1 //===--- CodeGenFunction.cpp - Emit LLVM Code from ASTs for a Function ----===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This coordinates the per-function state used while generating code.
12 //===----------------------------------------------------------------------===//
14 #include "CodeGenFunction.h"
16 #include "CGCleanup.h"
17 #include "CGCUDARuntime.h"
19 #include "CGDebugInfo.h"
20 #include "CGOpenMPRuntime.h"
21 #include "CodeGenModule.h"
22 #include "CodeGenPGO.h"
23 #include "TargetInfo.h"
24 #include "clang/AST/ASTContext.h"
25 #include "clang/AST/Decl.h"
26 #include "clang/AST/DeclCXX.h"
27 #include "clang/AST/StmtCXX.h"
28 #include "clang/AST/StmtObjC.h"
29 #include "clang/Basic/Builtins.h"
30 #include "clang/Basic/TargetInfo.h"
31 #include "clang/CodeGen/CGFunctionInfo.h"
32 #include "clang/Frontend/CodeGenOptions.h"
33 #include "clang/Sema/SemaDiagnostic.h"
34 #include "llvm/IR/DataLayout.h"
35 #include "llvm/IR/Intrinsics.h"
36 #include "llvm/IR/MDBuilder.h"
37 #include "llvm/IR/Operator.h"
38 using namespace clang;
39 using namespace CodeGen;
41 /// shouldEmitLifetimeMarkers - Decide whether we need emit the life-time
43 static bool shouldEmitLifetimeMarkers(const CodeGenOptions &CGOpts,
44 const LangOptions &LangOpts) {
45 // Asan uses markers for use-after-scope checks.
46 if (CGOpts.SanitizeAddressUseAfterScope)
49 // Disable lifetime markers in msan builds.
50 // FIXME: Remove this when msan works with lifetime markers.
51 if (LangOpts.Sanitize.has(SanitizerKind::Memory))
54 // For now, only in optimized builds.
55 return CGOpts.OptimizationLevel != 0;
58 CodeGenFunction::CodeGenFunction(CodeGenModule &cgm, bool suppressNewContext)
59 : CodeGenTypeCache(cgm), CGM(cgm), Target(cgm.getTarget()),
60 Builder(cgm, cgm.getModule().getContext(), llvm::ConstantFolder(),
61 CGBuilderInserterTy(this)),
62 CurFn(nullptr), ReturnValue(Address::invalid()),
63 CapturedStmtInfo(nullptr), SanOpts(CGM.getLangOpts().Sanitize),
64 IsSanitizerScope(false), CurFuncIsThunk(false), AutoreleaseResult(false),
65 SawAsmBlock(false), IsOutlinedSEHHelper(false), BlockInfo(nullptr),
66 BlockPointer(nullptr), LambdaThisCaptureField(nullptr),
67 NormalCleanupDest(nullptr), NextCleanupDestIndex(1),
68 FirstBlockInfo(nullptr), EHResumeBlock(nullptr), ExceptionSlot(nullptr),
69 EHSelectorSlot(nullptr), DebugInfo(CGM.getModuleDebugInfo()),
70 DisableDebugInfo(false), DidCallStackSave(false), IndirectBranch(nullptr),
71 PGO(cgm), SwitchInsn(nullptr), SwitchWeights(nullptr),
72 CaseRangeBlock(nullptr), UnreachableBlock(nullptr), NumReturnExprs(0),
73 NumSimpleReturnExprs(0), CXXABIThisDecl(nullptr),
74 CXXABIThisValue(nullptr), CXXThisValue(nullptr),
75 CXXStructorImplicitParamDecl(nullptr),
76 CXXStructorImplicitParamValue(nullptr), OutermostConditional(nullptr),
77 CurLexicalScope(nullptr), TerminateLandingPad(nullptr),
78 TerminateHandler(nullptr), TrapBB(nullptr),
79 ShouldEmitLifetimeMarkers(
80 shouldEmitLifetimeMarkers(CGM.getCodeGenOpts(), CGM.getLangOpts())) {
81 if (!suppressNewContext)
82 CGM.getCXXABI().getMangleContext().startNewFunction();
84 llvm::FastMathFlags FMF;
85 if (CGM.getLangOpts().FastMath)
86 FMF.setUnsafeAlgebra();
87 if (CGM.getLangOpts().FiniteMathOnly) {
91 if (CGM.getCodeGenOpts().NoNaNsFPMath) {
94 if (CGM.getCodeGenOpts().NoSignedZeros) {
95 FMF.setNoSignedZeros();
97 if (CGM.getCodeGenOpts().ReciprocalMath) {
98 FMF.setAllowReciprocal();
100 Builder.setFastMathFlags(FMF);
103 CodeGenFunction::~CodeGenFunction() {
104 assert(LifetimeExtendedCleanupStack.empty() && "failed to emit a cleanup");
106 // If there are any unclaimed block infos, go ahead and destroy them
107 // now. This can happen if IR-gen gets clever and skips evaluating
110 destroyBlockInfos(FirstBlockInfo);
112 if (getLangOpts().OpenMP) {
113 CGM.getOpenMPRuntime().functionFinished(*this);
117 CharUnits CodeGenFunction::getNaturalPointeeTypeAlignment(QualType T,
118 AlignmentSource *Source) {
119 return getNaturalTypeAlignment(T->getPointeeType(), Source,
120 /*forPointee*/ true);
123 CharUnits CodeGenFunction::getNaturalTypeAlignment(QualType T,
124 AlignmentSource *Source,
125 bool forPointeeType) {
126 // Honor alignment typedef attributes even on incomplete types.
127 // We also honor them straight for C++ class types, even as pointees;
128 // there's an expressivity gap here.
129 if (auto TT = T->getAs<TypedefType>()) {
130 if (auto Align = TT->getDecl()->getMaxAlignment()) {
131 if (Source) *Source = AlignmentSource::AttributedType;
132 return getContext().toCharUnitsFromBits(Align);
136 if (Source) *Source = AlignmentSource::Type;
139 if (T->isIncompleteType()) {
140 Alignment = CharUnits::One(); // Shouldn't be used, but pessimistic is best.
142 // For C++ class pointees, we don't know whether we're pointing at a
143 // base or a complete object, so we generally need to use the
144 // non-virtual alignment.
145 const CXXRecordDecl *RD;
146 if (forPointeeType && (RD = T->getAsCXXRecordDecl())) {
147 Alignment = CGM.getClassPointerAlignment(RD);
149 Alignment = getContext().getTypeAlignInChars(T);
152 // Cap to the global maximum type alignment unless the alignment
153 // was somehow explicit on the type.
154 if (unsigned MaxAlign = getLangOpts().MaxTypeAlign) {
155 if (Alignment.getQuantity() > MaxAlign &&
156 !getContext().isAlignmentRequired(T))
157 Alignment = CharUnits::fromQuantity(MaxAlign);
163 LValue CodeGenFunction::MakeNaturalAlignAddrLValue(llvm::Value *V, QualType T) {
164 AlignmentSource AlignSource;
165 CharUnits Alignment = getNaturalTypeAlignment(T, &AlignSource);
166 return LValue::MakeAddr(Address(V, Alignment), T, getContext(), AlignSource,
170 /// Given a value of type T* that may not be to a complete object,
171 /// construct an l-value with the natural pointee alignment of T.
173 CodeGenFunction::MakeNaturalAlignPointeeAddrLValue(llvm::Value *V, QualType T) {
174 AlignmentSource AlignSource;
175 CharUnits Align = getNaturalTypeAlignment(T, &AlignSource, /*pointee*/ true);
176 return MakeAddrLValue(Address(V, Align), T, AlignSource);
180 llvm::Type *CodeGenFunction::ConvertTypeForMem(QualType T) {
181 return CGM.getTypes().ConvertTypeForMem(T);
184 llvm::Type *CodeGenFunction::ConvertType(QualType T) {
185 return CGM.getTypes().ConvertType(T);
188 TypeEvaluationKind CodeGenFunction::getEvaluationKind(QualType type) {
189 type = type.getCanonicalType();
191 switch (type->getTypeClass()) {
192 #define TYPE(name, parent)
193 #define ABSTRACT_TYPE(name, parent)
194 #define NON_CANONICAL_TYPE(name, parent) case Type::name:
195 #define DEPENDENT_TYPE(name, parent) case Type::name:
196 #define NON_CANONICAL_UNLESS_DEPENDENT_TYPE(name, parent) case Type::name:
197 #include "clang/AST/TypeNodes.def"
198 llvm_unreachable("non-canonical or dependent type in IR-generation");
201 llvm_unreachable("undeduced auto type in IR-generation");
203 // Various scalar types.
206 case Type::BlockPointer:
207 case Type::LValueReference:
208 case Type::RValueReference:
209 case Type::MemberPointer:
211 case Type::ExtVector:
212 case Type::FunctionProto:
213 case Type::FunctionNoProto:
215 case Type::ObjCObjectPointer:
223 // Arrays, records, and Objective-C objects.
224 case Type::ConstantArray:
225 case Type::IncompleteArray:
226 case Type::VariableArray:
228 case Type::ObjCObject:
229 case Type::ObjCInterface:
230 return TEK_Aggregate;
232 // We operate on atomic values according to their underlying type.
234 type = cast<AtomicType>(type)->getValueType();
237 llvm_unreachable("unknown type kind!");
241 llvm::DebugLoc CodeGenFunction::EmitReturnBlock() {
242 // For cleanliness, we try to avoid emitting the return block for
244 llvm::BasicBlock *CurBB = Builder.GetInsertBlock();
247 assert(!CurBB->getTerminator() && "Unexpected terminated block.");
249 // We have a valid insert point, reuse it if it is empty or there are no
250 // explicit jumps to the return block.
251 if (CurBB->empty() || ReturnBlock.getBlock()->use_empty()) {
252 ReturnBlock.getBlock()->replaceAllUsesWith(CurBB);
253 delete ReturnBlock.getBlock();
255 EmitBlock(ReturnBlock.getBlock());
256 return llvm::DebugLoc();
259 // Otherwise, if the return block is the target of a single direct
260 // branch then we can just put the code in that block instead. This
261 // cleans up functions which started with a unified return block.
262 if (ReturnBlock.getBlock()->hasOneUse()) {
263 llvm::BranchInst *BI =
264 dyn_cast<llvm::BranchInst>(*ReturnBlock.getBlock()->user_begin());
265 if (BI && BI->isUnconditional() &&
266 BI->getSuccessor(0) == ReturnBlock.getBlock()) {
267 // Record/return the DebugLoc of the simple 'return' expression to be used
268 // later by the actual 'ret' instruction.
269 llvm::DebugLoc Loc = BI->getDebugLoc();
270 Builder.SetInsertPoint(BI->getParent());
271 BI->eraseFromParent();
272 delete ReturnBlock.getBlock();
277 // FIXME: We are at an unreachable point, there is no reason to emit the block
278 // unless it has uses. However, we still need a place to put the debug
279 // region.end for now.
