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 if (CGOpts.DisableLifetimeMarkers)
48 // Disable lifetime markers in msan builds.
49 // FIXME: Remove this when msan works with lifetime markers.
50 if (LangOpts.Sanitize.has(SanitizerKind::Memory))
53 // Asan uses markers for use-after-scope checks.
54 if (CGOpts.SanitizeAddressUseAfterScope)
57 // For now, only in optimized builds.
58 return CGOpts.OptimizationLevel != 0;
61 CodeGenFunction::CodeGenFunction(CodeGenModule &cgm, bool suppressNewContext)
62 : CodeGenTypeCache(cgm), CGM(cgm), Target(cgm.getTarget()),
63 Builder(cgm, cgm.getModule().getContext(), llvm::ConstantFolder(),
64 CGBuilderInserterTy(this)),
65 CurFn(nullptr), ReturnValue(Address::invalid()),
66 CapturedStmtInfo(nullptr), SanOpts(CGM.getLangOpts().Sanitize),
67 IsSanitizerScope(false), CurFuncIsThunk(false), AutoreleaseResult(false),
68 SawAsmBlock(false), IsOutlinedSEHHelper(false), BlockInfo(nullptr),
69 BlockPointer(nullptr), LambdaThisCaptureField(nullptr),
70 NormalCleanupDest(nullptr), NextCleanupDestIndex(1),
71 FirstBlockInfo(nullptr), EHResumeBlock(nullptr), ExceptionSlot(nullptr),
72 EHSelectorSlot(nullptr), DebugInfo(CGM.getModuleDebugInfo()),
73 DisableDebugInfo(false), DidCallStackSave(false), IndirectBranch(nullptr),
74 PGO(cgm), SwitchInsn(nullptr), SwitchWeights(nullptr),
75 CaseRangeBlock(nullptr), UnreachableBlock(nullptr), NumReturnExprs(0),
76 NumSimpleReturnExprs(0), CXXABIThisDecl(nullptr),
77 CXXABIThisValue(nullptr), CXXThisValue(nullptr),
78 CXXStructorImplicitParamDecl(nullptr),
79 CXXStructorImplicitParamValue(nullptr), OutermostConditional(nullptr),
80 CurLexicalScope(nullptr), TerminateLandingPad(nullptr),
81 TerminateHandler(nullptr), TrapBB(nullptr),
82 ShouldEmitLifetimeMarkers(
83 shouldEmitLifetimeMarkers(CGM.getCodeGenOpts(), CGM.getLangOpts())) {
84 if (!suppressNewContext)
85 CGM.getCXXABI().getMangleContext().startNewFunction();
87 llvm::FastMathFlags FMF;
88 if (CGM.getLangOpts().FastMath)
89 FMF.setUnsafeAlgebra();
90 if (CGM.getLangOpts().FiniteMathOnly) {
94 if (CGM.getCodeGenOpts().NoNaNsFPMath) {
97 if (CGM.getCodeGenOpts().NoSignedZeros) {
98 FMF.setNoSignedZeros();
100 if (CGM.getCodeGenOpts().ReciprocalMath) {
101 FMF.setAllowReciprocal();
103 Builder.setFastMathFlags(FMF);
106 CodeGenFunction::~CodeGenFunction() {
107 assert(LifetimeExtendedCleanupStack.empty() && "failed to emit a cleanup");
109 // If there are any unclaimed block infos, go ahead and destroy them
110 // now. This can happen if IR-gen gets clever and skips evaluating
113 destroyBlockInfos(FirstBlockInfo);
115 if (getLangOpts().OpenMP && CurFn)
116 CGM.getOpenMPRuntime().functionFinished(*this);
119 CharUnits CodeGenFunction::getNaturalPointeeTypeAlignment(QualType T,
120 AlignmentSource *Source) {
121 return getNaturalTypeAlignment(T->getPointeeType(), Source,
122 /*forPointee*/ true);
125 CharUnits CodeGenFunction::getNaturalTypeAlignment(QualType T,
126 AlignmentSource *Source,
127 bool forPointeeType) {
128 // Honor alignment typedef attributes even on incomplete types.
129 // We also honor them straight for C++ class types, even as pointees;
130 // there's an expressivity gap here.
131 if (auto TT = T->getAs<TypedefType>()) {
132 if (auto Align = TT->getDecl()->getMaxAlignment()) {
133 if (Source) *Source = AlignmentSource::AttributedType;
134 return getContext().toCharUnitsFromBits(Align);
138 if (Source) *Source = AlignmentSource::Type;
141 if (T->isIncompleteType()) {
142 Alignment = CharUnits::One(); // Shouldn't be used, but pessimistic is best.
144 // For C++ class pointees, we don't know whether we're pointing at a
145 // base or a complete object, so we generally need to use the
146 // non-virtual alignment.
147 const CXXRecordDecl *RD;
148 if (forPointeeType && (RD = T->getAsCXXRecordDecl())) {
149 Alignment = CGM.getClassPointerAlignment(RD);
151 Alignment = getContext().getTypeAlignInChars(T);
152 if (T.getQualifiers().hasUnaligned())
153 Alignment = CharUnits::One();
156 // Cap to the global maximum type alignment unless the alignment
157 // was somehow explicit on the type.
158 if (unsigned MaxAlign = getLangOpts().MaxTypeAlign) {
159 if (Alignment.getQuantity() > MaxAlign &&
160 !getContext().isAlignmentRequired(T))
161 Alignment = CharUnits::fromQuantity(MaxAlign);
167 LValue CodeGenFunction::MakeNaturalAlignAddrLValue(llvm::Value *V, QualType T) {
168 AlignmentSource AlignSource;
169 CharUnits Alignment = getNaturalTypeAlignment(T, &AlignSource);
170 return LValue::MakeAddr(Address(V, Alignment), T, getContext(), AlignSource,
174 /// Given a value of type T* that may not be to a complete object,
175 /// construct an l-value with the natural pointee alignment of T.
177 CodeGenFunction::MakeNaturalAlignPointeeAddrLValue(llvm::Value *V, QualType T) {
178 AlignmentSource AlignSource;
179 CharUnits Align = getNaturalTypeAlignment(T, &AlignSource, /*pointee*/ true);
180 return MakeAddrLValue(Address(V, Align), T, AlignSource);
184 llvm::Type *CodeGenFunction::ConvertTypeForMem(QualType T) {
185 return CGM.getTypes().ConvertTypeForMem(T);
188 llvm::Type *CodeGenFunction::ConvertType(QualType T) {
189 return CGM.getTypes().ConvertType(T);
192 TypeEvaluationKind CodeGenFunction::getEvaluationKind(QualType type) {
193 type = type.getCanonicalType();
195 switch (type->getTypeClass()) {
196 #define TYPE(name, parent)
197 #define ABSTRACT_TYPE(name, parent)
198 #define NON_CANONICAL_TYPE(name, parent) case Type::name:
199 #define DEPENDENT_TYPE(name, parent) case Type::name:
200 #define NON_CANONICAL_UNLESS_DEPENDENT_TYPE(name, parent) case Type::name:
201 #include "clang/AST/TypeNodes.def"
202 llvm_unreachable("non-canonical or dependent type in IR-generation");
205 case Type::DeducedTemplateSpecialization:
206 llvm_unreachable("undeduced type in IR-generation");
208 // Various scalar types.
211 case Type::BlockPointer:
212 case Type::LValueReference:
213 case Type::RValueReference:
214 case Type::MemberPointer:
216 case Type::ExtVector:
217 case Type::FunctionProto:
218 case Type::FunctionNoProto:
220 case Type::ObjCObjectPointer:
228 // Arrays, records, and Objective-C objects.
229 case Type::ConstantArray:
230 case Type::IncompleteArray:
231 case Type::VariableArray:
233 case Type::ObjCObject:
234 case Type::ObjCInterface:
235 return TEK_Aggregate;
237 // We operate on atomic values according to their underlying type.
239 type = cast<AtomicType>(type)->getValueType();
242 llvm_unreachable("unknown type kind!");
246 llvm::DebugLoc CodeGenFunction::EmitReturnBlock() {
247 // For cleanliness, we try to avoid emitting the return block for
249 llvm::BasicBlock *CurBB = Builder.GetInsertBlock();
252 assert(!CurBB->getTerminator() && "Unexpected terminated block.");
254 // We have a valid insert point, reuse it if it is empty or there are no
255 // explicit jumps to the return block.
256 if (CurBB->empty() || ReturnBlock.getBlock()->use_empty()) {
257 ReturnBlock.getBlock()->replaceAllUsesWith(CurBB);
258 delete ReturnBlock.getBlock();
260 EmitBlock(ReturnBlock.getBlock());
261 return llvm::DebugLoc();
264 // Otherwise, if the return block is the target of a single direct
265 // branch then we can just put the code in that block instead. This
266 // cleans up functions which started with a unified return block.
267 if (ReturnBlock.getBlock()->hasOneUse()) {
268 llvm::BranchInst *BI =
269 dyn_cast<llvm::BranchInst>(*ReturnBlock.getBlock()->user_begin());
270 if (BI && BI->isUnconditional() &&
271 BI->getSuccessor(0) == ReturnBlock.getBlock()) {
272 // Record/return the DebugLoc of the simple 'return' expression to be used
273 // later by the actual 'ret' instruction.
274 llvm::DebugLoc Loc = BI->getDebugLoc();
275 Builder.SetInsertPoint(BI->getParent());
276 BI->eraseFromParent();
277 delete ReturnBlock.getBlock();
282 // FIXME: We are at an unreachable point, there is no reason to emit the block
283 // unless it has uses. However, we still need a place to put the debug
284 // region.end for now.
286 EmitBlock(ReturnBlock.getBlock());
287 return llvm::DebugLoc();
290 static void EmitIfUsed(CodeGenFunction &CGF, llvm::BasicBlock *BB) {
292 if (!BB->use_empty())
293 return CGF.CurFn->getBasicBlockList().push_back(BB);
297 void CodeGenFunction::FinishFunction(SourceLocation EndLoc) {
298 assert(BreakContinueStack.empty() &&
299 "mismatched push/pop in break/continue stack!");
301 bool OnlySimpleReturnStmts = NumSimpleReturnExprs > 0
302 && NumSimpleReturnExprs == NumReturnExprs
303 && ReturnBlock.getBlock()->use_empty();
304 // Usually the return expression is evaluated before the cleanup
305 // code. If the function contains only a simple return statement,
306 // such as a constant, the location before the cleanup code becomes
307 // the last useful breakpoint in the function, because the simple
308 // return expression will be evaluated after the cleanup code. To be
309 // safe, set the debug location for cleanup code to the location of
310 // the return statement. Otherwise the cleanup code should be at the
311 // end of the function's lexical scope.
