_MatrixCast:
1530 case CK_NonAtomicToAtomic:
1531 case CK_AtomicToNonAtomic:
1532 case CK_HLSLVectorTruncation:
1533 case CK_HLSLMatrixTruncation:
1534 case CK_HLSLElementwiseCast:
1535 case CK_HLSLAggregateSplatCast:
1536 return true;
1537
1538 case CK_BaseToDerivedMemberPointer:
1539 case CK_DerivedToBaseMemberPointer:
1540 case CK_MemberPointerToBoolean:
1541 case CK_NullToMemberPointer:
1542 case CK_ReinterpretMemberPointer:
1543 // FIXME: ABI-dependent.
1544 return false;
1545
1546 case CK_AnyPointerToBlockPointerCast:
1547 case CK_BlockPointerToObjCPointerCast:
1548 case CK_CPointerToObjCPointerCast:
1549 case CK_ObjCObjectLValueCast:
1550 case CK_IntToOCLSampler:
1551 case CK_ZeroToOCLOpaqueType:
1552 // FIXME: Check these.
1553 return false;
1554
1555 case CK_FixedPointCast:
1556 case CK_FixedPointToBoolean:
1557 case CK_FixedPointToFloating:
1558 case CK_FixedPointToIntegral:
1559 case CK_FloatingToFixedPoint:
1560 case CK_IntegralToFixedPoint:
1561 // FIXME: Do all fixed-point types represent zero as all 0 bits?
1562 return false;
1563
1564 case CK_AddressSpaceConversion:
1565 case CK_BaseToDerived:
1566 case CK_DerivedToBase:
1567 case CK_Dynamic:
1568 case CK_NullToPointer:
1569 case CK_PointerToBoolean:
1570 // FIXME: Preserves zeroes only if zero pointers and null pointers have the
1571 // same representation in all involved address spaces.
1572 return false;
1573
1574 case CK_ARCConsumeObject:
1575 case CK_ARCExtendBlockObject:
1576 case CK_ARCProduceObject:
1577 case CK_ARCReclaimReturnedObject:
1578 case CK_CopyAndAutoreleaseBlockObject:
1579 case CK_ArrayToPointerDecay:
1580 case CK_FunctionToPointerDecay:
1581 case CK_BuiltinFnToFnPtr:
1582 case CK_Dependent:
1583 case CK_LValueBitCast:
1584 case CK_LValueToRValue:
1585 case CK_LValueToRValueBitCast:
1586 case CK_UncheckedDerivedToBase:
1587 case CK_HLSLArrayRValue:
1588 return false;
1589 }
1590 llvm_unreachable("Unhandled clang::CastKind enum");
1591}
1592
1593/// isSimpleZero - If emitting this value will obviously just cause a store of
1594/// zero to memory, return true. This can return false if uncertain, so it just
1595/// handles simple cases.
1596static bool isSimpleZero(const Expr *E, CodeGenFunction &CGF) {
1597 E = E->IgnoreParens();
1598 while (auto *CE = dyn_cast<CastExpr>(Val: E)) {
1599 if (!castPreservesZero(CE))
1600 break;
1601 E = CE->getSubExpr()->IgnoreParens();
1602 }
1603
1604 // 0
1605 if (const IntegerLiteral *IL = dyn_cast<IntegerLiteral>(Val: E))
1606 return IL->getValue() == 0;
1607 // +0.0
1608 if (const FloatingLiteral *FL = dyn_cast<FloatingLiteral>(Val: E))
1609 return FL->getValue().isPosZero();
1610 // int()
1611 if ((isa<ImplicitValueInitExpr>(Val: E) || isa<CXXScalarValueInitExpr>(Val: E)) &&
1612 CGF.getTypes().isZeroInitializable(T: E->getType()))
1613 return true;
1614 // (int*)0 - Null pointer expressions.
1615 if (const CastExpr *ICE = dyn_cast<CastExpr>(Val: E))
1616 return ICE->getCastKind() == CK_NullToPointer &&
1617 CGF.getTypes().isPointerZeroInitializable(T: E->getType()) &&
1618 !E->HasSideEffects(Ctx: CGF.getContext());
1619 // '\0'
1620 if (const CharacterLiteral *CL = dyn_cast<CharacterLiteral>(Val: E))
1621 return CL->getValue() == 0;
1622
1623 // Otherwise, hard case: conservatively return false.
1624 return false;
1625}
1626
1627void AggExprEmitter::EmitInitializationToLValue(Expr *E, LValue LV) {
1628 QualType type = LV.getType();
1629 // FIXME: Ignore result?
1630 // FIXME: Are initializers affected by volatile?
1631 if (Dest.isZeroed() && isSimpleZero(E, CGF)) {
1632 // Storing "i32 0" to a zero'd memory location is a noop.
1633 return;
1634 } else if (isa<ImplicitValueInitExpr>(Val: E) || isa<CXXScalarValueInitExpr>(Val: E)) {
1635 return EmitNullInitializationToLValue(Address: LV);
1636 } else if (isa<NoInitExpr>(Val: E)) {
1637 // Do nothing.
1638 return;
1639 } else if (type->isReferenceType()) {
1640 RValue RV = CGF.EmitReferenceBindingToExpr(E);
1641 return CGF.EmitStoreThroughLValue(Src: RV, Dst: LV);
1642 }
1643
1644 CGF.EmitInitializationToLValue(E, LV, IsZeroed: Dest.isZeroed());
1645}
1646
1647void AggExprEmitter::EmitNullInitializationToLValue(LValue lv) {
1648 QualType type = lv.getType();
1649
1650 // If the destination slot is already zeroed out before the aggregate is
1651 // copied into it, we don't have to emit any zeros here.
