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/ASTLambda.h"
26 #include "clang/AST/Decl.h"
27 #include "clang/AST/DeclCXX.h"
28 #include "clang/AST/StmtCXX.h"
29 #include "clang/AST/StmtObjC.h"
30 #include "clang/Basic/Builtins.h"
31 #include "clang/Basic/TargetInfo.h"
32 #include "clang/CodeGen/CGFunctionInfo.h"
33 #include "clang/Frontend/CodeGenOptions.h"
34 #include "clang/Sema/SemaDiagnostic.h"
35 #include "llvm/IR/DataLayout.h"
36 #include "llvm/IR/Intrinsics.h"
37 #include "llvm/IR/MDBuilder.h"
38 #include "llvm/IR/Operator.h"
39 using namespace clang;
40 using namespace CodeGen;
42 /// shouldEmitLifetimeMarkers - Decide whether we need emit the life-time
44 static bool shouldEmitLifetimeMarkers(const CodeGenOptions &CGOpts,
45 const LangOptions &LangOpts) {
46 if (CGOpts.DisableLifetimeMarkers)
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 // Asan uses markers for use-after-scope checks.
55 if (CGOpts.SanitizeAddressUseAfterScope)
58 // For now, only in optimized builds.
59 return CGOpts.OptimizationLevel != 0;
62 CodeGenFunction::CodeGenFunction(CodeGenModule &cgm, bool suppressNewContext)
63 : CodeGenTypeCache(cgm), CGM(cgm), Target(cgm.getTarget()),
64 Builder(cgm, cgm.getModule().getContext(), llvm::ConstantFolder(),
65 CGBuilderInserterTy(this)),
66 CurFn(nullptr), ReturnValue(Address::invalid()),
67 CapturedStmtInfo(nullptr), SanOpts(CGM.getLangOpts().Sanitize),
68 IsSanitizerScope(false), CurFuncIsThunk(false), AutoreleaseResult(false),
69 SawAsmBlock(false), IsOutlinedSEHHelper(false), BlockInfo(nullptr),
70 BlockPointer(nullptr), LambdaThisCaptureField(nullptr),
71 NormalCleanupDest(nullptr), NextCleanupDestIndex(1),
72 FirstBlockInfo(nullptr), EHResumeBlock(nullptr), ExceptionSlot(nullptr),
73 EHSelectorSlot(nullptr), DebugInfo(CGM.getModuleDebugInfo()),
74 DisableDebugInfo(false), DidCallStackSave(false), IndirectBranch(nullptr),
75 PGO(cgm), SwitchInsn(nullptr), SwitchWeights(nullptr),
76 CaseRangeBlock(nullptr), UnreachableBlock(nullptr), NumReturnExprs(0),
77 NumSimpleReturnExprs(0), CXXABIThisDecl(nullptr),
78 CXXABIThisValue(nullptr), CXXThisValue(nullptr),
79 CXXStructorImplicitParamDecl(nullptr),
80 CXXStructorImplicitParamValue(nullptr), OutermostConditional(nullptr),
81 CurLexicalScope(nullptr), TerminateLandingPad(nullptr),
82 TerminateHandler(nullptr), TrapBB(nullptr),
83 ShouldEmitLifetimeMarkers(
84 shouldEmitLifetimeMarkers(CGM.getCodeGenOpts(), CGM.getLangOpts())) {
85 if (!suppressNewContext)
86 CGM.getCXXABI().getMangleContext().startNewFunction();
88 llvm::FastMathFlags FMF;
89 if (CGM.getLangOpts().FastMath)
90 FMF.setUnsafeAlgebra();
91 if (CGM.getLangOpts().FiniteMathOnly) {
95 if (CGM.getCodeGenOpts().NoNaNsFPMath) {
98 if (CGM.getCodeGenOpts().NoSignedZeros) {
99 FMF.setNoSignedZeros();
101 if (CGM.getCodeGenOpts().ReciprocalMath) {
102 FMF.setAllowReciprocal();
104 Builder.setFastMathFlags(FMF);
107 CodeGenFunction::~CodeGenFunction() {
108 assert(LifetimeExtendedCleanupStack.empty() && "failed to emit a cleanup");
110 // If there are any unclaimed block infos, go ahead and destroy them
111 // now. This can happen if IR-gen gets clever and skips evaluating
114 destroyBlockInfos(FirstBlockInfo);
116 if (getLangOpts().OpenMP && CurFn)
117 CGM.getOpenMPRuntime().functionFinished(*this);
120 CharUnits CodeGenFunction::getNaturalPointeeTypeAlignment(QualType T,
121 LValueBaseInfo *BaseInfo) {
122 return getNaturalTypeAlignment(T->getPointeeType(), BaseInfo,
123 /*forPointee*/ true);
126 CharUnits CodeGenFunction::getNaturalTypeAlignment(QualType T,
127 LValueBaseInfo *BaseInfo,
128 bool forPointeeType) {
129 // Honor alignment typedef attributes even on incomplete types.
130 // We also honor them straight for C++ class types, even as pointees;
131 // there's an expressivity gap here.
132 if (auto TT = T->getAs<TypedefType>()) {
133 if (auto Align = TT->getDecl()->getMaxAlignment()) {
135 *BaseInfo = LValueBaseInfo(AlignmentSource::AttributedType, false);
136 return getContext().toCharUnitsFromBits(Align);
141 *BaseInfo = LValueBaseInfo(AlignmentSource::Type, false);
144 if (T->isIncompleteType()) {
145 Alignment = CharUnits::One(); // Shouldn't be used, but pessimistic is best.
147 // For C++ class pointees, we don't know whether we're pointing at a
148 // base or a complete object, so we generally need to use the
149 // non-virtual alignment.
150 const CXXRecordDecl *RD;
151 if (forPointeeType && (RD = T->getAsCXXRecordDecl())) {
152 Alignment = CGM.getClassPointerAlignment(RD);
154 Alignment = getContext().getTypeAlignInChars(T);
155 if (T.getQualifiers().hasUnaligned())
156 Alignment = CharUnits::One();
159 // Cap to the global maximum type alignment unless the alignment
160 // was somehow explicit on the type.
161 if (unsigned MaxAlign = getLangOpts().MaxTypeAlign) {
162 if (Alignment.getQuantity() > MaxAlign &&
163 !getContext().isAlignmentRequired(T))
164 Alignment = CharUnits::fromQuantity(MaxAlign);
170 LValue CodeGenFunction::MakeNaturalAlignAddrLValue(llvm::Value *V, QualType T) {
171 LValueBaseInfo BaseInfo;
172 CharUnits Alignment = getNaturalTypeAlignment(T, &BaseInfo);
173 return LValue::MakeAddr(Address(V, Alignment), T, getContext(), BaseInfo,
177 /// Given a value of type T* that may not be to a complete object,
178 /// construct an l-value with the natural pointee alignment of T.
180 CodeGenFunction::MakeNaturalAlignPointeeAddrLValue(llvm::Value *V, QualType T) {
181 LValueBaseInfo BaseInfo;
182 CharUnits Align = getNaturalTypeAlignment(T, &BaseInfo, /*pointee*/ true);
183 return MakeAddrLValue(Address(V, Align), T, BaseInfo);
187 llvm::Type *CodeGenFunction::ConvertTypeForMem(QualType T) {
188 return CGM.getTypes().ConvertTypeForMem(T);
191 llvm::Type *CodeGenFunction::ConvertType(QualType T) {
192 return CGM.getTypes().ConvertType(T);
195 TypeEvaluationKind CodeGenFunction::getEvaluationKind(QualType type) {
196 type = type.getCanonicalType();
198 switch (type->getTypeClass()) {
199 #define TYPE(name, parent)
200 #define ABSTRACT_TYPE(name, parent)
201 #define NON_CANONICAL_TYPE(name, parent) case Type::name:
202 #define DEPENDENT_TYPE(name, parent) case Type::name:
203 #define NON_CANONICAL_UNLESS_DEPENDENT_TYPE(name, parent) case Type::name:
204 #include "clang/AST/TypeNodes.def"
205 llvm_unreachable("non-canonical or dependent type in IR-generation");
208 case Type::DeducedTemplateSpecialization:
209 llvm_unreachable("undeduced type in IR-generation");
211 // Various scalar types.
214 case Type::BlockPointer:
215 case Type::LValueReference:
216 case Type::RValueReference:
217 case Type::MemberPointer:
219 case Type::ExtVector:
220 case Type::FunctionProto:
221 case Type::FunctionNoProto:
223 case Type::ObjCObjectPointer:
231 // Arrays, records, and Objective-C objects.
232 case Type::ConstantArray:
233 case Type::IncompleteArray:
234 case Type::VariableArray:
236 case Type::ObjCObject:
237 case Type::ObjCInterface:
238 return TEK_Aggregate;
240 // We operate on atomic values according to their underlying type.
242 type = cast<AtomicType>(type)->getValueType();
245 llvm_unreachable("unknown type kind!");
249 llvm::DebugLoc CodeGenFunction::EmitReturnBlock() {
250 // For cleanliness, we try to avoid emitting the return block for
252 llvm::BasicBlock *CurBB = Builder.GetInsertBlock();
255 assert(!CurBB->getTerminator() && "Unexpected terminated block.");
257 // We have a valid insert point, reuse it if it is empty or there are no
258 // explicit jumps to the return block.
259 if (CurBB->empty() || ReturnBlock.getBlock()->use_empty()) {
260 ReturnBlock.getBlock()->replaceAllUsesWith(CurBB);
261 delete ReturnBlock.getBlock();
263 EmitBlock(ReturnBlock.getBlock());
264 return llvm::DebugLoc();
267 // Otherwise, if the return block is the target of a single direct
268 // branch then we can just put the code in that block instead. This
269 // cleans up functions which started with a unified return block.
270 if (ReturnBlock.getBlock()->hasOneUse()) {
271 llvm::BranchInst *BI =
272 dyn_cast<llvm::BranchInst>(*ReturnBlock.getBlock()->user_begin());
273 if (BI && BI->isUnconditional() &&
274 BI->getSuccessor(0) == ReturnBlock.getBlock()) {
275 // Record/return the DebugLoc of the simple 'return' expression to be used
276 // later by the actual 'ret' instruction.
277 llvm::DebugLoc Loc = BI->getDebugLoc();
278 Builder.SetInsertPoint(BI->getParent());
279 BI->eraseFromParent();
280 delete ReturnBlock.getBlock();
285 // FIXME: We are at an unreachable point, there is no reason to emit the block
286 // unless it has uses. However, we still need a place to put the debug
287 // region.end for now.
289 EmitBlock(ReturnBlock.getBlock());
290 return llvm::DebugLoc();
293 static void EmitIfUsed(CodeGenFunction &CGF, llvm::BasicBlock *BB) {
295 if (!BB->use_empty())
296 return CGF.CurFn->getBasicBlockList().push_back(BB);
300 void CodeGenFunction::FinishFunction(SourceLocation EndLoc) {
301 assert(BreakContinueStack.empty() &&
302 "mismatched push/pop in break/continue stack!");
304 bool OnlySimpleReturnStmts = NumSimpleReturnExprs > 0
305 && NumSimpleReturnExprs == NumReturnExprs
306 && ReturnBlock.getBlock()->use_empty();
307 // Usually the return expression is evaluated before the cleanup
308 // code. If the function contains only a simple return statement,
309 // such as a constant, the location before the cleanup code becomes
310 // the last useful breakpoint in the function, because the simple
311 // return expression will be evaluated after the cleanup code. To be
312 // safe, set the debug location for cleanup code to the location of
313 // the return statement. Otherwise the cleanup code should be at the
314 // end of the function's lexical scope.
