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 LValueBaseInfo *BaseInfo) {
121 return getNaturalTypeAlignment(T->getPointeeType(), BaseInfo,
122 /*forPointee*/ true);
125 CharUnits CodeGenFunction::getNaturalTypeAlignment(QualType T,
126 LValueBaseInfo *BaseInfo,
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()) {
134 *BaseInfo = LValueBaseInfo(AlignmentSource::AttributedType, false);
135 return getContext().toCharUnitsFromBits(Align);
140 *BaseInfo = LValueBaseInfo(AlignmentSource::Type, false);
143 if (T->isIncompleteType()) {
144 Alignment = CharUnits::One(); // Shouldn't be used, but pessimistic is best.
146 // For C++ class pointees, we don't know whether we're pointing at a
147 // base or a complete object, so we generally need to use the
148 // non-virtual alignment.
149 const CXXRecordDecl *RD;
150 if (forPointeeType && (RD = T->getAsCXXRecordDecl())) {
151 Alignment = CGM.getClassPointerAlignment(RD);
153 Alignment = getContext().getTypeAlignInChars(T);
154 if (T.getQualifiers().hasUnaligned())
155 Alignment = CharUnits::One();
158 // Cap to the global maximum type alignment unless the alignment
159 // was somehow explicit on the type.
160 if (unsigned MaxAlign = getLangOpts().MaxTypeAlign) {
161 if (Alignment.getQuantity() > MaxAlign &&
162 !getContext().isAlignmentRequired(T))
163 Alignment = CharUnits::fromQuantity(MaxAlign);
169 LValue CodeGenFunction::MakeNaturalAlignAddrLValue(llvm::Value *V, QualType T) {
170 LValueBaseInfo BaseInfo;
171 CharUnits Alignment = getNaturalTypeAlignment(T, &BaseInfo);
172 return LValue::MakeAddr(Address(V, Alignment), T, getContext(), BaseInfo,
176 /// Given a value of type T* that may not be to a complete object,
177 /// construct an l-value with the natural pointee alignment of T.
179 CodeGenFunction::MakeNaturalAlignPointeeAddrLValue(llvm::Value *V, QualType T) {
180 LValueBaseInfo BaseInfo;
181 CharUnits Align = getNaturalTypeAlignment(T, &BaseInfo, /*pointee*/ true);
182 return MakeAddrLValue(Address(V, Align), T, BaseInfo);
186 llvm::Type *CodeGenFunction::ConvertTypeForMem(QualType T) {
187 return CGM.getTypes().ConvertTypeForMem(T);
190 llvm::Type *CodeGenFunction::ConvertType(QualType T) {
191 return CGM.getTypes().ConvertType(T);
194 TypeEvaluationKind CodeGenFunction::getEvaluationKind(QualType type) {
195 type = type.getCanonicalType();
197 switch (type->getTypeClass()) {
198 #define TYPE(name, parent)
199 #define ABSTRACT_TYPE(name, parent)
200 #define NON_CANONICAL_TYPE(name, parent) case Type::name:
201 #define DEPENDENT_TYPE(name, parent) case Type::name:
202 #define NON_CANONICAL_UNLESS_DEPENDENT_TYPE(name, parent) case Type::name:
203 #include "clang/AST/TypeNodes.def"
204 llvm_unreachable("non-canonical or dependent type in IR-generation");
207 case Type::DeducedTemplateSpecialization:
208 llvm_unreachable("undeduced type in IR-generation");
210 // Various scalar types.
213 case Type::BlockPointer:
214 case Type::LValueReference:
215 case Type::RValueReference:
216 case Type::MemberPointer:
218 case Type::ExtVector:
219 case Type::FunctionProto:
220 case Type::FunctionNoProto:
222 case Type::ObjCObjectPointer:
230 // Arrays, records, and Objective-C objects.
231 case Type::ConstantArray:
232 case Type::IncompleteArray:
233 case Type::VariableArray:
235 case Type::ObjCObject:
236 case Type::ObjCInterface:
237 return TEK_Aggregate;
239 // We operate on atomic values according to their underlying type.
241 type = cast<AtomicType>(type)->getValueType();
244 llvm_unreachable("unknown type kind!");
248 llvm::DebugLoc CodeGenFunction::EmitReturnBlock() {
249 // For cleanliness, we try to avoid emitting the return block for
251 llvm::BasicBlock *CurBB = Builder.GetInsertBlock();
254 assert(!CurBB->getTerminator() && "Unexpected terminated block.");
256 // We have a valid insert point, reuse it if it is empty or there are no
257 // explicit jumps to the return block.
258 if (CurBB->empty() || ReturnBlock.getBlock()->use_empty()) {
259 ReturnBlock.getBlock()->replaceAllUsesWith(CurBB);
260 delete ReturnBlock.getBlock();
262 EmitBlock(ReturnBlock.getBlock());
263 return llvm::DebugLoc();
266 // Otherwise, if the return block is the target of a single direct
267 // branch then we can just put the code in that block instead. This
268 // cleans up functions which started with a unified return block.
269 if (ReturnBlock.getBlock()->hasOneUse()) {
270 llvm::BranchInst *BI =
271 dyn_cast<llvm::BranchInst>(*ReturnBlock.getBlock()->user_begin());
272 if (BI && BI->isUnconditional() &&
273 BI->getSuccessor(0) == ReturnBlock.getBlock()) {
274 // Record/return the DebugLoc of the simple 'return' expression to be used
275 // later by the actual 'ret' instruction.
276 llvm::DebugLoc Loc = BI->getDebugLoc();
277 Builder.SetInsertPoint(BI->getParent());
278 BI->eraseFromParent();
279 delete ReturnBlock.getBlock();
284 // FIXME: We are at an unreachable point, there is no reason to emit the block
285 // unless it has uses. However, we still need a place to put the debug
286 // region.end for now.
288 EmitBlock(ReturnBlock.getBlock());
289 return llvm::DebugLoc();
292 static void EmitIfUsed(CodeGenFunction &CGF, llvm::BasicBlock *BB) {
294 if (!BB->use_empty())
295 return CGF.CurFn->getBasicBlockList().push_back(BB);
299 void CodeGenFunction::FinishFunction(SourceLocation EndLoc) {
300 assert(BreakContinueStack.empty() &&
301 "mismatched push/pop in break/continue stack!");
303 bool OnlySimpleReturnStmts = NumSimpleReturnExprs > 0
304 && NumSimpleReturnExprs == NumReturnExprs
305 && ReturnBlock.getBlock()->use_empty();
306 // Usually the return expression is evaluated before the cleanup
307 // code. If the function contains only a simple return statement,
308 // such as a constant, the location before the cleanup code becomes
309 // the last useful breakpoint in the function, because the simple
310 // return expression will be evaluated after the cleanup code. To be
311 // safe, set the debug location for cleanup code to the location of
312 // the return statement. Otherwise the cleanup code should be at the
313 // end of the function's lexical scope.
315 // If there are multiple branches to the return block, the branch
316 // instructions will get the location of the return statements and
318 if (CGDebugInfo *DI = getDebugInfo()) {
319 if (OnlySimpleReturnStmts)
320 DI->EmitLocation(Builder, LastStopPoint);
322 DI->EmitLocation(Builder, EndLoc);
325 // Pop any cleanups that might have been associated with the
326 // parameters. Do this in whatever block we're currently in; it's
327 // important to do this before we enter the return block or return
328 // edges will be *really* confused.
329 bool HasCleanups = EHStack.stable_begin() != PrologueCleanupDepth;
330 bool HasOnlyLifetimeMarkers =
331 HasCleanups && EHStack.containsOnlyLifetimeMarkers(PrologueCleanupDepth);
332 bool EmitRetDbgLoc = !HasCleanups || HasOnlyLifetimeMarkers;
334 // Make sure the line table doesn't jump back into the body for
335 // the ret after it's been at EndLoc.
336 if (CGDebugInfo *DI = getDebugInfo())
337 if (OnlySimpleReturnStmts)
338 DI->EmitLocation(Builder, EndLoc);
340 PopCleanupBlocks(PrologueCleanupDepth);
343 // Emit function epilog (to return).
344 llvm::DebugLoc Loc = EmitReturnBlock();
346 if (ShouldInstrumentFunction())
347 EmitFunctionInstrumentation("__cyg_profile_func_exit");
349 // Emit debug descriptor for function end.
350 if (CGDebugInfo *DI = getDebugInfo())
351 DI->EmitFunctionEnd(Builder, CurFn);
353 // Reset the debug location to that of the simple 'return' expression, if any
354 // rather than that of the end of the function's scope '}'.
355 ApplyDebugLocation AL(*this, Loc);
356 EmitFunctionEpilog(*CurFnInfo, EmitRetDbgLoc, EndLoc);
357 EmitEndEHSpec(CurCodeDecl);
359 assert(EHStack.empty() &&
360 "did not remove all scopes from cleanup stack!");
362 // If someone did an indirect goto, emit the indirect goto block at the end of
364 if (IndirectBranch) {
365 EmitBlock(IndirectBranch->getParent());
366 Builder.ClearInsertionPoint();
369 // If some of our locals escaped, insert a call to llvm.localescape in the
371 if (!EscapedLocals.empty()) {
372 // Invert the map from local to index into a simple vector. There should be
374 SmallVector<llvm::Value *, 4> EscapeArgs;
375 EscapeArgs.resize(EscapedLocals.size());
376 for (auto &Pair : EscapedLocals)
377 EscapeArgs[Pair.second] = Pair.first;
378 llvm::Function *FrameEscapeFn = llvm::Intrinsic::getDeclaration(
379 &CGM.getModule(), llvm::Intrinsic::localescape);
380 CGBuilderTy(*this, AllocaInsertPt).CreateCall(FrameEscapeFn, EscapeArgs);
383 // Remove the AllocaInsertPt instruction, which is just a convenience for us.
384 llvm::Instruction *Ptr = AllocaInsertPt;
385 AllocaInsertPt = nullptr;
386 Ptr->eraseFromParent();
388 // If someone took the address of a label but never did an indirect goto, we
389 // made a zero entry PHI node, which is illegal, zap it now.