313 // If there are multiple branches to the return block, the branch
314 // instructions will get the location of the return statements and
316 if (CGDebugInfo *DI = getDebugInfo()) {
317 if (OnlySimpleReturnStmts)
318 DI->EmitLocation(Builder, LastStopPoint);
320 DI->EmitLocation(Builder, EndLoc);
323 // Pop any cleanups that might have been associated with the
324 // parameters. Do this in whatever block we're currently in; it's
325 // important to do this before we enter the return block or return
326 // edges will be *really* confused.
327 bool HasCleanups = EHStack.stable_begin() != PrologueCleanupDepth;
328 bool HasOnlyLifetimeMarkers =
329 HasCleanups && EHStack.containsOnlyLifetimeMarkers(PrologueCleanupDepth);
330 bool EmitRetDbgLoc = !HasCleanups || HasOnlyLifetimeMarkers;
332 // Make sure the line table doesn't jump back into the body for
333 // the ret after it's been at EndLoc.
334 if (CGDebugInfo *DI = getDebugInfo())
335 if (OnlySimpleReturnStmts)
336 DI->EmitLocation(Builder, EndLoc);
338 PopCleanupBlocks(PrologueCleanupDepth);
341 // Emit function epilog (to return).
342 llvm::DebugLoc Loc = EmitReturnBlock();
344 if (ShouldInstrumentFunction())
345 EmitFunctionInstrumentation("__cyg_profile_func_exit");
347 // Emit debug descriptor for function end.
348 if (CGDebugInfo *DI = getDebugInfo())
349 DI->EmitFunctionEnd(Builder);
351 // Reset the debug location to that of the simple 'return' expression, if any
352 // rather than that of the end of the function's scope '}'.
353 ApplyDebugLocation AL(*this, Loc);
354 EmitFunctionEpilog(*CurFnInfo, EmitRetDbgLoc, EndLoc);
355 EmitEndEHSpec(CurCodeDecl);
357 assert(EHStack.empty() &&
358 "did not remove all scopes from cleanup stack!");
360 // If someone did an indirect goto, emit the indirect goto block at the end of
362 if (IndirectBranch) {
363 EmitBlock(IndirectBranch->getParent());
364 Builder.ClearInsertionPoint();
367 // If some of our locals escaped, insert a call to llvm.localescape in the
369 if (!EscapedLocals.empty()) {
370 // Invert the map from local to index into a simple vector. There should be
372 SmallVector<llvm::Value *, 4> EscapeArgs;
373 EscapeArgs.resize(EscapedLocals.size());
374 for (auto &Pair : EscapedLocals)
375 EscapeArgs[Pair.second] = Pair.first;
376 llvm::Function *FrameEscapeFn = llvm::Intrinsic::getDeclaration(
377 &CGM.getModule(), llvm::Intrinsic::localescape);
378 CGBuilderTy(*this, AllocaInsertPt).CreateCall(FrameEscapeFn, EscapeArgs);
381 // Remove the AllocaInsertPt instruction, which is just a convenience for us.
382 llvm::Instruction *Ptr = AllocaInsertPt;
383 AllocaInsertPt = nullptr;
384 Ptr->eraseFromParent();
386 // If someone took the address of a label but never did an indirect goto, we
387 // made a zero entry PHI node, which is illegal, zap it now.
388 if (IndirectBranch) {
389 llvm::PHINode *PN = cast<llvm::PHINode>(IndirectBranch->getAddress());
390 if (PN->getNumIncomingValues() == 0) {
391 PN->replaceAllUsesWith(llvm::UndefValue::get(PN->getType()));
392 PN->eraseFromParent();
396 EmitIfUsed(*this, EHResumeBlock);
397 EmitIfUsed(*this, TerminateLandingPad);
398 EmitIfUsed(*this, TerminateHandler);
399 EmitIfUsed(*this, UnreachableBlock);
401 if (CGM.getCodeGenOpts().EmitDeclMetadata)
404 for (SmallVectorImpl<std::pair<llvm::Instruction *, llvm::Value *> >::iterator
405 I = DeferredReplacements.begin(),
406 E = DeferredReplacements.end();
408 I->first->replaceAllUsesWith(I->second);
409 I->first->eraseFromParent();
413 /// ShouldInstrumentFunction - Return true if the current function should be
414 /// instrumented with __cyg_profile_func_* calls
415 bool CodeGenFunction::ShouldInstrumentFunction() {
416 if (!CGM.getCodeGenOpts().InstrumentFunctions)
418 if (!CurFuncDecl || CurFuncDecl->hasAttr<NoInstrumentFunctionAttr>())
423 /// ShouldXRayInstrument - Return true if the current function should be
424 /// instrumented with XRay nop sleds.
425 bool CodeGenFunction::ShouldXRayInstrumentFunction() const {
426 return CGM.getCodeGenOpts().XRayInstrumentFunctions;
429 /// EmitFunctionInstrumentation - Emit LLVM code to call the specified
430 /// instrumentation function with the current function and the call site, if
431 /// function instrumentation is enabled.
432 void CodeGenFunction::EmitFunctionInstrumentation(const char *Fn) {
433 auto NL = ApplyDebugLocation::CreateArtificial(*this);
434 // void __cyg_profile_func_{enter,exit} (void *this_fn, void *call_site);
435 llvm::PointerType *PointerTy = Int8PtrTy;
436 llvm::Type *ProfileFuncArgs[] = { PointerTy, PointerTy };
437 llvm::FunctionType *FunctionTy =
438 llvm::FunctionType::get(VoidTy, ProfileFuncArgs, false);
440 llvm::Constant *F = CGM.CreateRuntimeFunction(FunctionTy, Fn);
441 llvm::CallInst *CallSite = Builder.CreateCall(
442 CGM.getIntrinsic(llvm::Intrinsic::returnaddress),
443 llvm::ConstantInt::get(Int32Ty, 0),
446 llvm::Value *args[] = {
447 llvm::ConstantExpr::getBitCast(CurFn, PointerTy),
451 EmitNounwindRuntimeCall(F, args);
454 static void removeImageAccessQualifier(std::string& TyName) {
455 std::string ReadOnlyQual("__read_only");
456 std::string::size_type ReadOnlyPos = TyName.find(ReadOnlyQual);
457 if (ReadOnlyPos != std::string::npos)
458 // "+ 1" for the space after access qualifier.
459 TyName.erase(ReadOnlyPos, ReadOnlyQual.size() + 1);
461 std::string WriteOnlyQual("__write_only");
462 std::string::size_type WriteOnlyPos = TyName.find(WriteOnlyQual);
463 if (WriteOnlyPos != std::string::npos)
464 TyName.erase(WriteOnlyPos, WriteOnlyQual.size() + 1);
466 std::string ReadWriteQual("__read_write");
467 std::string::size_type ReadWritePos = TyName.find(ReadWriteQual);
468 if (ReadWritePos != std::string::npos)
469 TyName.erase(ReadWritePos, ReadWriteQual.size() + 1);
474 // Returns the address space id that should be produced to the
475 // kernel_arg_addr_space metadata. This is always fixed to the ids
476 // as specified in the SPIR 2.0 specification in order to differentiate
477 // for example in clGetKernelArgInfo() implementation between the address
478 // spaces with targets without unique mapping to the OpenCL address spaces
479 // (basically all single AS CPUs).
480 static unsigned ArgInfoAddressSpace(unsigned LangAS) {
482 case LangAS::opencl_global: return 1;
483 case LangAS::opencl_constant: return 2;
484 case LangAS::opencl_local: return 3;
485 case LangAS::opencl_generic: return 4; // Not in SPIR 2.0 specs.
487 return 0; // Assume private.
491 // OpenCL v1.2 s5.6.4.6 allows the compiler to store kernel argument
492 // information in the program executable. The argument information stored
493 // includes the argument name, its type, the address and access qualifiers used.
494 static void GenOpenCLArgMetadata(const FunctionDecl *FD, llvm::Function *Fn,
495 CodeGenModule &CGM, llvm::LLVMContext &Context,
496 CGBuilderTy &Builder, ASTContext &ASTCtx) {
497 // Create MDNodes that represent the kernel arg metadata.
498 // Each MDNode is a list in the form of "key", N number of values which is
499 // the same number of values as their are kernel arguments.
501 const PrintingPolicy &Policy = ASTCtx.getPrintingPolicy();
503 // MDNode for the kernel argument address space qualifiers.
504 SmallVector<llvm::Metadata *, 8> addressQuals;
506 // MDNode for the kernel argument access qualifiers (images only).
507 SmallVector<llvm::Metadata *, 8> accessQuals;
509 // MDNode for the kernel argument type names.
510 SmallVector<llvm::Metadata *, 8> argTypeNames;
512 // MDNode for the kernel argument base type names.
513 SmallVector<llvm::Metadata *, 8> argBaseTypeNames;
515 // MDNode for the kernel argument type qualifiers.
516 SmallVector<llvm::Metadata *, 8> argTypeQuals;
518 // MDNode for the kernel argument names.
519 SmallVector<llvm::Metadata *, 8> argNames;
521 for (unsigned i = 0, e = FD->getNumParams(); i != e; ++i) {
522 const ParmVarDecl *parm = FD->getParamDecl(i);
523 QualType ty = parm->getType();
524 std::string typeQuals;
526 if (ty->isPointerType()) {
527 QualType pointeeTy = ty->getPointeeType();
529 // Get address qualifier.
530 addressQuals.push_back(llvm::ConstantAsMetadata::get(Builder.getInt32(
531 ArgInfoAddressSpace(pointeeTy.getAddressSpace()))));
533 // Get argument type name.