316 // If there are multiple branches to the return block, the branch
317 // instructions will get the location of the return statements and
319 if (CGDebugInfo *DI = getDebugInfo()) {
320 if (OnlySimpleReturnStmts)
321 DI->EmitLocation(Builder, LastStopPoint);
323 DI->EmitLocation(Builder, EndLoc);
326 // Pop any cleanups that might have been associated with the
327 // parameters. Do this in whatever block we're currently in; it's
328 // important to do this before we enter the return block or return
329 // edges will be *really* confused.
330 bool HasCleanups = EHStack.stable_begin() != PrologueCleanupDepth;
331 bool HasOnlyLifetimeMarkers =
332 HasCleanups && EHStack.containsOnlyLifetimeMarkers(PrologueCleanupDepth);
333 bool EmitRetDbgLoc = !HasCleanups || HasOnlyLifetimeMarkers;
335 // Make sure the line table doesn't jump back into the body for
336 // the ret after it's been at EndLoc.
337 if (CGDebugInfo *DI = getDebugInfo())
338 if (OnlySimpleReturnStmts)
339 DI->EmitLocation(Builder, EndLoc);
341 PopCleanupBlocks(PrologueCleanupDepth);
344 // Emit function epilog (to return).
345 llvm::DebugLoc Loc = EmitReturnBlock();
347 if (ShouldInstrumentFunction())
348 EmitFunctionInstrumentation("__cyg_profile_func_exit");
350 // Emit debug descriptor for function end.
351 if (CGDebugInfo *DI = getDebugInfo())
352 DI->EmitFunctionEnd(Builder, CurFn);
354 // Reset the debug location to that of the simple 'return' expression, if any
355 // rather than that of the end of the function's scope '}'.
356 ApplyDebugLocation AL(*this, Loc);
357 EmitFunctionEpilog(*CurFnInfo, EmitRetDbgLoc, EndLoc);
358 EmitEndEHSpec(CurCodeDecl);
360 assert(EHStack.empty() &&
361 "did not remove all scopes from cleanup stack!");
363 // If someone did an indirect goto, emit the indirect goto block at the end of
365 if (IndirectBranch) {
366 EmitBlock(IndirectBranch->getParent());
367 Builder.ClearInsertionPoint();
370 // If some of our locals escaped, insert a call to llvm.localescape in the
372 if (!EscapedLocals.empty()) {
373 // Invert the map from local to index into a simple vector. There should be
375 SmallVector<llvm::Value *, 4> EscapeArgs;
376 EscapeArgs.resize(EscapedLocals.size());
377 for (auto &Pair : EscapedLocals)
378 EscapeArgs[Pair.second] = Pair.first;
379 llvm::Function *FrameEscapeFn = llvm::Intrinsic::getDeclaration(
380 &CGM.getModule(), llvm::Intrinsic::localescape);
381 CGBuilderTy(*this, AllocaInsertPt).CreateCall(FrameEscapeFn, EscapeArgs);
384 // Remove the AllocaInsertPt instruction, which is just a convenience for us.
385 llvm::Instruction *Ptr = AllocaInsertPt;
386 AllocaInsertPt = nullptr;
387 Ptr->eraseFromParent();
389 // If someone took the address of a label but never did an indirect goto, we
390 // made a zero entry PHI node, which is illegal, zap it now.
391 if (IndirectBranch) {
392 llvm::PHINode *PN = cast<llvm::PHINode>(IndirectBranch->getAddress());
393 if (PN->getNumIncomingValues() == 0) {
394 PN->replaceAllUsesWith(llvm::UndefValue::get(PN->getType()));
395 PN->eraseFromParent();
399 EmitIfUsed(*this, EHResumeBlock);
400 EmitIfUsed(*this, TerminateLandingPad);
401 EmitIfUsed(*this, TerminateHandler);
402 EmitIfUsed(*this, UnreachableBlock);
404 if (CGM.getCodeGenOpts().EmitDeclMetadata)
407 for (SmallVectorImpl<std::pair<llvm::Instruction *, llvm::Value *> >::iterator
408 I = DeferredReplacements.begin(),
409 E = DeferredReplacements.end();
411 I->first->replaceAllUsesWith(I->second);
412 I->first->eraseFromParent();
416 /// ShouldInstrumentFunction - Return true if the current function should be
417 /// instrumented with __cyg_profile_func_* calls
418 bool CodeGenFunction::ShouldInstrumentFunction() {
419 if (!CGM.getCodeGenOpts().InstrumentFunctions)
421 if (!CurFuncDecl || CurFuncDecl->hasAttr<NoInstrumentFunctionAttr>())
426 /// ShouldXRayInstrument - Return true if the current function should be
427 /// instrumented with XRay nop sleds.
428 bool CodeGenFunction::ShouldXRayInstrumentFunction() const {
429 return CGM.getCodeGenOpts().XRayInstrumentFunctions;
432 /// EmitFunctionInstrumentation - Emit LLVM code to call the specified
433 /// instrumentation function with the current function and the call site, if
434 /// function instrumentation is enabled.
435 void CodeGenFunction::EmitFunctionInstrumentation(const char *Fn) {
436 auto NL = ApplyDebugLocation::CreateArtificial(*this);
437 // void __cyg_profile_func_{enter,exit} (void *this_fn, void *call_site);
438 llvm::PointerType *PointerTy = Int8PtrTy;
439 llvm::Type *ProfileFuncArgs[] = { PointerTy, PointerTy };
440 llvm::FunctionType *FunctionTy =
441 llvm::FunctionType::get(VoidTy, ProfileFuncArgs, false);
443 llvm::Constant *F = CGM.CreateRuntimeFunction(FunctionTy, Fn);
444 llvm::CallInst *CallSite = Builder.CreateCall(
445 CGM.getIntrinsic(llvm::Intrinsic::returnaddress),
446 llvm::ConstantInt::get(Int32Ty, 0),
449 llvm::Value *args[] = {
450 llvm::ConstantExpr::getBitCast(CurFn, PointerTy),
454 EmitNounwindRuntimeCall(F, args);
457 static void removeImageAccessQualifier(std::string& TyName) {
458 std::string ReadOnlyQual("__read_only");
459 std::string::size_type ReadOnlyPos = TyName.find(ReadOnlyQual);
460 if (ReadOnlyPos != std::string::npos)
461 // "+ 1" for the space after access qualifier.
462 TyName.erase(ReadOnlyPos, ReadOnlyQual.size() + 1);
464 std::string WriteOnlyQual("__write_only");
465 std::string::size_type WriteOnlyPos = TyName.find(WriteOnlyQual);
466 if (WriteOnlyPos != std::string::npos)
467 TyName.erase(WriteOnlyPos, WriteOnlyQual.size() + 1);
469 std::string ReadWriteQual("__read_write");
470 std::string::size_type ReadWritePos = TyName.find(ReadWriteQual);
471 if (ReadWritePos != std::string::npos)
472 TyName.erase(ReadWritePos, ReadWriteQual.size() + 1);
477 // Returns the address space id that should be produced to the
478 // kernel_arg_addr_space metadata. This is always fixed to the ids
479 // as specified in the SPIR 2.0 specification in order to differentiate
480 // for example in clGetKernelArgInfo() implementation between the address
481 // spaces with targets without unique mapping to the OpenCL address spaces
482 // (basically all single AS CPUs).
483 static unsigned ArgInfoAddressSpace(unsigned LangAS) {
485 case LangAS::opencl_global: return 1;
486 case LangAS::opencl_constant: return 2;
487 case LangAS::opencl_local: return 3;
488 case LangAS::opencl_generic: return 4; // Not in SPIR 2.0 specs.
490 return 0; // Assume private.
494 // OpenCL v1.2 s5.6.4.6 allows the compiler to store kernel argument
495 // information in the program executable. The argument information stored
496 // includes the argument name, its type, the address and access qualifiers used.
497 static void GenOpenCLArgMetadata(const FunctionDecl *FD, llvm::Function *Fn,
498 CodeGenModule &CGM, llvm::LLVMContext &Context,
499 CGBuilderTy &Builder, ASTContext &ASTCtx) {
500 // Create MDNodes that represent the kernel arg metadata.
501 // Each MDNode is a list in the form of "key", N number of values which is
502 // the same number of values as their are kernel arguments.
504 const PrintingPolicy &Policy = ASTCtx.getPrintingPolicy();
506 // MDNode for the kernel argument address space qualifiers.
507 SmallVector<llvm::Metadata *, 8> addressQuals;
509 // MDNode for the kernel argument access qualifiers (images only).
510 SmallVector<llvm::Metadata *, 8> accessQuals;
512 // MDNode for the kernel argument type names.
513 SmallVector<llvm::Metadata *, 8> argTypeNames;
515 // MDNode for the kernel argument base type names.
516 SmallVector<llvm::Metadata *, 8> argBaseTypeNames;
518 // MDNode for the kernel argument type qualifiers.
519 SmallVector<llvm::Metadata *, 8> argTypeQuals;
521 // MDNode for the kernel argument names.
522 SmallVector<llvm::Metadata *, 8> argNames;
524 for (unsigned i = 0, e = FD->getNumParams(); i != e; ++i) {
525 const ParmVarDecl *parm = FD->getParamDecl(i);
526 QualType ty = parm->getType();
527 std::string typeQuals;
529 if (ty->isPointerType()) {
530 QualType pointeeTy = ty->getPointeeType();
532 // Get address qualifier.
533 addressQuals.push_back(llvm::ConstantAsMetadata::get(Builder.getInt32(
534 ArgInfoAddressSpace(pointeeTy.getAddressSpace()))));
536 // Get argument type name.
537 std::string typeName =
538 pointeeTy.getUnqualifiedType().getAsString(Policy) + "*";
540 // Turn "unsigned type" to "utype"
541 std::string::size_type pos = typeName.find("unsigned");
542 if (pointeeTy.isCanonical() && pos != std::string::npos)
543 typeName.erase(pos+1, 8);
545 argTypeNames.push_back(llvm::MDString::get(Context, typeName));
547 std::string baseTypeName =
548 pointeeTy.getUnqualifiedType().getCanonicalType().getAsString(
552 // Turn "unsigned type" to "utype"
553 pos = baseTypeName.find("unsigned");
554 if (pos != std::string::npos)
555 baseTypeName.erase(pos+1, 8);
557 argBaseTypeNames.push_back(llvm::MDString::get(Context, baseTypeName));
559 // Get argument type qualifiers:
560 if (ty.isRestrictQualified())
561 typeQuals = "restrict";
562 if (pointeeTy.isConstQualified() ||
563 (pointeeTy.getAddressSpace() == LangAS::opencl_constant))
564 typeQuals += typeQuals.empty() ? "const" : " const";
565 if (pointeeTy.isVolatileQualified())
566 typeQuals += typeQuals.empty() ? "volatile" : " volatile";
568 uint32_t AddrSpc = 0;
569 bool isPipe = ty->isPipeType();
570 if (ty->isImageType() || isPipe)
571 AddrSpc = ArgInfoAddressSpace(LangAS::opencl_global);
573 addressQuals.push_back(
574 llvm::ConstantAsMetadata::get(Builder.getInt32(AddrSpc)));
576 // Get argument type name.