390 if (IndirectBranch) {
391 llvm::PHINode *PN = cast<llvm::PHINode>(IndirectBranch->getAddress());
392 if (PN->getNumIncomingValues() == 0) {
393 PN->replaceAllUsesWith(llvm::UndefValue::get(PN->getType()));
394 PN->eraseFromParent();
398 EmitIfUsed(*this, EHResumeBlock);
399 EmitIfUsed(*this, TerminateLandingPad);
400 EmitIfUsed(*this, TerminateHandler);
401 EmitIfUsed(*this, UnreachableBlock);
403 if (CGM.getCodeGenOpts().EmitDeclMetadata)
406 for (SmallVectorImpl<std::pair<llvm::Instruction *, llvm::Value *> >::iterator
407 I = DeferredReplacements.begin(),
408 E = DeferredReplacements.end();
410 I->first->replaceAllUsesWith(I->second);
411 I->first->eraseFromParent();
415 /// ShouldInstrumentFunction - Return true if the current function should be
416 /// instrumented with __cyg_profile_func_* calls
417 bool CodeGenFunction::ShouldInstrumentFunction() {
418 if (!CGM.getCodeGenOpts().InstrumentFunctions)
420 if (!CurFuncDecl || CurFuncDecl->hasAttr<NoInstrumentFunctionAttr>())
425 /// ShouldXRayInstrument - Return true if the current function should be
426 /// instrumented with XRay nop sleds.
427 bool CodeGenFunction::ShouldXRayInstrumentFunction() const {
428 return CGM.getCodeGenOpts().XRayInstrumentFunctions;
431 /// EmitFunctionInstrumentation - Emit LLVM code to call the specified
432 /// instrumentation function with the current function and the call site, if
433 /// function instrumentation is enabled.
434 void CodeGenFunction::EmitFunctionInstrumentation(const char *Fn) {
435 auto NL = ApplyDebugLocation::CreateArtificial(*this);
436 // void __cyg_profile_func_{enter,exit} (void *this_fn, void *call_site);
437 llvm::PointerType *PointerTy = Int8PtrTy;
438 llvm::Type *ProfileFuncArgs[] = { PointerTy, PointerTy };
439 llvm::FunctionType *FunctionTy =
440 llvm::FunctionType::get(VoidTy, ProfileFuncArgs, false);
442 llvm::Constant *F = CGM.CreateRuntimeFunction(FunctionTy, Fn);
443 llvm::CallInst *CallSite = Builder.CreateCall(
444 CGM.getIntrinsic(llvm::Intrinsic::returnaddress),
445 llvm::ConstantInt::get(Int32Ty, 0),
448 llvm::Value *args[] = {
449 llvm::ConstantExpr::getBitCast(CurFn, PointerTy),
453 EmitNounwindRuntimeCall(F, args);
456 static void removeImageAccessQualifier(std::string& TyName) {
457 std::string ReadOnlyQual("__read_only");
458 std::string::size_type ReadOnlyPos = TyName.find(ReadOnlyQual);
459 if (ReadOnlyPos != std::string::npos)
460 // "+ 1" for the space after access qualifier.
461 TyName.erase(ReadOnlyPos, ReadOnlyQual.size() + 1);
463 std::string WriteOnlyQual("__write_only");
464 std::string::size_type WriteOnlyPos = TyName.find(WriteOnlyQual);
465 if (WriteOnlyPos != std::string::npos)
466 TyName.erase(WriteOnlyPos, WriteOnlyQual.size() + 1);
468 std::string ReadWriteQual("__read_write");
469 std::string::size_type ReadWritePos = TyName.find(ReadWriteQual);
470 if (ReadWritePos != std::string::npos)
471 TyName.erase(ReadWritePos, ReadWriteQual.size() + 1);
476 // Returns the address space id that should be produced to the
477 // kernel_arg_addr_space metadata. This is always fixed to the ids
478 // as specified in the SPIR 2.0 specification in order to differentiate
479 // for example in clGetKernelArgInfo() implementation between the address
480 // spaces with targets without unique mapping to the OpenCL address spaces
481 // (basically all single AS CPUs).
482 static unsigned ArgInfoAddressSpace(unsigned LangAS) {
484 case LangAS::opencl_global: return 1;
485 case LangAS::opencl_constant: return 2;
486 case LangAS::opencl_local: return 3;
487 case LangAS::opencl_generic: return 4; // Not in SPIR 2.0 specs.
489 return 0; // Assume private.
493 // OpenCL v1.2 s5.6.4.6 allows the compiler to store kernel argument
494 // information in the program executable. The argument information stored
495 // includes the argument name, its type, the address and access qualifiers used.
496 static void GenOpenCLArgMetadata(const FunctionDecl *FD, llvm::Function *Fn,
497 CodeGenModule &CGM, llvm::LLVMContext &Context,
498 CGBuilderTy &Builder, ASTContext &ASTCtx) {
499 // Create MDNodes that represent the kernel arg metadata.
500 // Each MDNode is a list in the form of "key", N number of values which is
501 // the same number of values as their are kernel arguments.
503 const PrintingPolicy &Policy = ASTCtx.getPrintingPolicy();
505 // MDNode for the kernel argument address space qualifiers.
506 SmallVector<llvm::Metadata *, 8> addressQuals;
508 // MDNode for the kernel argument access qualifiers (images only).
509 SmallVector<llvm::Metadata *, 8> accessQuals;
511 // MDNode for the kernel argument type names.
512 SmallVector<llvm::Metadata *, 8> argTypeNames;
514 // MDNode for the kernel argument base type names.
515 SmallVector<llvm::Metadata *, 8> argBaseTypeNames;
517 // MDNode for the kernel argument type qualifiers.
518 SmallVector<llvm::Metadata *, 8> argTypeQuals;
520 // MDNode for the kernel argument names.
521 SmallVector<llvm::Metadata *, 8> argNames;
523 for (unsigned i = 0, e = FD->getNumParams(); i != e; ++i) {
524 const ParmVarDecl *parm = FD->getParamDecl(i);
525 QualType ty = parm->getType();
526 std::string typeQuals;
528 if (ty->isPointerType()) {
529 QualType pointeeTy = ty->getPointeeType();
531 // Get address qualifier.
532 addressQuals.push_back(llvm::ConstantAsMetadata::get(Builder.getInt32(
533 ArgInfoAddressSpace(pointeeTy.getAddressSpace()))));
535 // Get argument type name.
536 std::string typeName =
537 pointeeTy.getUnqualifiedType().getAsString(Policy) + "*";
539 // Turn "unsigned type" to "utype"
540 std::string::size_type pos = typeName.find("unsigned");
541 if (pointeeTy.isCanonical() && pos != std::string::npos)
542 typeName.erase(pos+1, 8);
544 argTypeNames.push_back(llvm::MDString::get(Context, typeName));
546 std::string baseTypeName =
547 pointeeTy.getUnqualifiedType().getCanonicalType().getAsString(
551 // Turn "unsigned type" to "utype"
552 pos = baseTypeName.find("unsigned");
553 if (pos != std::string::npos)
554 baseTypeName.erase(pos+1, 8);
556 argBaseTypeNames.push_back(llvm::MDString::get(Context, baseTypeName));
558 // Get argument type qualifiers:
559 if (ty.isRestrictQualified())
560 typeQuals = "restrict";
561 if (pointeeTy.isConstQualified() ||
562 (pointeeTy.getAddressSpace() == LangAS::opencl_constant))
563 typeQuals += typeQuals.empty() ? "const" : " const";
564 if (pointeeTy.isVolatileQualified())
565 typeQuals += typeQuals.empty() ? "volatile" : " volatile";
567 uint32_t AddrSpc = 0;
568 bool isPipe = ty->isPipeType();
569 if (ty->isImageType() || isPipe)
570 AddrSpc = ArgInfoAddressSpace(LangAS::opencl_global);
572 addressQuals.push_back(
573 llvm::ConstantAsMetadata::get(Builder.getInt32(AddrSpc)));
575 // Get argument type name.
576 std::string typeName;
578 typeName = ty.getCanonicalType()->getAs<PipeType>()->getElementType()
579 .getAsString(Policy);
581 typeName = ty.getUnqualifiedType().getAsString(Policy);
583 // Turn "unsigned type" to "utype"
584 std::string::size_type pos = typeName.find("unsigned");
585 if (ty.isCanonical() && pos != std::string::npos)
586 typeName.erase(pos+1, 8);
588 std::string baseTypeName;
590 baseTypeName = ty.getCanonicalType()->getAs<PipeType>()
591 ->getElementType().getCanonicalType()
592 .getAsString(Policy);
595 ty.getUnqualifiedType().getCanonicalType().getAsString(Policy);
597 // Remove access qualifiers on images
598 // (as they are inseparable from type in clang implementation,
599 // but OpenCL spec provides a special query to get access qualifier
600 // via clGetKernelArgInfo with CL_KERNEL_ARG_ACCESS_QUALIFIER):
601 if (ty->isImageType()) {
602 removeImageAccessQualifier(typeName);
603 removeImageAccessQualifier(baseTypeName);
606 argTypeNames.push_back(llvm::MDString::get(Context, typeName));
608 // Turn "unsigned type" to "utype"
609 pos = baseTypeName.find("unsigned");
610 if (pos != std::string::npos)
611 baseTypeName.erase(pos+1, 8);
613 argBaseTypeNames.push_back(llvm::MDString::get(Context, baseTypeName));
619 argTypeQuals.push_back(llvm::MDString::get(Context, typeQuals));
621 // Get image and pipe access qualifier:
622 if (ty->isImageType()|| ty->isPipeType()) {
623 const OpenCLAccessAttr *A = parm->getAttr<OpenCLAccessAttr>();
624 if (A && A->isWriteOnly())
625 accessQuals.push_back(llvm::MDString::get(Context, "write_only"));
626 else if (A && A->isReadWrite())
627 accessQuals.push_back(llvm::MDString::get(Context, "read_write"));
629 accessQuals.push_back(llvm::MDString::get(Context, "read_only"));
631 accessQuals.push_back(llvm::MDString::get(Context, "none"));
633 // Get argument name.