534 std::string typeName =
535 pointeeTy.getUnqualifiedType().getAsString(Policy) + "*";
537 // Turn "unsigned type" to "utype"
538 std::string::size_type pos = typeName.find("unsigned");
539 if (pointeeTy.isCanonical() && pos != std::string::npos)
540 typeName.erase(pos+1, 8);
542 argTypeNames.push_back(llvm::MDString::get(Context, typeName));
544 std::string baseTypeName =
545 pointeeTy.getUnqualifiedType().getCanonicalType().getAsString(
549 // Turn "unsigned type" to "utype"
550 pos = baseTypeName.find("unsigned");
551 if (pos != std::string::npos)
552 baseTypeName.erase(pos+1, 8);
554 argBaseTypeNames.push_back(llvm::MDString::get(Context, baseTypeName));
556 // Get argument type qualifiers:
557 if (ty.isRestrictQualified())
558 typeQuals = "restrict";
559 if (pointeeTy.isConstQualified() ||
560 (pointeeTy.getAddressSpace() == LangAS::opencl_constant))
561 typeQuals += typeQuals.empty() ? "const" : " const";
562 if (pointeeTy.isVolatileQualified())
563 typeQuals += typeQuals.empty() ? "volatile" : " volatile";
565 uint32_t AddrSpc = 0;
566 bool isPipe = ty->isPipeType();
567 if (ty->isImageType() || isPipe)
568 AddrSpc = ArgInfoAddressSpace(LangAS::opencl_global);
570 addressQuals.push_back(
571 llvm::ConstantAsMetadata::get(Builder.getInt32(AddrSpc)));
573 // Get argument type name.
574 std::string typeName;
576 typeName = ty.getCanonicalType()->getAs<PipeType>()->getElementType()
577 .getAsString(Policy);
579 typeName = ty.getUnqualifiedType().getAsString(Policy);
581 // Turn "unsigned type" to "utype"
582 std::string::size_type pos = typeName.find("unsigned");
583 if (ty.isCanonical() && pos != std::string::npos)
584 typeName.erase(pos+1, 8);
586 std::string baseTypeName;
588 baseTypeName = ty.getCanonicalType()->getAs<PipeType>()
589 ->getElementType().getCanonicalType()
590 .getAsString(Policy);
593 ty.getUnqualifiedType().getCanonicalType().getAsString(Policy);
595 // Remove access qualifiers on images
596 // (as they are inseparable from type in clang implementation,
597 // but OpenCL spec provides a special query to get access qualifier
598 // via clGetKernelArgInfo with CL_KERNEL_ARG_ACCESS_QUALIFIER):
599 if (ty->isImageType()) {
600 removeImageAccessQualifier(typeName);
601 removeImageAccessQualifier(baseTypeName);
604 argTypeNames.push_back(llvm::MDString::get(Context, typeName));
606 // Turn "unsigned type" to "utype"
607 pos = baseTypeName.find("unsigned");
608 if (pos != std::string::npos)
609 baseTypeName.erase(pos+1, 8);
611 argBaseTypeNames.push_back(llvm::MDString::get(Context, baseTypeName));
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 static void markAsIgnoreThreadCheckingAtRuntime(llvm::Function *Fn) {
709 Fn->addFnAttr("sanitize_thread_no_checking_at_run_time");
710 Fn->removeFnAttr(llvm::Attribute::SanitizeThread);
713 void CodeGenFunction::StartFunction(GlobalDecl GD,
716 const CGFunctionInfo &FnInfo,
717 const FunctionArgList &Args,
719 SourceLocation StartLoc) {
721 "Do not use a CodeGenFunction object for more than one function");
723 const Decl *D = GD.getDecl();
725 DidCallStackSave = false;
727 if (const auto *FD = dyn_cast_or_null<FunctionDecl>(D))
728 if (FD->usesSEHTry())
730 CurFuncDecl = (D ? D->getNonClosureContext() : nullptr);
734 assert(CurFn->isDeclaration() && "Function already has body?");
736 if (CGM.isInSanitizerBlacklist(Fn, Loc))
740 // Apply the no_sanitize* attributes to SanOpts.
741 for (auto Attr : D->specific_attrs<NoSanitizeAttr>())
742 SanOpts.Mask &= ~Attr->getMask();
745 // Apply sanitizer attributes to the function.
746 if (SanOpts.hasOneOf(SanitizerKind::Address | SanitizerKind::KernelAddress))
747 Fn->addFnAttr(llvm::Attribute::SanitizeAddress);
748 if (SanOpts.has(SanitizerKind::Thread))
749 Fn->addFnAttr(llvm::Attribute::SanitizeThread);
750 if (SanOpts.has(SanitizerKind::Memory))
751 Fn->addFnAttr(llvm::Attribute::SanitizeMemory);
752 if (SanOpts.has(SanitizerKind::SafeStack))
753 Fn->addFnAttr(llvm::Attribute::SafeStack);
755 // Ignore TSan memory acesses from within ObjC/ObjC++ dealloc, initialize,
756 // .cxx_destruct, __destroy_helper_block_ and all of their calees at run time.
757 if (SanOpts.has(SanitizerKind::Thread)) {
758 if (const auto *OMD = dyn_cast_or_null<ObjCMethodDecl>(D)) {
759 IdentifierInfo *II = OMD->getSelector().getIdentifierInfoForSlot(0);
760 if (OMD->getMethodFamily() == OMF_dealloc ||
761 OMD->getMethodFamily() == OMF_initialize ||
762 (OMD->getSelector().isUnarySelector() && II->isStr(".cxx_destruct"))) {
763 markAsIgnoreThreadCheckingAtRuntime(Fn);
765 } else if (const auto *FD = dyn_cast_or_null<FunctionDecl>(D)) {
766 IdentifierInfo *II = FD->getIdentifier();
767 if (II && II->isStr("__destroy_helper_block_"))
768 markAsIgnoreThreadCheckingAtRuntime(Fn);
772 // Apply xray attributes to the function (as a string, for now)
773 if (D && ShouldXRayInstrumentFunction()) {
774 if (const auto *XRayAttr = D->getAttr<XRayInstrumentAttr>()) {
775 if (XRayAttr->alwaysXRayInstrument())
776 Fn->addFnAttr("function-instrument", "xray-always");
777 if (XRayAttr->neverXRayInstrument())
778 Fn->addFnAttr("function-instrument", "xray-never");
779 if (const auto *LogArgs = D->getAttr<XRayLogArgsAttr>()) {
780 Fn->addFnAttr("xray-log-args",
781 llvm::utostr(LogArgs->getArgumentCount()));
784 if (!CGM.imbueXRayAttrs(Fn, Loc))
786 "xray-instruction-threshold",
787 llvm::itostr(CGM.getCodeGenOpts().XRayInstructionThreshold));
791 if (const FunctionDecl *FD = dyn_cast_or_null<FunctionDecl>(D))
792 if (CGM.getLangOpts().OpenMP && FD->hasAttr<OMPDeclareSimdDeclAttr>())
793 CGM.getOpenMPRuntime().emitDeclareSimdFunction(FD, Fn);
795 // Add no-jump-tables value.
796 Fn->addFnAttr("no-jump-tables",
797 llvm::toStringRef(CGM.getCodeGenOpts().NoUseJumpTables));
799 if (getLangOpts().OpenCL) {
800 // Add metadata for a kernel function.
801 if (const FunctionDecl *FD = dyn_cast_or_null<FunctionDecl>(D))
802 EmitOpenCLKernelMetadata(FD, Fn);
805 // If we are checking function types, emit a function type signature as
807 if (getLangOpts().CPlusPlus && SanOpts.has(SanitizerKind::Function)) {
808 if (const FunctionDecl *FD = dyn_cast_or_null<FunctionDecl>(D)) {
809 if (llvm::Constant *PrologueSig =
810 CGM.getTargetCodeGenInfo().getUBSanFunctionSignature(CGM)) {
811 llvm::Constant *FTRTTIConst =
812 CGM.GetAddrOfRTTIDescriptor(FD->getType(), /*ForEH=*/true);
813 llvm::Constant *PrologueStructElems[] = { PrologueSig, FTRTTIConst };
814 llvm::Constant *PrologueStructConst =
815 llvm::ConstantStruct::getAnon(PrologueStructElems, /*Packed=*/true);
816 Fn->setPrologueData(PrologueStructConst);
821 // If we're checking nullability, we need to know whether we can check the
822 // return value. Initialize the flag to 'true' and refine it in EmitParmDecl.
823 if (SanOpts.has(SanitizerKind::NullabilityReturn)) {
824 auto Nullability = FnRetTy->getNullability(getContext());
825 if (Nullability && *Nullability == NullabilityKind::NonNull) {
826 if (!(SanOpts.has(SanitizerKind::ReturnsNonnullAttribute) &&
827 CurCodeDecl && CurCodeDecl->getAttr<ReturnsNonNullAttr>()))
828 RetValNullabilityPrecondition =
829 llvm::ConstantInt::getTrue(getLLVMContext());
833 // If we're in C++ mode and the function name is "main", it is guaranteed
834 // to be norecurse by the standard (3.6.1.3 "The function main shall not be
835 // used within a program").
836 if (getLangOpts().CPlusPlus)
837 if (const FunctionDecl *FD = dyn_cast_or_null<FunctionDecl>(D))
839 Fn->addFnAttr(llvm::Attribute::NoRecurse);
841 llvm::BasicBlock *EntryBB = createBasicBlock("entry", CurFn);
843 // Create a marker to make it easy to insert allocas into the entryblock
844 // later. Don't create this with the builder, because we don't want it
846 llvm::Value *Undef = llvm::UndefValue::get(Int32Ty);
847 AllocaInsertPt = new llvm::BitCastInst(Undef, Int32Ty, "allocapt", EntryBB);
849 ReturnBlock = getJumpDestInCurrentScope("return");
851 Builder.SetInsertPoint(EntryBB);
853 // Emit subprogram debug descriptor.
854 if (CGDebugInfo *DI = getDebugInfo()) {
855 // Reconstruct the type from the argument list so that implicit parameters,
856 // such as 'this' and 'vtt', show up in the debug info. Preserve the calling
858 CallingConv CC = CallingConv::CC_C;
859 if (auto *FD = dyn_cast_or_null<FunctionDecl>(D))
860 if (const auto *SrcFnTy = FD->getType()->getAs<FunctionType>())
861 CC = SrcFnTy->getCallConv();
862 SmallVector<QualType, 16> ArgTypes;
863 for (const VarDecl *VD : Args)
864 ArgTypes.push_back(VD->getType());
865 QualType FnType = getContext().getFunctionType(
866 RetTy, ArgTypes, FunctionProtoType::ExtProtoInfo(CC));
867 DI->EmitFunctionStart(GD, Loc, StartLoc, FnType, CurFn, Builder);
870 if (ShouldInstrumentFunction())
871 EmitFunctionInstrumentation("__cyg_profile_func_enter");
873 // Since emitting the mcount call here impacts optimizations such as function
874 // inlining, we just add an attribute to insert a mcount call in backend.
875 // The attribute "counting-function" is set to mcount function name which is
876 // architecture dependent.
