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/Dominators.h"
37 #include "llvm/IR/Intrinsics.h"
38 #include "llvm/IR/MDBuilder.h"
39 #include "llvm/IR/Operator.h"
40 #include "llvm/Transforms/Utils/PromoteMemToReg.h"
41 using namespace clang;
42 using namespace CodeGen;
44 /// shouldEmitLifetimeMarkers - Decide whether we need emit the life-time
46 static bool shouldEmitLifetimeMarkers(const CodeGenOptions &CGOpts,
47 const LangOptions &LangOpts) {
48 if (CGOpts.DisableLifetimeMarkers)
51 // Disable lifetime markers in msan builds.
52 // FIXME: Remove this when msan works with lifetime markers.
53 if (LangOpts.Sanitize.has(SanitizerKind::Memory))
56 // Asan uses markers for use-after-scope checks.
57 if (CGOpts.SanitizeAddressUseAfterScope)
60 // For now, only in optimized builds.
61 return CGOpts.OptimizationLevel != 0;
64 CodeGenFunction::CodeGenFunction(CodeGenModule &cgm, bool suppressNewContext)
65 : CodeGenTypeCache(cgm), CGM(cgm), Target(cgm.getTarget()),
66 Builder(cgm, cgm.getModule().getContext(), llvm::ConstantFolder(),
67 CGBuilderInserterTy(this)),
68 CurFn(nullptr), ReturnValue(Address::invalid()),
69 CapturedStmtInfo(nullptr), SanOpts(CGM.getLangOpts().Sanitize),
70 IsSanitizerScope(false), CurFuncIsThunk(false), AutoreleaseResult(false),
71 SawAsmBlock(false), IsOutlinedSEHHelper(false), BlockInfo(nullptr),
72 BlockPointer(nullptr), LambdaThisCaptureField(nullptr),
73 NormalCleanupDest(nullptr), NextCleanupDestIndex(1),
74 FirstBlockInfo(nullptr), EHResumeBlock(nullptr), ExceptionSlot(nullptr),
75 EHSelectorSlot(nullptr), DebugInfo(CGM.getModuleDebugInfo()),
76 DisableDebugInfo(false), DidCallStackSave(false), IndirectBranch(nullptr),
77 PGO(cgm), SwitchInsn(nullptr), SwitchWeights(nullptr),
78 CaseRangeBlock(nullptr), UnreachableBlock(nullptr), NumReturnExprs(0),
79 NumSimpleReturnExprs(0), CXXABIThisDecl(nullptr),
80 CXXABIThisValue(nullptr), CXXThisValue(nullptr),
81 CXXStructorImplicitParamDecl(nullptr),
82 CXXStructorImplicitParamValue(nullptr), OutermostConditional(nullptr),
83 CurLexicalScope(nullptr), TerminateLandingPad(nullptr),
84 TerminateHandler(nullptr), TrapBB(nullptr),
85 ShouldEmitLifetimeMarkers(
86 shouldEmitLifetimeMarkers(CGM.getCodeGenOpts(), CGM.getLangOpts())) {
87 if (!suppressNewContext)
88 CGM.getCXXABI().getMangleContext().startNewFunction();
90 llvm::FastMathFlags FMF;
91 if (CGM.getLangOpts().FastMath)
93 if (CGM.getLangOpts().FiniteMathOnly) {
97 if (CGM.getCodeGenOpts().NoNaNsFPMath) {
100 if (CGM.getCodeGenOpts().NoSignedZeros) {
101 FMF.setNoSignedZeros();
103 if (CGM.getCodeGenOpts().ReciprocalMath) {
104 FMF.setAllowReciprocal();
106 if (CGM.getCodeGenOpts().Reassociate) {
107 FMF.setAllowReassoc();
109 Builder.setFastMathFlags(FMF);
112 CodeGenFunction::~CodeGenFunction() {
113 assert(LifetimeExtendedCleanupStack.empty() && "failed to emit a cleanup");
115 // If there are any unclaimed block infos, go ahead and destroy them
116 // now. This can happen if IR-gen gets clever and skips evaluating
119 destroyBlockInfos(FirstBlockInfo);
121 if (getLangOpts().OpenMP && CurFn)
122 CGM.getOpenMPRuntime().functionFinished(*this);
125 CharUnits CodeGenFunction::getNaturalPointeeTypeAlignment(QualType T,
126 LValueBaseInfo *BaseInfo,
127 TBAAAccessInfo *TBAAInfo) {
128 return getNaturalTypeAlignment(T->getPointeeType(), BaseInfo, TBAAInfo,
129 /* forPointeeType= */ true);
132 CharUnits CodeGenFunction::getNaturalTypeAlignment(QualType T,
133 LValueBaseInfo *BaseInfo,
134 TBAAAccessInfo *TBAAInfo,
135 bool forPointeeType) {
137 *TBAAInfo = CGM.getTBAAAccessInfo(T);
139 // Honor alignment typedef attributes even on incomplete types.
140 // We also honor them straight for C++ class types, even as pointees;
141 // there's an expressivity gap here.
142 if (auto TT = T->getAs<TypedefType>()) {
143 if (auto Align = TT->getDecl()->getMaxAlignment()) {
145 *BaseInfo = LValueBaseInfo(AlignmentSource::AttributedType);
146 return getContext().toCharUnitsFromBits(Align);
151 *BaseInfo = LValueBaseInfo(AlignmentSource::Type);
154 if (T->isIncompleteType()) {
155 Alignment = CharUnits::One(); // Shouldn't be used, but pessimistic is best.
157 // For C++ class pointees, we don't know whether we're pointing at a
158 // base or a complete object, so we generally need to use the
159 // non-virtual alignment.
160 const CXXRecordDecl *RD;
161 if (forPointeeType && (RD = T->getAsCXXRecordDecl())) {
162 Alignment = CGM.getClassPointerAlignment(RD);
164 Alignment = getContext().getTypeAlignInChars(T);
165 if (T.getQualifiers().hasUnaligned())
166 Alignment = CharUnits::One();
169 // Cap to the global maximum type alignment unless the alignment
170 // was somehow explicit on the type.
171 if (unsigned MaxAlign = getLangOpts().MaxTypeAlign) {
172 if (Alignment.getQuantity() > MaxAlign &&
173 !getContext().isAlignmentRequired(T))
174 Alignment = CharUnits::fromQuantity(MaxAlign);
180 LValue CodeGenFunction::MakeNaturalAlignAddrLValue(llvm::Value *V, QualType T) {
181 LValueBaseInfo BaseInfo;
182 TBAAAccessInfo TBAAInfo;
183 CharUnits Alignment = getNaturalTypeAlignment(T, &BaseInfo, &TBAAInfo);
184 return LValue::MakeAddr(Address(V, Alignment), T, getContext(), BaseInfo,
188 /// Given a value of type T* that may not be to a complete object,
189 /// construct an l-value with the natural pointee alignment of T.
191 CodeGenFunction::MakeNaturalAlignPointeeAddrLValue(llvm::Value *V, QualType T) {
192 LValueBaseInfo BaseInfo;
193 TBAAAccessInfo TBAAInfo;
194 CharUnits Align = getNaturalTypeAlignment(T, &BaseInfo, &TBAAInfo,
195 /* forPointeeType= */ true);
196 return MakeAddrLValue(Address(V, Align), T, BaseInfo, TBAAInfo);
200 llvm::Type *CodeGenFunction::ConvertTypeForMem(QualType T) {
201 return CGM.getTypes().ConvertTypeForMem(T);
204 llvm::Type *CodeGenFunction::ConvertType(QualType T) {
205 return CGM.getTypes().ConvertType(T);
208 TypeEvaluationKind CodeGenFunction::getEvaluationKind(QualType type) {
209 type = type.getCanonicalType();
211 switch (type->getTypeClass()) {
212 #define TYPE(name, parent)
213 #define ABSTRACT_TYPE(name, parent)
214 #define NON_CANONICAL_TYPE(name, parent) case Type::name:
215 #define DEPENDENT_TYPE(name, parent) case Type::name:
216 #define NON_CANONICAL_UNLESS_DEPENDENT_TYPE(name, parent) case Type::name:
217 #include "clang/AST/TypeNodes.def"
218 llvm_unreachable("non-canonical or dependent type in IR-generation");
221 case Type::DeducedTemplateSpecialization:
222 llvm_unreachable("undeduced type in IR-generation");
224 // Various scalar types.
227 case Type::BlockPointer:
228 case Type::LValueReference:
229 case Type::RValueReference:
230 case Type::MemberPointer:
232 case Type::ExtVector:
233 case Type::FunctionProto:
234 case Type::FunctionNoProto:
236 case Type::ObjCObjectPointer:
244 // Arrays, records, and Objective-C objects.
245 case Type::ConstantArray:
246 case Type::IncompleteArray:
247 case Type::VariableArray:
249 case Type::ObjCObject:
250 case Type::ObjCInterface:
251 return TEK_Aggregate;
253 // We operate on atomic values according to their underlying type.
255 type = cast<AtomicType>(type)->getValueType();
258 llvm_unreachable("unknown type kind!");
262 llvm::DebugLoc CodeGenFunction::EmitReturnBlock() {
263 // For cleanliness, we try to avoid emitting the return block for
265 llvm::BasicBlock *CurBB = Builder.GetInsertBlock();
268 assert(!CurBB->getTerminator() && "Unexpected terminated block.");
270 // We have a valid insert point, reuse it if it is empty or there are no
271 // explicit jumps to the return block.
272 if (CurBB->empty() || ReturnBlock.getBlock()->use_empty()) {
273 ReturnBlock.getBlock()->replaceAllUsesWith(CurBB);
274 delete ReturnBlock.getBlock();
276 EmitBlock(ReturnBlock.getBlock());
277 return llvm::DebugLoc();
280 // Otherwise, if the return block is the target of a single direct
281 // branch then we can just put the code in that block instead. This
282 // cleans up functions which started with a unified return block.
283 if (ReturnBlock.getBlock()->hasOneUse()) {
284 llvm::BranchInst *BI =
285 dyn_cast<llvm::BranchInst>(*ReturnBlock.getBlock()->user_begin());
286 if (BI && BI->isUnconditional() &&
287 BI->getSuccessor(0) == ReturnBlock.getBlock()) {
288 // Record/return the DebugLoc of the simple 'return' expression to be used
289 // later by the actual 'ret' instruction.
290 llvm::DebugLoc Loc = BI->getDebugLoc();
291 Builder.SetInsertPoint(BI->getParent());
292 BI->eraseFromParent();
293 delete ReturnBlock.getBlock();
298 // FIXME: We are at an unreachable point, there is no reason to emit the block
299 // unless it has uses. However, we still need a place to put the debug
300 // region.end for now.
302 EmitBlock(ReturnBlock.getBlock());
303 return llvm::DebugLoc();
306 static void EmitIfUsed(CodeGenFunction &CGF, llvm::BasicBlock *BB) {
308 if (!BB->use_empty())
309 return CGF.CurFn->getBasicBlockList().push_back(BB);
313 void CodeGenFunction::FinishFunction(SourceLocation EndLoc) {
314 assert(BreakContinueStack.empty() &&
315 "mismatched push/pop in break/continue stack!");
317 bool OnlySimpleReturnStmts = NumSimpleReturnExprs > 0
318 && NumSimpleReturnExprs == NumReturnExprs
319 && ReturnBlock.getBlock()->use_empty();
320 // Usually the return expression is evaluated before the cleanup
321 // code. If the function contains only a simple return statement,
322 // such as a constant, the location before the cleanup code becomes
323 // the last useful breakpoint in the function, because the simple
324 // return expression will be evaluated after the cleanup code. To be
325 // safe, set the debug location for cleanup code to the location of
326 // the return statement. Otherwise the cleanup code should be at the
327 // end of the function's lexical scope.