634 argNames.push_back(llvm::MDString::get(Context, parm->getName()));
637 Fn->setMetadata("kernel_arg_addr_space",
638 llvm::MDNode::get(Context, addressQuals));
639 Fn->setMetadata("kernel_arg_access_qual",
640 llvm::MDNode::get(Context, accessQuals));
641 Fn->setMetadata("kernel_arg_type",
642 llvm::MDNode::get(Context, argTypeNames));
643 Fn->setMetadata("kernel_arg_base_type",
644 llvm::MDNode::get(Context, argBaseTypeNames));
645 Fn->setMetadata("kernel_arg_type_qual",
646 llvm::MDNode::get(Context, argTypeQuals));
647 if (CGM.getCodeGenOpts().EmitOpenCLArgMetadata)
648 Fn->setMetadata("kernel_arg_name",
649 llvm::MDNode::get(Context, argNames));
652 void CodeGenFunction::EmitOpenCLKernelMetadata(const FunctionDecl *FD,
655 if (!FD->hasAttr<OpenCLKernelAttr>())
658 llvm::LLVMContext &Context = getLLVMContext();
660 GenOpenCLArgMetadata(FD, Fn, CGM, Context, Builder, getContext());
662 if (const VecTypeHintAttr *A = FD->getAttr<VecTypeHintAttr>()) {
663 QualType HintQTy = A->getTypeHint();
664 const ExtVectorType *HintEltQTy = HintQTy->getAs<ExtVectorType>();
665 bool IsSignedInteger =
666 HintQTy->isSignedIntegerType() ||
667 (HintEltQTy && HintEltQTy->getElementType()->isSignedIntegerType());
668 llvm::Metadata *AttrMDArgs[] = {
669 llvm::ConstantAsMetadata::get(llvm::UndefValue::get(
670 CGM.getTypes().ConvertType(A->getTypeHint()))),
671 llvm::ConstantAsMetadata::get(llvm::ConstantInt::get(
672 llvm::IntegerType::get(Context, 32),
673 llvm::APInt(32, (uint64_t)(IsSignedInteger ? 1 : 0))))};
674 Fn->setMetadata("vec_type_hint", llvm::MDNode::get(Context, AttrMDArgs));
677 if (const WorkGroupSizeHintAttr *A = FD->getAttr<WorkGroupSizeHintAttr>()) {
678 llvm::Metadata *AttrMDArgs[] = {
679 llvm::ConstantAsMetadata::get(Builder.getInt32(A->getXDim())),
680 llvm::ConstantAsMetadata::get(Builder.getInt32(A->getYDim())),
681 llvm::ConstantAsMetadata::get(Builder.getInt32(A->getZDim()))};
682 Fn->setMetadata("work_group_size_hint", llvm::MDNode::get(Context, AttrMDArgs));
685 if (const ReqdWorkGroupSizeAttr *A = FD->getAttr<ReqdWorkGroupSizeAttr>()) {
686 llvm::Metadata *AttrMDArgs[] = {
687 llvm::ConstantAsMetadata::get(Builder.getInt32(A->getXDim())),
688 llvm::ConstantAsMetadata::get(Builder.getInt32(A->getYDim())),
689 llvm::ConstantAsMetadata::get(Builder.getInt32(A->getZDim()))};
690 Fn->setMetadata("reqd_work_group_size", llvm::MDNode::get(Context, AttrMDArgs));
693 if (const OpenCLIntelReqdSubGroupSizeAttr *A =
694 FD->getAttr<OpenCLIntelReqdSubGroupSizeAttr>()) {
695 llvm::Metadata *AttrMDArgs[] = {
696 llvm::ConstantAsMetadata::get(Builder.getInt32(A->getSubGroupSize()))};
697 Fn->setMetadata("intel_reqd_sub_group_size",
698 llvm::MDNode::get(Context, AttrMDArgs));
702 /// Determine whether the function F ends with a return stmt.
703 static bool endsWithReturn(const Decl* F) {
704 const Stmt *Body = nullptr;
705 if (auto *FD = dyn_cast_or_null<FunctionDecl>(F))
706 Body = FD->getBody();
707 else if (auto *OMD = dyn_cast_or_null<ObjCMethodDecl>(F))
708 Body = OMD->getBody();
710 if (auto *CS = dyn_cast_or_null<CompoundStmt>(Body)) {
711 auto LastStmt = CS->body_rbegin();
712 if (LastStmt != CS->body_rend())
713 return isa<ReturnStmt>(*LastStmt);
718 static void markAsIgnoreThreadCheckingAtRuntime(llvm::Function *Fn) {
719 Fn->addFnAttr("sanitize_thread_no_checking_at_run_time");
720 Fn->removeFnAttr(llvm::Attribute::SanitizeThread);
723 void CodeGenFunction::StartFunction(GlobalDecl GD,
726 const CGFunctionInfo &FnInfo,
727 const FunctionArgList &Args,
729 SourceLocation StartLoc) {
731 "Do not use a CodeGenFunction object for more than one function");
733 const Decl *D = GD.getDecl();
735 DidCallStackSave = false;
737 if (const auto *FD = dyn_cast_or_null<FunctionDecl>(D))
738 if (FD->usesSEHTry())
740 CurFuncDecl = (D ? D->getNonClosureContext() : nullptr);
744 assert(CurFn->isDeclaration() && "Function already has body?");
746 if (CGM.isInSanitizerBlacklist(Fn, Loc))
750 // Apply the no_sanitize* attributes to SanOpts.
751 for (auto Attr : D->specific_attrs<NoSanitizeAttr>())
752 SanOpts.Mask &= ~Attr->getMask();
755 // Apply sanitizer attributes to the function.
756 if (SanOpts.hasOneOf(SanitizerKind::Address | SanitizerKind::KernelAddress))
757 Fn->addFnAttr(llvm::Attribute::SanitizeAddress);
758 if (SanOpts.has(SanitizerKind::Thread))
759 Fn->addFnAttr(llvm::Attribute::SanitizeThread);
760 if (SanOpts.has(SanitizerKind::Memory))
761 Fn->addFnAttr(llvm::Attribute::SanitizeMemory);
762 if (SanOpts.has(SanitizerKind::SafeStack))
763 Fn->addFnAttr(llvm::Attribute::SafeStack);
765 // Ignore TSan memory acesses from within ObjC/ObjC++ dealloc, initialize,
766 // .cxx_destruct, __destroy_helper_block_ and all of their calees at run time.
767 if (SanOpts.has(SanitizerKind::Thread)) {
768 if (const auto *OMD = dyn_cast_or_null<ObjCMethodDecl>(D)) {
769 IdentifierInfo *II = OMD->getSelector().getIdentifierInfoForSlot(0);
770 if (OMD->getMethodFamily() == OMF_dealloc ||
771 OMD->getMethodFamily() == OMF_initialize ||
772 (OMD->getSelector().isUnarySelector() && II->isStr(".cxx_destruct"))) {
773 markAsIgnoreThreadCheckingAtRuntime(Fn);
775 } else if (const auto *FD = dyn_cast_or_null<FunctionDecl>(D)) {
776 IdentifierInfo *II = FD->getIdentifier();
777 if (II && II->isStr("__destroy_helper_block_"))
778 markAsIgnoreThreadCheckingAtRuntime(Fn);
782 // Apply xray attributes to the function (as a string, for now)
783 if (D && ShouldXRayInstrumentFunction()) {
784 if (const auto *XRayAttr = D->getAttr<XRayInstrumentAttr>()) {
785 if (XRayAttr->alwaysXRayInstrument())
786 Fn->addFnAttr("function-instrument", "xray-always");
787 if (XRayAttr->neverXRayInstrument())
788 Fn->addFnAttr("function-instrument", "xray-never");
789 if (const auto *LogArgs = D->getAttr<XRayLogArgsAttr>()) {
790 Fn->addFnAttr("xray-log-args",
791 llvm::utostr(LogArgs->getArgumentCount()));
794 if (!CGM.imbueXRayAttrs(Fn, Loc))
796 "xray-instruction-threshold",
797 llvm::itostr(CGM.getCodeGenOpts().XRayInstructionThreshold));
801 if (const FunctionDecl *FD = dyn_cast_or_null<FunctionDecl>(D))
802 if (CGM.getLangOpts().OpenMP && FD->hasAttr<OMPDeclareSimdDeclAttr>())
803 CGM.getOpenMPRuntime().emitDeclareSimdFunction(FD, Fn);
805 // Add no-jump-tables value.
806 Fn->addFnAttr("no-jump-tables",
807 llvm::toStringRef(CGM.getCodeGenOpts().NoUseJumpTables));
809 if (getLangOpts().OpenCL) {
810 // Add metadata for a kernel function.
811 if (const FunctionDecl *FD = dyn_cast_or_null<FunctionDecl>(D))
812 EmitOpenCLKernelMetadata(FD, Fn);
815 // If we are checking function types, emit a function type signature as
817 if (getLangOpts().CPlusPlus && SanOpts.has(SanitizerKind::Function)) {
818 if (const FunctionDecl *FD = dyn_cast_or_null<FunctionDecl>(D)) {
819 if (llvm::Constant *PrologueSig =
820 CGM.getTargetCodeGenInfo().getUBSanFunctionSignature(CGM)) {
821 llvm::Constant *FTRTTIConst =
822 CGM.GetAddrOfRTTIDescriptor(FD->getType(), /*ForEH=*/true);
823 llvm::Constant *PrologueStructElems[] = { PrologueSig, FTRTTIConst };
824 llvm::Constant *PrologueStructConst =
825 llvm::ConstantStruct::getAnon(PrologueStructElems, /*Packed=*/true);
826 Fn->setPrologueData(PrologueStructConst);
831 // If we're checking nullability, we need to know whether we can check the
832 // return value. Initialize the flag to 'true' and refine it in EmitParmDecl.
833 if (SanOpts.has(SanitizerKind::NullabilityReturn)) {
834 auto Nullability = FnRetTy->getNullability(getContext());
835 if (Nullability && *Nullability == NullabilityKind::NonNull) {
836 if (!(SanOpts.has(SanitizerKind::ReturnsNonnullAttribute) &&
837 CurCodeDecl && CurCodeDecl->getAttr<ReturnsNonNullAttr>()))
838 RetValNullabilityPrecondition =
839 llvm::ConstantInt::getTrue(getLLVMContext());
843 // If we're in C++ mode and the function name is "main", it is guaranteed
844 // to be norecurse by the standard (3.6.1.3 "The function main shall not be
845 // used within a program").
846 if (getLangOpts().CPlusPlus)
847 if (const FunctionDecl *FD = dyn_cast_or_null<FunctionDecl>(D))
849 Fn->addFnAttr(llvm::Attribute::NoRecurse);
851 llvm::BasicBlock *EntryBB = createBasicBlock("entry", CurFn);
853 // Create a marker to make it easy to insert allocas into the entryblock
854 // later. Don't create this with the builder, because we don't want it
856 llvm::Value *Undef = llvm::UndefValue::get(Int32Ty);
857 AllocaInsertPt = new llvm::BitCastInst(Undef, Int32Ty, "allocapt", EntryBB);
859 ReturnBlock = getJumpDestInCurrentScope("return");
861 Builder.SetInsertPoint(EntryBB);
863 // Emit subprogram debug descriptor.