281 EmitBlock(ReturnBlock.getBlock());
282 return llvm::DebugLoc();
285 static void EmitIfUsed(CodeGenFunction &CGF, llvm::BasicBlock *BB) {
287 if (!BB->use_empty())
288 return CGF.CurFn->getBasicBlockList().push_back(BB);
292 void CodeGenFunction::FinishFunction(SourceLocation EndLoc) {
293 assert(BreakContinueStack.empty() &&
294 "mismatched push/pop in break/continue stack!");
296 bool OnlySimpleReturnStmts = NumSimpleReturnExprs > 0
297 && NumSimpleReturnExprs == NumReturnExprs
298 && ReturnBlock.getBlock()->use_empty();
299 // Usually the return expression is evaluated before the cleanup
300 // code. If the function contains only a simple return statement,
301 // such as a constant, the location before the cleanup code becomes
302 // the last useful breakpoint in the function, because the simple
303 // return expression will be evaluated after the cleanup code. To be
304 // safe, set the debug location for cleanup code to the location of
305 // the return statement. Otherwise the cleanup code should be at the
306 // end of the function's lexical scope.
308 // If there are multiple branches to the return block, the branch
309 // instructions will get the location of the return statements and
311 if (CGDebugInfo *DI = getDebugInfo()) {
312 if (OnlySimpleReturnStmts)
313 DI->EmitLocation(Builder, LastStopPoint);
315 DI->EmitLocation(Builder, EndLoc);
318 // Pop any cleanups that might have been associated with the
319 // parameters. Do this in whatever block we're currently in; it's
320 // important to do this before we enter the return block or return
321 // edges will be *really* confused.
322 bool HasCleanups = EHStack.stable_begin() != PrologueCleanupDepth;
323 bool HasOnlyLifetimeMarkers =
324 HasCleanups && EHStack.containsOnlyLifetimeMarkers(PrologueCleanupDepth);
325 bool EmitRetDbgLoc = !HasCleanups || HasOnlyLifetimeMarkers;
327 // Make sure the line table doesn't jump back into the body for
328 // the ret after it's been at EndLoc.
329 if (CGDebugInfo *DI = getDebugInfo())
330 if (OnlySimpleReturnStmts)
331 DI->EmitLocation(Builder, EndLoc);
333 PopCleanupBlocks(PrologueCleanupDepth);
336 // Emit function epilog (to return).
337 llvm::DebugLoc Loc = EmitReturnBlock();
339 if (ShouldInstrumentFunction())
340 EmitFunctionInstrumentation("__cyg_profile_func_exit");
342 // Emit debug descriptor for function end.
343 if (CGDebugInfo *DI = getDebugInfo())
344 DI->EmitFunctionEnd(Builder);
346 // Reset the debug location to that of the simple 'return' expression, if any
347 // rather than that of the end of the function's scope '}'.
348 ApplyDebugLocation AL(*this, Loc);
349 EmitFunctionEpilog(*CurFnInfo, EmitRetDbgLoc, EndLoc);
350 EmitEndEHSpec(CurCodeDecl);
352 assert(EHStack.empty() &&
353 "did not remove all scopes from cleanup stack!");
355 // If someone did an indirect goto, emit the indirect goto block at the end of
357 if (IndirectBranch) {
358 EmitBlock(IndirectBranch->getParent());
359 Builder.ClearInsertionPoint();
362 // If some of our locals escaped, insert a call to llvm.localescape in the
364 if (!EscapedLocals.empty()) {
365 // Invert the map from local to index into a simple vector. There should be
367 SmallVector<llvm::Value *, 4> EscapeArgs;
368 EscapeArgs.resize(EscapedLocals.size());
369 for (auto &Pair : EscapedLocals)
370 EscapeArgs[Pair.second] = Pair.first;
371 llvm::Function *FrameEscapeFn = llvm::Intrinsic::getDeclaration(
372 &CGM.getModule(), llvm::Intrinsic::localescape);
373 CGBuilderTy(*this, AllocaInsertPt).CreateCall(FrameEscapeFn, EscapeArgs);
376 // Remove the AllocaInsertPt instruction, which is just a convenience for us.
377 llvm::Instruction *Ptr = AllocaInsertPt;
378 AllocaInsertPt = nullptr;
379 Ptr->eraseFromParent();
381 // If someone took the address of a label but never did an indirect goto, we
382 // made a zero entry PHI node, which is illegal, zap it now.
383 if (IndirectBranch) {
384 llvm::PHINode *PN = cast<llvm::PHINode>(IndirectBranch->getAddress());
385 if (PN->getNumIncomingValues() == 0) {
386 PN->replaceAllUsesWith(llvm::UndefValue::get(PN->getType()));
387 PN->eraseFromParent();
391 EmitIfUsed(*this, EHResumeBlock);
392 EmitIfUsed(*this, TerminateLandingPad);
393 EmitIfUsed(*this, TerminateHandler);
394 EmitIfUsed(*this, UnreachableBlock);
396 if (CGM.getCodeGenOpts().EmitDeclMetadata)
399 for (SmallVectorImpl<std::pair<llvm::Instruction *, llvm::Value *> >::iterator
400 I = DeferredReplacements.begin(),
401 E = DeferredReplacements.end();
403 I->first->replaceAllUsesWith(I->second);
404 I->first->eraseFromParent();
408 /// ShouldInstrumentFunction - Return true if the current function should be
409 /// instrumented with __cyg_profile_func_* calls
410 bool CodeGenFunction::ShouldInstrumentFunction() {
411 if (!CGM.getCodeGenOpts().InstrumentFunctions)
413 if (!CurFuncDecl || CurFuncDecl->hasAttr<NoInstrumentFunctionAttr>())
418 /// ShouldXRayInstrument - Return true if the current function should be
419 /// instrumented with XRay nop sleds.
420 bool CodeGenFunction::ShouldXRayInstrumentFunction() const {
421 return CGM.getCodeGenOpts().XRayInstrumentFunctions;
424 /// EmitFunctionInstrumentation - Emit LLVM code to call the specified
425 /// instrumentation function with the current function and the call site, if
426 /// function instrumentation is enabled.
427 void CodeGenFunction::EmitFunctionInstrumentation(const char *Fn) {
428 auto NL = ApplyDebugLocation::CreateArtificial(*this);
429 // void __cyg_profile_func_{enter,exit} (void *this_fn, void *call_site);
430 llvm::PointerType *PointerTy = Int8PtrTy;
431 llvm::Type *ProfileFuncArgs[] = { PointerTy, PointerTy };
432 llvm::FunctionType *FunctionTy =
433 llvm::FunctionType::get(VoidTy, ProfileFuncArgs, false);
435 llvm::Constant *F = CGM.CreateRuntimeFunction(FunctionTy, Fn);
436 llvm::CallInst *CallSite = Builder.CreateCall(
437 CGM.getIntrinsic(llvm::Intrinsic::returnaddress),
438 llvm::ConstantInt::get(Int32Ty, 0),
441 llvm::Value *args[] = {
442 llvm::ConstantExpr::getBitCast(CurFn, PointerTy),
446 EmitNounwindRuntimeCall(F, args);
449 static void removeImageAccessQualifier(std::string& TyName) {
450 std::string ReadOnlyQual("__read_only");
451 std::string::size_type ReadOnlyPos = TyName.find(ReadOnlyQual);
452 if (ReadOnlyPos != std::string::npos)
453 // "+ 1" for the space after access qualifier.
454 TyName.erase(ReadOnlyPos, ReadOnlyQual.size() + 1);
456 std::string WriteOnlyQual("__write_only");
457 std::string::size_type WriteOnlyPos = TyName.find(WriteOnlyQual);
458 if (WriteOnlyPos != std::string::npos)
459 TyName.erase(WriteOnlyPos, WriteOnlyQual.size() + 1);
461 std::string ReadWriteQual("__read_write");
462 std::string::size_type ReadWritePos = TyName.find(ReadWriteQual);
463 if (ReadWritePos != std::string::npos)
464 TyName.erase(ReadWritePos, ReadWriteQual.size() + 1);
469 // Returns the address space id that should be produced to the
470 // kernel_arg_addr_space metadata. This is always fixed to the ids
471 // as specified in the SPIR 2.0 specification in order to differentiate
472 // for example in clGetKernelArgInfo() implementation between the address
473 // spaces with targets without unique mapping to the OpenCL address spaces
474 // (basically all single AS CPUs).
475 static unsigned ArgInfoAddressSpace(unsigned LangAS) {
477 case LangAS::opencl_global: return 1;
478 case LangAS::opencl_constant: return 2;
479 case LangAS::opencl_local: return 3;
480 case LangAS::opencl_generic: return 4; // Not in SPIR 2.0 specs.
482 return 0; // Assume private.
486 // OpenCL v1.2 s5.6.4.6 allows the compiler to store kernel argument
487 // information in the program executable. The argument information stored
488 // includes the argument name, its type, the address and access qualifiers used.
489 static void GenOpenCLArgMetadata(const FunctionDecl *FD, llvm::Function *Fn,
490 CodeGenModule &CGM, llvm::LLVMContext &Context,
491 CGBuilderTy &Builder, ASTContext &ASTCtx) {
492 // Create MDNodes that represent the kernel arg metadata.
493 // Each MDNode is a list in the form of "key", N number of values which is
494 // the same number of values as their are kernel arguments.
496 const PrintingPolicy &Policy = ASTCtx.getPrintingPolicy();
498 // MDNode for the kernel argument address space qualifiers.
499 SmallVector<llvm::Metadata *, 8> addressQuals;
501 // MDNode for the kernel argument access qualifiers (images only).
502 SmallVector<llvm::Metadata *, 8> accessQuals;
504 // MDNode for the kernel argument type names.
505 SmallVector<llvm::Metadata *, 8> argTypeNames;
507 // MDNode for the kernel argument base type names.
508 SmallVector<llvm::Metadata *, 8> argBaseTypeNames;
510 // MDNode for the kernel argument type qualifiers.
511 SmallVector<llvm::Metadata *, 8> argTypeQuals;
513 // MDNode for the kernel argument names.
514 SmallVector<llvm::Metadata *, 8> argNames;
516 for (unsigned i = 0, e = FD->getNumParams(); i != e; ++i) {
517 const ParmVarDecl *parm = FD->getParamDecl(i);
518 QualType ty = parm->getType();
519 std::string typeQuals;
521 if (ty->isPointerType()) {
522 QualType pointeeTy = ty->getPointeeType();
524 // Get address qualifier.
525 addressQuals.push_back(llvm::ConstantAsMetadata::get(Builder.getInt32(
526 ArgInfoAddressSpace(pointeeTy.getAddressSpace()))));
528 // Get argument type name.