329 // If there are multiple branches to the return block, the branch
330 // instructions will get the location of the return statements and
332 if (CGDebugInfo *DI = getDebugInfo()) {
333 if (OnlySimpleReturnStmts)
334 DI->EmitLocation(Builder, LastStopPoint);
336 DI->EmitLocation(Builder, EndLoc);
339 // Pop any cleanups that might have been associated with the
340 // parameters. Do this in whatever block we're currently in; it's
341 // important to do this before we enter the return block or return
342 // edges will be *really* confused.
343 bool HasCleanups = EHStack.stable_begin() != PrologueCleanupDepth;
344 bool HasOnlyLifetimeMarkers =
345 HasCleanups && EHStack.containsOnlyLifetimeMarkers(PrologueCleanupDepth);
346 bool EmitRetDbgLoc = !HasCleanups || HasOnlyLifetimeMarkers;
348 // Make sure the line table doesn't jump back into the body for
349 // the ret after it's been at EndLoc.
350 if (CGDebugInfo *DI = getDebugInfo())
351 if (OnlySimpleReturnStmts)
352 DI->EmitLocation(Builder, EndLoc);
354 PopCleanupBlocks(PrologueCleanupDepth);
357 // Emit function epilog (to return).
358 llvm::DebugLoc Loc = EmitReturnBlock();
360 if (ShouldInstrumentFunction()) {
361 if (CGM.getCodeGenOpts().InstrumentFunctions)
362 CurFn->addFnAttr("instrument-function-exit", "__cyg_profile_func_exit");
363 if (CGM.getCodeGenOpts().InstrumentFunctionsAfterInlining)
364 CurFn->addFnAttr("instrument-function-exit-inlined",
365 "__cyg_profile_func_exit");
368 // Emit debug descriptor for function end.
369 if (CGDebugInfo *DI = getDebugInfo())
370 DI->EmitFunctionEnd(Builder, CurFn);
372 // Reset the debug location to that of the simple 'return' expression, if any
373 // rather than that of the end of the function's scope '}'.
374 ApplyDebugLocation AL(*this, Loc);
375 EmitFunctionEpilog(*CurFnInfo, EmitRetDbgLoc, EndLoc);
376 EmitEndEHSpec(CurCodeDecl);
378 assert(EHStack.empty() &&
379 "did not remove all scopes from cleanup stack!");
381 // If someone did an indirect goto, emit the indirect goto block at the end of
383 if (IndirectBranch) {
384 EmitBlock(IndirectBranch->getParent());
385 Builder.ClearInsertionPoint();
388 // If some of our locals escaped, insert a call to llvm.localescape in the
390 if (!EscapedLocals.empty()) {
391 // Invert the map from local to index into a simple vector. There should be
393 SmallVector<llvm::Value *, 4> EscapeArgs;
394 EscapeArgs.resize(EscapedLocals.size());
395 for (auto &Pair : EscapedLocals)
396 EscapeArgs[Pair.second] = Pair.first;
397 llvm::Function *FrameEscapeFn = llvm::Intrinsic::getDeclaration(
398 &CGM.getModule(), llvm::Intrinsic::localescape);
399 CGBuilderTy(*this, AllocaInsertPt).CreateCall(FrameEscapeFn, EscapeArgs);
402 // Remove the AllocaInsertPt instruction, which is just a convenience for us.
403 llvm::Instruction *Ptr = AllocaInsertPt;
404 AllocaInsertPt = nullptr;
405 Ptr->eraseFromParent();
407 // If someone took the address of a label but never did an indirect goto, we
408 // made a zero entry PHI node, which is illegal, zap it now.
409 if (IndirectBranch) {
410 llvm::PHINode *PN = cast<llvm::PHINode>(IndirectBranch->getAddress());
411 if (PN->getNumIncomingValues() == 0) {
412 PN->replaceAllUsesWith(llvm::UndefValue::get(PN->getType()));
413 PN->eraseFromParent();
417 EmitIfUsed(*this, EHResumeBlock);
418 EmitIfUsed(*this, TerminateLandingPad);
419 EmitIfUsed(*this, TerminateHandler);
420 EmitIfUsed(*this, UnreachableBlock);
422 if (CGM.getCodeGenOpts().EmitDeclMetadata)
425 for (SmallVectorImpl<std::pair<llvm::Instruction *, llvm::Value *> >::iterator
426 I = DeferredReplacements.begin(),
427 E = DeferredReplacements.end();
429 I->first->replaceAllUsesWith(I->second);
430 I->first->eraseFromParent();
433 // Eliminate CleanupDestSlot alloca by replacing it with SSA values and
434 // PHIs if the current function is a coroutine. We don't do it for all
435 // functions as it may result in slight increase in numbers of instructions
436 // if compiled with no optimizations. We do it for coroutine as the lifetime
437 // of CleanupDestSlot alloca make correct coroutine frame building very
439 if (NormalCleanupDest && isCoroutine()) {
440 llvm::DominatorTree DT(*CurFn);
441 llvm::PromoteMemToReg(NormalCleanupDest, DT);
442 NormalCleanupDest = nullptr;
446 /// ShouldInstrumentFunction - Return true if the current function should be
447 /// instrumented with __cyg_profile_func_* calls
448 bool CodeGenFunction::ShouldInstrumentFunction() {
449 if (!CGM.getCodeGenOpts().InstrumentFunctions &&
450 !CGM.getCodeGenOpts().InstrumentFunctionsAfterInlining &&
451 !CGM.getCodeGenOpts().InstrumentFunctionEntryBare)
453 if (!CurFuncDecl || CurFuncDecl->hasAttr<NoInstrumentFunctionAttr>())
458 /// ShouldXRayInstrument - Return true if the current function should be
459 /// instrumented with XRay nop sleds.
460 bool CodeGenFunction::ShouldXRayInstrumentFunction() const {
461 return CGM.getCodeGenOpts().XRayInstrumentFunctions;
464 /// AlwaysEmitXRayCustomEvents - Return true if we should emit IR for calls to
465 /// the __xray_customevent(...) builin calls, when doing XRay instrumentation.
466 bool CodeGenFunction::AlwaysEmitXRayCustomEvents() const {
467 return CGM.getCodeGenOpts().XRayAlwaysEmitCustomEvents;
471 CodeGenFunction::EncodeAddrForUseInPrologue(llvm::Function *F,
472 llvm::Constant *Addr) {
473 // Addresses stored in prologue data can't require run-time fixups and must
474 // be PC-relative. Run-time fixups are undesirable because they necessitate
475 // writable text segments, which are unsafe. And absolute addresses are
476 // undesirable because they break PIE mode.
478 // Add a layer of indirection through a private global. Taking its address
479 // won't result in a run-time fixup, even if Addr has linkonce_odr linkage.
480 auto *GV = new llvm::GlobalVariable(CGM.getModule(), Addr->getType(),
482 llvm::GlobalValue::PrivateLinkage, Addr);
484 // Create a PC-relative address.
485 auto *GOTAsInt = llvm::ConstantExpr::getPtrToInt(GV, IntPtrTy);
486 auto *FuncAsInt = llvm::ConstantExpr::getPtrToInt(F, IntPtrTy);
487 auto *PCRelAsInt = llvm::ConstantExpr::getSub(GOTAsInt, FuncAsInt);
488 return (IntPtrTy == Int32Ty)
490 : llvm::ConstantExpr::getTrunc(PCRelAsInt, Int32Ty);
494 CodeGenFunction::DecodeAddrUsedInPrologue(llvm::Value *F,
495 llvm::Value *EncodedAddr) {
496 // Reconstruct the address of the global.
497 auto *PCRelAsInt = Builder.CreateSExt(EncodedAddr, IntPtrTy);
498 auto *FuncAsInt = Builder.CreatePtrToInt(F, IntPtrTy, "func_addr.int");
499 auto *GOTAsInt = Builder.CreateAdd(PCRelAsInt, FuncAsInt, "global_addr.int");
500 auto *GOTAddr = Builder.CreateIntToPtr(GOTAsInt, Int8PtrPtrTy, "global_addr");
502 // Load the original pointer through the global.
503 return Builder.CreateLoad(Address(GOTAddr, getPointerAlign()),
507 static void removeImageAccessQualifier(std::string& TyName) {
508 std::string ReadOnlyQual("__read_only");
509 std::string::size_type ReadOnlyPos = TyName.find(ReadOnlyQual);
510 if (ReadOnlyPos != std::string::npos)
511 // "+ 1" for the space after access qualifier.
512 TyName.erase(ReadOnlyPos, ReadOnlyQual.size() + 1);
514 std::string WriteOnlyQual("__write_only");
515 std::string::size_type WriteOnlyPos = TyName.find(WriteOnlyQual);
516 if (WriteOnlyPos != std::string::npos)
517 TyName.erase(WriteOnlyPos, WriteOnlyQual.size() + 1);
519 std::string ReadWriteQual("__read_write");
520 std::string::size_type ReadWritePos = TyName.find(ReadWriteQual);
521 if (ReadWritePos != std::string::npos)
522 TyName.erase(ReadWritePos, ReadWriteQual.size() + 1);
527 // Returns the address space id that should be produced to the
528 // kernel_arg_addr_space metadata. This is always fixed to the ids
529 // as specified in the SPIR 2.0 specification in order to differentiate
530 // for example in clGetKernelArgInfo() implementation between the address
531 // spaces with targets without unique mapping to the OpenCL address spaces
532 // (basically all single AS CPUs).
533 static unsigned ArgInfoAddressSpace(LangAS AS) {
535 case LangAS::opencl_global: return 1;
536 case LangAS::opencl_constant: return 2;
537 case LangAS::opencl_local: return 3;
538 case LangAS::opencl_generic: return 4; // Not in SPIR 2.0 specs.
540 return 0; // Assume private.
544 // OpenCL v1.2 s5.6.4.6 allows the compiler to store kernel argument
545 // information in the program executable. The argument information stored
546 // includes the argument name, its type, the address and access qualifiers used.
547 static void GenOpenCLArgMetadata(const FunctionDecl *FD, llvm::Function *Fn,
548 CodeGenModule &CGM, llvm::LLVMContext &Context,
549 CGBuilderTy &Builder, ASTContext &ASTCtx) {
550 // Create MDNodes that represent the kernel arg metadata.
551 // Each MDNode is a list in the form of "key", N number of values which is
552 // the same number of values as their are kernel arguments.
554 const PrintingPolicy &Policy = ASTCtx.getPrintingPolicy();
556 // MDNode for the kernel argument address space qualifiers.
557 SmallVector<llvm::Metadata *, 8> addressQuals;
559 // MDNode for the kernel argument access qualifiers (images only).
560 SmallVector<llvm::Metadata *, 8> accessQuals;
562 // MDNode for the kernel argument type names.
563 SmallVector<llvm::Metadata *, 8> argTypeNames;
565 // MDNode for the kernel argument base type names.
566 SmallVector<llvm::Metadata *, 8> argBaseTypeNames;
568 // MDNode for the kernel argument type qualifiers.
569 SmallVector<llvm::Metadata *, 8> argTypeQuals;
571 // MDNode for the kernel argument names.
572 SmallVector<llvm::Metadata *, 8> argNames;
574 for (unsigned i = 0, e = FD->getNumParams(); i != e; ++i) {
575 const ParmVarDecl *parm = FD->getParamDecl(i);
576 QualType ty = parm->getType();
577 std::string typeQuals;
579 if (ty->isPointerType()) {
580 QualType pointeeTy = ty->getPointeeType();
582 // Get address qualifier.
583 addressQuals.push_back(llvm::ConstantAsMetadata::get(Builder.getInt32(
584 ArgInfoAddressSpace(pointeeTy.getAddressSpace()))));
586 // Get argument type name.