577 std::string typeName;
579 typeName = ty.getCanonicalType()->getAs<PipeType>()->getElementType()
580 .getAsString(Policy);
582 typeName = ty.getUnqualifiedType().getAsString(Policy);
584 // Turn "unsigned type" to "utype"
585 std::string::size_type pos = typeName.find("unsigned");
586 if (ty.isCanonical() && pos != std::string::npos)
587 typeName.erase(pos+1, 8);
589 std::string baseTypeName;
591 baseTypeName = ty.getCanonicalType()->getAs<PipeType>()
592 ->getElementType().getCanonicalType()
593 .getAsString(Policy);
596 ty.getUnqualifiedType().getCanonicalType().getAsString(Policy);
598 // Remove access qualifiers on images
599 // (as they are inseparable from type in clang implementation,
600 // but OpenCL spec provides a special query to get access qualifier
601 // via clGetKernelArgInfo with CL_KERNEL_ARG_ACCESS_QUALIFIER):
602 if (ty->isImageType()) {
603 removeImageAccessQualifier(typeName);
604 removeImageAccessQualifier(baseTypeName);
607 argTypeNames.push_back(llvm::MDString::get(Context, typeName));
609 // Turn "unsigned type" to "utype"
610 pos = baseTypeName.find("unsigned");
611 if (pos != std::string::npos)
612 baseTypeName.erase(pos+1, 8);
614 argBaseTypeNames.push_back(llvm::MDString::get(Context, baseTypeName));
620 argTypeQuals.push_back(llvm::MDString::get(Context, typeQuals));
622 // Get image and pipe access qualifier:
623 if (ty->isImageType()|| ty->isPipeType()) {
624 const OpenCLAccessAttr *A = parm->getAttr<OpenCLAccessAttr>();
625 if (A && A->isWriteOnly())
626 accessQuals.push_back(llvm::MDString::get(Context, "write_only"));
627 else if (A && A->isReadWrite())
628 accessQuals.push_back(llvm::MDString::get(Context, "read_write"));
630 accessQuals.push_back(llvm::MDString::get(Context, "read_only"));
632 accessQuals.push_back(llvm::MDString::get(Context, "none"));
634 // Get argument name.
635 argNames.push_back(llvm::MDString::get(Context, parm->getName()));
638 Fn->setMetadata("kernel_arg_addr_space",
639 llvm::MDNode::get(Context, addressQuals));
640 Fn->setMetadata("kernel_arg_access_qual",
641 llvm::MDNode::get(Context, accessQuals));
642 Fn->setMetadata("kernel_arg_type",
643 llvm::MDNode::get(Context, argTypeNames));
644 Fn->setMetadata("kernel_arg_base_type",
645 llvm::MDNode::get(Context, argBaseTypeNames));
646 Fn->setMetadata("kernel_arg_type_qual",
647 llvm::MDNode::get(Context, argTypeQuals));
648 if (CGM.getCodeGenOpts().EmitOpenCLArgMetadata)
649 Fn->setMetadata("kernel_arg_name",
650 llvm::MDNode::get(Context, argNames));
653 void CodeGenFunction::EmitOpenCLKernelMetadata(const FunctionDecl *FD,
656 if (!FD->hasAttr<OpenCLKernelAttr>())
659 llvm::LLVMContext &Context = getLLVMContext();
661 GenOpenCLArgMetadata(FD, Fn, CGM, Context, Builder, getContext());
663 if (const VecTypeHintAttr *A = FD->getAttr<VecTypeHintAttr>()) {
664 QualType HintQTy = A->getTypeHint();
665 const ExtVectorType *HintEltQTy = HintQTy->getAs<ExtVectorType>();
666 bool IsSignedInteger =
667 HintQTy->isSignedIntegerType() ||
668 (HintEltQTy && HintEltQTy->getElementType()->isSignedIntegerType());
669 llvm::Metadata *AttrMDArgs[] = {
670 llvm::ConstantAsMetadata::get(llvm::UndefValue::get(
671 CGM.getTypes().ConvertType(A->getTypeHint()))),
672 llvm::ConstantAsMetadata::get(llvm::ConstantInt::get(
673 llvm::IntegerType::get(Context, 32),
674 llvm::APInt(32, (uint64_t)(IsSignedInteger ? 1 : 0))))};
675 Fn->setMetadata("vec_type_hint", llvm::MDNode::get(Context, AttrMDArgs));
678 if (const WorkGroupSizeHintAttr *A = FD->getAttr<WorkGroupSizeHintAttr>()) {
679 llvm::Metadata *AttrMDArgs[] = {
680 llvm::ConstantAsMetadata::get(Builder.getInt32(A->getXDim())),
681 llvm::ConstantAsMetadata::get(Builder.getInt32(A->getYDim())),
682 llvm::ConstantAsMetadata::get(Builder.getInt32(A->getZDim()))};
683 Fn->setMetadata("work_group_size_hint", llvm::MDNode::get(Context, AttrMDArgs));
686 if (const ReqdWorkGroupSizeAttr *A = FD->getAttr<ReqdWorkGroupSizeAttr>()) {
687 llvm::Metadata *AttrMDArgs[] = {
688 llvm::ConstantAsMetadata::get(Builder.getInt32(A->getXDim())),
689 llvm::ConstantAsMetadata::get(Builder.getInt32(A->getYDim())),
690 llvm::ConstantAsMetadata::get(Builder.getInt32(A->getZDim()))};
691 Fn->setMetadata("reqd_work_group_size", llvm::MDNode::get(Context, AttrMDArgs));
694 if (const OpenCLIntelReqdSubGroupSizeAttr *A =
695 FD->getAttr<OpenCLIntelReqdSubGroupSizeAttr>()) {
696 llvm::Metadata *AttrMDArgs[] = {
697 llvm::ConstantAsMetadata::get(Builder.getInt32(A->getSubGroupSize()))};
698 Fn->setMetadata("intel_reqd_sub_group_size",
699 llvm::MDNode::get(Context, AttrMDArgs));
703 /// Determine whether the function F ends with a return stmt.
704 static bool endsWithReturn(const Decl* F) {
705 const Stmt *Body = nullptr;
706 if (auto *FD = dyn_cast_or_null<FunctionDecl>(F))
707 Body = FD->getBody();
708 else if (auto *OMD = dyn_cast_or_null<ObjCMethodDecl>(F))
709 Body = OMD->getBody();
711 if (auto *CS = dyn_cast_or_null<CompoundStmt>(Body)) {
712 auto LastStmt = CS->body_rbegin();
713 if (LastStmt != CS->body_rend())
714 return isa<ReturnStmt>(*LastStmt);
719 static void markAsIgnoreThreadCheckingAtRuntime(llvm::Function *Fn) {
720 Fn->addFnAttr("sanitize_thread_no_checking_at_run_time");
721 Fn->removeFnAttr(llvm::Attribute::SanitizeThread);
724 void CodeGenFunction::StartFunction(GlobalDecl GD,
727 const CGFunctionInfo &FnInfo,
728 const FunctionArgList &Args,
730 SourceLocation StartLoc) {
732 "Do not use a CodeGenFunction object for more than one function");
734 const Decl *D = GD.getDecl();
736 DidCallStackSave = false;
738 if (const auto *FD = dyn_cast_or_null<FunctionDecl>(D))
739 if (FD->usesSEHTry())
741 CurFuncDecl = (D ? D->getNonClosureContext() : nullptr);
745 assert(CurFn->isDeclaration() && "Function already has body?");
747 if (CGM.isInSanitizerBlacklist(Fn, Loc))
751 // Apply the no_sanitize* attributes to SanOpts.
752 for (auto Attr : D->specific_attrs<NoSanitizeAttr>())
753 SanOpts.Mask &= ~Attr->getMask();
756 // Apply sanitizer attributes to the function.
757 if (SanOpts.hasOneOf(SanitizerKind::Address | SanitizerKind::KernelAddress))
758 Fn->addFnAttr(llvm::Attribute::SanitizeAddress);
759 if (SanOpts.has(SanitizerKind::Thread))
760 Fn->addFnAttr(llvm::Attribute::SanitizeThread);
761 if (SanOpts.has(SanitizerKind::Memory))
762 Fn->addFnAttr(llvm::Attribute::SanitizeMemory);
763 if (SanOpts.has(SanitizerKind::SafeStack))
764 Fn->addFnAttr(llvm::Attribute::SafeStack);
766 // Ignore TSan memory acesses from within ObjC/ObjC++ dealloc, initialize,
767 // .cxx_destruct, __destroy_helper_block_ and all of their calees at run time.
768 if (SanOpts.has(SanitizerKind::Thread)) {
769 if (const auto *OMD = dyn_cast_or_null<ObjCMethodDecl>(D)) {
770 IdentifierInfo *II = OMD->getSelector().getIdentifierInfoForSlot(0);
771 if (OMD->getMethodFamily() == OMF_dealloc ||
772 OMD->getMethodFamily() == OMF_initialize ||
773 (OMD->getSelector().isUnarySelector() && II->isStr(".cxx_destruct"))) {
774 markAsIgnoreThreadCheckingAtRuntime(Fn);
776 } else if (const auto *FD = dyn_cast_or_null<FunctionDecl>(D)) {
777 IdentifierInfo *II = FD->getIdentifier();
778 if (II && II->isStr("__destroy_helper_block_"))
779 markAsIgnoreThreadCheckingAtRuntime(Fn);
783 // Apply xray attributes to the function (as a string, for now)
784 if (D && ShouldXRayInstrumentFunction()) {
785 if (const auto *XRayAttr = D->getAttr<XRayInstrumentAttr>()) {
786 if (XRayAttr->alwaysXRayInstrument())
787 Fn->addFnAttr("function-instrument", "xray-always");
788 if (XRayAttr->neverXRayInstrument())
789 Fn->addFnAttr("function-instrument", "xray-never");
790 if (const auto *LogArgs = D->getAttr<XRayLogArgsAttr>()) {
791 Fn->addFnAttr("xray-log-args",
792 llvm::utostr(LogArgs->getArgumentCount()));
795 if (!CGM.imbueXRayAttrs(Fn, Loc))
797 "xray-instruction-threshold",
798 llvm::itostr(CGM.getCodeGenOpts().XRayInstructionThreshold));
802 if (const FunctionDecl *FD = dyn_cast_or_null<FunctionDecl>(D))
803 if (CGM.getLangOpts().OpenMP && FD->hasAttr<OMPDeclareSimdDeclAttr>())
804 CGM.getOpenMPRuntime().emitDeclareSimdFunction(FD, Fn);
806 // Add no-jump-tables value.
807 Fn->addFnAttr("no-jump-tables",
808 llvm::toStringRef(CGM.getCodeGenOpts().NoUseJumpTables));
810 if (getLangOpts().OpenCL) {
811 // Add metadata for a kernel function.