529 std::string typeName =
530 pointeeTy.getUnqualifiedType().getAsString(Policy) + "*";
532 // Turn "unsigned type" to "utype"
533 std::string::size_type pos = typeName.find("unsigned");
534 if (pointeeTy.isCanonical() && pos != std::string::npos)
535 typeName.erase(pos+1, 8);
537 argTypeNames.push_back(llvm::MDString::get(Context, typeName));
539 std::string baseTypeName =
540 pointeeTy.getUnqualifiedType().getCanonicalType().getAsString(
544 // Turn "unsigned type" to "utype"
545 pos = baseTypeName.find("unsigned");
546 if (pos != std::string::npos)
547 baseTypeName.erase(pos+1, 8);
549 argBaseTypeNames.push_back(llvm::MDString::get(Context, baseTypeName));
551 // Get argument type qualifiers:
552 if (ty.isRestrictQualified())
553 typeQuals = "restrict";
554 if (pointeeTy.isConstQualified() ||
555 (pointeeTy.getAddressSpace() == LangAS::opencl_constant))
556 typeQuals += typeQuals.empty() ? "const" : " const";
557 if (pointeeTy.isVolatileQualified())
558 typeQuals += typeQuals.empty() ? "volatile" : " volatile";
560 uint32_t AddrSpc = 0;
561 bool isPipe = ty->isPipeType();
562 if (ty->isImageType() || isPipe)
563 AddrSpc = ArgInfoAddressSpace(LangAS::opencl_global);
565 addressQuals.push_back(
566 llvm::ConstantAsMetadata::get(Builder.getInt32(AddrSpc)));
568 // Get argument type name.
569 std::string typeName;
571 typeName = ty.getCanonicalType()->getAs<PipeType>()->getElementType()
572 .getAsString(Policy);
574 typeName = ty.getUnqualifiedType().getAsString(Policy);
576 // Turn "unsigned type" to "utype"
577 std::string::size_type pos = typeName.find("unsigned");
578 if (ty.isCanonical() && pos != std::string::npos)
579 typeName.erase(pos+1, 8);
581 std::string baseTypeName;
583 baseTypeName = ty.getCanonicalType()->getAs<PipeType>()
584 ->getElementType().getCanonicalType()
585 .getAsString(Policy);
588 ty.getUnqualifiedType().getCanonicalType().getAsString(Policy);
590 // Remove access qualifiers on images
591 // (as they are inseparable from type in clang implementation,
592 // but OpenCL spec provides a special query to get access qualifier
593 // via clGetKernelArgInfo with CL_KERNEL_ARG_ACCESS_QUALIFIER):
594 if (ty->isImageType()) {
595 removeImageAccessQualifier(typeName);
596 removeImageAccessQualifier(baseTypeName);
599 argTypeNames.push_back(llvm::MDString::get(Context, typeName));
601 // Turn "unsigned type" to "utype"
602 pos = baseTypeName.find("unsigned");
603 if (pos != std::string::npos)
604 baseTypeName.erase(pos+1, 8);
606 argBaseTypeNames.push_back(llvm::MDString::get(Context, baseTypeName));
608 // Get argument type qualifiers:
609 if (ty.isConstQualified())
611 if (ty.isVolatileQualified())
612 typeQuals += typeQuals.empty() ? "volatile" : " volatile";
617 argTypeQuals.push_back(llvm::MDString::get(Context, typeQuals));
619 // Get image and pipe access qualifier:
620 if (ty->isImageType()|| ty->isPipeType()) {
621 const OpenCLAccessAttr *A = parm->getAttr<OpenCLAccessAttr>();
622 if (A && A->isWriteOnly())
623 accessQuals.push_back(llvm::MDString::get(Context, "write_only"));
624 else if (A && A->isReadWrite())
625 accessQuals.push_back(llvm::MDString::get(Context, "read_write"));
627 accessQuals.push_back(llvm::MDString::get(Context, "read_only"));
629 accessQuals.push_back(llvm::MDString::get(Context, "none"));
631 // Get argument name.
632 argNames.push_back(llvm::MDString::get(Context, parm->getName()));
635 Fn->setMetadata("kernel_arg_addr_space",
636 llvm::MDNode::get(Context, addressQuals));
637 Fn->setMetadata("kernel_arg_access_qual",
638 llvm::MDNode::get(Context, accessQuals));
639 Fn->setMetadata("kernel_arg_type",
640 llvm::MDNode::get(Context, argTypeNames));
641 Fn->setMetadata("kernel_arg_base_type",
642 llvm::MDNode::get(Context, argBaseTypeNames));
643 Fn->setMetadata("kernel_arg_type_qual",
644 llvm::MDNode::get(Context, argTypeQuals));
645 if (CGM.getCodeGenOpts().EmitOpenCLArgMetadata)
646 Fn->setMetadata("kernel_arg_name",
647 llvm::MDNode::get(Context, argNames));
650 void CodeGenFunction::EmitOpenCLKernelMetadata(const FunctionDecl *FD,
653 if (!FD->hasAttr<OpenCLKernelAttr>())
656 llvm::LLVMContext &Context = getLLVMContext();
658 GenOpenCLArgMetadata(FD, Fn, CGM, Context, Builder, getContext());
660 if (const VecTypeHintAttr *A = FD->getAttr<VecTypeHintAttr>()) {
661 QualType hintQTy = A->getTypeHint();
662 const ExtVectorType *hintEltQTy = hintQTy->getAs<ExtVectorType>();
663 bool isSignedInteger =
664 hintQTy->isSignedIntegerType() ||
665 (hintEltQTy && hintEltQTy->getElementType()->isSignedIntegerType());
666 llvm::Metadata *attrMDArgs[] = {
667 llvm::ConstantAsMetadata::get(llvm::UndefValue::get(
668 CGM.getTypes().ConvertType(A->getTypeHint()))),
669 llvm::ConstantAsMetadata::get(llvm::ConstantInt::get(
670 llvm::IntegerType::get(Context, 32),
671 llvm::APInt(32, (uint64_t)(isSignedInteger ? 1 : 0))))};
672 Fn->setMetadata("vec_type_hint", llvm::MDNode::get(Context, attrMDArgs));
675 if (const WorkGroupSizeHintAttr *A = FD->getAttr<WorkGroupSizeHintAttr>()) {
676 llvm::Metadata *attrMDArgs[] = {
677 llvm::ConstantAsMetadata::get(Builder.getInt32(A->getXDim())),
678 llvm::ConstantAsMetadata::get(Builder.getInt32(A->getYDim())),
679 llvm::ConstantAsMetadata::get(Builder.getInt32(A->getZDim()))};
680 Fn->setMetadata("work_group_size_hint", llvm::MDNode::get(Context, attrMDArgs));
683 if (const ReqdWorkGroupSizeAttr *A = FD->getAttr<ReqdWorkGroupSizeAttr>()) {
684 llvm::Metadata *attrMDArgs[] = {
685 llvm::ConstantAsMetadata::get(Builder.getInt32(A->getXDim())),
686 llvm::ConstantAsMetadata::get(Builder.getInt32(A->getYDim())),
687 llvm::ConstantAsMetadata::get(Builder.getInt32(A->getZDim()))};
688 Fn->setMetadata("reqd_work_group_size", llvm::MDNode::get(Context, attrMDArgs));
692 /// Determine whether the function F ends with a return stmt.
693 static bool endsWithReturn(const Decl* F) {
694 const Stmt *Body = nullptr;
695 if (auto *FD = dyn_cast_or_null<FunctionDecl>(F))
696 Body = FD->getBody();
697 else if (auto *OMD = dyn_cast_or_null<ObjCMethodDecl>(F))
698 Body = OMD->getBody();
700 if (auto *CS = dyn_cast_or_null<CompoundStmt>(Body)) {
701 auto LastStmt = CS->body_rbegin();
702 if (LastStmt != CS->body_rend())
703 return isa<ReturnStmt>(*LastStmt);
708 void CodeGenFunction::StartFunction(GlobalDecl GD,
711 const CGFunctionInfo &FnInfo,
712 const FunctionArgList &Args,
714 SourceLocation StartLoc) {
716 "Do not use a CodeGenFunction object for more than one function");
718 const Decl *D = GD.getDecl();
720 DidCallStackSave = false;
722 if (const auto *FD = dyn_cast_or_null<FunctionDecl>(D))
723 if (FD->usesSEHTry())
725 CurFuncDecl = (D ? D->getNonClosureContext() : nullptr);
729 assert(CurFn->isDeclaration() && "Function already has body?");
731 if (CGM.isInSanitizerBlacklist(Fn, Loc))
735 // Apply the no_sanitize* attributes to SanOpts.
736 for (auto Attr : D->specific_attrs<NoSanitizeAttr>())
737 SanOpts.Mask &= ~Attr->getMask();
740 // Apply sanitizer attributes to the function.
741 if (SanOpts.hasOneOf(SanitizerKind::Address | SanitizerKind::KernelAddress))
742 Fn->addFnAttr(llvm::Attribute::SanitizeAddress);
743 if (SanOpts.has(SanitizerKind::Thread))
744 Fn->addFnAttr(llvm::Attribute::SanitizeThread);
745 if (SanOpts.has(SanitizerKind::Memory))
746 Fn->addFnAttr(llvm::Attribute::SanitizeMemory);
747 if (SanOpts.has(SanitizerKind::SafeStack))
748 Fn->addFnAttr(llvm::Attribute::SafeStack);
750 // Ignore TSan memory acesses from within ObjC/ObjC++ dealloc, initialize,
751 // .cxx_destruct and all of their calees at run time.
752 if (SanOpts.has(SanitizerKind::Thread)) {
753 if (const auto *OMD = dyn_cast_or_null<ObjCMethodDecl>(D)) {
754 IdentifierInfo *II = OMD->getSelector().getIdentifierInfoForSlot(0);
755 if (OMD->getMethodFamily() == OMF_dealloc ||
756 OMD->getMethodFamily() == OMF_initialize ||
757 (OMD->getSelector().isUnarySelector() && II->isStr(".cxx_destruct"))) {
758 Fn->addFnAttr("sanitize_thread_no_checking_at_run_time");
759 Fn->removeFnAttr(llvm::Attribute::SanitizeThread);
764 // Apply xray attributes to the function (as a string, for now)
765 if (D && ShouldXRayInstrumentFunction()) {
766 if (const auto *XRayAttr = D->getAttr<XRayInstrumentAttr>()) {
767 if (XRayAttr->alwaysXRayInstrument())
768 Fn->addFnAttr("function-instrument", "xray-always");
769 if (XRayAttr->neverXRayInstrument())
770 Fn->addFnAttr("function-instrument", "xray-never");
773 "xray-instruction-threshold",
774 llvm::itostr(CGM.getCodeGenOpts().XRayInstructionThreshold));
778 if (const FunctionDecl *FD = dyn_cast_or_null<FunctionDecl>(D))
779 if (CGM.getLangOpts().OpenMP && FD->hasAttr<OMPDeclareSimdDeclAttr>())
780 CGM.getOpenMPRuntime().emitDeclareSimdFunction(FD, Fn);
782 // Add no-jump-tables value.
783 Fn->addFnAttr("no-jump-tables",
784 llvm::toStringRef(CGM.getCodeGenOpts().NoUseJumpTables));
786 if (getLangOpts().OpenCL) {
787 // Add metadata for a kernel function.