587 std::string typeName =
588 pointeeTy.getUnqualifiedType().getAsString(Policy) + "*";
590 // Turn "unsigned type" to "utype"
591 std::string::size_type pos = typeName.find("unsigned");
592 if (pointeeTy.isCanonical() && pos != std::string::npos)
593 typeName.erase(pos+1, 8);
595 argTypeNames.push_back(llvm::MDString::get(Context, typeName));
597 std::string baseTypeName =
598 pointeeTy.getUnqualifiedType().getCanonicalType().getAsString(
602 // Turn "unsigned type" to "utype"
603 pos = baseTypeName.find("unsigned");
604 if (pos != std::string::npos)
605 baseTypeName.erase(pos+1, 8);
607 argBaseTypeNames.push_back(llvm::MDString::get(Context, baseTypeName));
609 // Get argument type qualifiers:
610 if (ty.isRestrictQualified())
611 typeQuals = "restrict";
612 if (pointeeTy.isConstQualified() ||
613 (pointeeTy.getAddressSpace() == LangAS::opencl_constant))
614 typeQuals += typeQuals.empty() ? "const" : " const";
615 if (pointeeTy.isVolatileQualified())
616 typeQuals += typeQuals.empty() ? "volatile" : " volatile";
618 uint32_t AddrSpc = 0;
619 bool isPipe = ty->isPipeType();
620 if (ty->isImageType() || isPipe)
621 AddrSpc = ArgInfoAddressSpace(LangAS::opencl_global);
623 addressQuals.push_back(
624 llvm::ConstantAsMetadata::get(Builder.getInt32(AddrSpc)));
626 // Get argument type name.
627 std::string typeName;
629 typeName = ty.getCanonicalType()->getAs<PipeType>()->getElementType()
630 .getAsString(Policy);
632 typeName = ty.getUnqualifiedType().getAsString(Policy);
634 // Turn "unsigned type" to "utype"
635 std::string::size_type pos = typeName.find("unsigned");
636 if (ty.isCanonical() && pos != std::string::npos)
637 typeName.erase(pos+1, 8);
639 std::string baseTypeName;
641 baseTypeName = ty.getCanonicalType()->getAs<PipeType>()
642 ->getElementType().getCanonicalType()
643 .getAsString(Policy);
646 ty.getUnqualifiedType().getCanonicalType().getAsString(Policy);
648 // Remove access qualifiers on images
649 // (as they are inseparable from type in clang implementation,
650 // but OpenCL spec provides a special query to get access qualifier
651 // via clGetKernelArgInfo with CL_KERNEL_ARG_ACCESS_QUALIFIER):
652 if (ty->isImageType()) {
653 removeImageAccessQualifier(typeName);
654 removeImageAccessQualifier(baseTypeName);
657 argTypeNames.push_back(llvm::MDString::get(Context, typeName));
659 // Turn "unsigned type" to "utype"
660 pos = baseTypeName.find("unsigned");
661 if (pos != std::string::npos)
662 baseTypeName.erase(pos+1, 8);
664 argBaseTypeNames.push_back(llvm::MDString::get(Context, baseTypeName));
670 argTypeQuals.push_back(llvm::MDString::get(Context, typeQuals));
672 // Get image and pipe access qualifier:
673 if (ty->isImageType()|| ty->isPipeType()) {
674 const Decl *PDecl = parm;
675 if (auto *TD = dyn_cast<TypedefType>(ty))
676 PDecl = TD->getDecl();
677 const OpenCLAccessAttr *A = PDecl->getAttr<OpenCLAccessAttr>();
678 if (A && A->isWriteOnly())
679 accessQuals.push_back(llvm::MDString::get(Context, "write_only"));
680 else if (A && A->isReadWrite())
681 accessQuals.push_back(llvm::MDString::get(Context, "read_write"));
683 accessQuals.push_back(llvm::MDString::get(Context, "read_only"));
685 accessQuals.push_back(llvm::MDString::get(Context, "none"));
687 // Get argument name.
688 argNames.push_back(llvm::MDString::get(Context, parm->getName()));
691 Fn->setMetadata("kernel_arg_addr_space",
692 llvm::MDNode::get(Context, addressQuals));
693 Fn->setMetadata("kernel_arg_access_qual",
694 llvm::MDNode::get(Context, accessQuals));
695 Fn->setMetadata("kernel_arg_type",
696 llvm::MDNode::get(Context, argTypeNames));
697 Fn->setMetadata("kernel_arg_base_type",
698 llvm::MDNode::get(Context, argBaseTypeNames));
699 Fn->setMetadata("kernel_arg_type_qual",
700 llvm::MDNode::get(Context, argTypeQuals));
701 if (CGM.getCodeGenOpts().EmitOpenCLArgMetadata)
702 Fn->setMetadata("kernel_arg_name",
703 llvm::MDNode::get(Context, argNames));
706 void CodeGenFunction::EmitOpenCLKernelMetadata(const FunctionDecl *FD,
709 if (!FD->hasAttr<OpenCLKernelAttr>())
712 llvm::LLVMContext &Context = getLLVMContext();
714 GenOpenCLArgMetadata(FD, Fn, CGM, Context, Builder, getContext());
716 if (const VecTypeHintAttr *A = FD->getAttr<VecTypeHintAttr>()) {
717 QualType HintQTy = A->getTypeHint();
718 const ExtVectorType *HintEltQTy = HintQTy->getAs<ExtVectorType>();
719 bool IsSignedInteger =
720 HintQTy->isSignedIntegerType() ||
721 (HintEltQTy && HintEltQTy->getElementType()->isSignedIntegerType());
722 llvm::Metadata *AttrMDArgs[] = {
723 llvm::ConstantAsMetadata::get(llvm::UndefValue::get(
724 CGM.getTypes().ConvertType(A->getTypeHint()))),
725 llvm::ConstantAsMetadata::get(llvm::ConstantInt::get(
726 llvm::IntegerType::get(Context, 32),
727 llvm::APInt(32, (uint64_t)(IsSignedInteger ? 1 : 0))))};
728 Fn->setMetadata("vec_type_hint", llvm::MDNode::get(Context, AttrMDArgs));
731 if (const WorkGroupSizeHintAttr *A = FD->getAttr<WorkGroupSizeHintAttr>()) {
732 llvm::Metadata *AttrMDArgs[] = {
733 llvm::ConstantAsMetadata::get(Builder.getInt32(A->getXDim())),
734 llvm::ConstantAsMetadata::get(Builder.getInt32(A->getYDim())),
735 llvm::ConstantAsMetadata::get(Builder.getInt32(A->getZDim()))};
736 Fn->setMetadata("work_group_size_hint", llvm::MDNode::get(Context, AttrMDArgs));
739 if (const ReqdWorkGroupSizeAttr *A = FD->getAttr<ReqdWorkGroupSizeAttr>()) {
740 llvm::Metadata *AttrMDArgs[] = {
741 llvm::ConstantAsMetadata::get(Builder.getInt32(A->getXDim())),
742 llvm::ConstantAsMetadata::get(Builder.getInt32(A->getYDim())),
743 llvm::ConstantAsMetadata::get(Builder.getInt32(A->getZDim()))};
744 Fn->setMetadata("reqd_work_group_size", llvm::MDNode::get(Context, AttrMDArgs));
747 if (const OpenCLIntelReqdSubGroupSizeAttr *A =
748 FD->getAttr<OpenCLIntelReqdSubGroupSizeAttr>()) {
749 llvm::Metadata *AttrMDArgs[] = {
750 llvm::ConstantAsMetadata::get(Builder.getInt32(A->getSubGroupSize()))};
751 Fn->setMetadata("intel_reqd_sub_group_size",
752 llvm::MDNode::get(Context, AttrMDArgs));
756 /// Determine whether the function F ends with a return stmt.
757 static bool endsWithReturn(const Decl* F) {
758 const Stmt *Body = nullptr;
759 if (auto *FD = dyn_cast_or_null<FunctionDecl>(F))
760 Body = FD->getBody();
761 else if (auto *OMD = dyn_cast_or_null<ObjCMethodDecl>(F))
762 Body = OMD->getBody();
764 if (auto *CS = dyn_cast_or_null<CompoundStmt>(Body)) {
765 auto LastStmt = CS->body_rbegin();
766 if (LastStmt != CS->body_rend())
767 return isa<ReturnStmt>(*LastStmt);
772 static void markAsIgnoreThreadCheckingAtRuntime(llvm::Function *Fn) {
773 Fn->addFnAttr("sanitize_thread_no_checking_at_run_time");
774 Fn->removeFnAttr(llvm::Attribute::SanitizeThread);
777 static bool matchesStlAllocatorFn(const Decl *D, const ASTContext &Ctx) {
778 auto *MD = dyn_cast_or_null<CXXMethodDecl>(D);
779 if (!MD || !MD->getDeclName().getAsIdentifierInfo() ||
780 !MD->getDeclName().getAsIdentifierInfo()->isStr("allocate") ||
781 (MD->getNumParams() != 1 && MD->getNumParams() != 2))
784 if (MD->parameters()[0]->getType().getCanonicalType() != Ctx.getSizeType())
787 if (MD->getNumParams() == 2) {
788 auto *PT = MD->parameters()[1]->getType()->getAs<PointerType>();
789 if (!PT || !PT->isVoidPointerType() ||
790 !PT->getPointeeType().isConstQualified())
797 /// Return the UBSan prologue signature for \p FD if one is available.
798 static llvm::Constant *getPrologueSignature(CodeGenModule &CGM,
799 const FunctionDecl *FD) {
800 if (const auto *MD = dyn_cast<CXXMethodDecl>(FD))
803 return CGM.getTargetCodeGenInfo().getUBSanFunctionSignature(CGM);
806 void CodeGenFunction::StartFunction(GlobalDecl GD,
809 const CGFunctionInfo &FnInfo,
810 const FunctionArgList &Args,
812 SourceLocation StartLoc) {
814 "Do not use a CodeGenFunction object for more than one function");
816 const Decl *D = GD.getDecl();
818 DidCallStackSave = false;
820 if (const auto *FD = dyn_cast_or_null<FunctionDecl>(D))
821 if (FD->usesSEHTry())
823 CurFuncDecl = (D ? D->getNonClosureContext() : nullptr);
827 assert(CurFn->isDeclaration() && "Function already has body?");
829 // If this function has been blacklisted for any of the enabled sanitizers,
830 // disable the sanitizer for the function.
832 #define SANITIZER(NAME, ID) \
833 if (SanOpts.empty()) \
835 if (SanOpts.has(SanitizerKind::ID)) \
836 if (CGM.isInSanitizerBlacklist(SanitizerKind::ID, Fn, Loc)) \
837 SanOpts.set(SanitizerKind::ID, false);
839 #include "clang/Basic/Sanitizers.def"
844 // Apply the no_sanitize* attributes to SanOpts.
845 for (auto Attr : D->specific_attrs<NoSanitizeAttr>())
846 SanOpts.Mask &= ~Attr->getMask();
849 // Apply sanitizer attributes to the function.
850 if (SanOpts.hasOneOf(SanitizerKind::Address | SanitizerKind::KernelAddress))
851 Fn->addFnAttr(llvm::Attribute::SanitizeAddress);
852 if (SanOpts.hasOneOf(SanitizerKind::HWAddress))
853 Fn->addFnAttr(llvm::Attribute::SanitizeHWAddress);
854 if (SanOpts.has(SanitizerKind::Thread))
855 Fn->addFnAttr(llvm::Attribute::SanitizeThread);
856 if (SanOpts.has(SanitizerKind::Memory))
857 Fn->addFnAttr(llvm::Attribute::SanitizeMemory);
858 if (SanOpts.has(SanitizerKind::SafeStack))
859 Fn->addFnAttr(llvm::Attribute::SafeStack);
861 // Ignore TSan memory acesses from within ObjC/ObjC++ dealloc, initialize,
862 // .cxx_destruct, __destroy_helper_block_ and all of their calees at run time.