1652 if (Dest.isZeroed() && CGF.getTypes().isZeroInitializable(T: type))
1653 return;
1654
1655 if (CGF.hasScalarEvaluationKind(T: type)) {
1656 // For non-aggregates, we can store the appropriate null constant.
1657 llvm::Value *null = CGF.CGM.EmitNullConstant(T: type);
1658 // Note that the following is not equivalent to
1659 // EmitStoreThroughBitfieldLValue for ARC types.
1660 if (lv.isBitField()) {
1661 CGF.EmitStoreThroughBitfieldLValue(Src: RValue::get(V: null), Dst: lv);
1662 } else {
1663 assert(lv.isSimple());
1664 CGF.EmitStoreOfScalar(value: null, lvalue: lv, /* isInitialization */ isInit: true);
1665 }
1666 } else {
1667 // There's a potential optimization opportunity in combining
1668 // memsets; that would be easy for arrays, but relatively
1669 // difficult for structures with the current code.
1670 CGF.EmitNullInitialization(DestPtr: lv.getAddress(), Ty: lv.getType());
1671 }
1672}
1673
1674void AggExprEmitter::VisitCXXParenListInitExpr(CXXParenListInitExpr *E) {
1675 VisitCXXParenListOrInitListExpr(ExprToVisit: E, Args: E->getInitExprs(),
1676 InitializedFieldInUnion: E->getInitializedFieldInUnion(),
1677 ArrayFiller: E->getArrayFiller());
1678}
1679
1680void AggExprEmitter::VisitInitListExpr(InitListExpr *E) {
1681 if (E->hadArrayRangeDesignator())
1682 CGF.ErrorUnsupported(S: E, Type: "GNU array range designator extension");
1683
1684 if (E->isTransparent())
1685 return Visit(E: E->getInit(Init: 0));
1686
1687 VisitCXXParenListOrInitListExpr(
1688 ExprToVisit: E, Args: E->inits(), InitializedFieldInUnion: E->getInitializedFieldInUnion(), ArrayFiller: E->getArrayFiller());
1689}
1690
1691void AggExprEmitter::VisitCXXParenListOrInitListExpr(
1692 Expr *ExprToVisit, ArrayRef<Expr *> InitExprs,
1693 FieldDecl *InitializedFieldInUnion, Expr *ArrayFiller) {
1694#if 0
1695 // FIXME: Assess perf here? Figure out what cases are worth optimizing here
1696 // (Length of globals? Chunks of zeroed-out space?).
1697 //
1698 // If we can, prefer a copy from a global; this is a lot less code for long
1699 // globals, and it's easier for the current optimizers to analyze.
1700 if (llvm::Constant *C =
1701 CGF.CGM.EmitConstantExpr(ExprToVisit, ExprToVisit->getType(), &CGF)) {
1702 llvm::GlobalVariable* GV =
1703 new llvm::GlobalVariable(CGF.CGM.getModule(), C->getType(), true,
1704 llvm::GlobalValue::InternalLinkage, C, "");
1705 EmitFinalDestCopy(ExprToVisit->getType(),
1706 CGF.MakeAddrLValue(GV, ExprToVisit->getType()));
1707 return;
1708 }
1709#endif
1710
1711 // HLSL initialization lists in the AST are an expansion which can contain
1712 // side-effecting expressions wrapped in opaque value expressions. To properly
1713 // emit these we need to emit the opaque values before we emit the argument
1714 // expressions themselves. This is a little hacky, but it prevents us needing
1715 // to do a bigger AST-level change for a language feature that we need
1716 // deprecate in the near future. See related HLSL language proposals:
1717 // * 0005-strict-initializer-lists.md
1718 // * https://github.com/microsoft/hlsl-specs/pull/325
1719 if (CGF.getLangOpts().HLSL && isa<InitListExpr>(Val: ExprToVisit))
1720 CGF.CGM.getHLSLRuntime().emitInitListOpaqueValues(
1721 CGF, E: cast<InitListExpr>(Val: ExprToVisit));
1722
1723 AggValueSlot Dest = EnsureSlot(T: ExprToVisit->getType());
1724
1725 LValue DestLV = CGF.MakeAddrLValue(Addr: Dest.getAddress(), T: ExprToVisit->getType());
1726
1727 // Handle initialization of an array.
1728 if (ExprToVisit->getType()->isConstantArrayType()) {
1729 auto AType = cast<llvm::ArrayType>(Val: Dest.getAddress().getElementType());
1730 EmitArrayInit(DestPtr: Dest.getAddress(), AType, ArrayQTy: ExprToVisit->getType(), ExprToVisit,
1731 Args: InitExprs, ArrayFiller);
1732 return;
1733 } else if (ExprToVisit->getType()->isVariableArrayType()) {
1734 // A variable array type that has an initializer can only do empty
1735 // initialization. And because this feature is not exposed as an extension
1736 // in C++, we can safely memset the array memory to zero.
1737 assert(InitExprs.size() == 0 &&
1738 "you can only use an empty initializer with VLAs");
1739 CGF.EmitNullInitialization(DestPtr: Dest.getAddress(), Ty: ExprToVisit->getType());
1740 return;
1741 }
1742
1743 assert(ExprToVisit->getType()->isRecordType() &&
1744 "Only support structs/unions here!");
1745
1746 // Do struct initialization; this code just sets each individual member
1747 // to the approprate value. This makes bitfield support automatic;
1748 // the disadvantage is that the generated code is more difficult for
1749 // the optimizer, especially with bitfields.