788 if (const FunctionDecl *FD = dyn_cast_or_null<FunctionDecl>(D))
789 EmitOpenCLKernelMetadata(FD, Fn);
792 // If we are checking function types, emit a function type signature as
794 if (getLangOpts().CPlusPlus && SanOpts.has(SanitizerKind::Function)) {
795 if (const FunctionDecl *FD = dyn_cast_or_null<FunctionDecl>(D)) {
796 if (llvm::Constant *PrologueSig =
797 CGM.getTargetCodeGenInfo().getUBSanFunctionSignature(CGM)) {
798 llvm::Constant *FTRTTIConst =
799 CGM.GetAddrOfRTTIDescriptor(FD->getType(), /*ForEH=*/true);
800 llvm::Constant *PrologueStructElems[] = { PrologueSig, FTRTTIConst };
801 llvm::Constant *PrologueStructConst =
802 llvm::ConstantStruct::getAnon(PrologueStructElems, /*Packed=*/true);
803 Fn->setPrologueData(PrologueStructConst);
808 // If we're in C++ mode and the function name is "main", it is guaranteed
809 // to be norecurse by the standard (3.6.1.3 "The function main shall not be
810 // used within a program").
811 if (getLangOpts().CPlusPlus)
812 if (const FunctionDecl *FD = dyn_cast_or_null<FunctionDecl>(D))
814 Fn->addFnAttr(llvm::Attribute::NoRecurse);
816 llvm::BasicBlock *EntryBB = createBasicBlock("entry", CurFn);
818 // Create a marker to make it easy to insert allocas into the entryblock
819 // later. Don't create this with the builder, because we don't want it
821 llvm::Value *Undef = llvm::UndefValue::get(Int32Ty);
822 AllocaInsertPt = new llvm::BitCastInst(Undef, Int32Ty, "allocapt", EntryBB);
824 ReturnBlock = getJumpDestInCurrentScope("return");
826 Builder.SetInsertPoint(EntryBB);
828 // Emit subprogram debug descriptor.
829 if (CGDebugInfo *DI = getDebugInfo()) {
830 // Reconstruct the type from the argument list so that implicit parameters,
831 // such as 'this' and 'vtt', show up in the debug info. Preserve the calling
833 CallingConv CC = CallingConv::CC_C;
834 if (auto *FD = dyn_cast_or_null<FunctionDecl>(D))
835 if (const auto *SrcFnTy = FD->getType()->getAs<FunctionType>())
836 CC = SrcFnTy->getCallConv();
837 SmallVector<QualType, 16> ArgTypes;
838 for (const VarDecl *VD : Args)
839 ArgTypes.push_back(VD->getType());
840 QualType FnType = getContext().getFunctionType(
841 RetTy, ArgTypes, FunctionProtoType::ExtProtoInfo(CC));
842 DI->EmitFunctionStart(GD, Loc, StartLoc, FnType, CurFn, Builder);
845 if (ShouldInstrumentFunction())
846 EmitFunctionInstrumentation("__cyg_profile_func_enter");
848 // Since emitting the mcount call here impacts optimizations such as function
849 // inlining, we just add an attribute to insert a mcount call in backend.
850 // The attribute "counting-function" is set to mcount function name which is
851 // architecture dependent.
852 if (CGM.getCodeGenOpts().InstrumentForProfiling)
853 Fn->addFnAttr("counting-function", getTarget().getMCountName());
855 if (RetTy->isVoidType()) {
856 // Void type; nothing to return.
857 ReturnValue = Address::invalid();
859 // Count the implicit return.
860 if (!endsWithReturn(D))
862 } else if (CurFnInfo->getReturnInfo().getKind() == ABIArgInfo::Indirect &&
863 !hasScalarEvaluationKind(CurFnInfo->getReturnType())) {
864 // Indirect aggregate return; emit returned value directly into sret slot.
865 // This reduces code size, and affects correctness in C++.
866 auto AI = CurFn->arg_begin();
867 if (CurFnInfo->getReturnInfo().isSRetAfterThis())
869 ReturnValue = Address(&*AI, CurFnInfo->getReturnInfo().getIndirectAlign());
870 } else if (CurFnInfo->getReturnInfo().getKind() == ABIArgInfo::InAlloca &&
871 !hasScalarEvaluationKind(CurFnInfo->getReturnType())) {
872 // Load the sret pointer from the argument struct and return into that.
873 unsigned Idx = CurFnInfo->getReturnInfo().getInAllocaFieldIndex();
874 llvm::Function::arg_iterator EI = CurFn->arg_end();
876 llvm::Value *Addr = Builder.CreateStructGEP(nullptr, &*EI, Idx);
877 Addr = Builder.CreateAlignedLoad(Addr, getPointerAlign(), "agg.result");
878 ReturnValue = Address(Addr, getNaturalTypeAlignment(RetTy));
880 ReturnValue = CreateIRTemp(RetTy, "retval");
882 // Tell the epilog emitter to autorelease the result. We do this
883 // now so that various specialized functions can suppress it
884 // during their IR-generation.
885 if (getLangOpts().ObjCAutoRefCount &&
886 !CurFnInfo->isReturnsRetained() &&
887 RetTy->isObjCRetainableType())
888 AutoreleaseResult = true;
891 EmitStartEHSpec(CurCodeDecl);
893 PrologueCleanupDepth = EHStack.stable_begin();
894 EmitFunctionProlog(*CurFnInfo, CurFn, Args);
896 if (D && isa<CXXMethodDecl>(D) && cast<CXXMethodDecl>(D)->isInstance()) {
897 CGM.getCXXABI().EmitInstanceFunctionProlog(*this);
898 const CXXMethodDecl *MD = cast<CXXMethodDecl>(D);
899 if (MD->getParent()->isLambda() &&
900 MD->getOverloadedOperator() == OO_Call) {
901 // We're in a lambda; figure out the captures.
902 MD->getParent()->getCaptureFields(LambdaCaptureFields,
903 LambdaThisCaptureField);
904 if (LambdaThisCaptureField) {
905 // If the lambda captures the object referred to by '*this' - either by
906 // value or by reference, make sure CXXThisValue points to the correct
909 // Get the lvalue for the field (which is a copy of the enclosing object
910 // or contains the address of the enclosing object).
911 LValue ThisFieldLValue = EmitLValueForLambdaField(LambdaThisCaptureField);
912 if (!LambdaThisCaptureField->getType()->isPointerType()) {
913 // If the enclosing object was captured by value, just use its address.
914 CXXThisValue = ThisFieldLValue.getAddress().getPointer();
916 // Load the lvalue pointed to by the field, since '*this' was captured
919 EmitLoadOfLValue(ThisFieldLValue, SourceLocation()).getScalarVal();
922 for (auto *FD : MD->getParent()->fields()) {
923 if (FD->hasCapturedVLAType()) {
924 auto *ExprArg = EmitLoadOfLValue(EmitLValueForLambdaField(FD),
925 SourceLocation()).getScalarVal();
926 auto VAT = FD->getCapturedVLAType();
927 VLASizeMap[VAT->getSizeExpr()] = ExprArg;
931 // Not in a lambda; just use 'this' from the method.
932 // FIXME: Should we generate a new load for each use of 'this'? The
933 // fast register allocator would be happier...
934 CXXThisValue = CXXABIThisValue;
938 // If any of the arguments have a variably modified type, make sure to
939 // emit the type size.
940 for (FunctionArgList::const_iterator i = Args.begin(), e = Args.end();
942 const VarDecl *VD = *i;
944 // Dig out the type as written from ParmVarDecls; it's unclear whether
945 // the standard (C99 6.9.1p10) requires this, but we're following the
946 // precedent set by gcc.
948 if (const ParmVarDecl *PVD = dyn_cast<ParmVarDecl>(VD))
949 Ty = PVD->getOriginalType();
953 if (Ty->isVariablyModifiedType())
954 EmitVariablyModifiedType(Ty);
956 // Emit a location at the end of the prologue.
957 if (CGDebugInfo *DI = getDebugInfo())
958 DI->EmitLocation(Builder, StartLoc);
961 void CodeGenFunction::EmitFunctionBody(FunctionArgList &Args,
963 incrementProfileCounter(Body);
964 if (const CompoundStmt *S = dyn_cast<CompoundStmt>(Body))
965 EmitCompoundStmtWithoutScope(*S);
970 /// When instrumenting to collect profile data, the counts for some blocks
971 /// such as switch cases need to not include the fall-through counts, so
972 /// emit a branch around the instrumentation code. When not instrumenting,
973 /// this just calls EmitBlock().
974 void CodeGenFunction::EmitBlockWithFallThrough(llvm::BasicBlock *BB,
976 llvm::BasicBlock *SkipCountBB = nullptr;
977 if (HaveInsertPoint() && CGM.getCodeGenOpts().hasProfileClangInstr()) {
978 // When instrumenting for profiling, the fallthrough to certain
979 // statements needs to skip over the instrumentation code so that we
980 // get an accurate count.
981 SkipCountBB = createBasicBlock("skipcount");
982 EmitBranch(SkipCountBB);
985 uint64_t CurrentCount = getCurrentProfileCount();
986 incrementProfileCounter(S);
987 setCurrentProfileCount(getCurrentProfileCount() + CurrentCount);
989 EmitBlock(SkipCountBB);
992 /// Tries to mark the given function nounwind based on the
993 /// non-existence of any throwing calls within it. We believe this is
994 /// lightweight enough to do at -O0.
995 static void TryMarkNoThrow(llvm::Function *F) {
996 // LLVM treats 'nounwind' on a function as part of the type, so we
997 // can't do this on functions that can be overwritten.
998 if (F->isInterposable()) return;
1000 for (llvm::BasicBlock &BB : *F)
1001 for (llvm::Instruction &I : BB)
1005 F->setDoesNotThrow();
1008 QualType CodeGenFunction::BuildFunctionArgList(GlobalDecl GD,
1009 FunctionArgList &Args) {
1010 const FunctionDecl *FD = cast<FunctionDecl>(GD.getDecl());
1011 QualType ResTy = FD->getReturnType();
1013 const CXXMethodDecl *MD = dyn_cast<CXXMethodDecl>(FD);
1014 if (MD && MD->isInstance()) {
1015 if (CGM.getCXXABI().HasThisReturn(GD))
1016 ResTy = MD->getThisType(getContext());
1017 else if (CGM.getCXXABI().hasMostDerivedReturn(GD))
1018 ResTy = CGM.getContext().VoidPtrTy;
1019 CGM.getCXXABI().buildThisParam(*this, Args);
1022 // The base version of an inheriting constructor whose constructed base is a
1023 // virtual base is not passed any arguments (because it doesn't actually call
1024 // the inherited constructor).