812 if (const FunctionDecl *FD = dyn_cast_or_null<FunctionDecl>(D))
813 EmitOpenCLKernelMetadata(FD, Fn);
816 // If we are checking function types, emit a function type signature as
818 if (getLangOpts().CPlusPlus && SanOpts.has(SanitizerKind::Function)) {
819 if (const FunctionDecl *FD = dyn_cast_or_null<FunctionDecl>(D)) {
820 if (llvm::Constant *PrologueSig =
821 CGM.getTargetCodeGenInfo().getUBSanFunctionSignature(CGM)) {
822 llvm::Constant *FTRTTIConst =
823 CGM.GetAddrOfRTTIDescriptor(FD->getType(), /*ForEH=*/true);
824 llvm::Constant *PrologueStructElems[] = { PrologueSig, FTRTTIConst };
825 llvm::Constant *PrologueStructConst =
826 llvm::ConstantStruct::getAnon(PrologueStructElems, /*Packed=*/true);
827 Fn->setPrologueData(PrologueStructConst);
832 // If we're checking nullability, we need to know whether we can check the
833 // return value. Initialize the flag to 'true' and refine it in EmitParmDecl.
834 if (SanOpts.has(SanitizerKind::NullabilityReturn)) {
835 auto Nullability = FnRetTy->getNullability(getContext());
836 if (Nullability && *Nullability == NullabilityKind::NonNull) {
837 if (!(SanOpts.has(SanitizerKind::ReturnsNonnullAttribute) &&
838 CurCodeDecl && CurCodeDecl->getAttr<ReturnsNonNullAttr>()))
839 RetValNullabilityPrecondition =
840 llvm::ConstantInt::getTrue(getLLVMContext());
844 // If we're in C++ mode and the function name is "main", it is guaranteed
845 // to be norecurse by the standard (3.6.1.3 "The function main shall not be
846 // used within a program").
847 if (getLangOpts().CPlusPlus)
848 if (const FunctionDecl *FD = dyn_cast_or_null<FunctionDecl>(D))
850 Fn->addFnAttr(llvm::Attribute::NoRecurse);
852 llvm::BasicBlock *EntryBB = createBasicBlock("entry", CurFn);
854 // Create a marker to make it easy to insert allocas into the entryblock
855 // later. Don't create this with the builder, because we don't want it
857 llvm::Value *Undef = llvm::UndefValue::get(Int32Ty);
858 AllocaInsertPt = new llvm::BitCastInst(Undef, Int32Ty, "allocapt", EntryBB);
860 ReturnBlock = getJumpDestInCurrentScope("return");
862 Builder.SetInsertPoint(EntryBB);
864 // If we're checking the return value, allocate space for a pointer to a
865 // precise source location of the checked return statement.
866 if (requiresReturnValueCheck()) {
867 ReturnLocation = CreateDefaultAlignTempAlloca(Int8PtrTy, "return.sloc.ptr");
868 InitTempAlloca(ReturnLocation, llvm::ConstantPointerNull::get(Int8PtrTy));
871 // Emit subprogram debug descriptor.
872 if (CGDebugInfo *DI = getDebugInfo()) {
873 // Reconstruct the type from the argument list so that implicit parameters,
874 // such as 'this' and 'vtt', show up in the debug info. Preserve the calling
876 CallingConv CC = CallingConv::CC_C;
877 if (auto *FD = dyn_cast_or_null<FunctionDecl>(D))
878 if (const auto *SrcFnTy = FD->getType()->getAs<FunctionType>())
879 CC = SrcFnTy->getCallConv();
880 SmallVector<QualType, 16> ArgTypes;
881 for (const VarDecl *VD : Args)
882 ArgTypes.push_back(VD->getType());
883 QualType FnType = getContext().getFunctionType(
884 RetTy, ArgTypes, FunctionProtoType::ExtProtoInfo(CC));
885 DI->EmitFunctionStart(GD, Loc, StartLoc, FnType, CurFn, Builder);
888 if (ShouldInstrumentFunction())
889 EmitFunctionInstrumentation("__cyg_profile_func_enter");
891 // Since emitting the mcount call here impacts optimizations such as function
892 // inlining, we just add an attribute to insert a mcount call in backend.
893 // The attribute "counting-function" is set to mcount function name which is
894 // architecture dependent.
895 if (CGM.getCodeGenOpts().InstrumentForProfiling) {
896 if (CGM.getCodeGenOpts().CallFEntry)
897 Fn->addFnAttr("fentry-call", "true");
899 if (!CurFuncDecl || !CurFuncDecl->hasAttr<NoInstrumentFunctionAttr>())
900 Fn->addFnAttr("counting-function", getTarget().getMCountName());
904 if (RetTy->isVoidType()) {
905 // Void type; nothing to return.
906 ReturnValue = Address::invalid();
908 // Count the implicit return.
909 if (!endsWithReturn(D))
911 } else if (CurFnInfo->getReturnInfo().getKind() == ABIArgInfo::Indirect &&
912 !hasScalarEvaluationKind(CurFnInfo->getReturnType())) {
913 // Indirect aggregate return; emit returned value directly into sret slot.
914 // This reduces code size, and affects correctness in C++.
915 auto AI = CurFn->arg_begin();
916 if (CurFnInfo->getReturnInfo().isSRetAfterThis())
918 ReturnValue = Address(&*AI, CurFnInfo->getReturnInfo().getIndirectAlign());
919 } else if (CurFnInfo->getReturnInfo().getKind() == ABIArgInfo::InAlloca &&
920 !hasScalarEvaluationKind(CurFnInfo->getReturnType())) {
921 // Load the sret pointer from the argument struct and return into that.
922 unsigned Idx = CurFnInfo->getReturnInfo().getInAllocaFieldIndex();
923 llvm::Function::arg_iterator EI = CurFn->arg_end();
925 llvm::Value *Addr = Builder.CreateStructGEP(nullptr, &*EI, Idx);
926 Addr = Builder.CreateAlignedLoad(Addr, getPointerAlign(), "agg.result");
927 ReturnValue = Address(Addr, getNaturalTypeAlignment(RetTy));
929 ReturnValue = CreateIRTemp(RetTy, "retval");
931 // Tell the epilog emitter to autorelease the result. We do this
932 // now so that various specialized functions can suppress it
933 // during their IR-generation.
934 if (getLangOpts().ObjCAutoRefCount &&
935 !CurFnInfo->isReturnsRetained() &&
936 RetTy->isObjCRetainableType())
937 AutoreleaseResult = true;
940 EmitStartEHSpec(CurCodeDecl);
942 PrologueCleanupDepth = EHStack.stable_begin();
943 EmitFunctionProlog(*CurFnInfo, CurFn, Args);
945 if (D && isa<CXXMethodDecl>(D) && cast<CXXMethodDecl>(D)->isInstance()) {
946 CGM.getCXXABI().EmitInstanceFunctionProlog(*this);
947 const CXXMethodDecl *MD = cast<CXXMethodDecl>(D);
948 if (MD->getParent()->isLambda() &&
949 MD->getOverloadedOperator() == OO_Call) {
950 // We're in a lambda; figure out the captures.
951 MD->getParent()->getCaptureFields(LambdaCaptureFields,
952 LambdaThisCaptureField);
953 if (LambdaThisCaptureField) {
954 // If the lambda captures the object referred to by '*this' - either by
955 // value or by reference, make sure CXXThisValue points to the correct
958 // Get the lvalue for the field (which is a copy of the enclosing object
959 // or contains the address of the enclosing object).
960 LValue ThisFieldLValue = EmitLValueForLambdaField(LambdaThisCaptureField);
961 if (!LambdaThisCaptureField->getType()->isPointerType()) {
962 // If the enclosing object was captured by value, just use its address.
963 CXXThisValue = ThisFieldLValue.getAddress().getPointer();
965 // Load the lvalue pointed to by the field, since '*this' was captured
968 EmitLoadOfLValue(ThisFieldLValue, SourceLocation()).getScalarVal();
971 for (auto *FD : MD->getParent()->fields()) {
972 if (FD->hasCapturedVLAType()) {
973 auto *ExprArg = EmitLoadOfLValue(EmitLValueForLambdaField(FD),
974 SourceLocation()).getScalarVal();
975 auto VAT = FD->getCapturedVLAType();
976 VLASizeMap[VAT->getSizeExpr()] = ExprArg;
980 // Not in a lambda; just use 'this' from the method.
981 // FIXME: Should we generate a new load for each use of 'this'? The
982 // fast register allocator would be happier...
983 CXXThisValue = CXXABIThisValue;
986 // Check the 'this' pointer once per function, if it's available.
987 if (CXXABIThisValue) {
988 SanitizerSet SkippedChecks;
989 SkippedChecks.set(SanitizerKind::ObjectSize, true);
990 QualType ThisTy = MD->getThisType(getContext());
992 // If this is the call operator of a lambda with no capture-default, it
993 // may have a static invoker function, which may call this operator with
994 // a null 'this' pointer.
995 if (isLambdaCallOperator(MD) &&
996 cast<CXXRecordDecl>(MD->getParent())->getLambdaCaptureDefault() ==
998 SkippedChecks.set(SanitizerKind::Null, true);
1000 EmitTypeCheck(isa<CXXConstructorDecl>(MD) ? TCK_ConstructorCall
1002 Loc, CXXABIThisValue, ThisTy,
1003 getContext().getTypeAlignInChars(ThisTy->getPointeeType()),
1008 // If any of the arguments have a variably modified type, make sure to
1009 // emit the type size.
1010 for (FunctionArgList::const_iterator i = Args.begin(), e = Args.end();
1012 const VarDecl *VD = *i;
1014 // Dig out the type as written from ParmVarDecls; it's unclear whether
1015 // the standard (C99 6.9.1p10) requires this, but we're following the
1016 // precedent set by gcc.
1018 if (const ParmVarDecl *PVD = dyn_cast<ParmVarDecl>(VD))
1019 Ty = PVD->getOriginalType();
1023 if (Ty->isVariablyModifiedType())
1024 EmitVariablyModifiedType(Ty);
1026 // Emit a location at the end of the prologue.
1027 if (CGDebugInfo *DI = getDebugInfo())
1028 DI->EmitLocation(Builder, StartLoc);
1031 void CodeGenFunction::EmitFunctionBody(FunctionArgList &Args,
1033 incrementProfileCounter(Body);
1034 if (const CompoundStmt *S = dyn_cast<CompoundStmt>(Body))
1035 EmitCompoundStmtWithoutScope(*S);
1040 /// When instrumenting to collect profile data, the counts for some blocks
1041 /// such as switch cases need to not include the fall-through counts, so
1042 /// emit a branch around the instrumentation code. When not instrumenting,
1043 /// this just calls EmitBlock().
1044 void CodeGenFunction::EmitBlockWithFallThrough(llvm::BasicBlock *BB,
1046 llvm::BasicBlock *SkipCountBB = nullptr;
1047 if (HaveInsertPoint() && CGM.getCodeGenOpts().hasProfileClangInstr()) {
1048 // When instrumenting for profiling, the fallthrough to certain
1049 // statements needs to skip over the instrumentation code so that we
1050 // get an accurate count.
1051 SkipCountBB = createBasicBlock("skipcount");
1052 EmitBranch(SkipCountBB);
1055 uint64_t CurrentCount = getCurrentProfileCount();
1056 incrementProfileCounter(S);
1057 setCurrentProfileCount(getCurrentProfileCount() + CurrentCount);
1059 EmitBlock(SkipCountBB);
1062 /// Tries to mark the given function nounwind based on the
1063 /// non-existence of any throwing calls within it. We believe this is
1064 /// lightweight enough to do at -O0.
1065 static void TryMarkNoThrow(llvm::Function *F) {
1066 // LLVM treats 'nounwind' on a function as part of the type, so we
1067 // can't do this on functions that can be overwritten.