1750 unsigned NumInitElements = InitExprs.size();
1751 RecordDecl *record = ExprToVisit->getType()->castAsRecordDecl();
1752
1753 // We'll need to enter cleanup scopes in case any of the element
1754 // initializers throws an exception.
1755 CodeGenFunction::CleanupDeactivationScope DeactivateCleanups(CGF);
1756
1757 unsigned curInitIndex = 0;
1758
1759 // Emit initialization of base classes.
1760 if (auto *CXXRD = dyn_cast<CXXRecordDecl>(Val: record)) {
1761 assert(NumInitElements >= CXXRD->getNumBases() &&
1762 "missing initializer for base class");
1763 for (auto &Base : CXXRD->bases()) {
1764 assert(!Base.isVirtual() && "should not see vbases here");
1765 auto *BaseRD = Base.getType()->getAsCXXRecordDecl();
1766 Address V = CGF.GetAddressOfDirectBaseInCompleteClass(
1767 Value: Dest.getAddress(), Derived: CXXRD, Base: BaseRD,
1768 /*isBaseVirtual*/ BaseIsVirtual: false);
1769 AggValueSlot AggSlot = AggValueSlot::forAddr(
1770 addr: V, quals: Qualifiers(), isDestructed: AggValueSlot::IsDestructed,
1771 needsGC: AggValueSlot::DoesNotNeedGCBarriers, isAliased: AggValueSlot::IsNotAliased,
1772 mayOverlap: CGF.getOverlapForBaseInit(RD: CXXRD, BaseRD, IsVirtual: Base.isVirtual()));
1773 CGF.EmitAggExpr(E: InitExprs[curInitIndex++], AS: AggSlot);
1774
1775 if (QualType::DestructionKind dtorKind =
1776 Base.getType().isDestructedType())
1777 CGF.pushDestroyAndDeferDeactivation(dtorKind, addr: V, type: Base.getType());
1778 }
1779 }
1780
1781 // Prepare a 'this' for CXXDefaultInitExprs.
1782 CodeGenFunction::FieldConstructionScope FCS(CGF, Dest.getAddress());
1783
1784 const bool ZeroInitPadding =
1785 CGF.CGM.shouldZeroInitPadding() && !Dest.isZeroed();
1786
1787 if (record->isUnion()) {
1788 // Only initialize one field of a union. The field itself is
1789 // specified by the initializer list.
1790 if (!InitializedFieldInUnion) {
1791 // Empty union; we have nothing to do.
1792
1793#ifndef NDEBUG
1794 // Make sure that it's really an empty and not a failure of
1795 // semantic analysis.
1796 for (const auto *Field : record->fields())
1797 assert(
1798 (Field->isUnnamedBitField() || Field->isAnonymousStructOrUnion()) &&
1799 "Only unnamed bitfields or anonymous class allowed");
1800#endif
1801 return;
1802 }
1803
1804 // FIXME: volatility
1805 FieldDecl *Field = InitializedFieldInUnion;
1806
1807 LValue FieldLoc = CGF.EmitLValueForFieldInitialization(Base: DestLV, Field);
1808 if (NumInitElements) {
1809 // Store the initializer into the field
1810 EmitInitializationToLValue(E: InitExprs[0], LV: FieldLoc);
1811 if (ZeroInitPadding) {
1812 uint64_t TotalSize = CGF.getContext().toBits(
1813 CharSize: Dest.getPreferredSize(Ctx&: CGF.getContext(), Type: DestLV.getType()));
1814 uint64_t FieldSize = CGF.getContext().getTypeSize(T: FieldLoc.getType());
1815 DoZeroInitPadding(PaddingStart&: FieldSize, PaddingEnd: TotalSize, NextField: nullptr);
1816 }
1817 } else {
1818 // Default-initialize to null.
1819 if (ZeroInitPadding)
1820 EmitNullInitializationToLValue(lv: DestLV);
1821 else
1822 EmitNullInitializationToLValue(lv: FieldLoc);
1823 }
1824 return;
1825 }
1826
1827 // Here we iterate over the fields; this makes it simpler to both
1828 // default-initialize fields and skip over unnamed fields.
1829 const ASTRecordLayout &Layout = CGF.getContext().getASTRecordLayout(D: record);
1830 uint64_t PaddingStart = 0;
1831
1832 for (const auto *field : record->fields()) {
1833 // We're done once we hit the flexible array member.
1834 if (field->getType()->isIncompleteArrayType())
1835 break;
1836
1837 // Always skip anonymous bitfields.
1838 if (field->isUnnamedBitField())
1839 continue;
1840
1841 // We're done if we reach the end of the explicit initializers, we
1842 // have a zeroed object, and the rest of the fields are
1843 // zero-initializable.
1844 if (curInitIndex == NumInitElements && Dest.isZeroed() &&
1845 CGF.getTypes().isZeroInitializable(T: ExprToVisit->getType()))
1846 break;
1847
1848 if (ZeroInitPadding)
1849 DoZeroInitPadding(PaddingStart,
1850 PaddingEnd: Layout.getFieldOffset(FieldNo: field->getFieldIndex()), NextField: field);
1851
1852 LValue LV = CGF.EmitLValueForFieldInitialization(Base: DestLV, Field: field);
1853 // We never generate write-barries for initialized fields.