1025 bool PassedParams = true;
1026 if (const CXXConstructorDecl *CD = dyn_cast<CXXConstructorDecl>(FD))
1027 if (auto Inherited = CD->getInheritedConstructor())
1029 getTypes().inheritingCtorHasParams(Inherited, GD.getCtorType());
1032 for (auto *Param : FD->parameters()) {
1033 Args.push_back(Param);
1034 if (!Param->hasAttr<PassObjectSizeAttr>())
1037 IdentifierInfo *NoID = nullptr;
1038 auto *Implicit = ImplicitParamDecl::Create(
1039 getContext(), Param->getDeclContext(), Param->getLocation(), NoID,
1040 getContext().getSizeType());
1041 SizeArguments[Param] = Implicit;
1042 Args.push_back(Implicit);
1046 if (MD && (isa<CXXConstructorDecl>(MD) || isa<CXXDestructorDecl>(MD)))
1047 CGM.getCXXABI().addImplicitStructorParams(*this, ResTy, Args);
1053 shouldUseUndefinedBehaviorReturnOptimization(const FunctionDecl *FD,
1054 const ASTContext &Context) {
1055 QualType T = FD->getReturnType();
1056 // Avoid the optimization for functions that return a record type with a
1057 // trivial destructor or another trivially copyable type.
1058 if (const RecordType *RT = T.getCanonicalType()->getAs<RecordType>()) {
1059 if (const auto *ClassDecl = dyn_cast<CXXRecordDecl>(RT->getDecl()))
1060 return !ClassDecl->hasTrivialDestructor();
1062 return !T.isTriviallyCopyableType(Context);
1065 void CodeGenFunction::GenerateCode(GlobalDecl GD, llvm::Function *Fn,
1066 const CGFunctionInfo &FnInfo) {
1067 const FunctionDecl *FD = cast<FunctionDecl>(GD.getDecl());
1070 FunctionArgList Args;
1071 QualType ResTy = BuildFunctionArgList(GD, Args);
1073 // Check if we should generate debug info for this function.
1074 if (FD->hasAttr<NoDebugAttr>())
1075 DebugInfo = nullptr; // disable debug info indefinitely for this function
1077 SourceRange BodyRange;
1078 if (Stmt *Body = FD->getBody()) BodyRange = Body->getSourceRange();
1079 CurEHLocation = BodyRange.getEnd();
1081 // Use the location of the start of the function to determine where
1082 // the function definition is located. By default use the location
1083 // of the declaration as the location for the subprogram. A function
1084 // may lack a declaration in the source code if it is created by code
1085 // gen. (examples: _GLOBAL__I_a, __cxx_global_array_dtor, thunk).
1086 SourceLocation Loc = FD->getLocation();
1088 // If this is a function specialization then use the pattern body
1089 // as the location for the function.
1090 if (const FunctionDecl *SpecDecl = FD->getTemplateInstantiationPattern())
1091 if (SpecDecl->hasBody(SpecDecl))
1092 Loc = SpecDecl->getLocation();
1094 Stmt *Body = FD->getBody();
1096 // Initialize helper which will detect jumps which can cause invalid lifetime
1098 if (Body && ShouldEmitLifetimeMarkers)
1099 Bypasses.Init(Body);
1101 // Emit the standard function prologue.
1102 StartFunction(GD, ResTy, Fn, FnInfo, Args, Loc, BodyRange.getBegin());
1104 // Generate the body of the function.
1105 PGO.assignRegionCounters(GD, CurFn);
1106 if (isa<CXXDestructorDecl>(FD))
1107 EmitDestructorBody(Args);
1108 else if (isa<CXXConstructorDecl>(FD))
1109 EmitConstructorBody(Args);
1110 else if (getLangOpts().CUDA &&
1111 !getLangOpts().CUDAIsDevice &&
1112 FD->hasAttr<CUDAGlobalAttr>())
1113 CGM.getCUDARuntime().emitDeviceStub(*this, Args);
1114 else if (isa<CXXConversionDecl>(FD) &&
1115 cast<CXXConversionDecl>(FD)->isLambdaToBlockPointerConversion()) {
1116 // The lambda conversion to block pointer is special; the semantics can't be
1117 // expressed in the AST, so IRGen needs to special-case it.
1118 EmitLambdaToBlockPointerBody(Args);
1119 } else if (isa<CXXMethodDecl>(FD) &&
1120 cast<CXXMethodDecl>(FD)->isLambdaStaticInvoker()) {
1121 // The lambda static invoker function is special, because it forwards or
1122 // clones the body of the function call operator (but is actually static).
1123 EmitLambdaStaticInvokeFunction(cast<CXXMethodDecl>(FD));
1124 } else if (FD->isDefaulted() && isa<CXXMethodDecl>(FD) &&
1125 (cast<CXXMethodDecl>(FD)->isCopyAssignmentOperator() ||
1126 cast<CXXMethodDecl>(FD)->isMoveAssignmentOperator())) {
1127 // Implicit copy-assignment gets the same special treatment as implicit
1128 // copy-constructors.
1129 emitImplicitAssignmentOperatorBody(Args);
1131 EmitFunctionBody(Args, Body);
1133 llvm_unreachable("no definition for emitted function");
1135 // C++11 [stmt.return]p2:
1136 // Flowing off the end of a function [...] results in undefined behavior in
1137 // a value-returning function.
1139 // If the '}' that terminates a function is reached, and the value of the
1140 // function call is used by the caller, the behavior is undefined.
1141 if (getLangOpts().CPlusPlus && !FD->hasImplicitReturnZero() && !SawAsmBlock &&
1142 !FD->getReturnType()->isVoidType() && Builder.GetInsertBlock()) {
1143 bool ShouldEmitUnreachable =
1144 CGM.getCodeGenOpts().StrictReturn ||
1145 shouldUseUndefinedBehaviorReturnOptimization(FD, getContext());
1146 if (SanOpts.has(SanitizerKind::Return)) {
1147 SanitizerScope SanScope(this);
1148 llvm::Value *IsFalse = Builder.getFalse();
1149 EmitCheck(std::make_pair(IsFalse, SanitizerKind::Return),
1150 SanitizerHandler::MissingReturn,
1151 EmitCheckSourceLocation(FD->getLocation()), None);
1152 } else if (ShouldEmitUnreachable) {
1153 if (CGM.getCodeGenOpts().OptimizationLevel == 0)
1154 EmitTrapCall(llvm::Intrinsic::trap);
1156 if (SanOpts.has(SanitizerKind::Return) || ShouldEmitUnreachable) {
1157 Builder.CreateUnreachable();
1158 Builder.ClearInsertionPoint();
1162 // Emit the standard function epilogue.
1163 FinishFunction(BodyRange.getEnd());
1165 // If we haven't marked the function nothrow through other means, do
1166 // a quick pass now to see if we can.
1167 if (!CurFn->doesNotThrow())
1168 TryMarkNoThrow(CurFn);
1171 /// ContainsLabel - Return true if the statement contains a label in it. If
1172 /// this statement is not executed normally, it not containing a label means
1173 /// that we can just remove the code.
1174 bool CodeGenFunction::ContainsLabel(const Stmt *S, bool IgnoreCaseStmts) {
1175 // Null statement, not a label!
1176 if (!S) return false;
1178 // If this is a label, we have to emit the code, consider something like:
1179 // if (0) { ... foo: bar(); } goto foo;
1181 // TODO: If anyone cared, we could track __label__'s, since we know that you
1182 // can't jump to one from outside their declared region.
1183 if (isa<LabelStmt>(S))
1186 // If this is a case/default statement, and we haven't seen a switch, we have
1187 // to emit the code.
1188 if (isa<SwitchCase>(S) && !IgnoreCaseStmts)
1191 // If this is a switch statement, we want to ignore cases below it.
1192 if (isa<SwitchStmt>(S))
1193 IgnoreCaseStmts = true;
1195 // Scan subexpressions for verboten labels.
1196 for (const Stmt *SubStmt : S->children())
1197 if (ContainsLabel(SubStmt, IgnoreCaseStmts))
1203 /// containsBreak - Return true if the statement contains a break out of it.
1204 /// If the statement (recursively) contains a switch or loop with a break
1205 /// inside of it, this is fine.
1206 bool CodeGenFunction::containsBreak(const Stmt *S) {
1207 // Null statement, not a label!
1208 if (!S) return false;
1210 // If this is a switch or loop that defines its own break scope, then we can
1211 // include it and anything inside of it.
1212 if (isa<SwitchStmt>(S) || isa<WhileStmt>(S) || isa<DoStmt>(S) ||
1216 if (isa<BreakStmt>(S))
1219 // Scan subexpressions for verboten breaks.
1220 for (const Stmt *SubStmt : S->children())
1221 if (containsBreak(SubStmt))
1227 bool CodeGenFunction::mightAddDeclToScope(const Stmt *S) {
1228 if (!S) return false;
1230 // Some statement kinds add a scope and thus never add a decl to the current
1231 // scope. Note, this list is longer than the list of statements that might
1232 // have an unscoped decl nested within them, but this way is conservatively
1233 // correct even if more statement kinds are added.
1234 if (isa<IfStmt>(S) || isa<SwitchStmt>(S) || isa<WhileStmt>(S) ||
1235 isa<DoStmt>(S) || isa<ForStmt>(S) || isa<CompoundStmt>(S) ||
1236 isa<CXXForRangeStmt>(S) || isa<CXXTryStmt>(S) ||
1237 isa<ObjCForCollectionStmt>(S) || isa<ObjCAtTryStmt>(S))
1240 if (isa<DeclStmt>(S))
1243 for (const Stmt *SubStmt : S->children())
1244 if (mightAddDeclToScope(SubStmt))
1250 /// ConstantFoldsToSimpleInteger - If the specified expression does not fold
1251 /// to a constant, or if it does but contains a label, return false. If it
1252 /// constant folds return true and set the boolean result in Result.
1253 bool CodeGenFunction::ConstantFoldsToSimpleInteger(const Expr *Cond,
1256 llvm::APSInt ResultInt;
1257 if (!ConstantFoldsToSimpleInteger(Cond, ResultInt, AllowLabels))
1260 ResultBool = ResultInt.getBoolValue();
1264 /// ConstantFoldsToSimpleInteger - If the specified expression does not fold
1265 /// to a constant, or if it does but contains a label, return false. If it
1266 /// constant folds return true and set the folded value.
1267 bool CodeGenFunction::ConstantFoldsToSimpleInteger(const Expr *Cond,
1268 llvm::APSInt &ResultInt,
1270 // FIXME: Rename and handle conversion of other evaluatable things
1273 if (!Cond->EvaluateAsInt(Int, getContext()))
1274 return false; // Not foldable, not integer or not fully evaluatable.
1276 if (!AllowLabels && CodeGenFunction::ContainsLabel(Cond))
1277 return false; // Contains a label.
1285 /// EmitBranchOnBoolExpr - Emit a branch on a boolean condition (e.g. for an if
1286 /// statement) to the specified blocks. Based on the condition, this might try
1287 /// to simplify the codegen of the conditional based on the branch.
1289 void CodeGenFunction::EmitBranchOnBoolExpr(const Expr *Cond,
1290 llvm::BasicBlock *TrueBlock,
1291 llvm::BasicBlock *FalseBlock,
1292 uint64_t TrueCount) {
1293 Cond = Cond->IgnoreParens();
1295 if (const BinaryOperator *CondBOp = dyn_cast<BinaryOperator>(Cond)) {
1297 // Handle X && Y in a condition.