1068 if (F->isInterposable()) return;
1070 for (llvm::BasicBlock &BB : *F)
1071 for (llvm::Instruction &I : BB)
1075 F->setDoesNotThrow();
1078 QualType CodeGenFunction::BuildFunctionArgList(GlobalDecl GD,
1079 FunctionArgList &Args) {
1080 const FunctionDecl *FD = cast<FunctionDecl>(GD.getDecl());
1081 QualType ResTy = FD->getReturnType();
1083 const CXXMethodDecl *MD = dyn_cast<CXXMethodDecl>(FD);
1084 if (MD && MD->isInstance()) {
1085 if (CGM.getCXXABI().HasThisReturn(GD))
1086 ResTy = MD->getThisType(getContext());
1087 else if (CGM.getCXXABI().hasMostDerivedReturn(GD))
1088 ResTy = CGM.getContext().VoidPtrTy;
1089 CGM.getCXXABI().buildThisParam(*this, Args);
1092 // The base version of an inheriting constructor whose constructed base is a
1093 // virtual base is not passed any arguments (because it doesn't actually call
1094 // the inherited constructor).
1095 bool PassedParams = true;
1096 if (const CXXConstructorDecl *CD = dyn_cast<CXXConstructorDecl>(FD))
1097 if (auto Inherited = CD->getInheritedConstructor())
1099 getTypes().inheritingCtorHasParams(Inherited, GD.getCtorType());
1102 for (auto *Param : FD->parameters()) {
1103 Args.push_back(Param);
1104 if (!Param->hasAttr<PassObjectSizeAttr>())
1107 auto *Implicit = ImplicitParamDecl::Create(
1108 getContext(), Param->getDeclContext(), Param->getLocation(),
1109 /*Id=*/nullptr, getContext().getSizeType(), ImplicitParamDecl::Other);
1110 SizeArguments[Param] = Implicit;
1111 Args.push_back(Implicit);
1115 if (MD && (isa<CXXConstructorDecl>(MD) || isa<CXXDestructorDecl>(MD)))
1116 CGM.getCXXABI().addImplicitStructorParams(*this, ResTy, Args);
1122 shouldUseUndefinedBehaviorReturnOptimization(const FunctionDecl *FD,
1123 const ASTContext &Context) {
1124 QualType T = FD->getReturnType();
1125 // Avoid the optimization for functions that return a record type with a
1126 // trivial destructor or another trivially copyable type.
1127 if (const RecordType *RT = T.getCanonicalType()->getAs<RecordType>()) {
1128 if (const auto *ClassDecl = dyn_cast<CXXRecordDecl>(RT->getDecl()))
1129 return !ClassDecl->hasTrivialDestructor();
1131 return !T.isTriviallyCopyableType(Context);
1134 void CodeGenFunction::GenerateCode(GlobalDecl GD, llvm::Function *Fn,
1135 const CGFunctionInfo &FnInfo) {
1136 const FunctionDecl *FD = cast<FunctionDecl>(GD.getDecl());
1139 FunctionArgList Args;
1140 QualType ResTy = BuildFunctionArgList(GD, Args);
1142 // Check if we should generate debug info for this function.
1143 if (FD->hasAttr<NoDebugAttr>())
1144 DebugInfo = nullptr; // disable debug info indefinitely for this function
1146 // The function might not have a body if we're generating thunks for a
1147 // function declaration.
1148 SourceRange BodyRange;
1149 if (Stmt *Body = FD->getBody())
1150 BodyRange = Body->getSourceRange();
1152 BodyRange = FD->getLocation();
1153 CurEHLocation = BodyRange.getEnd();
1155 // Use the location of the start of the function to determine where
1156 // the function definition is located. By default use the location
1157 // of the declaration as the location for the subprogram. A function
1158 // may lack a declaration in the source code if it is created by code
1159 // gen. (examples: _GLOBAL__I_a, __cxx_global_array_dtor, thunk).
1160 SourceLocation Loc = FD->getLocation();
1162 // If this is a function specialization then use the pattern body
1163 // as the location for the function.
1164 if (const FunctionDecl *SpecDecl = FD->getTemplateInstantiationPattern())
1165 if (SpecDecl->hasBody(SpecDecl))
1166 Loc = SpecDecl->getLocation();
1168 Stmt *Body = FD->getBody();
1170 // Initialize helper which will detect jumps which can cause invalid lifetime
1172 if (Body && ShouldEmitLifetimeMarkers)
1173 Bypasses.Init(Body);
1175 // Emit the standard function prologue.
1176 StartFunction(GD, ResTy, Fn, FnInfo, Args, Loc, BodyRange.getBegin());
1178 // Generate the body of the function.
1179 PGO.assignRegionCounters(GD, CurFn);
1180 if (isa<CXXDestructorDecl>(FD))
1181 EmitDestructorBody(Args);
1182 else if (isa<CXXConstructorDecl>(FD))
1183 EmitConstructorBody(Args);
1184 else if (getLangOpts().CUDA &&
1185 !getLangOpts().CUDAIsDevice &&
1186 FD->hasAttr<CUDAGlobalAttr>())
1187 CGM.getCUDARuntime().emitDeviceStub(*this, Args);
1188 else if (isa<CXXConversionDecl>(FD) &&
1189 cast<CXXConversionDecl>(FD)->isLambdaToBlockPointerConversion()) {
1190 // The lambda conversion to block pointer is special; the semantics can't be
1191 // expressed in the AST, so IRGen needs to special-case it.
1192 EmitLambdaToBlockPointerBody(Args);
1193 } else if (isa<CXXMethodDecl>(FD) &&
1194 cast<CXXMethodDecl>(FD)->isLambdaStaticInvoker()) {
1195 // The lambda static invoker function is special, because it forwards or
1196 // clones the body of the function call operator (but is actually static).
1197 EmitLambdaStaticInvokeFunction(cast<CXXMethodDecl>(FD));
1198 } else if (FD->isDefaulted() && isa<CXXMethodDecl>(FD) &&
1199 (cast<CXXMethodDecl>(FD)->isCopyAssignmentOperator() ||
1200 cast<CXXMethodDecl>(FD)->isMoveAssignmentOperator())) {
1201 // Implicit copy-assignment gets the same special treatment as implicit
1202 // copy-constructors.
1203 emitImplicitAssignmentOperatorBody(Args);
1205 EmitFunctionBody(Args, Body);
1207 llvm_unreachable("no definition for emitted function");
1209 // C++11 [stmt.return]p2:
1210 // Flowing off the end of a function [...] results in undefined behavior in
1211 // a value-returning function.
1213 // If the '}' that terminates a function is reached, and the value of the
1214 // function call is used by the caller, the behavior is undefined.
1215 if (getLangOpts().CPlusPlus && !FD->hasImplicitReturnZero() && !SawAsmBlock &&
1216 !FD->getReturnType()->isVoidType() && Builder.GetInsertBlock()) {
1217 bool ShouldEmitUnreachable =
1218 CGM.getCodeGenOpts().StrictReturn ||
1219 shouldUseUndefinedBehaviorReturnOptimization(FD, getContext());
1220 if (SanOpts.has(SanitizerKind::Return)) {
1221 SanitizerScope SanScope(this);
1222 llvm::Value *IsFalse = Builder.getFalse();
1223 EmitCheck(std::make_pair(IsFalse, SanitizerKind::Return),
1224 SanitizerHandler::MissingReturn,
1225 EmitCheckSourceLocation(FD->getLocation()), None);
1226 } else if (ShouldEmitUnreachable) {
1227 if (CGM.getCodeGenOpts().OptimizationLevel == 0)
1228 EmitTrapCall(llvm::Intrinsic::trap);
1230 if (SanOpts.has(SanitizerKind::Return) || ShouldEmitUnreachable) {
1231 Builder.CreateUnreachable();
1232 Builder.ClearInsertionPoint();
1236 // Emit the standard function epilogue.
1237 FinishFunction(BodyRange.getEnd());
1239 // If we haven't marked the function nothrow through other means, do
1240 // a quick pass now to see if we can.
1241 if (!CurFn->doesNotThrow())
1242 TryMarkNoThrow(CurFn);
1245 /// ContainsLabel - Return true if the statement contains a label in it. If
1246 /// this statement is not executed normally, it not containing a label means
1247 /// that we can just remove the code.
1248 bool CodeGenFunction::ContainsLabel(const Stmt *S, bool IgnoreCaseStmts) {
1249 // Null statement, not a label!
1250 if (!S) return false;
1252 // If this is a label, we have to emit the code, consider something like:
1253 // if (0) { ... foo: bar(); } goto foo;
1255 // TODO: If anyone cared, we could track __label__'s, since we know that you
1256 // can't jump to one from outside their declared region.
1257 if (isa<LabelStmt>(S))
1260 // If this is a case/default statement, and we haven't seen a switch, we have
1261 // to emit the code.
1262 if (isa<SwitchCase>(S) && !IgnoreCaseStmts)
1265 // If this is a switch statement, we want to ignore cases below it.
1266 if (isa<SwitchStmt>(S))
1267 IgnoreCaseStmts = true;
1269 // Scan subexpressions for verboten labels.
1270 for (const Stmt *SubStmt : S->children())
1271 if (ContainsLabel(SubStmt, IgnoreCaseStmts))
1277 /// containsBreak - Return true if the statement contains a break out of it.
1278 /// If the statement (recursively) contains a switch or loop with a break
1279 /// inside of it, this is fine.
1280 bool CodeGenFunction::containsBreak(const Stmt *S) {
1281 // Null statement, not a label!
1282 if (!S) return false;
1284 // If this is a switch or loop that defines its own break scope, then we can
1285 // include it and anything inside of it.
1286 if (isa<SwitchStmt>(S) || isa<WhileStmt>(S) || isa<DoStmt>(S) ||
1290 if (isa<BreakStmt>(S))
1293 // Scan subexpressions for verboten breaks.
1294 for (const Stmt *SubStmt : S->children())
1295 if (containsBreak(SubStmt))
1301 bool CodeGenFunction::mightAddDeclToScope(const Stmt *S) {
1302 if (!S) return false;
1304 // Some statement kinds add a scope and thus never add a decl to the current
1305 // scope. Note, this list is longer than the list of statements that might
1306 // have an unscoped decl nested within them, but this way is conservatively
1307 // correct even if more statement kinds are added.
1308 if (isa<IfStmt>(S) || isa<SwitchStmt>(S) || isa<WhileStmt>(S) ||
1309 isa<DoStmt>(S) || isa<ForStmt>(S) || isa<CompoundStmt>(S) ||
1310 isa<CXXForRangeStmt>(S) || isa<CXXTryStmt>(S) ||
1311 isa<ObjCForCollectionStmt>(S) || isa<ObjCAtTryStmt>(S))
1314 if (isa<DeclStmt>(S))
1317 for (const Stmt *SubStmt : S->children())
1318 if (mightAddDeclToScope(SubStmt))
1324 /// ConstantFoldsToSimpleInteger - If the specified expression does not fold
1325 /// to a constant, or if it does but contains a label, return false. If it
1326 /// constant folds return true and set the boolean result in Result.
1327 bool CodeGenFunction::ConstantFoldsToSimpleInteger(const Expr *Cond,
1330 llvm::APSInt ResultInt;
1331 if (!ConstantFoldsToSimpleInteger(Cond, ResultInt, AllowLabels))
1334 ResultBool = ResultInt.getBoolValue();
1338 /// ConstantFoldsToSimpleInteger - If the specified expression does not fold
1339 /// to a constant, or if it does but contains a label, return false. If it
1340 /// constant folds return true and set the folded value.