863 if (SanOpts.has(SanitizerKind::Thread)) {
864 if (const auto *OMD = dyn_cast_or_null<ObjCMethodDecl>(D)) {
865 IdentifierInfo *II = OMD->getSelector().getIdentifierInfoForSlot(0);
866 if (OMD->getMethodFamily() == OMF_dealloc ||
867 OMD->getMethodFamily() == OMF_initialize ||
868 (OMD->getSelector().isUnarySelector() && II->isStr(".cxx_destruct"))) {
869 markAsIgnoreThreadCheckingAtRuntime(Fn);
871 } else if (const auto *FD = dyn_cast_or_null<FunctionDecl>(D)) {
872 IdentifierInfo *II = FD->getIdentifier();
873 if (II && II->isStr("__destroy_helper_block_"))
874 markAsIgnoreThreadCheckingAtRuntime(Fn);
878 // Ignore unrelated casts in STL allocate() since the allocator must cast
879 // from void* to T* before object initialization completes. Don't match on the
880 // namespace because not all allocators are in std::
881 if (D && SanOpts.has(SanitizerKind::CFIUnrelatedCast)) {
882 if (matchesStlAllocatorFn(D, getContext()))
883 SanOpts.Mask &= ~SanitizerKind::CFIUnrelatedCast;
886 // Apply xray attributes to the function (as a string, for now)
887 if (D && ShouldXRayInstrumentFunction()) {
888 if (const auto *XRayAttr = D->getAttr<XRayInstrumentAttr>()) {
889 if (XRayAttr->alwaysXRayInstrument())
890 Fn->addFnAttr("function-instrument", "xray-always");
891 if (XRayAttr->neverXRayInstrument())
892 Fn->addFnAttr("function-instrument", "xray-never");
893 if (const auto *LogArgs = D->getAttr<XRayLogArgsAttr>()) {
894 Fn->addFnAttr("xray-log-args",
895 llvm::utostr(LogArgs->getArgumentCount()));
898 if (!CGM.imbueXRayAttrs(Fn, Loc))
900 "xray-instruction-threshold",
901 llvm::itostr(CGM.getCodeGenOpts().XRayInstructionThreshold));
905 // Add no-jump-tables value.
906 Fn->addFnAttr("no-jump-tables",
907 llvm::toStringRef(CGM.getCodeGenOpts().NoUseJumpTables));
909 // Add profile-sample-accurate value.
910 if (CGM.getCodeGenOpts().ProfileSampleAccurate)
911 Fn->addFnAttr("profile-sample-accurate");
913 if (getLangOpts().OpenCL) {
914 // Add metadata for a kernel function.
915 if (const FunctionDecl *FD = dyn_cast_or_null<FunctionDecl>(D))
916 EmitOpenCLKernelMetadata(FD, Fn);
919 // If we are checking function types, emit a function type signature as
921 if (getLangOpts().CPlusPlus && SanOpts.has(SanitizerKind::Function)) {
922 if (const FunctionDecl *FD = dyn_cast_or_null<FunctionDecl>(D)) {
923 if (llvm::Constant *PrologueSig = getPrologueSignature(CGM, FD)) {
924 llvm::Constant *FTRTTIConst =
925 CGM.GetAddrOfRTTIDescriptor(FD->getType(), /*ForEH=*/true);
926 llvm::Constant *FTRTTIConstEncoded =
927 EncodeAddrForUseInPrologue(Fn, FTRTTIConst);
928 llvm::Constant *PrologueStructElems[] = {PrologueSig,
930 llvm::Constant *PrologueStructConst =
931 llvm::ConstantStruct::getAnon(PrologueStructElems, /*Packed=*/true);
932 Fn->setPrologueData(PrologueStructConst);
937 // If we're checking nullability, we need to know whether we can check the
938 // return value. Initialize the flag to 'true' and refine it in EmitParmDecl.
939 if (SanOpts.has(SanitizerKind::NullabilityReturn)) {
940 auto Nullability = FnRetTy->getNullability(getContext());
941 if (Nullability && *Nullability == NullabilityKind::NonNull) {
942 if (!(SanOpts.has(SanitizerKind::ReturnsNonnullAttribute) &&
943 CurCodeDecl && CurCodeDecl->getAttr<ReturnsNonNullAttr>()))
944 RetValNullabilityPrecondition =
945 llvm::ConstantInt::getTrue(getLLVMContext());
949 // If we're in C++ mode and the function name is "main", it is guaranteed
950 // to be norecurse by the standard (3.6.1.3 "The function main shall not be
951 // used within a program").
952 if (getLangOpts().CPlusPlus)
953 if (const FunctionDecl *FD = dyn_cast_or_null<FunctionDecl>(D))
955 Fn->addFnAttr(llvm::Attribute::NoRecurse);
957 llvm::BasicBlock *EntryBB = createBasicBlock("entry", CurFn);
959 // Create a marker to make it easy to insert allocas into the entryblock
960 // later. Don't create this with the builder, because we don't want it
962 llvm::Value *Undef = llvm::UndefValue::get(Int32Ty);
963 AllocaInsertPt = new llvm::BitCastInst(Undef, Int32Ty, "allocapt", EntryBB);
965 ReturnBlock = getJumpDestInCurrentScope("return");
967 Builder.SetInsertPoint(EntryBB);
969 // If we're checking the return value, allocate space for a pointer to a
970 // precise source location of the checked return statement.
971 if (requiresReturnValueCheck()) {
972 ReturnLocation = CreateDefaultAlignTempAlloca(Int8PtrTy, "return.sloc.ptr");
973 InitTempAlloca(ReturnLocation, llvm::ConstantPointerNull::get(Int8PtrTy));
976 // Emit subprogram debug descriptor.
977 if (CGDebugInfo *DI = getDebugInfo()) {
978 // Reconstruct the type from the argument list so that implicit parameters,
979 // such as 'this' and 'vtt', show up in the debug info. Preserve the calling
981 CallingConv CC = CallingConv::CC_C;
982 if (auto *FD = dyn_cast_or_null<FunctionDecl>(D))
983 if (const auto *SrcFnTy = FD->getType()->getAs<FunctionType>())
984 CC = SrcFnTy->getCallConv();
985 SmallVector<QualType, 16> ArgTypes;
986 for (const VarDecl *VD : Args)
987 ArgTypes.push_back(VD->getType());
988 QualType FnType = getContext().getFunctionType(
989 RetTy, ArgTypes, FunctionProtoType::ExtProtoInfo(CC));
990 DI->EmitFunctionStart(GD, Loc, StartLoc, FnType, CurFn, Builder);
993 if (ShouldInstrumentFunction()) {
994 if (CGM.getCodeGenOpts().InstrumentFunctions)
995 CurFn->addFnAttr("instrument-function-entry", "__cyg_profile_func_enter");
996 if (CGM.getCodeGenOpts().InstrumentFunctionsAfterInlining)
997 CurFn->addFnAttr("instrument-function-entry-inlined",
998 "__cyg_profile_func_enter");
999 if (CGM.getCodeGenOpts().InstrumentFunctionEntryBare)
1000 CurFn->addFnAttr("instrument-function-entry-inlined",
1001 "__cyg_profile_func_enter_bare");
1004 // Since emitting the mcount call here impacts optimizations such as function
1005 // inlining, we just add an attribute to insert a mcount call in backend.
1006 // The attribute "counting-function" is set to mcount function name which is
1007 // architecture dependent.
1008 if (CGM.getCodeGenOpts().InstrumentForProfiling) {
1009 if (CGM.getCodeGenOpts().CallFEntry)
1010 Fn->addFnAttr("fentry-call", "true");
1012 if (!CurFuncDecl || !CurFuncDecl->hasAttr<NoInstrumentFunctionAttr>()) {
1013 Fn->addFnAttr("instrument-function-entry-inlined",
1014 getTarget().getMCountName());
1019 if (RetTy->isVoidType()) {
1020 // Void type; nothing to return.
1021 ReturnValue = Address::invalid();
1023 // Count the implicit return.
1024 if (!endsWithReturn(D))
1026 } else if (CurFnInfo->getReturnInfo().getKind() == ABIArgInfo::Indirect &&
1027 !hasScalarEvaluationKind(CurFnInfo->getReturnType())) {
1028 // Indirect aggregate return; emit returned value directly into sret slot.
1029 // This reduces code size, and affects correctness in C++.
1030 auto AI = CurFn->arg_begin();
1031 if (CurFnInfo->getReturnInfo().isSRetAfterThis())
1033 ReturnValue = Address(&*AI, CurFnInfo->getReturnInfo().getIndirectAlign());
1034 } else if (CurFnInfo->getReturnInfo().getKind() == ABIArgInfo::InAlloca &&
1035 !hasScalarEvaluationKind(CurFnInfo->getReturnType())) {
1036 // Load the sret pointer from the argument struct and return into that.
1037 unsigned Idx = CurFnInfo->getReturnInfo().getInAllocaFieldIndex();
1038 llvm::Function::arg_iterator EI = CurFn->arg_end();
1040 llvm::Value *Addr = Builder.CreateStructGEP(nullptr, &*EI, Idx);
1041 Addr = Builder.CreateAlignedLoad(Addr, getPointerAlign(), "agg.result");
1042 ReturnValue = Address(Addr, getNaturalTypeAlignment(RetTy));
1044 ReturnValue = CreateIRTemp(RetTy, "retval");
1046 // Tell the epilog emitter to autorelease the result. We do this
1047 // now so that various specialized functions can suppress it
1048 // during their IR-generation.
1049 if (getLangOpts().ObjCAutoRefCount &&
1050 !CurFnInfo->isReturnsRetained() &&
1051 RetTy->isObjCRetainableType())
1052 AutoreleaseResult = true;
1055 EmitStartEHSpec(CurCodeDecl);
1057 PrologueCleanupDepth = EHStack.stable_begin();
1058 EmitFunctionProlog(*CurFnInfo, CurFn, Args);
1060 if (D && isa<CXXMethodDecl>(D) && cast<CXXMethodDecl>(D)->isInstance()) {
1061 CGM.getCXXABI().EmitInstanceFunctionProlog(*this);
1062 const CXXMethodDecl *MD = cast<CXXMethodDecl>(D);
1063 if (MD->getParent()->isLambda() &&
1064 MD->getOverloadedOperator() == OO_Call) {
1065 // We're in a lambda; figure out the captures.
1066 MD->getParent()->getCaptureFields(LambdaCaptureFields,
1067 LambdaThisCaptureField);
1068 if (LambdaThisCaptureField) {
1069 // If the lambda captures the object referred to by '*this' - either by
1070 // value or by reference, make sure CXXThisValue points to the correct
1073 // Get the lvalue for the field (which is a copy of the enclosing object
1074 // or contains the address of the enclosing object).
1075 LValue ThisFieldLValue = EmitLValueForLambdaField(LambdaThisCaptureField);
1076 if (!LambdaThisCaptureField->getType()->isPointerType()) {
1077 // If the enclosing object was captured by value, just use its address.
1078 CXXThisValue = ThisFieldLValue.getAddress().getPointer();
1080 // Load the lvalue pointed to by the field, since '*this' was captured
1083 EmitLoadOfLValue(ThisFieldLValue, SourceLocation()).getScalarVal();
1086 for (auto *FD : MD->getParent()->fields()) {
1087 if (FD->hasCapturedVLAType()) {
1088 auto *ExprArg = EmitLoadOfLValue(EmitLValueForLambdaField(FD),
1089 SourceLocation()).getScalarVal();
1090 auto VAT = FD->getCapturedVLAType();
1091 VLASizeMap[VAT->getSizeExpr()] = ExprArg;
1095 // Not in a lambda; just use 'this' from the method.
1096 // FIXME: Should we generate a new load for each use of 'this'? The
1097 // fast register allocator would be happier...
1098 CXXThisValue = CXXABIThisValue;
1101 // Check the 'this' pointer once per function, if it's available.
1102 if (CXXABIThisValue) {
1103 SanitizerSet SkippedChecks;
1104 SkippedChecks.set(SanitizerKind::ObjectSize, true);
1105 QualType ThisTy = MD->getThisType(getContext());
1107 // If this is the call operator of a lambda with no capture-default, it
1108 // may have a static invoker function, which may call this operator with
1109 // a null 'this' pointer.