1298 if (CondBOp->getOpcode() == BO_LAnd) {
1299 // If we have "1 && X", simplify the code. "0 && X" would have constant
1300 // folded if the case was simple enough.
1301 bool ConstantBool = false;
1302 if (ConstantFoldsToSimpleInteger(CondBOp->getLHS(), ConstantBool) &&
1304 // br(1 && X) -> br(X).
1305 incrementProfileCounter(CondBOp);
1306 return EmitBranchOnBoolExpr(CondBOp->getRHS(), TrueBlock, FalseBlock,
1310 // If we have "X && 1", simplify the code to use an uncond branch.
1311 // "X && 0" would have been constant folded to 0.
1312 if (ConstantFoldsToSimpleInteger(CondBOp->getRHS(), ConstantBool) &&
1314 // br(X && 1) -> br(X).
1315 return EmitBranchOnBoolExpr(CondBOp->getLHS(), TrueBlock, FalseBlock,
1319 // Emit the LHS as a conditional. If the LHS conditional is false, we
1320 // want to jump to the FalseBlock.
1321 llvm::BasicBlock *LHSTrue = createBasicBlock("land.lhs.true");
1322 // The counter tells us how often we evaluate RHS, and all of TrueCount
1323 // can be propagated to that branch.
1324 uint64_t RHSCount = getProfileCount(CondBOp->getRHS());
1326 ConditionalEvaluation eval(*this);
1328 ApplyDebugLocation DL(*this, Cond);
1329 EmitBranchOnBoolExpr(CondBOp->getLHS(), LHSTrue, FalseBlock, RHSCount);
1333 incrementProfileCounter(CondBOp);
1334 setCurrentProfileCount(getProfileCount(CondBOp->getRHS()));
1336 // Any temporaries created here are conditional.
1338 EmitBranchOnBoolExpr(CondBOp->getRHS(), TrueBlock, FalseBlock, TrueCount);
1344 if (CondBOp->getOpcode() == BO_LOr) {
1345 // If we have "0 || X", simplify the code. "1 || X" would have constant
1346 // folded if the case was simple enough.
1347 bool ConstantBool = false;
1348 if (ConstantFoldsToSimpleInteger(CondBOp->getLHS(), ConstantBool) &&
1350 // br(0 || X) -> br(X).
1351 incrementProfileCounter(CondBOp);
1352 return EmitBranchOnBoolExpr(CondBOp->getRHS(), TrueBlock, FalseBlock,
1356 // If we have "X || 0", simplify the code to use an uncond branch.
1357 // "X || 1" would have been constant folded to 1.
1358 if (ConstantFoldsToSimpleInteger(CondBOp->getRHS(), ConstantBool) &&
1360 // br(X || 0) -> br(X).
1361 return EmitBranchOnBoolExpr(CondBOp->getLHS(), TrueBlock, FalseBlock,
1365 // Emit the LHS as a conditional. If the LHS conditional is true, we
1366 // want to jump to the TrueBlock.
1367 llvm::BasicBlock *LHSFalse = createBasicBlock("lor.lhs.false");
1368 // We have the count for entry to the RHS and for the whole expression
1369 // being true, so we can divy up True count between the short circuit and
1372 getCurrentProfileCount() - getProfileCount(CondBOp->getRHS());
1373 uint64_t RHSCount = TrueCount - LHSCount;
1375 ConditionalEvaluation eval(*this);
1377 ApplyDebugLocation DL(*this, Cond);
1378 EmitBranchOnBoolExpr(CondBOp->getLHS(), TrueBlock, LHSFalse, LHSCount);
1379 EmitBlock(LHSFalse);
1382 incrementProfileCounter(CondBOp);
1383 setCurrentProfileCount(getProfileCount(CondBOp->getRHS()));
1385 // Any temporaries created here are conditional.
1387 EmitBranchOnBoolExpr(CondBOp->getRHS(), TrueBlock, FalseBlock, RHSCount);
1395 if (const UnaryOperator *CondUOp = dyn_cast<UnaryOperator>(Cond)) {
1396 // br(!x, t, f) -> br(x, f, t)
1397 if (CondUOp->getOpcode() == UO_LNot) {
1398 // Negate the count.
1399 uint64_t FalseCount = getCurrentProfileCount() - TrueCount;
1400 // Negate the condition and swap the destination blocks.
1401 return EmitBranchOnBoolExpr(CondUOp->getSubExpr(), FalseBlock, TrueBlock,
1406 if (const ConditionalOperator *CondOp = dyn_cast<ConditionalOperator>(Cond)) {
1407 // br(c ? x : y, t, f) -> br(c, br(x, t, f), br(y, t, f))
1408 llvm::BasicBlock *LHSBlock = createBasicBlock("cond.true");
1409 llvm::BasicBlock *RHSBlock = createBasicBlock("cond.false");
1411 ConditionalEvaluation cond(*this);
1412 EmitBranchOnBoolExpr(CondOp->getCond(), LHSBlock, RHSBlock,
1413 getProfileCount(CondOp));
1415 // When computing PGO branch weights, we only know the overall count for
1416 // the true block. This code is essentially doing tail duplication of the
1417 // naive code-gen, introducing new edges for which counts are not
1418 // available. Divide the counts proportionally between the LHS and RHS of
1419 // the conditional operator.
1420 uint64_t LHSScaledTrueCount = 0;
1423 getProfileCount(CondOp) / (double)getCurrentProfileCount();
1424 LHSScaledTrueCount = TrueCount * LHSRatio;
1428 EmitBlock(LHSBlock);
1429 incrementProfileCounter(CondOp);
1431 ApplyDebugLocation DL(*this, Cond);
1432 EmitBranchOnBoolExpr(CondOp->getLHS(), TrueBlock, FalseBlock,
1433 LHSScaledTrueCount);
1438 EmitBlock(RHSBlock);
1439 EmitBranchOnBoolExpr(CondOp->getRHS(), TrueBlock, FalseBlock,
1440 TrueCount - LHSScaledTrueCount);
1446 if (const CXXThrowExpr *Throw = dyn_cast<CXXThrowExpr>(Cond)) {
1447 // Conditional operator handling can give us a throw expression as a
1448 // condition for a case like:
1449 // br(c ? throw x : y, t, f) -> br(c, br(throw x, t, f), br(y, t, f)
1451 // br(c, throw x, br(y, t, f))
1452 EmitCXXThrowExpr(Throw, /*KeepInsertionPoint*/false);
1456 // If the branch has a condition wrapped by __builtin_unpredictable,
1457 // create metadata that specifies that the branch is unpredictable.
1458 // Don't bother if not optimizing because that metadata would not be used.
1459 llvm::MDNode *Unpredictable = nullptr;
1460 auto *Call = dyn_cast<CallExpr>(Cond);
1461 if (Call && CGM.getCodeGenOpts().OptimizationLevel != 0) {
1462 auto *FD = dyn_cast_or_null<FunctionDecl>(Call->getCalleeDecl());
1463 if (FD && FD->getBuiltinID() == Builtin::BI__builtin_unpredictable) {
1464 llvm::MDBuilder MDHelper(getLLVMContext());
1465 Unpredictable = MDHelper.createUnpredictable();
1469 // Create branch weights based on the number of times we get here and the
1470 // number of times the condition should be true.
1471 uint64_t CurrentCount = std::max(getCurrentProfileCount(), TrueCount);
1472 llvm::MDNode *Weights =
1473 createProfileWeights(TrueCount, CurrentCount - TrueCount);
1475 // Emit the code with the fully general case.
1478 ApplyDebugLocation DL(*this, Cond);
1479 CondV = EvaluateExprAsBool(Cond);
1481 Builder.CreateCondBr(CondV, TrueBlock, FalseBlock, Weights, Unpredictable);
1484 /// ErrorUnsupported - Print out an error that codegen doesn't support the
1485 /// specified stmt yet.
1486 void CodeGenFunction::ErrorUnsupported(const Stmt *S, const char *Type) {
1487 CGM.ErrorUnsupported(S, Type);
1490 /// emitNonZeroVLAInit - Emit the "zero" initialization of a
1491 /// variable-length array whose elements have a non-zero bit-pattern.
1493 /// \param baseType the inner-most element type of the array
1494 /// \param src - a char* pointing to the bit-pattern for a single
1495 /// base element of the array
1496 /// \param sizeInChars - the total size of the VLA, in chars
1497 static void emitNonZeroVLAInit(CodeGenFunction &CGF, QualType baseType,
1498 Address dest, Address src,
1499 llvm::Value *sizeInChars) {
1500 CGBuilderTy &Builder = CGF.Builder;
1502 CharUnits baseSize = CGF.getContext().getTypeSizeInChars(baseType);
1503 llvm::Value *baseSizeInChars
1504 = llvm::ConstantInt::get(CGF.IntPtrTy, baseSize.getQuantity());
1507 Builder.CreateElementBitCast(dest, CGF.Int8Ty, "vla.begin");
1509 Builder.CreateInBoundsGEP(begin.getPointer(), sizeInChars, "vla.end");
1511 llvm::BasicBlock *originBB = CGF.Builder.GetInsertBlock();
1512 llvm::BasicBlock *loopBB = CGF.createBasicBlock("vla-init.loop");
1513 llvm::BasicBlock *contBB = CGF.createBasicBlock("vla-init.cont");
1515 // Make a loop over the VLA. C99 guarantees that the VLA element
1516 // count must be nonzero.
1517 CGF.EmitBlock(loopBB);
1519 llvm::PHINode *cur = Builder.CreatePHI(begin.getType(), 2, "vla.cur");
1520 cur->addIncoming(begin.getPointer(), originBB);
1522 CharUnits curAlign =
1523 dest.getAlignment().alignmentOfArrayElement(baseSize);
1525 // memcpy the individual element bit-pattern.
1526 Builder.CreateMemCpy(Address(cur, curAlign), src, baseSizeInChars,
1527 /*volatile*/ false);
1529 // Go to the next element.
1531 Builder.CreateInBoundsGEP(CGF.Int8Ty, cur, baseSizeInChars, "vla.next");
1533 // Leave if that's the end of the VLA.
1534 llvm::Value *done = Builder.CreateICmpEQ(next, end, "vla-init.isdone");
1535 Builder.CreateCondBr(done, contBB, loopBB);
1536 cur->addIncoming(next, loopBB);
1538 CGF.EmitBlock(contBB);
1542 CodeGenFunction::EmitNullInitialization(Address DestPtr, QualType Ty) {
1543 // Ignore empty classes in C++.
1544 if (getLangOpts().CPlusPlus) {
1545 if (const RecordType *RT = Ty->getAs<RecordType>()) {
1546 if (cast<CXXRecordDecl>(RT->getDecl())->isEmpty())
1551 // Cast the dest ptr to the appropriate i8 pointer type.