1341 bool CodeGenFunction::ConstantFoldsToSimpleInteger(const Expr *Cond,
1342 llvm::APSInt &ResultInt,
1344 // FIXME: Rename and handle conversion of other evaluatable things
1347 if (!Cond->EvaluateAsInt(Int, getContext()))
1348 return false; // Not foldable, not integer or not fully evaluatable.
1350 if (!AllowLabels && CodeGenFunction::ContainsLabel(Cond))
1351 return false; // Contains a label.
1359 /// EmitBranchOnBoolExpr - Emit a branch on a boolean condition (e.g. for an if
1360 /// statement) to the specified blocks. Based on the condition, this might try
1361 /// to simplify the codegen of the conditional based on the branch.
1363 void CodeGenFunction::EmitBranchOnBoolExpr(const Expr *Cond,
1364 llvm::BasicBlock *TrueBlock,
1365 llvm::BasicBlock *FalseBlock,
1366 uint64_t TrueCount) {
1367 Cond = Cond->IgnoreParens();
1369 if (const BinaryOperator *CondBOp = dyn_cast<BinaryOperator>(Cond)) {
1371 // Handle X && Y in a condition.
1372 if (CondBOp->getOpcode() == BO_LAnd) {
1373 // If we have "1 && X", simplify the code. "0 && X" would have constant
1374 // folded if the case was simple enough.
1375 bool ConstantBool = false;
1376 if (ConstantFoldsToSimpleInteger(CondBOp->getLHS(), ConstantBool) &&
1378 // br(1 && X) -> br(X).
1379 incrementProfileCounter(CondBOp);
1380 return EmitBranchOnBoolExpr(CondBOp->getRHS(), TrueBlock, FalseBlock,
1384 // If we have "X && 1", simplify the code to use an uncond branch.
1385 // "X && 0" would have been constant folded to 0.
1386 if (ConstantFoldsToSimpleInteger(CondBOp->getRHS(), ConstantBool) &&
1388 // br(X && 1) -> br(X).
1389 return EmitBranchOnBoolExpr(CondBOp->getLHS(), TrueBlock, FalseBlock,
1393 // Emit the LHS as a conditional. If the LHS conditional is false, we
1394 // want to jump to the FalseBlock.
1395 llvm::BasicBlock *LHSTrue = createBasicBlock("land.lhs.true");
1396 // The counter tells us how often we evaluate RHS, and all of TrueCount
1397 // can be propagated to that branch.
1398 uint64_t RHSCount = getProfileCount(CondBOp->getRHS());
1400 ConditionalEvaluation eval(*this);
1402 ApplyDebugLocation DL(*this, Cond);
1403 EmitBranchOnBoolExpr(CondBOp->getLHS(), LHSTrue, FalseBlock, RHSCount);
1407 incrementProfileCounter(CondBOp);
1408 setCurrentProfileCount(getProfileCount(CondBOp->getRHS()));
1410 // Any temporaries created here are conditional.
1412 EmitBranchOnBoolExpr(CondBOp->getRHS(), TrueBlock, FalseBlock, TrueCount);
1418 if (CondBOp->getOpcode() == BO_LOr) {
1419 // If we have "0 || X", simplify the code. "1 || X" would have constant
1420 // folded if the case was simple enough.
1421 bool ConstantBool = false;
1422 if (ConstantFoldsToSimpleInteger(CondBOp->getLHS(), ConstantBool) &&
1424 // br(0 || X) -> br(X).
1425 incrementProfileCounter(CondBOp);
1426 return EmitBranchOnBoolExpr(CondBOp->getRHS(), TrueBlock, FalseBlock,
1430 // If we have "X || 0", simplify the code to use an uncond branch.
1431 // "X || 1" would have been constant folded to 1.
1432 if (ConstantFoldsToSimpleInteger(CondBOp->getRHS(), ConstantBool) &&
1434 // br(X || 0) -> br(X).
1435 return EmitBranchOnBoolExpr(CondBOp->getLHS(), TrueBlock, FalseBlock,
1439 // Emit the LHS as a conditional. If the LHS conditional is true, we
1440 // want to jump to the TrueBlock.
1441 llvm::BasicBlock *LHSFalse = createBasicBlock("lor.lhs.false");
1442 // We have the count for entry to the RHS and for the whole expression
1443 // being true, so we can divy up True count between the short circuit and
1446 getCurrentProfileCount() - getProfileCount(CondBOp->getRHS());
1447 uint64_t RHSCount = TrueCount - LHSCount;
1449 ConditionalEvaluation eval(*this);
1451 ApplyDebugLocation DL(*this, Cond);
1452 EmitBranchOnBoolExpr(CondBOp->getLHS(), TrueBlock, LHSFalse, LHSCount);
1453 EmitBlock(LHSFalse);
1456 incrementProfileCounter(CondBOp);
1457 setCurrentProfileCount(getProfileCount(CondBOp->getRHS()));
1459 // Any temporaries created here are conditional.
1461 EmitBranchOnBoolExpr(CondBOp->getRHS(), TrueBlock, FalseBlock, RHSCount);
1469 if (const UnaryOperator *CondUOp = dyn_cast<UnaryOperator>(Cond)) {
1470 // br(!x, t, f) -> br(x, f, t)
1471 if (CondUOp->getOpcode() == UO_LNot) {
1472 // Negate the count.
1473 uint64_t FalseCount = getCurrentProfileCount() - TrueCount;
1474 // Negate the condition and swap the destination blocks.
1475 return EmitBranchOnBoolExpr(CondUOp->getSubExpr(), FalseBlock, TrueBlock,
1480 if (const ConditionalOperator *CondOp = dyn_cast<ConditionalOperator>(Cond)) {
1481 // br(c ? x : y, t, f) -> br(c, br(x, t, f), br(y, t, f))
1482 llvm::BasicBlock *LHSBlock = createBasicBlock("cond.true");
1483 llvm::BasicBlock *RHSBlock = createBasicBlock("cond.false");
1485 ConditionalEvaluation cond(*this);
1486 EmitBranchOnBoolExpr(CondOp->getCond(), LHSBlock, RHSBlock,
1487 getProfileCount(CondOp));
1489 // When computing PGO branch weights, we only know the overall count for
1490 // the true block. This code is essentially doing tail duplication of the
1491 // naive code-gen, introducing new edges for which counts are not
1492 // available. Divide the counts proportionally between the LHS and RHS of
1493 // the conditional operator.
1494 uint64_t LHSScaledTrueCount = 0;
1497 getProfileCount(CondOp) / (double)getCurrentProfileCount();
1498 LHSScaledTrueCount = TrueCount * LHSRatio;
1502 EmitBlock(LHSBlock);
1503 incrementProfileCounter(CondOp);
1505 ApplyDebugLocation DL(*this, Cond);
1506 EmitBranchOnBoolExpr(CondOp->getLHS(), TrueBlock, FalseBlock,
1507 LHSScaledTrueCount);
1512 EmitBlock(RHSBlock);
1513 EmitBranchOnBoolExpr(CondOp->getRHS(), TrueBlock, FalseBlock,
1514 TrueCount - LHSScaledTrueCount);
1520 if (const CXXThrowExpr *Throw = dyn_cast<CXXThrowExpr>(Cond)) {
1521 // Conditional operator handling can give us a throw expression as a
1522 // condition for a case like:
1523 // br(c ? throw x : y, t, f) -> br(c, br(throw x, t, f), br(y, t, f)
1525 // br(c, throw x, br(y, t, f))
1526 EmitCXXThrowExpr(Throw, /*KeepInsertionPoint*/false);
1530 // If the branch has a condition wrapped by __builtin_unpredictable,
1531 // create metadata that specifies that the branch is unpredictable.
1532 // Don't bother if not optimizing because that metadata would not be used.
1533 llvm::MDNode *Unpredictable = nullptr;
1534 auto *Call = dyn_cast<CallExpr>(Cond);
1535 if (Call && CGM.getCodeGenOpts().OptimizationLevel != 0) {
1536 auto *FD = dyn_cast_or_null<FunctionDecl>(Call->getCalleeDecl());
1537 if (FD && FD->getBuiltinID() == Builtin::BI__builtin_unpredictable) {
1538 llvm::MDBuilder MDHelper(getLLVMContext());
1539 Unpredictable = MDHelper.createUnpredictable();
1543 // Create branch weights based on the number of times we get here and the
1544 // number of times the condition should be true.
1545 uint64_t CurrentCount = std::max(getCurrentProfileCount(), TrueCount);
1546 llvm::MDNode *Weights =
1547 createProfileWeights(TrueCount, CurrentCount - TrueCount);
1549 // Emit the code with the fully general case.
1552 ApplyDebugLocation DL(*this, Cond);
1553 CondV = EvaluateExprAsBool(Cond);
1555 Builder.CreateCondBr(CondV, TrueBlock, FalseBlock, Weights, Unpredictable);
1558 /// ErrorUnsupported - Print out an error that codegen doesn't support the
1559 /// specified stmt yet.
1560 void CodeGenFunction::ErrorUnsupported(const Stmt *S, const char *Type) {
1561 CGM.ErrorUnsupported(S, Type);
1564 /// emitNonZeroVLAInit - Emit the "zero" initialization of a
1565 /// variable-length array whose elements have a non-zero bit-pattern.
1567 /// \param baseType the inner-most element type of the array
1568 /// \param src - a char* pointing to the bit-pattern for a single
1569 /// base element of the array
1570 /// \param sizeInChars - the total size of the VLA, in chars
1571 static void emitNonZeroVLAInit(CodeGenFunction &CGF, QualType baseType,
1572 Address dest, Address src,
1573 llvm::Value *sizeInChars) {
1574 CGBuilderTy &Builder = CGF.Builder;
1576 CharUnits baseSize = CGF.getContext().getTypeSizeInChars(baseType);
1577 llvm::Value *baseSizeInChars
1578 = llvm::ConstantInt::get(CGF.IntPtrTy, baseSize.getQuantity());
1581 Builder.CreateElementBitCast(dest, CGF.Int8Ty, "vla.begin");
1583 Builder.CreateInBoundsGEP(begin.getPointer(), sizeInChars, "vla.end");
1585 llvm::BasicBlock *originBB = CGF.Builder.GetInsertBlock();
1586 llvm::BasicBlock *loopBB = CGF.createBasicBlock("vla-init.loop");
1587 llvm::BasicBlock *contBB = CGF.createBasicBlock("vla-init.cont");
1589 // Make a loop over the VLA. C99 guarantees that the VLA element
1590 // count must be nonzero.
1591 CGF.EmitBlock(loopBB);
1593 llvm::PHINode *cur = Builder.CreatePHI(begin.getType(), 2, "vla.cur");
1594 cur->addIncoming(begin.getPointer(), originBB);
1596 CharUnits curAlign =
1597 dest.getAlignment().alignmentOfArrayElement(baseSize);
1599 // memcpy the individual element bit-pattern.
1600 Builder.CreateMemCpy(Address(cur, curAlign), src, baseSizeInChars,
1601 /*volatile*/ false);
1603 // Go to the next element.
1605 Builder.CreateInBoundsGEP(CGF.Int8Ty, cur, baseSizeInChars, "vla.next");
1607 // Leave if that's the end of the VLA.