864 if (CGDebugInfo *DI = getDebugInfo()) {
865 // Reconstruct the type from the argument list so that implicit parameters,
866 // such as 'this' and 'vtt', show up in the debug info. Preserve the calling
868 CallingConv CC = CallingConv::CC_C;
869 if (auto *FD = dyn_cast_or_null<FunctionDecl>(D))
870 if (const auto *SrcFnTy = FD->getType()->getAs<FunctionType>())
871 CC = SrcFnTy->getCallConv();
872 SmallVector<QualType, 16> ArgTypes;
873 for (const VarDecl *VD : Args)
874 ArgTypes.push_back(VD->getType());
875 QualType FnType = getContext().getFunctionType(
876 RetTy, ArgTypes, FunctionProtoType::ExtProtoInfo(CC));
877 DI->EmitFunctionStart(GD, Loc, StartLoc, FnType, CurFn, Builder);
880 if (ShouldInstrumentFunction())
881 EmitFunctionInstrumentation("__cyg_profile_func_enter");
883 // Since emitting the mcount call here impacts optimizations such as function
884 // inlining, we just add an attribute to insert a mcount call in backend.
885 // The attribute "counting-function" is set to mcount function name which is
886 // architecture dependent.
887 if (CGM.getCodeGenOpts().InstrumentForProfiling) {
888 if (CGM.getCodeGenOpts().CallFEntry)
889 Fn->addFnAttr("fentry-call", "true");
891 Fn->addFnAttr("counting-function", getTarget().getMCountName());
894 if (RetTy->isVoidType()) {
895 // Void type; nothing to return.
896 ReturnValue = Address::invalid();
898 // Count the implicit return.
899 if (!endsWithReturn(D))
901 } else if (CurFnInfo->getReturnInfo().getKind() == ABIArgInfo::Indirect &&
902 !hasScalarEvaluationKind(CurFnInfo->getReturnType())) {
903 // Indirect aggregate return; emit returned value directly into sret slot.
904 // This reduces code size, and affects correctness in C++.
905 auto AI = CurFn->arg_begin();
906 if (CurFnInfo->getReturnInfo().isSRetAfterThis())
908 ReturnValue = Address(&*AI, CurFnInfo->getReturnInfo().getIndirectAlign());
909 } else if (CurFnInfo->getReturnInfo().getKind() == ABIArgInfo::InAlloca &&
910 !hasScalarEvaluationKind(CurFnInfo->getReturnType())) {
911 // Load the sret pointer from the argument struct and return into that.
912 unsigned Idx = CurFnInfo->getReturnInfo().getInAllocaFieldIndex();
913 llvm::Function::arg_iterator EI = CurFn->arg_end();
915 llvm::Value *Addr = Builder.CreateStructGEP(nullptr, &*EI, Idx);
916 Addr = Builder.CreateAlignedLoad(Addr, getPointerAlign(), "agg.result");
917 ReturnValue = Address(Addr, getNaturalTypeAlignment(RetTy));
919 ReturnValue = CreateIRTemp(RetTy, "retval");
921 // Tell the epilog emitter to autorelease the result. We do this
922 // now so that various specialized functions can suppress it
923 // during their IR-generation.
924 if (getLangOpts().ObjCAutoRefCount &&
925 !CurFnInfo->isReturnsRetained() &&
926 RetTy->isObjCRetainableType())
927 AutoreleaseResult = true;
930 EmitStartEHSpec(CurCodeDecl);
932 PrologueCleanupDepth = EHStack.stable_begin();
933 EmitFunctionProlog(*CurFnInfo, CurFn, Args);
935 if (D && isa<CXXMethodDecl>(D) && cast<CXXMethodDecl>(D)->isInstance()) {
936 CGM.getCXXABI().EmitInstanceFunctionProlog(*this);
937 const CXXMethodDecl *MD = cast<CXXMethodDecl>(D);
938 if (MD->getParent()->isLambda() &&
939 MD->getOverloadedOperator() == OO_Call) {
940 // We're in a lambda; figure out the captures.
941 MD->getParent()->getCaptureFields(LambdaCaptureFields,
942 LambdaThisCaptureField);
943 if (LambdaThisCaptureField) {
944 // If the lambda captures the object referred to by '*this' - either by
945 // value or by reference, make sure CXXThisValue points to the correct
948 // Get the lvalue for the field (which is a copy of the enclosing object
949 // or contains the address of the enclosing object).
950 LValue ThisFieldLValue = EmitLValueForLambdaField(LambdaThisCaptureField);
951 if (!LambdaThisCaptureField->getType()->isPointerType()) {
952 // If the enclosing object was captured by value, just use its address.
953 CXXThisValue = ThisFieldLValue.getAddress().getPointer();
955 // Load the lvalue pointed to by the field, since '*this' was captured
958 EmitLoadOfLValue(ThisFieldLValue, SourceLocation()).getScalarVal();
961 for (auto *FD : MD->getParent()->fields()) {
962 if (FD->hasCapturedVLAType()) {
963 auto *ExprArg = EmitLoadOfLValue(EmitLValueForLambdaField(FD),
964 SourceLocation()).getScalarVal();
965 auto VAT = FD->getCapturedVLAType();
966 VLASizeMap[VAT->getSizeExpr()] = ExprArg;
970 // Not in a lambda; just use 'this' from the method.
971 // FIXME: Should we generate a new load for each use of 'this'? The
972 // fast register allocator would be happier...
973 CXXThisValue = CXXABIThisValue;
976 // Check the 'this' pointer once per function, if it's available.
978 SanitizerSet SkippedChecks;
979 SkippedChecks.set(SanitizerKind::ObjectSize, true);
980 QualType ThisTy = MD->getThisType(getContext());
981 EmitTypeCheck(TCK_Load, Loc, CXXThisValue, ThisTy,
982 getContext().getTypeAlignInChars(ThisTy->getPointeeType()),
987 // If any of the arguments have a variably modified type, make sure to
988 // emit the type size.
989 for (FunctionArgList::const_iterator i = Args.begin(), e = Args.end();
991 const VarDecl *VD = *i;
993 // Dig out the type as written from ParmVarDecls; it's unclear whether
994 // the standard (C99 6.9.1p10) requires this, but we're following the
995 // precedent set by gcc.
997 if (const ParmVarDecl *PVD = dyn_cast<ParmVarDecl>(VD))
998 Ty = PVD->getOriginalType();
1002 if (Ty->isVariablyModifiedType())
1003 EmitVariablyModifiedType(Ty);
1005 // Emit a location at the end of the prologue.
1006 if (CGDebugInfo *DI = getDebugInfo())
1007 DI->EmitLocation(Builder, StartLoc);
1010 void CodeGenFunction::EmitFunctionBody(FunctionArgList &Args,
1012 incrementProfileCounter(Body);
1013 if (const CompoundStmt *S = dyn_cast<CompoundStmt>(Body))
1014 EmitCompoundStmtWithoutScope(*S);
1019 /// When instrumenting to collect profile data, the counts for some blocks
1020 /// such as switch cases need to not include the fall-through counts, so
1021 /// emit a branch around the instrumentation code. When not instrumenting,
1022 /// this just calls EmitBlock().
1023 void CodeGenFunction::EmitBlockWithFallThrough(llvm::BasicBlock *BB,
1025 llvm::BasicBlock *SkipCountBB = nullptr;
1026 if (HaveInsertPoint() && CGM.getCodeGenOpts().hasProfileClangInstr()) {
1027 // When instrumenting for profiling, the fallthrough to certain
1028 // statements needs to skip over the instrumentation code so that we
1029 // get an accurate count.
1030 SkipCountBB = createBasicBlock("skipcount");
1031 EmitBranch(SkipCountBB);
1034 uint64_t CurrentCount = getCurrentProfileCount();
1035 incrementProfileCounter(S);
1036 setCurrentProfileCount(getCurrentProfileCount() + CurrentCount);
1038 EmitBlock(SkipCountBB);
1041 /// Tries to mark the given function nounwind based on the
1042 /// non-existence of any throwing calls within it. We believe this is
1043 /// lightweight enough to do at -O0.
1044 static void TryMarkNoThrow(llvm::Function *F) {
1045 // LLVM treats 'nounwind' on a function as part of the type, so we
1046 // can't do this on functions that can be overwritten.
1047 if (F->isInterposable()) return;
1049 for (llvm::BasicBlock &BB : *F)
1050 for (llvm::Instruction &I : BB)
1054 F->setDoesNotThrow();
1057 QualType CodeGenFunction::BuildFunctionArgList(GlobalDecl GD,
1058 FunctionArgList &Args) {
1059 const FunctionDecl *FD = cast<FunctionDecl>(GD.getDecl());
1060 QualType ResTy = FD->getReturnType();
1062 const CXXMethodDecl *MD = dyn_cast<CXXMethodDecl>(FD);
1063 if (MD && MD->isInstance()) {
1064 if (CGM.getCXXABI().HasThisReturn(GD))
1065 ResTy = MD->getThisType(getContext());
1066 else if (CGM.getCXXABI().hasMostDerivedReturn(GD))
1067 ResTy = CGM.getContext().VoidPtrTy;
1068 CGM.getCXXABI().buildThisParam(*this, Args);
1071 // The base version of an inheriting constructor whose constructed base is a
1072 // virtual base is not passed any arguments (because it doesn't actually call
1073 // the inherited constructor).
1074 bool PassedParams = true;
1075 if (const CXXConstructorDecl *CD = dyn_cast<CXXConstructorDecl>(FD))
1076 if (auto Inherited = CD->getInheritedConstructor())
1078 getTypes().inheritingCtorHasParams(Inherited, GD.getCtorType());
1081 for (auto *Param : FD->parameters()) {
1082 Args.push_back(Param);
1083 if (!Param->hasAttr<PassObjectSizeAttr>())
1086 IdentifierInfo *NoID = nullptr;
1087 auto *Implicit = ImplicitParamDecl::Create(
1088 getContext(), Param->getDeclContext(), Param->getLocation(), NoID,
1089 getContext().getSizeType());
1090 SizeArguments[Param] = Implicit;
1091 Args.push_back(Implicit);
1095 if (MD && (isa<CXXConstructorDecl>(MD) || isa<CXXDestructorDecl>(MD)))
1096 CGM.getCXXABI().addImplicitStructorParams(*this, ResTy, Args);
1102 shouldUseUndefinedBehaviorReturnOptimization(const FunctionDecl *FD,
1103 const ASTContext &Context) {
1104 QualType T = FD->getReturnType();
1105 // Avoid the optimization for functions that return a record type with a
1106 // trivial destructor or another trivially copyable type.
1107 if (const RecordType *RT = T.getCanonicalType()->getAs<RecordType>()) {
1108 if (const auto *ClassDecl = dyn_cast<CXXRecordDecl>(RT->getDecl()))
1109 return !ClassDecl->hasTrivialDestructor();
1111 return !T.isTriviallyCopyableType(Context);
1114 void CodeGenFunction::GenerateCode(GlobalDecl GD, llvm::Function *Fn,
1115 const CGFunctionInfo &FnInfo) {
1116 const FunctionDecl *FD = cast<FunctionDecl>(GD.getDecl());
1119 FunctionArgList Args;
1120 QualType ResTy = BuildFunctionArgList(GD, Args);
1122 // Check if we should generate debug info for this function.