1552 if (DestPtr.getElementType() != Int8Ty)
1553 DestPtr = Builder.CreateElementBitCast(DestPtr, Int8Ty);
1555 // Get size and alignment info for this aggregate.
1556 CharUnits size = getContext().getTypeSizeInChars(Ty);
1558 llvm::Value *SizeVal;
1559 const VariableArrayType *vla;
1561 // Don't bother emitting a zero-byte memset.
1562 if (size.isZero()) {
1563 // But note that getTypeInfo returns 0 for a VLA.
1564 if (const VariableArrayType *vlaType =
1565 dyn_cast_or_null<VariableArrayType>(
1566 getContext().getAsArrayType(Ty))) {
1568 llvm::Value *numElts;
1569 std::tie(numElts, eltType) = getVLASize(vlaType);
1572 CharUnits eltSize = getContext().getTypeSizeInChars(eltType);
1573 if (!eltSize.isOne())
1574 SizeVal = Builder.CreateNUWMul(SizeVal, CGM.getSize(eltSize));
1580 SizeVal = CGM.getSize(size);
1584 // If the type contains a pointer to data member we can't memset it to zero.
1585 // Instead, create a null constant and copy it to the destination.
1586 // TODO: there are other patterns besides zero that we can usefully memset,
1587 // like -1, which happens to be the pattern used by member-pointers.
1588 if (!CGM.getTypes().isZeroInitializable(Ty)) {
1589 // For a VLA, emit a single element, then splat that over the VLA.
1590 if (vla) Ty = getContext().getBaseElementType(vla);
1592 llvm::Constant *NullConstant = CGM.EmitNullConstant(Ty);
1594 llvm::GlobalVariable *NullVariable =
1595 new llvm::GlobalVariable(CGM.getModule(), NullConstant->getType(),
1596 /*isConstant=*/true,
1597 llvm::GlobalVariable::PrivateLinkage,
1598 NullConstant, Twine());
1599 CharUnits NullAlign = DestPtr.getAlignment();
1600 NullVariable->setAlignment(NullAlign.getQuantity());
1601 Address SrcPtr(Builder.CreateBitCast(NullVariable, Builder.getInt8PtrTy()),
1604 if (vla) return emitNonZeroVLAInit(*this, Ty, DestPtr, SrcPtr, SizeVal);
1606 // Get and call the appropriate llvm.memcpy overload.
1607 Builder.CreateMemCpy(DestPtr, SrcPtr, SizeVal, false);
1611 // Otherwise, just memset the whole thing to zero. This is legal
1612 // because in LLVM, all default initializers (other than the ones we just
1613 // handled above) are guaranteed to have a bit pattern of all zeros.
1614 Builder.CreateMemSet(DestPtr, Builder.getInt8(0), SizeVal, false);
1617 llvm::BlockAddress *CodeGenFunction::GetAddrOfLabel(const LabelDecl *L) {
1618 // Make sure that there is a block for the indirect goto.
1619 if (!IndirectBranch)
1620 GetIndirectGotoBlock();
1622 llvm::BasicBlock *BB = getJumpDestForLabel(L).getBlock();
1624 // Make sure the indirect branch includes all of the address-taken blocks.
1625 IndirectBranch->addDestination(BB);
1626 return llvm::BlockAddress::get(CurFn, BB);
1629 llvm::BasicBlock *CodeGenFunction::GetIndirectGotoBlock() {
1630 // If we already made the indirect branch for indirect goto, return its block.
1631 if (IndirectBranch) return IndirectBranch->getParent();
1633 CGBuilderTy TmpBuilder(*this, createBasicBlock("indirectgoto"));
1635 // Create the PHI node that indirect gotos will add entries to.
1636 llvm::Value *DestVal = TmpBuilder.CreatePHI(Int8PtrTy, 0,
1637 "indirect.goto.dest");
1639 // Create the indirect branch instruction.
1640 IndirectBranch = TmpBuilder.CreateIndirectBr(DestVal);
1641 return IndirectBranch->getParent();
1644 /// Computes the length of an array in elements, as well as the base
1645 /// element type and a properly-typed first element pointer.
1646 llvm::Value *CodeGenFunction::emitArrayLength(const ArrayType *origArrayType,
1649 const ArrayType *arrayType = origArrayType;
1651 // If it's a VLA, we have to load the stored size. Note that
1652 // this is the size of the VLA in bytes, not its size in elements.
1653 llvm::Value *numVLAElements = nullptr;
1654 if (isa<VariableArrayType>(arrayType)) {
1655 numVLAElements = getVLASize(cast<VariableArrayType>(arrayType)).first;
1657 // Walk into all VLAs. This doesn't require changes to addr,
1658 // which has type T* where T is the first non-VLA element type.
1660 QualType elementType = arrayType->getElementType();
1661 arrayType = getContext().getAsArrayType(elementType);
1663 // If we only have VLA components, 'addr' requires no adjustment.
1665 baseType = elementType;
1666 return numVLAElements;
1668 } while (isa<VariableArrayType>(arrayType));
1670 // We get out here only if we find a constant array type
1674 // We have some number of constant-length arrays, so addr should
1675 // have LLVM type [M x [N x [...]]]*. Build a GEP that walks
1676 // down to the first element of addr.
1677 SmallVector<llvm::Value*, 8> gepIndices;
1679 // GEP down to the array type.
1680 llvm::ConstantInt *zero = Builder.getInt32(0);
1681 gepIndices.push_back(zero);
1683 uint64_t countFromCLAs = 1;
1686 llvm::ArrayType *llvmArrayType =
1687 dyn_cast<llvm::ArrayType>(addr.getElementType());
1688 while (llvmArrayType) {
1689 assert(isa<ConstantArrayType>(arrayType));
1690 assert(cast<ConstantArrayType>(arrayType)->getSize().getZExtValue()
1691 == llvmArrayType->getNumElements());
1693 gepIndices.push_back(zero);
1694 countFromCLAs *= llvmArrayType->getNumElements();
1695 eltType = arrayType->getElementType();
1698 dyn_cast<llvm::ArrayType>(llvmArrayType->getElementType());
1699 arrayType = getContext().getAsArrayType(arrayType->getElementType());
1700 assert((!llvmArrayType || arrayType) &&
1701 "LLVM and Clang types are out-of-synch");
1705 // From this point onwards, the Clang array type has been emitted
1706 // as some other type (probably a packed struct). Compute the array
1707 // size, and just emit the 'begin' expression as a bitcast.
1710 cast<ConstantArrayType>(arrayType)->getSize().getZExtValue();
1711 eltType = arrayType->getElementType();
1712 arrayType = getContext().getAsArrayType(eltType);
1715 llvm::Type *baseType = ConvertType(eltType);
1716 addr = Builder.CreateElementBitCast(addr, baseType, "array.begin");
1718 // Create the actual GEP.
1719 addr = Address(Builder.CreateInBoundsGEP(addr.getPointer(),
1720 gepIndices, "array.begin"),
1721 addr.getAlignment());
1726 llvm::Value *numElements
1727 = llvm::ConstantInt::get(SizeTy, countFromCLAs);
1729 // If we had any VLA dimensions, factor them in.
1731 numElements = Builder.CreateNUWMul(numVLAElements, numElements);
1736 std::pair<llvm::Value*, QualType>
1737 CodeGenFunction::getVLASize(QualType type) {
1738 const VariableArrayType *vla = getContext().getAsVariableArrayType(type);
1739 assert(vla && "type was not a variable array type!");
1740 return getVLASize(vla);
1743 std::pair<llvm::Value*, QualType>
1744 CodeGenFunction::getVLASize(const VariableArrayType *type) {
1745 // The number of elements so far; always size_t.
1746 llvm::Value *numElements = nullptr;
1748 QualType elementType;
1750 elementType = type->getElementType();
1751 llvm::Value *vlaSize = VLASizeMap[type->getSizeExpr()];
1752 assert(vlaSize && "no size for VLA!");
1753 assert(vlaSize->getType() == SizeTy);
1756 numElements = vlaSize;
1758 // It's undefined behavior if this wraps around, so mark it that way.
1759 // FIXME: Teach -fsanitize=undefined to trap this.
1760 numElements = Builder.CreateNUWMul(numElements, vlaSize);
1762 } while ((type = getContext().getAsVariableArrayType(elementType)));
1764 return std::pair<llvm::Value*,QualType>(numElements, elementType);
1767 void CodeGenFunction::EmitVariablyModifiedType(QualType type) {
1768 assert(type->isVariablyModifiedType() &&
1769 "Must pass variably modified type to EmitVLASizes!");
1771 EnsureInsertPoint();
1773 // We're going to walk down into the type and look for VLA
1776 assert(type->isVariablyModifiedType());
1778 const Type *ty = type.getTypePtr();
1779 switch (ty->getTypeClass()) {
1781 #define TYPE(Class, Base)
1782 #define ABSTRACT_TYPE(Class, Base)
1783 #define NON_CANONICAL_TYPE(Class, Base)
1784 #define DEPENDENT_TYPE(Class, Base) case Type::Class:
1785 #define NON_CANONICAL_UNLESS_DEPENDENT_TYPE(Class, Base)
1786 #include "clang/AST/TypeNodes.def"
1787 llvm_unreachable("unexpected dependent type!");
1789 // These types are never variably-modified.
1793 case Type::ExtVector:
1796 case Type::Elaborated:
1797 case Type::TemplateSpecialization:
1798 case Type::ObjCTypeParam:
1799 case Type::ObjCObject:
1800 case Type::ObjCInterface:
1801 case Type::ObjCObjectPointer:
1802 llvm_unreachable("type class is never variably-modified!");
1804 case Type::Adjusted:
1805 type = cast<AdjustedType>(ty)->getAdjustedType();
1809 type = cast<DecayedType>(ty)->getPointeeType();
1813 type = cast<PointerType>(ty)->getPointeeType();
1816 case Type::BlockPointer:
1817 type = cast<BlockPointerType>(ty)->getPointeeType();
1820 case Type::LValueReference:
1821 case Type::RValueReference:
1822 type = cast<ReferenceType>(ty)->getPointeeType();
1825 case Type::MemberPointer:
1826 type = cast<MemberPointerType>(ty)->getPointeeType();
1829 case Type::ConstantArray:
1830 case Type::IncompleteArray:
1831 // Losing element qualification here is fine.
1832 type = cast<ArrayType>(ty)->getElementType();
1835 case Type::VariableArray: {
1836 // Losing element qualification here is fine.
1837 const VariableArrayType *vat = cast<VariableArrayType>(ty);
1839 // Unknown size indication requires no size computation.
1840 // Otherwise, evaluate and record it.
1841 if (const Expr *size = vat->getSizeExpr()) {
1842 // It's possible that we might have emitted this already,
1843 // e.g. with a typedef and a pointer to it.
1844 llvm::Value *&entry = VLASizeMap[size];
1846 llvm::Value *Size = EmitScalarExpr(size);
1849 // If the size is an expression that is not an integer constant
1850 // expression [...] each time it is evaluated it shall have a value
1851 // greater than zero.