877 if (CGM.getCodeGenOpts().InstrumentForProfiling) {
878 if (CGM.getCodeGenOpts().CallFEntry)
879 Fn->addFnAttr("fentry-call", "true");
881 Fn->addFnAttr("counting-function", getTarget().getMCountName());
884 if (RetTy->isVoidType()) {
885 // Void type; nothing to return.
886 ReturnValue = Address::invalid();
888 // Count the implicit return.
889 if (!endsWithReturn(D))
891 } else if (CurFnInfo->getReturnInfo().getKind() == ABIArgInfo::Indirect &&
892 !hasScalarEvaluationKind(CurFnInfo->getReturnType())) {
893 // Indirect aggregate return; emit returned value directly into sret slot.
894 // This reduces code size, and affects correctness in C++.
895 auto AI = CurFn->arg_begin();
896 if (CurFnInfo->getReturnInfo().isSRetAfterThis())
898 ReturnValue = Address(&*AI, CurFnInfo->getReturnInfo().getIndirectAlign());
899 } else if (CurFnInfo->getReturnInfo().getKind() == ABIArgInfo::InAlloca &&
900 !hasScalarEvaluationKind(CurFnInfo->getReturnType())) {
901 // Load the sret pointer from the argument struct and return into that.
902 unsigned Idx = CurFnInfo->getReturnInfo().getInAllocaFieldIndex();
903 llvm::Function::arg_iterator EI = CurFn->arg_end();
905 llvm::Value *Addr = Builder.CreateStructGEP(nullptr, &*EI, Idx);
906 Addr = Builder.CreateAlignedLoad(Addr, getPointerAlign(), "agg.result");
907 ReturnValue = Address(Addr, getNaturalTypeAlignment(RetTy));
909 ReturnValue = CreateIRTemp(RetTy, "retval");
911 // Tell the epilog emitter to autorelease the result. We do this
912 // now so that various specialized functions can suppress it
913 // during their IR-generation.
914 if (getLangOpts().ObjCAutoRefCount &&
915 !CurFnInfo->isReturnsRetained() &&
916 RetTy->isObjCRetainableType())
917 AutoreleaseResult = true;
920 EmitStartEHSpec(CurCodeDecl);
922 PrologueCleanupDepth = EHStack.stable_begin();
923 EmitFunctionProlog(*CurFnInfo, CurFn, Args);
925 if (D && isa<CXXMethodDecl>(D) && cast<CXXMethodDecl>(D)->isInstance()) {
926 CGM.getCXXABI().EmitInstanceFunctionProlog(*this);
927 const CXXMethodDecl *MD = cast<CXXMethodDecl>(D);
928 if (MD->getParent()->isLambda() &&
929 MD->getOverloadedOperator() == OO_Call) {
930 // We're in a lambda; figure out the captures.
931 MD->getParent()->getCaptureFields(LambdaCaptureFields,
932 LambdaThisCaptureField);
933 if (LambdaThisCaptureField) {
934 // If the lambda captures the object referred to by '*this' - either by
935 // value or by reference, make sure CXXThisValue points to the correct
938 // Get the lvalue for the field (which is a copy of the enclosing object
939 // or contains the address of the enclosing object).
940 LValue ThisFieldLValue = EmitLValueForLambdaField(LambdaThisCaptureField);
941 if (!LambdaThisCaptureField->getType()->isPointerType()) {
942 // If the enclosing object was captured by value, just use its address.
943 CXXThisValue = ThisFieldLValue.getAddress().getPointer();
945 // Load the lvalue pointed to by the field, since '*this' was captured
948 EmitLoadOfLValue(ThisFieldLValue, SourceLocation()).getScalarVal();
951 for (auto *FD : MD->getParent()->fields()) {
952 if (FD->hasCapturedVLAType()) {
953 auto *ExprArg = EmitLoadOfLValue(EmitLValueForLambdaField(FD),
954 SourceLocation()).getScalarVal();
955 auto VAT = FD->getCapturedVLAType();
956 VLASizeMap[VAT->getSizeExpr()] = ExprArg;
960 // Not in a lambda; just use 'this' from the method.
961 // FIXME: Should we generate a new load for each use of 'this'? The
962 // fast register allocator would be happier...
963 CXXThisValue = CXXABIThisValue;
966 // Check the 'this' pointer once per function, if it's available.
968 SanitizerSet SkippedChecks;
969 SkippedChecks.set(SanitizerKind::ObjectSize, true);
970 QualType ThisTy = MD->getThisType(getContext());
971 EmitTypeCheck(TCK_Load, Loc, CXXThisValue, ThisTy,
972 getContext().getTypeAlignInChars(ThisTy->getPointeeType()),
977 // If any of the arguments have a variably modified type, make sure to
978 // emit the type size.
979 for (FunctionArgList::const_iterator i = Args.begin(), e = Args.end();
981 const VarDecl *VD = *i;
983 // Dig out the type as written from ParmVarDecls; it's unclear whether
984 // the standard (C99 6.9.1p10) requires this, but we're following the
985 // precedent set by gcc.
987 if (const ParmVarDecl *PVD = dyn_cast<ParmVarDecl>(VD))
988 Ty = PVD->getOriginalType();
992 if (Ty->isVariablyModifiedType())
993 EmitVariablyModifiedType(Ty);
995 // Emit a location at the end of the prologue.
996 if (CGDebugInfo *DI = getDebugInfo())
997 DI->EmitLocation(Builder, StartLoc);
1000 void CodeGenFunction::EmitFunctionBody(FunctionArgList &Args,
1002 incrementProfileCounter(Body);
1003 if (const CompoundStmt *S = dyn_cast<CompoundStmt>(Body))
1004 EmitCompoundStmtWithoutScope(*S);
1009 /// When instrumenting to collect profile data, the counts for some blocks
1010 /// such as switch cases need to not include the fall-through counts, so
1011 /// emit a branch around the instrumentation code. When not instrumenting,
1012 /// this just calls EmitBlock().
1013 void CodeGenFunction::EmitBlockWithFallThrough(llvm::BasicBlock *BB,
1015 llvm::BasicBlock *SkipCountBB = nullptr;
1016 if (HaveInsertPoint() && CGM.getCodeGenOpts().hasProfileClangInstr()) {
1017 // When instrumenting for profiling, the fallthrough to certain
1018 // statements needs to skip over the instrumentation code so that we
1019 // get an accurate count.
1020 SkipCountBB = createBasicBlock("skipcount");
1021 EmitBranch(SkipCountBB);
1024 uint64_t CurrentCount = getCurrentProfileCount();
1025 incrementProfileCounter(S);
1026 setCurrentProfileCount(getCurrentProfileCount() + CurrentCount);
1028 EmitBlock(SkipCountBB);
1031 /// Tries to mark the given function nounwind based on the
1032 /// non-existence of any throwing calls within it. We believe this is
1033 /// lightweight enough to do at -O0.
1034 static void TryMarkNoThrow(llvm::Function *F) {
1035 // LLVM treats 'nounwind' on a function as part of the type, so we
1036 // can't do this on functions that can be overwritten.
1037 if (F->isInterposable()) return;
1039 for (llvm::BasicBlock &BB : *F)
1040 for (llvm::Instruction &I : BB)
1044 F->setDoesNotThrow();
1047 QualType CodeGenFunction::BuildFunctionArgList(GlobalDecl GD,
1048 FunctionArgList &Args) {
1049 const FunctionDecl *FD = cast<FunctionDecl>(GD.getDecl());
1050 QualType ResTy = FD->getReturnType();
1052 const CXXMethodDecl *MD = dyn_cast<CXXMethodDecl>(FD);
1053 if (MD && MD->isInstance()) {
1054 if (CGM.getCXXABI().HasThisReturn(GD))
1055 ResTy = MD->getThisType(getContext());
1056 else if (CGM.getCXXABI().hasMostDerivedReturn(GD))
1057 ResTy = CGM.getContext().VoidPtrTy;
1058 CGM.getCXXABI().buildThisParam(*this, Args);
1061 // The base version of an inheriting constructor whose constructed base is a
1062 // virtual base is not passed any arguments (because it doesn't actually call
1063 // the inherited constructor).
1064 bool PassedParams = true;
1065 if (const CXXConstructorDecl *CD = dyn_cast<CXXConstructorDecl>(FD))
1066 if (auto Inherited = CD->getInheritedConstructor())
1068 getTypes().inheritingCtorHasParams(Inherited, GD.getCtorType());
1071 for (auto *Param : FD->parameters()) {
1072 Args.push_back(Param);
1073 if (!Param->hasAttr<PassObjectSizeAttr>())
1076 IdentifierInfo *NoID = nullptr;
1077 auto *Implicit = ImplicitParamDecl::Create(
1078 getContext(), Param->getDeclContext(), Param->getLocation(), NoID,
1079 getContext().getSizeType());
1080 SizeArguments[Param] = Implicit;
1081 Args.push_back(Implicit);
1085 if (MD && (isa<CXXConstructorDecl>(MD) || isa<CXXDestructorDecl>(MD)))
1086 CGM.getCXXABI().addImplicitStructorParams(*this, ResTy, Args);
1092 shouldUseUndefinedBehaviorReturnOptimization(const FunctionDecl *FD,
1093 const ASTContext &Context) {
1094 QualType T = FD->getReturnType();
1095 // Avoid the optimization for functions that return a record type with a
1096 // trivial destructor or another trivially copyable type.
1097 if (const RecordType *RT = T.getCanonicalType()->getAs<RecordType>()) {
1098 if (const auto *ClassDecl = dyn_cast<CXXRecordDecl>(RT->getDecl()))
1099 return !ClassDecl->hasTrivialDestructor();
1101 return !T.isTriviallyCopyableType(Context);
1104 void CodeGenFunction::GenerateCode(GlobalDecl GD, llvm::Function *Fn,
1105 const CGFunctionInfo &FnInfo) {
1106 const FunctionDecl *FD = cast<FunctionDecl>(GD.getDecl());
1109 FunctionArgList Args;
1110 QualType ResTy = BuildFunctionArgList(GD, Args);
1112 // Check if we should generate debug info for this function.
1113 if (FD->hasAttr<NoDebugAttr>())
1114 DebugInfo = nullptr; // disable debug info indefinitely for this function
1116 // The function might not have a body if we're generating thunks for a
1117 // function declaration.
1118 SourceRange BodyRange;
1119 if (Stmt *Body = FD->getBody())
1120 BodyRange = Body->getSourceRange();
1122 BodyRange = FD->getLocation();
1123 CurEHLocation = BodyRange.getEnd();
1125 // Use the location of the start of the function to determine where
1126 // the function definition is located. By default use the location
1127 // of the declaration as the location for the subprogram. A function
1128 // may lack a declaration in the source code if it is created by code
1129 // gen. (examples: _GLOBAL__I_a, __cxx_global_array_dtor, thunk).