1854 LV.setNonGC(true);
1855
1856 if (curInitIndex < NumInitElements) {
1857 // Store the initializer into the field.
1858 EmitInitializationToLValue(E: InitExprs[curInitIndex++], LV);
1859 } else {
1860 // We're out of initializers; default-initialize to null
1861 EmitNullInitializationToLValue(lv: LV);
1862 }
1863
1864 // Push a destructor if necessary.
1865 // FIXME: if we have an array of structures, all explicitly
1866 // initialized, we can end up pushing a linear number of cleanups.
1867 if (QualType::DestructionKind dtorKind =
1868 field->getType().isDestructedType()) {
1869 assert(LV.isSimple());
1870 if (dtorKind) {
1871 CGF.pushDestroyAndDeferDeactivation(cleanupKind: NormalAndEHCleanup, addr: LV.getAddress(),
1872 type: field->getType(),
1873 destroyer: CGF.getDestroyer(destructionKind: dtorKind), useEHCleanupForArray: false);
1874 }
1875 }
1876 }
1877 if (ZeroInitPadding) {
1878 uint64_t TotalSize = CGF.getContext().toBits(
1879 CharSize: Dest.getPreferredSize(Ctx&: CGF.getContext(), Type: DestLV.getType()));
1880 DoZeroInitPadding(PaddingStart, PaddingEnd: TotalSize, NextField: nullptr);
1881 }
1882}
1883
1884void AggExprEmitter::DoZeroInitPadding(uint64_t &PaddingStart,
1885 uint64_t PaddingEnd,
1886 const FieldDecl *NextField) {
1887
1888 auto InitBytes = [&](uint64_t StartBit, uint64_t EndBit) {
1889 CharUnits Start = CGF.getContext().toCharUnitsFromBits(BitSize: StartBit);
1890 CharUnits End = CGF.getContext().toCharUnitsFromBits(BitSize: EndBit);
1891 Address Addr = Dest.getAddress().withElementType(ElemTy: CGF.CharTy);
1892 if (!Start.isZero())
1893 Addr = Builder.CreateConstGEP(Addr, Index: Start.getQuantity());
1894 llvm::Constant *SizeVal = Builder.getInt64(C: (End - Start).getQuantity());
1895 CGF.Builder.CreateMemSet(Dest: Addr, Value: Builder.getInt8(C: 0), Size: SizeVal, IsVolatile: false);
1896 };
1897
1898 if (NextField != nullptr && NextField->isBitField()) {
1899 // For bitfield, zero init StorageSize before storing the bits. So we don't
1900 // need to handle big/little endian.
1901 const CGRecordLayout &RL =
1902 CGF.getTypes().getCGRecordLayout(NextField->getParent());
1903 const CGBitFieldInfo &Info = RL.getBitFieldInfo(FD: NextField);
1904 uint64_t StorageStart = CGF.getContext().toBits(CharSize: Info.StorageOffset);
1905 if (StorageStart + Info.StorageSize > PaddingStart) {
1906 if (StorageStart > PaddingStart)
1907 InitBytes(PaddingStart, StorageStart);
1908 Address Addr = Dest.getAddress();
1909 if (!Info.StorageOffset.isZero())
1910 Addr = Builder.CreateConstGEP(Addr: Addr.withElementType(ElemTy: CGF.CharTy),
1911 Index: Info.StorageOffset.getQuantity());
1912 Addr = Addr.withElementType(
1913 ElemTy: llvm::Type::getIntNTy(C&: CGF.getLLVMContext(), N: Info.StorageSize));
1914 Builder.CreateStore(Val: Builder.getIntN(N: Info.StorageSize, C: 0), Addr);
1915 PaddingStart = StorageStart + Info.StorageSize;
1916 }
1917 return;
1918 }
1919
1920 if (PaddingStart < PaddingEnd)
1921 InitBytes(PaddingStart, PaddingEnd);
1922 if (NextField != nullptr)
1923 PaddingStart =
1924 PaddingEnd + CGF.getContext().getTypeSize(T: NextField->getType());
1925}
1926
1927void AggExprEmitter::VisitArrayInitLoopExpr(const ArrayInitLoopExpr *E,
1928 llvm::Value *outerBegin) {
1929 // Emit the common subexpression.
1930 CodeGenFunction::OpaqueValueMapping binding(CGF, E->getCommonExpr());
1931
1932 Address destPtr = EnsureSlot(T: E->getType()).getAddress();
1933 uint64_t numElements = E->getArraySize().getZExtValue();
1934
1935 if (!numElements)
1936 return;
1937
1938 // destPtr is an array*. Construct an elementType* by drilling down a level.
1939 llvm::Value *zero = llvm::ConstantInt::get(Ty: CGF.SizeTy, V: 0);
1940 llvm::Value *indices[] = {zero, zero};
1941 llvm::Value *begin = Builder.CreateInBoundsGEP(Ty: destPtr.getElementType(),
1942 Ptr: destPtr.emitRawPointer(CGF),
1943 IdxList: indices, Name: "arrayinit.begin");
1944
1945 // Prepare to special-case multidimensional array initialization: we avoid
1946 // emitting multiple destructor loops in that case.