1852 if (SanOpts.has(SanitizerKind::VLABound) &&
1853 size->getType()->isSignedIntegerType()) {
1854 SanitizerScope SanScope(this);
1855 llvm::Value *Zero = llvm::Constant::getNullValue(Size->getType());
1856 llvm::Constant *StaticArgs[] = {
1857 EmitCheckSourceLocation(size->getLocStart()),
1858 EmitCheckTypeDescriptor(size->getType())
1860 EmitCheck(std::make_pair(Builder.CreateICmpSGT(Size, Zero),
1861 SanitizerKind::VLABound),
1862 SanitizerHandler::VLABoundNotPositive, StaticArgs, Size);
1865 // Always zexting here would be wrong if it weren't
1866 // undefined behavior to have a negative bound.
1867 entry = Builder.CreateIntCast(Size, SizeTy, /*signed*/ false);
1870 type = vat->getElementType();
1874 case Type::FunctionProto:
1875 case Type::FunctionNoProto:
1876 type = cast<FunctionType>(ty)->getReturnType();
1881 case Type::UnaryTransform:
1882 case Type::Attributed:
1883 case Type::SubstTemplateTypeParm:
1884 case Type::PackExpansion:
1885 // Keep walking after single level desugaring.
1886 type = type.getSingleStepDesugaredType(getContext());
1890 case Type::Decltype:
1892 // Stop walking: nothing to do.
1895 case Type::TypeOfExpr:
1896 // Stop walking: emit typeof expression.
1897 EmitIgnoredExpr(cast<TypeOfExprType>(ty)->getUnderlyingExpr());
1901 type = cast<AtomicType>(ty)->getValueType();
1905 type = cast<PipeType>(ty)->getElementType();
1908 } while (type->isVariablyModifiedType());
1911 Address CodeGenFunction::EmitVAListRef(const Expr* E) {
1912 if (getContext().getBuiltinVaListType()->isArrayType())
1913 return EmitPointerWithAlignment(E);
1914 return EmitLValue(E).getAddress();
1917 Address CodeGenFunction::EmitMSVAListRef(const Expr *E) {
1918 return EmitLValue(E).getAddress();
1921 void CodeGenFunction::EmitDeclRefExprDbgValue(const DeclRefExpr *E,
1922 const APValue &Init) {
1923 assert(!Init.isUninit() && "Invalid DeclRefExpr initializer!");
1924 if (CGDebugInfo *Dbg = getDebugInfo())
1925 if (CGM.getCodeGenOpts().getDebugInfo() >= codegenoptions::LimitedDebugInfo)
1926 Dbg->EmitGlobalVariable(E->getDecl(), Init);
1929 CodeGenFunction::PeepholeProtection
1930 CodeGenFunction::protectFromPeepholes(RValue rvalue) {
1931 // At the moment, the only aggressive peephole we do in IR gen
1932 // is trunc(zext) folding, but if we add more, we can easily
1933 // extend this protection.
1935 if (!rvalue.isScalar()) return PeepholeProtection();
1936 llvm::Value *value = rvalue.getScalarVal();
1937 if (!isa<llvm::ZExtInst>(value)) return PeepholeProtection();
1939 // Just make an extra bitcast.
1940 assert(HaveInsertPoint());
1941 llvm::Instruction *inst = new llvm::BitCastInst(value, value->getType(), "",
1942 Builder.GetInsertBlock());
1944 PeepholeProtection protection;
1945 protection.Inst = inst;
1949 void CodeGenFunction::unprotectFromPeepholes(PeepholeProtection protection) {
1950 if (!protection.Inst) return;
1952 // In theory, we could try to duplicate the peepholes now, but whatever.
1953 protection.Inst->eraseFromParent();
1956 llvm::Value *CodeGenFunction::EmitAnnotationCall(llvm::Value *AnnotationFn,
1957 llvm::Value *AnnotatedVal,
1958 StringRef AnnotationStr,
1959 SourceLocation Location) {
1960 llvm::Value *Args[4] = {
1962 Builder.CreateBitCast(CGM.EmitAnnotationString(AnnotationStr), Int8PtrTy),
1963 Builder.CreateBitCast(CGM.EmitAnnotationUnit(Location), Int8PtrTy),
1964 CGM.EmitAnnotationLineNo(Location)
1966 return Builder.CreateCall(AnnotationFn, Args);
1969 void CodeGenFunction::EmitVarAnnotations(const VarDecl *D, llvm::Value *V) {
1970 assert(D->hasAttr<AnnotateAttr>() && "no annotate attribute");
1971 // FIXME We create a new bitcast for every annotation because that's what
1972 // llvm-gcc was doing.
1973 for (const auto *I : D->specific_attrs<AnnotateAttr>())
1974 EmitAnnotationCall(CGM.getIntrinsic(llvm::Intrinsic::var_annotation),
1975 Builder.CreateBitCast(V, CGM.Int8PtrTy, V->getName()),
1976 I->getAnnotation(), D->getLocation());
1979 Address CodeGenFunction::EmitFieldAnnotations(const FieldDecl *D,
1981 assert(D->hasAttr<AnnotateAttr>() && "no annotate attribute");
1982 llvm::Value *V = Addr.getPointer();
1983 llvm::Type *VTy = V->getType();
1984 llvm::Value *F = CGM.getIntrinsic(llvm::Intrinsic::ptr_annotation,
1987 for (const auto *I : D->specific_attrs<AnnotateAttr>()) {
1988 // FIXME Always emit the cast inst so we can differentiate between
1989 // annotation on the first field of a struct and annotation on the struct
1991 if (VTy != CGM.Int8PtrTy)
1992 V = Builder.Insert(new llvm::BitCastInst(V, CGM.Int8PtrTy));
1993 V = EmitAnnotationCall(F, V, I->getAnnotation(), D->getLocation());
1994 V = Builder.CreateBitCast(V, VTy);
1997 return Address(V, Addr.getAlignment());
2000 CodeGenFunction::CGCapturedStmtInfo::~CGCapturedStmtInfo() { }
2002 CodeGenFunction::SanitizerScope::SanitizerScope(CodeGenFunction *CGF)
2004 assert(!CGF->IsSanitizerScope);
2005 CGF->IsSanitizerScope = true;
2008 CodeGenFunction::SanitizerScope::~SanitizerScope() {
2009 CGF->IsSanitizerScope = false;
2012 void CodeGenFunction::InsertHelper(llvm::Instruction *I,
2013 const llvm::Twine &Name,
2014 llvm::BasicBlock *BB,
2015 llvm::BasicBlock::iterator InsertPt) const {
2016 LoopStack.InsertHelper(I);
2017 if (IsSanitizerScope)
2018 CGM.getSanitizerMetadata()->disableSanitizerForInstruction(I);
2021 void CGBuilderInserter::InsertHelper(
2022 llvm::Instruction *I, const llvm::Twine &Name, llvm::BasicBlock *BB,
2023 llvm::BasicBlock::iterator InsertPt) const {
2024 llvm::IRBuilderDefaultInserter::InsertHelper(I, Name, BB, InsertPt);
2026 CGF->InsertHelper(I, Name, BB, InsertPt);
2029 static bool hasRequiredFeatures(const SmallVectorImpl<StringRef> &ReqFeatures,
2030 CodeGenModule &CGM, const FunctionDecl *FD,
2031 std::string &FirstMissing) {
2032 // If there aren't any required features listed then go ahead and return.
2033 if (ReqFeatures.empty())
2036 // Now build up the set of caller features and verify that all the required
2037 // features are there.
2038 llvm::StringMap<bool> CallerFeatureMap;
2039 CGM.getFunctionFeatureMap(CallerFeatureMap, FD);
2041 // If we have at least one of the features in the feature list return
2042 // true, otherwise return false.
2044 ReqFeatures.begin(), ReqFeatures.end(), [&](StringRef Feature) {
2045 SmallVector<StringRef, 1> OrFeatures;
2046 Feature.split(OrFeatures, "|");
2047 return std::any_of(OrFeatures.begin(), OrFeatures.end(),
2048 [&](StringRef Feature) {
2049 if (!CallerFeatureMap.lookup(Feature)) {
2050 FirstMissing = Feature.str();
2058 // Emits an error if we don't have a valid set of target features for the
2060 void CodeGenFunction::checkTargetFeatures(const CallExpr *E,
2061 const FunctionDecl *TargetDecl) {
2062 // Early exit if this is an indirect call.
2066 // Get the current enclosing function if it exists. If it doesn't
2067 // we can't check the target features anyhow.
2068 const FunctionDecl *FD = dyn_cast_or_null<FunctionDecl>(CurFuncDecl);
2072 // Grab the required features for the call. For a builtin this is listed in
2073 // the td file with the default cpu, for an always_inline function this is any
2074 // listed cpu and any listed features.
2075 unsigned BuiltinID = TargetDecl->getBuiltinID();
2076 std::string MissingFeature;
2078 SmallVector<StringRef, 1> ReqFeatures;
2079 const char *FeatureList =
2080 CGM.getContext().BuiltinInfo.getRequiredFeatures(BuiltinID);
2081 // Return if the builtin doesn't have any required features.
2082 if (!FeatureList || StringRef(FeatureList) == "")
2084 StringRef(FeatureList).split(ReqFeatures, ",");
2085 if (!hasRequiredFeatures(ReqFeatures, CGM, FD, MissingFeature))
2086 CGM.getDiags().Report(E->getLocStart(), diag::err_builtin_needs_feature)
2087 << TargetDecl->getDeclName()
2088 << CGM.getContext().BuiltinInfo.getRequiredFeatures(BuiltinID);
2090 } else if (TargetDecl->hasAttr<TargetAttr>()) {
2091 // Get the required features for the callee.
2092 SmallVector<StringRef, 1> ReqFeatures;
2093 llvm::StringMap<bool> CalleeFeatureMap;
2094 CGM.getFunctionFeatureMap(CalleeFeatureMap, TargetDecl);
2095 for (const auto &F : CalleeFeatureMap) {
2096 // Only positive features are "required".
2098 ReqFeatures.push_back(F.getKey());
2100 if (!hasRequiredFeatures(ReqFeatures, CGM, FD, MissingFeature))
2101 CGM.getDiags().Report(E->getLocStart(), diag::err_function_needs_feature)
2102 << FD->getDeclName() << TargetDecl->getDeclName() << MissingFeature;
2106 void CodeGenFunction::EmitSanitizerStatReport(llvm::SanitizerStatKind SSK) {
2107 if (!CGM.getCodeGenOpts().SanitizeStats)
2110 llvm::IRBuilder<> IRB(Builder.GetInsertBlock(), Builder.GetInsertPoint());
2111 IRB.SetCurrentDebugLocation(Builder.getCurrentDebugLocation());
2112 CGM.getSanStats().create(IRB, SSK);
2115 llvm::DebugLoc CodeGenFunction::SourceLocToDebugLoc(SourceLocation Location) {
2116 if (CGDebugInfo *DI = getDebugInfo())
2117 return DI->SourceLocToDebugLoc(Location);
2119 return llvm::DebugLoc();