1130 SourceLocation Loc = FD->getLocation();
1132 // If this is a function specialization then use the pattern body
1133 // as the location for the function.
1134 if (const FunctionDecl *SpecDecl = FD->getTemplateInstantiationPattern())
1135 if (SpecDecl->hasBody(SpecDecl))
1136 Loc = SpecDecl->getLocation();
1138 Stmt *Body = FD->getBody();
1140 // Initialize helper which will detect jumps which can cause invalid lifetime
1142 if (Body && ShouldEmitLifetimeMarkers)
1143 Bypasses.Init(Body);
1145 // Emit the standard function prologue.
1146 StartFunction(GD, ResTy, Fn, FnInfo, Args, Loc, BodyRange.getBegin());
1148 // Generate the body of the function.
1149 PGO.assignRegionCounters(GD, CurFn);
1150 if (isa<CXXDestructorDecl>(FD))
1151 EmitDestructorBody(Args);
1152 else if (isa<CXXConstructorDecl>(FD))
1153 EmitConstructorBody(Args);
1154 else if (getLangOpts().CUDA &&
1155 !getLangOpts().CUDAIsDevice &&
1156 FD->hasAttr<CUDAGlobalAttr>())
1157 CGM.getCUDARuntime().emitDeviceStub(*this, Args);
1158 else if (isa<CXXConversionDecl>(FD) &&
1159 cast<CXXConversionDecl>(FD)->isLambdaToBlockPointerConversion()) {
1160 // The lambda conversion to block pointer is special; the semantics can't be
1161 // expressed in the AST, so IRGen needs to special-case it.
1162 EmitLambdaToBlockPointerBody(Args);
1163 } else if (isa<CXXMethodDecl>(FD) &&
1164 cast<CXXMethodDecl>(FD)->isLambdaStaticInvoker()) {
1165 // The lambda static invoker function is special, because it forwards or
1166 // clones the body of the function call operator (but is actually static).
1167 EmitLambdaStaticInvokeFunction(cast<CXXMethodDecl>(FD));
1168 } else if (FD->isDefaulted() && isa<CXXMethodDecl>(FD) &&
1169 (cast<CXXMethodDecl>(FD)->isCopyAssignmentOperator() ||
1170 cast<CXXMethodDecl>(FD)->isMoveAssignmentOperator())) {
1171 // Implicit copy-assignment gets the same special treatment as implicit
1172 // copy-constructors.
1173 emitImplicitAssignmentOperatorBody(Args);
1175 EmitFunctionBody(Args, Body);
1177 llvm_unreachable("no definition for emitted function");
1179 // C++11 [stmt.return]p2:
1180 // Flowing off the end of a function [...] results in undefined behavior in
1181 // a value-returning function.
1183 // If the '}' that terminates a function is reached, and the value of the
1184 // function call is used by the caller, the behavior is undefined.
1185 if (getLangOpts().CPlusPlus && !FD->hasImplicitReturnZero() && !SawAsmBlock &&
1186 !FD->getReturnType()->isVoidType() && Builder.GetInsertBlock()) {
1187 bool ShouldEmitUnreachable =
1188 CGM.getCodeGenOpts().StrictReturn ||
1189 shouldUseUndefinedBehaviorReturnOptimization(FD, getContext());
1190 if (SanOpts.has(SanitizerKind::Return)) {
1191 SanitizerScope SanScope(this);
1192 llvm::Value *IsFalse = Builder.getFalse();
1193 EmitCheck(std::make_pair(IsFalse, SanitizerKind::Return),
1194 SanitizerHandler::MissingReturn,
1195 EmitCheckSourceLocation(FD->getLocation()), None);
1196 } else if (ShouldEmitUnreachable) {
1197 if (CGM.getCodeGenOpts().OptimizationLevel == 0)
1198 EmitTrapCall(llvm::Intrinsic::trap);
1200 if (SanOpts.has(SanitizerKind::Return) || ShouldEmitUnreachable) {
1201 Builder.CreateUnreachable();
1202 Builder.ClearInsertionPoint();
1206 // Emit the standard function epilogue.
1207 FinishFunction(BodyRange.getEnd());
1209 // If we haven't marked the function nothrow through other means, do
1210 // a quick pass now to see if we can.
1211 if (!CurFn->doesNotThrow())
1212 TryMarkNoThrow(CurFn);
1215 /// ContainsLabel - Return true if the statement contains a label in it. If
1216 /// this statement is not executed normally, it not containing a label means
1217 /// that we can just remove the code.
1218 bool CodeGenFunction::ContainsLabel(const Stmt *S, bool IgnoreCaseStmts) {
1219 // Null statement, not a label!
1220 if (!S) return false;
1222 // If this is a label, we have to emit the code, consider something like:
1223 // if (0) { ... foo: bar(); } goto foo;
1225 // TODO: If anyone cared, we could track __label__'s, since we know that you
1226 // can't jump to one from outside their declared region.
1227 if (isa<LabelStmt>(S))
1230 // If this is a case/default statement, and we haven't seen a switch, we have
1231 // to emit the code.
1232 if (isa<SwitchCase>(S) && !IgnoreCaseStmts)
1235 // If this is a switch statement, we want to ignore cases below it.
1236 if (isa<SwitchStmt>(S))
1237 IgnoreCaseStmts = true;
1239 // Scan subexpressions for verboten labels.
1240 for (const Stmt *SubStmt : S->children())
1241 if (ContainsLabel(SubStmt, IgnoreCaseStmts))
1247 /// containsBreak - Return true if the statement contains a break out of it.
1248 /// If the statement (recursively) contains a switch or loop with a break
1249 /// inside of it, this is fine.
1250 bool CodeGenFunction::containsBreak(const Stmt *S) {
1251 // Null statement, not a label!
1252 if (!S) return false;
1254 // If this is a switch or loop that defines its own break scope, then we can
1255 // include it and anything inside of it.
1256 if (isa<SwitchStmt>(S) || isa<WhileStmt>(S) || isa<DoStmt>(S) ||
1260 if (isa<BreakStmt>(S))
1263 // Scan subexpressions for verboten breaks.
1264 for (const Stmt *SubStmt : S->children())
1265 if (containsBreak(SubStmt))
1271 bool CodeGenFunction::mightAddDeclToScope(const Stmt *S) {
1272 if (!S) return false;
1274 // Some statement kinds add a scope and thus never add a decl to the current
1275 // scope. Note, this list is longer than the list of statements that might
1276 // have an unscoped decl nested within them, but this way is conservatively
1277 // correct even if more statement kinds are added.
1278 if (isa<IfStmt>(S) || isa<SwitchStmt>(S) || isa<WhileStmt>(S) ||
1279 isa<DoStmt>(S) || isa<ForStmt>(S) || isa<CompoundStmt>(S) ||
1280 isa<CXXForRangeStmt>(S) || isa<CXXTryStmt>(S) ||
1281 isa<ObjCForCollectionStmt>(S) || isa<ObjCAtTryStmt>(S))
1284 if (isa<DeclStmt>(S))
1287 for (const Stmt *SubStmt : S->children())
1288 if (mightAddDeclToScope(SubStmt))
1294 /// ConstantFoldsToSimpleInteger - If the specified expression does not fold
1295 /// to a constant, or if it does but contains a label, return false. If it
1296 /// constant folds return true and set the boolean result in Result.
1297 bool CodeGenFunction::ConstantFoldsToSimpleInteger(const Expr *Cond,
1300 llvm::APSInt ResultInt;
1301 if (!ConstantFoldsToSimpleInteger(Cond, ResultInt, AllowLabels))
1304 ResultBool = ResultInt.getBoolValue();
1308 /// ConstantFoldsToSimpleInteger - If the specified expression does not fold
1309 /// to a constant, or if it does but contains a label, return false. If it
1310 /// constant folds return true and set the folded value.
1311 bool CodeGenFunction::ConstantFoldsToSimpleInteger(const Expr *Cond,
1312 llvm::APSInt &ResultInt,
1314 // FIXME: Rename and handle conversion of other evaluatable things
1317 if (!Cond->EvaluateAsInt(Int, getContext()))
1318 return false; // Not foldable, not integer or not fully evaluatable.
1320 if (!AllowLabels && CodeGenFunction::ContainsLabel(Cond))
1321 return false; // Contains a label.
1329 /// EmitBranchOnBoolExpr - Emit a branch on a boolean condition (e.g. for an if
1330 /// statement) to the specified blocks. Based on the condition, this might try
1331 /// to simplify the codegen of the conditional based on the branch.
1333 void CodeGenFunction::EmitBranchOnBoolExpr(const Expr *Cond,
1334 llvm::BasicBlock *TrueBlock,
1335 llvm::BasicBlock *FalseBlock,
1336 uint64_t TrueCount) {
1337 Cond = Cond->IgnoreParens();
1339 if (const BinaryOperator *CondBOp = dyn_cast<BinaryOperator>(Cond)) {
1341 // Handle X && Y in a condition.
1342 if (CondBOp->getOpcode() == BO_LAnd) {
1343 // If we have "1 && X", simplify the code. "0 && X" would have constant
1344 // folded if the case was simple enough.
1345 bool ConstantBool = false;
1346 if (ConstantFoldsToSimpleInteger(CondBOp->getLHS(), ConstantBool) &&
1348 // br(1 && X) -> br(X).
1349 incrementProfileCounter(CondBOp);
1350 return EmitBranchOnBoolExpr(CondBOp->getRHS(), TrueBlock, FalseBlock,
1354 // If we have "X && 1", simplify the code to use an uncond branch.
1355 // "X && 0" would have been constant folded to 0.
1356 if (ConstantFoldsToSimpleInteger(CondBOp->getRHS(), ConstantBool) &&
1358 // br(X && 1) -> br(X).
1359 return EmitBranchOnBoolExpr(CondBOp->getLHS(), TrueBlock, FalseBlock,
1363 // Emit the LHS as a conditional. If the LHS conditional is false, we
1364 // want to jump to the FalseBlock.
1365 llvm::BasicBlock *LHSTrue = createBasicBlock("land.lhs.true");
1366 // The counter tells us how often we evaluate RHS, and all of TrueCount
1367 // can be propagated to that branch.