1608 llvm::Value *done = Builder.CreateICmpEQ(next, end, "vla-init.isdone");
1609 Builder.CreateCondBr(done, contBB, loopBB);
1610 cur->addIncoming(next, loopBB);
1612 CGF.EmitBlock(contBB);
1616 CodeGenFunction::EmitNullInitialization(Address DestPtr, QualType Ty) {
1617 // Ignore empty classes in C++.
1618 if (getLangOpts().CPlusPlus) {
1619 if (const RecordType *RT = Ty->getAs<RecordType>()) {
1620 if (cast<CXXRecordDecl>(RT->getDecl())->isEmpty())
1625 // Cast the dest ptr to the appropriate i8 pointer type.
1626 if (DestPtr.getElementType() != Int8Ty)
1627 DestPtr = Builder.CreateElementBitCast(DestPtr, Int8Ty);
1629 // Get size and alignment info for this aggregate.
1630 CharUnits size = getContext().getTypeSizeInChars(Ty);
1632 llvm::Value *SizeVal;
1633 const VariableArrayType *vla;
1635 // Don't bother emitting a zero-byte memset.
1636 if (size.isZero()) {
1637 // But note that getTypeInfo returns 0 for a VLA.
1638 if (const VariableArrayType *vlaType =
1639 dyn_cast_or_null<VariableArrayType>(
1640 getContext().getAsArrayType(Ty))) {
1642 llvm::Value *numElts;
1643 std::tie(numElts, eltType) = getVLASize(vlaType);
1646 CharUnits eltSize = getContext().getTypeSizeInChars(eltType);
1647 if (!eltSize.isOne())
1648 SizeVal = Builder.CreateNUWMul(SizeVal, CGM.getSize(eltSize));
1654 SizeVal = CGM.getSize(size);
1658 // If the type contains a pointer to data member we can't memset it to zero.
1659 // Instead, create a null constant and copy it to the destination.
1660 // TODO: there are other patterns besides zero that we can usefully memset,
1661 // like -1, which happens to be the pattern used by member-pointers.
1662 if (!CGM.getTypes().isZeroInitializable(Ty)) {
1663 // For a VLA, emit a single element, then splat that over the VLA.
1664 if (vla) Ty = getContext().getBaseElementType(vla);
1666 llvm::Constant *NullConstant = CGM.EmitNullConstant(Ty);
1668 llvm::GlobalVariable *NullVariable =
1669 new llvm::GlobalVariable(CGM.getModule(), NullConstant->getType(),
1670 /*isConstant=*/true,
1671 llvm::GlobalVariable::PrivateLinkage,
1672 NullConstant, Twine());
1673 CharUnits NullAlign = DestPtr.getAlignment();
1674 NullVariable->setAlignment(NullAlign.getQuantity());
1675 Address SrcPtr(Builder.CreateBitCast(NullVariable, Builder.getInt8PtrTy()),
1678 if (vla) return emitNonZeroVLAInit(*this, Ty, DestPtr, SrcPtr, SizeVal);
1680 // Get and call the appropriate llvm.memcpy overload.
1681 Builder.CreateMemCpy(DestPtr, SrcPtr, SizeVal, false);
1685 // Otherwise, just memset the whole thing to zero. This is legal
1686 // because in LLVM, all default initializers (other than the ones we just
1687 // handled above) are guaranteed to have a bit pattern of all zeros.
1688 Builder.CreateMemSet(DestPtr, Builder.getInt8(0), SizeVal, false);
1691 llvm::BlockAddress *CodeGenFunction::GetAddrOfLabel(const LabelDecl *L) {
1692 // Make sure that there is a block for the indirect goto.
1693 if (!IndirectBranch)
1694 GetIndirectGotoBlock();
1696 llvm::BasicBlock *BB = getJumpDestForLabel(L).getBlock();
1698 // Make sure the indirect branch includes all of the address-taken blocks.
1699 IndirectBranch->addDestination(BB);
1700 return llvm::BlockAddress::get(CurFn, BB);
1703 llvm::BasicBlock *CodeGenFunction::GetIndirectGotoBlock() {
1704 // If we already made the indirect branch for indirect goto, return its block.
1705 if (IndirectBranch) return IndirectBranch->getParent();
1707 CGBuilderTy TmpBuilder(*this, createBasicBlock("indirectgoto"));
1709 // Create the PHI node that indirect gotos will add entries to.
1710 llvm::Value *DestVal = TmpBuilder.CreatePHI(Int8PtrTy, 0,
1711 "indirect.goto.dest");
1713 // Create the indirect branch instruction.
1714 IndirectBranch = TmpBuilder.CreateIndirectBr(DestVal);
1715 return IndirectBranch->getParent();
1718 /// Computes the length of an array in elements, as well as the base
1719 /// element type and a properly-typed first element pointer.
1720 llvm::Value *CodeGenFunction::emitArrayLength(const ArrayType *origArrayType,
1723 const ArrayType *arrayType = origArrayType;
1725 // If it's a VLA, we have to load the stored size. Note that
1726 // this is the size of the VLA in bytes, not its size in elements.
1727 llvm::Value *numVLAElements = nullptr;
1728 if (isa<VariableArrayType>(arrayType)) {
1729 numVLAElements = getVLASize(cast<VariableArrayType>(arrayType)).first;
1731 // Walk into all VLAs. This doesn't require changes to addr,
1732 // which has type T* where T is the first non-VLA element type.
1734 QualType elementType = arrayType->getElementType();
1735 arrayType = getContext().getAsArrayType(elementType);
1737 // If we only have VLA components, 'addr' requires no adjustment.
1739 baseType = elementType;
1740 return numVLAElements;
1742 } while (isa<VariableArrayType>(arrayType));
1744 // We get out here only if we find a constant array type
1748 // We have some number of constant-length arrays, so addr should
1749 // have LLVM type [M x [N x [...]]]*. Build a GEP that walks
1750 // down to the first element of addr.
1751 SmallVector<llvm::Value*, 8> gepIndices;
1753 // GEP down to the array type.
1754 llvm::ConstantInt *zero = Builder.getInt32(0);
1755 gepIndices.push_back(zero);
1757 uint64_t countFromCLAs = 1;
1760 llvm::ArrayType *llvmArrayType =
1761 dyn_cast<llvm::ArrayType>(addr.getElementType());
1762 while (llvmArrayType) {
1763 assert(isa<ConstantArrayType>(arrayType));
1764 assert(cast<ConstantArrayType>(arrayType)->getSize().getZExtValue()
1765 == llvmArrayType->getNumElements());
1767 gepIndices.push_back(zero);
1768 countFromCLAs *= llvmArrayType->getNumElements();
1769 eltType = arrayType->getElementType();
1772 dyn_cast<llvm::ArrayType>(llvmArrayType->getElementType());
1773 arrayType = getContext().getAsArrayType(arrayType->getElementType());
1774 assert((!llvmArrayType || arrayType) &&
1775 "LLVM and Clang types are out-of-synch");
1779 // From this point onwards, the Clang array type has been emitted
1780 // as some other type (probably a packed struct). Compute the array
1781 // size, and just emit the 'begin' expression as a bitcast.
1784 cast<ConstantArrayType>(arrayType)->getSize().getZExtValue();
1785 eltType = arrayType->getElementType();
1786 arrayType = getContext().getAsArrayType(eltType);
1789 llvm::Type *baseType = ConvertType(eltType);
1790 addr = Builder.CreateElementBitCast(addr, baseType, "array.begin");
1792 // Create the actual GEP.
1793 addr = Address(Builder.CreateInBoundsGEP(addr.getPointer(),
1794 gepIndices, "array.begin"),
1795 addr.getAlignment());
1800 llvm::Value *numElements
1801 = llvm::ConstantInt::get(SizeTy, countFromCLAs);
1803 // If we had any VLA dimensions, factor them in.
1805 numElements = Builder.CreateNUWMul(numVLAElements, numElements);
1810 std::pair<llvm::Value*, QualType>
1811 CodeGenFunction::getVLASize(QualType type) {
1812 const VariableArrayType *vla = getContext().getAsVariableArrayType(type);
1813 assert(vla && "type was not a variable array type!");
1814 return getVLASize(vla);
1817 std::pair<llvm::Value*, QualType>
1818 CodeGenFunction::getVLASize(const VariableArrayType *type) {
1819 // The number of elements so far; always size_t.
1820 llvm::Value *numElements = nullptr;
1822 QualType elementType;
1824 elementType = type->getElementType();
1825 llvm::Value *vlaSize = VLASizeMap[type->getSizeExpr()];
1826 assert(vlaSize && "no size for VLA!");
1827 assert(vlaSize->getType() == SizeTy);
1830 numElements = vlaSize;
1832 // It's undefined behavior if this wraps around, so mark it that way.
1833 // FIXME: Teach -fsanitize=undefined to trap this.
1834 numElements = Builder.CreateNUWMul(numElements, vlaSize);
1836 } while ((type = getContext().getAsVariableArrayType(elementType)));
1838 return std::pair<llvm::Value*,QualType>(numElements, elementType);
1841 void CodeGenFunction::EmitVariablyModifiedType(QualType type) {
1842 assert(type->isVariablyModifiedType() &&
1843 "Must pass variably modified type to EmitVLASizes!");
1845 EnsureInsertPoint();
1847 // We're going to walk down into the type and look for VLA
1850 assert(type->isVariablyModifiedType());
1852 const Type *ty = type.getTypePtr();
1853 switch (ty->getTypeClass()) {
1855 #define TYPE(Class, Base)
1856 #define ABSTRACT_TYPE(Class, Base)
1857 #define NON_CANONICAL_TYPE(Class, Base)
1858 #define DEPENDENT_TYPE(Class, Base) case Type::Class:
1859 #define NON_CANONICAL_UNLESS_DEPENDENT_TYPE(Class, Base)
1860 #include "clang/AST/TypeNodes.def"
1861 llvm_unreachable("unexpected dependent type!");
1863 // These types are never variably-modified.
1867 case Type::ExtVector:
1870 case Type::Elaborated:
1871 case Type::TemplateSpecialization:
1872 case Type::ObjCTypeParam:
1873 case Type::ObjCObject:
1874 case Type::ObjCInterface:
1875 case Type::ObjCObjectPointer:
1876 llvm_unreachable("type class is never variably-modified!");
1878 case Type::Adjusted:
1879 type = cast<AdjustedType>(ty)->getAdjustedType();
1883 type = cast<DecayedType>(ty)->getPointeeType();
1887 type = cast<PointerType>(ty)->getPointeeType();
1890 case Type::BlockPointer:
1891 type = cast<BlockPointerType>(ty)->getPointeeType();
1894 case Type::LValueReference:
1895 case Type::RValueReference:
1896 type = cast<ReferenceType>(ty)->getPointeeType();
1899 case Type::MemberPointer:
1900 type = cast<MemberPointerType>(ty)->getPointeeType();
1903 case Type::ConstantArray:
1904 case Type::IncompleteArray:
1905 // Losing element qualification here is fine.
1906 type = cast<ArrayType>(ty)->getElementType();
1909 case Type::VariableArray: {
1910 // Losing element qualification here is fine.
1911 const VariableArrayType *vat = cast<VariableArrayType>(ty);
1913 // Unknown size indication requires no size computation.
1914 // Otherwise, evaluate and record it.
1915 if (const Expr *size = vat->getSizeExpr()) {
1916 // It's possible that we might have emitted this already,
1917 // e.g. with a typedef and a pointer to it.