1110 if (isLambdaCallOperator(MD) &&
1111 cast<CXXRecordDecl>(MD->getParent())->getLambdaCaptureDefault() ==
1113 SkippedChecks.set(SanitizerKind::Null, true);
1115 EmitTypeCheck(isa<CXXConstructorDecl>(MD) ? TCK_ConstructorCall
1117 Loc, CXXABIThisValue, ThisTy,
1118 getContext().getTypeAlignInChars(ThisTy->getPointeeType()),
1123 // If any of the arguments have a variably modified type, make sure to
1124 // emit the type size.
1125 for (FunctionArgList::const_iterator i = Args.begin(), e = Args.end();
1127 const VarDecl *VD = *i;
1129 // Dig out the type as written from ParmVarDecls; it's unclear whether
1130 // the standard (C99 6.9.1p10) requires this, but we're following the
1131 // precedent set by gcc.
1133 if (const ParmVarDecl *PVD = dyn_cast<ParmVarDecl>(VD))
1134 Ty = PVD->getOriginalType();
1138 if (Ty->isVariablyModifiedType())
1139 EmitVariablyModifiedType(Ty);
1141 // Emit a location at the end of the prologue.
1142 if (CGDebugInfo *DI = getDebugInfo())
1143 DI->EmitLocation(Builder, StartLoc);
1146 void CodeGenFunction::EmitFunctionBody(FunctionArgList &Args,
1148 incrementProfileCounter(Body);
1149 if (const CompoundStmt *S = dyn_cast<CompoundStmt>(Body))
1150 EmitCompoundStmtWithoutScope(*S);
1155 /// When instrumenting to collect profile data, the counts for some blocks
1156 /// such as switch cases need to not include the fall-through counts, so
1157 /// emit a branch around the instrumentation code. When not instrumenting,
1158 /// this just calls EmitBlock().
1159 void CodeGenFunction::EmitBlockWithFallThrough(llvm::BasicBlock *BB,
1161 llvm::BasicBlock *SkipCountBB = nullptr;
1162 if (HaveInsertPoint() && CGM.getCodeGenOpts().hasProfileClangInstr()) {
1163 // When instrumenting for profiling, the fallthrough to certain
1164 // statements needs to skip over the instrumentation code so that we
1165 // get an accurate count.
1166 SkipCountBB = createBasicBlock("skipcount");
1167 EmitBranch(SkipCountBB);
1170 uint64_t CurrentCount = getCurrentProfileCount();
1171 incrementProfileCounter(S);
1172 setCurrentProfileCount(getCurrentProfileCount() + CurrentCount);
1174 EmitBlock(SkipCountBB);
1177 /// Tries to mark the given function nounwind based on the
1178 /// non-existence of any throwing calls within it. We believe this is
1179 /// lightweight enough to do at -O0.
1180 static void TryMarkNoThrow(llvm::Function *F) {
1181 // LLVM treats 'nounwind' on a function as part of the type, so we
1182 // can't do this on functions that can be overwritten.
1183 if (F->isInterposable()) return;
1185 for (llvm::BasicBlock &BB : *F)
1186 for (llvm::Instruction &I : BB)
1190 F->setDoesNotThrow();
1193 QualType CodeGenFunction::BuildFunctionArgList(GlobalDecl GD,
1194 FunctionArgList &Args) {
1195 const FunctionDecl *FD = cast<FunctionDecl>(GD.getDecl());
1196 QualType ResTy = FD->getReturnType();
1198 const CXXMethodDecl *MD = dyn_cast<CXXMethodDecl>(FD);
1199 if (MD && MD->isInstance()) {
1200 if (CGM.getCXXABI().HasThisReturn(GD))
1201 ResTy = MD->getThisType(getContext());
1202 else if (CGM.getCXXABI().hasMostDerivedReturn(GD))
1203 ResTy = CGM.getContext().VoidPtrTy;
1204 CGM.getCXXABI().buildThisParam(*this, Args);
1207 // The base version of an inheriting constructor whose constructed base is a
1208 // virtual base is not passed any arguments (because it doesn't actually call
1209 // the inherited constructor).
1210 bool PassedParams = true;
1211 if (const CXXConstructorDecl *CD = dyn_cast<CXXConstructorDecl>(FD))
1212 if (auto Inherited = CD->getInheritedConstructor())
1214 getTypes().inheritingCtorHasParams(Inherited, GD.getCtorType());
1217 for (auto *Param : FD->parameters()) {
1218 Args.push_back(Param);
1219 if (!Param->hasAttr<PassObjectSizeAttr>())
1222 auto *Implicit = ImplicitParamDecl::Create(
1223 getContext(), Param->getDeclContext(), Param->getLocation(),
1224 /*Id=*/nullptr, getContext().getSizeType(), ImplicitParamDecl::Other);
1225 SizeArguments[Param] = Implicit;
1226 Args.push_back(Implicit);
1230 if (MD && (isa<CXXConstructorDecl>(MD) || isa<CXXDestructorDecl>(MD)))
1231 CGM.getCXXABI().addImplicitStructorParams(*this, ResTy, Args);
1237 shouldUseUndefinedBehaviorReturnOptimization(const FunctionDecl *FD,
1238 const ASTContext &Context) {
1239 QualType T = FD->getReturnType();
1240 // Avoid the optimization for functions that return a record type with a
1241 // trivial destructor or another trivially copyable type.
1242 if (const RecordType *RT = T.getCanonicalType()->getAs<RecordType>()) {
1243 if (const auto *ClassDecl = dyn_cast<CXXRecordDecl>(RT->getDecl()))
1244 return !ClassDecl->hasTrivialDestructor();
1246 return !T.isTriviallyCopyableType(Context);
1249 void CodeGenFunction::GenerateCode(GlobalDecl GD, llvm::Function *Fn,
1250 const CGFunctionInfo &FnInfo) {
1251 const FunctionDecl *FD = cast<FunctionDecl>(GD.getDecl());
1254 FunctionArgList Args;
1255 QualType ResTy = BuildFunctionArgList(GD, Args);
1257 // Check if we should generate debug info for this function.
1258 if (FD->hasAttr<NoDebugAttr>())
1259 DebugInfo = nullptr; // disable debug info indefinitely for this function
1261 // The function might not have a body if we're generating thunks for a
1262 // function declaration.
1263 SourceRange BodyRange;
1264 if (Stmt *Body = FD->getBody())
1265 BodyRange = Body->getSourceRange();
1267 BodyRange = FD->getLocation();
1268 CurEHLocation = BodyRange.getEnd();
1270 // Use the location of the start of the function to determine where
1271 // the function definition is located. By default use the location
1272 // of the declaration as the location for the subprogram. A function
1273 // may lack a declaration in the source code if it is created by code
1274 // gen. (examples: _GLOBAL__I_a, __cxx_global_array_dtor, thunk).
1275 SourceLocation Loc = FD->getLocation();
1277 // If this is a function specialization then use the pattern body
1278 // as the location for the function.
1279 if (const FunctionDecl *SpecDecl = FD->getTemplateInstantiationPattern())
1280 if (SpecDecl->hasBody(SpecDecl))
1281 Loc = SpecDecl->getLocation();
1283 Stmt *Body = FD->getBody();
1285 // Initialize helper which will detect jumps which can cause invalid lifetime
1287 if (Body && ShouldEmitLifetimeMarkers)
1288 Bypasses.Init(Body);
1290 // Emit the standard function prologue.
1291 StartFunction(GD, ResTy, Fn, FnInfo, Args, Loc, BodyRange.getBegin());
1293 // Generate the body of the function.
1294 PGO.assignRegionCounters(GD, CurFn);
1295 if (isa<CXXDestructorDecl>(FD))
1296 EmitDestructorBody(Args);
1297 else if (isa<CXXConstructorDecl>(FD))
1298 EmitConstructorBody(Args);
1299 else if (getLangOpts().CUDA &&
1300 !getLangOpts().CUDAIsDevice &&
1301 FD->hasAttr<CUDAGlobalAttr>())
1302 CGM.getCUDARuntime().emitDeviceStub(*this, Args);
1303 else if (isa<CXXMethodDecl>(FD) &&
1304 cast<CXXMethodDecl>(FD)->isLambdaStaticInvoker()) {
1305 // The lambda static invoker function is special, because it forwards or
1306 // clones the body of the function call operator (but is actually static).
1307 EmitLambdaStaticInvokeBody(cast<CXXMethodDecl>(FD));
1308 } else if (FD->isDefaulted() && isa<CXXMethodDecl>(FD) &&
1309 (cast<CXXMethodDecl>(FD)->isCopyAssignmentOperator() ||
1310 cast<CXXMethodDecl>(FD)->isMoveAssignmentOperator())) {
1311 // Implicit copy-assignment gets the same special treatment as implicit
1312 // copy-constructors.
1313 emitImplicitAssignmentOperatorBody(Args);
1315 EmitFunctionBody(Args, Body);
1317 llvm_unreachable("no definition for emitted function");
1319 // C++11 [stmt.return]p2:
1320 // Flowing off the end of a function [...] results in undefined behavior in
1321 // a value-returning function.
1323 // If the '}' that terminates a function is reached, and the value of the
1324 // function call is used by the caller, the behavior is undefined.
1325 if (getLangOpts().CPlusPlus && !FD->hasImplicitReturnZero() && !SawAsmBlock &&
1326 !FD->getReturnType()->isVoidType() && Builder.GetInsertBlock()) {
1327 bool ShouldEmitUnreachable =
1328 CGM.getCodeGenOpts().StrictReturn ||
1329 shouldUseUndefinedBehaviorReturnOptimization(FD, getContext());
1330 if (SanOpts.has(SanitizerKind::Return)) {
1331 SanitizerScope SanScope(this);
1332 llvm::Value *IsFalse = Builder.getFalse();
1333 EmitCheck(std::make_pair(IsFalse, SanitizerKind::Return),
1334 SanitizerHandler::MissingReturn,
1335 EmitCheckSourceLocation(FD->getLocation()), None);
1336 } else if (ShouldEmitUnreachable) {
1337 if (CGM.getCodeGenOpts().OptimizationLevel == 0)
1338 EmitTrapCall(llvm::Intrinsic::trap);
1340 if (SanOpts.has(SanitizerKind::Return) || ShouldEmitUnreachable) {
1341 Builder.CreateUnreachable();
1342 Builder.ClearInsertionPoint();
1346 // Emit the standard function epilogue.
1347 FinishFunction(BodyRange.getEnd());
1349 // If we haven't marked the function nothrow through other means, do
1350 // a quick pass now to see if we can.
1351 if (!CurFn->doesNotThrow())
1352 TryMarkNoThrow(CurFn);
1355 /// ContainsLabel - Return true if the statement contains a label in it. If
1356 /// this statement is not executed normally, it not containing a label means
1357 /// that we can just remove the code.
1358 bool CodeGenFunction::ContainsLabel(const Stmt *S, bool IgnoreCaseStmts) {
1359 // Null statement, not a label!
1360 if (!S) return false;
1362 // If this is a label, we have to emit the code, consider something like:
1363 // if (0) { ... foo: bar(); } goto foo;
1365 // TODO: If anyone cared, we could track __label__'s, since we know that you
1366 // can't jump to one from outside their declared region.
1367 if (isa<LabelStmt>(S))
1370 // If this is a case/default statement, and we haven't seen a switch, we have
1371 // to emit the code.
1372 if (isa<SwitchCase>(S) && !IgnoreCaseStmts)
1375 // If this is a switch statement, we want to ignore cases below it.
1376 if (isa<SwitchStmt>(S))
1377 IgnoreCaseStmts = true;
1379 // Scan subexpressions for verboten labels.
1380 for (const Stmt *SubStmt : S->children())
1381 if (ContainsLabel(SubStmt, IgnoreCaseStmts))
1387 /// containsBreak - Return true if the statement contains a break out of it.
1388 /// If the statement (recursively) contains a switch or loop with a break
1389 /// inside of it, this is fine.
1390 bool CodeGenFunction::containsBreak(const Stmt *S) {
1391 // Null statement, not a label!
1392 if (!S) return false;
1394 // If this is a switch or loop that defines its own break scope, then we can
1395 // include it and anything inside of it.