1947 if (!outerBegin)
1948 outerBegin = begin;
1949 ArrayInitLoopExpr *InnerLoop = dyn_cast<ArrayInitLoopExpr>(Val: E->getSubExpr());
1950
1951 QualType elementType =
1952 CGF.getContext().getAsArrayType(T: E->getType())->getElementType();
1953 CharUnits elementSize = CGF.getContext().getTypeSizeInChars(T: elementType);
1954 CharUnits elementAlign =
1955 destPtr.getAlignment().alignmentOfArrayElement(elementSize);
1956 llvm::Type *llvmElementType = CGF.ConvertTypeForMem(T: elementType);
1957
1958 llvm::BasicBlock *entryBB = Builder.GetInsertBlock();
1959 llvm::BasicBlock *bodyBB = CGF.createBasicBlock(name: "arrayinit.body");
1960
1961 // Jump into the body.
1962 CGF.EmitBlock(BB: bodyBB);
1963 llvm::PHINode *index =
1964 Builder.CreatePHI(Ty: zero->getType(), NumReservedValues: 2, Name: "arrayinit.index");
1965 index->addIncoming(V: zero, BB: entryBB);
1966 llvm::Value *element =
1967 Builder.CreateInBoundsGEP(Ty: llvmElementType, Ptr: begin, IdxList: index);
1968
1969 // Prepare for a cleanup.
1970 QualType::DestructionKind dtorKind = elementType.isDestructedType();
1971 EHScopeStack::stable_iterator cleanup;
1972 if (CGF.needsEHCleanup(kind: dtorKind) && !InnerLoop) {
1973 if (outerBegin->getType() != element->getType())
1974 outerBegin = Builder.CreateBitCast(V: outerBegin, DestTy: element->getType());
1975 CGF.pushRegularPartialArrayCleanup(arrayBegin: outerBegin, arrayEnd: element, elementType,
1976 elementAlignment: elementAlign,
1977 destroyer: CGF.getDestroyer(destructionKind: dtorKind));
1978 cleanup = CGF.EHStack.stable_begin();
1979 } else {
1980 dtorKind = QualType::DK_none;
1981 }
1982
1983 // Emit the actual filler expression.
1984 {
1985 // Temporaries created in an array initialization loop are destroyed
1986 // at the end of each iteration.
1987 CodeGenFunction::RunCleanupsScope CleanupsScope(CGF);
1988 CodeGenFunction::ArrayInitLoopExprScope Scope(CGF, index);
1989 LValue elementLV = CGF.MakeAddrLValue(
1990 Addr: Address(element, llvmElementType, elementAlign), T: elementType);
1991
1992 if (InnerLoop) {
1993 // If the subexpression is an ArrayInitLoopExpr, share its cleanup.
1994 auto elementSlot = AggValueSlot::forLValue(
1995 LV: elementLV, isDestructed: AggValueSlot::IsDestructed,
1996 needsGC: AggValueSlot::DoesNotNeedGCBarriers, isAliased: AggValueSlot::IsNotAliased,
1997 mayOverlap: AggValueSlot::DoesNotOverlap);
1998 AggExprEmitter(CGF, elementSlot, false)
1999 .VisitArrayInitLoopExpr(E: InnerLoop, outerBegin);
2000 } else
2001 EmitInitializationToLValue(E: E->getSubExpr(), LV: elementLV);
2002 }
2003
2004 // Move on to the next element.
2005 llvm::Value *nextIndex = Builder.CreateNUWAdd(
2006 LHS: index, RHS: llvm::ConstantInt::get(Ty: CGF.SizeTy, V: 1), Name: "arrayinit.next");
2007 index->addIncoming(V: nextIndex, BB: Builder.GetInsertBlock());
2008
2009 // Leave the loop if we're done.
2010 llvm::Value *done = Builder.CreateICmpEQ(
2011 LHS: nextIndex, RHS: llvm::ConstantInt::get(Ty: CGF.SizeTy, V: numElements),
2012 Name: "arrayinit.done");
2013 llvm::BasicBlock *endBB = CGF.createBasicBlock(name: "arrayinit.end");
2014 Builder.CreateCondBr(Cond: done, True: endBB, False: bodyBB);
2015
2016 CGF.EmitBlock(BB: endBB);
2017
2018 // Leave the partial-array cleanup if we entered one.
2019 if (dtorKind)
2020 CGF.DeactivateCleanupBlock(Cleanup: cleanup, DominatingIP: index);
2021}
2022
2023void AggExprEmitter::VisitDesignatedInitUpdateExpr(
2024 DesignatedInitUpdateExpr *E) {
2025 AggValueSlot Dest = EnsureSlot(T: E->getType());
2026
2027 LValue DestLV = CGF.MakeAddrLValue(Addr: Dest.getAddress(), T: E->getType());
2028 EmitInitializationToLValue(E: E->getBase(), LV: DestLV);
2029 VisitInitListExpr(E: E->getUpdater());
2030}
2031
2032//===----------------------------------------------------------------------===//
2033// Entry Points into this File
2034//===----------------------------------------------------------------------===//
2035
2036/// GetNumNonZeroBytesInInit - Get an approximate count of the number of
2037/// non-zero bytes that will be stored when outputting the initializer for the
2038/// specified initializer expression.
2039static CharUnits GetNumNonZeroBytesInInit(const Expr *E, CodeGenFunction &CGF) {
2040 if (auto *MTE = dyn_cast<MaterializeTemporaryExpr>(Val: E))
2041 E = MTE->getSubExpr();
2042 E = E->IgnoreParenNoopCasts(Ctx: CGF.getContext());
2043
2044 // 0 and 0.0 won't require any non-zero stores!