1918 llvm::Value *&entry = VLASizeMap[size];
1920 llvm::Value *Size = EmitScalarExpr(size);
1923 // If the size is an expression that is not an integer constant
1924 // expression [...] each time it is evaluated it shall have a value
1925 // greater than zero.
1926 if (SanOpts.has(SanitizerKind::VLABound) &&
1927 size->getType()->isSignedIntegerType()) {
1928 SanitizerScope SanScope(this);
1929 llvm::Value *Zero = llvm::Constant::getNullValue(Size->getType());
1930 llvm::Constant *StaticArgs[] = {
1931 EmitCheckSourceLocation(size->getLocStart()),
1932 EmitCheckTypeDescriptor(size->getType())
1934 EmitCheck(std::make_pair(Builder.CreateICmpSGT(Size, Zero),
1935 SanitizerKind::VLABound),
1936 SanitizerHandler::VLABoundNotPositive, StaticArgs, Size);
1939 // Always zexting here would be wrong if it weren't
1940 // undefined behavior to have a negative bound.
1941 entry = Builder.CreateIntCast(Size, SizeTy, /*signed*/ false);
1944 type = vat->getElementType();
1948 case Type::FunctionProto:
1949 case Type::FunctionNoProto:
1950 type = cast<FunctionType>(ty)->getReturnType();
1955 case Type::UnaryTransform:
1956 case Type::Attributed:
1957 case Type::SubstTemplateTypeParm:
1958 case Type::PackExpansion:
1959 // Keep walking after single level desugaring.
1960 type = type.getSingleStepDesugaredType(getContext());
1964 case Type::Decltype:
1966 case Type::DeducedTemplateSpecialization:
1967 // Stop walking: nothing to do.
1970 case Type::TypeOfExpr:
1971 // Stop walking: emit typeof expression.
1972 EmitIgnoredExpr(cast<TypeOfExprType>(ty)->getUnderlyingExpr());
1976 type = cast<AtomicType>(ty)->getValueType();
1980 type = cast<PipeType>(ty)->getElementType();
1983 } while (type->isVariablyModifiedType());
1986 Address CodeGenFunction::EmitVAListRef(const Expr* E) {
1987 if (getContext().getBuiltinVaListType()->isArrayType())
1988 return EmitPointerWithAlignment(E);
1989 return EmitLValue(E).getAddress();
1992 Address CodeGenFunction::EmitMSVAListRef(const Expr *E) {
1993 return EmitLValue(E).getAddress();
1996 void CodeGenFunction::EmitDeclRefExprDbgValue(const DeclRefExpr *E,
1997 const APValue &Init) {
1998 assert(!Init.isUninit() && "Invalid DeclRefExpr initializer!");
1999 if (CGDebugInfo *Dbg = getDebugInfo())
2000 if (CGM.getCodeGenOpts().getDebugInfo() >= codegenoptions::LimitedDebugInfo)
2001 Dbg->EmitGlobalVariable(E->getDecl(), Init);
2004 CodeGenFunction::PeepholeProtection
2005 CodeGenFunction::protectFromPeepholes(RValue rvalue) {
2006 // At the moment, the only aggressive peephole we do in IR gen
2007 // is trunc(zext) folding, but if we add more, we can easily
2008 // extend this protection.
2010 if (!rvalue.isScalar()) return PeepholeProtection();
2011 llvm::Value *value = rvalue.getScalarVal();
2012 if (!isa<llvm::ZExtInst>(value)) return PeepholeProtection();
2014 // Just make an extra bitcast.
2015 assert(HaveInsertPoint());
2016 llvm::Instruction *inst = new llvm::BitCastInst(value, value->getType(), "",
2017 Builder.GetInsertBlock());
2019 PeepholeProtection protection;
2020 protection.Inst = inst;
2024 void CodeGenFunction::unprotectFromPeepholes(PeepholeProtection protection) {
2025 if (!protection.Inst) return;
2027 // In theory, we could try to duplicate the peepholes now, but whatever.
2028 protection.Inst->eraseFromParent();
2031 llvm::Value *CodeGenFunction::EmitAnnotationCall(llvm::Value *AnnotationFn,
2032 llvm::Value *AnnotatedVal,
2033 StringRef AnnotationStr,
2034 SourceLocation Location) {
2035 llvm::Value *Args[4] = {
2037 Builder.CreateBitCast(CGM.EmitAnnotationString(AnnotationStr), Int8PtrTy),
2038 Builder.CreateBitCast(CGM.EmitAnnotationUnit(Location), Int8PtrTy),
2039 CGM.EmitAnnotationLineNo(Location)
2041 return Builder.CreateCall(AnnotationFn, Args);
2044 void CodeGenFunction::EmitVarAnnotations(const VarDecl *D, llvm::Value *V) {
2045 assert(D->hasAttr<AnnotateAttr>() && "no annotate attribute");
2046 // FIXME We create a new bitcast for every annotation because that's what
2047 // llvm-gcc was doing.
2048 for (const auto *I : D->specific_attrs<AnnotateAttr>())
2049 EmitAnnotationCall(CGM.getIntrinsic(llvm::Intrinsic::var_annotation),
2050 Builder.CreateBitCast(V, CGM.Int8PtrTy, V->getName()),
2051 I->getAnnotation(), D->getLocation());
2054 Address CodeGenFunction::EmitFieldAnnotations(const FieldDecl *D,
2056 assert(D->hasAttr<AnnotateAttr>() && "no annotate attribute");
2057 llvm::Value *V = Addr.getPointer();
2058 llvm::Type *VTy = V->getType();
2059 llvm::Value *F = CGM.getIntrinsic(llvm::Intrinsic::ptr_annotation,
2062 for (const auto *I : D->specific_attrs<AnnotateAttr>()) {
2063 // FIXME Always emit the cast inst so we can differentiate between
2064 // annotation on the first field of a struct and annotation on the struct
2066 if (VTy != CGM.Int8PtrTy)
2067 V = Builder.Insert(new llvm::BitCastInst(V, CGM.Int8PtrTy));
2068 V = EmitAnnotationCall(F, V, I->getAnnotation(), D->getLocation());
2069 V = Builder.CreateBitCast(V, VTy);
2072 return Address(V, Addr.getAlignment());
2075 CodeGenFunction::CGCapturedStmtInfo::~CGCapturedStmtInfo() { }
2077 CodeGenFunction::SanitizerScope::SanitizerScope(CodeGenFunction *CGF)
2079 assert(!CGF->IsSanitizerScope);
2080 CGF->IsSanitizerScope = true;
2083 CodeGenFunction::SanitizerScope::~SanitizerScope() {
2084 CGF->IsSanitizerScope = false;
2087 void CodeGenFunction::InsertHelper(llvm::Instruction *I,
2088 const llvm::Twine &Name,
2089 llvm::BasicBlock *BB,
2090 llvm::BasicBlock::iterator InsertPt) const {
2091 LoopStack.InsertHelper(I);
2092 if (IsSanitizerScope)
2093 CGM.getSanitizerMetadata()->disableSanitizerForInstruction(I);
2096 void CGBuilderInserter::InsertHelper(
2097 llvm::Instruction *I, const llvm::Twine &Name, llvm::BasicBlock *BB,
2098 llvm::BasicBlock::iterator InsertPt) const {
2099 llvm::IRBuilderDefaultInserter::InsertHelper(I, Name, BB, InsertPt);
2101 CGF->InsertHelper(I, Name, BB, InsertPt);
2104 static bool hasRequiredFeatures(const SmallVectorImpl<StringRef> &ReqFeatures,
2105 CodeGenModule &CGM, const FunctionDecl *FD,
2106 std::string &FirstMissing) {
2107 // If there aren't any required features listed then go ahead and return.
2108 if (ReqFeatures.empty())
2111 // Now build up the set of caller features and verify that all the required
2112 // features are there.
2113 llvm::StringMap<bool> CallerFeatureMap;
2114 CGM.getFunctionFeatureMap(CallerFeatureMap, FD);
2116 // If we have at least one of the features in the feature list return
2117 // true, otherwise return false.
2119 ReqFeatures.begin(), ReqFeatures.end(), [&](StringRef Feature) {
2120 SmallVector<StringRef, 1> OrFeatures;
2121 Feature.split(OrFeatures, "|");
2122 return std::any_of(OrFeatures.begin(), OrFeatures.end(),
2123 [&](StringRef Feature) {
2124 if (!CallerFeatureMap.lookup(Feature)) {
2125 FirstMissing = Feature.str();
2133 // Emits an error if we don't have a valid set of target features for the
2135 void CodeGenFunction::checkTargetFeatures(const CallExpr *E,
2136 const FunctionDecl *TargetDecl) {
2137 // Early exit if this is an indirect call.
2141 // Get the current enclosing function if it exists. If it doesn't
2142 // we can't check the target features anyhow.
2143 const FunctionDecl *FD = dyn_cast_or_null<FunctionDecl>(CurFuncDecl);
2147 // Grab the required features for the call. For a builtin this is listed in
2148 // the td file with the default cpu, for an always_inline function this is any
2149 // listed cpu and any listed features.
2150 unsigned BuiltinID = TargetDecl->getBuiltinID();
2151 std::string MissingFeature;
2153 SmallVector<StringRef, 1> ReqFeatures;
2154 const char *FeatureList =
2155 CGM.getContext().BuiltinInfo.getRequiredFeatures(BuiltinID);
2156 // Return if the builtin doesn't have any required features.
2157 if (!FeatureList || StringRef(FeatureList) == "")
2159 StringRef(FeatureList).split(ReqFeatures, ",");
2160 if (!hasRequiredFeatures(ReqFeatures, CGM, FD, MissingFeature))
2161 CGM.getDiags().Report(E->getLocStart(), diag::err_builtin_needs_feature)
2162 << TargetDecl->getDeclName()
2163 << CGM.getContext().BuiltinInfo.getRequiredFeatures(BuiltinID);
2165 } else if (TargetDecl->hasAttr<TargetAttr>()) {
2166 // Get the required features for the callee.
2167 SmallVector<StringRef, 1> ReqFeatures;
2168 llvm::StringMap<bool> CalleeFeatureMap;
2169 CGM.getFunctionFeatureMap(CalleeFeatureMap, TargetDecl);
2170 for (const auto &F : CalleeFeatureMap) {
2171 // Only positive features are "required".
2173 ReqFeatures.push_back(F.getKey());
2175 if (!hasRequiredFeatures(ReqFeatures, CGM, FD, MissingFeature))
2176 CGM.getDiags().Report(E->getLocStart(), diag::err_function_needs_feature)
2177 << FD->getDeclName() << TargetDecl->getDeclName() << MissingFeature;
2181 void CodeGenFunction::EmitSanitizerStatReport(llvm::SanitizerStatKind SSK) {
2182 if (!CGM.getCodeGenOpts().SanitizeStats)
2185 llvm::IRBuilder<> IRB(Builder.GetInsertBlock(), Builder.GetInsertPoint());
2186 IRB.SetCurrentDebugLocation(Builder.getCurrentDebugLocation());
2187 CGM.getSanStats().create(IRB, SSK);
2190 llvm::DebugLoc CodeGenFunction::SourceLocToDebugLoc(SourceLocation Location) {
2191 if (CGDebugInfo *DI = getDebugInfo())
2192 return DI->SourceLocToDebugLoc(Location);
2194 return llvm::DebugLoc();