1396 if (isa<SwitchStmt>(S) || isa<WhileStmt>(S) || isa<DoStmt>(S) ||
1400 if (isa<BreakStmt>(S))
1403 // Scan subexpressions for verboten breaks.
1404 for (const Stmt *SubStmt : S->children())
1405 if (containsBreak(SubStmt))
1411 bool CodeGenFunction::mightAddDeclToScope(const Stmt *S) {
1412 if (!S) return false;
1414 // Some statement kinds add a scope and thus never add a decl to the current
1415 // scope. Note, this list is longer than the list of statements that might
1416 // have an unscoped decl nested within them, but this way is conservatively
1417 // correct even if more statement kinds are added.
1418 if (isa<IfStmt>(S) || isa<SwitchStmt>(S) || isa<WhileStmt>(S) ||
1419 isa<DoStmt>(S) || isa<ForStmt>(S) || isa<CompoundStmt>(S) ||
1420 isa<CXXForRangeStmt>(S) || isa<CXXTryStmt>(S) ||
1421 isa<ObjCForCollectionStmt>(S) || isa<ObjCAtTryStmt>(S))
1424 if (isa<DeclStmt>(S))
1427 for (const Stmt *SubStmt : S->children())
1428 if (mightAddDeclToScope(SubStmt))
1434 /// ConstantFoldsToSimpleInteger - If the specified expression does not fold
1435 /// to a constant, or if it does but contains a label, return false. If it
1436 /// constant folds return true and set the boolean result in Result.
1437 bool CodeGenFunction::ConstantFoldsToSimpleInteger(const Expr *Cond,
1440 llvm::APSInt ResultInt;
1441 if (!ConstantFoldsToSimpleInteger(Cond, ResultInt, AllowLabels))
1444 ResultBool = ResultInt.getBoolValue();
1448 /// ConstantFoldsToSimpleInteger - If the specified expression does not fold
1449 /// to a constant, or if it does but contains a label, return false. If it
1450 /// constant folds return true and set the folded value.
1451 bool CodeGenFunction::ConstantFoldsToSimpleInteger(const Expr *Cond,
1452 llvm::APSInt &ResultInt,
1454 // FIXME: Rename and handle conversion of other evaluatable things
1457 if (!Cond->EvaluateAsInt(Int, getContext()))
1458 return false; // Not foldable, not integer or not fully evaluatable.
1460 if (!AllowLabels && CodeGenFunction::ContainsLabel(Cond))
1461 return false; // Contains a label.
1469 /// EmitBranchOnBoolExpr - Emit a branch on a boolean condition (e.g. for an if
1470 /// statement) to the specified blocks. Based on the condition, this might try
1471 /// to simplify the codegen of the conditional based on the branch.
1473 void CodeGenFunction::EmitBranchOnBoolExpr(const Expr *Cond,
1474 llvm::BasicBlock *TrueBlock,
1475 llvm::BasicBlock *FalseBlock,
1476 uint64_t TrueCount) {
1477 Cond = Cond->IgnoreParens();
1479 if (const BinaryOperator *CondBOp = dyn_cast<BinaryOperator>(Cond)) {
1481 // Handle X && Y in a condition.
1482 if (CondBOp->getOpcode() == BO_LAnd) {
1483 // If we have "1 && X", simplify the code. "0 && X" would have constant
1484 // folded if the case was simple enough.
1485 bool ConstantBool = false;
1486 if (ConstantFoldsToSimpleInteger(CondBOp->getLHS(), ConstantBool) &&
1488 // br(1 && X) -> br(X).
1489 incrementProfileCounter(CondBOp);
1490 return EmitBranchOnBoolExpr(CondBOp->getRHS(), TrueBlock, FalseBlock,
1494 // If we have "X && 1", simplify the code to use an uncond branch.
1495 // "X && 0" would have been constant folded to 0.
1496 if (ConstantFoldsToSimpleInteger(CondBOp->getRHS(), ConstantBool) &&
1498 // br(X && 1) -> br(X).
1499 return EmitBranchOnBoolExpr(CondBOp->getLHS(), TrueBlock, FalseBlock,
1503 // Emit the LHS as a conditional. If the LHS conditional is false, we
1504 // want to jump to the FalseBlock.
1505 llvm::BasicBlock *LHSTrue = createBasicBlock("land.lhs.true");
1506 // The counter tells us how often we evaluate RHS, and all of TrueCount
1507 // can be propagated to that branch.
1508 uint64_t RHSCount = getProfileCount(CondBOp->getRHS());
1510 ConditionalEvaluation eval(*this);
1512 ApplyDebugLocation DL(*this, Cond);
1513 EmitBranchOnBoolExpr(CondBOp->getLHS(), LHSTrue, FalseBlock, RHSCount);
1517 incrementProfileCounter(CondBOp);
1518 setCurrentProfileCount(getProfileCount(CondBOp->getRHS()));
1520 // Any temporaries created here are conditional.
1522 EmitBranchOnBoolExpr(CondBOp->getRHS(), TrueBlock, FalseBlock, TrueCount);
1528 if (CondBOp->getOpcode() == BO_LOr) {
1529 // If we have "0 || X", simplify the code. "1 || X" would have constant
1530 // folded if the case was simple enough.
1531 bool ConstantBool = false;
1532 if (ConstantFoldsToSimpleInteger(CondBOp->getLHS(), ConstantBool) &&
1534 // br(0 || X) -> br(X).
1535 incrementProfileCounter(CondBOp);
1536 return EmitBranchOnBoolExpr(CondBOp->getRHS(), TrueBlock, FalseBlock,
1540 // If we have "X || 0", simplify the code to use an uncond branch.
1541 // "X || 1" would have been constant folded to 1.
1542 if (ConstantFoldsToSimpleInteger(CondBOp->getRHS(), ConstantBool) &&
1544 // br(X || 0) -> br(X).
1545 return EmitBranchOnBoolExpr(CondBOp->getLHS(), TrueBlock, FalseBlock,
1549 // Emit the LHS as a conditional. If the LHS conditional is true, we
1550 // want to jump to the TrueBlock.
1551 llvm::BasicBlock *LHSFalse = createBasicBlock("lor.lhs.false");
1552 // We have the count for entry to the RHS and for the whole expression
1553 // being true, so we can divy up True count between the short circuit and
1556 getCurrentProfileCount() - getProfileCount(CondBOp->getRHS());
1557 uint64_t RHSCount = TrueCount - LHSCount;
1559 ConditionalEvaluation eval(*this);
1561 ApplyDebugLocation DL(*this, Cond);
1562 EmitBranchOnBoolExpr(CondBOp->getLHS(), TrueBlock, LHSFalse, LHSCount);
1563 EmitBlock(LHSFalse);
1566 incrementProfileCounter(CondBOp);
1567 setCurrentProfileCount(getProfileCount(CondBOp->getRHS()));
1569 // Any temporaries created here are conditional.
1571 EmitBranchOnBoolExpr(CondBOp->getRHS(), TrueBlock, FalseBlock, RHSCount);
1579 if (const UnaryOperator *CondUOp = dyn_cast<UnaryOperator>(Cond)) {
1580 // br(!x, t, f) -> br(x, f, t)
1581 if (CondUOp->getOpcode() == UO_LNot) {
1582 // Negate the count.
1583 uint64_t FalseCount = getCurrentProfileCount() - TrueCount;
1584 // Negate the condition and swap the destination blocks.
1585 return EmitBranchOnBoolExpr(CondUOp->getSubExpr(), FalseBlock, TrueBlock,
1590 if (const ConditionalOperator *CondOp = dyn_cast<ConditionalOperator>(Cond)) {
1591 // br(c ? x : y, t, f) -> br(c, br(x, t, f), br(y, t, f))
1592 llvm::BasicBlock *LHSBlock = createBasicBlock("cond.true");
1593 llvm::BasicBlock *RHSBlock = createBasicBlock("cond.false");
1595 ConditionalEvaluation cond(*this);
1596 EmitBranchOnBoolExpr(CondOp->getCond(), LHSBlock, RHSBlock,
1597 getProfileCount(CondOp));
1599 // When computing PGO branch weights, we only know the overall count for
1600 // the true block. This code is essentially doing tail duplication of the
1601 // naive code-gen, introducing new edges for which counts are not
1602 // available. Divide the counts proportionally between the LHS and RHS of
1603 // the conditional operator.
1604 uint64_t LHSScaledTrueCount = 0;
1607 getProfileCount(CondOp) / (double)getCurrentProfileCount();
1608 LHSScaledTrueCount = TrueCount * LHSRatio;
1612 EmitBlock(LHSBlock);
1613 incrementProfileCounter(CondOp);
1615 ApplyDebugLocation DL(*this, Cond);
1616 EmitBranchOnBoolExpr(CondOp->getLHS(), TrueBlock, FalseBlock,
1617 LHSScaledTrueCount);
1622 EmitBlock(RHSBlock);
1623 EmitBranchOnBoolExpr(CondOp->getRHS(), TrueBlock, FalseBlock,
1624 TrueCount - LHSScaledTrueCount);
1630 if (const CXXThrowExpr *Throw = dyn_cast<CXXThrowExpr>(Cond)) {
1631 // Conditional operator handling can give us a throw expression as a
1632 // condition for a case like:
1633 // br(c ? throw x : y, t, f) -> br(c, br(throw x, t, f), br(y, t, f)
1635 // br(c, throw x, br(y, t, f))
1636 EmitCXXThrowExpr(Throw, /*KeepInsertionPoint*/false);
1640 // If the branch has a condition wrapped by __builtin_unpredictable,
1641 // create metadata that specifies that the branch is unpredictable.
1642 // Don't bother if not optimizing because that metadata would not be used.
1643 llvm::MDNode *Unpredictable = nullptr;
1644 auto *Call = dyn_cast<CallExpr>(Cond);
1645 if (Call && CGM.getCodeGenOpts().OptimizationLevel != 0) {
1646 auto *FD = dyn_cast_or_null<FunctionDecl>(Call->getCalleeDecl());
1647 if (FD && FD->getBuiltinID() == Builtin::BI__builtin_unpredictable) {
1648 llvm::MDBuilder MDHelper(getLLVMContext());
1649 Unpredictable = MDHelper.createUnpredictable();
1653 // Create branch weights based on the number of times we get here and the
1654 // number of times the condition should be true.
1655 uint64_t CurrentCount = std::max(getCurrentProfileCount(), TrueCount);
1656 llvm::MDNode *Weights =
1657 createProfileWeights(TrueCount, CurrentCount - TrueCount);
1659 // Emit the code with the fully general case.
1662 ApplyDebugLocation DL(*this, Cond);
1663 CondV = EvaluateExprAsBool(Cond);
1665 Builder.CreateCondBr(CondV, TrueBlock, FalseBlock, Weights, Unpredictable);
1668 /// ErrorUnsupported - Print out an error that codegen doesn't support the
1669 /// specified stmt yet.
1670 void CodeGenFunction::ErrorUnsupported(const Stmt *S, const char *Type) {
1671 CGM.ErrorUnsupported(S, Type);
1674 /// emitNonZeroVLAInit - Emit the "zero" initialization of a
1675 /// variable-length array whose elements have a non-zero bit-pattern.
1677 /// \param baseType the inner-most element type of the array
1678 /// \param src - a char* pointing to the bit-pattern for a single
1679 /// base element of the array
1680 /// \param sizeInChars - the total size of the VLA, in chars
1681 static void emitNonZeroVLAInit(CodeGenFunction &CGF, QualType baseType,
1682 Address dest, Address src,
1683 llvm::Value *sizeInChars) {
1684 CGBuilderTy &Builder = CGF.Builder;
1686 CharUnits baseSize = CGF.getContext().getTypeSizeInChars(baseType);
1687 llvm::Value *baseSizeInChars
1688 = llvm::ConstantInt::get(CGF.IntPtrTy, baseSize.getQuantity());
1691 Builder.CreateElementBitCast(dest, CGF.Int8Ty, "vla.begin");
1693 Builder.CreateInBoundsGEP(begin.getPointer(), sizeInChars, "vla.end");
1695 llvm::BasicBlock *originBB = CGF.Builder.GetInsertBlock();
1696 llvm::BasicBlock *loopBB = CGF.createBasicBlock("vla-init.loop");
1697 llvm::BasicBlock *contBB = CGF.createBasicBlock("vla-init.cont");
1699 // Make a loop over the VLA. C99 guarantees that the VLA element
1700 // count must be nonzero.