1123 if (FD->hasAttr<NoDebugAttr>())
1124 DebugInfo = nullptr; // disable debug info indefinitely for this function
1126 // The function might not have a body if we're generating thunks for a
1127 // function declaration.
1128 SourceRange BodyRange;
1129 if (Stmt *Body = FD->getBody())
1130 BodyRange = Body->getSourceRange();
1132 BodyRange = FD->getLocation();
1133 CurEHLocation = BodyRange.getEnd();
1135 // Use the location of the start of the function to determine where
1136 // the function definition is located. By default use the location
1137 // of the declaration as the location for the subprogram. A function
1138 // may lack a declaration in the source code if it is created by code
1139 // gen. (examples: _GLOBAL__I_a, __cxx_global_array_dtor, thunk).
1140 SourceLocation Loc = FD->getLocation();
1142 // If this is a function specialization then use the pattern body
1143 // as the location for the function.
1144 if (const FunctionDecl *SpecDecl = FD->getTemplateInstantiationPattern())
1145 if (SpecDecl->hasBody(SpecDecl))
1146 Loc = SpecDecl->getLocation();
1148 Stmt *Body = FD->getBody();
1150 // Initialize helper which will detect jumps which can cause invalid lifetime
1152 if (Body && ShouldEmitLifetimeMarkers)
1153 Bypasses.Init(Body);
1155 // Emit the standard function prologue.
1156 StartFunction(GD, ResTy, Fn, FnInfo, Args, Loc, BodyRange.getBegin());
1158 // Generate the body of the function.
1159 PGO.assignRegionCounters(GD, CurFn);
1160 if (isa<CXXDestructorDecl>(FD))
1161 EmitDestructorBody(Args);
1162 else if (isa<CXXConstructorDecl>(FD))
1163 EmitConstructorBody(Args);
1164 else if (getLangOpts().CUDA &&
1165 !getLangOpts().CUDAIsDevice &&
1166 FD->hasAttr<CUDAGlobalAttr>())
1167 CGM.getCUDARuntime().emitDeviceStub(*this, Args);
1168 else if (isa<CXXConversionDecl>(FD) &&
1169 cast<CXXConversionDecl>(FD)->isLambdaToBlockPointerConversion()) {
1170 // The lambda conversion to block pointer is special; the semantics can't be
1171 // expressed in the AST, so IRGen needs to special-case it.
1172 EmitLambdaToBlockPointerBody(Args);
1173 } else if (isa<CXXMethodDecl>(FD) &&
1174 cast<CXXMethodDecl>(FD)->isLambdaStaticInvoker()) {
1175 // The lambda static invoker function is special, because it forwards or
1176 // clones the body of the function call operator (but is actually static).
1177 EmitLambdaStaticInvokeFunction(cast<CXXMethodDecl>(FD));
1178 } else if (FD->isDefaulted() && isa<CXXMethodDecl>(FD) &&
1179 (cast<CXXMethodDecl>(FD)->isCopyAssignmentOperator() ||
1180 cast<CXXMethodDecl>(FD)->isMoveAssignmentOperator())) {
1181 // Implicit copy-assignment gets the same special treatment as implicit
1182 // copy-constructors.
1183 emitImplicitAssignmentOperatorBody(Args);
1185 EmitFunctionBody(Args, Body);
1187 llvm_unreachable("no definition for emitted function");
1189 // C++11 [stmt.return]p2:
1190 // Flowing off the end of a function [...] results in undefined behavior in
1191 // a value-returning function.
1193 // If the '}' that terminates a function is reached, and the value of the
1194 // function call is used by the caller, the behavior is undefined.
1195 if (getLangOpts().CPlusPlus && !FD->hasImplicitReturnZero() && !SawAsmBlock &&
1196 !FD->getReturnType()->isVoidType() && Builder.GetInsertBlock()) {
1197 bool ShouldEmitUnreachable =
1198 CGM.getCodeGenOpts().StrictReturn ||
1199 shouldUseUndefinedBehaviorReturnOptimization(FD, getContext());
1200 if (SanOpts.has(SanitizerKind::Return)) {
1201 SanitizerScope SanScope(this);
1202 llvm::Value *IsFalse = Builder.getFalse();
1203 EmitCheck(std::make_pair(IsFalse, SanitizerKind::Return),
1204 SanitizerHandler::MissingReturn,
1205 EmitCheckSourceLocation(FD->getLocation()), None);
1206 } else if (ShouldEmitUnreachable) {
1207 if (CGM.getCodeGenOpts().OptimizationLevel == 0)
1208 EmitTrapCall(llvm::Intrinsic::trap);
1210 if (SanOpts.has(SanitizerKind::Return) || ShouldEmitUnreachable) {
1211 Builder.CreateUnreachable();
1212 Builder.ClearInsertionPoint();
1216 // Emit the standard function epilogue.
1217 FinishFunction(BodyRange.getEnd());
1219 // If we haven't marked the function nothrow through other means, do
1220 // a quick pass now to see if we can.
1221 if (!CurFn->doesNotThrow())
1222 TryMarkNoThrow(CurFn);
1225 /// ContainsLabel - Return true if the statement contains a label in it. If
1226 /// this statement is not executed normally, it not containing a label means
1227 /// that we can just remove the code.
1228 bool CodeGenFunction::ContainsLabel(const Stmt *S, bool IgnoreCaseStmts) {
1229 // Null statement, not a label!
1230 if (!S) return false;
1232 // If this is a label, we have to emit the code, consider something like:
1233 // if (0) { ... foo: bar(); } goto foo;
1235 // TODO: If anyone cared, we could track __label__'s, since we know that you
1236 // can't jump to one from outside their declared region.
1237 if (isa<LabelStmt>(S))
1240 // If this is a case/default statement, and we haven't seen a switch, we have
1241 // to emit the code.
1242 if (isa<SwitchCase>(S) && !IgnoreCaseStmts)
1245 // If this is a switch statement, we want to ignore cases below it.
1246 if (isa<SwitchStmt>(S))
1247 IgnoreCaseStmts = true;
1249 // Scan subexpressions for verboten labels.
1250 for (const Stmt *SubStmt : S->children())
1251 if (ContainsLabel(SubStmt, IgnoreCaseStmts))
1257 /// containsBreak - Return true if the statement contains a break out of it.
1258 /// If the statement (recursively) contains a switch or loop with a break
1259 /// inside of it, this is fine.
1260 bool CodeGenFunction::containsBreak(const Stmt *S) {
1261 // Null statement, not a label!
1262 if (!S) return false;
1264 // If this is a switch or loop that defines its own break scope, then we can
1265 // include it and anything inside of it.
1266 if (isa<SwitchStmt>(S) || isa<WhileStmt>(S) || isa<DoStmt>(S) ||
1270 if (isa<BreakStmt>(S))
1273 // Scan subexpressions for verboten breaks.
1274 for (const Stmt *SubStmt : S->children())
1275 if (containsBreak(SubStmt))
1281 bool CodeGenFunction::mightAddDeclToScope(const Stmt *S) {
1282 if (!S) return false;
1284 // Some statement kinds add a scope and thus never add a decl to the current
1285 // scope. Note, this list is longer than the list of statements that might
1286 // have an unscoped decl nested within them, but this way is conservatively
1287 // correct even if more statement kinds are added.
1288 if (isa<IfStmt>(S) || isa<SwitchStmt>(S) || isa<WhileStmt>(S) ||
1289 isa<DoStmt>(S) || isa<ForStmt>(S) || isa<CompoundStmt>(S) ||
1290 isa<CXXForRangeStmt>(S) || isa<CXXTryStmt>(S) ||
1291 isa<ObjCForCollectionStmt>(S) || isa<ObjCAtTryStmt>(S))
1294 if (isa<DeclStmt>(S))
1297 for (const Stmt *SubStmt : S->children())
1298 if (mightAddDeclToScope(SubStmt))
1304 /// ConstantFoldsToSimpleInteger - If the specified expression does not fold
1305 /// to a constant, or if it does but contains a label, return false. If it
1306 /// constant folds return true and set the boolean result in Result.
1307 bool CodeGenFunction::ConstantFoldsToSimpleInteger(const Expr *Cond,
1310 llvm::APSInt ResultInt;
1311 if (!ConstantFoldsToSimpleInteger(Cond, ResultInt, AllowLabels))
1314 ResultBool = ResultInt.getBoolValue();
1318 /// ConstantFoldsToSimpleInteger - If the specified expression does not fold
1319 /// to a constant, or if it does but contains a label, return false. If it
1320 /// constant folds return true and set the folded value.
1321 bool CodeGenFunction::ConstantFoldsToSimpleInteger(const Expr *Cond,
1322 llvm::APSInt &ResultInt,
1324 // FIXME: Rename and handle conversion of other evaluatable things
1327 if (!Cond->EvaluateAsInt(Int, getContext()))
1328 return false; // Not foldable, not integer or not fully evaluatable.
1330 if (!AllowLabels && CodeGenFunction::ContainsLabel(Cond))
1331 return false; // Contains a label.
1339 /// EmitBranchOnBoolExpr - Emit a branch on a boolean condition (e.g. for an if
1340 /// statement) to the specified blocks. Based on the condition, this might try
1341 /// to simplify the codegen of the conditional based on the branch.
1343 void CodeGenFunction::EmitBranchOnBoolExpr(const Expr *Cond,
1344 llvm::BasicBlock *TrueBlock,
1345 llvm::BasicBlock *FalseBlock,
1346 uint64_t TrueCount) {
1347 Cond = Cond->IgnoreParens();
1349 if (const BinaryOperator *CondBOp = dyn_cast<BinaryOperator>(Cond)) {
1351 // Handle X && Y in a condition.
1352 if (CondBOp->getOpcode() == BO_LAnd) {
1353 // If we have "1 && X", simplify the code. "0 && X" would have constant
1354 // folded if the case was simple enough.
1355 bool ConstantBool = false;
1356 if (ConstantFoldsToSimpleInteger(CondBOp->getLHS(), ConstantBool) &&
1358 // br(1 && X) -> br(X).
1359 incrementProfileCounter(CondBOp);
1360 return EmitBranchOnBoolExpr(CondBOp->getRHS(), TrueBlock, FalseBlock,
1364 // If we have "X && 1", simplify the code to use an uncond branch.
1365 // "X && 0" would have been constant folded to 0.
1366 if (ConstantFoldsToSimpleInteger(CondBOp->getRHS(), ConstantBool) &&
1368 // br(X && 1) -> br(X).