1368 uint64_t RHSCount = getProfileCount(CondBOp->getRHS());
1370 ConditionalEvaluation eval(*this);
1372 ApplyDebugLocation DL(*this, Cond);
1373 EmitBranchOnBoolExpr(CondBOp->getLHS(), LHSTrue, FalseBlock, RHSCount);
1377 incrementProfileCounter(CondBOp);
1378 setCurrentProfileCount(getProfileCount(CondBOp->getRHS()));
1380 // Any temporaries created here are conditional.
1382 EmitBranchOnBoolExpr(CondBOp->getRHS(), TrueBlock, FalseBlock, TrueCount);
1388 if (CondBOp->getOpcode() == BO_LOr) {
1389 // If we have "0 || X", simplify the code. "1 || X" would have constant
1390 // folded if the case was simple enough.
1391 bool ConstantBool = false;
1392 if (ConstantFoldsToSimpleInteger(CondBOp->getLHS(), ConstantBool) &&
1394 // br(0 || X) -> br(X).
1395 incrementProfileCounter(CondBOp);
1396 return EmitBranchOnBoolExpr(CondBOp->getRHS(), TrueBlock, FalseBlock,
1400 // If we have "X || 0", simplify the code to use an uncond branch.
1401 // "X || 1" would have been constant folded to 1.
1402 if (ConstantFoldsToSimpleInteger(CondBOp->getRHS(), ConstantBool) &&
1404 // br(X || 0) -> br(X).
1405 return EmitBranchOnBoolExpr(CondBOp->getLHS(), TrueBlock, FalseBlock,
1409 // Emit the LHS as a conditional. If the LHS conditional is true, we
1410 // want to jump to the TrueBlock.
1411 llvm::BasicBlock *LHSFalse = createBasicBlock("lor.lhs.false");
1412 // We have the count for entry to the RHS and for the whole expression
1413 // being true, so we can divy up True count between the short circuit and
1416 getCurrentProfileCount() - getProfileCount(CondBOp->getRHS());
1417 uint64_t RHSCount = TrueCount - LHSCount;
1419 ConditionalEvaluation eval(*this);
1421 ApplyDebugLocation DL(*this, Cond);
1422 EmitBranchOnBoolExpr(CondBOp->getLHS(), TrueBlock, LHSFalse, LHSCount);
1423 EmitBlock(LHSFalse);
1426 incrementProfileCounter(CondBOp);
1427 setCurrentProfileCount(getProfileCount(CondBOp->getRHS()));
1429 // Any temporaries created here are conditional.
1431 EmitBranchOnBoolExpr(CondBOp->getRHS(), TrueBlock, FalseBlock, RHSCount);
1439 if (const UnaryOperator *CondUOp = dyn_cast<UnaryOperator>(Cond)) {
1440 // br(!x, t, f) -> br(x, f, t)
1441 if (CondUOp->getOpcode() == UO_LNot) {
1442 // Negate the count.
1443 uint64_t FalseCount = getCurrentProfileCount() - TrueCount;
1444 // Negate the condition and swap the destination blocks.
1445 return EmitBranchOnBoolExpr(CondUOp->getSubExpr(), FalseBlock, TrueBlock,
1450 if (const ConditionalOperator *CondOp = dyn_cast<ConditionalOperator>(Cond)) {
1451 // br(c ? x : y, t, f) -> br(c, br(x, t, f), br(y, t, f))
1452 llvm::BasicBlock *LHSBlock = createBasicBlock("cond.true");
1453 llvm::BasicBlock *RHSBlock = createBasicBlock("cond.false");
1455 ConditionalEvaluation cond(*this);
1456 EmitBranchOnBoolExpr(CondOp->getCond(), LHSBlock, RHSBlock,
1457 getProfileCount(CondOp));
1459 // When computing PGO branch weights, we only know the overall count for
1460 // the true block. This code is essentially doing tail duplication of the
1461 // naive code-gen, introducing new edges for which counts are not
1462 // available. Divide the counts proportionally between the LHS and RHS of
1463 // the conditional operator.
1464 uint64_t LHSScaledTrueCount = 0;
1467 getProfileCount(CondOp) / (double)getCurrentProfileCount();
1468 LHSScaledTrueCount = TrueCount * LHSRatio;
1472 EmitBlock(LHSBlock);
1473 incrementProfileCounter(CondOp);
1475 ApplyDebugLocation DL(*this, Cond);
1476 EmitBranchOnBoolExpr(CondOp->getLHS(), TrueBlock, FalseBlock,
1477 LHSScaledTrueCount);
1482 EmitBlock(RHSBlock);
1483 EmitBranchOnBoolExpr(CondOp->getRHS(), TrueBlock, FalseBlock,
1484 TrueCount - LHSScaledTrueCount);
1490 if (const CXXThrowExpr *Throw = dyn_cast<CXXThrowExpr>(Cond)) {
1491 // Conditional operator handling can give us a throw expression as a
1492 // condition for a case like:
1493 // br(c ? throw x : y, t, f) -> br(c, br(throw x, t, f), br(y, t, f)
1495 // br(c, throw x, br(y, t, f))
1496 EmitCXXThrowExpr(Throw, /*KeepInsertionPoint*/false);
1500 // If the branch has a condition wrapped by __builtin_unpredictable,
1501 // create metadata that specifies that the branch is unpredictable.
1502 // Don't bother if not optimizing because that metadata would not be used.
1503 llvm::MDNode *Unpredictable = nullptr;
1504 auto *Call = dyn_cast<CallExpr>(Cond);
1505 if (Call && CGM.getCodeGenOpts().OptimizationLevel != 0) {
1506 auto *FD = dyn_cast_or_null<FunctionDecl>(Call->getCalleeDecl());
1507 if (FD && FD->getBuiltinID() == Builtin::BI__builtin_unpredictable) {
1508 llvm::MDBuilder MDHelper(getLLVMContext());
1509 Unpredictable = MDHelper.createUnpredictable();
1513 // Create branch weights based on the number of times we get here and the
1514 // number of times the condition should be true.
1515 uint64_t CurrentCount = std::max(getCurrentProfileCount(), TrueCount);
1516 llvm::MDNode *Weights =
1517 createProfileWeights(TrueCount, CurrentCount - TrueCount);
1519 // Emit the code with the fully general case.
1522 ApplyDebugLocation DL(*this, Cond);
1523 CondV = EvaluateExprAsBool(Cond);
1525 Builder.CreateCondBr(CondV, TrueBlock, FalseBlock, Weights, Unpredictable);
1528 /// ErrorUnsupported - Print out an error that codegen doesn't support the
1529 /// specified stmt yet.
1530 void CodeGenFunction::ErrorUnsupported(const Stmt *S, const char *Type) {
1531 CGM.ErrorUnsupported(S, Type);
1534 /// emitNonZeroVLAInit - Emit the "zero" initialization of a
1535 /// variable-length array whose elements have a non-zero bit-pattern.
1537 /// \param baseType the inner-most element type of the array
1538 /// \param src - a char* pointing to the bit-pattern for a single
1539 /// base element of the array
1540 /// \param sizeInChars - the total size of the VLA, in chars
1541 static void emitNonZeroVLAInit(CodeGenFunction &CGF, QualType baseType,
1542 Address dest, Address src,
1543 llvm::Value *sizeInChars) {
1544 CGBuilderTy &Builder = CGF.Builder;
1546 CharUnits baseSize = CGF.getContext().getTypeSizeInChars(baseType);
1547 llvm::Value *baseSizeInChars
1548 = llvm::ConstantInt::get(CGF.IntPtrTy, baseSize.getQuantity());
1551 Builder.CreateElementBitCast(dest, CGF.Int8Ty, "vla.begin");
1553 Builder.CreateInBoundsGEP(begin.getPointer(), sizeInChars, "vla.end");
1555 llvm::BasicBlock *originBB = CGF.Builder.GetInsertBlock();
1556 llvm::BasicBlock *loopBB = CGF.createBasicBlock("vla-init.loop");
1557 llvm::BasicBlock *contBB = CGF.createBasicBlock("vla-init.cont");
1559 // Make a loop over the VLA. C99 guarantees that the VLA element
1560 // count must be nonzero.
1561 CGF.EmitBlock(loopBB);
1563 llvm::PHINode *cur = Builder.CreatePHI(begin.getType(), 2, "vla.cur");
1564 cur->addIncoming(begin.getPointer(), originBB);
1566 CharUnits curAlign =
1567 dest.getAlignment().alignmentOfArrayElement(baseSize);
1569 // memcpy the individual element bit-pattern.
1570 Builder.CreateMemCpy(Address(cur, curAlign), src, baseSizeInChars,
1571 /*volatile*/ false);
1573 // Go to the next element.
1575 Builder.CreateInBoundsGEP(CGF.Int8Ty, cur, baseSizeInChars, "vla.next");
1577 // Leave if that's the end of the VLA.
1578 llvm::Value *done = Builder.CreateICmpEQ(next, end, "vla-init.isdone");
1579 Builder.CreateCondBr(done, contBB, loopBB);
1580 cur->addIncoming(next, loopBB);
1582 CGF.EmitBlock(contBB);
1586 CodeGenFunction::EmitNullInitialization(Address DestPtr, QualType Ty) {
1587 // Ignore empty classes in C++.
1588 if (getLangOpts().CPlusPlus) {
1589 if (const RecordType *RT = Ty->getAs<RecordType>()) {
1590 if (cast<CXXRecordDecl>(RT->getDecl())->isEmpty())
1595 // Cast the dest ptr to the appropriate i8 pointer type.
1596 if (DestPtr.getElementType() != Int8Ty)
1597 DestPtr = Builder.CreateElementBitCast(DestPtr, Int8Ty);
1599 // Get size and alignment info for this aggregate.
1600 CharUnits size = getContext().getTypeSizeInChars(Ty);
1602 llvm::Value *SizeVal;
1603 const VariableArrayType *vla;
1605 // Don't bother emitting a zero-byte memset.
1606 if (size.isZero()) {
1607 // But note that getTypeInfo returns 0 for a VLA.
1608 if (const VariableArrayType *vlaType =
1609 dyn_cast_or_null<VariableArrayType>(
1610 getContext().getAsArrayType(Ty))) {
1612 llvm::Value *numElts;
1613 std::tie(numElts, eltType) = getVLASize(vlaType);
1616 CharUnits eltSize = getContext().getTypeSizeInChars(eltType);
1617 if (!eltSize.isOne())
1618 SizeVal = Builder.CreateNUWMul(SizeVal, CGM.getSize(eltSize));
1624 SizeVal = CGM.getSize(size);
1628 // If the type contains a pointer to data member we can't memset it to zero.