2045 if (isSimpleZero(E, CGF))
2046 return CharUnits::Zero();
2047
2048 // If this is an initlist expr, sum up the size of sizes of the (present)
2049 // elements. If this is something weird, assume the whole thing is non-zero.
2050 const InitListExpr *ILE = dyn_cast<InitListExpr>(Val: E);
2051 while (ILE && ILE->isTransparent())
2052 ILE = dyn_cast<InitListExpr>(Val: ILE->getInit(Init: 0));
2053 if (!ILE || !CGF.getTypes().isZeroInitializable(T: ILE->getType()))
2054 return CGF.getContext().getTypeSizeInChars(T: E->getType());
2055
2056 // InitListExprs for structs have to be handled carefully. If there are
2057 // reference members, we need to consider the size of the reference, not the
2058 // referencee. InitListExprs for unions and arrays can't have references.
2059 if (const RecordType *RT = E->getType()->getAsCanonical<RecordType>()) {
2060 if (!RT->isUnionType()) {
2061 RecordDecl *SD = RT->getDecl()->getDefinitionOrSelf();
2062 CharUnits NumNonZeroBytes = CharUnits::Zero();
2063
2064 unsigned ILEElement = 0;
2065 if (auto *CXXRD = dyn_cast<CXXRecordDecl>(Val: SD))
2066 while (ILEElement != CXXRD->getNumBases())
2067 NumNonZeroBytes +=
2068 GetNumNonZeroBytesInInit(E: ILE->getInit(Init: ILEElement++), CGF);
2069 for (const auto *Field : SD->fields()) {
2070 // We're done once we hit the flexible array member or run out of
2071 // InitListExpr elements.
2072 if (Field->getType()->isIncompleteArrayType() ||
2073 ILEElement == ILE->getNumInits())
2074 break;
2075 if (Field->isUnnamedBitField())
2076 continue;
2077
2078 const Expr *E = ILE->getInit(Init: ILEElement++);
2079
2080 // Reference values are always non-null and have the width of a pointer.
2081 if (Field->getType()->isReferenceType())
2082 NumNonZeroBytes += CGF.getContext().toCharUnitsFromBits(
2083 BitSize: CGF.getTarget().getPointerWidth(AddrSpace: LangAS::Default));
2084 else
2085 NumNonZeroBytes += GetNumNonZeroBytesInInit(E, CGF);
2086 }
2087
2088 return NumNonZeroBytes;
2089 }
2090 }
2091
2092 // FIXME: This overestimates the number of non-zero bytes for bit-fields.
2093 CharUnits NumNonZeroBytes = CharUnits::Zero();
2094 for (unsigned i = 0, e = ILE->getNumInits(); i != e; ++i)
2095 NumNonZeroBytes += GetNumNonZeroBytesInInit(E: ILE->getInit(Init: i), CGF);
2096 return NumNonZeroBytes;
2097}
2098
2099/// CheckAggExprForMemSetUse - If the initializer is large and has a lot of
2100/// zeros in it, emit a memset and avoid storing the individual zeros.
2101///
2102static void CheckAggExprForMemSetUse(AggValueSlot &Slot, const Expr *E,
2103 CodeGenFunction &CGF) {
2104 // If the slot is already known to be zeroed, nothing to do. Don't mess with
2105 // volatile stores.
2106 if (Slot.isZeroed() || Slot.isVolatile() || !Slot.getAddress().isValid())
2107 return;
2108
2109 // C++ objects with a user-declared constructor don't need zero'ing.
2110 if (CGF.getLangOpts().CPlusPlus)
2111 if (const RecordType *RT = CGF.getContext()
2112 .getBaseElementType(QT: E->getType())
2113 ->getAsCanonical<RecordType>()) {
2114 const auto *RD = cast<CXXRecordDecl>(Val: RT->getDecl());
2115 if (RD->hasUserDeclaredConstructor())
2116 return;
2117 }
2118
2119 // If the type is 16-bytes or smaller, prefer individual stores over memset.
2120 CharUnits Size = Slot.getPreferredSize(Ctx&: CGF.getContext(), Type: E->getType());
2121 if (Size <= CharUnits::fromQuantity(Quantity: 16))
2122 return;
2123
2124 // Check to see if over 3/4 of the initializer are known to be zero. If so,
2125 // we prefer to emit memset + individual stores for the rest.
2126 CharUnits NumNonZeroBytes = GetNumNonZeroBytesInInit(E, CGF);
2127 if (NumNonZeroBytes * 4 > Size)
2128 return;
2129
2130 // Okay, it seems like a good idea to use an initial memset, emit the call.
2131 llvm::Constant *SizeVal = CGF.Builder.getInt64(C: Size.getQuantity());
2132
2133 Address Loc = Slot.getAddress().withElementType(ElemTy: CGF.Int8Ty);
2134 CGF.Builder.CreateMemSet(Dest: Loc, Value: CGF.Builder.getInt8(C: 0), Size: SizeVal, IsVolatile: false);
2135
2136 // Tell the AggExprEmitter that the slot is known zero.
2137 Slot.setZeroed();
2138}
2139
2140/// EmitAggExpr - Emit the computation of the specified expression of aggregate
2141/// type. The result is computed into DestPtr. Note that if DestPtr is null,
2142/// the value of the aggregate expression is not needed. If VolatileDest is
2143/// true, DestPtr cannot be 0.