1369 return EmitBranchOnBoolExpr(CondBOp->getLHS(), TrueBlock, FalseBlock,
1373 // Emit the LHS as a conditional. If the LHS conditional is false, we
1374 // want to jump to the FalseBlock.
1375 llvm::BasicBlock *LHSTrue = createBasicBlock("land.lhs.true");
1376 // The counter tells us how often we evaluate RHS, and all of TrueCount
1377 // can be propagated to that branch.
1378 uint64_t RHSCount = getProfileCount(CondBOp->getRHS());
1380 ConditionalEvaluation eval(*this);
1382 ApplyDebugLocation DL(*this, Cond);
1383 EmitBranchOnBoolExpr(CondBOp->getLHS(), LHSTrue, FalseBlock, RHSCount);
1387 incrementProfileCounter(CondBOp);
1388 setCurrentProfileCount(getProfileCount(CondBOp->getRHS()));
1390 // Any temporaries created here are conditional.
1392 EmitBranchOnBoolExpr(CondBOp->getRHS(), TrueBlock, FalseBlock, TrueCount);
1398 if (CondBOp->getOpcode() == BO_LOr) {
1399 // If we have "0 || X", simplify the code. "1 || X" would have constant
1400 // folded if the case was simple enough.
1401 bool ConstantBool = false;
1402 if (ConstantFoldsToSimpleInteger(CondBOp->getLHS(), ConstantBool) &&
1404 // br(0 || X) -> br(X).
1405 incrementProfileCounter(CondBOp);
1406 return EmitBranchOnBoolExpr(CondBOp->getRHS(), TrueBlock, FalseBlock,
1410 // If we have "X || 0", simplify the code to use an uncond branch.
1411 // "X || 1" would have been constant folded to 1.
1412 if (ConstantFoldsToSimpleInteger(CondBOp->getRHS(), ConstantBool) &&
1414 // br(X || 0) -> br(X).
1415 return EmitBranchOnBoolExpr(CondBOp->getLHS(), TrueBlock, FalseBlock,
1419 // Emit the LHS as a conditional. If the LHS conditional is true, we
1420 // want to jump to the TrueBlock.
1421 llvm::BasicBlock *LHSFalse = createBasicBlock("lor.lhs.false");
1422 // We have the count for entry to the RHS and for the whole expression
1423 // being true, so we can divy up True count between the short circuit and
1426 getCurrentProfileCount() - getProfileCount(CondBOp->getRHS());
1427 uint64_t RHSCount = TrueCount - LHSCount;
1429 ConditionalEvaluation eval(*this);
1431 ApplyDebugLocation DL(*this, Cond);
1432 EmitBranchOnBoolExpr(CondBOp->getLHS(), TrueBlock, LHSFalse, LHSCount);
1433 EmitBlock(LHSFalse);
1436 incrementProfileCounter(CondBOp);
1437 setCurrentProfileCount(getProfileCount(CondBOp->getRHS()));
1439 // Any temporaries created here are conditional.
1441 EmitBranchOnBoolExpr(CondBOp->getRHS(), TrueBlock, FalseBlock, RHSCount);
1449 if (const UnaryOperator *CondUOp = dyn_cast<UnaryOperator>(Cond)) {
1450 // br(!x, t, f) -> br(x, f, t)
1451 if (CondUOp->getOpcode() == UO_LNot) {
1452 // Negate the count.
1453 uint64_t FalseCount = getCurrentProfileCount() - TrueCount;
1454 // Negate the condition and swap the destination blocks.
1455 return EmitBranchOnBoolExpr(CondUOp->getSubExpr(), FalseBlock, TrueBlock,
1460 if (const ConditionalOperator *CondOp = dyn_cast<ConditionalOperator>(Cond)) {
1461 // br(c ? x : y, t, f) -> br(c, br(x, t, f), br(y, t, f))
1462 llvm::BasicBlock *LHSBlock = createBasicBlock("cond.true");
1463 llvm::BasicBlock *RHSBlock = createBasicBlock("cond.false");
1465 ConditionalEvaluation cond(*this);
1466 EmitBranchOnBoolExpr(CondOp->getCond(), LHSBlock, RHSBlock,
1467 getProfileCount(CondOp));
1469 // When computing PGO branch weights, we only know the overall count for
1470 // the true block. This code is essentially doing tail duplication of the
1471 // naive code-gen, introducing new edges for which counts are not
1472 // available. Divide the counts proportionally between the LHS and RHS of
1473 // the conditional operator.
1474 uint64_t LHSScaledTrueCount = 0;
1477 getProfileCount(CondOp) / (double)getCurrentProfileCount();
1478 LHSScaledTrueCount = TrueCount * LHSRatio;
1482 EmitBlock(LHSBlock);
1483 incrementProfileCounter(CondOp);
1485 ApplyDebugLocation DL(*this, Cond);
1486 EmitBranchOnBoolExpr(CondOp->getLHS(), TrueBlock, FalseBlock,
1487 LHSScaledTrueCount);
1492 EmitBlock(RHSBlock);
1493 EmitBranchOnBoolExpr(CondOp->getRHS(), TrueBlock, FalseBlock,
1494 TrueCount - LHSScaledTrueCount);
1500 if (const CXXThrowExpr *Throw = dyn_cast<CXXThrowExpr>(Cond)) {
1501 // Conditional operator handling can give us a throw expression as a
1502 // condition for a case like:
1503 // br(c ? throw x : y, t, f) -> br(c, br(throw x, t, f), br(y, t, f)
1505 // br(c, throw x, br(y, t, f))
1506 EmitCXXThrowExpr(Throw, /*KeepInsertionPoint*/false);
1510 // If the branch has a condition wrapped by __builtin_unpredictable,
1511 // create metadata that specifies that the branch is unpredictable.
1512 // Don't bother if not optimizing because that metadata would not be used.
1513 llvm::MDNode *Unpredictable = nullptr;
1514 auto *Call = dyn_cast<CallExpr>(Cond);
1515 if (Call && CGM.getCodeGenOpts().OptimizationLevel != 0) {
1516 auto *FD = dyn_cast_or_null<FunctionDecl>(Call->getCalleeDecl());
1517 if (FD && FD->getBuiltinID() == Builtin::BI__builtin_unpredictable) {
1518 llvm::MDBuilder MDHelper(getLLVMContext());
1519 Unpredictable = MDHelper.createUnpredictable();
1523 // Create branch weights based on the number of times we get here and the
1524 // number of times the condition should be true.
1525 uint64_t CurrentCount = std::max(getCurrentProfileCount(), TrueCount);
1526 llvm::MDNode *Weights =
1527 createProfileWeights(TrueCount, CurrentCount - TrueCount);
1529 // Emit the code with the fully general case.
1532 ApplyDebugLocation DL(*this, Cond);
1533 CondV = EvaluateExprAsBool(Cond);
1535 Builder.CreateCondBr(CondV, TrueBlock, FalseBlock, Weights, Unpredictable);
1538 /// ErrorUnsupported - Print out an error that codegen doesn't support the
1539 /// specified stmt yet.
1540 void CodeGenFunction::ErrorUnsupported(const Stmt *S, const char *Type) {
1541 CGM.ErrorUnsupported(S, Type);
1544 /// emitNonZeroVLAInit - Emit the "zero" initialization of a
1545 /// variable-length array whose elements have a non-zero bit-pattern.
1547 /// \param baseType the inner-most element type of the array
1548 /// \param src - a char* pointing to the bit-pattern for a single
1549 /// base element of the array
1550 /// \param sizeInChars - the total size of the VLA, in chars
1551 static void emitNonZeroVLAInit(CodeGenFunction &CGF, QualType baseType,
1552 Address dest, Address src,
1553 llvm::Value *sizeInChars) {
1554 CGBuilderTy &Builder = CGF.Builder;
1556 CharUnits baseSize = CGF.getContext().getTypeSizeInChars(baseType);
1557 llvm::Value *baseSizeInChars
1558 = llvm::ConstantInt::get(CGF.IntPtrTy, baseSize.getQuantity());
1561 Builder.CreateElementBitCast(dest, CGF.Int8Ty, "vla.begin");
1563 Builder.CreateInBoundsGEP(begin.getPointer(), sizeInChars, "vla.end");
1565 llvm::BasicBlock *originBB = CGF.Builder.GetInsertBlock();
1566 llvm::BasicBlock *loopBB = CGF.createBasicBlock("vla-init.loop");
1567 llvm::BasicBlock *contBB = CGF.createBasicBlock("vla-init.cont");
1569 // Make a loop over the VLA. C99 guarantees that the VLA element
1570 // count must be nonzero.
1571 CGF.EmitBlock(loopBB);
1573 llvm::PHINode *cur = Builder.CreatePHI(begin.getType(), 2, "vla.cur");
1574 cur->addIncoming(begin.getPointer(), originBB);
1576 CharUnits curAlign =
1577 dest.getAlignment().alignmentOfArrayElement(baseSize);
1579 // memcpy the individual element bit-pattern.
1580 Builder.CreateMemCpy(Address(cur, curAlign), src, baseSizeInChars,
1581 /*volatile*/ false);
1583 // Go to the next element.
1585 Builder.CreateInBoundsGEP(CGF.Int8Ty, cur, baseSizeInChars, "vla.next");
1587 // Leave if that's the end of the VLA.
1588 llvm::Value *done = Builder.CreateICmpEQ(next, end, "vla-init.isdone");
1589 Builder.CreateCondBr(done, contBB, loopBB);
1590 cur->addIncoming(next, loopBB);
1592 CGF.EmitBlock(contBB);
1596 CodeGenFunction::EmitNullInitialization(Address DestPtr, QualType Ty) {
1597 // Ignore empty classes in C++.
1598 if (getLangOpts().CPlusPlus) {
1599 if (const RecordType *RT = Ty->getAs<RecordType>()) {
1600 if (cast<CXXRecordDecl>(RT->getDecl())->isEmpty())
1605 // Cast the dest ptr to the appropriate i8 pointer type.
1606 if (DestPtr.getElementType() != Int8Ty)
1607 DestPtr = Builder.CreateElementBitCast(DestPtr, Int8Ty);
1609 // Get size and alignment info for this aggregate.
1610 CharUnits size = getContext().getTypeSizeInChars(Ty);
1612 llvm::Value *SizeVal;
1613 const VariableArrayType *vla;
1615 // Don't bother emitting a zero-byte memset.
1616 if (size.isZero()) {
1617 // But note that getTypeInfo returns 0 for a VLA.