1629 // Instead, create a null constant and copy it to the destination.
1630 // TODO: there are other patterns besides zero that we can usefully memset,
1631 // like -1, which happens to be the pattern used by member-pointers.
1632 if (!CGM.getTypes().isZeroInitializable(Ty)) {
1633 // For a VLA, emit a single element, then splat that over the VLA.
1634 if (vla) Ty = getContext().getBaseElementType(vla);
1636 llvm::Constant *NullConstant = CGM.EmitNullConstant(Ty);
1638 llvm::GlobalVariable *NullVariable =
1639 new llvm::GlobalVariable(CGM.getModule(), NullConstant->getType(),
1640 /*isConstant=*/true,
1641 llvm::GlobalVariable::PrivateLinkage,
1642 NullConstant, Twine());
1643 CharUnits NullAlign = DestPtr.getAlignment();
1644 NullVariable->setAlignment(NullAlign.getQuantity());
1645 Address SrcPtr(Builder.CreateBitCast(NullVariable, Builder.getInt8PtrTy()),
1648 if (vla) return emitNonZeroVLAInit(*this, Ty, DestPtr, SrcPtr, SizeVal);
1650 // Get and call the appropriate llvm.memcpy overload.
1651 Builder.CreateMemCpy(DestPtr, SrcPtr, SizeVal, false);
1655 // Otherwise, just memset the whole thing to zero. This is legal
1656 // because in LLVM, all default initializers (other than the ones we just
1657 // handled above) are guaranteed to have a bit pattern of all zeros.
1658 Builder.CreateMemSet(DestPtr, Builder.getInt8(0), SizeVal, false);
1661 llvm::BlockAddress *CodeGenFunction::GetAddrOfLabel(const LabelDecl *L) {
1662 // Make sure that there is a block for the indirect goto.
1663 if (!IndirectBranch)
1664 GetIndirectGotoBlock();
1666 llvm::BasicBlock *BB = getJumpDestForLabel(L).getBlock();
1668 // Make sure the indirect branch includes all of the address-taken blocks.
1669 IndirectBranch->addDestination(BB);
1670 return llvm::BlockAddress::get(CurFn, BB);
1673 llvm::BasicBlock *CodeGenFunction::GetIndirectGotoBlock() {
1674 // If we already made the indirect branch for indirect goto, return its block.
1675 if (IndirectBranch) return IndirectBranch->getParent();
1677 CGBuilderTy TmpBuilder(*this, createBasicBlock("indirectgoto"));
1679 // Create the PHI node that indirect gotos will add entries to.
1680 llvm::Value *DestVal = TmpBuilder.CreatePHI(Int8PtrTy, 0,
1681 "indirect.goto.dest");
1683 // Create the indirect branch instruction.
1684 IndirectBranch = TmpBuilder.CreateIndirectBr(DestVal);
1685 return IndirectBranch->getParent();
1688 /// Computes the length of an array in elements, as well as the base
1689 /// element type and a properly-typed first element pointer.
1690 llvm::Value *CodeGenFunction::emitArrayLength(const ArrayType *origArrayType,
1693 const ArrayType *arrayType = origArrayType;
1695 // If it's a VLA, we have to load the stored size. Note that
1696 // this is the size of the VLA in bytes, not its size in elements.
1697 llvm::Value *numVLAElements = nullptr;
1698 if (isa<VariableArrayType>(arrayType)) {
1699 numVLAElements = getVLASize(cast<VariableArrayType>(arrayType)).first;
1701 // Walk into all VLAs. This doesn't require changes to addr,
1702 // which has type T* where T is the first non-VLA element type.
1704 QualType elementType = arrayType->getElementType();
1705 arrayType = getContext().getAsArrayType(elementType);
1707 // If we only have VLA components, 'addr' requires no adjustment.
1709 baseType = elementType;
1710 return numVLAElements;
1712 } while (isa<VariableArrayType>(arrayType));
1714 // We get out here only if we find a constant array type
1718 // We have some number of constant-length arrays, so addr should
1719 // have LLVM type [M x [N x [...]]]*. Build a GEP that walks
1720 // down to the first element of addr.
1721 SmallVector<llvm::Value*, 8> gepIndices;
1723 // GEP down to the array type.
1724 llvm::ConstantInt *zero = Builder.getInt32(0);
1725 gepIndices.push_back(zero);
1727 uint64_t countFromCLAs = 1;
1730 llvm::ArrayType *llvmArrayType =
1731 dyn_cast<llvm::ArrayType>(addr.getElementType());
1732 while (llvmArrayType) {
1733 assert(isa<ConstantArrayType>(arrayType));
1734 assert(cast<ConstantArrayType>(arrayType)->getSize().getZExtValue()
1735 == llvmArrayType->getNumElements());
1737 gepIndices.push_back(zero);
1738 countFromCLAs *= llvmArrayType->getNumElements();
1739 eltType = arrayType->getElementType();
1742 dyn_cast<llvm::ArrayType>(llvmArrayType->getElementType());
1743 arrayType = getContext().getAsArrayType(arrayType->getElementType());
1744 assert((!llvmArrayType || arrayType) &&
1745 "LLVM and Clang types are out-of-synch");
1749 // From this point onwards, the Clang array type has been emitted
1750 // as some other type (probably a packed struct). Compute the array
1751 // size, and just emit the 'begin' expression as a bitcast.
1754 cast<ConstantArrayType>(arrayType)->getSize().getZExtValue();
1755 eltType = arrayType->getElementType();
1756 arrayType = getContext().getAsArrayType(eltType);
1759 llvm::Type *baseType = ConvertType(eltType);
1760 addr = Builder.CreateElementBitCast(addr, baseType, "array.begin");
1762 // Create the actual GEP.
1763 addr = Address(Builder.CreateInBoundsGEP(addr.getPointer(),
1764 gepIndices, "array.begin"),
1765 addr.getAlignment());
1770 llvm::Value *numElements
1771 = llvm::ConstantInt::get(SizeTy, countFromCLAs);
1773 // If we had any VLA dimensions, factor them in.
1775 numElements = Builder.CreateNUWMul(numVLAElements, numElements);
1780 std::pair<llvm::Value*, QualType>
1781 CodeGenFunction::getVLASize(QualType type) {
1782 const VariableArrayType *vla = getContext().getAsVariableArrayType(type);
1783 assert(vla && "type was not a variable array type!");
1784 return getVLASize(vla);
1787 std::pair<llvm::Value*, QualType>
1788 CodeGenFunction::getVLASize(const VariableArrayType *type) {
1789 // The number of elements so far; always size_t.
1790 llvm::Value *numElements = nullptr;
1792 QualType elementType;
1794 elementType = type->getElementType();
1795 llvm::Value *vlaSize = VLASizeMap[type->getSizeExpr()];
1796 assert(vlaSize && "no size for VLA!");
1797 assert(vlaSize->getType() == SizeTy);
1800 numElements = vlaSize;
1802 // It's undefined behavior if this wraps around, so mark it that way.
1803 // FIXME: Teach -fsanitize=undefined to trap this.
1804 numElements = Builder.CreateNUWMul(numElements, vlaSize);
1806 } while ((type = getContext().getAsVariableArrayType(elementType)));
1808 return std::pair<llvm::Value*,QualType>(numElements, elementType);
1811 void CodeGenFunction::EmitVariablyModifiedType(QualType type) {
1812 assert(type->isVariablyModifiedType() &&
1813 "Must pass variably modified type to EmitVLASizes!");
1815 EnsureInsertPoint();
1817 // We're going to walk down into the type and look for VLA
1820 assert(type->isVariablyModifiedType());
1822 const Type *ty = type.getTypePtr();
1823 switch (ty->getTypeClass()) {
1825 #define TYPE(Class, Base)
1826 #define ABSTRACT_TYPE(Class, Base)
1827 #define NON_CANONICAL_TYPE(Class, Base)
1828 #define DEPENDENT_TYPE(Class, Base) case Type::Class:
1829 #define NON_CANONICAL_UNLESS_DEPENDENT_TYPE(Class, Base)
1830 #include "clang/AST/TypeNodes.def"
1831 llvm_unreachable("unexpected dependent type!");
1833 // These types are never variably-modified.
1837 case Type::ExtVector:
1840 case Type::Elaborated:
1841 case Type::TemplateSpecialization:
1842 case Type::ObjCTypeParam:
1843 case Type::ObjCObject:
1844 case Type::ObjCInterface:
1845 case Type::ObjCObjectPointer:
1846 llvm_unreachable("type class is never variably-modified!");
1848 case Type::Adjusted:
1849 type = cast<AdjustedType>(ty)->getAdjustedType();
1853 type = cast<DecayedType>(ty)->getPointeeType();
1857 type = cast<PointerType>(ty)->getPointeeType();
1860 case Type::BlockPointer:
1861 type = cast<BlockPointerType>(ty)->getPointeeType();
1864 case Type::LValueReference:
1865 case Type::RValueReference:
1866 type = cast<ReferenceType>(ty)->getPointeeType();
1869 case Type::MemberPointer:
1870 type = cast<MemberPointerType>(ty)->getPointeeType();
1873 case Type::ConstantArray:
1874 case Type::IncompleteArray:
1875 // Losing element qualification here is fine.
1876 type = cast<ArrayType>(ty)->getElementType();
1879 case Type::VariableArray: {
1880 // Losing element qualification here is fine.
1881 const VariableArrayType *vat = cast<VariableArrayType>(ty);
1883 // Unknown size indication requires no size computation.
1884 // Otherwise, evaluate and record it.
1885 if (const Expr *size = vat->getSizeExpr()) {
1886 // It's possible that we might have emitted this already,
1887 // e.g. with a typedef and a pointer to it.
1888 llvm::Value *&entry = VLASizeMap[size];
1890 llvm::Value *Size = EmitScalarExpr(size);
1893 // If the size is an expression that is not an integer constant
1894 // expression [...] each time it is evaluated it shall have a value
1895 // greater than zero.
1896 if (SanOpts.has(SanitizerKind::VLABound) &&
1897 size->getType()->isSignedIntegerType()) {
1898 SanitizerScope SanScope(this);
1899 llvm::Value *Zero = llvm::Constant::getNullValue(Size->getType());
1900 llvm::Constant *StaticArgs[] = {
1901 EmitCheckSourceLocation(size->getLocStart()),
1902 EmitCheckTypeDescriptor(size->getType())
1904 EmitCheck(std::make_pair(Builder.CreateICmpSGT(Size, Zero),
1905 SanitizerKind::VLABound),
1906 SanitizerHandler::VLABoundNotPositive, StaticArgs, Size);
1909 // Always zexting here would be wrong if it weren't
1910 // undefined behavior to have a negative bound.