1701 CGF.EmitBlock(loopBB);
1703 llvm::PHINode *cur = Builder.CreatePHI(begin.getType(), 2, "vla.cur");
1704 cur->addIncoming(begin.getPointer(), originBB);
1706 CharUnits curAlign =
1707 dest.getAlignment().alignmentOfArrayElement(baseSize);
1709 // memcpy the individual element bit-pattern.
1710 Builder.CreateMemCpy(Address(cur, curAlign), src, baseSizeInChars,
1711 /*volatile*/ false);
1713 // Go to the next element.
1715 Builder.CreateInBoundsGEP(CGF.Int8Ty, cur, baseSizeInChars, "vla.next");
1717 // Leave if that's the end of the VLA.
1718 llvm::Value *done = Builder.CreateICmpEQ(next, end, "vla-init.isdone");
1719 Builder.CreateCondBr(done, contBB, loopBB);
1720 cur->addIncoming(next, loopBB);
1722 CGF.EmitBlock(contBB);
1726 CodeGenFunction::EmitNullInitialization(Address DestPtr, QualType Ty) {
1727 // Ignore empty classes in C++.
1728 if (getLangOpts().CPlusPlus) {
1729 if (const RecordType *RT = Ty->getAs<RecordType>()) {
1730 if (cast<CXXRecordDecl>(RT->getDecl())->isEmpty())
1735 // Cast the dest ptr to the appropriate i8 pointer type.
1736 if (DestPtr.getElementType() != Int8Ty)
1737 DestPtr = Builder.CreateElementBitCast(DestPtr, Int8Ty);
1739 // Get size and alignment info for this aggregate.
1740 CharUnits size = getContext().getTypeSizeInChars(Ty);
1742 llvm::Value *SizeVal;
1743 const VariableArrayType *vla;
1745 // Don't bother emitting a zero-byte memset.
1746 if (size.isZero()) {
1747 // But note that getTypeInfo returns 0 for a VLA.
1748 if (const VariableArrayType *vlaType =
1749 dyn_cast_or_null<VariableArrayType>(
1750 getContext().getAsArrayType(Ty))) {
1752 llvm::Value *numElts;
1753 std::tie(numElts, eltType) = getVLASize(vlaType);
1756 CharUnits eltSize = getContext().getTypeSizeInChars(eltType);
1757 if (!eltSize.isOne())
1758 SizeVal = Builder.CreateNUWMul(SizeVal, CGM.getSize(eltSize));
1764 SizeVal = CGM.getSize(size);
1768 // If the type contains a pointer to data member we can't memset it to zero.
1769 // Instead, create a null constant and copy it to the destination.
1770 // TODO: there are other patterns besides zero that we can usefully memset,
1771 // like -1, which happens to be the pattern used by member-pointers.
1772 if (!CGM.getTypes().isZeroInitializable(Ty)) {
1773 // For a VLA, emit a single element, then splat that over the VLA.
1774 if (vla) Ty = getContext().getBaseElementType(vla);
1776 llvm::Constant *NullConstant = CGM.EmitNullConstant(Ty);
1778 llvm::GlobalVariable *NullVariable =
1779 new llvm::GlobalVariable(CGM.getModule(), NullConstant->getType(),
1780 /*isConstant=*/true,
1781 llvm::GlobalVariable::PrivateLinkage,
1782 NullConstant, Twine());
1783 CharUnits NullAlign = DestPtr.getAlignment();
1784 NullVariable->setAlignment(NullAlign.getQuantity());
1785 Address SrcPtr(Builder.CreateBitCast(NullVariable, Builder.getInt8PtrTy()),
1788 if (vla) return emitNonZeroVLAInit(*this, Ty, DestPtr, SrcPtr, SizeVal);
1790 // Get and call the appropriate llvm.memcpy overload.
1791 Builder.CreateMemCpy(DestPtr, SrcPtr, SizeVal, false);
1795 // Otherwise, just memset the whole thing to zero. This is legal
1796 // because in LLVM, all default initializers (other than the ones we just
1797 // handled above) are guaranteed to have a bit pattern of all zeros.
1798 Builder.CreateMemSet(DestPtr, Builder.getInt8(0), SizeVal, false);
1801 llvm::BlockAddress *CodeGenFunction::GetAddrOfLabel(const LabelDecl *L) {
1802 // Make sure that there is a block for the indirect goto.
1803 if (!IndirectBranch)
1804 GetIndirectGotoBlock();
1806 llvm::BasicBlock *BB = getJumpDestForLabel(L).getBlock();
1808 // Make sure the indirect branch includes all of the address-taken blocks.
1809 IndirectBranch->addDestination(BB);
1810 return llvm::BlockAddress::get(CurFn, BB);
1813 llvm::BasicBlock *CodeGenFunction::GetIndirectGotoBlock() {
1814 // If we already made the indirect branch for indirect goto, return its block.
1815 if (IndirectBranch) return IndirectBranch->getParent();
1817 CGBuilderTy TmpBuilder(*this, createBasicBlock("indirectgoto"));
1819 // Create the PHI node that indirect gotos will add entries to.
1820 llvm::Value *DestVal = TmpBuilder.CreatePHI(Int8PtrTy, 0,
1821 "indirect.goto.dest");
1823 // Create the indirect branch instruction.
1824 IndirectBranch = TmpBuilder.CreateIndirectBr(DestVal);
1825 return IndirectBranch->getParent();
1828 /// Computes the length of an array in elements, as well as the base
1829 /// element type and a properly-typed first element pointer.
1830 llvm::Value *CodeGenFunction::emitArrayLength(const ArrayType *origArrayType,
1833 const ArrayType *arrayType = origArrayType;
1835 // If it's a VLA, we have to load the stored size. Note that
1836 // this is the size of the VLA in bytes, not its size in elements.
1837 llvm::Value *numVLAElements = nullptr;
1838 if (isa<VariableArrayType>(arrayType)) {
1839 numVLAElements = getVLASize(cast<VariableArrayType>(arrayType)).first;
1841 // Walk into all VLAs. This doesn't require changes to addr,
1842 // which has type T* where T is the first non-VLA element type.
1844 QualType elementType = arrayType->getElementType();
1845 arrayType = getContext().getAsArrayType(elementType);
1847 // If we only have VLA components, 'addr' requires no adjustment.
1849 baseType = elementType;
1850 return numVLAElements;
1852 } while (isa<VariableArrayType>(arrayType));
1854 // We get out here only if we find a constant array type
1858 // We have some number of constant-length arrays, so addr should
1859 // have LLVM type [M x [N x [...]]]*. Build a GEP that walks
1860 // down to the first element of addr.
1861 SmallVector<llvm::Value*, 8> gepIndices;
1863 // GEP down to the array type.
1864 llvm::ConstantInt *zero = Builder.getInt32(0);
1865 gepIndices.push_back(zero);
1867 uint64_t countFromCLAs = 1;
1870 llvm::ArrayType *llvmArrayType =
1871 dyn_cast<llvm::ArrayType>(addr.getElementType());
1872 while (llvmArrayType) {
1873 assert(isa<ConstantArrayType>(arrayType));
1874 assert(cast<ConstantArrayType>(arrayType)->getSize().getZExtValue()
1875 == llvmArrayType->getNumElements());
1877 gepIndices.push_back(zero);
1878 countFromCLAs *= llvmArrayType->getNumElements();
1879 eltType = arrayType->getElementType();
1882 dyn_cast<llvm::ArrayType>(llvmArrayType->getElementType());
1883 arrayType = getContext().getAsArrayType(arrayType->getElementType());
1884 assert((!llvmArrayType || arrayType) &&
1885 "LLVM and Clang types are out-of-synch");
1889 // From this point onwards, the Clang array type has been emitted
1890 // as some other type (probably a packed struct). Compute the array
1891 // size, and just emit the 'begin' expression as a bitcast.
1894 cast<ConstantArrayType>(arrayType)->getSize().getZExtValue();
1895 eltType = arrayType->getElementType();
1896 arrayType = getContext().getAsArrayType(eltType);
1899 llvm::Type *baseType = ConvertType(eltType);
1900 addr = Builder.CreateElementBitCast(addr, baseType, "array.begin");
1902 // Create the actual GEP.
1903 addr = Address(Builder.CreateInBoundsGEP(addr.getPointer(),
1904 gepIndices, "array.begin"),
1905 addr.getAlignment());
1910 llvm::Value *numElements
1911 = llvm::ConstantInt::get(SizeTy, countFromCLAs);
1913 // If we had any VLA dimensions, factor them in.
1915 numElements = Builder.CreateNUWMul(numVLAElements, numElements);
1920 std::pair<llvm::Value*, QualType>
1921 CodeGenFunction::getVLASize(QualType type) {
1922 const VariableArrayType *vla = getContext().getAsVariableArrayType(type);
1923 assert(vla && "type was not a variable array type!");
1924 return getVLASize(vla);
1927 std::pair<llvm::Value*, QualType>
1928 CodeGenFunction::getVLASize(const VariableArrayType *type) {
1929 // The number of elements so far; always size_t.
1930 llvm::Value *numElements = nullptr;
1932 QualType elementType;
1934 elementType = type->getElementType();
1935 llvm::Value *vlaSize = VLASizeMap[type->getSizeExpr()];
1936 assert(vlaSize && "no size for VLA!");
1937 assert(vlaSize->getType() == SizeTy);
1940 numElements = vlaSize;
1942 // It's undefined behavior if this wraps around, so mark it that way.
1943 // FIXME: Teach -fsanitize=undefined to trap this.
1944 numElements = Builder.CreateNUWMul(numElements, vlaSize);
1946 } while ((type = getContext().getAsVariableArrayType(elementType)));
1948 return std::pair<llvm::Value*,QualType>(numElements, elementType);
1951 void CodeGenFunction::EmitVariablyModifiedType(QualType type) {
1952 assert(type->isVariablyModifiedType() &&
1953 "Must pass variably modified type to EmitVLASizes!");
1955 EnsureInsertPoint();
1957 // We're going to walk down into the type and look for VLA
1960 assert(type->isVariablyModifiedType());
1962 const Type *ty = type.getTypePtr();
1963 switch (ty->getTypeClass()) {
1965 #define TYPE(Class, Base)
1966 #define ABSTRACT_TYPE(Class, Base)
1967 #define NON_CANONICAL_TYPE(Class, Base)
1968 #define DEPENDENT_TYPE(Class, Base) case Type::Class:
1969 #define NON_CANONICAL_UNLESS_DEPENDENT_TYPE(Class, Base)
1970 #include "clang/AST/TypeNodes.def"
1971 llvm_unreachable("unexpected dependent type!");
1973 // These types are never variably-modified.
1977 case Type::ExtVector:
1980 case Type::Elaborated:
1981 case Type::TemplateSpecialization:
1982 case Type::ObjCTypeParam:
1983 case Type::ObjCObject:
1984 case Type::ObjCInterface:
1985 case Type::ObjCObjectPointer:
1986 llvm_unreachable("type class is never variably-modified!");
1988 case Type::Adjusted:
1989 type = cast<AdjustedType>(ty)->getAdjustedType();
1993 type = cast<DecayedType>(ty)->getPointeeType();
1997 type = cast<PointerType>(ty)->getPointeeType();
2000 case Type::BlockPointer:
2001 type = cast<BlockPointerType>(ty)->getPointeeType();
2004 case Type::LValueReference:
2005 case Type::RValueReference:
2006 type = cast<ReferenceType>(ty)->getPointeeType();
2009 case Type::MemberPointer:
2010 type = cast<MemberPointerType>(ty)->getPointeeType();
2013 case Type::ConstantArray:
2014 case Type::IncompleteArray:
2015 // Losing element qualification here is fine.