2144void CodeGenFunction::EmitAggExpr(const Expr *E, AggValueSlot Slot) {
2145 assert(E && hasAggregateEvaluationKind(E->getType()) &&
2146 "Invalid aggregate expression to emit");
2147 assert((Slot.getAddress().isValid() || Slot.isIgnored()) &&
2148 "slot has bits but no address");
2149
2150 // Optimize the slot if possible.
2151 CheckAggExprForMemSetUse(Slot, E, CGF&: *this);
2152
2153 AggExprEmitter(*this, Slot, Slot.isIgnored()).Visit(E: const_cast<Expr *>(E));
2154}
2155
2156LValue CodeGenFunction::EmitAggExprToLValue(const Expr *E) {
2157 assert(hasAggregateEvaluationKind(E->getType()) && "Invalid argument!");
2158 Address Temp = CreateMemTemp(T: E->getType());
2159 LValue LV = MakeAddrLValue(Addr: Temp, T: E->getType());
2160 EmitAggExpr(E, Slot: AggValueSlot::forLValue(LV, isDestructed: AggValueSlot::IsNotDestructed,
2161 needsGC: AggValueSlot::DoesNotNeedGCBarriers,
2162 isAliased: AggValueSlot::IsNotAliased,
2163 mayOverlap: AggValueSlot::DoesNotOverlap));
2164 return LV;
2165}
2166
2167void CodeGenFunction::EmitAggFinalDestCopy(QualType Type, AggValueSlot Dest,
2168 const LValue &Src,
2169 ExprValueKind SrcKind) {
2170 return AggExprEmitter(*this, Dest, Dest.isIgnored())
2171 .EmitFinalDestCopy(type: Type, src: Src, SrcValueKind: SrcKind);
2172}
2173
2174AggValueSlot::Overlap_t
2175CodeGenFunction::getOverlapForFieldInit(const FieldDecl *FD) {
2176 if (!FD->hasAttr<NoUniqueAddressAttr>() || !FD->getType()->isRecordType())
2177 return AggValueSlot::DoesNotOverlap;
2178
2179 // Empty fields can overlap earlier fields.
2180 if (FD->getType()->getAsCXXRecordDecl()->isEmpty())
2181 return AggValueSlot::MayOverlap;
2182
2183 // If the field lies entirely within the enclosing class's nvsize, its tail
2184 // padding cannot overlap any already-initialized object. (The only subobjects
2185 // with greater addresses that might already be initialized are vbases.)
2186 const RecordDecl *ClassRD = FD->getParent();
2187 const ASTRecordLayout &Layout = getContext().getASTRecordLayout(D: ClassRD);
2188 if (Layout.getFieldOffset(FieldNo: FD->getFieldIndex()) +
2189 getContext().getTypeSize(T: FD->getType()) <=
2190 (uint64_t)getContext().toBits(CharSize: Layout.getNonVirtualSize()))
2191 return AggValueSlot::DoesNotOverlap;
2192
2193 // The tail padding may contain values we need to preserve.
2194 return AggValueSlot::MayOverlap;
2195}
2196
2197AggValueSlot::Overlap_t CodeGenFunction::getOverlapForBaseInit(
2198 const CXXRecordDecl *RD, const CXXRecordDecl *BaseRD, bool IsVirtual) {
2199 // If the most-derived object is a field declared with [[no_unique_address]],
2200 // the tail padding of any virtual base could be reused for other subobjects
2201 // of that field's class.
2202 if (IsVirtual)
2203 return AggValueSlot::MayOverlap;
2204
2205 // Empty bases can overlap earlier bases.
2206 if (BaseRD->isEmpty())
2207 return AggValueSlot::MayOverlap;
2208
2209 // If the base class is laid out entirely within the nvsize of the derived
2210 // class, its tail padding cannot yet be initialized, so we can issue
2211 // stores at the full width of the base class.
2212 const ASTRecordLayout &Layout = getContext().getASTRecordLayout(D: RD);
2213 if (Layout.getBaseClassOffset(Base: BaseRD) +
2214 getContext().getASTRecordLayout(D: BaseRD).getSize() <=
2215 Layout.getNonVirtualSize())
2216 return AggValueSlot::DoesNotOverlap;
2217
2218 // The tail padding may contain values we need to preserve.
2219 return AggValueSlot::MayOverlap;
2220}
2221
2222void CodeGenFunction::EmitAggregateCopy(LValue Dest, LValue Src, QualType Ty,
2223 AggValueSlot::Overlap_t MayOverlap,
2224 bool isVolatile) {
2225 assert(!Ty->isAnyComplexType() && "Shouldn't happen for complex");
2226
2227 Address DestPtr = Dest.getAddress();
2228 Address SrcPtr = Src.getAddress();
2229
2230 if (getLangOpts().CPlusPlus) {
2231 if (const auto *Record = Ty->getAsCXXRecordDecl()) {
2232 assert((Record->hasTrivialCopyConstructor() ||
2233 Record->hasTrivialCopyAssignment() ||
2234 Record->hasTrivialMoveConstructor() ||
2235 Record->hasTrivialMoveAssignment() ||
2236 Record->hasAttr<TrivialABIAttr>() || Record->isUnion()) &&
2237 "Trying to aggregate-copy a type without a trivial copy/move "
2238 "constructor or assignment operator");
2239 // Ignore empty classes in C++.