1618 if (const VariableArrayType *vlaType =
1619 dyn_cast_or_null<VariableArrayType>(
1620 getContext().getAsArrayType(Ty))) {
1622 llvm::Value *numElts;
1623 std::tie(numElts, eltType) = getVLASize(vlaType);
1626 CharUnits eltSize = getContext().getTypeSizeInChars(eltType);
1627 if (!eltSize.isOne())
1628 SizeVal = Builder.CreateNUWMul(SizeVal, CGM.getSize(eltSize));
1634 SizeVal = CGM.getSize(size);
1638 // If the type contains a pointer to data member we can't memset it to zero.
1639 // Instead, create a null constant and copy it to the destination.
1640 // TODO: there are other patterns besides zero that we can usefully memset,
1641 // like -1, which happens to be the pattern used by member-pointers.
1642 if (!CGM.getTypes().isZeroInitializable(Ty)) {
1643 // For a VLA, emit a single element, then splat that over the VLA.
1644 if (vla) Ty = getContext().getBaseElementType(vla);
1646 llvm::Constant *NullConstant = CGM.EmitNullConstant(Ty);
1648 llvm::GlobalVariable *NullVariable =
1649 new llvm::GlobalVariable(CGM.getModule(), NullConstant->getType(),
1650 /*isConstant=*/true,
1651 llvm::GlobalVariable::PrivateLinkage,
1652 NullConstant, Twine());
1653 CharUnits NullAlign = DestPtr.getAlignment();
1654 NullVariable->setAlignment(NullAlign.getQuantity());
1655 Address SrcPtr(Builder.CreateBitCast(NullVariable, Builder.getInt8PtrTy()),
1658 if (vla) return emitNonZeroVLAInit(*this, Ty, DestPtr, SrcPtr, SizeVal);
1660 // Get and call the appropriate llvm.memcpy overload.
1661 Builder.CreateMemCpy(DestPtr, SrcPtr, SizeVal, false);
1665 // Otherwise, just memset the whole thing to zero. This is legal
1666 // because in LLVM, all default initializers (other than the ones we just
1667 // handled above) are guaranteed to have a bit pattern of all zeros.
1668 Builder.CreateMemSet(DestPtr, Builder.getInt8(0), SizeVal, false);
1671 llvm::BlockAddress *CodeGenFunction::GetAddrOfLabel(const LabelDecl *L) {
1672 // Make sure that there is a block for the indirect goto.
1673 if (!IndirectBranch)
1674 GetIndirectGotoBlock();
1676 llvm::BasicBlock *BB = getJumpDestForLabel(L).getBlock();
1678 // Make sure the indirect branch includes all of the address-taken blocks.
1679 IndirectBranch->addDestination(BB);
1680 return llvm::BlockAddress::get(CurFn, BB);
1683 llvm::BasicBlock *CodeGenFunction::GetIndirectGotoBlock() {
1684 // If we already made the indirect branch for indirect goto, return its block.
1685 if (IndirectBranch) return IndirectBranch->getParent();
1687 CGBuilderTy TmpBuilder(*this, createBasicBlock("indirectgoto"));
1689 // Create the PHI node that indirect gotos will add entries to.
1690 llvm::Value *DestVal = TmpBuilder.CreatePHI(Int8PtrTy, 0,
1691 "indirect.goto.dest");
1693 // Create the indirect branch instruction.
1694 IndirectBranch = TmpBuilder.CreateIndirectBr(DestVal);
1695 return IndirectBranch->getParent();
1698 /// Computes the length of an array in elements, as well as the base
1699 /// element type and a properly-typed first element pointer.
1700 llvm::Value *CodeGenFunction::emitArrayLength(const ArrayType *origArrayType,
1703 const ArrayType *arrayType = origArrayType;
1705 // If it's a VLA, we have to load the stored size. Note that
1706 // this is the size of the VLA in bytes, not its size in elements.
1707 llvm::Value *numVLAElements = nullptr;
1708 if (isa<VariableArrayType>(arrayType)) {
1709 numVLAElements = getVLASize(cast<VariableArrayType>(arrayType)).first;
1711 // Walk into all VLAs. This doesn't require changes to addr,
1712 // which has type T* where T is the first non-VLA element type.
1714 QualType elementType = arrayType->getElementType();
1715 arrayType = getContext().getAsArrayType(elementType);
1717 // If we only have VLA components, 'addr' requires no adjustment.
1719 baseType = elementType;
1720 return numVLAElements;
1722 } while (isa<VariableArrayType>(arrayType));
1724 // We get out here only if we find a constant array type
1728 // We have some number of constant-length arrays, so addr should
1729 // have LLVM type [M x [N x [...]]]*. Build a GEP that walks
1730 // down to the first element of addr.
1731 SmallVector<llvm::Value*, 8> gepIndices;
1733 // GEP down to the array type.
1734 llvm::ConstantInt *zero = Builder.getInt32(0);
1735 gepIndices.push_back(zero);
1737 uint64_t countFromCLAs = 1;
1740 llvm::ArrayType *llvmArrayType =
1741 dyn_cast<llvm::ArrayType>(addr.getElementType());
1742 while (llvmArrayType) {
1743 assert(isa<ConstantArrayType>(arrayType));
1744 assert(cast<ConstantArrayType>(arrayType)->getSize().getZExtValue()
1745 == llvmArrayType->getNumElements());
1747 gepIndices.push_back(zero);
1748 countFromCLAs *= llvmArrayType->getNumElements();
1749 eltType = arrayType->getElementType();
1752 dyn_cast<llvm::ArrayType>(llvmArrayType->getElementType());
1753 arrayType = getContext().getAsArrayType(arrayType->getElementType());
1754 assert((!llvmArrayType || arrayType) &&
1755 "LLVM and Clang types are out-of-synch");
1759 // From this point onwards, the Clang array type has been emitted
1760 // as some other type (probably a packed struct). Compute the array
1761 // size, and just emit the 'begin' expression as a bitcast.
1764 cast<ConstantArrayType>(arrayType)->getSize().getZExtValue();
1765 eltType = arrayType->getElementType();
1766 arrayType = getContext().getAsArrayType(eltType);
1769 llvm::Type *baseType = ConvertType(eltType);
1770 addr = Builder.CreateElementBitCast(addr, baseType, "array.begin");
1772 // Create the actual GEP.
1773 addr = Address(Builder.CreateInBoundsGEP(addr.getPointer(),
1774 gepIndices, "array.begin"),
1775 addr.getAlignment());
1780 llvm::Value *numElements
1781 = llvm::ConstantInt::get(SizeTy, countFromCLAs);
1783 // If we had any VLA dimensions, factor them in.
1785 numElements = Builder.CreateNUWMul(numVLAElements, numElements);
1790 std::pair<llvm::Value*, QualType>
1791 CodeGenFunction::getVLASize(QualType type) {
1792 const VariableArrayType *vla = getContext().getAsVariableArrayType(type);
1793 assert(vla && "type was not a variable array type!");
1794 return getVLASize(vla);
1797 std::pair<llvm::Value*, QualType>
1798 CodeGenFunction::getVLASize(const VariableArrayType *type) {
1799 // The number of elements so far; always size_t.
1800 llvm::Value *numElements = nullptr;
1802 QualType elementType;
1804 elementType = type->getElementType();
1805 llvm::Value *vlaSize = VLASizeMap[type->getSizeExpr()];
1806 assert(vlaSize && "no size for VLA!");
1807 assert(vlaSize->getType() == SizeTy);
1810 numElements = vlaSize;
1812 // It's undefined behavior if this wraps around, so mark it that way.
1813 // FIXME: Teach -fsanitize=undefined to trap this.
1814 numElements = Builder.CreateNUWMul(numElements, vlaSize);
1816 } while ((type = getContext().getAsVariableArrayType(elementType)));
1818 return std::pair<llvm::Value*,QualType>(numElements, elementType);
1821 void CodeGenFunction::EmitVariablyModifiedType(QualType type) {
1822 assert(type->isVariablyModifiedType() &&
1823 "Must pass variably modified type to EmitVLASizes!");
1825 EnsureInsertPoint();
1827 // We're going to walk down into the type and look for VLA
1830 assert(type->isVariablyModifiedType());
1832 const Type *ty = type.getTypePtr();
1833 switch (ty->getTypeClass()) {
1835 #define TYPE(Class, Base)
1836 #define ABSTRACT_TYPE(Class, Base)
1837 #define NON_CANONICAL_TYPE(Class, Base)
1838 #define DEPENDENT_TYPE(Class, Base) case Type::Class:
1839 #define NON_CANONICAL_UNLESS_DEPENDENT_TYPE(Class, Base)
1840 #include "clang/AST/TypeNodes.def"
1841 llvm_unreachable("unexpected dependent type!");
1843 // These types are never variably-modified.
1847 case Type::ExtVector:
1850 case Type::Elaborated:
1851 case Type::TemplateSpecialization:
1852 case Type::ObjCTypeParam:
1853 case Type::ObjCObject:
1854 case Type::ObjCInterface:
1855 case Type::ObjCObjectPointer:
1856 llvm_unreachable("type class is never variably-modified!");
1858 case Type::Adjusted:
1859 type = cast<AdjustedType>(ty)->getAdjustedType();
1863 type = cast<DecayedType>(ty)->getPointeeType();
1867 type = cast<PointerType>(ty)->getPointeeType();
1870 case Type::BlockPointer:
1871 type = cast<BlockPointerType>(ty)->getPointeeType();
1874 case Type::LValueReference:
1875 case Type::RValueReference:
1876 type = cast<ReferenceType>(ty)->getPointeeType();
1879 case Type::MemberPointer:
1880 type = cast<MemberPointerType>(ty)->getPointeeType();
1883 case Type::ConstantArray:
1884 case Type::IncompleteArray:
1885 // Losing element qualification here is fine.
1886 type = cast<ArrayType>(ty)->getElementType();
1889 case Type::VariableArray: {
1890 // Losing element qualification here is fine.
1891 const VariableArrayType *vat = cast<VariableArrayType>(ty);
1893 // Unknown size indication requires no size computation.
1894 // Otherwise, evaluate and record it.
1895 if (const Expr *size = vat->getSizeExpr()) {
1896 // It's possible that we might have emitted this already,
1897 // e.g. with a typedef and a pointer to it.
1898 llvm::Value *&entry = VLASizeMap[size];
1900 llvm::Value *Size = EmitScalarExpr(size);
1903 // If the size is an expression that is not an integer constant
1904 // expression [...] each time it is evaluated it shall have a value
1905 // greater than zero.
1906 if (SanOpts.has(SanitizerKind::VLABound) &&
1907 size->getType()->isSignedIntegerType()) {
1908 SanitizerScope SanScope(this);
1909 llvm::Value *Zero = llvm::Constant::getNullValue(Size->getType());
1910 llvm::Constant *StaticArgs[] = {
1911 EmitCheckSourceLocation(size->getLocStart()),
1912 EmitCheckTypeDescriptor(size->getType())
1914 EmitCheck(std::make_pair(Builder.CreateICmpSGT(Size, Zero),
1915 SanitizerKind::VLABound),
1916 SanitizerHandler::VLABoundNotPositive, StaticArgs, Size);
1919 // Always zexting here would be wrong if it weren't
1920 // undefined behavior to have a negative bound.