1911 entry = Builder.CreateIntCast(Size, SizeTy, /*signed*/ false);
1914 type = vat->getElementType();
1918 case Type::FunctionProto:
1919 case Type::FunctionNoProto:
1920 type = cast<FunctionType>(ty)->getReturnType();
1925 case Type::UnaryTransform:
1926 case Type::Attributed:
1927 case Type::SubstTemplateTypeParm:
1928 case Type::PackExpansion:
1929 // Keep walking after single level desugaring.
1930 type = type.getSingleStepDesugaredType(getContext());
1934 case Type::Decltype:
1936 case Type::DeducedTemplateSpecialization:
1937 // Stop walking: nothing to do.
1940 case Type::TypeOfExpr:
1941 // Stop walking: emit typeof expression.
1942 EmitIgnoredExpr(cast<TypeOfExprType>(ty)->getUnderlyingExpr());
1946 type = cast<AtomicType>(ty)->getValueType();
1950 type = cast<PipeType>(ty)->getElementType();
1953 } while (type->isVariablyModifiedType());
1956 Address CodeGenFunction::EmitVAListRef(const Expr* E) {
1957 if (getContext().getBuiltinVaListType()->isArrayType())
1958 return EmitPointerWithAlignment(E);
1959 return EmitLValue(E).getAddress();
1962 Address CodeGenFunction::EmitMSVAListRef(const Expr *E) {
1963 return EmitLValue(E).getAddress();
1966 void CodeGenFunction::EmitDeclRefExprDbgValue(const DeclRefExpr *E,
1967 const APValue &Init) {
1968 assert(!Init.isUninit() && "Invalid DeclRefExpr initializer!");
1969 if (CGDebugInfo *Dbg = getDebugInfo())
1970 if (CGM.getCodeGenOpts().getDebugInfo() >= codegenoptions::LimitedDebugInfo)
1971 Dbg->EmitGlobalVariable(E->getDecl(), Init);
1974 CodeGenFunction::PeepholeProtection
1975 CodeGenFunction::protectFromPeepholes(RValue rvalue) {
1976 // At the moment, the only aggressive peephole we do in IR gen
1977 // is trunc(zext) folding, but if we add more, we can easily
1978 // extend this protection.
1980 if (!rvalue.isScalar()) return PeepholeProtection();
1981 llvm::Value *value = rvalue.getScalarVal();
1982 if (!isa<llvm::ZExtInst>(value)) return PeepholeProtection();
1984 // Just make an extra bitcast.
1985 assert(HaveInsertPoint());
1986 llvm::Instruction *inst = new llvm::BitCastInst(value, value->getType(), "",
1987 Builder.GetInsertBlock());
1989 PeepholeProtection protection;
1990 protection.Inst = inst;
1994 void CodeGenFunction::unprotectFromPeepholes(PeepholeProtection protection) {
1995 if (!protection.Inst) return;
1997 // In theory, we could try to duplicate the peepholes now, but whatever.
1998 protection.Inst->eraseFromParent();
2001 llvm::Value *CodeGenFunction::EmitAnnotationCall(llvm::Value *AnnotationFn,
2002 llvm::Value *AnnotatedVal,
2003 StringRef AnnotationStr,
2004 SourceLocation Location) {
2005 llvm::Value *Args[4] = {
2007 Builder.CreateBitCast(CGM.EmitAnnotationString(AnnotationStr), Int8PtrTy),
2008 Builder.CreateBitCast(CGM.EmitAnnotationUnit(Location), Int8PtrTy),
2009 CGM.EmitAnnotationLineNo(Location)
2011 return Builder.CreateCall(AnnotationFn, Args);
2014 void CodeGenFunction::EmitVarAnnotations(const VarDecl *D, llvm::Value *V) {
2015 assert(D->hasAttr<AnnotateAttr>() && "no annotate attribute");
2016 // FIXME We create a new bitcast for every annotation because that's what
2017 // llvm-gcc was doing.
2018 for (const auto *I : D->specific_attrs<AnnotateAttr>())
2019 EmitAnnotationCall(CGM.getIntrinsic(llvm::Intrinsic::var_annotation),
2020 Builder.CreateBitCast(V, CGM.Int8PtrTy, V->getName()),
2021 I->getAnnotation(), D->getLocation());
2024 Address CodeGenFunction::EmitFieldAnnotations(const FieldDecl *D,
2026 assert(D->hasAttr<AnnotateAttr>() && "no annotate attribute");
2027 llvm::Value *V = Addr.getPointer();
2028 llvm::Type *VTy = V->getType();
2029 llvm::Value *F = CGM.getIntrinsic(llvm::Intrinsic::ptr_annotation,
2032 for (const auto *I : D->specific_attrs<AnnotateAttr>()) {
2033 // FIXME Always emit the cast inst so we can differentiate between
2034 // annotation on the first field of a struct and annotation on the struct
2036 if (VTy != CGM.Int8PtrTy)
2037 V = Builder.Insert(new llvm::BitCastInst(V, CGM.Int8PtrTy));
2038 V = EmitAnnotationCall(F, V, I->getAnnotation(), D->getLocation());
2039 V = Builder.CreateBitCast(V, VTy);
2042 return Address(V, Addr.getAlignment());
2045 CodeGenFunction::CGCapturedStmtInfo::~CGCapturedStmtInfo() { }
2047 CodeGenFunction::SanitizerScope::SanitizerScope(CodeGenFunction *CGF)
2049 assert(!CGF->IsSanitizerScope);
2050 CGF->IsSanitizerScope = true;
2053 CodeGenFunction::SanitizerScope::~SanitizerScope() {
2054 CGF->IsSanitizerScope = false;
2057 void CodeGenFunction::InsertHelper(llvm::Instruction *I,
2058 const llvm::Twine &Name,
2059 llvm::BasicBlock *BB,
2060 llvm::BasicBlock::iterator InsertPt) const {
2061 LoopStack.InsertHelper(I);
2062 if (IsSanitizerScope)
2063 CGM.getSanitizerMetadata()->disableSanitizerForInstruction(I);
2066 void CGBuilderInserter::InsertHelper(
2067 llvm::Instruction *I, const llvm::Twine &Name, llvm::BasicBlock *BB,
2068 llvm::BasicBlock::iterator InsertPt) const {
2069 llvm::IRBuilderDefaultInserter::InsertHelper(I, Name, BB, InsertPt);
2071 CGF->InsertHelper(I, Name, BB, InsertPt);
2074 static bool hasRequiredFeatures(const SmallVectorImpl<StringRef> &ReqFeatures,
2075 CodeGenModule &CGM, const FunctionDecl *FD,
2076 std::string &FirstMissing) {
2077 // If there aren't any required features listed then go ahead and return.
2078 if (ReqFeatures.empty())
2081 // Now build up the set of caller features and verify that all the required
2082 // features are there.
2083 llvm::StringMap<bool> CallerFeatureMap;
2084 CGM.getFunctionFeatureMap(CallerFeatureMap, FD);
2086 // If we have at least one of the features in the feature list return
2087 // true, otherwise return false.
2089 ReqFeatures.begin(), ReqFeatures.end(), [&](StringRef Feature) {
2090 SmallVector<StringRef, 1> OrFeatures;
2091 Feature.split(OrFeatures, "|");
2092 return std::any_of(OrFeatures.begin(), OrFeatures.end(),
2093 [&](StringRef Feature) {
2094 if (!CallerFeatureMap.lookup(Feature)) {
2095 FirstMissing = Feature.str();
2103 // Emits an error if we don't have a valid set of target features for the
2105 void CodeGenFunction::checkTargetFeatures(const CallExpr *E,
2106 const FunctionDecl *TargetDecl) {
2107 // Early exit if this is an indirect call.
2111 // Get the current enclosing function if it exists. If it doesn't
2112 // we can't check the target features anyhow.
2113 const FunctionDecl *FD = dyn_cast_or_null<FunctionDecl>(CurFuncDecl);
2117 // Grab the required features for the call. For a builtin this is listed in
2118 // the td file with the default cpu, for an always_inline function this is any
2119 // listed cpu and any listed features.
2120 unsigned BuiltinID = TargetDecl->getBuiltinID();
2121 std::string MissingFeature;
2123 SmallVector<StringRef, 1> ReqFeatures;
2124 const char *FeatureList =
2125 CGM.getContext().BuiltinInfo.getRequiredFeatures(BuiltinID);
2126 // Return if the builtin doesn't have any required features.
2127 if (!FeatureList || StringRef(FeatureList) == "")
2129 StringRef(FeatureList).split(ReqFeatures, ",");
2130 if (!hasRequiredFeatures(ReqFeatures, CGM, FD, MissingFeature))
2131 CGM.getDiags().Report(E->getLocStart(), diag::err_builtin_needs_feature)
2132 << TargetDecl->getDeclName()
2133 << CGM.getContext().BuiltinInfo.getRequiredFeatures(BuiltinID);
2135 } else if (TargetDecl->hasAttr<TargetAttr>()) {
2136 // Get the required features for the callee.
2137 SmallVector<StringRef, 1> ReqFeatures;
2138 llvm::StringMap<bool> CalleeFeatureMap;
2139 CGM.getFunctionFeatureMap(CalleeFeatureMap, TargetDecl);
2140 for (const auto &F : CalleeFeatureMap) {
2141 // Only positive features are "required".
2143 ReqFeatures.push_back(F.getKey());
2145 if (!hasRequiredFeatures(ReqFeatures, CGM, FD, MissingFeature))
2146 CGM.getDiags().Report(E->getLocStart(), diag::err_function_needs_feature)
2147 << FD->getDeclName() << TargetDecl->getDeclName() << MissingFeature;
2151 void CodeGenFunction::EmitSanitizerStatReport(llvm::SanitizerStatKind SSK) {
2152 if (!CGM.getCodeGenOpts().SanitizeStats)
2155 llvm::IRBuilder<> IRB(Builder.GetInsertBlock(), Builder.GetInsertPoint());
2156 IRB.SetCurrentDebugLocation(Builder.getCurrentDebugLocation());
2157 CGM.getSanStats().create(IRB, SSK);
2160 llvm::DebugLoc CodeGenFunction::SourceLocToDebugLoc(SourceLocation Location) {
2161 if (CGDebugInfo *DI = getDebugInfo())
2162 return DI->SourceLocToDebugLoc(Location);
2164 return llvm::DebugLoc();