2240 if (Record->isEmpty())
2241 return;
2242 }
2243 }
2244
2245 if (getLangOpts().CUDAIsDevice) {
2246 if (Ty->isCUDADeviceBuiltinSurfaceType()) {
2247 if (getTargetHooks().emitCUDADeviceBuiltinSurfaceDeviceCopy(CGF&: *this, Dst: Dest,
2248 Src))
2249 return;
2250 } else if (Ty->isCUDADeviceBuiltinTextureType()) {
2251 if (getTargetHooks().emitCUDADeviceBuiltinTextureDeviceCopy(CGF&: *this, Dst: Dest,
2252 Src))
2253 return;
2254 }
2255 }
2256
2257 if (getLangOpts().HLSL && Ty.getAddressSpace() == LangAS::hlsl_constant)
2258 if (CGM.getHLSLRuntime().emitBufferCopy(CGF&: *this, DestPtr, SrcPtr, CType: Ty))
2259 return;
2260
2261 // Aggregate assignment turns into llvm.memcpy. This is almost valid per
2262 // C99 6.5.16.1p3, which states "If the value being stored in an object is
2263 // read from another object that overlaps in anyway the storage of the first
2264 // object, then the overlap shall be exact and the two objects shall have
2265 // qualified or unqualified versions of a compatible type."
2266 //
2267 // memcpy is not defined if the source and destination pointers are exactly
2268 // equal, but other compilers do this optimization, and almost every memcpy
2269 // implementation handles this case safely. If there is a libc that does not
2270 // safely handle this, we can add a target hook.
2271
2272 // Get data size info for this aggregate. Don't copy the tail padding if this
2273 // might be a potentially-overlapping subobject, since the tail padding might
2274 // be occupied by a different object. Otherwise, copying it is fine.
2275 TypeInfoChars TypeInfo;
2276 if (MayOverlap)
2277 TypeInfo = getContext().getTypeInfoDataSizeInChars(T: Ty);
2278 else
2279 TypeInfo = getContext().getTypeInfoInChars(T: Ty);
2280
2281 llvm::Value *SizeVal = nullptr;
2282 if (TypeInfo.Width.isZero()) {
2283 // But note that getTypeInfo returns 0 for a VLA.
2284 if (auto *VAT = dyn_cast_or_null<VariableArrayType>(
2285 Val: getContext().getAsArrayType(T: Ty))) {
2286 QualType BaseEltTy;
2287 SizeVal = emitArrayLength(arrayType: VAT, baseType&: BaseEltTy, addr&: DestPtr);
2288 TypeInfo = getContext().getTypeInfoInChars(T: BaseEltTy);
2289 assert(!TypeInfo.Width.isZero());
2290 SizeVal = Builder.CreateNUWMul(
2291 LHS: SizeVal,
2292 RHS: llvm::ConstantInt::get(Ty: SizeTy, V: TypeInfo.Width.getQuantity()));
2293 }
2294 }
2295 if (!SizeVal) {
2296 SizeVal = llvm::ConstantInt::get(Ty: SizeTy, V: TypeInfo.Width.getQuantity());
2297 }
2298
2299 // FIXME: If we have a volatile struct, the optimizer can remove what might
2300 // appear to be `extra' memory ops:
2301 //
2302 // volatile struct { int i; } a, b;
2303 //
2304 // int main() {
2305 // a = b;
2306 // a = b;
2307 // }
2308 //
2309 // we need to use a different call here. We use isVolatile to indicate when
2310 // either the source or the destination is volatile.
2311
2312 DestPtr = DestPtr.withElementType(ElemTy: Int8Ty);
2313 SrcPtr = SrcPtr.withElementType(ElemTy: Int8Ty);
2314
2315 // Don't do any of the memmove_collectable tests if GC isn't set.
2316 if (CGM.getLangOpts().getGC() == LangOptions::NonGC) {
2317 // fall through
2318 } else if (const auto *Record = Ty->getAsRecordDecl()) {
2319 if (Record->hasObjectMember()) {
2320 CGM.getObjCRuntime().EmitGCMemmoveCollectable(CGF&: *this, DestPtr, SrcPtr,
2321 Size: SizeVal);
2322 return;
2323 }
2324 } else if (Ty->isArrayType()) {
2325 QualType BaseType = getContext().getBaseElementType(QT: Ty);
2326 if (const auto *Record = BaseType->getAsRecordDecl()) {
2327 if (Record->hasObjectMember()) {
2328 CGM.getObjCRuntime().EmitGCMemmoveCollectable(CGF&: *this, DestPtr, SrcPtr,
2329 Size: SizeVal);
2330 return;
2331 }
2332 }
2333 }
2334
2335 auto *Inst = Builder.CreateMemCpy(Dest: DestPtr, Src: SrcPtr, Size: SizeVal, IsVolatile: isVolatile);
2336 addInstToCurrentSourceAtom(KeyInstruction: Inst, Backup: nullptr);
2337
2338 // Determine the metadata to describe the position of any padding in this
2339 // memcpy, as well as the TBAA tags for the members of the struct, in case
2340 // the optimizer wishes to expand it in to scalar memory operations.
2341 if (llvm::MDNode *TBAAStructTag = CGM.getTBAAStructInfo(QTy: Ty))
2342 Inst->setMetadata(KindID: llvm::LLVMContext::MD_tbaa_struct, Node: TBAAStructTag);
2343
2344 if (CGM.getCodeGenOpts().NewStructPathTBAA) {
2345 TBAAAccessInfo TBAAInfo = CGM.mergeTBAAInfoForMemoryTransfer(
2346 DestInfo: Dest.getTBAAInfo(), SrcInfo: Src.getTBAAInfo());
2347 CGM.DecorateInstructionWithTBAA(Inst, TBAAInfo);
2348 }
2349}
2350