1921 entry = Builder.CreateIntCast(Size, SizeTy, /*signed*/ false);
1924 type = vat->getElementType();
1928 case Type::FunctionProto:
1929 case Type::FunctionNoProto:
1930 type = cast<FunctionType>(ty)->getReturnType();
1935 case Type::UnaryTransform:
1936 case Type::Attributed:
1937 case Type::SubstTemplateTypeParm:
1938 case Type::PackExpansion:
1939 // Keep walking after single level desugaring.
1940 type = type.getSingleStepDesugaredType(getContext());
1944 case Type::Decltype:
1946 case Type::DeducedTemplateSpecialization:
1947 // Stop walking: nothing to do.
1950 case Type::TypeOfExpr:
1951 // Stop walking: emit typeof expression.
1952 EmitIgnoredExpr(cast<TypeOfExprType>(ty)->getUnderlyingExpr());
1956 type = cast<AtomicType>(ty)->getValueType();
1960 type = cast<PipeType>(ty)->getElementType();
1963 } while (type->isVariablyModifiedType());
1966 Address CodeGenFunction::EmitVAListRef(const Expr* E) {
1967 if (getContext().getBuiltinVaListType()->isArrayType())
1968 return EmitPointerWithAlignment(E);
1969 return EmitLValue(E).getAddress();
1972 Address CodeGenFunction::EmitMSVAListRef(const Expr *E) {
1973 return EmitLValue(E).getAddress();
1976 void CodeGenFunction::EmitDeclRefExprDbgValue(const DeclRefExpr *E,
1977 const APValue &Init) {
1978 assert(!Init.isUninit() && "Invalid DeclRefExpr initializer!");
1979 if (CGDebugInfo *Dbg = getDebugInfo())
1980 if (CGM.getCodeGenOpts().getDebugInfo() >= codegenoptions::LimitedDebugInfo)
1981 Dbg->EmitGlobalVariable(E->getDecl(), Init);
1984 CodeGenFunction::PeepholeProtection
1985 CodeGenFunction::protectFromPeepholes(RValue rvalue) {
1986 // At the moment, the only aggressive peephole we do in IR gen
1987 // is trunc(zext) folding, but if we add more, we can easily
1988 // extend this protection.
1990 if (!rvalue.isScalar()) return PeepholeProtection();
1991 llvm::Value *value = rvalue.getScalarVal();
1992 if (!isa<llvm::ZExtInst>(value)) return PeepholeProtection();
1994 // Just make an extra bitcast.
1995 assert(HaveInsertPoint());
1996 llvm::Instruction *inst = new llvm::BitCastInst(value, value->getType(), "",
1997 Builder.GetInsertBlock());
1999 PeepholeProtection protection;
2000 protection.Inst = inst;
2004 void CodeGenFunction::unprotectFromPeepholes(PeepholeProtection protection) {
2005 if (!protection.Inst) return;
2007 // In theory, we could try to duplicate the peepholes now, but whatever.
2008 protection.Inst->eraseFromParent();
2011 llvm::Value *CodeGenFunction::EmitAnnotationCall(llvm::Value *AnnotationFn,
2012 llvm::Value *AnnotatedVal,
2013 StringRef AnnotationStr,
2014 SourceLocation Location) {
2015 llvm::Value *Args[4] = {
2017 Builder.CreateBitCast(CGM.EmitAnnotationString(AnnotationStr), Int8PtrTy),
2018 Builder.CreateBitCast(CGM.EmitAnnotationUnit(Location), Int8PtrTy),
2019 CGM.EmitAnnotationLineNo(Location)
2021 return Builder.CreateCall(AnnotationFn, Args);
2024 void CodeGenFunction::EmitVarAnnotations(const VarDecl *D, llvm::Value *V) {
2025 assert(D->hasAttr<AnnotateAttr>() && "no annotate attribute");
2026 // FIXME We create a new bitcast for every annotation because that's what
2027 // llvm-gcc was doing.
2028 for (const auto *I : D->specific_attrs<AnnotateAttr>())
2029 EmitAnnotationCall(CGM.getIntrinsic(llvm::Intrinsic::var_annotation),
2030 Builder.CreateBitCast(V, CGM.Int8PtrTy, V->getName()),
2031 I->getAnnotation(), D->getLocation());
2034 Address CodeGenFunction::EmitFieldAnnotations(const FieldDecl *D,
2036 assert(D->hasAttr<AnnotateAttr>() && "no annotate attribute");
2037 llvm::Value *V = Addr.getPointer();
2038 llvm::Type *VTy = V->getType();
2039 llvm::Value *F = CGM.getIntrinsic(llvm::Intrinsic::ptr_annotation,
2042 for (const auto *I : D->specific_attrs<AnnotateAttr>()) {
2043 // FIXME Always emit the cast inst so we can differentiate between
2044 // annotation on the first field of a struct and annotation on the struct
2046 if (VTy != CGM.Int8PtrTy)
2047 V = Builder.Insert(new llvm::BitCastInst(V, CGM.Int8PtrTy));
2048 V = EmitAnnotationCall(F, V, I->getAnnotation(), D->getLocation());
2049 V = Builder.CreateBitCast(V, VTy);
2052 return Address(V, Addr.getAlignment());
2055 CodeGenFunction::CGCapturedStmtInfo::~CGCapturedStmtInfo() { }
2057 CodeGenFunction::SanitizerScope::SanitizerScope(CodeGenFunction *CGF)
2059 assert(!CGF->IsSanitizerScope);
2060 CGF->IsSanitizerScope = true;
2063 CodeGenFunction::SanitizerScope::~SanitizerScope() {
2064 CGF->IsSanitizerScope = false;
2067 void CodeGenFunction::InsertHelper(llvm::Instruction *I,
2068 const llvm::Twine &Name,
2069 llvm::BasicBlock *BB,
2070 llvm::BasicBlock::iterator InsertPt) const {
2071 LoopStack.InsertHelper(I);
2072 if (IsSanitizerScope)
2073 CGM.getSanitizerMetadata()->disableSanitizerForInstruction(I);
2076 void CGBuilderInserter::InsertHelper(
2077 llvm::Instruction *I, const llvm::Twine &Name, llvm::BasicBlock *BB,
2078 llvm::BasicBlock::iterator InsertPt) const {
2079 llvm::IRBuilderDefaultInserter::InsertHelper(I, Name, BB, InsertPt);
2081 CGF->InsertHelper(I, Name, BB, InsertPt);
2084 static bool hasRequiredFeatures(const SmallVectorImpl<StringRef> &ReqFeatures,
2085 CodeGenModule &CGM, const FunctionDecl *FD,
2086 std::string &FirstMissing) {
2087 // If there aren't any required features listed then go ahead and return.
2088 if (ReqFeatures.empty())
2091 // Now build up the set of caller features and verify that all the required
2092 // features are there.
2093 llvm::StringMap<bool> CallerFeatureMap;
2094 CGM.getFunctionFeatureMap(CallerFeatureMap, FD);
2096 // If we have at least one of the features in the feature list return
2097 // true, otherwise return false.
2099 ReqFeatures.begin(), ReqFeatures.end(), [&](StringRef Feature) {
2100 SmallVector<StringRef, 1> OrFeatures;
2101 Feature.split(OrFeatures, "|");
2102 return std::any_of(OrFeatures.begin(), OrFeatures.end(),
2103 [&](StringRef Feature) {
2104 if (!CallerFeatureMap.lookup(Feature)) {
2105 FirstMissing = Feature.str();
2113 // Emits an error if we don't have a valid set of target features for the
2115 void CodeGenFunction::checkTargetFeatures(const CallExpr *E,
2116 const FunctionDecl *TargetDecl) {
2117 // Early exit if this is an indirect call.
2121 // Get the current enclosing function if it exists. If it doesn't
2122 // we can't check the target features anyhow.
2123 const FunctionDecl *FD = dyn_cast_or_null<FunctionDecl>(CurFuncDecl);
2127 // Grab the required features for the call. For a builtin this is listed in
2128 // the td file with the default cpu, for an always_inline function this is any
2129 // listed cpu and any listed features.
2130 unsigned BuiltinID = TargetDecl->getBuiltinID();
2131 std::string MissingFeature;
2133 SmallVector<StringRef, 1> ReqFeatures;
2134 const char *FeatureList =
2135 CGM.getContext().BuiltinInfo.getRequiredFeatures(BuiltinID);
2136 // Return if the builtin doesn't have any required features.
2137 if (!FeatureList || StringRef(FeatureList) == "")
2139 StringRef(FeatureList).split(ReqFeatures, ",");
2140 if (!hasRequiredFeatures(ReqFeatures, CGM, FD, MissingFeature))
2141 CGM.getDiags().Report(E->getLocStart(), diag::err_builtin_needs_feature)
2142 << TargetDecl->getDeclName()
2143 << CGM.getContext().BuiltinInfo.getRequiredFeatures(BuiltinID);
2145 } else if (TargetDecl->hasAttr<TargetAttr>()) {
2146 // Get the required features for the callee.
2147 SmallVector<StringRef, 1> ReqFeatures;
2148 llvm::StringMap<bool> CalleeFeatureMap;
2149 CGM.getFunctionFeatureMap(CalleeFeatureMap, TargetDecl);
2150 for (const auto &F : CalleeFeatureMap) {
2151 // Only positive features are "required".
2153 ReqFeatures.push_back(F.getKey());
2155 if (!hasRequiredFeatures(ReqFeatures, CGM, FD, MissingFeature))
2156 CGM.getDiags().Report(E->getLocStart(), diag::err_function_needs_feature)
2157 << FD->getDeclName() << TargetDecl->getDeclName() << MissingFeature;
2161 void CodeGenFunction::EmitSanitizerStatReport(llvm::SanitizerStatKind SSK) {
2162 if (!CGM.getCodeGenOpts().SanitizeStats)
2165 llvm::IRBuilder<> IRB(Builder.GetInsertBlock(), Builder.GetInsertPoint());
2166 IRB.SetCurrentDebugLocation(Builder.getCurrentDebugLocation());
2167 CGM.getSanStats().create(IRB, SSK);
2170 llvm::DebugLoc CodeGenFunction::SourceLocToDebugLoc(SourceLocation Location) {
2171 if (CGDebugInfo *DI = getDebugInfo())
2172 return DI->SourceLocToDebugLoc(Location);
2174 return llvm::DebugLoc();