2016 type = cast<ArrayType>(ty)->getElementType();
2019 case Type::VariableArray: {
2020 // Losing element qualification here is fine.
2021 const VariableArrayType *vat = cast<VariableArrayType>(ty);
2023 // Unknown size indication requires no size computation.
2024 // Otherwise, evaluate and record it.
2025 if (const Expr *size = vat->getSizeExpr()) {
2026 // It's possible that we might have emitted this already,
2027 // e.g. with a typedef and a pointer to it.
2028 llvm::Value *&entry = VLASizeMap[size];
2030 llvm::Value *Size = EmitScalarExpr(size);
2033 // If the size is an expression that is not an integer constant
2034 // expression [...] each time it is evaluated it shall have a value
2035 // greater than zero.
2036 if (SanOpts.has(SanitizerKind::VLABound) &&
2037 size->getType()->isSignedIntegerType()) {
2038 SanitizerScope SanScope(this);
2039 llvm::Value *Zero = llvm::Constant::getNullValue(Size->getType());
2040 llvm::Constant *StaticArgs[] = {
2041 EmitCheckSourceLocation(size->getLocStart()),
2042 EmitCheckTypeDescriptor(size->getType())
2044 EmitCheck(std::make_pair(Builder.CreateICmpSGT(Size, Zero),
2045 SanitizerKind::VLABound),
2046 SanitizerHandler::VLABoundNotPositive, StaticArgs, Size);
2049 // Always zexting here would be wrong if it weren't
2050 // undefined behavior to have a negative bound.
2051 entry = Builder.CreateIntCast(Size, SizeTy, /*signed*/ false);
2054 type = vat->getElementType();
2058 case Type::FunctionProto:
2059 case Type::FunctionNoProto:
2060 type = cast<FunctionType>(ty)->getReturnType();
2065 case Type::UnaryTransform:
2066 case Type::Attributed:
2067 case Type::SubstTemplateTypeParm:
2068 case Type::PackExpansion:
2069 // Keep walking after single level desugaring.
2070 type = type.getSingleStepDesugaredType(getContext());
2074 case Type::Decltype:
2076 case Type::DeducedTemplateSpecialization:
2077 // Stop walking: nothing to do.
2080 case Type::TypeOfExpr:
2081 // Stop walking: emit typeof expression.
2082 EmitIgnoredExpr(cast<TypeOfExprType>(ty)->getUnderlyingExpr());
2086 type = cast<AtomicType>(ty)->getValueType();
2090 type = cast<PipeType>(ty)->getElementType();
2093 } while (type->isVariablyModifiedType());
2096 Address CodeGenFunction::EmitVAListRef(const Expr* E) {
2097 if (getContext().getBuiltinVaListType()->isArrayType())
2098 return EmitPointerWithAlignment(E);
2099 return EmitLValue(E).getAddress();
2102 Address CodeGenFunction::EmitMSVAListRef(const Expr *E) {
2103 return EmitLValue(E).getAddress();
2106 void CodeGenFunction::EmitDeclRefExprDbgValue(const DeclRefExpr *E,
2107 const APValue &Init) {
2108 assert(!Init.isUninit() && "Invalid DeclRefExpr initializer!");
2109 if (CGDebugInfo *Dbg = getDebugInfo())
2110 if (CGM.getCodeGenOpts().getDebugInfo() >= codegenoptions::LimitedDebugInfo)
2111 Dbg->EmitGlobalVariable(E->getDecl(), Init);
2114 CodeGenFunction::PeepholeProtection
2115 CodeGenFunction::protectFromPeepholes(RValue rvalue) {
2116 // At the moment, the only aggressive peephole we do in IR gen
2117 // is trunc(zext) folding, but if we add more, we can easily
2118 // extend this protection.
2120 if (!rvalue.isScalar()) return PeepholeProtection();
2121 llvm::Value *value = rvalue.getScalarVal();
2122 if (!isa<llvm::ZExtInst>(value)) return PeepholeProtection();
2124 // Just make an extra bitcast.
2125 assert(HaveInsertPoint());
2126 llvm::Instruction *inst = new llvm::BitCastInst(value, value->getType(), "",
2127 Builder.GetInsertBlock());
2129 PeepholeProtection protection;
2130 protection.Inst = inst;
2134 void CodeGenFunction::unprotectFromPeepholes(PeepholeProtection protection) {
2135 if (!protection.Inst) return;
2137 // In theory, we could try to duplicate the peepholes now, but whatever.
2138 protection.Inst->eraseFromParent();
2141 llvm::Value *CodeGenFunction::EmitAnnotationCall(llvm::Value *AnnotationFn,
2142 llvm::Value *AnnotatedVal,
2143 StringRef AnnotationStr,
2144 SourceLocation Location) {
2145 llvm::Value *Args[4] = {
2147 Builder.CreateBitCast(CGM.EmitAnnotationString(AnnotationStr), Int8PtrTy),
2148 Builder.CreateBitCast(CGM.EmitAnnotationUnit(Location), Int8PtrTy),
2149 CGM.EmitAnnotationLineNo(Location)
2151 return Builder.CreateCall(AnnotationFn, Args);
2154 void CodeGenFunction::EmitVarAnnotations(const VarDecl *D, llvm::Value *V) {
2155 assert(D->hasAttr<AnnotateAttr>() && "no annotate attribute");
2156 // FIXME We create a new bitcast for every annotation because that's what
2157 // llvm-gcc was doing.
2158 for (const auto *I : D->specific_attrs<AnnotateAttr>())
2159 EmitAnnotationCall(CGM.getIntrinsic(llvm::Intrinsic::var_annotation),
2160 Builder.CreateBitCast(V, CGM.Int8PtrTy, V->getName()),
2161 I->getAnnotation(), D->getLocation());
2164 Address CodeGenFunction::EmitFieldAnnotations(const FieldDecl *D,
2166 assert(D->hasAttr<AnnotateAttr>() && "no annotate attribute");
2167 llvm::Value *V = Addr.getPointer();
2168 llvm::Type *VTy = V->getType();
2169 llvm::Value *F = CGM.getIntrinsic(llvm::Intrinsic::ptr_annotation,
2172 for (const auto *I : D->specific_attrs<AnnotateAttr>()) {
2173 // FIXME Always emit the cast inst so we can differentiate between
2174 // annotation on the first field of a struct and annotation on the struct
2176 if (VTy != CGM.Int8PtrTy)
2177 V = Builder.Insert(new llvm::BitCastInst(V, CGM.Int8PtrTy));
2178 V = EmitAnnotationCall(F, V, I->getAnnotation(), D->getLocation());
2179 V = Builder.CreateBitCast(V, VTy);
2182 return Address(V, Addr.getAlignment());
2185 CodeGenFunction::CGCapturedStmtInfo::~CGCapturedStmtInfo() { }
2187 CodeGenFunction::SanitizerScope::SanitizerScope(CodeGenFunction *CGF)
2189 assert(!CGF->IsSanitizerScope);
2190 CGF->IsSanitizerScope = true;
2193 CodeGenFunction::SanitizerScope::~SanitizerScope() {
2194 CGF->IsSanitizerScope = false;
2197 void CodeGenFunction::InsertHelper(llvm::Instruction *I,
2198 const llvm::Twine &Name,
2199 llvm::BasicBlock *BB,
2200 llvm::BasicBlock::iterator InsertPt) const {
2201 LoopStack.InsertHelper(I);
2202 if (IsSanitizerScope)
2203 CGM.getSanitizerMetadata()->disableSanitizerForInstruction(I);
2206 void CGBuilderInserter::InsertHelper(
2207 llvm::Instruction *I, const llvm::Twine &Name, llvm::BasicBlock *BB,
2208 llvm::BasicBlock::iterator InsertPt) const {
2209 llvm::IRBuilderDefaultInserter::InsertHelper(I, Name, BB, InsertPt);
2211 CGF->InsertHelper(I, Name, BB, InsertPt);
2214 static bool hasRequiredFeatures(const SmallVectorImpl<StringRef> &ReqFeatures,
2215 CodeGenModule &CGM, const FunctionDecl *FD,
2216 std::string &FirstMissing) {
2217 // If there aren't any required features listed then go ahead and return.
2218 if (ReqFeatures.empty())
2221 // Now build up the set of caller features and verify that all the required
2222 // features are there.
2223 llvm::StringMap<bool> CallerFeatureMap;
2224 CGM.getFunctionFeatureMap(CallerFeatureMap, FD);
2226 // If we have at least one of the features in the feature list return
2227 // true, otherwise return false.
2229 ReqFeatures.begin(), ReqFeatures.end(), [&](StringRef Feature) {
2230 SmallVector<StringRef, 1> OrFeatures;
2231 Feature.split(OrFeatures, "|");
2232 return std::any_of(OrFeatures.begin(), OrFeatures.end(),
2233 [&](StringRef Feature) {
2234 if (!CallerFeatureMap.lookup(Feature)) {
2235 FirstMissing = Feature.str();
2243 // Emits an error if we don't have a valid set of target features for the
2245 void CodeGenFunction::checkTargetFeatures(const CallExpr *E,
2246 const FunctionDecl *TargetDecl) {
2247 // Early exit if this is an indirect call.
2251 // Get the current enclosing function if it exists. If it doesn't
2252 // we can't check the target features anyhow.
2253 const FunctionDecl *FD = dyn_cast_or_null<FunctionDecl>(CurFuncDecl);
2257 // Grab the required features for the call. For a builtin this is listed in
2258 // the td file with the default cpu, for an always_inline function this is any
2259 // listed cpu and any listed features.
2260 unsigned BuiltinID = TargetDecl->getBuiltinID();
2261 std::string MissingFeature;
2263 SmallVector<StringRef, 1> ReqFeatures;
2264 const char *FeatureList =
2265 CGM.getContext().BuiltinInfo.getRequiredFeatures(BuiltinID);
2266 // Return if the builtin doesn't have any required features.
2267 if (!FeatureList || StringRef(FeatureList) == "")
2269 StringRef(FeatureList).split(ReqFeatures, ",");
2270 if (!hasRequiredFeatures(ReqFeatures, CGM, FD, MissingFeature))
2271 CGM.getDiags().Report(E->getLocStart(), diag::err_builtin_needs_feature)
2272 << TargetDecl->getDeclName()
2273 << CGM.getContext().BuiltinInfo.getRequiredFeatures(BuiltinID);
2275 } else if (TargetDecl->hasAttr<TargetAttr>()) {
2276 // Get the required features for the callee.
2277 SmallVector<StringRef, 1> ReqFeatures;
2278 llvm::StringMap<bool> CalleeFeatureMap;
2279 CGM.getFunctionFeatureMap(CalleeFeatureMap, TargetDecl);
2280 for (const auto &F : CalleeFeatureMap) {
2281 // Only positive features are "required".
2283 ReqFeatures.push_back(F.getKey());
2285 if (!hasRequiredFeatures(ReqFeatures, CGM, FD, MissingFeature))
2286 CGM.getDiags().Report(E->getLocStart(), diag::err_function_needs_feature)
2287 << FD->getDeclName() << TargetDecl->getDeclName() << MissingFeature;
2291 void CodeGenFunction::EmitSanitizerStatReport(llvm::SanitizerStatKind SSK) {
2292 if (!CGM.getCodeGenOpts().SanitizeStats)
2295 llvm::IRBuilder<> IRB(Builder.GetInsertBlock(), Builder.GetInsertPoint());
2296 IRB.SetCurrentDebugLocation(Builder.getCurrentDebugLocation());
2297 CGM.getSanStats().create(IRB, SSK);
2300 llvm::DebugLoc CodeGenFunction::SourceLocToDebugLoc(SourceLocation Location) {
2301 if (CGDebugInfo *DI = getDebugInfo())
2302 return DI->SourceLocToDebugLoc(Location);
2304 return llvm::DebugLoc();